xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 2f923a0c) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30 #include "opt_ipsec.h"
31 #include "opt_ratelimit.h"
32 #include "opt_kern_tls.h"
33 #if defined(INET) || defined(INET6)
34 #include <sys/param.h>
35 #include <sys/arb.h>
36 #include <sys/module.h>
37 #include <sys/kernel.h>
38 #ifdef TCP_HHOOK
39 #include <sys/hhook.h>
40 #endif
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/mbuf.h>
46 #include <sys/proc.h>		/* for proc0 declaration */
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 #ifdef STATS
52 #include <sys/qmath.h>
53 #include <sys/tree.h>
54 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
55 #else
56 #include <sys/tree.h>
57 #endif
58 #include <sys/refcount.h>
59 #include <sys/queue.h>
60 #include <sys/tim_filter.h>
61 #include <sys/smp.h>
62 #include <sys/kthread.h>
63 #include <sys/kern_prefetch.h>
64 #include <sys/protosw.h>
65 #ifdef TCP_ACCOUNTING
66 #include <sys/sched.h>
67 #include <machine/cpu.h>
68 #endif
69 #include <vm/uma.h>
70 
71 #include <net/route.h>
72 #include <net/route/nhop.h>
73 #include <net/vnet.h>
74 
75 #define TCPSTATES		/* for logging */
76 
77 #include <netinet/in.h>
78 #include <netinet/in_kdtrace.h>
79 #include <netinet/in_pcb.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
82 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
83 #include <netinet/ip_var.h>
84 #include <netinet/ip6.h>
85 #include <netinet6/in6_pcb.h>
86 #include <netinet6/ip6_var.h>
87 #include <netinet/tcp.h>
88 #define	TCPOUTFLAGS
89 #include <netinet/tcp_fsm.h>
90 #include <netinet/tcp_seq.h>
91 #include <netinet/tcp_timer.h>
92 #include <netinet/tcp_var.h>
93 #include <netinet/tcp_log_buf.h>
94 #include <netinet/tcp_syncache.h>
95 #include <netinet/tcp_hpts.h>
96 #include <netinet/tcp_ratelimit.h>
97 #include <netinet/tcp_accounting.h>
98 #include <netinet/tcpip.h>
99 #include <netinet/cc/cc.h>
100 #include <netinet/cc/cc_newreno.h>
101 #include <netinet/tcp_fastopen.h>
102 #include <netinet/tcp_lro.h>
103 #ifdef NETFLIX_SHARED_CWND
104 #include <netinet/tcp_shared_cwnd.h>
105 #endif
106 #ifdef TCP_OFFLOAD
107 #include <netinet/tcp_offload.h>
108 #endif
109 #ifdef INET6
110 #include <netinet6/tcp6_var.h>
111 #endif
112 #include <netinet/tcp_ecn.h>
113 
114 #include <netipsec/ipsec_support.h>
115 
116 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
117 #include <netipsec/ipsec.h>
118 #include <netipsec/ipsec6.h>
119 #endif				/* IPSEC */
120 
121 #include <netinet/udp.h>
122 #include <netinet/udp_var.h>
123 #include <machine/in_cksum.h>
124 
125 #ifdef MAC
126 #include <security/mac/mac_framework.h>
127 #endif
128 #include "sack_filter.h"
129 #include "tcp_rack.h"
130 #include "tailq_hash.h"
131 #include "rack_bbr_common.h"
132 
133 uma_zone_t rack_zone;
134 uma_zone_t rack_pcb_zone;
135 
136 #ifndef TICKS2SBT
137 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
138 #endif
139 
140 VNET_DECLARE(uint32_t, newreno_beta);
141 VNET_DECLARE(uint32_t, newreno_beta_ecn);
142 #define V_newreno_beta VNET(newreno_beta)
143 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
144 
145 #define	M_TCPFSB	__CONCAT(M_TCPFSB, STACKNAME)
146 #define	M_TCPDO		__CONCAT(M_TCPDO, STACKNAME)
147 
148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb_" __XSTRING(STACKNAME), "TCP fast send block");
149 MALLOC_DEFINE(M_TCPDO, "tcp_do_" __XSTRING(STACKNAME), "TCP deferred options");
150 MALLOC_DEFINE(M_TCPPCM, "tcp_pcm_" __XSTRING(STACKNAME), "TCP PCM measurement information");
151 
152 struct sysctl_ctx_list rack_sysctl_ctx;
153 struct sysctl_oid *rack_sysctl_root;
154 
155 #define CUM_ACKED 1
156 #define SACKED 2
157 
158 /*
159  * The RACK module incorporates a number of
160  * TCP ideas that have been put out into the IETF
161  * over the last few years:
162  * - Matt Mathis's Rate Halving which slowly drops
163  *    the congestion window so that the ack clock can
164  *    be maintained during a recovery.
165  * - Yuchung Cheng's RACK TCP (for which its named) that
166  *    will stop us using the number of dup acks and instead
167  *    use time as the gage of when we retransmit.
168  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
169  *    of Dukkipati et.al.
170  * RACK depends on SACK, so if an endpoint arrives that
171  * cannot do SACK the state machine below will shuttle the
172  * connection back to using the "default" TCP stack that is
173  * in FreeBSD.
174  *
175  * To implement RACK the original TCP stack was first decomposed
176  * into a functional state machine with individual states
177  * for each of the possible TCP connection states. The do_segment
178  * functions role in life is to mandate the connection supports SACK
179  * initially and then assure that the RACK state matches the conenction
180  * state before calling the states do_segment function. Each
181  * state is simplified due to the fact that the original do_segment
182  * has been decomposed and we *know* what state we are in (no
183  * switches on the state) and all tests for SACK are gone. This
184  * greatly simplifies what each state does.
185  *
186  * TCP output is also over-written with a new version since it
187  * must maintain the new rack scoreboard.
188  *
189  */
190 static int32_t rack_tlp_thresh = 1;
191 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
192 static int32_t rack_tlp_use_greater = 1;
193 static int32_t rack_reorder_thresh = 2;
194 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
195 						 * - 60 seconds */
196 static uint16_t rack_policer_rxt_thresh= 0;	/* 499 = 49.9%, 0 is off  */
197 static uint8_t rack_policer_avg_thresh = 0; /* 3.2 */
198 static uint8_t rack_policer_med_thresh = 0; /* 1 - 16 */
199 static uint16_t rack_policer_bucket_reserve = 20; /* How much % is reserved in the bucket */
200 static uint64_t rack_pol_min_bw = 125000;	/* 1mbps in Bytes per sec */
201 static uint32_t rack_policer_data_thresh = 64000;	/* 64,000 bytes must be sent before we engage */
202 static uint32_t rack_policing_do_bw_comp = 1;
203 static uint32_t rack_pcm_every_n_rounds = 100;
204 static uint32_t rack_pcm_blast = 0;
205 static uint32_t rack_pcm_is_enabled = 1;
206 static uint8_t rack_req_del_mss = 18;	/* How many segments need to be sent in a recovery episode to do policer_detection */
207 static uint8_t rack_ssthresh_rest_rto_rec = 0; /* Do we restore ssthresh when we have rec -> rto -> rec */
208 
209 static uint32_t rack_gp_gain_req = 1200;		/* Amount percent wise required to gain to record a round has "gaining" */
210 static uint32_t rack_rnd_cnt_req = 0x10005;		/* Default number of rounds if we are below rack_gp_gain_req where we exit ss */
211 
212 
213 static int32_t rack_rxt_scoreboard_clear_thresh = 2;
214 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
215 static int32_t rack_rxt_controls = 0;
216 static int32_t rack_fill_cw_state = 0;
217 static uint8_t rack_req_measurements = 1;
218 /* Attack threshold detections */
219 static uint32_t rack_highest_sack_thresh_seen = 0;
220 static uint32_t rack_highest_move_thresh_seen = 0;
221 static uint32_t rack_merge_out_sacks_on_attack = 0;
222 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
223 static int32_t rack_hw_pace_extra_slots = 0;	/* 2 extra MSS time betweens */
224 static int32_t rack_hw_rate_caps = 0; /* 1; */
225 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
226 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
227 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
228 static int32_t rack_hw_up_only = 0;
229 static int32_t rack_stats_gets_ms_rtt = 1;
230 static int32_t rack_prr_addbackmax = 2;
231 static int32_t rack_do_hystart = 0;
232 static int32_t rack_apply_rtt_with_reduced_conf = 0;
233 static int32_t rack_hibeta_setting = 0;
234 static int32_t rack_default_pacing_divisor = 250;
235 static uint16_t rack_pacing_min_seg = 0;
236 static int32_t rack_timely_off = 0;
237 
238 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
239 static int32_t rack_pkt_delay = 1000;
240 static int32_t rack_send_a_lot_in_prr = 1;
241 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
242 static int32_t rack_verbose_logging = 0;
243 static int32_t rack_ignore_data_after_close = 1;
244 static int32_t rack_enable_shared_cwnd = 1;
245 static int32_t rack_use_cmp_acks = 1;
246 static int32_t rack_use_fsb = 1;
247 static int32_t rack_use_rfo = 1;
248 static int32_t rack_use_rsm_rfo = 1;
249 static int32_t rack_max_abc_post_recovery = 2;
250 static int32_t rack_client_low_buf = 0;
251 static int32_t rack_dsack_std_based = 0x3;	/* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
252 static int32_t rack_bw_multipler = 0;		/* Limit on fill cw's jump up to be this x gp_est */
253 #ifdef TCP_ACCOUNTING
254 static int32_t rack_tcp_accounting = 0;
255 #endif
256 static int32_t rack_limits_scwnd = 1;
257 static int32_t rack_enable_mqueue_for_nonpaced = 0;
258 static int32_t rack_hybrid_allow_set_maxseg = 0;
259 static int32_t rack_disable_prr = 0;
260 static int32_t use_rack_rr = 1;
261 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
262 static int32_t rack_persist_min = 250000;	/* 250usec */
263 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
264 static int32_t rack_honors_hpts_min_to =  1;	/* Do we honor the hpts minimum time out for pacing timers */
265 static uint32_t rack_max_reduce = 10;		/* Percent we can reduce slot by */
266 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
267 static int32_t rack_limit_time_with_srtt = 0;
268 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
269 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
270 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
271 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
272 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
273 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
274 static int32_t rack_full_buffer_discount = 10;
275 /*
276  * Currently regular tcp has a rto_min of 30ms
277  * the backoff goes 12 times so that ends up
278  * being a total of 122.850 seconds before a
279  * connection is killed.
280  */
281 static uint32_t rack_def_data_window = 20;
282 static uint32_t rack_goal_bdp = 2;
283 static uint32_t rack_min_srtts = 1;
284 static uint32_t rack_min_measure_usec = 0;
285 static int32_t rack_tlp_min = 10000;	/* 10ms */
286 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
287 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
288 static const int32_t rack_free_cache = 2;
289 static int32_t rack_hptsi_segments = 40;
290 static int32_t rack_rate_sample_method = USE_RTT_LOW;
291 static int32_t rack_pace_every_seg = 0;
292 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
293 static int32_t rack_slot_reduction = 4;
294 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
295 static int32_t rack_cwnd_block_ends_measure = 0;
296 static int32_t rack_rwnd_block_ends_measure = 0;
297 static int32_t rack_def_profile = 0;
298 
299 static int32_t rack_lower_cwnd_at_tlp = 0;
300 static int32_t rack_always_send_oldest = 0;
301 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
302 
303 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
304 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
305 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
306 
307 /* Probertt */
308 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
309 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
310 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
311 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
312 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
313 
314 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
315 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
316 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
317 static uint32_t rack_probertt_use_min_rtt_exit = 0;
318 static uint32_t rack_probe_rtt_sets_cwnd = 0;
319 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
320 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
321 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
322 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
323 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
324 static uint32_t rack_probertt_filter_life = 10000000;
325 static uint32_t rack_probertt_lower_within = 10;
326 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
327 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
328 static int32_t rack_probertt_clear_is = 1;
329 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
330 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
331 
332 /* Part of pacing */
333 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
334 
335 /* Timely information:
336  *
337  * Here we have various control parameters on how
338  * timely may change the multiplier. rack_gain_p5_ub
339  * is associated with timely but not directly influencing
340  * the rate decision like the other variables. It controls
341  * the way fill-cw interacts with timely and caps how much
342  * timely can boost the fill-cw b/w.
343  *
344  * The other values are various boost/shrink numbers as well
345  * as potential caps when adjustments are made to the timely
346  * gain (returned by rack_get_output_gain(). Remember too that
347  * the gain returned can be overriden by other factors such as
348  * probeRTT as well as fixed-rate-pacing.
349  */
350 static int32_t rack_gain_p5_ub = 250;
351 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
352 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
353 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
354 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
355 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
356 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
357 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
358 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
359 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
360 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
361 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
362 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
363 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
364 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
365 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
366 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
367 static int32_t rack_use_max_for_nobackoff = 0;
368 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
369 static int32_t rack_timely_no_stopping = 0;
370 static int32_t rack_down_raise_thresh = 100;
371 static int32_t rack_req_segs = 1;
372 static uint64_t rack_bw_rate_cap = 0;
373 static uint64_t rack_fillcw_bw_cap = 3750000;	/* Cap fillcw at 30Mbps */
374 
375 
376 /* Rack specific counters */
377 counter_u64_t rack_saw_enobuf;
378 counter_u64_t rack_saw_enobuf_hw;
379 counter_u64_t rack_saw_enetunreach;
380 counter_u64_t rack_persists_sends;
381 counter_u64_t rack_persists_acks;
382 counter_u64_t rack_persists_loss;
383 counter_u64_t rack_persists_lost_ends;
384 counter_u64_t rack_total_bytes;
385 #ifdef INVARIANTS
386 counter_u64_t rack_adjust_map_bw;
387 #endif
388 /* Tail loss probe counters */
389 counter_u64_t rack_tlp_tot;
390 counter_u64_t rack_tlp_newdata;
391 counter_u64_t rack_tlp_retran;
392 counter_u64_t rack_tlp_retran_bytes;
393 counter_u64_t rack_to_tot;
394 counter_u64_t rack_hot_alloc;
395 counter_u64_t tcp_policer_detected;
396 counter_u64_t rack_to_alloc;
397 counter_u64_t rack_to_alloc_hard;
398 counter_u64_t rack_to_alloc_emerg;
399 counter_u64_t rack_to_alloc_limited;
400 counter_u64_t rack_alloc_limited_conns;
401 counter_u64_t rack_split_limited;
402 counter_u64_t rack_rxt_clamps_cwnd;
403 counter_u64_t rack_rxt_clamps_cwnd_uniq;
404 
405 counter_u64_t rack_multi_single_eq;
406 counter_u64_t rack_proc_non_comp_ack;
407 
408 counter_u64_t rack_fto_send;
409 counter_u64_t rack_fto_rsm_send;
410 counter_u64_t rack_nfto_resend;
411 counter_u64_t rack_non_fto_send;
412 counter_u64_t rack_extended_rfo;
413 
414 counter_u64_t rack_sack_proc_all;
415 counter_u64_t rack_sack_proc_short;
416 counter_u64_t rack_sack_proc_restart;
417 counter_u64_t rack_sack_attacks_detected;
418 counter_u64_t rack_sack_attacks_reversed;
419 counter_u64_t rack_sack_attacks_suspect;
420 counter_u64_t rack_sack_used_next_merge;
421 counter_u64_t rack_sack_splits;
422 counter_u64_t rack_sack_used_prev_merge;
423 counter_u64_t rack_sack_skipped_acked;
424 counter_u64_t rack_ack_total;
425 counter_u64_t rack_express_sack;
426 counter_u64_t rack_sack_total;
427 counter_u64_t rack_move_none;
428 counter_u64_t rack_move_some;
429 
430 counter_u64_t rack_input_idle_reduces;
431 counter_u64_t rack_collapsed_win;
432 counter_u64_t rack_collapsed_win_seen;
433 counter_u64_t rack_collapsed_win_rxt;
434 counter_u64_t rack_collapsed_win_rxt_bytes;
435 counter_u64_t rack_try_scwnd;
436 counter_u64_t rack_hw_pace_init_fail;
437 counter_u64_t rack_hw_pace_lost;
438 
439 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
440 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
441 
442 
443 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
444 
445 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
446 	(tv) = (value) + slop;	 \
447 	if ((u_long)(tv) < (u_long)(tvmin)) \
448 		(tv) = (tvmin); \
449 	if ((u_long)(tv) > (u_long)(tvmax)) \
450 		(tv) = (tvmax); \
451 } while (0)
452 
453 static void
454 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
455 
456 static int
457 rack_process_ack(struct mbuf *m, struct tcphdr *th,
458     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
459     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val, int32_t orig_tlen);
460 static int
461 rack_process_data(struct mbuf *m, struct tcphdr *th,
462     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
463     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
464 static void
465 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
466    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
467 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
468 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
469     uint8_t limit_type);
470 static struct rack_sendmap *
471 rack_check_recovery_mode(struct tcpcb *tp,
472     uint32_t tsused);
473 static uint32_t
474 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack);
475 static void
476 rack_cong_signal(struct tcpcb *tp,
477 		 uint32_t type, uint32_t ack, int );
478 static void rack_counter_destroy(void);
479 static int
480 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
481 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
482 static void
483 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
484 static void
485 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
486     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
487 static void rack_dtor(void *mem, int32_t size, void *arg);
488 static void
489 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
490     uint32_t flex1, uint32_t flex2,
491     uint32_t flex3, uint32_t flex4,
492     uint32_t flex5, uint32_t flex6,
493     uint16_t flex7, uint8_t mod);
494 
495 static void
496 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
497    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
498    struct rack_sendmap *rsm, uint8_t quality);
499 static struct rack_sendmap *
500 rack_find_high_nonack(struct tcp_rack *rack,
501     struct rack_sendmap *rsm);
502 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
503 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
504 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
505 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
506 static void
507 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
508 			    tcp_seq th_ack, int line, uint8_t quality);
509 static void
510 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
511 
512 static uint32_t
513 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
514 static int32_t rack_handoff_ok(struct tcpcb *tp);
515 static int32_t rack_init(struct tcpcb *tp, void **ptr);
516 static void rack_init_sysctls(void);
517 
518 static void
519 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
520     struct tcphdr *th, int entered_rec, int dup_ack_struck,
521     int *dsack_seen, int *sacks_seen);
522 static void
523 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
524     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
525     struct rack_sendmap *hintrsm, uint32_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
526 
527 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
528 
529 
530 static void
531 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
532     struct rack_sendmap *rsm, uint32_t cts);
533 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
534 static int32_t rack_output(struct tcpcb *tp);
535 
536 static uint32_t
537 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
538     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
539     uint32_t cts, int *no_extra, int *moved_two, uint32_t segsiz);
540 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
541 static void rack_remxt_tmr(struct tcpcb *tp);
542 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
543 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
544 static int32_t rack_stopall(struct tcpcb *tp);
545 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
546 static uint32_t
547 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
548     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint32_t add_flag, int segsiz);
549 static void
550 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
551     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz);
552 static int
553 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
554     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
555 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
556 static int
557 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
558     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
559     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
560 
561 static void
562 rack_peg_rxt(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t segsiz);
563 
564 static int
565 rack_do_closing(struct mbuf *m, struct tcphdr *th,
566     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
567     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
568 static int
569 rack_do_established(struct mbuf *m, struct tcphdr *th,
570     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
571     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
572 static int
573 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
574     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
575     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
576 static int
577 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
578     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
579     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
580 static int
581 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
582     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
583     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
584 static int
585 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
586     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
587     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
588 static int
589 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
590     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
591     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
592 static int
593 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
594     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
595     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
596 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
597 struct rack_sendmap *
598 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
599     uint32_t tsused);
600 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
601     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
602 static void
603      tcp_rack_partialack(struct tcpcb *tp);
604 static int
605 rack_set_profile(struct tcp_rack *rack, int prof);
606 static void
607 rack_apply_deferred_options(struct tcp_rack *rack);
608 
609 int32_t rack_clear_counter=0;
610 
611 static uint64_t
612 rack_get_lt_bw(struct tcp_rack *rack)
613 {
614 	struct timeval tv;
615 	uint64_t tim, bytes;
616 
617 	tim = rack->r_ctl.lt_bw_time;
618 	bytes = rack->r_ctl.lt_bw_bytes;
619 	if (rack->lt_bw_up) {
620 		/* Include all the current bytes too */
621 		microuptime(&tv);
622 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
623 		tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
624 	}
625 	if ((bytes != 0) && (tim != 0))
626 		return ((bytes * (uint64_t)1000000) / tim);
627 	else
628 		return (0);
629 }
630 
631 static void
632 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
633 {
634 	struct sockopt sopt;
635 	struct cc_newreno_opts opt;
636 	struct newreno old;
637 	struct tcpcb *tp;
638 	int error, failed = 0;
639 
640 	tp = rack->rc_tp;
641 	if (tp->t_cc == NULL) {
642 		/* Tcb is leaving */
643 		return;
644 	}
645 	rack->rc_pacing_cc_set = 1;
646 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
647 		/* Not new-reno we can't play games with beta! */
648 		failed = 1;
649 		goto out;
650 
651 	}
652 	if (CC_ALGO(tp)->ctl_output == NULL)  {
653 		/* Huh, not using new-reno so no swaps.? */
654 		failed = 2;
655 		goto out;
656 	}
657 	/* Get the current values out */
658 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
659 	sopt.sopt_dir = SOPT_GET;
660 	opt.name = CC_NEWRENO_BETA;
661 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
662 	if (error)  {
663 		failed = 3;
664 		goto out;
665 	}
666 	old.beta = opt.val;
667 	opt.name = CC_NEWRENO_BETA_ECN;
668 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
669 	if (error)  {
670 		failed = 4;
671 		goto out;
672 	}
673 	old.beta_ecn = opt.val;
674 
675 	/* Now lets set in the values we have stored */
676 	sopt.sopt_dir = SOPT_SET;
677 	opt.name = CC_NEWRENO_BETA;
678 	opt.val = rack->r_ctl.rc_saved_beta.beta;
679 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
680 	if (error)  {
681 		failed = 5;
682 		goto out;
683 	}
684 	opt.name = CC_NEWRENO_BETA_ECN;
685 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
686 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
687 	if (error) {
688 		failed = 6;
689 		goto out;
690 	}
691 	/* Save off the values for restoral */
692 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
693 out:
694 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
695 		union tcp_log_stackspecific log;
696 		struct timeval tv;
697 		struct newreno *ptr;
698 
699 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
700 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
701 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
702 		log.u_bbr.flex1 = ptr->beta;
703 		log.u_bbr.flex2 = ptr->beta_ecn;
704 		log.u_bbr.flex3 = ptr->newreno_flags;
705 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
706 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
707 		log.u_bbr.flex6 = failed;
708 		log.u_bbr.flex7 = rack->gp_ready;
709 		log.u_bbr.flex7 <<= 1;
710 		log.u_bbr.flex7 |= rack->use_fixed_rate;
711 		log.u_bbr.flex7 <<= 1;
712 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
713 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
714 		log.u_bbr.flex8 = flex8;
715 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
716 			       0, &log, false, NULL, NULL, 0, &tv);
717 	}
718 }
719 
720 static void
721 rack_set_cc_pacing(struct tcp_rack *rack)
722 {
723 	if (rack->rc_pacing_cc_set)
724 		return;
725 	/*
726 	 * Use the swap utility placing in 3 for flex8 to id a
727 	 * set of a new set of values.
728 	 */
729 	rack->rc_pacing_cc_set = 1;
730 	rack_swap_beta_values(rack, 3);
731 }
732 
733 static void
734 rack_undo_cc_pacing(struct tcp_rack *rack)
735 {
736 	if (rack->rc_pacing_cc_set == 0)
737 		return;
738 	/*
739 	 * Use the swap utility placing in 4 for flex8 to id a
740 	 * restoral of the old values.
741 	 */
742 	rack->rc_pacing_cc_set = 0;
743 	rack_swap_beta_values(rack, 4);
744 }
745 
746 static void
747 rack_remove_pacing(struct tcp_rack *rack)
748 {
749 	if (rack->rc_pacing_cc_set)
750 		rack_undo_cc_pacing(rack);
751 	if (rack->r_ctl.pacing_method & RACK_REG_PACING)
752 		tcp_decrement_paced_conn();
753 	if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
754 		tcp_dec_dgp_pacing_cnt();
755 	rack->rc_always_pace = 0;
756 	rack->r_ctl.pacing_method = RACK_PACING_NONE;
757 	rack->dgp_on = 0;
758 	rack->rc_hybrid_mode = 0;
759 	rack->use_fixed_rate = 0;
760 }
761 
762 static void
763 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
764 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
765 {
766 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
767 		union tcp_log_stackspecific log;
768 		struct timeval tv;
769 
770 		memset(&log, 0, sizeof(log));
771 		log.u_bbr.flex1 = seq_end;
772 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
773 		log.u_bbr.flex3 = ack_end_t;
774 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
775 		log.u_bbr.flex5 = send_end_t;
776 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
777 		log.u_bbr.flex7 = mode;
778 		log.u_bbr.flex8 = 69;
779 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
780 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
781 		log.u_bbr.pkts_out = line;
782 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
783 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
784 		log.u_bbr.epoch = rack->r_ctl.current_round;
785 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
786 		if (rsm != NULL) {
787 			log.u_bbr.applimited = rsm->r_start;
788 			log.u_bbr.delivered = rsm->r_end;
789 			log.u_bbr.epoch = rsm->r_flags;
790 		}
791 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
792 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
793 		    &rack->rc_inp->inp_socket->so_rcv,
794 		    &rack->rc_inp->inp_socket->so_snd,
795 		    BBR_LOG_HPTSI_CALC, 0,
796 		    0, &log, false, &tv);
797 	}
798 }
799 
800 static int
801 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
802 {
803 	uint32_t stat;
804 	int32_t error;
805 
806 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
807 	if (error || req->newptr == NULL)
808 		return error;
809 
810 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
811 	if (error)
812 		return (error);
813 	if (stat == 1) {
814 #ifdef INVARIANTS
815 		printf("Clearing RACK counters\n");
816 #endif
817 		counter_u64_zero(rack_tlp_tot);
818 		counter_u64_zero(rack_tlp_newdata);
819 		counter_u64_zero(rack_tlp_retran);
820 		counter_u64_zero(rack_tlp_retran_bytes);
821 		counter_u64_zero(rack_to_tot);
822 		counter_u64_zero(rack_saw_enobuf);
823 		counter_u64_zero(rack_saw_enobuf_hw);
824 		counter_u64_zero(rack_saw_enetunreach);
825 		counter_u64_zero(rack_persists_sends);
826 		counter_u64_zero(rack_total_bytes);
827 		counter_u64_zero(rack_persists_acks);
828 		counter_u64_zero(rack_persists_loss);
829 		counter_u64_zero(rack_persists_lost_ends);
830 #ifdef INVARIANTS
831 		counter_u64_zero(rack_adjust_map_bw);
832 #endif
833 		counter_u64_zero(rack_to_alloc_hard);
834 		counter_u64_zero(rack_to_alloc_emerg);
835 		counter_u64_zero(rack_sack_proc_all);
836 		counter_u64_zero(rack_fto_send);
837 		counter_u64_zero(rack_fto_rsm_send);
838 		counter_u64_zero(rack_extended_rfo);
839 		counter_u64_zero(rack_hw_pace_init_fail);
840 		counter_u64_zero(rack_hw_pace_lost);
841 		counter_u64_zero(rack_non_fto_send);
842 		counter_u64_zero(rack_nfto_resend);
843 		counter_u64_zero(rack_sack_proc_short);
844 		counter_u64_zero(rack_sack_proc_restart);
845 		counter_u64_zero(rack_to_alloc);
846 		counter_u64_zero(rack_to_alloc_limited);
847 		counter_u64_zero(rack_alloc_limited_conns);
848 		counter_u64_zero(rack_split_limited);
849 		counter_u64_zero(rack_rxt_clamps_cwnd);
850 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
851 		counter_u64_zero(rack_multi_single_eq);
852 		counter_u64_zero(rack_proc_non_comp_ack);
853 		counter_u64_zero(rack_sack_attacks_detected);
854 		counter_u64_zero(rack_sack_attacks_reversed);
855 		counter_u64_zero(rack_sack_attacks_suspect);
856 		counter_u64_zero(rack_sack_used_next_merge);
857 		counter_u64_zero(rack_sack_used_prev_merge);
858 		counter_u64_zero(rack_sack_splits);
859 		counter_u64_zero(rack_sack_skipped_acked);
860 		counter_u64_zero(rack_ack_total);
861 		counter_u64_zero(rack_express_sack);
862 		counter_u64_zero(rack_sack_total);
863 		counter_u64_zero(rack_move_none);
864 		counter_u64_zero(rack_move_some);
865 		counter_u64_zero(rack_try_scwnd);
866 		counter_u64_zero(rack_collapsed_win);
867 		counter_u64_zero(rack_collapsed_win_rxt);
868 		counter_u64_zero(rack_collapsed_win_seen);
869 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
870 	} else if (stat == 2) {
871 #ifdef INVARIANTS
872 		printf("Clearing RACK option array\n");
873 #endif
874 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
875 	} else if (stat == 3) {
876 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
877 	} else if (stat == 4) {
878 #ifdef INVARIANTS
879 		printf("Clearing RACK out size array\n");
880 #endif
881 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
882 	}
883 	rack_clear_counter = 0;
884 	return (0);
885 }
886 
887 static void
888 rack_init_sysctls(void)
889 {
890 	struct sysctl_oid *rack_counters;
891 	struct sysctl_oid *rack_attack;
892 	struct sysctl_oid *rack_pacing;
893 	struct sysctl_oid *rack_timely;
894 	struct sysctl_oid *rack_timers;
895 	struct sysctl_oid *rack_tlp;
896 	struct sysctl_oid *rack_misc;
897 	struct sysctl_oid *rack_features;
898 	struct sysctl_oid *rack_measure;
899 	struct sysctl_oid *rack_probertt;
900 	struct sysctl_oid *rack_hw_pacing;
901 	struct sysctl_oid *rack_policing;
902 
903 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
904 	    SYSCTL_CHILDREN(rack_sysctl_root),
905 	    OID_AUTO,
906 	    "sack_attack",
907 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
908 	    "Rack Sack Attack Counters and Controls");
909 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_sysctl_root),
911 	    OID_AUTO,
912 	    "stats",
913 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
914 	    "Rack Counters");
915 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_sysctl_root),
917 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
918 	    &rack_rate_sample_method , USE_RTT_LOW,
919 	    "What method should we use for rate sampling 0=high, 1=low ");
920 	/* Probe rtt related controls */
921 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
922 	    SYSCTL_CHILDREN(rack_sysctl_root),
923 	    OID_AUTO,
924 	    "probertt",
925 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
926 	    "ProbeRTT related Controls");
927 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
928 	    SYSCTL_CHILDREN(rack_probertt),
929 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
930 	    &rack_atexit_prtt_hbp, 130,
931 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
932 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
933 	    SYSCTL_CHILDREN(rack_probertt),
934 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
935 	    &rack_atexit_prtt, 130,
936 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
937 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
938 	    SYSCTL_CHILDREN(rack_probertt),
939 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
940 	    &rack_per_of_gp_probertt, 60,
941 	    "What percentage of goodput do we pace at in probertt");
942 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
943 	    SYSCTL_CHILDREN(rack_probertt),
944 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
945 	    &rack_per_of_gp_probertt_reduce, 10,
946 	    "What percentage of goodput do we reduce every gp_srtt");
947 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
948 	    SYSCTL_CHILDREN(rack_probertt),
949 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
950 	    &rack_per_of_gp_lowthresh, 40,
951 	    "What percentage of goodput do we allow the multiplier to fall to");
952 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
953 	    SYSCTL_CHILDREN(rack_probertt),
954 	    OID_AUTO, "time_between", CTLFLAG_RW,
955 	    & rack_time_between_probertt, 96000000,
956 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
957 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
958 	    SYSCTL_CHILDREN(rack_probertt),
959 	    OID_AUTO, "safety", CTLFLAG_RW,
960 	    &rack_probe_rtt_safety_val, 2000000,
961 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
962 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
963 	    SYSCTL_CHILDREN(rack_probertt),
964 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
965 	    &rack_probe_rtt_sets_cwnd, 0,
966 	    "Do we set the cwnd too (if always_lower is on)");
967 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
968 	    SYSCTL_CHILDREN(rack_probertt),
969 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
970 	    &rack_max_drain_wait, 2,
971 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
972 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
973 	    SYSCTL_CHILDREN(rack_probertt),
974 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
975 	    &rack_must_drain, 1,
976 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
977 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
978 	    SYSCTL_CHILDREN(rack_probertt),
979 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
980 	    &rack_probertt_use_min_rtt_entry, 1,
981 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
982 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
983 	    SYSCTL_CHILDREN(rack_probertt),
984 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
985 	    &rack_probertt_use_min_rtt_exit, 0,
986 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
987 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_probertt),
989 	    OID_AUTO, "length_div", CTLFLAG_RW,
990 	    &rack_probertt_gpsrtt_cnt_div, 0,
991 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
992 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
993 	    SYSCTL_CHILDREN(rack_probertt),
994 	    OID_AUTO, "length_mul", CTLFLAG_RW,
995 	    &rack_probertt_gpsrtt_cnt_mul, 0,
996 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
997 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_probertt),
999 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
1000 	    &rack_min_probertt_hold, 200000,
1001 	    "What is the minimum time we hold probertt at target");
1002 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_probertt),
1004 	    OID_AUTO, "filter_life", CTLFLAG_RW,
1005 	    &rack_probertt_filter_life, 10000000,
1006 	    "What is the time for the filters life in useconds");
1007 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_probertt),
1009 	    OID_AUTO, "lower_within", CTLFLAG_RW,
1010 	    &rack_probertt_lower_within, 10,
1011 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
1012 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_probertt),
1014 	    OID_AUTO, "must_move", CTLFLAG_RW,
1015 	    &rack_min_rtt_movement, 250,
1016 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1017 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_probertt),
1019 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1020 	    &rack_probertt_clear_is, 1,
1021 	    "Do we clear I/S counts on exiting probe-rtt");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_probertt),
1024 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1025 	    &rack_max_drain_hbp, 1,
1026 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
1027 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_probertt),
1029 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1030 	    &rack_hbp_thresh, 3,
1031 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1032 	/* Pacing related sysctls */
1033 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1034 	    SYSCTL_CHILDREN(rack_sysctl_root),
1035 	    OID_AUTO,
1036 	    "pacing",
1037 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1038 	    "Pacing related Controls");
1039 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1040 	    SYSCTL_CHILDREN(rack_pacing),
1041 	    OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1042 	    &rack_pcm_is_enabled, 1,
1043 	    "Do we by default do PCM measurements?");
1044 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1045 	    SYSCTL_CHILDREN(rack_pacing),
1046 	    OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1047 	    &rack_pcm_every_n_rounds, 100,
1048 	    "How many rounds before we need to do a PCM measurement");
1049 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_pacing),
1051 	    OID_AUTO, "pcm_blast", CTLFLAG_RW,
1052 	    &rack_pcm_blast, 0,
1053 	    "Blast out the full cwnd/rwnd when doing a PCM measurement");
1054 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1055 	    SYSCTL_CHILDREN(rack_pacing),
1056 	    OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1057 	    &rack_gp_gain_req, 1200,
1058 	    "How much do we have to increase the GP to record the round 1200 = 120.0");
1059 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1060 	    SYSCTL_CHILDREN(rack_pacing),
1061 	    OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1062 	    &rack_rnd_cnt_req, 0x10005,
1063 	    "How many rounds less than rnd_gp_gain will drop us out of SS");
1064 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 	    SYSCTL_CHILDREN(rack_pacing),
1066 	    OID_AUTO, "no_timely", CTLFLAG_RW,
1067 	    &rack_timely_off, 0,
1068 	    "Do we not use timely in DGP?");
1069 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 	    SYSCTL_CHILDREN(rack_pacing),
1071 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1072 	    &rack_full_buffer_discount, 10,
1073 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1074 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 	    SYSCTL_CHILDREN(rack_pacing),
1076 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1077 	    &rack_fill_cw_state, 0,
1078 	    "Enable fillcw on new connections (default=0 off)?");
1079 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1080 	    SYSCTL_CHILDREN(rack_pacing),
1081 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1082 	    &rack_pacing_min_seg, 0,
1083 	    "What is the min burst size for pacing (0 disables)?");
1084 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085 	    SYSCTL_CHILDREN(rack_pacing),
1086 	    OID_AUTO, "divisor", CTLFLAG_RW,
1087 	    &rack_default_pacing_divisor, 250,
1088 	    "What is the default divisor given to the rl code?");
1089 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1090 	    SYSCTL_CHILDREN(rack_pacing),
1091 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1092 	    &rack_bw_multipler, 0,
1093 	    "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1094 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_pacing),
1096 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1097 	    &rack_max_per_above, 30,
1098 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1099 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100 	    SYSCTL_CHILDREN(rack_pacing),
1101 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1102 	    &rack_pace_one_seg, 0,
1103 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1104 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 	    SYSCTL_CHILDREN(rack_pacing),
1106 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1107 	    &rack_limit_time_with_srtt, 0,
1108 	    "Do we limit pacing time based on srtt");
1109 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1110 	    SYSCTL_CHILDREN(rack_pacing),
1111 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1112 	    &rack_per_of_gp_ss, 250,
1113 	    "If non zero, what percentage of goodput to pace at in slow start");
1114 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1115 	    SYSCTL_CHILDREN(rack_pacing),
1116 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1117 	    &rack_per_of_gp_ca, 150,
1118 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1119 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1120 	    SYSCTL_CHILDREN(rack_pacing),
1121 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1122 	    &rack_per_of_gp_rec, 200,
1123 	    "If non zero, what percentage of goodput to pace at in recovery");
1124 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125 	    SYSCTL_CHILDREN(rack_pacing),
1126 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1127 	    &rack_hptsi_segments, 40,
1128 	    "What size is the max for TSO segments in pacing and burst mitigation");
1129 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130 	    SYSCTL_CHILDREN(rack_pacing),
1131 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1132 	    &rack_slot_reduction, 4,
1133 	    "When doing only burst mitigation what is the reduce divisor");
1134 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135 	    SYSCTL_CHILDREN(rack_sysctl_root),
1136 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1137 	    &rack_pace_every_seg, 0,
1138 	    "If set we use pacing, if clear we use only the original burst mitigation");
1139 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_pacing),
1141 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1142 	    &rack_bw_rate_cap, 0,
1143 	    "If set we apply this value to the absolute rate cap used by pacing");
1144 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_pacing),
1146 	    OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1147 	    &rack_fillcw_bw_cap, 3750000,
1148 	    "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1149 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_sysctl_root),
1151 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1152 	    &rack_req_measurements, 1,
1153 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1154 	/* Hardware pacing */
1155 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1156 	    SYSCTL_CHILDREN(rack_sysctl_root),
1157 	    OID_AUTO,
1158 	    "hdwr_pacing",
1159 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1160 	    "Pacing related Controls");
1161 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162 	    SYSCTL_CHILDREN(rack_hw_pacing),
1163 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1164 	    &rack_hw_rwnd_factor, 2,
1165 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1166 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167 	    SYSCTL_CHILDREN(rack_hw_pacing),
1168 	    OID_AUTO, "precheck", CTLFLAG_RW,
1169 	    &rack_hw_check_queue, 0,
1170 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1171 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(rack_hw_pacing),
1173 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1174 	    &rack_enobuf_hw_boost_mult, 0,
1175 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1176 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177 	    SYSCTL_CHILDREN(rack_hw_pacing),
1178 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1179 	    &rack_enobuf_hw_max, 2,
1180 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1181 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_hw_pacing),
1183 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1184 	    &rack_enobuf_hw_min, 2,
1185 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1186 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187 	    SYSCTL_CHILDREN(rack_hw_pacing),
1188 	    OID_AUTO, "enable", CTLFLAG_RW,
1189 	    &rack_enable_hw_pacing, 0,
1190 	    "Should RACK attempt to use hw pacing?");
1191 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192 	    SYSCTL_CHILDREN(rack_hw_pacing),
1193 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1194 	    &rack_hw_rate_caps, 0,
1195 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1196 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1197 	    SYSCTL_CHILDREN(rack_hw_pacing),
1198 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1199 	    &rack_hw_rate_cap_per, 0,
1200 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1201 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(rack_hw_pacing),
1203 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1204 	    &rack_hw_rate_min, 0,
1205 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1206 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1207 	    SYSCTL_CHILDREN(rack_hw_pacing),
1208 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1209 	    &rack_hw_rate_to_low, 0,
1210 	    "If we fall below this rate, dis-engage hw pacing?");
1211 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1212 	    SYSCTL_CHILDREN(rack_hw_pacing),
1213 	    OID_AUTO, "up_only", CTLFLAG_RW,
1214 	    &rack_hw_up_only, 0,
1215 	    "Do we allow hw pacing to lower the rate selected?");
1216 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(rack_hw_pacing),
1218 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1219 	    &rack_hw_pace_extra_slots, 0,
1220 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1221 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1222 	    SYSCTL_CHILDREN(rack_sysctl_root),
1223 	    OID_AUTO,
1224 	    "timely",
1225 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1226 	    "Rack Timely RTT Controls");
1227 	/* Timely based GP dynmics */
1228 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_timely),
1230 	    OID_AUTO, "upper", CTLFLAG_RW,
1231 	    &rack_gp_per_bw_mul_up, 2,
1232 	    "Rack timely upper range for equal b/w (in percentage)");
1233 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234 	    SYSCTL_CHILDREN(rack_timely),
1235 	    OID_AUTO, "lower", CTLFLAG_RW,
1236 	    &rack_gp_per_bw_mul_down, 4,
1237 	    "Rack timely lower range for equal b/w (in percentage)");
1238 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239 	    SYSCTL_CHILDREN(rack_timely),
1240 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1241 	    &rack_gp_rtt_maxmul, 3,
1242 	    "Rack timely multiplier of lowest rtt for rtt_max");
1243 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244 	    SYSCTL_CHILDREN(rack_timely),
1245 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1246 	    &rack_gp_rtt_mindiv, 4,
1247 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1248 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249 	    SYSCTL_CHILDREN(rack_timely),
1250 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1251 	    &rack_gp_rtt_minmul, 1,
1252 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1253 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254 	    SYSCTL_CHILDREN(rack_timely),
1255 	    OID_AUTO, "decrease", CTLFLAG_RW,
1256 	    &rack_gp_decrease_per, 80,
1257 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1258 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259 	    SYSCTL_CHILDREN(rack_timely),
1260 	    OID_AUTO, "increase", CTLFLAG_RW,
1261 	    &rack_gp_increase_per, 2,
1262 	    "Rack timely increase perentage of our GP multiplication factor");
1263 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264 	    SYSCTL_CHILDREN(rack_timely),
1265 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1266 	    &rack_per_lower_bound, 50,
1267 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1268 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269 	    SYSCTL_CHILDREN(rack_timely),
1270 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1271 	    &rack_gain_p5_ub, 250,
1272 	    "Profile 5 upper bound to timely gain");
1273 
1274 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1275 	    SYSCTL_CHILDREN(rack_timely),
1276 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1277 	    &rack_per_upper_bound_ss, 0,
1278 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1279 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1280 	    SYSCTL_CHILDREN(rack_timely),
1281 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1282 	    &rack_per_upper_bound_ca, 0,
1283 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1284 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1285 	    SYSCTL_CHILDREN(rack_timely),
1286 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1287 	    &rack_do_dyn_mul, 0,
1288 	    "Rack timely do we enable dynmaic timely goodput by default");
1289 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1290 	    SYSCTL_CHILDREN(rack_timely),
1291 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1292 	    &rack_gp_no_rec_chg, 1,
1293 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1294 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1295 	    SYSCTL_CHILDREN(rack_timely),
1296 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1297 	    &rack_timely_dec_clear, 6,
1298 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1299 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1300 	    SYSCTL_CHILDREN(rack_timely),
1301 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1302 	    &rack_timely_max_push_rise, 3,
1303 	    "Rack timely how many times do we push up with b/w increase");
1304 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1305 	    SYSCTL_CHILDREN(rack_timely),
1306 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1307 	    &rack_timely_max_push_drop, 3,
1308 	    "Rack timely how many times do we push back on b/w decent");
1309 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1310 	    SYSCTL_CHILDREN(rack_timely),
1311 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1312 	    &rack_timely_min_segs, 4,
1313 	    "Rack timely when setting the cwnd what is the min num segments");
1314 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1315 	    SYSCTL_CHILDREN(rack_timely),
1316 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1317 	    &rack_use_max_for_nobackoff, 0,
1318 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1319 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1320 	    SYSCTL_CHILDREN(rack_timely),
1321 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1322 	    &rack_timely_int_timely_only, 0,
1323 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1324 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(rack_timely),
1326 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1327 	    &rack_timely_no_stopping, 0,
1328 	    "Rack timely don't stop increase");
1329 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(rack_timely),
1331 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1332 	    &rack_down_raise_thresh, 100,
1333 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1334 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335 	    SYSCTL_CHILDREN(rack_timely),
1336 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1337 	    &rack_req_segs, 1,
1338 	    "Bottom dragging if not these many segments outstanding and room");
1339 
1340 	/* TLP and Rack related parameters */
1341 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1342 	    SYSCTL_CHILDREN(rack_sysctl_root),
1343 	    OID_AUTO,
1344 	    "tlp",
1345 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1346 	    "TLP and Rack related Controls");
1347 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1348 	    SYSCTL_CHILDREN(rack_tlp),
1349 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1350 	    &use_rack_rr, 1,
1351 	    "Do we use Rack Rapid Recovery");
1352 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1353 	    SYSCTL_CHILDREN(rack_tlp),
1354 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1355 	    &rack_max_abc_post_recovery, 2,
1356 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1357 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1358 	    SYSCTL_CHILDREN(rack_tlp),
1359 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1360 	    &rack_non_rxt_use_cr, 0,
1361 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1362 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1363 	    SYSCTL_CHILDREN(rack_tlp),
1364 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1365 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1366 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1367 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1368 	    SYSCTL_CHILDREN(rack_tlp),
1369 	    OID_AUTO, "limit", CTLFLAG_RW,
1370 	    &rack_tlp_limit, 2,
1371 	    "How many TLP's can be sent without sending new data");
1372 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373 	    SYSCTL_CHILDREN(rack_tlp),
1374 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1375 	    &rack_tlp_use_greater, 1,
1376 	    "Should we use the rack_rtt time if its greater than srtt");
1377 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378 	    SYSCTL_CHILDREN(rack_tlp),
1379 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1380 	    &rack_tlp_min, 10000,
1381 	    "TLP minimum timeout per the specification (in microseconds)");
1382 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383 	    SYSCTL_CHILDREN(rack_tlp),
1384 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1385 	    &rack_always_send_oldest, 0,
1386 	    "Should we always send the oldest TLP and RACK-TLP");
1387 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388 	    SYSCTL_CHILDREN(rack_tlp),
1389 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1390 	    &rack_lower_cwnd_at_tlp, 0,
1391 	    "When a TLP completes a retran should we enter recovery");
1392 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393 	    SYSCTL_CHILDREN(rack_tlp),
1394 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1395 	    &rack_reorder_thresh, 2,
1396 	    "What factor for rack will be added when seeing reordering (shift right)");
1397 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398 	    SYSCTL_CHILDREN(rack_tlp),
1399 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1400 	    &rack_tlp_thresh, 1,
1401 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1402 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_tlp),
1404 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1405 	    &rack_reorder_fade, 60000000,
1406 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1407 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1408 	    SYSCTL_CHILDREN(rack_tlp),
1409 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1410 	    &rack_pkt_delay, 1000,
1411 	    "Extra RACK time (in microseconds) besides reordering thresh");
1412 
1413 	/* Timer related controls */
1414 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1415 	    SYSCTL_CHILDREN(rack_sysctl_root),
1416 	    OID_AUTO,
1417 	    "timers",
1418 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1419 	    "Timer related controls");
1420 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_timers),
1422 	    OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1423 	    &rack_ssthresh_rest_rto_rec, 0,
1424 	    "When doing recovery -> rto -> recovery do we reset SSthresh?");
1425 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1426 	    SYSCTL_CHILDREN(rack_timers),
1427 	    OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1428 	    &rack_rxt_scoreboard_clear_thresh, 2,
1429 	    "How many RTO's are allowed before we clear the scoreboard");
1430 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1431 	    SYSCTL_CHILDREN(rack_timers),
1432 	    OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1433 	    &rack_honors_hpts_min_to, 1,
1434 	    "Do rack pacing timers honor hpts min timeout");
1435 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1436 	    SYSCTL_CHILDREN(rack_timers),
1437 	    OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1438 	    &rack_max_reduce, 10,
1439 	    "Max percentage we will reduce slot by for pacing when we are behind");
1440 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_timers),
1442 	    OID_AUTO, "persmin", CTLFLAG_RW,
1443 	    &rack_persist_min, 250000,
1444 	    "What is the minimum time in microseconds between persists");
1445 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1446 	    SYSCTL_CHILDREN(rack_timers),
1447 	    OID_AUTO, "persmax", CTLFLAG_RW,
1448 	    &rack_persist_max, 2000000,
1449 	    "What is the largest delay in microseconds between persists");
1450 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1451 	    SYSCTL_CHILDREN(rack_timers),
1452 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1453 	    &rack_delayed_ack_time, 40000,
1454 	    "Delayed ack time (40ms in microseconds)");
1455 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1456 	    SYSCTL_CHILDREN(rack_timers),
1457 	    OID_AUTO, "minrto", CTLFLAG_RW,
1458 	    &rack_rto_min, 30000,
1459 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1460 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1461 	    SYSCTL_CHILDREN(rack_timers),
1462 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1463 	    &rack_rto_max, 4000000,
1464 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1465 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1466 	    SYSCTL_CHILDREN(rack_timers),
1467 	    OID_AUTO, "minto", CTLFLAG_RW,
1468 	    &rack_min_to, 1000,
1469 	    "Minimum rack timeout in microseconds");
1470 	/* Measure controls */
1471 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1472 	    SYSCTL_CHILDREN(rack_sysctl_root),
1473 	    OID_AUTO,
1474 	    "measure",
1475 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1476 	    "Measure related controls");
1477 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1478 	    SYSCTL_CHILDREN(rack_measure),
1479 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1480 	    &rack_wma_divisor, 8,
1481 	    "When doing b/w calculation what is the  divisor for the WMA");
1482 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1483 	    SYSCTL_CHILDREN(rack_measure),
1484 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1485 	    &rack_cwnd_block_ends_measure, 0,
1486 	    "Does a cwnd just-return end the measurement window (app limited)");
1487 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1488 	    SYSCTL_CHILDREN(rack_measure),
1489 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1490 	    &rack_rwnd_block_ends_measure, 0,
1491 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1492 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1493 	    SYSCTL_CHILDREN(rack_measure),
1494 	    OID_AUTO, "min_target", CTLFLAG_RW,
1495 	    &rack_def_data_window, 20,
1496 	    "What is the minimum target window (in mss) for a GP measurements");
1497 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1498 	    SYSCTL_CHILDREN(rack_measure),
1499 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1500 	    &rack_goal_bdp, 2,
1501 	    "What is the goal BDP to measure");
1502 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1503 	    SYSCTL_CHILDREN(rack_measure),
1504 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1505 	    &rack_min_srtts, 1,
1506 	    "What is the goal BDP to measure");
1507 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(rack_measure),
1509 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1510 	    &rack_min_measure_usec, 0,
1511 	    "What is the Minimum time time for a measurement if 0, this is off");
1512 	/* Features */
1513 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1514 	    SYSCTL_CHILDREN(rack_sysctl_root),
1515 	    OID_AUTO,
1516 	    "features",
1517 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1518 	    "Feature controls");
1519 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1520 	    SYSCTL_CHILDREN(rack_features),
1521 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1522 	    &rack_hybrid_allow_set_maxseg, 0,
1523 	    "Should hybrid pacing allow the setmss command");
1524 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1525 	    SYSCTL_CHILDREN(rack_features),
1526 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1527 	    &rack_use_cmp_acks, 1,
1528 	    "Should RACK have LRO send compressed acks");
1529 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1530 	    SYSCTL_CHILDREN(rack_features),
1531 	    OID_AUTO, "fsb", CTLFLAG_RW,
1532 	    &rack_use_fsb, 1,
1533 	    "Should RACK use the fast send block?");
1534 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1535 	    SYSCTL_CHILDREN(rack_features),
1536 	    OID_AUTO, "rfo", CTLFLAG_RW,
1537 	    &rack_use_rfo, 1,
1538 	    "Should RACK use rack_fast_output()?");
1539 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1540 	    SYSCTL_CHILDREN(rack_features),
1541 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1542 	    &rack_use_rsm_rfo, 1,
1543 	    "Should RACK use rack_fast_rsm_output()?");
1544 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1545 	    SYSCTL_CHILDREN(rack_features),
1546 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1547 	    &rack_enable_mqueue_for_nonpaced, 0,
1548 	    "Should RACK use mbuf queuing for non-paced connections");
1549 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1550 	    SYSCTL_CHILDREN(rack_features),
1551 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1552 	    &rack_do_hystart, 0,
1553 	    "Should RACK enable HyStart++ on connections?");
1554 	/* Policer detection */
1555 	rack_policing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1556 	    SYSCTL_CHILDREN(rack_sysctl_root),
1557 	    OID_AUTO,
1558 	    "policing",
1559 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1560 	    "policer detection");
1561 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_policing),
1563 	    OID_AUTO, "rxt_thresh", CTLFLAG_RW,
1564 	    &rack_policer_rxt_thresh, 0,
1565 	   "Percentage of retransmits we need to be a possible policer (499 = 49.9 percent)");
1566 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_policing),
1568 	    OID_AUTO, "avg_thresh", CTLFLAG_RW,
1569 	    &rack_policer_avg_thresh, 0,
1570 	    "What threshold of average retransmits needed to recover a lost packet (1 - 169 aka 21 = 2.1)?");
1571 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_policing),
1573 	    OID_AUTO, "med_thresh", CTLFLAG_RW,
1574 	    &rack_policer_med_thresh, 0,
1575 	    "What threshold of Median retransmits needed to recover a lost packet (1 - 16)?");
1576 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_policing),
1578 	    OID_AUTO, "data_thresh", CTLFLAG_RW,
1579 	    &rack_policer_data_thresh, 64000,
1580 	    "How many bytes must have gotten through before we can start doing policer detection?");
1581 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_policing),
1583 	    OID_AUTO, "bwcomp", CTLFLAG_RW,
1584 	    &rack_policing_do_bw_comp, 1,
1585 	    "Do we raise up low b/w so that at least pace_max_seg can be sent in the srtt?");
1586 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_policing),
1588 	    OID_AUTO, "recmss", CTLFLAG_RW,
1589 	    &rack_req_del_mss, 18,
1590 	    "How many MSS must be delivered during recovery to engage policer detection?");
1591 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_policing),
1593 	    OID_AUTO, "res_div", CTLFLAG_RW,
1594 	    &rack_policer_bucket_reserve, 20,
1595 	    "What percentage is reserved in the policer bucket?");
1596 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_policing),
1598 	    OID_AUTO, "min_comp_bw", CTLFLAG_RW,
1599 	    &rack_pol_min_bw, 125000,
1600 	    "Do we have a min b/w for b/w compensation (0 = no)?");
1601 	/* Misc rack controls */
1602 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1603 	    SYSCTL_CHILDREN(rack_sysctl_root),
1604 	    OID_AUTO,
1605 	    "misc",
1606 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1607 	    "Misc related controls");
1608 #ifdef TCP_ACCOUNTING
1609 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1610 	    SYSCTL_CHILDREN(rack_misc),
1611 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1612 	    &rack_tcp_accounting, 0,
1613 	    "Should we turn on TCP accounting for all rack sessions?");
1614 #endif
1615 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1616 	    SYSCTL_CHILDREN(rack_misc),
1617 	    OID_AUTO, "dnd", CTLFLAG_RW,
1618 	    &rack_dnd_default, 0,
1619 	    "Do not disturb default for rack_rrr = 3");
1620 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1621 	    SYSCTL_CHILDREN(rack_misc),
1622 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1623 	    &sad_seg_size_per, 800,
1624 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1625 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1626 	    SYSCTL_CHILDREN(rack_misc),
1627 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1628 	    &rack_rxt_controls, 0,
1629 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1630 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_misc),
1632 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1633 	    &rack_hibeta_setting, 0,
1634 	    "Do we ue a high beta (80 instead of 50)?");
1635 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(rack_misc),
1637 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1638 	    &rack_apply_rtt_with_reduced_conf, 0,
1639 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1640 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1641 	    SYSCTL_CHILDREN(rack_misc),
1642 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1643 	    &rack_dsack_std_based, 3,
1644 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1645 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1646 	    SYSCTL_CHILDREN(rack_misc),
1647 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1648 	    &rack_prr_addbackmax, 2,
1649 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1650 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1651 	    SYSCTL_CHILDREN(rack_misc),
1652 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1653 	    &rack_stats_gets_ms_rtt, 1,
1654 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1655 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1656 	    SYSCTL_CHILDREN(rack_misc),
1657 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1658 	    &rack_client_low_buf, 0,
1659 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1660 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(rack_misc),
1662 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1663 	    &rack_def_profile, 0,
1664 	    "Should RACK use a default profile (0=no, num == profile num)?");
1665 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_misc),
1667 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1668 	    &rack_enable_shared_cwnd, 1,
1669 	    "Should RACK try to use the shared cwnd on connections where allowed");
1670 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1671 	    SYSCTL_CHILDREN(rack_misc),
1672 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1673 	    &rack_limits_scwnd, 1,
1674 	    "Should RACK place low end time limits on the shared cwnd feature");
1675 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1676 	    SYSCTL_CHILDREN(rack_misc),
1677 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1678 	    &rack_disable_prr, 0,
1679 	    "Should RACK not use prr and only pace (must have pacing on)");
1680 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1681 	    SYSCTL_CHILDREN(rack_misc),
1682 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1683 	    &rack_verbose_logging, 0,
1684 	    "Should RACK black box logging be verbose");
1685 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1686 	    SYSCTL_CHILDREN(rack_misc),
1687 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1688 	    &rack_ignore_data_after_close, 1,
1689 	    "Do we hold off sending a RST until all pending data is ack'd");
1690 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1691 	    SYSCTL_CHILDREN(rack_misc),
1692 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1693 	    &rack_sack_not_required, 1,
1694 	    "Do we allow rack to run on connections not supporting SACK");
1695 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1696 	    SYSCTL_CHILDREN(rack_misc),
1697 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1698 	    &rack_send_a_lot_in_prr, 1,
1699 	    "Send a lot in prr");
1700 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1701 	    SYSCTL_CHILDREN(rack_misc),
1702 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1703 	    &rack_autosndbuf_inc, 20,
1704 	    "What percentage should rack scale up its snd buffer by?");
1705 
1706 
1707 	/* Sack Attacker detection stuff */
1708 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1709 	    SYSCTL_CHILDREN(rack_attack),
1710 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1711 	    &rack_merge_out_sacks_on_attack, 0,
1712 	    "Do we merge the sendmap when we decide we are being attacked?");
1713 
1714 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1715 	    SYSCTL_CHILDREN(rack_attack),
1716 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1717 	    &rack_highest_sack_thresh_seen, 0,
1718 	    "Highest sack to ack ratio seen");
1719 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1720 	    SYSCTL_CHILDREN(rack_attack),
1721 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1722 	    &rack_highest_move_thresh_seen, 0,
1723 	    "Highest move to non-move ratio seen");
1724 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1725 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1726 	    SYSCTL_CHILDREN(rack_attack),
1727 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1728 	    &rack_ack_total,
1729 	    "Total number of Ack's");
1730 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1731 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732 	    SYSCTL_CHILDREN(rack_attack),
1733 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1734 	    &rack_express_sack,
1735 	    "Total expresss number of Sack's");
1736 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1737 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 	    SYSCTL_CHILDREN(rack_attack),
1739 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1740 	    &rack_sack_total,
1741 	    "Total number of SACKs");
1742 	rack_move_none = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_attack),
1745 	    OID_AUTO, "move_none", CTLFLAG_RD,
1746 	    &rack_move_none,
1747 	    "Total number of SACK index reuse of positions under threshold");
1748 	rack_move_some = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_attack),
1751 	    OID_AUTO, "move_some", CTLFLAG_RD,
1752 	    &rack_move_some,
1753 	    "Total number of SACK index reuse of positions over threshold");
1754 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1755 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 	    SYSCTL_CHILDREN(rack_attack),
1757 	    OID_AUTO, "attacks", CTLFLAG_RD,
1758 	    &rack_sack_attacks_detected,
1759 	    "Total number of SACK attackers that had sack disabled");
1760 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1761 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762 	    SYSCTL_CHILDREN(rack_attack),
1763 	    OID_AUTO, "reversed", CTLFLAG_RD,
1764 	    &rack_sack_attacks_reversed,
1765 	    "Total number of SACK attackers that were later determined false positive");
1766 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1767 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1768 	    SYSCTL_CHILDREN(rack_attack),
1769 	    OID_AUTO, "suspect", CTLFLAG_RD,
1770 	    &rack_sack_attacks_suspect,
1771 	    "Total number of SACKs that triggered early detection");
1772 
1773 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1774 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1775 	    SYSCTL_CHILDREN(rack_attack),
1776 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1777 	    &rack_sack_used_next_merge,
1778 	    "Total number of times we used the next merge");
1779 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1780 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1781 	    SYSCTL_CHILDREN(rack_attack),
1782 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1783 	    &rack_sack_used_prev_merge,
1784 	    "Total number of times we used the prev merge");
1785 	/* Counters */
1786 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1787 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1788 	    SYSCTL_CHILDREN(rack_counters),
1789 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1790 	    &rack_total_bytes,
1791 	    "Total number of bytes sent");
1792 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1793 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1794 	    SYSCTL_CHILDREN(rack_counters),
1795 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1796 	    &rack_fto_send, "Total number of rack_fast_output sends");
1797 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1798 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1799 	    SYSCTL_CHILDREN(rack_counters),
1800 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1801 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1802 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1803 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804 	    SYSCTL_CHILDREN(rack_counters),
1805 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1806 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1807 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1808 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1809 	    SYSCTL_CHILDREN(rack_counters),
1810 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1811 	    &rack_non_fto_send, "Total number of rack_output first sends");
1812 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1813 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814 	    SYSCTL_CHILDREN(rack_counters),
1815 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1816 	    &rack_extended_rfo, "Total number of times we extended rfo");
1817 
1818 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1819 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820 	    SYSCTL_CHILDREN(rack_counters),
1821 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1822 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1823 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1824 
1825 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826 	    SYSCTL_CHILDREN(rack_counters),
1827 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1828 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1829 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1831 	    SYSCTL_CHILDREN(rack_counters),
1832 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1833 	    &rack_tlp_tot,
1834 	    "Total number of tail loss probe expirations");
1835 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1836 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1837 	    SYSCTL_CHILDREN(rack_counters),
1838 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1839 	    &rack_tlp_newdata,
1840 	    "Total number of tail loss probe sending new data");
1841 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1842 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1843 	    SYSCTL_CHILDREN(rack_counters),
1844 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1845 	    &rack_tlp_retran,
1846 	    "Total number of tail loss probe sending retransmitted data");
1847 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1848 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1849 	    SYSCTL_CHILDREN(rack_counters),
1850 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1851 	    &rack_tlp_retran_bytes,
1852 	    "Total bytes of tail loss probe sending retransmitted data");
1853 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1854 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1855 	    SYSCTL_CHILDREN(rack_counters),
1856 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1857 	    &rack_to_tot,
1858 	    "Total number of times the rack to expired");
1859 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1860 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1861 	    SYSCTL_CHILDREN(rack_counters),
1862 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1863 	    &rack_saw_enobuf,
1864 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1865 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1866 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1867 	    SYSCTL_CHILDREN(rack_counters),
1868 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1869 	    &rack_saw_enobuf_hw,
1870 	    "Total number of times a send returned enobuf for hdwr paced connections");
1871 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1872 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1873 	    SYSCTL_CHILDREN(rack_counters),
1874 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1875 	    &rack_saw_enetunreach,
1876 	    "Total number of times a send received a enetunreachable");
1877 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1878 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1879 	    SYSCTL_CHILDREN(rack_counters),
1880 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1881 	    &rack_hot_alloc,
1882 	    "Total allocations from the top of our list");
1883 	tcp_policer_detected = counter_u64_alloc(M_WAITOK);
1884 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1885 	    SYSCTL_CHILDREN(rack_counters),
1886 	    OID_AUTO, "policer_detected", CTLFLAG_RD,
1887 	    &tcp_policer_detected,
1888 	    "Total policer_detections");
1889 
1890 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1891 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1892 	    SYSCTL_CHILDREN(rack_counters),
1893 	    OID_AUTO, "allocs", CTLFLAG_RD,
1894 	    &rack_to_alloc,
1895 	    "Total allocations of tracking structures");
1896 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1897 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1898 	    SYSCTL_CHILDREN(rack_counters),
1899 	    OID_AUTO, "allochard", CTLFLAG_RD,
1900 	    &rack_to_alloc_hard,
1901 	    "Total allocations done with sleeping the hard way");
1902 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1903 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1904 	    SYSCTL_CHILDREN(rack_counters),
1905 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1906 	    &rack_to_alloc_emerg,
1907 	    "Total allocations done from emergency cache");
1908 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1909 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1910 	    SYSCTL_CHILDREN(rack_counters),
1911 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1912 	    &rack_to_alloc_limited,
1913 	    "Total allocations dropped due to limit");
1914 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1915 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1916 	    SYSCTL_CHILDREN(rack_counters),
1917 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1918 	    &rack_alloc_limited_conns,
1919 	    "Connections with allocations dropped due to limit");
1920 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1921 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1922 	    SYSCTL_CHILDREN(rack_counters),
1923 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1924 	    &rack_split_limited,
1925 	    "Split allocations dropped due to limit");
1926 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1927 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1928 	    SYSCTL_CHILDREN(rack_counters),
1929 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1930 	    &rack_rxt_clamps_cwnd,
1931 	    "Number of times that excessive rxt clamped the cwnd down");
1932 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1933 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1934 	    SYSCTL_CHILDREN(rack_counters),
1935 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1936 	    &rack_rxt_clamps_cwnd_uniq,
1937 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1938 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1939 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1940 	    SYSCTL_CHILDREN(rack_counters),
1941 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1942 	    &rack_persists_sends,
1943 	    "Number of times we sent a persist probe");
1944 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1945 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1946 	    SYSCTL_CHILDREN(rack_counters),
1947 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1948 	    &rack_persists_acks,
1949 	    "Number of times a persist probe was acked");
1950 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1951 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1952 	    SYSCTL_CHILDREN(rack_counters),
1953 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1954 	    &rack_persists_loss,
1955 	    "Number of times we detected a lost persist probe (no ack)");
1956 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1957 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1958 	    SYSCTL_CHILDREN(rack_counters),
1959 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1960 	    &rack_persists_lost_ends,
1961 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1962 #ifdef INVARIANTS
1963 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1964 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1965 	    SYSCTL_CHILDREN(rack_counters),
1966 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1967 	    &rack_adjust_map_bw,
1968 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1969 #endif
1970 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1971 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1972 	    SYSCTL_CHILDREN(rack_counters),
1973 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1974 	    &rack_multi_single_eq,
1975 	    "Number of compressed acks total represented");
1976 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1977 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1978 	    SYSCTL_CHILDREN(rack_counters),
1979 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1980 	    &rack_proc_non_comp_ack,
1981 	    "Number of non compresseds acks that we processed");
1982 
1983 
1984 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1985 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1986 	    SYSCTL_CHILDREN(rack_counters),
1987 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1988 	    &rack_sack_proc_all,
1989 	    "Total times we had to walk whole list for sack processing");
1990 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1991 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1992 	    SYSCTL_CHILDREN(rack_counters),
1993 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1994 	    &rack_sack_proc_restart,
1995 	    "Total times we had to walk whole list due to a restart");
1996 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1997 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1998 	    SYSCTL_CHILDREN(rack_counters),
1999 	    OID_AUTO, "sack_short", CTLFLAG_RD,
2000 	    &rack_sack_proc_short,
2001 	    "Total times we took shortcut for sack processing");
2002 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
2003 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2004 	    SYSCTL_CHILDREN(rack_attack),
2005 	    OID_AUTO, "skipacked", CTLFLAG_RD,
2006 	    &rack_sack_skipped_acked,
2007 	    "Total number of times we skipped previously sacked");
2008 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
2009 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2010 	    SYSCTL_CHILDREN(rack_attack),
2011 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
2012 	    &rack_sack_splits,
2013 	    "Total number of times we did the old fashion tree split");
2014 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
2015 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2016 	    SYSCTL_CHILDREN(rack_counters),
2017 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
2018 	    &rack_input_idle_reduces,
2019 	    "Total number of idle reductions on input");
2020 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
2021 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2022 	    SYSCTL_CHILDREN(rack_counters),
2023 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
2024 	    &rack_collapsed_win_seen,
2025 	    "Total number of collapsed window events seen (where our window shrinks)");
2026 
2027 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
2028 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2029 	    SYSCTL_CHILDREN(rack_counters),
2030 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
2031 	    &rack_collapsed_win,
2032 	    "Total number of collapsed window events where we mark packets");
2033 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
2034 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2035 	    SYSCTL_CHILDREN(rack_counters),
2036 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
2037 	    &rack_collapsed_win_rxt,
2038 	    "Total number of packets that were retransmitted");
2039 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
2040 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2041 	    SYSCTL_CHILDREN(rack_counters),
2042 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
2043 	    &rack_collapsed_win_rxt_bytes,
2044 	    "Total number of bytes that were retransmitted");
2045 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
2046 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
2047 	    SYSCTL_CHILDREN(rack_counters),
2048 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
2049 	    &rack_try_scwnd,
2050 	    "Total number of scwnd attempts");
2051 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
2052 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
2053 	    OID_AUTO, "outsize", CTLFLAG_RD,
2054 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
2055 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
2056 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
2057 	    OID_AUTO, "opts", CTLFLAG_RD,
2058 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
2059 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
2060 	    SYSCTL_CHILDREN(rack_sysctl_root),
2061 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2062 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
2063 }
2064 
2065 static uint32_t
2066 rc_init_window(struct tcp_rack *rack)
2067 {
2068 	return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
2069 
2070 }
2071 
2072 static uint64_t
2073 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
2074 {
2075 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
2076 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
2077 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2078 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
2079 	else
2080 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
2081 }
2082 
2083 static void
2084 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
2085 	uint64_t data, uint8_t mod, uint16_t aux,
2086 	struct tcp_sendfile_track *cur, int line)
2087 {
2088 #ifdef TCP_REQUEST_TRK
2089 	int do_log = 0;
2090 
2091 	/*
2092 	 * The rate cap one is noisy and only should come out when normal BB logging
2093 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2094 	 * once per chunk and make up the BBpoint that can be turned on by the client.
2095 	 */
2096 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2097 		/*
2098 		 * The very noisy two need to only come out when
2099 		 * we have verbose logging on.
2100 		 */
2101 		if (rack_verbose_logging != 0)
2102 			do_log = tcp_bblogging_on(rack->rc_tp);
2103 		else
2104 			do_log = 0;
2105 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2106 		/*
2107 		 * All other less noisy logs here except the measure which
2108 		 * also needs to come out on the point and the log.
2109 		 */
2110 		do_log = tcp_bblogging_on(rack->rc_tp);
2111 	} else {
2112 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2113 	}
2114 
2115 	if (do_log) {
2116 		union tcp_log_stackspecific log;
2117 		struct timeval tv;
2118 		uint64_t lt_bw;
2119 
2120 		/* Convert our ms to a microsecond */
2121 		memset(&log, 0, sizeof(log));
2122 
2123 		log.u_bbr.cwnd_gain = line;
2124 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2125 		log.u_bbr.rttProp = tim;
2126 		log.u_bbr.bw_inuse = cbw;
2127 		log.u_bbr.delRate = rack_get_gp_est(rack);
2128 		lt_bw = rack_get_lt_bw(rack);
2129 		log.u_bbr.flex1 = seq;
2130 		log.u_bbr.pacing_gain = aux;
2131 		/* lt_bw = < flex3 | flex2 > */
2132 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2133 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2134 		/* Record the last obtained us rtt in inflight */
2135 		if (cur == NULL) {
2136 			/* Make sure we are looking at the right log if an overide comes in */
2137 			cur = rack->r_ctl.rc_last_sft;
2138 		}
2139 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2140 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2141 		else {
2142 			/* Use the last known rtt i.e. the rack-rtt */
2143 			log.u_bbr.inflight = rack->rc_rack_rtt;
2144 		}
2145 		if (cur != NULL) {
2146 			uint64_t off;
2147 
2148 			log.u_bbr.cur_del_rate = cur->deadline;
2149 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2150 				/* start = < lost | pkt_epoch > */
2151 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2152 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2153 				log.u_bbr.flex6 = cur->start_seq;
2154 				log.u_bbr.pkts_out = cur->end_seq;
2155 			} else {
2156 				/* start = < lost | pkt_epoch > */
2157 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2158 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2159 				/* end = < pkts_out | flex6 > */
2160 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2161 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2162 			}
2163 			/* first_send = <lt_epoch | epoch> */
2164 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2165 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2166 			/* localtime = <delivered | applimited>*/
2167 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2168 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2169 #ifdef TCP_REQUEST_TRK
2170 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2171 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2172 #endif
2173 			log.u_bbr.inhpts = 1;
2174 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2175 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2176 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2177 		} else {
2178 			log.u_bbr.flex7 = 0xffff;
2179 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2180 		}
2181 		/*
2182 		 * Compose bbr_state to be a bit wise 0000ADHF
2183 		 * where A is the always_pace flag
2184 		 * where D is the dgp_on flag
2185 		 * where H is the hybrid_mode on flag
2186 		 * where F is the use_fixed_rate flag.
2187 		 */
2188 		log.u_bbr.bbr_state = rack->rc_always_pace;
2189 		log.u_bbr.bbr_state <<= 1;
2190 		log.u_bbr.bbr_state |= rack->dgp_on;
2191 		log.u_bbr.bbr_state <<= 1;
2192 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2193 		log.u_bbr.bbr_state <<= 1;
2194 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2195 		log.u_bbr.flex8 = mod;
2196 		tcp_log_event(rack->rc_tp, NULL,
2197 		    &rack->rc_inp->inp_socket->so_rcv,
2198 		    &rack->rc_inp->inp_socket->so_snd,
2199 		    TCP_HYBRID_PACING_LOG, 0,
2200 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2201 
2202 	}
2203 #endif
2204 }
2205 
2206 #ifdef TCP_REQUEST_TRK
2207 static void
2208 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2209 {
2210 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2211 		union tcp_log_stackspecific log;
2212 		struct timeval tv;
2213 		uint64_t off;
2214 
2215 		/* Convert our ms to a microsecond */
2216 		memset(&log, 0, sizeof(log));
2217 
2218 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2219 		log.u_bbr.delRate = cur->sent_at_fs;
2220 
2221 		if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2222 			/*
2223 			 * We did not get a new Rules Applied to set so
2224 			 * no overlapping send occured, this means the
2225 			 * current byte counts are correct.
2226 			 */
2227 			log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2228 			log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2229 		} else {
2230 			/*
2231 			 * Overlapping send case, we switched to a new
2232 			 * send and did a rules applied.
2233 			 */
2234 			log.u_bbr.cur_del_rate = cur->sent_at_ls;
2235 			log.u_bbr.rttProp = cur->rxt_at_ls;
2236 		}
2237 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2238 		log.u_bbr.cwnd_gain = line;
2239 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2240 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2241 		/* start = < flex1 | flex2 > */
2242 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2243 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2244 		/* end = < flex3 | flex4 > */
2245 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2246 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2247 
2248 		/* localtime = <delivered | applimited>*/
2249 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2250 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2251 		/* client timestamp = <lt_epoch | epoch>*/
2252 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2253 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2254 		/* now set all the flags in */
2255 		log.u_bbr.pkts_out = cur->hybrid_flags;
2256 		log.u_bbr.lost = cur->playout_ms;
2257 		log.u_bbr.flex6 = cur->flags;
2258 		/*
2259 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2260 		 * where a false retransmit occurred so first_send  <-> lastsend may
2261 		 * include longer time then it actually took if we have a false rxt.
2262 		 */
2263 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2264 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2265 		/*
2266 		 * Compose bbr_state to be a bit wise 0000ADHF
2267 		 * where A is the always_pace flag
2268 		 * where D is the dgp_on flag
2269 		 * where H is the hybrid_mode on flag
2270 		 * where F is the use_fixed_rate flag.
2271 		 */
2272 		log.u_bbr.bbr_state = rack->rc_always_pace;
2273 		log.u_bbr.bbr_state <<= 1;
2274 		log.u_bbr.bbr_state |= rack->dgp_on;
2275 		log.u_bbr.bbr_state <<= 1;
2276 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2277 		log.u_bbr.bbr_state <<= 1;
2278 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2279 
2280 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2281 		tcp_log_event(rack->rc_tp, NULL,
2282 		    &rack->rc_inp->inp_socket->so_rcv,
2283 		    &rack->rc_inp->inp_socket->so_snd,
2284 		    TCP_HYBRID_PACING_LOG, 0,
2285 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2286 	}
2287 }
2288 #endif
2289 
2290 static inline uint64_t
2291 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2292 {
2293 	uint64_t ret_bw, ether;
2294 	uint64_t u_segsiz;
2295 
2296 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2297 	if (rack->r_is_v6){
2298 #ifdef INET6
2299 		ether += sizeof(struct ip6_hdr);
2300 #endif
2301 		ether += 14;	/* eheader size 6+6+2 */
2302 	} else {
2303 #ifdef INET
2304 		ether += sizeof(struct ip);
2305 #endif
2306 		ether += 14;	/* eheader size 6+6+2 */
2307 	}
2308 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2309 	ret_bw = bw;
2310 	ret_bw *= ether;
2311 	ret_bw /= u_segsiz;
2312 	return (ret_bw);
2313 }
2314 
2315 static void
2316 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2317 {
2318 #ifdef TCP_REQUEST_TRK
2319 	struct timeval tv;
2320 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2321 #endif
2322 
2323 	if (rack->r_ctl.bw_rate_cap == 0)
2324 		return;
2325 #ifdef TCP_REQUEST_TRK
2326 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2327 	    (rack->r_ctl.rc_last_sft != NULL)) {
2328 		/*
2329 		 * We have a dynamic cap. The original target
2330 		 * is in bw_rate_cap, but we need to look at
2331 		 * how long it is until we hit the deadline.
2332 		 */
2333 		struct tcp_sendfile_track *ent;
2334 
2335       		ent = rack->r_ctl.rc_last_sft;
2336 		microuptime(&tv);
2337 		timenow = tcp_tv_to_lusectick(&tv);
2338 		if (timenow >= ent->deadline) {
2339 			/* No time left we do DGP only */
2340 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2341 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2342 			rack->r_ctl.bw_rate_cap = 0;
2343 			return;
2344 		}
2345 		/* We have the time */
2346 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2347 		if (timeleft < HPTS_MSEC_IN_SEC) {
2348 			/* If there is less than a ms left just use DGPs rate */
2349 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2350 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2351 			rack->r_ctl.bw_rate_cap = 0;
2352 			return;
2353 		}
2354 		/*
2355 		 * Now lets find the amount of data left to send.
2356 		 *
2357 		 * Now ideally we want to use the end_seq to figure out how much more
2358 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2359 		 */
2360 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2361 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2362 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2363 			else {
2364 				/* TSNH, we should catch it at the send */
2365 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2366 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2367 				rack->r_ctl.bw_rate_cap = 0;
2368 				return;
2369 			}
2370 		} else {
2371 			/*
2372 			 * The hard way, figure out how much is gone and then
2373 			 * take that away from the total the client asked for
2374 			 * (thats off by tls overhead if this is tls).
2375 			 */
2376 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2377 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2378 			else
2379 				lengone = 0;
2380 			if (lengone < (ent->end - ent->start))
2381 				lenleft = (ent->end - ent->start) - lengone;
2382 			else {
2383 				/* TSNH, we should catch it at the send */
2384 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2385 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2386 				rack->r_ctl.bw_rate_cap = 0;
2387 				return;
2388 			}
2389 		}
2390 		if (lenleft == 0) {
2391 			/* We have it all sent */
2392 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2393 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2394 			if (rack->r_ctl.bw_rate_cap)
2395 				goto normal_ratecap;
2396 			else
2397 				return;
2398 		}
2399 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2400 		calcbw /= timeleft;
2401 		/* Now we must compensate for IP/TCP overhead */
2402 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2403 		/* Update the bit rate cap */
2404 		rack->r_ctl.bw_rate_cap = calcbw;
2405 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2406 		    (rack_hybrid_allow_set_maxseg == 1) &&
2407 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2408 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2409 			uint32_t orig_max;
2410 
2411 			orig_max = rack->r_ctl.rc_pace_max_segs;
2412 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2413 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2414 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2415 		}
2416 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2417 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2418 		if ((calcbw > 0) && (*bw > calcbw)) {
2419 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2420 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2421 			*capped = 1;
2422 			*bw = calcbw;
2423 		}
2424 		return;
2425 	}
2426 normal_ratecap:
2427 #endif
2428 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2429 #ifdef TCP_REQUEST_TRK
2430 		if (rack->rc_hybrid_mode &&
2431 		    rack->rc_catch_up &&
2432 		    (rack->r_ctl.rc_last_sft != NULL) &&
2433 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2434 		    (rack_hybrid_allow_set_maxseg == 1) &&
2435 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2436 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2437 			uint32_t orig_max;
2438 
2439 			orig_max = rack->r_ctl.rc_pace_max_segs;
2440 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2441 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2442 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2443 		}
2444 #endif
2445 		*capped = 1;
2446 		*bw = rack->r_ctl.bw_rate_cap;
2447 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2448 				   *bw, 0, 0,
2449 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2450 	}
2451 }
2452 
2453 static uint64_t
2454 rack_get_gp_est(struct tcp_rack *rack)
2455 {
2456 	uint64_t bw, lt_bw, ret_bw;
2457 
2458 	if (rack->rc_gp_filled == 0) {
2459 		/*
2460 		 * We have yet no b/w measurement,
2461 		 * if we have a user set initial bw
2462 		 * return it. If we don't have that and
2463 		 * we have an srtt, use the tcp IW (10) to
2464 		 * calculate a fictional b/w over the SRTT
2465 		 * which is more or less a guess. Note
2466 		 * we don't use our IW from rack on purpose
2467 		 * so if we have like IW=30, we are not
2468 		 * calculating a "huge" b/w.
2469 		 */
2470 		uint64_t srtt;
2471 
2472 		if (rack->dis_lt_bw == 1)
2473 			lt_bw = 0;
2474 		else
2475 			lt_bw = rack_get_lt_bw(rack);
2476 		if (lt_bw) {
2477 			/*
2478 			 * No goodput bw but a long-term b/w does exist
2479 			 * lets use that.
2480 			 */
2481 			ret_bw = lt_bw;
2482 			goto compensate;
2483 		}
2484 		if (rack->r_ctl.init_rate)
2485 			return (rack->r_ctl.init_rate);
2486 
2487 		/* Ok lets come up with the IW guess, if we have a srtt */
2488 		if (rack->rc_tp->t_srtt == 0) {
2489 			/*
2490 			 * Go with old pacing method
2491 			 * i.e. burst mitigation only.
2492 			 */
2493 			return (0);
2494 		}
2495 		/* Ok lets get the initial TCP win (not racks) */
2496 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2497 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2498 		bw *= (uint64_t)USECS_IN_SECOND;
2499 		bw /= srtt;
2500 		ret_bw = bw;
2501 		goto compensate;
2502 
2503 	}
2504 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2505 		/* Averaging is done, we can return the value */
2506 		bw = rack->r_ctl.gp_bw;
2507 	} else {
2508 		/* Still doing initial average must calculate */
2509 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2510 	}
2511 	if (rack->dis_lt_bw) {
2512 		/* We are not using lt-bw */
2513 		ret_bw = bw;
2514 		goto compensate;
2515 	}
2516 	lt_bw = rack_get_lt_bw(rack);
2517 	if (lt_bw == 0) {
2518 		/* If we don't have one then equate it to the gp_bw */
2519 		lt_bw = rack->r_ctl.gp_bw;
2520 	}
2521 	if (rack->use_lesser_lt_bw) {
2522 		if (lt_bw < bw)
2523 			ret_bw = lt_bw;
2524 		else
2525 			ret_bw = bw;
2526 	} else {
2527 		if (lt_bw > bw)
2528 			ret_bw = lt_bw;
2529 		else
2530 			ret_bw = bw;
2531 	}
2532 	/*
2533 	 * Now lets compensate based on the TCP/IP overhead. Our
2534 	 * Goodput estimate does not include this so we must pace out
2535 	 * a bit faster since our pacing calculations do. The pacing
2536 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2537 	 * we are using to do this, so we do that here in the opposite
2538 	 * direction as well. This means that if we are tunneled and the
2539 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2540 	 * compensated for in the pacing time the opposite way.
2541 	 */
2542 compensate:
2543 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2544 	return(ret_bw);
2545 }
2546 
2547 
2548 static uint64_t
2549 rack_get_bw(struct tcp_rack *rack)
2550 {
2551 	uint64_t bw;
2552 
2553 	if (rack->use_fixed_rate) {
2554 		/* Return the fixed pacing rate */
2555 		return (rack_get_fixed_pacing_bw(rack));
2556 	}
2557 	bw = rack_get_gp_est(rack);
2558 	return (bw);
2559 }
2560 
2561 static uint16_t
2562 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2563 {
2564 	if (rack->use_fixed_rate) {
2565 		return (100);
2566 	} else if (rack->in_probe_rtt && (rsm == NULL))
2567 		return (rack->r_ctl.rack_per_of_gp_probertt);
2568 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2569 		  rack->r_ctl.rack_per_of_gp_rec)) {
2570 		if (rsm) {
2571 			/* a retransmission always use the recovery rate */
2572 			return (rack->r_ctl.rack_per_of_gp_rec);
2573 		} else if (rack->rack_rec_nonrxt_use_cr) {
2574 			/* Directed to use the configured rate */
2575 			goto configured_rate;
2576 		} else if (rack->rack_no_prr &&
2577 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2578 			/* No PRR, lets just use the b/w estimate only */
2579 			return (100);
2580 		} else {
2581 			/*
2582 			 * Here we may have a non-retransmit but we
2583 			 * have no overrides, so just use the recovery
2584 			 * rate (prr is in effect).
2585 			 */
2586 			return (rack->r_ctl.rack_per_of_gp_rec);
2587 		}
2588 	}
2589 configured_rate:
2590 	/* For the configured rate we look at our cwnd vs the ssthresh */
2591 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2592 		return (rack->r_ctl.rack_per_of_gp_ss);
2593 	else
2594 		return (rack->r_ctl.rack_per_of_gp_ca);
2595 }
2596 
2597 static void
2598 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2599 {
2600 	/*
2601 	 * Types of logs (mod value)
2602 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2603 	 * 2 = a dsack round begins, persist is reset to 16.
2604 	 * 3 = a dsack round ends
2605 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2606 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2607 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2608 	 */
2609 	if (tcp_bblogging_on(rack->rc_tp)) {
2610 		union tcp_log_stackspecific log;
2611 		struct timeval tv;
2612 
2613 		memset(&log, 0, sizeof(log));
2614 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2615 		log.u_bbr.flex1 <<= 1;
2616 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2617 		log.u_bbr.flex1 <<= 1;
2618 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2619 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2620 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2621 		log.u_bbr.flex4 = flex4;
2622 		log.u_bbr.flex5 = flex5;
2623 		log.u_bbr.flex6 = flex6;
2624 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2625 		log.u_bbr.flex8 = mod;
2626 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2627 		log.u_bbr.epoch = rack->r_ctl.current_round;
2628 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2629 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2630 		    &rack->rc_inp->inp_socket->so_rcv,
2631 		    &rack->rc_inp->inp_socket->so_snd,
2632 		    RACK_DSACK_HANDLING, 0,
2633 		    0, &log, false, &tv);
2634 	}
2635 }
2636 
2637 static void
2638 rack_log_hdwr_pacing(struct tcp_rack *rack,
2639 		     uint64_t rate, uint64_t hw_rate, int line,
2640 		     int error, uint16_t mod)
2641 {
2642 	if (tcp_bblogging_on(rack->rc_tp)) {
2643 		union tcp_log_stackspecific log;
2644 		struct timeval tv;
2645 		const struct ifnet *ifp;
2646 
2647 		memset(&log, 0, sizeof(log));
2648 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2649 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2650 		if (rack->r_ctl.crte) {
2651 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2652 		} else if (rack->rc_inp->inp_route.ro_nh &&
2653 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2654 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2655 		} else
2656 			ifp = NULL;
2657 		if (ifp) {
2658 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2659 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2660 		}
2661 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2662 		log.u_bbr.bw_inuse = rate;
2663 		log.u_bbr.flex5 = line;
2664 		log.u_bbr.flex6 = error;
2665 		log.u_bbr.flex7 = mod;
2666 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2667 		log.u_bbr.flex8 = rack->use_fixed_rate;
2668 		log.u_bbr.flex8 <<= 1;
2669 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2670 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2671 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2672 		if (rack->r_ctl.crte)
2673 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2674 		else
2675 			log.u_bbr.cur_del_rate = 0;
2676 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2677 		log.u_bbr.epoch = rack->r_ctl.current_round;
2678 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2679 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2680 		    &rack->rc_inp->inp_socket->so_rcv,
2681 		    &rack->rc_inp->inp_socket->so_snd,
2682 		    BBR_LOG_HDWR_PACE, 0,
2683 		    0, &log, false, &tv);
2684 	}
2685 }
2686 
2687 static uint64_t
2688 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2689 {
2690 	/*
2691 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2692 	 */
2693 	uint64_t bw_est, high_rate;
2694 	uint64_t gain;
2695 
2696 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2697 	bw_est = bw * gain;
2698 	bw_est /= (uint64_t)100;
2699 	/* Never fall below the minimum (def 64kbps) */
2700 	if (bw_est < RACK_MIN_BW)
2701 		bw_est = RACK_MIN_BW;
2702 	if (rack->r_rack_hw_rate_caps) {
2703 		/* Rate caps are in place */
2704 		if (rack->r_ctl.crte != NULL) {
2705 			/* We have a hdwr rate already */
2706 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2707 			if (bw_est >= high_rate) {
2708 				/* We are capping bw at the highest rate table entry */
2709 				if (rack_hw_rate_cap_per &&
2710 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2711 					rack->r_rack_hw_rate_caps = 0;
2712 					goto done;
2713 				}
2714 				rack_log_hdwr_pacing(rack,
2715 						     bw_est, high_rate, __LINE__,
2716 						     0, 3);
2717 				bw_est = high_rate;
2718 				if (capped)
2719 					*capped = 1;
2720 			}
2721 		} else if ((rack->rack_hdrw_pacing == 0) &&
2722 			   (rack->rack_hdw_pace_ena) &&
2723 			   (rack->rack_attempt_hdwr_pace == 0) &&
2724 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2725 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2726 			/*
2727 			 * Special case, we have not yet attempted hardware
2728 			 * pacing, and yet we may, when we do, find out if we are
2729 			 * above the highest rate. We need to know the maxbw for the interface
2730 			 * in question (if it supports ratelimiting). We get back
2731 			 * a 0, if the interface is not found in the RL lists.
2732 			 */
2733 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2734 			if (high_rate) {
2735 				/* Yep, we have a rate is it above this rate? */
2736 				if (bw_est > high_rate) {
2737 					bw_est = high_rate;
2738 					if (capped)
2739 						*capped = 1;
2740 				}
2741 			}
2742 		}
2743 	}
2744 done:
2745 	return (bw_est);
2746 }
2747 
2748 static void
2749 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2750 {
2751 	if (tcp_bblogging_on(rack->rc_tp)) {
2752 		union tcp_log_stackspecific log;
2753 		struct timeval tv;
2754 
2755 		if (rack->sack_attack_disable > 0)
2756 			goto log_anyway;
2757 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2758 			/*
2759 			 * We get 3 values currently for mod
2760 			 * 1 - We are retransmitting and this tells the reason.
2761 			 * 2 - We are clearing a dup-ack count.
2762 			 * 3 - We are incrementing a dup-ack count.
2763 			 *
2764 			 * The clear/increment are only logged
2765 			 * if you have BBverbose on.
2766 			 */
2767 			return;
2768 		}
2769 log_anyway:
2770 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2771 		log.u_bbr.flex1 = tsused;
2772 		log.u_bbr.flex2 = thresh;
2773 		log.u_bbr.flex3 = rsm->r_flags;
2774 		log.u_bbr.flex4 = rsm->r_dupack;
2775 		log.u_bbr.flex5 = rsm->r_start;
2776 		log.u_bbr.flex6 = rsm->r_end;
2777 		log.u_bbr.flex8 = mod;
2778 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2779 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2780 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2781 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2782 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2783 		log.u_bbr.pacing_gain = rack->r_must_retran;
2784 		log.u_bbr.epoch = rack->r_ctl.current_round;
2785 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2786 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2787 		    &rack->rc_inp->inp_socket->so_rcv,
2788 		    &rack->rc_inp->inp_socket->so_snd,
2789 		    BBR_LOG_SETTINGS_CHG, 0,
2790 		    0, &log, false, &tv);
2791 	}
2792 }
2793 
2794 static void
2795 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2796 {
2797 	if (tcp_bblogging_on(rack->rc_tp)) {
2798 		union tcp_log_stackspecific log;
2799 		struct timeval tv;
2800 
2801 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2802 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2803 		log.u_bbr.flex2 = to;
2804 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2805 		log.u_bbr.flex4 = slot;
2806 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2807 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2808 		log.u_bbr.flex7 = rack->rc_in_persist;
2809 		log.u_bbr.flex8 = which;
2810 		if (rack->rack_no_prr)
2811 			log.u_bbr.pkts_out = 0;
2812 		else
2813 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2814 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2815 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2816 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2817 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2818 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2819 		log.u_bbr.pacing_gain = rack->r_must_retran;
2820 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2821 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2822 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2823 		log.u_bbr.lost = rack_rto_min;
2824 		log.u_bbr.epoch = rack->r_ctl.roundends;
2825 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2826 		log.u_bbr.bw_inuse <<= 32;
2827 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2828 		log.u_bbr.applimited = rack->rc_tp->t_flags2;
2829 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2830 		    &rack->rc_inp->inp_socket->so_rcv,
2831 		    &rack->rc_inp->inp_socket->so_snd,
2832 		    BBR_LOG_TIMERSTAR, 0,
2833 		    0, &log, false, &tv);
2834 	}
2835 }
2836 
2837 static void
2838 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2839 {
2840 	if (tcp_bblogging_on(rack->rc_tp)) {
2841 		union tcp_log_stackspecific log;
2842 		struct timeval tv;
2843 
2844 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2845 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2846 		log.u_bbr.flex8 = to_num;
2847 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2848 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2849 		if (rsm == NULL)
2850 			log.u_bbr.flex3 = 0;
2851 		else
2852 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2853 		if (rack->rack_no_prr)
2854 			log.u_bbr.flex5 = 0;
2855 		else
2856 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2857 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2858 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2859 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2860 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2861 		log.u_bbr.pacing_gain = rack->r_must_retran;
2862 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2863 		log.u_bbr.bw_inuse <<= 32;
2864 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2865 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2866 		    &rack->rc_inp->inp_socket->so_rcv,
2867 		    &rack->rc_inp->inp_socket->so_snd,
2868 		    BBR_LOG_RTO, 0,
2869 		    0, &log, false, &tv);
2870 	}
2871 }
2872 
2873 static void
2874 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2875 		 struct rack_sendmap *prev,
2876 		 struct rack_sendmap *rsm,
2877 		 struct rack_sendmap *next,
2878 		 int flag, uint32_t th_ack, int line)
2879 {
2880 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2881 		union tcp_log_stackspecific log;
2882 		struct timeval tv;
2883 
2884 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2885 		log.u_bbr.flex8 = flag;
2886 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2887 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2888 		log.u_bbr.delRate = (uint64_t)rsm;
2889 		log.u_bbr.rttProp = (uint64_t)next;
2890 		log.u_bbr.flex7 = 0;
2891 		if (prev) {
2892 			log.u_bbr.flex1 = prev->r_start;
2893 			log.u_bbr.flex2 = prev->r_end;
2894 			log.u_bbr.flex7 |= 0x4;
2895 		}
2896 		if (rsm) {
2897 			log.u_bbr.flex3 = rsm->r_start;
2898 			log.u_bbr.flex4 = rsm->r_end;
2899 			log.u_bbr.flex7 |= 0x2;
2900 		}
2901 		if (next) {
2902 			log.u_bbr.flex5 = next->r_start;
2903 			log.u_bbr.flex6 = next->r_end;
2904 			log.u_bbr.flex7 |= 0x1;
2905 		}
2906 		log.u_bbr.applimited = line;
2907 		log.u_bbr.pkts_out = th_ack;
2908 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2909 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2910 		if (rack->rack_no_prr)
2911 			log.u_bbr.lost = 0;
2912 		else
2913 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2914 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2915 		log.u_bbr.bw_inuse <<= 32;
2916 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2917 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2918 		    &rack->rc_inp->inp_socket->so_rcv,
2919 		    &rack->rc_inp->inp_socket->so_snd,
2920 		    TCP_LOG_MAPCHG, 0,
2921 		    0, &log, false, &tv);
2922 	}
2923 }
2924 
2925 static void
2926 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2927 		 struct rack_sendmap *rsm, int conf)
2928 {
2929 	if (tcp_bblogging_on(tp)) {
2930 		union tcp_log_stackspecific log;
2931 		struct timeval tv;
2932 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2933 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2934 		log.u_bbr.flex1 = t;
2935 		log.u_bbr.flex2 = len;
2936 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2937 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2938 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2939 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2940 		log.u_bbr.flex7 = conf;
2941 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2942 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2943 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2944 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2945 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2946 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2947 		if (rsm) {
2948 			log.u_bbr.pkt_epoch = rsm->r_start;
2949 			log.u_bbr.lost = rsm->r_end;
2950 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2951 			/* We loose any upper of the 24 bits */
2952 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2953 		} else {
2954 			/* Its a SYN */
2955 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2956 			log.u_bbr.lost = 0;
2957 			log.u_bbr.cwnd_gain = 0;
2958 			log.u_bbr.pacing_gain = 0;
2959 		}
2960 		/* Write out general bits of interest rrs here */
2961 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2962 		log.u_bbr.use_lt_bw <<= 1;
2963 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2964 		log.u_bbr.use_lt_bw <<= 1;
2965 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2966 		log.u_bbr.use_lt_bw <<= 1;
2967 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2968 		log.u_bbr.use_lt_bw <<= 1;
2969 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2970 		log.u_bbr.use_lt_bw <<= 1;
2971 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2972 		log.u_bbr.use_lt_bw <<= 1;
2973 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2974 		log.u_bbr.use_lt_bw <<= 1;
2975 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2976 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2977 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2978 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2979 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2980 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2981 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2982 		log.u_bbr.bw_inuse <<= 32;
2983 		if (rsm)
2984 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2985 		TCP_LOG_EVENTP(tp, NULL,
2986 		    &rack->rc_inp->inp_socket->so_rcv,
2987 		    &rack->rc_inp->inp_socket->so_snd,
2988 		    BBR_LOG_BBRRTT, 0,
2989 		    0, &log, false, &tv);
2990 
2991 
2992 	}
2993 }
2994 
2995 static void
2996 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2997 {
2998 	/*
2999 	 * Log the rtt sample we are
3000 	 * applying to the srtt algorithm in
3001 	 * useconds.
3002 	 */
3003 	if (tcp_bblogging_on(rack->rc_tp)) {
3004 		union tcp_log_stackspecific log;
3005 		struct timeval tv;
3006 
3007 		/* Convert our ms to a microsecond */
3008 		memset(&log, 0, sizeof(log));
3009 		log.u_bbr.flex1 = rtt;
3010 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
3011 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
3012 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
3013 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
3014 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3015 		log.u_bbr.flex7 = 1;
3016 		log.u_bbr.flex8 = rack->sack_attack_disable;
3017 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3018 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3019 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3020 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3021 		log.u_bbr.pacing_gain = rack->r_must_retran;
3022 		/*
3023 		 * We capture in delRate the upper 32 bits as
3024 		 * the confidence level we had declared, and the
3025 		 * lower 32 bits as the actual RTT using the arrival
3026 		 * timestamp.
3027 		 */
3028 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
3029 		log.u_bbr.delRate <<= 32;
3030 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
3031 		/* Lets capture all the things that make up t_rtxcur */
3032 		log.u_bbr.applimited = rack_rto_min;
3033 		log.u_bbr.epoch = rack_rto_max;
3034 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
3035 		log.u_bbr.lost = rack_rto_min;
3036 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
3037 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
3038 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
3039 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
3040 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
3041 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3042 		    &rack->rc_inp->inp_socket->so_rcv,
3043 		    &rack->rc_inp->inp_socket->so_snd,
3044 		    TCP_LOG_RTT, 0,
3045 		    0, &log, false, &tv);
3046 	}
3047 }
3048 
3049 static void
3050 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
3051 {
3052 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3053 		union tcp_log_stackspecific log;
3054 		struct timeval tv;
3055 
3056 		/* Convert our ms to a microsecond */
3057 		memset(&log, 0, sizeof(log));
3058 		log.u_bbr.flex1 = rtt;
3059 		log.u_bbr.flex2 = send_time;
3060 		log.u_bbr.flex3 = ack_time;
3061 		log.u_bbr.flex4 = where;
3062 		log.u_bbr.flex7 = 2;
3063 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3064 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3065 		log.u_bbr.bw_inuse <<= 32;
3066 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3067 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3068 		    &rack->rc_inp->inp_socket->so_rcv,
3069 		    &rack->rc_inp->inp_socket->so_snd,
3070 		    TCP_LOG_RTT, 0,
3071 		    0, &log, false, &tv);
3072 	}
3073 }
3074 
3075 
3076 static void
3077 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
3078 {
3079 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3080 		union tcp_log_stackspecific log;
3081 		struct timeval tv;
3082 
3083 		/* Convert our ms to a microsecond */
3084 		memset(&log, 0, sizeof(log));
3085 		log.u_bbr.flex1 = idx;
3086 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
3087 		log.u_bbr.flex3 = tsecho;
3088 		log.u_bbr.flex7 = 3;
3089 		log.u_bbr.rttProp = tsv;
3090 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3091 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3092 		log.u_bbr.bw_inuse <<= 32;
3093 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3094 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3095 		    &rack->rc_inp->inp_socket->so_rcv,
3096 		    &rack->rc_inp->inp_socket->so_snd,
3097 		    TCP_LOG_RTT, 0,
3098 		    0, &log, false, &tv);
3099 	}
3100 }
3101 
3102 
3103 static inline void
3104 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
3105 {
3106 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3107 		union tcp_log_stackspecific log;
3108 		struct timeval tv;
3109 
3110 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3111 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3112 		log.u_bbr.flex1 = line;
3113 		log.u_bbr.flex2 = tick;
3114 		log.u_bbr.flex3 = tp->t_maxunacktime;
3115 		log.u_bbr.flex4 = tp->t_acktime;
3116 		log.u_bbr.flex8 = event;
3117 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3118 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3119 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3120 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3121 		log.u_bbr.pacing_gain = rack->r_must_retran;
3122 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3123 		log.u_bbr.bw_inuse <<= 32;
3124 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3125 		TCP_LOG_EVENTP(tp, NULL,
3126 		    &rack->rc_inp->inp_socket->so_rcv,
3127 		    &rack->rc_inp->inp_socket->so_snd,
3128 		    BBR_LOG_PROGRESS, 0,
3129 		    0, &log, false, &tv);
3130 	}
3131 }
3132 
3133 static void
3134 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3135 {
3136 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3137 		union tcp_log_stackspecific log;
3138 
3139 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3140 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3141 		log.u_bbr.flex1 = slot;
3142 		if (rack->rack_no_prr)
3143 			log.u_bbr.flex2 = 0;
3144 		else
3145 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3146 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3147 		log.u_bbr.flex5 = rack->r_ctl.ack_during_sd;
3148 		log.u_bbr.flex6 = line;
3149 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3150 		log.u_bbr.flex8 = rack->rc_in_persist;
3151 		log.u_bbr.timeStamp = cts;
3152 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3153 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3154 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3155 		log.u_bbr.pacing_gain = rack->r_must_retran;
3156 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3157 		    &rack->rc_inp->inp_socket->so_rcv,
3158 		    &rack->rc_inp->inp_socket->so_snd,
3159 		    BBR_LOG_BBRSND, 0,
3160 		    0, &log, false, tv);
3161 	}
3162 }
3163 
3164 static void
3165 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3166 {
3167 	if (tcp_bblogging_on(rack->rc_tp)) {
3168 		union tcp_log_stackspecific log;
3169 		struct timeval tv;
3170 
3171 		memset(&log, 0, sizeof(log));
3172 		log.u_bbr.flex1 = did_out;
3173 		log.u_bbr.flex2 = nxt_pkt;
3174 		log.u_bbr.flex3 = way_out;
3175 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3176 		if (rack->rack_no_prr)
3177 			log.u_bbr.flex5 = 0;
3178 		else
3179 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3180 		log.u_bbr.flex6 = nsegs;
3181 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3182 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3183 		log.u_bbr.flex7 <<= 1;
3184 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3185 		log.u_bbr.flex7 <<= 1;
3186 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3187 		log.u_bbr.flex8 = rack->rc_in_persist;
3188 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3189 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3190 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3191 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3192 		log.u_bbr.use_lt_bw <<= 1;
3193 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3194 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3195 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3196 		log.u_bbr.pacing_gain = rack->r_must_retran;
3197 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3198 		log.u_bbr.bw_inuse <<= 32;
3199 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3200 		log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3201 		log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3202 		log.u_bbr.lost = rack->rc_tp->t_srtt;
3203 		log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3204 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3205 		    &rack->rc_inp->inp_socket->so_rcv,
3206 		    &rack->rc_inp->inp_socket->so_snd,
3207 		    BBR_LOG_DOSEG_DONE, 0,
3208 		    0, &log, false, &tv);
3209 	}
3210 }
3211 
3212 static void
3213 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3214 {
3215 	if (tcp_bblogging_on(rack->rc_tp)) {
3216 		union tcp_log_stackspecific log;
3217 		struct timeval tv;
3218 
3219 		memset(&log, 0, sizeof(log));
3220 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3221 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3222 		log.u_bbr.flex4 = arg1;
3223 		log.u_bbr.flex5 = arg2;
3224 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3225 		log.u_bbr.flex6 = arg3;
3226 		log.u_bbr.flex8 = frm;
3227 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3228 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3229 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3230 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3231 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3232 		log.u_bbr.pacing_gain = rack->r_must_retran;
3233 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3234 		    &tptosocket(tp)->so_snd,
3235 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3236 	}
3237 }
3238 
3239 static void
3240 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3241 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3242 {
3243 	if (tcp_bblogging_on(rack->rc_tp)) {
3244 		union tcp_log_stackspecific log;
3245 		struct timeval tv;
3246 
3247 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3248 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3249 		log.u_bbr.flex1 = slot;
3250 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3251 		log.u_bbr.flex4 = reason;
3252 		if (rack->rack_no_prr)
3253 			log.u_bbr.flex5 = 0;
3254 		else
3255 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3256 		log.u_bbr.flex7 = hpts_calling;
3257 		log.u_bbr.flex8 = rack->rc_in_persist;
3258 		log.u_bbr.lt_epoch = cwnd_to_use;
3259 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3260 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3261 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3262 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3263 		log.u_bbr.pacing_gain = rack->r_must_retran;
3264 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3265 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3266 		log.u_bbr.bw_inuse <<= 32;
3267 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3268 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3269 		    &rack->rc_inp->inp_socket->so_rcv,
3270 		    &rack->rc_inp->inp_socket->so_snd,
3271 		    BBR_LOG_JUSTRET, 0,
3272 		    tlen, &log, false, &tv);
3273 	}
3274 }
3275 
3276 static void
3277 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3278 		   struct timeval *tv, uint32_t flags_on_entry)
3279 {
3280 	if (tcp_bblogging_on(rack->rc_tp)) {
3281 		union tcp_log_stackspecific log;
3282 
3283 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3284 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3285 		log.u_bbr.flex1 = line;
3286 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3287 		log.u_bbr.flex3 = flags_on_entry;
3288 		log.u_bbr.flex4 = us_cts;
3289 		if (rack->rack_no_prr)
3290 			log.u_bbr.flex5 = 0;
3291 		else
3292 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3293 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3294 		log.u_bbr.flex7 = hpts_removed;
3295 		log.u_bbr.flex8 = 1;
3296 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3297 		log.u_bbr.timeStamp = us_cts;
3298 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3299 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3300 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3301 		log.u_bbr.pacing_gain = rack->r_must_retran;
3302 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3303 		log.u_bbr.bw_inuse <<= 32;
3304 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3305 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3306 		    &rack->rc_inp->inp_socket->so_rcv,
3307 		    &rack->rc_inp->inp_socket->so_snd,
3308 		    BBR_LOG_TIMERCANC, 0,
3309 		    0, &log, false, tv);
3310 	}
3311 }
3312 
3313 static void
3314 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3315 			  uint32_t flex1, uint32_t flex2,
3316 			  uint32_t flex3, uint32_t flex4,
3317 			  uint32_t flex5, uint32_t flex6,
3318 			  uint16_t flex7, uint8_t mod)
3319 {
3320 	if (tcp_bblogging_on(rack->rc_tp)) {
3321 		union tcp_log_stackspecific log;
3322 		struct timeval tv;
3323 
3324 		if (mod == 1) {
3325 			/* No you can't use 1, its for the real to cancel */
3326 			return;
3327 		}
3328 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3329 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3330 		log.u_bbr.flex1 = flex1;
3331 		log.u_bbr.flex2 = flex2;
3332 		log.u_bbr.flex3 = flex3;
3333 		log.u_bbr.flex4 = flex4;
3334 		log.u_bbr.flex5 = flex5;
3335 		log.u_bbr.flex6 = flex6;
3336 		log.u_bbr.flex7 = flex7;
3337 		log.u_bbr.flex8 = mod;
3338 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3339 		    &rack->rc_inp->inp_socket->so_rcv,
3340 		    &rack->rc_inp->inp_socket->so_snd,
3341 		    BBR_LOG_TIMERCANC, 0,
3342 		    0, &log, false, &tv);
3343 	}
3344 }
3345 
3346 static void
3347 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3348 {
3349 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3350 		union tcp_log_stackspecific log;
3351 		struct timeval tv;
3352 
3353 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3354 		log.u_bbr.flex1 = timers;
3355 		log.u_bbr.flex2 = ret;
3356 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3357 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3358 		log.u_bbr.flex5 = cts;
3359 		if (rack->rack_no_prr)
3360 			log.u_bbr.flex6 = 0;
3361 		else
3362 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3363 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3364 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3365 		log.u_bbr.pacing_gain = rack->r_must_retran;
3366 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3367 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3368 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3369 		    &rack->rc_inp->inp_socket->so_rcv,
3370 		    &rack->rc_inp->inp_socket->so_snd,
3371 		    BBR_LOG_TO_PROCESS, 0,
3372 		    0, &log, false, &tv);
3373 	}
3374 }
3375 
3376 static void
3377 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3378 {
3379 	if (tcp_bblogging_on(rack->rc_tp)) {
3380 		union tcp_log_stackspecific log;
3381 		struct timeval tv;
3382 
3383 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3384 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3385 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3386 		if (rack->rack_no_prr)
3387 			log.u_bbr.flex3 = 0;
3388 		else
3389 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3390 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3391 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3392 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3393 		log.u_bbr.flex7 = line;
3394 		log.u_bbr.flex8 = frm;
3395 		log.u_bbr.pkts_out = orig_cwnd;
3396 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3397 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3398 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3399 		log.u_bbr.use_lt_bw <<= 1;
3400 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3401 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3402 		    &rack->rc_inp->inp_socket->so_rcv,
3403 		    &rack->rc_inp->inp_socket->so_snd,
3404 		    BBR_LOG_BBRUPD, 0,
3405 		    0, &log, false, &tv);
3406 	}
3407 }
3408 
3409 #ifdef TCP_SAD_DETECTION
3410 static void
3411 rack_log_sad(struct tcp_rack *rack, int event)
3412 {
3413 	if (tcp_bblogging_on(rack->rc_tp)) {
3414 		union tcp_log_stackspecific log;
3415 		struct timeval tv;
3416 
3417 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3418 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
3419 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
3420 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
3421 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
3422 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
3423 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
3424 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
3425 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
3426 		log.u_bbr.lt_epoch |= rack->do_detection;
3427 		log.u_bbr.applimited = tcp_map_minimum;
3428 		log.u_bbr.flex7 = rack->sack_attack_disable;
3429 		log.u_bbr.flex8 = event;
3430 		log.u_bbr.bbr_state = rack->rc_suspicious;
3431 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3432 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3433 		log.u_bbr.delivered = tcp_sad_decay_val;
3434 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3435 		    &rack->rc_inp->inp_socket->so_rcv,
3436 		    &rack->rc_inp->inp_socket->so_snd,
3437 		    TCP_SAD_DETECT, 0,
3438 		    0, &log, false, &tv);
3439 	}
3440 }
3441 #endif
3442 
3443 static void
3444 rack_counter_destroy(void)
3445 {
3446 	counter_u64_free(rack_total_bytes);
3447 	counter_u64_free(rack_fto_send);
3448 	counter_u64_free(rack_fto_rsm_send);
3449 	counter_u64_free(rack_nfto_resend);
3450 	counter_u64_free(rack_hw_pace_init_fail);
3451 	counter_u64_free(rack_hw_pace_lost);
3452 	counter_u64_free(rack_non_fto_send);
3453 	counter_u64_free(rack_extended_rfo);
3454 	counter_u64_free(rack_ack_total);
3455 	counter_u64_free(rack_express_sack);
3456 	counter_u64_free(rack_sack_total);
3457 	counter_u64_free(rack_move_none);
3458 	counter_u64_free(rack_move_some);
3459 	counter_u64_free(rack_sack_attacks_detected);
3460 	counter_u64_free(rack_sack_attacks_reversed);
3461 	counter_u64_free(rack_sack_attacks_suspect);
3462 	counter_u64_free(rack_sack_used_next_merge);
3463 	counter_u64_free(rack_sack_used_prev_merge);
3464 	counter_u64_free(rack_tlp_tot);
3465 	counter_u64_free(rack_tlp_newdata);
3466 	counter_u64_free(rack_tlp_retran);
3467 	counter_u64_free(rack_tlp_retran_bytes);
3468 	counter_u64_free(rack_to_tot);
3469 	counter_u64_free(rack_saw_enobuf);
3470 	counter_u64_free(rack_saw_enobuf_hw);
3471 	counter_u64_free(rack_saw_enetunreach);
3472 	counter_u64_free(rack_hot_alloc);
3473 	counter_u64_free(tcp_policer_detected);
3474 	counter_u64_free(rack_to_alloc);
3475 	counter_u64_free(rack_to_alloc_hard);
3476 	counter_u64_free(rack_to_alloc_emerg);
3477 	counter_u64_free(rack_to_alloc_limited);
3478 	counter_u64_free(rack_alloc_limited_conns);
3479 	counter_u64_free(rack_split_limited);
3480 	counter_u64_free(rack_multi_single_eq);
3481 	counter_u64_free(rack_rxt_clamps_cwnd);
3482 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3483 	counter_u64_free(rack_proc_non_comp_ack);
3484 	counter_u64_free(rack_sack_proc_all);
3485 	counter_u64_free(rack_sack_proc_restart);
3486 	counter_u64_free(rack_sack_proc_short);
3487 	counter_u64_free(rack_sack_skipped_acked);
3488 	counter_u64_free(rack_sack_splits);
3489 	counter_u64_free(rack_input_idle_reduces);
3490 	counter_u64_free(rack_collapsed_win);
3491 	counter_u64_free(rack_collapsed_win_rxt);
3492 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3493 	counter_u64_free(rack_collapsed_win_seen);
3494 	counter_u64_free(rack_try_scwnd);
3495 	counter_u64_free(rack_persists_sends);
3496 	counter_u64_free(rack_persists_acks);
3497 	counter_u64_free(rack_persists_loss);
3498 	counter_u64_free(rack_persists_lost_ends);
3499 #ifdef INVARIANTS
3500 	counter_u64_free(rack_adjust_map_bw);
3501 #endif
3502 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3503 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3504 }
3505 
3506 static struct rack_sendmap *
3507 rack_alloc(struct tcp_rack *rack)
3508 {
3509 	struct rack_sendmap *rsm;
3510 
3511 	/*
3512 	 * First get the top of the list it in
3513 	 * theory is the "hottest" rsm we have,
3514 	 * possibly just freed by ack processing.
3515 	 */
3516 	if (rack->rc_free_cnt > rack_free_cache) {
3517 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3518 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3519 		counter_u64_add(rack_hot_alloc, 1);
3520 		rack->rc_free_cnt--;
3521 		return (rsm);
3522 	}
3523 	/*
3524 	 * Once we get under our free cache we probably
3525 	 * no longer have a "hot" one available. Lets
3526 	 * get one from UMA.
3527 	 */
3528 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3529 	if (rsm) {
3530 		rack->r_ctl.rc_num_maps_alloced++;
3531 		counter_u64_add(rack_to_alloc, 1);
3532 		return (rsm);
3533 	}
3534 	/*
3535 	 * Dig in to our aux rsm's (the last two) since
3536 	 * UMA failed to get us one.
3537 	 */
3538 	if (rack->rc_free_cnt) {
3539 		counter_u64_add(rack_to_alloc_emerg, 1);
3540 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3541 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3542 		rack->rc_free_cnt--;
3543 		return (rsm);
3544 	}
3545 	return (NULL);
3546 }
3547 
3548 static struct rack_sendmap *
3549 rack_alloc_full_limit(struct tcp_rack *rack)
3550 {
3551 	if ((V_tcp_map_entries_limit > 0) &&
3552 	    (rack->do_detection == 0) &&
3553 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3554 		counter_u64_add(rack_to_alloc_limited, 1);
3555 		if (!rack->alloc_limit_reported) {
3556 			rack->alloc_limit_reported = 1;
3557 			counter_u64_add(rack_alloc_limited_conns, 1);
3558 		}
3559 		return (NULL);
3560 	}
3561 	return (rack_alloc(rack));
3562 }
3563 
3564 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3565 static struct rack_sendmap *
3566 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3567 {
3568 	struct rack_sendmap *rsm;
3569 
3570 	if (limit_type) {
3571 		/* currently there is only one limit type */
3572 		if (rack->r_ctl.rc_split_limit > 0 &&
3573 		    (rack->do_detection == 0) &&
3574 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3575 			counter_u64_add(rack_split_limited, 1);
3576 			if (!rack->alloc_limit_reported) {
3577 				rack->alloc_limit_reported = 1;
3578 				counter_u64_add(rack_alloc_limited_conns, 1);
3579 			}
3580 			return (NULL);
3581 #ifdef TCP_SAD_DETECTION
3582 		} else if ((tcp_sad_limit != 0) &&
3583 			   (rack->do_detection == 1) &&
3584 			   (rack->r_ctl.rc_num_split_allocs >= tcp_sad_limit)) {
3585 			counter_u64_add(rack_split_limited, 1);
3586 			if (!rack->alloc_limit_reported) {
3587 				rack->alloc_limit_reported = 1;
3588 				counter_u64_add(rack_alloc_limited_conns, 1);
3589 			}
3590 			return (NULL);
3591 #endif
3592 		}
3593 	}
3594 
3595 	/* allocate and mark in the limit type, if set */
3596 	rsm = rack_alloc(rack);
3597 	if (rsm != NULL && limit_type) {
3598 		rsm->r_limit_type = limit_type;
3599 		rack->r_ctl.rc_num_split_allocs++;
3600 	}
3601 	return (rsm);
3602 }
3603 
3604 static void
3605 rack_free_trim(struct tcp_rack *rack)
3606 {
3607 	struct rack_sendmap *rsm;
3608 
3609 	/*
3610 	 * Free up all the tail entries until
3611 	 * we get our list down to the limit.
3612 	 */
3613 	while (rack->rc_free_cnt > rack_free_cache) {
3614 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3615 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3616 		rack->rc_free_cnt--;
3617 		rack->r_ctl.rc_num_maps_alloced--;
3618 		uma_zfree(rack_zone, rsm);
3619 	}
3620 }
3621 
3622 static void
3623 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3624 {
3625 	if (rsm->r_flags & RACK_APP_LIMITED) {
3626 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3627 			rack->r_ctl.rc_app_limited_cnt--;
3628 		}
3629 	}
3630 	if (rsm->r_limit_type) {
3631 		/* currently there is only one limit type */
3632 		rack->r_ctl.rc_num_split_allocs--;
3633 	}
3634 	if (rsm == rack->r_ctl.rc_first_appl) {
3635 		rack->r_ctl.cleared_app_ack_seq = rsm->r_start + (rsm->r_end - rsm->r_start);
3636 		rack->r_ctl.cleared_app_ack = 1;
3637 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3638 			rack->r_ctl.rc_first_appl = NULL;
3639 		else
3640 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3641 	}
3642 	if (rsm == rack->r_ctl.rc_resend)
3643 		rack->r_ctl.rc_resend = NULL;
3644 	if (rsm == rack->r_ctl.rc_end_appl)
3645 		rack->r_ctl.rc_end_appl = NULL;
3646 	if (rack->r_ctl.rc_tlpsend == rsm)
3647 		rack->r_ctl.rc_tlpsend = NULL;
3648 	if (rack->r_ctl.rc_sacklast == rsm)
3649 		rack->r_ctl.rc_sacklast = NULL;
3650 	memset(rsm, 0, sizeof(struct rack_sendmap));
3651 	/* Make sure we are not going to overrun our count limit of 0xff */
3652 	if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3653 		rack_free_trim(rack);
3654 	}
3655 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3656 	rack->rc_free_cnt++;
3657 }
3658 
3659 static uint32_t
3660 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3661 {
3662 	uint64_t srtt, bw, len, tim;
3663 	uint32_t segsiz, def_len, minl;
3664 
3665 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3666 	def_len = rack_def_data_window * segsiz;
3667 	if (rack->rc_gp_filled == 0) {
3668 		/*
3669 		 * We have no measurement (IW is in flight?) so
3670 		 * we can only guess using our data_window sysctl
3671 		 * value (usually 20MSS).
3672 		 */
3673 		return (def_len);
3674 	}
3675 	/*
3676 	 * Now we have a number of factors to consider.
3677 	 *
3678 	 * 1) We have a desired BDP which is usually
3679 	 *    at least 2.
3680 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3681 	 *    but we allow it too to be more.
3682 	 * 3) We want to make sure a measurement last N useconds (if
3683 	 *    we have set rack_min_measure_usec.
3684 	 *
3685 	 * We handle the first concern here by trying to create a data
3686 	 * window of max(rack_def_data_window, DesiredBDP). The
3687 	 * second concern we handle in not letting the measurement
3688 	 * window end normally until at least the required SRTT's
3689 	 * have gone by which is done further below in
3690 	 * rack_enough_for_measurement(). Finally the third concern
3691 	 * we also handle here by calculating how long that time
3692 	 * would take at the current BW and then return the
3693 	 * max of our first calculation and that length. Note
3694 	 * that if rack_min_measure_usec is 0, we don't deal
3695 	 * with concern 3. Also for both Concern 1 and 3 an
3696 	 * application limited period could end the measurement
3697 	 * earlier.
3698 	 *
3699 	 * So lets calculate the BDP with the "known" b/w using
3700 	 * the SRTT has our rtt and then multiply it by the
3701 	 * goal.
3702 	 */
3703 	bw = rack_get_bw(rack);
3704 	srtt = (uint64_t)tp->t_srtt;
3705 	len = bw * srtt;
3706 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3707 	len *= max(1, rack_goal_bdp);
3708 	/* Now we need to round up to the nearest MSS */
3709 	len = roundup(len, segsiz);
3710 	if (rack_min_measure_usec) {
3711 		/* Now calculate our min length for this b/w */
3712 		tim = rack_min_measure_usec;
3713 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3714 		if (minl == 0)
3715 			minl = 1;
3716 		minl = roundup(minl, segsiz);
3717 		if (len < minl)
3718 			len = minl;
3719 	}
3720 	/*
3721 	 * Now if we have a very small window we want
3722 	 * to attempt to get the window that is
3723 	 * as small as possible. This happens on
3724 	 * low b/w connections and we don't want to
3725 	 * span huge numbers of rtt's between measurements.
3726 	 *
3727 	 * We basically include 2 over our "MIN window" so
3728 	 * that the measurement can be shortened (possibly) by
3729 	 * an ack'ed packet.
3730 	 */
3731 	if (len < def_len)
3732 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3733 	else
3734 		return (max((uint32_t)len, def_len));
3735 
3736 }
3737 
3738 static int
3739 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3740 {
3741 	uint32_t tim, srtts, segsiz;
3742 
3743 	/*
3744 	 * Has enough time passed for the GP measurement to be valid?
3745 	 */
3746 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3747 		/* Not enough bytes yet */
3748 		return (0);
3749 	}
3750 	if ((tp->snd_max == tp->snd_una) ||
3751 	    (th_ack == tp->snd_max)){
3752 		/*
3753 		 * All is acked quality of all acked is
3754 		 * usually low or medium, but we in theory could split
3755 		 * all acked into two cases, where you got
3756 		 * a signifigant amount of your window and
3757 		 * where you did not. For now we leave it
3758 		 * but it is something to contemplate in the
3759 		 * future. The danger here is that delayed ack
3760 		 * is effecting the last byte (which is a 50:50 chance).
3761 		 */
3762 		*quality = RACK_QUALITY_ALLACKED;
3763 		return (1);
3764 	}
3765 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3766 		/*
3767 		 * We obtained our entire window of data we wanted
3768 		 * no matter if we are in recovery or not then
3769 		 * its ok since expanding the window does not
3770 		 * make things fuzzy (or at least not as much).
3771 		 */
3772 		*quality = RACK_QUALITY_HIGH;
3773 		return (1);
3774 	}
3775 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3776 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3777 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3778 		/* Not enough bytes yet */
3779 		return (0);
3780 	}
3781 	if (rack->r_ctl.rc_first_appl &&
3782 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3783 		/*
3784 		 * We are up to the app limited send point
3785 		 * we have to measure irrespective of the time..
3786 		 */
3787 		*quality = RACK_QUALITY_APPLIMITED;
3788 		return (1);
3789 	}
3790 	/* Now what about time? */
3791 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3792 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3793 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3794 		/*
3795 		 * We do not allow a measurement if we are in recovery
3796 		 * that would shrink the goodput window we wanted.
3797 		 * This is to prevent cloudyness of when the last send
3798 		 * was actually made.
3799 		 */
3800 		*quality = RACK_QUALITY_HIGH;
3801 		return (1);
3802 	}
3803 	/* Nope not even a full SRTT has passed */
3804 	return (0);
3805 }
3806 
3807 static void
3808 rack_log_timely(struct tcp_rack *rack,
3809 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3810 		uint64_t up_bnd, int line, uint8_t method)
3811 {
3812 	if (tcp_bblogging_on(rack->rc_tp)) {
3813 		union tcp_log_stackspecific log;
3814 		struct timeval tv;
3815 
3816 		memset(&log, 0, sizeof(log));
3817 		log.u_bbr.flex1 = logged;
3818 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3819 		log.u_bbr.flex2 <<= 4;
3820 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3821 		log.u_bbr.flex2 <<= 4;
3822 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3823 		log.u_bbr.flex2 <<= 4;
3824 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3825 		log.u_bbr.flex3 = rack->rc_gp_incr;
3826 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3827 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3828 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3829 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3830 		log.u_bbr.flex8 = method;
3831 		log.u_bbr.cur_del_rate = cur_bw;
3832 		log.u_bbr.delRate = low_bnd;
3833 		log.u_bbr.bw_inuse = up_bnd;
3834 		log.u_bbr.rttProp = rack_get_bw(rack);
3835 		log.u_bbr.pkt_epoch = line;
3836 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3837 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3838 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3839 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3840 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3841 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3842 		log.u_bbr.cwnd_gain <<= 1;
3843 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3844 		log.u_bbr.cwnd_gain <<= 1;
3845 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3846 		log.u_bbr.cwnd_gain <<= 1;
3847 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3848 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3849 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3850 		    &rack->rc_inp->inp_socket->so_rcv,
3851 		    &rack->rc_inp->inp_socket->so_snd,
3852 		    TCP_TIMELY_WORK, 0,
3853 		    0, &log, false, &tv);
3854 	}
3855 }
3856 
3857 static int
3858 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3859 {
3860 	/*
3861 	 * Before we increase we need to know if
3862 	 * the estimate just made was less than
3863 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3864 	 *
3865 	 * If we already are pacing at a fast enough
3866 	 * rate to push us faster there is no sense of
3867 	 * increasing.
3868 	 *
3869 	 * We first caculate our actual pacing rate (ss or ca multiplier
3870 	 * times our cur_bw).
3871 	 *
3872 	 * Then we take the last measured rate and multipy by our
3873 	 * maximum pacing overage to give us a max allowable rate.
3874 	 *
3875 	 * If our act_rate is smaller than our max_allowable rate
3876 	 * then we should increase. Else we should hold steady.
3877 	 *
3878 	 */
3879 	uint64_t act_rate, max_allow_rate;
3880 
3881 	if (rack_timely_no_stopping)
3882 		return (1);
3883 
3884 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3885 		/*
3886 		 * Initial startup case or
3887 		 * everything is acked case.
3888 		 */
3889 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3890 				__LINE__, 9);
3891 		return (1);
3892 	}
3893 	if (mult <= 100) {
3894 		/*
3895 		 * We can always pace at or slightly above our rate.
3896 		 */
3897 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3898 				__LINE__, 9);
3899 		return (1);
3900 	}
3901 	act_rate = cur_bw * (uint64_t)mult;
3902 	act_rate /= 100;
3903 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3904 	max_allow_rate /= 100;
3905 	if (act_rate < max_allow_rate) {
3906 		/*
3907 		 * Here the rate we are actually pacing at
3908 		 * is smaller than 10% above our last measurement.
3909 		 * This means we are pacing below what we would
3910 		 * like to try to achieve (plus some wiggle room).
3911 		 */
3912 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3913 				__LINE__, 9);
3914 		return (1);
3915 	} else {
3916 		/*
3917 		 * Here we are already pacing at least rack_max_per_above(10%)
3918 		 * what we are getting back. This indicates most likely
3919 		 * that we are being limited (cwnd/rwnd/app) and can't
3920 		 * get any more b/w. There is no sense of trying to
3921 		 * raise up the pacing rate its not speeding us up
3922 		 * and we already are pacing faster than we are getting.
3923 		 */
3924 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3925 				__LINE__, 8);
3926 		return (0);
3927 	}
3928 }
3929 
3930 static void
3931 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3932 {
3933 	/*
3934 	 * When we drag bottom, we want to assure
3935 	 * that no multiplier is below 1.0, if so
3936 	 * we want to restore it to at least that.
3937 	 */
3938 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3939 		/* This is unlikely we usually do not touch recovery */
3940 		rack->r_ctl.rack_per_of_gp_rec = 100;
3941 	}
3942 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3943 		rack->r_ctl.rack_per_of_gp_ca = 100;
3944 	}
3945 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3946 		rack->r_ctl.rack_per_of_gp_ss = 100;
3947 	}
3948 }
3949 
3950 static void
3951 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3952 {
3953 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3954 		rack->r_ctl.rack_per_of_gp_ca = 100;
3955 	}
3956 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3957 		rack->r_ctl.rack_per_of_gp_ss = 100;
3958 	}
3959 }
3960 
3961 static void
3962 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3963 {
3964 	int32_t  calc, logged, plus;
3965 
3966 	logged = 0;
3967 
3968 	if (rack->rc_skip_timely)
3969 		return;
3970 	if (override) {
3971 		/*
3972 		 * override is passed when we are
3973 		 * loosing b/w and making one last
3974 		 * gasp at trying to not loose out
3975 		 * to a new-reno flow.
3976 		 */
3977 		goto extra_boost;
3978 	}
3979 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3980 	if (rack->rc_gp_incr &&
3981 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3982 		/*
3983 		 * Reset and get 5 strokes more before the boost. Note
3984 		 * that the count is 0 based so we have to add one.
3985 		 */
3986 extra_boost:
3987 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3988 		rack->rc_gp_timely_inc_cnt = 0;
3989 	} else
3990 		plus = (uint32_t)rack_gp_increase_per;
3991 	/* Must be at least 1% increase for true timely increases */
3992 	if ((plus < 1) &&
3993 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3994 		plus = 1;
3995 	if (rack->rc_gp_saw_rec &&
3996 	    (rack->rc_gp_no_rec_chg == 0) &&
3997 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3998 				  rack->r_ctl.rack_per_of_gp_rec)) {
3999 		/* We have been in recovery ding it too */
4000 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
4001 		if (calc > 0xffff)
4002 			calc = 0xffff;
4003 		logged |= 1;
4004 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
4005 		if (rack->r_ctl.rack_per_upper_bound_ca &&
4006 		    (rack->rc_dragged_bottom == 0) &&
4007 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
4008 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
4009 	}
4010 	if (rack->rc_gp_saw_ca &&
4011 	    (rack->rc_gp_saw_ss == 0) &&
4012 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
4013 				  rack->r_ctl.rack_per_of_gp_ca)) {
4014 		/* In CA */
4015 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
4016 		if (calc > 0xffff)
4017 			calc = 0xffff;
4018 		logged |= 2;
4019 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
4020 		if (rack->r_ctl.rack_per_upper_bound_ca &&
4021 		    (rack->rc_dragged_bottom == 0) &&
4022 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
4023 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
4024 	}
4025 	if (rack->rc_gp_saw_ss &&
4026 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
4027 				  rack->r_ctl.rack_per_of_gp_ss)) {
4028 		/* In SS */
4029 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
4030 		if (calc > 0xffff)
4031 			calc = 0xffff;
4032 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
4033 		if (rack->r_ctl.rack_per_upper_bound_ss &&
4034 		    (rack->rc_dragged_bottom == 0) &&
4035 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
4036 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
4037 		logged |= 4;
4038 	}
4039 	if (logged &&
4040 	    (rack->rc_gp_incr == 0)){
4041 		/* Go into increment mode */
4042 		rack->rc_gp_incr = 1;
4043 		rack->rc_gp_timely_inc_cnt = 0;
4044 	}
4045 	if (rack->rc_gp_incr &&
4046 	    logged &&
4047 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
4048 		rack->rc_gp_timely_inc_cnt++;
4049 	}
4050 	rack_log_timely(rack,  logged, plus, 0, 0,
4051 			__LINE__, 1);
4052 }
4053 
4054 static uint32_t
4055 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
4056 {
4057 	/*-
4058 	 * norm_grad = rtt_diff / minrtt;
4059 	 * new_per = curper * (1 - B * norm_grad)
4060 	 *
4061 	 * B = rack_gp_decrease_per (default 80%)
4062 	 * rtt_dif = input var current rtt-diff
4063 	 * curper = input var current percentage
4064 	 * minrtt = from rack filter
4065 	 *
4066 	 * In order to do the floating point calculations above we
4067 	 * do an integer conversion. The code looks confusing so let me
4068 	 * translate it into something that use more variables and
4069 	 * is clearer for us humans :)
4070 	 *
4071 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
4072 	 * uint32_t perf;
4073 	 *
4074 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
4075 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
4076 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
4077 	 * inverse /= 1000000;
4078 	 * reduce_by = (1000000 - inverse);
4079 	 * final_result = (cur_per * reduce_by) / 1000000;
4080 	 * perf = (uint32_t)final_result;
4081 	 */
4082 	uint64_t perf;
4083 
4084 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
4085 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
4086 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
4087 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
4088 		     (uint64_t)1000000)) /
4089 		(uint64_t)1000000);
4090 	if (perf > curper) {
4091 		/* TSNH */
4092 		perf = curper - 1;
4093 	}
4094 	return ((uint32_t)perf);
4095 }
4096 
4097 static uint32_t
4098 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
4099 {
4100 	/*
4101 	 *                                   highrttthresh
4102 	 * result = curper * (1 - (B * ( 1 -  ------          ))
4103 	 *                                     gp_srtt
4104 	 *
4105 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
4106 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
4107 	 */
4108 	uint64_t perf;
4109 	uint32_t highrttthresh;
4110 
4111 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4112 
4113 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
4114 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
4115 					((uint64_t)highrttthresh * (uint64_t)1000000) /
4116 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
4117 	if (tcp_bblogging_on(rack->rc_tp)) {
4118 		uint64_t log1;
4119 
4120 		log1 = rtt;
4121 		log1 <<= 32;
4122 		log1 |= highrttthresh;
4123 		rack_log_timely(rack,
4124 				rack_gp_decrease_per,
4125 				(uint64_t)curper,
4126 				log1,
4127 				perf,
4128 				__LINE__,
4129 				15);
4130 	}
4131 	return (perf);
4132 }
4133 
4134 static void
4135 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
4136 {
4137 	uint64_t logvar, logvar2, logvar3;
4138 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
4139 
4140 	if (rack->rc_skip_timely)
4141 		return;
4142 	if (rack->rc_gp_incr) {
4143 		/* Turn off increment counting */
4144 		rack->rc_gp_incr = 0;
4145 		rack->rc_gp_timely_inc_cnt = 0;
4146 	}
4147 	ss_red = ca_red = rec_red = 0;
4148 	logged = 0;
4149 	/* Calculate the reduction value */
4150 	if (rtt_diff < 0) {
4151 		rtt_diff *= -1;
4152 	}
4153 	/* Must be at least 1% reduction */
4154 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4155 		/* We have been in recovery ding it too */
4156 		if (timely_says == 2) {
4157 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4158 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4159 			if (alt < new_per)
4160 				val = alt;
4161 			else
4162 				val = new_per;
4163 		} else
4164 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4165 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4166 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4167 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4168 		} else {
4169 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4170 			rec_red = 0;
4171 		}
4172 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4173 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4174 		logged |= 1;
4175 	}
4176 	if (rack->rc_gp_saw_ss) {
4177 		/* Sent in SS */
4178 		if (timely_says == 2) {
4179 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4180 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4181 			if (alt < new_per)
4182 				val = alt;
4183 			else
4184 				val = new_per;
4185 		} else
4186 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4187 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4188 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4189 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4190 		} else {
4191 			ss_red = new_per;
4192 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4193 			logvar = new_per;
4194 			logvar <<= 32;
4195 			logvar |= alt;
4196 			logvar2 = (uint32_t)rtt;
4197 			logvar2 <<= 32;
4198 			logvar2 |= (uint32_t)rtt_diff;
4199 			logvar3 = rack_gp_rtt_maxmul;
4200 			logvar3 <<= 32;
4201 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4202 			rack_log_timely(rack, timely_says,
4203 					logvar2, logvar3,
4204 					logvar, __LINE__, 10);
4205 		}
4206 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4207 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4208 		logged |= 4;
4209 	} else if (rack->rc_gp_saw_ca) {
4210 		/* Sent in CA */
4211 		if (timely_says == 2) {
4212 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4213 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4214 			if (alt < new_per)
4215 				val = alt;
4216 			else
4217 				val = new_per;
4218 		} else
4219 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4220 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4221 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4222 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4223 		} else {
4224 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4225 			ca_red = 0;
4226 			logvar = new_per;
4227 			logvar <<= 32;
4228 			logvar |= alt;
4229 			logvar2 = (uint32_t)rtt;
4230 			logvar2 <<= 32;
4231 			logvar2 |= (uint32_t)rtt_diff;
4232 			logvar3 = rack_gp_rtt_maxmul;
4233 			logvar3 <<= 32;
4234 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4235 			rack_log_timely(rack, timely_says,
4236 					logvar2, logvar3,
4237 					logvar, __LINE__, 10);
4238 		}
4239 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4240 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4241 		logged |= 2;
4242 	}
4243 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4244 		rack->rc_gp_timely_dec_cnt++;
4245 		if (rack_timely_dec_clear &&
4246 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4247 			rack->rc_gp_timely_dec_cnt = 0;
4248 	}
4249 	logvar = ss_red;
4250 	logvar <<= 32;
4251 	logvar |= ca_red;
4252 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4253 			__LINE__, 2);
4254 }
4255 
4256 static void
4257 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4258 		     uint32_t rtt, uint32_t line, uint8_t reas)
4259 {
4260 	if (tcp_bblogging_on(rack->rc_tp)) {
4261 		union tcp_log_stackspecific log;
4262 		struct timeval tv;
4263 
4264 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4265 		log.u_bbr.flex1 = line;
4266 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4267 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4268 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4269 		log.u_bbr.flex5 = rtt;
4270 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4271 		log.u_bbr.flex6 <<= 1;
4272 		log.u_bbr.flex6 |= rack->forced_ack;
4273 		log.u_bbr.flex6 <<= 1;
4274 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4275 		log.u_bbr.flex6 <<= 1;
4276 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4277 		log.u_bbr.flex6 <<= 1;
4278 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4279 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4280 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4281 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4282 		log.u_bbr.flex8 = reas;
4283 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4284 		log.u_bbr.delRate = rack_get_bw(rack);
4285 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4286 		log.u_bbr.cur_del_rate <<= 32;
4287 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4288 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4289 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4290 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4291 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4292 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4293 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4294 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4295 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4296 		log.u_bbr.rttProp = us_cts;
4297 		log.u_bbr.rttProp <<= 32;
4298 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4299 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4300 		    &rack->rc_inp->inp_socket->so_rcv,
4301 		    &rack->rc_inp->inp_socket->so_snd,
4302 		    BBR_LOG_RTT_SHRINKS, 0,
4303 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4304 	}
4305 }
4306 
4307 static void
4308 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4309 {
4310 	uint64_t bwdp;
4311 
4312 	bwdp = rack_get_bw(rack);
4313 	bwdp *= (uint64_t)rtt;
4314 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4315 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4316 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4317 		/*
4318 		 * A window protocol must be able to have 4 packets
4319 		 * outstanding as the floor in order to function
4320 		 * (especially considering delayed ack :D).
4321 		 */
4322 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4323 	}
4324 }
4325 
4326 static void
4327 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4328 {
4329 	/**
4330 	 * ProbeRTT is a bit different in rack_pacing than in
4331 	 * BBR. It is like BBR in that it uses the lowering of
4332 	 * the RTT as a signal that we saw something new and
4333 	 * counts from there for how long between. But it is
4334 	 * different in that its quite simple. It does not
4335 	 * play with the cwnd and wait until we get down
4336 	 * to N segments outstanding and hold that for
4337 	 * 200ms. Instead it just sets the pacing reduction
4338 	 * rate to a set percentage (70 by default) and hold
4339 	 * that for a number of recent GP Srtt's.
4340 	 */
4341 	uint32_t segsiz;
4342 
4343 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4344 	if (rack->rc_gp_dyn_mul == 0)
4345 		return;
4346 
4347 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4348 		/* We are idle */
4349 		return;
4350 	}
4351 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4352 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4353 		/*
4354 		 * Stop the goodput now, the idea here is
4355 		 * that future measurements with in_probe_rtt
4356 		 * won't register if they are not greater so
4357 		 * we want to get what info (if any) is available
4358 		 * now.
4359 		 */
4360 		rack_do_goodput_measurement(rack->rc_tp, rack,
4361 					    rack->rc_tp->snd_una, __LINE__,
4362 					    RACK_QUALITY_PROBERTT);
4363 	}
4364 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4365 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4366 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4367 		     rack->r_ctl.rc_pace_min_segs);
4368 	rack->in_probe_rtt = 1;
4369 	rack->measure_saw_probe_rtt = 1;
4370 	rack->r_ctl.rc_time_probertt_starts = 0;
4371 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4372 	if (rack_probertt_use_min_rtt_entry)
4373 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4374 	else
4375 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4376 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4377 			     __LINE__, RACK_RTTS_ENTERPROBE);
4378 }
4379 
4380 static void
4381 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4382 {
4383 	struct rack_sendmap *rsm;
4384 	uint32_t segsiz;
4385 
4386 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4387 		     rack->r_ctl.rc_pace_min_segs);
4388 	rack->in_probe_rtt = 0;
4389 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4390 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4391 		/*
4392 		 * Stop the goodput now, the idea here is
4393 		 * that future measurements with in_probe_rtt
4394 		 * won't register if they are not greater so
4395 		 * we want to get what info (if any) is available
4396 		 * now.
4397 		 */
4398 		rack_do_goodput_measurement(rack->rc_tp, rack,
4399 					    rack->rc_tp->snd_una, __LINE__,
4400 					    RACK_QUALITY_PROBERTT);
4401 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4402 		/*
4403 		 * We don't have enough data to make a measurement.
4404 		 * So lets just stop and start here after exiting
4405 		 * probe-rtt. We probably are not interested in
4406 		 * the results anyway.
4407 		 */
4408 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4409 	}
4410 	/*
4411 	 * Measurements through the current snd_max are going
4412 	 * to be limited by the slower pacing rate.
4413 	 *
4414 	 * We need to mark these as app-limited so we
4415 	 * don't collapse the b/w.
4416 	 */
4417 	rsm = tqhash_max(rack->r_ctl.tqh);
4418 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4419 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4420 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4421 		else {
4422 			/*
4423 			 * Go out to the end app limited and mark
4424 			 * this new one as next and move the end_appl up
4425 			 * to this guy.
4426 			 */
4427 			if (rack->r_ctl.rc_end_appl)
4428 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4429 			rack->r_ctl.rc_end_appl = rsm;
4430 		}
4431 		rsm->r_flags |= RACK_APP_LIMITED;
4432 		rack->r_ctl.rc_app_limited_cnt++;
4433 	}
4434 	/*
4435 	 * Now, we need to examine our pacing rate multipliers.
4436 	 * If its under 100%, we need to kick it back up to
4437 	 * 100%. We also don't let it be over our "max" above
4438 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4439 	 * Note setting clamp_atexit_prtt to 0 has the effect
4440 	 * of setting CA/SS to 100% always at exit (which is
4441 	 * the default behavior).
4442 	 */
4443 	if (rack_probertt_clear_is) {
4444 		rack->rc_gp_incr = 0;
4445 		rack->rc_gp_bwred = 0;
4446 		rack->rc_gp_timely_inc_cnt = 0;
4447 		rack->rc_gp_timely_dec_cnt = 0;
4448 	}
4449 	/* Do we do any clamping at exit? */
4450 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4451 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4452 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4453 	}
4454 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4455 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4456 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4457 	}
4458 	/*
4459 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4460 	 * after exiting.
4461 	 */
4462 	rack->r_ctl.rc_rtt_diff = 0;
4463 
4464 	/* Clear all flags so we start fresh */
4465 	rack->rc_tp->t_bytes_acked = 0;
4466 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4467 	/*
4468 	 * If configured to, set the cwnd and ssthresh to
4469 	 * our targets.
4470 	 */
4471 	if (rack_probe_rtt_sets_cwnd) {
4472 		uint64_t ebdp;
4473 		uint32_t setto;
4474 
4475 		/* Set ssthresh so we get into CA once we hit our target */
4476 		if (rack_probertt_use_min_rtt_exit == 1) {
4477 			/* Set to min rtt */
4478 			rack_set_prtt_target(rack, segsiz,
4479 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4480 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4481 			/* Set to current gp rtt */
4482 			rack_set_prtt_target(rack, segsiz,
4483 					     rack->r_ctl.rc_gp_srtt);
4484 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4485 			/* Set to entry gp rtt */
4486 			rack_set_prtt_target(rack, segsiz,
4487 					     rack->r_ctl.rc_entry_gp_rtt);
4488 		} else {
4489 			uint64_t sum;
4490 			uint32_t setval;
4491 
4492 			sum = rack->r_ctl.rc_entry_gp_rtt;
4493 			sum *= 10;
4494 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4495 			if (sum >= 20) {
4496 				/*
4497 				 * A highly buffered path needs
4498 				 * cwnd space for timely to work.
4499 				 * Lets set things up as if
4500 				 * we are heading back here again.
4501 				 */
4502 				setval = rack->r_ctl.rc_entry_gp_rtt;
4503 			} else if (sum >= 15) {
4504 				/*
4505 				 * Lets take the smaller of the
4506 				 * two since we are just somewhat
4507 				 * buffered.
4508 				 */
4509 				setval = rack->r_ctl.rc_gp_srtt;
4510 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4511 					setval = rack->r_ctl.rc_entry_gp_rtt;
4512 			} else {
4513 				/*
4514 				 * Here we are not highly buffered
4515 				 * and should pick the min we can to
4516 				 * keep from causing loss.
4517 				 */
4518 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4519 			}
4520 			rack_set_prtt_target(rack, segsiz,
4521 					     setval);
4522 		}
4523 		if (rack_probe_rtt_sets_cwnd > 1) {
4524 			/* There is a percentage here to boost */
4525 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4526 			ebdp *= rack_probe_rtt_sets_cwnd;
4527 			ebdp /= 100;
4528 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4529 		} else
4530 			setto = rack->r_ctl.rc_target_probertt_flight;
4531 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4532 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4533 			/* Enforce a min */
4534 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4535 		}
4536 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4537 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4538 	}
4539 	rack_log_rtt_shrinks(rack,  us_cts,
4540 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4541 			     __LINE__, RACK_RTTS_EXITPROBE);
4542 	/* Clear times last so log has all the info */
4543 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4544 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4545 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4546 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4547 }
4548 
4549 static void
4550 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4551 {
4552 	/* Check in on probe-rtt */
4553 
4554 	if (rack->rc_gp_filled == 0) {
4555 		/* We do not do p-rtt unless we have gp measurements */
4556 		return;
4557 	}
4558 	if (rack->in_probe_rtt) {
4559 		uint64_t no_overflow;
4560 		uint32_t endtime, must_stay;
4561 
4562 		if (rack->r_ctl.rc_went_idle_time &&
4563 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4564 			/*
4565 			 * We went idle during prtt, just exit now.
4566 			 */
4567 			rack_exit_probertt(rack, us_cts);
4568 		} else if (rack_probe_rtt_safety_val &&
4569 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4570 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4571 			/*
4572 			 * Probe RTT safety value triggered!
4573 			 */
4574 			rack_log_rtt_shrinks(rack,  us_cts,
4575 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4576 					     __LINE__, RACK_RTTS_SAFETY);
4577 			rack_exit_probertt(rack, us_cts);
4578 		}
4579 		/* Calculate the max we will wait */
4580 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4581 		if (rack->rc_highly_buffered)
4582 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4583 		/* Calculate the min we must wait */
4584 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4585 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4586 		    TSTMP_LT(us_cts, endtime)) {
4587 			uint32_t calc;
4588 			/* Do we lower more? */
4589 no_exit:
4590 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4591 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4592 			else
4593 				calc = 0;
4594 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4595 			if (calc) {
4596 				/* Maybe */
4597 				calc *= rack_per_of_gp_probertt_reduce;
4598 				if (calc > rack_per_of_gp_probertt)
4599 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4600 				else
4601 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4602 				/* Limit it too */
4603 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4604 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4605 			}
4606 			/* We must reach target or the time set */
4607 			return;
4608 		}
4609 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4610 			if ((TSTMP_LT(us_cts, must_stay) &&
4611 			     rack->rc_highly_buffered) ||
4612 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4613 			      rack->r_ctl.rc_target_probertt_flight)) {
4614 				/* We are not past the must_stay time */
4615 				goto no_exit;
4616 			}
4617 			rack_log_rtt_shrinks(rack,  us_cts,
4618 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4619 					     __LINE__, RACK_RTTS_REACHTARGET);
4620 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4621 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4622 				rack->r_ctl.rc_time_probertt_starts = 1;
4623 			/* Restore back to our rate we want to pace at in prtt */
4624 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4625 		}
4626 		/*
4627 		 * Setup our end time, some number of gp_srtts plus 200ms.
4628 		 */
4629 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4630 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4631 		if (rack_probertt_gpsrtt_cnt_div)
4632 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4633 		else
4634 			endtime = 0;
4635 		endtime += rack_min_probertt_hold;
4636 		endtime += rack->r_ctl.rc_time_probertt_starts;
4637 		if (TSTMP_GEQ(us_cts,  endtime)) {
4638 			/* yes, exit probertt */
4639 			rack_exit_probertt(rack, us_cts);
4640 		}
4641 
4642 	} else if ((rack->rc_skip_timely == 0) &&
4643 		   (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4644 		   ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4645 		/* Go into probertt, its been too long since we went lower */
4646 		rack_enter_probertt(rack, us_cts);
4647 	}
4648 }
4649 
4650 static void
4651 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4652 		       uint32_t rtt, int32_t rtt_diff)
4653 {
4654 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4655 	uint32_t losses;
4656 
4657 	if ((rack->rc_gp_dyn_mul == 0) ||
4658 	    (rack->use_fixed_rate) ||
4659 	    (rack->in_probe_rtt) ||
4660 	    (rack->rc_always_pace == 0)) {
4661 		/* No dynamic GP multiplier in play */
4662 		return;
4663 	}
4664 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4665 	cur_bw = rack_get_bw(rack);
4666 	/* Calculate our up and down range */
4667 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4668 	up_bnd /= 100;
4669 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4670 
4671 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4672 	subfr /= 100;
4673 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4674 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4675 		/*
4676 		 * This is the case where our RTT is above
4677 		 * the max target and we have been configured
4678 		 * to just do timely no bonus up stuff in that case.
4679 		 *
4680 		 * There are two configurations, set to 1, and we
4681 		 * just do timely if we are over our max. If its
4682 		 * set above 1 then we slam the multipliers down
4683 		 * to 100 and then decrement per timely.
4684 		 */
4685 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4686 				__LINE__, 3);
4687 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4688 			rack_validate_multipliers_at_or_below_100(rack);
4689 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4690 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4691 		/*
4692 		 * We are decreasing this is a bit complicated this
4693 		 * means we are loosing ground. This could be
4694 		 * because another flow entered and we are competing
4695 		 * for b/w with it. This will push the RTT up which
4696 		 * makes timely unusable unless we want to get shoved
4697 		 * into a corner and just be backed off (the age
4698 		 * old problem with delay based CC).
4699 		 *
4700 		 * On the other hand if it was a route change we
4701 		 * would like to stay somewhat contained and not
4702 		 * blow out the buffers.
4703 		 */
4704 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4705 				__LINE__, 3);
4706 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4707 		if (rack->rc_gp_bwred == 0) {
4708 			/* Go into reduction counting */
4709 			rack->rc_gp_bwred = 1;
4710 			rack->rc_gp_timely_dec_cnt = 0;
4711 		}
4712 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4713 			/*
4714 			 * Push another time with a faster pacing
4715 			 * to try to gain back (we include override to
4716 			 * get a full raise factor).
4717 			 */
4718 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4719 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4720 			    (timely_says == 0) ||
4721 			    (rack_down_raise_thresh == 0)) {
4722 				/*
4723 				 * Do an override up in b/w if we were
4724 				 * below the threshold or if the threshold
4725 				 * is zero we always do the raise.
4726 				 */
4727 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4728 			} else {
4729 				/* Log it stays the same */
4730 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4731 						__LINE__, 11);
4732 			}
4733 			rack->rc_gp_timely_dec_cnt++;
4734 			/* We are not incrementing really no-count */
4735 			rack->rc_gp_incr = 0;
4736 			rack->rc_gp_timely_inc_cnt = 0;
4737 		} else {
4738 			/*
4739 			 * Lets just use the RTT
4740 			 * information and give up
4741 			 * pushing.
4742 			 */
4743 			goto use_timely;
4744 		}
4745 	} else if ((timely_says != 2) &&
4746 		    !losses &&
4747 		    (last_bw_est > up_bnd)) {
4748 		/*
4749 		 * We are increasing b/w lets keep going, updating
4750 		 * our b/w and ignoring any timely input, unless
4751 		 * of course we are at our max raise (if there is one).
4752 		 */
4753 
4754 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4755 				__LINE__, 3);
4756 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4757 		if (rack->rc_gp_saw_ss &&
4758 		    rack->r_ctl.rack_per_upper_bound_ss &&
4759 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4760 			    /*
4761 			     * In cases where we can't go higher
4762 			     * we should just use timely.
4763 			     */
4764 			    goto use_timely;
4765 		}
4766 		if (rack->rc_gp_saw_ca &&
4767 		    rack->r_ctl.rack_per_upper_bound_ca &&
4768 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4769 			    /*
4770 			     * In cases where we can't go higher
4771 			     * we should just use timely.
4772 			     */
4773 			    goto use_timely;
4774 		}
4775 		rack->rc_gp_bwred = 0;
4776 		rack->rc_gp_timely_dec_cnt = 0;
4777 		/* You get a set number of pushes if timely is trying to reduce */
4778 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4779 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4780 		} else {
4781 			/* Log it stays the same */
4782 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4783 			    __LINE__, 12);
4784 		}
4785 		return;
4786 	} else {
4787 		/*
4788 		 * We are staying between the lower and upper range bounds
4789 		 * so use timely to decide.
4790 		 */
4791 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4792 				__LINE__, 3);
4793 use_timely:
4794 		if (timely_says) {
4795 			rack->rc_gp_incr = 0;
4796 			rack->rc_gp_timely_inc_cnt = 0;
4797 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4798 			    !losses &&
4799 			    (last_bw_est < low_bnd)) {
4800 				/* We are loosing ground */
4801 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4802 				rack->rc_gp_timely_dec_cnt++;
4803 				/* We are not incrementing really no-count */
4804 				rack->rc_gp_incr = 0;
4805 				rack->rc_gp_timely_inc_cnt = 0;
4806 			} else
4807 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4808 		} else {
4809 			rack->rc_gp_bwred = 0;
4810 			rack->rc_gp_timely_dec_cnt = 0;
4811 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4812 		}
4813 	}
4814 }
4815 
4816 static int32_t
4817 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4818 {
4819 	int32_t timely_says;
4820 	uint64_t log_mult, log_rtt_a_diff;
4821 
4822 	log_rtt_a_diff = rtt;
4823 	log_rtt_a_diff <<= 32;
4824 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4825 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4826 		    rack_gp_rtt_maxmul)) {
4827 		/* Reduce the b/w multiplier */
4828 		timely_says = 2;
4829 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4830 		log_mult <<= 32;
4831 		log_mult |= prev_rtt;
4832 		rack_log_timely(rack,  timely_says, log_mult,
4833 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4834 				log_rtt_a_diff, __LINE__, 4);
4835 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4836 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4837 			    max(rack_gp_rtt_mindiv , 1)))) {
4838 		/* Increase the b/w multiplier */
4839 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4840 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4841 			 max(rack_gp_rtt_mindiv , 1));
4842 		log_mult <<= 32;
4843 		log_mult |= prev_rtt;
4844 		timely_says = 0;
4845 		rack_log_timely(rack,  timely_says, log_mult ,
4846 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4847 				log_rtt_a_diff, __LINE__, 5);
4848 	} else {
4849 		/*
4850 		 * Use a gradient to find it the timely gradient
4851 		 * is:
4852 		 * grad = rc_rtt_diff / min_rtt;
4853 		 *
4854 		 * anything below or equal to 0 will be
4855 		 * a increase indication. Anything above
4856 		 * zero is a decrease. Note we take care
4857 		 * of the actual gradient calculation
4858 		 * in the reduction (its not needed for
4859 		 * increase).
4860 		 */
4861 		log_mult = prev_rtt;
4862 		if (rtt_diff <= 0) {
4863 			/*
4864 			 * Rttdiff is less than zero, increase the
4865 			 * b/w multiplier (its 0 or negative)
4866 			 */
4867 			timely_says = 0;
4868 			rack_log_timely(rack,  timely_says, log_mult,
4869 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4870 		} else {
4871 			/* Reduce the b/w multiplier */
4872 			timely_says = 1;
4873 			rack_log_timely(rack,  timely_says, log_mult,
4874 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4875 		}
4876 	}
4877 	return (timely_says);
4878 }
4879 
4880 static __inline int
4881 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4882 {
4883 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4884 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4885 		/**
4886 		 * This covers the case that the
4887 		 * resent is completely inside
4888 		 * the gp range or up to it.
4889 		 *      |----------------|
4890 		 *      |-----| <or>
4891 		 *            |----|
4892 		 *            <or>   |---|
4893 		 */
4894 		return (1);
4895 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4896 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4897 		/**
4898 		 * This covers the case of
4899 		 *      |--------------|
4900 		 *  |-------->|
4901 		 */
4902 		return (1);
4903 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4904 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4905 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4906 
4907 		/**
4908 		 * This covers the case of
4909 		 *      |--------------|
4910 		 *              |-------->|
4911 		 */
4912 		return (1);
4913 	}
4914 	return (0);
4915 }
4916 
4917 static __inline void
4918 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4919 {
4920 
4921 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4922 		return;
4923 	/*
4924 	 * We have a Goodput measurement in progress. Mark
4925 	 * the send if its within the window. If its not
4926 	 * in the window make sure it does not have the mark.
4927 	 */
4928 	if (rack_in_gp_window(tp, rsm))
4929 		rsm->r_flags |= RACK_IN_GP_WIN;
4930 	else
4931 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4932 }
4933 
4934 static __inline void
4935 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4936 {
4937 	/* A GP measurement is ending, clear all marks on the send map*/
4938 	struct rack_sendmap *rsm = NULL;
4939 
4940 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4941 	if (rsm == NULL) {
4942 		rsm = tqhash_min(rack->r_ctl.tqh);
4943 	}
4944 	/* Nothing left? */
4945 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4946 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4947 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4948 	}
4949 }
4950 
4951 
4952 static __inline void
4953 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4954 {
4955 	struct rack_sendmap *rsm = NULL;
4956 
4957 	if (tp->snd_una == tp->snd_max) {
4958 		/* Nothing outstanding yet, nothing to do here */
4959 		return;
4960 	}
4961 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4962 		/*
4963 		 * We are measuring ahead of some outstanding
4964 		 * data. We need to walk through up until we get
4965 		 * to gp_seq marking so that no rsm is set incorrectly
4966 		 * with RACK_IN_GP_WIN.
4967 		 */
4968 		rsm = tqhash_min(rack->r_ctl.tqh);
4969 		while (rsm != NULL) {
4970 			rack_mark_in_gp_win(tp, rsm);
4971 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4972 				break;
4973 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4974 		}
4975 	}
4976 	if (rsm == NULL) {
4977 		/*
4978 		 * Need to find the GP seq, if rsm is
4979 		 * set we stopped as we hit it.
4980 		 */
4981 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4982 		if (rsm == NULL)
4983 			return;
4984 		rack_mark_in_gp_win(tp, rsm);
4985 	}
4986 	/*
4987 	 * Now we may need to mark already sent rsm, ahead of
4988 	 * gput_seq in the window since they may have been sent
4989 	 * *before* we started our measurment. The rsm, if non-null
4990 	 * has been marked (note if rsm would have been NULL we would have
4991 	 * returned in the previous block). So we go to the next, and continue
4992 	 * until we run out of entries or we exceed the gp_ack value.
4993 	 */
4994 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4995 	while (rsm) {
4996 		rack_mark_in_gp_win(tp, rsm);
4997 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4998 			break;
4999 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
5000 	}
5001 }
5002 
5003 static void
5004 rack_log_gp_calc(struct tcp_rack *rack, uint32_t add_part, uint32_t sub_part, uint32_t srtt, uint64_t meas_bw, uint64_t utim, uint8_t meth, uint32_t line)
5005 {
5006 	if (tcp_bblogging_on(rack->rc_tp)) {
5007 		union tcp_log_stackspecific log;
5008 		struct timeval tv;
5009 
5010 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5011 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5012 		log.u_bbr.flex1 = add_part;
5013 		log.u_bbr.flex2 = sub_part;
5014 		log.u_bbr.flex3 = rack_wma_divisor;
5015 		log.u_bbr.flex4 = srtt;
5016 		log.u_bbr.flex7 = (uint16_t)line;
5017 		log.u_bbr.flex8 = meth;
5018 		log.u_bbr.delRate = rack->r_ctl.gp_bw;
5019 		log.u_bbr.cur_del_rate = meas_bw;
5020 		log.u_bbr.rttProp = utim;
5021 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5022 		    &rack->rc_inp->inp_socket->so_rcv,
5023 		    &rack->rc_inp->inp_socket->so_snd,
5024 		    BBR_LOG_THRESH_CALC, 0,
5025 		    0, &log, false, &rack->r_ctl.act_rcv_time);
5026 	}
5027 }
5028 
5029 static void
5030 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
5031 			    tcp_seq th_ack, int line, uint8_t quality)
5032 {
5033 	uint64_t tim, bytes_ps, stim, utim;
5034 	uint32_t segsiz, bytes, reqbytes, us_cts;
5035 	int32_t gput, new_rtt_diff, timely_says;
5036 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
5037 	int did_add = 0;
5038 
5039 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5040 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5041 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
5042 		tim = us_cts - tp->gput_ts;
5043 	else
5044 		tim = 0;
5045 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
5046 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
5047 	else
5048 		stim = 0;
5049 	/*
5050 	 * Use the larger of the send time or ack time. This prevents us
5051 	 * from being influenced by ack artifacts to come up with too
5052 	 * high of measurement. Note that since we are spanning over many more
5053 	 * bytes in most of our measurements hopefully that is less likely to
5054 	 * occur.
5055 	 */
5056 	if (tim > stim)
5057 		utim = max(tim, 1);
5058 	else
5059 		utim = max(stim, 1);
5060 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
5061 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
5062 	if ((tim == 0) && (stim == 0)) {
5063 		/*
5064 		 * Invalid measurement time, maybe
5065 		 * all on one ack/one send?
5066 		 */
5067 		bytes = 0;
5068 		bytes_ps = 0;
5069 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5070 					   0, 0, 0, 10, __LINE__, NULL, quality);
5071 		goto skip_measurement;
5072 	}
5073 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
5074 		/* We never made a us_rtt measurement? */
5075 		bytes = 0;
5076 		bytes_ps = 0;
5077 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5078 					   0, 0, 0, 10, __LINE__, NULL, quality);
5079 		goto skip_measurement;
5080 	}
5081 	/*
5082 	 * Calculate the maximum possible b/w this connection
5083 	 * could have. We base our calculation on the lowest
5084 	 * rtt we have seen during the measurement and the
5085 	 * largest rwnd the client has given us in that time. This
5086 	 * forms a BDP that is the maximum that we could ever
5087 	 * get to the client. Anything larger is not valid.
5088 	 *
5089 	 * I originally had code here that rejected measurements
5090 	 * where the time was less than 1/2 the latest us_rtt.
5091 	 * But after thinking on that I realized its wrong since
5092 	 * say you had a 150Mbps or even 1Gbps link, and you
5093 	 * were a long way away.. example I am in Europe (100ms rtt)
5094 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
5095 	 * bytes my time would be 1.2ms, and yet my rtt would say
5096 	 * the measurement was invalid the time was < 50ms. The
5097 	 * same thing is true for 150Mb (8ms of time).
5098 	 *
5099 	 * A better way I realized is to look at what the maximum
5100 	 * the connection could possibly do. This is gated on
5101 	 * the lowest RTT we have seen and the highest rwnd.
5102 	 * We should in theory never exceed that, if we are
5103 	 * then something on the path is storing up packets
5104 	 * and then feeding them all at once to our endpoint
5105 	 * messing up our measurement.
5106 	 */
5107 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
5108 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
5109 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
5110 	if (SEQ_LT(th_ack, tp->gput_seq)) {
5111 		/* No measurement can be made */
5112 		bytes = 0;
5113 		bytes_ps = 0;
5114 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5115 					   0, 0, 0, 10, __LINE__, NULL, quality);
5116 		goto skip_measurement;
5117 	} else
5118 		bytes = (th_ack - tp->gput_seq);
5119 	bytes_ps = (uint64_t)bytes;
5120 	/*
5121 	 * Don't measure a b/w for pacing unless we have gotten at least
5122 	 * an initial windows worth of data in this measurement interval.
5123 	 *
5124 	 * Small numbers of bytes get badly influenced by delayed ack and
5125 	 * other artifacts. Note we take the initial window or our
5126 	 * defined minimum GP (defaulting to 10 which hopefully is the
5127 	 * IW).
5128 	 */
5129 	if (rack->rc_gp_filled == 0) {
5130 		/*
5131 		 * The initial estimate is special. We
5132 		 * have blasted out an IW worth of packets
5133 		 * without a real valid ack ts results. We
5134 		 * then setup the app_limited_needs_set flag,
5135 		 * this should get the first ack in (probably 2
5136 		 * MSS worth) to be recorded as the timestamp.
5137 		 * We thus allow a smaller number of bytes i.e.
5138 		 * IW - 2MSS.
5139 		 */
5140 		reqbytes -= (2 * segsiz);
5141 		/* Also lets fill previous for our first measurement to be neutral */
5142 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5143 	}
5144 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
5145 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5146 					   rack->r_ctl.rc_app_limited_cnt,
5147 					   0, 0, 10, __LINE__, NULL, quality);
5148 		goto skip_measurement;
5149 	}
5150 	/*
5151 	 * We now need to calculate the Timely like status so
5152 	 * we can update (possibly) the b/w multipliers.
5153 	 */
5154 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5155 	if (rack->rc_gp_filled == 0) {
5156 		/* No previous reading */
5157 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5158 	} else {
5159 		if (rack->measure_saw_probe_rtt == 0) {
5160 			/*
5161 			 * We don't want a probertt to be counted
5162 			 * since it will be negative incorrectly. We
5163 			 * expect to be reducing the RTT when we
5164 			 * pace at a slower rate.
5165 			 */
5166 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5167 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5168 		}
5169 	}
5170 	timely_says = rack_make_timely_judgement(rack,
5171 	    rack->r_ctl.rc_gp_srtt,
5172 	    rack->r_ctl.rc_rtt_diff,
5173 	    rack->r_ctl.rc_prev_gp_srtt
5174 	);
5175 	bytes_ps *= HPTS_USEC_IN_SEC;
5176 	bytes_ps /= utim;
5177 	if (bytes_ps > rack->r_ctl.last_max_bw) {
5178 		/*
5179 		 * Something is on path playing
5180 		 * since this b/w is not possible based
5181 		 * on our BDP (highest rwnd and lowest rtt
5182 		 * we saw in the measurement window).
5183 		 *
5184 		 * Another option here would be to
5185 		 * instead skip the measurement.
5186 		 */
5187 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5188 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
5189 					   11, __LINE__, NULL, quality);
5190 		bytes_ps = rack->r_ctl.last_max_bw;
5191 	}
5192 	/* We store gp for b/w in bytes per second */
5193 	if (rack->rc_gp_filled == 0) {
5194 		/* Initial measurement */
5195 		if (bytes_ps) {
5196 			rack->r_ctl.gp_bw = bytes_ps;
5197 			rack->rc_gp_filled = 1;
5198 			rack->r_ctl.num_measurements = 1;
5199 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5200 		} else {
5201 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5202 						   rack->r_ctl.rc_app_limited_cnt,
5203 						   0, 0, 10, __LINE__, NULL, quality);
5204 		}
5205 		if (tcp_in_hpts(rack->rc_tp) &&
5206 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5207 			/*
5208 			 * Ok we can't trust the pacer in this case
5209 			 * where we transition from un-paced to paced.
5210 			 * Or for that matter when the burst mitigation
5211 			 * was making a wild guess and got it wrong.
5212 			 * Stop the pacer and clear up all the aggregate
5213 			 * delays etc.
5214 			 */
5215 			tcp_hpts_remove(rack->rc_tp);
5216 			rack->r_ctl.rc_hpts_flags = 0;
5217 			rack->r_ctl.rc_last_output_to = 0;
5218 		}
5219 		did_add = 2;
5220 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5221 		/* Still a small number run an average */
5222 		rack->r_ctl.gp_bw += bytes_ps;
5223 		addpart = rack->r_ctl.num_measurements;
5224 		rack->r_ctl.num_measurements++;
5225 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5226 			/* We have collected enough to move forward */
5227 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5228 		}
5229 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5230 		did_add = 3;
5231 	} else {
5232 		/*
5233 		 * We want to take 1/wma of the goodput and add in to 7/8th
5234 		 * of the old value weighted by the srtt. So if your measurement
5235 		 * period is say 2 SRTT's long you would get 1/4 as the
5236 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5237 		 *
5238 		 * But we must be careful not to take too much i.e. if the
5239 		 * srtt is say 20ms and the measurement is taken over
5240 		 * 400ms our weight would be 400/20 i.e. 20. On the
5241 		 * other hand if we get a measurement over 1ms with a
5242 		 * 10ms rtt we only want to take a much smaller portion.
5243 		 */
5244 		uint8_t meth;
5245 
5246 		if (rack->r_ctl.num_measurements < 0xff) {
5247 			rack->r_ctl.num_measurements++;
5248 		}
5249 		srtt = (uint64_t)tp->t_srtt;
5250 		if (srtt == 0) {
5251 			/*
5252 			 * Strange why did t_srtt go back to zero?
5253 			 */
5254 			if (rack->r_ctl.rc_rack_min_rtt)
5255 				srtt = rack->r_ctl.rc_rack_min_rtt;
5256 			else
5257 				srtt = HPTS_USEC_IN_MSEC;
5258 		}
5259 		/*
5260 		 * XXXrrs: Note for reviewers, in playing with
5261 		 * dynamic pacing I discovered this GP calculation
5262 		 * as done originally leads to some undesired results.
5263 		 * Basically you can get longer measurements contributing
5264 		 * too much to the WMA. Thus I changed it if you are doing
5265 		 * dynamic adjustments to only do the aportioned adjustment
5266 		 * if we have a very small (time wise) measurement. Longer
5267 		 * measurements just get there weight (defaulting to 1/8)
5268 		 * add to the WMA. We may want to think about changing
5269 		 * this to always do that for both sides i.e. dynamic
5270 		 * and non-dynamic... but considering lots of folks
5271 		 * were playing with this I did not want to change the
5272 		 * calculation per.se. without your thoughts.. Lawerence?
5273 		 * Peter??
5274 		 */
5275 		if (rack->rc_gp_dyn_mul == 0) {
5276 			subpart = rack->r_ctl.gp_bw * utim;
5277 			subpart /= (srtt * 8);
5278 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5279 				/*
5280 				 * The b/w update takes no more
5281 				 * away then 1/2 our running total
5282 				 * so factor it in.
5283 				 */
5284 				addpart = bytes_ps * utim;
5285 				addpart /= (srtt * 8);
5286 				meth = 1;
5287 			} else {
5288 				/*
5289 				 * Don't allow a single measurement
5290 				 * to account for more than 1/2 of the
5291 				 * WMA. This could happen on a retransmission
5292 				 * where utim becomes huge compared to
5293 				 * srtt (multiple retransmissions when using
5294 				 * the sending rate which factors in all the
5295 				 * transmissions from the first one).
5296 				 */
5297 				subpart = rack->r_ctl.gp_bw / 2;
5298 				addpart = bytes_ps / 2;
5299 				meth = 2;
5300 			}
5301 			rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5302 			resid_bw = rack->r_ctl.gp_bw - subpart;
5303 			rack->r_ctl.gp_bw = resid_bw + addpart;
5304 			did_add = 1;
5305 		} else {
5306 			if ((utim / srtt) <= 1) {
5307 				/*
5308 				 * The b/w update was over a small period
5309 				 * of time. The idea here is to prevent a small
5310 				 * measurement time period from counting
5311 				 * too much. So we scale it based on the
5312 				 * time so it attributes less than 1/rack_wma_divisor
5313 				 * of its measurement.
5314 				 */
5315 				subpart = rack->r_ctl.gp_bw * utim;
5316 				subpart /= (srtt * rack_wma_divisor);
5317 				addpart = bytes_ps * utim;
5318 				addpart /= (srtt * rack_wma_divisor);
5319 				meth = 3;
5320 			} else {
5321 				/*
5322 				 * The scaled measurement was long
5323 				 * enough so lets just add in the
5324 				 * portion of the measurement i.e. 1/rack_wma_divisor
5325 				 */
5326 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5327 				addpart = bytes_ps / rack_wma_divisor;
5328 				meth = 4;
5329 			}
5330 			if ((rack->measure_saw_probe_rtt == 0) ||
5331 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5332 				/*
5333 				 * For probe-rtt we only add it in
5334 				 * if its larger, all others we just
5335 				 * add in.
5336 				 */
5337 				did_add = 1;
5338 				rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5339 				resid_bw = rack->r_ctl.gp_bw - subpart;
5340 				rack->r_ctl.gp_bw = resid_bw + addpart;
5341 			}
5342 		}
5343 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5344 	}
5345 	/*
5346 	 * We only watch the growth of the GP during the initial startup
5347 	 * or first-slowstart that ensues. If we ever needed to watch
5348 	 * growth of gp outside of that period all we need to do is
5349 	 * remove the first clause of this if (rc_initial_ss_comp).
5350 	 */
5351 	if ((rack->rc_initial_ss_comp == 0) &&
5352 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5353 		uint64_t gp_est;
5354 
5355 		gp_est = bytes_ps;
5356 		if (tcp_bblogging_on(rack->rc_tp)) {
5357 			union tcp_log_stackspecific log;
5358 			struct timeval tv;
5359 
5360 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5361 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5362 			log.u_bbr.flex1 = rack->r_ctl.current_round;
5363 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5364 			log.u_bbr.delRate = gp_est;
5365 			log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5366 			log.u_bbr.flex8 = 41;
5367 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5368 					    0, &log, false, NULL, __func__, __LINE__,&tv);
5369 		}
5370 		if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5371 		    (rack->r_ctl.last_gpest == 0)) {
5372 			/*
5373 			 * The round we get our measurement averaging going
5374 			 * is the base round so it always is the source point
5375 			 * for when we had our first increment. From there on
5376 			 * we only record the round that had a rise.
5377 			 */
5378 			rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5379 			rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5380 		} else if (gp_est >= rack->r_ctl.last_gpest) {
5381 			/*
5382 			 * Test to see if its gone up enough
5383 			 * to set the round count up to now. Note
5384 			 * that on the seeding of the 4th measurement we
5385 			 */
5386 			gp_est *= 1000;
5387 			gp_est /= rack->r_ctl.last_gpest;
5388 			if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5389 				/*
5390 				 * We went up enough to record the round.
5391 				 */
5392 				if (tcp_bblogging_on(rack->rc_tp)) {
5393 					union tcp_log_stackspecific log;
5394 					struct timeval tv;
5395 
5396 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5397 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5398 					log.u_bbr.flex1 = rack->r_ctl.current_round;
5399 					log.u_bbr.flex2 = (uint32_t)gp_est;
5400 					log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5401 					log.u_bbr.delRate = gp_est;
5402 					log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5403 					log.u_bbr.flex8 = 42;
5404 					(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5405 							    0, &log, false, NULL, __func__, __LINE__,&tv);
5406 				}
5407 				rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5408 				if (rack->r_ctl.use_gp_not_last == 1)
5409 					rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5410 				else
5411 					rack->r_ctl.last_gpest = bytes_ps;
5412 			}
5413 		}
5414 	}
5415 	if ((rack->gp_ready == 0) &&
5416 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5417 		/* We have enough measurements now */
5418 		rack->gp_ready = 1;
5419 		if (rack->dgp_on ||
5420 		    rack->rack_hibeta)
5421 			rack_set_cc_pacing(rack);
5422 		if (rack->defer_options)
5423 			rack_apply_deferred_options(rack);
5424 	}
5425 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5426 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5427 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5428 
5429 	if ((rack->measure_saw_probe_rtt == 0) &&
5430 	    rack->rc_gp_rtt_set) {
5431 		if (rack->rc_skip_timely == 0) {
5432 			rack_update_multiplier(rack, timely_says, bytes_ps,
5433 					       rack->r_ctl.rc_gp_srtt,
5434 					       rack->r_ctl.rc_rtt_diff);
5435 		}
5436 	}
5437 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5438 				   rack_get_bw(rack), 3, line, NULL, quality);
5439 	rack_log_pacing_delay_calc(rack,
5440 				   bytes, /* flex2 */
5441 				   tim, /* flex1 */
5442 				   bytes_ps, /* bw_inuse */
5443 				   rack->r_ctl.gp_bw, /* delRate */
5444 				   rack_get_lt_bw(rack), /* rttProp */
5445 				   20, line, NULL, 0);
5446 	/* reset the gp srtt and setup the new prev */
5447 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5448 	/* Record the lost count for the next measurement */
5449 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5450 skip_measurement:
5451 	/*
5452 	 * We restart our diffs based on the gpsrtt in the
5453 	 * measurement window.
5454 	 */
5455 	rack->rc_gp_rtt_set = 0;
5456 	rack->rc_gp_saw_rec = 0;
5457 	rack->rc_gp_saw_ca = 0;
5458 	rack->rc_gp_saw_ss = 0;
5459 	rack->rc_dragged_bottom = 0;
5460 	if (quality == RACK_QUALITY_HIGH) {
5461 		/*
5462 		 * Gput in the stats world is in kbps where bytes_ps is
5463 		 * bytes per second so we do ((x * 8)/ 1000).
5464 		 */
5465 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5466 #ifdef STATS
5467 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5468 					 gput);
5469 		/*
5470 		 * XXXLAS: This is a temporary hack, and should be
5471 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5472 		 * API to deal with chained VOIs.
5473 		 */
5474 		if (tp->t_stats_gput_prev > 0)
5475 			stats_voi_update_abs_s32(tp->t_stats,
5476 						 VOI_TCP_GPUT_ND,
5477 						 ((gput - tp->t_stats_gput_prev) * 100) /
5478 						 tp->t_stats_gput_prev);
5479 #endif
5480 		tp->t_stats_gput_prev = gput;
5481 	}
5482 	tp->t_flags &= ~TF_GPUTINPROG;
5483 	/*
5484 	 * Now are we app limited now and there is space from where we
5485 	 * were to where we want to go?
5486 	 *
5487 	 * We don't do the other case i.e. non-applimited here since
5488 	 * the next send will trigger us picking up the missing data.
5489 	 */
5490 	if (rack->r_ctl.rc_first_appl &&
5491 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5492 	    rack->r_ctl.rc_app_limited_cnt &&
5493 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5494 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5495 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5496 		/*
5497 		 * Yep there is enough outstanding to make a measurement here.
5498 		 */
5499 		struct rack_sendmap *rsm;
5500 
5501 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5502 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5503 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5504 		rack->app_limited_needs_set = 0;
5505 		tp->gput_seq = th_ack;
5506 		if (rack->in_probe_rtt)
5507 			rack->measure_saw_probe_rtt = 1;
5508 		else if ((rack->measure_saw_probe_rtt) &&
5509 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5510 			rack->measure_saw_probe_rtt = 0;
5511 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5512 			/* There is a full window to gain info from */
5513 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5514 		} else {
5515 			/* We can only measure up to the applimited point */
5516 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5517 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5518 				/*
5519 				 * We don't have enough to make a measurement.
5520 				 */
5521 				tp->t_flags &= ~TF_GPUTINPROG;
5522 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5523 							   0, 0, 0, 6, __LINE__, NULL, quality);
5524 				return;
5525 			}
5526 		}
5527 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5528 			/*
5529 			 * We will get no more data into the SB
5530 			 * this means we need to have the data available
5531 			 * before we start a measurement.
5532 			 */
5533 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5534 				/* Nope not enough data. */
5535 				return;
5536 			}
5537 		}
5538 		tp->t_flags |= TF_GPUTINPROG;
5539 		/*
5540 		 * Now we need to find the timestamp of the send at tp->gput_seq
5541 		 * for the send based measurement.
5542 		 */
5543 		rack->r_ctl.rc_gp_cumack_ts = 0;
5544 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5545 		if (rsm) {
5546 			/* Ok send-based limit is set */
5547 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5548 				/*
5549 				 * Move back to include the earlier part
5550 				 * so our ack time lines up right (this may
5551 				 * make an overlapping measurement but thats
5552 				 * ok).
5553 				 */
5554 				tp->gput_seq = rsm->r_start;
5555 			}
5556 			if (rsm->r_flags & RACK_ACKED) {
5557 				struct rack_sendmap *nrsm;
5558 
5559 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5560 				tp->gput_seq = rsm->r_end;
5561 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5562 				if (nrsm)
5563 					rsm = nrsm;
5564 				else {
5565 					rack->app_limited_needs_set = 1;
5566 				}
5567 			} else
5568 				rack->app_limited_needs_set = 1;
5569 			/* We always go from the first send */
5570 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5571 		} else {
5572 			/*
5573 			 * If we don't find the rsm due to some
5574 			 * send-limit set the current time, which
5575 			 * basically disables the send-limit.
5576 			 */
5577 			struct timeval tv;
5578 
5579 			microuptime(&tv);
5580 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5581 		}
5582 		rack_tend_gp_marks(tp, rack);
5583 		rack_log_pacing_delay_calc(rack,
5584 					   tp->gput_seq,
5585 					   tp->gput_ack,
5586 					   (uint64_t)rsm,
5587 					   tp->gput_ts,
5588 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5589 					   9,
5590 					   __LINE__, rsm, quality);
5591 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5592 	} else {
5593 		/*
5594 		 * To make sure proper timestamp merging occurs, we need to clear
5595 		 * all GP marks if we don't start a measurement.
5596 		 */
5597 		rack_clear_gp_marks(tp, rack);
5598 	}
5599 }
5600 
5601 /*
5602  * CC wrapper hook functions
5603  */
5604 static void
5605 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5606     uint16_t type, int32_t post_recovery)
5607 {
5608 	uint32_t prior_cwnd, acked;
5609 	struct tcp_log_buffer *lgb = NULL;
5610 	uint8_t labc_to_use, quality;
5611 
5612 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5613 	tp->t_ccv.nsegs = nsegs;
5614 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5615 	if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5616 		uint32_t max;
5617 
5618 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5619 		if (tp->t_ccv.bytes_this_ack > max) {
5620 			tp->t_ccv.bytes_this_ack = max;
5621 		}
5622 	}
5623 #ifdef STATS
5624 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5625 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5626 #endif
5627 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5628 		/*
5629 		 * We will ack all the data, time to end any
5630 		 * lt_bw_up we have running until something
5631 		 * new is sent. Note we need to use the actual
5632 		 * ack_rcv_time which with pacing may be different.
5633 		 */
5634 		uint64_t tmark;
5635 
5636 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5637 		rack->r_ctl.lt_seq = tp->snd_max;
5638 		tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
5639 		if (tmark >= rack->r_ctl.lt_timemark) {
5640 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5641 		}
5642 		rack->r_ctl.lt_timemark = tmark;
5643 		rack->lt_bw_up = 0;
5644 	}
5645 	quality = RACK_QUALITY_NONE;
5646 	if ((tp->t_flags & TF_GPUTINPROG) &&
5647 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5648 		/* Measure the Goodput */
5649 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5650 	}
5651 	/* Which way our we limited, if not cwnd limited no advance in CA */
5652 	if (tp->snd_cwnd <= tp->snd_wnd)
5653 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5654 	else
5655 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5656 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5657 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5658 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5659 		/* For the setting of a window past use the actual scwnd we are using */
5660 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5661 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5662 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5663 		}
5664 	} else {
5665 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5666 		tp->t_bytes_acked = 0;
5667 	}
5668 	prior_cwnd = tp->snd_cwnd;
5669 	if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5670 	    (rack_client_low_buf && rack->client_bufferlvl &&
5671 	    (rack->client_bufferlvl < rack_client_low_buf)))
5672 		labc_to_use = rack->rc_labc;
5673 	else
5674 		labc_to_use = rack_max_abc_post_recovery;
5675 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5676 		union tcp_log_stackspecific log;
5677 		struct timeval tv;
5678 
5679 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5680 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5681 		log.u_bbr.flex1 = th_ack;
5682 		log.u_bbr.flex2 = tp->t_ccv.flags;
5683 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5684 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5685 		log.u_bbr.flex5 = labc_to_use;
5686 		log.u_bbr.flex6 = prior_cwnd;
5687 		log.u_bbr.flex7 = V_tcp_do_newsack;
5688 		log.u_bbr.flex8 = 1;
5689 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5690 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5691 	}
5692 	if (CC_ALGO(tp)->ack_received != NULL) {
5693 		/* XXXLAS: Find a way to live without this */
5694 		tp->t_ccv.curack = th_ack;
5695 		tp->t_ccv.labc = labc_to_use;
5696 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5697 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5698 	}
5699 	if (lgb) {
5700 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5701 	}
5702 	if (rack->r_must_retran) {
5703 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5704 			/*
5705 			 * We now are beyond the rxt point so lets disable
5706 			 * the flag.
5707 			 */
5708 			rack->r_ctl.rc_out_at_rto = 0;
5709 			rack->r_must_retran = 0;
5710 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5711 			/*
5712 			 * Only decrement the rc_out_at_rto if the cwnd advances
5713 			 * at least a whole segment. Otherwise next time the peer
5714 			 * acks, we won't be able to send this generaly happens
5715 			 * when we are in Congestion Avoidance.
5716 			 */
5717 			if (acked <= rack->r_ctl.rc_out_at_rto){
5718 				rack->r_ctl.rc_out_at_rto -= acked;
5719 			} else {
5720 				rack->r_ctl.rc_out_at_rto = 0;
5721 			}
5722 		}
5723 	}
5724 #ifdef STATS
5725 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5726 #endif
5727 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5728 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5729 	}
5730 	if ((rack->rc_initial_ss_comp == 0) &&
5731 	    (tp->snd_cwnd >= tp->snd_ssthresh)) {
5732 		/*
5733 		 * The cwnd has grown beyond ssthresh we have
5734 		 * entered ca and completed our first Slowstart.
5735 		 */
5736 		rack->rc_initial_ss_comp = 1;
5737 	}
5738 }
5739 
5740 static void
5741 tcp_rack_partialack(struct tcpcb *tp)
5742 {
5743 	struct tcp_rack *rack;
5744 
5745 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5746 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5747 	/*
5748 	 * If we are doing PRR and have enough
5749 	 * room to send <or> we are pacing and prr
5750 	 * is disabled we will want to see if we
5751 	 * can send data (by setting r_wanted_output to
5752 	 * true).
5753 	 */
5754 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5755 	    rack->rack_no_prr)
5756 		rack->r_wanted_output = 1;
5757 }
5758 
5759 static inline uint64_t
5760 rack_get_rxt_per(uint64_t snds,  uint64_t rxts)
5761 {
5762 	uint64_t rxt_per;
5763 
5764 	if (snds > 0) {
5765 		rxt_per = rxts * 1000;
5766 		rxt_per /= snds;
5767 	} else {
5768 		/* This is an unlikely path */
5769 		if (rxts) {
5770 			/* Its the max it was all re-transmits */
5771 			rxt_per = 0xffffffffffffffff;
5772 		} else {
5773 			rxt_per = 0;
5774 		}
5775 	}
5776 	return (rxt_per);
5777 }
5778 
5779 static void
5780 policer_detection_log(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint8_t flex8)
5781 {
5782 	if (tcp_bblogging_on(rack->rc_tp)) {
5783 		union tcp_log_stackspecific log;
5784 		struct timeval tv;
5785 
5786 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5787 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5788 		log.u_bbr.flex1 = flex1;
5789 		log.u_bbr.flex2 = flex2;
5790 		log.u_bbr.flex3 = flex3;
5791 		log.u_bbr.flex4 = flex4;
5792 		log.u_bbr.flex5 = rack->r_ctl.current_policer_bucket;
5793 		log.u_bbr.flex6 = rack->r_ctl.policer_bucket_size;
5794 		log.u_bbr.flex7 = 0;
5795 		log.u_bbr.flex8 = flex8;
5796 		log.u_bbr.bw_inuse = rack->r_ctl.policer_bw;
5797 		log.u_bbr.applimited = rack->r_ctl.current_round;
5798 		log.u_bbr.epoch = rack->r_ctl.policer_max_seg;
5799 		log.u_bbr.delivered = (uint32_t)rack->r_ctl.bytes_acked_in_recovery;
5800 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
5801 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
5802 		log.u_bbr.rttProp = rack->r_ctl.gp_bw;
5803 		log.u_bbr.bbr_state = rack->rc_policer_detected;
5804 		log.u_bbr.bbr_substate = 0;
5805 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5806 		log.u_bbr.use_lt_bw = rack->policer_detect_on;
5807 		log.u_bbr.lt_epoch = 0;
5808 		log.u_bbr.pkts_out = 0;
5809 		tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
5810 			      0, &log, false, NULL, NULL, 0, &tv);
5811 	}
5812 
5813 }
5814 
5815 static void
5816 policer_detection(struct tcpcb *tp, struct tcp_rack *rack, int post_recovery)
5817 {
5818 	/*
5819 	 * Rack excess rxt accounting is turned on. If we
5820 	 * are above a threshold of rxt's in at least N
5821 	 * rounds, then back off the cwnd and ssthresh
5822 	 * to fit into the long-term b/w.
5823 	 */
5824 
5825 	uint32_t pkts, mid, med, alt_med, avg, segsiz, tot_retran_pkt_count = 0;
5826 	uint32_t cnt_of_mape_rxt = 0;
5827 	uint64_t snds, rxts, rxt_per, tim, del, del_bw;
5828 	int i;
5829 	struct timeval tv;
5830 
5831 
5832 	/*
5833 	 * First is there enough packets delivered during recovery to make
5834 	 * a determiniation of b/w?
5835 	 */
5836 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5837 	if ((rack->rc_policer_detected == 0) &&
5838 	    (rack->r_ctl.policer_del_mss > 0) &&
5839 	    ((uint32_t)rack->r_ctl.policer_del_mss > ((rack->r_ctl.bytes_acked_in_recovery + segsiz - 1)/segsiz))) {
5840 		/*
5841 		 * Not enough data sent in recovery for initial detection. Once
5842 		 * we have deteced a policer we allow less than the threshold (polcer_del_mss)
5843 		 * amount of data in a recovery to let us fall through and double check
5844 		 * our policer settings and possibly expand or collapse the bucket size and
5845 		 * the polcier b/w.
5846 		 *
5847 		 * Once you are declared to be policed. this block of code cannot be
5848 		 * reached, instead blocks further down will re-check the policer detection
5849 		 * triggers and possibly reset the measurements if somehow we have let the
5850 		 * policer bucket size grow too large.
5851 		 */
5852 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5853 			policer_detection_log(rack, rack->r_ctl.policer_del_mss,
5854 					      ((rack->r_ctl.bytes_acked_in_recovery + segsiz - 1)/segsiz),
5855 					      rack->r_ctl.bytes_acked_in_recovery, segsiz, 18);
5856 		}
5857 		return;
5858 	}
5859 	tcp_get_usecs(&tv);
5860 	tim = tcp_tv_to_lusectick(&tv) - rack->r_ctl.time_entered_recovery;
5861 	del = rack->r_ctl.bytes_acked_in_recovery;
5862 	if (tim > 0)
5863 		del_bw = (del * (uint64_t)1000000) / tim;
5864 	else
5865 		del_bw = 0;
5866 	/* B/W compensation? */
5867 
5868 	if (rack->r_ctl.pol_bw_comp && ((rack->r_ctl.policer_bw > 0) ||
5869 					(del_bw > 0))) {
5870 		/*
5871 		 * Sanity check now that the data is in. How long does it
5872 		 * take for us to pace out two of our policer_max_seg's?
5873 		 *
5874 		 * If it is longer than the RTT then we are set
5875 		 * too slow, maybe because of not enough data
5876 		 * sent during recovery.
5877 		 */
5878 		uint64_t lentime, res, srtt, max_delbw, alt_bw;
5879 
5880 		srtt = (uint64_t)rack_grab_rtt(tp, rack);
5881 		if ((tp->t_srtt > 0) && (srtt > tp->t_srtt))
5882 			srtt = tp->t_srtt;
5883 		lentime = rack->r_ctl.policer_max_seg * (uint64_t)HPTS_USEC_IN_SEC * 2;
5884 		if (del_bw > rack->r_ctl.policer_bw) {
5885 			max_delbw = del_bw;
5886 		} else {
5887 			max_delbw = rack->r_ctl.policer_bw;
5888 		}
5889 		res = lentime / max_delbw;
5890 		if ((srtt > 0) && (res > srtt)) {
5891 			/*
5892 			 * At this rate we can not get two policer_maxsegs
5893 			 * out before the ack arrives back.
5894 			 *
5895 			 * Lets at least get it raised up so that
5896 			 * we can be a bit faster than that if possible.
5897 			 */
5898 			lentime = (rack->r_ctl.policer_max_seg * 2);
5899 			tim = srtt;
5900 			alt_bw = (lentime * (uint64_t)HPTS_USEC_IN_SEC) / tim;
5901 			if (alt_bw > max_delbw) {
5902 				uint64_t cap_alt_bw;
5903 
5904 				cap_alt_bw = (max_delbw + (max_delbw * rack->r_ctl.pol_bw_comp));
5905 				if ((rack_pol_min_bw > 0) && (cap_alt_bw < rack_pol_min_bw)) {
5906 					/* We place a min on the cap which defaults to 1Mbps */
5907 					cap_alt_bw = rack_pol_min_bw;
5908 				}
5909 				if (alt_bw <= cap_alt_bw) {
5910 					/* It should be */
5911 					del_bw = alt_bw;
5912 					policer_detection_log(rack,
5913 							      (uint32_t)tim,
5914 							      rack->r_ctl.policer_max_seg,
5915 							      0,
5916 							      0,
5917 							      16);
5918 				} else {
5919 					/*
5920 					 * This is an odd case where likely the RTT is very very
5921 					 * low. And yet it is still being policed. We don't want
5922 					 * to get more than (rack_policing_do_bw_comp+1) x del-rate
5923 					 * where del-rate is what we got in recovery for either the
5924 					 * first Policer Detection(PD) or this PD we are on now.
5925 					 */
5926 					del_bw = cap_alt_bw;
5927 					policer_detection_log(rack,
5928 							      (uint32_t)tim,
5929 							      rack->r_ctl.policer_max_seg,
5930 							      (uint32_t)max_delbw,
5931 							      (rack->r_ctl.pol_bw_comp + 1),
5932 							      16);
5933 				}
5934 			}
5935 		}
5936 	}
5937 	snds = tp->t_sndbytes - rack->r_ctl.last_policer_sndbytes;
5938 	rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_policer_snd_rxt_bytes;
5939 	rxt_per = rack_get_rxt_per(snds,  rxts);
5940 	/* Figure up the average  and median */
5941 	for(i = 0; i < RETRAN_CNT_SIZE; i++) {
5942 		if (rack->r_ctl.rc_cnt_of_retran[i] > 0) {
5943 			tot_retran_pkt_count += (i + 1) * rack->r_ctl.rc_cnt_of_retran[i];
5944 			cnt_of_mape_rxt  += rack->r_ctl.rc_cnt_of_retran[i];
5945 		}
5946 	}
5947 	if (cnt_of_mape_rxt)
5948 		avg = (tot_retran_pkt_count * 10)/cnt_of_mape_rxt;
5949 	else
5950 		avg = 0;
5951 	alt_med = med = 0;
5952 	mid = tot_retran_pkt_count/2;
5953 	for(i = 0; i < RETRAN_CNT_SIZE; i++) {
5954 		pkts = (i + 1) * rack->r_ctl.rc_cnt_of_retran[i];
5955 		if (mid > pkts) {
5956 			mid -= pkts;
5957 			continue;
5958 		}
5959 		med = (i + 1);
5960 		break;
5961 	}
5962 	mid = cnt_of_mape_rxt / 2;
5963 	for(i = 0; i < RETRAN_CNT_SIZE; i++) {
5964 		if (mid > rack->r_ctl.rc_cnt_of_retran[i]) {
5965 			mid -= rack->r_ctl.rc_cnt_of_retran[i];
5966 			continue;
5967 		}
5968 		alt_med = (i + 1);
5969 		break;
5970 	}
5971 	if (rack->r_ctl.policer_alt_median) {
5972 		/* Swap the medians */
5973 		uint32_t swap;
5974 
5975 		swap = med;
5976 		med = alt_med;
5977 		alt_med = swap;
5978 	}
5979 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5980 		union tcp_log_stackspecific log;
5981 		struct timeval tv;
5982 
5983 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5984 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5985 		log.u_bbr.flex1 = avg;
5986 		log.u_bbr.flex2 = med;
5987 		log.u_bbr.flex3 = (uint32_t)rxt_per;
5988 		log.u_bbr.flex4 = rack->r_ctl.policer_avg_threshold;
5989 		log.u_bbr.flex5 = rack->r_ctl.policer_med_threshold;
5990 		log.u_bbr.flex6 = rack->r_ctl.policer_rxt_threshold;
5991 		log.u_bbr.flex7 = rack->r_ctl.policer_alt_median;
5992 		log.u_bbr.flex8 = 1;
5993 		log.u_bbr.delivered = rack->r_ctl.policer_bucket_size;
5994 		log.u_bbr.applimited = rack->r_ctl.current_round;
5995 		log.u_bbr.epoch = rack->r_ctl.policer_max_seg;
5996 		log.u_bbr.bw_inuse = del_bw;
5997 		log.u_bbr.cur_del_rate = rxts;
5998 		log.u_bbr.delRate = snds;
5999 		log.u_bbr.rttProp = rack->r_ctl.gp_bw;
6000 		log.u_bbr.bbr_state = rack->rc_policer_detected;
6001 		log.u_bbr.bbr_substate = 0;
6002 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
6003 		log.u_bbr.use_lt_bw = rack->policer_detect_on;
6004 		log.u_bbr.lt_epoch = (uint32_t)tim;
6005 		log.u_bbr.pkts_out = rack->r_ctl.bytes_acked_in_recovery;
6006 		tcp_log_event(tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
6007 			      0, &log, false, NULL, NULL, 0, &tv);
6008 	}
6009 	if (med == RETRAN_CNT_SIZE) {
6010 		/*
6011 		 * If the median is the maximum, then what we
6012 		 * likely have here is a network breakage. Either that
6013 		 * or we are so unlucky that all of our traffic is being
6014 		 * dropped and having to be retransmitted the maximum times
6015 		 * and this just is not how a policer works.
6016 		 *
6017 		 * If it is truely a policer eventually we will come
6018 		 * through and it won't be the maximum.
6019 		 */
6020 		return;
6021 	}
6022 	/* Has enough rounds progressed for us to re-measure? */
6023 	if ((rxt_per >= (uint64_t)rack->r_ctl.policer_rxt_threshold) &&
6024 	    (avg >= rack->r_ctl.policer_avg_threshold) &&
6025 	    (med >= rack->r_ctl.policer_med_threshold)) {
6026 		/*
6027 		 * We hit all thresholds that indicate we are
6028 		 * being policed. Now we may be doing this from a rack timeout
6029 		 * which then means the rest of recovery will hopefully go
6030 		 * smoother as we pace. At the end of recovery we will
6031 		 * fall back in here and reset the values using the
6032 		 * results of the entire recovery episode (we could also
6033 		 * hit this as we exit recovery as well which means only
6034 		 * one time in here).
6035 		 *
6036 		 * This is done explicitly that if we hit the thresholds
6037 		 * again in a second recovery we overwrite the values. We do
6038 		 * that because over time, as we pace the policer_bucket_size may
6039 		 * continue to grow. This then provides more and more times when
6040 		 * we are not pacing to the policer rate. This lets us compensate
6041 		 * for when we hit a false positive and those flows continue to
6042 		 * increase. However if its a real policer we will then get over its
6043 		 * limit, over time, again and thus end up back here hitting the
6044 		 * thresholds again.
6045 		 *
6046 		 * The alternative to this is to instead whenever we pace due to
6047 		 * policing in rack_policed_sending we could add the amount len paced to the
6048 		 * idle_snd_una value (which decreases the amount in last_amount_before_rec
6049 		 * since that is always [th_ack - idle_snd_una]). This would then prevent
6050 		 * the polcier_bucket_size from growing in additional recovery episodes
6051 		 * Which would then mean false  postives would be pretty much stuck
6052 		 * after things got back to normal (assuming that what caused the
6053 		 * false positive was a small network outage).
6054 		 *
6055 		 */
6056 		tcp_trace_point(rack->rc_tp, TCP_TP_POLICER_DET);
6057 		if (rack->rc_policer_detected == 0) {
6058 			/*
6059 			 * Increment the stat that tells us we identified
6060 			 * a policer only once. Note that if we ever allow
6061 			 * the flag to be cleared (reverted) then we need
6062 			 * to adjust this to not do multi-counting.
6063 			 */
6064 			counter_u64_add(tcp_policer_detected, 1);
6065 		}
6066 		rack->r_ctl.last_policer_sndbytes = tp->t_sndbytes;
6067 		rack->r_ctl.last_policer_snd_rxt_bytes = tp->t_snd_rxt_bytes;
6068 		rack->r_ctl.policer_bw = del_bw;
6069 		rack->r_ctl.policer_max_seg = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp,
6070 										  rack->r_ctl.policer_bw,
6071 										  min(ctf_fixed_maxseg(rack->rc_tp),
6072 										      rack->r_ctl.rc_pace_min_segs),
6073 										  0, NULL,
6074 										  NULL, rack->r_ctl.pace_len_divisor);
6075 		/* Now what about the policer bucket size */
6076 		rack->r_ctl.policer_bucket_size = rack->r_ctl.last_amount_before_rec;
6077 		if (rack->r_ctl.policer_bucket_size < rack->r_ctl.policer_max_seg) {
6078 			/* We must be able to send our max-seg or else chaos ensues */
6079 			rack->r_ctl.policer_bucket_size = rack->r_ctl.policer_max_seg * 2;
6080 		}
6081 		if (rack->rc_policer_detected == 0)
6082 			rack->r_ctl.current_policer_bucket = 0;
6083 		if (tcp_bblogging_on(rack->rc_tp)) {
6084 			union tcp_log_stackspecific log;
6085 			struct timeval tv;
6086 
6087 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6088 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6089 			log.u_bbr.flex1 = avg;
6090 			log.u_bbr.flex2 = med;
6091 			log.u_bbr.flex3 = rxt_per;
6092 			log.u_bbr.flex4 = rack->r_ctl.policer_avg_threshold;
6093 			log.u_bbr.flex5 = rack->r_ctl.policer_med_threshold;
6094 			log.u_bbr.flex6 = rack->r_ctl.policer_rxt_threshold;
6095 			log.u_bbr.flex7 = rack->r_ctl.policer_alt_median;
6096 			log.u_bbr.flex8 = 2;
6097 			log.u_bbr.applimited = rack->r_ctl.current_round;
6098 			log.u_bbr.bw_inuse = del_bw;
6099 			log.u_bbr.delivered = rack->r_ctl.policer_bucket_size;
6100 			log.u_bbr.cur_del_rate = rxts;
6101 			log.u_bbr.delRate = snds;
6102 			log.u_bbr.rttProp = rack->r_ctl.gp_bw;
6103 			log.u_bbr.bbr_state = rack->rc_policer_detected;
6104 			log.u_bbr.bbr_substate = 0;
6105 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
6106 			log.u_bbr.use_lt_bw = rack->policer_detect_on;
6107 			log.u_bbr.epoch = rack->r_ctl.policer_max_seg;
6108 			log.u_bbr.lt_epoch = (uint32_t)tim;
6109 			log.u_bbr.pkts_out = rack->r_ctl.bytes_acked_in_recovery;
6110 			tcp_log_event(tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
6111 				      0, &log, false, NULL, NULL, 0, &tv);
6112 			/*
6113 			 * Put out an added log, 19, for the sole purpose
6114 			 * of getting the txt/rxt so that we can benchmark
6115 			 * in read-bbrlog the ongoing rxt rate after our
6116 			 * policer invocation in the HYSTART announcments.
6117 			 */
6118 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6119 			log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
6120 			log.u_bbr.flex1 = alt_med;
6121 			log.u_bbr.flex8 = 19;
6122 			log.u_bbr.cur_del_rate = tp->t_sndbytes;
6123 			log.u_bbr.delRate = tp->t_snd_rxt_bytes;
6124 			tcp_log_event(tp, NULL, NULL, NULL, TCP_POLICER_DET, 0,
6125 				      0, &log, false, NULL, NULL, 0, &tv);
6126 		}
6127 		/* Turn off any fast output, thats ended */
6128 		rack->r_fast_output = 0;
6129 		/* Mark the time for credits */
6130 		rack->r_ctl.last_sendtime = tcp_get_u64_usecs(NULL);
6131 		if (rack->r_rr_config < 2) {
6132 			/*
6133 			 * We need to be stricter on the RR config so
6134 			 * the pacing has priority.
6135 			 */
6136 			rack->r_rr_config = 2;
6137 		}
6138 		policer_detection_log(rack,
6139 				      rack->r_ctl.idle_snd_una,
6140 				      rack->r_ctl.ack_for_idle,
6141 				      0,
6142 				      (uint32_t)tim,
6143 				      14);
6144 		rack->rc_policer_detected = 1;
6145 	} else if ((rack->rc_policer_detected == 1) &&
6146 		   (post_recovery == 1)) {
6147 		/*
6148 		 * If we are exiting recovery and have already detected
6149 		 * we need to possibly update the values.
6150 		 *
6151 		 * First: Update the idle -> recovery sent value.
6152 		 */
6153 		uint32_t srtt;
6154 
6155 		if (rack->r_ctl.last_amount_before_rec > rack->r_ctl.policer_bucket_size) {
6156 			rack->r_ctl.policer_bucket_size = rack->r_ctl.last_amount_before_rec;
6157 		}
6158 		srtt = (uint64_t)rack_grab_rtt(tp, rack);
6159 		if ((tp->t_srtt > 0) && (srtt > tp->t_srtt))
6160 			srtt = tp->t_srtt;
6161 		if ((srtt != 0) &&
6162 		    (tim < (uint64_t)srtt)) {
6163 			/*
6164 			 * Not long enough.
6165 			 */
6166 			if (rack_verbose_logging)
6167 				policer_detection_log(rack,
6168 						      (uint32_t)tim,
6169 						      0,
6170 						      0,
6171 						      0,
6172 						      15);
6173 			return;
6174 		}
6175 		/*
6176 		 * Finally update the b/w if its grown.
6177 		 */
6178 		if (del_bw > rack->r_ctl.policer_bw) {
6179 			rack->r_ctl.policer_bw = del_bw;
6180 			rack->r_ctl.policer_max_seg = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp,
6181 											  rack->r_ctl.policer_bw,
6182 											  min(ctf_fixed_maxseg(rack->rc_tp),
6183 											      rack->r_ctl.rc_pace_min_segs),
6184 											  0, NULL,
6185 											  NULL, rack->r_ctl.pace_len_divisor);
6186 			if (rack->r_ctl.policer_bucket_size < rack->r_ctl.policer_max_seg) {
6187 				/* We must be able to send our max-seg or else chaos ensues */
6188 				rack->r_ctl.policer_bucket_size = rack->r_ctl.policer_max_seg * 2;
6189 			}
6190 		}
6191 		policer_detection_log(rack,
6192 				      rack->r_ctl.idle_snd_una,
6193 				      rack->r_ctl.ack_for_idle,
6194 				      0,
6195 				      (uint32_t)tim,
6196 				      3);
6197 	}
6198 }
6199 
6200 static void
6201 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
6202 {
6203 	/* now check with the policer if on */
6204 	if (rack->policer_detect_on == 1) {
6205 		policer_detection(tp, rack, 1);
6206 	}
6207 	/*
6208 	 * Now exit recovery, note we must do the idle set after the policer_detection
6209 	 * to get the amount acked prior to recovery correct.
6210 	 */
6211 	rack->r_ctl.idle_snd_una = tp->snd_una;
6212 	EXIT_RECOVERY(tp->t_flags);
6213 }
6214 
6215 static void
6216 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
6217 {
6218 	struct tcp_rack *rack;
6219 	uint32_t orig_cwnd;
6220 
6221 	orig_cwnd = tp->snd_cwnd;
6222 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6223 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6224 	/* only alert CC if we alerted when we entered */
6225 	if (CC_ALGO(tp)->post_recovery != NULL) {
6226 		tp->t_ccv.curack = th_ack;
6227 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
6228 		if (tp->snd_cwnd < tp->snd_ssthresh) {
6229 			/*
6230 			 * Rack has burst control and pacing
6231 			 * so lets not set this any lower than
6232 			 * snd_ssthresh per RFC-6582 (option 2).
6233 			 */
6234 			tp->snd_cwnd = tp->snd_ssthresh;
6235 		}
6236 	}
6237 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6238 		union tcp_log_stackspecific log;
6239 		struct timeval tv;
6240 
6241 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6242 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6243 		log.u_bbr.flex1 = th_ack;
6244 		log.u_bbr.flex2 = tp->t_ccv.flags;
6245 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
6246 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
6247 		log.u_bbr.flex5 = V_tcp_abc_l_var;
6248 		log.u_bbr.flex6 = orig_cwnd;
6249 		log.u_bbr.flex7 = V_tcp_do_newsack;
6250 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
6251 		log.u_bbr.flex8 = 2;
6252 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
6253 			       0, &log, false, NULL, __func__, __LINE__, &tv);
6254 	}
6255 	if ((rack->rack_no_prr == 0) &&
6256 	    (rack->no_prr_addback == 0) &&
6257 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
6258 		/*
6259 		 * Suck the next prr cnt back into cwnd, but
6260 		 * only do that if we are not application limited.
6261 		 */
6262 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
6263 			/*
6264 			 * We are allowed to add back to the cwnd the amount we did
6265 			 * not get out if:
6266 			 * a) no_prr_addback is off.
6267 			 * b) we are not app limited
6268 			 * c) we are doing prr
6269 			 * <and>
6270 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
6271 			 */
6272 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
6273 					    rack->r_ctl.rc_prr_sndcnt);
6274 		}
6275 		rack->r_ctl.rc_prr_sndcnt = 0;
6276 		rack_log_to_prr(rack, 1, 0, __LINE__);
6277 	}
6278 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
6279 	tp->snd_recover = tp->snd_una;
6280 	if (rack->r_ctl.dsack_persist) {
6281 		rack->r_ctl.dsack_persist--;
6282 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6283 			rack->r_ctl.num_dsack = 0;
6284 		}
6285 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6286 	}
6287 	if (rack->rto_from_rec == 1) {
6288 		rack->rto_from_rec = 0;
6289 		if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
6290 			tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
6291 	}
6292 	rack_exit_recovery(tp, rack, 1);
6293 }
6294 
6295 static void
6296 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
6297 {
6298 	struct tcp_rack *rack;
6299 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
6300 
6301 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6302 #ifdef STATS
6303 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
6304 #endif
6305 	if (IN_RECOVERY(tp->t_flags) == 0) {
6306 		in_rec_at_entry = 0;
6307 		ssthresh_enter = tp->snd_ssthresh;
6308 		cwnd_enter = tp->snd_cwnd;
6309 	} else
6310 		in_rec_at_entry = 1;
6311 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6312 	switch (type) {
6313 	case CC_NDUPACK:
6314 		tp->t_flags &= ~TF_WASFRECOVERY;
6315 		tp->t_flags &= ~TF_WASCRECOVERY;
6316 		if (!IN_FASTRECOVERY(tp->t_flags)) {
6317 			struct rack_sendmap *rsm;
6318 			struct timeval tv;
6319 			uint32_t segsiz;
6320 
6321 			/* Check if this is the end of the initial Start-up i.e. initial slow-start */
6322 			if (rack->rc_initial_ss_comp == 0) {
6323 				/* Yep it is the end of the initial slowstart */
6324 				rack->rc_initial_ss_comp = 1;
6325 			}
6326 			microuptime(&tv);
6327 			rack->r_ctl.time_entered_recovery = tcp_tv_to_lusectick(&tv);
6328 			if (SEQ_GEQ(ack, tp->snd_una)) {
6329 				/*
6330 				 * The ack is above snd_una. Lets see
6331 				 * if we can establish a postive distance from
6332 				 * our idle mark.
6333 				 */
6334 				rack->r_ctl.ack_for_idle = ack;
6335 				if (SEQ_GT(ack, rack->r_ctl.idle_snd_una)) {
6336 					rack->r_ctl.last_amount_before_rec = ack - rack->r_ctl.idle_snd_una;
6337 				} else {
6338 					/* No data thru yet */
6339 					rack->r_ctl.last_amount_before_rec = 0;
6340 				}
6341 			} else if (SEQ_GT(tp->snd_una, rack->r_ctl.idle_snd_una)) {
6342 				/*
6343 				 * The ack is out of order and behind the snd_una. It may
6344 				 * have contained SACK information which we processed else
6345 				 * we would have rejected it.
6346 				 */
6347 				rack->r_ctl.ack_for_idle = tp->snd_una;
6348 				rack->r_ctl.last_amount_before_rec = tp->snd_una - rack->r_ctl.idle_snd_una;
6349 			} else {
6350 				rack->r_ctl.ack_for_idle = ack;
6351 				rack->r_ctl.last_amount_before_rec = 0;
6352 			}
6353 			if (rack->rc_policer_detected) {
6354 				/*
6355 				 * If we are being policed and we have a loss, it
6356 				 * means our bucket is now empty. This can happen
6357 				 * where some other flow on the same host sends
6358 				 * that this connection is not aware of.
6359 				 */
6360 				rack->r_ctl.current_policer_bucket = 0;
6361 				if (rack_verbose_logging)
6362 					policer_detection_log(rack, rack->r_ctl.last_amount_before_rec, 0, 0, 0, 4);
6363 				if (rack->r_ctl.last_amount_before_rec > rack->r_ctl.policer_bucket_size) {
6364 					rack->r_ctl.policer_bucket_size = rack->r_ctl.last_amount_before_rec;
6365 				}
6366 			}
6367 			memset(rack->r_ctl.rc_cnt_of_retran, 0, sizeof(rack->r_ctl.rc_cnt_of_retran));
6368 			segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6369 			TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6370 				/*
6371 				 * Go through the outstanding and re-peg
6372 				 * any that should have been left in the
6373 				 * retransmit list (on a double recovery).
6374 				 */
6375 				if (rsm->r_act_rxt_cnt > 0) {
6376 					rack_peg_rxt(rack, rsm, segsiz);
6377 				}
6378 			}
6379 			rack->r_ctl.bytes_acked_in_recovery = 0;
6380 			rack->r_ctl.rc_prr_delivered = 0;
6381 			rack->r_ctl.rc_prr_out = 0;
6382 			rack->r_fast_output = 0;
6383 			if (rack->rack_no_prr == 0) {
6384 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
6385 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
6386 			}
6387 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
6388 			tp->snd_recover = tp->snd_max;
6389 			if (tp->t_flags2 & TF2_ECN_PERMIT)
6390 				tp->t_flags2 |= TF2_ECN_SND_CWR;
6391 		}
6392 		break;
6393 	case CC_ECN:
6394 		if (!IN_CONGRECOVERY(tp->t_flags) ||
6395 		    /*
6396 		     * Allow ECN reaction on ACK to CWR, if
6397 		     * that data segment was also CE marked.
6398 		     */
6399 		    SEQ_GEQ(ack, tp->snd_recover)) {
6400 			EXIT_CONGRECOVERY(tp->t_flags);
6401 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
6402 			rack->r_fast_output = 0;
6403 			tp->snd_recover = tp->snd_max + 1;
6404 			if (tp->t_flags2 & TF2_ECN_PERMIT)
6405 				tp->t_flags2 |= TF2_ECN_SND_CWR;
6406 		}
6407 		break;
6408 	case CC_RTO:
6409 		tp->t_dupacks = 0;
6410 		tp->t_bytes_acked = 0;
6411 		rack->r_fast_output = 0;
6412 		if (IN_RECOVERY(tp->t_flags))
6413 			rack_exit_recovery(tp, rack, 2);
6414 		rack->r_ctl.bytes_acked_in_recovery = 0;
6415 		rack->r_ctl.time_entered_recovery = 0;
6416 		orig_cwnd = tp->snd_cwnd;
6417 		rack_log_to_prr(rack, 16, orig_cwnd, line);
6418 		if (CC_ALGO(tp)->cong_signal == NULL) {
6419 			/* TSNH */
6420 			tp->snd_ssthresh = max(2,
6421 			    min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
6422 			    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
6423 			tp->snd_cwnd = ctf_fixed_maxseg(tp);
6424 		}
6425 		if (tp->t_flags2 & TF2_ECN_PERMIT)
6426 			tp->t_flags2 |= TF2_ECN_SND_CWR;
6427 		break;
6428 	case CC_RTO_ERR:
6429 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
6430 		/* RTO was unnecessary, so reset everything. */
6431 		tp->snd_cwnd = tp->snd_cwnd_prev;
6432 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
6433 		tp->snd_recover = tp->snd_recover_prev;
6434 		if (tp->t_flags & TF_WASFRECOVERY) {
6435 			ENTER_FASTRECOVERY(tp->t_flags);
6436 			tp->t_flags &= ~TF_WASFRECOVERY;
6437 		}
6438 		if (tp->t_flags & TF_WASCRECOVERY) {
6439 			ENTER_CONGRECOVERY(tp->t_flags);
6440 			tp->t_flags &= ~TF_WASCRECOVERY;
6441 		}
6442 		tp->snd_nxt = tp->snd_max;
6443 		tp->t_badrxtwin = 0;
6444 		break;
6445 	}
6446 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
6447 	    (type != CC_RTO)){
6448 		tp->t_ccv.curack = ack;
6449 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
6450 	}
6451 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
6452 		rack_log_to_prr(rack, 15, cwnd_enter, line);
6453 		rack->r_ctl.dsack_byte_cnt = 0;
6454 		rack->r_ctl.retran_during_recovery = 0;
6455 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
6456 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
6457 		rack->r_ent_rec_ns = 1;
6458 	}
6459 }
6460 
6461 static inline void
6462 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
6463 {
6464 	uint32_t i_cwnd;
6465 
6466 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6467 
6468 	if (CC_ALGO(tp)->after_idle != NULL)
6469 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
6470 
6471 	if (tp->snd_cwnd == 1)
6472 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
6473 	else
6474 		i_cwnd = rc_init_window(rack);
6475 
6476 	/*
6477 	 * Being idle is no different than the initial window. If the cc
6478 	 * clamps it down below the initial window raise it to the initial
6479 	 * window.
6480 	 */
6481 	if (tp->snd_cwnd < i_cwnd) {
6482 		tp->snd_cwnd = i_cwnd;
6483 	}
6484 }
6485 
6486 /*
6487  * Indicate whether this ack should be delayed.  We can delay the ack if
6488  * following conditions are met:
6489  *	- There is no delayed ack timer in progress.
6490  *	- Our last ack wasn't a 0-sized window. We never want to delay
6491  *	  the ack that opens up a 0-sized window.
6492  *	- LRO wasn't used for this segment. We make sure by checking that the
6493  *	  segment size is not larger than the MSS.
6494  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
6495  *	  connection.
6496  */
6497 #define DELAY_ACK(tp, tlen)			 \
6498 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
6499 	((tp->t_flags & TF_DELACK) == 0) &&	 \
6500 	(tlen <= tp->t_maxseg) &&		 \
6501 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
6502 
6503 static struct rack_sendmap *
6504 rack_find_lowest_rsm(struct tcp_rack *rack)
6505 {
6506 	struct rack_sendmap *rsm;
6507 
6508 	/*
6509 	 * Walk the time-order transmitted list looking for an rsm that is
6510 	 * not acked. This will be the one that was sent the longest time
6511 	 * ago that is still outstanding.
6512 	 */
6513 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6514 		if (rsm->r_flags & RACK_ACKED) {
6515 			continue;
6516 		}
6517 		goto finish;
6518 	}
6519 finish:
6520 	return (rsm);
6521 }
6522 
6523 static struct rack_sendmap *
6524 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
6525 {
6526 	struct rack_sendmap *prsm;
6527 
6528 	/*
6529 	 * Walk the sequence order list backward until we hit and arrive at
6530 	 * the highest seq not acked. In theory when this is called it
6531 	 * should be the last segment (which it was not).
6532 	 */
6533 	prsm = rsm;
6534 
6535 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
6536 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
6537 			continue;
6538 		}
6539 		return (prsm);
6540 	}
6541 	return (NULL);
6542 }
6543 
6544 static uint32_t
6545 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
6546 {
6547 	int32_t lro;
6548 	uint32_t thresh;
6549 
6550 	/*
6551 	 * lro is the flag we use to determine if we have seen reordering.
6552 	 * If it gets set we have seen reordering. The reorder logic either
6553 	 * works in one of two ways:
6554 	 *
6555 	 * If reorder-fade is configured, then we track the last time we saw
6556 	 * re-ordering occur. If we reach the point where enough time as
6557 	 * passed we no longer consider reordering has occuring.
6558 	 *
6559 	 * Or if reorder-face is 0, then once we see reordering we consider
6560 	 * the connection to alway be subject to reordering and just set lro
6561 	 * to 1.
6562 	 *
6563 	 * In the end if lro is non-zero we add the extra time for
6564 	 * reordering in.
6565 	 */
6566 	if (srtt == 0)
6567 		srtt = 1;
6568 	if (rack->r_ctl.rc_reorder_ts) {
6569 		if (rack->r_ctl.rc_reorder_fade) {
6570 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
6571 				lro = cts - rack->r_ctl.rc_reorder_ts;
6572 				if (lro == 0) {
6573 					/*
6574 					 * No time as passed since the last
6575 					 * reorder, mark it as reordering.
6576 					 */
6577 					lro = 1;
6578 				}
6579 			} else {
6580 				/* Negative time? */
6581 				lro = 0;
6582 			}
6583 			if (lro > rack->r_ctl.rc_reorder_fade) {
6584 				/* Turn off reordering seen too */
6585 				rack->r_ctl.rc_reorder_ts = 0;
6586 				lro = 0;
6587 			}
6588 		} else {
6589 			/* Reodering does not fade */
6590 			lro = 1;
6591 		}
6592 	} else {
6593 		lro = 0;
6594 	}
6595 	if (rack->rc_rack_tmr_std_based == 0) {
6596 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
6597 	} else {
6598 		/* Standards based pkt-delay is 1/4 srtt */
6599 		thresh = srtt +  (srtt >> 2);
6600 	}
6601 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
6602 		/* It must be set, if not you get 1/4 rtt */
6603 		if (rack->r_ctl.rc_reorder_shift)
6604 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
6605 		else
6606 			thresh += (srtt >> 2);
6607 	}
6608 	if (rack->rc_rack_use_dsack &&
6609 	    lro &&
6610 	    (rack->r_ctl.num_dsack > 0)) {
6611 		/*
6612 		 * We only increase the reordering window if we
6613 		 * have seen reordering <and> we have a DSACK count.
6614 		 */
6615 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
6616 		if (log_allowed)
6617 			rack_log_dsack_event(rack, 4, line, srtt, thresh);
6618 	}
6619 	/* SRTT * 2 is the ceiling */
6620 	if (thresh > (srtt * 2)) {
6621 		thresh = srtt * 2;
6622 	}
6623 	/* And we don't want it above the RTO max either */
6624 	if (thresh > rack_rto_max) {
6625 		thresh = rack_rto_max;
6626 	}
6627 	if (log_allowed)
6628 		rack_log_dsack_event(rack, 6, line,  srtt, thresh);
6629 	return (thresh);
6630 }
6631 
6632 static uint32_t
6633 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6634 		     struct rack_sendmap *rsm, uint32_t srtt)
6635 {
6636 	struct rack_sendmap *prsm;
6637 	uint32_t thresh, len;
6638 	int segsiz;
6639 
6640 	if (srtt == 0)
6641 		srtt = 1;
6642 	if (rack->r_ctl.rc_tlp_threshold)
6643 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6644 	else
6645 		thresh = (srtt * 2);
6646 
6647 	/* Get the previous sent packet, if any */
6648 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6649 	len = rsm->r_end - rsm->r_start;
6650 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6651 		/* Exactly like the ID */
6652 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6653 			uint32_t alt_thresh;
6654 			/*
6655 			 * Compensate for delayed-ack with the d-ack time.
6656 			 */
6657 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6658 			if (alt_thresh > thresh)
6659 				thresh = alt_thresh;
6660 		}
6661 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6662 		/* 2.1 behavior */
6663 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6664 		if (prsm && (len <= segsiz)) {
6665 			/*
6666 			 * Two packets outstanding, thresh should be (2*srtt) +
6667 			 * possible inter-packet delay (if any).
6668 			 */
6669 			uint32_t inter_gap = 0;
6670 			int idx, nidx;
6671 
6672 			idx = rsm->r_rtr_cnt - 1;
6673 			nidx = prsm->r_rtr_cnt - 1;
6674 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6675 				/* Yes it was sent later (or at the same time) */
6676 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6677 			}
6678 			thresh += inter_gap;
6679 		} else if (len <= segsiz) {
6680 			/*
6681 			 * Possibly compensate for delayed-ack.
6682 			 */
6683 			uint32_t alt_thresh;
6684 
6685 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6686 			if (alt_thresh > thresh)
6687 				thresh = alt_thresh;
6688 		}
6689 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6690 		/* 2.2 behavior */
6691 		if (len <= segsiz) {
6692 			uint32_t alt_thresh;
6693 			/*
6694 			 * Compensate for delayed-ack with the d-ack time.
6695 			 */
6696 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6697 			if (alt_thresh > thresh)
6698 				thresh = alt_thresh;
6699 		}
6700 	}
6701 	/* Not above an RTO */
6702 	if (thresh > tp->t_rxtcur) {
6703 		thresh = tp->t_rxtcur;
6704 	}
6705 	/* Not above a RTO max */
6706 	if (thresh > rack_rto_max) {
6707 		thresh = rack_rto_max;
6708 	}
6709 	/* Apply user supplied min TLP */
6710 	if (thresh < rack_tlp_min) {
6711 		thresh = rack_tlp_min;
6712 	}
6713 	return (thresh);
6714 }
6715 
6716 static uint32_t
6717 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6718 {
6719 	/*
6720 	 * We want the rack_rtt which is the
6721 	 * last rtt we measured. However if that
6722 	 * does not exist we fallback to the srtt (which
6723 	 * we probably will never do) and then as a last
6724 	 * resort we use RACK_INITIAL_RTO if no srtt is
6725 	 * yet set.
6726 	 */
6727 	if (rack->rc_rack_rtt)
6728 		return (rack->rc_rack_rtt);
6729 	else if (tp->t_srtt == 0)
6730 		return (RACK_INITIAL_RTO);
6731 	return (tp->t_srtt);
6732 }
6733 
6734 static struct rack_sendmap *
6735 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6736 {
6737 	/*
6738 	 * Check to see that we don't need to fall into recovery. We will
6739 	 * need to do so if our oldest transmit is past the time we should
6740 	 * have had an ack.
6741 	 */
6742 	struct tcp_rack *rack;
6743 	struct rack_sendmap *rsm;
6744 	int32_t idx;
6745 	uint32_t srtt, thresh;
6746 
6747 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6748 	if (tqhash_empty(rack->r_ctl.tqh)) {
6749 		return (NULL);
6750 	}
6751 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6752 	if (rsm == NULL)
6753 		return (NULL);
6754 
6755 
6756 	if (rsm->r_flags & RACK_ACKED) {
6757 		rsm = rack_find_lowest_rsm(rack);
6758 		if (rsm == NULL)
6759 			return (NULL);
6760 	}
6761 	idx = rsm->r_rtr_cnt - 1;
6762 	srtt = rack_grab_rtt(tp, rack);
6763 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6764 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6765 		return (NULL);
6766 	}
6767 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6768 		return (NULL);
6769 	}
6770 	/* Ok if we reach here we are over-due and this guy can be sent */
6771 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6772 	return (rsm);
6773 }
6774 
6775 static uint32_t
6776 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6777 {
6778 	int32_t t;
6779 	int32_t tt;
6780 	uint32_t ret_val;
6781 
6782 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6783 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6784  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6785 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6786 	ret_val = (uint32_t)tt;
6787 	return (ret_val);
6788 }
6789 
6790 static uint32_t
6791 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6792 {
6793 	/*
6794 	 * Start the FR timer, we do this based on getting the first one in
6795 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6796 	 * events we need to stop the running timer (if its running) before
6797 	 * starting the new one.
6798 	 */
6799 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6800 	uint32_t srtt_cur;
6801 	int32_t idx;
6802 	int32_t is_tlp_timer = 0;
6803 	struct rack_sendmap *rsm;
6804 
6805 	if (rack->t_timers_stopped) {
6806 		/* All timers have been stopped none are to run */
6807 		return (0);
6808 	}
6809 	if (rack->rc_in_persist) {
6810 		/* We can't start any timer in persists */
6811 		return (rack_get_persists_timer_val(tp, rack));
6812 	}
6813 	rack->rc_on_min_to = 0;
6814 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6815 	    (rack->sack_attack_disable > 0) ||
6816 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6817 		goto activate_rxt;
6818 	}
6819 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6820 	if ((rsm == NULL) || sup_rack) {
6821 		/* Nothing on the send map or no rack */
6822 activate_rxt:
6823 		time_since_sent = 0;
6824 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6825 		if (rsm) {
6826 			/*
6827 			 * Should we discount the RTX timer any?
6828 			 *
6829 			 * We want to discount it the smallest amount.
6830 			 * If a timer (Rack/TLP or RXT) has gone off more
6831 			 * recently thats the discount we want to use (now - timer time).
6832 			 * If the retransmit of the oldest packet was more recent then
6833 			 * we want to use that (now - oldest-packet-last_transmit_time).
6834 			 *
6835 			 */
6836 			idx = rsm->r_rtr_cnt - 1;
6837 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6838 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6839 			else
6840 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6841 			if (TSTMP_GT(cts, tstmp_touse))
6842 			    time_since_sent = cts - tstmp_touse;
6843 		}
6844 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6845 		    sbavail(&tptosocket(tp)->so_snd)) {
6846 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6847 			to = tp->t_rxtcur;
6848 			if (to > time_since_sent)
6849 				to -= time_since_sent;
6850 			else
6851 				to = rack->r_ctl.rc_min_to;
6852 			if (to == 0)
6853 				to = 1;
6854 			/* Special case for KEEPINIT */
6855 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6856 			    (TP_KEEPINIT(tp) != 0) &&
6857 			    rsm) {
6858 				/*
6859 				 * We have to put a ceiling on the rxt timer
6860 				 * of the keep-init timeout.
6861 				 */
6862 				uint32_t max_time, red;
6863 
6864 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6865 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6866 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6867 					if (red < max_time)
6868 						max_time -= red;
6869 					else
6870 						max_time = 1;
6871 				}
6872 				/* Reduce timeout to the keep value if needed */
6873 				if (max_time < to)
6874 					to = max_time;
6875 			}
6876 			return (to);
6877 		}
6878 		return (0);
6879 	}
6880 	if (rsm->r_flags & RACK_ACKED) {
6881 		rsm = rack_find_lowest_rsm(rack);
6882 		if (rsm == NULL) {
6883 			/* No lowest? */
6884 			goto activate_rxt;
6885 		}
6886 	}
6887 	if (rack->sack_attack_disable) {
6888 		/*
6889 		 * We don't want to do
6890 		 * any TLP's if you are an attacker.
6891 		 * Though if you are doing what
6892 		 * is expected you may still have
6893 		 * SACK-PASSED marks.
6894 		 */
6895 		goto activate_rxt;
6896 	}
6897 	/* Convert from ms to usecs */
6898 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6899 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6900 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6901 		if ((tp->t_flags & TF_SENTFIN) &&
6902 		    ((tp->snd_max - tp->snd_una) == 1) &&
6903 		    (rsm->r_flags & RACK_HAS_FIN)) {
6904 			/*
6905 			 * We don't start a rack timer if all we have is a
6906 			 * FIN outstanding.
6907 			 */
6908 			goto activate_rxt;
6909 		}
6910 		if ((rack->use_rack_rr == 0) &&
6911 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6912 		    (rack->rack_no_prr == 0) &&
6913 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6914 			/*
6915 			 * We are not cheating, in recovery  and
6916 			 * not enough ack's to yet get our next
6917 			 * retransmission out.
6918 			 *
6919 			 * Note that classified attackers do not
6920 			 * get to use the rack-cheat.
6921 			 */
6922 			goto activate_tlp;
6923 		}
6924 		srtt = rack_grab_rtt(tp, rack);
6925 		thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6926 		idx = rsm->r_rtr_cnt - 1;
6927 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6928 		if (SEQ_GEQ(exp, cts)) {
6929 			to = exp - cts;
6930 			if (to < rack->r_ctl.rc_min_to) {
6931 				to = rack->r_ctl.rc_min_to;
6932 				if (rack->r_rr_config == 3)
6933 					rack->rc_on_min_to = 1;
6934 			}
6935 		} else {
6936 			to = rack->r_ctl.rc_min_to;
6937 			if (rack->r_rr_config == 3)
6938 				rack->rc_on_min_to = 1;
6939 		}
6940 	} else {
6941 		/* Ok we need to do a TLP not RACK */
6942 activate_tlp:
6943 		if ((rack->rc_tlp_in_progress != 0) &&
6944 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6945 			/*
6946 			 * The previous send was a TLP and we have sent
6947 			 * N TLP's without sending new data.
6948 			 */
6949 			goto activate_rxt;
6950 		}
6951 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6952 		if (rsm == NULL) {
6953 			/* We found no rsm to TLP with. */
6954 			goto activate_rxt;
6955 		}
6956 		if (rsm->r_flags & RACK_HAS_FIN) {
6957 			/* If its a FIN we dont do TLP */
6958 			rsm = NULL;
6959 			goto activate_rxt;
6960 		}
6961 		idx = rsm->r_rtr_cnt - 1;
6962 		time_since_sent = 0;
6963 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6964 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6965 		else
6966 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6967 		if (TSTMP_GT(cts, tstmp_touse))
6968 		    time_since_sent = cts - tstmp_touse;
6969 		is_tlp_timer = 1;
6970 		if (tp->t_srtt) {
6971 			if ((rack->rc_srtt_measure_made == 0) &&
6972 			    (tp->t_srtt == 1)) {
6973 				/*
6974 				 * If another stack as run and set srtt to 1,
6975 				 * then the srtt was 0, so lets use the initial.
6976 				 */
6977 				srtt = RACK_INITIAL_RTO;
6978 			} else {
6979 				srtt_cur = tp->t_srtt;
6980 				srtt = srtt_cur;
6981 			}
6982 		} else
6983 			srtt = RACK_INITIAL_RTO;
6984 		/*
6985 		 * If the SRTT is not keeping up and the
6986 		 * rack RTT has spiked we want to use
6987 		 * the last RTT not the smoothed one.
6988 		 */
6989 		if (rack_tlp_use_greater &&
6990 		    tp->t_srtt &&
6991 		    (srtt < rack_grab_rtt(tp, rack))) {
6992 			srtt = rack_grab_rtt(tp, rack);
6993 		}
6994 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6995 		if (thresh > time_since_sent) {
6996 			to = thresh - time_since_sent;
6997 		} else {
6998 			to = rack->r_ctl.rc_min_to;
6999 			rack_log_alt_to_to_cancel(rack,
7000 						  thresh,		/* flex1 */
7001 						  time_since_sent,	/* flex2 */
7002 						  tstmp_touse,		/* flex3 */
7003 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
7004 						  (uint32_t)rsm->r_tim_lastsent[idx],
7005 						  srtt,
7006 						  idx, 99);
7007 		}
7008 		if (to < rack_tlp_min) {
7009 			to = rack_tlp_min;
7010 		}
7011 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
7012 			/*
7013 			 * If the TLP time works out to larger than the max
7014 			 * RTO lets not do TLP.. just RTO.
7015 			 */
7016 			goto activate_rxt;
7017 		}
7018 	}
7019 	if (is_tlp_timer == 0) {
7020 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
7021 	} else {
7022 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
7023 	}
7024 	if (to == 0)
7025 		to = 1;
7026 	return (to);
7027 }
7028 
7029 static void
7030 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
7031 {
7032 	if (rack->rc_in_persist == 0) {
7033 		if (tp->t_flags & TF_GPUTINPROG) {
7034 			/*
7035 			 * Stop the goodput now, the calling of the
7036 			 * measurement function clears the flag.
7037 			 */
7038 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
7039 						    RACK_QUALITY_PERSIST);
7040 		}
7041 #ifdef NETFLIX_SHARED_CWND
7042 		if (rack->r_ctl.rc_scw) {
7043 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
7044 			rack->rack_scwnd_is_idle = 1;
7045 		}
7046 #endif
7047 		rack->r_ctl.rc_went_idle_time = cts;
7048 		if (rack->r_ctl.rc_went_idle_time == 0)
7049 			rack->r_ctl.rc_went_idle_time = 1;
7050 		if (rack->lt_bw_up) {
7051 			/* Suspend our LT BW measurement */
7052 			uint64_t tmark;
7053 
7054 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
7055 			rack->r_ctl.lt_seq = snd_una;
7056 			tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
7057 			if (tmark >= rack->r_ctl.lt_timemark) {
7058 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
7059 			}
7060 			rack->r_ctl.lt_timemark = tmark;
7061 			rack->lt_bw_up = 0;
7062 			rack->r_persist_lt_bw_off = 1;
7063 		}
7064 		rack_timer_cancel(tp, rack, cts, __LINE__);
7065 		rack->r_ctl.persist_lost_ends = 0;
7066 		rack->probe_not_answered = 0;
7067 		rack->forced_ack = 0;
7068 		tp->t_rxtshift = 0;
7069 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7070 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7071 		rack->rc_in_persist = 1;
7072 	}
7073 }
7074 
7075 static void
7076 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7077 {
7078 	if (tcp_in_hpts(rack->rc_tp)) {
7079 		tcp_hpts_remove(rack->rc_tp);
7080 		rack->r_ctl.rc_hpts_flags = 0;
7081 	}
7082 #ifdef NETFLIX_SHARED_CWND
7083 	if (rack->r_ctl.rc_scw) {
7084 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
7085 		rack->rack_scwnd_is_idle = 0;
7086 	}
7087 #endif
7088 	if (rack->rc_gp_dyn_mul &&
7089 	    (rack->use_fixed_rate == 0) &&
7090 	    (rack->rc_always_pace)) {
7091 		/*
7092 		 * Do we count this as if a probe-rtt just
7093 		 * finished?
7094 		 */
7095 		uint32_t time_idle, idle_min;
7096 
7097 		time_idle = cts - rack->r_ctl.rc_went_idle_time;
7098 		idle_min = rack_min_probertt_hold;
7099 		if (rack_probertt_gpsrtt_cnt_div) {
7100 			uint64_t extra;
7101 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
7102 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
7103 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
7104 			idle_min += (uint32_t)extra;
7105 		}
7106 		if (time_idle >= idle_min) {
7107 			/* Yes, we count it as a probe-rtt. */
7108 			uint32_t us_cts;
7109 
7110 			us_cts = tcp_get_usecs(NULL);
7111 			if (rack->in_probe_rtt == 0) {
7112 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7113 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
7114 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
7115 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
7116 			} else {
7117 				rack_exit_probertt(rack, us_cts);
7118 			}
7119 		}
7120 	}
7121 	if (rack->r_persist_lt_bw_off) {
7122 		/* Continue where we left off */
7123 		rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
7124 		rack->lt_bw_up = 1;
7125 		rack->r_persist_lt_bw_off = 0;
7126 	}
7127 	rack->r_ctl.idle_snd_una = tp->snd_una;
7128 	rack->rc_in_persist = 0;
7129 	rack->r_ctl.rc_went_idle_time = 0;
7130 	tp->t_rxtshift = 0;
7131 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7132 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7133 	rack->r_ctl.rc_agg_delayed = 0;
7134 	rack->r_early = 0;
7135 	rack->r_late = 0;
7136 	rack->r_ctl.rc_agg_early = 0;
7137 }
7138 
7139 static void
7140 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
7141 		   struct hpts_diag *diag, struct timeval *tv)
7142 {
7143 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
7144 		union tcp_log_stackspecific log;
7145 
7146 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
7147 		log.u_bbr.flex1 = diag->p_nxt_slot;
7148 		log.u_bbr.flex2 = diag->p_cur_slot;
7149 		log.u_bbr.flex3 = diag->slot_req;
7150 		log.u_bbr.flex4 = diag->inp_hptsslot;
7151 		log.u_bbr.flex5 = diag->slot_remaining;
7152 		log.u_bbr.flex6 = diag->need_new_to;
7153 		log.u_bbr.flex7 = diag->p_hpts_active;
7154 		log.u_bbr.flex8 = diag->p_on_min_sleep;
7155 		/* Hijack other fields as needed */
7156 		log.u_bbr.epoch = diag->have_slept;
7157 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
7158 		log.u_bbr.pkts_out = diag->co_ret;
7159 		log.u_bbr.applimited = diag->hpts_sleep_time;
7160 		log.u_bbr.delivered = diag->p_prev_slot;
7161 		log.u_bbr.inflight = diag->p_runningslot;
7162 		log.u_bbr.bw_inuse = diag->wheel_slot;
7163 		log.u_bbr.rttProp = diag->wheel_cts;
7164 		log.u_bbr.timeStamp = cts;
7165 		log.u_bbr.delRate = diag->maxslots;
7166 		log.u_bbr.cur_del_rate = diag->p_curtick;
7167 		log.u_bbr.cur_del_rate <<= 32;
7168 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
7169 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
7170 		    &rack->rc_inp->inp_socket->so_rcv,
7171 		    &rack->rc_inp->inp_socket->so_snd,
7172 		    BBR_LOG_HPTSDIAG, 0,
7173 		    0, &log, false, tv);
7174 	}
7175 
7176 }
7177 
7178 static void
7179 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
7180 {
7181 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
7182 		union tcp_log_stackspecific log;
7183 		struct timeval tv;
7184 
7185 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
7186 		log.u_bbr.flex1 = sb->sb_flags;
7187 		log.u_bbr.flex2 = len;
7188 		log.u_bbr.flex3 = sb->sb_state;
7189 		log.u_bbr.flex8 = type;
7190 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
7191 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
7192 		    &rack->rc_inp->inp_socket->so_rcv,
7193 		    &rack->rc_inp->inp_socket->so_snd,
7194 		    TCP_LOG_SB_WAKE, 0,
7195 		    len, &log, false, &tv);
7196 	}
7197 }
7198 
7199 static void
7200 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
7201       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
7202 {
7203 	struct hpts_diag diag;
7204 	struct inpcb *inp = tptoinpcb(tp);
7205 	struct timeval tv;
7206 	uint32_t delayed_ack = 0;
7207 	uint32_t hpts_timeout;
7208 	uint32_t entry_slot = slot;
7209 	uint8_t stopped;
7210 	uint32_t left = 0;
7211 	uint32_t us_cts;
7212 
7213 	if ((tp->t_state == TCPS_CLOSED) ||
7214 	    (tp->t_state == TCPS_LISTEN)) {
7215 		return;
7216 	}
7217 	if (tcp_in_hpts(tp)) {
7218 		/* Already on the pacer */
7219 		return;
7220 	}
7221 	stopped = rack->rc_tmr_stopped;
7222 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
7223 		left = rack->r_ctl.rc_timer_exp - cts;
7224 	}
7225 	rack->r_ctl.rc_timer_exp = 0;
7226 	rack->r_ctl.rc_hpts_flags = 0;
7227 	us_cts = tcp_get_usecs(&tv);
7228 	/* Now early/late accounting */
7229 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
7230 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
7231 		/*
7232 		 * We have a early carry over set,
7233 		 * we can always add more time so we
7234 		 * can always make this compensation.
7235 		 *
7236 		 * Note if ack's are allowed to wake us do not
7237 		 * penalize the next timer for being awoke
7238 		 * by an ack aka the rc_agg_early (non-paced mode).
7239 		 */
7240 		slot += rack->r_ctl.rc_agg_early;
7241 		rack->r_early = 0;
7242 		rack->r_ctl.rc_agg_early = 0;
7243 	}
7244 	if ((rack->r_late) &&
7245 	    ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
7246 		/*
7247 		 * This is harder, we can
7248 		 * compensate some but it
7249 		 * really depends on what
7250 		 * the current pacing time is.
7251 		 */
7252 		if (rack->r_ctl.rc_agg_delayed >= slot) {
7253 			/*
7254 			 * We can't compensate for it all.
7255 			 * And we have to have some time
7256 			 * on the clock. We always have a min
7257 			 * 10 slots (10 x 10 i.e. 100 usecs).
7258 			 */
7259 			if (slot <= HPTS_TICKS_PER_SLOT) {
7260 				/* We gain delay */
7261 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
7262 				slot = HPTS_TICKS_PER_SLOT;
7263 			} else {
7264 				/* We take off some */
7265 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
7266 				slot = HPTS_TICKS_PER_SLOT;
7267 			}
7268 		} else {
7269 			slot -= rack->r_ctl.rc_agg_delayed;
7270 			rack->r_ctl.rc_agg_delayed = 0;
7271 			/* Make sure we have 100 useconds at minimum */
7272 			if (slot < HPTS_TICKS_PER_SLOT) {
7273 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
7274 				slot = HPTS_TICKS_PER_SLOT;
7275 			}
7276 			if (rack->r_ctl.rc_agg_delayed == 0)
7277 				rack->r_late = 0;
7278 		}
7279 	} else if (rack->r_late) {
7280 		/* r_use_hpts_min is on and so is DGP */
7281 		uint32_t max_red;
7282 
7283 		max_red = (slot * rack->r_ctl.max_reduction) / 100;
7284 		if (max_red >= rack->r_ctl.rc_agg_delayed) {
7285 			slot -= rack->r_ctl.rc_agg_delayed;
7286 			rack->r_ctl.rc_agg_delayed = 0;
7287 		} else {
7288 			slot -= max_red;
7289 			rack->r_ctl.rc_agg_delayed -= max_red;
7290 		}
7291 	}
7292 	if ((rack->r_use_hpts_min == 1) &&
7293 	    (slot > 0) &&
7294 	    (rack->dgp_on == 1)) {
7295 		/*
7296 		 * We are enforcing a min pacing timer
7297 		 * based on our hpts min timeout.
7298 		 */
7299 		uint32_t min;
7300 
7301 		min = get_hpts_min_sleep_time();
7302 		if (min > slot) {
7303 			slot = min;
7304 		}
7305 	}
7306 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
7307 #ifdef TCP_SAD_DETECTION
7308 	if (rack->sack_attack_disable &&
7309 	    (rack->r_ctl.ack_during_sd > 0) &&
7310 	    (slot < tcp_sad_pacing_interval)) {
7311 		/*
7312 		 * We have a potential attacker on
7313 		 * the line. We have possibly some
7314 		 * (or now) pacing time set. We want to
7315 		 * slow down the processing of sacks by some
7316 		 * amount (if it is an attacker). Set the default
7317 		 * slot for attackers in place (unless the original
7318 		 * interval is longer). Its stored in
7319 		 * micro-seconds, so lets convert to msecs.
7320 		 */
7321 		slot = tcp_sad_pacing_interval;
7322 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
7323 		rack->r_ctl.ack_during_sd = 0;
7324 	}
7325 #endif
7326 	if (tp->t_flags & TF_DELACK) {
7327 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
7328 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
7329 	}
7330 	if (delayed_ack && ((hpts_timeout == 0) ||
7331 			    (delayed_ack < hpts_timeout)))
7332 		hpts_timeout = delayed_ack;
7333 	else
7334 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7335 	/*
7336 	 * If no timers are going to run and we will fall off the hptsi
7337 	 * wheel, we resort to a keep-alive timer if its configured.
7338 	 */
7339 	if ((hpts_timeout == 0) &&
7340 	    (slot == 0)) {
7341 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7342 		    (tp->t_state <= TCPS_CLOSING)) {
7343 			/*
7344 			 * Ok we have no timer (persists, rack, tlp, rxt  or
7345 			 * del-ack), we don't have segments being paced. So
7346 			 * all that is left is the keepalive timer.
7347 			 */
7348 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7349 				/* Get the established keep-alive time */
7350 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
7351 			} else {
7352 				/*
7353 				 * Get the initial setup keep-alive time,
7354 				 * note that this is probably not going to
7355 				 * happen, since rack will be running a rxt timer
7356 				 * if a SYN of some sort is outstanding. It is
7357 				 * actually handled in rack_timeout_rxt().
7358 				 */
7359 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
7360 			}
7361 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
7362 			if (rack->in_probe_rtt) {
7363 				/*
7364 				 * We want to instead not wake up a long time from
7365 				 * now but to wake up about the time we would
7366 				 * exit probe-rtt and initiate a keep-alive ack.
7367 				 * This will get us out of probe-rtt and update
7368 				 * our min-rtt.
7369 				 */
7370 				hpts_timeout = rack_min_probertt_hold;
7371 			}
7372 		}
7373 	}
7374 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
7375 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
7376 		/*
7377 		 * RACK, TLP, persists and RXT timers all are restartable
7378 		 * based on actions input .. i.e we received a packet (ack
7379 		 * or sack) and that changes things (rw, or snd_una etc).
7380 		 * Thus we can restart them with a new value. For
7381 		 * keep-alive, delayed_ack we keep track of what was left
7382 		 * and restart the timer with a smaller value.
7383 		 */
7384 		if (left < hpts_timeout)
7385 			hpts_timeout = left;
7386 	}
7387 	if (hpts_timeout) {
7388 		/*
7389 		 * Hack alert for now we can't time-out over 2,147,483
7390 		 * seconds (a bit more than 596 hours), which is probably ok
7391 		 * :).
7392 		 */
7393 		if (hpts_timeout > 0x7ffffffe)
7394 			hpts_timeout = 0x7ffffffe;
7395 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
7396 	}
7397 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
7398 	if ((rack->gp_ready == 0) &&
7399 	    (rack->use_fixed_rate == 0) &&
7400 	    (hpts_timeout < slot) &&
7401 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
7402 		/*
7403 		 * We have no good estimate yet for the
7404 		 * old clunky burst mitigation or the
7405 		 * real pacing. And the tlp or rxt is smaller
7406 		 * than the pacing calculation. Lets not
7407 		 * pace that long since we know the calculation
7408 		 * so far is not accurate.
7409 		 */
7410 		slot = hpts_timeout;
7411 	}
7412 	/**
7413 	 * Turn off all the flags for queuing by default. The
7414 	 * flags have important meanings to what happens when
7415 	 * LRO interacts with the transport. Most likely (by default now)
7416 	 * mbuf_queueing and ack compression are on. So the transport
7417 	 * has a couple of flags that control what happens (if those
7418 	 * are not on then these flags won't have any effect since it
7419 	 * won't go through the queuing LRO path).
7420 	 *
7421 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
7422 	 *                        pacing output, so don't disturb. But
7423 	 *                        it also means LRO can wake me if there
7424 	 *                        is a SACK arrival.
7425 	 *
7426 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
7427 	 *                       with the above flag (QUEUE_READY) and
7428 	 *                       when present it says don't even wake me
7429 	 *                       if a SACK arrives.
7430 	 *
7431 	 * The idea behind these flags is that if we are pacing we
7432 	 * set the MBUF_QUEUE_READY and only get woken up if
7433 	 * a SACK arrives (which could change things) or if
7434 	 * our pacing timer expires. If, however, we have a rack
7435 	 * timer running, then we don't even want a sack to wake
7436 	 * us since the rack timer has to expire before we can send.
7437 	 *
7438 	 * Other cases should usually have none of the flags set
7439 	 * so LRO can call into us.
7440 	 */
7441 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
7442 	if (slot) {
7443 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
7444 		rack->r_ctl.rc_last_output_to = us_cts + slot;
7445 		/*
7446 		 * A pacing timer (slot) is being set, in
7447 		 * such a case we cannot send (we are blocked by
7448 		 * the timer). So lets tell LRO that it should not
7449 		 * wake us unless there is a SACK. Note this only
7450 		 * will be effective if mbuf queueing is on or
7451 		 * compressed acks are being processed.
7452 		 */
7453 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
7454 		/*
7455 		 * But wait if we have a Rack timer running
7456 		 * even a SACK should not disturb us (with
7457 		 * the exception of r_rr_config 3).
7458 		 */
7459 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
7460 		    (IN_RECOVERY(tp->t_flags))) {
7461 			if (rack->r_rr_config != 3)
7462 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
7463 			else if (rack->rc_pace_dnd) {
7464 				/*
7465 				 * When DND is on, we only let a sack
7466 				 * interrupt us if we are not in recovery.
7467 				 *
7468 				 * If DND is off, then we never hit here
7469 				 * and let all sacks wake us up.
7470 				 *
7471 				 */
7472 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
7473 			}
7474 		}
7475 		/* For sack attackers we want to ignore sack */
7476 		if (rack->sack_attack_disable == 1) {
7477 			tp->t_flags2 |= (TF2_DONT_SACK_QUEUE |
7478 			    TF2_MBUF_QUEUE_READY);
7479 		} else if (rack->rc_ack_can_sendout_data) {
7480 			/*
7481 			 * Ahh but wait, this is that special case
7482 			 * where the pacing timer can be disturbed
7483 			 * backout the changes (used for non-paced
7484 			 * burst limiting).
7485 			 */
7486 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
7487 			    TF2_MBUF_QUEUE_READY);
7488 		}
7489 		if ((rack->use_rack_rr) &&
7490 		    (rack->r_rr_config < 2) &&
7491 		    ((hpts_timeout) && (hpts_timeout < slot))) {
7492 			/*
7493 			 * Arrange for the hpts to kick back in after the
7494 			 * t-o if the t-o does not cause a send.
7495 			 */
7496 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7497 						   __LINE__, &diag);
7498 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7499 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7500 		} else {
7501 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
7502 						   __LINE__, &diag);
7503 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7504 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
7505 		}
7506 	} else if (hpts_timeout) {
7507 		/*
7508 		 * With respect to t_flags2(?) here, lets let any new acks wake
7509 		 * us up here. Since we are not pacing (no pacing timer), output
7510 		 * can happen so we should let it. If its a Rack timer, then any inbound
7511 		 * packet probably won't change the sending (we will be blocked)
7512 		 * but it may change the prr stats so letting it in (the set defaults
7513 		 * at the start of this block) are good enough.
7514 		 */
7515 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7516 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7517 					   __LINE__, &diag);
7518 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7519 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7520 	} else {
7521 		/* No timer starting */
7522 #ifdef INVARIANTS
7523 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
7524 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
7525 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
7526 		}
7527 #endif
7528 	}
7529 	rack->rc_tmr_stopped = 0;
7530 	if (slot)
7531 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
7532 }
7533 
7534 static void
7535 rack_mark_lost(struct tcpcb *tp,
7536     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
7537 {
7538 	struct rack_sendmap *nrsm;
7539 	uint32_t thresh,  exp;
7540 
7541 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
7542 	nrsm = rsm;
7543 	TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
7544 		if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
7545 			/* Got up to all that were marked sack-passed */
7546 			break;
7547 		}
7548 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
7549 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
7550 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
7551 				/* We now consider it lost */
7552 				nrsm->r_flags |= RACK_WAS_LOST;
7553 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
7554 			} else {
7555 				/* Past here it won't be lost so stop */
7556 				break;
7557 			}
7558 		}
7559 	}
7560 }
7561 
7562 /*
7563  * RACK Timer, here we simply do logging and house keeping.
7564  * the normal rack_output() function will call the
7565  * appropriate thing to check if we need to do a RACK retransmit.
7566  * We return 1, saying don't proceed with rack_output only
7567  * when all timers have been stopped (destroyed PCB?).
7568  */
7569 static int
7570 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7571 {
7572 	/*
7573 	 * This timer simply provides an internal trigger to send out data.
7574 	 * The check_recovery_mode call will see if there are needed
7575 	 * retransmissions, if so we will enter fast-recovery. The output
7576 	 * call may or may not do the same thing depending on sysctl
7577 	 * settings.
7578 	 */
7579 	struct rack_sendmap *rsm;
7580 
7581 	counter_u64_add(rack_to_tot, 1);
7582 	if (rack->r_state && (rack->r_state != tp->t_state))
7583 		rack_set_state(tp, rack);
7584 	rack->rc_on_min_to = 0;
7585 	rsm = rack_check_recovery_mode(tp, cts);
7586 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
7587 	if (rsm) {
7588 		/* We need to stroke any lost that are now declared as lost */
7589 		rack_mark_lost(tp, rack, rsm, cts);
7590 		rack->r_ctl.rc_resend = rsm;
7591 		rack->r_timer_override = 1;
7592 		if (rack->use_rack_rr) {
7593 			/*
7594 			 * Don't accumulate extra pacing delay
7595 			 * we are allowing the rack timer to
7596 			 * over-ride pacing i.e. rrr takes precedence
7597 			 * if the pacing interval is longer than the rrr
7598 			 * time (in other words we get the min pacing
7599 			 * time versus rrr pacing time).
7600 			 */
7601 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7602 		}
7603 	}
7604 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
7605 	if (rsm == NULL) {
7606 		/* restart a timer and return 1 */
7607 		rack_start_hpts_timer(rack, tp, cts,
7608 				      0, 0, 0);
7609 		return (1);
7610 	}
7611 	if ((rack->policer_detect_on == 1) &&
7612 	    (rack->rc_policer_detected == 0)) {
7613 		/*
7614 		 * We do this early if we have not
7615 		 * deteceted to attempt to detect
7616 		 * quicker. Normally we want to do this
7617 		 * as recovery exits (and we will again).
7618 		 */
7619 		policer_detection(tp, rack, 0);
7620 	}
7621 	return (0);
7622 }
7623 
7624 
7625 
7626 static void
7627 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
7628 {
7629 
7630 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
7631 		/*
7632 		 * The trailing space changed, mbufs can grow
7633 		 * at the tail but they can't shrink from
7634 		 * it, KASSERT that. Adjust the orig_m_len to
7635 		 * compensate for this change.
7636 		 */
7637 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7638 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7639 			 rsm->m,
7640 			 rsm,
7641 			 (intmax_t)M_TRAILINGROOM(rsm->m),
7642 			 rsm->orig_t_space,
7643 			 rsm->orig_m_len,
7644 			 rsm->m->m_len));
7645 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7646 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7647 	}
7648 	if (rsm->m->m_len < rsm->orig_m_len) {
7649 		/*
7650 		 * Mbuf shrank, trimmed off the top by an ack, our
7651 		 * offset changes.
7652 		 */
7653 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7654 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7655 			 rsm->m, rsm->m->m_len,
7656 			 rsm, rsm->orig_m_len,
7657 			 rsm->soff));
7658 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7659 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7660 		else
7661 			rsm->soff = 0;
7662 		rsm->orig_m_len = rsm->m->m_len;
7663 #ifdef INVARIANTS
7664 	} else if (rsm->m->m_len > rsm->orig_m_len) {
7665 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7666 		      rsm, rsm->m);
7667 #endif
7668 	}
7669 }
7670 
7671 static void
7672 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7673 {
7674 	struct mbuf *m;
7675 	uint32_t soff;
7676 
7677 	if (src_rsm->m &&
7678 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7679 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7680 		/* Fix up the orig_m_len and possibly the mbuf offset */
7681 		rack_adjust_orig_mlen(src_rsm);
7682 	}
7683 	m = src_rsm->m;
7684 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7685 	while (soff >= m->m_len) {
7686 		/* Move out past this mbuf */
7687 		soff -= m->m_len;
7688 		m = m->m_next;
7689 		KASSERT((m != NULL),
7690 			("rsm:%p nrsm:%p hit at soff:%u null m",
7691 			 src_rsm, rsm, soff));
7692 		if (m == NULL) {
7693 			/* This should *not* happen which is why there is a kassert */
7694 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7695 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7696 					       &src_rsm->soff);
7697 			src_rsm->orig_m_len = src_rsm->m->m_len;
7698 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7699 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7700 					   (rsm->r_start - rack->rc_tp->snd_una),
7701 					   &rsm->soff);
7702 			rsm->orig_m_len = rsm->m->m_len;
7703 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7704 			return;
7705 		}
7706 	}
7707 	rsm->m = m;
7708 	rsm->soff = soff;
7709 	rsm->orig_m_len = m->m_len;
7710 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7711 }
7712 
7713 static __inline void
7714 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7715 	       struct rack_sendmap *rsm, uint32_t start)
7716 {
7717 	int idx;
7718 
7719 	nrsm->r_start = start;
7720 	nrsm->r_end = rsm->r_end;
7721 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7722 	nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7723 	nrsm->r_flags = rsm->r_flags;
7724 	nrsm->r_dupack = rsm->r_dupack;
7725 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7726 	nrsm->r_rtr_bytes = 0;
7727 	nrsm->r_fas = rsm->r_fas;
7728 	nrsm->r_bas = rsm->r_bas;
7729 	tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7730 	nrsm->r_just_ret = rsm->r_just_ret;
7731 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7732 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7733 	}
7734 	/* Now if we have SYN flag we keep it on the left edge */
7735 	if (nrsm->r_flags & RACK_HAS_SYN)
7736 		nrsm->r_flags &= ~RACK_HAS_SYN;
7737 	/* Now if we have a FIN flag we keep it on the right edge */
7738 	if (rsm->r_flags & RACK_HAS_FIN)
7739 		rsm->r_flags &= ~RACK_HAS_FIN;
7740 	/* Push bit must go to the right edge as well */
7741 	if (rsm->r_flags & RACK_HAD_PUSH)
7742 		rsm->r_flags &= ~RACK_HAD_PUSH;
7743 	/* Clone over the state of the hw_tls flag */
7744 	nrsm->r_hw_tls = rsm->r_hw_tls;
7745 	/*
7746 	 * Now we need to find nrsm's new location in the mbuf chain
7747 	 * we basically calculate a new offset, which is soff +
7748 	 * how much is left in original rsm. Then we walk out the mbuf
7749 	 * chain to find the righ position, it may be the same mbuf
7750 	 * or maybe not.
7751 	 */
7752 	KASSERT(((rsm->m != NULL) ||
7753 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7754 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7755 	if (rsm->m)
7756 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7757 }
7758 
7759 static struct rack_sendmap *
7760 rack_merge_rsm(struct tcp_rack *rack,
7761 	       struct rack_sendmap *l_rsm,
7762 	       struct rack_sendmap *r_rsm)
7763 {
7764 	/*
7765 	 * We are merging two ack'd RSM's,
7766 	 * the l_rsm is on the left (lower seq
7767 	 * values) and the r_rsm is on the right
7768 	 * (higher seq value). The simplest way
7769 	 * to merge these is to move the right
7770 	 * one into the left. I don't think there
7771 	 * is any reason we need to try to find
7772 	 * the oldest (or last oldest retransmitted).
7773 	 */
7774 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7775 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7776 	tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7777 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7778 		l_rsm->r_dupack = r_rsm->r_dupack;
7779 	if (r_rsm->r_rtr_bytes)
7780 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7781 	if (r_rsm->r_in_tmap) {
7782 		/* This really should not happen */
7783 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7784 		r_rsm->r_in_tmap = 0;
7785 	}
7786 
7787 	/* Now the flags */
7788 	if (r_rsm->r_flags & RACK_HAS_FIN)
7789 		l_rsm->r_flags |= RACK_HAS_FIN;
7790 	if (r_rsm->r_flags & RACK_TLP)
7791 		l_rsm->r_flags |= RACK_TLP;
7792 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7793 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7794 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7795 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7796 		/*
7797 		 * If both are app-limited then let the
7798 		 * free lower the count. If right is app
7799 		 * limited and left is not, transfer.
7800 		 */
7801 		l_rsm->r_flags |= RACK_APP_LIMITED;
7802 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7803 		if (r_rsm == rack->r_ctl.rc_first_appl)
7804 			rack->r_ctl.rc_first_appl = l_rsm;
7805 	}
7806 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7807 	/*
7808 	 * We keep the largest value, which is the newest
7809 	 * send. We do this in case a segment that is
7810 	 * joined together and not part of a GP estimate
7811 	 * later gets expanded into the GP estimate.
7812 	 *
7813 	 * We prohibit the merging of unlike kinds i.e.
7814 	 * all pieces that are in the GP estimate can be
7815 	 * merged and all pieces that are not in a GP estimate
7816 	 * can be merged, but not disimilar pieces. Combine
7817 	 * this with taking the highest here and we should
7818 	 * be ok unless of course the client reneges. Then
7819 	 * all bets are off.
7820 	 */
7821 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7822 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7823 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7824 	}
7825 	/*
7826 	 * When merging two RSM's we also need to consider the ack time and keep
7827 	 * newest. If the ack gets merged into a measurement then that is the
7828 	 * one we will want to be using.
7829 	 */
7830 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7831 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7832 
7833 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7834 		/* Transfer the split limit to the map we free */
7835 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7836 		l_rsm->r_limit_type = 0;
7837 	}
7838 	rack_free(rack, r_rsm);
7839 	l_rsm->r_flags |= RACK_MERGED;
7840 	return (l_rsm);
7841 }
7842 
7843 /*
7844  * TLP Timer, here we simply setup what segment we want to
7845  * have the TLP expire on, the normal rack_output() will then
7846  * send it out.
7847  *
7848  * We return 1, saying don't proceed with rack_output only
7849  * when all timers have been stopped (destroyed PCB?).
7850  */
7851 static int
7852 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7853 {
7854 	/*
7855 	 * Tail Loss Probe.
7856 	 */
7857 	struct rack_sendmap *rsm = NULL;
7858 	int insret __diagused;
7859 	struct socket *so = tptosocket(tp);
7860 	uint32_t amm;
7861 	uint32_t out, avail;
7862 	int collapsed_win = 0;
7863 
7864 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7865 		/* Its not time yet */
7866 		return (0);
7867 	}
7868 	if (ctf_progress_timeout_check(tp, true)) {
7869 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7870 		return (-ETIMEDOUT);	/* tcp_drop() */
7871 	}
7872 	/*
7873 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7874 	 * need to figure out how to force a full MSS segment out.
7875 	 */
7876 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7877 	rack->r_ctl.retran_during_recovery = 0;
7878 	rack->r_might_revert = 0;
7879 	rack->r_ctl.dsack_byte_cnt = 0;
7880 	counter_u64_add(rack_tlp_tot, 1);
7881 	if (rack->r_state && (rack->r_state != tp->t_state))
7882 		rack_set_state(tp, rack);
7883 	avail = sbavail(&so->so_snd);
7884 	out = tp->snd_max - tp->snd_una;
7885 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7886 		/* special case, we need a retransmission */
7887 		collapsed_win = 1;
7888 		goto need_retran;
7889 	}
7890 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7891 		rack->r_ctl.dsack_persist--;
7892 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7893 			rack->r_ctl.num_dsack = 0;
7894 		}
7895 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7896 	}
7897 	if ((tp->t_flags & TF_GPUTINPROG) &&
7898 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7899 		/*
7900 		 * If this is the second in a row
7901 		 * TLP and we are doing a measurement
7902 		 * its time to abandon the measurement.
7903 		 * Something is likely broken on
7904 		 * the clients network and measuring a
7905 		 * broken network does us no good.
7906 		 */
7907 		tp->t_flags &= ~TF_GPUTINPROG;
7908 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7909 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7910 					   tp->gput_seq,
7911 					   0, 0, 18, __LINE__, NULL, 0);
7912 	}
7913 	/*
7914 	 * Check our send oldest always settings, and if
7915 	 * there is an oldest to send jump to the need_retran.
7916 	 */
7917 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7918 		goto need_retran;
7919 
7920 	if (avail > out) {
7921 		/* New data is available */
7922 		amm = avail - out;
7923 		if (amm > ctf_fixed_maxseg(tp)) {
7924 			amm = ctf_fixed_maxseg(tp);
7925 			if ((amm + out) > tp->snd_wnd) {
7926 				/* We are rwnd limited */
7927 				goto need_retran;
7928 			}
7929 		} else if (amm < ctf_fixed_maxseg(tp)) {
7930 			/* not enough to fill a MTU */
7931 			goto need_retran;
7932 		}
7933 		if (IN_FASTRECOVERY(tp->t_flags)) {
7934 			/* Unlikely */
7935 			if (rack->rack_no_prr == 0) {
7936 				if (out + amm <= tp->snd_wnd) {
7937 					rack->r_ctl.rc_prr_sndcnt = amm;
7938 					rack->r_ctl.rc_tlp_new_data = amm;
7939 					rack_log_to_prr(rack, 4, 0, __LINE__);
7940 				}
7941 			} else
7942 				goto need_retran;
7943 		} else {
7944 			/* Set the send-new override */
7945 			if (out + amm <= tp->snd_wnd)
7946 				rack->r_ctl.rc_tlp_new_data = amm;
7947 			else
7948 				goto need_retran;
7949 		}
7950 		rack->r_ctl.rc_tlpsend = NULL;
7951 		counter_u64_add(rack_tlp_newdata, 1);
7952 		goto send;
7953 	}
7954 need_retran:
7955 	/*
7956 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7957 	 * optionally the first un-acked segment.
7958 	 */
7959 	if (collapsed_win == 0) {
7960 		if (rack_always_send_oldest)
7961 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7962 		else {
7963 			rsm = tqhash_max(rack->r_ctl.tqh);
7964 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7965 				rsm = rack_find_high_nonack(rack, rsm);
7966 			}
7967 		}
7968 		if (rsm == NULL) {
7969 #ifdef TCP_BLACKBOX
7970 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7971 #endif
7972 			goto out;
7973 		}
7974 	} else {
7975 		/*
7976 		 * We had a collapsed window, lets find
7977 		 * the point before the collapse.
7978 		 */
7979 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7980 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7981 		else {
7982 			rsm = tqhash_min(rack->r_ctl.tqh);
7983 		}
7984 		if (rsm == NULL) {
7985 			/* Huh */
7986 			goto out;
7987 		}
7988 	}
7989 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7990 		/*
7991 		 * We need to split this the last segment in two.
7992 		 */
7993 		struct rack_sendmap *nrsm;
7994 
7995 		nrsm = rack_alloc_full_limit(rack);
7996 		if (nrsm == NULL) {
7997 			/*
7998 			 * No memory to split, we will just exit and punt
7999 			 * off to the RXT timer.
8000 			 */
8001 			goto out;
8002 		}
8003 		rack_clone_rsm(rack, nrsm, rsm,
8004 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
8005 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8006 #ifndef INVARIANTS
8007 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8008 #else
8009 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8010 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8011 			      nrsm, insret, rack, rsm);
8012 		}
8013 #endif
8014 		if (rsm->r_in_tmap) {
8015 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8016 			nrsm->r_in_tmap = 1;
8017 		}
8018 		rsm = nrsm;
8019 	}
8020 	rack->r_ctl.rc_tlpsend = rsm;
8021 send:
8022 	/* Make sure output path knows we are doing a TLP */
8023 	*doing_tlp = 1;
8024 	rack->r_timer_override = 1;
8025 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
8026 	return (0);
8027 out:
8028 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
8029 	return (0);
8030 }
8031 
8032 /*
8033  * Delayed ack Timer, here we simply need to setup the
8034  * ACK_NOW flag and remove the DELACK flag. From there
8035  * the output routine will send the ack out.
8036  *
8037  * We only return 1, saying don't proceed, if all timers
8038  * are stopped (destroyed PCB?).
8039  */
8040 static int
8041 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8042 {
8043 
8044 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
8045 	tp->t_flags &= ~TF_DELACK;
8046 	tp->t_flags |= TF_ACKNOW;
8047 	KMOD_TCPSTAT_INC(tcps_delack);
8048 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
8049 	return (0);
8050 }
8051 
8052 static inline int
8053 rack_send_ack_challange(struct tcp_rack *rack)
8054 {
8055 	struct tcptemp *t_template;
8056 
8057 	t_template = tcpip_maketemplate(rack->rc_inp);
8058 	if (t_template) {
8059 		if (rack->forced_ack == 0) {
8060 			rack->forced_ack = 1;
8061 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
8062 		} else {
8063 			rack->probe_not_answered = 1;
8064 		}
8065 		tcp_respond(rack->rc_tp, t_template->tt_ipgen,
8066 			    &t_template->tt_t, (struct mbuf *)NULL,
8067 			    rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
8068 		free(t_template, M_TEMP);
8069 		/* This does send an ack so kill any D-ack timer */
8070 		if (rack->rc_tp->t_flags & TF_DELACK)
8071 			rack->rc_tp->t_flags &= ~TF_DELACK;
8072 		return(1);
8073 	} else
8074 		return (0);
8075 
8076 }
8077 
8078 /*
8079  * Persists timer, here we simply send the
8080  * same thing as a keepalive will.
8081  * the one byte send.
8082  *
8083  * We only return 1, saying don't proceed, if all timers
8084  * are stopped (destroyed PCB?).
8085  */
8086 static int
8087 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8088 {
8089 	int32_t retval = 1;
8090 
8091 	if (rack->rc_in_persist == 0)
8092 		return (0);
8093 	if (ctf_progress_timeout_check(tp, false)) {
8094 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
8095 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
8096 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
8097 		return (-ETIMEDOUT);	/* tcp_drop() */
8098 	}
8099 	/*
8100 	 * Persistence timer into zero window. Force a byte to be output, if
8101 	 * possible.
8102 	 */
8103 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
8104 	/*
8105 	 * Hack: if the peer is dead/unreachable, we do not time out if the
8106 	 * window is closed.  After a full backoff, drop the connection if
8107 	 * the idle time (no responses to probes) reaches the maximum
8108 	 * backoff that we would use if retransmitting.
8109 	 */
8110 	if (tp->t_rxtshift >= V_tcp_retries &&
8111 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
8112 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
8113 		KMOD_TCPSTAT_INC(tcps_persistdrop);
8114 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
8115 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
8116 		retval = -ETIMEDOUT;	/* tcp_drop() */
8117 		goto out;
8118 	}
8119 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
8120 	    tp->snd_una == tp->snd_max)
8121 		rack_exit_persist(tp, rack, cts);
8122 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
8123 	/*
8124 	 * If the user has closed the socket then drop a persisting
8125 	 * connection after a much reduced timeout.
8126 	 */
8127 	if (tp->t_state > TCPS_CLOSE_WAIT &&
8128 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
8129 		KMOD_TCPSTAT_INC(tcps_persistdrop);
8130 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
8131 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
8132 		retval = -ETIMEDOUT;	/* tcp_drop() */
8133 		goto out;
8134 	}
8135 	if (rack_send_ack_challange(rack)) {
8136 		/* only set it if we were answered */
8137 		if (rack->probe_not_answered) {
8138 			counter_u64_add(rack_persists_loss, 1);
8139 			rack->r_ctl.persist_lost_ends++;
8140 		}
8141 		counter_u64_add(rack_persists_sends, 1);
8142 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
8143 	}
8144 	if (tp->t_rxtshift < V_tcp_retries)
8145 		tp->t_rxtshift++;
8146 out:
8147 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
8148 	rack_start_hpts_timer(rack, tp, cts,
8149 			      0, 0, 0);
8150 	return (retval);
8151 }
8152 
8153 /*
8154  * If a keepalive goes off, we had no other timers
8155  * happening. We always return 1 here since this
8156  * routine either drops the connection or sends
8157  * out a segment with respond.
8158  */
8159 static int
8160 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8161 {
8162 	struct inpcb *inp = tptoinpcb(tp);
8163 
8164 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
8165 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
8166 	/*
8167 	 * Keep-alive timer went off; send something or drop connection if
8168 	 * idle for too long.
8169 	 */
8170 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
8171 	if (tp->t_state < TCPS_ESTABLISHED)
8172 		goto dropit;
8173 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
8174 	    tp->t_state <= TCPS_CLOSING) {
8175 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
8176 			goto dropit;
8177 		/*
8178 		 * Send a packet designed to force a response if the peer is
8179 		 * up and reachable: either an ACK if the connection is
8180 		 * still alive, or an RST if the peer has closed the
8181 		 * connection due to timeout or reboot. Using sequence
8182 		 * number tp->snd_una-1 causes the transmitted zero-length
8183 		 * segment to lie outside the receive window; by the
8184 		 * protocol spec, this requires the correspondent TCP to
8185 		 * respond.
8186 		 */
8187 		KMOD_TCPSTAT_INC(tcps_keepprobe);
8188 		rack_send_ack_challange(rack);
8189 	}
8190 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
8191 	return (1);
8192 dropit:
8193 	KMOD_TCPSTAT_INC(tcps_keepdrops);
8194 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
8195 	return (-ETIMEDOUT);	/* tcp_drop() */
8196 }
8197 
8198 /*
8199  * Retransmit helper function, clear up all the ack
8200  * flags and take care of important book keeping.
8201  */
8202 static void
8203 rack_remxt_tmr(struct tcpcb *tp)
8204 {
8205 	/*
8206 	 * The retransmit timer went off, all sack'd blocks must be
8207 	 * un-acked.
8208 	 */
8209 	struct rack_sendmap *rsm, *trsm = NULL;
8210 	struct tcp_rack *rack;
8211 
8212 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8213 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
8214 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
8215 	rack->r_timer_override = 1;
8216 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
8217 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
8218 	rack->r_late = 0;
8219 	rack->r_early = 0;
8220 	rack->r_ctl.rc_agg_delayed = 0;
8221 	rack->r_ctl.rc_agg_early = 0;
8222 	if (rack->r_state && (rack->r_state != tp->t_state))
8223 		rack_set_state(tp, rack);
8224 	if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
8225 		/*
8226 		 * We do not clear the scoreboard until we have had
8227 		 * more than rack_rxt_scoreboard_clear_thresh time-outs.
8228 		 */
8229 		rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
8230 		if (rack->r_ctl.rc_resend != NULL)
8231 			rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
8232 
8233 		return;
8234 	}
8235 	/*
8236 	 * Ideally we would like to be able to
8237 	 * mark SACK-PASS on anything not acked here.
8238 	 *
8239 	 * However, if we do that we would burst out
8240 	 * all that data 1ms apart. This would be unwise,
8241 	 * so for now we will just let the normal rxt timer
8242 	 * and tlp timer take care of it.
8243 	 *
8244 	 * Also we really need to stick them back in sequence
8245 	 * order. This way we send in the proper order and any
8246 	 * sacks that come floating in will "re-ack" the data.
8247 	 * To do this we zap the tmap with an INIT and then
8248 	 * walk through and place every rsm in the tail queue
8249 	 * hash table back in its seq ordered place.
8250 	 */
8251 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
8252 
8253 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8254 		rsm->r_dupack = 0;
8255 		if (rack_verbose_logging)
8256 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8257 		/* We must re-add it back to the tlist */
8258 		if (trsm == NULL) {
8259 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8260 		} else {
8261 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
8262 		}
8263 		rsm->r_in_tmap = 1;
8264 		trsm = rsm;
8265 		if (rsm->r_flags & RACK_ACKED)
8266 			rsm->r_flags |= RACK_WAS_ACKED;
8267 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
8268 		rsm->r_flags |= RACK_MUST_RXT;
8269 	}
8270 	/* zero the lost since it's all gone */
8271 	rack->r_ctl.rc_considered_lost = 0;
8272 	/* Clear the count (we just un-acked them) */
8273 	rack->r_ctl.rc_sacked = 0;
8274 	rack->r_ctl.rc_sacklast = NULL;
8275 	/* Clear the tlp rtx mark */
8276 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
8277 	if (rack->r_ctl.rc_resend != NULL)
8278 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
8279 	rack->r_ctl.rc_prr_sndcnt = 0;
8280 	rack_log_to_prr(rack, 6, 0, __LINE__);
8281 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
8282 	if (rack->r_ctl.rc_resend != NULL)
8283 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
8284 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
8285 #ifdef TCP_SAD_DETECTION
8286 	     || (rack->sack_attack_disable != 0)
8287 #endif
8288 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
8289 		/*
8290 		 * For non-sack customers new data
8291 		 * needs to go out as retransmits until
8292 		 * we retransmit up to snd_max.
8293 		 */
8294 		rack->r_must_retran = 1;
8295 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
8296 							    rack->r_ctl.rc_sacked);
8297 	}
8298 }
8299 
8300 static void
8301 rack_convert_rtts(struct tcpcb *tp)
8302 {
8303 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
8304 	tp->t_rxtcur = RACK_REXMTVAL(tp);
8305 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
8306 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
8307 	}
8308 	if (tp->t_rxtcur > rack_rto_max) {
8309 		tp->t_rxtcur = rack_rto_max;
8310 	}
8311 }
8312 
8313 static void
8314 rack_cc_conn_init(struct tcpcb *tp)
8315 {
8316 	struct tcp_rack *rack;
8317 	uint32_t srtt;
8318 
8319 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8320 	srtt = tp->t_srtt;
8321 	cc_conn_init(tp);
8322 	/*
8323 	 * Now convert to rack's internal format,
8324 	 * if required.
8325 	 */
8326 	if ((srtt == 0) && (tp->t_srtt != 0))
8327 		rack_convert_rtts(tp);
8328 	/*
8329 	 * We want a chance to stay in slowstart as
8330 	 * we create a connection. TCP spec says that
8331 	 * initially ssthresh is infinite. For our
8332 	 * purposes that is the snd_wnd.
8333 	 */
8334 	if (tp->snd_ssthresh < tp->snd_wnd) {
8335 		tp->snd_ssthresh = tp->snd_wnd;
8336 	}
8337 	/*
8338 	 * We also want to assure a IW worth of
8339 	 * data can get inflight.
8340 	 */
8341 	if (rc_init_window(rack) < tp->snd_cwnd)
8342 		tp->snd_cwnd = rc_init_window(rack);
8343 }
8344 
8345 /*
8346  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
8347  * we will setup to retransmit the lowest seq number outstanding.
8348  */
8349 static int
8350 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
8351 {
8352 	struct inpcb *inp = tptoinpcb(tp);
8353 	int32_t rexmt;
8354 	int32_t retval = 0;
8355 	bool isipv6;
8356 
8357 	if ((tp->t_flags & TF_GPUTINPROG) &&
8358 	    (tp->t_rxtshift)) {
8359 		/*
8360 		 * We have had a second timeout
8361 		 * measurements on successive rxt's are not profitable.
8362 		 * It is unlikely to be of any use (the network is
8363 		 * broken or the client went away).
8364 		 */
8365 		tp->t_flags &= ~TF_GPUTINPROG;
8366 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8367 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
8368 					   tp->gput_seq,
8369 					   0, 0, 18, __LINE__, NULL, 0);
8370 	}
8371 	if (ctf_progress_timeout_check(tp, false)) {
8372 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
8373 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
8374 		return (-ETIMEDOUT);	/* tcp_drop() */
8375 	}
8376 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
8377 	rack->r_ctl.retran_during_recovery = 0;
8378 	rack->rc_ack_required = 1;
8379 	rack->r_ctl.dsack_byte_cnt = 0;
8380 	if (IN_RECOVERY(tp->t_flags) &&
8381 	    (rack->rto_from_rec == 0)) {
8382 		/*
8383 		 * Mark that we had a rto while in recovery
8384 		 * and save the ssthresh so if we go back
8385 		 * into recovery we will have a chance
8386 		 * to slowstart back to the level.
8387 		 */
8388 		rack->rto_from_rec = 1;
8389 		rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
8390 	}
8391 	if (IN_FASTRECOVERY(tp->t_flags))
8392 		tp->t_flags |= TF_WASFRECOVERY;
8393 	else
8394 		tp->t_flags &= ~TF_WASFRECOVERY;
8395 	if (IN_CONGRECOVERY(tp->t_flags))
8396 		tp->t_flags |= TF_WASCRECOVERY;
8397 	else
8398 		tp->t_flags &= ~TF_WASCRECOVERY;
8399 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
8400 	    (tp->snd_una == tp->snd_max)) {
8401 		/* Nothing outstanding .. nothing to do */
8402 		return (0);
8403 	}
8404 	if (rack->r_ctl.dsack_persist) {
8405 		rack->r_ctl.dsack_persist--;
8406 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
8407 			rack->r_ctl.num_dsack = 0;
8408 		}
8409 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
8410 	}
8411 	/*
8412 	 * Rack can only run one timer  at a time, so we cannot
8413 	 * run a KEEPINIT (gating SYN sending) and a retransmit
8414 	 * timer for the SYN. So if we are in a front state and
8415 	 * have a KEEPINIT timer we need to check the first transmit
8416 	 * against now to see if we have exceeded the KEEPINIT time
8417 	 * (if one is set).
8418 	 */
8419 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
8420 	    (TP_KEEPINIT(tp) != 0)) {
8421 		struct rack_sendmap *rsm;
8422 
8423 		rsm = tqhash_min(rack->r_ctl.tqh);
8424 		if (rsm) {
8425 			/* Ok we have something outstanding to test keepinit with */
8426 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
8427 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
8428 				/* We have exceeded the KEEPINIT time */
8429 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
8430 				goto drop_it;
8431 			}
8432 		}
8433 	}
8434 	/*
8435 	 * Retransmission timer went off.  Message has not been acked within
8436 	 * retransmit interval.  Back off to a longer retransmit interval
8437 	 * and retransmit one segment.
8438 	 */
8439 	if ((rack->r_ctl.rc_resend == NULL) ||
8440 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
8441 		/*
8442 		 * If the rwnd collapsed on
8443 		 * the one we are retransmitting
8444 		 * it does not count against the
8445 		 * rxt count.
8446 		 */
8447 		tp->t_rxtshift++;
8448 	}
8449 	rack_remxt_tmr(tp);
8450 	if (tp->t_rxtshift > V_tcp_retries) {
8451 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
8452 drop_it:
8453 		tp->t_rxtshift = V_tcp_retries;
8454 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
8455 		/* XXXGL: previously t_softerror was casted to uint16_t */
8456 		MPASS(tp->t_softerror >= 0);
8457 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
8458 		goto out;	/* tcp_drop() */
8459 	}
8460 	if (tp->t_state == TCPS_SYN_SENT) {
8461 		/*
8462 		 * If the SYN was retransmitted, indicate CWND to be limited
8463 		 * to 1 segment in cc_conn_init().
8464 		 */
8465 		tp->snd_cwnd = 1;
8466 	} else if (tp->t_rxtshift == 1) {
8467 		/*
8468 		 * first retransmit; record ssthresh and cwnd so they can be
8469 		 * recovered if this turns out to be a "bad" retransmit. A
8470 		 * retransmit is considered "bad" if an ACK for this segment
8471 		 * is received within RTT/2 interval; the assumption here is
8472 		 * that the ACK was already in flight.  See "On Estimating
8473 		 * End-to-End Network Path Properties" by Allman and Paxson
8474 		 * for more details.
8475 		 */
8476 		tp->snd_cwnd_prev = tp->snd_cwnd;
8477 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
8478 		tp->snd_recover_prev = tp->snd_recover;
8479 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
8480 		tp->t_flags |= TF_PREVVALID;
8481 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
8482 		tp->t_flags &= ~TF_PREVVALID;
8483 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
8484 	if ((tp->t_state == TCPS_SYN_SENT) ||
8485 	    (tp->t_state == TCPS_SYN_RECEIVED))
8486 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
8487 	else
8488 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
8489 
8490 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
8491 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
8492 	/*
8493 	 * We enter the path for PLMTUD if connection is established or, if
8494 	 * connection is FIN_WAIT_1 status, reason for the last is that if
8495 	 * amount of data we send is very small, we could send it in couple
8496 	 * of packets and process straight to FIN. In that case we won't
8497 	 * catch ESTABLISHED state.
8498 	 */
8499 #ifdef INET6
8500 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
8501 #else
8502 	isipv6 = false;
8503 #endif
8504 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
8505 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
8506 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
8507 	    ((tp->t_state == TCPS_ESTABLISHED) ||
8508 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
8509 		/*
8510 		 * Idea here is that at each stage of mtu probe (usually,
8511 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
8512 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
8513 		 * should take care of that.
8514 		 */
8515 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
8516 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
8517 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
8518 		    tp->t_rxtshift % 2 == 0)) {
8519 			/*
8520 			 * Enter Path MTU Black-hole Detection mechanism: -
8521 			 * Disable Path MTU Discovery (IP "DF" bit). -
8522 			 * Reduce MTU to lower value than what we negotiated
8523 			 * with peer.
8524 			 */
8525 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
8526 				/* Record that we may have found a black hole. */
8527 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
8528 				/* Keep track of previous MSS. */
8529 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
8530 			}
8531 
8532 			/*
8533 			 * Reduce the MSS to blackhole value or to the
8534 			 * default in an attempt to retransmit.
8535 			 */
8536 #ifdef INET6
8537 			if (isipv6 &&
8538 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
8539 				/* Use the sysctl tuneable blackhole MSS. */
8540 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
8541 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
8542 			} else if (isipv6) {
8543 				/* Use the default MSS. */
8544 				tp->t_maxseg = V_tcp_v6mssdflt;
8545 				/*
8546 				 * Disable Path MTU Discovery when we switch
8547 				 * to minmss.
8548 				 */
8549 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8550 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8551 			}
8552 #endif
8553 #if defined(INET6) && defined(INET)
8554 			else
8555 #endif
8556 #ifdef INET
8557 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
8558 				/* Use the sysctl tuneable blackhole MSS. */
8559 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
8560 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
8561 			} else {
8562 				/* Use the default MSS. */
8563 				tp->t_maxseg = V_tcp_mssdflt;
8564 				/*
8565 				 * Disable Path MTU Discovery when we switch
8566 				 * to minmss.
8567 				 */
8568 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8569 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8570 			}
8571 #endif
8572 		} else {
8573 			/*
8574 			 * If further retransmissions are still unsuccessful
8575 			 * with a lowered MTU, maybe this isn't a blackhole
8576 			 * and we restore the previous MSS and blackhole
8577 			 * detection flags. The limit '6' is determined by
8578 			 * giving each probe stage (1448, 1188, 524) 2
8579 			 * chances to recover.
8580 			 */
8581 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
8582 			    (tp->t_rxtshift >= 6)) {
8583 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8584 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
8585 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
8586 				if (tp->t_maxseg < V_tcp_mssdflt) {
8587 					/*
8588 					 * The MSS is so small we should not
8589 					 * process incoming SACK's since we are
8590 					 * subject to attack in such a case.
8591 					 */
8592 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
8593 				} else {
8594 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
8595 				}
8596 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
8597 			}
8598 		}
8599 	}
8600 	/*
8601 	 * Disable RFC1323 and SACK if we haven't got any response to
8602 	 * our third SYN to work-around some broken terminal servers
8603 	 * (most of which have hopefully been retired) that have bad VJ
8604 	 * header compression code which trashes TCP segments containing
8605 	 * unknown-to-them TCP options.
8606 	 */
8607 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
8608 	    (tp->t_rxtshift == 3))
8609 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
8610 	/*
8611 	 * If we backed off this far, our srtt estimate is probably bogus.
8612 	 * Clobber it so we'll take the next rtt measurement as our srtt;
8613 	 * move the current srtt into rttvar to keep the current retransmit
8614 	 * times until then.
8615 	 */
8616 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
8617 #ifdef INET6
8618 		if ((inp->inp_vflag & INP_IPV6) != 0)
8619 			in6_losing(inp);
8620 		else
8621 #endif
8622 			in_losing(inp);
8623 		tp->t_rttvar += tp->t_srtt;
8624 		tp->t_srtt = 0;
8625 	}
8626 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8627 	tp->snd_recover = tp->snd_max;
8628 	tp->t_flags |= TF_ACKNOW;
8629 	tp->t_rtttime = 0;
8630 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
8631 out:
8632 	return (retval);
8633 }
8634 
8635 static int
8636 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
8637 {
8638 	int32_t ret = 0;
8639 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
8640 
8641 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8642 	    (tp->t_flags & TF_GPUTINPROG)) {
8643 		/*
8644 		 * We have a goodput in progress
8645 		 * and we have entered a late state.
8646 		 * Do we have enough data in the sb
8647 		 * to handle the GPUT request?
8648 		 */
8649 		uint32_t bytes;
8650 
8651 		bytes = tp->gput_ack - tp->gput_seq;
8652 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
8653 			bytes += tp->gput_seq - tp->snd_una;
8654 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
8655 			/*
8656 			 * There are not enough bytes in the socket
8657 			 * buffer that have been sent to cover this
8658 			 * measurement. Cancel it.
8659 			 */
8660 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8661 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
8662 						   tp->gput_seq,
8663 						   0, 0, 18, __LINE__, NULL, 0);
8664 			tp->t_flags &= ~TF_GPUTINPROG;
8665 		}
8666 	}
8667 	if (timers == 0) {
8668 		return (0);
8669 	}
8670 	if (tp->t_state == TCPS_LISTEN) {
8671 		/* no timers on listen sockets */
8672 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
8673 			return (0);
8674 		return (1);
8675 	}
8676 	if ((timers & PACE_TMR_RACK) &&
8677 	    rack->rc_on_min_to) {
8678 		/*
8679 		 * For the rack timer when we
8680 		 * are on a min-timeout (which means rrr_conf = 3)
8681 		 * we don't want to check the timer. It may
8682 		 * be going off for a pace and thats ok we
8683 		 * want to send the retransmit (if its ready).
8684 		 *
8685 		 * If its on a normal rack timer (non-min) then
8686 		 * we will check if its expired.
8687 		 */
8688 		goto skip_time_check;
8689 	}
8690 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8691 		uint32_t left;
8692 
8693 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8694 			ret = -1;
8695 			rack_log_to_processing(rack, cts, ret, 0);
8696 			return (0);
8697 		}
8698 		if (hpts_calling == 0) {
8699 			/*
8700 			 * A user send or queued mbuf (sack) has called us? We
8701 			 * return 0 and let the pacing guards
8702 			 * deal with it if they should or
8703 			 * should not cause a send.
8704 			 */
8705 			ret = -2;
8706 			rack_log_to_processing(rack, cts, ret, 0);
8707 			return (0);
8708 		}
8709 		/*
8710 		 * Ok our timer went off early and we are not paced false
8711 		 * alarm, go back to sleep. We make sure we don't have
8712 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8713 		 * flag in place.
8714 		 */
8715 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8716 		ret = -3;
8717 		left = rack->r_ctl.rc_timer_exp - cts;
8718 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8719 		rack_log_to_processing(rack, cts, ret, left);
8720 		return (1);
8721 	}
8722 skip_time_check:
8723 	rack->rc_tmr_stopped = 0;
8724 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8725 	if (timers & PACE_TMR_DELACK) {
8726 		ret = rack_timeout_delack(tp, rack, cts);
8727 	} else if (timers & PACE_TMR_RACK) {
8728 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8729 		rack->r_fast_output = 0;
8730 		ret = rack_timeout_rack(tp, rack, cts);
8731 	} else if (timers & PACE_TMR_TLP) {
8732 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8733 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8734 	} else if (timers & PACE_TMR_RXT) {
8735 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8736 		rack->r_fast_output = 0;
8737 		ret = rack_timeout_rxt(tp, rack, cts);
8738 	} else if (timers & PACE_TMR_PERSIT) {
8739 		ret = rack_timeout_persist(tp, rack, cts);
8740 	} else if (timers & PACE_TMR_KEEP) {
8741 		ret = rack_timeout_keepalive(tp, rack, cts);
8742 	}
8743 	rack_log_to_processing(rack, cts, ret, timers);
8744 	return (ret);
8745 }
8746 
8747 static void
8748 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8749 {
8750 	struct timeval tv;
8751 	uint32_t us_cts, flags_on_entry;
8752 	uint8_t hpts_removed = 0;
8753 
8754 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8755 	us_cts = tcp_get_usecs(&tv);
8756 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8757 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8758 	     ((tp->snd_max - tp->snd_una) == 0))) {
8759 		tcp_hpts_remove(rack->rc_tp);
8760 		hpts_removed = 1;
8761 		/* If we were not delayed cancel out the flag. */
8762 		if ((tp->snd_max - tp->snd_una) == 0)
8763 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8764 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8765 	}
8766 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8767 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8768 		if (tcp_in_hpts(rack->rc_tp) &&
8769 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8770 			/*
8771 			 * Canceling timer's when we have no output being
8772 			 * paced. We also must remove ourselves from the
8773 			 * hpts.
8774 			 */
8775 			tcp_hpts_remove(rack->rc_tp);
8776 			hpts_removed = 1;
8777 		}
8778 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8779 	}
8780 	if (hpts_removed == 0)
8781 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8782 }
8783 
8784 static int
8785 rack_stopall(struct tcpcb *tp)
8786 {
8787 	struct tcp_rack *rack;
8788 
8789 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8790 	rack->t_timers_stopped = 1;
8791 
8792 	tcp_hpts_remove(tp);
8793 
8794 	return (0);
8795 }
8796 
8797 static void
8798 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8799 {
8800 	/*
8801 	 * Assure no timers are running.
8802 	 */
8803 	if (tcp_timer_active(tp, TT_PERSIST)) {
8804 		/* We enter in persists, set the flag appropriately */
8805 		rack->rc_in_persist = 1;
8806 	}
8807 	if (tcp_in_hpts(rack->rc_tp)) {
8808 		tcp_hpts_remove(rack->rc_tp);
8809 	}
8810 }
8811 
8812 /*
8813  * We maintain an array fo 16 (RETRAN_CNT_SIZE) entries. This
8814  * array is zeroed at the start of recovery. Each time a segment
8815  * is retransmitted, we translate that into a number of packets
8816  * (based on segsiz) and based on how many times its been retransmitted
8817  * increment by the number of packets the counter that represents
8818  * retansmitted N times. Index 0 is retransmitted 1 time, index 1
8819  * is retransmitted 2 times etc.
8820  *
8821  * So for example when we send a 4344 byte transmission with a 1448
8822  * byte segsize, and its the third time we have retransmitted this
8823  * segment, we would add to the rc_cnt_of_retran[2] the value of
8824  * 3. That represents 3 MSS were retransmitted 3 times (index is
8825  * the number of times retranmitted minus 1).
8826  */
8827 static void
8828 rack_peg_rxt(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t segsiz)
8829 {
8830 	int idx;
8831 	uint32_t peg;
8832 
8833 	peg = ((rsm->r_end - rsm->r_start) + segsiz) - 1;
8834 	peg /= segsiz;
8835 	idx = rsm->r_act_rxt_cnt - 1;
8836 	if (idx >= RETRAN_CNT_SIZE)
8837 		idx = RETRAN_CNT_SIZE - 1;
8838 	/* Max of a uint16_t retransmits in a bucket */
8839 	if ((rack->r_ctl.rc_cnt_of_retran[idx] + peg) < 0xffff)
8840 		rack->r_ctl.rc_cnt_of_retran[idx] += peg;
8841 	else
8842 		rack->r_ctl.rc_cnt_of_retran[idx] = 0xffff;
8843 }
8844 
8845 /*
8846  * We maintain an array fo 16 (RETRAN_CNT_SIZE) entries. This
8847  * array is zeroed at the start of recovery. Each time a segment
8848  * is retransmitted, we translate that into a number of packets
8849  * (based on segsiz) and based on how many times its been retransmitted
8850  * increment by the number of packets the counter that represents
8851  * retansmitted N times. Index 0 is retransmitted 1 time, index 1
8852  * is retransmitted 2 times etc.
8853  *
8854  * The rack_unpeg_rxt is used when we go to retransmit a segment
8855  * again. Basically if the segment had previously been retransmitted
8856  * say 3 times (as our previous example illustrated in the comment
8857  * above rack_peg_rxt() prior to calling that and incrementing
8858  * r_ack_rxt_cnt we would have called rack_unpeg_rxt() that would
8859  * subtract back the previous add from its last rxt (in this
8860  * example r_act_cnt would have been 2 for 2 retransmissions. So
8861  * we would have subtracted 3 from rc_cnt_of_reetran[1] to remove
8862  * those 3 segments. You will see this in the rack_update_rsm()
8863  * below where we do:
8864  *	if (rsm->r_act_rxt_cnt > 0) {
8865  *		rack_unpeg_rxt(rack, rsm, segsiz);
8866  *	}
8867  *	rsm->r_act_rxt_cnt++;
8868  *	rack_peg_rxt(rack, rsm, segsiz);
8869  *
8870  * This effectively moves the count from rc_cnt_of_retran[1] to
8871  * rc_cnt_of_retran[2].
8872  */
8873 static void
8874 rack_unpeg_rxt(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t segsiz)
8875 {
8876 	int idx;
8877 	uint32_t peg;
8878 
8879 	idx = rsm->r_act_rxt_cnt - 1;
8880 	if (idx >= RETRAN_CNT_SIZE)
8881 		idx = RETRAN_CNT_SIZE - 1;
8882 	peg = ((rsm->r_end - rsm->r_start) + segsiz) - 1;
8883 	peg /= segsiz;
8884 	if (peg < rack->r_ctl.rc_cnt_of_retran[idx])
8885 		rack->r_ctl.rc_cnt_of_retran[idx] -= peg;
8886 	else {
8887 		/* TSNH */
8888 		rack->r_ctl.rc_cnt_of_retran[idx] = 0;
8889 	}
8890 }
8891 
8892 static void
8893 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8894     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8895 {
8896 	int32_t idx;
8897 
8898 	rsm->r_rtr_cnt++;
8899 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8900 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8901 		rsm->r_flags |= RACK_OVERMAX;
8902 	}
8903 	if (rsm->r_act_rxt_cnt > 0) {
8904 		/* Drop the count back for this, its retransmitting again */
8905 		rack_unpeg_rxt(rack, rsm, segsiz);
8906 	}
8907 	rsm->r_act_rxt_cnt++;
8908 	/* Peg the count/index */
8909 	rack_peg_rxt(rack, rsm, segsiz);
8910 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8911 	rsm->r_dupack = 0;
8912 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8913 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8914 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8915 	}
8916 	if (rsm->r_flags & RACK_WAS_LOST) {
8917 		/*
8918 		 * We retransmitted it putting it back in flight
8919 		 * remove the lost desgination and reduce the
8920 		 * bytes considered lost.
8921 		 */
8922 		rsm->r_flags  &= ~RACK_WAS_LOST;
8923 		KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8924 			("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
8925 		if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8926 			rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8927 		else
8928 			rack->r_ctl.rc_considered_lost = 0;
8929 	}
8930 	idx = rsm->r_rtr_cnt - 1;
8931 	rsm->r_tim_lastsent[idx] = ts;
8932 	/*
8933 	 * Here we don't add in the len of send, since its already
8934 	 * in snduna <->snd_max.
8935 	 */
8936 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8937 				     rack->r_ctl.rc_sacked);
8938 	if (rsm->r_flags & RACK_ACKED) {
8939 		/* Problably MTU discovery messing with us */
8940 		rsm->r_flags &= ~RACK_ACKED;
8941 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8942 	}
8943 	if (rsm->r_in_tmap) {
8944 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8945 		rsm->r_in_tmap = 0;
8946 	}
8947 	/* Lets make sure it really is in or not the GP window */
8948 	rack_mark_in_gp_win(tp, rsm);
8949 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8950 	rsm->r_in_tmap = 1;
8951 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8952 	/* Take off the must retransmit flag, if its on */
8953 	if (rsm->r_flags & RACK_MUST_RXT) {
8954 		if (rack->r_must_retran)
8955 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8956 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8957 			/*
8958 			 * We have retransmitted all we need. Clear
8959 			 * any must retransmit flags.
8960 			 */
8961 			rack->r_must_retran = 0;
8962 			rack->r_ctl.rc_out_at_rto = 0;
8963 		}
8964 		rsm->r_flags &= ~RACK_MUST_RXT;
8965 	}
8966 	/* Remove any collapsed flag */
8967 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8968 	if (rsm->r_flags & RACK_SACK_PASSED) {
8969 		/* We have retransmitted due to the SACK pass */
8970 		rsm->r_flags &= ~RACK_SACK_PASSED;
8971 		rsm->r_flags |= RACK_WAS_SACKPASS;
8972 	}
8973 }
8974 
8975 static uint32_t
8976 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8977     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8978 {
8979 	/*
8980 	 * We (re-)transmitted starting at rsm->r_start for some length
8981 	 * (possibly less than r_end.
8982 	 */
8983 	struct rack_sendmap *nrsm;
8984 	int insret __diagused;
8985 	uint32_t c_end;
8986 	int32_t len;
8987 
8988 	len = *lenp;
8989 	c_end = rsm->r_start + len;
8990 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8991 		/*
8992 		 * We retransmitted the whole piece or more than the whole
8993 		 * slopping into the next rsm.
8994 		 */
8995 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8996 		if (c_end == rsm->r_end) {
8997 			*lenp = 0;
8998 			return (0);
8999 		} else {
9000 			int32_t act_len;
9001 
9002 			/* Hangs over the end return whats left */
9003 			act_len = rsm->r_end - rsm->r_start;
9004 			*lenp = (len - act_len);
9005 			return (rsm->r_end);
9006 		}
9007 		/* We don't get out of this block. */
9008 	}
9009 	/*
9010 	 * Here we retransmitted less than the whole thing which means we
9011 	 * have to split this into what was transmitted and what was not.
9012 	 */
9013 	nrsm = rack_alloc_full_limit(rack);
9014 	if (nrsm == NULL) {
9015 		/*
9016 		 * We can't get memory, so lets not proceed.
9017 		 */
9018 		*lenp = 0;
9019 		return (0);
9020 	}
9021 	/*
9022 	 * So here we are going to take the original rsm and make it what we
9023 	 * retransmitted. nrsm will be the tail portion we did not
9024 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
9025 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
9026 	 * 1, 6 and the new piece will be 6, 11.
9027 	 */
9028 	rack_clone_rsm(rack, nrsm, rsm, c_end);
9029 	nrsm->r_dupack = 0;
9030 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
9031 #ifndef INVARIANTS
9032 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9033 #else
9034 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9035 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9036 		      nrsm, insret, rack, rsm);
9037 	}
9038 #endif
9039 	if (rsm->r_in_tmap) {
9040 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9041 		nrsm->r_in_tmap = 1;
9042 	}
9043 	rsm->r_flags &= (~RACK_HAS_FIN);
9044 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
9045 	/* Log a split of rsm into rsm and nrsm */
9046 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
9047 	*lenp = 0;
9048 	return (0);
9049 }
9050 
9051 static void
9052 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
9053 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
9054 		struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
9055 		uint32_t s_moff, int hw_tls, int segsiz)
9056 {
9057 	struct tcp_rack *rack;
9058 	struct rack_sendmap *rsm, *nrsm;
9059 	int insret __diagused;
9060 
9061 	register uint32_t snd_max, snd_una;
9062 
9063 	/*
9064 	 * Add to the RACK log of packets in flight or retransmitted. If
9065 	 * there is a TS option we will use the TS echoed, if not we will
9066 	 * grab a TS.
9067 	 *
9068 	 * Retransmissions will increment the count and move the ts to its
9069 	 * proper place. Note that if options do not include TS's then we
9070 	 * won't be able to effectively use the ACK for an RTT on a retran.
9071 	 *
9072 	 * Notes about r_start and r_end. Lets consider a send starting at
9073 	 * sequence 1 for 10 bytes. In such an example the r_start would be
9074 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
9075 	 * This means that r_end is actually the first sequence for the next
9076 	 * slot (11).
9077 	 *
9078 	 */
9079 	/*
9080 	 * If err is set what do we do XXXrrs? should we not add the thing?
9081 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
9082 	 * i.e. proceed with add ** do this for now.
9083 	 */
9084 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9085 	if (err)
9086 		/*
9087 		 * We don't log errors -- we could but snd_max does not
9088 		 * advance in this case either.
9089 		 */
9090 		return;
9091 
9092 	if (th_flags & TH_RST) {
9093 		/*
9094 		 * We don't log resets and we return immediately from
9095 		 * sending
9096 		 */
9097 		return;
9098 	}
9099 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9100 	snd_una = tp->snd_una;
9101 	snd_max = tp->snd_max;
9102 	if (th_flags & (TH_SYN | TH_FIN)) {
9103 		/*
9104 		 * The call to rack_log_output is made before bumping
9105 		 * snd_max. This means we can record one extra byte on a SYN
9106 		 * or FIN if seq_out is adding more on and a FIN is present
9107 		 * (and we are not resending).
9108 		 */
9109 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
9110 			len++;
9111 		if (th_flags & TH_FIN)
9112 			len++;
9113 	}
9114 	if (SEQ_LEQ((seq_out + len), snd_una)) {
9115 		/* Are sending an old segment to induce an ack (keep-alive)? */
9116 		return;
9117 	}
9118 	if (SEQ_LT(seq_out, snd_una)) {
9119 		/* huh? should we panic? */
9120 		uint32_t end;
9121 
9122 		end = seq_out + len;
9123 		seq_out = snd_una;
9124 		if (SEQ_GEQ(end, seq_out))
9125 			len = end - seq_out;
9126 		else
9127 			len = 0;
9128 	}
9129 	if (len == 0) {
9130 		/* We don't log zero window probes */
9131 		return;
9132 	}
9133 	if (IN_FASTRECOVERY(tp->t_flags)) {
9134 		rack->r_ctl.rc_prr_out += len;
9135 	}
9136 	/* First question is it a retransmission or new? */
9137 	if (seq_out == snd_max) {
9138 		/* Its new */
9139 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
9140 again:
9141 		rsm = rack_alloc(rack);
9142 		if (rsm == NULL) {
9143 			/*
9144 			 * Hmm out of memory and the tcb got destroyed while
9145 			 * we tried to wait.
9146 			 */
9147 			return;
9148 		}
9149 		if (th_flags & TH_FIN) {
9150 			rsm->r_flags = RACK_HAS_FIN|add_flag;
9151 		} else {
9152 			rsm->r_flags = add_flag;
9153 		}
9154 		if (hw_tls)
9155 			rsm->r_hw_tls = 1;
9156 		rsm->r_tim_lastsent[0] = cts;
9157 		rsm->r_rtr_cnt = 1;
9158  		rsm->r_act_rxt_cnt = 0;
9159 		rsm->r_rtr_bytes = 0;
9160 		if (th_flags & TH_SYN) {
9161 			/* The data space is one beyond snd_una */
9162 			rsm->r_flags |= RACK_HAS_SYN;
9163 		}
9164 		rsm->r_start = seq_out;
9165 		rsm->r_end = rsm->r_start + len;
9166 		rack_mark_in_gp_win(tp, rsm);
9167 		rsm->r_dupack = 0;
9168 		/*
9169 		 * save off the mbuf location that
9170 		 * sndmbuf_noadv returned (which is
9171 		 * where we started copying from)..
9172 		 */
9173 		rsm->m = s_mb;
9174 		rsm->soff = s_moff;
9175 		/*
9176 		 * Here we do add in the len of send, since its not yet
9177 		 * reflected in in snduna <->snd_max
9178 		 */
9179 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
9180 					      rack->r_ctl.rc_sacked) +
9181 			      (rsm->r_end - rsm->r_start));
9182 		if ((rack->rc_initial_ss_comp == 0) &&
9183 		    (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
9184 			   rack->r_ctl.ss_hi_fs = rsm->r_fas;
9185 		}
9186 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
9187 		if (rsm->m) {
9188 			if (rsm->m->m_len <= rsm->soff) {
9189 				/*
9190 				 * XXXrrs Question, will this happen?
9191 				 *
9192 				 * If sbsndptr is set at the correct place
9193 				 * then s_moff should always be somewhere
9194 				 * within rsm->m. But if the sbsndptr was
9195 				 * off then that won't be true. If it occurs
9196 				 * we need to walkout to the correct location.
9197 				 */
9198 				struct mbuf *lm;
9199 
9200 				lm = rsm->m;
9201 				while (lm->m_len <= rsm->soff) {
9202 					rsm->soff -= lm->m_len;
9203 					lm = lm->m_next;
9204 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
9205 							     __func__, rack, s_moff, s_mb, rsm->soff));
9206 				}
9207 				rsm->m = lm;
9208 			}
9209 			rsm->orig_m_len = rsm->m->m_len;
9210 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
9211 		} else {
9212 			rsm->orig_m_len = 0;
9213 			rsm->orig_t_space = 0;
9214 		}
9215 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
9216 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9217 		/* Log a new rsm */
9218 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
9219 #ifndef INVARIANTS
9220 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
9221 #else
9222 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
9223 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9224 			      nrsm, insret, rack, rsm);
9225 		}
9226 #endif
9227 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9228 		rsm->r_in_tmap = 1;
9229 		if (rsm->r_flags & RACK_IS_PCM) {
9230 			rack->r_ctl.pcm_i.send_time = cts;
9231 			rack->r_ctl.pcm_i.eseq = rsm->r_end;
9232 			/* First time through we set the start too */
9233 			if (rack->pcm_in_progress == 0)
9234 				rack->r_ctl.pcm_i.sseq = rsm->r_start;
9235 		}
9236 		/*
9237 		 * Special case detection, is there just a single
9238 		 * packet outstanding when we are not in recovery?
9239 		 *
9240 		 * If this is true mark it so.
9241 		 */
9242 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9243 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
9244 			struct rack_sendmap *prsm;
9245 
9246 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
9247 			if (prsm)
9248 				prsm->r_one_out_nr = 1;
9249 		}
9250 		return;
9251 	}
9252 	/*
9253 	 * If we reach here its a retransmission and we need to find it.
9254 	 */
9255 more:
9256 	if (hintrsm && (hintrsm->r_start == seq_out)) {
9257 		rsm = hintrsm;
9258 		hintrsm = NULL;
9259 	} else {
9260 		/* No hints sorry */
9261 		rsm = NULL;
9262 	}
9263 	if ((rsm) && (rsm->r_start == seq_out)) {
9264 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
9265 		if (len == 0) {
9266 			return;
9267 		} else {
9268 			goto more;
9269 		}
9270 	}
9271 	/* Ok it was not the last pointer go through it the hard way. */
9272 refind:
9273 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
9274 	if (rsm) {
9275 		if (rsm->r_start == seq_out) {
9276 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
9277 			if (len == 0) {
9278 				return;
9279 			} else {
9280 				goto refind;
9281 			}
9282 		}
9283 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
9284 			/* Transmitted within this piece */
9285 			/*
9286 			 * Ok we must split off the front and then let the
9287 			 * update do the rest
9288 			 */
9289 			nrsm = rack_alloc_full_limit(rack);
9290 			if (nrsm == NULL) {
9291 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
9292 				return;
9293 			}
9294 			/*
9295 			 * copy rsm to nrsm and then trim the front of rsm
9296 			 * to not include this part.
9297 			 */
9298 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
9299 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
9300 #ifndef INVARIANTS
9301 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9302 #else
9303 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9304 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9305 				      nrsm, insret, rack, rsm);
9306 			}
9307 #endif
9308 			if (rsm->r_in_tmap) {
9309 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9310 				nrsm->r_in_tmap = 1;
9311 			}
9312 			rsm->r_flags &= (~RACK_HAS_FIN);
9313 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
9314 			if (len == 0) {
9315 				return;
9316 			} else if (len > 0)
9317 				goto refind;
9318 		}
9319 	}
9320 	/*
9321 	 * Hmm not found in map did they retransmit both old and on into the
9322 	 * new?
9323 	 */
9324 	if (seq_out == tp->snd_max) {
9325 		goto again;
9326 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
9327 #ifdef INVARIANTS
9328 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
9329 		       seq_out, len, tp->snd_una, tp->snd_max);
9330 		printf("Starting Dump of all rack entries\n");
9331 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
9332 			printf("rsm:%p start:%u end:%u\n",
9333 			       rsm, rsm->r_start, rsm->r_end);
9334 		}
9335 		printf("Dump complete\n");
9336 		panic("seq_out not found rack:%p tp:%p",
9337 		      rack, tp);
9338 #endif
9339 	} else {
9340 #ifdef INVARIANTS
9341 		/*
9342 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
9343 		 * flag)
9344 		 */
9345 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
9346 		      seq_out, len, tp->snd_max, tp);
9347 #endif
9348 	}
9349 }
9350 
9351 /*
9352  * Record one of the RTT updates from an ack into
9353  * our sample structure.
9354  */
9355 
9356 static void
9357 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
9358 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
9359 {
9360 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
9361 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
9362 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
9363 	}
9364 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
9365 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
9366 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
9367 	}
9368 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
9369 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
9370 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
9371 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
9372 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
9373 	}
9374 	if ((confidence == 1) &&
9375 	    ((rsm == NULL) ||
9376 	     (rsm->r_just_ret) ||
9377 	     (rsm->r_one_out_nr &&
9378 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
9379 		/*
9380 		 * If the rsm had a just return
9381 		 * hit it then we can't trust the
9382 		 * rtt measurement for buffer deterimination
9383 		 * Note that a confidence of 2, indicates
9384 		 * SACK'd which overrides the r_just_ret or
9385 		 * the r_one_out_nr. If it was a CUM-ACK and
9386 		 * we had only two outstanding, but get an
9387 		 * ack for only 1. Then that also lowers our
9388 		 * confidence.
9389 		 */
9390 		confidence = 0;
9391 	}
9392 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
9393 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
9394 		if (rack->r_ctl.rack_rs.confidence == 0) {
9395 			/*
9396 			 * We take anything with no current confidence
9397 			 * saved.
9398 			 */
9399 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
9400 			rack->r_ctl.rack_rs.confidence = confidence;
9401 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
9402 		} else if (confidence != 0) {
9403 			/*
9404 			 * Once we have a confident number,
9405 			 * we can update it with a smaller
9406 			 * value since this confident number
9407 			 * may include the DSACK time until
9408 			 * the next segment (the second one) arrived.
9409 			 */
9410 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
9411 			rack->r_ctl.rack_rs.confidence = confidence;
9412 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
9413 		}
9414 	}
9415 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
9416 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
9417 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
9418 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
9419 }
9420 
9421 /*
9422  * Collect new round-trip time estimate
9423  * and update averages and current timeout.
9424  */
9425 static void
9426 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
9427 {
9428 	int32_t delta;
9429 	int32_t rtt;
9430 
9431 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
9432 		/* No valid sample */
9433 		return;
9434 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
9435 		/* We are to use the lowest RTT seen in a single ack */
9436 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9437 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
9438 		/* We are to use the highest RTT seen in a single ack */
9439 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
9440 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
9441 		/* We are to use the average RTT seen in a single ack */
9442 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
9443 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
9444 	} else {
9445 #ifdef INVARIANTS
9446 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
9447 #endif
9448 		return;
9449 	}
9450 	if (rtt == 0)
9451 		rtt = 1;
9452 	if (rack->rc_gp_rtt_set == 0) {
9453 		/*
9454 		 * With no RTT we have to accept
9455 		 * even one we are not confident of.
9456 		 */
9457 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
9458 		rack->rc_gp_rtt_set = 1;
9459 	} else if (rack->r_ctl.rack_rs.confidence) {
9460 		/* update the running gp srtt */
9461 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
9462 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
9463 	}
9464 	if (rack->r_ctl.rack_rs.confidence) {
9465 		/*
9466 		 * record the low and high for highly buffered path computation,
9467 		 * we only do this if we are confident (not a retransmission).
9468 		 */
9469 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
9470 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9471 		}
9472 		if (rack->rc_highly_buffered == 0) {
9473 			/*
9474 			 * Currently once we declare a path has
9475 			 * highly buffered there is no going
9476 			 * back, which may be a problem...
9477 			 */
9478 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
9479 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
9480 						     rack->r_ctl.rc_highest_us_rtt,
9481 						     rack->r_ctl.rc_lowest_us_rtt,
9482 						     RACK_RTTS_SEEHBP);
9483 				rack->rc_highly_buffered = 1;
9484 			}
9485 		}
9486 	}
9487 	if ((rack->r_ctl.rack_rs.confidence) ||
9488 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
9489 		/*
9490 		 * If we are highly confident of it <or> it was
9491 		 * never retransmitted we accept it as the last us_rtt.
9492 		 */
9493 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9494 		/* The lowest rtt can be set if its was not retransmited */
9495 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
9496 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9497 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
9498 				rack->r_ctl.rc_lowest_us_rtt = 1;
9499 		}
9500 	}
9501 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9502 	if (tp->t_srtt != 0) {
9503 		/*
9504 		 * We keep a simple srtt in microseconds, like our rtt
9505 		 * measurement. We don't need to do any tricks with shifting
9506 		 * etc. Instead we just add in 1/8th of the new measurement
9507 		 * and subtract out 1/8 of the old srtt. We do the same with
9508 		 * the variance after finding the absolute value of the
9509 		 * difference between this sample and the current srtt.
9510 		 */
9511 		delta = tp->t_srtt - rtt;
9512 		/* Take off 1/8th of the current sRTT */
9513 		tp->t_srtt -= (tp->t_srtt >> 3);
9514 		/* Add in 1/8th of the new RTT just measured */
9515 		tp->t_srtt += (rtt >> 3);
9516 		if (tp->t_srtt <= 0)
9517 			tp->t_srtt = 1;
9518 		/* Now lets make the absolute value of the variance */
9519 		if (delta < 0)
9520 			delta = -delta;
9521 		/* Subtract out 1/8th */
9522 		tp->t_rttvar -= (tp->t_rttvar >> 3);
9523 		/* Add in 1/8th of the new variance we just saw */
9524 		tp->t_rttvar += (delta >> 3);
9525 		if (tp->t_rttvar <= 0)
9526 			tp->t_rttvar = 1;
9527 	} else {
9528 		/*
9529 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
9530 		 * variance to half the rtt (so our first retransmit happens
9531 		 * at 3*rtt).
9532 		 */
9533 		tp->t_srtt = rtt;
9534 		tp->t_rttvar = rtt >> 1;
9535 	}
9536 	rack->rc_srtt_measure_made = 1;
9537 	KMOD_TCPSTAT_INC(tcps_rttupdated);
9538 	if (tp->t_rttupdated < UCHAR_MAX)
9539 		tp->t_rttupdated++;
9540 #ifdef STATS
9541 	if (rack_stats_gets_ms_rtt == 0) {
9542 		/* Send in the microsecond rtt used for rxt timeout purposes */
9543 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
9544 	} else if (rack_stats_gets_ms_rtt == 1) {
9545 		/* Send in the millisecond rtt used for rxt timeout purposes */
9546 		int32_t ms_rtt;
9547 
9548 		/* Round up */
9549 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
9550 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
9551 	} else if (rack_stats_gets_ms_rtt == 2) {
9552 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
9553 		int32_t ms_rtt;
9554 
9555 		/* Round up */
9556 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
9557 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
9558 	}  else {
9559 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
9560 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
9561 	}
9562 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
9563 #endif
9564 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
9565 	/*
9566 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
9567 	 * way we do the smoothing, srtt and rttvar will each average +1/2
9568 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
9569 	 * tick of rounding and 1 extra tick because of +-1/2 tick
9570 	 * uncertainty in the firing of the timer.  The bias will give us
9571 	 * exactly the 1.5 tick we need.  But, because the bias is
9572 	 * statistical, we have to test that we don't drop below the minimum
9573 	 * feasible timer (which is 2 ticks).
9574 	 */
9575 	tp->t_rxtshift = 0;
9576 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9577 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
9578 	rack_log_rtt_sample(rack, rtt);
9579 	tp->t_softerror = 0;
9580 }
9581 
9582 
9583 static void
9584 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
9585 {
9586 	/*
9587 	 * Apply to filter the inbound us-rtt at us_cts.
9588 	 */
9589 	uint32_t old_rtt;
9590 
9591 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
9592 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
9593 			       us_rtt, us_cts);
9594 	if (old_rtt > us_rtt) {
9595 		/* We just hit a new lower rtt time */
9596 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
9597 				     __LINE__, RACK_RTTS_NEWRTT);
9598 		/*
9599 		 * Only count it if its lower than what we saw within our
9600 		 * calculated range.
9601 		 */
9602 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
9603 			if (rack_probertt_lower_within &&
9604 			    rack->rc_gp_dyn_mul &&
9605 			    (rack->use_fixed_rate == 0) &&
9606 			    (rack->rc_always_pace)) {
9607 				/*
9608 				 * We are seeing a new lower rtt very close
9609 				 * to the time that we would have entered probe-rtt.
9610 				 * This is probably due to the fact that a peer flow
9611 				 * has entered probe-rtt. Lets go in now too.
9612 				 */
9613 				uint32_t val;
9614 
9615 				val = rack_probertt_lower_within * rack_time_between_probertt;
9616 				val /= 100;
9617 				if ((rack->in_probe_rtt == 0)  &&
9618 				    (rack->rc_skip_timely == 0) &&
9619 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
9620 					rack_enter_probertt(rack, us_cts);
9621 				}
9622 			}
9623 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
9624 		}
9625 	}
9626 }
9627 
9628 static int
9629 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
9630     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
9631 {
9632 	uint32_t us_rtt;
9633 	int32_t i, all;
9634 	uint32_t t, len_acked;
9635 
9636 	if ((rsm->r_flags & RACK_ACKED) ||
9637 	    (rsm->r_flags & RACK_WAS_ACKED))
9638 		/* Already done */
9639 		return (0);
9640 	if (rsm->r_no_rtt_allowed) {
9641 		/* Not allowed */
9642 		return (0);
9643 	}
9644 	if (ack_type == CUM_ACKED) {
9645 		if (SEQ_GT(th_ack, rsm->r_end)) {
9646 			len_acked = rsm->r_end - rsm->r_start;
9647 			all = 1;
9648 		} else {
9649 			len_acked = th_ack - rsm->r_start;
9650 			all = 0;
9651 		}
9652 	} else {
9653 		len_acked = rsm->r_end - rsm->r_start;
9654 		all = 0;
9655 	}
9656 	if (rsm->r_rtr_cnt == 1) {
9657 
9658 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9659 		if ((int)t <= 0)
9660 			t = 1;
9661 		if (!tp->t_rttlow || tp->t_rttlow > t)
9662 			tp->t_rttlow = t;
9663 		if (!rack->r_ctl.rc_rack_min_rtt ||
9664 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9665 			rack->r_ctl.rc_rack_min_rtt = t;
9666 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
9667 				rack->r_ctl.rc_rack_min_rtt = 1;
9668 			}
9669 		}
9670 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
9671 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9672 		else
9673 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9674 		if (us_rtt == 0)
9675 			us_rtt = 1;
9676 		if (CC_ALGO(tp)->rttsample != NULL) {
9677 			/* Kick the RTT to the CC */
9678 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9679 		}
9680 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
9681 		if (ack_type == SACKED) {
9682 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
9683 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
9684 		} else {
9685 			/*
9686 			 * We need to setup what our confidence
9687 			 * is in this ack.
9688 			 *
9689 			 * If the rsm was app limited and it is
9690 			 * less than a mss in length (the end
9691 			 * of the send) then we have a gap. If we
9692 			 * were app limited but say we were sending
9693 			 * multiple MSS's then we are more confident
9694 			 * int it.
9695 			 *
9696 			 * When we are not app-limited then we see if
9697 			 * the rsm is being included in the current
9698 			 * measurement, we tell this by the app_limited_needs_set
9699 			 * flag.
9700 			 *
9701 			 * Note that being cwnd blocked is not applimited
9702 			 * as well as the pacing delay between packets which
9703 			 * are sending only 1 or 2 MSS's also will show up
9704 			 * in the RTT. We probably need to examine this algorithm
9705 			 * a bit more and enhance it to account for the delay
9706 			 * between rsm's. We could do that by saving off the
9707 			 * pacing delay of each rsm (in an rsm) and then
9708 			 * factoring that in somehow though for now I am
9709 			 * not sure how :)
9710 			 */
9711 			int calc_conf = 0;
9712 
9713 			if (rsm->r_flags & RACK_APP_LIMITED) {
9714 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
9715 					calc_conf = 0;
9716 				else
9717 					calc_conf = 1;
9718 			} else if (rack->app_limited_needs_set == 0) {
9719 				calc_conf = 1;
9720 			} else {
9721 				calc_conf = 0;
9722 			}
9723 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
9724 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
9725 					    calc_conf, rsm, rsm->r_rtr_cnt);
9726 		}
9727 		if ((rsm->r_flags & RACK_TLP) &&
9728 		    (!IN_FASTRECOVERY(tp->t_flags))) {
9729 			/* Segment was a TLP and our retrans matched */
9730 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
9731 				rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
9732 			}
9733 		}
9734 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9735 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9736 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9737 			/* New more recent rack_tmit_time */
9738 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9739 			if (rack->r_ctl.rc_rack_tmit_time == 0)
9740 				rack->r_ctl.rc_rack_tmit_time = 1;
9741 			rack->rc_rack_rtt = t;
9742 		}
9743 		return (1);
9744 	}
9745 	/*
9746 	 * We clear the soft/rxtshift since we got an ack.
9747 	 * There is no assurance we will call the commit() function
9748 	 * so we need to clear these to avoid incorrect handling.
9749 	 */
9750 	tp->t_rxtshift = 0;
9751 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9752 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9753 	tp->t_softerror = 0;
9754 	if (to && (to->to_flags & TOF_TS) &&
9755 	    (ack_type == CUM_ACKED) &&
9756 	    (to->to_tsecr) &&
9757 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
9758 		/*
9759 		 * Now which timestamp does it match? In this block the ACK
9760 		 * must be coming from a previous transmission.
9761 		 */
9762 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9763 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9764 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9765 				if ((int)t <= 0)
9766 					t = 1;
9767 				if (CC_ALGO(tp)->rttsample != NULL) {
9768 					/*
9769 					 * Kick the RTT to the CC, here
9770 					 * we lie a bit in that we know the
9771 					 * retransmission is correct even though
9772 					 * we retransmitted. This is because
9773 					 * we match the timestamps.
9774 					 */
9775 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9776 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9777 					else
9778 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9779 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9780 				}
9781 				if ((i + 1) < rsm->r_rtr_cnt) {
9782 					/*
9783 					 * The peer ack'd from our previous
9784 					 * transmission. We have a spurious
9785 					 * retransmission and thus we dont
9786 					 * want to update our rack_rtt.
9787 					 *
9788 					 * Hmm should there be a CC revert here?
9789 					 *
9790 					 */
9791 					return (0);
9792 				}
9793 				if (!tp->t_rttlow || tp->t_rttlow > t)
9794 					tp->t_rttlow = t;
9795 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9796 					rack->r_ctl.rc_rack_min_rtt = t;
9797 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9798 						rack->r_ctl.rc_rack_min_rtt = 1;
9799 					}
9800 				}
9801 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9802 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9803 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9804 					/* New more recent rack_tmit_time */
9805 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9806 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9807 						rack->r_ctl.rc_rack_tmit_time = 1;
9808 					rack->rc_rack_rtt = t;
9809 				}
9810 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9811 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9812 						    rsm->r_rtr_cnt);
9813 				return (1);
9814 			}
9815 		}
9816 		/* If we are logging log out the sendmap */
9817 		if (tcp_bblogging_on(rack->rc_tp)) {
9818 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9819 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9820 			}
9821 		}
9822 		goto ts_not_found;
9823 	} else {
9824 		/*
9825 		 * Ok its a SACK block that we retransmitted. or a windows
9826 		 * machine without timestamps. We can tell nothing from the
9827 		 * time-stamp since its not there or the time the peer last
9828 		 * received a segment that moved forward its cum-ack point.
9829 		 */
9830 ts_not_found:
9831 		i = rsm->r_rtr_cnt - 1;
9832 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9833 		if ((int)t <= 0)
9834 			t = 1;
9835 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9836 			/*
9837 			 * We retransmitted and the ack came back in less
9838 			 * than the smallest rtt we have observed. We most
9839 			 * likely did an improper retransmit as outlined in
9840 			 * 6.2 Step 2 point 2 in the rack-draft so we
9841 			 * don't want to update our rack_rtt. We in
9842 			 * theory (in future) might want to think about reverting our
9843 			 * cwnd state but we won't for now.
9844 			 */
9845 			return (0);
9846 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9847 			/*
9848 			 * We retransmitted it and the retransmit did the
9849 			 * job.
9850 			 */
9851 			if (!rack->r_ctl.rc_rack_min_rtt ||
9852 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9853 				rack->r_ctl.rc_rack_min_rtt = t;
9854 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9855 					rack->r_ctl.rc_rack_min_rtt = 1;
9856 				}
9857 			}
9858 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9859 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9860 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9861 				/* New more recent rack_tmit_time */
9862 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9863 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9864 					rack->r_ctl.rc_rack_tmit_time = 1;
9865 				rack->rc_rack_rtt = t;
9866 			}
9867 			return (1);
9868 		}
9869 	}
9870 	return (0);
9871 }
9872 
9873 /*
9874  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9875  */
9876 static void
9877 rack_log_sack_passed(struct tcpcb *tp,
9878     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9879 {
9880 	struct rack_sendmap *nrsm;
9881 	uint32_t thresh;
9882 
9883 	/* Get our rxt threshold for lost consideration */
9884 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9885 	/* Now start looking at rsm's */
9886 	nrsm = rsm;
9887 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9888 	    rack_head, r_tnext) {
9889 		if (nrsm == rsm) {
9890 			/* Skip original segment he is acked */
9891 			continue;
9892 		}
9893 		if (nrsm->r_flags & RACK_ACKED) {
9894 			/*
9895 			 * Skip ack'd segments, though we
9896 			 * should not see these, since tmap
9897 			 * should not have ack'd segments.
9898 			 */
9899 			continue;
9900 		}
9901 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9902 			/*
9903 			 * If the peer dropped the rwnd on
9904 			 * these then we don't worry about them.
9905 			 */
9906 			continue;
9907 		}
9908 		/* Check lost state */
9909 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9910 			uint32_t exp;
9911 
9912 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9913 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
9914 				/* We consider it lost */
9915 				nrsm->r_flags |= RACK_WAS_LOST;
9916 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9917 			}
9918 		}
9919 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9920 			/*
9921 			 * We found one that is already marked
9922 			 * passed, we have been here before and
9923 			 * so all others below this are marked.
9924 			 */
9925 			break;
9926 		}
9927 		nrsm->r_flags |= RACK_SACK_PASSED;
9928 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9929 	}
9930 }
9931 
9932 static void
9933 rack_need_set_test(struct tcpcb *tp,
9934 		   struct tcp_rack *rack,
9935 		   struct rack_sendmap *rsm,
9936 		   tcp_seq th_ack,
9937 		   int line,
9938 		   int use_which)
9939 {
9940 	struct rack_sendmap *s_rsm;
9941 
9942 	if ((tp->t_flags & TF_GPUTINPROG) &&
9943 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9944 		/*
9945 		 * We were app limited, and this ack
9946 		 * butts up or goes beyond the point where we want
9947 		 * to start our next measurement. We need
9948 		 * to record the new gput_ts as here and
9949 		 * possibly update the start sequence.
9950 		 */
9951 		uint32_t seq, ts;
9952 
9953 		if (rsm->r_rtr_cnt > 1) {
9954 			/*
9955 			 * This is a retransmit, can we
9956 			 * really make any assessment at this
9957 			 * point?  We are not really sure of
9958 			 * the timestamp, is it this or the
9959 			 * previous transmission?
9960 			 *
9961 			 * Lets wait for something better that
9962 			 * is not retransmitted.
9963 			 */
9964 			return;
9965 		}
9966 		seq = tp->gput_seq;
9967 		ts = tp->gput_ts;
9968 		rack->app_limited_needs_set = 0;
9969 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9970 		/* Do we start at a new end? */
9971 		if ((use_which == RACK_USE_BEG) &&
9972 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9973 			/*
9974 			 * When we get an ACK that just eats
9975 			 * up some of the rsm, we set RACK_USE_BEG
9976 			 * since whats at r_start (i.e. th_ack)
9977 			 * is left unacked and thats where the
9978 			 * measurement now starts.
9979 			 */
9980 			tp->gput_seq = rsm->r_start;
9981 		}
9982 		if ((use_which == RACK_USE_END) &&
9983 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9984 			/*
9985 			 * We use the end when the cumack
9986 			 * is moving forward and completely
9987 			 * deleting the rsm passed so basically
9988 			 * r_end holds th_ack.
9989 			 *
9990 			 * For SACK's we also want to use the end
9991 			 * since this piece just got sacked and
9992 			 * we want to target anything after that
9993 			 * in our measurement.
9994 			 */
9995 			tp->gput_seq = rsm->r_end;
9996 		}
9997 		if (use_which == RACK_USE_END_OR_THACK) {
9998 			/*
9999 			 * special case for ack moving forward,
10000 			 * not a sack, we need to move all the
10001 			 * way up to where this ack cum-ack moves
10002 			 * to.
10003 			 */
10004 			if (SEQ_GT(th_ack, rsm->r_end))
10005 				tp->gput_seq = th_ack;
10006 			else
10007 				tp->gput_seq = rsm->r_end;
10008 		}
10009 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
10010 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
10011 		else
10012 			s_rsm = NULL;
10013 		/*
10014 		 * Pick up the correct send time if we can the rsm passed in
10015 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
10016 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
10017 		 * find a different seq i.e. the next send up.
10018 		 *
10019 		 * If that has not been sent, s_rsm will be NULL and we must
10020 		 * arrange it so this function will get called again by setting
10021 		 * app_limited_needs_set.
10022 		 */
10023 		if (s_rsm)
10024 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
10025 		else {
10026 			/* If we hit here we have to have *not* sent tp->gput_seq */
10027 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
10028 			/* Set it up so we will go through here again */
10029 			rack->app_limited_needs_set = 1;
10030 		}
10031 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
10032 			/*
10033 			 * We moved beyond this guy's range, re-calculate
10034 			 * the new end point.
10035 			 */
10036 			if (rack->rc_gp_filled == 0) {
10037 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
10038 			} else {
10039 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
10040 			}
10041 		}
10042 		/*
10043 		 * We are moving the goal post, we may be able to clear the
10044 		 * measure_saw_probe_rtt flag.
10045 		 */
10046 		if ((rack->in_probe_rtt == 0) &&
10047 		    (rack->measure_saw_probe_rtt) &&
10048 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
10049 			rack->measure_saw_probe_rtt = 0;
10050 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
10051 					   seq, tp->gput_seq,
10052 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
10053 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
10054 					   5, line, NULL, 0);
10055 		if (rack->rc_gp_filled &&
10056 		    ((tp->gput_ack - tp->gput_seq) <
10057 		     max(rc_init_window(rack), (MIN_GP_WIN *
10058 						ctf_fixed_maxseg(tp))))) {
10059 			uint32_t ideal_amount;
10060 
10061 			ideal_amount = rack_get_measure_window(tp, rack);
10062 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
10063 				/*
10064 				 * There is no sense of continuing this measurement
10065 				 * because its too small to gain us anything we
10066 				 * trust. Skip it and that way we can start a new
10067 				 * measurement quicker.
10068 				 */
10069 				tp->t_flags &= ~TF_GPUTINPROG;
10070 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
10071 							   0, 0,
10072 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
10073 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
10074 							   6, __LINE__, NULL, 0);
10075 			} else {
10076 				/*
10077 				 * Reset the window further out.
10078 				 */
10079 				tp->gput_ack = tp->gput_seq + ideal_amount;
10080 			}
10081 		}
10082 		rack_tend_gp_marks(tp, rack);
10083 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
10084 	}
10085 }
10086 
10087 static inline int
10088 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
10089 {
10090 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
10091 		/* Behind our TLP definition or right at */
10092 		return (0);
10093 	}
10094 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
10095 		/* The start is beyond or right at our end of TLP definition */
10096 		return (0);
10097 	}
10098 	/* It has to be a sub-part of the original TLP recorded */
10099 	return (1);
10100 }
10101 
10102 static uint32_t
10103 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
10104 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
10105 		   int *no_extra,
10106 		   int *moved_two, uint32_t segsiz)
10107 {
10108 	uint32_t start, end, changed = 0;
10109 	struct rack_sendmap stack_map;
10110 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
10111 	int insret __diagused;
10112 	int32_t used_ref = 1;
10113 	int moved = 0;
10114 #ifdef TCP_SAD_DETECTION
10115 	int allow_segsiz;
10116 	int first_time_through = 1;
10117 #endif
10118 	int noextra = 0;
10119 	int can_use_hookery = 0;
10120 
10121 	start = sack->start;
10122 	end = sack->end;
10123 	rsm = *prsm;
10124 
10125 #ifdef TCP_SAD_DETECTION
10126 	/*
10127 	 * There are a strange number of proxys and meddle boxes in the world
10128 	 * that seem to cut up segments on different boundaries. This gets us
10129 	 * smaller sacks that are still ok in terms of it being an attacker.
10130 	 * We use the base segsiz to calculate an allowable smallness but
10131 	 * also enforce a min on the segsiz in case it is an attacker playing
10132 	 * games with MSS. So basically if the sack arrives and it is
10133 	 * larger than a worse case 960 bytes, we don't classify the guy
10134 	 * as supicious.
10135 	 */
10136 	allow_segsiz = max(segsiz, 1200) * sad_seg_size_per;
10137 	allow_segsiz /= 1000;
10138 #endif
10139 do_rest_ofb:
10140 	if ((rsm == NULL) ||
10141 	    (SEQ_LT(end, rsm->r_start)) ||
10142 	    (SEQ_GEQ(start, rsm->r_end)) ||
10143 	    (SEQ_LT(start, rsm->r_start))) {
10144 		/*
10145 		 * We are not in the right spot,
10146 		 * find the correct spot in the tree.
10147 		 */
10148 		used_ref = 0;
10149 		rsm = tqhash_find(rack->r_ctl.tqh, start);
10150 		moved++;
10151 	}
10152 	if (rsm == NULL) {
10153 		/* TSNH */
10154 		goto out;
10155 	}
10156 #ifdef TCP_SAD_DETECTION
10157 	/* Now we must check for suspicous activity */
10158 	if ((first_time_through == 1) &&
10159 	    ((end - start) < min((rsm->r_end - rsm->r_start), allow_segsiz)) &&
10160 	    ((rsm->r_flags & RACK_PMTU_CHG) == 0) &&
10161 	    ((rsm->r_flags & RACK_TLP) == 0)) {
10162 		/*
10163 		 * Its less than a full MSS or the segment being acked
10164 		 * this should only happen if the rsm in question had the
10165 		 * r_just_ret flag set <and> the end matches the end of
10166 		 * the rsm block.
10167 		 *
10168 		 * Note we do not look at segments that have had TLP's on
10169 		 * them since we can get un-reported rwnd collapses that
10170 		 * basically we TLP on and then we get back a sack block
10171 		 * that goes from the start to only a small way.
10172 		 *
10173 		 */
10174 		int loss, ok;
10175 
10176 		ok = 0;
10177 		if (SEQ_GEQ(end, rsm->r_end)) {
10178 			if (rsm->r_just_ret == 1) {
10179 				/* This was at the end of a send which is ok */
10180 				ok = 1;
10181 			} else {
10182 				/* A bit harder was it the end of our segment */
10183 				int segs, len;
10184 
10185 				len = (rsm->r_end - rsm->r_start);
10186 				segs = len / segsiz;
10187 				segs *= segsiz;
10188 				if ((segs + (rsm->r_end - start)) == len) {
10189 					/*
10190 					 * So this last bit was the
10191 					 * end of our send if we cut it
10192 					 * up into segsiz pieces so its ok.
10193 					 */
10194 					ok = 1;
10195 				}
10196 			}
10197 		}
10198 		if (ok == 0) {
10199 			/*
10200 			 * This guy is doing something suspicious
10201 			 * lets start detection.
10202 			 */
10203 			if (rack->rc_suspicious == 0) {
10204 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_SUSPECT);
10205 				counter_u64_add(rack_sack_attacks_suspect, 1);
10206 				rack->rc_suspicious = 1;
10207 				rack_log_sad(rack, 4);
10208 				if (tcp_bblogging_on(rack->rc_tp)) {
10209 					union tcp_log_stackspecific log;
10210 					struct timeval tv;
10211 
10212 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
10213 					log.u_bbr.flex1 = end;
10214 					log.u_bbr.flex2 = start;
10215 					log.u_bbr.flex3 = rsm->r_end;
10216 					log.u_bbr.flex4 = rsm->r_start;
10217 					log.u_bbr.flex5 = segsiz;
10218 					log.u_bbr.flex6 = rsm->r_fas;
10219 					log.u_bbr.flex7 = rsm->r_bas;
10220 					log.u_bbr.flex8 = 5;
10221 					log.u_bbr.pkts_out = rsm->r_flags;
10222 					log.u_bbr.bbr_state = rack->rc_suspicious;
10223 					log.u_bbr.bbr_substate = rsm->r_just_ret;
10224 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10225 					log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10226 					TCP_LOG_EVENTP(rack->rc_tp, NULL,
10227 						       &rack->rc_inp->inp_socket->so_rcv,
10228 						       &rack->rc_inp->inp_socket->so_snd,
10229 						       TCP_SAD_DETECTION, 0,
10230 						       0, &log, false, &tv);
10231 				}
10232 			}
10233 			/* You loose some ack count every time you sack
10234 			 * a small bit that is not butting to the end of
10235 			 * what we have sent. This is because we never
10236 			 * send small bits unless its the end of the sb.
10237 			 * Anyone sending a sack that is not at the end
10238 			 * is thus very very suspicious.
10239 			 */
10240 			loss = (segsiz/2) / (end - start);
10241 			if (loss < rack->r_ctl.ack_count)
10242 				rack->r_ctl.ack_count -= loss;
10243 			else
10244 				rack->r_ctl.ack_count = 0;
10245 		}
10246 	}
10247 	first_time_through = 0;
10248 #endif
10249 	/* Ok we have an ACK for some piece of this rsm */
10250 	if (rsm->r_start != start) {
10251 		if ((rsm->r_flags & RACK_ACKED) == 0) {
10252 			/*
10253 			 * Before any splitting or hookery is
10254 			 * done is it a TLP of interest i.e. rxt?
10255 			 */
10256 			if ((rsm->r_flags & RACK_TLP) &&
10257 			    (rsm->r_rtr_cnt > 1)) {
10258 				/*
10259 				 * We are splitting a rxt TLP, check
10260 				 * if we need to save off the start/end
10261 				 */
10262 				if (rack->rc_last_tlp_acked_set &&
10263 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10264 					/*
10265 					 * We already turned this on since we are inside
10266 					 * the previous one was a partially sack now we
10267 					 * are getting another one (maybe all of it).
10268 					 *
10269 					 */
10270 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10271 					/*
10272 					 * Lets make sure we have all of it though.
10273 					 */
10274 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10275 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10276 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10277 								     rack->r_ctl.last_tlp_acked_end);
10278 					}
10279 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10280 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10281 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10282 								     rack->r_ctl.last_tlp_acked_end);
10283 					}
10284 				} else {
10285 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10286 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10287 					rack->rc_last_tlp_past_cumack = 0;
10288 					rack->rc_last_tlp_acked_set = 1;
10289 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10290 				}
10291 			}
10292 			/**
10293 			 * Need to split this in two pieces the before and after,
10294 			 * the before remains in the map, the after must be
10295 			 * added. In other words we have:
10296 			 * rsm        |--------------|
10297 			 * sackblk        |------->
10298 			 * rsm will become
10299 			 *     rsm    |---|
10300 			 * and nrsm will be  the sacked piece
10301 			 *     nrsm       |----------|
10302 			 *
10303 			 * But before we start down that path lets
10304 			 * see if the sack spans over on top of
10305 			 * the next guy and it is already sacked.
10306 			 *
10307 			 */
10308 			/*
10309 			 * Hookery can only be used if the two entries
10310 			 * are in the same bucket and neither one of
10311 			 * them staddle the bucket line.
10312 			 */
10313 			next = tqhash_next(rack->r_ctl.tqh, rsm);
10314 			if (next &&
10315 			    (rsm->bindex == next->bindex) &&
10316 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
10317 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
10318 			    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
10319 			    ((next->r_flags & RACK_IS_PCM) == 0) &&
10320 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
10321 			    (next->r_flags & RACK_IN_GP_WIN))
10322 				can_use_hookery = 1;
10323 			else
10324 				can_use_hookery = 0;
10325 			if (next && can_use_hookery &&
10326 			    (next->r_flags & RACK_ACKED) &&
10327 			    SEQ_GEQ(end, next->r_start)) {
10328 				/**
10329 				 * So the next one is already acked, and
10330 				 * we can thus by hookery use our stack_map
10331 				 * to reflect the piece being sacked and
10332 				 * then adjust the two tree entries moving
10333 				 * the start and ends around. So we start like:
10334 				 *  rsm     |------------|             (not-acked)
10335 				 *  next                 |-----------| (acked)
10336 				 *  sackblk        |-------->
10337 				 *  We want to end like so:
10338 				 *  rsm     |------|                   (not-acked)
10339 				 *  next           |-----------------| (acked)
10340 				 *  nrsm           |-----|
10341 				 * Where nrsm is a temporary stack piece we
10342 				 * use to update all the gizmos.
10343 				 */
10344 				/* Copy up our fudge block */
10345 				noextra++;
10346 				nrsm = &stack_map;
10347 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10348 				/* Now adjust our tree blocks */
10349 				tqhash_update_end(rack->r_ctl.tqh, rsm, start);
10350 				next->r_start = start;
10351  				rsm->r_flags |= RACK_SHUFFLED;
10352 				next->r_flags |= RACK_SHUFFLED;
10353 				/* Now we must adjust back where next->m is */
10354 				rack_setup_offset_for_rsm(rack, rsm, next);
10355 				/*
10356 				 * Which timestamp do we keep? It is rather
10357 				 * important in GP measurements to have the
10358 				 * accurate end of the send window.
10359 				 *
10360 				 * We keep the largest value, which is the newest
10361 				 * send. We do this in case a segment that is
10362 				 * joined together and not part of a GP estimate
10363 				 * later gets expanded into the GP estimate.
10364 				 *
10365 				 * We prohibit the merging of unlike kinds i.e.
10366 				 * all pieces that are in the GP estimate can be
10367 				 * merged and all pieces that are not in a GP estimate
10368 				 * can be merged, but not disimilar pieces. Combine
10369 				 * this with taking the highest here and we should
10370 				 * be ok unless of course the client reneges. Then
10371 				 * all bets are off.
10372 				 */
10373 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
10374 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
10375 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
10376 				/*
10377 				 * And we must keep the newest ack arrival time.
10378 				 */
10379 				if (next->r_ack_arrival <
10380 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10381 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10382 
10383 
10384 				/* We don't need to adjust rsm, it did not change */
10385 				/* Clear out the dup ack count of the remainder */
10386 				rsm->r_dupack = 0;
10387 				rsm->r_just_ret = 0;
10388 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10389 				/* Now lets make sure our fudge block is right */
10390 				nrsm->r_start = start;
10391 				/* Now lets update all the stats and such */
10392 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10393 				if (rack->app_limited_needs_set)
10394 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10395 				changed += (nrsm->r_end - nrsm->r_start);
10396 				/* You get a count for acking a whole segment or more */
10397 				if ((nrsm->r_end - nrsm->r_start) >= segsiz)
10398 					rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
10399 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10400 				if (rsm->r_flags & RACK_WAS_LOST) {
10401 					int my_chg;
10402 
10403 					my_chg = (nrsm->r_end - nrsm->r_start);
10404 					KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10405 						("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10406 					if (my_chg <= rack->r_ctl.rc_considered_lost)
10407 						rack->r_ctl.rc_considered_lost -= my_chg;
10408 					else
10409 						rack->r_ctl.rc_considered_lost = 0;
10410 				}
10411 				if (nrsm->r_flags & RACK_SACK_PASSED) {
10412 					rack->r_ctl.rc_reorder_ts = cts;
10413 					if (rack->r_ctl.rc_reorder_ts == 0)
10414 						rack->r_ctl.rc_reorder_ts = 1;
10415 				}
10416 				/*
10417 				 * Now we want to go up from rsm (the
10418 				 * one left un-acked) to the next one
10419 				 * in the tmap. We do this so when
10420 				 * we walk backwards we include marking
10421 				 * sack-passed on rsm (The one passed in
10422 				 * is skipped since it is generally called
10423 				 * on something sacked before removing it
10424 				 * from the tmap).
10425 				 */
10426 				if (rsm->r_in_tmap) {
10427 					nrsm = TAILQ_NEXT(rsm, r_tnext);
10428 					/*
10429 					 * Now that we have the next
10430 					 * one walk backwards from there.
10431 					 */
10432 					if (nrsm && nrsm->r_in_tmap)
10433 						rack_log_sack_passed(tp, rack, nrsm, cts);
10434 				}
10435 				/* Now are we done? */
10436 				if (SEQ_LT(end, next->r_end) ||
10437 				    (end == next->r_end)) {
10438 					/* Done with block */
10439 					goto out;
10440 				}
10441 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
10442 				counter_u64_add(rack_sack_used_next_merge, 1);
10443 				/* Postion for the next block */
10444 				start = next->r_end;
10445 				rsm = tqhash_next(rack->r_ctl.tqh, next);
10446 				if (rsm == NULL)
10447 					goto out;
10448 			} else {
10449 				/**
10450 				 * We can't use any hookery here, so we
10451 				 * need to split the map. We enter like
10452 				 * so:
10453 				 *  rsm      |--------|
10454 				 *  sackblk       |----->
10455 				 * We will add the new block nrsm and
10456 				 * that will be the new portion, and then
10457 				 * fall through after reseting rsm. So we
10458 				 * split and look like this:
10459 				 *  rsm      |----|
10460 				 *  sackblk       |----->
10461 				 *  nrsm          |---|
10462 				 * We then fall through reseting
10463 				 * rsm to nrsm, so the next block
10464 				 * picks it up.
10465 				 */
10466 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10467 				if (nrsm == NULL) {
10468 					/*
10469 					 * failed XXXrrs what can we do but loose the sack
10470 					 * info?
10471 					 */
10472 					goto out;
10473 				}
10474 				counter_u64_add(rack_sack_splits, 1);
10475 				rack_clone_rsm(rack, nrsm, rsm, start);
10476 				moved++;
10477 				rsm->r_just_ret = 0;
10478 #ifndef INVARIANTS
10479 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10480 #else
10481 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10482 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
10483 					      nrsm, insret, rack, rsm);
10484 				}
10485 #endif
10486 				if (rsm->r_in_tmap) {
10487 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10488 					nrsm->r_in_tmap = 1;
10489 				}
10490 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
10491 				rsm->r_flags &= (~RACK_HAS_FIN);
10492 				/* Position us to point to the new nrsm that starts the sack blk */
10493 				rsm = nrsm;
10494 			}
10495 		} else {
10496 			/* Already sacked this piece */
10497 			counter_u64_add(rack_sack_skipped_acked, 1);
10498 			moved++;
10499 			if (end == rsm->r_end) {
10500 				/* Done with block */
10501 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10502 				goto out;
10503 			} else if (SEQ_LT(end, rsm->r_end)) {
10504 				/* A partial sack to a already sacked block */
10505 				moved++;
10506 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10507 				goto out;
10508 			} else {
10509 				/*
10510 				 * The end goes beyond this guy
10511 				 * reposition the start to the
10512 				 * next block.
10513 				 */
10514 				start = rsm->r_end;
10515 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10516 				if (rsm == NULL)
10517 					goto out;
10518 			}
10519 		}
10520 	}
10521 	if (SEQ_GEQ(end, rsm->r_end)) {
10522 		/**
10523 		 * The end of this block is either beyond this guy or right
10524 		 * at this guy. I.e.:
10525 		 *  rsm ---                 |-----|
10526 		 *  end                     |-----|
10527 		 *  <or>
10528 		 *  end                     |---------|
10529 		 */
10530 		if ((rsm->r_flags & RACK_ACKED) == 0) {
10531 			/*
10532 			 * Is it a TLP of interest?
10533 			 */
10534 			if ((rsm->r_flags & RACK_TLP) &&
10535 			    (rsm->r_rtr_cnt > 1)) {
10536 				/*
10537 				 * We are splitting a rxt TLP, check
10538 				 * if we need to save off the start/end
10539 				 */
10540 				if (rack->rc_last_tlp_acked_set &&
10541 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10542 					/*
10543 					 * We already turned this on since we are inside
10544 					 * the previous one was a partially sack now we
10545 					 * are getting another one (maybe all of it).
10546 					 */
10547 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10548 					/*
10549 					 * Lets make sure we have all of it though.
10550 					 */
10551 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10552 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10553 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10554 								     rack->r_ctl.last_tlp_acked_end);
10555 					}
10556 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10557 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10558 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10559 								     rack->r_ctl.last_tlp_acked_end);
10560 					}
10561 				} else {
10562 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10563 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10564 					rack->rc_last_tlp_past_cumack = 0;
10565 					rack->rc_last_tlp_acked_set = 1;
10566 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10567 				}
10568 			}
10569 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10570 			changed += (rsm->r_end - rsm->r_start);
10571 			/* You get a count for acking a whole segment or more */
10572 			if ((rsm->r_end - rsm->r_start) >= segsiz)
10573 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10574 			if (rsm->r_flags & RACK_WAS_LOST) {
10575 				int my_chg;
10576 
10577 				my_chg = (rsm->r_end - rsm->r_start);
10578 				rsm->r_flags &= ~RACK_WAS_LOST;
10579 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10580 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10581 				if (my_chg <= rack->r_ctl.rc_considered_lost)
10582 					rack->r_ctl.rc_considered_lost -= my_chg;
10583 				else
10584 					rack->r_ctl.rc_considered_lost = 0;
10585 			}
10586 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10587 			if (rsm->r_in_tmap) /* should be true */
10588 				rack_log_sack_passed(tp, rack, rsm, cts);
10589 			/* Is Reordering occuring? */
10590 			if (rsm->r_flags & RACK_SACK_PASSED) {
10591 				rsm->r_flags &= ~RACK_SACK_PASSED;
10592 				rack->r_ctl.rc_reorder_ts = cts;
10593 				if (rack->r_ctl.rc_reorder_ts == 0)
10594 					rack->r_ctl.rc_reorder_ts = 1;
10595 			}
10596 			if (rack->app_limited_needs_set)
10597 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10598 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10599 			rsm->r_flags |= RACK_ACKED;
10600 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
10601 			if (rsm->r_in_tmap) {
10602 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10603 				rsm->r_in_tmap = 0;
10604 			}
10605 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
10606 		} else {
10607 			counter_u64_add(rack_sack_skipped_acked, 1);
10608 			moved++;
10609 		}
10610 		if (end == rsm->r_end) {
10611 			/* This block only - done, setup for next */
10612 			goto out;
10613 		}
10614 		/*
10615 		 * There is more not coverend by this rsm move on
10616 		 * to the next block in the tail queue hash table.
10617 		 */
10618 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
10619 		start = rsm->r_end;
10620 		rsm = nrsm;
10621 		if (rsm == NULL)
10622 			goto out;
10623 		goto do_rest_ofb;
10624 	}
10625 	/**
10626 	 * The end of this sack block is smaller than
10627 	 * our rsm i.e.:
10628 	 *  rsm ---                 |-----|
10629 	 *  end                     |--|
10630 	 */
10631 	if ((rsm->r_flags & RACK_ACKED) == 0) {
10632 		/*
10633 		 * Is it a TLP of interest?
10634 		 */
10635 		if ((rsm->r_flags & RACK_TLP) &&
10636 		    (rsm->r_rtr_cnt > 1)) {
10637 			/*
10638 			 * We are splitting a rxt TLP, check
10639 			 * if we need to save off the start/end
10640 			 */
10641 			if (rack->rc_last_tlp_acked_set &&
10642 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10643 				/*
10644 				 * We already turned this on since we are inside
10645 				 * the previous one was a partially sack now we
10646 				 * are getting another one (maybe all of it).
10647 				 */
10648 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10649 				/*
10650 				 * Lets make sure we have all of it though.
10651 				 */
10652 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10653 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10654 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10655 							     rack->r_ctl.last_tlp_acked_end);
10656 				}
10657 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10658 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10659 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10660 							     rack->r_ctl.last_tlp_acked_end);
10661 				}
10662 			} else {
10663 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10664 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10665 				rack->rc_last_tlp_past_cumack = 0;
10666 				rack->rc_last_tlp_acked_set = 1;
10667 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10668 			}
10669 		}
10670 		/*
10671 		 * Hookery can only be used if the two entries
10672 		 * are in the same bucket and neither one of
10673 		 * them staddle the bucket line.
10674 		 */
10675 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10676 		if (prev &&
10677 		    (rsm->bindex == prev->bindex) &&
10678 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
10679 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
10680 		    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
10681 		    ((prev->r_flags & RACK_IS_PCM) == 0) &&
10682 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
10683 		    (prev->r_flags & RACK_IN_GP_WIN))
10684 			can_use_hookery = 1;
10685 		else
10686 			can_use_hookery = 0;
10687 		if (prev && can_use_hookery &&
10688 		    (prev->r_flags & RACK_ACKED)) {
10689 			/**
10690 			 * Goal, we want the right remainder of rsm to shrink
10691 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
10692 			 * We want to expand prev to go all the way
10693 			 * to prev->r_end <- end.
10694 			 * so in the tree we have before:
10695 			 *   prev     |--------|         (acked)
10696 			 *   rsm               |-------| (non-acked)
10697 			 *   sackblk           |-|
10698 			 * We churn it so we end up with
10699 			 *   prev     |----------|       (acked)
10700 			 *   rsm                 |-----| (non-acked)
10701 			 *   nrsm              |-| (temporary)
10702 			 *
10703 			 * Note if either prev/rsm is a TLP we don't
10704 			 * do this.
10705 			 */
10706 			noextra++;
10707 			nrsm = &stack_map;
10708 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10709 			tqhash_update_end(rack->r_ctl.tqh, prev, end);
10710 			rsm->r_start = end;
10711 			rsm->r_flags |= RACK_SHUFFLED;
10712 			prev->r_flags |= RACK_SHUFFLED;
10713 			/* Now adjust nrsm (stack copy) to be
10714 			 * the one that is the small
10715 			 * piece that was "sacked".
10716 			 */
10717 			nrsm->r_end = end;
10718 			rsm->r_dupack = 0;
10719 			/*
10720 			 * Which timestamp do we keep? It is rather
10721 			 * important in GP measurements to have the
10722 			 * accurate end of the send window.
10723 			 *
10724 			 * We keep the largest value, which is the newest
10725 			 * send. We do this in case a segment that is
10726 			 * joined together and not part of a GP estimate
10727 			 * later gets expanded into the GP estimate.
10728 			 *
10729 			 * We prohibit the merging of unlike kinds i.e.
10730 			 * all pieces that are in the GP estimate can be
10731 			 * merged and all pieces that are not in a GP estimate
10732 			 * can be merged, but not disimilar pieces. Combine
10733 			 * this with taking the highest here and we should
10734 			 * be ok unless of course the client reneges. Then
10735 			 * all bets are off.
10736 			 */
10737 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
10738 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
10739 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
10740 			}
10741 			/*
10742 			 * And we must keep the newest ack arrival time.
10743 			 */
10744 
10745 			if(prev->r_ack_arrival <
10746 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10747 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10748 
10749 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10750 			/*
10751 			 * Now that the rsm has had its start moved forward
10752 			 * lets go ahead and get its new place in the world.
10753 			 */
10754 			rack_setup_offset_for_rsm(rack, prev, rsm);
10755 			/*
10756 			 * Now nrsm is our new little piece
10757 			 * that is acked (which was merged
10758 			 * to prev). Update the rtt and changed
10759 			 * based on that. Also check for reordering.
10760 			 */
10761 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10762 			if (rack->app_limited_needs_set)
10763 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10764 			changed += (nrsm->r_end - nrsm->r_start);
10765 			/* You get a count for acking a whole segment or more */
10766 			if ((nrsm->r_end - nrsm->r_start) >= segsiz)
10767 				rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
10768 
10769 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10770 			if (rsm->r_flags & RACK_WAS_LOST) {
10771 				int my_chg;
10772 
10773 				my_chg = (nrsm->r_end - nrsm->r_start);
10774 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10775 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10776 				if (my_chg <= rack->r_ctl.rc_considered_lost)
10777 					rack->r_ctl.rc_considered_lost -= my_chg;
10778 				else
10779 					rack->r_ctl.rc_considered_lost = 0;
10780 			}
10781 			if (nrsm->r_flags & RACK_SACK_PASSED) {
10782 				rack->r_ctl.rc_reorder_ts = cts;
10783 				if (rack->r_ctl.rc_reorder_ts == 0)
10784 					rack->r_ctl.rc_reorder_ts = 1;
10785 			}
10786 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
10787 			rsm = prev;
10788 			counter_u64_add(rack_sack_used_prev_merge, 1);
10789 		} else {
10790 			/**
10791 			 * This is the case where our previous
10792 			 * block is not acked either, so we must
10793 			 * split the block in two.
10794 			 */
10795 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10796 			if (nrsm == NULL) {
10797 				/* failed rrs what can we do but loose the sack info? */
10798 				goto out;
10799 			}
10800 			if ((rsm->r_flags & RACK_TLP) &&
10801 			    (rsm->r_rtr_cnt > 1)) {
10802 				/*
10803 				 * We are splitting a rxt TLP, check
10804 				 * if we need to save off the start/end
10805 				 */
10806 				if (rack->rc_last_tlp_acked_set &&
10807 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10808 					/*
10809 					 * We already turned this on since this block is inside
10810 					 * the previous one was a partially sack now we
10811 					 * are getting another one (maybe all of it).
10812 					 */
10813 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10814 					/*
10815 					 * Lets make sure we have all of it though.
10816 					 */
10817 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10818 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10819 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10820 								     rack->r_ctl.last_tlp_acked_end);
10821 					}
10822 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10823 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10824 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10825 								     rack->r_ctl.last_tlp_acked_end);
10826 					}
10827 				} else {
10828 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10829 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10830 					rack->rc_last_tlp_acked_set = 1;
10831 					rack->rc_last_tlp_past_cumack = 0;
10832 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10833 				}
10834 			}
10835 			/**
10836 			 * In this case nrsm becomes
10837 			 * nrsm->r_start = end;
10838 			 * nrsm->r_end = rsm->r_end;
10839 			 * which is un-acked.
10840 			 * <and>
10841 			 * rsm->r_end = nrsm->r_start;
10842 			 * i.e. the remaining un-acked
10843 			 * piece is left on the left
10844 			 * hand side.
10845 			 *
10846 			 * So we start like this
10847 			 * rsm      |----------| (not acked)
10848 			 * sackblk  |---|
10849 			 * build it so we have
10850 			 * rsm      |---|         (acked)
10851 			 * nrsm         |------|  (not acked)
10852 			 */
10853 			counter_u64_add(rack_sack_splits, 1);
10854 			rack_clone_rsm(rack, nrsm, rsm, end);
10855 			moved++;
10856 			rsm->r_flags &= (~RACK_HAS_FIN);
10857 			rsm->r_just_ret = 0;
10858 #ifndef INVARIANTS
10859 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10860 #else
10861 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10862 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
10863 				      nrsm, insret, rack, rsm);
10864 			}
10865 #endif
10866 			if (rsm->r_in_tmap) {
10867 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10868 				nrsm->r_in_tmap = 1;
10869 			}
10870 			nrsm->r_dupack = 0;
10871 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10872 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10873 			changed += (rsm->r_end - rsm->r_start);
10874 			/* You get a count for acking a whole segment or more */
10875 			if ((rsm->r_end - rsm->r_start) >= segsiz)
10876 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10877 			if (rsm->r_flags & RACK_WAS_LOST) {
10878 				int my_chg;
10879 
10880 				my_chg = (rsm->r_end - rsm->r_start);
10881 				rsm->r_flags &= ~RACK_WAS_LOST;
10882 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
10883 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10884 				if (my_chg <= rack->r_ctl.rc_considered_lost)
10885 					rack->r_ctl.rc_considered_lost -= my_chg;
10886 				else
10887 					rack->r_ctl.rc_considered_lost = 0;
10888 			}
10889 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10890 
10891 			if (rsm->r_in_tmap) /* should be true */
10892 				rack_log_sack_passed(tp, rack, rsm, cts);
10893 			/* Is Reordering occuring? */
10894 			if (rsm->r_flags & RACK_SACK_PASSED) {
10895 				rsm->r_flags &= ~RACK_SACK_PASSED;
10896 				rack->r_ctl.rc_reorder_ts = cts;
10897 				if (rack->r_ctl.rc_reorder_ts == 0)
10898 					rack->r_ctl.rc_reorder_ts = 1;
10899 			}
10900 			if (rack->app_limited_needs_set)
10901 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10902 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10903 			rsm->r_flags |= RACK_ACKED;
10904 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
10905 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10906 			if (rsm->r_in_tmap) {
10907 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10908 				rsm->r_in_tmap = 0;
10909 			}
10910 		}
10911 	} else if (start != end){
10912 		/*
10913 		 * The block was already acked.
10914 		 */
10915 		counter_u64_add(rack_sack_skipped_acked, 1);
10916 		moved++;
10917 	}
10918 out:
10919 	if (rsm &&
10920 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10921 	    (rsm->r_flags & RACK_ACKED)) {
10922 		/*
10923 		 * Now can we merge where we worked
10924 		 * with either the previous or
10925 		 * next block?
10926 		 */
10927 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10928 		while (next) {
10929 			if (next->r_flags & RACK_TLP)
10930 				break;
10931 			/* Only allow merges between ones in or out of GP window */
10932 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10933 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10934 				break;
10935 			}
10936 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10937 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10938 				break;
10939 			}
10940 			if (rsm->bindex != next->bindex)
10941 				break;
10942 			if (rsm->r_flags & RACK_STRADDLE)
10943 				break;
10944 			if (rsm->r_flags & RACK_IS_PCM)
10945 				break;
10946 			if (next->r_flags & RACK_STRADDLE)
10947 				break;
10948 			if (next->r_flags & RACK_IS_PCM)
10949 				break;
10950 			if (next->r_flags & RACK_ACKED) {
10951 				/* yep this and next can be merged */
10952 				rsm = rack_merge_rsm(rack, rsm, next);
10953 				noextra++;
10954 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10955 			} else
10956 				break;
10957 		}
10958 		/* Now what about the previous? */
10959 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10960 		while (prev) {
10961 			if (prev->r_flags & RACK_TLP)
10962 				break;
10963 			/* Only allow merges between ones in or out of GP window */
10964 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10965 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10966 				break;
10967 			}
10968 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10969 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10970 				break;
10971 			}
10972 			if (rsm->bindex != prev->bindex)
10973 				break;
10974 			if (rsm->r_flags & RACK_STRADDLE)
10975 				break;
10976 			if (rsm->r_flags & RACK_IS_PCM)
10977 				break;
10978 			if (prev->r_flags & RACK_STRADDLE)
10979 				break;
10980 			if (prev->r_flags & RACK_IS_PCM)
10981 				break;
10982 			if (prev->r_flags & RACK_ACKED) {
10983 				/* yep the previous and this can be merged */
10984 				rsm = rack_merge_rsm(rack, prev, rsm);
10985 				noextra++;
10986 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10987 			} else
10988 				break;
10989 		}
10990 	}
10991 	if (used_ref == 0) {
10992 		counter_u64_add(rack_sack_proc_all, 1);
10993 	} else {
10994 		counter_u64_add(rack_sack_proc_short, 1);
10995 	}
10996 	/* Save off the next one for quick reference. */
10997 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10998 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10999 	/* Pass back the moved. */
11000 	*moved_two = moved;
11001 	*no_extra = noextra;
11002 	if (IN_RECOVERY(tp->t_flags)) {
11003 		rack->r_ctl.bytes_acked_in_recovery += changed;
11004 	}
11005 	return (changed);
11006 }
11007 
11008 static void inline
11009 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
11010 {
11011 	struct rack_sendmap *tmap;
11012 
11013 	tmap = NULL;
11014 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
11015 		/* Its no longer sacked, mark it so */
11016 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
11017 #ifdef INVARIANTS
11018 		if (rsm->r_in_tmap) {
11019 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
11020 			      rack, rsm, rsm->r_flags);
11021 		}
11022 #endif
11023 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
11024 		/* Rebuild it into our tmap */
11025 		if (tmap == NULL) {
11026 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
11027 			tmap = rsm;
11028 		} else {
11029 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
11030 			tmap = rsm;
11031 		}
11032 		tmap->r_in_tmap = 1;
11033 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11034 	}
11035 	/*
11036 	 * Now lets possibly clear the sack filter so we start
11037 	 * recognizing sacks that cover this area.
11038 	 */
11039 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
11040 
11041 }
11042 
11043 static void
11044 rack_do_decay(struct tcp_rack *rack)
11045 {
11046 	struct timeval res;
11047 
11048 #define	timersub(tvp, uvp, vvp)						\
11049 	do {								\
11050 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
11051 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
11052 		if ((vvp)->tv_usec < 0) {				\
11053 			(vvp)->tv_sec--;				\
11054 			(vvp)->tv_usec += 1000000;			\
11055 		}							\
11056 	} while (0)
11057 
11058 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
11059 #undef timersub
11060 
11061 	rack->r_ctl.input_pkt++;
11062 	if ((rack->rc_in_persist) ||
11063 	    (res.tv_sec >= 1) ||
11064 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
11065 		/*
11066 		 * Check for decay of non-SAD,
11067 		 * we want all SAD detection metrics to
11068 		 * decay 1/4 per second (or more) passed.
11069 		 * Current default is 800 so it decays
11070 		 * 80% every second.
11071 		 */
11072 #ifdef TCP_SAD_DETECTION
11073 		uint32_t pkt_delta;
11074 
11075 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
11076 #endif
11077 		/* Update our saved tracking values */
11078 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
11079 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
11080 		/* Now do we escape without decay? */
11081 #ifdef TCP_SAD_DETECTION
11082 		if (rack->rc_in_persist ||
11083 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
11084 		    (pkt_delta < tcp_sad_low_pps)){
11085 			/*
11086 			 * We don't decay idle connections
11087 			 * or ones that have a low input pps.
11088 			 */
11089 			return;
11090 		}
11091 		/* Decay the counters */
11092 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
11093 							tcp_sad_decay_val);
11094 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
11095 							 tcp_sad_decay_val);
11096 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
11097 							       tcp_sad_decay_val);
11098 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
11099 								tcp_sad_decay_val);
11100 #endif
11101 	}
11102 }
11103 
11104 static void inline
11105 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
11106 {
11107 	/*
11108 	 * We look at advancing the end send time for our GP
11109 	 * measurement tracking only as the cumulative acknowledgment
11110 	 * moves forward. You might wonder about this, why not
11111 	 * at every transmission or retransmission within the
11112 	 * GP window update the rc_gp_cumack_ts? Well its rather
11113 	 * nuanced but basically the GP window *may* expand (as
11114 	 * it does below) or worse and harder to track it may shrink.
11115 	 *
11116 	 * This last makes it impossible to track at the time of
11117 	 * the send, since you may set forward your rc_gp_cumack_ts
11118 	 * when you send, because that send *is* in your currently
11119 	 * "guessed" window, but then it shrinks. Now which was
11120 	 * the send time of the last bytes in the window, by the
11121 	 * time you ask that question that part of the sendmap
11122 	 * is freed. So you don't know and you will have too
11123 	 * long of send window. Instead by updating the time
11124 	 * marker only when the cumack advances this assures us
11125 	 * that we will have only the sends in the window of our
11126 	 * GP measurement.
11127 	 *
11128 	 * Another complication from this is the
11129 	 * merging of sendmap entries. During SACK processing this
11130 	 * can happen to conserve the sendmap size. That breaks
11131 	 * everything down in tracking the send window of the GP
11132 	 * estimate. So to prevent that and keep it working with
11133 	 * a tiny bit more limited merging, we only allow like
11134 	 * types to be merged. I.e. if two sends are in the GP window
11135 	 * then its ok to merge them together. If two sends are not
11136 	 * in the GP window its ok to merge them together too. Though
11137 	 * one send in and one send out cannot be merged. We combine
11138 	 * this with never allowing the shrinking of the GP window when
11139 	 * we are in recovery so that we can properly calculate the
11140 	 * sending times.
11141 	 *
11142 	 * This all of course seems complicated, because it is.. :)
11143 	 *
11144 	 * The cum-ack is being advanced upon the sendmap.
11145 	 * If we are not doing a GP estimate don't
11146 	 * proceed.
11147 	 */
11148 	uint64_t ts;
11149 
11150 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
11151 		return;
11152 	/*
11153 	 * If this sendmap entry is going
11154 	 * beyond the measurement window we had picked,
11155 	 * expand the measurement window by that much.
11156 	 */
11157 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
11158 		tp->gput_ack = rsm->r_end;
11159 	}
11160 	/*
11161 	 * If we have not setup a ack, then we
11162 	 * have no idea if the newly acked pieces
11163 	 * will be "in our seq measurement range". If
11164 	 * it is when we clear the app_limited_needs_set
11165 	 * flag the timestamp will be updated.
11166 	 */
11167 	if (rack->app_limited_needs_set)
11168 		return;
11169 	/*
11170 	 * Finally, we grab out the latest timestamp
11171 	 * that this packet was sent and then see
11172 	 * if:
11173 	 *  a) The packet touches are newly defined GP range.
11174 	 *  b) The time is greater than (newer) than the
11175 	 *     one we currently have. If so we update
11176 	 *     our sending end time window.
11177 	 *
11178 	 * Note we *do not* do this at send time. The reason
11179 	 * is that if you do you *may* pick up a newer timestamp
11180 	 * for a range you are not going to measure. We project
11181 	 * out how far and then sometimes modify that to be
11182 	 * smaller. If that occurs then you will have a send
11183 	 * that does not belong to the range included.
11184 	 */
11185 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
11186 	    rack->r_ctl.rc_gp_cumack_ts)
11187 		return;
11188 	if (rack_in_gp_window(tp, rsm)) {
11189 		rack->r_ctl.rc_gp_cumack_ts = ts;
11190 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
11191 			       __LINE__, from, rsm);
11192 	}
11193 }
11194 
11195 static void
11196 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to, uint64_t acktime)
11197 {
11198 	struct rack_sendmap *rsm;
11199 	/*
11200 	 * The ACK point is advancing to th_ack, we must drop off
11201 	 * the packets in the rack log and calculate any eligble
11202 	 * RTT's.
11203 	 */
11204 
11205 	if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
11206 		/*
11207 		 * If we have some sack blocks in the filter
11208 		 * lets prune them out by calling sfb with no blocks.
11209 		 */
11210 		sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
11211 	}
11212 	if (SEQ_GT(th_ack, tp->snd_una)) {
11213 		/* Clear any app ack remembered settings */
11214 		rack->r_ctl.cleared_app_ack = 0;
11215 	}
11216 	rack->r_wanted_output = 1;
11217 	if (SEQ_GT(th_ack, tp->snd_una))
11218 		rack->r_ctl.last_cumack_advance = acktime;
11219 
11220 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
11221 	if ((rack->rc_last_tlp_acked_set == 1)&&
11222 	    (rack->rc_last_tlp_past_cumack == 1) &&
11223 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
11224 		/*
11225 		 * We have reached the point where our last rack
11226 		 * tlp retransmit sequence is ahead of the cum-ack.
11227 		 * This can only happen when the cum-ack moves all
11228 		 * the way around (its been a full 2^^31+1 bytes
11229 		 * or more since we sent a retransmitted TLP). Lets
11230 		 * turn off the valid flag since its not really valid.
11231 		 *
11232 		 * Note since sack's also turn on this event we have
11233 		 * a complication, we have to wait to age it out until
11234 		 * the cum-ack is by the TLP before checking which is
11235 		 * what the next else clause does.
11236 		 */
11237 		rack_log_dsack_event(rack, 9, __LINE__,
11238 				     rack->r_ctl.last_tlp_acked_start,
11239 				     rack->r_ctl.last_tlp_acked_end);
11240 		rack->rc_last_tlp_acked_set = 0;
11241 		rack->rc_last_tlp_past_cumack = 0;
11242 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
11243 		   (rack->rc_last_tlp_past_cumack == 0) &&
11244 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
11245 		/*
11246 		 * It is safe to start aging TLP's out.
11247 		 */
11248 		rack->rc_last_tlp_past_cumack = 1;
11249 	}
11250 	/* We do the same for the tlp send seq as well */
11251 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
11252 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
11253 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
11254 		rack_log_dsack_event(rack, 9, __LINE__,
11255 				     rack->r_ctl.last_sent_tlp_seq,
11256 				     (rack->r_ctl.last_sent_tlp_seq +
11257 				      rack->r_ctl.last_sent_tlp_len));
11258 		rack->rc_last_sent_tlp_seq_valid = 0;
11259 		rack->rc_last_sent_tlp_past_cumack = 0;
11260 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
11261 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
11262 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
11263 		/*
11264 		 * It is safe to start aging TLP's send.
11265 		 */
11266 		rack->rc_last_sent_tlp_past_cumack = 1;
11267 	}
11268 more:
11269 	rsm = tqhash_min(rack->r_ctl.tqh);
11270 	if (rsm == NULL) {
11271 		if ((th_ack - 1) == tp->iss) {
11272 			/*
11273 			 * For the SYN incoming case we will not
11274 			 * have called tcp_output for the sending of
11275 			 * the SYN, so there will be no map. All
11276 			 * other cases should probably be a panic.
11277 			 */
11278 			return;
11279 		}
11280 		if (tp->t_flags & TF_SENTFIN) {
11281 			/* if we sent a FIN we often will not have map */
11282 			return;
11283 		}
11284 #ifdef INVARIANTS
11285 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
11286 		      tp,
11287 		      tp->t_state, th_ack, rack,
11288 		      tp->snd_una, tp->snd_max);
11289 #endif
11290 		return;
11291 	}
11292 	if (SEQ_LT(th_ack, rsm->r_start)) {
11293 		/* Huh map is missing this */
11294 #ifdef INVARIANTS
11295 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
11296 		       rsm->r_start,
11297 		       th_ack, tp->t_state, rack->r_state);
11298 #endif
11299 		return;
11300 	}
11301 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
11302 
11303 	/* Now was it a retransmitted TLP? */
11304 	if ((rsm->r_flags & RACK_TLP) &&
11305 	    (rsm->r_rtr_cnt > 1)) {
11306 		/*
11307 		 * Yes, this rsm was a TLP and retransmitted, remember that
11308 		 * since if a DSACK comes back on this we don't want
11309 		 * to think of it as a reordered segment. This may
11310 		 * get updated again with possibly even other TLPs
11311 		 * in flight, but thats ok. Only when we don't send
11312 		 * a retransmitted TLP for 1/2 the sequences space
11313 		 * will it get turned off (above).
11314 		 */
11315 		if (rack->rc_last_tlp_acked_set &&
11316 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
11317 			/*
11318 			 * We already turned this on since the end matches,
11319 			 * the previous one was a partially ack now we
11320 			 * are getting another one (maybe all of it).
11321 			 */
11322 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
11323 			/*
11324 			 * Lets make sure we have all of it though.
11325 			 */
11326 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
11327 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
11328 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
11329 						     rack->r_ctl.last_tlp_acked_end);
11330 			}
11331 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
11332 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
11333 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
11334 						     rack->r_ctl.last_tlp_acked_end);
11335 			}
11336 		} else {
11337 			rack->rc_last_tlp_past_cumack = 1;
11338 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
11339 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
11340 			rack->rc_last_tlp_acked_set = 1;
11341 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
11342 		}
11343 	}
11344 	/* Now do we consume the whole thing? */
11345 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
11346 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
11347 		/* Its all consumed. */
11348 		uint32_t left;
11349 		uint8_t newly_acked;
11350 
11351 		if (rsm->r_flags & RACK_WAS_LOST) {
11352 			/*
11353 			 * This can happen when we marked it as lost
11354 			 * and yet before retransmitting we get an ack
11355 			 * which can happen due to reordering.
11356 			 */
11357 			rsm->r_flags  &= ~RACK_WAS_LOST;
11358 			KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
11359 				("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
11360 			if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
11361 				rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
11362 			else
11363 				rack->r_ctl.rc_considered_lost = 0;
11364 		}
11365 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
11366 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
11367 		rsm->r_rtr_bytes = 0;
11368 		/*
11369 		 * Record the time of highest cumack sent if its in our measurement
11370 		 * window and possibly bump out the end.
11371 		 */
11372 		rack_rsm_sender_update(rack, tp, rsm, 4);
11373 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
11374 		if (rsm->r_in_tmap) {
11375 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
11376 			rsm->r_in_tmap = 0;
11377 		}
11378 		newly_acked = 1;
11379 		if (((rsm->r_flags & RACK_ACKED) == 0) &&
11380 		    (IN_RECOVERY(tp->t_flags))) {
11381 			rack->r_ctl.bytes_acked_in_recovery += (rsm->r_end - rsm->r_start);
11382 		}
11383 		if (rsm->r_flags & RACK_ACKED) {
11384 			/*
11385 			 * It was acked on the scoreboard -- remove
11386 			 * it from total
11387 			 */
11388 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
11389 			newly_acked = 0;
11390 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
11391 			/*
11392 			 * There are segments ACKED on the
11393 			 * scoreboard further up. We are seeing
11394 			 * reordering.
11395 			 */
11396 			rsm->r_flags &= ~RACK_SACK_PASSED;
11397 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
11398 			rsm->r_flags |= RACK_ACKED;
11399 			rack->r_ctl.rc_reorder_ts = cts;
11400 			if (rack->r_ctl.rc_reorder_ts == 0)
11401 				rack->r_ctl.rc_reorder_ts = 1;
11402 			if (rack->r_ent_rec_ns) {
11403 				/*
11404 				 * We have sent no more, and we saw an sack
11405 				 * then ack arrive.
11406 				 */
11407 				rack->r_might_revert = 1;
11408 			}
11409 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
11410 		} else {
11411 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
11412 		}
11413 		if ((rsm->r_flags & RACK_TO_REXT) &&
11414 		    (tp->t_flags & TF_RCVD_TSTMP) &&
11415 		    (to->to_flags & TOF_TS) &&
11416 		    (to->to_tsecr != 0) &&
11417 		    (tp->t_flags & TF_PREVVALID)) {
11418 			/*
11419 			 * We can use the timestamp to see
11420 			 * if this retransmission was from the
11421 			 * first transmit. If so we made a mistake.
11422 			 */
11423 			tp->t_flags &= ~TF_PREVVALID;
11424 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
11425 				/* The first transmit is what this ack is for */
11426 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
11427 			}
11428 		}
11429 		left = th_ack - rsm->r_end;
11430 		if (rack->app_limited_needs_set && newly_acked)
11431 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
11432 		/* Free back to zone */
11433 		rack_free(rack, rsm);
11434 		if (left) {
11435 			goto more;
11436 		}
11437 		/* Check for reneging */
11438 		rsm = tqhash_min(rack->r_ctl.tqh);
11439 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
11440 			/*
11441 			 * The peer has moved snd_una up to
11442 			 * the edge of this send, i.e. one
11443 			 * that it had previously acked. The only
11444 			 * way that can be true if the peer threw
11445 			 * away data (space issues) that it had
11446 			 * previously sacked (else it would have
11447 			 * given us snd_una up to (rsm->r_end).
11448 			 * We need to undo the acked markings here.
11449 			 *
11450 			 * Note we have to look to make sure th_ack is
11451 			 * our rsm->r_start in case we get an old ack
11452 			 * where th_ack is behind snd_una.
11453 			 */
11454 			rack_peer_reneges(rack, rsm, th_ack);
11455 		}
11456 		return;
11457 	}
11458 	if (rsm->r_flags & RACK_ACKED) {
11459 		/*
11460 		 * It was acked on the scoreboard -- remove it from
11461 		 * total for the part being cum-acked.
11462 		 */
11463 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
11464 	} else {
11465 		if (((rsm->r_flags & RACK_ACKED) == 0) &&
11466 		    (IN_RECOVERY(tp->t_flags))) {
11467 			rack->r_ctl.bytes_acked_in_recovery += (th_ack - rsm->r_start);
11468 		}
11469 		rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
11470 	}
11471 	/* And what about the lost flag? */
11472 	if (rsm->r_flags & RACK_WAS_LOST) {
11473 		/*
11474 		 * This can happen when we marked it as lost
11475 		 * and yet before retransmitting we get an ack
11476 		 * which can happen due to reordering. In this
11477 		 * case its only a partial ack of the send.
11478 		 */
11479 		KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
11480 			("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm,  rack, th_ack));
11481 		if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
11482 			rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
11483 		else
11484 			rack->r_ctl.rc_considered_lost = 0;
11485 	}
11486 	/*
11487 	 * Clear the dup ack count for
11488 	 * the piece that remains.
11489 	 */
11490 	rsm->r_dupack = 0;
11491 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
11492 	if (rsm->r_rtr_bytes) {
11493 		/*
11494 		 * It was retransmitted adjust the
11495 		 * sack holes for what was acked.
11496 		 */
11497 		int ack_am;
11498 
11499 		ack_am = (th_ack - rsm->r_start);
11500 		if (ack_am >= rsm->r_rtr_bytes) {
11501 			rack->r_ctl.rc_holes_rxt -= ack_am;
11502 			rsm->r_rtr_bytes -= ack_am;
11503 		}
11504 	}
11505 	/*
11506 	 * Update where the piece starts and record
11507 	 * the time of send of highest cumack sent if
11508 	 * its in our GP range.
11509 	 */
11510 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
11511 	/* Now we need to move our offset forward too */
11512 	if (rsm->m &&
11513 	    ((rsm->orig_m_len != rsm->m->m_len) ||
11514 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
11515 		/* Fix up the orig_m_len and possibly the mbuf offset */
11516 		rack_adjust_orig_mlen(rsm);
11517 	}
11518 	rsm->soff += (th_ack - rsm->r_start);
11519 	rack_rsm_sender_update(rack, tp, rsm, 5);
11520 	/* The trim will move th_ack into r_start for us */
11521 	tqhash_trim(rack->r_ctl.tqh, th_ack);
11522 	/* Now do we need to move the mbuf fwd too? */
11523 	{
11524 		struct mbuf *m;
11525 		uint32_t soff;
11526 
11527 		m = rsm->m;
11528 		soff = rsm->soff;
11529 		if (m) {
11530 			while (soff >= m->m_len) {
11531 				soff -= m->m_len;
11532 				KASSERT((m->m_next != NULL),
11533 					(" rsm:%p  off:%u soff:%u m:%p",
11534 					 rsm, rsm->soff, soff, m));
11535 				m = m->m_next;
11536 				if (m == NULL) {
11537 					/*
11538 					 * This is a fall-back that prevents a panic. In reality
11539 					 * we should be able to walk the mbuf's and find our place.
11540 					 * At this point snd_una has not been updated with the sbcut() yet
11541 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
11542 					 * should work fine. This is undesirable since we will take cache
11543 					 * hits to access the socket buffer. And even more puzzling is that
11544 					 * it happens occasionally. It should not :(
11545 					 */
11546 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
11547 						      (rsm->r_start - tp->snd_una),
11548 						      &soff);
11549 					break;
11550 				}
11551 			}
11552 			/*
11553 			 * Now save in our updated values.
11554 			 */
11555 			rsm->m = m;
11556 			rsm->soff = soff;
11557 			rsm->orig_m_len = rsm->m->m_len;
11558 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11559 		}
11560 	}
11561 	if (rack->app_limited_needs_set &&
11562 	    SEQ_GEQ(th_ack, tp->gput_seq))
11563 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
11564 }
11565 
11566 static void
11567 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
11568 {
11569 	struct rack_sendmap *rsm;
11570 	int sack_pass_fnd = 0;
11571 
11572 	if (rack->r_might_revert) {
11573 		/*
11574 		 * Ok we have reordering, have not sent anything, we
11575 		 * might want to revert the congestion state if nothing
11576 		 * further has SACK_PASSED on it. Lets check.
11577 		 *
11578 		 * We also get here when we have DSACKs come in for
11579 		 * all the data that we FR'd. Note that a rxt or tlp
11580 		 * timer clears this from happening.
11581 		 */
11582 
11583 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
11584 			if (rsm->r_flags & RACK_SACK_PASSED) {
11585 				sack_pass_fnd = 1;
11586 				break;
11587 			}
11588 		}
11589 		if (sack_pass_fnd == 0) {
11590 			/*
11591 			 * We went into recovery
11592 			 * incorrectly due to reordering!
11593 			 */
11594 			int orig_cwnd;
11595 
11596 			rack->r_ent_rec_ns = 0;
11597 			orig_cwnd = tp->snd_cwnd;
11598 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
11599 			tp->snd_recover = tp->snd_una;
11600 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
11601 			if (IN_RECOVERY(tp->t_flags)) {
11602 				rack_exit_recovery(tp, rack, 3);
11603 				if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
11604 					/*
11605 					 * We were in recovery, had an RTO
11606 					 * and then re-entered recovery (more sack's arrived)
11607 					 * and we have properly recorded the old ssthresh from
11608 					 * the first recovery. We want to be able to slow-start
11609 					 * back to this level. The ssthresh from the timeout
11610 					 * and then back into recovery will end up most likely
11611 					 * to be min(cwnd=1mss, 2mss). Which makes it basically
11612 					 * so we get no slow-start after our RTO.
11613 					 */
11614 					rack->rto_from_rec = 0;
11615 					if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
11616 						tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
11617 				}
11618 			}
11619 			rack->r_ctl.bytes_acked_in_recovery = 0;
11620 			rack->r_ctl.time_entered_recovery = 0;
11621 		}
11622 		rack->r_might_revert = 0;
11623 	}
11624 }
11625 
11626 #ifdef TCP_SAD_DETECTION
11627 
11628 static void
11629 rack_merge_out_sacks(struct tcp_rack *rack)
11630 {
11631 	struct rack_sendmap *cur, *next, *rsm, *trsm = NULL;
11632 
11633 	cur = tqhash_min(rack->r_ctl.tqh);
11634 	while(cur) {
11635 		next = tqhash_next(rack->r_ctl.tqh, cur);
11636 		/*
11637 		 * The idea is to go through all and merge back
11638 		 * together the pieces sent together,
11639 		 */
11640 		if ((next != NULL) &&
11641 		    (cur->r_tim_lastsent[0] == next->r_tim_lastsent[0])) {
11642 			rack_merge_rsm(rack, cur, next);
11643 		} else {
11644 			cur = next;
11645 		}
11646 	}
11647 	/*
11648 	 * now treat it like a rxt event, everything is outstanding
11649 	 * and sent nothing acvked and dupacks are all zero. If this
11650 	 * is not an attacker it will have to dupack its way through
11651 	 * it all.
11652 	 */
11653 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
11654 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
11655 		rsm->r_dupack = 0;
11656 		/* We must re-add it back to the tlist */
11657 		if (trsm == NULL) {
11658 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
11659 		} else {
11660 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
11661 		}
11662 		rsm->r_in_tmap = 1;
11663 		trsm = rsm;
11664 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
11665 	}
11666 	sack_filter_clear(&rack->r_ctl.rack_sf, rack->rc_tp->snd_una);
11667 }
11668 
11669 static void
11670 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
11671 {
11672 	int do_detection = 0;
11673 
11674 	if (rack->sack_attack_disable || rack->rc_suspicious) {
11675 		/*
11676 		 * If we have been disabled we must detect
11677 		 * to possibly reverse it. Or if the guy has
11678 		 * sent in suspicious sacks we want to do detection too.
11679 		 */
11680 		do_detection = 1;
11681 
11682 	} else if  ((rack->do_detection || tcp_force_detection) &&
11683 		    (tcp_sack_to_ack_thresh > 0) &&
11684 		    (tcp_sack_to_move_thresh > 0) &&
11685 		    (rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum)) {
11686 		/*
11687 		 * We only detect here if:
11688 		 * 1) System wide forcing is on <or> do_detection is on
11689 		 *   <and>
11690 		 * 2) We have thresholds for move and ack (set one to 0 and we are off)
11691 		 *   <and>
11692 		 * 3) We have maps allocated larger than our min (500).
11693 		 */
11694 		do_detection = 1;
11695 	}
11696 	if (do_detection > 0) {
11697 		/*
11698 		 * We have thresholds set to find
11699 		 * possible attackers and disable sack.
11700 		 * Check them.
11701 		 */
11702 		uint64_t ackratio, moveratio, movetotal;
11703 
11704 		/* Log detecting */
11705 		rack_log_sad(rack, 1);
11706 		/* Do we establish a ack ratio */
11707 		if ((rack->r_ctl.sack_count > tcp_map_minimum)  ||
11708 		    (rack->rc_suspicious == 1) ||
11709 		    (rack->sack_attack_disable > 0)) {
11710 			ackratio = (uint64_t)(rack->r_ctl.sack_count);
11711 			ackratio *= (uint64_t)(1000);
11712 			if (rack->r_ctl.ack_count)
11713 				ackratio /= (uint64_t)(rack->r_ctl.ack_count);
11714 			else {
11715 				/* We can hit this due to ack totals degregation (via small sacks) */
11716 				ackratio = 1000;
11717 			}
11718 		} else {
11719 			/*
11720 			 * No ack ratio needed if we have not
11721 			 * seen more sacks then the number of map entries.
11722 			 * The exception to that is if we have disabled sack then
11723 			 * we need to find a ratio.
11724 			 */
11725 			ackratio = 0;
11726 		}
11727 
11728 		if ((rack->sack_attack_disable == 0) &&
11729 		    (ackratio > rack_highest_sack_thresh_seen))
11730 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
11731 		/* Do we establish a move ratio? */
11732 		if ((rack->r_ctl.sack_moved_extra > tcp_map_minimum) ||
11733 		    (rack->rc_suspicious == 1) ||
11734 		    (rack->sack_attack_disable > 0)) {
11735 			/*
11736 			 * We need to have more sack moves than maps
11737 			 * allocated to have a move ratio considered.
11738 			 */
11739 			movetotal = rack->r_ctl.sack_moved_extra;
11740 			movetotal += rack->r_ctl.sack_noextra_move;
11741 			moveratio = rack->r_ctl.sack_moved_extra;
11742 			moveratio *= (uint64_t)1000;
11743 			if (movetotal)
11744 				moveratio /= movetotal;
11745 			else {
11746 				/* No moves, thats pretty good */
11747 				moveratio = 0;
11748 			}
11749 		} else {
11750 			/*
11751 			 * Not enough moves have occured to consider
11752 			 * if we are out of whack in that ratio.
11753 			 * The exception to that is if we have disabled sack then
11754 			 * we need to find a ratio.
11755 			 */
11756 			moveratio = 0;
11757 		}
11758 		if ((rack->sack_attack_disable == 0) &&
11759 		    (moveratio > rack_highest_move_thresh_seen))
11760 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
11761 		/* Now the tests */
11762 		if (rack->sack_attack_disable == 0) {
11763 			/* Not disabled, do we need to disable? */
11764 			if ((ackratio > tcp_sack_to_ack_thresh) &&
11765 			    (moveratio > tcp_sack_to_move_thresh)) {
11766 				/* Disable sack processing */
11767 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
11768 				rack->sack_attack_disable = 1;
11769 				/* set it so we have the built in delay */
11770 				rack->r_ctl.ack_during_sd = 1;
11771 				if (rack_merge_out_sacks_on_attack)
11772 					rack_merge_out_sacks(rack);
11773 				counter_u64_add(rack_sack_attacks_detected, 1);
11774 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
11775 				/* Clamp the cwnd at flight size */
11776 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
11777 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11778 				rack_log_sad(rack, 2);
11779 			}
11780 		} else {
11781 			/* We are sack-disabled check for false positives */
11782 			if ((ackratio <= tcp_restoral_thresh) ||
11783 			    ((rack_merge_out_sacks_on_attack == 0) &&
11784 			     (rack->rc_suspicious == 0) &&
11785 			     (rack->r_ctl.rc_num_maps_alloced <= (tcp_map_minimum/2)))) {
11786 				rack->sack_attack_disable = 0;
11787 				rack_log_sad(rack, 3);
11788 				/* Restart counting */
11789 				rack->r_ctl.sack_count = 0;
11790 				rack->r_ctl.sack_moved_extra = 0;
11791 				rack->r_ctl.sack_noextra_move = 1;
11792 				rack->rc_suspicious = 0;
11793 				rack->r_ctl.ack_count = max(1,
11794 							    (bytes_this_ack / segsiz));
11795 
11796 				counter_u64_add(rack_sack_attacks_reversed, 1);
11797 				/* Restore the cwnd */
11798 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
11799 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
11800 			}
11801 		}
11802 	}
11803 }
11804 #endif
11805 
11806 static int
11807 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
11808 {
11809 
11810 	uint32_t am, l_end;
11811 	int was_tlp = 0;
11812 
11813 	if (SEQ_GT(end, start))
11814 		am = end - start;
11815 	else
11816 		am = 0;
11817 	if ((rack->rc_last_tlp_acked_set ) &&
11818 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
11819 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
11820 		/*
11821 		 * The DSACK is because of a TLP which we don't
11822 		 * do anything with the reordering window over since
11823 		 * it was not reordering that caused the DSACK but
11824 		 * our previous retransmit TLP.
11825 		 */
11826 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
11827 		was_tlp = 1;
11828 		goto skip_dsack_round;
11829 	}
11830 	if (rack->rc_last_sent_tlp_seq_valid) {
11831 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
11832 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
11833 		    (SEQ_LEQ(end, l_end))) {
11834 			/*
11835 			 * This dsack is from the last sent TLP, ignore it
11836 			 * for reordering purposes.
11837 			 */
11838 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
11839 			was_tlp = 1;
11840 			goto skip_dsack_round;
11841 		}
11842 	}
11843 	if (rack->rc_dsack_round_seen == 0) {
11844 		rack->rc_dsack_round_seen = 1;
11845 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
11846 		rack->r_ctl.num_dsack++;
11847 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
11848 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
11849 	}
11850 skip_dsack_round:
11851 	/*
11852 	 * We keep track of how many DSACK blocks we get
11853 	 * after a recovery incident.
11854 	 */
11855 	rack->r_ctl.dsack_byte_cnt += am;
11856 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
11857 	    rack->r_ctl.retran_during_recovery &&
11858 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
11859 		/*
11860 		 * False recovery most likely culprit is reordering. If
11861 		 * nothing else is missing we need to revert.
11862 		 */
11863 		rack->r_might_revert = 1;
11864 		rack_handle_might_revert(rack->rc_tp, rack);
11865 		rack->r_might_revert = 0;
11866 		rack->r_ctl.retran_during_recovery = 0;
11867 		rack->r_ctl.dsack_byte_cnt = 0;
11868 	}
11869 	return (was_tlp);
11870 }
11871 
11872 static uint32_t
11873 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
11874 {
11875 	return (((tp->snd_max - snd_una) -
11876 		 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
11877 }
11878 
11879 static int32_t
11880 rack_compute_pipe(struct tcpcb *tp)
11881 {
11882 	return ((int32_t)do_rack_compute_pipe(tp,
11883 					      (struct tcp_rack *)tp->t_fb_ptr,
11884 					      tp->snd_una));
11885 }
11886 
11887 static void
11888 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
11889 {
11890 	/* Deal with changed and PRR here (in recovery only) */
11891 	uint32_t pipe, snd_una;
11892 
11893 	rack->r_ctl.rc_prr_delivered += changed;
11894 
11895 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
11896 		/*
11897 		 * It is all outstanding, we are application limited
11898 		 * and thus we don't need more room to send anything.
11899 		 * Note we use tp->snd_una here and not th_ack because
11900 		 * the data as yet not been cut from the sb.
11901 		 */
11902 		rack->r_ctl.rc_prr_sndcnt = 0;
11903 		return;
11904 	}
11905 	/* Compute prr_sndcnt */
11906 	if (SEQ_GT(tp->snd_una, th_ack)) {
11907 		snd_una = tp->snd_una;
11908 	} else {
11909 		snd_una = th_ack;
11910 	}
11911 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
11912 	if (pipe > tp->snd_ssthresh) {
11913 		long sndcnt;
11914 
11915 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
11916 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
11917 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
11918 		else {
11919 			rack->r_ctl.rc_prr_sndcnt = 0;
11920 			rack_log_to_prr(rack, 9, 0, __LINE__);
11921 			sndcnt = 0;
11922 		}
11923 		sndcnt++;
11924 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
11925 			sndcnt -= rack->r_ctl.rc_prr_out;
11926 		else
11927 			sndcnt = 0;
11928 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
11929 		rack_log_to_prr(rack, 10, 0, __LINE__);
11930 	} else {
11931 		uint32_t limit;
11932 
11933 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
11934 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
11935 		else
11936 			limit = 0;
11937 		if (changed > limit)
11938 			limit = changed;
11939 		limit += ctf_fixed_maxseg(tp);
11940 		if (tp->snd_ssthresh > pipe) {
11941 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
11942 			rack_log_to_prr(rack, 11, 0, __LINE__);
11943 		} else {
11944 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
11945 			rack_log_to_prr(rack, 12, 0, __LINE__);
11946 		}
11947 	}
11948 }
11949 
11950 static void
11951 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11952 	     int *dsack_seen, int *sacks_seen)
11953 {
11954 	uint32_t changed;
11955 	struct tcp_rack *rack;
11956 	struct rack_sendmap *rsm;
11957 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11958 	register uint32_t th_ack;
11959 	int32_t i, j, k, num_sack_blks = 0;
11960 	uint32_t cts, acked, ack_point;
11961 	int loop_start = 0, moved_two = 0, no_extra = 0;
11962 	uint32_t tsused;
11963 	uint32_t segsiz, o_cnt;
11964 
11965 
11966 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11967 	if (tcp_get_flags(th) & TH_RST) {
11968 		/* We don't log resets */
11969 		return;
11970 	}
11971 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11972 	cts = tcp_get_usecs(NULL);
11973 	rsm = tqhash_min(rack->r_ctl.tqh);
11974 	changed = 0;
11975 	th_ack = th->th_ack;
11976 	if (rack->sack_attack_disable == 0)
11977 		rack_do_decay(rack);
11978 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
11979 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11980 		/*
11981 		 * You only get credit for
11982 		 * MSS and greater (and you get extra
11983 		 * credit for larger cum-ack moves).
11984 		 */
11985 		int ac;
11986 
11987 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11988 		rack->r_ctl.ack_count += ac;
11989 		counter_u64_add(rack_ack_total, ac);
11990 	}
11991 	if (rack->r_ctl.ack_count > 0xfff00000) {
11992 		/*
11993 		 * reduce the number to keep us under
11994 		 * a uint32_t.
11995 		 */
11996 		rack->r_ctl.ack_count /= 2;
11997 		rack->r_ctl.sack_count /= 2;
11998 	}
11999 	if (SEQ_GT(th_ack, tp->snd_una)) {
12000 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
12001 		tp->t_acktime = ticks;
12002 	}
12003 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
12004 		changed = th_ack - rsm->r_start;
12005 	if (changed) {
12006 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
12007 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
12008 	}
12009 	if ((to->to_flags & TOF_SACK) == 0) {
12010 		/* We are done nothing left and no sack. */
12011 		rack_handle_might_revert(tp, rack);
12012 		/*
12013 		 * For cases where we struck a dup-ack
12014 		 * with no SACK, add to the changes so
12015 		 * PRR will work right.
12016 		 */
12017 		if (dup_ack_struck && (changed == 0)) {
12018 			changed += ctf_fixed_maxseg(rack->rc_tp);
12019 		}
12020 		goto out;
12021 	}
12022 	/* Sack block processing */
12023 	if (SEQ_GT(th_ack, tp->snd_una))
12024 		ack_point = th_ack;
12025 	else
12026 		ack_point = tp->snd_una;
12027 	for (i = 0; i < to->to_nsacks; i++) {
12028 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
12029 		      &sack, sizeof(sack));
12030 		sack.start = ntohl(sack.start);
12031 		sack.end = ntohl(sack.end);
12032 		if (SEQ_GT(sack.end, sack.start) &&
12033 		    SEQ_GT(sack.start, ack_point) &&
12034 		    SEQ_LT(sack.start, tp->snd_max) &&
12035 		    SEQ_GT(sack.end, ack_point) &&
12036 		    SEQ_LEQ(sack.end, tp->snd_max)) {
12037 			sack_blocks[num_sack_blks] = sack;
12038 			num_sack_blks++;
12039 		} else if (SEQ_LEQ(sack.start, th_ack) &&
12040 			   SEQ_LEQ(sack.end, th_ack)) {
12041 			int was_tlp;
12042 
12043 			if (dsack_seen != NULL)
12044 				*dsack_seen = 1;
12045 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
12046 			/*
12047 			 * Its a D-SACK block.
12048 			 */
12049 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
12050 		}
12051 	}
12052 	if (rack->rc_dsack_round_seen) {
12053 		/* Is the dsack roound over? */
12054 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
12055 			/* Yes it is */
12056 			rack->rc_dsack_round_seen = 0;
12057 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
12058 		}
12059 	}
12060 	/*
12061 	 * Sort the SACK blocks so we can update the rack scoreboard with
12062 	 * just one pass.
12063 	 */
12064 	o_cnt = num_sack_blks;
12065 	num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
12066 					 num_sack_blks, th->th_ack);
12067 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
12068 	if (sacks_seen != NULL)
12069 		*sacks_seen = num_sack_blks;
12070 	if (num_sack_blks == 0) {
12071 		/* Nothing to sack, but we need to update counts */
12072 		if ((o_cnt == 1) &&
12073 		    (*dsack_seen != 1))
12074 			rack->r_ctl.sack_count++;
12075 		else if (o_cnt > 1)
12076 			rack->r_ctl.sack_count++;
12077 		goto out_with_totals;
12078 	}
12079 	if (rack->sack_attack_disable) {
12080 		/*
12081 		 * An attacker disablement is in place, for
12082 		 * every sack block that is not at least a full MSS
12083 		 * count up sack_count.
12084 		 */
12085 		for (i = 0; i < num_sack_blks; i++) {
12086 			if ((sack_blocks[i].end - sack_blocks[i].start) < segsiz) {
12087 				rack->r_ctl.sack_count++;
12088 			}
12089 			if (rack->r_ctl.sack_count > 0xfff00000) {
12090 				/*
12091 				 * reduce the number to keep us under
12092 				 * a uint32_t.
12093 				 */
12094 				rack->r_ctl.ack_count /= 2;
12095 				rack->r_ctl.sack_count /= 2;
12096 			}
12097 		}
12098 		goto out;
12099 	}
12100 	/* Its a sack of some sort */
12101 	rack->r_ctl.sack_count += num_sack_blks;
12102 	if (rack->r_ctl.sack_count > 0xfff00000) {
12103 		/*
12104 		 * reduce the number to keep us under
12105 		 * a uint32_t.
12106 		 */
12107 		rack->r_ctl.ack_count /= 2;
12108 		rack->r_ctl.sack_count /= 2;
12109 	}
12110 	if (num_sack_blks < 2) {
12111 		/* Only one, we don't need to sort */
12112 		goto do_sack_work;
12113 	}
12114 	/* Sort the sacks */
12115 	for (i = 0; i < num_sack_blks; i++) {
12116 		for (j = i + 1; j < num_sack_blks; j++) {
12117 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
12118 				sack = sack_blocks[i];
12119 				sack_blocks[i] = sack_blocks[j];
12120 				sack_blocks[j] = sack;
12121 			}
12122 		}
12123 	}
12124 	/*
12125 	 * Now are any of the sack block ends the same (yes some
12126 	 * implementations send these)?
12127 	 */
12128 again:
12129 	if (num_sack_blks == 0)
12130 		goto out_with_totals;
12131 	if (num_sack_blks > 1) {
12132 		for (i = 0; i < num_sack_blks; i++) {
12133 			for (j = i + 1; j < num_sack_blks; j++) {
12134 				if (sack_blocks[i].end == sack_blocks[j].end) {
12135 					/*
12136 					 * Ok these two have the same end we
12137 					 * want the smallest end and then
12138 					 * throw away the larger and start
12139 					 * again.
12140 					 */
12141 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
12142 						/*
12143 						 * The second block covers
12144 						 * more area use that
12145 						 */
12146 						sack_blocks[i].start = sack_blocks[j].start;
12147 					}
12148 					/*
12149 					 * Now collapse out the dup-sack and
12150 					 * lower the count
12151 					 */
12152 					for (k = (j + 1); k < num_sack_blks; k++) {
12153 						sack_blocks[j].start = sack_blocks[k].start;
12154 						sack_blocks[j].end = sack_blocks[k].end;
12155 						j++;
12156 					}
12157 					num_sack_blks--;
12158 					goto again;
12159 				}
12160 			}
12161 		}
12162 	}
12163 do_sack_work:
12164 	/*
12165 	 * First lets look to see if
12166 	 * we have retransmitted and
12167 	 * can use the transmit next?
12168 	 */
12169 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12170 	if (rsm &&
12171 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
12172 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
12173 		/*
12174 		 * We probably did the FR and the next
12175 		 * SACK in continues as we would expect.
12176 		 */
12177 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &no_extra, &moved_two, segsiz);
12178 		if (acked) {
12179 			rack->r_wanted_output = 1;
12180 			changed += acked;
12181 		}
12182 		if (num_sack_blks == 1) {
12183 			/*
12184 			 * This is what we would expect from
12185 			 * a normal implementation to happen
12186 			 * after we have retransmitted the FR,
12187 			 * i.e the sack-filter pushes down
12188 			 * to 1 block and the next to be retransmitted
12189 			 * is the sequence in the sack block (has more
12190 			 * are acked). Count this as ACK'd data to boost
12191 			 * up the chances of recovering any false positives.
12192 			 */
12193 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
12194 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
12195 			counter_u64_add(rack_express_sack, 1);
12196 			if (rack->r_ctl.ack_count > 0xfff00000) {
12197 				/*
12198 				 * reduce the number to keep us under
12199 				 * a uint32_t.
12200 				 */
12201 				rack->r_ctl.ack_count /= 2;
12202 				rack->r_ctl.sack_count /= 2;
12203 			}
12204 			if (moved_two) {
12205 				/*
12206 				 * If we did not get a SACK for at least a MSS and
12207 				 * had to move at all, or if we moved more than our
12208 				 * threshold, it counts against the "extra" move.
12209 				 */
12210 				rack->r_ctl.sack_moved_extra += moved_two;
12211 				rack->r_ctl.sack_noextra_move += no_extra;
12212 				counter_u64_add(rack_move_some, 1);
12213 			} else {
12214 				/*
12215 				 * else we did not have to move
12216 				 * any more than we would expect.
12217 				 */
12218 				rack->r_ctl.sack_noextra_move += no_extra;
12219 				rack->r_ctl.sack_noextra_move++;
12220 				counter_u64_add(rack_move_none, 1);
12221 			}
12222 			if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
12223 			    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
12224 				rack->r_ctl.sack_moved_extra /= 2;
12225 				rack->r_ctl.sack_noextra_move /= 2;
12226 			}
12227 			goto out_with_totals;
12228 		} else {
12229 			/*
12230 			 * Start the loop through the
12231 			 * rest of blocks, past the first block.
12232 			 */
12233 			loop_start = 1;
12234 		}
12235 	}
12236 	counter_u64_add(rack_sack_total, 1);
12237 	rsm = rack->r_ctl.rc_sacklast;
12238 	for (i = loop_start; i < num_sack_blks; i++) {
12239 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &no_extra, &moved_two, segsiz);
12240 		if (acked) {
12241 			rack->r_wanted_output = 1;
12242 			changed += acked;
12243 		}
12244 		if (moved_two) {
12245 			/*
12246 			 * If we did not get a SACK for at least a MSS and
12247 			 * had to move at all, or if we moved more than our
12248 			 * threshold, it counts against the "extra" move.
12249 			 */
12250 			rack->r_ctl.sack_moved_extra += moved_two;
12251 			rack->r_ctl.sack_noextra_move += no_extra;
12252 			counter_u64_add(rack_move_some, 1);
12253 		} else {
12254 			/*
12255 			 * else we did not have to move
12256 			 * any more than we would expect.
12257 			 */
12258 			rack->r_ctl.sack_noextra_move += no_extra;
12259 			rack->r_ctl.sack_noextra_move++;
12260 			counter_u64_add(rack_move_none, 1);
12261 		}
12262 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
12263 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
12264 			rack->r_ctl.sack_moved_extra /= 2;
12265 			rack->r_ctl.sack_noextra_move /= 2;
12266 		}
12267 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
12268 			/*
12269 			 * If the SACK was not a full MSS then
12270 			 * we add to sack_count the number of
12271 			 * MSS's (or possibly more than
12272 			 * a MSS if its a TSO send) we had to skip by.
12273 			 */
12274 			rack->r_ctl.sack_count += moved_two;
12275 			if (rack->r_ctl.sack_count > 0xfff00000) {
12276 				rack->r_ctl.ack_count /= 2;
12277 				rack->r_ctl.sack_count /= 2;
12278 			}
12279 			counter_u64_add(rack_sack_total, moved_two);
12280 		}
12281 		/*
12282 		 * Now we need to setup for the next
12283 		 * round. First we make sure we won't
12284 		 * exceed the size of our uint32_t on
12285 		 * the various counts, and then clear out
12286 		 * moved_two.
12287 		 */
12288 		moved_two = 0;
12289 		no_extra = 0;
12290 	}
12291 out_with_totals:
12292 	if (num_sack_blks > 1) {
12293 		/*
12294 		 * You get an extra stroke if
12295 		 * you have more than one sack-blk, this
12296 		 * could be where we are skipping forward
12297 		 * and the sack-filter is still working, or
12298 		 * it could be an attacker constantly
12299 		 * moving us.
12300 		 */
12301 		rack->r_ctl.sack_moved_extra++;
12302 		counter_u64_add(rack_move_some, 1);
12303 	}
12304 out:
12305 #ifdef TCP_SAD_DETECTION
12306 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
12307 #endif
12308 	if (changed) {
12309 		/* Something changed cancel the rack timer */
12310 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12311 	}
12312 	tsused = tcp_get_usecs(NULL);
12313 	rsm = tcp_rack_output(tp, rack, tsused);
12314 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
12315 	    rsm &&
12316 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
12317 		/* Enter recovery */
12318 		entered_recovery = 1;
12319 		rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
12320 		/*
12321 		 * When we enter recovery we need to assure we send
12322 		 * one packet.
12323 		 */
12324 		if (rack->rack_no_prr == 0) {
12325 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
12326 			rack_log_to_prr(rack, 8, 0, __LINE__);
12327 		}
12328 		rack->r_timer_override = 1;
12329 		rack->r_early = 0;
12330 		rack->r_ctl.rc_agg_early = 0;
12331 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
12332 		   rsm &&
12333 		   (rack->r_rr_config == 3)) {
12334 		/*
12335 		 * Assure we can output and we get no
12336 		 * remembered pace time except the retransmit.
12337 		 */
12338 		rack->r_timer_override = 1;
12339 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12340 		rack->r_ctl.rc_resend = rsm;
12341 	}
12342 	if (IN_FASTRECOVERY(tp->t_flags) &&
12343 	    (rack->rack_no_prr == 0) &&
12344 	    (entered_recovery == 0)) {
12345 		rack_update_prr(tp, rack, changed, th_ack);
12346 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
12347 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
12348 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
12349 			/*
12350 			 * If you are pacing output you don't want
12351 			 * to override.
12352 			 */
12353 			rack->r_early = 0;
12354 			rack->r_ctl.rc_agg_early = 0;
12355 			rack->r_timer_override = 1;
12356 		}
12357 	}
12358 }
12359 
12360 static void
12361 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
12362 {
12363 	struct rack_sendmap *rsm;
12364 
12365 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12366 	while (rsm) {
12367 		/*
12368 		 * We need to skip anything already set
12369 		 * to be retransmitted.
12370 		 */
12371 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
12372 		    (rsm->r_flags & RACK_MUST_RXT)) {
12373 			rsm = TAILQ_NEXT(rsm, r_tnext);
12374 			continue;
12375 		}
12376 		break;
12377 	}
12378 	if (rsm && (rsm->r_dupack < 0xff)) {
12379 		rsm->r_dupack++;
12380 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
12381 			struct timeval tv;
12382 			uint32_t cts;
12383 			/*
12384 			 * Here we see if we need to retransmit. For
12385 			 * a SACK type connection if enough time has passed
12386 			 * we will get a return of the rsm. For a non-sack
12387 			 * connection we will get the rsm returned if the
12388 			 * dupack value is 3 or more.
12389 			 */
12390 			cts = tcp_get_usecs(&tv);
12391 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
12392 			if (rack->r_ctl.rc_resend != NULL) {
12393 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
12394 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
12395 							 th_ack,  __LINE__);
12396 				}
12397 				rack->r_wanted_output = 1;
12398 				rack->r_timer_override = 1;
12399 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
12400 			}
12401 		} else {
12402 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
12403 		}
12404 	}
12405 }
12406 
12407 static void
12408 rack_check_bottom_drag(struct tcpcb *tp,
12409 		       struct tcp_rack *rack,
12410 		       struct socket *so)
12411 {
12412 	/*
12413 	 * So what is dragging bottom?
12414 	 *
12415 	 * Dragging bottom means you were under pacing and had a
12416 	 * delay in processing inbound acks waiting on our pacing
12417 	 * timer to expire. While you were waiting all of the acknowledgments
12418 	 * for the packets you sent have arrived. This means we are pacing
12419 	 * way underneath the bottleneck to the point where our Goodput
12420 	 * measurements stop working, since they require more than one
12421 	 * ack (usually at least 8 packets worth with multiple acks so we can
12422 	 * gauge the inter-ack times). If that occurs we have a real problem
12423 	 * since we are stuck in a hole that we can't get out of without
12424 	 * something speeding us up.
12425 	 *
12426 	 * We also check to see if we are widdling down to just one segment
12427 	 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
12428 	 * then we are adding the delayed ack interval into our measurments and
12429 	 * we need to speed up slightly.
12430 	 */
12431 	uint32_t segsiz, minseg;
12432 
12433 	segsiz = ctf_fixed_maxseg(tp);
12434 	minseg = segsiz;
12435 	if (tp->snd_max == tp->snd_una) {
12436 		/*
12437 		 * We are doing dynamic pacing and we are way
12438 		 * under. Basically everything got acked while
12439 		 * we were still waiting on the pacer to expire.
12440 		 *
12441 		 * This means we need to boost the b/w in
12442 		 * addition to any earlier boosting of
12443 		 * the multiplier.
12444 		 */
12445 		uint64_t lt_bw;
12446 
12447 		tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
12448 		lt_bw = rack_get_lt_bw(rack);
12449 		rack->rc_dragged_bottom = 1;
12450 		rack_validate_multipliers_at_or_above100(rack);
12451 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
12452 		    (rack->dis_lt_bw == 0) &&
12453 		    (rack->use_lesser_lt_bw == 0) &&
12454 		    (lt_bw > 0)) {
12455 			/*
12456 			 * Lets use the long-term b/w we have
12457 			 * been getting as a base.
12458 			 */
12459 			if (rack->rc_gp_filled == 0) {
12460 				if (lt_bw > ONE_POINT_TWO_MEG) {
12461 					/*
12462 					 * If we have no measurement
12463 					 * don't let us set in more than
12464 					 * 1.2Mbps. If we are still too
12465 					 * low after pacing with this we
12466 					 * will hopefully have a max b/w
12467 					 * available to sanity check things.
12468 					 */
12469 					lt_bw = ONE_POINT_TWO_MEG;
12470 				}
12471 				rack->r_ctl.rc_rtt_diff = 0;
12472 				rack->r_ctl.gp_bw = lt_bw;
12473 				rack->rc_gp_filled = 1;
12474 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
12475 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
12476 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
12477 			} else if (lt_bw > rack->r_ctl.gp_bw) {
12478 				rack->r_ctl.rc_rtt_diff = 0;
12479 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
12480 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
12481 				rack->r_ctl.gp_bw = lt_bw;
12482 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
12483 			} else
12484 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
12485 			if ((rack->gp_ready == 0) &&
12486 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
12487 				/* We have enough measurements now */
12488 				rack->gp_ready = 1;
12489 				if (rack->dgp_on ||
12490 				    rack->rack_hibeta)
12491 					rack_set_cc_pacing(rack);
12492 				if (rack->defer_options)
12493 					rack_apply_deferred_options(rack);
12494 			}
12495 		} else {
12496 			/*
12497 			 * zero rtt possibly?, settle for just an old increase.
12498 			 */
12499 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
12500 		}
12501 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
12502 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
12503 					       minseg)) &&
12504 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
12505 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
12506 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
12507 		    (segsiz * rack_req_segs))) {
12508 		/*
12509 		 * We are doing dynamic GP pacing and
12510 		 * we have everything except 1MSS or less
12511 		 * bytes left out. We are still pacing away.
12512 		 * And there is data that could be sent, This
12513 		 * means we are inserting delayed ack time in
12514 		 * our measurements because we are pacing too slow.
12515 		 */
12516 		rack_validate_multipliers_at_or_above100(rack);
12517 		rack->rc_dragged_bottom = 1;
12518 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
12519 	}
12520 }
12521 
12522 #ifdef TCP_REQUEST_TRK
12523 static void
12524 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
12525 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
12526 {
12527 	int do_log;
12528 
12529 	do_log = tcp_bblogging_on(rack->rc_tp);
12530 	if (do_log == 0) {
12531 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
12532 			return;
12533 		/* We only allow the three below with point logging on */
12534 		if ((mod != HYBRID_LOG_RULES_APP) &&
12535 		    (mod != HYBRID_LOG_RULES_SET) &&
12536 		    (mod != HYBRID_LOG_REQ_COMP))
12537 			return;
12538 
12539 	}
12540 	if (do_log) {
12541 		union tcp_log_stackspecific log;
12542 		struct timeval tv;
12543 
12544 		/* Convert our ms to a microsecond */
12545 		memset(&log, 0, sizeof(log));
12546 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12547 		log.u_bbr.flex1 = seq;
12548 		log.u_bbr.cwnd_gain = line;
12549 		if (cur != NULL) {
12550 			uint64_t off;
12551 
12552 			log.u_bbr.flex2 = cur->start_seq;
12553 			log.u_bbr.flex3 = cur->end_seq;
12554 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
12555 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
12556 			log.u_bbr.flex6 = cur->flags;
12557 			log.u_bbr.pkts_out = cur->hybrid_flags;
12558 			log.u_bbr.rttProp = cur->timestamp;
12559 			log.u_bbr.cur_del_rate = cur->cspr;
12560 			log.u_bbr.bw_inuse = cur->start;
12561 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
12562 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
12563 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
12564 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
12565 			log.u_bbr.inhpts = 1;
12566 #ifdef TCP_REQUEST_TRK
12567 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
12568 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
12569 #endif
12570 		} else {
12571 			log.u_bbr.flex2 = err;
12572 		}
12573 		/*
12574 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
12575 		 */
12576 		log.u_bbr.flex7 = rack->rc_catch_up;
12577 		log.u_bbr.flex7 <<= 1;
12578 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
12579 		log.u_bbr.flex7 <<= 1;
12580 		log.u_bbr.flex7 |= rack->dgp_on;
12581 		/*
12582 		 * Compose bbr_state to be a bit wise 0000ADHF
12583 		 * where A is the always_pace flag
12584 		 * where D is the dgp_on flag
12585 		 * where H is the hybrid_mode on flag
12586 		 * where F is the use_fixed_rate flag.
12587 		 */
12588 		log.u_bbr.bbr_state = rack->rc_always_pace;
12589 		log.u_bbr.bbr_state <<= 1;
12590 		log.u_bbr.bbr_state |= rack->dgp_on;
12591 		log.u_bbr.bbr_state <<= 1;
12592 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
12593 		log.u_bbr.bbr_state <<= 1;
12594 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
12595 		log.u_bbr.flex8 = mod;
12596 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
12597 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
12598 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12599 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
12600 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
12601 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
12602 		tcp_log_event(rack->rc_tp, NULL,
12603 		    &rack->rc_inp->inp_socket->so_rcv,
12604 		    &rack->rc_inp->inp_socket->so_snd,
12605 		    TCP_HYBRID_PACING_LOG, 0,
12606 	            0, &log, false, NULL, __func__, __LINE__, &tv);
12607 	}
12608 }
12609 #endif
12610 
12611 #ifdef TCP_REQUEST_TRK
12612 static void
12613 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
12614 {
12615 	struct tcp_sendfile_track *rc_cur, *orig_ent;
12616 	struct tcpcb *tp;
12617 	int err = 0;
12618 
12619 	orig_ent = rack->r_ctl.rc_last_sft;
12620 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
12621 	if (rc_cur == NULL) {
12622 		/* If not in the beginning what about the end piece */
12623 		if (rack->rc_hybrid_mode)
12624 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
12625 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
12626 	} else {
12627 		err = 12345;
12628 	}
12629 	/* If we find no parameters we are in straight DGP mode */
12630 	if(rc_cur == NULL) {
12631 		/* None found for this seq, just DGP for now */
12632 		if (rack->rc_hybrid_mode) {
12633 			rack->r_ctl.client_suggested_maxseg = 0;
12634 			rack->rc_catch_up = 0;
12635 			if (rack->cspr_is_fcc == 0)
12636 				rack->r_ctl.bw_rate_cap = 0;
12637 			else
12638 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
12639 		}
12640 		if (rack->rc_hybrid_mode) {
12641 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
12642 		}
12643 		if (rack->r_ctl.rc_last_sft) {
12644 			rack->r_ctl.rc_last_sft = NULL;
12645 		}
12646 		return;
12647 	}
12648 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
12649 		/* This entry was never setup for hybrid pacing on/off etc */
12650 		if (rack->rc_hybrid_mode) {
12651 			rack->r_ctl.client_suggested_maxseg = 0;
12652 			rack->rc_catch_up = 0;
12653 			rack->r_ctl.bw_rate_cap = 0;
12654 		}
12655 		if (rack->r_ctl.rc_last_sft) {
12656 			rack->r_ctl.rc_last_sft = NULL;
12657 		}
12658 		if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
12659 			rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
12660 			rc_cur->first_send = cts;
12661 			rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
12662 			rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
12663 		}
12664 		return;
12665 	}
12666 	/*
12667 	 * Ok if we have a new entry *or* have never
12668 	 * set up an entry we need to proceed. If
12669 	 * we have already set it up this entry we
12670 	 * just continue along with what we already
12671 	 * setup.
12672 	 */
12673 	tp = rack->rc_tp;
12674 	if ((rack->r_ctl.rc_last_sft != NULL) &&
12675 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
12676 		/* Its already in place */
12677 		if (rack->rc_hybrid_mode)
12678 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
12679 		return;
12680 	}
12681 	if (rack->rc_hybrid_mode == 0) {
12682 		rack->r_ctl.rc_last_sft = rc_cur;
12683 		if (orig_ent) {
12684 			orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
12685 			orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
12686 			orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
12687 		}
12688 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
12689 		return;
12690 	}
12691 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
12692 		/* Compensate for all the header overhead's */
12693 		if (rack->cspr_is_fcc == 0)
12694 			rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
12695 		else
12696 			rack->r_ctl.fillcw_cap =  rack_compensate_for_linerate(rack, rc_cur->cspr);
12697 	} else {
12698 		if (rack->rc_hybrid_mode) {
12699 			if (rack->cspr_is_fcc == 0)
12700 				rack->r_ctl.bw_rate_cap = 0;
12701 			else
12702 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
12703 		}
12704 	}
12705 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
12706 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
12707 	else
12708 		rack->r_ctl.client_suggested_maxseg = 0;
12709 	if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
12710 		/*
12711 		 * It is the same timestamp as the previous one
12712 		 * add the hybrid flag that will indicate we use
12713 		 * sendtime not arrival time for catch-up mode.
12714 		 */
12715 		rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
12716 	}
12717 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
12718 	    (rc_cur->cspr > 0)) {
12719 		uint64_t len;
12720 
12721 		rack->rc_catch_up = 1;
12722 		/*
12723 		 * Calculate the deadline time, first set the
12724 		 * time to when the request arrived.
12725 		 */
12726 		if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
12727 			/*
12728 			 * For cases where its a duplicate tm (we received more
12729 			 * than one request for a tm) we want to use now, the point
12730 			 * where we are just sending the first bit of the request.
12731 			 */
12732 			rc_cur->deadline = cts;
12733 		} else {
12734 			/*
12735 			 * Here we have a different tm from the last request
12736 			 * so we want to use arrival time as our base.
12737 			 */
12738 			rc_cur->deadline = rc_cur->localtime;
12739 		}
12740 		/*
12741 		 * Next calculate the length and compensate for
12742 		 * TLS if need be.
12743 		 */
12744 		len = rc_cur->end - rc_cur->start;
12745 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
12746 			/*
12747 			 * This session is doing TLS. Take a swag guess
12748 			 * at the overhead.
12749 			 */
12750 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
12751 		}
12752 		/*
12753 		 * Now considering the size, and the cspr, what is the time that
12754 		 * would be required at the cspr rate. Here we use the raw
12755 		 * cspr value since the client only looks at the raw data. We
12756 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
12757 		 * That will get made up for in the CU pacing rate set.
12758 		 */
12759 		len *= HPTS_USEC_IN_SEC;
12760 		len /= rc_cur->cspr;
12761 		rc_cur->deadline += len;
12762 	} else {
12763 		rack->rc_catch_up = 0;
12764 		rc_cur->deadline = 0;
12765 	}
12766 	if (rack->r_ctl.client_suggested_maxseg != 0) {
12767 		/*
12768 		 * We need to reset the max pace segs if we have a
12769 		 * client_suggested_maxseg.
12770 		 */
12771 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12772 	}
12773 	if (orig_ent) {
12774 		orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
12775 		orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
12776 		orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
12777 	}
12778 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
12779 	/* Remember it for next time and for CU mode */
12780 	rack->r_ctl.rc_last_sft = rc_cur;
12781 	rack->r_ctl.last_tm_mark = rc_cur->timestamp;
12782 }
12783 #endif
12784 
12785 static void
12786 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
12787 {
12788 #ifdef TCP_REQUEST_TRK
12789 	struct tcp_sendfile_track *ent;
12790 
12791 	ent = rack->r_ctl.rc_last_sft;
12792 	if ((ent == NULL) ||
12793 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
12794 	    (SEQ_GEQ(seq, ent->end_seq))) {
12795 		/* Time to update the track. */
12796 		rack_set_dgp_hybrid_mode(rack, seq, len, cts);
12797 		ent = rack->r_ctl.rc_last_sft;
12798 	}
12799 	/* Out of all */
12800 	if (ent == NULL) {
12801 		return;
12802 	}
12803 	if (SEQ_LT(ent->end_seq, (seq + len))) {
12804 		/*
12805 		 * This is the case where our end_seq guess
12806 		 * was wrong. This is usually due to TLS having
12807 		 * more bytes then our guess. It could also be the
12808 		 * case that the client sent in two requests closely
12809 		 * and the SB is full of both so we are sending part
12810 		 * of each (end|beg). In such a case lets move this
12811 		 * guys end to match the end of this send. That
12812 		 * way it will complete when all of it is acked.
12813 		 */
12814 		ent->end_seq = (seq + len);
12815 		if (rack->rc_hybrid_mode)
12816 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
12817 	}
12818 	/* Now validate we have set the send time of this one */
12819 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
12820 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
12821 		ent->first_send = cts;
12822 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
12823 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
12824 	}
12825 #endif
12826 }
12827 
12828 static void
12829 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
12830 {
12831 	/*
12832 	 * The fast output path is enabled and we
12833 	 * have moved the cumack forward. Lets see if
12834 	 * we can expand forward the fast path length by
12835 	 * that amount. What we would ideally like to
12836 	 * do is increase the number of bytes in the
12837 	 * fast path block (left_to_send) by the
12838 	 * acked amount. However we have to gate that
12839 	 * by two factors:
12840 	 * 1) The amount outstanding and the rwnd of the peer
12841 	 *    (i.e. we don't want to exceed the rwnd of the peer).
12842 	 *    <and>
12843 	 * 2) The amount of data left in the socket buffer (i.e.
12844 	 *    we can't send beyond what is in the buffer).
12845 	 *
12846 	 * Note that this does not take into account any increase
12847 	 * in the cwnd. We will only extend the fast path by
12848 	 * what was acked.
12849 	 */
12850 	uint32_t new_total, gating_val;
12851 
12852 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
12853 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
12854 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
12855 	if (new_total <= gating_val) {
12856 		/* We can increase left_to_send by the acked amount */
12857 		counter_u64_add(rack_extended_rfo, 1);
12858 		rack->r_ctl.fsb.left_to_send = new_total;
12859 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
12860 			("rack:%p left_to_send:%u sbavail:%u out:%u",
12861 			 rack, rack->r_ctl.fsb.left_to_send,
12862 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
12863 			 (tp->snd_max - tp->snd_una)));
12864 
12865 	}
12866 }
12867 
12868 static void
12869 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
12870 {
12871 	/*
12872 	 * Here any sendmap entry that points to the
12873 	 * beginning mbuf must be adjusted to the correct
12874 	 * offset. This must be called with:
12875 	 * 1) The socket buffer locked
12876 	 * 2) snd_una adjusted to its new position.
12877 	 *
12878 	 * Note that (2) implies rack_ack_received has also
12879 	 * been called and all the sbcut's have been done.
12880 	 *
12881 	 * We grab the first mbuf in the socket buffer and
12882 	 * then go through the front of the sendmap, recalculating
12883 	 * the stored offset for any sendmap entry that has
12884 	 * that mbuf. We must use the sb functions to do this
12885 	 * since its possible an add was done has well as
12886 	 * the subtraction we may have just completed. This should
12887 	 * not be a penalty though, since we just referenced the sb
12888 	 * to go in and trim off the mbufs that we freed (of course
12889 	 * there will be a penalty for the sendmap references though).
12890 	 *
12891 	 * Note also with INVARIANT on, we validate with a KASSERT
12892 	 * that the first sendmap entry has a soff of 0.
12893 	 *
12894 	 */
12895 	struct mbuf *m;
12896 	struct rack_sendmap *rsm;
12897 	tcp_seq snd_una;
12898 #ifdef INVARIANTS
12899 	int first_processed = 0;
12900 #endif
12901 
12902 	snd_una = rack->rc_tp->snd_una;
12903 	SOCKBUF_LOCK_ASSERT(sb);
12904 	m = sb->sb_mb;
12905 	rsm = tqhash_min(rack->r_ctl.tqh);
12906 	if ((rsm == NULL) || (m == NULL)) {
12907 		/* Nothing outstanding */
12908 		return;
12909 	}
12910 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
12911 	KASSERT((rsm->m == m),
12912 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
12913 		 rack, sb, rsm));
12914 	while (rsm->m && (rsm->m == m)) {
12915 		/* one to adjust */
12916 #ifdef INVARIANTS
12917 		struct mbuf *tm;
12918 		uint32_t soff;
12919 
12920 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
12921 		if ((rsm->orig_m_len != m->m_len) ||
12922 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
12923 			rack_adjust_orig_mlen(rsm);
12924 		}
12925 		if (first_processed == 0) {
12926 			KASSERT((rsm->soff == 0),
12927 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
12928 				 rack, rsm));
12929 			first_processed = 1;
12930 		}
12931 		if ((rsm->soff != soff) || (rsm->m != tm)) {
12932 			/*
12933 			 * This is not a fatal error, we anticipate it
12934 			 * might happen (the else code), so we count it here
12935 			 * so that under invariant we can see that it really
12936 			 * does happen.
12937 			 */
12938 			counter_u64_add(rack_adjust_map_bw, 1);
12939 		}
12940 		rsm->m = tm;
12941 		rsm->soff = soff;
12942 		if (tm) {
12943 			rsm->orig_m_len = rsm->m->m_len;
12944 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12945 		} else {
12946 			rsm->orig_m_len = 0;
12947 			rsm->orig_t_space = 0;
12948 		}
12949 #else
12950 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
12951 		if (rsm->m) {
12952 			rsm->orig_m_len = rsm->m->m_len;
12953 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12954 		} else {
12955 			rsm->orig_m_len = 0;
12956 			rsm->orig_t_space = 0;
12957 		}
12958 #endif
12959 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
12960 		if (rsm == NULL)
12961 			break;
12962 	}
12963 }
12964 
12965 #ifdef TCP_REQUEST_TRK
12966 static inline void
12967 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
12968 {
12969 	struct tcp_sendfile_track *ent;
12970 	int i;
12971 
12972 	if ((rack->rc_hybrid_mode == 0) &&
12973 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
12974 		/*
12975 		 * Just do normal completions hybrid pacing is not on
12976 		 * and CLDL is off as well.
12977 		 */
12978 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
12979 		return;
12980 	}
12981 	/*
12982 	 * Originally I was just going to find the th_ack associated
12983 	 * with an entry. But then I realized a large strech ack could
12984 	 * in theory ack two or more requests at once. So instead we
12985 	 * need to find all entries that are completed by th_ack not
12986 	 * just a single entry and do our logging.
12987 	 */
12988 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12989 	while (ent != NULL) {
12990 		/*
12991 		 * We may be doing hybrid pacing or CLDL and need more details possibly
12992 		 * so we do it manually instead of calling
12993 		 * tcp_req_check_for_comp()
12994 		 */
12995 		uint64_t laa, tim, data, cbw, ftim;
12996 
12997 		/* Ok this ack frees it */
12998 		rack_log_hybrid(rack, th_ack,
12999 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
13000 		rack_log_hybrid_sends(rack, ent, __LINE__);
13001 		/* calculate the time based on the ack arrival */
13002 		data = ent->end - ent->start;
13003 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
13004 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
13005 			if (ent->first_send > ent->localtime)
13006 				ftim = ent->first_send;
13007 			else
13008 				ftim = ent->localtime;
13009 		} else {
13010 			/* TSNH */
13011 			ftim = ent->localtime;
13012 		}
13013 		if (laa > ent->localtime)
13014 			tim = laa - ftim;
13015 		else
13016 			tim = 0;
13017 		cbw = data * HPTS_USEC_IN_SEC;
13018 		if (tim > 0)
13019 			cbw /= tim;
13020 		else
13021 			cbw = 0;
13022 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
13023 		/*
13024 		 * Check to see if we are freeing what we are pointing to send wise
13025 		 * if so be sure to NULL the pointer so we know we are no longer
13026 		 * set to anything.
13027 		 */
13028 		if (ent == rack->r_ctl.rc_last_sft) {
13029 			rack->r_ctl.rc_last_sft = NULL;
13030 			if (rack->rc_hybrid_mode) {
13031 				rack->rc_catch_up = 0;
13032 				if (rack->cspr_is_fcc == 0)
13033 					rack->r_ctl.bw_rate_cap = 0;
13034 				else
13035 					rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
13036 				rack->r_ctl.client_suggested_maxseg = 0;
13037 			}
13038 		}
13039 		/* Generate the log that the tcp_netflix call would have */
13040 		tcp_req_log_req_info(rack->rc_tp, ent,
13041 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
13042 		/* Free it and see if there is another one */
13043 		tcp_req_free_a_slot(rack->rc_tp, ent);
13044 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
13045 	}
13046 }
13047 #endif
13048 
13049 
13050 /*
13051  * Return value of 1, we do not need to call rack_process_data().
13052  * return value of 0, rack_process_data can be called.
13053  * For ret_val if its 0 the TCP is locked, if its non-zero
13054  * its unlocked and probably unsafe to touch the TCB.
13055  */
13056 static int
13057 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13058     struct tcpcb *tp, struct tcpopt *to,
13059     uint32_t tiwin, int32_t tlen,
13060     int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
13061 {
13062 	int32_t ourfinisacked = 0;
13063 	int32_t nsegs, acked_amount;
13064 	int32_t acked;
13065 	struct mbuf *mfree;
13066 	struct tcp_rack *rack;
13067 	int32_t under_pacing = 0;
13068 	int32_t post_recovery = 0;
13069 	uint32_t p_cwnd;
13070 
13071 	INP_WLOCK_ASSERT(tptoinpcb(tp));
13072 
13073 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13074 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
13075 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
13076 				      &rack->r_ctl.challenge_ack_ts,
13077 				      &rack->r_ctl.challenge_ack_cnt);
13078 		rack->r_wanted_output = 1;
13079 		return (1);
13080 	}
13081 	if (rack->gp_ready &&
13082 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13083 		under_pacing = 1;
13084 	}
13085 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
13086 		int in_rec, dup_ack_struck = 0;
13087 		int dsack_seen = 0, sacks_seen = 0;
13088 
13089 		in_rec = IN_FASTRECOVERY(tp->t_flags);
13090 		if (rack->rc_in_persist) {
13091 			tp->t_rxtshift = 0;
13092 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13093 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13094 		}
13095 
13096 		if ((th->th_ack == tp->snd_una) &&
13097 		    (tiwin == tp->snd_wnd) &&
13098 		    (orig_tlen == 0) &&
13099 		    ((to->to_flags & TOF_SACK) == 0)) {
13100 			rack_strike_dupack(rack, th->th_ack);
13101 			dup_ack_struck = 1;
13102 		}
13103 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
13104 			     dup_ack_struck, &dsack_seen, &sacks_seen);
13105 		if ((rack->sack_attack_disable > 0) &&
13106 		    (th->th_ack == tp->snd_una) &&
13107 		    (tiwin == tp->snd_wnd) &&
13108 		    (orig_tlen == 0) &&
13109 		    (dsack_seen == 0) &&
13110 		    (sacks_seen > 0)) {
13111 			/*
13112 			 * If sacks have been disabled we may
13113 			 * want to strike a dup-ack "ignoring" the
13114 			 * sack as long as the sack was not a "dsack". Note
13115 			 * that if no sack is sent (TOF_SACK is off) then the
13116 			 * normal dsack code above rack_log_ack() would have
13117 			 * already struck. So this is just to catch the case
13118 			 * were we are ignoring sacks from this guy due to
13119 			 * it being a suspected attacker.
13120 			 */
13121 			rack_strike_dupack(rack, th->th_ack);
13122 		}
13123 
13124 	}
13125 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
13126 		/*
13127 		 * Old ack, behind (or duplicate to) the last one rcv'd
13128 		 * Note: We mark reordering is occuring if its
13129 		 * less than and we have not closed our window.
13130 		 */
13131 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
13132 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13133 			if (rack->r_ctl.rc_reorder_ts == 0)
13134 				rack->r_ctl.rc_reorder_ts = 1;
13135 		}
13136 		return (0);
13137 	}
13138 	/*
13139 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
13140 	 * something we sent.
13141 	 */
13142 	if (tp->t_flags & TF_NEEDSYN) {
13143 		/*
13144 		 * T/TCP: Connection was half-synchronized, and our SYN has
13145 		 * been ACK'd (so connection is now fully synchronized).  Go
13146 		 * to non-starred state, increment snd_una for ACK of SYN,
13147 		 * and check if we can do window scaling.
13148 		 */
13149 		tp->t_flags &= ~TF_NEEDSYN;
13150 		tp->snd_una++;
13151 		/* Do window scaling? */
13152 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13153 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13154 			tp->rcv_scale = tp->request_r_scale;
13155 			/* Send window already scaled. */
13156 		}
13157 	}
13158 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13159 
13160 	acked = BYTES_THIS_ACK(tp, th);
13161 	if (acked) {
13162 		/*
13163 		 * Any time we move the cum-ack forward clear
13164 		 * keep-alive tied probe-not-answered. The
13165 		 * persists clears its own on entry.
13166 		 */
13167 		rack->probe_not_answered = 0;
13168 	}
13169 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
13170 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13171 	/*
13172 	 * If we just performed our first retransmit, and the ACK arrives
13173 	 * within our recovery window, then it was a mistake to do the
13174 	 * retransmit in the first place.  Recover our original cwnd and
13175 	 * ssthresh, and proceed to transmit where we left off.
13176 	 */
13177 	if ((tp->t_flags & TF_PREVVALID) &&
13178 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13179 		tp->t_flags &= ~TF_PREVVALID;
13180 		if (tp->t_rxtshift == 1 &&
13181 		    (int)(ticks - tp->t_badrxtwin) < 0)
13182 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
13183 	}
13184 	if (acked) {
13185 		/* assure we are not backed off */
13186 		tp->t_rxtshift = 0;
13187 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13188 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13189 		rack->rc_tlp_in_progress = 0;
13190 		rack->r_ctl.rc_tlp_cnt_out = 0;
13191 		/*
13192 		 * If it is the RXT timer we want to
13193 		 * stop it, so we can restart a TLP.
13194 		 */
13195 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13196 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13197 #ifdef TCP_REQUEST_TRK
13198 		rack_req_check_for_comp(rack, th->th_ack);
13199 #endif
13200 	}
13201 	/*
13202 	 * If we have a timestamp reply, update smoothed round trip time. If
13203 	 * no timestamp is present but transmit timer is running and timed
13204 	 * sequence number was acked, update smoothed round trip time. Since
13205 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
13206 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
13207 	 * timer.
13208 	 *
13209 	 * Some boxes send broken timestamp replies during the SYN+ACK
13210 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
13211 	 * and blow up the retransmit timer.
13212 	 */
13213 	/*
13214 	 * If all outstanding data is acked, stop retransmit timer and
13215 	 * remember to restart (more output or persist). If there is more
13216 	 * data to be acked, restart retransmit timer, using current
13217 	 * (possibly backed-off) value.
13218 	 */
13219 	if (acked == 0) {
13220 		if (ofia)
13221 			*ofia = ourfinisacked;
13222 		return (0);
13223 	}
13224 	if (IN_RECOVERY(tp->t_flags)) {
13225 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
13226 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
13227 			tcp_rack_partialack(tp);
13228 		} else {
13229 			rack_post_recovery(tp, th->th_ack);
13230 			post_recovery = 1;
13231 			/*
13232 			 * Grab the segsiz, multiply by 2 and add the snd_cwnd
13233 			 * that is the max the CC should add if we are exiting
13234 			 * recovery and doing a late add.
13235 			 */
13236 			p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
13237 			p_cwnd <<= 1;
13238 			p_cwnd += tp->snd_cwnd;
13239 		}
13240 	} else if ((rack->rto_from_rec == 1) &&
13241 		   SEQ_GEQ(th->th_ack, tp->snd_recover)) {
13242 		/*
13243 		 * We were in recovery, hit a rxt timeout
13244 		 * and never re-entered recovery. The timeout(s)
13245 		 * made up all the lost data. In such a case
13246 		 * we need to clear the rto_from_rec flag.
13247 		 */
13248 		rack->rto_from_rec = 0;
13249 	}
13250 	/*
13251 	 * Let the congestion control algorithm update congestion control
13252 	 * related information. This typically means increasing the
13253 	 * congestion window.
13254 	 */
13255 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
13256 	if (post_recovery &&
13257 	    (tp->snd_cwnd > p_cwnd)) {
13258 		/* Must be non-newreno (cubic) getting too ahead of itself */
13259 		tp->snd_cwnd = p_cwnd;
13260 	}
13261 	SOCKBUF_LOCK(&so->so_snd);
13262 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
13263 	tp->snd_wnd -= acked_amount;
13264 	mfree = sbcut_locked(&so->so_snd, acked_amount);
13265 	if ((sbused(&so->so_snd) == 0) &&
13266 	    (acked > acked_amount) &&
13267 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13268 	    (tp->t_flags & TF_SENTFIN)) {
13269 		/*
13270 		 * We must be sure our fin
13271 		 * was sent and acked (we can be
13272 		 * in FIN_WAIT_1 without having
13273 		 * sent the fin).
13274 		 */
13275 		ourfinisacked = 1;
13276 	}
13277 	tp->snd_una = th->th_ack;
13278 	/* wakeups? */
13279 	if (acked_amount && sbavail(&so->so_snd))
13280 		rack_adjust_sendmap_head(rack, &so->so_snd);
13281 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13282 	/* NB: sowwakeup_locked() does an implicit unlock. */
13283 	sowwakeup_locked(so);
13284 	m_freem(mfree);
13285 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
13286 		tp->snd_recover = tp->snd_una;
13287 
13288 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
13289 		tp->snd_nxt = tp->snd_max;
13290 	}
13291 	if (under_pacing &&
13292 	    (rack->use_fixed_rate == 0) &&
13293 	    (rack->in_probe_rtt == 0) &&
13294 	    rack->rc_gp_dyn_mul &&
13295 	    rack->rc_always_pace) {
13296 		/* Check if we are dragging bottom */
13297 		rack_check_bottom_drag(tp, rack, so);
13298 	}
13299 	if (tp->snd_una == tp->snd_max) {
13300 		/* Nothing left outstanding */
13301 		tp->t_flags &= ~TF_PREVVALID;
13302 		rack->r_ctl.idle_snd_una = tp->snd_una;
13303 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13304 		if (rack->r_ctl.rc_went_idle_time == 0)
13305 			rack->r_ctl.rc_went_idle_time = 1;
13306 		rack->r_ctl.retran_during_recovery = 0;
13307 		rack->r_ctl.dsack_byte_cnt = 0;
13308 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13309 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
13310 			tp->t_acktime = 0;
13311 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13312 		rack->rc_suspicious = 0;
13313 		/* Set need output so persist might get set */
13314 		rack->r_wanted_output = 1;
13315 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13316 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13317 		    (sbavail(&so->so_snd) == 0) &&
13318 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13319 			/*
13320 			 * The socket was gone and the
13321 			 * peer sent data (now or in the past), time to
13322 			 * reset him.
13323 			 */
13324 			*ret_val = 1;
13325 			/* tcp_close will kill the inp pre-log the Reset */
13326 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13327 			tp = tcp_close(tp);
13328 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
13329 			return (1);
13330 		}
13331 	}
13332 	if (ofia)
13333 		*ofia = ourfinisacked;
13334 	return (0);
13335 }
13336 
13337 
13338 static void
13339 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
13340 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
13341 {
13342 	if (tcp_bblogging_on(rack->rc_tp)) {
13343 		union tcp_log_stackspecific log;
13344 		struct timeval tv;
13345 
13346 		memset(&log, 0, sizeof(log));
13347 		log.u_bbr.flex1 = cnt;
13348 		log.u_bbr.flex2 = split;
13349 		log.u_bbr.flex3 = out;
13350 		log.u_bbr.flex4 = line;
13351 		log.u_bbr.flex5 = rack->r_must_retran;
13352 		log.u_bbr.flex6 = flags;
13353 		log.u_bbr.flex7 = rack->rc_has_collapsed;
13354 		log.u_bbr.flex8 = dir;	/*
13355 					 * 1 is collapsed, 0 is uncollapsed,
13356 					 * 2 is log of a rsm being marked, 3 is a split.
13357 					 */
13358 		if (rsm == NULL)
13359 			log.u_bbr.rttProp = 0;
13360 		else
13361 			log.u_bbr.rttProp = (uint64_t)rsm;
13362 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13363 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13364 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
13365 		    &rack->rc_inp->inp_socket->so_rcv,
13366 		    &rack->rc_inp->inp_socket->so_snd,
13367 		    TCP_RACK_LOG_COLLAPSE, 0,
13368 		    0, &log, false, &tv);
13369 	}
13370 }
13371 
13372 static void
13373 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
13374 {
13375 	/*
13376 	 * Here all we do is mark the collapsed point and set the flag.
13377 	 * This may happen again and again, but there is no
13378 	 * sense splitting our map until we know where the
13379 	 * peer finally lands in the collapse.
13380 	 */
13381 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
13382 	if ((rack->rc_has_collapsed == 0) ||
13383 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
13384 		counter_u64_add(rack_collapsed_win_seen, 1);
13385 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
13386 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
13387 	rack->rc_has_collapsed = 1;
13388 	rack->r_collapse_point_valid = 1;
13389 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
13390 }
13391 
13392 static void
13393 rack_un_collapse_window(struct tcp_rack *rack, int line)
13394 {
13395 	struct rack_sendmap *nrsm, *rsm;
13396 	int cnt = 0, split = 0;
13397 	int insret __diagused;
13398 
13399 
13400 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
13401 	rack->rc_has_collapsed = 0;
13402 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
13403 	if (rsm == NULL) {
13404 		/* Nothing to do maybe the peer ack'ed it all */
13405 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
13406 		return;
13407 	}
13408 	/* Now do we need to split this one? */
13409 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
13410 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
13411 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
13412 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
13413 		if (nrsm == NULL) {
13414 			/* We can't get a rsm, mark all? */
13415 			nrsm = rsm;
13416 			goto no_split;
13417 		}
13418 		/* Clone it */
13419 		split = 1;
13420 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
13421 #ifndef INVARIANTS
13422 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
13423 #else
13424 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
13425 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
13426 			      nrsm, insret, rack, rsm);
13427 		}
13428 #endif
13429 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
13430 				 rack->r_ctl.last_collapse_point, __LINE__);
13431 		if (rsm->r_in_tmap) {
13432 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
13433 			nrsm->r_in_tmap = 1;
13434 		}
13435 		/*
13436 		 * Set in the new RSM as the
13437 		 * collapsed starting point
13438 		 */
13439 		rsm = nrsm;
13440 	}
13441 
13442 no_split:
13443 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
13444 		cnt++;
13445 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
13446 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
13447 		cnt++;
13448 	}
13449 	if (cnt) {
13450 		counter_u64_add(rack_collapsed_win, 1);
13451 	}
13452 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
13453 }
13454 
13455 static void
13456 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
13457 			int32_t tlen, int32_t tfo_syn)
13458 {
13459 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
13460 		rack_timer_cancel(tp, rack,
13461 				  rack->r_ctl.rc_rcvtime, __LINE__);
13462 		tp->t_flags |= TF_DELACK;
13463 	} else {
13464 		rack->r_wanted_output = 1;
13465 		tp->t_flags |= TF_ACKNOW;
13466 	}
13467 }
13468 
13469 static void
13470 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
13471 {
13472 	/*
13473 	 * If fast output is in progress, lets validate that
13474 	 * the new window did not shrink on us and make it
13475 	 * so fast output should end.
13476 	 */
13477 	if (rack->r_fast_output) {
13478 		uint32_t out;
13479 
13480 		/*
13481 		 * Calculate what we will send if left as is
13482 		 * and compare that to our send window.
13483 		 */
13484 		out = ctf_outstanding(tp);
13485 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
13486 			/* ok we have an issue */
13487 			if (out >= tp->snd_wnd) {
13488 				/* Turn off fast output the window is met or collapsed */
13489 				rack->r_fast_output = 0;
13490 			} else {
13491 				/* we have some room left */
13492 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
13493 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
13494 					/* If not at least 1 full segment never mind */
13495 					rack->r_fast_output = 0;
13496 				}
13497 			}
13498 		}
13499 	}
13500 }
13501 
13502 /*
13503  * Return value of 1, the TCB is unlocked and most
13504  * likely gone, return value of 0, the TCP is still
13505  * locked.
13506  */
13507 static int
13508 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
13509     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
13510     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
13511 {
13512 	/*
13513 	 * Update window information. Don't look at window if no ACK: TAC's
13514 	 * send garbage on first SYN.
13515 	 */
13516 	int32_t nsegs;
13517 	int32_t tfo_syn;
13518 	struct tcp_rack *rack;
13519 
13520 	INP_WLOCK_ASSERT(tptoinpcb(tp));
13521 
13522 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13523 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13524 	if ((thflags & TH_ACK) &&
13525 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
13526 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
13527 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
13528 		/* keep track of pure window updates */
13529 		if (tlen == 0 &&
13530 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
13531 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13532 		tp->snd_wnd = tiwin;
13533 		rack_validate_fo_sendwin_up(tp, rack);
13534 		tp->snd_wl1 = th->th_seq;
13535 		tp->snd_wl2 = th->th_ack;
13536 		if (tp->snd_wnd > tp->max_sndwnd)
13537 			tp->max_sndwnd = tp->snd_wnd;
13538 		rack->r_wanted_output = 1;
13539 	} else if (thflags & TH_ACK) {
13540 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
13541 			tp->snd_wnd = tiwin;
13542 			rack_validate_fo_sendwin_up(tp, rack);
13543 			tp->snd_wl1 = th->th_seq;
13544 			tp->snd_wl2 = th->th_ack;
13545 		}
13546 	}
13547 	if (tp->snd_wnd < ctf_outstanding(tp))
13548 		/* The peer collapsed the window */
13549 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
13550 	else if (rack->rc_has_collapsed)
13551 		rack_un_collapse_window(rack, __LINE__);
13552 	if ((rack->r_collapse_point_valid) &&
13553 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
13554 		rack->r_collapse_point_valid = 0;
13555 	/* Was persist timer active and now we have window space? */
13556 	if ((rack->rc_in_persist != 0) &&
13557 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13558 				rack->r_ctl.rc_pace_min_segs))) {
13559 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13560 		tp->snd_nxt = tp->snd_max;
13561 		/* Make sure we output to start the timer */
13562 		rack->r_wanted_output = 1;
13563 	}
13564 	/* Do we enter persists? */
13565 	if ((rack->rc_in_persist == 0) &&
13566 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13567 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13568 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13569 	    sbavail(&tptosocket(tp)->so_snd) &&
13570 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13571 		/*
13572 		 * Here the rwnd is less than
13573 		 * the pacing size, we are established,
13574 		 * nothing is outstanding, and there is
13575 		 * data to send. Enter persists.
13576 		 */
13577 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
13578 	}
13579 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
13580 		m_freem(m);
13581 		return (0);
13582 	}
13583 	/*
13584 	 * don't process the URG bit, ignore them drag
13585 	 * along the up.
13586 	 */
13587 	tp->rcv_up = tp->rcv_nxt;
13588 
13589 	/*
13590 	 * Process the segment text, merging it into the TCP sequencing
13591 	 * queue, and arranging for acknowledgment of receipt if necessary.
13592 	 * This process logically involves adjusting tp->rcv_wnd as data is
13593 	 * presented to the user (this happens in tcp_usrreq.c, case
13594 	 * PRU_RCVD).  If a FIN has already been received on this connection
13595 	 * then we just ignore the text.
13596 	 */
13597 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
13598 	    (tp->t_flags & TF_FASTOPEN));
13599 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
13600 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
13601 		tcp_seq save_start = th->th_seq;
13602 		tcp_seq save_rnxt  = tp->rcv_nxt;
13603 		int     save_tlen  = tlen;
13604 
13605 		m_adj(m, drop_hdrlen);	/* delayed header drop */
13606 		/*
13607 		 * Insert segment which includes th into TCP reassembly
13608 		 * queue with control block tp.  Set thflags to whether
13609 		 * reassembly now includes a segment with FIN.  This handles
13610 		 * the common case inline (segment is the next to be
13611 		 * received on an established connection, and the queue is
13612 		 * empty), avoiding linkage into and removal from the queue
13613 		 * and repetition of various conversions. Set DELACK for
13614 		 * segments received in order, but ack immediately when
13615 		 * segments are out of order (so fast retransmit can work).
13616 		 */
13617 		if (th->th_seq == tp->rcv_nxt &&
13618 		    SEGQ_EMPTY(tp) &&
13619 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
13620 		    tfo_syn)) {
13621 #ifdef NETFLIX_SB_LIMITS
13622 			u_int mcnt, appended;
13623 
13624 			if (so->so_rcv.sb_shlim) {
13625 				mcnt = m_memcnt(m);
13626 				appended = 0;
13627 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
13628 				    CFO_NOSLEEP, NULL) == false) {
13629 					counter_u64_add(tcp_sb_shlim_fails, 1);
13630 					m_freem(m);
13631 					return (0);
13632 				}
13633 			}
13634 #endif
13635 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
13636 			tp->rcv_nxt += tlen;
13637 			if (tlen &&
13638 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
13639 			    (tp->t_fbyte_in == 0)) {
13640 				tp->t_fbyte_in = ticks;
13641 				if (tp->t_fbyte_in == 0)
13642 					tp->t_fbyte_in = 1;
13643 				if (tp->t_fbyte_out && tp->t_fbyte_in)
13644 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
13645 			}
13646 			thflags = tcp_get_flags(th) & TH_FIN;
13647 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
13648 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
13649 			SOCKBUF_LOCK(&so->so_rcv);
13650 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13651 				m_freem(m);
13652 			} else {
13653 				int32_t newsize;
13654 
13655 				if (tlen > 0) {
13656 					newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
13657 					if (newsize)
13658 						if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
13659 							so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
13660 				}
13661 #ifdef NETFLIX_SB_LIMITS
13662 				appended =
13663 #endif
13664 					sbappendstream_locked(&so->so_rcv, m, 0);
13665 			}
13666 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
13667 			/* NB: sorwakeup_locked() does an implicit unlock. */
13668 			sorwakeup_locked(so);
13669 #ifdef NETFLIX_SB_LIMITS
13670 			if (so->so_rcv.sb_shlim && appended != mcnt)
13671 				counter_fo_release(so->so_rcv.sb_shlim,
13672 				    mcnt - appended);
13673 #endif
13674 		} else {
13675 			/*
13676 			 * XXX: Due to the header drop above "th" is
13677 			 * theoretically invalid by now.  Fortunately
13678 			 * m_adj() doesn't actually frees any mbufs when
13679 			 * trimming from the head.
13680 			 */
13681 			tcp_seq temp = save_start;
13682 
13683 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
13684 			tp->t_flags |= TF_ACKNOW;
13685 			if (tp->t_flags & TF_WAKESOR) {
13686 				tp->t_flags &= ~TF_WAKESOR;
13687 				/* NB: sorwakeup_locked() does an implicit unlock. */
13688 				sorwakeup_locked(so);
13689 			}
13690 		}
13691 		if ((tp->t_flags & TF_SACK_PERMIT) &&
13692 		    (save_tlen > 0) &&
13693 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
13694 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
13695 				/*
13696 				 * DSACK actually handled in the fastpath
13697 				 * above.
13698 				 */
13699 				tcp_update_sack_list(tp, save_start,
13700 				    save_start + save_tlen);
13701 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
13702 				if ((tp->rcv_numsacks >= 1) &&
13703 				    (tp->sackblks[0].end == save_start)) {
13704 					/*
13705 					 * Partial overlap, recorded at todrop
13706 					 * above.
13707 					 */
13708 					tcp_update_sack_list(tp,
13709 					    tp->sackblks[0].start,
13710 					    tp->sackblks[0].end);
13711 				} else {
13712 					tcp_update_dsack_list(tp, save_start,
13713 					    save_start + save_tlen);
13714 				}
13715 			} else if (tlen >= save_tlen) {
13716 				/* Update of sackblks. */
13717 				tcp_update_dsack_list(tp, save_start,
13718 				    save_start + save_tlen);
13719 			} else if (tlen > 0) {
13720 				tcp_update_dsack_list(tp, save_start,
13721 				    save_start + tlen);
13722 			}
13723 		}
13724 	} else {
13725 		m_freem(m);
13726 		thflags &= ~TH_FIN;
13727 	}
13728 
13729 	/*
13730 	 * If FIN is received ACK the FIN and let the user know that the
13731 	 * connection is closing.
13732 	 */
13733 	if (thflags & TH_FIN) {
13734 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
13735 			/* The socket upcall is handled by socantrcvmore. */
13736 			socantrcvmore(so);
13737 			/*
13738 			 * If connection is half-synchronized (ie NEEDSYN
13739 			 * flag on) then delay ACK, so it may be piggybacked
13740 			 * when SYN is sent. Otherwise, since we received a
13741 			 * FIN then no more input can be expected, send ACK
13742 			 * now.
13743 			 */
13744 			if (tp->t_flags & TF_NEEDSYN) {
13745 				rack_timer_cancel(tp, rack,
13746 				    rack->r_ctl.rc_rcvtime, __LINE__);
13747 				tp->t_flags |= TF_DELACK;
13748 			} else {
13749 				tp->t_flags |= TF_ACKNOW;
13750 			}
13751 			tp->rcv_nxt++;
13752 		}
13753 		switch (tp->t_state) {
13754 			/*
13755 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
13756 			 * CLOSE_WAIT state.
13757 			 */
13758 		case TCPS_SYN_RECEIVED:
13759 			tp->t_starttime = ticks;
13760 			/* FALLTHROUGH */
13761 		case TCPS_ESTABLISHED:
13762 			rack_timer_cancel(tp, rack,
13763 			    rack->r_ctl.rc_rcvtime, __LINE__);
13764 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
13765 			break;
13766 
13767 			/*
13768 			 * If still in FIN_WAIT_1 STATE FIN has not been
13769 			 * acked so enter the CLOSING state.
13770 			 */
13771 		case TCPS_FIN_WAIT_1:
13772 			rack_timer_cancel(tp, rack,
13773 			    rack->r_ctl.rc_rcvtime, __LINE__);
13774 			tcp_state_change(tp, TCPS_CLOSING);
13775 			break;
13776 
13777 			/*
13778 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
13779 			 * starting the time-wait timer, turning off the
13780 			 * other standard timers.
13781 			 */
13782 		case TCPS_FIN_WAIT_2:
13783 			rack_timer_cancel(tp, rack,
13784 			    rack->r_ctl.rc_rcvtime, __LINE__);
13785 			tcp_twstart(tp);
13786 			return (1);
13787 		}
13788 	}
13789 	/*
13790 	 * Return any desired output.
13791 	 */
13792 	if ((tp->t_flags & TF_ACKNOW) ||
13793 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
13794 		rack->r_wanted_output = 1;
13795 	}
13796 	return (0);
13797 }
13798 
13799 /*
13800  * Here nothing is really faster, its just that we
13801  * have broken out the fast-data path also just like
13802  * the fast-ack.
13803  */
13804 static int
13805 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
13806     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13807     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
13808 {
13809 	int32_t nsegs;
13810 	int32_t newsize = 0;	/* automatic sockbuf scaling */
13811 	struct tcp_rack *rack;
13812 #ifdef NETFLIX_SB_LIMITS
13813 	u_int mcnt, appended;
13814 #endif
13815 
13816 	/*
13817 	 * If last ACK falls within this segment's sequence numbers, record
13818 	 * the timestamp. NOTE that the test is modified according to the
13819 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
13820 	 */
13821 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
13822 		return (0);
13823 	}
13824 	if (tiwin && tiwin != tp->snd_wnd) {
13825 		return (0);
13826 	}
13827 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
13828 		return (0);
13829 	}
13830 	if (__predict_false((to->to_flags & TOF_TS) &&
13831 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
13832 		return (0);
13833 	}
13834 	if (__predict_false((th->th_ack != tp->snd_una))) {
13835 		return (0);
13836 	}
13837 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
13838 		return (0);
13839 	}
13840 	if ((to->to_flags & TOF_TS) != 0 &&
13841 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
13842 		tp->ts_recent_age = tcp_ts_getticks();
13843 		tp->ts_recent = to->to_tsval;
13844 	}
13845 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13846 	/*
13847 	 * This is a pure, in-sequence data packet with nothing on the
13848 	 * reassembly queue and we have enough buffer space to take it.
13849 	 */
13850 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13851 
13852 #ifdef NETFLIX_SB_LIMITS
13853 	if (so->so_rcv.sb_shlim) {
13854 		mcnt = m_memcnt(m);
13855 		appended = 0;
13856 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
13857 		    CFO_NOSLEEP, NULL) == false) {
13858 			counter_u64_add(tcp_sb_shlim_fails, 1);
13859 			m_freem(m);
13860 			return (1);
13861 		}
13862 	}
13863 #endif
13864 	/* Clean receiver SACK report if present */
13865 	if (tp->rcv_numsacks)
13866 		tcp_clean_sackreport(tp);
13867 	KMOD_TCPSTAT_INC(tcps_preddat);
13868 	tp->rcv_nxt += tlen;
13869 	if (tlen &&
13870 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
13871 	    (tp->t_fbyte_in == 0)) {
13872 		tp->t_fbyte_in = ticks;
13873 		if (tp->t_fbyte_in == 0)
13874 			tp->t_fbyte_in = 1;
13875 		if (tp->t_fbyte_out && tp->t_fbyte_in)
13876 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
13877 	}
13878 	/*
13879 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
13880 	 */
13881 	tp->snd_wl1 = th->th_seq;
13882 	/*
13883 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
13884 	 */
13885 	tp->rcv_up = tp->rcv_nxt;
13886 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
13887 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
13888 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
13889 
13890 	/* Add data to socket buffer. */
13891 	SOCKBUF_LOCK(&so->so_rcv);
13892 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13893 		m_freem(m);
13894 	} else {
13895 		/*
13896 		 * Set new socket buffer size. Give up when limit is
13897 		 * reached.
13898 		 */
13899 		if (newsize)
13900 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
13901 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
13902 		m_adj(m, drop_hdrlen);	/* delayed header drop */
13903 #ifdef NETFLIX_SB_LIMITS
13904 		appended =
13905 #endif
13906 			sbappendstream_locked(&so->so_rcv, m, 0);
13907 		ctf_calc_rwin(so, tp);
13908 	}
13909 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
13910 	/* NB: sorwakeup_locked() does an implicit unlock. */
13911 	sorwakeup_locked(so);
13912 #ifdef NETFLIX_SB_LIMITS
13913 	if (so->so_rcv.sb_shlim && mcnt != appended)
13914 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
13915 #endif
13916 	rack_handle_delayed_ack(tp, rack, tlen, 0);
13917 	if (tp->snd_una == tp->snd_max)
13918 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13919 	return (1);
13920 }
13921 
13922 /*
13923  * This subfunction is used to try to highly optimize the
13924  * fast path. We again allow window updates that are
13925  * in sequence to remain in the fast-path. We also add
13926  * in the __predict's to attempt to help the compiler.
13927  * Note that if we return a 0, then we can *not* process
13928  * it and the caller should push the packet into the
13929  * slow-path.
13930  */
13931 static int
13932 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13933     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13934     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
13935 {
13936 	int32_t acked;
13937 	int32_t nsegs;
13938 	int32_t under_pacing = 0;
13939 	struct tcp_rack *rack;
13940 
13941 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
13942 		/* Old ack, behind (or duplicate to) the last one rcv'd */
13943 		return (0);
13944 	}
13945 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
13946 		/* Above what we have sent? */
13947 		return (0);
13948 	}
13949 	if (__predict_false(tiwin == 0)) {
13950 		/* zero window */
13951 		return (0);
13952 	}
13953 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
13954 		/* We need a SYN or a FIN, unlikely.. */
13955 		return (0);
13956 	}
13957 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
13958 		/* Timestamp is behind .. old ack with seq wrap? */
13959 		return (0);
13960 	}
13961 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
13962 		/* Still recovering */
13963 		return (0);
13964 	}
13965 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13966 	if (rack->r_ctl.rc_sacked) {
13967 		/* We have sack holes on our scoreboard */
13968 		return (0);
13969 	}
13970 	/* Ok if we reach here, we can process a fast-ack */
13971 	if (rack->gp_ready &&
13972 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13973 		under_pacing = 1;
13974 	}
13975 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13976 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
13977 	/* Did the window get updated? */
13978 	if (tiwin != tp->snd_wnd) {
13979 		tp->snd_wnd = tiwin;
13980 		rack_validate_fo_sendwin_up(tp, rack);
13981 		tp->snd_wl1 = th->th_seq;
13982 		if (tp->snd_wnd > tp->max_sndwnd)
13983 			tp->max_sndwnd = tp->snd_wnd;
13984 	}
13985 	/* Do we exit persists? */
13986 	if ((rack->rc_in_persist != 0) &&
13987 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13988 			       rack->r_ctl.rc_pace_min_segs))) {
13989 		rack_exit_persist(tp, rack, cts);
13990 	}
13991 	/* Do we enter persists? */
13992 	if ((rack->rc_in_persist == 0) &&
13993 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13994 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13995 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13996 	    sbavail(&tptosocket(tp)->so_snd) &&
13997 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13998 		/*
13999 		 * Here the rwnd is less than
14000 		 * the pacing size, we are established,
14001 		 * nothing is outstanding, and there is
14002 		 * data to send. Enter persists.
14003 		 */
14004 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
14005 	}
14006 	/*
14007 	 * If last ACK falls within this segment's sequence numbers, record
14008 	 * the timestamp. NOTE that the test is modified according to the
14009 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
14010 	 */
14011 	if ((to->to_flags & TOF_TS) != 0 &&
14012 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
14013 		tp->ts_recent_age = tcp_ts_getticks();
14014 		tp->ts_recent = to->to_tsval;
14015 	}
14016 	/*
14017 	 * This is a pure ack for outstanding data.
14018 	 */
14019 	KMOD_TCPSTAT_INC(tcps_predack);
14020 
14021 	/*
14022 	 * "bad retransmit" recovery.
14023 	 */
14024 	if ((tp->t_flags & TF_PREVVALID) &&
14025 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
14026 		tp->t_flags &= ~TF_PREVVALID;
14027 		if (tp->t_rxtshift == 1 &&
14028 		    (int)(ticks - tp->t_badrxtwin) < 0)
14029 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
14030 	}
14031 	/*
14032 	 * Recalculate the transmit timer / rtt.
14033 	 *
14034 	 * Some boxes send broken timestamp replies during the SYN+ACK
14035 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
14036 	 * and blow up the retransmit timer.
14037 	 */
14038 	acked = BYTES_THIS_ACK(tp, th);
14039 
14040 #ifdef TCP_HHOOK
14041 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
14042 	hhook_run_tcp_est_in(tp, th, to);
14043 #endif
14044 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
14045 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
14046 	if (acked) {
14047 		struct mbuf *mfree;
14048 
14049 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
14050 		SOCKBUF_LOCK(&so->so_snd);
14051 		mfree = sbcut_locked(&so->so_snd, acked);
14052 		tp->snd_una = th->th_ack;
14053 		/* Note we want to hold the sb lock through the sendmap adjust */
14054 		rack_adjust_sendmap_head(rack, &so->so_snd);
14055 		/* Wake up the socket if we have room to write more */
14056 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
14057 		sowwakeup_locked(so);
14058 		m_freem(mfree);
14059 		tp->t_rxtshift = 0;
14060 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
14061 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
14062 		rack->rc_tlp_in_progress = 0;
14063 		rack->r_ctl.rc_tlp_cnt_out = 0;
14064 		/*
14065 		 * If it is the RXT timer we want to
14066 		 * stop it, so we can restart a TLP.
14067 		 */
14068 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
14069 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
14070 
14071 #ifdef TCP_REQUEST_TRK
14072 		rack_req_check_for_comp(rack, th->th_ack);
14073 #endif
14074 	}
14075 	/*
14076 	 * Let the congestion control algorithm update congestion control
14077 	 * related information. This typically means increasing the
14078 	 * congestion window.
14079 	 */
14080 	if (tp->snd_wnd < ctf_outstanding(tp)) {
14081 		/* The peer collapsed the window */
14082 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
14083 	} else if (rack->rc_has_collapsed)
14084 		rack_un_collapse_window(rack, __LINE__);
14085 	if ((rack->r_collapse_point_valid) &&
14086 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
14087 		rack->r_collapse_point_valid = 0;
14088 	/*
14089 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
14090 	 */
14091 	tp->snd_wl2 = th->th_ack;
14092 	tp->t_dupacks = 0;
14093 	m_freem(m);
14094 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
14095 
14096 	/*
14097 	 * If all outstanding data are acked, stop retransmit timer,
14098 	 * otherwise restart timer using current (possibly backed-off)
14099 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
14100 	 * If data are ready to send, let tcp_output decide between more
14101 	 * output or persist.
14102 	 */
14103 	if (under_pacing &&
14104 	    (rack->use_fixed_rate == 0) &&
14105 	    (rack->in_probe_rtt == 0) &&
14106 	    rack->rc_gp_dyn_mul &&
14107 	    rack->rc_always_pace) {
14108 		/* Check if we are dragging bottom */
14109 		rack_check_bottom_drag(tp, rack, so);
14110 	}
14111 	if (tp->snd_una == tp->snd_max) {
14112 		tp->t_flags &= ~TF_PREVVALID;
14113 		rack->r_ctl.retran_during_recovery = 0;
14114 		rack->rc_suspicious = 0;
14115 		rack->r_ctl.dsack_byte_cnt = 0;
14116 		rack->r_ctl.idle_snd_una = tp->snd_una;
14117 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
14118 		if (rack->r_ctl.rc_went_idle_time == 0)
14119 			rack->r_ctl.rc_went_idle_time = 1;
14120 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
14121 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
14122 			tp->t_acktime = 0;
14123 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
14124 	}
14125 	if (acked && rack->r_fast_output)
14126 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
14127 	if (sbavail(&so->so_snd)) {
14128 		rack->r_wanted_output = 1;
14129 	}
14130 	return (1);
14131 }
14132 
14133 /*
14134  * Return value of 1, the TCB is unlocked and most
14135  * likely gone, return value of 0, the TCP is still
14136  * locked.
14137  */
14138 static int
14139 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
14140     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14141     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14142 {
14143 	int32_t ret_val = 0;
14144 	int32_t orig_tlen = tlen;
14145 	int32_t todrop;
14146 	int32_t ourfinisacked = 0;
14147 	struct tcp_rack *rack;
14148 
14149 	INP_WLOCK_ASSERT(tptoinpcb(tp));
14150 
14151 	ctf_calc_rwin(so, tp);
14152 	/*
14153 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
14154 	 * SYN, drop the input. if seg contains a RST, then drop the
14155 	 * connection. if seg does not contain SYN, then drop it. Otherwise
14156 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
14157 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
14158 	 * contains an ECE and ECN support is enabled, the stream is ECN
14159 	 * capable. if SYN has been acked change to ESTABLISHED else
14160 	 * SYN_RCVD state arrange for segment to be acked (eventually)
14161 	 * continue processing rest of data/controls.
14162 	 */
14163 	if ((thflags & TH_ACK) &&
14164 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
14165 	    SEQ_GT(th->th_ack, tp->snd_max))) {
14166 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14167 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14168 		return (1);
14169 	}
14170 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
14171 		TCP_PROBE5(connect__refused, NULL, tp,
14172 		    mtod(m, const char *), tp, th);
14173 		tp = tcp_drop(tp, ECONNREFUSED);
14174 		ctf_do_drop(m, tp);
14175 		return (1);
14176 	}
14177 	if (thflags & TH_RST) {
14178 		ctf_do_drop(m, tp);
14179 		return (1);
14180 	}
14181 	if (!(thflags & TH_SYN)) {
14182 		ctf_do_drop(m, tp);
14183 		return (1);
14184 	}
14185 	tp->irs = th->th_seq;
14186 	tcp_rcvseqinit(tp);
14187 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14188 	if (thflags & TH_ACK) {
14189 		int tfo_partial = 0;
14190 
14191 		KMOD_TCPSTAT_INC(tcps_connects);
14192 		soisconnected(so);
14193 #ifdef MAC
14194 		mac_socketpeer_set_from_mbuf(m, so);
14195 #endif
14196 		/* Do window scaling on this connection? */
14197 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
14198 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
14199 			tp->rcv_scale = tp->request_r_scale;
14200 		}
14201 		tp->rcv_adv += min(tp->rcv_wnd,
14202 		    TCP_MAXWIN << tp->rcv_scale);
14203 		/*
14204 		 * If not all the data that was sent in the TFO SYN
14205 		 * has been acked, resend the remainder right away.
14206 		 */
14207 		if ((tp->t_flags & TF_FASTOPEN) &&
14208 		    (tp->snd_una != tp->snd_max)) {
14209 			/* Was it a partial ack? */
14210 			if (SEQ_LT(th->th_ack, tp->snd_max))
14211 				tfo_partial = 1;
14212 		}
14213 		/*
14214 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
14215 		 * will be turned on later.
14216 		 */
14217 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
14218 			rack_timer_cancel(tp, rack,
14219 					  rack->r_ctl.rc_rcvtime, __LINE__);
14220 			tp->t_flags |= TF_DELACK;
14221 		} else {
14222 			rack->r_wanted_output = 1;
14223 			tp->t_flags |= TF_ACKNOW;
14224 		}
14225 
14226 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
14227 
14228 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14229 			/*
14230 			 * We advance snd_una for the
14231 			 * fast open case. If th_ack is
14232 			 * acknowledging data beyond
14233 			 * snd_una we can't just call
14234 			 * ack-processing since the
14235 			 * data stream in our send-map
14236 			 * will start at snd_una + 1 (one
14237 			 * beyond the SYN). If its just
14238 			 * equal we don't need to do that
14239 			 * and there is no send_map.
14240 			 */
14241 			tp->snd_una++;
14242 			if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
14243 				/*
14244 				 * We sent a SYN with data, and thus have a
14245 				 * sendmap entry with a SYN set. Lets find it
14246 				 * and take off the send bit and the byte and
14247 				 * set it up to be what we send (send it next).
14248 				 */
14249 				struct rack_sendmap *rsm;
14250 
14251 				rsm = tqhash_min(rack->r_ctl.tqh);
14252 				if (rsm) {
14253 					if (rsm->r_flags & RACK_HAS_SYN) {
14254 						rsm->r_flags &= ~RACK_HAS_SYN;
14255 						rsm->r_start++;
14256 					}
14257 					rack->r_ctl.rc_resend = rsm;
14258 				}
14259 			}
14260 		}
14261 		/*
14262 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
14263 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
14264 		 */
14265 		tp->t_starttime = ticks;
14266 		if (tp->t_flags & TF_NEEDFIN) {
14267 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
14268 			tp->t_flags &= ~TF_NEEDFIN;
14269 			thflags &= ~TH_SYN;
14270 		} else {
14271 			tcp_state_change(tp, TCPS_ESTABLISHED);
14272 			TCP_PROBE5(connect__established, NULL, tp,
14273 			    mtod(m, const char *), tp, th);
14274 			rack_cc_conn_init(tp);
14275 		}
14276 	} else {
14277 		/*
14278 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
14279 		 * open.  If segment contains CC option and there is a
14280 		 * cached CC, apply TAO test. If it succeeds, connection is *
14281 		 * half-synchronized. Otherwise, do 3-way handshake:
14282 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
14283 		 * there was no CC option, clear cached CC value.
14284 		 */
14285 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
14286 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
14287 	}
14288 	/*
14289 	 * Advance th->th_seq to correspond to first data byte. If data,
14290 	 * trim to stay within window, dropping FIN if necessary.
14291 	 */
14292 	th->th_seq++;
14293 	if (tlen > tp->rcv_wnd) {
14294 		todrop = tlen - tp->rcv_wnd;
14295 		m_adj(m, -todrop);
14296 		tlen = tp->rcv_wnd;
14297 		thflags &= ~TH_FIN;
14298 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
14299 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
14300 	}
14301 	tp->snd_wl1 = th->th_seq - 1;
14302 	tp->rcv_up = th->th_seq;
14303 	/*
14304 	 * Client side of transaction: already sent SYN and data. If the
14305 	 * remote host used T/TCP to validate the SYN, our data will be
14306 	 * ACK'd; if so, enter normal data segment processing in the middle
14307 	 * of step 5, ack processing. Otherwise, goto step 6.
14308 	 */
14309 	if (thflags & TH_ACK) {
14310 		/* For syn-sent we need to possibly update the rtt */
14311 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
14312 			uint32_t t, mcts;
14313 
14314 			mcts = tcp_ts_getticks();
14315 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
14316 			if (!tp->t_rttlow || tp->t_rttlow > t)
14317 				tp->t_rttlow = t;
14318 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
14319 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
14320 			tcp_rack_xmit_timer_commit(rack, tp);
14321 		}
14322 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
14323 			return (ret_val);
14324 		/* We may have changed to FIN_WAIT_1 above */
14325 		if (tp->t_state == TCPS_FIN_WAIT_1) {
14326 			/*
14327 			 * In FIN_WAIT_1 STATE in addition to the processing
14328 			 * for the ESTABLISHED state if our FIN is now
14329 			 * acknowledged then enter FIN_WAIT_2.
14330 			 */
14331 			if (ourfinisacked) {
14332 				/*
14333 				 * If we can't receive any more data, then
14334 				 * closing user can proceed. Starting the
14335 				 * timer is contrary to the specification,
14336 				 * but if we don't get a FIN we'll hang
14337 				 * forever.
14338 				 *
14339 				 * XXXjl: we should release the tp also, and
14340 				 * use a compressed state.
14341 				 */
14342 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
14343 					soisdisconnected(so);
14344 					tcp_timer_activate(tp, TT_2MSL,
14345 					    (tcp_fast_finwait2_recycle ?
14346 					    tcp_finwait2_timeout :
14347 					    TP_MAXIDLE(tp)));
14348 				}
14349 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
14350 			}
14351 		}
14352 	}
14353 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14354 	   tiwin, thflags, nxt_pkt));
14355 }
14356 
14357 /*
14358  * Return value of 1, the TCB is unlocked and most
14359  * likely gone, return value of 0, the TCP is still
14360  * locked.
14361  */
14362 static int
14363 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
14364     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14365     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14366 {
14367 	struct tcp_rack *rack;
14368 	int32_t orig_tlen = tlen;
14369 	int32_t ret_val = 0;
14370 	int32_t ourfinisacked = 0;
14371 
14372 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14373 	ctf_calc_rwin(so, tp);
14374 	if ((thflags & TH_RST) ||
14375 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14376 		return (__ctf_process_rst(m, th, so, tp,
14377 					  &rack->r_ctl.challenge_ack_ts,
14378 					  &rack->r_ctl.challenge_ack_cnt));
14379 	if ((thflags & TH_ACK) &&
14380 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
14381 	    SEQ_GT(th->th_ack, tp->snd_max))) {
14382 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14383 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14384 		return (1);
14385 	}
14386 	if (tp->t_flags & TF_FASTOPEN) {
14387 		/*
14388 		 * When a TFO connection is in SYN_RECEIVED, the
14389 		 * only valid packets are the initial SYN, a
14390 		 * retransmit/copy of the initial SYN (possibly with
14391 		 * a subset of the original data), a valid ACK, a
14392 		 * FIN, or a RST.
14393 		 */
14394 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
14395 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14396 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14397 			return (1);
14398 		} else if (thflags & TH_SYN) {
14399 			/* non-initial SYN is ignored */
14400 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
14401 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
14402 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
14403 				ctf_do_drop(m, NULL);
14404 				return (0);
14405 			}
14406 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
14407 			ctf_do_drop(m, NULL);
14408 			return (0);
14409 		}
14410 	}
14411 
14412 	/*
14413 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14414 	 * it's less than ts_recent, drop it.
14415 	 */
14416 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14417 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14418 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14419 			return (ret_val);
14420 	}
14421 	/*
14422 	 * In the SYN-RECEIVED state, validate that the packet belongs to
14423 	 * this connection before trimming the data to fit the receive
14424 	 * window.  Check the sequence number versus IRS since we know the
14425 	 * sequence numbers haven't wrapped.  This is a partial fix for the
14426 	 * "LAND" DoS attack.
14427 	 */
14428 	if (SEQ_LT(th->th_seq, tp->irs)) {
14429 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14430 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14431 		return (1);
14432 	}
14433 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14434 			      &rack->r_ctl.challenge_ack_ts,
14435 			      &rack->r_ctl.challenge_ack_cnt)) {
14436 		return (ret_val);
14437 	}
14438 	/*
14439 	 * If last ACK falls within this segment's sequence numbers, record
14440 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14441 	 * from the latest proposal of the tcplw@cray.com list (Braden
14442 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14443 	 * with our earlier PAWS tests, so this check should be solely
14444 	 * predicated on the sequence space of this segment. 3) That we
14445 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14446 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14447 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14448 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14449 	 * p.869. In such cases, we can still calculate the RTT correctly
14450 	 * when RCV.NXT == Last.ACK.Sent.
14451 	 */
14452 	if ((to->to_flags & TOF_TS) != 0 &&
14453 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14454 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14455 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14456 		tp->ts_recent_age = tcp_ts_getticks();
14457 		tp->ts_recent = to->to_tsval;
14458 	}
14459 	tp->snd_wnd = tiwin;
14460 	rack_validate_fo_sendwin_up(tp, rack);
14461 	/*
14462 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14463 	 * is on (half-synchronized state), then queue data for later
14464 	 * processing; else drop segment and return.
14465 	 */
14466 	if ((thflags & TH_ACK) == 0) {
14467 		if (tp->t_flags & TF_FASTOPEN) {
14468 			rack_cc_conn_init(tp);
14469 		}
14470 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14471 		    tiwin, thflags, nxt_pkt));
14472 	}
14473 	KMOD_TCPSTAT_INC(tcps_connects);
14474 	if (tp->t_flags & TF_SONOTCONN) {
14475 		tp->t_flags &= ~TF_SONOTCONN;
14476 		soisconnected(so);
14477 	}
14478 	/* Do window scaling? */
14479 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
14480 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
14481 		tp->rcv_scale = tp->request_r_scale;
14482 	}
14483 	/*
14484 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
14485 	 * FIN-WAIT-1
14486 	 */
14487 	tp->t_starttime = ticks;
14488 	if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
14489 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
14490 		tp->t_tfo_pending = NULL;
14491 	}
14492 	if (tp->t_flags & TF_NEEDFIN) {
14493 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
14494 		tp->t_flags &= ~TF_NEEDFIN;
14495 	} else {
14496 		tcp_state_change(tp, TCPS_ESTABLISHED);
14497 		TCP_PROBE5(accept__established, NULL, tp,
14498 		    mtod(m, const char *), tp, th);
14499 		/*
14500 		 * TFO connections call cc_conn_init() during SYN
14501 		 * processing.  Calling it again here for such connections
14502 		 * is not harmless as it would undo the snd_cwnd reduction
14503 		 * that occurs when a TFO SYN|ACK is retransmitted.
14504 		 */
14505 		if (!(tp->t_flags & TF_FASTOPEN))
14506 			rack_cc_conn_init(tp);
14507 	}
14508 	/*
14509 	 * Account for the ACK of our SYN prior to
14510 	 * regular ACK processing below, except for
14511 	 * simultaneous SYN, which is handled later.
14512 	 */
14513 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
14514 		tp->snd_una++;
14515 	/*
14516 	 * If segment contains data or ACK, will call tcp_reass() later; if
14517 	 * not, do so now to pass queued data to user.
14518 	 */
14519 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
14520 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
14521 		    (struct mbuf *)0);
14522 		if (tp->t_flags & TF_WAKESOR) {
14523 			tp->t_flags &= ~TF_WAKESOR;
14524 			/* NB: sorwakeup_locked() does an implicit unlock. */
14525 			sorwakeup_locked(so);
14526 		}
14527 	}
14528 	tp->snd_wl1 = th->th_seq - 1;
14529 	/* For syn-recv we need to possibly update the rtt */
14530 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
14531 		uint32_t t, mcts;
14532 
14533 		mcts = tcp_ts_getticks();
14534 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
14535 		if (!tp->t_rttlow || tp->t_rttlow > t)
14536 			tp->t_rttlow = t;
14537 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
14538 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
14539 		tcp_rack_xmit_timer_commit(rack, tp);
14540 	}
14541 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
14542 		return (ret_val);
14543 	}
14544 	if (tp->t_state == TCPS_FIN_WAIT_1) {
14545 		/* We could have went to FIN_WAIT_1 (or EST) above */
14546 		/*
14547 		 * In FIN_WAIT_1 STATE in addition to the processing for the
14548 		 * ESTABLISHED state if our FIN is now acknowledged then
14549 		 * enter FIN_WAIT_2.
14550 		 */
14551 		if (ourfinisacked) {
14552 			/*
14553 			 * If we can't receive any more data, then closing
14554 			 * user can proceed. Starting the timer is contrary
14555 			 * to the specification, but if we don't get a FIN
14556 			 * we'll hang forever.
14557 			 *
14558 			 * XXXjl: we should release the tp also, and use a
14559 			 * compressed state.
14560 			 */
14561 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
14562 				soisdisconnected(so);
14563 				tcp_timer_activate(tp, TT_2MSL,
14564 				    (tcp_fast_finwait2_recycle ?
14565 				    tcp_finwait2_timeout :
14566 				    TP_MAXIDLE(tp)));
14567 			}
14568 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
14569 		}
14570 	}
14571 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14572 	    tiwin, thflags, nxt_pkt));
14573 }
14574 
14575 /*
14576  * Return value of 1, the TCB is unlocked and most
14577  * likely gone, return value of 0, the TCP is still
14578  * locked.
14579  */
14580 static int
14581 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
14582     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14583     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14584 {
14585 	int32_t ret_val = 0;
14586 	int32_t orig_tlen = tlen;
14587 	struct tcp_rack *rack;
14588 
14589 	/*
14590 	 * Header prediction: check for the two common cases of a
14591 	 * uni-directional data xfer.  If the packet has no control flags,
14592 	 * is in-sequence, the window didn't change and we're not
14593 	 * retransmitting, it's a candidate.  If the length is zero and the
14594 	 * ack moved forward, we're the sender side of the xfer.  Just free
14595 	 * the data acked & wake any higher level process that was blocked
14596 	 * waiting for space.  If the length is non-zero and the ack didn't
14597 	 * move, we're the receiver side.  If we're getting packets in-order
14598 	 * (the reassembly queue is empty), add the data toc The socket
14599 	 * buffer and note that we need a delayed ack. Make sure that the
14600 	 * hidden state-flags are also off. Since we check for
14601 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
14602 	 */
14603 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14604 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
14605 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
14606 	    __predict_true(SEGQ_EMPTY(tp)) &&
14607 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
14608 		if (tlen == 0) {
14609 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
14610 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
14611 				return (0);
14612 			}
14613 		} else {
14614 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
14615 			    tiwin, nxt_pkt, iptos)) {
14616 				return (0);
14617 			}
14618 		}
14619 	}
14620 	ctf_calc_rwin(so, tp);
14621 
14622 	if ((thflags & TH_RST) ||
14623 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14624 		return (__ctf_process_rst(m, th, so, tp,
14625 					  &rack->r_ctl.challenge_ack_ts,
14626 					  &rack->r_ctl.challenge_ack_cnt));
14627 
14628 	/*
14629 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14630 	 * synchronized state.
14631 	 */
14632 	if (thflags & TH_SYN) {
14633 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14634 		return (ret_val);
14635 	}
14636 	/*
14637 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14638 	 * it's less than ts_recent, drop it.
14639 	 */
14640 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14641 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14642 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14643 			return (ret_val);
14644 	}
14645 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14646 			      &rack->r_ctl.challenge_ack_ts,
14647 			      &rack->r_ctl.challenge_ack_cnt)) {
14648 		return (ret_val);
14649 	}
14650 	/*
14651 	 * If last ACK falls within this segment's sequence numbers, record
14652 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14653 	 * from the latest proposal of the tcplw@cray.com list (Braden
14654 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14655 	 * with our earlier PAWS tests, so this check should be solely
14656 	 * predicated on the sequence space of this segment. 3) That we
14657 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14658 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14659 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14660 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14661 	 * p.869. In such cases, we can still calculate the RTT correctly
14662 	 * when RCV.NXT == Last.ACK.Sent.
14663 	 */
14664 	if ((to->to_flags & TOF_TS) != 0 &&
14665 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14666 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14667 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14668 		tp->ts_recent_age = tcp_ts_getticks();
14669 		tp->ts_recent = to->to_tsval;
14670 	}
14671 	/*
14672 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14673 	 * is on (half-synchronized state), then queue data for later
14674 	 * processing; else drop segment and return.
14675 	 */
14676 	if ((thflags & TH_ACK) == 0) {
14677 		if (tp->t_flags & TF_NEEDSYN) {
14678 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14679 			    tiwin, thflags, nxt_pkt));
14680 
14681 		} else if (tp->t_flags & TF_ACKNOW) {
14682 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14683 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14684 			return (ret_val);
14685 		} else {
14686 			ctf_do_drop(m, NULL);
14687 			return (0);
14688 		}
14689 	}
14690 	/*
14691 	 * Ack processing.
14692 	 */
14693 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
14694 		return (ret_val);
14695 	}
14696 	if (sbavail(&so->so_snd)) {
14697 		if (ctf_progress_timeout_check(tp, true)) {
14698 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
14699 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14700 			return (1);
14701 		}
14702 	}
14703 	/* State changes only happen in rack_process_data() */
14704 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14705 	    tiwin, thflags, nxt_pkt));
14706 }
14707 
14708 /*
14709  * Return value of 1, the TCB is unlocked and most
14710  * likely gone, return value of 0, the TCP is still
14711  * locked.
14712  */
14713 static int
14714 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
14715     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14716     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14717 {
14718 	int32_t ret_val = 0;
14719 	int32_t orig_tlen = tlen;
14720 	struct tcp_rack *rack;
14721 
14722 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14723 	ctf_calc_rwin(so, tp);
14724 	if ((thflags & TH_RST) ||
14725 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14726 		return (__ctf_process_rst(m, th, so, tp,
14727 					  &rack->r_ctl.challenge_ack_ts,
14728 					  &rack->r_ctl.challenge_ack_cnt));
14729 	/*
14730 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14731 	 * synchronized state.
14732 	 */
14733 	if (thflags & TH_SYN) {
14734 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14735 		return (ret_val);
14736 	}
14737 	/*
14738 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14739 	 * it's less than ts_recent, drop it.
14740 	 */
14741 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14742 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14743 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14744 			return (ret_val);
14745 	}
14746 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14747 			      &rack->r_ctl.challenge_ack_ts,
14748 			      &rack->r_ctl.challenge_ack_cnt)) {
14749 		return (ret_val);
14750 	}
14751 	/*
14752 	 * If last ACK falls within this segment's sequence numbers, record
14753 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14754 	 * from the latest proposal of the tcplw@cray.com list (Braden
14755 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14756 	 * with our earlier PAWS tests, so this check should be solely
14757 	 * predicated on the sequence space of this segment. 3) That we
14758 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14759 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14760 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14761 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14762 	 * p.869. In such cases, we can still calculate the RTT correctly
14763 	 * when RCV.NXT == Last.ACK.Sent.
14764 	 */
14765 	if ((to->to_flags & TOF_TS) != 0 &&
14766 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14767 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14768 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14769 		tp->ts_recent_age = tcp_ts_getticks();
14770 		tp->ts_recent = to->to_tsval;
14771 	}
14772 	/*
14773 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14774 	 * is on (half-synchronized state), then queue data for later
14775 	 * processing; else drop segment and return.
14776 	 */
14777 	if ((thflags & TH_ACK) == 0) {
14778 		if (tp->t_flags & TF_NEEDSYN) {
14779 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14780 			    tiwin, thflags, nxt_pkt));
14781 
14782 		} else if (tp->t_flags & TF_ACKNOW) {
14783 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14784 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14785 			return (ret_val);
14786 		} else {
14787 			ctf_do_drop(m, NULL);
14788 			return (0);
14789 		}
14790 	}
14791 	/*
14792 	 * Ack processing.
14793 	 */
14794 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
14795 		return (ret_val);
14796 	}
14797 	if (sbavail(&so->so_snd)) {
14798 		if (ctf_progress_timeout_check(tp, true)) {
14799 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14800 						tp, tick, PROGRESS_DROP, __LINE__);
14801 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14802 			return (1);
14803 		}
14804 	}
14805 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14806 	    tiwin, thflags, nxt_pkt));
14807 }
14808 
14809 static int
14810 rack_check_data_after_close(struct mbuf *m,
14811     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
14812 {
14813 	struct tcp_rack *rack;
14814 
14815 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14816 	if (rack->rc_allow_data_af_clo == 0) {
14817 	close_now:
14818 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
14819 		/* tcp_close will kill the inp pre-log the Reset */
14820 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14821 		tp = tcp_close(tp);
14822 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
14823 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
14824 		return (1);
14825 	}
14826 	if (sbavail(&so->so_snd) == 0)
14827 		goto close_now;
14828 	/* Ok we allow data that is ignored and a followup reset */
14829 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
14830 	tp->rcv_nxt = th->th_seq + *tlen;
14831 	tp->t_flags2 |= TF2_DROP_AF_DATA;
14832 	rack->r_wanted_output = 1;
14833 	*tlen = 0;
14834 	return (0);
14835 }
14836 
14837 /*
14838  * Return value of 1, the TCB is unlocked and most
14839  * likely gone, return value of 0, the TCP is still
14840  * locked.
14841  */
14842 static int
14843 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
14844     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14845     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14846 {
14847 	int32_t ret_val = 0;
14848 	int32_t orig_tlen = tlen;
14849 	int32_t ourfinisacked = 0;
14850 	struct tcp_rack *rack;
14851 
14852 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14853 	ctf_calc_rwin(so, tp);
14854 
14855 	if ((thflags & TH_RST) ||
14856 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14857 		return (__ctf_process_rst(m, th, so, tp,
14858 					  &rack->r_ctl.challenge_ack_ts,
14859 					  &rack->r_ctl.challenge_ack_cnt));
14860 	/*
14861 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14862 	 * synchronized state.
14863 	 */
14864 	if (thflags & TH_SYN) {
14865 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14866 		return (ret_val);
14867 	}
14868 	/*
14869 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14870 	 * it's less than ts_recent, drop it.
14871 	 */
14872 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14873 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14874 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14875 			return (ret_val);
14876 	}
14877 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14878 			      &rack->r_ctl.challenge_ack_ts,
14879 			      &rack->r_ctl.challenge_ack_cnt)) {
14880 		return (ret_val);
14881 	}
14882 	/*
14883 	 * If new data are received on a connection after the user processes
14884 	 * are gone, then RST the other end.
14885 	 */
14886 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14887 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14888 		return (1);
14889 	/*
14890 	 * If last ACK falls within this segment's sequence numbers, record
14891 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14892 	 * from the latest proposal of the tcplw@cray.com list (Braden
14893 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14894 	 * with our earlier PAWS tests, so this check should be solely
14895 	 * predicated on the sequence space of this segment. 3) That we
14896 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14897 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14898 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14899 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14900 	 * p.869. In such cases, we can still calculate the RTT correctly
14901 	 * when RCV.NXT == Last.ACK.Sent.
14902 	 */
14903 	if ((to->to_flags & TOF_TS) != 0 &&
14904 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14905 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14906 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14907 		tp->ts_recent_age = tcp_ts_getticks();
14908 		tp->ts_recent = to->to_tsval;
14909 	}
14910 	/*
14911 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14912 	 * is on (half-synchronized state), then queue data for later
14913 	 * processing; else drop segment and return.
14914 	 */
14915 	if ((thflags & TH_ACK) == 0) {
14916 		if (tp->t_flags & TF_NEEDSYN) {
14917 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14918 			    tiwin, thflags, nxt_pkt));
14919 		} else if (tp->t_flags & TF_ACKNOW) {
14920 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14921 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14922 			return (ret_val);
14923 		} else {
14924 			ctf_do_drop(m, NULL);
14925 			return (0);
14926 		}
14927 	}
14928 	/*
14929 	 * Ack processing.
14930 	 */
14931 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
14932 		return (ret_val);
14933 	}
14934 	if (ourfinisacked) {
14935 		/*
14936 		 * If we can't receive any more data, then closing user can
14937 		 * proceed. Starting the timer is contrary to the
14938 		 * specification, but if we don't get a FIN we'll hang
14939 		 * forever.
14940 		 *
14941 		 * XXXjl: we should release the tp also, and use a
14942 		 * compressed state.
14943 		 */
14944 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
14945 			soisdisconnected(so);
14946 			tcp_timer_activate(tp, TT_2MSL,
14947 			    (tcp_fast_finwait2_recycle ?
14948 			    tcp_finwait2_timeout :
14949 			    TP_MAXIDLE(tp)));
14950 		}
14951 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
14952 	}
14953 	if (sbavail(&so->so_snd)) {
14954 		if (ctf_progress_timeout_check(tp, true)) {
14955 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14956 						tp, tick, PROGRESS_DROP, __LINE__);
14957 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14958 			return (1);
14959 		}
14960 	}
14961 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14962 	    tiwin, thflags, nxt_pkt));
14963 }
14964 
14965 /*
14966  * Return value of 1, the TCB is unlocked and most
14967  * likely gone, return value of 0, the TCP is still
14968  * locked.
14969  */
14970 static int
14971 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
14972     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14973     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14974 {
14975 	int32_t ret_val = 0;
14976 	int32_t orig_tlen = tlen;
14977 	int32_t ourfinisacked = 0;
14978 	struct tcp_rack *rack;
14979 
14980 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14981 	ctf_calc_rwin(so, tp);
14982 
14983 	if ((thflags & TH_RST) ||
14984 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14985 		return (__ctf_process_rst(m, th, so, tp,
14986 					  &rack->r_ctl.challenge_ack_ts,
14987 					  &rack->r_ctl.challenge_ack_cnt));
14988 	/*
14989 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14990 	 * synchronized state.
14991 	 */
14992 	if (thflags & TH_SYN) {
14993 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14994 		return (ret_val);
14995 	}
14996 	/*
14997 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14998 	 * it's less than ts_recent, drop it.
14999 	 */
15000 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
15001 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
15002 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
15003 			return (ret_val);
15004 	}
15005 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
15006 			      &rack->r_ctl.challenge_ack_ts,
15007 			      &rack->r_ctl.challenge_ack_cnt)) {
15008 		return (ret_val);
15009 	}
15010 	/*
15011 	 * If last ACK falls within this segment's sequence numbers, record
15012 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
15013 	 * from the latest proposal of the tcplw@cray.com list (Braden
15014 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
15015 	 * with our earlier PAWS tests, so this check should be solely
15016 	 * predicated on the sequence space of this segment. 3) That we
15017 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
15018 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
15019 	 * SEG.Len, This modified check allows us to overcome RFC1323's
15020 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
15021 	 * p.869. In such cases, we can still calculate the RTT correctly
15022 	 * when RCV.NXT == Last.ACK.Sent.
15023 	 */
15024 	if ((to->to_flags & TOF_TS) != 0 &&
15025 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
15026 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
15027 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
15028 		tp->ts_recent_age = tcp_ts_getticks();
15029 		tp->ts_recent = to->to_tsval;
15030 	}
15031 	/*
15032 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
15033 	 * is on (half-synchronized state), then queue data for later
15034 	 * processing; else drop segment and return.
15035 	 */
15036 	if ((thflags & TH_ACK) == 0) {
15037 		if (tp->t_flags & TF_NEEDSYN) {
15038 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
15039 			    tiwin, thflags, nxt_pkt));
15040 		} else if (tp->t_flags & TF_ACKNOW) {
15041 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
15042 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
15043 			return (ret_val);
15044 		} else {
15045 			ctf_do_drop(m, NULL);
15046 			return (0);
15047 		}
15048 	}
15049 	/*
15050 	 * Ack processing.
15051 	 */
15052 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
15053 		return (ret_val);
15054 	}
15055 	if (ourfinisacked) {
15056 		tcp_twstart(tp);
15057 		m_freem(m);
15058 		return (1);
15059 	}
15060 	if (sbavail(&so->so_snd)) {
15061 		if (ctf_progress_timeout_check(tp, true)) {
15062 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
15063 						tp, tick, PROGRESS_DROP, __LINE__);
15064 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
15065 			return (1);
15066 		}
15067 	}
15068 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
15069 	    tiwin, thflags, nxt_pkt));
15070 }
15071 
15072 /*
15073  * Return value of 1, the TCB is unlocked and most
15074  * likely gone, return value of 0, the TCP is still
15075  * locked.
15076  */
15077 static int
15078 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
15079     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
15080     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
15081 {
15082 	int32_t ret_val = 0;
15083 	int32_t orig_tlen;
15084 	int32_t ourfinisacked = 0;
15085 	struct tcp_rack *rack;
15086 
15087 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15088 	ctf_calc_rwin(so, tp);
15089 
15090 	if ((thflags & TH_RST) ||
15091 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
15092 		return (__ctf_process_rst(m, th, so, tp,
15093 					  &rack->r_ctl.challenge_ack_ts,
15094 					  &rack->r_ctl.challenge_ack_cnt));
15095 	/*
15096 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
15097 	 * synchronized state.
15098 	 */
15099 	if (thflags & TH_SYN) {
15100 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
15101 		return (ret_val);
15102 	}
15103 	/*
15104 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
15105 	 * it's less than ts_recent, drop it.
15106 	 */
15107 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
15108 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
15109 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
15110 			return (ret_val);
15111 	}
15112 	orig_tlen = tlen;
15113 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
15114 			      &rack->r_ctl.challenge_ack_ts,
15115 			      &rack->r_ctl.challenge_ack_cnt)) {
15116 		return (ret_val);
15117 	}
15118 	/*
15119 	 * If last ACK falls within this segment's sequence numbers, record
15120 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
15121 	 * from the latest proposal of the tcplw@cray.com list (Braden
15122 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
15123 	 * with our earlier PAWS tests, so this check should be solely
15124 	 * predicated on the sequence space of this segment. 3) That we
15125 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
15126 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
15127 	 * SEG.Len, This modified check allows us to overcome RFC1323's
15128 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
15129 	 * p.869. In such cases, we can still calculate the RTT correctly
15130 	 * when RCV.NXT == Last.ACK.Sent.
15131 	 */
15132 	if ((to->to_flags & TOF_TS) != 0 &&
15133 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
15134 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
15135 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
15136 		tp->ts_recent_age = tcp_ts_getticks();
15137 		tp->ts_recent = to->to_tsval;
15138 	}
15139 	/*
15140 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
15141 	 * is on (half-synchronized state), then queue data for later
15142 	 * processing; else drop segment and return.
15143 	 */
15144 	if ((thflags & TH_ACK) == 0) {
15145 		if (tp->t_flags & TF_NEEDSYN) {
15146 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
15147 			    tiwin, thflags, nxt_pkt));
15148 		} else if (tp->t_flags & TF_ACKNOW) {
15149 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
15150 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
15151 			return (ret_val);
15152 		} else {
15153 			ctf_do_drop(m, NULL);
15154 			return (0);
15155 		}
15156 	}
15157 	/*
15158 	 * case TCPS_LAST_ACK: Ack processing.
15159 	 */
15160 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
15161 		return (ret_val);
15162 	}
15163 	if (ourfinisacked) {
15164 		tp = tcp_close(tp);
15165 		ctf_do_drop(m, tp);
15166 		return (1);
15167 	}
15168 	if (sbavail(&so->so_snd)) {
15169 		if (ctf_progress_timeout_check(tp, true)) {
15170 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
15171 						tp, tick, PROGRESS_DROP, __LINE__);
15172 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
15173 			return (1);
15174 		}
15175 	}
15176 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
15177 	    tiwin, thflags, nxt_pkt));
15178 }
15179 
15180 /*
15181  * Return value of 1, the TCB is unlocked and most
15182  * likely gone, return value of 0, the TCP is still
15183  * locked.
15184  */
15185 static int
15186 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
15187     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
15188     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
15189 {
15190 	int32_t ret_val = 0;
15191 	int32_t orig_tlen = tlen;
15192 	int32_t ourfinisacked = 0;
15193 	struct tcp_rack *rack;
15194 
15195 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15196 	ctf_calc_rwin(so, tp);
15197 
15198 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
15199 	if ((thflags & TH_RST) ||
15200 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
15201 		return (__ctf_process_rst(m, th, so, tp,
15202 					  &rack->r_ctl.challenge_ack_ts,
15203 					  &rack->r_ctl.challenge_ack_cnt));
15204 	/*
15205 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
15206 	 * synchronized state.
15207 	 */
15208 	if (thflags & TH_SYN) {
15209 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
15210 		return (ret_val);
15211 	}
15212 	/*
15213 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
15214 	 * it's less than ts_recent, drop it.
15215 	 */
15216 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
15217 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
15218 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
15219 			return (ret_val);
15220 	}
15221 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
15222 			      &rack->r_ctl.challenge_ack_ts,
15223 			      &rack->r_ctl.challenge_ack_cnt)) {
15224 		return (ret_val);
15225 	}
15226 	/*
15227 	 * If new data are received on a connection after the user processes
15228 	 * are gone, then RST the other end.
15229 	 */
15230 	if ((tp->t_flags & TF_CLOSED) && tlen &&
15231 	    rack_check_data_after_close(m, tp, &tlen, th, so))
15232 		return (1);
15233 	/*
15234 	 * If last ACK falls within this segment's sequence numbers, record
15235 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
15236 	 * from the latest proposal of the tcplw@cray.com list (Braden
15237 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
15238 	 * with our earlier PAWS tests, so this check should be solely
15239 	 * predicated on the sequence space of this segment. 3) That we
15240 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
15241 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
15242 	 * SEG.Len, This modified check allows us to overcome RFC1323's
15243 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
15244 	 * p.869. In such cases, we can still calculate the RTT correctly
15245 	 * when RCV.NXT == Last.ACK.Sent.
15246 	 */
15247 	if ((to->to_flags & TOF_TS) != 0 &&
15248 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
15249 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
15250 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
15251 		tp->ts_recent_age = tcp_ts_getticks();
15252 		tp->ts_recent = to->to_tsval;
15253 	}
15254 	/*
15255 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
15256 	 * is on (half-synchronized state), then queue data for later
15257 	 * processing; else drop segment and return.
15258 	 */
15259 	if ((thflags & TH_ACK) == 0) {
15260 		if (tp->t_flags & TF_NEEDSYN) {
15261 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
15262 			    tiwin, thflags, nxt_pkt));
15263 		} else if (tp->t_flags & TF_ACKNOW) {
15264 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
15265 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
15266 			return (ret_val);
15267 		} else {
15268 			ctf_do_drop(m, NULL);
15269 			return (0);
15270 		}
15271 	}
15272 	/*
15273 	 * Ack processing.
15274 	 */
15275 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
15276 		return (ret_val);
15277 	}
15278 	if (sbavail(&so->so_snd)) {
15279 		if (ctf_progress_timeout_check(tp, true)) {
15280 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
15281 						tp, tick, PROGRESS_DROP, __LINE__);
15282 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
15283 			return (1);
15284 		}
15285 	}
15286 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
15287 	    tiwin, thflags, nxt_pkt));
15288 }
15289 
15290 static void inline
15291 rack_clear_rate_sample(struct tcp_rack *rack)
15292 {
15293 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
15294 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
15295 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
15296 }
15297 
15298 static void
15299 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
15300 {
15301 	uint64_t bw_est, rate_wanted;
15302 	int chged = 0;
15303 	uint32_t user_max, orig_min, orig_max;
15304 
15305 #ifdef TCP_REQUEST_TRK
15306 	if (rack->rc_hybrid_mode &&
15307 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
15308 	    (rack_hybrid_allow_set_maxseg == 1) &&
15309 	    (rack->r_ctl.rc_last_sft != NULL)) {
15310 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
15311 		return;
15312 	}
15313 #endif
15314 	orig_min = rack->r_ctl.rc_pace_min_segs;
15315 	orig_max = rack->r_ctl.rc_pace_max_segs;
15316 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
15317 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
15318 		chged = 1;
15319 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
15320 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
15321 		if (user_max != rack->r_ctl.rc_pace_max_segs)
15322 			chged = 1;
15323 	}
15324 	if (rack->rc_force_max_seg) {
15325 		rack->r_ctl.rc_pace_max_segs = user_max;
15326 	} else if (rack->use_fixed_rate) {
15327 		bw_est = rack_get_bw(rack);
15328 		if ((rack->r_ctl.crte == NULL) ||
15329 		    (bw_est != rack->r_ctl.crte->rate)) {
15330 			rack->r_ctl.rc_pace_max_segs = user_max;
15331 		} else {
15332 			/* We are pacing right at the hardware rate */
15333 			uint32_t segsiz, pace_one;
15334 
15335 			if (rack_pace_one_seg ||
15336 			    (rack->r_ctl.rc_user_set_min_segs == 1))
15337 				pace_one = 1;
15338 			else
15339 				pace_one = 0;
15340 			segsiz = min(ctf_fixed_maxseg(tp),
15341 				     rack->r_ctl.rc_pace_min_segs);
15342 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
15343 				tp, bw_est, segsiz, pace_one,
15344 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
15345 		}
15346 	} else if (rack->rc_always_pace) {
15347 		if (rack->r_ctl.gp_bw ||
15348 		    rack->r_ctl.init_rate) {
15349 			/* We have a rate of some sort set */
15350 			uint32_t  orig;
15351 
15352 			bw_est = rack_get_bw(rack);
15353 			orig = rack->r_ctl.rc_pace_max_segs;
15354 			if (fill_override)
15355 				rate_wanted = *fill_override;
15356 			else
15357 				rate_wanted = rack_get_gp_est(rack);
15358 			if (rate_wanted) {
15359 				/* We have something */
15360 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
15361 										   rate_wanted,
15362 										   ctf_fixed_maxseg(rack->rc_tp));
15363 			} else
15364 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
15365 			if (orig != rack->r_ctl.rc_pace_max_segs)
15366 				chged = 1;
15367 		} else if ((rack->r_ctl.gp_bw == 0) &&
15368 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
15369 			/*
15370 			 * If we have nothing limit us to bursting
15371 			 * out IW sized pieces.
15372 			 */
15373 			chged = 1;
15374 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
15375 		}
15376 	}
15377 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
15378 		chged = 1;
15379 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
15380 	}
15381 	if (chged)
15382 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
15383 }
15384 
15385 
15386 static void
15387 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
15388 {
15389 #ifdef INET6
15390 	struct ip6_hdr *ip6 = NULL;
15391 #endif
15392 #ifdef INET
15393 	struct ip *ip = NULL;
15394 #endif
15395 	struct udphdr *udp = NULL;
15396 
15397 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
15398 #ifdef INET6
15399 	if (rack->r_is_v6) {
15400 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15401 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15402 		if (tp->t_port) {
15403 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
15404 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
15405 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
15406 			udp->uh_dport = tp->t_port;
15407 			rack->r_ctl.fsb.udp = udp;
15408 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
15409 		} else
15410 		{
15411 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
15412 			rack->r_ctl.fsb.udp = NULL;
15413 		}
15414 		tcpip_fillheaders(rack->rc_inp,
15415 				  tp->t_port,
15416 				  ip6, rack->r_ctl.fsb.th);
15417 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
15418 	} else
15419 #endif				/* INET6 */
15420 #ifdef INET
15421 	{
15422 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
15423 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15424 		if (tp->t_port) {
15425 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
15426 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
15427 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
15428 			udp->uh_dport = tp->t_port;
15429 			rack->r_ctl.fsb.udp = udp;
15430 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
15431 		} else
15432 		{
15433 			rack->r_ctl.fsb.udp = NULL;
15434 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
15435 		}
15436 		tcpip_fillheaders(rack->rc_inp,
15437 				  tp->t_port,
15438 				  ip, rack->r_ctl.fsb.th);
15439 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
15440 	}
15441 #endif
15442 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
15443 	    (long)TCP_MAXWIN << tp->rcv_scale);
15444 	rack->r_fsb_inited = 1;
15445 }
15446 
15447 static int
15448 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
15449 {
15450 	/*
15451 	 * Allocate the larger of spaces V6 if available else just
15452 	 * V4 and include udphdr (overbook)
15453 	 */
15454 #ifdef INET6
15455 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
15456 #else
15457 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
15458 #endif
15459 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
15460 					    M_TCPFSB, M_NOWAIT|M_ZERO);
15461 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
15462 		return (ENOMEM);
15463 	}
15464 	rack->r_fsb_inited = 0;
15465 	return (0);
15466 }
15467 
15468 static void
15469 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
15470 {
15471 	/*
15472 	 * Types of logs (mod value)
15473 	 * 20 - Initial round setup
15474 	 * 21 - Rack declares a new round.
15475 	 */
15476 	struct tcpcb *tp;
15477 
15478 	tp = rack->rc_tp;
15479 	if (tcp_bblogging_on(tp)) {
15480 		union tcp_log_stackspecific log;
15481 		struct timeval tv;
15482 
15483 		memset(&log, 0, sizeof(log));
15484 		log.u_bbr.flex1 = rack->r_ctl.current_round;
15485 		log.u_bbr.flex2 = rack->r_ctl.roundends;
15486 		log.u_bbr.flex3 = high_seq;
15487 		log.u_bbr.flex4 = tp->snd_max;
15488 		log.u_bbr.flex8 = mod;
15489 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15490 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
15491 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
15492 		TCP_LOG_EVENTP(tp, NULL,
15493 		    &tptosocket(tp)->so_rcv,
15494 		    &tptosocket(tp)->so_snd,
15495 		    TCP_HYSTART, 0,
15496 		    0, &log, false, &tv);
15497 	}
15498 }
15499 
15500 static void
15501 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
15502 {
15503 	rack->rack_deferred_inited = 1;
15504 	rack->r_ctl.roundends = tp->snd_max;
15505 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
15506 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
15507 }
15508 
15509 static void
15510 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
15511 {
15512 	/* Retransmit bit controls.
15513 	 *
15514 	 * The setting of these values control one of
15515 	 * three settings you can have and dictate
15516 	 * how rack does retransmissions. Note this
15517 	 * is in *any* mode i.e. pacing on or off DGP
15518 	 * fixed rate pacing, or just bursting rack.
15519 	 *
15520 	 * 1 - Use full sized retransmits i.e. limit
15521 	 *     the size to whatever the pace_max_segments
15522 	 *     size is.
15523 	 *
15524 	 * 2 - Use pacer min granularity as a guide to
15525 	 *     the size combined with the current calculated
15526 	 *     goodput b/w measurement. So for example if
15527 	 *     the goodput is measured at 20Mbps we would
15528 	 *     calculate 8125 (pacer minimum 250usec in
15529 	 *     that b/w) and then round it up to the next
15530 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
15531 	 *
15532 	 * 0 - The rack default 1 MSS (anything not 0/1/2
15533 	 *     fall here too if we are setting via rack_init()).
15534 	 *
15535 	 */
15536 	if (ctl == 1) {
15537 		rack->full_size_rxt = 1;
15538 		rack->shape_rxt_to_pacing_min  = 0;
15539 	} else if (ctl == 2) {
15540 		rack->full_size_rxt = 0;
15541 		rack->shape_rxt_to_pacing_min  = 1;
15542 	} else {
15543 		rack->full_size_rxt = 0;
15544 		rack->shape_rxt_to_pacing_min  = 0;
15545 	}
15546 }
15547 
15548 static void
15549 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
15550 		  uint32_t flex1,
15551 		  uint32_t flex2,
15552 		  uint32_t flex3)
15553 {
15554 	if (tcp_bblogging_on(rack->rc_tp)) {
15555 		union tcp_log_stackspecific log;
15556 		struct timeval tv;
15557 
15558 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15559 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15560 		log.u_bbr.flex8 = mod;
15561 		log.u_bbr.flex1 = flex1;
15562 		log.u_bbr.flex2 = flex2;
15563 		log.u_bbr.flex3 = flex3;
15564 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
15565 			       0, &log, false, NULL, __func__, __LINE__, &tv);
15566 	}
15567 }
15568 
15569 static int
15570 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
15571 {
15572 	struct tcp_rack *rack;
15573 	struct rack_sendmap *rsm;
15574 	int i;
15575 
15576 
15577 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15578 	switch (reqr->req) {
15579 	case TCP_QUERY_SENDMAP:
15580 		if ((reqr->req_param == tp->snd_max) ||
15581 		    (tp->snd_max == tp->snd_una)){
15582 			/* Unlikely */
15583 			return (0);
15584 		}
15585 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
15586 		if (rsm == NULL) {
15587 			/* Can't find that seq -- unlikely */
15588 			return (0);
15589 		}
15590 		reqr->sendmap_start = rsm->r_start;
15591 		reqr->sendmap_end = rsm->r_end;
15592 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
15593 		reqr->sendmap_fas = rsm->r_fas;
15594 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
15595 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
15596 		for(i=0; i<reqr->sendmap_send_cnt; i++)
15597 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
15598 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
15599 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
15600 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
15601 		reqr->sendmap_dupacks = rsm->r_dupack;
15602 		rack_log_chg_info(tp, rack, 1,
15603 				  rsm->r_start,
15604 				  rsm->r_end,
15605 				  rsm->r_flags);
15606 		return(1);
15607 		break;
15608 	case TCP_QUERY_TIMERS_UP:
15609 		if (rack->r_ctl.rc_hpts_flags == 0) {
15610 			/* no timers up */
15611 			return (0);
15612 		}
15613 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
15614 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15615 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
15616 		}
15617 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
15618 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
15619 		}
15620 		rack_log_chg_info(tp, rack, 2,
15621 				  rack->r_ctl.rc_hpts_flags,
15622 				  rack->r_ctl.rc_last_output_to,
15623 				  rack->r_ctl.rc_timer_exp);
15624 		return (1);
15625 		break;
15626 	case TCP_QUERY_RACK_TIMES:
15627 		/* Reordering items */
15628 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
15629 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
15630 		/* Timerstamps and timers */
15631 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
15632 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
15633 		reqr->rack_rtt = rack->rc_rack_rtt;
15634 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
15635 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
15636 		/* PRR data */
15637 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
15638 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
15639 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
15640 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
15641 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
15642 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
15643 		/* TLP and persists info */
15644 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
15645 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
15646 		if (rack->rc_in_persist) {
15647 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
15648 			reqr->rack_in_persist = 1;
15649 		} else {
15650 			reqr->rack_time_went_idle = 0;
15651 			reqr->rack_in_persist = 0;
15652 		}
15653 		if (rack->r_wanted_output)
15654 			reqr->rack_wanted_output = 1;
15655 		else
15656 			reqr->rack_wanted_output = 0;
15657 		return (1);
15658 		break;
15659 	default:
15660 		return (-EINVAL);
15661 	}
15662 }
15663 
15664 static void
15665 rack_switch_failed(struct tcpcb *tp)
15666 {
15667 	/*
15668 	 * This method gets called if a stack switch was
15669 	 * attempted and it failed. We are left
15670 	 * but our hpts timers were stopped and we
15671 	 * need to validate time units and t_flags2.
15672 	 */
15673 	struct tcp_rack *rack;
15674 	struct timeval tv;
15675 	uint32_t cts;
15676 	uint32_t toval;
15677 	struct hpts_diag diag;
15678 
15679 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15680 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
15681 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15682 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15683 	else
15684 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15685 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15686 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
15687 	if (tp->t_in_hpts > IHPTS_NONE) {
15688 		/* Strange */
15689 		return;
15690 	}
15691 	cts = tcp_get_usecs(&tv);
15692 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15693 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
15694 			toval = rack->r_ctl.rc_last_output_to - cts;
15695 		} else {
15696 			/* one slot please */
15697 			toval = HPTS_TICKS_PER_SLOT;
15698 		}
15699 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
15700 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
15701 			toval = rack->r_ctl.rc_timer_exp - cts;
15702 		} else {
15703 			/* one slot please */
15704 			toval = HPTS_TICKS_PER_SLOT;
15705 		}
15706 	} else
15707 		toval = HPTS_TICKS_PER_SLOT;
15708 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
15709 				   __LINE__, &diag);
15710 	rack_log_hpts_diag(rack, cts, &diag, &tv);
15711 }
15712 
15713 static int
15714 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
15715 {
15716 	struct rack_sendmap *rsm, *ersm;
15717 	int insret __diagused;
15718 	/*
15719 	 * When initing outstanding, we must be quite careful
15720 	 * to not refer to tp->t_fb_ptr. This has the old rack
15721 	 * pointer in it, not the "new" one (when we are doing
15722 	 * a stack switch).
15723 	 */
15724 
15725 
15726 	if (tp->t_fb->tfb_chg_query == NULL) {
15727 		/* Create a send map for the current outstanding data */
15728 
15729 		rsm = rack_alloc(rack);
15730 		if (rsm == NULL) {
15731 			uma_zfree(rack_pcb_zone, ptr);
15732 			return (ENOMEM);
15733 		}
15734 		rsm->r_no_rtt_allowed = 1;
15735 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
15736 		rsm->r_rtr_cnt = 1;
15737 		rsm->r_rtr_bytes = 0;
15738 		if (tp->t_flags & TF_SENTFIN)
15739 			rsm->r_flags |= RACK_HAS_FIN;
15740 		rsm->r_end = tp->snd_max;
15741 		if (tp->snd_una == tp->iss) {
15742 			/* The data space is one beyond snd_una */
15743 			rsm->r_flags |= RACK_HAS_SYN;
15744 			rsm->r_start = tp->iss;
15745 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
15746 		} else
15747 			rsm->r_start = tp->snd_una;
15748 		rsm->r_dupack = 0;
15749 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
15750 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
15751 			if (rsm->m) {
15752 				rsm->orig_m_len = rsm->m->m_len;
15753 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
15754 			} else {
15755 				rsm->orig_m_len = 0;
15756 				rsm->orig_t_space = 0;
15757 			}
15758 		} else {
15759 			/*
15760 			 * This can happen if we have a stand-alone FIN or
15761 			 *  SYN.
15762 			 */
15763 			rsm->m = NULL;
15764 			rsm->orig_m_len = 0;
15765 			rsm->orig_t_space = 0;
15766 			rsm->soff = 0;
15767 		}
15768 #ifdef INVARIANTS
15769 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
15770 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
15771 			      insret, rack, rsm);
15772 		}
15773 #else
15774 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
15775 #endif
15776 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
15777 		rsm->r_in_tmap = 1;
15778 	} else {
15779 		/* We have a query mechanism, lets use it */
15780 		struct tcp_query_resp qr;
15781 		int i;
15782 		tcp_seq at;
15783 
15784 		at = tp->snd_una;
15785 		while (at != tp->snd_max) {
15786 			memset(&qr, 0, sizeof(qr));
15787 			qr.req = TCP_QUERY_SENDMAP;
15788 			qr.req_param = at;
15789 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
15790 				break;
15791 			/* Move forward */
15792 			at = qr.sendmap_end;
15793 			/* Now lets build the entry for this one */
15794 			rsm = rack_alloc(rack);
15795 			if (rsm == NULL) {
15796 				uma_zfree(rack_pcb_zone, ptr);
15797 				return (ENOMEM);
15798 			}
15799 			memset(rsm, 0, sizeof(struct rack_sendmap));
15800 			/* Now configure the rsm and insert it */
15801 			rsm->r_dupack = qr.sendmap_dupacks;
15802 			rsm->r_start = qr.sendmap_start;
15803 			rsm->r_end = qr.sendmap_end;
15804 			if (qr.sendmap_fas)
15805 				rsm->r_fas = qr.sendmap_end;
15806 			else
15807 				rsm->r_fas = rsm->r_start - tp->snd_una;
15808 			/*
15809 			 * We have carefully aligned the bits
15810 			 * so that all we have to do is copy over
15811 			 * the bits with the mask.
15812 			 */
15813 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
15814 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
15815 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
15816 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
15817 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
15818 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
15819 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
15820 					   (rsm->r_start - tp->snd_una), &rsm->soff);
15821 			if (rsm->m) {
15822 				rsm->orig_m_len = rsm->m->m_len;
15823 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
15824 			} else {
15825 				rsm->orig_m_len = 0;
15826 				rsm->orig_t_space = 0;
15827 			}
15828 #ifdef INVARIANTS
15829 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
15830 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
15831 				      insret, rack, rsm);
15832 			}
15833 #else
15834 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
15835 #endif
15836 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
15837 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
15838 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
15839 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
15840 						/*
15841 						 * If the existing ersm was sent at
15842 						 * a later time than the new one, then
15843 						 * the new one should appear ahead of this
15844 						 * ersm.
15845 						 */
15846 						rsm->r_in_tmap = 1;
15847 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
15848 						break;
15849 					}
15850 				}
15851 				if (rsm->r_in_tmap == 0) {
15852 					/*
15853 					 * Not found so shove it on the tail.
15854 					 */
15855 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
15856 					rsm->r_in_tmap = 1;
15857 				}
15858  			} else {
15859 				if ((rack->r_ctl.rc_sacklast == NULL) ||
15860 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
15861 					rack->r_ctl.rc_sacklast = rsm;
15862 				}
15863 			}
15864 			rack_log_chg_info(tp, rack, 3,
15865 					  rsm->r_start,
15866 					  rsm->r_end,
15867 					  rsm->r_flags);
15868 		}
15869 	}
15870 	return (0);
15871 }
15872 
15873 static void
15874 rack_translate_policer_detect(struct tcp_rack *rack, uint32_t optval)
15875 {
15876 	/*
15877 	 * P = Percent of retransmits 499 = 49.9%
15878 	 * A = Average number 1 (.1%) -> 169 (16.9%)
15879 	 * M = Median number of retrans 1 - 16
15880 	 * MMMM MMMM AAAA AAAA PPPP PPPP PPPP PPPP
15881 	 *
15882 	 */
15883 	uint16_t per, upp;
15884 
15885 	per = optval & 0x0000ffff;
15886 	rack->r_ctl.policer_rxt_threshold = (uint32_t)(per & 0xffff);
15887 	upp = ((optval & 0xffff0000) >> 16);
15888 	rack->r_ctl.policer_avg_threshold = (0x00ff & upp);
15889 	rack->r_ctl.policer_med_threshold = ((upp >> 8) & 0x00ff);
15890 	if ((rack->r_ctl.policer_rxt_threshold > 0) &&
15891 	    (rack->r_ctl.policer_avg_threshold > 0) &&
15892 	    (rack->r_ctl.policer_med_threshold > 0)) {
15893 		rack->policer_detect_on = 1;
15894 	} else {
15895 		rack->policer_detect_on = 0;
15896 	}
15897 	rack->r_ctl.saved_policer_val = optval;
15898 	policer_detection_log(rack, optval,
15899 			      rack->r_ctl.policer_avg_threshold,
15900 			      rack->r_ctl.policer_med_threshold,
15901 			      rack->r_ctl.policer_rxt_threshold, 11);
15902 }
15903 
15904 static int32_t
15905 rack_init(struct tcpcb *tp, void **ptr)
15906 {
15907 	struct inpcb *inp = tptoinpcb(tp);
15908 	struct tcp_rack *rack = NULL;
15909 	uint32_t iwin, snt, us_cts;
15910 	size_t sz;
15911 	int err, no_query;
15912 
15913 	tcp_hpts_init(tp);
15914 
15915 	/*
15916 	 * First are we the initial or are we a switched stack?
15917 	 * If we are initing via tcp_newtcppcb the ptr passed
15918 	 * will be tp->t_fb_ptr. If its a stack switch that
15919 	 * has a previous stack we can query it will be a local
15920 	 * var that will in the end be set into t_fb_ptr.
15921 	 */
15922 	if (ptr == &tp->t_fb_ptr)
15923 		no_query = 1;
15924 	else
15925 		no_query = 0;
15926 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
15927 	if (*ptr == NULL) {
15928 		/*
15929 		 * We need to allocate memory but cant. The INP and INP_INFO
15930 		 * locks and they are recursive (happens during setup. So a
15931 		 * scheme to drop the locks fails :(
15932 		 *
15933 		 */
15934 		return(ENOMEM);
15935 	}
15936 	memset(*ptr, 0, sizeof(struct tcp_rack));
15937 	rack = (struct tcp_rack *)*ptr;
15938 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
15939 	if (rack->r_ctl.tqh == NULL) {
15940 		uma_zfree(rack_pcb_zone, rack);
15941 		return(ENOMEM);
15942 	}
15943 	tqhash_init(rack->r_ctl.tqh);
15944 	TAILQ_INIT(&rack->r_ctl.rc_free);
15945 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
15946 	rack->rc_tp = tp;
15947 	rack->rc_inp = inp;
15948 	/* Set the flag */
15949 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
15950 	/* Probably not needed but lets be sure */
15951 	rack_clear_rate_sample(rack);
15952 	/*
15953 	 * Save off the default values, socket options will poke
15954 	 * at these if pacing is not on or we have not yet
15955 	 * reached where pacing is on (gp_ready/fixed enabled).
15956 	 * When they get set into the CC module (when gp_ready
15957 	 * is enabled or we enable fixed) then we will set these
15958 	 * values into the CC and place in here the old values
15959 	 * so we have a restoral. Then we will set the flag
15960 	 * rc_pacing_cc_set. That way whenever we turn off pacing
15961 	 * or switch off this stack, we will know to go restore
15962 	 * the saved values.
15963 	 *
15964 	 * We specifically put into the beta the ecn value for pacing.
15965 	 */
15966 	rack->rc_new_rnd_needed = 1;
15967 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
15968 	/* We want abe like behavior as well */
15969 
15970 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
15971 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
15972 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
15973 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
15974 	rack->r_ctl.policer_del_mss = rack_req_del_mss;
15975 	if ((rack_policer_rxt_thresh > 0) &&
15976 	    (rack_policer_avg_thresh > 0) &&
15977 	    (rack_policer_med_thresh > 0)) {
15978 		rack->r_ctl.policer_rxt_threshold = rack_policer_rxt_thresh;
15979 		rack->r_ctl.policer_avg_threshold = rack_policer_avg_thresh;
15980 		rack->r_ctl.policer_med_threshold = rack_policer_med_thresh;
15981 		rack->policer_detect_on = 1;
15982 	} else {
15983 		rack->policer_detect_on = 0;
15984 	}
15985 	if (rack_fill_cw_state)
15986 		rack->rc_pace_to_cwnd = 1;
15987 	if (rack_pacing_min_seg)
15988 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
15989 	if (use_rack_rr)
15990 		rack->use_rack_rr = 1;
15991 	if (rack_dnd_default) {
15992 		rack->rc_pace_dnd = 1;
15993 	}
15994 	if (V_tcp_delack_enabled)
15995 		tp->t_delayed_ack = 1;
15996 	else
15997 		tp->t_delayed_ack = 0;
15998 #ifdef TCP_ACCOUNTING
15999 	if (rack_tcp_accounting) {
16000 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
16001 	}
16002 #endif
16003 	rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
16004 	sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
16005 	rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
16006 	if (rack->r_ctl.pcm_s == NULL) {
16007 		rack->r_ctl.pcm_i.cnt_alloc = 0;
16008 	}
16009 #ifdef NETFLIX_STATS
16010 	rack->r_ctl.side_chan_dis_mask = tcp_sidechannel_disable_mask;
16011 #endif
16012 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
16013 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
16014 	if (rack_enable_shared_cwnd)
16015 		rack->rack_enable_scwnd = 1;
16016 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
16017 	rack->rc_user_set_max_segs = rack_hptsi_segments;
16018 	rack->r_ctl.max_reduction = rack_max_reduce;
16019 	rack->rc_force_max_seg = 0;
16020 	TAILQ_INIT(&rack->r_ctl.opt_list);
16021 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
16022 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
16023 	if (rack_hibeta_setting) {
16024 		rack->rack_hibeta = 1;
16025 		if ((rack_hibeta_setting >= 50) &&
16026 		    (rack_hibeta_setting <= 100)) {
16027 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
16028 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
16029 		}
16030 	} else {
16031 		rack->r_ctl.saved_hibeta = 50;
16032 	}
16033 	/*
16034 	 * We initialize to all ones so we never match 0
16035 	 * just in case the client sends in 0, it hopefully
16036 	 * will never have all 1's in ms :-)
16037 	 */
16038 	rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
16039 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
16040 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
16041 	rack->r_ctl.pol_bw_comp = rack_policing_do_bw_comp;
16042 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
16043 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
16044 	rack->r_ctl.rc_highest_us_rtt = 0;
16045 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
16046 	rack->pcm_enabled = rack_pcm_is_enabled;
16047 	if (rack_fillcw_bw_cap)
16048 		rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
16049 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
16050 	if (rack_use_cmp_acks)
16051 		rack->r_use_cmp_ack = 1;
16052 	if (rack_disable_prr)
16053 		rack->rack_no_prr = 1;
16054 	if (rack_gp_no_rec_chg)
16055 		rack->rc_gp_no_rec_chg = 1;
16056 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
16057 		rack->r_ctl.pacing_method |= RACK_REG_PACING;
16058 		rack->rc_always_pace = 1;
16059 		if (rack->rack_hibeta)
16060 			rack_set_cc_pacing(rack);
16061 	} else
16062 		rack->rc_always_pace = 0;
16063 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
16064 		rack->r_mbuf_queue = 1;
16065 	else
16066 		rack->r_mbuf_queue = 0;
16067 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
16068 	if (rack_limits_scwnd)
16069 		rack->r_limit_scw = 1;
16070 	else
16071 		rack->r_limit_scw = 0;
16072 	rack_init_retransmit_value(rack, rack_rxt_controls);
16073 	rack->rc_labc = V_tcp_abc_l_var;
16074 	if (rack_honors_hpts_min_to)
16075 		rack->r_use_hpts_min = 1;
16076 	if (tp->snd_una != 0) {
16077 		rack->r_ctl.idle_snd_una = tp->snd_una;
16078 		rack->rc_sendvars_notset = 0;
16079 		/*
16080 		 * Make sure any TCP timers are not running.
16081 		 */
16082 		tcp_timer_stop(tp);
16083 	} else {
16084 		/*
16085 		 * Server side, we are called from the
16086 		 * syn-cache. This means none of the
16087 		 * snd_una/max are set yet so we have
16088 		 * to defer this until the first send.
16089 		 */
16090 		rack->rc_sendvars_notset = 1;
16091 	}
16092 
16093 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
16094 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
16095 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
16096 	rack->r_ctl.rc_min_to = rack_min_to;
16097 	microuptime(&rack->r_ctl.act_rcv_time);
16098 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
16099 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
16100 	if (rack_hw_up_only)
16101 		rack->r_up_only = 1;
16102 	if (rack_do_dyn_mul) {
16103 		/* When dynamic adjustment is on CA needs to start at 100% */
16104 		rack->rc_gp_dyn_mul = 1;
16105 		if (rack_do_dyn_mul >= 100)
16106 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
16107 	} else
16108 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
16109 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
16110 	if (rack_timely_off) {
16111 		rack->rc_skip_timely = 1;
16112 	}
16113 	if (rack->rc_skip_timely) {
16114 		rack->r_ctl.rack_per_of_gp_rec = 90;
16115 		rack->r_ctl.rack_per_of_gp_ca = 100;
16116 		rack->r_ctl.rack_per_of_gp_ss = 250;
16117 	}
16118 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
16119 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
16120 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
16121 
16122 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
16123 				rack_probertt_filter_life);
16124 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16125 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
16126 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
16127 	rack->r_ctl.rc_went_idle_time = us_cts;
16128 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks() - (tcp_ack_war_time_window + 1);
16129 	rack->r_ctl.rc_time_probertt_starts = 0;
16130 
16131 	rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
16132 	if (rack_rnd_cnt_req  & 0x10000)
16133 		rack->r_ctl.gate_to_fs = 1;
16134 	rack->r_ctl.gp_gain_req = rack_gp_gain_req;
16135 	if ((rack_rnd_cnt_req & 0x100) > 0) {
16136 
16137 	}
16138 	if (rack_dsack_std_based & 0x1) {
16139 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
16140 		rack->rc_rack_tmr_std_based = 1;
16141 	}
16142 	if (rack_dsack_std_based & 0x2) {
16143 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
16144 		rack->rc_rack_use_dsack = 1;
16145 	}
16146 	/* We require at least one measurement, even if the sysctl is 0 */
16147 	if (rack_req_measurements)
16148 		rack->r_ctl.req_measurements = rack_req_measurements;
16149 	else
16150 		rack->r_ctl.req_measurements = 1;
16151 	if (rack_enable_hw_pacing)
16152 		rack->rack_hdw_pace_ena = 1;
16153 	if (rack_hw_rate_caps)
16154 		rack->r_rack_hw_rate_caps = 1;
16155 #ifdef TCP_SAD_DETECTION
16156 	rack->do_detection = 1;
16157 #else
16158 	rack->do_detection = 0;
16159 #endif
16160 	if (rack_non_rxt_use_cr)
16161 		rack->rack_rec_nonrxt_use_cr = 1;
16162 	/* Lets setup the fsb block */
16163 	err = rack_init_fsb(tp, rack);
16164 	if (err) {
16165 		uma_zfree(rack_pcb_zone, *ptr);
16166 		*ptr = NULL;
16167 		return (err);
16168 	}
16169 	if (rack_do_hystart) {
16170 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
16171 		if (rack_do_hystart > 1)
16172 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
16173 		if (rack_do_hystart > 2)
16174 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
16175 	}
16176 	/* Log what we will do with queries */
16177 	rack_log_chg_info(tp, rack, 7,
16178 			  no_query, 0, 0);
16179 	if (rack_def_profile)
16180 		rack_set_profile(rack, rack_def_profile);
16181 	/* Cancel the GP measurement in progress */
16182 	tp->t_flags &= ~TF_GPUTINPROG;
16183 	if ((tp->t_state != TCPS_CLOSED) &&
16184 	    (tp->t_state != TCPS_TIME_WAIT)) {
16185 		/*
16186 		 * We are already open, we may
16187 		 * need to adjust a few things.
16188 		 */
16189 		if (SEQ_GT(tp->snd_max, tp->iss))
16190 			snt = tp->snd_max - tp->iss;
16191 		else
16192 			snt = 0;
16193 		iwin = rc_init_window(rack);
16194 		if ((snt < iwin) &&
16195 		    (no_query == 1)) {
16196 			/* We are not past the initial window
16197 			 * on the first init (i.e. a stack switch
16198 			 * has not yet occured) so we need to make
16199 			 * sure cwnd and ssthresh is correct.
16200 			 */
16201 			if (tp->snd_cwnd < iwin)
16202 				tp->snd_cwnd = iwin;
16203 			/*
16204 			 * If we are within the initial window
16205 			 * we want ssthresh to be unlimited. Setting
16206 			 * it to the rwnd (which the default stack does
16207 			 * and older racks) is not really a good idea
16208 			 * since we want to be in SS and grow both the
16209 			 * cwnd and the rwnd (via dynamic rwnd growth). If
16210 			 * we set it to the rwnd then as the peer grows its
16211 			 * rwnd we will be stuck in CA and never hit SS.
16212 			 *
16213 			 * Its far better to raise it up high (this takes the
16214 			 * risk that there as been a loss already, probably
16215 			 * we should have an indicator in all stacks of loss
16216 			 * but we don't), but considering the normal use this
16217 			 * is a risk worth taking. The consequences of not
16218 			 * hitting SS are far worse than going one more time
16219 			 * into it early on (before we have sent even a IW).
16220 			 * It is highly unlikely that we will have had a loss
16221 			 * before getting the IW out.
16222 			 */
16223 			tp->snd_ssthresh = 0xffffffff;
16224 		}
16225 		/*
16226 		 * Any init based on sequence numbers
16227 		 * should be done in the deferred init path
16228 		 * since we can be CLOSED and not have them
16229 		 * inited when rack_init() is called. We
16230 		 * are not closed so lets call it.
16231 		 */
16232 		rack_deferred_init(tp, rack);
16233 	}
16234 	if ((tp->t_state != TCPS_CLOSED) &&
16235 	    (tp->t_state != TCPS_TIME_WAIT) &&
16236 	    (no_query == 0) &&
16237 	    (tp->snd_una != tp->snd_max))  {
16238 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
16239 		if (err) {
16240 			*ptr = NULL;
16241 			return(err);
16242 		}
16243 	}
16244 	rack_stop_all_timers(tp, rack);
16245 	/* Setup all the t_flags2 */
16246 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
16247 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
16248 	else
16249 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
16250 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
16251 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
16252 	/*
16253 	 * Timers in Rack are kept in microseconds so lets
16254 	 * convert any initial incoming variables
16255 	 * from ticks into usecs. Note that we
16256 	 * also change the values of t_srtt and t_rttvar, if
16257 	 * they are non-zero. They are kept with a 5
16258 	 * bit decimal so we have to carefully convert
16259 	 * these to get the full precision.
16260 	 */
16261 	rack_convert_rtts(tp);
16262 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
16263 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
16264 		/* We do not start any timers on DROPPED connections */
16265 		if (tp->t_fb->tfb_chg_query == NULL) {
16266 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16267 		} else {
16268 			struct tcp_query_resp qr;
16269 			int ret;
16270 
16271 			memset(&qr, 0, sizeof(qr));
16272 
16273 			/* Get the misc time stamps and such for rack */
16274 			qr.req = TCP_QUERY_RACK_TIMES;
16275 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
16276 			if (ret == 1) {
16277 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
16278 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
16279 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
16280 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
16281 				rack->rc_rack_rtt = qr.rack_rtt;
16282 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
16283 				rack->r_ctl.rc_sacked = qr.rack_sacked;
16284 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
16285 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
16286 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
16287 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
16288 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
16289 				if (qr.rack_tlp_out) {
16290 					rack->rc_tlp_in_progress = 1;
16291 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
16292 				} else {
16293 					rack->rc_tlp_in_progress = 0;
16294 					rack->r_ctl.rc_tlp_cnt_out = 0;
16295 				}
16296 				if (qr.rack_srtt_measured)
16297 					rack->rc_srtt_measure_made = 1;
16298 				if (qr.rack_in_persist == 1) {
16299 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
16300 #ifdef NETFLIX_SHARED_CWND
16301 					if (rack->r_ctl.rc_scw) {
16302 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16303 						rack->rack_scwnd_is_idle = 1;
16304 					}
16305 #endif
16306 					rack->r_ctl.persist_lost_ends = 0;
16307 					rack->probe_not_answered = 0;
16308 					rack->forced_ack = 0;
16309 					tp->t_rxtshift = 0;
16310 					rack->rc_in_persist = 1;
16311 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16312 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16313 				}
16314 				if (qr.rack_wanted_output)
16315 					rack->r_wanted_output = 1;
16316 				rack_log_chg_info(tp, rack, 6,
16317 						  qr.rack_min_rtt,
16318 						  qr.rack_rtt,
16319 						  qr.rack_reorder_ts);
16320 			}
16321 			/* Get the old stack timers */
16322 			qr.req_param = 0;
16323 			qr.req = TCP_QUERY_TIMERS_UP;
16324 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
16325 			if (ret) {
16326 				/*
16327 				 * non-zero return means we have a timer('s)
16328 				 * to start. Zero means no timer (no keepalive
16329 				 * I suppose).
16330 				 */
16331 				uint32_t tov = 0;
16332 
16333 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
16334 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
16335 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
16336 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
16337 						tov = qr.timer_pacing_to - us_cts;
16338 					else
16339 						tov = HPTS_TICKS_PER_SLOT;
16340 				}
16341 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
16342 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
16343 					if (tov == 0) {
16344 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
16345 							tov = qr.timer_timer_exp - us_cts;
16346 						else
16347 							tov = HPTS_TICKS_PER_SLOT;
16348 					}
16349 				}
16350 				rack_log_chg_info(tp, rack, 4,
16351 						  rack->r_ctl.rc_hpts_flags,
16352 						  rack->r_ctl.rc_last_output_to,
16353 						  rack->r_ctl.rc_timer_exp);
16354 				if (tov) {
16355 					struct hpts_diag diag;
16356 
16357 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
16358 								   __LINE__, &diag);
16359 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
16360 				}
16361 			}
16362 		}
16363 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
16364 				     __LINE__, RACK_RTTS_INIT);
16365 	}
16366 	return (0);
16367 }
16368 
16369 static int
16370 rack_handoff_ok(struct tcpcb *tp)
16371 {
16372 	if ((tp->t_state == TCPS_CLOSED) ||
16373 	    (tp->t_state == TCPS_LISTEN)) {
16374 		/* Sure no problem though it may not stick */
16375 		return (0);
16376 	}
16377 	if ((tp->t_state == TCPS_SYN_SENT) ||
16378 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
16379 		/*
16380 		 * We really don't know if you support sack,
16381 		 * you have to get to ESTAB or beyond to tell.
16382 		 */
16383 		return (EAGAIN);
16384 	}
16385 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
16386 		/*
16387 		 * Rack will only send a FIN after all data is acknowledged.
16388 		 * So in this case we have more data outstanding. We can't
16389 		 * switch stacks until either all data and only the FIN
16390 		 * is left (in which case rack_init() now knows how
16391 		 * to deal with that) <or> all is acknowledged and we
16392 		 * are only left with incoming data, though why you
16393 		 * would want to switch to rack after all data is acknowledged
16394 		 * I have no idea (rrs)!
16395 		 */
16396 		return (EAGAIN);
16397 	}
16398 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
16399 		return (0);
16400 	}
16401 	/*
16402 	 * If we reach here we don't do SACK on this connection so we can
16403 	 * never do rack.
16404 	 */
16405 	return (EINVAL);
16406 }
16407 
16408 static void
16409 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
16410 {
16411 
16412 	if (tp->t_fb_ptr) {
16413 		uint32_t cnt_free = 0;
16414 		struct tcp_rack *rack;
16415 		struct rack_sendmap *rsm;
16416 
16417 		tcp_handle_orphaned_packets(tp);
16418 		tp->t_flags &= ~TF_FORCEDATA;
16419 		rack = (struct tcp_rack *)tp->t_fb_ptr;
16420 		rack_log_pacing_delay_calc(rack,
16421 					   0,
16422 					   0,
16423 					   0,
16424 					   rack_get_gp_est(rack), /* delRate */
16425 					   rack_get_lt_bw(rack), /* rttProp */
16426 					   20, __LINE__, NULL, 0);
16427 #ifdef NETFLIX_SHARED_CWND
16428 		if (rack->r_ctl.rc_scw) {
16429 			uint32_t limit;
16430 
16431 			if (rack->r_limit_scw)
16432 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
16433 			else
16434 				limit = 0;
16435 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
16436 						  rack->r_ctl.rc_scw_index,
16437 						  limit);
16438 			rack->r_ctl.rc_scw = NULL;
16439 		}
16440 #endif
16441 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
16442 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
16443 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
16444 			rack->r_ctl.fsb.th = NULL;
16445 		}
16446 		if (rack->rc_always_pace == 1) {
16447 			rack_remove_pacing(rack);
16448 		}
16449 		/* Clean up any options if they were not applied */
16450 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
16451 			struct deferred_opt_list *dol;
16452 
16453 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
16454 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
16455 			free(dol, M_TCPDO);
16456 		}
16457 		/* rack does not use force data but other stacks may clear it */
16458 		if (rack->r_ctl.crte != NULL) {
16459 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
16460 			rack->rack_hdrw_pacing = 0;
16461 			rack->r_ctl.crte = NULL;
16462 		}
16463 #ifdef TCP_BLACKBOX
16464 		tcp_log_flowend(tp);
16465 #endif
16466 		/*
16467 		 * Lets take a different approach to purging just
16468 		 * get each one and free it like a cum-ack would and
16469 		 * not use a foreach loop.
16470 		 */
16471 		rsm = tqhash_min(rack->r_ctl.tqh);
16472 		while (rsm) {
16473 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
16474 			rack->r_ctl.rc_num_maps_alloced--;
16475 			uma_zfree(rack_zone, rsm);
16476 			rsm = tqhash_min(rack->r_ctl.tqh);
16477 		}
16478 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
16479 		while (rsm) {
16480 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
16481 			rack->r_ctl.rc_num_maps_alloced--;
16482 			rack->rc_free_cnt--;
16483 			cnt_free++;
16484 			uma_zfree(rack_zone, rsm);
16485 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
16486 		}
16487 		if (rack->r_ctl.pcm_s != NULL) {
16488 			free(rack->r_ctl.pcm_s, M_TCPPCM);
16489 			rack->r_ctl.pcm_s = NULL;
16490 			rack->r_ctl.pcm_i.cnt_alloc = 0;
16491 			rack->r_ctl.pcm_i.cnt = 0;
16492 		}
16493 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
16494 		    (tcp_bblogging_on(tp))) {
16495 			union tcp_log_stackspecific log;
16496 			struct timeval tv;
16497 
16498 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16499 			log.u_bbr.flex8 = 10;
16500 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
16501 			log.u_bbr.flex2 = rack->rc_free_cnt;
16502 			log.u_bbr.flex3 = cnt_free;
16503 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16504 			rsm = tqhash_min(rack->r_ctl.tqh);
16505 			log.u_bbr.delRate = (uint64_t)rsm;
16506 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
16507 			log.u_bbr.cur_del_rate = (uint64_t)rsm;
16508 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
16509 			log.u_bbr.pkt_epoch = __LINE__;
16510 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16511 					     0, &log, false, NULL, NULL, 0, &tv);
16512 		}
16513 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
16514 			("rack:%p num_aloc:%u after freeing all?",
16515 			 rack,
16516 			 rack->r_ctl.rc_num_maps_alloced));
16517 		rack->rc_free_cnt = 0;
16518 		free(rack->r_ctl.tqh, M_TCPFSB);
16519 		rack->r_ctl.tqh = NULL;
16520 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
16521 		tp->t_fb_ptr = NULL;
16522 	}
16523 	/* Make sure snd_nxt is correctly set */
16524 	tp->snd_nxt = tp->snd_max;
16525 }
16526 
16527 static void
16528 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
16529 {
16530 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
16531 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
16532 	}
16533 	switch (tp->t_state) {
16534 	case TCPS_SYN_SENT:
16535 		rack->r_state = TCPS_SYN_SENT;
16536 		rack->r_substate = rack_do_syn_sent;
16537 		break;
16538 	case TCPS_SYN_RECEIVED:
16539 		rack->r_state = TCPS_SYN_RECEIVED;
16540 		rack->r_substate = rack_do_syn_recv;
16541 		break;
16542 	case TCPS_ESTABLISHED:
16543 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
16544 		rack->r_state = TCPS_ESTABLISHED;
16545 		rack->r_substate = rack_do_established;
16546 		break;
16547 	case TCPS_CLOSE_WAIT:
16548 		rack->r_state = TCPS_CLOSE_WAIT;
16549 		rack->r_substate = rack_do_close_wait;
16550 		break;
16551 	case TCPS_FIN_WAIT_1:
16552 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
16553 		rack->r_state = TCPS_FIN_WAIT_1;
16554 		rack->r_substate = rack_do_fin_wait_1;
16555 		break;
16556 	case TCPS_CLOSING:
16557 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
16558 		rack->r_state = TCPS_CLOSING;
16559 		rack->r_substate = rack_do_closing;
16560 		break;
16561 	case TCPS_LAST_ACK:
16562 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
16563 		rack->r_state = TCPS_LAST_ACK;
16564 		rack->r_substate = rack_do_lastack;
16565 		break;
16566 	case TCPS_FIN_WAIT_2:
16567 		rack->r_state = TCPS_FIN_WAIT_2;
16568 		rack->r_substate = rack_do_fin_wait_2;
16569 		break;
16570 	case TCPS_LISTEN:
16571 	case TCPS_CLOSED:
16572 	case TCPS_TIME_WAIT:
16573 	default:
16574 		break;
16575 	};
16576 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
16577 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
16578 
16579 }
16580 
16581 static void
16582 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
16583 {
16584 	/*
16585 	 * We received an ack, and then did not
16586 	 * call send or were bounced out due to the
16587 	 * hpts was running. Now a timer is up as well, is
16588 	 * it the right timer?
16589 	 */
16590 	struct rack_sendmap *rsm;
16591 	int tmr_up;
16592 
16593 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
16594 	if (tcp_in_hpts(rack->rc_tp) == 0) {
16595 		/*
16596 		 * Ok we probably need some timer up, but no
16597 		 * matter what the mask we are not in hpts. We
16598 		 * may have received an old ack and thus did nothing.
16599 		 */
16600 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16601 		rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16602 		return;
16603 	}
16604 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
16605 		return;
16606 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16607 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
16608 	    (tmr_up == PACE_TMR_RXT)) {
16609 		/* Should be an RXT */
16610 		return;
16611 	}
16612 	if (rsm == NULL) {
16613 		/* Nothing outstanding? */
16614 		if (tp->t_flags & TF_DELACK) {
16615 			if (tmr_up == PACE_TMR_DELACK)
16616 				/* We are supposed to have delayed ack up and we do */
16617 				return;
16618 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
16619 			/*
16620 			 * if we hit enobufs then we would expect the possibility
16621 			 * of nothing outstanding and the RXT up (and the hptsi timer).
16622 			 */
16623 			return;
16624 		} else if (((V_tcp_always_keepalive ||
16625 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16626 			    (tp->t_state <= TCPS_CLOSING)) &&
16627 			   (tmr_up == PACE_TMR_KEEP) &&
16628 			   (tp->snd_max == tp->snd_una)) {
16629 			/* We should have keep alive up and we do */
16630 			return;
16631 		}
16632 	}
16633 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
16634 		   ((tmr_up == PACE_TMR_TLP) ||
16635 		    (tmr_up == PACE_TMR_RACK) ||
16636 		    (tmr_up == PACE_TMR_RXT))) {
16637 		/*
16638 		 * Either a Rack, TLP or RXT is fine if  we
16639 		 * have outstanding data.
16640 		 */
16641 		return;
16642 	} else if (tmr_up == PACE_TMR_DELACK) {
16643 		/*
16644 		 * If the delayed ack was going to go off
16645 		 * before the rtx/tlp/rack timer were going to
16646 		 * expire, then that would be the timer in control.
16647 		 * Note we don't check the time here trusting the
16648 		 * code is correct.
16649 		 */
16650 		return;
16651 	}
16652 	/*
16653 	 * Ok the timer originally started is not what we want now.
16654 	 * We will force the hpts to be stopped if any, and restart
16655 	 * with the slot set to what was in the saved slot.
16656 	 */
16657 	if (tcp_in_hpts(rack->rc_tp)) {
16658 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16659 			uint32_t us_cts;
16660 
16661 			us_cts = tcp_get_usecs(NULL);
16662 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
16663 				rack->r_early = 1;
16664 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
16665 			}
16666 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16667 		}
16668 		tcp_hpts_remove(rack->rc_tp);
16669 	}
16670 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16671 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16672 }
16673 
16674 
16675 static void
16676 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
16677 {
16678 	if ((SEQ_LT(tp->snd_wl1, seq) ||
16679 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
16680 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
16681 		/* keep track of pure window updates */
16682 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
16683 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
16684 		tp->snd_wnd = tiwin;
16685 		rack_validate_fo_sendwin_up(tp, rack);
16686 		tp->snd_wl1 = seq;
16687 		tp->snd_wl2 = ack;
16688 		if (tp->snd_wnd > tp->max_sndwnd)
16689 			tp->max_sndwnd = tp->snd_wnd;
16690 	    rack->r_wanted_output = 1;
16691 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
16692 		tp->snd_wnd = tiwin;
16693 		rack_validate_fo_sendwin_up(tp, rack);
16694 		tp->snd_wl1 = seq;
16695 		tp->snd_wl2 = ack;
16696 	} else {
16697 		/* Not a valid win update */
16698 		return;
16699 	}
16700 	if (tp->snd_wnd > tp->max_sndwnd)
16701 		tp->max_sndwnd = tp->snd_wnd;
16702 	/* Do we exit persists? */
16703 	if ((rack->rc_in_persist != 0) &&
16704 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
16705 				rack->r_ctl.rc_pace_min_segs))) {
16706 		rack_exit_persist(tp, rack, cts);
16707 	}
16708 	/* Do we enter persists? */
16709 	if ((rack->rc_in_persist == 0) &&
16710 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
16711 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
16712 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
16713 	    sbavail(&tptosocket(tp)->so_snd) &&
16714 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
16715 		/*
16716 		 * Here the rwnd is less than
16717 		 * the pacing size, we are established,
16718 		 * nothing is outstanding, and there is
16719 		 * data to send. Enter persists.
16720 		 */
16721 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
16722 	}
16723 }
16724 
16725 static void
16726 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
16727 {
16728 
16729 	if (tcp_bblogging_on(rack->rc_tp)) {
16730 		struct inpcb *inp = tptoinpcb(tp);
16731 		union tcp_log_stackspecific log;
16732 		struct timeval ltv;
16733 		char tcp_hdr_buf[60];
16734 		struct tcphdr *th;
16735 		struct timespec ts;
16736 		uint32_t orig_snd_una;
16737 		uint8_t xx = 0;
16738 
16739 #ifdef TCP_REQUEST_TRK
16740 		struct tcp_sendfile_track *tcp_req;
16741 
16742 		if (SEQ_GT(ae->ack, tp->snd_una)) {
16743 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
16744 		} else {
16745 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
16746 		}
16747 #endif
16748 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16749 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16750 		if (rack->rack_no_prr == 0)
16751 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16752 		else
16753 			log.u_bbr.flex1 = 0;
16754 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16755 		log.u_bbr.use_lt_bw <<= 1;
16756 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16757 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16758 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16759 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
16760 		log.u_bbr.pkts_out = tp->t_maxseg;
16761 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16762 		log.u_bbr.flex7 = 1;
16763 		log.u_bbr.lost = ae->flags;
16764 		log.u_bbr.cwnd_gain = ackval;
16765 		log.u_bbr.pacing_gain = 0x2;
16766 		if (ae->flags & TSTMP_HDWR) {
16767 			/* Record the hardware timestamp if present */
16768 			log.u_bbr.flex3 = M_TSTMP;
16769 			ts.tv_sec = ae->timestamp / 1000000000;
16770 			ts.tv_nsec = ae->timestamp % 1000000000;
16771 			ltv.tv_sec = ts.tv_sec;
16772 			ltv.tv_usec = ts.tv_nsec / 1000;
16773 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16774 		} else if (ae->flags & TSTMP_LRO) {
16775 			/* Record the LRO the arrival timestamp */
16776 			log.u_bbr.flex3 = M_TSTMP_LRO;
16777 			ts.tv_sec = ae->timestamp / 1000000000;
16778 			ts.tv_nsec = ae->timestamp % 1000000000;
16779 			ltv.tv_sec = ts.tv_sec;
16780 			ltv.tv_usec = ts.tv_nsec / 1000;
16781 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16782 		}
16783 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16784 		/* Log the rcv time */
16785 		log.u_bbr.delRate = ae->timestamp;
16786 #ifdef TCP_REQUEST_TRK
16787 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16788 		log.u_bbr.applimited <<= 8;
16789 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16790 		log.u_bbr.applimited <<= 8;
16791 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16792 		if (tcp_req) {
16793 			/* Copy out any client req info */
16794 			/* seconds */
16795 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16796 			/* useconds */
16797 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16798 			log.u_bbr.rttProp = tcp_req->timestamp;
16799 			log.u_bbr.cur_del_rate = tcp_req->start;
16800 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16801 				log.u_bbr.flex8 |= 1;
16802 			} else {
16803 				log.u_bbr.flex8 |= 2;
16804 				log.u_bbr.bw_inuse = tcp_req->end;
16805 			}
16806 			log.u_bbr.flex6 = tcp_req->start_seq;
16807 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16808 				log.u_bbr.flex8 |= 4;
16809 				log.u_bbr.epoch = tcp_req->end_seq;
16810 			}
16811 		}
16812 #endif
16813 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
16814 		th = (struct tcphdr *)tcp_hdr_buf;
16815 		th->th_seq = ae->seq;
16816 		th->th_ack = ae->ack;
16817 		th->th_win = ae->win;
16818 		/* Now fill in the ports */
16819 		th->th_sport = inp->inp_fport;
16820 		th->th_dport = inp->inp_lport;
16821 		tcp_set_flags(th, ae->flags);
16822 		/* Now do we have a timestamp option? */
16823 		if (ae->flags & HAS_TSTMP) {
16824 			u_char *cp;
16825 			uint32_t val;
16826 
16827 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
16828 			cp = (u_char *)(th + 1);
16829 			*cp = TCPOPT_NOP;
16830 			cp++;
16831 			*cp = TCPOPT_NOP;
16832 			cp++;
16833 			*cp = TCPOPT_TIMESTAMP;
16834 			cp++;
16835 			*cp = TCPOLEN_TIMESTAMP;
16836 			cp++;
16837 			val = htonl(ae->ts_value);
16838 			bcopy((char *)&val,
16839 			      (char *)cp, sizeof(uint32_t));
16840 			val = htonl(ae->ts_echo);
16841 			bcopy((char *)&val,
16842 			      (char *)(cp + 4), sizeof(uint32_t));
16843 		} else
16844 			th->th_off = (sizeof(struct tcphdr) >> 2);
16845 
16846 		/*
16847 		 * For sane logging we need to play a little trick.
16848 		 * If the ack were fully processed we would have moved
16849 		 * snd_una to high_seq, but since compressed acks are
16850 		 * processed in two phases, at this point (logging) snd_una
16851 		 * won't be advanced. So we would see multiple acks showing
16852 		 * the advancement. We can prevent that by "pretending" that
16853 		 * snd_una was advanced and then un-advancing it so that the
16854 		 * logging code has the right value for tlb_snd_una.
16855 		 */
16856 		if (tp->snd_una != high_seq) {
16857 			orig_snd_una = tp->snd_una;
16858 			tp->snd_una = high_seq;
16859 			xx = 1;
16860 		} else
16861 			xx = 0;
16862 		TCP_LOG_EVENTP(tp, th,
16863 			       &tptosocket(tp)->so_rcv,
16864 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
16865 			       0, &log, true, &ltv);
16866 		if (xx) {
16867 			tp->snd_una = orig_snd_una;
16868 		}
16869 	}
16870 
16871 }
16872 
16873 static void
16874 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
16875 {
16876 	uint32_t us_rtt;
16877 	/*
16878 	 * A persist or keep-alive was forced out, update our
16879 	 * min rtt time. Note now worry about lost responses.
16880 	 * When a subsequent keep-alive or persist times out
16881 	 * and forced_ack is still on, then the last probe
16882 	 * was not responded to. In such cases we have a
16883 	 * sysctl that controls the behavior. Either we apply
16884 	 * the rtt but with reduced confidence (0). Or we just
16885 	 * plain don't apply the rtt estimate. Having data flow
16886 	 * will clear the probe_not_answered flag i.e. cum-ack
16887 	 * move forward <or> exiting and reentering persists.
16888 	 */
16889 
16890 	rack->forced_ack = 0;
16891 	rack->rc_tp->t_rxtshift = 0;
16892 	if ((rack->rc_in_persist &&
16893 	     (tiwin == rack->rc_tp->snd_wnd)) ||
16894 	    (rack->rc_in_persist == 0)) {
16895 		/*
16896 		 * In persists only apply the RTT update if this is
16897 		 * a response to our window probe. And that
16898 		 * means the rwnd sent must match the current
16899 		 * snd_wnd. If it does not, then we got a
16900 		 * window update ack instead. For keepalive
16901 		 * we allow the answer no matter what the window.
16902 		 *
16903 		 * Note that if the probe_not_answered is set then
16904 		 * the forced_ack_ts is the oldest one i.e. the first
16905 		 * probe sent that might have been lost. This assures
16906 		 * us that if we do calculate an RTT it is longer not
16907 		 * some short thing.
16908 		 */
16909 		if (rack->rc_in_persist)
16910 			counter_u64_add(rack_persists_acks, 1);
16911 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
16912 		if (us_rtt == 0)
16913 			us_rtt = 1;
16914 		if (rack->probe_not_answered == 0) {
16915 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
16916 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
16917 		} else {
16918 			/* We have a retransmitted probe here too */
16919 			if (rack_apply_rtt_with_reduced_conf) {
16920 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
16921 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
16922 			}
16923 		}
16924 	}
16925 }
16926 
16927 static void
16928 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
16929 {
16930 	/*
16931 	 * The next send has occurred mark the end of the round
16932 	 * as when that data gets acknowledged. We can
16933 	 * also do common things we might need to do when
16934 	 * a round begins.
16935 	 */
16936 	rack->r_ctl.roundends = tp->snd_max;
16937 	rack->rc_new_rnd_needed = 0;
16938 	rack_log_hystart_event(rack, tp->snd_max, 4);
16939 }
16940 
16941 
16942 static void
16943 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
16944 	     uint32_t flex3)
16945 {
16946 	if (tcp_bblogging_on(rack->rc_tp)) {
16947 		union tcp_log_stackspecific log;
16948 		struct timeval tv;
16949 
16950 		(void)tcp_get_usecs(&tv);
16951 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16952 		log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
16953 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16954 		log.u_bbr.flex8 = mod;
16955 		log.u_bbr.flex1 = flex1;
16956 		log.u_bbr.flex2 = flex2;
16957 		log.u_bbr.flex3 = flex3;
16958 		log.u_bbr.flex4 = rack_pcm_every_n_rounds;
16959 		log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
16960 		log.u_bbr.bbr_substate = rack->pcm_needed;
16961 		log.u_bbr.bbr_substate <<= 1;
16962 		log.u_bbr.bbr_substate |= rack->pcm_in_progress;
16963 		log.u_bbr.bbr_substate <<= 1;
16964 		log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
16965 		(void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
16966 				    0, &log, false, NULL, NULL, 0, &tv);
16967 	}
16968 }
16969 
16970 static void
16971 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
16972 {
16973 	/*
16974 	 * The round (current_round) has ended. We now
16975 	 * setup for the next round by incrementing the
16976 	 * round numnber and doing any round specific
16977 	 * things.
16978 	 */
16979 	rack_log_hystart_event(rack, high_seq, 21);
16980 	rack->r_ctl.current_round++;
16981 	/* New round (current_round) begins at next send */
16982 	rack->rc_new_rnd_needed = 1;
16983 	if ((rack->pcm_enabled == 1) &&
16984 	    (rack->pcm_needed == 0) &&
16985 	    (rack->pcm_in_progress == 0)) {
16986 		/*
16987 		 * If we have enabled PCM, then we need to
16988 		 * check if the round has adanced to the state
16989 		 * where one is required.
16990 		 */
16991 		int rnds;
16992 
16993 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
16994 		if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
16995 			rack->pcm_needed = 1;
16996 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
16997 		} else if (rack_verbose_logging) {
16998 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
16999 		}
17000 	}
17001 	if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
17002 		/* We have hystart enabled send the round info in */
17003 		if (CC_ALGO(tp)->newround != NULL) {
17004 			CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
17005 		}
17006 	}
17007 	/*
17008 	 * For DGP an initial startup check. We want to validate
17009 	 * that we are not just pushing on slow-start and just
17010 	 * not gaining.. i.e. filling buffers without getting any
17011 	 * boost in b/w during the inital slow-start.
17012 	 */
17013 	if (rack->dgp_on &&
17014 	    (rack->rc_initial_ss_comp == 0) &&
17015 	    (tp->snd_cwnd < tp->snd_ssthresh) &&
17016 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
17017 	    (rack->r_ctl.gp_rnd_thresh > 0) &&
17018 	    ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
17019 
17020 		/*
17021 		 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
17022 		 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
17023 		 * exit SS.
17024 		 *
17025 		 * Pick up the flight size now as we enter slowstart (not the
17026 		 * cwnd which may be inflated).
17027 		 */
17028 		rack->rc_initial_ss_comp = 1;
17029 
17030 		if (tcp_bblogging_on(rack->rc_tp)) {
17031 			union tcp_log_stackspecific log;
17032 			struct timeval tv;
17033 
17034 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17035 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17036 			log.u_bbr.flex1 = rack->r_ctl.current_round;
17037 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
17038 			log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
17039 			log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
17040 			log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
17041 			log.u_bbr.flex8 = 40;
17042 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
17043 					    0, &log, false, NULL, __func__, __LINE__,&tv);
17044 		}
17045 		if ((rack->r_ctl.gate_to_fs == 1) &&
17046 		     (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
17047 			tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
17048 		}
17049 		tp->snd_ssthresh = tp->snd_cwnd - 1;
17050 		/* Turn off any fast output running */
17051 		rack->r_fast_output = 0;
17052 	}
17053 }
17054 
17055 static int
17056 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
17057 {
17058 	/*
17059 	 * Handle a "special" compressed ack mbuf. Each incoming
17060 	 * ack has only four possible dispositions:
17061 	 *
17062 	 * A) It moves the cum-ack forward
17063 	 * B) It is behind the cum-ack.
17064 	 * C) It is a window-update ack.
17065 	 * D) It is a dup-ack.
17066 	 *
17067 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
17068 	 * in the incoming mbuf. We also need to still pay attention
17069 	 * to nxt_pkt since there may be another packet after this
17070 	 * one.
17071 	 */
17072 #ifdef TCP_ACCOUNTING
17073 	uint64_t ts_val;
17074 	uint64_t rdstc;
17075 #endif
17076 	int segsiz;
17077 	struct timespec ts;
17078 	struct tcp_rack *rack;
17079 	struct tcp_ackent *ae;
17080 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
17081 	int cnt, i, did_out, ourfinisacked = 0;
17082 	struct tcpopt to_holder, *to = NULL;
17083 #ifdef TCP_ACCOUNTING
17084 	int win_up_req = 0;
17085 #endif
17086 	int nsegs = 0;
17087 	int under_pacing = 0;
17088 	int post_recovery = 0;
17089 #ifdef TCP_ACCOUNTING
17090 	sched_pin();
17091 #endif
17092 	rack = (struct tcp_rack *)tp->t_fb_ptr;
17093 	if (rack->gp_ready &&
17094 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
17095 		under_pacing = 1;
17096 
17097 	if (rack->r_state != tp->t_state)
17098 		rack_set_state(tp, rack);
17099 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
17100 	    (tp->t_flags & TF_GPUTINPROG)) {
17101 		/*
17102 		 * We have a goodput in progress
17103 		 * and we have entered a late state.
17104 		 * Do we have enough data in the sb
17105 		 * to handle the GPUT request?
17106 		 */
17107 		uint32_t bytes;
17108 
17109 		bytes = tp->gput_ack - tp->gput_seq;
17110 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
17111 			bytes += tp->gput_seq - tp->snd_una;
17112 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
17113 			/*
17114 			 * There are not enough bytes in the socket
17115 			 * buffer that have been sent to cover this
17116 			 * measurement. Cancel it.
17117 			 */
17118 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17119 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
17120 						   tp->gput_seq,
17121 						   0, 0, 18, __LINE__, NULL, 0);
17122 			tp->t_flags &= ~TF_GPUTINPROG;
17123 		}
17124 	}
17125 	to = &to_holder;
17126 	to->to_flags = 0;
17127 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
17128 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
17129 	cnt = m->m_len / sizeof(struct tcp_ackent);
17130 	counter_u64_add(rack_multi_single_eq, cnt);
17131 	high_seq = tp->snd_una;
17132 	the_win = tp->snd_wnd;
17133 	win_seq = tp->snd_wl1;
17134 	win_upd_ack = tp->snd_wl2;
17135 	cts = tcp_tv_to_usectick(tv);
17136 	ms_cts = tcp_tv_to_mssectick(tv);
17137 	rack->r_ctl.rc_rcvtime = cts;
17138 	segsiz = ctf_fixed_maxseg(tp);
17139 	if ((rack->rc_gp_dyn_mul) &&
17140 	    (rack->use_fixed_rate == 0) &&
17141 	    (rack->rc_always_pace)) {
17142 		/* Check in on probertt */
17143 		rack_check_probe_rtt(rack, cts);
17144 	}
17145 	for (i = 0; i < cnt; i++) {
17146 #ifdef TCP_ACCOUNTING
17147 		ts_val = get_cyclecount();
17148 #endif
17149 		rack_clear_rate_sample(rack);
17150 		ae = ((mtod(m, struct tcp_ackent *)) + i);
17151 		if (ae->flags & TH_FIN)
17152 			rack_log_pacing_delay_calc(rack,
17153 						   0,
17154 						   0,
17155 						   0,
17156 						   rack_get_gp_est(rack), /* delRate */
17157 						   rack_get_lt_bw(rack), /* rttProp */
17158 						   20, __LINE__, NULL, 0);
17159 		/* Setup the window */
17160 		tiwin = ae->win << tp->snd_scale;
17161 		if (tiwin > rack->r_ctl.rc_high_rwnd)
17162 			rack->r_ctl.rc_high_rwnd = tiwin;
17163 		/* figure out the type of ack */
17164 		if (SEQ_LT(ae->ack, high_seq)) {
17165 			/* Case B*/
17166 			ae->ack_val_set = ACK_BEHIND;
17167 		} else if (SEQ_GT(ae->ack, high_seq)) {
17168 			/* Case A */
17169 			ae->ack_val_set = ACK_CUMACK;
17170 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
17171 			/* Case D */
17172 			ae->ack_val_set = ACK_DUPACK;
17173 		} else {
17174 			/* Case C */
17175 			ae->ack_val_set = ACK_RWND;
17176 		}
17177 		if (rack->sack_attack_disable > 0) {
17178 			rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
17179 			rack->r_ctl.ack_during_sd++;
17180 		}
17181 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
17182 		/* Validate timestamp */
17183 		if (ae->flags & HAS_TSTMP) {
17184 			/* Setup for a timestamp */
17185 			to->to_flags = TOF_TS;
17186 			ae->ts_echo -= tp->ts_offset;
17187 			to->to_tsecr = ae->ts_echo;
17188 			to->to_tsval = ae->ts_value;
17189 			/*
17190 			 * If echoed timestamp is later than the current time, fall back to
17191 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
17192 			 * were used when this connection was established.
17193 			 */
17194 			if (TSTMP_GT(ae->ts_echo, ms_cts))
17195 				to->to_tsecr = 0;
17196 			if (tp->ts_recent &&
17197 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
17198 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
17199 #ifdef TCP_ACCOUNTING
17200 					rdstc = get_cyclecount();
17201 					if (rdstc > ts_val) {
17202 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17203 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
17204 						}
17205 					}
17206 #endif
17207 					continue;
17208 				}
17209 			}
17210 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
17211 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
17212 				tp->ts_recent_age = tcp_ts_getticks();
17213 				tp->ts_recent = ae->ts_value;
17214 			}
17215 		} else {
17216 			/* Setup for a no options */
17217 			to->to_flags = 0;
17218 		}
17219 		/* Update the rcv time and perform idle reduction possibly */
17220 		if  (tp->t_idle_reduce &&
17221 		     (tp->snd_max == tp->snd_una) &&
17222 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
17223 			counter_u64_add(rack_input_idle_reduces, 1);
17224 			rack_cc_after_idle(rack, tp);
17225 		}
17226 		tp->t_rcvtime = ticks;
17227 		/* Now what about ECN of a chain of pure ACKs? */
17228 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
17229 			tcp_packets_this_ack(tp, ae->ack),
17230 			ae->codepoint))
17231 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
17232 #ifdef TCP_ACCOUNTING
17233 		/* Count for the specific type of ack in */
17234 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17235 			tp->tcp_cnt_counters[ae->ack_val_set]++;
17236 		}
17237 #endif
17238 		/*
17239 		 * Note how we could move up these in the determination
17240 		 * above, but we don't so that way the timestamp checks (and ECN)
17241 		 * is done first before we do any processing on the ACK.
17242 		 * The non-compressed path through the code has this
17243 		 * weakness (noted by @jtl) that it actually does some
17244 		 * processing before verifying the timestamp information.
17245 		 * We don't take that path here which is why we set
17246 		 * the ack_val_set first, do the timestamp and ecn
17247 		 * processing, and then look at what we have setup.
17248 		 */
17249 		if (ae->ack_val_set == ACK_BEHIND) {
17250 			/*
17251 			 * Case B flag reordering, if window is not closed
17252 			 * or it could be a keep-alive or persists
17253 			 */
17254 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
17255 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
17256 				if (rack->r_ctl.rc_reorder_ts == 0)
17257 					rack->r_ctl.rc_reorder_ts = 1;
17258 			}
17259 		} else if (ae->ack_val_set == ACK_DUPACK) {
17260 			/* Case D */
17261 			rack_strike_dupack(rack, ae->ack);
17262 		} else if (ae->ack_val_set == ACK_RWND) {
17263 			/* Case C */
17264 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
17265 				ts.tv_sec = ae->timestamp / 1000000000;
17266 				ts.tv_nsec = ae->timestamp % 1000000000;
17267 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
17268 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
17269 			} else {
17270 				rack->r_ctl.act_rcv_time = *tv;
17271 			}
17272 			if (rack->forced_ack) {
17273 				rack_handle_probe_response(rack, tiwin,
17274 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
17275 			}
17276 #ifdef TCP_ACCOUNTING
17277 			win_up_req = 1;
17278 #endif
17279 			win_upd_ack = ae->ack;
17280 			win_seq = ae->seq;
17281 			the_win = tiwin;
17282 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
17283 		} else {
17284 			/* Case A */
17285 			if (SEQ_GT(ae->ack, tp->snd_max)) {
17286 				/*
17287 				 * We just send an ack since the incoming
17288 				 * ack is beyond the largest seq we sent.
17289 				 */
17290 				if ((tp->t_flags & TF_ACKNOW) == 0) {
17291 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
17292 					if (tp->t_flags && TF_ACKNOW)
17293 						rack->r_wanted_output = 1;
17294 				}
17295 			} else {
17296 				nsegs++;
17297 				/* If the window changed setup to update */
17298 				if (tiwin != tp->snd_wnd) {
17299 					win_upd_ack = ae->ack;
17300 					win_seq = ae->seq;
17301 					the_win = tiwin;
17302 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
17303 				}
17304 #ifdef TCP_ACCOUNTING
17305 				/* Account for the acks */
17306 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17307 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
17308 				}
17309 #endif
17310 				high_seq = ae->ack;
17311 				/* Setup our act_rcv_time */
17312 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
17313 					ts.tv_sec = ae->timestamp / 1000000000;
17314 					ts.tv_nsec = ae->timestamp % 1000000000;
17315 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
17316 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
17317 				} else {
17318 					rack->r_ctl.act_rcv_time = *tv;
17319 				}
17320 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
17321 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
17322 #ifdef TCP_REQUEST_TRK
17323 				rack_req_check_for_comp(rack, high_seq);
17324 #endif
17325 				if (rack->rc_dsack_round_seen) {
17326 					/* Is the dsack round over? */
17327 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
17328 						/* Yes it is */
17329 						rack->rc_dsack_round_seen = 0;
17330 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
17331 					}
17332 				}
17333 			}
17334 		}
17335 		/* And lets be sure to commit the rtt measurements for this ack */
17336 		tcp_rack_xmit_timer_commit(rack, tp);
17337 #ifdef TCP_ACCOUNTING
17338 		rdstc = get_cyclecount();
17339 		if (rdstc > ts_val) {
17340 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17341 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
17342 				if (ae->ack_val_set == ACK_CUMACK)
17343 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
17344 			}
17345 		}
17346 #endif
17347 	}
17348 #ifdef TCP_ACCOUNTING
17349 	ts_val = get_cyclecount();
17350 #endif
17351 	/* Tend to any collapsed window */
17352 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
17353 		/* The peer collapsed the window */
17354 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
17355 	} else if (rack->rc_has_collapsed)
17356 		rack_un_collapse_window(rack, __LINE__);
17357 	if ((rack->r_collapse_point_valid) &&
17358 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
17359 		rack->r_collapse_point_valid = 0;
17360 	acked_amount = acked = (high_seq - tp->snd_una);
17361 	if (acked) {
17362 		/*
17363 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
17364 		 * causes issues when we are just going app limited. Lets
17365 		 * instead use SEQ_GT <or> where its equal but more data
17366 		 * is outstanding.
17367 		 *
17368 		 * Also make sure we are on the last ack of a series. We
17369 		 * have to have all the ack's processed in queue to know
17370 		 * if there is something left outstanding.
17371 		 *
17372 		 */
17373 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
17374 		    (rack->rc_new_rnd_needed == 0) &&
17375 		    (nxt_pkt == 0)) {
17376 			/*
17377 			 * We have crossed into a new round with
17378 			 * this th_ack value.
17379 			 */
17380 			rack_new_round_setup(tp, rack, high_seq);
17381 		}
17382 		/*
17383 		 * Clear the probe not answered flag
17384 		 * since cum-ack moved forward.
17385 		 */
17386 		rack->probe_not_answered = 0;
17387 		if (rack->sack_attack_disable == 0)
17388 			rack_do_decay(rack);
17389 		if (acked >= segsiz) {
17390 			/*
17391 			 * You only get credit for
17392 			 * MSS and greater (and you get extra
17393 			 * credit for larger cum-ack moves).
17394 			 */
17395 			int ac;
17396 
17397 			ac = acked / segsiz;
17398 			rack->r_ctl.ack_count += ac;
17399 			counter_u64_add(rack_ack_total, ac);
17400 		}
17401 		if (rack->r_ctl.ack_count > 0xfff00000) {
17402 			/*
17403 			 * reduce the number to keep us under
17404 			 * a uint32_t.
17405 			 */
17406 			rack->r_ctl.ack_count /= 2;
17407 			rack->r_ctl.sack_count /= 2;
17408 		}
17409 		if (tp->t_flags & TF_NEEDSYN) {
17410 			/*
17411 			 * T/TCP: Connection was half-synchronized, and our SYN has
17412 			 * been ACK'd (so connection is now fully synchronized).  Go
17413 			 * to non-starred state, increment snd_una for ACK of SYN,
17414 			 * and check if we can do window scaling.
17415 			 */
17416 			tp->t_flags &= ~TF_NEEDSYN;
17417 			tp->snd_una++;
17418 			acked_amount = acked = (high_seq - tp->snd_una);
17419 		}
17420 		if (acked > sbavail(&so->so_snd))
17421 			acked_amount = sbavail(&so->so_snd);
17422 #ifdef TCP_SAD_DETECTION
17423 		/*
17424 		 * We only care on a cum-ack move if we are in a sack-disabled
17425 		 * state. We have already added in to the ack_count, and we never
17426 		 * would disable on a cum-ack move, so we only care to do the
17427 		 * detection if it may "undo" it, i.e. we were in disabled already.
17428 		 */
17429 		if (rack->sack_attack_disable)
17430 			rack_do_detection(tp, rack, acked_amount, segsiz);
17431 #endif
17432 		if (IN_FASTRECOVERY(tp->t_flags) &&
17433 		    (rack->rack_no_prr == 0))
17434 			rack_update_prr(tp, rack, acked_amount, high_seq);
17435 		if (IN_RECOVERY(tp->t_flags)) {
17436 			if (SEQ_LT(high_seq, tp->snd_recover) &&
17437 			    (SEQ_LT(high_seq, tp->snd_max))) {
17438 				tcp_rack_partialack(tp);
17439 			} else {
17440 				rack_post_recovery(tp, high_seq);
17441 				post_recovery = 1;
17442 			}
17443 		}  else if ((rack->rto_from_rec == 1) &&
17444 			    SEQ_GEQ(high_seq, tp->snd_recover)) {
17445 			/*
17446 			 * We were in recovery, hit a rxt timeout
17447 			 * and never re-entered recovery. The timeout(s)
17448 			 * made up all the lost data. In such a case
17449 			 * we need to clear the rto_from_rec flag.
17450 			 */
17451 			rack->rto_from_rec = 0;
17452 		}
17453 		/* Handle the rack-log-ack part (sendmap) */
17454 		if ((sbused(&so->so_snd) == 0) &&
17455 		    (acked > acked_amount) &&
17456 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
17457 		    (tp->t_flags & TF_SENTFIN)) {
17458 			/*
17459 			 * We must be sure our fin
17460 			 * was sent and acked (we can be
17461 			 * in FIN_WAIT_1 without having
17462 			 * sent the fin).
17463 			 */
17464 			ourfinisacked = 1;
17465 			/*
17466 			 * Lets make sure snd_una is updated
17467 			 * since most likely acked_amount = 0 (it
17468 			 * should be).
17469 			 */
17470 			tp->snd_una = high_seq;
17471 		}
17472 		/* Did we make a RTO error? */
17473 		if ((tp->t_flags & TF_PREVVALID) &&
17474 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
17475 			tp->t_flags &= ~TF_PREVVALID;
17476 			if (tp->t_rxtshift == 1 &&
17477 			    (int)(ticks - tp->t_badrxtwin) < 0)
17478 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
17479 		}
17480 		/* Handle the data in the socket buffer */
17481 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
17482 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
17483 		if (acked_amount > 0) {
17484 			uint32_t p_cwnd;
17485 			struct mbuf *mfree;
17486 
17487 			if (post_recovery) {
17488 				/*
17489 				 * Grab the segsiz, multiply by 2 and add the snd_cwnd
17490 				 * that is the max the CC should add if we are exiting
17491 				 * recovery and doing a late add.
17492 				 */
17493 				p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
17494 				p_cwnd <<= 1;
17495 				p_cwnd += tp->snd_cwnd;
17496 			}
17497 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
17498 			if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
17499 				/* Must be non-newreno (cubic) getting too ahead of itself */
17500 				tp->snd_cwnd = p_cwnd;
17501 			}
17502 			SOCKBUF_LOCK(&so->so_snd);
17503 			mfree = sbcut_locked(&so->so_snd, acked_amount);
17504 			tp->snd_una = high_seq;
17505 			/* Note we want to hold the sb lock through the sendmap adjust */
17506 			rack_adjust_sendmap_head(rack, &so->so_snd);
17507 			/* Wake up the socket if we have room to write more */
17508 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
17509 			sowwakeup_locked(so);
17510 			m_freem(mfree);
17511 		}
17512 		/* update progress */
17513 		tp->t_acktime = ticks;
17514 		rack_log_progress_event(rack, tp, tp->t_acktime,
17515 					PROGRESS_UPDATE, __LINE__);
17516 		/* Clear out shifts and such */
17517 		tp->t_rxtshift = 0;
17518 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
17519 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
17520 		rack->rc_tlp_in_progress = 0;
17521 		rack->r_ctl.rc_tlp_cnt_out = 0;
17522 		/* Send recover and snd_nxt must be dragged along */
17523 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
17524 			tp->snd_recover = tp->snd_una;
17525 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17526 			tp->snd_nxt = tp->snd_max;
17527 		/*
17528 		 * If the RXT timer is running we want to
17529 		 * stop it, so we can restart a TLP (or new RXT).
17530 		 */
17531 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
17532 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
17533 		tp->snd_wl2 = high_seq;
17534 		tp->t_dupacks = 0;
17535 		if (under_pacing &&
17536 		    (rack->use_fixed_rate == 0) &&
17537 		    (rack->in_probe_rtt == 0) &&
17538 		    rack->rc_gp_dyn_mul &&
17539 		    rack->rc_always_pace) {
17540 			/* Check if we are dragging bottom */
17541 			rack_check_bottom_drag(tp, rack, so);
17542 		}
17543 		if (tp->snd_una == tp->snd_max) {
17544 			tp->t_flags &= ~TF_PREVVALID;
17545 			rack->r_ctl.retran_during_recovery = 0;
17546 			rack->rc_suspicious = 0;
17547 			rack->r_ctl.dsack_byte_cnt = 0;
17548 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
17549 			if (rack->r_ctl.rc_went_idle_time == 0)
17550 				rack->r_ctl.rc_went_idle_time = 1;
17551 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
17552 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
17553 				tp->t_acktime = 0;
17554 			/* Set so we might enter persists... */
17555 			rack->r_wanted_output = 1;
17556 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
17557 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
17558 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
17559 			    (sbavail(&so->so_snd) == 0) &&
17560 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
17561 				/*
17562 				 * The socket was gone and the
17563 				 * peer sent data (not now in the past), time to
17564 				 * reset him.
17565 				 */
17566 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
17567 				/* tcp_close will kill the inp pre-log the Reset */
17568 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
17569 #ifdef TCP_ACCOUNTING
17570 				rdstc = get_cyclecount();
17571 				if (rdstc > ts_val) {
17572 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17573 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17574 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17575 					}
17576 				}
17577 #endif
17578 				m_freem(m);
17579 				tp = tcp_close(tp);
17580 				if (tp == NULL) {
17581 #ifdef TCP_ACCOUNTING
17582 					sched_unpin();
17583 #endif
17584 					return (1);
17585 				}
17586 				/*
17587 				 * We would normally do drop-with-reset which would
17588 				 * send back a reset. We can't since we don't have
17589 				 * all the needed bits. Instead lets arrange for
17590 				 * a call to tcp_output(). That way since we
17591 				 * are in the closed state we will generate a reset.
17592 				 *
17593 				 * Note if tcp_accounting is on we don't unpin since
17594 				 * we do that after the goto label.
17595 				 */
17596 				goto send_out_a_rst;
17597 			}
17598 			if ((sbused(&so->so_snd) == 0) &&
17599 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
17600 			    (tp->t_flags & TF_SENTFIN)) {
17601 				/*
17602 				 * If we can't receive any more data, then closing user can
17603 				 * proceed. Starting the timer is contrary to the
17604 				 * specification, but if we don't get a FIN we'll hang
17605 				 * forever.
17606 				 *
17607 				 */
17608 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
17609 					soisdisconnected(so);
17610 					tcp_timer_activate(tp, TT_2MSL,
17611 							   (tcp_fast_finwait2_recycle ?
17612 							    tcp_finwait2_timeout :
17613 							    TP_MAXIDLE(tp)));
17614 				}
17615 				if (ourfinisacked == 0) {
17616 					/*
17617 					 * We don't change to fin-wait-2 if we have our fin acked
17618 					 * which means we are probably in TCPS_CLOSING.
17619 					 */
17620 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
17621 				}
17622 			}
17623 		}
17624 		/* Wake up the socket if we have room to write more */
17625 		if (sbavail(&so->so_snd)) {
17626 			rack->r_wanted_output = 1;
17627 			if (ctf_progress_timeout_check(tp, true)) {
17628 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
17629 							tp, tick, PROGRESS_DROP, __LINE__);
17630 				/*
17631 				 * We cheat here and don't send a RST, we should send one
17632 				 * when the pacer drops the connection.
17633 				 */
17634 #ifdef TCP_ACCOUNTING
17635 				rdstc = get_cyclecount();
17636 				if (rdstc > ts_val) {
17637 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17638 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17639 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17640 					}
17641 				}
17642 				sched_unpin();
17643 #endif
17644 				(void)tcp_drop(tp, ETIMEDOUT);
17645 				m_freem(m);
17646 				return (1);
17647 			}
17648 		}
17649 		if (ourfinisacked) {
17650 			switch(tp->t_state) {
17651 			case TCPS_CLOSING:
17652 #ifdef TCP_ACCOUNTING
17653 				rdstc = get_cyclecount();
17654 				if (rdstc > ts_val) {
17655 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17656 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17657 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17658 					}
17659 				}
17660 				sched_unpin();
17661 #endif
17662 				tcp_twstart(tp);
17663 				m_freem(m);
17664 				return (1);
17665 				break;
17666 			case TCPS_LAST_ACK:
17667 #ifdef TCP_ACCOUNTING
17668 				rdstc = get_cyclecount();
17669 				if (rdstc > ts_val) {
17670 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17671 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17672 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17673 					}
17674 				}
17675 				sched_unpin();
17676 #endif
17677 				tp = tcp_close(tp);
17678 				ctf_do_drop(m, tp);
17679 				return (1);
17680 				break;
17681 			case TCPS_FIN_WAIT_1:
17682 #ifdef TCP_ACCOUNTING
17683 				rdstc = get_cyclecount();
17684 				if (rdstc > ts_val) {
17685 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17686 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17687 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17688 					}
17689 				}
17690 #endif
17691 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
17692 					soisdisconnected(so);
17693 					tcp_timer_activate(tp, TT_2MSL,
17694 							   (tcp_fast_finwait2_recycle ?
17695 							    tcp_finwait2_timeout :
17696 							    TP_MAXIDLE(tp)));
17697 				}
17698 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
17699 				break;
17700 			default:
17701 				break;
17702 			}
17703 		}
17704 		if (rack->r_fast_output) {
17705 			/*
17706 			 * We re doing fast output.. can we expand that?
17707 			 */
17708 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
17709 		}
17710 #ifdef TCP_ACCOUNTING
17711 		rdstc = get_cyclecount();
17712 		if (rdstc > ts_val) {
17713 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17714 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
17715 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
17716 			}
17717 		}
17718 
17719 	} else if (win_up_req) {
17720 		rdstc = get_cyclecount();
17721 		if (rdstc > ts_val) {
17722 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17723 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
17724 			}
17725 		}
17726 #endif
17727 	}
17728 	/* Now is there a next packet, if so we are done */
17729 	m_freem(m);
17730 	did_out = 0;
17731 	if (nxt_pkt) {
17732 #ifdef TCP_ACCOUNTING
17733 		sched_unpin();
17734 #endif
17735 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
17736 		return (0);
17737 	}
17738 	rack_handle_might_revert(tp, rack);
17739 	ctf_calc_rwin(so, tp);
17740 	if ((rack->r_wanted_output != 0) ||
17741 	    (rack->r_fast_output != 0) ||
17742 	    (tp->t_flags & TF_ACKNOW )) {
17743 	send_out_a_rst:
17744 		if (tcp_output(tp) < 0) {
17745 #ifdef TCP_ACCOUNTING
17746 			sched_unpin();
17747 #endif
17748 			return (1);
17749 		}
17750 		did_out = 1;
17751 	}
17752 	if (tp->t_flags2 & TF2_HPTS_CALLS)
17753 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
17754 	rack_free_trim(rack);
17755 #ifdef TCP_ACCOUNTING
17756 	sched_unpin();
17757 #endif
17758 	rack_timer_audit(tp, rack, &so->so_snd);
17759 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
17760 	return (0);
17761 }
17762 
17763 #define	TCP_LRO_TS_OPTION \
17764     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
17765 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
17766 
17767 static int
17768 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17769     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
17770     struct timeval *tv)
17771 {
17772 	struct inpcb *inp = tptoinpcb(tp);
17773 	struct socket *so = tptosocket(tp);
17774 #ifdef TCP_ACCOUNTING
17775 	uint64_t ts_val;
17776 #endif
17777 	int32_t thflags, retval, did_out = 0;
17778 	int32_t way_out = 0;
17779 	/*
17780 	 * cts - is the current time from tv (caller gets ts) in microseconds.
17781 	 * ms_cts - is the current time from tv in milliseconds.
17782 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
17783 	 */
17784 	uint32_t cts, us_cts, ms_cts;
17785 	uint32_t tiwin;
17786 	struct timespec ts;
17787 	struct tcpopt to;
17788 	struct tcp_rack *rack;
17789 	struct rack_sendmap *rsm;
17790 	int32_t prev_state = 0;
17791 	int no_output = 0;
17792 	int slot_remaining = 0;
17793 #ifdef TCP_ACCOUNTING
17794 	int ack_val_set = 0xf;
17795 #endif
17796 	int nsegs;
17797 
17798 	NET_EPOCH_ASSERT();
17799 	INP_WLOCK_ASSERT(inp);
17800 
17801 	/*
17802 	 * tv passed from common code is from either M_TSTMP_LRO or
17803 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
17804 	 */
17805 	rack = (struct tcp_rack *)tp->t_fb_ptr;
17806 	if (rack->rack_deferred_inited == 0) {
17807 		/*
17808 		 * If we are the connecting socket we will
17809 		 * hit rack_init() when no sequence numbers
17810 		 * are setup. This makes it so we must defer
17811 		 * some initialization. Call that now.
17812 		 */
17813 		rack_deferred_init(tp, rack);
17814 	}
17815 	/*
17816 	 * Check to see if we need to skip any output plans. This
17817 	 * can happen in the non-LRO path where we are pacing and
17818 	 * must process the ack coming in but need to defer sending
17819 	 * anything becase a pacing timer is running.
17820 	 */
17821 	us_cts = tcp_tv_to_usectick(tv);
17822 	if (m->m_flags & M_ACKCMP) {
17823 		/*
17824 		 * All compressed ack's are ack's by definition so
17825 		 * remove any ack required flag and then do the processing.
17826 		 */
17827 		rack->rc_ack_required = 0;
17828 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
17829 	}
17830 	thflags = tcp_get_flags(th);
17831 	if ((rack->rc_always_pace == 1) &&
17832 	    (rack->rc_ack_can_sendout_data == 0) &&
17833 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
17834 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
17835 		/*
17836 		 * Ok conditions are right for queuing the packets
17837 		 * but we do have to check the flags in the inp, it
17838 		 * could be, if a sack is present, we want to be awoken and
17839 		 * so should process the packets.
17840 		 */
17841 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
17842 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
17843 			no_output = 1;
17844 		} else {
17845 			/*
17846 			 * If there is no options, or just a
17847 			 * timestamp option, we will want to queue
17848 			 * the packets. This is the same that LRO does
17849 			 * and will need to change with accurate ECN.
17850 			 */
17851 			uint32_t *ts_ptr;
17852 			int optlen;
17853 
17854 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
17855 			ts_ptr = (uint32_t *)(th + 1);
17856 			if ((optlen == 0) ||
17857 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
17858 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
17859 				no_output = 1;
17860 		}
17861 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
17862 			/*
17863 			 * It is unrealistic to think we can pace in less than
17864 			 * the minimum granularity of the pacer (def:250usec). So
17865 			 * if we have less than that time remaining we should go
17866 			 * ahead and allow output to be "early". We will attempt to
17867 			 * make up for it in any pacing time we try to apply on
17868 			 * the outbound packet.
17869 			 */
17870 			no_output = 0;
17871 		}
17872 	}
17873 	/*
17874 	 * If there is a RST or FIN lets dump out the bw
17875 	 * with a FIN the connection may go on but we
17876 	 * may not.
17877 	 */
17878 	if ((thflags & TH_FIN) || (thflags & TH_RST))
17879 		rack_log_pacing_delay_calc(rack,
17880 					   rack->r_ctl.gp_bw,
17881 					   0,
17882 					   0,
17883 					   rack_get_gp_est(rack), /* delRate */
17884 					   rack_get_lt_bw(rack), /* rttProp */
17885 					   20, __LINE__, NULL, 0);
17886 	if (m->m_flags & M_ACKCMP) {
17887 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
17888 	}
17889 	cts = tcp_tv_to_usectick(tv);
17890 	ms_cts =  tcp_tv_to_mssectick(tv);
17891 	nsegs = m->m_pkthdr.lro_nsegs;
17892 	counter_u64_add(rack_proc_non_comp_ack, 1);
17893 #ifdef TCP_ACCOUNTING
17894 	sched_pin();
17895 	if (thflags & TH_ACK)
17896 		ts_val = get_cyclecount();
17897 #endif
17898 	if ((m->m_flags & M_TSTMP) ||
17899 	    (m->m_flags & M_TSTMP_LRO)) {
17900 		mbuf_tstmp2timespec(m, &ts);
17901 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
17902 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
17903 	} else
17904 		rack->r_ctl.act_rcv_time = *tv;
17905 	kern_prefetch(rack, &prev_state);
17906 	prev_state = 0;
17907 	/*
17908 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
17909 	 * the scale is zero.
17910 	 */
17911 	tiwin = th->th_win << tp->snd_scale;
17912 #ifdef TCP_ACCOUNTING
17913 	if (thflags & TH_ACK) {
17914 		/*
17915 		 * We have a tradeoff here. We can either do what we are
17916 		 * doing i.e. pinning to this CPU and then doing the accounting
17917 		 * <or> we could do a critical enter, setup the rdtsc and cpu
17918 		 * as in below, and then validate we are on the same CPU on
17919 		 * exit. I have choosen to not do the critical enter since
17920 		 * that often will gain you a context switch, and instead lock
17921 		 * us (line above this if) to the same CPU with sched_pin(). This
17922 		 * means we may be context switched out for a higher priority
17923 		 * interupt but we won't be moved to another CPU.
17924 		 *
17925 		 * If this occurs (which it won't very often since we most likely
17926 		 * are running this code in interupt context and only a higher
17927 		 * priority will bump us ... clock?) we will falsely add in
17928 		 * to the time the interupt processing time plus the ack processing
17929 		 * time. This is ok since its a rare event.
17930 		 */
17931 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
17932 						    ctf_fixed_maxseg(tp));
17933 	}
17934 #endif
17935 	/*
17936 	 * Parse options on any incoming segment.
17937 	 */
17938 	memset(&to, 0, sizeof(to));
17939 	tcp_dooptions(&to, (u_char *)(th + 1),
17940 	    (th->th_off << 2) - sizeof(struct tcphdr),
17941 	    (thflags & TH_SYN) ? TO_SYN : 0);
17942 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
17943 	    __func__));
17944 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
17945 	    __func__));
17946 	if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
17947 		/*
17948 		 * We don't look at sack's from the
17949 		 * peer because the MSS is too small which
17950 		 * can subject us to an attack.
17951 		 */
17952 		to.to_flags &= ~TOF_SACK;
17953 	}
17954 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
17955 	    (tp->t_flags & TF_GPUTINPROG)) {
17956 		/*
17957 		 * We have a goodput in progress
17958 		 * and we have entered a late state.
17959 		 * Do we have enough data in the sb
17960 		 * to handle the GPUT request?
17961 		 */
17962 		uint32_t bytes;
17963 
17964 		bytes = tp->gput_ack - tp->gput_seq;
17965 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
17966 			bytes += tp->gput_seq - tp->snd_una;
17967 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
17968 			/*
17969 			 * There are not enough bytes in the socket
17970 			 * buffer that have been sent to cover this
17971 			 * measurement. Cancel it.
17972 			 */
17973 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17974 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
17975 						   tp->gput_seq,
17976 						   0, 0, 18, __LINE__, NULL, 0);
17977 			tp->t_flags &= ~TF_GPUTINPROG;
17978 		}
17979 	}
17980 	if (tcp_bblogging_on(rack->rc_tp)) {
17981 		union tcp_log_stackspecific log;
17982 		struct timeval ltv;
17983 #ifdef TCP_REQUEST_TRK
17984 		struct tcp_sendfile_track *tcp_req;
17985 
17986 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
17987 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
17988 		} else {
17989 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
17990 		}
17991 #endif
17992 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17993 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
17994 		if (rack->rack_no_prr == 0)
17995 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17996 		else
17997 			log.u_bbr.flex1 = 0;
17998 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
17999 		log.u_bbr.use_lt_bw <<= 1;
18000 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
18001 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
18002 		log.u_bbr.bbr_state = rack->rc_free_cnt;
18003 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18004 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
18005 		log.u_bbr.flex3 = m->m_flags;
18006 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
18007 		log.u_bbr.lost = thflags;
18008 		log.u_bbr.pacing_gain = 0x1;
18009 #ifdef TCP_ACCOUNTING
18010 		log.u_bbr.cwnd_gain = ack_val_set;
18011 #endif
18012 		log.u_bbr.flex7 = 2;
18013 		if (m->m_flags & M_TSTMP) {
18014 			/* Record the hardware timestamp if present */
18015 			mbuf_tstmp2timespec(m, &ts);
18016 			ltv.tv_sec = ts.tv_sec;
18017 			ltv.tv_usec = ts.tv_nsec / 1000;
18018 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
18019 		} else if (m->m_flags & M_TSTMP_LRO) {
18020 			/* Record the LRO the arrival timestamp */
18021 			mbuf_tstmp2timespec(m, &ts);
18022 			ltv.tv_sec = ts.tv_sec;
18023 			ltv.tv_usec = ts.tv_nsec / 1000;
18024 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
18025 		}
18026 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
18027 		/* Log the rcv time */
18028 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
18029 #ifdef TCP_REQUEST_TRK
18030 		log.u_bbr.applimited = tp->t_tcpreq_closed;
18031 		log.u_bbr.applimited <<= 8;
18032 		log.u_bbr.applimited |= tp->t_tcpreq_open;
18033 		log.u_bbr.applimited <<= 8;
18034 		log.u_bbr.applimited |= tp->t_tcpreq_req;
18035 		if (tcp_req) {
18036 			/* Copy out any client req info */
18037 			/* seconds */
18038 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
18039 			/* useconds */
18040 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
18041 			log.u_bbr.rttProp = tcp_req->timestamp;
18042 			log.u_bbr.cur_del_rate = tcp_req->start;
18043 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
18044 				log.u_bbr.flex8 |= 1;
18045 			} else {
18046 				log.u_bbr.flex8 |= 2;
18047 				log.u_bbr.bw_inuse = tcp_req->end;
18048 			}
18049 			log.u_bbr.flex6 = tcp_req->start_seq;
18050 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
18051 				log.u_bbr.flex8 |= 4;
18052 				log.u_bbr.epoch = tcp_req->end_seq;
18053 			}
18054 		}
18055 #endif
18056 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
18057 		    tlen, &log, true, &ltv);
18058 	}
18059 	/* Remove ack required flag if set, we have one  */
18060 	if (thflags & TH_ACK)
18061 		rack->rc_ack_required = 0;
18062 	if (rack->sack_attack_disable > 0) {
18063 		rack->r_ctl.ack_during_sd++;
18064 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
18065 	}
18066 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
18067 		way_out = 4;
18068 		retval = 0;
18069 		m_freem(m);
18070 		goto done_with_input;
18071 	}
18072 	/*
18073 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
18074 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
18075 	 */
18076 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
18077 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
18078 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
18079 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
18080 #ifdef TCP_ACCOUNTING
18081 		sched_unpin();
18082 #endif
18083 		return (1);
18084 	}
18085 	/*
18086 	 * If timestamps were negotiated during SYN/ACK and a
18087 	 * segment without a timestamp is received, silently drop
18088 	 * the segment, unless it is a RST segment or missing timestamps are
18089 	 * tolerated.
18090 	 * See section 3.2 of RFC 7323.
18091 	 */
18092 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
18093 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
18094 		way_out = 5;
18095 		retval = 0;
18096 		m_freem(m);
18097 		goto done_with_input;
18098 	}
18099 	/*
18100 	 * Segment received on connection. Reset idle time and keep-alive
18101 	 * timer. XXX: This should be done after segment validation to
18102 	 * ignore broken/spoofed segs.
18103 	 */
18104 	if  (tp->t_idle_reduce &&
18105 	     (tp->snd_max == tp->snd_una) &&
18106 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
18107 		counter_u64_add(rack_input_idle_reduces, 1);
18108 		rack_cc_after_idle(rack, tp);
18109 	}
18110 	tp->t_rcvtime = ticks;
18111 #ifdef STATS
18112 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
18113 #endif
18114 	if (tiwin > rack->r_ctl.rc_high_rwnd)
18115 		rack->r_ctl.rc_high_rwnd = tiwin;
18116 	/*
18117 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
18118 	 * this to occur after we've validated the segment.
18119 	 */
18120 	if (tcp_ecn_input_segment(tp, thflags, tlen,
18121 	    tcp_packets_this_ack(tp, th->th_ack),
18122 	    iptos))
18123 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
18124 
18125 	/*
18126 	 * If echoed timestamp is later than the current time, fall back to
18127 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
18128 	 * were used when this connection was established.
18129 	 */
18130 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
18131 		to.to_tsecr -= tp->ts_offset;
18132 		if (TSTMP_GT(to.to_tsecr, ms_cts))
18133 			to.to_tsecr = 0;
18134 	}
18135 	if ((rack->r_rcvpath_rtt_up == 1) &&
18136 	    (to.to_flags & TOF_TS) &&
18137 	    (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
18138 		uint32_t rtt = 0;
18139 
18140 		/*
18141 		 * We are receiving only and thus not sending
18142 		 * data to do an RTT. We set a flag when we first
18143 		 * sent this TS to the peer. We now have it back
18144 		 * and have an RTT to share. We log it as a conf
18145 		 * 4, we are not so sure about it.. since we
18146 		 * may have lost an ack.
18147 		 */
18148 		if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
18149 		    rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
18150 		rack->r_rcvpath_rtt_up = 0;
18151 		/* Submit and commit the timer */
18152 		if (rtt > 0) {
18153 			tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
18154 			tcp_rack_xmit_timer_commit(rack, tp);
18155 		}
18156 	}
18157 	/*
18158 	 * If its the first time in we need to take care of options and
18159 	 * verify we can do SACK for rack!
18160 	 */
18161 	if (rack->r_state == 0) {
18162 		/* Should be init'd by rack_init() */
18163 		KASSERT(rack->rc_inp != NULL,
18164 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
18165 		if (rack->rc_inp == NULL) {
18166 			rack->rc_inp = inp;
18167 		}
18168 
18169 		/*
18170 		 * Process options only when we get SYN/ACK back. The SYN
18171 		 * case for incoming connections is handled in tcp_syncache.
18172 		 * According to RFC1323 the window field in a SYN (i.e., a
18173 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
18174 		 * this is traditional behavior, may need to be cleaned up.
18175 		 */
18176 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
18177 			/* Handle parallel SYN for ECN */
18178 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
18179 			if ((to.to_flags & TOF_SCALE) &&
18180 			    (tp->t_flags & TF_REQ_SCALE)) {
18181 				tp->t_flags |= TF_RCVD_SCALE;
18182 				tp->snd_scale = to.to_wscale;
18183 			} else
18184 				tp->t_flags &= ~TF_REQ_SCALE;
18185 			/*
18186 			 * Initial send window.  It will be updated with the
18187 			 * next incoming segment to the scaled value.
18188 			 */
18189 			tp->snd_wnd = th->th_win;
18190 			rack_validate_fo_sendwin_up(tp, rack);
18191 			if ((to.to_flags & TOF_TS) &&
18192 			    (tp->t_flags & TF_REQ_TSTMP)) {
18193 				tp->t_flags |= TF_RCVD_TSTMP;
18194 				tp->ts_recent = to.to_tsval;
18195 				tp->ts_recent_age = cts;
18196 			} else
18197 				tp->t_flags &= ~TF_REQ_TSTMP;
18198 			if (to.to_flags & TOF_MSS) {
18199 				tcp_mss(tp, to.to_mss);
18200 			}
18201 			if ((tp->t_flags & TF_SACK_PERMIT) &&
18202 			    (to.to_flags & TOF_SACKPERM) == 0)
18203 				tp->t_flags &= ~TF_SACK_PERMIT;
18204 			if (tp->t_flags & TF_FASTOPEN) {
18205 				if (to.to_flags & TOF_FASTOPEN) {
18206 					uint16_t mss;
18207 
18208 					if (to.to_flags & TOF_MSS)
18209 						mss = to.to_mss;
18210 					else
18211 						if ((inp->inp_vflag & INP_IPV6) != 0)
18212 							mss = TCP6_MSS;
18213 						else
18214 							mss = TCP_MSS;
18215 					tcp_fastopen_update_cache(tp, mss,
18216 					    to.to_tfo_len, to.to_tfo_cookie);
18217 				} else
18218 					tcp_fastopen_disable_path(tp);
18219 			}
18220 		}
18221 		/*
18222 		 * At this point we are at the initial call. Here we decide
18223 		 * if we are doing RACK or not. We do this by seeing if
18224 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
18225 		 * The code now does do dup-ack counting so if you don't
18226 		 * switch back you won't get rack & TLP, but you will still
18227 		 * get this stack.
18228 		 */
18229 
18230 		if ((rack_sack_not_required == 0) &&
18231 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
18232 			tcp_switch_back_to_default(tp);
18233 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
18234 			    tlen, iptos);
18235 #ifdef TCP_ACCOUNTING
18236 			sched_unpin();
18237 #endif
18238 			return (1);
18239 		}
18240 		tcp_set_hpts(tp);
18241 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
18242 	}
18243 	if (thflags & TH_FIN)
18244 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
18245 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
18246 	if ((rack->rc_gp_dyn_mul) &&
18247 	    (rack->use_fixed_rate == 0) &&
18248 	    (rack->rc_always_pace)) {
18249 		/* Check in on probertt */
18250 		rack_check_probe_rtt(rack, cts);
18251 	}
18252 	rack_clear_rate_sample(rack);
18253 	if ((rack->forced_ack) &&
18254 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
18255 		rack_handle_probe_response(rack, tiwin, us_cts);
18256 	}
18257 	/*
18258 	 * This is the one exception case where we set the rack state
18259 	 * always. All other times (timers etc) we must have a rack-state
18260 	 * set (so we assure we have done the checks above for SACK).
18261 	 */
18262 	rack->r_ctl.rc_rcvtime = cts;
18263 	if (rack->r_state != tp->t_state)
18264 		rack_set_state(tp, rack);
18265 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
18266 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
18267 		kern_prefetch(rsm, &prev_state);
18268 	prev_state = rack->r_state;
18269 	if ((thflags & TH_RST) &&
18270 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
18271 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
18272 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
18273 		/* The connection will be killed by a reset check the tracepoint */
18274 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
18275 	}
18276 	retval = (*rack->r_substate) (m, th, so,
18277 	    tp, &to, drop_hdrlen,
18278 	    tlen, tiwin, thflags, nxt_pkt, iptos);
18279 	if (retval == 0) {
18280 		/*
18281 		 * If retval is 1 the tcb is unlocked and most likely the tp
18282 		 * is gone.
18283 		 */
18284 		INP_WLOCK_ASSERT(inp);
18285 		if ((rack->rc_gp_dyn_mul) &&
18286 		    (rack->rc_always_pace) &&
18287 		    (rack->use_fixed_rate == 0) &&
18288 		    rack->in_probe_rtt &&
18289 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
18290 			/*
18291 			 * If we are going for target, lets recheck before
18292 			 * we output.
18293 			 */
18294 			rack_check_probe_rtt(rack, cts);
18295 		}
18296 		if (rack->set_pacing_done_a_iw == 0) {
18297 			/* How much has been acked? */
18298 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
18299 				/* We have enough to set in the pacing segment size */
18300 				rack->set_pacing_done_a_iw = 1;
18301 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
18302 			}
18303 		}
18304 		tcp_rack_xmit_timer_commit(rack, tp);
18305 #ifdef TCP_ACCOUNTING
18306 		/*
18307 		 * If we set the ack_val_se to what ack processing we are doing
18308 		 * we also want to track how many cycles we burned. Note
18309 		 * the bits after tcp_output we let be "free". This is because
18310 		 * we are also tracking the tcp_output times as well. Note the
18311 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
18312 		 * 0xf cannot be returned and is what we initialize it too to
18313 		 * indicate we are not doing the tabulations.
18314 		 */
18315 		if (ack_val_set != 0xf) {
18316 			uint64_t crtsc;
18317 
18318 			crtsc = get_cyclecount();
18319 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18320 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
18321 			}
18322 		}
18323 #endif
18324 		if ((nxt_pkt == 0) && (no_output == 0)) {
18325 			if ((rack->r_wanted_output != 0) ||
18326 			    (tp->t_flags & TF_ACKNOW) ||
18327 			    (rack->r_fast_output != 0)) {
18328 
18329 do_output_now:
18330 				if (tcp_output(tp) < 0) {
18331 #ifdef TCP_ACCOUNTING
18332 					sched_unpin();
18333 #endif
18334 					return (1);
18335 				}
18336 				did_out = 1;
18337 			}
18338 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
18339 			rack_free_trim(rack);
18340 		} else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
18341 			goto do_output_now;
18342 		} else if ((no_output == 1) &&
18343 			   (nxt_pkt == 0)  &&
18344 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
18345 			/*
18346 			 * We are not in hpts and we had a pacing timer up. Use
18347 			 * the remaining time (slot_remaining) to restart the timer.
18348 			 */
18349 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
18350 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
18351 			rack_free_trim(rack);
18352 		}
18353 		/* Clear the flag, it may have been cleared by output but we may not have  */
18354 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
18355 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
18356 		/*
18357 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
18358 		 * causes issues when we are just going app limited. Lets
18359 		 * instead use SEQ_GT <or> where its equal but more data
18360 		 * is outstanding.
18361 		 *
18362 		 * Also make sure we are on the last ack of a series. We
18363 		 * have to have all the ack's processed in queue to know
18364 		 * if there is something left outstanding.
18365 		 */
18366 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
18367 		    (rack->rc_new_rnd_needed == 0) &&
18368 		    (nxt_pkt == 0)) {
18369 			/*
18370 			 * We have crossed into a new round with
18371 			 * the new snd_unae.
18372 			 */
18373 			rack_new_round_setup(tp, rack, tp->snd_una);
18374 		}
18375 		if ((nxt_pkt == 0) &&
18376 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
18377 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
18378 		     (tp->t_flags & TF_DELACK) ||
18379 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
18380 		      (tp->t_state <= TCPS_CLOSING)))) {
18381 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
18382 			if ((tp->snd_max == tp->snd_una) &&
18383 			    ((tp->t_flags & TF_DELACK) == 0) &&
18384 			    (tcp_in_hpts(rack->rc_tp)) &&
18385 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
18386 				/* keep alive not needed if we are hptsi output yet */
18387 				;
18388 			} else {
18389 				int late = 0;
18390 				if (tcp_in_hpts(tp)) {
18391 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
18392 						us_cts = tcp_get_usecs(NULL);
18393 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
18394 							rack->r_early = 1;
18395 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
18396 						} else
18397 							late = 1;
18398 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
18399 					}
18400 					tcp_hpts_remove(tp);
18401 				}
18402 				if (late && (did_out == 0)) {
18403 					/*
18404 					 * We are late in the sending
18405 					 * and we did not call the output
18406 					 * (this probably should not happen).
18407 					 */
18408 					goto do_output_now;
18409 				}
18410 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
18411 			}
18412 			way_out = 1;
18413 		} else if (nxt_pkt == 0) {
18414 			/* Do we have the correct timer running? */
18415 			rack_timer_audit(tp, rack, &so->so_snd);
18416 			way_out = 2;
18417 		}
18418 	done_with_input:
18419 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
18420 		if (did_out)
18421 			rack->r_wanted_output = 0;
18422 	}
18423 
18424 #ifdef TCP_ACCOUNTING
18425 	sched_unpin();
18426 #endif
18427 	return (retval);
18428 }
18429 
18430 static void
18431 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
18432     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
18433 {
18434 	struct timeval tv;
18435 
18436 	/* First lets see if we have old packets */
18437 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
18438 		if (ctf_do_queued_segments(tp, 1)) {
18439 			m_freem(m);
18440 			return;
18441 		}
18442 	}
18443 	if (m->m_flags & M_TSTMP_LRO) {
18444 		mbuf_tstmp2timeval(m, &tv);
18445 	} else {
18446 		/* Should not be should we kassert instead? */
18447 		tcp_get_usecs(&tv);
18448 	}
18449 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
18450 	    &tv) == 0) {
18451 		INP_WUNLOCK(tptoinpcb(tp));
18452 	}
18453 }
18454 
18455 struct rack_sendmap *
18456 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
18457 {
18458 	struct rack_sendmap *rsm = NULL;
18459 	int32_t idx;
18460 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
18461 	int no_sack = 0;
18462 
18463 	/* Return the next guy to be re-transmitted */
18464 	if (tqhash_empty(rack->r_ctl.tqh)) {
18465 		return (NULL);
18466 	}
18467 	if (tp->t_flags & TF_SENTFIN) {
18468 		/* retran the end FIN? */
18469 		return (NULL);
18470 	}
18471 	/* ok lets look at this one */
18472 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
18473 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
18474 		return (rsm);
18475 	}
18476 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
18477 		goto check_it;
18478 	}
18479 	rsm = rack_find_lowest_rsm(rack);
18480 	if (rsm == NULL) {
18481 		return (NULL);
18482 	}
18483 check_it:
18484 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) ||
18485 	    (rack->sack_attack_disable > 0)) {
18486 		no_sack = 1;
18487 	}
18488 	if ((no_sack > 0) &&
18489 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
18490 		/*
18491 		 * No sack so we automatically do the 3 strikes and
18492 		 * retransmit (no rack timer would be started).
18493 		 */
18494 		return (rsm);
18495 	}
18496 	if (rsm->r_flags & RACK_ACKED) {
18497 		return (NULL);
18498 	}
18499 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
18500 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
18501 		/* Its not yet ready */
18502 		return (NULL);
18503 	}
18504 	srtt = rack_grab_rtt(tp, rack);
18505 	idx = rsm->r_rtr_cnt - 1;
18506 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
18507 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
18508 	if ((tsused == ts_low) ||
18509 	    (TSTMP_LT(tsused, ts_low))) {
18510 		/* No time since sending */
18511 		return (NULL);
18512 	}
18513 	if ((tsused - ts_low) < thresh) {
18514 		/* It has not been long enough yet */
18515 		return (NULL);
18516 	}
18517 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
18518 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
18519 	     (rack->sack_attack_disable == 0))) {
18520 		/*
18521 		 * We have passed the dup-ack threshold <or>
18522 		 * a SACK has indicated this is missing.
18523 		 * Note that if you are a declared attacker
18524 		 * it is only the dup-ack threshold that
18525 		 * will cause retransmits.
18526 		 */
18527 		/* log retransmit reason */
18528 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
18529 		rack->r_fast_output = 0;
18530 		return (rsm);
18531 	}
18532 	return (NULL);
18533 }
18534 
18535 static void
18536 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
18537 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
18538 			   int line, struct rack_sendmap *rsm, uint8_t quality)
18539 {
18540 	if (tcp_bblogging_on(rack->rc_tp)) {
18541 		union tcp_log_stackspecific log;
18542 		struct timeval tv;
18543 
18544 		if (rack_verbose_logging == 0) {
18545 			/*
18546 			 * We are not verbose screen out all but
18547 			 * ones we always want.
18548 			 */
18549 			if ((method != 2) &&
18550 			    (method != 3) &&
18551 			    (method != 7) &&
18552 			    (method != 89) &&
18553 			    (method != 14) &&
18554 			    (method != 20)) {
18555 				return;
18556 			}
18557 		}
18558 		memset(&log, 0, sizeof(log));
18559 		log.u_bbr.flex1 = slot;
18560 		log.u_bbr.flex2 = len;
18561 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
18562 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
18563 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
18564 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
18565 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
18566 		log.u_bbr.use_lt_bw <<= 1;
18567 		log.u_bbr.use_lt_bw |= rack->r_late;
18568 		log.u_bbr.use_lt_bw <<= 1;
18569 		log.u_bbr.use_lt_bw |= rack->r_early;
18570 		log.u_bbr.use_lt_bw <<= 1;
18571 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
18572 		log.u_bbr.use_lt_bw <<= 1;
18573 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
18574 		log.u_bbr.use_lt_bw <<= 1;
18575 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
18576 		log.u_bbr.use_lt_bw <<= 1;
18577 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
18578 		log.u_bbr.use_lt_bw <<= 1;
18579 		log.u_bbr.use_lt_bw |= rack->gp_ready;
18580 		log.u_bbr.pkt_epoch = line;
18581 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
18582 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
18583 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
18584 		log.u_bbr.bw_inuse = bw_est;
18585 		log.u_bbr.delRate = bw;
18586 		if (rack->r_ctl.gp_bw == 0)
18587 			log.u_bbr.cur_del_rate = 0;
18588 		else
18589 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
18590 		log.u_bbr.rttProp = len_time;
18591 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
18592 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
18593 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18594 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
18595 			/* We are in slow start */
18596 			log.u_bbr.flex7 = 1;
18597 		} else {
18598 			/* we are on congestion avoidance */
18599 			log.u_bbr.flex7 = 0;
18600 		}
18601 		log.u_bbr.flex8 = method;
18602 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18603 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18604 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
18605 		log.u_bbr.cwnd_gain <<= 1;
18606 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
18607 		log.u_bbr.cwnd_gain <<= 1;
18608 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
18609 		log.u_bbr.bbr_substate = quality;
18610 		log.u_bbr.bbr_state = rack->dgp_on;
18611 		log.u_bbr.bbr_state <<= 1;
18612 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
18613 		log.u_bbr.bbr_state <<= 2;
18614 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
18615 		    &rack->rc_inp->inp_socket->so_rcv,
18616 		    &rack->rc_inp->inp_socket->so_snd,
18617 		    BBR_LOG_HPTSI_CALC, 0,
18618 		    0, &log, false, &tv);
18619 	}
18620 }
18621 
18622 static uint32_t
18623 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
18624 {
18625 	uint32_t new_tso, user_max, pace_one;
18626 
18627 	user_max = rack->rc_user_set_max_segs * mss;
18628 	if (rack->rc_force_max_seg) {
18629 		return (user_max);
18630 	}
18631 	if (rack->use_fixed_rate &&
18632 	    ((rack->r_ctl.crte == NULL) ||
18633 	     (bw != rack->r_ctl.crte->rate))) {
18634 		/* Use the user mss since we are not exactly matched */
18635 		return (user_max);
18636 	}
18637 	if (rack_pace_one_seg ||
18638 	    (rack->r_ctl.rc_user_set_min_segs == 1))
18639 		pace_one = 1;
18640 	else
18641 		pace_one = 0;
18642 
18643 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
18644 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
18645 	if (new_tso > user_max)
18646 		new_tso = user_max;
18647 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
18648 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
18649 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
18650 	}
18651 	if (rack->r_ctl.rc_user_set_min_segs &&
18652 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
18653 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
18654 	return (new_tso);
18655 }
18656 
18657 static uint64_t
18658 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
18659 {
18660 	uint64_t reduced_win;
18661 	uint32_t gain;
18662 
18663 	if (window_input < rc_init_window(rack)) {
18664 		/*
18665 		 * The cwnd is collapsed to
18666 		 * nearly zero, maybe because of a time-out?
18667 		 * Lets drop back to the lt-bw.
18668 		 */
18669 		reduced_win = rack_get_lt_bw(rack);
18670 		/* Set the flag so the caller knows its a rate and not a reduced window */
18671 		*rate_set = 1;
18672 		gain = 100;
18673 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
18674 		/*
18675 		 * If we are in recover our cwnd needs to be less for
18676 		 * our pacing consideration.
18677 		 */
18678 		if (rack->rack_hibeta == 0) {
18679 			reduced_win = window_input / 2;
18680 			gain = 50;
18681 		} else {
18682 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
18683 			reduced_win /= 100;
18684 			gain = rack->r_ctl.saved_hibeta;
18685 		}
18686 	} else {
18687 		/*
18688 		 * Apply Timely factor to increase/decrease the
18689 		 * amount we are pacing at.
18690 		 */
18691 		gain = rack_get_output_gain(rack, NULL);
18692 		if (gain > rack_gain_p5_ub) {
18693 			gain = rack_gain_p5_ub;
18694 		}
18695 		reduced_win = window_input * gain;
18696 		reduced_win /= 100;
18697 	}
18698 	if (gain_b != NULL)
18699 		*gain_b = gain;
18700 	/*
18701 	 * What is being returned here is a trimmed down
18702 	 * window values in all cases where rate_set is left
18703 	 * at 0. In one case we actually return the rate (lt_bw).
18704 	 * the "reduced_win" is returned as a slimmed down cwnd that
18705 	 * is then calculated by the caller into a rate when rate_set
18706 	 * is 0.
18707 	 */
18708 	return (reduced_win);
18709 }
18710 
18711 static int32_t
18712 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
18713 {
18714 	uint64_t lentim, fill_bw;
18715 
18716 	rack->r_via_fill_cw = 0;
18717 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
18718 		return (slot);
18719 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
18720 		return (slot);
18721 	if (rack->r_ctl.rc_last_us_rtt == 0)
18722 		return (slot);
18723 	if (rack->rc_pace_fill_if_rttin_range &&
18724 	    (rack->r_ctl.rc_last_us_rtt >=
18725 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
18726 		/* The rtt is huge, N * smallest, lets not fill */
18727 		return (slot);
18728 	}
18729 	if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
18730 		return (slot);
18731 	/*
18732 	 * first lets calculate the b/w based on the last us-rtt
18733 	 * and the the smallest send window.
18734 	 */
18735 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
18736 	if (rack->rc_fillcw_apply_discount) {
18737 		uint32_t rate_set = 0;
18738 
18739 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
18740 		if (rate_set) {
18741 			goto at_lt_bw;
18742 		}
18743 	}
18744 	/* Take the rwnd if its smaller */
18745 	if (fill_bw > rack->rc_tp->snd_wnd)
18746 		fill_bw = rack->rc_tp->snd_wnd;
18747 	/* Now lets make it into a b/w */
18748 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
18749 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
18750 	/* Adjust to any cap */
18751 	if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
18752 		fill_bw = rack->r_ctl.fillcw_cap;
18753 
18754 at_lt_bw:
18755 	if (rack_bw_multipler > 0) {
18756 		/*
18757 		 * We want to limit fill-cw to the some multiplier
18758 		 * of the max(lt_bw, gp_est). The normal default
18759 		 * is 0 for off, so a sysctl has enabled it.
18760 		 */
18761 		uint64_t lt_bw, gp, rate;
18762 
18763 		gp = rack_get_gp_est(rack);
18764 		lt_bw = rack_get_lt_bw(rack);
18765 		if (lt_bw > gp)
18766 			rate = lt_bw;
18767 		else
18768 			rate = gp;
18769 		rate *= rack_bw_multipler;
18770 		rate /= 100;
18771 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18772 			union tcp_log_stackspecific log;
18773 			struct timeval tv;
18774 
18775 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18776 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18777 			log.u_bbr.flex1 = rack_bw_multipler;
18778 			log.u_bbr.flex2 = len;
18779 			log.u_bbr.cur_del_rate = gp;
18780 			log.u_bbr.delRate = lt_bw;
18781 			log.u_bbr.bw_inuse = rate;
18782 			log.u_bbr.rttProp = fill_bw;
18783 			log.u_bbr.flex8 = 44;
18784 			tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
18785 				      BBR_LOG_CWND, 0,
18786 				      0, &log, false, NULL,
18787 				      __func__, __LINE__, &tv);
18788 		}
18789 		if (fill_bw > rate)
18790 			fill_bw = rate;
18791 	}
18792 	/* We are below the min b/w */
18793 	if (non_paced)
18794 		*rate_wanted = fill_bw;
18795 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
18796 		return (slot);
18797 	rack->r_via_fill_cw = 1;
18798 	if (rack->r_rack_hw_rate_caps &&
18799 	    (rack->r_ctl.crte != NULL)) {
18800 		uint64_t high_rate;
18801 
18802 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
18803 		if (fill_bw > high_rate) {
18804 			/* We are capping bw at the highest rate table entry */
18805 			if (*rate_wanted > high_rate) {
18806 				/* The original rate was also capped */
18807 				rack->r_via_fill_cw = 0;
18808 			}
18809 			rack_log_hdwr_pacing(rack,
18810 					     fill_bw, high_rate, __LINE__,
18811 					     0, 3);
18812 			fill_bw = high_rate;
18813 			if (capped)
18814 				*capped = 1;
18815 		}
18816 	} else if ((rack->r_ctl.crte == NULL) &&
18817 		   (rack->rack_hdrw_pacing == 0) &&
18818 		   (rack->rack_hdw_pace_ena) &&
18819 		   rack->r_rack_hw_rate_caps &&
18820 		   (rack->rack_attempt_hdwr_pace == 0) &&
18821 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
18822 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
18823 		/*
18824 		 * Ok we may have a first attempt that is greater than our top rate
18825 		 * lets check.
18826 		 */
18827 		uint64_t high_rate;
18828 
18829 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
18830 		if (high_rate) {
18831 			if (fill_bw > high_rate) {
18832 				fill_bw = high_rate;
18833 				if (capped)
18834 					*capped = 1;
18835 			}
18836 		}
18837 	}
18838 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
18839 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
18840 				   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
18841 		fill_bw = rack->r_ctl.bw_rate_cap;
18842 	}
18843 	/*
18844 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
18845 	 * in an rtt (unless it was capped), what does that
18846 	 * time wise equate too?
18847 	 */
18848 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
18849 	lentim /= fill_bw;
18850 	*rate_wanted = fill_bw;
18851 	if (non_paced || (lentim < slot)) {
18852 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
18853 					   0, lentim, 12, __LINE__, NULL, 0);
18854 		return ((int32_t)lentim);
18855 	} else
18856 		return (slot);
18857 }
18858 
18859 static uint32_t
18860 rack_policer_check_send(struct tcp_rack *rack, uint32_t len, uint32_t segsiz, uint32_t *needs)
18861 {
18862 	uint64_t calc;
18863 
18864 	rack->rc_policer_should_pace = 0;
18865 	calc = rack_policer_bucket_reserve * rack->r_ctl.policer_bucket_size;
18866 	calc /= 100;
18867 	/*
18868 	 * Now lets look at if we want more than is in the bucket <or>
18869 	 * we want more than is reserved in the bucket.
18870 	 */
18871 	if (rack_verbose_logging > 0)
18872 		policer_detection_log(rack, len, segsiz, calc, rack->r_ctl.current_policer_bucket, 8);
18873 	if ((calc > rack->r_ctl.current_policer_bucket) ||
18874 	    (len >= (rack->r_ctl.current_policer_bucket - calc))) {
18875 		/*
18876 		 * We may want to pace depending on if we are going
18877 		 * into the reserve or not.
18878 		 */
18879 		uint32_t newlen;
18880 
18881 		if (calc > rack->r_ctl.current_policer_bucket) {
18882 			/*
18883 			 * This will eat into the reserve if we
18884 			 * don't have room at all some lines
18885 			 * below will catch it.
18886 			 */
18887 			newlen = rack->r_ctl.policer_max_seg;
18888 			rack->rc_policer_should_pace = 1;
18889 		} else {
18890 			/*
18891 			 * We have all of the reserve plus something in the bucket
18892 			 * that we can give out.
18893 			 */
18894 			newlen = rack->r_ctl.current_policer_bucket - calc;
18895 			if (newlen < rack->r_ctl.policer_max_seg) {
18896 				/*
18897 				 * Into the reserve to get a full policer_max_seg
18898 				 * so we set the len to that and eat into
18899 				 * the reserve. If we go over the code
18900 				 * below will make us wait.
18901 				 */
18902 				newlen = rack->r_ctl.policer_max_seg;
18903 				rack->rc_policer_should_pace = 1;
18904 			}
18905 		}
18906 		if (newlen > rack->r_ctl.current_policer_bucket) {
18907 			/* We have to wait some */
18908 			*needs = newlen - rack->r_ctl.current_policer_bucket;
18909 			return (0);
18910 		}
18911 		if (rack_verbose_logging > 0)
18912 			policer_detection_log(rack, len, segsiz, newlen, 0, 9);
18913 		len = newlen;
18914 	} /* else we have all len available above the reserve */
18915 	if (rack_verbose_logging > 0)
18916 		policer_detection_log(rack, len, segsiz, calc, 0, 10);
18917 	return (len);
18918 }
18919 
18920 static uint32_t
18921 rack_policed_sending(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, uint32_t segsiz, int call_line)
18922 {
18923 	/*
18924 	 * Given a send of len, and a token bucket set at current_policer_bucket_size
18925 	 * are we close enough to the end of the bucket that we need to pace? If so
18926 	 * calculate out a time and return it. Otherwise subtract the tokens from
18927 	 * the bucket.
18928 	 */
18929 	uint64_t calc;
18930 
18931 	if ((rack->r_ctl.policer_bw == 0) ||
18932 	    (rack->r_ctl.policer_bucket_size < segsiz)) {
18933 		/*
18934 		 * We should have an estimate here...
18935 		 */
18936 		return (0);
18937 	}
18938 	calc = (uint64_t)rack_policer_bucket_reserve * (uint64_t)rack->r_ctl.policer_bucket_size;
18939 	calc /= 100;
18940 	if ((rack->r_ctl.current_policer_bucket < len) ||
18941 	    (rack->rc_policer_should_pace == 1) ||
18942 	    ((rack->r_ctl.current_policer_bucket - len) <= (uint32_t)calc)) {
18943 		/* we need to pace */
18944 		uint64_t lentim, res;
18945 		uint32_t slot;
18946 
18947 		lentim = (uint64_t)len * (uint64_t)HPTS_USEC_IN_SEC;
18948 		res = lentim / rack->r_ctl.policer_bw;
18949 		slot = (uint32_t)res;
18950 		if (rack->r_ctl.current_policer_bucket > len)
18951 			rack->r_ctl.current_policer_bucket -= len;
18952 		else
18953 			rack->r_ctl.current_policer_bucket = 0;
18954 		policer_detection_log(rack, len, slot, (uint32_t)rack_policer_bucket_reserve, call_line, 5);
18955 		rack->rc_policer_should_pace = 0;
18956 		return(slot);
18957 	}
18958 	/* Just take tokens out of the bucket and let rack do whatever it would have */
18959 	policer_detection_log(rack, len, 0, (uint32_t)rack_policer_bucket_reserve, call_line, 6);
18960 	if (len < rack->r_ctl.current_policer_bucket) {
18961 		rack->r_ctl.current_policer_bucket -= len;
18962 	} else {
18963 		rack->r_ctl.current_policer_bucket = 0;
18964 	}
18965 	return (0);
18966 }
18967 
18968 
18969 static int32_t
18970 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
18971 {
18972 	uint64_t srtt;
18973 	int32_t slot = 0;
18974 	int32_t minslot = 0;
18975 	int can_start_hw_pacing = 1;
18976 	int err;
18977 	int pace_one;
18978 
18979 	if (rack_pace_one_seg ||
18980 	    (rack->r_ctl.rc_user_set_min_segs == 1))
18981 		pace_one = 1;
18982 	else
18983 		pace_one = 0;
18984 	if (rack->rc_policer_detected == 1) {
18985 		/*
18986 		 * A policer has been detected and we
18987 		 * have all of our data (policer-bw and
18988 		 * policer bucket size) calculated. Call
18989 		 * into the function to find out if we are
18990 		 * overriding the time.
18991 		 */
18992 		slot = rack_policed_sending(rack, tp, len, segsiz, line);
18993 		if (slot) {
18994 			uint64_t logbw;
18995 
18996 			logbw = rack->r_ctl.current_policer_bucket;
18997 			logbw <<= 32;
18998 			logbw |= rack->r_ctl.policer_bucket_size;
18999 			rack_log_pacing_delay_calc(rack, len, slot, rack->r_ctl.policer_bw, logbw, 0, 89, __LINE__, NULL, 0);
19000 			return(slot);
19001 		}
19002 	}
19003 	if (rack->rc_always_pace == 0) {
19004 		/*
19005 		 * We use the most optimistic possible cwnd/srtt for
19006 		 * sending calculations. This will make our
19007 		 * calculation anticipate getting more through
19008 		 * quicker then possible. But thats ok we don't want
19009 		 * the peer to have a gap in data sending.
19010 		 */
19011 		uint64_t cwnd, tr_perms = 0;
19012 		int32_t reduce = 0;
19013 
19014 	old_method:
19015 		/*
19016 		 * We keep no precise pacing with the old method
19017 		 * instead we use the pacer to mitigate bursts.
19018 		 */
19019 		if (rack->r_ctl.rc_rack_min_rtt)
19020 			srtt = rack->r_ctl.rc_rack_min_rtt;
19021 		else
19022 			srtt = max(tp->t_srtt, 1);
19023 		if (rack->r_ctl.rc_rack_largest_cwnd)
19024 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
19025 		else
19026 			cwnd = rack->r_ctl.cwnd_to_use;
19027 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
19028 		tr_perms = (cwnd * 1000) / srtt;
19029 		if (tr_perms == 0) {
19030 			tr_perms = ctf_fixed_maxseg(tp);
19031 		}
19032 		/*
19033 		 * Calculate how long this will take to drain, if
19034 		 * the calculation comes out to zero, thats ok we
19035 		 * will use send_a_lot to possibly spin around for
19036 		 * more increasing tot_len_this_send to the point
19037 		 * that its going to require a pace, or we hit the
19038 		 * cwnd. Which in that case we are just waiting for
19039 		 * a ACK.
19040 		 */
19041 		slot = len / tr_perms;
19042 		/* Now do we reduce the time so we don't run dry? */
19043 		if (slot && rack_slot_reduction) {
19044 			reduce = (slot / rack_slot_reduction);
19045 			if (reduce < slot) {
19046 				slot -= reduce;
19047 			} else
19048 				slot = 0;
19049 		}
19050 		slot *= HPTS_USEC_IN_MSEC;
19051 		if (rack->rc_pace_to_cwnd) {
19052 			uint64_t rate_wanted = 0;
19053 
19054 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
19055 			rack->rc_ack_can_sendout_data = 1;
19056 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
19057 		} else
19058 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
19059 		/*******************************************************/
19060 		/* RRS: We insert non-paced call to stats here for len */
19061 		/*******************************************************/
19062 	} else {
19063 		uint64_t bw_est, res, lentim, rate_wanted;
19064 		uint32_t segs, oh;
19065 		int capped = 0;
19066 		int prev_fill;
19067 
19068 		if ((rack->r_rr_config == 1) && rsm) {
19069 			return (rack->r_ctl.rc_min_to);
19070 		}
19071 		if (rack->use_fixed_rate) {
19072 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
19073 		} else if ((rack->r_ctl.init_rate == 0) &&
19074 			   (rack->r_ctl.gp_bw == 0)) {
19075 			/* no way to yet do an estimate */
19076 			bw_est = rate_wanted = 0;
19077 		} else if (rack->dgp_on) {
19078 			bw_est = rack_get_bw(rack);
19079 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
19080 		} else {
19081 			uint32_t gain, rate_set = 0;
19082 
19083 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
19084 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
19085 			if (rate_set == 0) {
19086 				if (rate_wanted > rack->rc_tp->snd_wnd)
19087 					rate_wanted = rack->rc_tp->snd_wnd;
19088 				/* Now lets make it into a b/w */
19089 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
19090 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
19091 			}
19092 			bw_est = rate_wanted;
19093 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
19094 						   rack->r_ctl.cwnd_to_use,
19095 						   rate_wanted, bw_est,
19096 						   rack->r_ctl.rc_last_us_rtt,
19097 						   88, __LINE__, NULL, gain);
19098 		}
19099 		if ((bw_est == 0) || (rate_wanted == 0) ||
19100 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
19101 			/*
19102 			 * No way yet to make a b/w estimate or
19103 			 * our raise is set incorrectly.
19104 			 */
19105 			goto old_method;
19106 		}
19107 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
19108 		/* We need to account for all the overheads */
19109 		segs = (len + segsiz - 1) / segsiz;
19110 		/*
19111 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
19112 		 * and how much data we put in each packet. Yes this
19113 		 * means we may be off if we are larger than 1500 bytes
19114 		 * or smaller. But this just makes us more conservative.
19115 		 */
19116 
19117 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
19118 		if (rack->r_is_v6) {
19119 #ifdef INET6
19120 			oh += sizeof(struct ip6_hdr);
19121 #endif
19122 		} else {
19123 #ifdef INET
19124 			oh += sizeof(struct ip);
19125 #endif
19126 		}
19127 		/* We add a fixed 14 for the ethernet header */
19128 		oh += 14;
19129 		segs *= oh;
19130 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
19131 		res = lentim / rate_wanted;
19132 		slot = (uint32_t)res;
19133 		if (rack_hw_rate_min &&
19134 		    (rate_wanted < rack_hw_rate_min)) {
19135 			can_start_hw_pacing = 0;
19136 			if (rack->r_ctl.crte) {
19137 				/*
19138 				 * Ok we need to release it, we
19139 				 * have fallen too low.
19140 				 */
19141 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
19142 				rack->r_ctl.crte = NULL;
19143 				rack->rack_attempt_hdwr_pace = 0;
19144 				rack->rack_hdrw_pacing = 0;
19145 			}
19146 		}
19147 		if (rack->r_ctl.crte &&
19148 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
19149 			/*
19150 			 * We want more than the hardware can give us,
19151 			 * don't start any hw pacing.
19152 			 */
19153 			can_start_hw_pacing = 0;
19154 			if (rack->r_rack_hw_rate_caps == 0) {
19155 				/*
19156 				 * Ok we need to release it, we
19157 				 * want more than the card can give us and
19158 				 * no rate cap is in place. Set it up so
19159 				 * when we want less we can retry.
19160 				 */
19161 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
19162 				rack->r_ctl.crte = NULL;
19163 				rack->rack_attempt_hdwr_pace = 0;
19164 				rack->rack_hdrw_pacing = 0;
19165 			}
19166 		}
19167 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
19168 			/*
19169 			 * We lost our rate somehow, this can happen
19170 			 * if the interface changed underneath us.
19171 			 */
19172 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
19173 			rack->r_ctl.crte = NULL;
19174 			/* Lets re-allow attempting to setup pacing */
19175 			rack->rack_hdrw_pacing = 0;
19176 			rack->rack_attempt_hdwr_pace = 0;
19177 			rack_log_hdwr_pacing(rack,
19178 					     rate_wanted, bw_est, __LINE__,
19179 					     0, 6);
19180 		}
19181 		prev_fill = rack->r_via_fill_cw;
19182 		if ((rack->rc_pace_to_cwnd) &&
19183 		    (capped == 0) &&
19184 		    (rack->dgp_on == 1) &&
19185 		    (rack->use_fixed_rate == 0) &&
19186 		    (rack->in_probe_rtt == 0) &&
19187 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
19188 			/*
19189 			 * We want to pace at our rate *or* faster to
19190 			 * fill the cwnd to the max if its not full.
19191 			 */
19192 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
19193 			/* Re-check to make sure we are not exceeding our max b/w */
19194 			if ((rack->r_ctl.crte != NULL) &&
19195 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
19196 				/*
19197 				 * We want more than the hardware can give us,
19198 				 * don't start any hw pacing.
19199 				 */
19200 				can_start_hw_pacing = 0;
19201 				if (rack->r_rack_hw_rate_caps == 0) {
19202 					/*
19203 					 * Ok we need to release it, we
19204 					 * want more than the card can give us and
19205 					 * no rate cap is in place. Set it up so
19206 					 * when we want less we can retry.
19207 					 */
19208 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
19209 					rack->r_ctl.crte = NULL;
19210 					rack->rack_attempt_hdwr_pace = 0;
19211 					rack->rack_hdrw_pacing = 0;
19212 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19213 				}
19214 			}
19215 		}
19216 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
19217 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
19218 			if ((rack->rack_hdw_pace_ena) &&
19219 			    (can_start_hw_pacing > 0) &&
19220 			    (rack->rack_hdrw_pacing == 0) &&
19221 			    (rack->rack_attempt_hdwr_pace == 0)) {
19222 				/*
19223 				 * Lets attempt to turn on hardware pacing
19224 				 * if we can.
19225 				 */
19226 				rack->rack_attempt_hdwr_pace = 1;
19227 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
19228 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
19229 								       rate_wanted,
19230 								       RS_PACING_GEQ,
19231 								       &err, &rack->r_ctl.crte_prev_rate);
19232 				if (rack->r_ctl.crte) {
19233 					rack->rack_hdrw_pacing = 1;
19234 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
19235 													   pace_one, rack->r_ctl.crte,
19236 													   NULL, rack->r_ctl.pace_len_divisor);
19237 					rack_log_hdwr_pacing(rack,
19238 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
19239 							     err, 0);
19240 					rack->r_ctl.last_hw_bw_req = rate_wanted;
19241 				} else {
19242 					counter_u64_add(rack_hw_pace_init_fail, 1);
19243 				}
19244 			} else if (rack->rack_hdrw_pacing &&
19245 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
19246 				/* Do we need to adjust our rate? */
19247 				const struct tcp_hwrate_limit_table *nrte;
19248 
19249 				if (rack->r_up_only &&
19250 				    (rate_wanted < rack->r_ctl.crte->rate)) {
19251 					/**
19252 					 * We have four possible states here
19253 					 * having to do with the previous time
19254 					 * and this time.
19255 					 *   previous  |  this-time
19256 					 * A)     0      |     0   -- fill_cw not in the picture
19257 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
19258 					 * C)     1      |     1   -- all rates from fill_cw
19259 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
19260 					 *
19261 					 * For case A, C and D we don't allow a drop. But for
19262 					 * case B where we now our on our steady rate we do
19263 					 * allow a drop.
19264 					 *
19265 					 */
19266 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
19267 						goto done_w_hdwr;
19268 				}
19269 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
19270 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
19271 					if (rack_hw_rate_to_low &&
19272 					    (bw_est < rack_hw_rate_to_low)) {
19273 						/*
19274 						 * The pacing rate is too low for hardware, but
19275 						 * do allow hardware pacing to be restarted.
19276 						 */
19277 						rack_log_hdwr_pacing(rack,
19278 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
19279 								     0, 5);
19280 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
19281 						rack->r_ctl.crte = NULL;
19282 						rack->rack_attempt_hdwr_pace = 0;
19283 						rack->rack_hdrw_pacing = 0;
19284 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
19285 						goto done_w_hdwr;
19286 					}
19287 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
19288 								   rack->rc_tp,
19289 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
19290 								   rate_wanted,
19291 								   RS_PACING_GEQ,
19292 								   &err, &rack->r_ctl.crte_prev_rate);
19293 					if (nrte == NULL) {
19294 						/*
19295 						 * Lost the rate, lets drop hardware pacing
19296 						 * period.
19297 						 */
19298 						rack->rack_hdrw_pacing = 0;
19299 						rack->r_ctl.crte = NULL;
19300 						rack_log_hdwr_pacing(rack,
19301 								     rate_wanted, 0, __LINE__,
19302 								     err, 1);
19303 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
19304 						counter_u64_add(rack_hw_pace_lost, 1);
19305 					} else if (nrte != rack->r_ctl.crte) {
19306 						rack->r_ctl.crte = nrte;
19307 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
19308 														   segsiz, pace_one, rack->r_ctl.crte,
19309 														   NULL, rack->r_ctl.pace_len_divisor);
19310 						rack_log_hdwr_pacing(rack,
19311 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
19312 								     err, 2);
19313 						rack->r_ctl.last_hw_bw_req = rate_wanted;
19314 					}
19315 				} else {
19316 					/* We just need to adjust the segment size */
19317 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
19318 					rack_log_hdwr_pacing(rack,
19319 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
19320 							     0, 4);
19321 					rack->r_ctl.last_hw_bw_req = rate_wanted;
19322 				}
19323 			}
19324 		}
19325 		if (minslot && (minslot > slot)) {
19326 			rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
19327 						   98, __LINE__, NULL, 0);
19328 			slot = minslot;
19329 		}
19330 	done_w_hdwr:
19331 		if (rack_limit_time_with_srtt &&
19332 		    (rack->use_fixed_rate == 0) &&
19333 		    (rack->rack_hdrw_pacing == 0)) {
19334 			/*
19335 			 * Sanity check, we do not allow the pacing delay
19336 			 * to be longer than the SRTT of the path. If it is
19337 			 * a slow path, then adding a packet should increase
19338 			 * the RTT and compensate for this i.e. the srtt will
19339 			 * be greater so the allowed pacing time will be greater.
19340 			 *
19341 			 * Note this restriction is not for where a peak rate
19342 			 * is set, we are doing fixed pacing or hardware pacing.
19343 			 */
19344 			if (rack->rc_tp->t_srtt)
19345 				srtt = rack->rc_tp->t_srtt;
19346 			else
19347 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
19348 			if (srtt < (uint64_t)slot) {
19349 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
19350 				slot = srtt;
19351 			}
19352 		}
19353 		/*******************************************************************/
19354 		/* RRS: We insert paced call to stats here for len and rate_wanted */
19355 		/*******************************************************************/
19356 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
19357 	}
19358 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
19359 		/*
19360 		 * If this rate is seeing enobufs when it
19361 		 * goes to send then either the nic is out
19362 		 * of gas or we are mis-estimating the time
19363 		 * somehow and not letting the queue empty
19364 		 * completely. Lets add to the pacing time.
19365 		 */
19366 		int hw_boost_delay;
19367 
19368 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
19369 		if (hw_boost_delay > rack_enobuf_hw_max)
19370 			hw_boost_delay = rack_enobuf_hw_max;
19371 		else if (hw_boost_delay < rack_enobuf_hw_min)
19372 			hw_boost_delay = rack_enobuf_hw_min;
19373 		slot += hw_boost_delay;
19374 	}
19375 	return (slot);
19376 }
19377 
19378 static void
19379 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
19380     tcp_seq startseq, uint32_t sb_offset)
19381 {
19382 	struct rack_sendmap *my_rsm = NULL;
19383 
19384 	if (tp->t_state < TCPS_ESTABLISHED) {
19385 		/*
19386 		 * We don't start any measurements if we are
19387 		 * not at least established.
19388 		 */
19389 		return;
19390 	}
19391 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
19392 		/*
19393 		 * We will get no more data into the SB
19394 		 * this means we need to have the data available
19395 		 * before we start a measurement.
19396 		 */
19397 
19398 		if (sbavail(&tptosocket(tp)->so_snd) <
19399 		    max(rc_init_window(rack),
19400 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
19401 			/* Nope not enough data */
19402 			return;
19403 		}
19404 	}
19405 	tp->t_flags |= TF_GPUTINPROG;
19406 	rack->r_ctl.rc_gp_cumack_ts = 0;
19407 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
19408 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
19409 	tp->gput_seq = startseq;
19410 	rack->app_limited_needs_set = 0;
19411 	if (rack->in_probe_rtt)
19412 		rack->measure_saw_probe_rtt = 1;
19413 	else if ((rack->measure_saw_probe_rtt) &&
19414 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
19415 		rack->measure_saw_probe_rtt = 0;
19416 	if (rack->rc_gp_filled)
19417 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
19418 	else {
19419 		/* Special case initial measurement */
19420 		struct timeval tv;
19421 
19422 		tp->gput_ts = tcp_get_usecs(&tv);
19423 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
19424 	}
19425 	/*
19426 	 * We take a guess out into the future,
19427 	 * if we have no measurement and no
19428 	 * initial rate, we measure the first
19429 	 * initial-windows worth of data to
19430 	 * speed up getting some GP measurement and
19431 	 * thus start pacing.
19432 	 */
19433 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
19434 		rack->app_limited_needs_set = 1;
19435 		tp->gput_ack = startseq + max(rc_init_window(rack),
19436 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
19437 		rack_log_pacing_delay_calc(rack,
19438 					   tp->gput_seq,
19439 					   tp->gput_ack,
19440 					   0,
19441 					   tp->gput_ts,
19442 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
19443 					   9,
19444 					   __LINE__, NULL, 0);
19445 		rack_tend_gp_marks(tp, rack);
19446 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
19447 		return;
19448 	}
19449 	if (sb_offset) {
19450 		/*
19451 		 * We are out somewhere in the sb
19452 		 * can we use the already outstanding data?
19453 		 */
19454 
19455 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
19456 			/*
19457 			 * Yes first one is good and in this case
19458 			 * the tp->gput_ts is correctly set based on
19459 			 * the last ack that arrived (no need to
19460 			 * set things up when an ack comes in).
19461 			 */
19462 			my_rsm = tqhash_min(rack->r_ctl.tqh);
19463 			if ((my_rsm == NULL) ||
19464 			    (my_rsm->r_rtr_cnt != 1)) {
19465 				/* retransmission? */
19466 				goto use_latest;
19467 			}
19468 		} else {
19469 			if (rack->r_ctl.rc_first_appl == NULL) {
19470 				/*
19471 				 * If rc_first_appl is NULL
19472 				 * then the cnt should be 0.
19473 				 * This is probably an error, maybe
19474 				 * a KASSERT would be approprate.
19475 				 */
19476 				goto use_latest;
19477 			}
19478 			/*
19479 			 * If we have a marker pointer to the last one that is
19480 			 * app limited we can use that, but we need to set
19481 			 * things up so that when it gets ack'ed we record
19482 			 * the ack time (if its not already acked).
19483 			 */
19484 			rack->app_limited_needs_set = 1;
19485 			/*
19486 			 * We want to get to the rsm that is either
19487 			 * next with space i.e. over 1 MSS or the one
19488 			 * after that (after the app-limited).
19489 			 */
19490 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
19491 			if (my_rsm) {
19492 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
19493 					/* Have to use the next one */
19494 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
19495 				else {
19496 					/* Use after the first MSS of it is acked */
19497 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
19498 					goto start_set;
19499 				}
19500 			}
19501 			if ((my_rsm == NULL) ||
19502 			    (my_rsm->r_rtr_cnt != 1)) {
19503 				/*
19504 				 * Either its a retransmit or
19505 				 * the last is the app-limited one.
19506 				 */
19507 				goto use_latest;
19508 			}
19509 		}
19510 		tp->gput_seq = my_rsm->r_start;
19511 start_set:
19512 		if (my_rsm->r_flags & RACK_ACKED) {
19513 			/*
19514 			 * This one has been acked use the arrival ack time
19515 			 */
19516 			struct rack_sendmap *nrsm;
19517 
19518 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
19519 			rack->app_limited_needs_set = 0;
19520 			/*
19521 			 * Ok in this path we need to use the r_end now
19522 			 * since this guy is the starting ack.
19523 			 */
19524 			tp->gput_seq = my_rsm->r_end;
19525 			/*
19526 			 * We also need to adjust up the sendtime
19527 			 * to the send of the next data after my_rsm.
19528 			 */
19529 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
19530 			if (nrsm != NULL)
19531 				my_rsm = nrsm;
19532 			else {
19533 				/*
19534 				 * The next as not been sent, thats the
19535 				 * case for using the latest.
19536 				 */
19537 				goto use_latest;
19538 			}
19539 		}
19540 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
19541 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
19542 		rack->r_ctl.rc_gp_cumack_ts = 0;
19543 		if ((rack->r_ctl.cleared_app_ack == 1) &&
19544 		    (SEQ_GEQ(rack->r_ctl.cleared_app_ack, tp->gput_seq))) {
19545 			/*
19546 			 * We just cleared an application limited period
19547 			 * so the next seq out needs to skip the first
19548 			 * ack.
19549 			 */
19550 			rack->app_limited_needs_set = 1;
19551 			rack->r_ctl.cleared_app_ack = 0;
19552 		}
19553 		rack_log_pacing_delay_calc(rack,
19554 					   tp->gput_seq,
19555 					   tp->gput_ack,
19556 					   (uint64_t)my_rsm,
19557 					   tp->gput_ts,
19558 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
19559 					   9,
19560 					   __LINE__, my_rsm, 0);
19561 		/* Now lets make sure all are marked as they should be */
19562 		rack_tend_gp_marks(tp, rack);
19563 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
19564 		return;
19565 	}
19566 
19567 use_latest:
19568 	/*
19569 	 * We don't know how long we may have been
19570 	 * idle or if this is the first-send. Lets
19571 	 * setup the flag so we will trim off
19572 	 * the first ack'd data so we get a true
19573 	 * measurement.
19574 	 */
19575 	rack->app_limited_needs_set = 1;
19576 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
19577 	rack->r_ctl.rc_gp_cumack_ts = 0;
19578 	/* Find this guy so we can pull the send time */
19579 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
19580 	if (my_rsm) {
19581 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
19582 		if (my_rsm->r_flags & RACK_ACKED) {
19583 			/*
19584 			 * Unlikely since its probably what was
19585 			 * just transmitted (but I am paranoid).
19586 			 */
19587 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
19588 			rack->app_limited_needs_set = 0;
19589 		}
19590 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
19591 			/* This also is unlikely */
19592 			tp->gput_seq = my_rsm->r_start;
19593 		}
19594 	} else {
19595 		/*
19596 		 * TSNH unless we have some send-map limit,
19597 		 * and even at that it should not be hitting
19598 		 * that limit (we should have stopped sending).
19599 		 */
19600 		struct timeval tv;
19601 
19602 		microuptime(&tv);
19603 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
19604 	}
19605 	rack_tend_gp_marks(tp, rack);
19606 	rack_log_pacing_delay_calc(rack,
19607 				   tp->gput_seq,
19608 				   tp->gput_ack,
19609 				   (uint64_t)my_rsm,
19610 				   tp->gput_ts,
19611 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
19612 				   9, __LINE__, NULL, 0);
19613 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
19614 }
19615 
19616 static inline uint32_t
19617 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
19618     uint32_t avail, int32_t sb_offset)
19619 {
19620 	uint32_t len;
19621 	uint32_t sendwin;
19622 
19623 	if (tp->snd_wnd > cwnd_to_use)
19624 		sendwin = cwnd_to_use;
19625 	else
19626 		sendwin = tp->snd_wnd;
19627 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
19628 		/* We never want to go over our peers rcv-window */
19629 		len = 0;
19630 	} else {
19631 		uint32_t flight;
19632 
19633 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
19634 		if (flight >= sendwin) {
19635 			/*
19636 			 * We have in flight what we are allowed by cwnd (if
19637 			 * it was rwnd blocking it would have hit above out
19638 			 * >= tp->snd_wnd).
19639 			 */
19640 			return (0);
19641 		}
19642 		len = sendwin - flight;
19643 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
19644 			/* We would send too much (beyond the rwnd) */
19645 			len = tp->snd_wnd - ctf_outstanding(tp);
19646 		}
19647 		if ((len + sb_offset) > avail) {
19648 			/*
19649 			 * We don't have that much in the SB, how much is
19650 			 * there?
19651 			 */
19652 			len = avail - sb_offset;
19653 		}
19654 	}
19655 	return (len);
19656 }
19657 
19658 static void
19659 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
19660 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
19661 	     int rsm_is_null, int optlen, int line, uint16_t mode)
19662 {
19663 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
19664 		union tcp_log_stackspecific log;
19665 		struct timeval tv;
19666 
19667 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19668 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19669 		log.u_bbr.flex1 = error;
19670 		log.u_bbr.flex2 = flags;
19671 		log.u_bbr.flex3 = rsm_is_null;
19672 		log.u_bbr.flex4 = ipoptlen;
19673 		log.u_bbr.flex5 = tp->rcv_numsacks;
19674 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19675 		log.u_bbr.flex7 = optlen;
19676 		log.u_bbr.flex8 = rack->r_fsb_inited;
19677 		log.u_bbr.applimited = rack->r_fast_output;
19678 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19679 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19680 		log.u_bbr.cwnd_gain = mode;
19681 		log.u_bbr.pkts_out = orig_len;
19682 		log.u_bbr.lt_epoch = len;
19683 		log.u_bbr.delivered = line;
19684 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
19685 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19686 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
19687 			       len, &log, false, NULL, __func__, __LINE__, &tv);
19688 	}
19689 }
19690 
19691 
19692 static struct mbuf *
19693 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
19694 		   struct rack_fast_send_blk *fsb,
19695 		   int32_t seglimit, int32_t segsize, int hw_tls)
19696 {
19697 #ifdef KERN_TLS
19698 	struct ktls_session *tls, *ntls;
19699 #ifdef INVARIANTS
19700 	struct mbuf *start;
19701 #endif
19702 #endif
19703 	struct mbuf *m, *n, **np, *smb;
19704 	struct mbuf *top;
19705 	int32_t off, soff;
19706 	int32_t len = *plen;
19707 	int32_t fragsize;
19708 	int32_t len_cp = 0;
19709 	uint32_t mlen, frags;
19710 
19711 	soff = off = the_off;
19712 	smb = m = the_m;
19713 	np = &top;
19714 	top = NULL;
19715 #ifdef KERN_TLS
19716 	if (hw_tls && (m->m_flags & M_EXTPG))
19717 		tls = m->m_epg_tls;
19718 	else
19719 		tls = NULL;
19720 #ifdef INVARIANTS
19721 	start = m;
19722 #endif
19723 #endif
19724 	while (len > 0) {
19725 		if (m == NULL) {
19726 			*plen = len_cp;
19727 			break;
19728 		}
19729 #ifdef KERN_TLS
19730 		if (hw_tls) {
19731 			if (m->m_flags & M_EXTPG)
19732 				ntls = m->m_epg_tls;
19733 			else
19734 				ntls = NULL;
19735 
19736 			/*
19737 			 * Avoid mixing TLS records with handshake
19738 			 * data or TLS records from different
19739 			 * sessions.
19740 			 */
19741 			if (tls != ntls) {
19742 				MPASS(m != start);
19743 				*plen = len_cp;
19744 				break;
19745 			}
19746 		}
19747 #endif
19748 		mlen = min(len, m->m_len - off);
19749 		if (seglimit) {
19750 			/*
19751 			 * For M_EXTPG mbufs, add 3 segments
19752 			 * + 1 in case we are crossing page boundaries
19753 			 * + 2 in case the TLS hdr/trailer are used
19754 			 * It is cheaper to just add the segments
19755 			 * than it is to take the cache miss to look
19756 			 * at the mbuf ext_pgs state in detail.
19757 			 */
19758 			if (m->m_flags & M_EXTPG) {
19759 				fragsize = min(segsize, PAGE_SIZE);
19760 				frags = 3;
19761 			} else {
19762 				fragsize = segsize;
19763 				frags = 0;
19764 			}
19765 
19766 			/* Break if we really can't fit anymore. */
19767 			if ((frags + 1) >= seglimit) {
19768 				*plen =	len_cp;
19769 				break;
19770 			}
19771 
19772 			/*
19773 			 * Reduce size if you can't copy the whole
19774 			 * mbuf. If we can't copy the whole mbuf, also
19775 			 * adjust len so the loop will end after this
19776 			 * mbuf.
19777 			 */
19778 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
19779 				mlen = (seglimit - frags - 1) * fragsize;
19780 				len = mlen;
19781 				*plen = len_cp + len;
19782 			}
19783 			frags += howmany(mlen, fragsize);
19784 			if (frags == 0)
19785 				frags++;
19786 			seglimit -= frags;
19787 			KASSERT(seglimit > 0,
19788 			    ("%s: seglimit went too low", __func__));
19789 		}
19790 		n = m_get(M_NOWAIT, m->m_type);
19791 		*np = n;
19792 		if (n == NULL)
19793 			goto nospace;
19794 		n->m_len = mlen;
19795 		soff += mlen;
19796 		len_cp += n->m_len;
19797 		if (m->m_flags & (M_EXT | M_EXTPG)) {
19798 			n->m_data = m->m_data + off;
19799 			mb_dupcl(n, m);
19800 		} else {
19801 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
19802 			    (u_int)n->m_len);
19803 		}
19804 		len -= n->m_len;
19805 		off = 0;
19806 		m = m->m_next;
19807 		np = &n->m_next;
19808 		if (len || (soff == smb->m_len)) {
19809 			/*
19810 			 * We have more so we move forward  or
19811 			 * we have consumed the entire mbuf and
19812 			 * len has fell to 0.
19813 			 */
19814 			soff = 0;
19815 			smb = m;
19816 		}
19817 
19818 	}
19819 	if (fsb != NULL) {
19820 		fsb->m = smb;
19821 		fsb->off = soff;
19822 		if (smb) {
19823 			/*
19824 			 * Save off the size of the mbuf. We do
19825 			 * this so that we can recognize when it
19826 			 * has been trimmed by sbcut() as acks
19827 			 * come in.
19828 			 */
19829 			fsb->o_m_len = smb->m_len;
19830 			fsb->o_t_len = M_TRAILINGROOM(smb);
19831 		} else {
19832 			/*
19833 			 * This is the case where the next mbuf went to NULL. This
19834 			 * means with this copy we have sent everything in the sb.
19835 			 * In theory we could clear the fast_output flag, but lets
19836 			 * not since its possible that we could get more added
19837 			 * and acks that call the extend function which would let
19838 			 * us send more.
19839 			 */
19840 			fsb->o_m_len = 0;
19841 			fsb->o_t_len = 0;
19842 		}
19843 	}
19844 	return (top);
19845 nospace:
19846 	if (top)
19847 		m_freem(top);
19848 	return (NULL);
19849 
19850 }
19851 
19852 /*
19853  * This is a copy of m_copym(), taking the TSO segment size/limit
19854  * constraints into account, and advancing the sndptr as it goes.
19855  */
19856 static struct mbuf *
19857 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
19858 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
19859 {
19860 	struct mbuf *m, *n;
19861 	int32_t soff;
19862 
19863 	m = rack->r_ctl.fsb.m;
19864 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
19865 		/*
19866 		 * The trailing space changed, mbufs can grow
19867 		 * at the tail but they can't shrink from
19868 		 * it, KASSERT that. Adjust the orig_m_len to
19869 		 * compensate for this change.
19870 		 */
19871 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
19872 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
19873 			 m,
19874 			 rack,
19875 			 (intmax_t)M_TRAILINGROOM(m),
19876 			 rack->r_ctl.fsb.o_t_len,
19877 			 rack->r_ctl.fsb.o_m_len,
19878 			 m->m_len));
19879 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
19880 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
19881 	}
19882 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
19883 		/*
19884 		 * Mbuf shrank, trimmed off the top by an ack, our
19885 		 * offset changes.
19886 		 */
19887 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
19888 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
19889 			 m, m->m_len,
19890 			 rack, rack->r_ctl.fsb.o_m_len,
19891 			 rack->r_ctl.fsb.off));
19892 
19893 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
19894 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
19895 		else
19896 			rack->r_ctl.fsb.off = 0;
19897 		rack->r_ctl.fsb.o_m_len = m->m_len;
19898 #ifdef INVARIANTS
19899 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
19900 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
19901 		      rack, m);
19902 #endif
19903 	}
19904 	soff = rack->r_ctl.fsb.off;
19905 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
19906 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
19907 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
19908 				 __FUNCTION__,
19909 				 rack, *plen, m, m->m_len));
19910 	/* Save off the right location before we copy and advance */
19911 	*s_soff = soff;
19912 	*s_mb = rack->r_ctl.fsb.m;
19913 	n = rack_fo_base_copym(m, soff, plen,
19914 			       &rack->r_ctl.fsb,
19915 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
19916 	return (n);
19917 }
19918 
19919 /* Log the buffer level */
19920 static void
19921 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
19922 		     int len, struct timeval *tv,
19923 		     uint32_t cts)
19924 {
19925 	uint32_t p_rate = 0, p_queue = 0, err = 0;
19926 	union tcp_log_stackspecific log;
19927 
19928 #ifdef RATELIMIT
19929 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
19930 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
19931 #endif
19932 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19933 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19934 	log.u_bbr.flex1 = p_rate;
19935 	log.u_bbr.flex2 = p_queue;
19936 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
19937 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
19938 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
19939 	log.u_bbr.flex7 = 99;
19940 	log.u_bbr.flex8 = 0;
19941 	log.u_bbr.pkts_out = err;
19942 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
19943 	log.u_bbr.timeStamp = cts;
19944 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19945 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
19946 		       len, &log, false, NULL, __func__, __LINE__, tv);
19947 
19948 }
19949 
19950 static uint32_t
19951 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
19952 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
19953 {
19954 	uint64_t lentime = 0;
19955 #ifdef RATELIMIT
19956 	uint32_t p_rate = 0, p_queue = 0, err;
19957 	union tcp_log_stackspecific log;
19958 	uint64_t bw;
19959 
19960 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
19961 	/* Failed or queue is zero */
19962 	if (err || (p_queue == 0)) {
19963 		lentime = 0;
19964 		goto out;
19965 	}
19966 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
19967 	if (err) {
19968 		lentime = 0;
19969 		goto out;
19970 	}
19971 	/*
19972 	 * If we reach here we have some bytes in
19973 	 * the queue. The number returned is a value
19974 	 * between 0 and 0xffff where ffff is full
19975 	 * and 0 is empty. So how best to make this into
19976 	 * something usable?
19977 	 *
19978 	 * The "safer" way is lets take the b/w gotten
19979 	 * from the query (which should be our b/w rate)
19980 	 * and pretend that a full send (our rc_pace_max_segs)
19981 	 * is outstanding. We factor it so its as if a full
19982 	 * number of our MSS segment is terms of full
19983 	 * ethernet segments are outstanding.
19984 	 */
19985 	bw = p_rate / 8;
19986 	if (bw) {
19987 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
19988 		lentime *= ETHERNET_SEGMENT_SIZE;
19989 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
19990 		lentime /= bw;
19991 	} else {
19992 		/* TSNH -- KASSERT? */
19993 		lentime = 0;
19994 	}
19995 out:
19996 	if (tcp_bblogging_on(tp)) {
19997 		memset(&log, 0, sizeof(log));
19998 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19999 		log.u_bbr.flex1 = p_rate;
20000 		log.u_bbr.flex2 = p_queue;
20001 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
20002 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
20003 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
20004 		log.u_bbr.flex7 = 99;
20005 		log.u_bbr.flex8 = 0;
20006 		log.u_bbr.pkts_out = err;
20007 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
20008 		log.u_bbr.cur_del_rate = lentime;
20009 		log.u_bbr.timeStamp = cts;
20010 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20011 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
20012 			       len, &log, false, NULL, __func__, __LINE__,tv);
20013 	}
20014 #endif
20015 	return ((uint32_t)lentime);
20016 }
20017 
20018 static int
20019 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
20020 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
20021 {
20022 	/*
20023 	 * Enter the fast retransmit path. We are given that a sched_pin is
20024 	 * in place (if accounting is compliled in) and the cycle count taken
20025 	 * at the entry is in the ts_val. The concept her is that the rsm
20026 	 * now holds the mbuf offsets and such so we can directly transmit
20027 	 * without a lot of overhead, the len field is already set for
20028 	 * us to prohibit us from sending too much (usually its 1MSS).
20029 	 */
20030 	struct ip *ip = NULL;
20031 	struct udphdr *udp = NULL;
20032 	struct tcphdr *th = NULL;
20033 	struct mbuf *m = NULL;
20034 	struct inpcb *inp;
20035 	uint8_t *cpto;
20036 	struct tcp_log_buffer *lgb;
20037 #ifdef TCP_ACCOUNTING
20038 	uint64_t crtsc;
20039 	int cnt_thru = 1;
20040 #endif
20041 	struct tcpopt to;
20042 	u_char opt[TCP_MAXOLEN];
20043 	uint32_t hdrlen, optlen;
20044 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
20045 	uint16_t flags;
20046 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
20047 	uint32_t if_hw_tsomaxsegsize;
20048 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
20049 
20050 #ifdef INET6
20051 	struct ip6_hdr *ip6 = NULL;
20052 
20053 	if (rack->r_is_v6) {
20054 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20055 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
20056 	} else
20057 #endif				/* INET6 */
20058 	{
20059 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20060 		hdrlen = sizeof(struct tcpiphdr);
20061 	}
20062 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
20063 		goto failed;
20064 	}
20065 	if (doing_tlp) {
20066 		/* Its a TLP add the flag, it may already be there but be sure */
20067 		rsm->r_flags |= RACK_TLP;
20068 	} else {
20069 		/* If it was a TLP it is not not on this retransmit */
20070 		rsm->r_flags &= ~RACK_TLP;
20071 	}
20072 	startseq = rsm->r_start;
20073 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20074 	inp = rack->rc_inp;
20075 	to.to_flags = 0;
20076 	flags = tcp_outflags[tp->t_state];
20077 	if (flags & (TH_SYN|TH_RST)) {
20078 		goto failed;
20079 	}
20080 	if (rsm->r_flags & RACK_HAS_FIN) {
20081 		/* We can't send a FIN here */
20082 		goto failed;
20083 	}
20084 	if (flags & TH_FIN) {
20085 		/* We never send a FIN */
20086 		flags &= ~TH_FIN;
20087 	}
20088 	if (tp->t_flags & TF_RCVD_TSTMP) {
20089 		to.to_tsval = ms_cts + tp->ts_offset;
20090 		to.to_tsecr = tp->ts_recent;
20091 		to.to_flags = TOF_TS;
20092 	}
20093 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
20094 	/* TCP-MD5 (RFC2385). */
20095 	if (tp->t_flags & TF_SIGNATURE)
20096 		to.to_flags |= TOF_SIGNATURE;
20097 #endif
20098 	optlen = tcp_addoptions(&to, opt);
20099 	hdrlen += optlen;
20100 	udp = rack->r_ctl.fsb.udp;
20101 	if (udp)
20102 		hdrlen += sizeof(struct udphdr);
20103 	if (rack->r_ctl.rc_pace_max_segs)
20104 		max_val = rack->r_ctl.rc_pace_max_segs;
20105 	else if (rack->rc_user_set_max_segs)
20106 		max_val = rack->rc_user_set_max_segs * segsiz;
20107 	else
20108 		max_val = len;
20109 	if ((tp->t_flags & TF_TSO) &&
20110 	    V_tcp_do_tso &&
20111 	    (len > segsiz) &&
20112 	    (tp->t_port == 0))
20113 		tso = 1;
20114 #ifdef INET6
20115 	if (MHLEN < hdrlen + max_linkhdr)
20116 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
20117 	else
20118 #endif
20119 		m = m_gethdr(M_NOWAIT, MT_DATA);
20120 	if (m == NULL)
20121 		goto failed;
20122 	m->m_data += max_linkhdr;
20123 	m->m_len = hdrlen;
20124 	th = rack->r_ctl.fsb.th;
20125 	/* Establish the len to send */
20126 	if (len > max_val)
20127 		len = max_val;
20128 	if ((tso) && (len + optlen > segsiz)) {
20129 		uint32_t if_hw_tsomax;
20130 		int32_t max_len;
20131 
20132 		/* extract TSO information */
20133 		if_hw_tsomax = tp->t_tsomax;
20134 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
20135 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
20136 		/*
20137 		 * Check if we should limit by maximum payload
20138 		 * length:
20139 		 */
20140 		if (if_hw_tsomax != 0) {
20141 			/* compute maximum TSO length */
20142 			max_len = (if_hw_tsomax - hdrlen -
20143 				   max_linkhdr);
20144 			if (max_len <= 0) {
20145 				goto failed;
20146 			} else if (len > max_len) {
20147 				len = max_len;
20148 			}
20149 		}
20150 		if (len <= segsiz) {
20151 			/*
20152 			 * In case there are too many small fragments don't
20153 			 * use TSO:
20154 			 */
20155 			tso = 0;
20156 		}
20157 	} else {
20158 		tso = 0;
20159 	}
20160 	if ((tso == 0) && (len > segsiz))
20161 		len = segsiz;
20162 	(void)tcp_get_usecs(tv);
20163 	if ((len == 0) ||
20164 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
20165 		goto failed;
20166 	}
20167 	th->th_seq = htonl(rsm->r_start);
20168 	th->th_ack = htonl(tp->rcv_nxt);
20169 	/*
20170 	 * The PUSH bit should only be applied
20171 	 * if the full retransmission is made. If
20172 	 * we are sending less than this is the
20173 	 * left hand edge and should not have
20174 	 * the PUSH bit.
20175 	 */
20176 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
20177 	    (len == (rsm->r_end - rsm->r_start)))
20178 		flags |= TH_PUSH;
20179 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
20180 	if (th->th_win == 0) {
20181 		tp->t_sndzerowin++;
20182 		tp->t_flags |= TF_RXWIN0SENT;
20183 	} else
20184 		tp->t_flags &= ~TF_RXWIN0SENT;
20185 	if (rsm->r_flags & RACK_TLP) {
20186 		/*
20187 		 * TLP should not count in retran count, but
20188 		 * in its own bin
20189 		 */
20190 		counter_u64_add(rack_tlp_retran, 1);
20191 		counter_u64_add(rack_tlp_retran_bytes, len);
20192 	} else {
20193 		tp->t_sndrexmitpack++;
20194 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
20195 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
20196 	}
20197 #ifdef STATS
20198 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
20199 				 len);
20200 #endif
20201 	if (rsm->m == NULL)
20202 		goto failed;
20203 	if (rsm->m &&
20204 	    ((rsm->orig_m_len != rsm->m->m_len) ||
20205 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
20206 		/* Fix up the orig_m_len and possibly the mbuf offset */
20207 		rack_adjust_orig_mlen(rsm);
20208 	}
20209 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
20210 	if (len <= segsiz) {
20211 		/*
20212 		 * Must have ran out of mbufs for the copy
20213 		 * shorten it to no longer need tso. Lets
20214 		 * not put on sendalot since we are low on
20215 		 * mbufs.
20216 		 */
20217 		tso = 0;
20218 	}
20219 	if ((m->m_next == NULL) || (len <= 0)){
20220 		goto failed;
20221 	}
20222 	if (udp) {
20223 		if (rack->r_is_v6)
20224 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
20225 		else
20226 			ulen = hdrlen + len - sizeof(struct ip);
20227 		udp->uh_ulen = htons(ulen);
20228 	}
20229 	m->m_pkthdr.rcvif = (struct ifnet *)0;
20230 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
20231 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
20232 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
20233 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
20234 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
20235 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
20236 #ifdef INET6
20237 		if (rack->r_is_v6) {
20238 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
20239 		    ip6->ip6_flow |= htonl(ect << 20);
20240 		}
20241 		else
20242 #endif
20243 		{
20244 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
20245 		    ip->ip_tos |= ect;
20246 		}
20247 	}
20248 	if (rack->r_ctl.crte != NULL) {
20249 		/* See if we can send via the hw queue */
20250 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
20251 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
20252 		if (slot == 0)
20253 			ip_sendflag = 0;
20254 	}
20255 	tcp_set_flags(th, flags);
20256 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
20257 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
20258 	if (to.to_flags & TOF_SIGNATURE) {
20259 		/*
20260 		 * Calculate MD5 signature and put it into the place
20261 		 * determined before.
20262 		 * NOTE: since TCP options buffer doesn't point into
20263 		 * mbuf's data, calculate offset and use it.
20264 		 */
20265 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
20266 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
20267 			/*
20268 			 * Do not send segment if the calculation of MD5
20269 			 * digest has failed.
20270 			 */
20271 			goto failed;
20272 		}
20273 	}
20274 #endif
20275 #ifdef INET6
20276 	if (rack->r_is_v6) {
20277 		if (tp->t_port) {
20278 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
20279 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
20280 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
20281 			th->th_sum = htons(0);
20282 			UDPSTAT_INC(udps_opackets);
20283 		} else {
20284 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
20285 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
20286 			th->th_sum = in6_cksum_pseudo(ip6,
20287 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
20288 						      0);
20289 		}
20290 	}
20291 #endif
20292 #if defined(INET6) && defined(INET)
20293 	else
20294 #endif
20295 #ifdef INET
20296 	{
20297 		if (tp->t_port) {
20298 			m->m_pkthdr.csum_flags = CSUM_UDP;
20299 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
20300 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
20301 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
20302 			th->th_sum = htons(0);
20303 			UDPSTAT_INC(udps_opackets);
20304 		} else {
20305 			m->m_pkthdr.csum_flags = CSUM_TCP;
20306 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
20307 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
20308 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
20309 									IPPROTO_TCP + len + optlen));
20310 		}
20311 		/* IP version must be set here for ipv4/ipv6 checking later */
20312 		KASSERT(ip->ip_v == IPVERSION,
20313 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
20314 	}
20315 #endif
20316 	if (tso) {
20317 		/*
20318 		 * Here we use segsiz since we have no added options besides
20319 		 * any standard timestamp options (no DSACKs or SACKS are sent
20320 		 * via either fast-path).
20321 		 */
20322 		KASSERT(len > segsiz,
20323 			("%s: len <= tso_segsz tp:%p", __func__, tp));
20324 		m->m_pkthdr.csum_flags |= CSUM_TSO;
20325 		m->m_pkthdr.tso_segsz = segsiz;
20326 	}
20327 #ifdef INET6
20328 	if (rack->r_is_v6) {
20329 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
20330 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
20331 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
20332 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
20333 		else
20334 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
20335 	}
20336 #endif
20337 #if defined(INET) && defined(INET6)
20338 	else
20339 #endif
20340 #ifdef INET
20341 	{
20342 		ip->ip_len = htons(m->m_pkthdr.len);
20343 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
20344 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
20345 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
20346 			if (tp->t_port == 0 || len < V_tcp_minmss) {
20347 				ip->ip_off |= htons(IP_DF);
20348 			}
20349 		} else {
20350 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
20351 		}
20352 	}
20353 #endif
20354 	if (doing_tlp == 0) {
20355 		/* Set we retransmitted */
20356 		rack->rc_gp_saw_rec = 1;
20357 	} else {
20358 		/* Its a TLP set ca or ss */
20359 		if (tp->snd_cwnd > tp->snd_ssthresh) {
20360 			/* Set we sent in CA */
20361 			rack->rc_gp_saw_ca = 1;
20362 		} else {
20363 			/* Set we sent in SS */
20364 			rack->rc_gp_saw_ss = 1;
20365 		}
20366 	}
20367 	/* Time to copy in our header */
20368 	cpto = mtod(m, uint8_t *);
20369 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
20370 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
20371 	if (optlen) {
20372 		bcopy(opt, th + 1, optlen);
20373 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
20374 	} else {
20375 		th->th_off = sizeof(struct tcphdr) >> 2;
20376 	}
20377 	if (tcp_bblogging_on(rack->rc_tp)) {
20378 		union tcp_log_stackspecific log;
20379 
20380 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
20381 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
20382 			counter_u64_add(rack_collapsed_win_rxt, 1);
20383 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
20384 		}
20385 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
20386 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
20387 		if (rack->rack_no_prr)
20388 			log.u_bbr.flex1 = 0;
20389 		else
20390 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
20391 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
20392 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
20393 		log.u_bbr.flex4 = max_val;
20394 		/* Save off the early/late values */
20395 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
20396 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
20397 		log.u_bbr.bw_inuse = rack_get_bw(rack);
20398 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
20399 		if (doing_tlp == 0)
20400 			log.u_bbr.flex8 = 1;
20401 		else
20402 			log.u_bbr.flex8 = 2;
20403 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
20404 		log.u_bbr.flex7 = 55;
20405 		log.u_bbr.pkts_out = tp->t_maxseg;
20406 		log.u_bbr.timeStamp = cts;
20407 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20408 		if (rsm && (rsm->r_rtr_cnt > 0)) {
20409 			/*
20410 			 * When we have a retransmit we want to log the
20411 			 * burst at send and flight at send from before.
20412 			 */
20413 			log.u_bbr.flex5 = rsm->r_fas;
20414 			log.u_bbr.bbr_substate = rsm->r_bas;
20415 		} else {
20416 			/*
20417 			 * This is currently unlikely until we do the
20418 			 * packet pair probes but I will add it for completeness.
20419 			 */
20420 			log.u_bbr.flex5 = log.u_bbr.inflight;
20421 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
20422 		}
20423 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
20424 		log.u_bbr.delivered = 0;
20425 		log.u_bbr.rttProp = (uint64_t)rsm;
20426 		log.u_bbr.delRate = rsm->r_flags;
20427 		log.u_bbr.delRate <<= 31;
20428 		log.u_bbr.delRate |= rack->r_must_retran;
20429 		log.u_bbr.delRate <<= 1;
20430 		log.u_bbr.delRate |= 1;
20431 		log.u_bbr.pkt_epoch = __LINE__;
20432 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
20433 				     len, &log, false, NULL, __func__, __LINE__, tv);
20434 	} else
20435 		lgb = NULL;
20436 	if ((rack->r_ctl.crte != NULL) &&
20437 	    tcp_bblogging_on(tp)) {
20438 		rack_log_queue_level(tp, rack, len, tv, cts);
20439 	}
20440 #ifdef INET6
20441 	if (rack->r_is_v6) {
20442 		error = ip6_output(m, inp->in6p_outputopts,
20443 				   &inp->inp_route6,
20444 				   ip_sendflag, NULL, NULL, inp);
20445 	}
20446 	else
20447 #endif
20448 #ifdef INET
20449 	{
20450 		error = ip_output(m, NULL,
20451 				  &inp->inp_route,
20452 				  ip_sendflag, 0, inp);
20453 	}
20454 #endif
20455 	m = NULL;
20456 	if (lgb) {
20457 		lgb->tlb_errno = error;
20458 		lgb = NULL;
20459 	}
20460 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
20461 	tp->snd_nxt = tp->snd_max;
20462 	if (error) {
20463 		goto failed;
20464 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
20465 		rack->rc_hw_nobuf = 0;
20466 		rack->r_ctl.rc_agg_delayed = 0;
20467 		rack->r_early = 0;
20468 		rack->r_late = 0;
20469 		rack->r_ctl.rc_agg_early = 0;
20470 	}
20471 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
20472 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
20473 	if (doing_tlp) {
20474 		rack->rc_tlp_in_progress = 1;
20475 		rack->r_ctl.rc_tlp_cnt_out++;
20476 	}
20477 	if (error == 0) {
20478 		counter_u64_add(rack_total_bytes, len);
20479 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
20480 		if (doing_tlp) {
20481 			rack->rc_last_sent_tlp_past_cumack = 0;
20482 			rack->rc_last_sent_tlp_seq_valid = 1;
20483 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
20484 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
20485 		}
20486 		if (rack->r_ctl.rc_prr_sndcnt >= len)
20487 			rack->r_ctl.rc_prr_sndcnt -= len;
20488 		else
20489 			rack->r_ctl.rc_prr_sndcnt = 0;
20490 	}
20491 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
20492 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
20493 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
20494 		rack->r_ctl.retran_during_recovery += len;
20495 	{
20496 		int idx;
20497 
20498 		idx = (len / segsiz) + 3;
20499 		if (idx >= TCP_MSS_ACCT_ATIMER)
20500 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
20501 		else
20502 			counter_u64_add(rack_out_size[idx], 1);
20503 	}
20504 	if (tp->t_rtttime == 0) {
20505 		tp->t_rtttime = ticks;
20506 		tp->t_rtseq = startseq;
20507 		KMOD_TCPSTAT_INC(tcps_segstimed);
20508 	}
20509 	counter_u64_add(rack_fto_rsm_send, 1);
20510 	if (error && (error == ENOBUFS)) {
20511 		if (rack->r_ctl.crte != NULL) {
20512 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
20513 			if (tcp_bblogging_on(rack->rc_tp))
20514 				rack_log_queue_level(tp, rack, len, tv, cts);
20515 		} else
20516 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
20517 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
20518 		if (rack->rc_enobuf < 0x7f)
20519 			rack->rc_enobuf++;
20520 		if (slot < (10 * HPTS_USEC_IN_MSEC))
20521 			slot = 10 * HPTS_USEC_IN_MSEC;
20522 		if (rack->r_ctl.crte != NULL) {
20523 			counter_u64_add(rack_saw_enobuf_hw, 1);
20524 			tcp_rl_log_enobuf(rack->r_ctl.crte);
20525 		}
20526 		counter_u64_add(rack_saw_enobuf, 1);
20527 	} else {
20528 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
20529 	}
20530 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
20531 #ifdef TCP_ACCOUNTING
20532 	crtsc = get_cyclecount();
20533 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20534 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
20535 	}
20536 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20537 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
20538 	}
20539 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20540 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
20541 	}
20542 	sched_unpin();
20543 #endif
20544 	return (0);
20545 failed:
20546 	if (m)
20547 		m_free(m);
20548 	return (-1);
20549 }
20550 
20551 static void
20552 rack_sndbuf_autoscale(struct tcp_rack *rack)
20553 {
20554 	/*
20555 	 * Automatic sizing of send socket buffer.  Often the send buffer
20556 	 * size is not optimally adjusted to the actual network conditions
20557 	 * at hand (delay bandwidth product).  Setting the buffer size too
20558 	 * small limits throughput on links with high bandwidth and high
20559 	 * delay (eg. trans-continental/oceanic links).  Setting the
20560 	 * buffer size too big consumes too much real kernel memory,
20561 	 * especially with many connections on busy servers.
20562 	 *
20563 	 * The criteria to step up the send buffer one notch are:
20564 	 *  1. receive window of remote host is larger than send buffer
20565 	 *     (with a fudge factor of 5/4th);
20566 	 *  2. send buffer is filled to 7/8th with data (so we actually
20567 	 *     have data to make use of it);
20568 	 *  3. send buffer fill has not hit maximal automatic size;
20569 	 *  4. our send window (slow start and cogestion controlled) is
20570 	 *     larger than sent but unacknowledged data in send buffer.
20571 	 *
20572 	 * Note that the rack version moves things much faster since
20573 	 * we want to avoid hitting cache lines in the rack_fast_output()
20574 	 * path so this is called much less often and thus moves
20575 	 * the SB forward by a percentage.
20576 	 */
20577 	struct socket *so;
20578 	struct tcpcb *tp;
20579 	uint32_t sendwin, scaleup;
20580 
20581 	tp = rack->rc_tp;
20582 	so = rack->rc_inp->inp_socket;
20583 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
20584 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
20585 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
20586 		    sbused(&so->so_snd) >=
20587 		    (so->so_snd.sb_hiwat / 8 * 7) &&
20588 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
20589 		    sendwin >= (sbused(&so->so_snd) -
20590 		    (tp->snd_max - tp->snd_una))) {
20591 			if (rack_autosndbuf_inc)
20592 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
20593 			else
20594 				scaleup = V_tcp_autosndbuf_inc;
20595 			if (scaleup < V_tcp_autosndbuf_inc)
20596 				scaleup = V_tcp_autosndbuf_inc;
20597 			scaleup += so->so_snd.sb_hiwat;
20598 			if (scaleup > V_tcp_autosndbuf_max)
20599 				scaleup = V_tcp_autosndbuf_max;
20600 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
20601 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
20602 		}
20603 	}
20604 }
20605 
20606 static int
20607 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
20608 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
20609 {
20610 	/*
20611 	 * Enter to do fast output. We are given that the sched_pin is
20612 	 * in place (if accounting is compiled in) and the cycle count taken
20613 	 * at entry is in place in ts_val. The idea here is that
20614 	 * we know how many more bytes needs to be sent (presumably either
20615 	 * during pacing or to fill the cwnd and that was greater than
20616 	 * the max-burst). We have how much to send and all the info we
20617 	 * need to just send.
20618 	 */
20619 #ifdef INET
20620 	struct ip *ip = NULL;
20621 #endif
20622 	struct udphdr *udp = NULL;
20623 	struct tcphdr *th = NULL;
20624 	struct mbuf *m, *s_mb;
20625 	struct inpcb *inp;
20626 	uint8_t *cpto;
20627 	struct tcp_log_buffer *lgb;
20628 #ifdef TCP_ACCOUNTING
20629 	uint64_t crtsc;
20630 #endif
20631 	struct tcpopt to;
20632 	u_char opt[TCP_MAXOLEN];
20633 	uint32_t hdrlen, optlen;
20634 #ifdef TCP_ACCOUNTING
20635 	int cnt_thru = 1;
20636 #endif
20637 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
20638 	uint16_t flags;
20639 	uint32_t s_soff;
20640 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
20641 	uint32_t if_hw_tsomaxsegsize;
20642 	uint32_t add_flag = RACK_SENT_FP;
20643 #ifdef INET6
20644 	struct ip6_hdr *ip6 = NULL;
20645 
20646 	if (rack->r_is_v6) {
20647 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20648 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
20649 	} else
20650 #endif				/* INET6 */
20651 	{
20652 #ifdef INET
20653 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20654 		hdrlen = sizeof(struct tcpiphdr);
20655 #endif
20656 	}
20657 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
20658 		m = NULL;
20659 		goto failed;
20660 	}
20661 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20662 	startseq = tp->snd_max;
20663 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20664 	inp = rack->rc_inp;
20665 	len = rack->r_ctl.fsb.left_to_send;
20666 	to.to_flags = 0;
20667 	flags = rack->r_ctl.fsb.tcp_flags;
20668 	if (tp->t_flags & TF_RCVD_TSTMP) {
20669 		to.to_tsval = ms_cts + tp->ts_offset;
20670 		to.to_tsecr = tp->ts_recent;
20671 		to.to_flags = TOF_TS;
20672 	}
20673 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
20674 	/* TCP-MD5 (RFC2385). */
20675 	if (tp->t_flags & TF_SIGNATURE)
20676 		to.to_flags |= TOF_SIGNATURE;
20677 #endif
20678 	optlen = tcp_addoptions(&to, opt);
20679 	hdrlen += optlen;
20680 	udp = rack->r_ctl.fsb.udp;
20681 	if (udp)
20682 		hdrlen += sizeof(struct udphdr);
20683 	if (rack->r_ctl.rc_pace_max_segs)
20684 		max_val = rack->r_ctl.rc_pace_max_segs;
20685 	else if (rack->rc_user_set_max_segs)
20686 		max_val = rack->rc_user_set_max_segs * segsiz;
20687 	else
20688 		max_val = len;
20689 	if ((tp->t_flags & TF_TSO) &&
20690 	    V_tcp_do_tso &&
20691 	    (len > segsiz) &&
20692 	    (tp->t_port == 0))
20693 		tso = 1;
20694 again:
20695 #ifdef INET6
20696 	if (MHLEN < hdrlen + max_linkhdr)
20697 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
20698 	else
20699 #endif
20700 		m = m_gethdr(M_NOWAIT, MT_DATA);
20701 	if (m == NULL)
20702 		goto failed;
20703 	m->m_data += max_linkhdr;
20704 	m->m_len = hdrlen;
20705 	th = rack->r_ctl.fsb.th;
20706 	/* Establish the len to send */
20707 	if (len > max_val)
20708 		len = max_val;
20709 	if ((tso) && (len + optlen > segsiz)) {
20710 		uint32_t if_hw_tsomax;
20711 		int32_t max_len;
20712 
20713 		/* extract TSO information */
20714 		if_hw_tsomax = tp->t_tsomax;
20715 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
20716 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
20717 		/*
20718 		 * Check if we should limit by maximum payload
20719 		 * length:
20720 		 */
20721 		if (if_hw_tsomax != 0) {
20722 			/* compute maximum TSO length */
20723 			max_len = (if_hw_tsomax - hdrlen -
20724 				   max_linkhdr);
20725 			if (max_len <= 0) {
20726 				goto failed;
20727 			} else if (len > max_len) {
20728 				len = max_len;
20729 			}
20730 		}
20731 		if (len <= segsiz) {
20732 			/*
20733 			 * In case there are too many small fragments don't
20734 			 * use TSO:
20735 			 */
20736 			tso = 0;
20737 		}
20738 	} else {
20739 		tso = 0;
20740 	}
20741 	if ((tso == 0) && (len > segsiz))
20742 		len = segsiz;
20743 	(void)tcp_get_usecs(tv);
20744 	if ((len == 0) ||
20745 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
20746 		goto failed;
20747 	}
20748 	sb_offset = tp->snd_max - tp->snd_una;
20749 	th->th_seq = htonl(tp->snd_max);
20750 	th->th_ack = htonl(tp->rcv_nxt);
20751 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
20752 	if (th->th_win == 0) {
20753 		tp->t_sndzerowin++;
20754 		tp->t_flags |= TF_RXWIN0SENT;
20755 	} else
20756 		tp->t_flags &= ~TF_RXWIN0SENT;
20757 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
20758 	KMOD_TCPSTAT_INC(tcps_sndpack);
20759 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
20760 #ifdef STATS
20761 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
20762 				 len);
20763 #endif
20764 	if (rack->r_ctl.fsb.m == NULL)
20765 		goto failed;
20766 
20767 	/* s_mb and s_soff are saved for rack_log_output */
20768 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
20769 				    &s_mb, &s_soff);
20770 	if (len <= segsiz) {
20771 		/*
20772 		 * Must have ran out of mbufs for the copy
20773 		 * shorten it to no longer need tso. Lets
20774 		 * not put on sendalot since we are low on
20775 		 * mbufs.
20776 		 */
20777 		tso = 0;
20778 	}
20779 	if (rack->r_ctl.fsb.rfo_apply_push &&
20780 	    (len == rack->r_ctl.fsb.left_to_send)) {
20781 		tcp_set_flags(th, flags | TH_PUSH);
20782 		add_flag |= RACK_HAD_PUSH;
20783 	}
20784 	if ((m->m_next == NULL) || (len <= 0)){
20785 		goto failed;
20786 	}
20787 	if (udp) {
20788 		if (rack->r_is_v6)
20789 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
20790 		else
20791 			ulen = hdrlen + len - sizeof(struct ip);
20792 		udp->uh_ulen = htons(ulen);
20793 	}
20794 	m->m_pkthdr.rcvif = (struct ifnet *)0;
20795 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
20796 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
20797 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
20798 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
20799 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
20800 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
20801 #ifdef INET6
20802 		if (rack->r_is_v6) {
20803 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
20804 			ip6->ip6_flow |= htonl(ect << 20);
20805 		}
20806 		else
20807 #endif
20808 		{
20809 #ifdef INET
20810 			ip->ip_tos &= ~IPTOS_ECN_MASK;
20811 			ip->ip_tos |= ect;
20812 #endif
20813 		}
20814 	}
20815 	tcp_set_flags(th, flags);
20816 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
20817 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
20818 	if (to.to_flags & TOF_SIGNATURE) {
20819 		/*
20820 		 * Calculate MD5 signature and put it into the place
20821 		 * determined before.
20822 		 * NOTE: since TCP options buffer doesn't point into
20823 		 * mbuf's data, calculate offset and use it.
20824 		 */
20825 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
20826 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
20827 			/*
20828 			 * Do not send segment if the calculation of MD5
20829 			 * digest has failed.
20830 			 */
20831 			goto failed;
20832 		}
20833 	}
20834 #endif
20835 #ifdef INET6
20836 	if (rack->r_is_v6) {
20837 		if (tp->t_port) {
20838 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
20839 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
20840 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
20841 			th->th_sum = htons(0);
20842 			UDPSTAT_INC(udps_opackets);
20843 		} else {
20844 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
20845 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
20846 			th->th_sum = in6_cksum_pseudo(ip6,
20847 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
20848 						      0);
20849 		}
20850 	}
20851 #endif
20852 #if defined(INET6) && defined(INET)
20853 	else
20854 #endif
20855 #ifdef INET
20856 	{
20857 		if (tp->t_port) {
20858 			m->m_pkthdr.csum_flags = CSUM_UDP;
20859 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
20860 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
20861 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
20862 			th->th_sum = htons(0);
20863 			UDPSTAT_INC(udps_opackets);
20864 		} else {
20865 			m->m_pkthdr.csum_flags = CSUM_TCP;
20866 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
20867 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
20868 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
20869 									IPPROTO_TCP + len + optlen));
20870 		}
20871 		/* IP version must be set here for ipv4/ipv6 checking later */
20872 		KASSERT(ip->ip_v == IPVERSION,
20873 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
20874 	}
20875 #endif
20876 	if (tso) {
20877 		/*
20878 		 * Here we use segsiz since we have no added options besides
20879 		 * any standard timestamp options (no DSACKs or SACKS are sent
20880 		 * via either fast-path).
20881 		 */
20882 		KASSERT(len > segsiz,
20883 			("%s: len <= tso_segsz tp:%p", __func__, tp));
20884 		m->m_pkthdr.csum_flags |= CSUM_TSO;
20885 		m->m_pkthdr.tso_segsz = segsiz;
20886 	}
20887 #ifdef INET6
20888 	if (rack->r_is_v6) {
20889 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
20890 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
20891 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
20892 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
20893 		else
20894 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
20895 	}
20896 #endif
20897 #if defined(INET) && defined(INET6)
20898 	else
20899 #endif
20900 #ifdef INET
20901 	{
20902 		ip->ip_len = htons(m->m_pkthdr.len);
20903 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
20904 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
20905 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
20906 			if (tp->t_port == 0 || len < V_tcp_minmss) {
20907 				ip->ip_off |= htons(IP_DF);
20908 			}
20909 		} else {
20910 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
20911 		}
20912 	}
20913 #endif
20914 	if (tp->snd_cwnd > tp->snd_ssthresh) {
20915 		/* Set we sent in CA */
20916 		rack->rc_gp_saw_ca = 1;
20917 	} else {
20918 		/* Set we sent in SS */
20919 		rack->rc_gp_saw_ss = 1;
20920 	}
20921 	/* Time to copy in our header */
20922 	cpto = mtod(m, uint8_t *);
20923 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
20924 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
20925 	if (optlen) {
20926 		bcopy(opt, th + 1, optlen);
20927 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
20928 	} else {
20929 		th->th_off = sizeof(struct tcphdr) >> 2;
20930 	}
20931 	if ((rack->r_ctl.crte != NULL) &&
20932 	    tcp_bblogging_on(tp)) {
20933 		rack_log_queue_level(tp, rack, len, tv, cts);
20934 	}
20935 	if (tcp_bblogging_on(rack->rc_tp)) {
20936 		union tcp_log_stackspecific log;
20937 
20938 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
20939 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
20940 		if (rack->rack_no_prr)
20941 			log.u_bbr.flex1 = 0;
20942 		else
20943 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
20944 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
20945 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
20946 		log.u_bbr.flex4 = max_val;
20947 		/* Save off the early/late values */
20948 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
20949 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
20950 		log.u_bbr.bw_inuse = rack_get_bw(rack);
20951 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
20952 		log.u_bbr.flex8 = 0;
20953 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
20954 		log.u_bbr.flex7 = 44;
20955 		log.u_bbr.pkts_out = tp->t_maxseg;
20956 		log.u_bbr.timeStamp = cts;
20957 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20958 		log.u_bbr.flex5 = log.u_bbr.inflight;
20959 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
20960 		log.u_bbr.delivered = 0;
20961 		log.u_bbr.rttProp = 0;
20962 		log.u_bbr.delRate = rack->r_must_retran;
20963 		log.u_bbr.delRate <<= 1;
20964 		log.u_bbr.pkt_epoch = __LINE__;
20965 		/* For fast output no retrans so just inflight and how many mss we send */
20966 		log.u_bbr.flex5 = log.u_bbr.inflight;
20967 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
20968 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
20969 				     len, &log, false, NULL, __func__, __LINE__, tv);
20970 	} else
20971 		lgb = NULL;
20972 #ifdef INET6
20973 	if (rack->r_is_v6) {
20974 		error = ip6_output(m, inp->in6p_outputopts,
20975 				   &inp->inp_route6,
20976 				   0, NULL, NULL, inp);
20977 	}
20978 #endif
20979 #if defined(INET) && defined(INET6)
20980 	else
20981 #endif
20982 #ifdef INET
20983 	{
20984 		error = ip_output(m, NULL,
20985 				  &inp->inp_route,
20986 				  0, 0, inp);
20987 	}
20988 #endif
20989 	if (lgb) {
20990 		lgb->tlb_errno = error;
20991 		lgb = NULL;
20992 	}
20993 	if (error) {
20994 		*send_err = error;
20995 		m = NULL;
20996 		goto failed;
20997 	} else if (rack->rc_hw_nobuf) {
20998 		rack->rc_hw_nobuf = 0;
20999 		rack->r_ctl.rc_agg_delayed = 0;
21000 		rack->r_early = 0;
21001 		rack->r_late = 0;
21002 		rack->r_ctl.rc_agg_early = 0;
21003 	}
21004 	if ((error == 0) && (rack->lt_bw_up == 0)) {
21005 		/* Unlikely */
21006 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
21007 		rack->r_ctl.lt_seq = tp->snd_una;
21008 		rack->lt_bw_up = 1;
21009 	} else if ((error == 0) &&
21010 		   (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
21011 		/*
21012 		 * Need to record what we have since we are
21013 		 * approaching seq wrap.
21014 		 */
21015 		struct timeval tv;
21016 		uint64_t tmark;
21017 
21018 		rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
21019 		rack->r_ctl.lt_seq = tp->snd_una;
21020 		tmark = tcp_get_u64_usecs(&tv);
21021 		if (tmark > rack->r_ctl.lt_timemark) {
21022 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
21023 			rack->r_ctl.lt_timemark = tmark;
21024 		}
21025 	}
21026 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
21027 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
21028 	m = NULL;
21029 	if (tp->snd_una == tp->snd_max) {
21030 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
21031 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
21032 		tp->t_acktime = ticks;
21033 	}
21034 	counter_u64_add(rack_total_bytes, len);
21035 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
21036 
21037 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
21038 	tot_len += len;
21039 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
21040 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
21041 	tp->snd_max += len;
21042 	tp->snd_nxt = tp->snd_max;
21043 	if (rack->rc_new_rnd_needed) {
21044 		rack_new_round_starts(tp, rack, tp->snd_max);
21045 	}
21046 	{
21047 		int idx;
21048 
21049 		idx = (len / segsiz) + 3;
21050 		if (idx >= TCP_MSS_ACCT_ATIMER)
21051 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
21052 		else
21053 			counter_u64_add(rack_out_size[idx], 1);
21054 	}
21055 	if (len <= rack->r_ctl.fsb.left_to_send)
21056 		rack->r_ctl.fsb.left_to_send -= len;
21057 	else
21058 		rack->r_ctl.fsb.left_to_send = 0;
21059 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
21060 		rack->r_fast_output = 0;
21061 		rack->r_ctl.fsb.left_to_send = 0;
21062 		/* At the end of fast_output scale up the sb */
21063 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
21064 		rack_sndbuf_autoscale(rack);
21065 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
21066 	}
21067 	if (tp->t_rtttime == 0) {
21068 		tp->t_rtttime = ticks;
21069 		tp->t_rtseq = startseq;
21070 		KMOD_TCPSTAT_INC(tcps_segstimed);
21071 	}
21072 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
21073 	    (max_val > len) &&
21074 	    (tso == 0)) {
21075 		max_val -= len;
21076 		len = segsiz;
21077 		th = rack->r_ctl.fsb.th;
21078 #ifdef TCP_ACCOUNTING
21079 		cnt_thru++;
21080 #endif
21081 		goto again;
21082 	}
21083 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
21084 	counter_u64_add(rack_fto_send, 1);
21085 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz, __LINE__);
21086 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
21087 #ifdef TCP_ACCOUNTING
21088 	crtsc = get_cyclecount();
21089 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21090 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
21091 	}
21092 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21093 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21094 	}
21095 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21096 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
21097 	}
21098 	sched_unpin();
21099 #endif
21100 	return (0);
21101 failed:
21102 	if (m)
21103 		m_free(m);
21104 	rack->r_fast_output = 0;
21105 	return (-1);
21106 }
21107 
21108 static inline void
21109 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
21110 		       struct sockbuf *sb,
21111 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
21112 		       bool hw_tls,
21113 		       uint16_t flags)
21114 {
21115 	rack->r_fast_output = 1;
21116 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
21117 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
21118 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
21119 	rack->r_ctl.fsb.tcp_flags = flags;
21120 	rack->r_ctl.fsb.left_to_send = orig_len - len;
21121 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
21122 		/* Less than a full sized pace, lets not  */
21123 		rack->r_fast_output = 0;
21124 		return;
21125 	} else {
21126 		/* Round down to the nearest pace_max_seg */
21127 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
21128 	}
21129 	if (hw_tls)
21130 		rack->r_ctl.fsb.hw_tls = 1;
21131 	else
21132 		rack->r_ctl.fsb.hw_tls = 0;
21133 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
21134 		("rack:%p left_to_send:%u sbavail:%u out:%u",
21135 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
21136 		 (tp->snd_max - tp->snd_una)));
21137 	if (rack->r_ctl.fsb.left_to_send < segsiz)
21138 		rack->r_fast_output = 0;
21139 	else {
21140 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
21141 			rack->r_ctl.fsb.rfo_apply_push = 1;
21142 		else
21143 			rack->r_ctl.fsb.rfo_apply_push = 0;
21144 	}
21145 }
21146 
21147 static uint32_t
21148 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
21149 {
21150 	uint64_t min_time;
21151 	uint32_t maxlen;
21152 
21153 	min_time = (uint64_t)get_hpts_min_sleep_time();
21154 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
21155 	maxlen = roundup(maxlen, segsiz);
21156 	return (maxlen);
21157 }
21158 
21159 static struct rack_sendmap *
21160 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
21161 {
21162 	struct rack_sendmap *rsm = NULL;
21163 	int thresh;
21164 
21165 restart:
21166 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
21167 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
21168 		/* Nothing, strange turn off validity  */
21169 		rack->r_collapse_point_valid = 0;
21170 		return (NULL);
21171 	}
21172 	/* Can we send it yet? */
21173 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
21174 		/*
21175 		 * Receiver window has not grown enough for
21176 		 * the segment to be put on the wire.
21177 		 */
21178 		return (NULL);
21179 	}
21180 	if (rsm->r_flags & RACK_ACKED) {
21181 		/*
21182 		 * It has been sacked, lets move to the
21183 		 * next one if possible.
21184 		 */
21185 		rack->r_ctl.last_collapse_point = rsm->r_end;
21186 		/* Are we done? */
21187 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
21188 			    rack->r_ctl.high_collapse_point)) {
21189 			rack->r_collapse_point_valid = 0;
21190 			return (NULL);
21191 		}
21192 		goto restart;
21193 	}
21194 	/* Now has it been long enough ? */
21195 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
21196 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
21197 		rack_log_collapse(rack, rsm->r_start,
21198 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
21199 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
21200 		return (rsm);
21201 	}
21202 	/* Not enough time */
21203 	rack_log_collapse(rack, rsm->r_start,
21204 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
21205 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
21206 	return (NULL);
21207 }
21208 
21209 static void
21210 rack_credit_back_policer_idle_time(struct tcp_rack *rack, uint64_t idle_t, int line)
21211 {
21212 	/*
21213 	 * We were idle some time (idle_t) and so our policer bucket
21214 	 * needs to grow. It can go no higher than policer_bucket_size.
21215 	 */
21216 	uint64_t len;
21217 
21218 	len = idle_t * rack->r_ctl.policer_bw;
21219 	len /= HPTS_USEC_IN_SEC;
21220 	rack->r_ctl.current_policer_bucket += (uint32_t)len;
21221 	if (rack->r_ctl.policer_bucket_size < rack->r_ctl.current_policer_bucket) {
21222 		rack->r_ctl.current_policer_bucket = rack->r_ctl.policer_bucket_size;
21223 	}
21224 	if (rack_verbose_logging > 0)
21225 		policer_detection_log(rack, (uint32_t)len, line, (uint32_t)idle_t, 0, 7);
21226 }
21227 
21228 static inline void
21229 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
21230 {
21231 	if ((rack->full_size_rxt == 0) &&
21232 	    (rack->shape_rxt_to_pacing_min == 0) &&
21233 	    (*len >= segsiz)) {
21234 		*len = segsiz;
21235 	} else if (rack->shape_rxt_to_pacing_min &&
21236 		 rack->gp_ready) {
21237 		/* We use pacing min as shaping len req */
21238 		uint32_t maxlen;
21239 
21240 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
21241 		if (*len > maxlen)
21242 			*len = maxlen;
21243 	} else {
21244 		/*
21245 		 * The else is full_size_rxt is on so send it all
21246 		 * note we do need to check this for exceeding
21247 		 * our max segment size due to the fact that
21248 		 * we do sometimes merge chunks together i.e.
21249 		 * we cannot just assume that we will never have
21250 		 * a chunk greater than pace_max_seg
21251 		 */
21252 		if (*len > pace_max_seg)
21253 			*len = pace_max_seg;
21254 	}
21255 }
21256 
21257 static int
21258 rack_output(struct tcpcb *tp)
21259 {
21260 	struct socket *so;
21261 	uint32_t recwin;
21262 	uint32_t sb_offset, s_moff = 0;
21263 	int32_t len, error = 0;
21264 	uint16_t flags;
21265 	struct mbuf *m, *s_mb = NULL;
21266 	struct mbuf *mb;
21267 	uint32_t if_hw_tsomaxsegcount = 0;
21268 	uint32_t if_hw_tsomaxsegsize;
21269 	int32_t segsiz, minseg;
21270 	long tot_len_this_send = 0;
21271 #ifdef INET
21272 	struct ip *ip = NULL;
21273 #endif
21274 	struct udphdr *udp = NULL;
21275 	struct tcp_rack *rack;
21276 	struct tcphdr *th;
21277 	uint8_t pass = 0;
21278 	uint8_t mark = 0;
21279 	uint8_t check_done = 0;
21280 	uint8_t wanted_cookie = 0;
21281 	u_char opt[TCP_MAXOLEN];
21282 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
21283 	uint32_t rack_seq;
21284 
21285 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21286 	unsigned ipsec_optlen = 0;
21287 
21288 #endif
21289 	int32_t idle, sendalot;
21290 	uint32_t tot_idle;
21291 	int32_t sub_from_prr = 0;
21292 	volatile int32_t sack_rxmit;
21293 	struct rack_sendmap *rsm = NULL;
21294 	int32_t tso, mtu;
21295 	struct tcpopt to;
21296 	int32_t slot = 0;
21297 	int32_t sup_rack = 0;
21298 	uint32_t cts, ms_cts, delayed, early;
21299 	uint32_t add_flag = RACK_SENT_SP;
21300 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
21301 	uint8_t doing_tlp = 0;
21302 	uint32_t cwnd_to_use, pace_max_seg;
21303 	int32_t do_a_prefetch = 0;
21304 	int32_t prefetch_rsm = 0;
21305 	int32_t orig_len = 0;
21306 	struct timeval tv;
21307 	int32_t prefetch_so_done = 0;
21308 	struct tcp_log_buffer *lgb;
21309 	struct inpcb *inp = tptoinpcb(tp);
21310 	struct sockbuf *sb;
21311 	uint64_t ts_val = 0;
21312 #ifdef TCP_ACCOUNTING
21313 	uint64_t crtsc;
21314 #endif
21315 #ifdef INET6
21316 	struct ip6_hdr *ip6 = NULL;
21317 	int32_t isipv6;
21318 #endif
21319 	bool hpts_calling, hw_tls = false;
21320 
21321 	NET_EPOCH_ASSERT();
21322 	INP_WLOCK_ASSERT(inp);
21323 
21324 	/* setup and take the cache hits here */
21325 	rack = (struct tcp_rack *)tp->t_fb_ptr;
21326 #ifdef TCP_ACCOUNTING
21327 	sched_pin();
21328 	ts_val = get_cyclecount();
21329 #endif
21330 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
21331 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
21332 #ifdef TCP_OFFLOAD
21333 	if (tp->t_flags & TF_TOE) {
21334 #ifdef TCP_ACCOUNTING
21335 		sched_unpin();
21336 #endif
21337 		return (tcp_offload_output(tp));
21338 	}
21339 #endif
21340 	if (rack->rack_deferred_inited == 0) {
21341 		/*
21342 		 * If we are the connecting socket we will
21343 		 * hit rack_init() when no sequence numbers
21344 		 * are setup. This makes it so we must defer
21345 		 * some initialization. Call that now.
21346 		 */
21347 		rack_deferred_init(tp, rack);
21348 	}
21349 	/*
21350 	 * For TFO connections in SYN_RECEIVED, only allow the initial
21351 	 * SYN|ACK and those sent by the retransmit timer.
21352 	 */
21353 	if ((tp->t_flags & TF_FASTOPEN) &&
21354 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
21355 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
21356 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
21357 #ifdef TCP_ACCOUNTING
21358 		sched_unpin();
21359 #endif
21360 		return (0);
21361 	}
21362 #ifdef INET6
21363 	if (rack->r_state) {
21364 		/* Use the cache line loaded if possible */
21365 		isipv6 = rack->r_is_v6;
21366 	} else {
21367 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
21368 	}
21369 #endif
21370 	early = 0;
21371 	cts = tcp_get_usecs(&tv);
21372 	ms_cts = tcp_tv_to_mssectick(&tv);
21373 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
21374 	    tcp_in_hpts(rack->rc_tp)) {
21375 		/*
21376 		 * We are on the hpts for some timer but not hptsi output.
21377 		 * Remove from the hpts unconditionally.
21378 		 */
21379 		rack_timer_cancel(tp, rack, cts, __LINE__);
21380 	}
21381 	/* Are we pacing and late? */
21382 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
21383 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
21384 		/* We are delayed */
21385 		delayed = cts - rack->r_ctl.rc_last_output_to;
21386 	} else {
21387 		delayed = 0;
21388 	}
21389 	/* Do the timers, which may override the pacer */
21390 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
21391 		int retval;
21392 
21393 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
21394 					     &doing_tlp);
21395 		if (retval != 0) {
21396 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
21397 #ifdef TCP_ACCOUNTING
21398 			sched_unpin();
21399 #endif
21400 			/*
21401 			 * If timers want tcp_drop(), then pass error out,
21402 			 * otherwise suppress it.
21403 			 */
21404 			return (retval < 0 ? retval : 0);
21405 		}
21406 	}
21407 	if (rack->rc_in_persist) {
21408 		if (tcp_in_hpts(rack->rc_tp) == 0) {
21409 			/* Timer is not running */
21410 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
21411 		}
21412 #ifdef TCP_ACCOUNTING
21413 		sched_unpin();
21414 #endif
21415 		return (0);
21416 	}
21417 	if ((rack->rc_ack_required == 1) &&
21418 	    (rack->r_timer_override == 0)){
21419 		/* A timeout occurred and no ack has arrived */
21420 		if (tcp_in_hpts(rack->rc_tp) == 0) {
21421 			/* Timer is not running */
21422 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
21423 		}
21424 #ifdef TCP_ACCOUNTING
21425 		sched_unpin();
21426 #endif
21427 		return (0);
21428 	}
21429 	if ((rack->r_timer_override) ||
21430 	    (rack->rc_ack_can_sendout_data) ||
21431 	    (delayed) ||
21432 	    (tp->t_state < TCPS_ESTABLISHED)) {
21433 		rack->rc_ack_can_sendout_data = 0;
21434 		if (tcp_in_hpts(rack->rc_tp))
21435 			tcp_hpts_remove(rack->rc_tp);
21436 	} else if (tcp_in_hpts(rack->rc_tp)) {
21437 		/*
21438 		 * On the hpts you can't pass even if ACKNOW is on, we will
21439 		 * when the hpts fires.
21440 		 */
21441 #ifdef TCP_ACCOUNTING
21442 		crtsc = get_cyclecount();
21443 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21444 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
21445 		}
21446 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21447 			tp->tcp_cnt_counters[SND_BLOCKED]++;
21448 		}
21449 		sched_unpin();
21450 #endif
21451 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
21452 		return (0);
21453 	}
21454 	/* Finish out both pacing early and late accounting */
21455 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
21456 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
21457 		early = rack->r_ctl.rc_last_output_to - cts;
21458 	} else
21459 		early = 0;
21460 	if (delayed && (rack->rc_always_pace == 1)) {
21461 		rack->r_ctl.rc_agg_delayed += delayed;
21462 		rack->r_late = 1;
21463 	} else if (early && (rack->rc_always_pace == 1)) {
21464 		rack->r_ctl.rc_agg_early += early;
21465 		rack->r_early = 1;
21466 	} else if (rack->rc_always_pace == 0) {
21467 		/* Non-paced we are not late */
21468 		rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
21469 		rack->r_early = rack->r_late = 0;
21470 	}
21471 	/* Now that early/late accounting is done turn off the flag */
21472 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
21473 	rack->r_wanted_output = 0;
21474 	rack->r_timer_override = 0;
21475 	if ((tp->t_state != rack->r_state) &&
21476 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
21477 		rack_set_state(tp, rack);
21478 	}
21479 	if ((rack->r_fast_output) &&
21480 	    (doing_tlp == 0) &&
21481 	    (tp->rcv_numsacks == 0)) {
21482 		int ret;
21483 
21484 		error = 0;
21485 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
21486 		if (ret >= 0)
21487 			return(ret);
21488 		else if (error) {
21489 			inp = rack->rc_inp;
21490 			so = inp->inp_socket;
21491 			sb = &so->so_snd;
21492 			goto nomore;
21493 		}
21494 	}
21495 	inp = rack->rc_inp;
21496 	/*
21497 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
21498 	 * only allow the initial SYN or SYN|ACK and those sent
21499 	 * by the retransmit timer.
21500 	 */
21501 	if ((tp->t_flags & TF_FASTOPEN) &&
21502 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
21503 	     (tp->t_state == TCPS_SYN_SENT)) &&
21504 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
21505 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
21506 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
21507 		so = inp->inp_socket;
21508 		sb = &so->so_snd;
21509 		goto just_return_nolock;
21510 	}
21511 	/*
21512 	 * Determine length of data that should be transmitted, and flags
21513 	 * that will be used. If there is some data or critical controls
21514 	 * (SYN, RST) to send, then transmit; otherwise, investigate
21515 	 * further.
21516 	 */
21517 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
21518 	if (tp->t_idle_reduce) {
21519 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
21520 			rack_cc_after_idle(rack, tp);
21521 	}
21522 	tp->t_flags &= ~TF_LASTIDLE;
21523 	if (idle) {
21524 		if (tp->t_flags & TF_MORETOCOME) {
21525 			tp->t_flags |= TF_LASTIDLE;
21526 			idle = 0;
21527 		}
21528 	}
21529 	if ((tp->snd_una == tp->snd_max) &&
21530 	    rack->r_ctl.rc_went_idle_time &&
21531 	    (cts > rack->r_ctl.rc_went_idle_time)) {
21532 		tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
21533 		if (tot_idle > rack_min_probertt_hold) {
21534 			/* Count as a probe rtt */
21535 			if (rack->in_probe_rtt == 0) {
21536 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
21537 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
21538 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
21539 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
21540 			} else {
21541 				rack_exit_probertt(rack, cts);
21542 			}
21543 		}
21544 	}
21545 	if(rack->policer_detect_on) {
21546 		/*
21547 		 * If we are doing policer detetion we at a minium
21548 		 * record the time but if possible add back to
21549 		 * the bucket based on the idle time.
21550 		 */
21551 		uint64_t idle_t, u64_cts;
21552 
21553 		segsiz = min(ctf_fixed_maxseg(tp),
21554 			     rack->r_ctl.rc_pace_min_segs);
21555 		u64_cts = tcp_tv_to_lusectick(&tv);
21556 		if ((rack->rc_policer_detected == 1) &&
21557 		    (rack->r_ctl.policer_bucket_size > segsiz) &&
21558 		    (rack->r_ctl.policer_bw > 0) &&
21559 		    (u64_cts > rack->r_ctl.last_sendtime)) {
21560 			/* We are being policed add back the time */
21561 			idle_t = u64_cts - rack->r_ctl.last_sendtime;
21562 			rack_credit_back_policer_idle_time(rack, idle_t, __LINE__);
21563 		}
21564 		rack->r_ctl.last_sendtime = u64_cts;
21565 	}
21566 	if (rack_use_fsb &&
21567 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
21568 	    (rack->r_fsb_inited == 0) &&
21569 	    (rack->r_state != TCPS_CLOSED))
21570 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
21571 	if (rack->rc_sendvars_notset == 1) {
21572 		rack->r_ctl.idle_snd_una = tp->snd_una;
21573 		rack->rc_sendvars_notset = 0;
21574 		/*
21575 		 * Make sure any TCP timers (keep-alive) is not running.
21576 		 */
21577 		tcp_timer_stop(tp);
21578 	}
21579 	if ((rack->rack_no_prr == 1) &&
21580 	    (rack->rc_always_pace == 0)) {
21581 		/*
21582 		 * Sanity check before sending, if we have
21583 		 * no-pacing enabled and prr is turned off that
21584 		 * is a logistics error. Correct this by turnning
21585 		 * prr back on. A user *must* set some form of
21586 		 * pacing in order to turn PRR off. We do this
21587 		 * in the output path so that we can avoid socket
21588 		 * option ordering issues that would occur if we
21589 		 * tried to do it while setting rack_no_prr on.
21590 		 */
21591 		rack->rack_no_prr = 0;
21592 	}
21593 	if ((rack->pcm_enabled == 1) &&
21594 	    (rack->pcm_needed == 0) &&
21595 	    (tot_idle > 0)) {
21596 		/*
21597 		 * We have been idle some micro seconds. We need
21598 		 * to factor this in to see if a PCM is needed.
21599 		 */
21600 		uint32_t rtts_idle, rnds;
21601 
21602 		if (tp->t_srtt)
21603 			rtts_idle = tot_idle / tp->t_srtt;
21604 		else
21605 			rtts_idle = 0;
21606 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
21607 		rack->r_ctl.pcm_idle_rounds += rtts_idle;
21608 		if ((rnds + rack->r_ctl.pcm_idle_rounds)  >= rack_pcm_every_n_rounds) {
21609 			rack->pcm_needed = 1;
21610 			rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
21611 		}
21612 	}
21613 again:
21614 	sendalot = 0;
21615 	cts = tcp_get_usecs(&tv);
21616 	ms_cts = tcp_tv_to_mssectick(&tv);
21617 	tso = 0;
21618 	mtu = 0;
21619 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
21620 	minseg = segsiz;
21621 	if (rack->r_ctl.rc_pace_max_segs == 0)
21622 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
21623 	else
21624 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
21625 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21626 	    (rack->r_ctl.pcm_max_seg == 0)) {
21627 		/*
21628 		 * We set in our first send so we know that the ctf_fixed_maxseg
21629 		 * has been fully set. If we do it in rack_init() we most likely
21630 		 * see 512 bytes so we end up at 5120, not desirable.
21631 		 */
21632 		rack->r_ctl.pcm_max_seg = rc_init_window(rack);
21633 		if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
21634 			/*
21635 			 * Assure our initial PCM probe is at least 10 MSS.
21636 			 */
21637 			rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
21638 		}
21639 	}
21640 	if ((rack->r_ctl.pcm_max_seg != 0)  && (rack->pcm_needed == 1)) {
21641 		uint32_t rw_avail, cwa;
21642 
21643 		if (tp->snd_wnd > ctf_outstanding(tp))
21644 			rw_avail = tp->snd_wnd - ctf_outstanding(tp);
21645 		else
21646 			rw_avail = 0;
21647 		if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
21648 			cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21649 		else
21650 			cwa = 0;
21651 		if ((cwa >= rack->r_ctl.pcm_max_seg) &&
21652 		    (rw_avail > rack->r_ctl.pcm_max_seg)) {
21653 			/* Raise up the max seg for this trip through */
21654 			pace_max_seg = rack->r_ctl.pcm_max_seg;
21655 			/* Disable any fast output */
21656 			rack->r_fast_output = 0;
21657 		}
21658 		if (rack_verbose_logging) {
21659 			rack_log_pcm(rack, 4,
21660 				     cwa, rack->r_ctl.pcm_max_seg, rw_avail);
21661 		}
21662 	}
21663 	sb_offset = tp->snd_max - tp->snd_una;
21664 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
21665 	flags = tcp_outflags[tp->t_state];
21666 	while (rack->rc_free_cnt < rack_free_cache) {
21667 		rsm = rack_alloc(rack);
21668 		if (rsm == NULL) {
21669 			if (hpts_calling)
21670 				/* Retry in a ms */
21671 				slot = (1 * HPTS_USEC_IN_MSEC);
21672 			so = inp->inp_socket;
21673 			sb = &so->so_snd;
21674 			goto just_return_nolock;
21675 		}
21676 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
21677 		rack->rc_free_cnt++;
21678 		rsm = NULL;
21679 	}
21680 	sack_rxmit = 0;
21681 	len = 0;
21682 	rsm = NULL;
21683 	if (flags & TH_RST) {
21684 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
21685 		so = inp->inp_socket;
21686 		sb = &so->so_snd;
21687 		goto send;
21688 	}
21689 	if (rack->r_ctl.rc_resend) {
21690 		/* Retransmit timer */
21691 		rsm = rack->r_ctl.rc_resend;
21692 		rack->r_ctl.rc_resend = NULL;
21693 		len = rsm->r_end - rsm->r_start;
21694 		sack_rxmit = 1;
21695 		sendalot = 0;
21696 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
21697 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
21698 			 __func__, __LINE__,
21699 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
21700 		sb_offset = rsm->r_start - tp->snd_una;
21701 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
21702 	} else if (rack->r_collapse_point_valid &&
21703 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
21704 		/*
21705 		 * If an RSM is returned then enough time has passed
21706 		 * for us to retransmit it. Move up the collapse point,
21707 		 * since this rsm has its chance to retransmit now.
21708 		 */
21709 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
21710 		rack->r_ctl.last_collapse_point = rsm->r_end;
21711 		/* Are we done? */
21712 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
21713 			    rack->r_ctl.high_collapse_point))
21714 			rack->r_collapse_point_valid = 0;
21715 		sack_rxmit = 1;
21716 		/* We are not doing a TLP */
21717 		doing_tlp = 0;
21718 		len = rsm->r_end - rsm->r_start;
21719 		sb_offset = rsm->r_start - tp->snd_una;
21720 		sendalot = 0;
21721 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
21722 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
21723 		/* We have a retransmit that takes precedence */
21724 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
21725 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
21726 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
21727 			/* Enter recovery if not induced by a time-out */
21728 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
21729 		}
21730 #ifdef INVARIANTS
21731 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
21732 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
21733 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
21734 		}
21735 #endif
21736 		len = rsm->r_end - rsm->r_start;
21737 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
21738 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
21739 			 __func__, __LINE__,
21740 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
21741 		sb_offset = rsm->r_start - tp->snd_una;
21742 		sendalot = 0;
21743 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
21744 		if (len > 0) {
21745 			sack_rxmit = 1;
21746 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
21747 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
21748 					 min(len, segsiz));
21749 		}
21750 	} else if (rack->r_ctl.rc_tlpsend) {
21751 		/* Tail loss probe */
21752 		long cwin;
21753 		long tlen;
21754 
21755 		/*
21756 		 * Check if we can do a TLP with a RACK'd packet
21757 		 * this can happen if we are not doing the rack
21758 		 * cheat and we skipped to a TLP and it
21759 		 * went off.
21760 		 */
21761 		rsm = rack->r_ctl.rc_tlpsend;
21762 		/* We are doing a TLP make sure the flag is preent */
21763 		rsm->r_flags |= RACK_TLP;
21764 		rack->r_ctl.rc_tlpsend = NULL;
21765 		sack_rxmit = 1;
21766 		tlen = rsm->r_end - rsm->r_start;
21767 		if (tlen > segsiz)
21768 			tlen = segsiz;
21769 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
21770 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
21771 			 __func__, __LINE__,
21772 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
21773 		sb_offset = rsm->r_start - tp->snd_una;
21774 		cwin = min(tp->snd_wnd, tlen);
21775 		len = cwin;
21776 	}
21777 	if (rack->r_must_retran &&
21778 	    (doing_tlp == 0) &&
21779 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
21780 	    (rsm == NULL)) {
21781 		/*
21782 		 * There are two different ways that we
21783 		 * can get into this block:
21784 		 * a) This is a non-sack connection, we had a time-out
21785 		 *    and thus r_must_retran was set and everything
21786 		 *    left outstanding as been marked for retransmit.
21787 		 * b) The MTU of the path shrank, so that everything
21788 		 *    was marked to be retransmitted with the smaller
21789 		 *    mtu and r_must_retran was set.
21790 		 *
21791 		 * This means that we expect the sendmap (outstanding)
21792 		 * to all be marked must. We can use the tmap to
21793 		 * look at them.
21794 		 *
21795 		 */
21796 		int sendwin, flight;
21797 
21798 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
21799 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
21800 		if (flight >= sendwin) {
21801 			/*
21802 			 * We can't send yet.
21803 			 */
21804 			so = inp->inp_socket;
21805 			sb = &so->so_snd;
21806 			goto just_return_nolock;
21807 		}
21808 		/*
21809 		 * This is the case a/b mentioned above. All
21810 		 * outstanding/not-acked should be marked.
21811 		 * We can use the tmap to find them.
21812 		 */
21813 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
21814 		if (rsm == NULL) {
21815 			/* TSNH */
21816 			rack->r_must_retran = 0;
21817 			rack->r_ctl.rc_out_at_rto = 0;
21818 			so = inp->inp_socket;
21819 			sb = &so->so_snd;
21820 			goto just_return_nolock;
21821 		}
21822 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
21823 			/*
21824 			 * The first one does not have the flag, did we collapse
21825 			 * further up in our list?
21826 			 */
21827 			rack->r_must_retran = 0;
21828 			rack->r_ctl.rc_out_at_rto = 0;
21829 			rsm = NULL;
21830 			sack_rxmit = 0;
21831 		} else {
21832 			sack_rxmit = 1;
21833 			len = rsm->r_end - rsm->r_start;
21834 			sb_offset = rsm->r_start - tp->snd_una;
21835 			sendalot = 0;
21836 			if ((rack->full_size_rxt == 0) &&
21837 			    (rack->shape_rxt_to_pacing_min == 0) &&
21838 			    (len >= segsiz))
21839 				len = segsiz;
21840 			else if (rack->shape_rxt_to_pacing_min &&
21841 				 rack->gp_ready) {
21842 				/* We use pacing min as shaping len req */
21843 				uint32_t maxlen;
21844 
21845 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
21846 				if (len > maxlen)
21847 					len = maxlen;
21848 			}
21849 			/*
21850 			 * Delay removing the flag RACK_MUST_RXT so
21851 			 * that the fastpath for retransmit will
21852 			 * work with this rsm.
21853 			 */
21854 		}
21855 	}
21856 	/*
21857 	 * Enforce a connection sendmap count limit if set
21858 	 * as long as we are not retransmiting.
21859 	 */
21860 	if ((rsm == NULL) &&
21861 	    (rack->do_detection == 0) &&
21862 	    (V_tcp_map_entries_limit > 0) &&
21863 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
21864 		counter_u64_add(rack_to_alloc_limited, 1);
21865 		if (!rack->alloc_limit_reported) {
21866 			rack->alloc_limit_reported = 1;
21867 			counter_u64_add(rack_alloc_limited_conns, 1);
21868 		}
21869 		so = inp->inp_socket;
21870 		sb = &so->so_snd;
21871 		goto just_return_nolock;
21872 	}
21873 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
21874 		/* we are retransmitting the fin */
21875 		len--;
21876 		if (len) {
21877 			/*
21878 			 * When retransmitting data do *not* include the
21879 			 * FIN. This could happen from a TLP probe.
21880 			 */
21881 			flags &= ~TH_FIN;
21882 		}
21883 	}
21884 	if (rsm && rack->r_fsb_inited &&
21885 	    rack_use_rsm_rfo &&
21886 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
21887 		int ret;
21888 
21889 		if ((rack->rc_policer_detected == 1) &&
21890 		    (rack->r_ctl.policer_bucket_size > segsiz) &&
21891 		    (rack->r_ctl.policer_bw > 0)) {
21892 			/* Check to see if there is room */
21893 			if (rack->r_ctl.current_policer_bucket < len) {
21894 				goto skip_fast_output;
21895 			}
21896 		}
21897 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
21898 		if (ret == 0)
21899 			return (0);
21900 	}
21901 skip_fast_output:
21902 	so = inp->inp_socket;
21903 	sb = &so->so_snd;
21904 	if (do_a_prefetch == 0) {
21905 		kern_prefetch(sb, &do_a_prefetch);
21906 		do_a_prefetch = 1;
21907 	}
21908 #ifdef NETFLIX_SHARED_CWND
21909 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
21910 	    rack->rack_enable_scwnd) {
21911 		/* We are doing cwnd sharing */
21912 		if (rack->gp_ready &&
21913 		    (rack->rack_attempted_scwnd == 0) &&
21914 		    (rack->r_ctl.rc_scw == NULL) &&
21915 		    tp->t_lib) {
21916 			/* The pcbid is in, lets make an attempt */
21917 			counter_u64_add(rack_try_scwnd, 1);
21918 			rack->rack_attempted_scwnd = 1;
21919 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
21920 								   &rack->r_ctl.rc_scw_index,
21921 								   segsiz);
21922 		}
21923 		if (rack->r_ctl.rc_scw &&
21924 		    (rack->rack_scwnd_is_idle == 1) &&
21925 		    sbavail(&so->so_snd)) {
21926 			/* we are no longer out of data */
21927 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21928 			rack->rack_scwnd_is_idle = 0;
21929 		}
21930 		if (rack->r_ctl.rc_scw) {
21931 			/* First lets update and get the cwnd */
21932 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
21933 										       rack->r_ctl.rc_scw_index,
21934 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
21935 		}
21936 	}
21937 #endif
21938 	/*
21939 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
21940 	 * state flags.
21941 	 */
21942 	if (tp->t_flags & TF_NEEDFIN)
21943 		flags |= TH_FIN;
21944 	if (tp->t_flags & TF_NEEDSYN)
21945 		flags |= TH_SYN;
21946 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
21947 		void *end_rsm;
21948 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
21949 		if (end_rsm)
21950 			kern_prefetch(end_rsm, &prefetch_rsm);
21951 		prefetch_rsm = 1;
21952 	}
21953 	SOCKBUF_LOCK(sb);
21954 	if ((sack_rxmit == 0) &&
21955 	    (TCPS_HAVEESTABLISHED(tp->t_state) ||
21956 	    (tp->t_flags & TF_FASTOPEN))) {
21957 		/*
21958 		 * We are not retransmitting (sack_rxmit is 0) so we
21959 		 * are sending new data. This is always based on snd_max.
21960 		 * Now in theory snd_max may be equal to snd_una, if so
21961 		 * then nothing is outstanding and the offset would be 0.
21962 		 */
21963 		uint32_t avail;
21964 
21965 		avail = sbavail(sb);
21966 		if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
21967 			sb_offset = tp->snd_max - tp->snd_una;
21968 		else
21969 			sb_offset = 0;
21970 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
21971 			if (rack->r_ctl.rc_tlp_new_data) {
21972 				/* TLP is forcing out new data */
21973 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
21974 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
21975 				}
21976 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
21977 					if (tp->snd_wnd > sb_offset)
21978 						len = tp->snd_wnd - sb_offset;
21979 					else
21980 						len = 0;
21981 				} else {
21982 					len = rack->r_ctl.rc_tlp_new_data;
21983 				}
21984 				rack->r_ctl.rc_tlp_new_data = 0;
21985 			}  else {
21986 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
21987 			}
21988 			if ((rack->r_ctl.crte == NULL) &&
21989 			    IN_FASTRECOVERY(tp->t_flags) &&
21990 			    (rack->full_size_rxt == 0) &&
21991 			    (rack->shape_rxt_to_pacing_min == 0) &&
21992 			    (len > segsiz)) {
21993 				/*
21994 				 * For prr=off, we need to send only 1 MSS
21995 				 * at a time. We do this because another sack could
21996 				 * be arriving that causes us to send retransmits and
21997 				 * we don't want to be on a long pace due to a larger send
21998 				 * that keeps us from sending out the retransmit.
21999 				 */
22000 				len = segsiz;
22001 			} else if (rack->shape_rxt_to_pacing_min &&
22002 				   rack->gp_ready) {
22003 				/* We use pacing min as shaping len req */
22004 				uint32_t maxlen;
22005 
22006 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
22007 				if (len > maxlen)
22008 					len = maxlen;
22009 			}/* The else is full_size_rxt is on so send it all */
22010 		} else {
22011 			uint32_t outstanding;
22012 			/*
22013 			 * We are inside of a Fast recovery episode, this
22014 			 * is caused by a SACK or 3 dup acks. At this point
22015 			 * we have sent all the retransmissions and we rely
22016 			 * on PRR to dictate what we will send in the form of
22017 			 * new data.
22018 			 */
22019 
22020 			outstanding = tp->snd_max - tp->snd_una;
22021 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
22022 				if (tp->snd_wnd > outstanding) {
22023 					len = tp->snd_wnd - outstanding;
22024 					/* Check to see if we have the data */
22025 					if ((sb_offset + len) > avail) {
22026 						/* It does not all fit */
22027 						if (avail > sb_offset)
22028 							len = avail - sb_offset;
22029 						else
22030 							len = 0;
22031 					}
22032 				} else {
22033 					len = 0;
22034 				}
22035 			} else if (avail > sb_offset) {
22036 				len = avail - sb_offset;
22037 			} else {
22038 				len = 0;
22039 			}
22040 			if (len > 0) {
22041 				if (len > rack->r_ctl.rc_prr_sndcnt) {
22042 					len = rack->r_ctl.rc_prr_sndcnt;
22043 				}
22044 				if (len > 0) {
22045 					sub_from_prr = 1;
22046 				}
22047 			}
22048 			if (len > segsiz) {
22049 				/*
22050 				 * We should never send more than a MSS when
22051 				 * retransmitting or sending new data in prr
22052 				 * mode unless the override flag is on. Most
22053 				 * likely the PRR algorithm is not going to
22054 				 * let us send a lot as well :-)
22055 				 */
22056 				if (rack->r_ctl.rc_prr_sendalot == 0) {
22057 					len = segsiz;
22058 				}
22059 			} else if (len < segsiz) {
22060 				/*
22061 				 * Do we send any? The idea here is if the
22062 				 * send empty's the socket buffer we want to
22063 				 * do it. However if not then lets just wait
22064 				 * for our prr_sndcnt to get bigger.
22065 				 */
22066 				long leftinsb;
22067 
22068 				leftinsb = sbavail(sb) - sb_offset;
22069 				if (leftinsb > len) {
22070 					/* This send does not empty the sb */
22071 					len = 0;
22072 				}
22073 			}
22074 		}
22075 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
22076 		/*
22077 		 * If you have not established
22078 		 * and are not doing FAST OPEN
22079 		 * no data please.
22080 		 */
22081 		if ((sack_rxmit == 0) &&
22082 		    !(tp->t_flags & TF_FASTOPEN)) {
22083 			len = 0;
22084 			sb_offset = 0;
22085 		}
22086 	}
22087 	if (prefetch_so_done == 0) {
22088 		kern_prefetch(so, &prefetch_so_done);
22089 		prefetch_so_done = 1;
22090 	}
22091 	orig_len = len;
22092 	if ((rack->rc_policer_detected == 1) &&
22093 	    (rack->r_ctl.policer_bucket_size > segsiz) &&
22094 	    (rack->r_ctl.policer_bw > 0) &&
22095 	    (len > 0)) {
22096 		/*
22097 		 * Ok we believe we have a policer watching
22098 		 * what we send, can we send len? If not can
22099 		 * we tune it down to a smaller value?
22100 		 */
22101 		uint32_t plen, buck_needs;
22102 
22103 		plen = rack_policer_check_send(rack, len, segsiz, &buck_needs);
22104 		if (plen == 0) {
22105 			/*
22106 			 * We are not allowed to send. How long
22107 			 * do we need to pace for i.e. how long
22108 			 * before len is available to send?
22109 			 */
22110 			uint64_t lentime;
22111 
22112 			lentime = buck_needs;
22113 			lentime *= HPTS_USEC_IN_SEC;
22114 			lentime /= rack->r_ctl.policer_bw;
22115 			slot = (uint32_t)lentime;
22116 			tot_len_this_send = 0;
22117 			SOCKBUF_UNLOCK(sb);
22118 			if (rack_verbose_logging > 0)
22119 				policer_detection_log(rack, len, slot, buck_needs, 0, 12);
22120 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22121 			rack_log_type_just_return(rack, cts, 0, slot, hpts_calling, 0, cwnd_to_use);
22122 			goto just_return_clean;
22123 		}
22124 		if (plen < len) {
22125 			sendalot = 0;
22126 			len = plen;
22127 		}
22128 	}
22129 	/*
22130 	 * Lop off SYN bit if it has already been sent.  However, if this is
22131 	 * SYN-SENT state and if segment contains data and if we don't know
22132 	 * that foreign host supports TAO, suppress sending segment.
22133 	 */
22134 	if ((flags & TH_SYN) &&
22135 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
22136 	    ((sack_rxmit == 0) &&
22137 	     (tp->t_rxtshift == 0))) {
22138 		/*
22139 		 * When sending additional segments following a TFO SYN|ACK,
22140 		 * do not include the SYN bit.
22141 		 */
22142 		if ((tp->t_flags & TF_FASTOPEN) &&
22143 		    (tp->t_state == TCPS_SYN_RECEIVED))
22144 			flags &= ~TH_SYN;
22145 	}
22146 	/*
22147 	 * Be careful not to send data and/or FIN on SYN segments. This
22148 	 * measure is needed to prevent interoperability problems with not
22149 	 * fully conformant TCP implementations.
22150 	 */
22151 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
22152 		len = 0;
22153 		flags &= ~TH_FIN;
22154 	}
22155 	/*
22156 	 * On TFO sockets, ensure no data is sent in the following cases:
22157 	 *
22158 	 *  - When retransmitting SYN|ACK on a passively-created socket
22159 	 *
22160 	 *  - When retransmitting SYN on an actively created socket
22161 	 *
22162 	 *  - When sending a zero-length cookie (cookie request) on an
22163 	 *    actively created socket
22164 	 *
22165 	 *  - When the socket is in the CLOSED state (RST is being sent)
22166 	 */
22167 	if ((tp->t_flags & TF_FASTOPEN) &&
22168 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
22169 	     ((tp->t_state == TCPS_SYN_SENT) &&
22170 	      (tp->t_tfo_client_cookie_len == 0)) ||
22171 	     (flags & TH_RST))) {
22172 		sack_rxmit = 0;
22173 		len = 0;
22174 	}
22175 	/* Without fast-open there should never be data sent on a SYN */
22176 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
22177 		len = 0;
22178 	}
22179 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
22180 		/* We only send 1 MSS if we have a DSACK block */
22181 		add_flag |= RACK_SENT_W_DSACK;
22182 		len = segsiz;
22183 	}
22184 	if (len <= 0) {
22185 		/*
22186 		 * We have nothing to send, or the window shrank, or
22187 		 * is closed, do we need to go into persists?
22188 		 */
22189 		len = 0;
22190 		if ((tp->snd_wnd == 0) &&
22191 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
22192 		    (tp->snd_una == tp->snd_max) &&
22193 		    (sb_offset < (int)sbavail(sb))) {
22194 			rack_enter_persist(tp, rack, cts, tp->snd_una);
22195 		}
22196 	} else if ((rsm == NULL) &&
22197 		   (doing_tlp == 0) &&
22198 		   (len < pace_max_seg)) {
22199 		/*
22200 		 * We are not sending a maximum sized segment for
22201 		 * some reason. Should we not send anything (think
22202 		 * sws or persists)?
22203 		 */
22204 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
22205 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
22206 		    (len < minseg) &&
22207 		    (len < (int)(sbavail(sb) - sb_offset))) {
22208 			/*
22209 			 * Here the rwnd is less than
22210 			 * the minimum pacing size, this is not a retransmit,
22211 			 * we are established and
22212 			 * the send is not the last in the socket buffer
22213 			 * we send nothing, and we may enter persists
22214 			 * if nothing is outstanding.
22215 			 */
22216 			len = 0;
22217 			if (tp->snd_max == tp->snd_una) {
22218 				/*
22219 				 * Nothing out we can
22220 				 * go into persists.
22221 				 */
22222 				rack_enter_persist(tp, rack, cts, tp->snd_una);
22223 			}
22224 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
22225 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
22226 			   (len < (int)(sbavail(sb) - sb_offset)) &&
22227 			   (len < minseg)) {
22228 			/*
22229 			 * Here we are not retransmitting, and
22230 			 * the cwnd is not so small that we could
22231 			 * not send at least a min size (rxt timer
22232 			 * not having gone off), We have 2 segments or
22233 			 * more already in flight, its not the tail end
22234 			 * of the socket buffer  and the cwnd is blocking
22235 			 * us from sending out a minimum pacing segment size.
22236 			 * Lets not send anything.
22237 			 */
22238 			len = 0;
22239 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
22240 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
22241 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
22242 			   (len < (int)(sbavail(sb) - sb_offset)) &&
22243 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
22244 			/*
22245 			 * Here we have a send window but we have
22246 			 * filled it up and we can't send another pacing segment.
22247 			 * We also have in flight more than 2 segments
22248 			 * and we are not completing the sb i.e. we allow
22249 			 * the last bytes of the sb to go out even if
22250 			 * its not a full pacing segment.
22251 			 */
22252 			len = 0;
22253 		} else if ((rack->r_ctl.crte != NULL) &&
22254 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
22255 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
22256 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
22257 			   (len < (int)(sbavail(sb) - sb_offset))) {
22258 			/*
22259 			 * Here we are doing hardware pacing, this is not a TLP,
22260 			 * we are not sending a pace max segment size, there is rwnd
22261 			 * room to send at least N pace_max_seg, the cwnd is greater
22262 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
22263 			 * more segments in flight and its not the tail of the socket buffer.
22264 			 *
22265 			 * We don't want to send instead we need to get more ack's in to
22266 			 * allow us to send a full pacing segment. Normally, if we are pacing
22267 			 * about the right speed, we should have finished our pacing
22268 			 * send as most of the acks have come back if we are at the
22269 			 * right rate. This is a bit fuzzy since return path delay
22270 			 * can delay the acks, which is why we want to make sure we
22271 			 * have cwnd space to have a bit more than a max pace segments in flight.
22272 			 *
22273 			 * If we have not gotten our acks back we are pacing at too high a
22274 			 * rate delaying will not hurt and will bring our GP estimate down by
22275 			 * injecting the delay. If we don't do this we will send
22276 			 * 2 MSS out in response to the acks being clocked in which
22277 			 * defeats the point of hw-pacing (i.e. to help us get
22278 			 * larger TSO's out).
22279 			 */
22280 			len = 0;
22281 		}
22282 
22283 	}
22284 	/* len will be >= 0 after this point. */
22285 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
22286 	rack_sndbuf_autoscale(rack);
22287 	/*
22288 	 * Decide if we can use TCP Segmentation Offloading (if supported by
22289 	 * hardware).
22290 	 *
22291 	 * TSO may only be used if we are in a pure bulk sending state.  The
22292 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
22293 	 * options prevent using TSO.  With TSO the TCP header is the same
22294 	 * (except for the sequence number) for all generated packets.  This
22295 	 * makes it impossible to transmit any options which vary per
22296 	 * generated segment or packet.
22297 	 *
22298 	 * IPv4 handling has a clear separation of ip options and ip header
22299 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
22300 	 * the right thing below to provide length of just ip options and thus
22301 	 * checking for ipoptlen is enough to decide if ip options are present.
22302 	 */
22303 	ipoptlen = 0;
22304 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22305 	/*
22306 	 * Pre-calculate here as we save another lookup into the darknesses
22307 	 * of IPsec that way and can actually decide if TSO is ok.
22308 	 */
22309 #ifdef INET6
22310 	if (isipv6 && IPSEC_ENABLED(ipv6))
22311 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
22312 #ifdef INET
22313 	else
22314 #endif
22315 #endif				/* INET6 */
22316 #ifdef INET
22317 		if (IPSEC_ENABLED(ipv4))
22318 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
22319 #endif				/* INET */
22320 #endif
22321 
22322 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22323 	ipoptlen += ipsec_optlen;
22324 #endif
22325 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
22326 	    (tp->t_port == 0) &&
22327 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
22328 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
22329 	    ipoptlen == 0)
22330 		tso = 1;
22331 	{
22332 		uint32_t outstanding __unused;
22333 
22334 		outstanding = tp->snd_max - tp->snd_una;
22335 		if (tp->t_flags & TF_SENTFIN) {
22336 			/*
22337 			 * If we sent a fin, snd_max is 1 higher than
22338 			 * snd_una
22339 			 */
22340 			outstanding--;
22341 		}
22342 		if (sack_rxmit) {
22343 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
22344 				flags &= ~TH_FIN;
22345 		}
22346 	}
22347 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
22348 		      (long)TCP_MAXWIN << tp->rcv_scale);
22349 
22350 	/*
22351 	 * Sender silly window avoidance.   We transmit under the following
22352 	 * conditions when len is non-zero:
22353 	 *
22354 	 * - We have a full segment (or more with TSO) - This is the last
22355 	 * buffer in a write()/send() and we are either idle or running
22356 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
22357 	 * then 1/2 the maximum send window's worth of data (receiver may be
22358 	 * limited the window size) - we need to retransmit
22359 	 */
22360 	if (len) {
22361 		if (len >= segsiz) {
22362 			goto send;
22363 		}
22364 		/*
22365 		 * NOTE! on localhost connections an 'ack' from the remote
22366 		 * end may occur synchronously with the output and cause us
22367 		 * to flush a buffer queued with moretocome.  XXX
22368 		 *
22369 		 */
22370 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
22371 		    (idle || (tp->t_flags & TF_NODELAY)) &&
22372 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
22373 		    (tp->t_flags & TF_NOPUSH) == 0) {
22374 			pass = 2;
22375 			goto send;
22376 		}
22377 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
22378 			pass = 22;
22379 			goto send;
22380 		}
22381 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
22382 			pass = 4;
22383 			goto send;
22384 		}
22385 		if (sack_rxmit) {
22386 			pass = 6;
22387 			goto send;
22388 		}
22389 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
22390 		    (ctf_outstanding(tp) < (segsiz * 2))) {
22391 			/*
22392 			 * We have less than two MSS outstanding (delayed ack)
22393 			 * and our rwnd will not let us send a full sized
22394 			 * MSS. Lets go ahead and let this small segment
22395 			 * out because we want to try to have at least two
22396 			 * packets inflight to not be caught by delayed ack.
22397 			 */
22398 			pass = 12;
22399 			goto send;
22400 		}
22401 	}
22402 	/*
22403 	 * Sending of standalone window updates.
22404 	 *
22405 	 * Window updates are important when we close our window due to a
22406 	 * full socket buffer and are opening it again after the application
22407 	 * reads data from it.  Once the window has opened again and the
22408 	 * remote end starts to send again the ACK clock takes over and
22409 	 * provides the most current window information.
22410 	 *
22411 	 * We must avoid the silly window syndrome whereas every read from
22412 	 * the receive buffer, no matter how small, causes a window update
22413 	 * to be sent.  We also should avoid sending a flurry of window
22414 	 * updates when the socket buffer had queued a lot of data and the
22415 	 * application is doing small reads.
22416 	 *
22417 	 * Prevent a flurry of pointless window updates by only sending an
22418 	 * update when we can increase the advertized window by more than
22419 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
22420 	 * full or is very small be more aggressive and send an update
22421 	 * whenever we can increase by two mss sized segments. In all other
22422 	 * situations the ACK's to new incoming data will carry further
22423 	 * window increases.
22424 	 *
22425 	 * Don't send an independent window update if a delayed ACK is
22426 	 * pending (it will get piggy-backed on it) or the remote side
22427 	 * already has done a half-close and won't send more data.  Skip
22428 	 * this if the connection is in T/TCP half-open state.
22429 	 */
22430 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
22431 	    !(tp->t_flags & TF_DELACK) &&
22432 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
22433 		/*
22434 		 * "adv" is the amount we could increase the window, taking
22435 		 * into account that we are limited by TCP_MAXWIN <<
22436 		 * tp->rcv_scale.
22437 		 */
22438 		int32_t adv;
22439 		int oldwin;
22440 
22441 		adv = recwin;
22442 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
22443 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
22444 			if (adv > oldwin)
22445 				adv -= oldwin;
22446 			else {
22447 				/* We can't increase the window */
22448 				adv = 0;
22449 			}
22450 		} else
22451 			oldwin = 0;
22452 
22453 		/*
22454 		 * If the new window size ends up being the same as or less
22455 		 * than the old size when it is scaled, then don't force
22456 		 * a window update.
22457 		 */
22458 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
22459 			goto dontupdate;
22460 
22461 		if (adv >= (int32_t)(2 * segsiz) &&
22462 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
22463 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
22464 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
22465 			pass = 7;
22466 			goto send;
22467 		}
22468 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
22469 			pass = 23;
22470 			goto send;
22471 		}
22472 	}
22473 dontupdate:
22474 
22475 	/*
22476 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
22477 	 * is also a catch-all for the retransmit timer timeout case.
22478 	 */
22479 	if (tp->t_flags & TF_ACKNOW) {
22480 		pass = 8;
22481 		goto send;
22482 	}
22483 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
22484 		pass = 9;
22485 		goto send;
22486 	}
22487 	/*
22488 	 * If our state indicates that FIN should be sent and we have not
22489 	 * yet done so, then we need to send.
22490 	 */
22491 	if ((flags & TH_FIN) &&
22492 	    (tp->snd_max == tp->snd_una)) {
22493 		pass = 11;
22494 		goto send;
22495 	}
22496 	/*
22497 	 * No reason to send a segment, just return.
22498 	 */
22499 just_return:
22500 	SOCKBUF_UNLOCK(sb);
22501 just_return_nolock:
22502 	{
22503 		int app_limited = CTF_JR_SENT_DATA;
22504 
22505 		if ((tp->t_flags & TF_FASTOPEN) == 0 &&
22506 		    (flags & TH_FIN) &&
22507 		    (len == 0) &&
22508 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
22509 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
22510 			/*
22511 			 * Ok less than or right at a MSS is
22512 			 * outstanding. The original FreeBSD stack would
22513 			 * have sent a FIN, which can speed things up for
22514 			 * a transactional application doing a MSG_WAITALL.
22515 			 * To speed things up since we do *not* send a FIN
22516 			 * if data is outstanding, we send a "challenge ack".
22517 			 * The idea behind that is instead of having to have
22518 			 * the peer wait for the delayed-ack timer to run off
22519 			 * we send an ack that makes the peer send us an ack.
22520 			 */
22521 			rack_send_ack_challange(rack);
22522 		}
22523 		if (tot_len_this_send > 0) {
22524 			rack->r_ctl.fsb.recwin = recwin;
22525 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
22526 			if ((error == 0) &&
22527 			    (rack->rc_policer_detected == 0)  &&
22528 			    rack_use_rfo &&
22529 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22530 			    (ipoptlen == 0) &&
22531 			    (tp->rcv_numsacks == 0) &&
22532 			    rack->r_fsb_inited &&
22533 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
22534 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22535 			    (rack->r_must_retran == 0) &&
22536 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22537 			    (len > 0) && (orig_len > 0) &&
22538 			    (orig_len > len) &&
22539 			    ((orig_len - len) >= segsiz) &&
22540 			    ((optlen == 0) ||
22541 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22542 				/* We can send at least one more MSS using our fsb */
22543 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
22544 						       segsiz, pace_max_seg, hw_tls, flags);
22545 			} else
22546 				rack->r_fast_output = 0;
22547 			rack_log_fsb(rack, tp, so, flags,
22548 				     ipoptlen, orig_len, len, 0,
22549 				     1, optlen, __LINE__, 1);
22550 			/* Assure when we leave that snd_nxt will point to top */
22551 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22552 				tp->snd_nxt = tp->snd_max;
22553 		} else {
22554 			int end_window = 0;
22555 			uint32_t seq = tp->gput_ack;
22556 
22557 			rsm = tqhash_max(rack->r_ctl.tqh);
22558 			if (rsm) {
22559 				/*
22560 				 * Mark the last sent that we just-returned (hinting
22561 				 * that delayed ack may play a role in any rtt measurement).
22562 				 */
22563 				rsm->r_just_ret = 1;
22564 			}
22565 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
22566 			rack->r_ctl.rc_agg_delayed = 0;
22567 			rack->r_early = 0;
22568 			rack->r_late = 0;
22569 			rack->r_ctl.rc_agg_early = 0;
22570 			if ((ctf_outstanding(tp) +
22571 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
22572 				 minseg)) >= tp->snd_wnd) {
22573 				/* We are limited by the rwnd */
22574 				app_limited = CTF_JR_RWND_LIMITED;
22575 				if (IN_FASTRECOVERY(tp->t_flags))
22576 					rack->r_ctl.rc_prr_sndcnt = 0;
22577 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
22578 				/* We are limited by whats available -- app limited */
22579 				app_limited = CTF_JR_APP_LIMITED;
22580 				if (IN_FASTRECOVERY(tp->t_flags))
22581 					rack->r_ctl.rc_prr_sndcnt = 0;
22582 			} else if ((idle == 0) &&
22583 				   ((tp->t_flags & TF_NODELAY) == 0) &&
22584 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
22585 				   (len < segsiz)) {
22586 				/*
22587 				 * No delay is not on and the
22588 				 * user is sending less than 1MSS. This
22589 				 * brings out SWS avoidance so we
22590 				 * don't send. Another app-limited case.
22591 				 */
22592 				app_limited = CTF_JR_APP_LIMITED;
22593 			} else if (tp->t_flags & TF_NOPUSH) {
22594 				/*
22595 				 * The user has requested no push of
22596 				 * the last segment and we are
22597 				 * at the last segment. Another app
22598 				 * limited case.
22599 				 */
22600 				app_limited = CTF_JR_APP_LIMITED;
22601 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
22602 				/* Its the cwnd */
22603 				app_limited = CTF_JR_CWND_LIMITED;
22604 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
22605 				   (rack->rack_no_prr == 0) &&
22606 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
22607 				app_limited = CTF_JR_PRR;
22608 			} else {
22609 				/* Now why here are we not sending? */
22610 #ifdef NOW
22611 #ifdef INVARIANTS
22612 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
22613 #endif
22614 #endif
22615 				app_limited = CTF_JR_ASSESSING;
22616 			}
22617 			/*
22618 			 * App limited in some fashion, for our pacing GP
22619 			 * measurements we don't want any gap (even cwnd).
22620 			 * Close  down the measurement window.
22621 			 */
22622 			if (rack_cwnd_block_ends_measure &&
22623 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
22624 			     (app_limited == CTF_JR_PRR))) {
22625 				/*
22626 				 * The reason we are not sending is
22627 				 * the cwnd (or prr). We have been configured
22628 				 * to end the measurement window in
22629 				 * this case.
22630 				 */
22631 				end_window = 1;
22632 			} else if (rack_rwnd_block_ends_measure &&
22633 				   (app_limited == CTF_JR_RWND_LIMITED)) {
22634 				/*
22635 				 * We are rwnd limited and have been
22636 				 * configured to end the measurement
22637 				 * window in this case.
22638 				 */
22639 				end_window = 1;
22640 			} else if (app_limited == CTF_JR_APP_LIMITED) {
22641 				/*
22642 				 * A true application limited period, we have
22643 				 * ran out of data.
22644 				 */
22645 				end_window = 1;
22646 			} else if (app_limited == CTF_JR_ASSESSING) {
22647 				/*
22648 				 * In the assessing case we hit the end of
22649 				 * the if/else and had no known reason
22650 				 * This will panic us under invariants..
22651 				 *
22652 				 * If we get this out in logs we need to
22653 				 * investagate which reason we missed.
22654 				 */
22655 				end_window = 1;
22656 			}
22657 			if (end_window) {
22658 				uint8_t log = 0;
22659 
22660 				/* Adjust the Gput measurement */
22661 				if ((tp->t_flags & TF_GPUTINPROG) &&
22662 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
22663 					tp->gput_ack = tp->snd_max;
22664 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
22665 						/*
22666 						 * There is not enough to measure.
22667 						 */
22668 						tp->t_flags &= ~TF_GPUTINPROG;
22669 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
22670 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
22671 									   tp->gput_seq,
22672 									   0, 0, 18, __LINE__, NULL, 0);
22673 					} else
22674 						log = 1;
22675 				}
22676 				/* Mark the last packet has app limited */
22677 				rsm = tqhash_max(rack->r_ctl.tqh);
22678 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
22679 					if (rack->r_ctl.rc_app_limited_cnt == 0)
22680 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
22681 					else {
22682 						/*
22683 						 * Go out to the end app limited and mark
22684 						 * this new one as next and move the end_appl up
22685 						 * to this guy.
22686 						 */
22687 						if (rack->r_ctl.rc_end_appl)
22688 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
22689 						rack->r_ctl.rc_end_appl = rsm;
22690 					}
22691 					rsm->r_flags |= RACK_APP_LIMITED;
22692 					rack->r_ctl.rc_app_limited_cnt++;
22693 				}
22694 				if (log)
22695 					rack_log_pacing_delay_calc(rack,
22696 								   rack->r_ctl.rc_app_limited_cnt, seq,
22697 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
22698 			}
22699 		}
22700 		/* Check if we need to go into persists or not */
22701 		if ((tp->snd_max == tp->snd_una) &&
22702 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22703 		    sbavail(sb) &&
22704 		    (sbavail(sb) > tp->snd_wnd) &&
22705 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
22706 			/* Yes lets make sure to move to persist before timer-start */
22707 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
22708 		}
22709 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
22710 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
22711 	}
22712 just_return_clean:
22713 #ifdef NETFLIX_SHARED_CWND
22714 	if ((sbavail(sb) == 0) &&
22715 	    rack->r_ctl.rc_scw) {
22716 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
22717 		rack->rack_scwnd_is_idle = 1;
22718 	}
22719 #endif
22720 #ifdef TCP_ACCOUNTING
22721 	if (tot_len_this_send > 0) {
22722 		crtsc = get_cyclecount();
22723 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22724 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22725 		}
22726 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22727 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
22728 		}
22729 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22730 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
22731 		}
22732 	} else {
22733 		crtsc = get_cyclecount();
22734 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22735 			tp->tcp_cnt_counters[SND_LIMITED]++;
22736 		}
22737 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22738 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
22739 		}
22740 	}
22741 	sched_unpin();
22742 #endif
22743 	return (0);
22744 
22745 send:
22746 	if ((rack->r_ctl.crte != NULL) &&
22747 	    (rsm == NULL) &&
22748 	    ((rack->rc_hw_nobuf == 1) ||
22749 	     (rack_hw_check_queue && (check_done == 0)))) {
22750 		/*
22751 		 * We only want to do this once with the hw_check_queue,
22752 		 * for the enobuf case we would only do it once if
22753 		 * we come around to again, the flag will be clear.
22754 		 */
22755 		check_done = 1;
22756 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
22757 		if (slot) {
22758 			rack->r_ctl.rc_agg_delayed = 0;
22759 			rack->r_ctl.rc_agg_early = 0;
22760 			rack->r_early = 0;
22761 			rack->r_late = 0;
22762 			SOCKBUF_UNLOCK(&so->so_snd);
22763 			goto skip_all_send;
22764 		}
22765 	}
22766 	if (rsm || sack_rxmit)
22767 		counter_u64_add(rack_nfto_resend, 1);
22768 	else
22769 		counter_u64_add(rack_non_fto_send, 1);
22770 	if ((flags & TH_FIN) &&
22771 	    sbavail(sb)) {
22772 		/*
22773 		 * We do not transmit a FIN
22774 		 * with data outstanding. We
22775 		 * need to make it so all data
22776 		 * is acked first.
22777 		 */
22778 		flags &= ~TH_FIN;
22779 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22780 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
22781 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
22782 			/*
22783 			 * Ok less than or right at a MSS is
22784 			 * outstanding. The original FreeBSD stack would
22785 			 * have sent a FIN, which can speed things up for
22786 			 * a transactional application doing a MSG_WAITALL.
22787 			 * To speed things up since we do *not* send a FIN
22788 			 * if data is outstanding, we send a "challenge ack".
22789 			 * The idea behind that is instead of having to have
22790 			 * the peer wait for the delayed-ack timer to run off
22791 			 * we send an ack that makes the peer send us an ack.
22792 			 */
22793 			rack_send_ack_challange(rack);
22794 		}
22795 	}
22796 	/* Enforce stack imposed max seg size if we have one */
22797 	if (pace_max_seg &&
22798 	    (len > pace_max_seg)) {
22799 		mark = 1;
22800 		len = pace_max_seg;
22801 	}
22802 	if ((rsm == NULL) &&
22803 	    (rack->pcm_in_progress == 0) &&
22804 	    (rack->r_ctl.pcm_max_seg > 0) &&
22805 	    (len >= rack->r_ctl.pcm_max_seg)) {
22806 		/* It is large enough for a measurement */
22807 		add_flag |= RACK_IS_PCM;
22808 		rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg,  add_flag);
22809 	} else if (rack_verbose_logging) {
22810 		rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg,  add_flag);
22811 	}
22812 
22813 	SOCKBUF_LOCK_ASSERT(sb);
22814 	if (len > 0) {
22815 		if (len >= segsiz)
22816 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
22817 		else
22818 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
22819 	}
22820 	/*
22821 	 * Before ESTABLISHED, force sending of initial options unless TCP
22822 	 * set not to do any options. NOTE: we assume that the IP/TCP header
22823 	 * plus TCP options always fit in a single mbuf, leaving room for a
22824 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
22825 	 * + optlen <= MCLBYTES
22826 	 */
22827 	optlen = 0;
22828 #ifdef INET6
22829 	if (isipv6)
22830 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
22831 	else
22832 #endif
22833 		hdrlen = sizeof(struct tcpiphdr);
22834 
22835 	/*
22836 	 * Ok what seq are we sending from. If we have
22837 	 * no rsm to use, then we look at various bits,
22838 	 * if we are putting out a SYN it will be ISS.
22839 	 * If we are retransmitting a FIN it will
22840 	 * be snd_max-1 else its snd_max.
22841 	 */
22842 	if (rsm == NULL) {
22843 		if (flags & TH_SYN)
22844 			rack_seq = tp->iss;
22845 		else if ((flags & TH_FIN) &&
22846 			 (tp->t_flags & TF_SENTFIN))
22847 			rack_seq = tp->snd_max - 1;
22848 		else
22849 			rack_seq = tp->snd_max;
22850 	} else {
22851 		rack_seq = rsm->r_start;
22852 	}
22853 	/*
22854 	 * Compute options for segment. We only have to care about SYN and
22855 	 * established connection segments.  Options for SYN-ACK segments
22856 	 * are handled in TCP syncache.
22857 	 */
22858 	to.to_flags = 0;
22859 	if ((tp->t_flags & TF_NOOPT) == 0) {
22860 		/* Maximum segment size. */
22861 		if (flags & TH_SYN) {
22862 			to.to_mss = tcp_mssopt(&inp->inp_inc);
22863 			if (tp->t_port)
22864 				to.to_mss -= V_tcp_udp_tunneling_overhead;
22865 			to.to_flags |= TOF_MSS;
22866 
22867 			/*
22868 			 * On SYN or SYN|ACK transmits on TFO connections,
22869 			 * only include the TFO option if it is not a
22870 			 * retransmit, as the presence of the TFO option may
22871 			 * have caused the original SYN or SYN|ACK to have
22872 			 * been dropped by a middlebox.
22873 			 */
22874 			if ((tp->t_flags & TF_FASTOPEN) &&
22875 			    (tp->t_rxtshift == 0)) {
22876 				if (tp->t_state == TCPS_SYN_RECEIVED) {
22877 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
22878 					to.to_tfo_cookie =
22879 						(u_int8_t *)&tp->t_tfo_cookie.server;
22880 					to.to_flags |= TOF_FASTOPEN;
22881 					wanted_cookie = 1;
22882 				} else if (tp->t_state == TCPS_SYN_SENT) {
22883 					to.to_tfo_len =
22884 						tp->t_tfo_client_cookie_len;
22885 					to.to_tfo_cookie =
22886 						tp->t_tfo_cookie.client;
22887 					to.to_flags |= TOF_FASTOPEN;
22888 					wanted_cookie = 1;
22889 					/*
22890 					 * If we wind up having more data to
22891 					 * send with the SYN than can fit in
22892 					 * one segment, don't send any more
22893 					 * until the SYN|ACK comes back from
22894 					 * the other end.
22895 					 */
22896 					sendalot = 0;
22897 				}
22898 			}
22899 		}
22900 		/* Window scaling. */
22901 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
22902 			to.to_wscale = tp->request_r_scale;
22903 			to.to_flags |= TOF_SCALE;
22904 		}
22905 		/* Timestamps. */
22906 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
22907 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
22908 			uint32_t ts_to_use;
22909 
22910 			if ((rack->r_rcvpath_rtt_up == 1) &&
22911 			    (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
22912 				/*
22913 				 * When we are doing a rcv_rtt probe all
22914 				 * other timestamps use the next msec. This
22915 				 * is safe since our previous ack is in the
22916 				 * air and we will just have a few more
22917 				 * on the next ms. This assures that only
22918 				 * the one ack has the ms_cts that was on
22919 				 * our ack-probe.
22920 				 */
22921 				ts_to_use = ms_cts + 1;
22922 			} else {
22923 				ts_to_use = ms_cts;
22924 			}
22925 			to.to_tsval = ts_to_use + tp->ts_offset;
22926 			to.to_tsecr = tp->ts_recent;
22927 			to.to_flags |= TOF_TS;
22928 			if ((len == 0) &&
22929 			    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
22930 			    ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
22931 			    (tp->snd_una == tp->snd_max) &&
22932 			    (flags & TH_ACK) &&
22933 			    (sbavail(sb) == 0) &&
22934 			    (rack->r_ctl.current_round != 0) &&
22935 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22936 			    (rack->r_rcvpath_rtt_up == 0)) {
22937 				rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
22938 				rack->r_ctl.last_time_of_arm_rcv = cts;
22939 				rack->r_rcvpath_rtt_up = 1;
22940 				/* Subtract 1 from seq to force a response */
22941 				rack_seq--;
22942 			}
22943 		}
22944 		/* Set receive buffer autosizing timestamp. */
22945 		if (tp->rfbuf_ts == 0 &&
22946 		    (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
22947 			tp->rfbuf_ts = ms_cts;
22948 		}
22949 		/* Selective ACK's. */
22950 		if (tp->t_flags & TF_SACK_PERMIT) {
22951 			if (flags & TH_SYN)
22952 				to.to_flags |= TOF_SACKPERM;
22953 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22954 				 tp->rcv_numsacks > 0) {
22955 				to.to_flags |= TOF_SACK;
22956 				to.to_nsacks = tp->rcv_numsacks;
22957 				to.to_sacks = (u_char *)tp->sackblks;
22958 			}
22959 		}
22960 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
22961 		/* TCP-MD5 (RFC2385). */
22962 		if (tp->t_flags & TF_SIGNATURE)
22963 			to.to_flags |= TOF_SIGNATURE;
22964 #endif
22965 
22966 		/* Processing the options. */
22967 		hdrlen += optlen = tcp_addoptions(&to, opt);
22968 		/*
22969 		 * If we wanted a TFO option to be added, but it was unable
22970 		 * to fit, ensure no data is sent.
22971 		 */
22972 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
22973 		    !(to.to_flags & TOF_FASTOPEN))
22974 			len = 0;
22975 	}
22976 	if (tp->t_port) {
22977 		if (V_tcp_udp_tunneling_port == 0) {
22978 			/* The port was removed?? */
22979 			SOCKBUF_UNLOCK(&so->so_snd);
22980 #ifdef TCP_ACCOUNTING
22981 			crtsc = get_cyclecount();
22982 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22983 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22984 			}
22985 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22986 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22987 			}
22988 			sched_unpin();
22989 #endif
22990 			return (EHOSTUNREACH);
22991 		}
22992 		hdrlen += sizeof(struct udphdr);
22993 	}
22994 #ifdef INET6
22995 	if (isipv6)
22996 		ipoptlen = ip6_optlen(inp);
22997 	else
22998 #endif
22999 		if (inp->inp_options)
23000 			ipoptlen = inp->inp_options->m_len -
23001 				offsetof(struct ipoption, ipopt_list);
23002 		else
23003 			ipoptlen = 0;
23004 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
23005 	ipoptlen += ipsec_optlen;
23006 #endif
23007 
23008 	/*
23009 	 * Adjust data length if insertion of options will bump the packet
23010 	 * length beyond the t_maxseg length. Clear the FIN bit because we
23011 	 * cut off the tail of the segment.
23012 	 */
23013 	if (len + optlen + ipoptlen > tp->t_maxseg) {
23014 		if (tso) {
23015 			uint32_t if_hw_tsomax;
23016 			uint32_t moff;
23017 			int32_t max_len;
23018 
23019 			/* extract TSO information */
23020 			if_hw_tsomax = tp->t_tsomax;
23021 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
23022 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
23023 			KASSERT(ipoptlen == 0,
23024 				("%s: TSO can't do IP options", __func__));
23025 
23026 			/*
23027 			 * Check if we should limit by maximum payload
23028 			 * length:
23029 			 */
23030 			if (if_hw_tsomax != 0) {
23031 				/* compute maximum TSO length */
23032 				max_len = (if_hw_tsomax - hdrlen -
23033 					   max_linkhdr);
23034 				if (max_len <= 0) {
23035 					len = 0;
23036 				} else if (len > max_len) {
23037 					sendalot = 1;
23038 					len = max_len;
23039 					mark = 2;
23040 				}
23041 			}
23042 			/*
23043 			 * Prevent the last segment from being fractional
23044 			 * unless the send sockbuf can be emptied:
23045 			 */
23046 			max_len = (tp->t_maxseg - optlen);
23047 			if ((sb_offset + len) < sbavail(sb)) {
23048 				moff = len % (u_int)max_len;
23049 				if (moff != 0) {
23050 					mark = 3;
23051 					len -= moff;
23052 				}
23053 			}
23054 			/*
23055 			 * In case there are too many small fragments don't
23056 			 * use TSO:
23057 			 */
23058 			if (len <= max_len) {
23059 				mark = 4;
23060 				tso = 0;
23061 			}
23062 			/*
23063 			 * Send the FIN in a separate segment after the bulk
23064 			 * sending is done. We don't trust the TSO
23065 			 * implementations to clear the FIN flag on all but
23066 			 * the last segment.
23067 			 */
23068 			if (tp->t_flags & TF_NEEDFIN) {
23069 				sendalot = 4;
23070 			}
23071 		} else {
23072 			mark = 5;
23073 			if (optlen + ipoptlen >= tp->t_maxseg) {
23074 				/*
23075 				 * Since we don't have enough space to put
23076 				 * the IP header chain and the TCP header in
23077 				 * one packet as required by RFC 7112, don't
23078 				 * send it. Also ensure that at least one
23079 				 * byte of the payload can be put into the
23080 				 * TCP segment.
23081 				 */
23082 				SOCKBUF_UNLOCK(&so->so_snd);
23083 				error = EMSGSIZE;
23084 				sack_rxmit = 0;
23085 				goto out;
23086 			}
23087 			len = tp->t_maxseg - optlen - ipoptlen;
23088 			sendalot = 5;
23089 		}
23090 	} else {
23091 		tso = 0;
23092 		mark = 6;
23093 	}
23094 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
23095 		("%s: len > IP_MAXPACKET", __func__));
23096 #ifdef DIAGNOSTIC
23097 #ifdef INET6
23098 	if (max_linkhdr + hdrlen > MCLBYTES)
23099 #else
23100 		if (max_linkhdr + hdrlen > MHLEN)
23101 #endif
23102 			panic("tcphdr too big");
23103 #endif
23104 
23105 	/*
23106 	 * This KASSERT is here to catch edge cases at a well defined place.
23107 	 * Before, those had triggered (random) panic conditions further
23108 	 * down.
23109 	 */
23110 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
23111 	if ((len == 0) &&
23112 	    (flags & TH_FIN) &&
23113 	    (sbused(sb))) {
23114 		/*
23115 		 * We have outstanding data, don't send a fin by itself!.
23116 		 *
23117 		 * Check to see if we need to send a challenge ack.
23118 		 */
23119 		if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
23120 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
23121 			/*
23122 			 * Ok less than or right at a MSS is
23123 			 * outstanding. The original FreeBSD stack would
23124 			 * have sent a FIN, which can speed things up for
23125 			 * a transactional application doing a MSG_WAITALL.
23126 			 * To speed things up since we do *not* send a FIN
23127 			 * if data is outstanding, we send a "challenge ack".
23128 			 * The idea behind that is instead of having to have
23129 			 * the peer wait for the delayed-ack timer to run off
23130 			 * we send an ack that makes the peer send us an ack.
23131 			 */
23132 			rack_send_ack_challange(rack);
23133 		}
23134 		goto just_return;
23135 	}
23136 	/*
23137 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
23138 	 * and initialize the header from the template for sends on this
23139 	 * connection.
23140 	 */
23141 	hw_tls = tp->t_nic_ktls_xmit != 0;
23142 	if (len) {
23143 		uint32_t max_val;
23144 		uint32_t moff;
23145 
23146 		if (pace_max_seg)
23147 			max_val = pace_max_seg;
23148 		else
23149 			max_val = len;
23150 		/*
23151 		 * We allow a limit on sending with hptsi.
23152 		 */
23153 		if (len > max_val) {
23154 			mark = 7;
23155 			len = max_val;
23156 		}
23157 #ifdef INET6
23158 		if (MHLEN < hdrlen + max_linkhdr)
23159 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
23160 		else
23161 #endif
23162 			m = m_gethdr(M_NOWAIT, MT_DATA);
23163 
23164 		if (m == NULL) {
23165 			SOCKBUF_UNLOCK(sb);
23166 			error = ENOBUFS;
23167 			sack_rxmit = 0;
23168 			goto out;
23169 		}
23170 		m->m_data += max_linkhdr;
23171 		m->m_len = hdrlen;
23172 
23173 		/*
23174 		 * Start the m_copy functions from the closest mbuf to the
23175 		 * sb_offset in the socket buffer chain.
23176 		 */
23177 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
23178 		s_mb = mb;
23179 		s_moff = moff;
23180 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
23181 			m_copydata(mb, moff, (int)len,
23182 				   mtod(m, caddr_t)+hdrlen);
23183 			/*
23184 			 * If we are not retransmitting advance the
23185 			 * sndptr to help remember the next place in
23186 			 * the sb.
23187 			 */
23188 			if (rsm == NULL)
23189 				sbsndptr_adv(sb, mb, len);
23190 			m->m_len += len;
23191 		} else {
23192 			struct sockbuf *msb;
23193 
23194 			/*
23195 			 * If we are not retransmitting pass in msb so
23196 			 * the socket buffer can be advanced. Otherwise
23197 			 * set it to NULL if its a retransmission since
23198 			 * we don't want to change the sb remembered
23199 			 * location.
23200 			 */
23201 			if (rsm == NULL)
23202 				msb = sb;
23203 			else
23204 				msb = NULL;
23205 			m->m_next = tcp_m_copym(
23206 				mb, moff, &len,
23207 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
23208 				((rsm == NULL) ? hw_tls : 0)
23209 #ifdef NETFLIX_COPY_ARGS
23210 				, &s_mb, &s_moff
23211 #endif
23212 				);
23213 			if (len <= (tp->t_maxseg - optlen)) {
23214 				/*
23215 				 * Must have ran out of mbufs for the copy
23216 				 * shorten it to no longer need tso. Lets
23217 				 * not put on sendalot since we are low on
23218 				 * mbufs.
23219 				 */
23220 				tso = 0;
23221 			}
23222 			if (m->m_next == NULL) {
23223 				SOCKBUF_UNLOCK(sb);
23224 				(void)m_free(m);
23225 				error = ENOBUFS;
23226 				sack_rxmit = 0;
23227 				goto out;
23228 			}
23229 		}
23230 		if (sack_rxmit) {
23231 			if (rsm && (rsm->r_flags & RACK_TLP)) {
23232 				/*
23233 				 * TLP should not count in retran count, but
23234 				 * in its own bin
23235 				 */
23236 				counter_u64_add(rack_tlp_retran, 1);
23237 				counter_u64_add(rack_tlp_retran_bytes, len);
23238 			} else {
23239 				tp->t_sndrexmitpack++;
23240 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
23241 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
23242 			}
23243 #ifdef STATS
23244 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
23245 						 len);
23246 #endif
23247 		} else {
23248 			KMOD_TCPSTAT_INC(tcps_sndpack);
23249 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
23250 #ifdef STATS
23251 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
23252 						 len);
23253 #endif
23254 		}
23255 		/*
23256 		 * If we're sending everything we've got, set PUSH. (This
23257 		 * will keep happy those implementations which only give
23258 		 * data to the user when a buffer fills or a PUSH comes in.)
23259 		 */
23260 		if (sb_offset + len == sbused(sb) &&
23261 		    sbused(sb) &&
23262 		    !(flags & TH_SYN)) {
23263 			flags |= TH_PUSH;
23264 			add_flag |= RACK_HAD_PUSH;
23265 		}
23266 
23267 		SOCKBUF_UNLOCK(sb);
23268 	} else {
23269 		SOCKBUF_UNLOCK(sb);
23270 		if (tp->t_flags & TF_ACKNOW)
23271 			KMOD_TCPSTAT_INC(tcps_sndacks);
23272 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
23273 			KMOD_TCPSTAT_INC(tcps_sndctrl);
23274 		else
23275 			KMOD_TCPSTAT_INC(tcps_sndwinup);
23276 
23277 		m = m_gethdr(M_NOWAIT, MT_DATA);
23278 		if (m == NULL) {
23279 			error = ENOBUFS;
23280 			sack_rxmit = 0;
23281 			goto out;
23282 		}
23283 #ifdef INET6
23284 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
23285 		    MHLEN >= hdrlen) {
23286 			M_ALIGN(m, hdrlen);
23287 		} else
23288 #endif
23289 			m->m_data += max_linkhdr;
23290 		m->m_len = hdrlen;
23291 	}
23292 	SOCKBUF_UNLOCK_ASSERT(sb);
23293 	m->m_pkthdr.rcvif = (struct ifnet *)0;
23294 #ifdef MAC
23295 	mac_inpcb_create_mbuf(inp, m);
23296 #endif
23297 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
23298 #ifdef INET6
23299 		if (isipv6)
23300 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
23301 		else
23302 #endif				/* INET6 */
23303 #ifdef INET
23304 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
23305 #endif
23306 		th = rack->r_ctl.fsb.th;
23307 		udp = rack->r_ctl.fsb.udp;
23308 		if (udp) {
23309 #ifdef INET6
23310 			if (isipv6)
23311 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
23312 			else
23313 #endif				/* INET6 */
23314 				ulen = hdrlen + len - sizeof(struct ip);
23315 			udp->uh_ulen = htons(ulen);
23316 		}
23317 	} else {
23318 #ifdef INET6
23319 		if (isipv6) {
23320 			ip6 = mtod(m, struct ip6_hdr *);
23321 			if (tp->t_port) {
23322 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
23323 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
23324 				udp->uh_dport = tp->t_port;
23325 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
23326 				udp->uh_ulen = htons(ulen);
23327 				th = (struct tcphdr *)(udp + 1);
23328 			} else
23329 				th = (struct tcphdr *)(ip6 + 1);
23330 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
23331 		} else
23332 #endif				/* INET6 */
23333 		{
23334 #ifdef INET
23335 			ip = mtod(m, struct ip *);
23336 			if (tp->t_port) {
23337 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
23338 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
23339 				udp->uh_dport = tp->t_port;
23340 				ulen = hdrlen + len - sizeof(struct ip);
23341 				udp->uh_ulen = htons(ulen);
23342 				th = (struct tcphdr *)(udp + 1);
23343 			} else
23344 				th = (struct tcphdr *)(ip + 1);
23345 			tcpip_fillheaders(inp, tp->t_port, ip, th);
23346 #endif
23347 		}
23348 	}
23349 	/*
23350 	 * If we are starting a connection, send ECN setup SYN packet. If we
23351 	 * are on a retransmit, we may resend those bits a number of times
23352 	 * as per RFC 3168.
23353 	 */
23354 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
23355 		flags |= tcp_ecn_output_syn_sent(tp);
23356 	}
23357 	/* Also handle parallel SYN for ECN */
23358 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
23359 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
23360 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
23361 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
23362 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
23363 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
23364 #ifdef INET6
23365 		if (isipv6) {
23366 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
23367 			ip6->ip6_flow |= htonl(ect << 20);
23368 		}
23369 		else
23370 #endif
23371 		{
23372 #ifdef INET
23373 			ip->ip_tos &= ~IPTOS_ECN_MASK;
23374 			ip->ip_tos |= ect;
23375 #endif
23376 		}
23377 	}
23378 	th->th_seq = htonl(rack_seq);
23379 	th->th_ack = htonl(tp->rcv_nxt);
23380 	tcp_set_flags(th, flags);
23381 	/*
23382 	 * Calculate receive window.  Don't shrink window, but avoid silly
23383 	 * window syndrome.
23384 	 * If a RST segment is sent, advertise a window of zero.
23385 	 */
23386 	if (flags & TH_RST) {
23387 		recwin = 0;
23388 	} else {
23389 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
23390 		    recwin < (long)segsiz) {
23391 			recwin = 0;
23392 		}
23393 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
23394 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
23395 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
23396 	}
23397 
23398 	/*
23399 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
23400 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
23401 	 * handled in syncache.
23402 	 */
23403 	if (flags & TH_SYN)
23404 		th->th_win = htons((u_short)
23405 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
23406 	else {
23407 		/* Avoid shrinking window with window scaling. */
23408 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
23409 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
23410 	}
23411 	/*
23412 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
23413 	 * window.  This may cause the remote transmitter to stall.  This
23414 	 * flag tells soreceive() to disable delayed acknowledgements when
23415 	 * draining the buffer.  This can occur if the receiver is
23416 	 * attempting to read more data than can be buffered prior to
23417 	 * transmitting on the connection.
23418 	 */
23419 	if (th->th_win == 0) {
23420 		tp->t_sndzerowin++;
23421 		tp->t_flags |= TF_RXWIN0SENT;
23422 	} else
23423 		tp->t_flags &= ~TF_RXWIN0SENT;
23424 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
23425 	/* Now are we using fsb?, if so copy the template data to the mbuf */
23426 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
23427 		uint8_t *cpto;
23428 
23429 		cpto = mtod(m, uint8_t *);
23430 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
23431 		/*
23432 		 * We have just copied in:
23433 		 * IP/IP6
23434 		 * <optional udphdr>
23435 		 * tcphdr (no options)
23436 		 *
23437 		 * We need to grab the correct pointers into the mbuf
23438 		 * for both the tcp header, and possibly the udp header (if tunneling).
23439 		 * We do this by using the offset in the copy buffer and adding it
23440 		 * to the mbuf base pointer (cpto).
23441 		 */
23442 #ifdef INET6
23443 		if (isipv6)
23444 			ip6 = mtod(m, struct ip6_hdr *);
23445 		else
23446 #endif				/* INET6 */
23447 #ifdef INET
23448 			ip = mtod(m, struct ip *);
23449 #endif
23450 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
23451 		/* If we have a udp header lets set it into the mbuf as well */
23452 		if (udp)
23453 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
23454 	}
23455 	if (optlen) {
23456 		bcopy(opt, th + 1, optlen);
23457 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
23458 	}
23459 	/*
23460 	 * Put TCP length in extended header, and then checksum extended
23461 	 * header and data.
23462 	 */
23463 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
23464 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
23465 	if (to.to_flags & TOF_SIGNATURE) {
23466 		/*
23467 		 * Calculate MD5 signature and put it into the place
23468 		 * determined before.
23469 		 * NOTE: since TCP options buffer doesn't point into
23470 		 * mbuf's data, calculate offset and use it.
23471 		 */
23472 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
23473 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
23474 			/*
23475 			 * Do not send segment if the calculation of MD5
23476 			 * digest has failed.
23477 			 */
23478 			goto out;
23479 		}
23480 	}
23481 #endif
23482 #ifdef INET6
23483 	if (isipv6) {
23484 		/*
23485 		 * ip6_plen is not need to be filled now, and will be filled
23486 		 * in ip6_output.
23487 		 */
23488 		if (tp->t_port) {
23489 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
23490 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
23491 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
23492 			th->th_sum = htons(0);
23493 			UDPSTAT_INC(udps_opackets);
23494 		} else {
23495 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
23496 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
23497 			th->th_sum = in6_cksum_pseudo(ip6,
23498 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
23499 						      0);
23500 		}
23501 	}
23502 #endif
23503 #if defined(INET6) && defined(INET)
23504 	else
23505 #endif
23506 #ifdef INET
23507 	{
23508 		if (tp->t_port) {
23509 			m->m_pkthdr.csum_flags = CSUM_UDP;
23510 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
23511 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
23512 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
23513 			th->th_sum = htons(0);
23514 			UDPSTAT_INC(udps_opackets);
23515 		} else {
23516 			m->m_pkthdr.csum_flags = CSUM_TCP;
23517 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
23518 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
23519 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
23520 									IPPROTO_TCP + len + optlen));
23521 		}
23522 		/* IP version must be set here for ipv4/ipv6 checking later */
23523 		KASSERT(ip->ip_v == IPVERSION,
23524 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
23525 	}
23526 #endif
23527 	/*
23528 	 * Enable TSO and specify the size of the segments. The TCP pseudo
23529 	 * header checksum is always provided. XXX: Fixme: This is currently
23530 	 * not the case for IPv6.
23531 	 */
23532 	if (tso) {
23533 		/*
23534 		 * Here we must use t_maxseg and the optlen since
23535 		 * the optlen may include SACK's (or DSACK).
23536 		 */
23537 		KASSERT(len > tp->t_maxseg - optlen,
23538 			("%s: len <= tso_segsz", __func__));
23539 		m->m_pkthdr.csum_flags |= CSUM_TSO;
23540 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
23541 	}
23542 	KASSERT(len + hdrlen == m_length(m, NULL),
23543 		("%s: mbuf chain different than expected: %d + %u != %u",
23544 		 __func__, len, hdrlen, m_length(m, NULL)));
23545 
23546 #ifdef TCP_HHOOK
23547 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
23548 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
23549 #endif
23550 	if ((rack->r_ctl.crte != NULL) &&
23551 	    (rack->rc_hw_nobuf == 0) &&
23552 	    tcp_bblogging_on(tp)) {
23553 		rack_log_queue_level(tp, rack, len, &tv, cts);
23554 	}
23555 	/* We're getting ready to send; log now. */
23556 	if (tcp_bblogging_on(rack->rc_tp)) {
23557 		union tcp_log_stackspecific log;
23558 
23559 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
23560 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
23561 		if (rack->rack_no_prr)
23562 			log.u_bbr.flex1 = 0;
23563 		else
23564 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
23565 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
23566 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
23567 		log.u_bbr.flex4 = orig_len;
23568 		/* Save off the early/late values */
23569 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
23570 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
23571 		log.u_bbr.bw_inuse = rack_get_bw(rack);
23572 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
23573 		log.u_bbr.flex8 = 0;
23574 		if (rsm) {
23575 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
23576 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
23577 				counter_u64_add(rack_collapsed_win_rxt, 1);
23578 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
23579 			}
23580 			if (doing_tlp)
23581 				log.u_bbr.flex8 = 2;
23582 			else
23583 				log.u_bbr.flex8 = 1;
23584 		} else {
23585 			if (doing_tlp)
23586 				log.u_bbr.flex8 = 3;
23587 		}
23588 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
23589 		log.u_bbr.flex7 = mark;
23590 		log.u_bbr.flex7 <<= 8;
23591 		log.u_bbr.flex7 |= pass;
23592 		log.u_bbr.pkts_out = tp->t_maxseg;
23593 		log.u_bbr.timeStamp = cts;
23594 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
23595 		if (rsm && (rsm->r_rtr_cnt > 0)) {
23596 			/*
23597 			 * When we have a retransmit we want to log the
23598 			 * burst at send and flight at send from before.
23599 			 */
23600 			log.u_bbr.flex5 = rsm->r_fas;
23601 			log.u_bbr.bbr_substate = rsm->r_bas;
23602 		} else {
23603 			/*
23604 			 * New transmits we log in flex5 the inflight again as
23605 			 * well as the number of segments in our send in the
23606 			 * substate field.
23607 			 */
23608 			log.u_bbr.flex5 = log.u_bbr.inflight;
23609 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
23610 		}
23611 		log.u_bbr.lt_epoch = cwnd_to_use;
23612 		log.u_bbr.delivered = sendalot;
23613 		log.u_bbr.rttProp = (uint64_t)rsm;
23614 		log.u_bbr.pkt_epoch = __LINE__;
23615 		if (rsm) {
23616 			log.u_bbr.delRate = rsm->r_flags;
23617 			log.u_bbr.delRate <<= 31;
23618 			log.u_bbr.delRate |= rack->r_must_retran;
23619 			log.u_bbr.delRate <<= 1;
23620 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
23621 		} else {
23622 			log.u_bbr.delRate = rack->r_must_retran;
23623 			log.u_bbr.delRate <<= 1;
23624 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
23625 		}
23626 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
23627 				    len, &log, false, NULL, __func__, __LINE__, &tv);
23628 	} else
23629 		lgb = NULL;
23630 
23631 	/*
23632 	 * Fill in IP length and desired time to live and send to IP level.
23633 	 * There should be a better way to handle ttl and tos; we could keep
23634 	 * them in the template, but need a way to checksum without them.
23635 	 */
23636 	/*
23637 	 * m->m_pkthdr.len should have been set before cksum calcuration,
23638 	 * because in6_cksum() need it.
23639 	 */
23640 #ifdef INET6
23641 	if (isipv6) {
23642 		/*
23643 		 * we separately set hoplimit for every segment, since the
23644 		 * user might want to change the value via setsockopt. Also,
23645 		 * desired default hop limit might be changed via Neighbor
23646 		 * Discovery.
23647 		 */
23648 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
23649 
23650 		/*
23651 		 * Set the packet size here for the benefit of DTrace
23652 		 * probes. ip6_output() will set it properly; it's supposed
23653 		 * to include the option header lengths as well.
23654 		 */
23655 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
23656 
23657 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
23658 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
23659 		else
23660 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
23661 
23662 		if (tp->t_state == TCPS_SYN_SENT)
23663 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
23664 
23665 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
23666 		/* TODO: IPv6 IP6TOS_ECT bit on */
23667 		error = ip6_output(m,
23668 				   inp->in6p_outputopts,
23669 				   &inp->inp_route6,
23670 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
23671 				   NULL, NULL, inp);
23672 
23673 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
23674 			mtu = inp->inp_route6.ro_nh->nh_mtu;
23675 	}
23676 #endif				/* INET6 */
23677 #if defined(INET) && defined(INET6)
23678 	else
23679 #endif
23680 #ifdef INET
23681 	{
23682 		ip->ip_len = htons(m->m_pkthdr.len);
23683 #ifdef INET6
23684 		if (inp->inp_vflag & INP_IPV6PROTO)
23685 			ip->ip_ttl = in6_selecthlim(inp, NULL);
23686 #endif				/* INET6 */
23687 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
23688 		/*
23689 		 * If we do path MTU discovery, then we set DF on every
23690 		 * packet. This might not be the best thing to do according
23691 		 * to RFC3390 Section 2. However the tcp hostcache migitates
23692 		 * the problem so it affects only the first tcp connection
23693 		 * with a host.
23694 		 *
23695 		 * NB: Don't set DF on small MTU/MSS to have a safe
23696 		 * fallback.
23697 		 */
23698 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
23699 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
23700 			if (tp->t_port == 0 || len < V_tcp_minmss) {
23701 				ip->ip_off |= htons(IP_DF);
23702 			}
23703 		} else {
23704 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
23705 		}
23706 
23707 		if (tp->t_state == TCPS_SYN_SENT)
23708 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
23709 
23710 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
23711 
23712 		error = ip_output(m,
23713 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
23714 				  inp->inp_options,
23715 #else
23716 				  NULL,
23717 #endif
23718 				  &inp->inp_route,
23719 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
23720 				  inp);
23721 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
23722 			mtu = inp->inp_route.ro_nh->nh_mtu;
23723 	}
23724 #endif				/* INET */
23725 	if (lgb) {
23726 		lgb->tlb_errno = error;
23727 		lgb = NULL;
23728 	}
23729 
23730 out:
23731 	/*
23732 	 * In transmit state, time the transmission and arrange for the
23733 	 * retransmit.  In persist state, just set snd_max.
23734 	 */
23735 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
23736 			rack_to_usec_ts(&tv),
23737 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
23738 	if (error == 0) {
23739 		if (add_flag & RACK_IS_PCM) {
23740 			/* We just launched a PCM */
23741 			/* rrs here log */
23742 			rack->pcm_in_progress = 1;
23743 			rack->pcm_needed = 0;
23744 			rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg,  add_flag);
23745 		}
23746 		if (rsm == NULL) {
23747 			if (rack->lt_bw_up == 0) {
23748 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
23749 				rack->r_ctl.lt_seq = tp->snd_una;
23750 				rack->lt_bw_up = 1;
23751 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
23752 				/*
23753 				 * Need to record what we have since we are
23754 				 * approaching seq wrap.
23755 				 */
23756 				uint64_t tmark;
23757 
23758 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
23759 				rack->r_ctl.lt_seq = tp->snd_una;
23760 				tmark = tcp_get_u64_usecs(&tv);
23761 				if (tmark > rack->r_ctl.lt_timemark) {
23762 					rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
23763 					rack->r_ctl.lt_timemark = tmark;
23764 				}
23765 			}
23766 		}
23767 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
23768 		counter_u64_add(rack_total_bytes, len);
23769 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
23770 		if (rsm && doing_tlp) {
23771 			rack->rc_last_sent_tlp_past_cumack = 0;
23772 			rack->rc_last_sent_tlp_seq_valid = 1;
23773 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
23774 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
23775 		}
23776 		if (rack->rc_hw_nobuf) {
23777 			rack->rc_hw_nobuf = 0;
23778 			rack->r_ctl.rc_agg_delayed = 0;
23779 			rack->r_early = 0;
23780 			rack->r_late = 0;
23781 			rack->r_ctl.rc_agg_early = 0;
23782 		}
23783 		if (rsm && (doing_tlp == 0)) {
23784 			/* Set we retransmitted */
23785 			rack->rc_gp_saw_rec = 1;
23786 		} else {
23787 			if (cwnd_to_use > tp->snd_ssthresh) {
23788 				/* Set we sent in CA */
23789 				rack->rc_gp_saw_ca = 1;
23790 			} else {
23791 				/* Set we sent in SS */
23792 				rack->rc_gp_saw_ss = 1;
23793 			}
23794 		}
23795 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
23796 		    (tp->t_flags & TF_SACK_PERMIT) &&
23797 		    tp->rcv_numsacks > 0)
23798 			tcp_clean_dsack_blocks(tp);
23799 		tot_len_this_send += len;
23800 		if (len == 0) {
23801 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
23802 		} else {
23803 			int idx;
23804 
23805 			idx = (len / segsiz) + 3;
23806 			if (idx >= TCP_MSS_ACCT_ATIMER)
23807 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
23808 			else
23809 				counter_u64_add(rack_out_size[idx], 1);
23810 		}
23811 	}
23812 	if ((rack->rack_no_prr == 0) &&
23813 	    sub_from_prr &&
23814 	    (error == 0)) {
23815 		if (rack->r_ctl.rc_prr_sndcnt >= len)
23816 			rack->r_ctl.rc_prr_sndcnt -= len;
23817 		else
23818 			rack->r_ctl.rc_prr_sndcnt = 0;
23819 	}
23820 	sub_from_prr = 0;
23821 	if (doing_tlp) {
23822 		/* Make sure the TLP is added */
23823 		add_flag |= RACK_TLP;
23824 	} else if (rsm) {
23825 		/* If its a resend without TLP then it must not have the flag */
23826 		rsm->r_flags &= ~RACK_TLP;
23827 	}
23828 
23829 
23830 	if ((error == 0) &&
23831 	    (len > 0) &&
23832 	    (tp->snd_una == tp->snd_max))
23833 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
23834 
23835 	{
23836 		/*
23837 		 * This block is not associated with the above error == 0 test.
23838 		 * It is used to advance snd_max if we have a new transmit.
23839 		 */
23840 		tcp_seq startseq = tp->snd_max;
23841 
23842 
23843 		if (rsm && (doing_tlp == 0))
23844 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
23845 		if (error)
23846 			/* We don't log or do anything with errors */
23847 			goto nomore;
23848 		if (doing_tlp == 0) {
23849 			if (rsm == NULL) {
23850 				/*
23851 				 * Not a retransmission of some
23852 				 * sort, new data is going out so
23853 				 * clear our TLP count and flag.
23854 				 */
23855 				rack->rc_tlp_in_progress = 0;
23856 				rack->r_ctl.rc_tlp_cnt_out = 0;
23857 			}
23858 		} else {
23859 			/*
23860 			 * We have just sent a TLP, mark that it is true
23861 			 * and make sure our in progress is set so we
23862 			 * continue to check the count.
23863 			 */
23864 			rack->rc_tlp_in_progress = 1;
23865 			rack->r_ctl.rc_tlp_cnt_out++;
23866 		}
23867 		/*
23868 		 * If we are retransmitting we are done, snd_max
23869 		 * does not get updated.
23870 		 */
23871 		if (sack_rxmit)
23872 			goto nomore;
23873 		if ((tp->snd_una == tp->snd_max) && (len > 0)) {
23874 			/*
23875 			 * Update the time we just added data since
23876 			 * nothing was outstanding.
23877 			 */
23878 			rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
23879 			tp->t_acktime = ticks;
23880 		}
23881 		/*
23882 		 * Now for special SYN/FIN handling.
23883 		 */
23884 		if (flags & (TH_SYN | TH_FIN)) {
23885 			if ((flags & TH_SYN) &&
23886 			    ((tp->t_flags & TF_SENTSYN) == 0)) {
23887 				tp->snd_max++;
23888 				tp->t_flags |= TF_SENTSYN;
23889 			}
23890 			if ((flags & TH_FIN) &&
23891 			    ((tp->t_flags & TF_SENTFIN) == 0)) {
23892 				tp->snd_max++;
23893 				tp->t_flags |= TF_SENTFIN;
23894 			}
23895 		}
23896 		tp->snd_max += len;
23897 		if (rack->rc_new_rnd_needed) {
23898 			rack_new_round_starts(tp, rack, tp->snd_max);
23899 		}
23900 		/*
23901 		 * Time this transmission if not a retransmission and
23902 		 * not currently timing anything.
23903 		 * This is only relevant in case of switching back to
23904 		 * the base stack.
23905 		 */
23906 		if (tp->t_rtttime == 0) {
23907 			tp->t_rtttime = ticks;
23908 			tp->t_rtseq = startseq;
23909 			KMOD_TCPSTAT_INC(tcps_segstimed);
23910 		}
23911 		if (len &&
23912 		    ((tp->t_flags & TF_GPUTINPROG) == 0))
23913 			rack_start_gp_measurement(tp, rack, startseq, sb_offset);
23914 		/*
23915 		 * If we are doing FO we need to update the mbuf position and subtract
23916 		 * this happens when the peer sends us duplicate information and
23917 		 * we thus want to send a DSACK.
23918 		 *
23919 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
23920 		 * turned off? If not then we are going to echo multiple DSACK blocks
23921 		 * out (with the TSO), which we should not be doing.
23922 		 */
23923 		if (rack->r_fast_output && len) {
23924 			if (rack->r_ctl.fsb.left_to_send > len)
23925 				rack->r_ctl.fsb.left_to_send -= len;
23926 			else
23927 				rack->r_ctl.fsb.left_to_send = 0;
23928 			if (rack->r_ctl.fsb.left_to_send < segsiz)
23929 				rack->r_fast_output = 0;
23930 			if (rack->r_fast_output) {
23931 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
23932 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
23933 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
23934 			}
23935 		}
23936 		if (rack_pcm_blast == 0) {
23937 			if ((orig_len > len) &&
23938 			    (add_flag & RACK_IS_PCM) &&
23939 			    (len < pace_max_seg) &&
23940 			    ((pace_max_seg - len) > segsiz)) {
23941 				/*
23942 				 * We are doing a PCM measurement and we did
23943 				 * not get enough data in the TSO to meet the
23944 				 * burst requirement.
23945 				 */
23946 				uint32_t n_len;
23947 
23948 				n_len = (orig_len - len);
23949 				orig_len -= len;
23950 				pace_max_seg -= len;
23951 				len = n_len;
23952 				sb_offset = tp->snd_max - tp->snd_una;
23953 				/* Re-lock for the next spin */
23954 				SOCKBUF_LOCK(sb);
23955 				goto send;
23956 			}
23957 		} else {
23958 			if ((orig_len > len) &&
23959 			    (add_flag & RACK_IS_PCM) &&
23960 			    ((orig_len - len) > segsiz)) {
23961 				/*
23962 				 * We are doing a PCM measurement and we did
23963 				 * not get enough data in the TSO to meet the
23964 				 * burst requirement.
23965 				 */
23966 				uint32_t n_len;
23967 
23968 				n_len = (orig_len - len);
23969 				orig_len -= len;
23970 				len = n_len;
23971 				sb_offset = tp->snd_max - tp->snd_una;
23972 				/* Re-lock for the next spin */
23973 				SOCKBUF_LOCK(sb);
23974 				goto send;
23975 			}
23976 		}
23977 	}
23978 nomore:
23979 	if (error) {
23980 		rack->r_ctl.rc_agg_delayed = 0;
23981 		rack->r_early = 0;
23982 		rack->r_late = 0;
23983 		rack->r_ctl.rc_agg_early = 0;
23984 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
23985 		/*
23986 		 * Failures do not advance the seq counter above. For the
23987 		 * case of ENOBUFS we will fall out and retry in 1ms with
23988 		 * the hpts. Everything else will just have to retransmit
23989 		 * with the timer.
23990 		 *
23991 		 * In any case, we do not want to loop around for another
23992 		 * send without a good reason.
23993 		 */
23994 		sendalot = 0;
23995 		switch (error) {
23996 		case EPERM:
23997 		case EACCES:
23998 			tp->t_softerror = error;
23999 #ifdef TCP_ACCOUNTING
24000 			crtsc = get_cyclecount();
24001 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24002 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
24003 			}
24004 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24005 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
24006 			}
24007 			sched_unpin();
24008 #endif
24009 			return (error);
24010 		case ENOBUFS:
24011 			/*
24012 			 * Pace us right away to retry in a some
24013 			 * time
24014 			 */
24015 			if (rack->r_ctl.crte != NULL) {
24016 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
24017 				if (tcp_bblogging_on(rack->rc_tp))
24018 					rack_log_queue_level(tp, rack, len, &tv, cts);
24019 			} else
24020 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
24021 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
24022 			if (rack->rc_enobuf < 0x7f)
24023 				rack->rc_enobuf++;
24024 			if (slot < (10 * HPTS_USEC_IN_MSEC))
24025 				slot = 10 * HPTS_USEC_IN_MSEC;
24026 			if (rack->r_ctl.crte != NULL) {
24027 				counter_u64_add(rack_saw_enobuf_hw, 1);
24028 				tcp_rl_log_enobuf(rack->r_ctl.crte);
24029 			}
24030 			counter_u64_add(rack_saw_enobuf, 1);
24031 			goto enobufs;
24032 		case EMSGSIZE:
24033 			/*
24034 			 * For some reason the interface we used initially
24035 			 * to send segments changed to another or lowered
24036 			 * its MTU. If TSO was active we either got an
24037 			 * interface without TSO capabilits or TSO was
24038 			 * turned off. If we obtained mtu from ip_output()
24039 			 * then update it and try again.
24040 			 */
24041 			if (tso)
24042 				tp->t_flags &= ~TF_TSO;
24043 			if (mtu != 0) {
24044 				int saved_mtu;
24045 
24046 				saved_mtu = tp->t_maxseg;
24047 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
24048 				if (saved_mtu > tp->t_maxseg) {
24049 					goto again;
24050 				}
24051 			}
24052 			slot = 10 * HPTS_USEC_IN_MSEC;
24053 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
24054 #ifdef TCP_ACCOUNTING
24055 			crtsc = get_cyclecount();
24056 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24057 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
24058 			}
24059 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24060 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
24061 			}
24062 			sched_unpin();
24063 #endif
24064 			return (error);
24065 		case ENETUNREACH:
24066 			counter_u64_add(rack_saw_enetunreach, 1);
24067 		case EHOSTDOWN:
24068 		case EHOSTUNREACH:
24069 		case ENETDOWN:
24070 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
24071 				tp->t_softerror = error;
24072 			}
24073 			/* FALLTHROUGH */
24074 		default:
24075 			slot = 10 * HPTS_USEC_IN_MSEC;
24076 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
24077 #ifdef TCP_ACCOUNTING
24078 			crtsc = get_cyclecount();
24079 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24080 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
24081 			}
24082 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24083 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
24084 			}
24085 			sched_unpin();
24086 #endif
24087 			return (error);
24088 		}
24089 	} else {
24090 		rack->rc_enobuf = 0;
24091 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
24092 			rack->r_ctl.retran_during_recovery += len;
24093 	}
24094 	KMOD_TCPSTAT_INC(tcps_sndtotal);
24095 
24096 	/*
24097 	 * Data sent (as far as we can tell). If this advertises a larger
24098 	 * window than any other segment, then remember the size of the
24099 	 * advertised window. Any pending ACK has now been sent.
24100 	 */
24101 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
24102 		tp->rcv_adv = tp->rcv_nxt + recwin;
24103 
24104 	tp->last_ack_sent = tp->rcv_nxt;
24105 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
24106 enobufs:
24107 	if (sendalot) {
24108 		/* Do we need to turn off sendalot? */
24109 		if (pace_max_seg &&
24110 		    (tot_len_this_send >= pace_max_seg)) {
24111 			/* We hit our max. */
24112 			sendalot = 0;
24113 		}
24114 	}
24115 	if ((error == 0) && (flags & TH_FIN))
24116 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
24117 	if (flags & TH_RST) {
24118 		/*
24119 		 * We don't send again after sending a RST.
24120 		 */
24121 		slot = 0;
24122 		sendalot = 0;
24123 		if (error == 0)
24124 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
24125 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
24126 		/*
24127 		 * Get our pacing rate, if an error
24128 		 * occurred in sending (ENOBUF) we would
24129 		 * hit the else if with slot preset. Other
24130 		 * errors return.
24131 		 */
24132 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
24133 	}
24134 	/* We have sent clear the flag */
24135 	rack->r_ent_rec_ns = 0;
24136 	if (rack->r_must_retran) {
24137 		if (rsm) {
24138 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
24139 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
24140 				/*
24141 				 * We have retransmitted all.
24142 				 */
24143 				rack->r_must_retran = 0;
24144 				rack->r_ctl.rc_out_at_rto = 0;
24145 			}
24146 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
24147 			/*
24148 			 * Sending new data will also kill
24149 			 * the loop.
24150 			 */
24151 			rack->r_must_retran = 0;
24152 			rack->r_ctl.rc_out_at_rto = 0;
24153 		}
24154 	}
24155 	rack->r_ctl.fsb.recwin = recwin;
24156 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
24157 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
24158 		/*
24159 		 * We hit an RTO and now have past snd_max at the RTO
24160 		 * clear all the WAS flags.
24161 		 */
24162 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
24163 	}
24164 	if (slot) {
24165 		/* set the rack tcb into the slot N */
24166 		if ((error == 0) &&
24167 		    rack_use_rfo &&
24168 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
24169 		    (rsm == NULL) &&
24170 		    (ipoptlen == 0) &&
24171 		    (tp->rcv_numsacks == 0) &&
24172 		    (rack->rc_policer_detected == 0)  &&
24173 		    rack->r_fsb_inited &&
24174 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
24175 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
24176 		    (rack->r_must_retran == 0) &&
24177 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
24178 		    (len > 0) && (orig_len > 0) &&
24179 		    (orig_len > len) &&
24180 		    ((orig_len - len) >= segsiz) &&
24181 		    ((optlen == 0) ||
24182 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
24183 			/* We can send at least one more MSS using our fsb */
24184 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
24185 					       segsiz, pace_max_seg, hw_tls, flags);
24186 		} else
24187 			rack->r_fast_output = 0;
24188 		rack_log_fsb(rack, tp, so, flags,
24189 			     ipoptlen, orig_len, len, error,
24190 			     (rsm == NULL), optlen, __LINE__, 2);
24191 	} else if (sendalot) {
24192 		int ret;
24193 
24194 		sack_rxmit = 0;
24195 		if ((error == 0) &&
24196 		    rack_use_rfo &&
24197 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
24198 		    (rsm == NULL) &&
24199 		    (ipoptlen == 0) &&
24200 		    (tp->rcv_numsacks == 0) &&
24201 		    (rack->r_must_retran == 0) &&
24202 		    rack->r_fsb_inited &&
24203 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
24204 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
24205 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
24206 		    (len > 0) && (orig_len > 0) &&
24207 		    (orig_len > len) &&
24208 		    ((orig_len - len) >= segsiz) &&
24209 		    ((optlen == 0) ||
24210 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
24211 			/* we can use fast_output for more */
24212 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
24213 					       segsiz, pace_max_seg, hw_tls, flags);
24214 			if (rack->r_fast_output) {
24215 				error = 0;
24216 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
24217 				if (ret >= 0)
24218 					return (ret);
24219 			        else if (error)
24220 					goto nomore;
24221 
24222 			}
24223 		}
24224 		goto again;
24225 	}
24226 skip_all_send:
24227 	/* Assure when we leave that snd_nxt will point to top */
24228 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
24229 		tp->snd_nxt = tp->snd_max;
24230 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
24231 #ifdef TCP_ACCOUNTING
24232 	crtsc = get_cyclecount() - ts_val;
24233 	if (tot_len_this_send) {
24234 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24235 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
24236 		}
24237 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24238 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
24239 		}
24240 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24241 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
24242 		}
24243 	} else {
24244 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24245 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
24246 		}
24247 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
24248 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
24249 		}
24250 	}
24251 	sched_unpin();
24252 #endif
24253 	if (error == ENOBUFS)
24254 		error = 0;
24255 	return (error);
24256 }
24257 
24258 static void
24259 rack_update_seg(struct tcp_rack *rack)
24260 {
24261 	uint32_t orig_val;
24262 
24263 	orig_val = rack->r_ctl.rc_pace_max_segs;
24264 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
24265 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
24266 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
24267 }
24268 
24269 static void
24270 rack_mtu_change(struct tcpcb *tp)
24271 {
24272 	/*
24273 	 * The MSS may have changed
24274 	 */
24275 	struct tcp_rack *rack;
24276 	struct rack_sendmap *rsm;
24277 
24278 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24279 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
24280 		/*
24281 		 * The MTU has changed we need to resend everything
24282 		 * since all we have sent is lost. We first fix
24283 		 * up the mtu though.
24284 		 */
24285 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
24286 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
24287 		rack_remxt_tmr(tp);
24288 		rack->r_fast_output = 0;
24289 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
24290 						rack->r_ctl.rc_sacked);
24291 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
24292 		rack->r_must_retran = 1;
24293 		/* Mark all inflight to needing to be rxt'd */
24294 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
24295 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
24296 		}
24297 	}
24298 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
24299 	/* We don't use snd_nxt to retransmit */
24300 	tp->snd_nxt = tp->snd_max;
24301 }
24302 
24303 static int
24304 rack_set_dgp(struct tcp_rack *rack)
24305 {
24306 	if (rack->dgp_on == 1)
24307 		return(0);
24308 	if ((rack->use_fixed_rate == 1) &&
24309 	    (rack->rc_always_pace == 1)) {
24310 		/*
24311 		 * We are already pacing another
24312 		 * way.
24313 		 */
24314 		return (EBUSY);
24315 	}
24316 	if (rack->rc_always_pace == 1) {
24317 		rack_remove_pacing(rack);
24318 	}
24319 	if (tcp_incr_dgp_pacing_cnt() == 0)
24320 		return (ENOSPC);
24321 	rack->r_ctl.pacing_method |= RACK_DGP_PACING;
24322 	rack->rc_fillcw_apply_discount = 0;
24323 	rack->dgp_on = 1;
24324 	rack->rc_always_pace = 1;
24325 	rack->rc_pace_dnd = 1;
24326 	rack->use_fixed_rate = 0;
24327 	if (rack->gp_ready)
24328 		rack_set_cc_pacing(rack);
24329 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24330 	rack->rack_attempt_hdwr_pace = 0;
24331 	/* rxt settings */
24332 	rack->full_size_rxt = 1;
24333 	rack->shape_rxt_to_pacing_min  = 0;
24334 	/* cmpack=1 */
24335 	rack->r_use_cmp_ack = 1;
24336 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
24337 	    rack->r_use_cmp_ack)
24338 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
24339 	/* scwnd=1 */
24340 	rack->rack_enable_scwnd = 1;
24341 	/* dynamic=100 */
24342 	rack->rc_gp_dyn_mul = 1;
24343 	/* gp_inc_ca */
24344 	rack->r_ctl.rack_per_of_gp_ca = 100;
24345 	/* rrr_conf=3 */
24346 	rack->r_rr_config = 3;
24347 	/* npush=2 */
24348 	rack->r_ctl.rc_no_push_at_mrtt = 2;
24349 	/* fillcw=1 */
24350 	rack->rc_pace_to_cwnd = 1;
24351 	rack->rc_pace_fill_if_rttin_range = 0;
24352 	rack->rtt_limit_mul = 0;
24353 	/* noprr=1 */
24354 	rack->rack_no_prr = 1;
24355 	/* lscwnd=1 */
24356 	rack->r_limit_scw = 1;
24357 	/* gp_inc_rec */
24358 	rack->r_ctl.rack_per_of_gp_rec = 90;
24359 	return (0);
24360 }
24361 
24362 static int
24363 rack_set_profile(struct tcp_rack *rack, int prof)
24364 {
24365 	int err = EINVAL;
24366 	if (prof == 1) {
24367 		/*
24368 		 * Profile 1 is "standard" DGP. It ignores
24369 		 * client buffer level.
24370 		 */
24371 		err = rack_set_dgp(rack);
24372 		if (err)
24373 			return (err);
24374 	} else if (prof == 6) {
24375 		err = rack_set_dgp(rack);
24376 		if (err)
24377 			return (err);
24378 		/*
24379 		 * Profile 6 tweaks DGP so that it will apply to
24380 		 * fill-cw the same settings that profile5 does
24381 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
24382 		 */
24383 		rack->rc_fillcw_apply_discount = 1;
24384 	} else if (prof == 0) {
24385 		/* This changes things back to the default settings */
24386 		if (rack->rc_always_pace == 1) {
24387 			rack_remove_pacing(rack);
24388 		} else {
24389 			/* Make sure any stray flags are off */
24390 			rack->dgp_on = 0;
24391 			rack->rc_hybrid_mode = 0;
24392 			rack->use_fixed_rate = 0;
24393 		}
24394 		err = 0;
24395 		if (rack_fill_cw_state)
24396 			rack->rc_pace_to_cwnd = 1;
24397 		else
24398 			rack->rc_pace_to_cwnd = 0;
24399 
24400 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
24401 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
24402 			rack->rc_always_pace = 1;
24403 			if (rack->rack_hibeta)
24404 				rack_set_cc_pacing(rack);
24405 		} else
24406 			rack->rc_always_pace = 0;
24407 		if (rack_dsack_std_based & 0x1) {
24408 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
24409 			rack->rc_rack_tmr_std_based = 1;
24410 		}
24411 		if (rack_dsack_std_based & 0x2) {
24412 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
24413 			rack->rc_rack_use_dsack = 1;
24414 		}
24415 		if (rack_use_cmp_acks)
24416 			rack->r_use_cmp_ack = 1;
24417 		else
24418 			rack->r_use_cmp_ack = 0;
24419 		if (rack_disable_prr)
24420 			rack->rack_no_prr = 1;
24421 		else
24422 			rack->rack_no_prr = 0;
24423 		if (rack_gp_no_rec_chg)
24424 			rack->rc_gp_no_rec_chg = 1;
24425 		else
24426 			rack->rc_gp_no_rec_chg = 0;
24427 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
24428 			rack->r_mbuf_queue = 1;
24429 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
24430 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
24431 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24432 		} else {
24433 			rack->r_mbuf_queue = 0;
24434 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
24435 		}
24436 		if (rack_enable_shared_cwnd)
24437 			rack->rack_enable_scwnd = 1;
24438 		else
24439 			rack->rack_enable_scwnd = 0;
24440 		if (rack_do_dyn_mul) {
24441 			/* When dynamic adjustment is on CA needs to start at 100% */
24442 			rack->rc_gp_dyn_mul = 1;
24443 			if (rack_do_dyn_mul >= 100)
24444 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
24445 		} else {
24446 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
24447 			rack->rc_gp_dyn_mul = 0;
24448 		}
24449 		rack->r_rr_config = 0;
24450 		rack->r_ctl.rc_no_push_at_mrtt = 0;
24451 		rack->rc_pace_fill_if_rttin_range = 0;
24452 		rack->rtt_limit_mul = 0;
24453 
24454 		if (rack_enable_hw_pacing)
24455 			rack->rack_hdw_pace_ena = 1;
24456 		else
24457 			rack->rack_hdw_pace_ena = 0;
24458 		if (rack_disable_prr)
24459 			rack->rack_no_prr = 1;
24460 		else
24461 			rack->rack_no_prr = 0;
24462 		if (rack_limits_scwnd)
24463 			rack->r_limit_scw  = 1;
24464 		else
24465 			rack->r_limit_scw  = 0;
24466 		rack_init_retransmit_value(rack, rack_rxt_controls);
24467 		err = 0;
24468 	}
24469 	return (err);
24470 }
24471 
24472 static int
24473 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
24474 {
24475 	struct deferred_opt_list *dol;
24476 
24477 	dol = malloc(sizeof(struct deferred_opt_list),
24478 		     M_TCPDO, M_NOWAIT|M_ZERO);
24479 	if (dol == NULL) {
24480 		/*
24481 		 * No space yikes -- fail out..
24482 		 */
24483 		return (0);
24484 	}
24485 	dol->optname = sopt_name;
24486 	dol->optval = loptval;
24487 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
24488 	return (1);
24489 }
24490 
24491 static int
24492 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
24493 {
24494 #ifdef TCP_REQUEST_TRK
24495 	struct tcp_sendfile_track *sft;
24496 	struct timeval tv;
24497 	tcp_seq seq;
24498 	int err;
24499 
24500 	microuptime(&tv);
24501 
24502 	/* Make sure no fixed rate is on */
24503 	rack->use_fixed_rate = 0;
24504 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
24505 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
24506 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
24507 	/* Now allocate or find our entry that will have these settings */
24508 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
24509 	if (sft == NULL) {
24510 		rack->rc_tp->tcp_hybrid_error++;
24511 		/* no space, where would it have gone? */
24512 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
24513 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
24514 		return (ENOSPC);
24515 	}
24516 	/* mask our internal flags */
24517 	hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
24518 	/* The seq will be snd_una + everything in the buffer */
24519 	seq = sft->start_seq;
24520 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
24521 		/* Disabling hybrid pacing */
24522 		if (rack->rc_hybrid_mode) {
24523 			rack_set_profile(rack, 0);
24524 			rack->rc_tp->tcp_hybrid_stop++;
24525 		}
24526 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
24527 		return (0);
24528 	}
24529 	if (rack->dgp_on == 0) {
24530 		/*
24531 		 * If we have not yet turned DGP on, do so
24532 		 * now setting pure DGP mode, no buffer level
24533 		 * response.
24534 		 */
24535 		if ((err = rack_set_profile(rack, 1)) != 0){
24536 			/* Failed to turn pacing on */
24537 			rack->rc_tp->tcp_hybrid_error++;
24538 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
24539 			return (err);
24540 		}
24541 	}
24542 	/*
24543 	 * Now we must switch to hybrid mode as well which also
24544 	 * means moving to regular pacing.
24545 	 */
24546 	if (rack->rc_hybrid_mode == 0) {
24547 		/* First time */
24548 		if (tcp_can_enable_pacing()) {
24549 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
24550 			rack->rc_hybrid_mode = 1;
24551 		} else {
24552 			return (ENOSPC);
24553 		}
24554 		if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
24555 			/*
24556 			 * This should be true.
24557 			 */
24558 			tcp_dec_dgp_pacing_cnt();
24559 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
24560 		}
24561 	}
24562 	/* Now set in our flags */
24563 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
24564 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
24565 		sft->cspr = hybrid->cspr;
24566 	else
24567 		sft->cspr = 0;
24568 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
24569 		sft->hint_maxseg = hybrid->hint_maxseg;
24570 	else
24571 		sft->hint_maxseg = 0;
24572 	rack->rc_tp->tcp_hybrid_start++;
24573 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
24574 	return (0);
24575 #else
24576 	return (ENOTSUP);
24577 #endif
24578 }
24579 
24580 static int
24581 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
24582 {
24583 	/*
24584 	 * Gather rack specific information.
24585 	 */
24586 	struct tcp_rack *rack;
24587 
24588 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24589 	/* We pulled a SSI info log out what was there */
24590 	policer_detection_log(rack, rack->rc_highly_buffered, 0, 0, 0, 20);
24591 	if (rack->policer_detect_on) {
24592 		si->policer_detection_enabled = 1;
24593 		if (rack->rc_policer_detected) {
24594 			si->policer_detected = 1;
24595 			si->policer_bucket_size = rack->r_ctl.policer_bucket_size;
24596 			si->policer_last_bw = rack->r_ctl.policer_bw;
24597 		} else {
24598 			si->policer_detected = 0;
24599 			si->policer_bucket_size = 0;
24600 			si->policer_last_bw = 0;
24601 		}
24602 		si->current_round = rack->r_ctl.current_round;
24603 		si->highly_buffered = rack->rc_highly_buffered;
24604 	}
24605 	si->bytes_transmitted = tp->t_sndbytes;
24606 	si->bytes_retransmitted = tp->t_snd_rxt_bytes;
24607 	return (0);
24608 }
24609 
24610 static int
24611 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
24612 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
24613 
24614 {
24615 	struct epoch_tracker et;
24616 	struct sockopt sopt;
24617 	struct cc_newreno_opts opt;
24618 	uint64_t val;
24619 	int error = 0;
24620 	uint16_t ca, ss;
24621 
24622 	switch (sopt_name) {
24623 	case TCP_RACK_SET_RXT_OPTIONS:
24624 		if ((optval >= 0) && (optval <= 2)) {
24625 			rack_init_retransmit_value(rack, optval);
24626 		} else {
24627 			/*
24628 			 * You must send in 0, 1 or 2 all else is
24629 			 * invalid.
24630 			 */
24631 			error = EINVAL;
24632 		}
24633 		break;
24634 	case TCP_RACK_DSACK_OPT:
24635 		RACK_OPTS_INC(tcp_rack_dsack_opt);
24636 		if (optval & 0x1) {
24637 			rack->rc_rack_tmr_std_based = 1;
24638 		} else {
24639 			rack->rc_rack_tmr_std_based = 0;
24640 		}
24641 		if (optval & 0x2) {
24642 			rack->rc_rack_use_dsack = 1;
24643 		} else {
24644 			rack->rc_rack_use_dsack = 0;
24645 		}
24646 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
24647 		break;
24648 	case TCP_RACK_PACING_DIVISOR:
24649 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
24650 		if (optval == 0) {
24651 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
24652 		} else {
24653 			if (optval < RL_MIN_DIVISOR)
24654 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
24655 			else
24656 				rack->r_ctl.pace_len_divisor = optval;
24657 		}
24658 		break;
24659 	case TCP_RACK_HI_BETA:
24660 		RACK_OPTS_INC(tcp_rack_hi_beta);
24661 		if (optval > 0) {
24662 			rack->rack_hibeta = 1;
24663 			if ((optval >= 50) &&
24664 			    (optval <= 100)) {
24665 				/*
24666 				 * User wants to set a custom beta.
24667 				 */
24668 				rack->r_ctl.saved_hibeta = optval;
24669 				if (rack->rc_pacing_cc_set)
24670 					rack_undo_cc_pacing(rack);
24671 				rack->r_ctl.rc_saved_beta.beta = optval;
24672 			}
24673 			if (rack->rc_pacing_cc_set == 0)
24674 				rack_set_cc_pacing(rack);
24675 		} else {
24676 			rack->rack_hibeta = 0;
24677 			if (rack->rc_pacing_cc_set)
24678 				rack_undo_cc_pacing(rack);
24679 		}
24680 		break;
24681 	case TCP_RACK_PACING_BETA:
24682 		error = EINVAL;
24683 		break;
24684 	case TCP_RACK_TIMER_SLOP:
24685 		RACK_OPTS_INC(tcp_rack_timer_slop);
24686 		rack->r_ctl.timer_slop = optval;
24687 		if (rack->rc_tp->t_srtt) {
24688 			/*
24689 			 * If we have an SRTT lets update t_rxtcur
24690 			 * to have the new slop.
24691 			 */
24692 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
24693 					   rack_rto_min, rack_rto_max,
24694 					   rack->r_ctl.timer_slop);
24695 		}
24696 		break;
24697 	case TCP_RACK_PACING_BETA_ECN:
24698 		RACK_OPTS_INC(tcp_rack_beta_ecn);
24699 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
24700 			/* This only works for newreno. */
24701 			error = EINVAL;
24702 			break;
24703 		}
24704 		if (rack->rc_pacing_cc_set) {
24705 			/*
24706 			 * Set them into the real CC module
24707 			 * whats in the rack pcb is the old values
24708 			 * to be used on restoral/
24709 			 */
24710 			sopt.sopt_dir = SOPT_SET;
24711 			opt.name = CC_NEWRENO_BETA_ECN;
24712 			opt.val = optval;
24713 			if (CC_ALGO(tp)->ctl_output != NULL)
24714 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
24715 			else
24716 				error = ENOENT;
24717 		} else {
24718 			/*
24719 			 * Not pacing yet so set it into our local
24720 			 * rack pcb storage.
24721 			 */
24722 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
24723 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
24724 		}
24725 		break;
24726 	case TCP_DEFER_OPTIONS:
24727 		RACK_OPTS_INC(tcp_defer_opt);
24728 		if (optval) {
24729 			if (rack->gp_ready) {
24730 				/* Too late */
24731 				error = EINVAL;
24732 				break;
24733 			}
24734 			rack->defer_options = 1;
24735 		} else
24736 			rack->defer_options = 0;
24737 		break;
24738 	case TCP_RACK_MEASURE_CNT:
24739 		RACK_OPTS_INC(tcp_rack_measure_cnt);
24740 		if (optval && (optval <= 0xff)) {
24741 			rack->r_ctl.req_measurements = optval;
24742 		} else
24743 			error = EINVAL;
24744 		break;
24745 	case TCP_REC_ABC_VAL:
24746 		RACK_OPTS_INC(tcp_rec_abc_val);
24747 		if (optval > 0)
24748 			rack->r_use_labc_for_rec = 1;
24749 		else
24750 			rack->r_use_labc_for_rec = 0;
24751 		break;
24752 	case TCP_RACK_ABC_VAL:
24753 		RACK_OPTS_INC(tcp_rack_abc_val);
24754 		if ((optval > 0) && (optval < 255))
24755 			rack->rc_labc = optval;
24756 		else
24757 			error = EINVAL;
24758 		break;
24759 	case TCP_HDWR_UP_ONLY:
24760 		RACK_OPTS_INC(tcp_pacing_up_only);
24761 		if (optval)
24762 			rack->r_up_only = 1;
24763 		else
24764 			rack->r_up_only = 0;
24765 		break;
24766 	case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
24767 		RACK_OPTS_INC(tcp_fillcw_rate_cap);
24768 		rack->r_ctl.fillcw_cap = loptval;
24769 		break;
24770 	case TCP_PACING_RATE_CAP:
24771 		RACK_OPTS_INC(tcp_pacing_rate_cap);
24772 		if ((rack->dgp_on == 1) &&
24773 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
24774 			/*
24775 			 * If we are doing DGP we need to switch
24776 			 * to using the pacing limit.
24777 			 */
24778 			if (tcp_can_enable_pacing() == 0) {
24779 				error = ENOSPC;
24780 				break;
24781 			}
24782 			/*
24783 			 * Now change up the flags and counts to be correct.
24784 			 */
24785 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
24786 			tcp_dec_dgp_pacing_cnt();
24787 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
24788 		}
24789 		rack->r_ctl.bw_rate_cap = loptval;
24790 		break;
24791 	case TCP_HYBRID_PACING:
24792 		if (hybrid == NULL) {
24793 			error = EINVAL;
24794 			break;
24795 		}
24796 		if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
24797 			error = EPERM;
24798 			break;
24799 		}
24800 		error = process_hybrid_pacing(rack, hybrid);
24801 		break;
24802 	case TCP_SIDECHAN_DIS:			/*  URL:scodm */
24803 		if (optval)
24804 			rack->r_ctl.side_chan_dis_mask = optval;
24805 		else
24806 			rack->r_ctl.side_chan_dis_mask = 0;
24807 		break;
24808 	case TCP_RACK_PROFILE:
24809 		RACK_OPTS_INC(tcp_profile);
24810 		error = rack_set_profile(rack, optval);
24811 		break;
24812 	case TCP_USE_CMP_ACKS:
24813 		RACK_OPTS_INC(tcp_use_cmp_acks);
24814 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
24815 			/* You can't turn it off once its on! */
24816 			error = EINVAL;
24817 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
24818 			rack->r_use_cmp_ack = 1;
24819 			rack->r_mbuf_queue = 1;
24820 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24821 		}
24822 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
24823 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
24824 		break;
24825 	case TCP_SHARED_CWND_TIME_LIMIT:
24826 		RACK_OPTS_INC(tcp_lscwnd);
24827 		if (optval)
24828 			rack->r_limit_scw = 1;
24829 		else
24830 			rack->r_limit_scw = 0;
24831 		break;
24832 	case TCP_RACK_DGP_IN_REC:
24833 		error = EINVAL;
24834 		break;
24835 	case TCP_POLICER_DETECT:		/*  URL:pol_det */
24836 		RACK_OPTS_INC(tcp_pol_detect);
24837 		rack_translate_policer_detect(rack, optval);
24838 		break;
24839 	case TCP_POLICER_MSS:
24840 		RACK_OPTS_INC(tcp_pol_mss);
24841 		rack->r_ctl.policer_del_mss = (uint8_t)optval;
24842 		if (optval & 0x00000100) {
24843 			/*
24844 			 * Value is setup like so:
24845 			 * VVVV VVVV VVVV VVVV VVVV VVAI MMMM MMMM
24846 			 * Where MMMM MMMM is MSS setting
24847 			 * I (9th bit) is the Postive value that
24848 			 * says it is being set (if its 0 then the
24849 			 * upper bits 11 - 32 have no meaning.
24850 			 * This allows setting it off with
24851 			 * 0x000001MM.
24852 			 *
24853 			 * The 10th bit is used to turn on the
24854 			 * alternate median (not the expanded one).
24855 			 *
24856 			 */
24857 			rack->r_ctl.pol_bw_comp = (optval >> 10);
24858 		}
24859 		if (optval & 0x00000200) {
24860 			rack->r_ctl.policer_alt_median = 1;
24861 		} else {
24862 			rack->r_ctl.policer_alt_median = 0;
24863 		}
24864 		break;
24865  	case TCP_RACK_PACE_TO_FILL:
24866 		RACK_OPTS_INC(tcp_fillcw);
24867 		if (optval == 0)
24868 			rack->rc_pace_to_cwnd = 0;
24869 		else {
24870 			rack->rc_pace_to_cwnd = 1;
24871 		}
24872 		if ((optval >= rack_gp_rtt_maxmul) &&
24873 		    rack_gp_rtt_maxmul &&
24874 		    (optval < 0xf)) {
24875 			rack->rc_pace_fill_if_rttin_range = 1;
24876 			rack->rtt_limit_mul = optval;
24877 		} else {
24878 			rack->rc_pace_fill_if_rttin_range = 0;
24879 			rack->rtt_limit_mul = 0;
24880 		}
24881 		break;
24882 	case TCP_RACK_NO_PUSH_AT_MAX:
24883 		RACK_OPTS_INC(tcp_npush);
24884 		if (optval == 0)
24885 			rack->r_ctl.rc_no_push_at_mrtt = 0;
24886 		else if (optval < 0xff)
24887 			rack->r_ctl.rc_no_push_at_mrtt = optval;
24888 		else
24889 			error = EINVAL;
24890 		break;
24891 	case TCP_SHARED_CWND_ENABLE:
24892 		RACK_OPTS_INC(tcp_rack_scwnd);
24893 		if (optval == 0)
24894 			rack->rack_enable_scwnd = 0;
24895 		else
24896 			rack->rack_enable_scwnd = 1;
24897 		break;
24898 	case TCP_RACK_MBUF_QUEUE:
24899 		/* Now do we use the LRO mbuf-queue feature */
24900 		RACK_OPTS_INC(tcp_rack_mbufq);
24901 		if (optval || rack->r_use_cmp_ack)
24902 			rack->r_mbuf_queue = 1;
24903 		else
24904 			rack->r_mbuf_queue = 0;
24905 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
24906 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
24907 		else
24908 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
24909 		break;
24910 	case TCP_RACK_NONRXT_CFG_RATE:
24911 		RACK_OPTS_INC(tcp_rack_cfg_rate);
24912 		if (optval == 0)
24913 			rack->rack_rec_nonrxt_use_cr = 0;
24914 		else
24915 			rack->rack_rec_nonrxt_use_cr = 1;
24916 		break;
24917 	case TCP_NO_PRR:
24918 		RACK_OPTS_INC(tcp_rack_noprr);
24919 		if (optval == 0)
24920 			rack->rack_no_prr = 0;
24921 		else if (optval == 1)
24922 			rack->rack_no_prr = 1;
24923 		else if (optval == 2)
24924 			rack->no_prr_addback = 1;
24925 		else
24926 			error = EINVAL;
24927 		break;
24928 	case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
24929 		if (optval > 0)
24930 			rack->cspr_is_fcc = 1;
24931 		else
24932 			rack->cspr_is_fcc = 0;
24933 		break;
24934 	case TCP_TIMELY_DYN_ADJ:
24935 		RACK_OPTS_INC(tcp_timely_dyn);
24936 		if (optval == 0)
24937 			rack->rc_gp_dyn_mul = 0;
24938 		else {
24939 			rack->rc_gp_dyn_mul = 1;
24940 			if (optval >= 100) {
24941 				/*
24942 				 * If the user sets something 100 or more
24943 				 * its the gp_ca value.
24944 				 */
24945 				rack->r_ctl.rack_per_of_gp_ca  = optval;
24946 			}
24947 		}
24948 		break;
24949 	case TCP_RACK_DO_DETECTION:
24950 		RACK_OPTS_INC(tcp_rack_do_detection);
24951 		if (optval == 0)
24952 			rack->do_detection = 0;
24953 		else
24954 			rack->do_detection = 1;
24955 		break;
24956 	case TCP_RACK_TLP_USE:
24957 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
24958 			error = EINVAL;
24959 			break;
24960 		}
24961 		RACK_OPTS_INC(tcp_tlp_use);
24962 		rack->rack_tlp_threshold_use = optval;
24963 		break;
24964 	case TCP_RACK_TLP_REDUCE:
24965 		/* RACK TLP cwnd reduction (bool) */
24966 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
24967 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
24968 		break;
24969 		/*  Pacing related ones */
24970 	case TCP_RACK_PACE_ALWAYS:
24971 		/*
24972 		 * zero is old rack method, 1 is new
24973 		 * method using a pacing rate.
24974 		 */
24975 		RACK_OPTS_INC(tcp_rack_pace_always);
24976 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
24977 			error = EPERM;
24978 			break;
24979 		}
24980 		if (optval > 0) {
24981 			if (rack->rc_always_pace) {
24982 				error = EALREADY;
24983 				break;
24984 			} else if (tcp_can_enable_pacing()) {
24985 				rack->r_ctl.pacing_method |= RACK_REG_PACING;
24986 				rack->rc_always_pace = 1;
24987 				if (rack->rack_hibeta)
24988 					rack_set_cc_pacing(rack);
24989 			}
24990 			else {
24991 				error = ENOSPC;
24992 				break;
24993 			}
24994 		} else {
24995 			if (rack->rc_always_pace == 1) {
24996 				rack_remove_pacing(rack);
24997 			}
24998 		}
24999 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
25000 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
25001 		else
25002 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
25003 		/* A rate may be set irate or other, if so set seg size */
25004 		rack_update_seg(rack);
25005 		break;
25006 	case TCP_BBR_RACK_INIT_RATE:
25007 		RACK_OPTS_INC(tcp_initial_rate);
25008 		val = optval;
25009 		/* Change from kbits per second to bytes per second */
25010 		val *= 1000;
25011 		val /= 8;
25012 		rack->r_ctl.init_rate = val;
25013 		if (rack->rc_always_pace)
25014 			rack_update_seg(rack);
25015 		break;
25016 	case TCP_BBR_IWINTSO:
25017 		error = EINVAL;
25018 		break;
25019 	case TCP_RACK_FORCE_MSEG:
25020 		RACK_OPTS_INC(tcp_rack_force_max_seg);
25021 		if (optval)
25022 			rack->rc_force_max_seg = 1;
25023 		else
25024 			rack->rc_force_max_seg = 0;
25025 		break;
25026 	case TCP_RACK_PACE_MIN_SEG:
25027 		RACK_OPTS_INC(tcp_rack_min_seg);
25028 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
25029 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
25030 		break;
25031 	case TCP_RACK_PACE_MAX_SEG:
25032 		/* Max segments size in a pace in bytes */
25033 		RACK_OPTS_INC(tcp_rack_max_seg);
25034 		if ((rack->dgp_on == 1) &&
25035 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
25036 			/*
25037 			 * If we set a max-seg and are doing DGP then
25038 			 * we now fall under the pacing limits not the
25039 			 * DGP ones.
25040 			 */
25041 			if (tcp_can_enable_pacing() == 0) {
25042 				error = ENOSPC;
25043 				break;
25044 			}
25045 			/*
25046 			 * Now change up the flags and counts to be correct.
25047 			 */
25048 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
25049 			tcp_dec_dgp_pacing_cnt();
25050 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
25051 		}
25052 		if (optval <= MAX_USER_SET_SEG)
25053 			rack->rc_user_set_max_segs = optval;
25054 		else
25055 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
25056 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
25057 		break;
25058 	case TCP_RACK_PACE_RATE_REC:
25059 		/* Set the fixed pacing rate in Bytes per second ca */
25060 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
25061 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
25062 			error = EPERM;
25063 			break;
25064 		}
25065 		if (rack->dgp_on) {
25066 			/*
25067 			 * We are already pacing another
25068 			 * way.
25069 			 */
25070 			error = EBUSY;
25071 			break;
25072 		}
25073 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
25074 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
25075 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
25076 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
25077 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
25078 		rack->use_fixed_rate = 1;
25079 		if (rack->rack_hibeta)
25080 			rack_set_cc_pacing(rack);
25081 		rack_log_pacing_delay_calc(rack,
25082 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
25083 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
25084 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
25085 					   __LINE__, NULL,0);
25086 		break;
25087 
25088 	case TCP_RACK_PACE_RATE_SS:
25089 		/* Set the fixed pacing rate in Bytes per second ca */
25090 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
25091 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
25092 			error = EPERM;
25093 			break;
25094 		}
25095 		if (rack->dgp_on) {
25096 			/*
25097 			 * We are already pacing another
25098 			 * way.
25099 			 */
25100 			error = EBUSY;
25101 			break;
25102 		}
25103 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
25104 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
25105 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
25106 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
25107 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
25108 		rack->use_fixed_rate = 1;
25109 		if (rack->rack_hibeta)
25110 			rack_set_cc_pacing(rack);
25111 		rack_log_pacing_delay_calc(rack,
25112 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
25113 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
25114 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
25115 					   __LINE__, NULL, 0);
25116 		break;
25117 
25118 	case TCP_RACK_PACE_RATE_CA:
25119 		/* Set the fixed pacing rate in Bytes per second ca */
25120 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
25121 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
25122 			error = EPERM;
25123 			break;
25124 		}
25125 		if (rack->dgp_on) {
25126 			/*
25127 			 * We are already pacing another
25128 			 * way.
25129 			 */
25130 			error = EBUSY;
25131 			break;
25132 		}
25133 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
25134 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
25135 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
25136 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
25137 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
25138 		rack->use_fixed_rate = 1;
25139 		if (rack->rack_hibeta)
25140 			rack_set_cc_pacing(rack);
25141 		rack_log_pacing_delay_calc(rack,
25142 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
25143 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
25144 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
25145 					   __LINE__, NULL, 0);
25146 		break;
25147 	case TCP_RACK_GP_INCREASE_REC:
25148 		RACK_OPTS_INC(tcp_gp_inc_rec);
25149 		rack->r_ctl.rack_per_of_gp_rec = optval;
25150 		rack_log_pacing_delay_calc(rack,
25151 					   rack->r_ctl.rack_per_of_gp_ss,
25152 					   rack->r_ctl.rack_per_of_gp_ca,
25153 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
25154 					   __LINE__, NULL, 0);
25155 		break;
25156 	case TCP_RACK_GP_INCREASE_CA:
25157 		RACK_OPTS_INC(tcp_gp_inc_ca);
25158 		ca = optval;
25159 		if (ca < 100) {
25160 			/*
25161 			 * We don't allow any reduction
25162 			 * over the GP b/w.
25163 			 */
25164 			error = EINVAL;
25165 			break;
25166 		}
25167 		rack->r_ctl.rack_per_of_gp_ca = ca;
25168 		rack_log_pacing_delay_calc(rack,
25169 					   rack->r_ctl.rack_per_of_gp_ss,
25170 					   rack->r_ctl.rack_per_of_gp_ca,
25171 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
25172 					   __LINE__, NULL, 0);
25173 		break;
25174 	case TCP_RACK_GP_INCREASE_SS:
25175 		RACK_OPTS_INC(tcp_gp_inc_ss);
25176 		ss = optval;
25177 		if (ss < 100) {
25178 			/*
25179 			 * We don't allow any reduction
25180 			 * over the GP b/w.
25181 			 */
25182 			error = EINVAL;
25183 			break;
25184 		}
25185 		rack->r_ctl.rack_per_of_gp_ss = ss;
25186 		rack_log_pacing_delay_calc(rack,
25187 					   rack->r_ctl.rack_per_of_gp_ss,
25188 					   rack->r_ctl.rack_per_of_gp_ca,
25189 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
25190 					   __LINE__, NULL, 0);
25191 		break;
25192 	case TCP_RACK_RR_CONF:
25193 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
25194 		if (optval && optval <= 3)
25195 			rack->r_rr_config = optval;
25196 		else
25197 			rack->r_rr_config = 0;
25198 		break;
25199 	case TCP_PACING_DND:			/*  URL:dnd */
25200 		if (optval > 0)
25201 			rack->rc_pace_dnd = 1;
25202 		else
25203 			rack->rc_pace_dnd = 0;
25204 		break;
25205 	case TCP_HDWR_RATE_CAP:
25206 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
25207 		if (optval) {
25208 			if (rack->r_rack_hw_rate_caps == 0)
25209 				rack->r_rack_hw_rate_caps = 1;
25210 			else
25211 				error = EALREADY;
25212 		} else {
25213 			rack->r_rack_hw_rate_caps = 0;
25214 		}
25215 		break;
25216 	case TCP_DGP_UPPER_BOUNDS:
25217 	{
25218 		uint8_t val;
25219 		val = optval & 0x0000ff;
25220 		rack->r_ctl.rack_per_upper_bound_ca = val;
25221 		val = (optval >> 16) & 0x0000ff;
25222 		rack->r_ctl.rack_per_upper_bound_ss = val;
25223 		break;
25224 	}
25225 	case TCP_SS_EEXIT:			/*  URL:eexit */
25226 		if (optval > 0) {
25227 			rack->r_ctl.gp_rnd_thresh =  optval & 0x0ff;
25228 			if (optval & 0x10000) {
25229 				rack->r_ctl.gate_to_fs = 1;
25230 			} else {
25231 				rack->r_ctl.gate_to_fs = 0;
25232 			}
25233 			if (optval & 0x20000) {
25234 				rack->r_ctl.use_gp_not_last = 1;
25235 			} else {
25236 				rack->r_ctl.use_gp_not_last = 0;
25237 			}
25238 			if (optval & 0xfffc0000) {
25239 				uint32_t v;
25240 
25241 				v = (optval >> 18) & 0x00003fff;
25242 				if (v >= 1000)
25243 					rack->r_ctl.gp_gain_req = v;
25244 			}
25245 		} else {
25246 			/* We do not do ss early exit at all */
25247 			rack->rc_initial_ss_comp = 1;
25248 			rack->r_ctl.gp_rnd_thresh = 0;
25249 		}
25250 		break;
25251 	case TCP_RACK_SPLIT_LIMIT:
25252 		RACK_OPTS_INC(tcp_split_limit);
25253 		rack->r_ctl.rc_split_limit = optval;
25254 		break;
25255 	case TCP_BBR_HDWR_PACE:
25256 		RACK_OPTS_INC(tcp_hdwr_pacing);
25257 		if (optval){
25258 			if (rack->rack_hdrw_pacing == 0) {
25259 				rack->rack_hdw_pace_ena = 1;
25260 				rack->rack_attempt_hdwr_pace = 0;
25261 			} else
25262 				error = EALREADY;
25263 		} else {
25264 			rack->rack_hdw_pace_ena = 0;
25265 #ifdef RATELIMIT
25266 			if (rack->r_ctl.crte != NULL) {
25267 				rack->rack_hdrw_pacing = 0;
25268 				rack->rack_attempt_hdwr_pace = 0;
25269 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
25270 				rack->r_ctl.crte = NULL;
25271 			}
25272 #endif
25273 		}
25274 		break;
25275 		/*  End Pacing related ones */
25276 	case TCP_RACK_PRR_SENDALOT:
25277 		/* Allow PRR to send more than one seg */
25278 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
25279 		rack->r_ctl.rc_prr_sendalot = optval;
25280 		break;
25281 	case TCP_RACK_MIN_TO:
25282 		/* Minimum time between rack t-o's in ms */
25283 		RACK_OPTS_INC(tcp_rack_min_to);
25284 		rack->r_ctl.rc_min_to = optval;
25285 		break;
25286 	case TCP_RACK_EARLY_SEG:
25287 		/* If early recovery max segments */
25288 		RACK_OPTS_INC(tcp_rack_early_seg);
25289 		rack->r_ctl.rc_early_recovery_segs = optval;
25290 		break;
25291 	case TCP_RACK_ENABLE_HYSTART:
25292 	{
25293 		if (optval) {
25294 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
25295 			if (rack_do_hystart > RACK_HYSTART_ON)
25296 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
25297 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
25298 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
25299 		} else {
25300 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
25301 		}
25302 	}
25303 	break;
25304 	case TCP_RACK_REORD_THRESH:
25305 		/* RACK reorder threshold (shift amount) */
25306 		RACK_OPTS_INC(tcp_rack_reord_thresh);
25307 		if ((optval > 0) && (optval < 31))
25308 			rack->r_ctl.rc_reorder_shift = optval;
25309 		else
25310 			error = EINVAL;
25311 		break;
25312 	case TCP_RACK_REORD_FADE:
25313 		/* Does reordering fade after ms time */
25314 		RACK_OPTS_INC(tcp_rack_reord_fade);
25315 		rack->r_ctl.rc_reorder_fade = optval;
25316 		break;
25317 	case TCP_RACK_TLP_THRESH:
25318 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
25319 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
25320 		if (optval)
25321 			rack->r_ctl.rc_tlp_threshold = optval;
25322 		else
25323 			error = EINVAL;
25324 		break;
25325 	case TCP_BBR_USE_RACK_RR:
25326 		RACK_OPTS_INC(tcp_rack_rr);
25327 		if (optval)
25328 			rack->use_rack_rr = 1;
25329 		else
25330 			rack->use_rack_rr = 0;
25331 		break;
25332 	case TCP_RACK_PKT_DELAY:
25333 		/* RACK added ms i.e. rack-rtt + reord + N */
25334 		RACK_OPTS_INC(tcp_rack_pkt_delay);
25335 		rack->r_ctl.rc_pkt_delay = optval;
25336 		break;
25337 	case TCP_DELACK:
25338 		RACK_OPTS_INC(tcp_rack_delayed_ack);
25339 		if (optval == 0)
25340 			tp->t_delayed_ack = 0;
25341 		else
25342 			tp->t_delayed_ack = 1;
25343 		if (tp->t_flags & TF_DELACK) {
25344 			tp->t_flags &= ~TF_DELACK;
25345 			tp->t_flags |= TF_ACKNOW;
25346 			NET_EPOCH_ENTER(et);
25347 			rack_output(tp);
25348 			NET_EPOCH_EXIT(et);
25349 		}
25350 		break;
25351 
25352 	case TCP_BBR_RACK_RTT_USE:
25353 		RACK_OPTS_INC(tcp_rack_rtt_use);
25354 		if ((optval != USE_RTT_HIGH) &&
25355 		    (optval != USE_RTT_LOW) &&
25356 		    (optval != USE_RTT_AVG))
25357 			error = EINVAL;
25358 		else
25359 			rack->r_ctl.rc_rate_sample_method = optval;
25360 		break;
25361 	case TCP_HONOR_HPTS_MIN:
25362 		RACK_OPTS_INC(tcp_honor_hpts);
25363 		if (optval) {
25364 			rack->r_use_hpts_min = 1;
25365 			/*
25366 			 * Must be between 2 - 80% to be a reduction else
25367 			 * we keep the default (10%).
25368 			 */
25369 			if ((optval > 1) && (optval <= 80)) {
25370 				rack->r_ctl.max_reduction = optval;
25371 			}
25372 		} else
25373 			rack->r_use_hpts_min = 0;
25374 		break;
25375 	case TCP_REC_IS_DYN:			/*  URL:dynrec */
25376 		RACK_OPTS_INC(tcp_dyn_rec);
25377 		if (optval)
25378 			rack->rc_gp_no_rec_chg = 1;
25379 		else
25380 			rack->rc_gp_no_rec_chg = 0;
25381 		break;
25382 	case TCP_NO_TIMELY:
25383 		RACK_OPTS_INC(tcp_notimely);
25384 		if (optval) {
25385 			rack->rc_skip_timely = 1;
25386 			rack->r_ctl.rack_per_of_gp_rec = 90;
25387 			rack->r_ctl.rack_per_of_gp_ca = 100;
25388 			rack->r_ctl.rack_per_of_gp_ss = 250;
25389 		} else {
25390 			rack->rc_skip_timely = 0;
25391 		}
25392 		break;
25393 	case TCP_GP_USE_LTBW:
25394 		if (optval == 0) {
25395 			rack->use_lesser_lt_bw = 0;
25396 			rack->dis_lt_bw = 1;
25397 		} else if (optval == 1) {
25398 			rack->use_lesser_lt_bw = 1;
25399 			rack->dis_lt_bw = 0;
25400 		} else if (optval == 2) {
25401 			rack->use_lesser_lt_bw = 0;
25402 			rack->dis_lt_bw = 0;
25403 		}
25404 		break;
25405 	case TCP_DATA_AFTER_CLOSE:
25406 		RACK_OPTS_INC(tcp_data_after_close);
25407 		if (optval)
25408 			rack->rc_allow_data_af_clo = 1;
25409 		else
25410 			rack->rc_allow_data_af_clo = 0;
25411 		break;
25412 	default:
25413 		break;
25414 	}
25415 	tcp_log_socket_option(tp, sopt_name, optval, error);
25416 	return (error);
25417 }
25418 
25419 static void
25420 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
25421 {
25422 	/*
25423 	 * A new connection has been created (tp) and
25424 	 * the parent is the inpcb given. We want to
25425 	 * apply a read-lock to the parent (we are already
25426 	 * holding a write lock on the tp) and copy anything
25427 	 * out of the rack specific data as long as its tfb is
25428 	 * the same as ours i.e. we are the same stack. Otherwise
25429 	 * we just return.
25430 	 */
25431 	struct tcpcb *par;
25432 	struct tcp_rack *dest, *src;
25433 	int cnt = 0;
25434 
25435 	par = intotcpcb(parent);
25436 	if (par->t_fb != tp->t_fb) {
25437 		/* Not the same stack */
25438 		tcp_log_socket_option(tp, 0, 0, 1);
25439 		return;
25440 	}
25441 	/* Ok if we reach here lets setup the two rack pointers */
25442 	dest = (struct tcp_rack *)tp->t_fb_ptr;
25443 	src = (struct tcp_rack *)par->t_fb_ptr;
25444 	if ((src == NULL) || (dest == NULL)) {
25445 		/* Huh? */
25446 		tcp_log_socket_option(tp, 0, 0, 2);
25447 		return;
25448 	}
25449 	/* Now copy out anything we wish to inherit i.e. things in socket-options */
25450 	/* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
25451 	if ((src->dgp_on) && (dest->dgp_on == 0)) {
25452 		/* Profile 1 had to be set via sock opt */
25453 		rack_set_dgp(dest);
25454 		cnt++;
25455 	}
25456 	/* TCP_RACK_SET_RXT_OPTIONS */
25457 	if (dest->full_size_rxt != src->full_size_rxt) {
25458 		dest->full_size_rxt = src->full_size_rxt;
25459 		cnt++;
25460 	}
25461 	if (dest->shape_rxt_to_pacing_min  != src->shape_rxt_to_pacing_min) {
25462 		dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
25463 		cnt++;
25464 	}
25465 	/* TCP_RACK_DSACK_OPT */
25466 	if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
25467 		dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
25468 		cnt++;
25469 	}
25470 	if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
25471 		dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
25472 		cnt++;
25473 	}
25474 	/* TCP_RACK_PACING_DIVISOR */
25475 	if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
25476 		dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
25477 		cnt++;
25478 	}
25479 	/* TCP_RACK_HI_BETA */
25480 	if (src->rack_hibeta != dest->rack_hibeta) {
25481 		cnt++;
25482 		if (src->rack_hibeta) {
25483 			dest->r_ctl.rc_saved_beta.beta = src->r_ctl.rc_saved_beta.beta;
25484 			dest->rack_hibeta = 1;
25485 		} else {
25486 			dest->rack_hibeta = 0;
25487 		}
25488 	}
25489 	/* TCP_RACK_TIMER_SLOP */
25490 	if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
25491 		dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
25492 		cnt++;
25493 	}
25494 	/* TCP_RACK_PACING_BETA_ECN */
25495 	if (dest->r_ctl.rc_saved_beta.beta_ecn != src->r_ctl.rc_saved_beta.beta_ecn) {
25496 		dest->r_ctl.rc_saved_beta.beta_ecn = src->r_ctl.rc_saved_beta.beta_ecn;
25497 		cnt++;
25498 	}
25499 	if (dest->r_ctl.rc_saved_beta.newreno_flags != src->r_ctl.rc_saved_beta.newreno_flags) {
25500 		dest->r_ctl.rc_saved_beta.newreno_flags = src->r_ctl.rc_saved_beta.newreno_flags;
25501 		cnt++;
25502 	}
25503 	/* We do not do TCP_DEFER_OPTIONS */
25504 	/* TCP_RACK_MEASURE_CNT */
25505 	if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
25506 		dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
25507 		cnt++;
25508 	}
25509 	/* TCP_HDWR_UP_ONLY */
25510 	if (dest->r_up_only != src->r_up_only) {
25511 		dest->r_up_only = src->r_up_only;
25512 		cnt++;
25513 	}
25514 	/* TCP_FILLCW_RATE_CAP */
25515 	if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
25516 		dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
25517 		cnt++;
25518 	}
25519 	/* TCP_PACING_RATE_CAP */
25520 	if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
25521 		dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
25522 		cnt++;
25523 	}
25524 	/* A listener can't set TCP_HYBRID_PACING */
25525 	/* TCP_SIDECHAN_DIS */
25526 	if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
25527 		dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
25528 		cnt++;
25529 	}
25530 	/* TCP_SHARED_CWND_TIME_LIMIT */
25531 	if (dest->r_limit_scw != src->r_limit_scw) {
25532 		dest->r_limit_scw = src->r_limit_scw;
25533 		cnt++;
25534 	}
25535 	/* TCP_POLICER_DETECT */
25536 	if (dest->r_ctl.policer_rxt_threshold != src->r_ctl.policer_rxt_threshold) {
25537 		dest->r_ctl.policer_rxt_threshold = src->r_ctl.policer_rxt_threshold;
25538 		cnt++;
25539 	}
25540 	if (dest->r_ctl.policer_avg_threshold != src->r_ctl.policer_avg_threshold) {
25541 		dest->r_ctl.policer_avg_threshold = src->r_ctl.policer_avg_threshold;
25542 		cnt++;
25543 	}
25544 	if (dest->r_ctl.policer_med_threshold != src->r_ctl.policer_med_threshold) {
25545 		dest->r_ctl.policer_med_threshold = src->r_ctl.policer_med_threshold;
25546 		cnt++;
25547 	}
25548 	if (dest->policer_detect_on != src->policer_detect_on) {
25549 		dest->policer_detect_on = src->policer_detect_on;
25550 		cnt++;
25551 	}
25552 
25553 	if (dest->r_ctl.saved_policer_val != src->r_ctl.saved_policer_val) {
25554 		dest->r_ctl.saved_policer_val = src->r_ctl.saved_policer_val;
25555 		cnt++;
25556 	}
25557 	/* TCP_POLICER_MSS */
25558 	if (dest->r_ctl.policer_del_mss != src->r_ctl.policer_del_mss) {
25559 		dest->r_ctl.policer_del_mss = src->r_ctl.policer_del_mss;
25560 		cnt++;
25561 	}
25562 
25563 	if (dest->r_ctl.pol_bw_comp != src->r_ctl.pol_bw_comp) {
25564 		dest->r_ctl.pol_bw_comp = src->r_ctl.pol_bw_comp;
25565 		cnt++;
25566 	}
25567 
25568 	if (dest->r_ctl.policer_alt_median != src->r_ctl.policer_alt_median) {
25569 		dest->r_ctl.policer_alt_median = src->r_ctl.policer_alt_median;
25570 		cnt++;
25571 	}
25572 	/* TCP_RACK_PACE_TO_FILL */
25573 	if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
25574 		dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
25575 		cnt++;
25576 	}
25577 	if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
25578 		dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
25579 		cnt++;
25580 	}
25581 	if (dest->rtt_limit_mul != src->rtt_limit_mul) {
25582 		dest->rtt_limit_mul = src->rtt_limit_mul;
25583 		cnt++;
25584 	}
25585 	/* TCP_RACK_NO_PUSH_AT_MAX */
25586 	if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
25587 		dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
25588 		cnt++;
25589 	}
25590 	/* TCP_SHARED_CWND_ENABLE */
25591 	if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
25592 		dest->rack_enable_scwnd = src->rack_enable_scwnd;
25593 		cnt++;
25594 	}
25595 	/* TCP_USE_CMP_ACKS */
25596 	if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
25597 		dest->r_use_cmp_ack = src->r_use_cmp_ack;
25598 		cnt++;
25599 	}
25600 
25601 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
25602 		dest->r_mbuf_queue = src->r_mbuf_queue;
25603 		cnt++;
25604 	}
25605 	/* TCP_RACK_MBUF_QUEUE */
25606 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
25607 		dest->r_mbuf_queue = src->r_mbuf_queue;
25608 		cnt++;
25609 	}
25610 	if  (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
25611 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
25612 	} else {
25613 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
25614 	}
25615 	if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
25616 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
25617 	}
25618 	/* TCP_RACK_NONRXT_CFG_RATE */
25619 	if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
25620 		dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
25621 		cnt++;
25622 	}
25623 	/* TCP_NO_PRR */
25624 	if (dest->rack_no_prr != src->rack_no_prr) {
25625 		dest->rack_no_prr = src->rack_no_prr;
25626 		cnt++;
25627 	}
25628 	if (dest->no_prr_addback != src->no_prr_addback) {
25629 		dest->no_prr_addback = src->no_prr_addback;
25630 		cnt++;
25631 	}
25632 	/* RACK_CSPR_IS_FCC */
25633 	if (dest->cspr_is_fcc != src->cspr_is_fcc) {
25634 		dest->cspr_is_fcc = src->cspr_is_fcc;
25635 		cnt++;
25636 	}
25637 	/* TCP_TIMELY_DYN_ADJ */
25638 	if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
25639 		dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
25640 		cnt++;
25641 	}
25642 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
25643 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
25644 		cnt++;
25645 	}
25646 	/* TCP_RACK_DO_DETECTION */
25647 	if (dest->do_detection != src->do_detection) {
25648 		dest->do_detection = src->do_detection;
25649 		cnt++;
25650 	}
25651 	/* TCP_RACK_TLP_USE */
25652 	if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
25653 		dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
25654 		cnt++;
25655 	}
25656 	/* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
25657 	/* TCP_BBR_RACK_INIT_RATE */
25658 	if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
25659 		dest->r_ctl.init_rate = src->r_ctl.init_rate;
25660 		cnt++;
25661 	}
25662 	/* TCP_RACK_FORCE_MSEG */
25663 	if (dest->rc_force_max_seg != src->rc_force_max_seg) {
25664 		dest->rc_force_max_seg = src->rc_force_max_seg;
25665 		cnt++;
25666 	}
25667 	/* TCP_RACK_PACE_MIN_SEG */
25668 	if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
25669 		dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
25670 		cnt++;
25671 	}
25672 	/* we don't allow TCP_RACK_PACE_MAX_SEG */
25673 	/* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS,  TCP_RACK_PACE_RATE_CA */
25674 	if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
25675 		dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
25676 		cnt++;
25677 	}
25678 	if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
25679 		dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
25680 		cnt++;
25681 	}
25682 	if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
25683 		dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
25684 		cnt++;
25685 	}
25686 	/* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
25687 	if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
25688 		dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
25689 		cnt++;
25690 	}
25691 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
25692 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
25693 		cnt++;
25694 	}
25695 
25696 	if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
25697 		dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
25698 		cnt++;
25699 	}
25700 	/* TCP_RACK_RR_CONF */
25701 	if (dest->r_rr_config != src->r_rr_config) {
25702 		dest->r_rr_config = src->r_rr_config;
25703 		cnt++;
25704 	}
25705 	/* TCP_PACING_DND */
25706 	if (dest->rc_pace_dnd != src->rc_pace_dnd) {
25707 		dest->rc_pace_dnd = src->rc_pace_dnd;
25708 		cnt++;
25709 	}
25710 	/* TCP_HDWR_RATE_CAP */
25711 	if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
25712 		dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
25713 		cnt++;
25714 	}
25715 	/* TCP_DGP_UPPER_BOUNDS */
25716 	if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
25717 		dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
25718 		cnt++;
25719 	}
25720 	if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
25721 		dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
25722 		cnt++;
25723 	}
25724 	/* TCP_SS_EEXIT */
25725 	if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
25726 		dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
25727 		cnt++;
25728 	}
25729 	if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
25730 		dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
25731 		cnt++;
25732 	}
25733 	if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
25734 		dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
25735 		cnt++;
25736 	}
25737 	if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
25738 		dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
25739 		cnt++;
25740 	}
25741 	/* TCP_BBR_HDWR_PACE */
25742 	if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
25743 		dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
25744 		cnt++;
25745 	}
25746 	if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
25747 		dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
25748 		cnt++;
25749 	}
25750 	/* TCP_RACK_PRR_SENDALOT */
25751 	if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
25752 		dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
25753 		cnt++;
25754 	}
25755 	/* TCP_RACK_MIN_TO */
25756 	if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
25757 		dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
25758 		cnt++;
25759 	}
25760 	/* TCP_RACK_EARLY_SEG */
25761 	if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
25762 		dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
25763 		cnt++;
25764 	}
25765 	/* TCP_RACK_ENABLE_HYSTART */
25766 	if (par->t_ccv.flags != tp->t_ccv.flags) {
25767 		cnt++;
25768 		if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
25769 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
25770 			if (rack_do_hystart > RACK_HYSTART_ON)
25771 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
25772 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
25773 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
25774 		} else {
25775 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
25776 		}
25777 	}
25778 	/* TCP_RACK_REORD_THRESH */
25779 	if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
25780 		dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
25781 		cnt++;
25782 	}
25783 	/* TCP_RACK_REORD_FADE */
25784 	if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
25785 		dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
25786 		cnt++;
25787 	}
25788 	/* TCP_RACK_TLP_THRESH */
25789 	if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
25790 		dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
25791 		cnt++;
25792 	}
25793 	/* TCP_BBR_USE_RACK_RR */
25794 	if (dest->use_rack_rr != src->use_rack_rr) {
25795 		dest->use_rack_rr = src->use_rack_rr;
25796 		cnt++;
25797 	}
25798 	/* TCP_RACK_PKT_DELAY */
25799 	if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
25800 		dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
25801 		cnt++;
25802 	}
25803 	/* TCP_DELACK will get copied via the main code if applicable */
25804 	/* TCP_BBR_RACK_RTT_USE */
25805 	if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
25806 		dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
25807 		cnt++;
25808 	}
25809 	/* TCP_HONOR_HPTS_MIN */
25810 	if (dest->r_use_hpts_min != src->r_use_hpts_min) {
25811 		dest->r_use_hpts_min = src->r_use_hpts_min;
25812 		cnt++;
25813 	}
25814 	if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
25815 		dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
25816 		cnt++;
25817 	}
25818 	/* TCP_REC_IS_DYN */
25819 	if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
25820 		dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
25821 		cnt++;
25822 	}
25823 	if (dest->rc_skip_timely != src->rc_skip_timely) {
25824 		dest->rc_skip_timely = src->rc_skip_timely;
25825 		cnt++;
25826 	}
25827 	/* TCP_DATA_AFTER_CLOSE */
25828 	if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
25829 		dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
25830 		cnt++;
25831 	}
25832 	/* TCP_GP_USE_LTBW */
25833 	if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
25834 		dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
25835 		cnt++;
25836 	}
25837 	if (dest->dis_lt_bw != src->dis_lt_bw) {
25838 		dest->dis_lt_bw = src->dis_lt_bw;
25839 		cnt++;
25840 	}
25841 	tcp_log_socket_option(tp, 0, cnt, 0);
25842 }
25843 
25844 
25845 static void
25846 rack_apply_deferred_options(struct tcp_rack *rack)
25847 {
25848 	struct deferred_opt_list *dol, *sdol;
25849 	uint32_t s_optval;
25850 
25851 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
25852 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
25853 		/* Disadvantage of deferal is you loose the error return */
25854 		s_optval = (uint32_t)dol->optval;
25855 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
25856 		free(dol, M_TCPDO);
25857 	}
25858 }
25859 
25860 static void
25861 rack_hw_tls_change(struct tcpcb *tp, int chg)
25862 {
25863 	/* Update HW tls state */
25864 	struct tcp_rack *rack;
25865 
25866 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25867 	if (chg)
25868 		rack->r_ctl.fsb.hw_tls = 1;
25869 	else
25870 		rack->r_ctl.fsb.hw_tls = 0;
25871 }
25872 
25873 static int
25874 rack_pru_options(struct tcpcb *tp, int flags)
25875 {
25876 	if (flags & PRUS_OOB)
25877 		return (EOPNOTSUPP);
25878 	return (0);
25879 }
25880 
25881 static bool
25882 rack_wake_check(struct tcpcb *tp)
25883 {
25884 	struct tcp_rack *rack;
25885 	struct timeval tv;
25886 	uint32_t cts;
25887 
25888 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25889 	if (rack->r_ctl.rc_hpts_flags) {
25890 		cts = tcp_get_usecs(&tv);
25891 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
25892 			/*
25893 			 * Pacing timer is up, check if we are ready.
25894 			 */
25895 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
25896 				return (true);
25897 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
25898 			/*
25899 			 * A timer is up, check if we are ready.
25900 			 */
25901 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
25902 				return (true);
25903 		}
25904 	}
25905 	return (false);
25906 }
25907 
25908 static struct tcp_function_block __tcp_rack = {
25909 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
25910 	.tfb_tcp_output = rack_output,
25911 	.tfb_do_queued_segments = ctf_do_queued_segments,
25912 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
25913 	.tfb_tcp_do_segment = rack_do_segment,
25914 	.tfb_tcp_ctloutput = rack_ctloutput,
25915 	.tfb_tcp_fb_init = rack_init,
25916 	.tfb_tcp_fb_fini = rack_fini,
25917 	.tfb_tcp_timer_stop_all = rack_stopall,
25918 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
25919 	.tfb_tcp_handoff_ok = rack_handoff_ok,
25920 	.tfb_tcp_mtu_chg = rack_mtu_change,
25921 	.tfb_pru_options = rack_pru_options,
25922 	.tfb_hwtls_change = rack_hw_tls_change,
25923 	.tfb_chg_query = rack_chg_query,
25924 	.tfb_switch_failed = rack_switch_failed,
25925 	.tfb_early_wake_check = rack_wake_check,
25926 	.tfb_compute_pipe = rack_compute_pipe,
25927 	.tfb_stack_info = rack_stack_information,
25928 	.tfb_inherit = rack_inherit,
25929 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
25930 
25931 };
25932 
25933 /*
25934  * rack_ctloutput() must drop the inpcb lock before performing copyin on
25935  * socket option arguments.  When it re-acquires the lock after the copy, it
25936  * has to revalidate that the connection is still valid for the socket
25937  * option.
25938  */
25939 static int
25940 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
25941 {
25942 	struct inpcb *inp = tptoinpcb(tp);
25943 #ifdef INET
25944 	struct ip *ip;
25945 #endif
25946 	struct tcp_rack *rack;
25947 	struct tcp_hybrid_req hybrid;
25948 	uint64_t loptval;
25949 	int32_t error = 0, optval;
25950 
25951 	rack = (struct tcp_rack *)tp->t_fb_ptr;
25952 	if (rack == NULL) {
25953 		INP_WUNLOCK(inp);
25954 		return (EINVAL);
25955 	}
25956 #ifdef INET
25957 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
25958 #endif
25959 
25960 	switch (sopt->sopt_level) {
25961 #ifdef INET6
25962 	case IPPROTO_IPV6:
25963 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
25964 		switch (sopt->sopt_name) {
25965 		case IPV6_USE_MIN_MTU:
25966 			tcp6_use_min_mtu(tp);
25967 			break;
25968 		}
25969 		INP_WUNLOCK(inp);
25970 		return (0);
25971 #endif
25972 #ifdef INET
25973 	case IPPROTO_IP:
25974 		switch (sopt->sopt_name) {
25975 		case IP_TOS:
25976 			/*
25977 			 * The DSCP codepoint has changed, update the fsb.
25978 			 */
25979 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
25980 			break;
25981 		case IP_TTL:
25982 			/*
25983 			 * The TTL has changed, update the fsb.
25984 			 */
25985 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
25986 			break;
25987 		}
25988 		INP_WUNLOCK(inp);
25989 		return (0);
25990 #endif
25991 #ifdef SO_PEERPRIO
25992 	case SOL_SOCKET:
25993 		switch (sopt->sopt_name) {
25994 		case SO_PEERPRIO:			/*  SC-URL:bs */
25995 			/* Already read in and sanity checked in sosetopt(). */
25996 			if (inp->inp_socket) {
25997 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
25998 			}
25999 			break;
26000 		}
26001 		INP_WUNLOCK(inp);
26002 		return (0);
26003 #endif
26004 	case IPPROTO_TCP:
26005 		switch (sopt->sopt_name) {
26006 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
26007 		/*  Pacing related ones */
26008 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
26009 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
26010 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
26011 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
26012 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
26013 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
26014 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
26015 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
26016 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
26017 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
26018 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
26019 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
26020 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
26021 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
26022 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
26023 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
26024 		case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
26025 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
26026 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
26027 			/* End pacing related */
26028 		case TCP_POLICER_DETECT:		/*  URL:pol_det */
26029 		case TCP_POLICER_MSS:			/*  URL:pol_mss */
26030 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
26031 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
26032 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
26033 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
26034 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
26035 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
26036 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
26037 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
26038 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
26039 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
26040 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
26041 		case TCP_RACK_DO_DETECTION:		/*  URL:detect */
26042 		case TCP_NO_PRR:			/*  URL:noprr */
26043 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
26044 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
26045 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
26046 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
26047 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
26048 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
26049 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
26050 		case TCP_RACK_PROFILE:			/*  URL:profile */
26051 		case TCP_SIDECHAN_DIS:			/*  URL:scodm */
26052 		case TCP_HYBRID_PACING:			/*  URL:pacing=hybrid */
26053 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
26054 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
26055 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
26056 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
26057 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
26058 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
26059 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
26060 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
26061 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
26062 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
26063 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
26064 		case TCP_SS_EEXIT:			/*  URL:eexit */
26065 		case TCP_DGP_UPPER_BOUNDS:		/*  URL:upper */
26066 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
26067 		case TCP_PACING_DND:			/*  URL:dnd */
26068 		case TCP_NO_TIMELY:			/*  URL:notimely */
26069 		case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
26070 		case TCP_HONOR_HPTS_MIN:		/*  URL:hptsmin */
26071 		case TCP_REC_IS_DYN:			/*  URL:dynrec */
26072 		case TCP_GP_USE_LTBW:			/*  URL:useltbw */
26073 			goto process_opt;
26074 			break;
26075 		default:
26076 			/* Filter off all unknown options to the base stack */
26077 			return (tcp_default_ctloutput(tp, sopt));
26078 			break;
26079 		}
26080 	default:
26081 		INP_WUNLOCK(inp);
26082 		return (0);
26083 	}
26084 process_opt:
26085 	INP_WUNLOCK(inp);
26086 	if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
26087 	    (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
26088 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
26089 		/*
26090 		 * We truncate it down to 32 bits for the socket-option trace this
26091 		 * means rates > 34Gbps won't show right, but thats probably ok.
26092 		 */
26093 		optval = (uint32_t)loptval;
26094 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
26095 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
26096 	} else {
26097 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
26098 		/* Save it in 64 bit form too */
26099 		loptval = optval;
26100 	}
26101 	if (error)
26102 		return (error);
26103 	INP_WLOCK(inp);
26104 	if (tp->t_fb != &__tcp_rack) {
26105 		INP_WUNLOCK(inp);
26106 		return (ENOPROTOOPT);
26107 	}
26108 	if (rack->defer_options && (rack->gp_ready == 0) &&
26109 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
26110 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
26111 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
26112 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
26113 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
26114 		/* Options are being deferred */
26115 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
26116 			INP_WUNLOCK(inp);
26117 			return (0);
26118 		} else {
26119 			/* No memory to defer, fail */
26120 			INP_WUNLOCK(inp);
26121 			return (ENOMEM);
26122 		}
26123 	}
26124 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
26125 	INP_WUNLOCK(inp);
26126 	return (error);
26127 }
26128 
26129 static void
26130 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
26131 {
26132 
26133 	INP_WLOCK_ASSERT(tptoinpcb(tp));
26134 	bzero(ti, sizeof(*ti));
26135 
26136 	ti->tcpi_state = tp->t_state;
26137 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
26138 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
26139 	if (tp->t_flags & TF_SACK_PERMIT)
26140 		ti->tcpi_options |= TCPI_OPT_SACK;
26141 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
26142 		ti->tcpi_options |= TCPI_OPT_WSCALE;
26143 		ti->tcpi_snd_wscale = tp->snd_scale;
26144 		ti->tcpi_rcv_wscale = tp->rcv_scale;
26145 	}
26146 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
26147 		ti->tcpi_options |= TCPI_OPT_ECN;
26148 	if (tp->t_flags & TF_FASTOPEN)
26149 		ti->tcpi_options |= TCPI_OPT_TFO;
26150 	/* still kept in ticks is t_rcvtime */
26151 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
26152 	/* Since we hold everything in precise useconds this is easy */
26153 	ti->tcpi_rtt = tp->t_srtt;
26154 	ti->tcpi_rttvar = tp->t_rttvar;
26155 	ti->tcpi_rto = tp->t_rxtcur;
26156 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
26157 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
26158 	/*
26159 	 * FreeBSD-specific extension fields for tcp_info.
26160 	 */
26161 	ti->tcpi_rcv_space = tp->rcv_wnd;
26162 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
26163 	ti->tcpi_snd_wnd = tp->snd_wnd;
26164 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
26165 	ti->tcpi_snd_nxt = tp->snd_nxt;
26166 	ti->tcpi_snd_mss = tp->t_maxseg;
26167 	ti->tcpi_rcv_mss = tp->t_maxseg;
26168 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
26169 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
26170 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
26171 	ti->tcpi_total_tlp = tp->t_sndtlppack;
26172 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
26173 	ti->tcpi_rttmin = tp->t_rttlow;
26174 #ifdef NETFLIX_STATS
26175 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
26176 #endif
26177 #ifdef TCP_OFFLOAD
26178 	if (tp->t_flags & TF_TOE) {
26179 		ti->tcpi_options |= TCPI_OPT_TOE;
26180 		tcp_offload_tcp_info(tp, ti);
26181 	}
26182 #endif
26183 }
26184 
26185 static int
26186 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
26187 {
26188 	struct inpcb *inp = tptoinpcb(tp);
26189 	struct tcp_rack *rack;
26190 	int32_t error, optval;
26191 	uint64_t val, loptval;
26192 	struct	tcp_info ti;
26193 	/*
26194 	 * Because all our options are either boolean or an int, we can just
26195 	 * pull everything into optval and then unlock and copy. If we ever
26196 	 * add a option that is not a int, then this will have quite an
26197 	 * impact to this routine.
26198 	 */
26199 	error = 0;
26200 	rack = (struct tcp_rack *)tp->t_fb_ptr;
26201 	if (rack == NULL) {
26202 		INP_WUNLOCK(inp);
26203 		return (EINVAL);
26204 	}
26205 	switch (sopt->sopt_name) {
26206 	case TCP_INFO:
26207 		/* First get the info filled */
26208 		rack_fill_info(tp, &ti);
26209 		/* Fix up the rtt related fields if needed */
26210 		INP_WUNLOCK(inp);
26211 		error = sooptcopyout(sopt, &ti, sizeof ti);
26212 		return (error);
26213 	/*
26214 	 * Beta is the congestion control value for NewReno that influences how
26215 	 * much of a backoff happens when loss is detected. It is normally set
26216 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
26217 	 * when you exit recovery.
26218 	 */
26219 	case TCP_RACK_PACING_BETA:
26220 		break;
26221 		/*
26222 		 * Beta_ecn is the congestion control value for NewReno that influences how
26223 		 * much of a backoff happens when a ECN mark is detected. It is normally set
26224 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
26225 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
26226 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
26227 		 */
26228 
26229 	case TCP_RACK_PACING_BETA_ECN:
26230 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
26231 			error = EINVAL;
26232 		else if (rack->rc_pacing_cc_set == 0)
26233 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
26234 		else {
26235 			/*
26236 			 * Reach out into the CC data and report back what
26237 			 * I have previously set. Yeah it looks hackish but
26238 			 * we don't want to report the saved values.
26239 			 */
26240 			if (tp->t_ccv.cc_data)
26241 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
26242 			else
26243 				error = EINVAL;
26244 		}
26245 		break;
26246 	case TCP_RACK_DSACK_OPT:
26247 		optval = 0;
26248 		if (rack->rc_rack_tmr_std_based) {
26249 			optval |= 1;
26250 		}
26251 		if (rack->rc_rack_use_dsack) {
26252 			optval |= 2;
26253 		}
26254 		break;
26255 	case TCP_RACK_ENABLE_HYSTART:
26256 	{
26257 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
26258 			optval = RACK_HYSTART_ON;
26259 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
26260 				optval = RACK_HYSTART_ON_W_SC;
26261 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
26262 				optval = RACK_HYSTART_ON_W_SC_C;
26263 		} else {
26264 			optval = RACK_HYSTART_OFF;
26265 		}
26266 	}
26267 	break;
26268 	case TCP_RACK_DGP_IN_REC:
26269 		error = EINVAL;
26270 		break;
26271 	case TCP_RACK_HI_BETA:
26272 		optval = rack->rack_hibeta;
26273 		break;
26274 	case TCP_POLICER_MSS:
26275 		optval = rack->r_ctl.policer_del_mss;
26276 		break;
26277 	case TCP_POLICER_DETECT:
26278 		optval = rack->r_ctl.saved_policer_val;
26279 		break;
26280 	case TCP_DEFER_OPTIONS:
26281 		optval = rack->defer_options;
26282 		break;
26283 	case TCP_RACK_MEASURE_CNT:
26284 		optval = rack->r_ctl.req_measurements;
26285 		break;
26286 	case TCP_REC_ABC_VAL:
26287 		optval = rack->r_use_labc_for_rec;
26288 		break;
26289 	case TCP_RACK_ABC_VAL:
26290 		optval = rack->rc_labc;
26291 		break;
26292 	case TCP_HDWR_UP_ONLY:
26293 		optval= rack->r_up_only;
26294 		break;
26295 	case TCP_FILLCW_RATE_CAP:
26296 		loptval = rack->r_ctl.fillcw_cap;
26297 		break;
26298 	case TCP_PACING_RATE_CAP:
26299 		loptval = rack->r_ctl.bw_rate_cap;
26300 		break;
26301 	case TCP_RACK_PROFILE:
26302 		/* You cannot retrieve a profile, its write only */
26303 		error = EINVAL;
26304 		break;
26305 	case TCP_SIDECHAN_DIS:
26306 		optval = rack->r_ctl.side_chan_dis_mask;
26307 		break;
26308 	case TCP_HYBRID_PACING:
26309 		/* You cannot retrieve hybrid pacing information, its write only */
26310 		error = EINVAL;
26311 		break;
26312 	case TCP_USE_CMP_ACKS:
26313 		optval = rack->r_use_cmp_ack;
26314 		break;
26315 	case TCP_RACK_PACE_TO_FILL:
26316 		optval = rack->rc_pace_to_cwnd;
26317 		break;
26318 	case TCP_RACK_NO_PUSH_AT_MAX:
26319 		optval = rack->r_ctl.rc_no_push_at_mrtt;
26320 		break;
26321 	case TCP_SHARED_CWND_ENABLE:
26322 		optval = rack->rack_enable_scwnd;
26323 		break;
26324 	case TCP_RACK_NONRXT_CFG_RATE:
26325 		optval = rack->rack_rec_nonrxt_use_cr;
26326 		break;
26327 	case TCP_NO_PRR:
26328 		if (rack->rack_no_prr  == 1)
26329 			optval = 1;
26330 		else if (rack->no_prr_addback == 1)
26331 			optval = 2;
26332 		else
26333 			optval = 0;
26334 		break;
26335 	case TCP_GP_USE_LTBW:
26336 		if (rack->dis_lt_bw) {
26337 			/* It is not used */
26338 			optval = 0;
26339 		} else if (rack->use_lesser_lt_bw) {
26340 			/* we use min() */
26341 			optval = 1;
26342 		} else {
26343 			/* we use max() */
26344 			optval = 2;
26345 		}
26346 		break;
26347 	case TCP_RACK_DO_DETECTION:
26348 		optval = rack->do_detection;
26349 		break;
26350 	case TCP_RACK_MBUF_QUEUE:
26351 		/* Now do we use the LRO mbuf-queue feature */
26352 		optval = rack->r_mbuf_queue;
26353 		break;
26354 	case RACK_CSPR_IS_FCC:
26355 		optval = rack->cspr_is_fcc;
26356 		break;
26357 	case TCP_TIMELY_DYN_ADJ:
26358 		optval = rack->rc_gp_dyn_mul;
26359 		break;
26360 	case TCP_BBR_IWINTSO:
26361 		error = EINVAL;
26362 		break;
26363 	case TCP_RACK_TLP_REDUCE:
26364 		/* RACK TLP cwnd reduction (bool) */
26365 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
26366 		break;
26367 	case TCP_BBR_RACK_INIT_RATE:
26368 		val = rack->r_ctl.init_rate;
26369 		/* convert to kbits per sec */
26370 		val *= 8;
26371 		val /= 1000;
26372 		optval = (uint32_t)val;
26373 		break;
26374 	case TCP_RACK_FORCE_MSEG:
26375 		optval = rack->rc_force_max_seg;
26376 		break;
26377 	case TCP_RACK_PACE_MIN_SEG:
26378 		optval = rack->r_ctl.rc_user_set_min_segs;
26379 		break;
26380 	case TCP_RACK_PACE_MAX_SEG:
26381 		/* Max segments in a pace */
26382 		optval = rack->rc_user_set_max_segs;
26383 		break;
26384 	case TCP_RACK_PACE_ALWAYS:
26385 		/* Use the always pace method */
26386 		optval = rack->rc_always_pace;
26387 		break;
26388 	case TCP_RACK_PRR_SENDALOT:
26389 		/* Allow PRR to send more than one seg */
26390 		optval = rack->r_ctl.rc_prr_sendalot;
26391 		break;
26392 	case TCP_RACK_MIN_TO:
26393 		/* Minimum time between rack t-o's in ms */
26394 		optval = rack->r_ctl.rc_min_to;
26395 		break;
26396 	case TCP_RACK_SPLIT_LIMIT:
26397 		optval = rack->r_ctl.rc_split_limit;
26398 		break;
26399 	case TCP_RACK_EARLY_SEG:
26400 		/* If early recovery max segments */
26401 		optval = rack->r_ctl.rc_early_recovery_segs;
26402 		break;
26403 	case TCP_RACK_REORD_THRESH:
26404 		/* RACK reorder threshold (shift amount) */
26405 		optval = rack->r_ctl.rc_reorder_shift;
26406 		break;
26407 	case TCP_SS_EEXIT:
26408 		if (rack->r_ctl.gp_rnd_thresh) {
26409 			uint32_t v;
26410 
26411 			v = rack->r_ctl.gp_gain_req;
26412 			v <<= 17;
26413 			optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
26414 			if (rack->r_ctl.gate_to_fs == 1)
26415 				optval |= 0x10000;
26416 		} else
26417 			optval = 0;
26418 		break;
26419 	case TCP_RACK_REORD_FADE:
26420 		/* Does reordering fade after ms time */
26421 		optval = rack->r_ctl.rc_reorder_fade;
26422 		break;
26423 	case TCP_BBR_USE_RACK_RR:
26424 		/* Do we use the rack cheat for rxt */
26425 		optval = rack->use_rack_rr;
26426 		break;
26427 	case TCP_RACK_RR_CONF:
26428 		optval = rack->r_rr_config;
26429 		break;
26430 	case TCP_HDWR_RATE_CAP:
26431 		optval = rack->r_rack_hw_rate_caps;
26432 		break;
26433 	case TCP_BBR_HDWR_PACE:
26434 		optval = rack->rack_hdw_pace_ena;
26435 		break;
26436 	case TCP_RACK_TLP_THRESH:
26437 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
26438 		optval = rack->r_ctl.rc_tlp_threshold;
26439 		break;
26440 	case TCP_RACK_PKT_DELAY:
26441 		/* RACK added ms i.e. rack-rtt + reord + N */
26442 		optval = rack->r_ctl.rc_pkt_delay;
26443 		break;
26444 	case TCP_RACK_TLP_USE:
26445 		optval = rack->rack_tlp_threshold_use;
26446 		break;
26447 	case TCP_PACING_DND:
26448 		optval = rack->rc_pace_dnd;
26449 		break;
26450 	case TCP_RACK_PACE_RATE_CA:
26451 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
26452 		break;
26453 	case TCP_RACK_PACE_RATE_SS:
26454 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
26455 		break;
26456 	case TCP_RACK_PACE_RATE_REC:
26457 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
26458 		break;
26459 	case TCP_DGP_UPPER_BOUNDS:
26460 		optval = rack->r_ctl.rack_per_upper_bound_ss;
26461 		optval <<= 16;
26462 		optval |= rack->r_ctl.rack_per_upper_bound_ca;
26463 		break;
26464 	case TCP_RACK_GP_INCREASE_SS:
26465 		optval = rack->r_ctl.rack_per_of_gp_ca;
26466 		break;
26467 	case TCP_RACK_GP_INCREASE_CA:
26468 		optval = rack->r_ctl.rack_per_of_gp_ss;
26469 		break;
26470 	case TCP_RACK_PACING_DIVISOR:
26471 		optval = rack->r_ctl.pace_len_divisor;
26472 		break;
26473 	case TCP_BBR_RACK_RTT_USE:
26474 		optval = rack->r_ctl.rc_rate_sample_method;
26475 		break;
26476 	case TCP_DELACK:
26477 		optval = tp->t_delayed_ack;
26478 		break;
26479 	case TCP_DATA_AFTER_CLOSE:
26480 		optval = rack->rc_allow_data_af_clo;
26481 		break;
26482 	case TCP_SHARED_CWND_TIME_LIMIT:
26483 		optval = rack->r_limit_scw;
26484 		break;
26485 	case TCP_HONOR_HPTS_MIN:
26486 		if (rack->r_use_hpts_min)
26487 			optval = rack->r_ctl.max_reduction;
26488 		else
26489 			optval = 0;
26490 		break;
26491 	case TCP_REC_IS_DYN:
26492 		optval = rack->rc_gp_no_rec_chg;
26493 		break;
26494 	case TCP_NO_TIMELY:
26495 		optval = rack->rc_skip_timely;
26496 		break;
26497 	case TCP_RACK_TIMER_SLOP:
26498 		optval = rack->r_ctl.timer_slop;
26499 		break;
26500 	default:
26501 		return (tcp_default_ctloutput(tp, sopt));
26502 		break;
26503 	}
26504 	INP_WUNLOCK(inp);
26505 	if (error == 0) {
26506 		if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
26507 		    (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
26508 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
26509 		else
26510 			error = sooptcopyout(sopt, &optval, sizeof optval);
26511 	}
26512 	return (error);
26513 }
26514 
26515 static int
26516 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
26517 {
26518 	if (sopt->sopt_dir == SOPT_SET) {
26519 		return (rack_set_sockopt(tp, sopt));
26520 	} else if (sopt->sopt_dir == SOPT_GET) {
26521 		return (rack_get_sockopt(tp, sopt));
26522 	} else {
26523 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
26524 	}
26525 }
26526 
26527 static const char *rack_stack_names[] = {
26528 	__XSTRING(STACKNAME),
26529 #ifdef STACKALIAS
26530 	__XSTRING(STACKALIAS),
26531 #endif
26532 };
26533 
26534 static int
26535 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
26536 {
26537 	memset(mem, 0, size);
26538 	return (0);
26539 }
26540 
26541 static void
26542 rack_dtor(void *mem, int32_t size, void *arg)
26543 {
26544 
26545 }
26546 
26547 static bool rack_mod_inited = false;
26548 
26549 static int
26550 tcp_addrack(module_t mod, int32_t type, void *data)
26551 {
26552 	int32_t err = 0;
26553 	int num_stacks;
26554 
26555 	switch (type) {
26556 	case MOD_LOAD:
26557 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
26558 		    sizeof(struct rack_sendmap),
26559 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
26560 
26561 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
26562 		    sizeof(struct tcp_rack),
26563 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
26564 
26565 		sysctl_ctx_init(&rack_sysctl_ctx);
26566 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
26567 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
26568 		    OID_AUTO,
26569 #ifdef STACKALIAS
26570 		    __XSTRING(STACKALIAS),
26571 #else
26572 		    __XSTRING(STACKNAME),
26573 #endif
26574 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
26575 		    "");
26576 		if (rack_sysctl_root == NULL) {
26577 			printf("Failed to add sysctl node\n");
26578 			err = EFAULT;
26579 			goto free_uma;
26580 		}
26581 		rack_init_sysctls();
26582 		num_stacks = nitems(rack_stack_names);
26583 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
26584 		    rack_stack_names, &num_stacks);
26585 		if (err) {
26586 			printf("Failed to register %s stack name for "
26587 			    "%s module\n", rack_stack_names[num_stacks],
26588 			    __XSTRING(MODNAME));
26589 			sysctl_ctx_free(&rack_sysctl_ctx);
26590 free_uma:
26591 			uma_zdestroy(rack_zone);
26592 			uma_zdestroy(rack_pcb_zone);
26593 			rack_counter_destroy();
26594 			printf("Failed to register rack module -- err:%d\n", err);
26595 			return (err);
26596 		}
26597 		tcp_lro_reg_mbufq();
26598 		rack_mod_inited = true;
26599 		break;
26600 	case MOD_QUIESCE:
26601 		err = deregister_tcp_functions(&__tcp_rack, true, false);
26602 		break;
26603 	case MOD_UNLOAD:
26604 		err = deregister_tcp_functions(&__tcp_rack, false, true);
26605 		if (err == EBUSY)
26606 			break;
26607 		if (rack_mod_inited) {
26608 			uma_zdestroy(rack_zone);
26609 			uma_zdestroy(rack_pcb_zone);
26610 			sysctl_ctx_free(&rack_sysctl_ctx);
26611 			rack_counter_destroy();
26612 			rack_mod_inited = false;
26613 		}
26614 		tcp_lro_dereg_mbufq();
26615 		err = 0;
26616 		break;
26617 	default:
26618 		return (EOPNOTSUPP);
26619 	}
26620 	return (err);
26621 }
26622 
26623 static moduledata_t tcp_rack = {
26624 	.name = __XSTRING(MODNAME),
26625 	.evhand = tcp_addrack,
26626 	.priv = 0
26627 };
26628 
26629 MODULE_VERSION(MODNAME, 1);
26630 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
26631 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
26632 
26633 #endif /* #if !defined(INET) && !defined(INET6) */
26634