xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 535af610) 
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 __FBSDID("$FreeBSD$");
29 
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_ratelimit.h"
34 #include "opt_kern_tls.h"
35 #if defined(INET) || defined(INET6)
36 #include <sys/param.h>
37 #include <sys/arb.h>
38 #include <sys/module.h>
39 #include <sys/kernel.h>
40 #ifdef TCP_HHOOK
41 #include <sys/hhook.h>
42 #endif
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/mbuf.h>
48 #include <sys/proc.h>		/* for proc0 declaration */
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #ifdef STATS
54 #include <sys/qmath.h>
55 #include <sys/tree.h>
56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
57 #else
58 #include <sys/tree.h>
59 #endif
60 #include <sys/refcount.h>
61 #include <sys/queue.h>
62 #include <sys/tim_filter.h>
63 #include <sys/smp.h>
64 #include <sys/kthread.h>
65 #include <sys/kern_prefetch.h>
66 #include <sys/protosw.h>
67 #ifdef TCP_ACCOUNTING
68 #include <sys/sched.h>
69 #include <machine/cpu.h>
70 #endif
71 #include <vm/uma.h>
72 
73 #include <net/route.h>
74 #include <net/route/nhop.h>
75 #include <net/vnet.h>
76 
77 #define TCPSTATES		/* for logging */
78 
79 #include <netinet/in.h>
80 #include <netinet/in_kdtrace.h>
81 #include <netinet/in_pcb.h>
82 #include <netinet/ip.h>
83 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
84 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
85 #include <netinet/ip_var.h>
86 #include <netinet/ip6.h>
87 #include <netinet6/in6_pcb.h>
88 #include <netinet6/ip6_var.h>
89 #include <netinet/tcp.h>
90 #define	TCPOUTFLAGS
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_seq.h>
93 #include <netinet/tcp_timer.h>
94 #include <netinet/tcp_var.h>
95 #include <netinet/tcp_log_buf.h>
96 #include <netinet/tcp_syncache.h>
97 #include <netinet/tcp_hpts.h>
98 #include <netinet/tcp_ratelimit.h>
99 #include <netinet/tcp_accounting.h>
100 #include <netinet/tcpip.h>
101 #include <netinet/cc/cc.h>
102 #include <netinet/cc/cc_newreno.h>
103 #include <netinet/tcp_fastopen.h>
104 #include <netinet/tcp_lro.h>
105 #ifdef NETFLIX_SHARED_CWND
106 #include <netinet/tcp_shared_cwnd.h>
107 #endif
108 #ifdef TCP_OFFLOAD
109 #include <netinet/tcp_offload.h>
110 #endif
111 #ifdef INET6
112 #include <netinet6/tcp6_var.h>
113 #endif
114 #include <netinet/tcp_ecn.h>
115 
116 #include <netipsec/ipsec_support.h>
117 
118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
119 #include <netipsec/ipsec.h>
120 #include <netipsec/ipsec6.h>
121 #endif				/* IPSEC */
122 
123 #include <netinet/udp.h>
124 #include <netinet/udp_var.h>
125 #include <machine/in_cksum.h>
126 
127 #ifdef MAC
128 #include <security/mac/mac_framework.h>
129 #endif
130 #include "sack_filter.h"
131 #include "tcp_rack.h"
132 #include "tailq_hash.h"
133 #include "rack_bbr_common.h"
134 
135 uma_zone_t rack_zone;
136 uma_zone_t rack_pcb_zone;
137 
138 #ifndef TICKS2SBT
139 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
140 #endif
141 
142 VNET_DECLARE(uint32_t, newreno_beta);
143 VNET_DECLARE(uint32_t, newreno_beta_ecn);
144 #define V_newreno_beta VNET(newreno_beta)
145 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
146 
147 
148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
149 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
150 
151 struct sysctl_ctx_list rack_sysctl_ctx;
152 struct sysctl_oid *rack_sysctl_root;
153 
154 #define CUM_ACKED 1
155 #define SACKED 2
156 
157 /*
158  * The RACK module incorporates a number of
159  * TCP ideas that have been put out into the IETF
160  * over the last few years:
161  * - Matt Mathis's Rate Halving which slowly drops
162  *    the congestion window so that the ack clock can
163  *    be maintained during a recovery.
164  * - Yuchung Cheng's RACK TCP (for which its named) that
165  *    will stop us using the number of dup acks and instead
166  *    use time as the gage of when we retransmit.
167  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
168  *    of Dukkipati et.al.
169  * RACK depends on SACK, so if an endpoint arrives that
170  * cannot do SACK the state machine below will shuttle the
171  * connection back to using the "default" TCP stack that is
172  * in FreeBSD.
173  *
174  * To implement RACK the original TCP stack was first decomposed
175  * into a functional state machine with individual states
176  * for each of the possible TCP connection states. The do_segment
177  * functions role in life is to mandate the connection supports SACK
178  * initially and then assure that the RACK state matches the conenction
179  * state before calling the states do_segment function. Each
180  * state is simplified due to the fact that the original do_segment
181  * has been decomposed and we *know* what state we are in (no
182  * switches on the state) and all tests for SACK are gone. This
183  * greatly simplifies what each state does.
184  *
185  * TCP output is also over-written with a new version since it
186  * must maintain the new rack scoreboard.
187  *
188  */
189 static int32_t rack_tlp_thresh = 1;
190 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
191 static int32_t rack_tlp_use_greater = 1;
192 static int32_t rack_reorder_thresh = 2;
193 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
194 						 * - 60 seconds */
195 static uint32_t rack_clamp_ss_upper = 110;
196 static uint32_t rack_clamp_ca_upper = 105;
197 static uint32_t rack_rxt_min_rnds = 10;	/* Min rounds if drastic rxt clamp is in place */
198 static uint32_t rack_unclamp_round_thresh = 100;	/* number of perfect rounds before we unclamp */
199 static uint32_t rack_unclamp_rxt_thresh = 5;	/* .5%  and under */
200 static uint64_t rack_rxt_clamp_thresh = 0;	/* Do we do the rxt clamp thing */
201 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
202 static int32_t rack_rxt_controls = 0;
203 static int32_t rack_fill_cw_state = 0;
204 static uint8_t rack_req_measurements = 1;
205 /* Attack threshold detections */
206 static uint32_t rack_highest_sack_thresh_seen = 0;
207 static uint32_t rack_highest_move_thresh_seen = 0;
208 static uint32_t rack_merge_out_sacks_on_attack = 0;
209 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
210 static int32_t rack_hw_pace_extra_slots = 0;	/* 2 extra MSS time betweens */
211 static int32_t rack_hw_rate_caps = 0; /* 1; */
212 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
213 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
214 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
215 static int32_t rack_hw_up_only = 0;
216 static int32_t rack_stats_gets_ms_rtt = 1;
217 static int32_t rack_prr_addbackmax = 2;
218 static int32_t rack_do_hystart = 0;
219 static int32_t rack_apply_rtt_with_reduced_conf = 0;
220 static int32_t rack_hibeta_setting = 0;
221 static int32_t rack_default_pacing_divisor = 250;
222 static int32_t rack_uses_full_dgp_in_rec = 1;
223 static uint16_t rack_pacing_min_seg = 0;
224 
225 
226 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
227 static int32_t rack_pkt_delay = 1000;
228 static int32_t rack_send_a_lot_in_prr = 1;
229 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
230 static int32_t rack_verbose_logging = 0;
231 static int32_t rack_ignore_data_after_close = 1;
232 static int32_t rack_enable_shared_cwnd = 1;
233 static int32_t rack_use_cmp_acks = 1;
234 static int32_t rack_use_fsb = 1;
235 static int32_t rack_use_rfo = 1;
236 static int32_t rack_use_rsm_rfo = 1;
237 static int32_t rack_max_abc_post_recovery = 2;
238 static int32_t rack_client_low_buf = 0;
239 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 */
240 static int32_t rack_bw_multipler = 2;		/* Limit on fill cw's jump up to be this x gp_est */
241 #ifdef TCP_ACCOUNTING
242 static int32_t rack_tcp_accounting = 0;
243 #endif
244 static int32_t rack_limits_scwnd = 1;
245 static int32_t rack_enable_mqueue_for_nonpaced = 0;
246 static int32_t rack_hybrid_allow_set_maxseg = 0;
247 static int32_t rack_disable_prr = 0;
248 static int32_t use_rack_rr = 1;
249 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
250 static int32_t rack_persist_min = 250000;	/* 250usec */
251 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
252 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
253 static int32_t rack_default_init_window = 0;	/* Use system default */
254 static int32_t rack_limit_time_with_srtt = 0;
255 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
256 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
257 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
258 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
259 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
260 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
261 static int32_t rack_full_buffer_discount = 10;
262 /*
263  * Currently regular tcp has a rto_min of 30ms
264  * the backoff goes 12 times so that ends up
265  * being a total of 122.850 seconds before a
266  * connection is killed.
267  */
268 static uint32_t rack_def_data_window = 20;
269 static uint32_t rack_goal_bdp = 2;
270 static uint32_t rack_min_srtts = 1;
271 static uint32_t rack_min_measure_usec = 0;
272 static int32_t rack_tlp_min = 10000;	/* 10ms */
273 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
274 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
275 static const int32_t rack_free_cache = 2;
276 static int32_t rack_hptsi_segments = 40;
277 static int32_t rack_rate_sample_method = USE_RTT_LOW;
278 static int32_t rack_pace_every_seg = 0;
279 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
280 static int32_t rack_slot_reduction = 4;
281 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
282 static int32_t rack_cwnd_block_ends_measure = 0;
283 static int32_t rack_rwnd_block_ends_measure = 0;
284 static int32_t rack_def_profile = 0;
285 
286 static int32_t rack_lower_cwnd_at_tlp = 0;
287 static int32_t rack_limited_retran = 0;
288 static int32_t rack_always_send_oldest = 0;
289 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
290 
291 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
292 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
293 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
294 
295 /* Probertt */
296 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
297 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
298 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
299 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
300 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
301 
302 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
303 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
304 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
305 static uint32_t rack_probertt_use_min_rtt_exit = 0;
306 static uint32_t rack_probe_rtt_sets_cwnd = 0;
307 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
308 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
309 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
310 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
311 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
312 static uint32_t rack_probertt_filter_life = 10000000;
313 static uint32_t rack_probertt_lower_within = 10;
314 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
315 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
316 static int32_t rack_probertt_clear_is = 1;
317 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
318 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
319 
320 /* Part of pacing */
321 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
322 
323 /* Timely information:
324  *
325  * Here we have various control parameters on how
326  * timely may change the multiplier. rack_gain_p5_ub
327  * is associated with timely but not directly influencing
328  * the rate decision like the other variables. It controls
329  * the way fill-cw interacts with timely and caps how much
330  * timely can boost the fill-cw b/w.
331  *
332  * The other values are various boost/shrink numbers as well
333  * as potential caps when adjustments are made to the timely
334  * gain (returned by rack_get_output_gain(). Remember too that
335  * the gain returned can be overriden by other factors such as
336  * probeRTT as well as fixed-rate-pacing.
337  */
338 static int32_t rack_gain_p5_ub = 250;
339 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
340 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
341 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
342 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
343 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
344 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
345 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
346 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
347 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
348 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
349 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
350 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
351 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
352 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
353 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
354 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
355 static int32_t rack_use_max_for_nobackoff = 0;
356 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
357 static int32_t rack_timely_no_stopping = 0;
358 static int32_t rack_down_raise_thresh = 100;
359 static int32_t rack_req_segs = 1;
360 static uint64_t rack_bw_rate_cap = 0;
361 
362 
363 /* Rack specific counters */
364 counter_u64_t rack_saw_enobuf;
365 counter_u64_t rack_saw_enobuf_hw;
366 counter_u64_t rack_saw_enetunreach;
367 counter_u64_t rack_persists_sends;
368 counter_u64_t rack_persists_acks;
369 counter_u64_t rack_persists_loss;
370 counter_u64_t rack_persists_lost_ends;
371 counter_u64_t rack_total_bytes;
372 #ifdef INVARIANTS
373 counter_u64_t rack_adjust_map_bw;
374 #endif
375 /* Tail loss probe counters */
376 counter_u64_t rack_tlp_tot;
377 counter_u64_t rack_tlp_newdata;
378 counter_u64_t rack_tlp_retran;
379 counter_u64_t rack_tlp_retran_bytes;
380 counter_u64_t rack_to_tot;
381 counter_u64_t rack_hot_alloc;
382 counter_u64_t rack_to_alloc;
383 counter_u64_t rack_to_alloc_hard;
384 counter_u64_t rack_to_alloc_emerg;
385 counter_u64_t rack_to_alloc_limited;
386 counter_u64_t rack_alloc_limited_conns;
387 counter_u64_t rack_split_limited;
388 counter_u64_t rack_rxt_clamps_cwnd;
389 counter_u64_t rack_rxt_clamps_cwnd_uniq;
390 
391 counter_u64_t rack_multi_single_eq;
392 counter_u64_t rack_proc_non_comp_ack;
393 
394 counter_u64_t rack_fto_send;
395 counter_u64_t rack_fto_rsm_send;
396 counter_u64_t rack_nfto_resend;
397 counter_u64_t rack_non_fto_send;
398 counter_u64_t rack_extended_rfo;
399 
400 counter_u64_t rack_sack_proc_all;
401 counter_u64_t rack_sack_proc_short;
402 counter_u64_t rack_sack_proc_restart;
403 counter_u64_t rack_sack_attacks_detected;
404 counter_u64_t rack_sack_attacks_reversed;
405 counter_u64_t rack_sack_attacks_suspect;
406 counter_u64_t rack_sack_used_next_merge;
407 counter_u64_t rack_sack_splits;
408 counter_u64_t rack_sack_used_prev_merge;
409 counter_u64_t rack_sack_skipped_acked;
410 counter_u64_t rack_ack_total;
411 counter_u64_t rack_express_sack;
412 counter_u64_t rack_sack_total;
413 counter_u64_t rack_move_none;
414 counter_u64_t rack_move_some;
415 
416 counter_u64_t rack_input_idle_reduces;
417 counter_u64_t rack_collapsed_win;
418 counter_u64_t rack_collapsed_win_seen;
419 counter_u64_t rack_collapsed_win_rxt;
420 counter_u64_t rack_collapsed_win_rxt_bytes;
421 counter_u64_t rack_try_scwnd;
422 counter_u64_t rack_hw_pace_init_fail;
423 counter_u64_t rack_hw_pace_lost;
424 
425 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
426 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
427 
428 
429 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
430 
431 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
432 	(tv) = (value) + slop;	 \
433 	if ((u_long)(tv) < (u_long)(tvmin)) \
434 		(tv) = (tvmin); \
435 	if ((u_long)(tv) > (u_long)(tvmax)) \
436 		(tv) = (tvmax); \
437 } while (0)
438 
439 static void
440 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
441 
442 static int
443 rack_process_ack(struct mbuf *m, struct tcphdr *th,
444     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
445     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
446 static int
447 rack_process_data(struct mbuf *m, struct tcphdr *th,
448     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
449     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
450 static void
451 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
452    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
453 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
454 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
455     uint8_t limit_type);
456 static struct rack_sendmap *
457 rack_check_recovery_mode(struct tcpcb *tp,
458     uint32_t tsused);
459 static void
460 rack_cong_signal(struct tcpcb *tp,
461 		 uint32_t type, uint32_t ack, int );
462 static void rack_counter_destroy(void);
463 static int
464 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
465 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
466 static void
467 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
468 static void
469 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
470     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
471 static void rack_dtor(void *mem, int32_t size, void *arg);
472 static void
473 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
474     uint32_t flex1, uint32_t flex2,
475     uint32_t flex3, uint32_t flex4,
476     uint32_t flex5, uint32_t flex6,
477     uint16_t flex7, uint8_t mod);
478 
479 static void
480 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
481    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
482    struct rack_sendmap *rsm, uint8_t quality);
483 static struct rack_sendmap *
484 rack_find_high_nonack(struct tcp_rack *rack,
485     struct rack_sendmap *rsm);
486 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
487 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
488 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
489 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
490 static void
491 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
492 			    tcp_seq th_ack, int line, uint8_t quality);
493 static void
494 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
495 
496 static uint32_t
497 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
498 static int32_t rack_handoff_ok(struct tcpcb *tp);
499 static int32_t rack_init(struct tcpcb *tp, void **ptr);
500 static void rack_init_sysctls(void);
501 
502 static void
503 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
504     struct tcphdr *th, int entered_rec, int dup_ack_struck,
505     int *dsack_seen, int *sacks_seen);
506 static void
507 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
508     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
509     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
510 
511 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
512 
513 static void
514 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
515     struct rack_sendmap *rsm);
516 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
517 static int32_t rack_output(struct tcpcb *tp);
518 
519 static uint32_t
520 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
521     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
522     uint32_t cts, int *no_extra, int *moved_two, uint32_t segsiz);
523 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
524 static void rack_remxt_tmr(struct tcpcb *tp);
525 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
526 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
527 static int32_t rack_stopall(struct tcpcb *tp);
528 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
529 static uint32_t
530 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
531     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag, int segsiz);
532 static void
533 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
534     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz);
535 static int
536 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
537     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
538 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
539 static int
540 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
541     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
542     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
543 static int
544 rack_do_closing(struct mbuf *m, struct tcphdr *th,
545     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
546     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
547 static int
548 rack_do_established(struct mbuf *m, struct tcphdr *th,
549     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
550     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
551 static int
552 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
553     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
554     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
555 static int
556 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
557     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
558     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
559 static int
560 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
561     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
562     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
563 static int
564 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
565     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
566     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
567 static int
568 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
569     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
570     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
571 static int
572 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
573     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
574     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
575 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
576 struct rack_sendmap *
577 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
578     uint32_t tsused);
579 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
580     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
581 static void
582      tcp_rack_partialack(struct tcpcb *tp);
583 static int
584 rack_set_profile(struct tcp_rack *rack, int prof);
585 static void
586 rack_apply_deferred_options(struct tcp_rack *rack);
587 
588 int32_t rack_clear_counter=0;
589 
590 static uint64_t
591 rack_get_lt_bw(struct tcp_rack *rack)
592 {
593 	struct timeval tv;
594 	uint64_t tim, bytes;
595 
596 	tim = rack->r_ctl.lt_bw_time;
597 	bytes = rack->r_ctl.lt_bw_bytes;
598 	if (rack->lt_bw_up) {
599 		/* Include all the current bytes too */
600 		microuptime(&tv);
601 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
602 		tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
603 	}
604 	if ((bytes != 0) && (tim != 0))
605 		return ((bytes * (uint64_t)1000000) / tim);
606 	else
607 		return (0);
608 }
609 
610 static void
611 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
612 {
613 	struct sockopt sopt;
614 	struct cc_newreno_opts opt;
615 	struct newreno old;
616 	struct tcpcb *tp;
617 	int error, failed = 0;
618 
619 	tp = rack->rc_tp;
620 	if (tp->t_cc == NULL) {
621 		/* Tcb is leaving */
622 		return;
623 	}
624 	rack->rc_pacing_cc_set = 1;
625 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
626 		/* Not new-reno we can't play games with beta! */
627 		failed = 1;
628 		goto out;
629 
630 	}
631 	if (CC_ALGO(tp)->ctl_output == NULL)  {
632 		/* Huh, not using new-reno so no swaps.? */
633 		failed = 2;
634 		goto out;
635 	}
636 	/* Get the current values out */
637 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
638 	sopt.sopt_dir = SOPT_GET;
639 	opt.name = CC_NEWRENO_BETA;
640 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
641 	if (error)  {
642 		failed = 3;
643 		goto out;
644 	}
645 	old.beta = opt.val;
646 	opt.name = CC_NEWRENO_BETA_ECN;
647 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
648 	if (error)  {
649 		failed = 4;
650 		goto out;
651 	}
652 	old.beta_ecn = opt.val;
653 
654 	/* Now lets set in the values we have stored */
655 	sopt.sopt_dir = SOPT_SET;
656 	opt.name = CC_NEWRENO_BETA;
657 	opt.val = rack->r_ctl.rc_saved_beta.beta;
658 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
659 	if (error)  {
660 		failed = 5;
661 		goto out;
662 	}
663 	opt.name = CC_NEWRENO_BETA_ECN;
664 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
665 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
666 	if (error) {
667 		failed = 6;
668 		goto out;
669 	}
670 	/* Save off the values for restoral */
671 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
672 out:
673 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
674 		union tcp_log_stackspecific log;
675 		struct timeval tv;
676 		struct newreno *ptr;
677 
678 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
679 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
680 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
681 		log.u_bbr.flex1 = ptr->beta;
682 		log.u_bbr.flex2 = ptr->beta_ecn;
683 		log.u_bbr.flex3 = ptr->newreno_flags;
684 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
685 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
686 		log.u_bbr.flex6 = failed;
687 		log.u_bbr.flex7 = rack->gp_ready;
688 		log.u_bbr.flex7 <<= 1;
689 		log.u_bbr.flex7 |= rack->use_fixed_rate;
690 		log.u_bbr.flex7 <<= 1;
691 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
692 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
693 		log.u_bbr.flex8 = flex8;
694 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
695 			       0, &log, false, NULL, NULL, 0, &tv);
696 	}
697 }
698 
699 static void
700 rack_set_cc_pacing(struct tcp_rack *rack)
701 {
702 	if (rack->rc_pacing_cc_set)
703 		return;
704 	/*
705 	 * Use the swap utility placing in 3 for flex8 to id a
706 	 * set of a new set of values.
707 	 */
708 	rack->rc_pacing_cc_set = 1;
709 	rack_swap_beta_values(rack, 3);
710 }
711 
712 static void
713 rack_undo_cc_pacing(struct tcp_rack *rack)
714 {
715 	if (rack->rc_pacing_cc_set == 0)
716 		return;
717 	/*
718 	 * Use the swap utility placing in 4 for flex8 to id a
719 	 * restoral of the old values.
720 	 */
721 	rack->rc_pacing_cc_set = 0;
722 	rack_swap_beta_values(rack, 4);
723 }
724 
725 static void
726 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
727 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
728 {
729 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
730 		union tcp_log_stackspecific log;
731 		struct timeval tv;
732 
733 		memset(&log, 0, sizeof(log));
734 		log.u_bbr.flex1 = seq_end;
735 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
736 		log.u_bbr.flex3 = ack_end_t;
737 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
738 		log.u_bbr.flex5 = send_end_t;
739 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
740 		log.u_bbr.flex7 = mode;
741 		log.u_bbr.flex8 = 69;
742 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
743 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
744 		log.u_bbr.pkts_out = line;
745 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
746 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
747 		if (rsm != NULL) {
748 			log.u_bbr.applimited = rsm->r_start;
749 			log.u_bbr.delivered = rsm->r_end;
750 			log.u_bbr.epoch = rsm->r_flags;
751 		}
752 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
753 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
754 		    &rack->rc_inp->inp_socket->so_rcv,
755 		    &rack->rc_inp->inp_socket->so_snd,
756 		    BBR_LOG_HPTSI_CALC, 0,
757 		    0, &log, false, &tv);
758 	}
759 }
760 
761 static int
762 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
763 {
764 	uint32_t stat;
765 	int32_t error;
766 
767 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
768 	if (error || req->newptr == NULL)
769 		return error;
770 
771 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
772 	if (error)
773 		return (error);
774 	if (stat == 1) {
775 #ifdef INVARIANTS
776 		printf("Clearing RACK counters\n");
777 #endif
778 		counter_u64_zero(rack_tlp_tot);
779 		counter_u64_zero(rack_tlp_newdata);
780 		counter_u64_zero(rack_tlp_retran);
781 		counter_u64_zero(rack_tlp_retran_bytes);
782 		counter_u64_zero(rack_to_tot);
783 		counter_u64_zero(rack_saw_enobuf);
784 		counter_u64_zero(rack_saw_enobuf_hw);
785 		counter_u64_zero(rack_saw_enetunreach);
786 		counter_u64_zero(rack_persists_sends);
787 		counter_u64_zero(rack_total_bytes);
788 		counter_u64_zero(rack_persists_acks);
789 		counter_u64_zero(rack_persists_loss);
790 		counter_u64_zero(rack_persists_lost_ends);
791 #ifdef INVARIANTS
792 		counter_u64_zero(rack_adjust_map_bw);
793 #endif
794 		counter_u64_zero(rack_to_alloc_hard);
795 		counter_u64_zero(rack_to_alloc_emerg);
796 		counter_u64_zero(rack_sack_proc_all);
797 		counter_u64_zero(rack_fto_send);
798 		counter_u64_zero(rack_fto_rsm_send);
799 		counter_u64_zero(rack_extended_rfo);
800 		counter_u64_zero(rack_hw_pace_init_fail);
801 		counter_u64_zero(rack_hw_pace_lost);
802 		counter_u64_zero(rack_non_fto_send);
803 		counter_u64_zero(rack_nfto_resend);
804 		counter_u64_zero(rack_sack_proc_short);
805 		counter_u64_zero(rack_sack_proc_restart);
806 		counter_u64_zero(rack_to_alloc);
807 		counter_u64_zero(rack_to_alloc_limited);
808 		counter_u64_zero(rack_alloc_limited_conns);
809 		counter_u64_zero(rack_split_limited);
810 		counter_u64_zero(rack_rxt_clamps_cwnd);
811 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
812 		counter_u64_zero(rack_multi_single_eq);
813 		counter_u64_zero(rack_proc_non_comp_ack);
814 		counter_u64_zero(rack_sack_attacks_detected);
815 		counter_u64_zero(rack_sack_attacks_reversed);
816 		counter_u64_zero(rack_sack_attacks_suspect);
817 		counter_u64_zero(rack_sack_used_next_merge);
818 		counter_u64_zero(rack_sack_used_prev_merge);
819 		counter_u64_zero(rack_sack_splits);
820 		counter_u64_zero(rack_sack_skipped_acked);
821 		counter_u64_zero(rack_ack_total);
822 		counter_u64_zero(rack_express_sack);
823 		counter_u64_zero(rack_sack_total);
824 		counter_u64_zero(rack_move_none);
825 		counter_u64_zero(rack_move_some);
826 		counter_u64_zero(rack_try_scwnd);
827 		counter_u64_zero(rack_collapsed_win);
828 		counter_u64_zero(rack_collapsed_win_rxt);
829 		counter_u64_zero(rack_collapsed_win_seen);
830 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
831 	} else if (stat == 2) {
832 #ifdef INVARIANTS
833 		printf("Clearing RACK option array\n");
834 #endif
835 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
836 	} else if (stat == 3) {
837 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
838 	} else if (stat == 4) {
839 #ifdef INVARIANTS
840 		printf("Clearing RACK out size array\n");
841 #endif
842 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
843 	}
844 	rack_clear_counter = 0;
845 	return (0);
846 }
847 
848 static void
849 rack_init_sysctls(void)
850 {
851 	struct sysctl_oid *rack_counters;
852 	struct sysctl_oid *rack_attack;
853 	struct sysctl_oid *rack_pacing;
854 	struct sysctl_oid *rack_timely;
855 	struct sysctl_oid *rack_timers;
856 	struct sysctl_oid *rack_tlp;
857 	struct sysctl_oid *rack_misc;
858 	struct sysctl_oid *rack_features;
859 	struct sysctl_oid *rack_measure;
860 	struct sysctl_oid *rack_probertt;
861 	struct sysctl_oid *rack_hw_pacing;
862 
863 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
864 	    SYSCTL_CHILDREN(rack_sysctl_root),
865 	    OID_AUTO,
866 	    "sack_attack",
867 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
868 	    "Rack Sack Attack Counters and Controls");
869 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
870 	    SYSCTL_CHILDREN(rack_sysctl_root),
871 	    OID_AUTO,
872 	    "stats",
873 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
874 	    "Rack Counters");
875 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
876 	    SYSCTL_CHILDREN(rack_sysctl_root),
877 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
878 	    &rack_rate_sample_method , USE_RTT_LOW,
879 	    "What method should we use for rate sampling 0=high, 1=low ");
880 	/* Probe rtt related controls */
881 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
882 	    SYSCTL_CHILDREN(rack_sysctl_root),
883 	    OID_AUTO,
884 	    "probertt",
885 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
886 	    "ProbeRTT related Controls");
887 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
888 	    SYSCTL_CHILDREN(rack_probertt),
889 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
890 	    &rack_atexit_prtt_hbp, 130,
891 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
892 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
893 	    SYSCTL_CHILDREN(rack_probertt),
894 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
895 	    &rack_atexit_prtt, 130,
896 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
897 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
898 	    SYSCTL_CHILDREN(rack_probertt),
899 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
900 	    &rack_per_of_gp_probertt, 60,
901 	    "What percentage of goodput do we pace at in probertt");
902 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
903 	    SYSCTL_CHILDREN(rack_probertt),
904 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
905 	    &rack_per_of_gp_probertt_reduce, 10,
906 	    "What percentage of goodput do we reduce every gp_srtt");
907 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
908 	    SYSCTL_CHILDREN(rack_probertt),
909 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
910 	    &rack_per_of_gp_lowthresh, 40,
911 	    "What percentage of goodput do we allow the multiplier to fall to");
912 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
913 	    SYSCTL_CHILDREN(rack_probertt),
914 	    OID_AUTO, "time_between", CTLFLAG_RW,
915 	    & rack_time_between_probertt, 96000000,
916 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
917 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
918 	    SYSCTL_CHILDREN(rack_probertt),
919 	    OID_AUTO, "safety", CTLFLAG_RW,
920 	    &rack_probe_rtt_safety_val, 2000000,
921 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
922 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
923 	    SYSCTL_CHILDREN(rack_probertt),
924 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
925 	    &rack_probe_rtt_sets_cwnd, 0,
926 	    "Do we set the cwnd too (if always_lower is on)");
927 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
928 	    SYSCTL_CHILDREN(rack_probertt),
929 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
930 	    &rack_max_drain_wait, 2,
931 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
932 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
933 	    SYSCTL_CHILDREN(rack_probertt),
934 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
935 	    &rack_must_drain, 1,
936 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
937 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
938 	    SYSCTL_CHILDREN(rack_probertt),
939 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
940 	    &rack_probertt_use_min_rtt_entry, 1,
941 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
942 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
943 	    SYSCTL_CHILDREN(rack_probertt),
944 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
945 	    &rack_probertt_use_min_rtt_exit, 0,
946 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
947 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
948 	    SYSCTL_CHILDREN(rack_probertt),
949 	    OID_AUTO, "length_div", CTLFLAG_RW,
950 	    &rack_probertt_gpsrtt_cnt_div, 0,
951 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
952 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
953 	    SYSCTL_CHILDREN(rack_probertt),
954 	    OID_AUTO, "length_mul", CTLFLAG_RW,
955 	    &rack_probertt_gpsrtt_cnt_mul, 0,
956 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
957 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
958 	    SYSCTL_CHILDREN(rack_probertt),
959 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
960 	    &rack_min_probertt_hold, 200000,
961 	    "What is the minimum time we hold probertt at target");
962 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
963 	    SYSCTL_CHILDREN(rack_probertt),
964 	    OID_AUTO, "filter_life", CTLFLAG_RW,
965 	    &rack_probertt_filter_life, 10000000,
966 	    "What is the time for the filters life in useconds");
967 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
968 	    SYSCTL_CHILDREN(rack_probertt),
969 	    OID_AUTO, "lower_within", CTLFLAG_RW,
970 	    &rack_probertt_lower_within, 10,
971 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
972 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
973 	    SYSCTL_CHILDREN(rack_probertt),
974 	    OID_AUTO, "must_move", CTLFLAG_RW,
975 	    &rack_min_rtt_movement, 250,
976 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
977 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
978 	    SYSCTL_CHILDREN(rack_probertt),
979 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
980 	    &rack_probertt_clear_is, 1,
981 	    "Do we clear I/S counts on exiting probe-rtt");
982 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
983 	    SYSCTL_CHILDREN(rack_probertt),
984 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
985 	    &rack_max_drain_hbp, 1,
986 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
987 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_probertt),
989 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
990 	    &rack_hbp_thresh, 3,
991 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
992 	/* Pacing related sysctls */
993 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
994 	    SYSCTL_CHILDREN(rack_sysctl_root),
995 	    OID_AUTO,
996 	    "pacing",
997 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
998 	    "Pacing related Controls");
999 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1000 	    SYSCTL_CHILDREN(rack_pacing),
1001 	    OID_AUTO, "fulldgpinrec", CTLFLAG_RW,
1002 	    &rack_uses_full_dgp_in_rec, 1,
1003 	    "Do we use all DGP features in recovery (fillcw, timely et.al.)?");
1004 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1005 	    SYSCTL_CHILDREN(rack_pacing),
1006 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1007 	    &rack_full_buffer_discount, 10,
1008 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1009 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1010 	    SYSCTL_CHILDREN(rack_pacing),
1011 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1012 	    &rack_fill_cw_state, 0,
1013 	    "Enable fillcw on new connections (default=0 off)?");
1014 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1015 	    SYSCTL_CHILDREN(rack_pacing),
1016 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1017 	    &rack_pacing_min_seg, 0,
1018 	    "What is the min burst size for pacing (0 disables)?");
1019 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1020 	    SYSCTL_CHILDREN(rack_pacing),
1021 	    OID_AUTO, "divisor", CTLFLAG_RW,
1022 	    &rack_default_pacing_divisor, 4,
1023 	    "What is the default divisor given to the rl code?");
1024 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1025 	    SYSCTL_CHILDREN(rack_pacing),
1026 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1027 	    &rack_bw_multipler, 2,
1028 	    "What is the multiplier of the current gp_est that fillcw can increase the b/w too?");
1029 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1030 	    SYSCTL_CHILDREN(rack_pacing),
1031 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1032 	    &rack_max_per_above, 30,
1033 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1034 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1035 	    SYSCTL_CHILDREN(rack_pacing),
1036 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1037 	    &rack_pace_one_seg, 0,
1038 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1039 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1040 	    SYSCTL_CHILDREN(rack_pacing),
1041 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1042 	    &rack_limit_time_with_srtt, 0,
1043 	    "Do we limit pacing time based on srtt");
1044 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1045 	    SYSCTL_CHILDREN(rack_pacing),
1046 	    OID_AUTO, "init_win", CTLFLAG_RW,
1047 	    &rack_default_init_window, 0,
1048 	    "Do we have a rack initial window 0 = system default");
1049 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_pacing),
1051 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1052 	    &rack_per_of_gp_ss, 250,
1053 	    "If non zero, what percentage of goodput to pace at in slow start");
1054 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1055 	    SYSCTL_CHILDREN(rack_pacing),
1056 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1057 	    &rack_per_of_gp_ca, 150,
1058 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1059 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1060 	    SYSCTL_CHILDREN(rack_pacing),
1061 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1062 	    &rack_per_of_gp_rec, 200,
1063 	    "If non zero, what percentage of goodput to pace at in recovery");
1064 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 	    SYSCTL_CHILDREN(rack_pacing),
1066 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1067 	    &rack_hptsi_segments, 40,
1068 	    "What size is the max for TSO segments in pacing and burst mitigation");
1069 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 	    SYSCTL_CHILDREN(rack_pacing),
1071 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1072 	    &rack_slot_reduction, 4,
1073 	    "When doing only burst mitigation what is the reduce divisor");
1074 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 	    SYSCTL_CHILDREN(rack_sysctl_root),
1076 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1077 	    &rack_pace_every_seg, 0,
1078 	    "If set we use pacing, if clear we use only the original burst mitigation");
1079 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1080 	    SYSCTL_CHILDREN(rack_pacing),
1081 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1082 	    &rack_bw_rate_cap, 0,
1083 	    "If set we apply this value to the absolute rate cap used by pacing");
1084 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1085 	    SYSCTL_CHILDREN(rack_sysctl_root),
1086 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1087 	    &rack_req_measurements, 1,
1088 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1089 	/* Hardware pacing */
1090 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1091 	    SYSCTL_CHILDREN(rack_sysctl_root),
1092 	    OID_AUTO,
1093 	    "hdwr_pacing",
1094 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1095 	    "Pacing related Controls");
1096 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1097 	    SYSCTL_CHILDREN(rack_hw_pacing),
1098 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1099 	    &rack_hw_rwnd_factor, 2,
1100 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1101 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1102 	    SYSCTL_CHILDREN(rack_hw_pacing),
1103 	    OID_AUTO, "precheck", CTLFLAG_RW,
1104 	    &rack_hw_check_queue, 0,
1105 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1106 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1107 	    SYSCTL_CHILDREN(rack_hw_pacing),
1108 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1109 	    &rack_enobuf_hw_boost_mult, 0,
1110 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1111 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1112 	    SYSCTL_CHILDREN(rack_hw_pacing),
1113 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1114 	    &rack_enobuf_hw_max, 2,
1115 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1116 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1117 	    SYSCTL_CHILDREN(rack_hw_pacing),
1118 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1119 	    &rack_enobuf_hw_min, 2,
1120 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1121 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1122 	    SYSCTL_CHILDREN(rack_hw_pacing),
1123 	    OID_AUTO, "enable", CTLFLAG_RW,
1124 	    &rack_enable_hw_pacing, 0,
1125 	    "Should RACK attempt to use hw pacing?");
1126 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1127 	    SYSCTL_CHILDREN(rack_hw_pacing),
1128 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1129 	    &rack_hw_rate_caps, 0,
1130 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1131 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1132 	    SYSCTL_CHILDREN(rack_hw_pacing),
1133 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1134 	    &rack_hw_rate_cap_per, 0,
1135 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1136 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1137 	    SYSCTL_CHILDREN(rack_hw_pacing),
1138 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1139 	    &rack_hw_rate_min, 0,
1140 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1141 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1142 	    SYSCTL_CHILDREN(rack_hw_pacing),
1143 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1144 	    &rack_hw_rate_to_low, 0,
1145 	    "If we fall below this rate, dis-engage hw pacing?");
1146 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1147 	    SYSCTL_CHILDREN(rack_hw_pacing),
1148 	    OID_AUTO, "up_only", CTLFLAG_RW,
1149 	    &rack_hw_up_only, 0,
1150 	    "Do we allow hw pacing to lower the rate selected?");
1151 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1152 	    SYSCTL_CHILDREN(rack_hw_pacing),
1153 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1154 	    &rack_hw_pace_extra_slots, 0,
1155 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1156 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1157 	    SYSCTL_CHILDREN(rack_sysctl_root),
1158 	    OID_AUTO,
1159 	    "timely",
1160 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1161 	    "Rack Timely RTT Controls");
1162 	/* Timely based GP dynmics */
1163 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1164 	    SYSCTL_CHILDREN(rack_timely),
1165 	    OID_AUTO, "upper", CTLFLAG_RW,
1166 	    &rack_gp_per_bw_mul_up, 2,
1167 	    "Rack timely upper range for equal b/w (in percentage)");
1168 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1169 	    SYSCTL_CHILDREN(rack_timely),
1170 	    OID_AUTO, "lower", CTLFLAG_RW,
1171 	    &rack_gp_per_bw_mul_down, 4,
1172 	    "Rack timely lower range for equal b/w (in percentage)");
1173 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1174 	    SYSCTL_CHILDREN(rack_timely),
1175 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1176 	    &rack_gp_rtt_maxmul, 3,
1177 	    "Rack timely multiplier of lowest rtt for rtt_max");
1178 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179 	    SYSCTL_CHILDREN(rack_timely),
1180 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1181 	    &rack_gp_rtt_mindiv, 4,
1182 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1183 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184 	    SYSCTL_CHILDREN(rack_timely),
1185 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1186 	    &rack_gp_rtt_minmul, 1,
1187 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1188 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189 	    SYSCTL_CHILDREN(rack_timely),
1190 	    OID_AUTO, "decrease", CTLFLAG_RW,
1191 	    &rack_gp_decrease_per, 80,
1192 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1193 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194 	    SYSCTL_CHILDREN(rack_timely),
1195 	    OID_AUTO, "increase", CTLFLAG_RW,
1196 	    &rack_gp_increase_per, 2,
1197 	    "Rack timely increase perentage of our GP multiplication factor");
1198 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199 	    SYSCTL_CHILDREN(rack_timely),
1200 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1201 	    &rack_per_lower_bound, 50,
1202 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1203 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1204 	    SYSCTL_CHILDREN(rack_timely),
1205 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1206 	    &rack_gain_p5_ub, 250,
1207 	    "Profile 5 upper bound to timely gain");
1208 
1209 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1210 	    SYSCTL_CHILDREN(rack_timely),
1211 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1212 	    &rack_per_upper_bound_ss, 0,
1213 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1214 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1215 	    SYSCTL_CHILDREN(rack_timely),
1216 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1217 	    &rack_per_upper_bound_ca, 0,
1218 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1219 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1220 	    SYSCTL_CHILDREN(rack_timely),
1221 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1222 	    &rack_do_dyn_mul, 0,
1223 	    "Rack timely do we enable dynmaic timely goodput by default");
1224 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1225 	    SYSCTL_CHILDREN(rack_timely),
1226 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1227 	    &rack_gp_no_rec_chg, 1,
1228 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1229 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1230 	    SYSCTL_CHILDREN(rack_timely),
1231 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1232 	    &rack_timely_dec_clear, 6,
1233 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1234 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1235 	    SYSCTL_CHILDREN(rack_timely),
1236 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1237 	    &rack_timely_max_push_rise, 3,
1238 	    "Rack timely how many times do we push up with b/w increase");
1239 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1240 	    SYSCTL_CHILDREN(rack_timely),
1241 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1242 	    &rack_timely_max_push_drop, 3,
1243 	    "Rack timely how many times do we push back on b/w decent");
1244 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1245 	    SYSCTL_CHILDREN(rack_timely),
1246 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1247 	    &rack_timely_min_segs, 4,
1248 	    "Rack timely when setting the cwnd what is the min num segments");
1249 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1250 	    SYSCTL_CHILDREN(rack_timely),
1251 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1252 	    &rack_use_max_for_nobackoff, 0,
1253 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1254 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1255 	    SYSCTL_CHILDREN(rack_timely),
1256 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1257 	    &rack_timely_int_timely_only, 0,
1258 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1259 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1260 	    SYSCTL_CHILDREN(rack_timely),
1261 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1262 	    &rack_timely_no_stopping, 0,
1263 	    "Rack timely don't stop increase");
1264 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1265 	    SYSCTL_CHILDREN(rack_timely),
1266 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1267 	    &rack_down_raise_thresh, 100,
1268 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1269 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1270 	    SYSCTL_CHILDREN(rack_timely),
1271 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1272 	    &rack_req_segs, 1,
1273 	    "Bottom dragging if not these many segments outstanding and room");
1274 
1275 	/* TLP and Rack related parameters */
1276 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1277 	    SYSCTL_CHILDREN(rack_sysctl_root),
1278 	    OID_AUTO,
1279 	    "tlp",
1280 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1281 	    "TLP and Rack related Controls");
1282 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(rack_tlp),
1284 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1285 	    &use_rack_rr, 1,
1286 	    "Do we use Rack Rapid Recovery");
1287 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(rack_tlp),
1289 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1290 	    &rack_max_abc_post_recovery, 2,
1291 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1292 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(rack_tlp),
1294 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1295 	    &rack_non_rxt_use_cr, 0,
1296 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1297 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(rack_tlp),
1299 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1300 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1301 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1302 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1303 	    SYSCTL_CHILDREN(rack_tlp),
1304 	    OID_AUTO, "limit", CTLFLAG_RW,
1305 	    &rack_tlp_limit, 2,
1306 	    "How many TLP's can be sent without sending new data");
1307 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1308 	    SYSCTL_CHILDREN(rack_tlp),
1309 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1310 	    &rack_tlp_use_greater, 1,
1311 	    "Should we use the rack_rtt time if its greater than srtt");
1312 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(rack_tlp),
1314 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1315 	    &rack_tlp_min, 10000,
1316 	    "TLP minimum timeout per the specification (in microseconds)");
1317 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1318 	    SYSCTL_CHILDREN(rack_tlp),
1319 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1320 	    &rack_always_send_oldest, 0,
1321 	    "Should we always send the oldest TLP and RACK-TLP");
1322 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1323 	    SYSCTL_CHILDREN(rack_tlp),
1324 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1325 	    &rack_limited_retran, 0,
1326 	    "How many times can a rack timeout drive out sends");
1327 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1328 	    SYSCTL_CHILDREN(rack_tlp),
1329 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1330 	    &rack_lower_cwnd_at_tlp, 0,
1331 	    "When a TLP completes a retran should we enter recovery");
1332 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1333 	    SYSCTL_CHILDREN(rack_tlp),
1334 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1335 	    &rack_reorder_thresh, 2,
1336 	    "What factor for rack will be added when seeing reordering (shift right)");
1337 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1338 	    SYSCTL_CHILDREN(rack_tlp),
1339 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1340 	    &rack_tlp_thresh, 1,
1341 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1342 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1343 	    SYSCTL_CHILDREN(rack_tlp),
1344 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1345 	    &rack_reorder_fade, 60000000,
1346 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1347 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1348 	    SYSCTL_CHILDREN(rack_tlp),
1349 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1350 	    &rack_pkt_delay, 1000,
1351 	    "Extra RACK time (in microseconds) besides reordering thresh");
1352 
1353 	/* Timer related controls */
1354 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(rack_sysctl_root),
1356 	    OID_AUTO,
1357 	    "timers",
1358 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1359 	    "Timer related controls");
1360 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_timers),
1362 	    OID_AUTO, "persmin", CTLFLAG_RW,
1363 	    &rack_persist_min, 250000,
1364 	    "What is the minimum time in microseconds between persists");
1365 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1366 	    SYSCTL_CHILDREN(rack_timers),
1367 	    OID_AUTO, "persmax", CTLFLAG_RW,
1368 	    &rack_persist_max, 2000000,
1369 	    "What is the largest delay in microseconds between persists");
1370 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1371 	    SYSCTL_CHILDREN(rack_timers),
1372 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1373 	    &rack_delayed_ack_time, 40000,
1374 	    "Delayed ack time (40ms in microseconds)");
1375 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1376 	    SYSCTL_CHILDREN(rack_timers),
1377 	    OID_AUTO, "minrto", CTLFLAG_RW,
1378 	    &rack_rto_min, 30000,
1379 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1380 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1381 	    SYSCTL_CHILDREN(rack_timers),
1382 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1383 	    &rack_rto_max, 4000000,
1384 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1385 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1386 	    SYSCTL_CHILDREN(rack_timers),
1387 	    OID_AUTO, "minto", CTLFLAG_RW,
1388 	    &rack_min_to, 1000,
1389 	    "Minimum rack timeout in microseconds");
1390 	/* Measure controls */
1391 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1392 	    SYSCTL_CHILDREN(rack_sysctl_root),
1393 	    OID_AUTO,
1394 	    "measure",
1395 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1396 	    "Measure related controls");
1397 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398 	    SYSCTL_CHILDREN(rack_measure),
1399 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1400 	    &rack_wma_divisor, 8,
1401 	    "When doing b/w calculation what is the  divisor for the WMA");
1402 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_measure),
1404 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1405 	    &rack_cwnd_block_ends_measure, 0,
1406 	    "Does a cwnd just-return end the measurement window (app limited)");
1407 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1408 	    SYSCTL_CHILDREN(rack_measure),
1409 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1410 	    &rack_rwnd_block_ends_measure, 0,
1411 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1412 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1413 	    SYSCTL_CHILDREN(rack_measure),
1414 	    OID_AUTO, "min_target", CTLFLAG_RW,
1415 	    &rack_def_data_window, 20,
1416 	    "What is the minimum target window (in mss) for a GP measurements");
1417 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1418 	    SYSCTL_CHILDREN(rack_measure),
1419 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1420 	    &rack_goal_bdp, 2,
1421 	    "What is the goal BDP to measure");
1422 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1423 	    SYSCTL_CHILDREN(rack_measure),
1424 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1425 	    &rack_min_srtts, 1,
1426 	    "What is the goal BDP to measure");
1427 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1428 	    SYSCTL_CHILDREN(rack_measure),
1429 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1430 	    &rack_min_measure_usec, 0,
1431 	    "What is the Minimum time time for a measurement if 0, this is off");
1432 	/* Features */
1433 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1434 	    SYSCTL_CHILDREN(rack_sysctl_root),
1435 	    OID_AUTO,
1436 	    "features",
1437 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1438 	    "Feature controls");
1439 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1440 	    SYSCTL_CHILDREN(rack_features),
1441 	    OID_AUTO, "rxt_clamp_thresh", CTLFLAG_RW,
1442 	    &rack_rxt_clamp_thresh, 0,
1443 	    "Bit encoded clamping setup bits CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP");
1444 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1445 	    SYSCTL_CHILDREN(rack_features),
1446 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1447 	    &rack_hybrid_allow_set_maxseg, 0,
1448 	    "Should hybrid pacing allow the setmss command");
1449 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1450 	    SYSCTL_CHILDREN(rack_features),
1451 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1452 	    &rack_use_cmp_acks, 1,
1453 	    "Should RACK have LRO send compressed acks");
1454 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1455 	    SYSCTL_CHILDREN(rack_features),
1456 	    OID_AUTO, "fsb", CTLFLAG_RW,
1457 	    &rack_use_fsb, 1,
1458 	    "Should RACK use the fast send block?");
1459 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1460 	    SYSCTL_CHILDREN(rack_features),
1461 	    OID_AUTO, "rfo", CTLFLAG_RW,
1462 	    &rack_use_rfo, 1,
1463 	    "Should RACK use rack_fast_output()?");
1464 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1465 	    SYSCTL_CHILDREN(rack_features),
1466 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1467 	    &rack_use_rsm_rfo, 1,
1468 	    "Should RACK use rack_fast_rsm_output()?");
1469 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1470 	    SYSCTL_CHILDREN(rack_features),
1471 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1472 	    &rack_enable_mqueue_for_nonpaced, 0,
1473 	    "Should RACK use mbuf queuing for non-paced connections");
1474 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1475 	    SYSCTL_CHILDREN(rack_features),
1476 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1477 	    &rack_do_hystart, 0,
1478 	    "Should RACK enable HyStart++ on connections?");
1479 	/* Misc rack controls */
1480 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1481 	    SYSCTL_CHILDREN(rack_sysctl_root),
1482 	    OID_AUTO,
1483 	    "misc",
1484 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1485 	    "Misc related controls");
1486 #ifdef TCP_ACCOUNTING
1487 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1488 	    SYSCTL_CHILDREN(rack_misc),
1489 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1490 	    &rack_tcp_accounting, 0,
1491 	    "Should we turn on TCP accounting for all rack sessions?");
1492 #endif
1493 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1494 	    SYSCTL_CHILDREN(rack_misc),
1495 	    OID_AUTO, "dnd", CTLFLAG_RW,
1496 	    &rack_dnd_default, 0,
1497 	    "Do not disturb default for rack_rrr = 3");
1498 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1499 	    SYSCTL_CHILDREN(rack_misc),
1500 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1501 	    &sad_seg_size_per, 800,
1502 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1503 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1504 	    SYSCTL_CHILDREN(rack_misc),
1505 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1506 	    &rack_rxt_controls, 0,
1507 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1508 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1509 	    SYSCTL_CHILDREN(rack_misc),
1510 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1511 	    &rack_hibeta_setting, 0,
1512 	    "Do we ue a high beta (80 instead of 50)?");
1513 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1514 	    SYSCTL_CHILDREN(rack_misc),
1515 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1516 	    &rack_apply_rtt_with_reduced_conf, 0,
1517 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1518 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1519 	    SYSCTL_CHILDREN(rack_misc),
1520 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1521 	    &rack_dsack_std_based, 3,
1522 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1523 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1524 	    SYSCTL_CHILDREN(rack_misc),
1525 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1526 	    &rack_prr_addbackmax, 2,
1527 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1528 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1529 	    SYSCTL_CHILDREN(rack_misc),
1530 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1531 	    &rack_stats_gets_ms_rtt, 1,
1532 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1533 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1534 	    SYSCTL_CHILDREN(rack_misc),
1535 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1536 	    &rack_client_low_buf, 0,
1537 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1538 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1539 	    SYSCTL_CHILDREN(rack_misc),
1540 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1541 	    &rack_def_profile, 0,
1542 	    "Should RACK use a default profile (0=no, num == profile num)?");
1543 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1544 	    SYSCTL_CHILDREN(rack_misc),
1545 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1546 	    &rack_enable_shared_cwnd, 1,
1547 	    "Should RACK try to use the shared cwnd on connections where allowed");
1548 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1549 	    SYSCTL_CHILDREN(rack_misc),
1550 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1551 	    &rack_limits_scwnd, 1,
1552 	    "Should RACK place low end time limits on the shared cwnd feature");
1553 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1554 	    SYSCTL_CHILDREN(rack_misc),
1555 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1556 	    &rack_disable_prr, 0,
1557 	    "Should RACK not use prr and only pace (must have pacing on)");
1558 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1559 	    SYSCTL_CHILDREN(rack_misc),
1560 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1561 	    &rack_verbose_logging, 0,
1562 	    "Should RACK black box logging be verbose");
1563 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1564 	    SYSCTL_CHILDREN(rack_misc),
1565 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1566 	    &rack_ignore_data_after_close, 1,
1567 	    "Do we hold off sending a RST until all pending data is ack'd");
1568 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1569 	    SYSCTL_CHILDREN(rack_misc),
1570 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1571 	    &rack_sack_not_required, 1,
1572 	    "Do we allow rack to run on connections not supporting SACK");
1573 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1574 	    SYSCTL_CHILDREN(rack_misc),
1575 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1576 	    &rack_send_a_lot_in_prr, 1,
1577 	    "Send a lot in prr");
1578 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1579 	    SYSCTL_CHILDREN(rack_misc),
1580 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1581 	    &rack_autosndbuf_inc, 20,
1582 	    "What percentage should rack scale up its snd buffer by?");
1583 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1584 	    SYSCTL_CHILDREN(rack_misc),
1585 	    OID_AUTO, "rnds_for_rxt_clamp", CTLFLAG_RW,
1586 	    &rack_rxt_min_rnds, 10,
1587 	    "Number of rounds needed between RTT clamps due to high loss rates");
1588 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1589 	    SYSCTL_CHILDREN(rack_misc),
1590 	    OID_AUTO, "rnds_for_unclamp", CTLFLAG_RW,
1591 	    &rack_unclamp_round_thresh, 100,
1592 	    "Number of rounds needed with no loss to unclamp");
1593 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1594 	    SYSCTL_CHILDREN(rack_misc),
1595 	    OID_AUTO, "rxt_threshs_for_unclamp", CTLFLAG_RW,
1596 	    &rack_unclamp_rxt_thresh, 5,
1597 	   "Percentage of retransmits we need to be under to unclamp (5 = .5 percent)\n");
1598 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1599 	    SYSCTL_CHILDREN(rack_misc),
1600 	    OID_AUTO, "clamp_ss_upper", CTLFLAG_RW,
1601 	    &rack_clamp_ss_upper, 110,
1602 	    "Clamp percentage ceiling in SS?");
1603 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1604 	    SYSCTL_CHILDREN(rack_misc),
1605 	    OID_AUTO, "clamp_ca_upper", CTLFLAG_RW,
1606 	    &rack_clamp_ca_upper, 110,
1607 	    "Clamp percentage ceiling in CA?");
1608 	/* Sack Attacker detection stuff */
1609 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1610 	    SYSCTL_CHILDREN(rack_attack),
1611 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1612 	    &rack_merge_out_sacks_on_attack, 0,
1613 	    "Do we merge the sendmap when we decide we are being attacked?");
1614 
1615 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1616 	    SYSCTL_CHILDREN(rack_attack),
1617 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1618 	    &rack_highest_sack_thresh_seen, 0,
1619 	    "Highest sack to ack ratio seen");
1620 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1621 	    SYSCTL_CHILDREN(rack_attack),
1622 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1623 	    &rack_highest_move_thresh_seen, 0,
1624 	    "Highest move to non-move ratio seen");
1625 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1626 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1627 	    SYSCTL_CHILDREN(rack_attack),
1628 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1629 	    &rack_ack_total,
1630 	    "Total number of Ack's");
1631 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1632 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1633 	    SYSCTL_CHILDREN(rack_attack),
1634 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1635 	    &rack_express_sack,
1636 	    "Total expresss number of Sack's");
1637 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1638 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1639 	    SYSCTL_CHILDREN(rack_attack),
1640 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1641 	    &rack_sack_total,
1642 	    "Total number of SACKs");
1643 	rack_move_none = counter_u64_alloc(M_WAITOK);
1644 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1645 	    SYSCTL_CHILDREN(rack_attack),
1646 	    OID_AUTO, "move_none", CTLFLAG_RD,
1647 	    &rack_move_none,
1648 	    "Total number of SACK index reuse of positions under threshold");
1649 	rack_move_some = counter_u64_alloc(M_WAITOK);
1650 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1651 	    SYSCTL_CHILDREN(rack_attack),
1652 	    OID_AUTO, "move_some", CTLFLAG_RD,
1653 	    &rack_move_some,
1654 	    "Total number of SACK index reuse of positions over threshold");
1655 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1656 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1657 	    SYSCTL_CHILDREN(rack_attack),
1658 	    OID_AUTO, "attacks", CTLFLAG_RD,
1659 	    &rack_sack_attacks_detected,
1660 	    "Total number of SACK attackers that had sack disabled");
1661 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1662 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1663 	    SYSCTL_CHILDREN(rack_attack),
1664 	    OID_AUTO, "reversed", CTLFLAG_RD,
1665 	    &rack_sack_attacks_reversed,
1666 	    "Total number of SACK attackers that were later determined false positive");
1667 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1668 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1669 	    SYSCTL_CHILDREN(rack_attack),
1670 	    OID_AUTO, "suspect", CTLFLAG_RD,
1671 	    &rack_sack_attacks_suspect,
1672 	    "Total number of SACKs that triggered early detection");
1673 
1674 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1675 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1676 	    SYSCTL_CHILDREN(rack_attack),
1677 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1678 	    &rack_sack_used_next_merge,
1679 	    "Total number of times we used the next merge");
1680 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1681 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1682 	    SYSCTL_CHILDREN(rack_attack),
1683 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1684 	    &rack_sack_used_prev_merge,
1685 	    "Total number of times we used the prev merge");
1686 	/* Counters */
1687 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1688 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1689 	    SYSCTL_CHILDREN(rack_counters),
1690 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1691 	    &rack_total_bytes,
1692 	    "Total number of bytes sent");
1693 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1694 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1695 	    SYSCTL_CHILDREN(rack_counters),
1696 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1697 	    &rack_fto_send, "Total number of rack_fast_output sends");
1698 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1699 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1700 	    SYSCTL_CHILDREN(rack_counters),
1701 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1702 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1703 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1704 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1705 	    SYSCTL_CHILDREN(rack_counters),
1706 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1707 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1708 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1709 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1710 	    SYSCTL_CHILDREN(rack_counters),
1711 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1712 	    &rack_non_fto_send, "Total number of rack_output first sends");
1713 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1714 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1715 	    SYSCTL_CHILDREN(rack_counters),
1716 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1717 	    &rack_extended_rfo, "Total number of times we extended rfo");
1718 
1719 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1720 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1721 	    SYSCTL_CHILDREN(rack_counters),
1722 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1723 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1724 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1725 
1726 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1727 	    SYSCTL_CHILDREN(rack_counters),
1728 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1729 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1730 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1731 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732 	    SYSCTL_CHILDREN(rack_counters),
1733 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1734 	    &rack_tlp_tot,
1735 	    "Total number of tail loss probe expirations");
1736 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1737 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 	    SYSCTL_CHILDREN(rack_counters),
1739 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1740 	    &rack_tlp_newdata,
1741 	    "Total number of tail loss probe sending new data");
1742 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_counters),
1745 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1746 	    &rack_tlp_retran,
1747 	    "Total number of tail loss probe sending retransmitted data");
1748 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_counters),
1751 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1752 	    &rack_tlp_retran_bytes,
1753 	    "Total bytes of tail loss probe sending retransmitted data");
1754 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1755 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 	    SYSCTL_CHILDREN(rack_counters),
1757 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1758 	    &rack_to_tot,
1759 	    "Total number of times the rack to expired");
1760 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1761 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762 	    SYSCTL_CHILDREN(rack_counters),
1763 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1764 	    &rack_saw_enobuf,
1765 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1766 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1767 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1768 	    SYSCTL_CHILDREN(rack_counters),
1769 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1770 	    &rack_saw_enobuf_hw,
1771 	    "Total number of times a send returned enobuf for hdwr paced connections");
1772 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1773 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1774 	    SYSCTL_CHILDREN(rack_counters),
1775 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1776 	    &rack_saw_enetunreach,
1777 	    "Total number of times a send received a enetunreachable");
1778 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1779 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1780 	    SYSCTL_CHILDREN(rack_counters),
1781 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1782 	    &rack_hot_alloc,
1783 	    "Total allocations from the top of our list");
1784 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1785 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1786 	    SYSCTL_CHILDREN(rack_counters),
1787 	    OID_AUTO, "allocs", CTLFLAG_RD,
1788 	    &rack_to_alloc,
1789 	    "Total allocations of tracking structures");
1790 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1791 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1792 	    SYSCTL_CHILDREN(rack_counters),
1793 	    OID_AUTO, "allochard", CTLFLAG_RD,
1794 	    &rack_to_alloc_hard,
1795 	    "Total allocations done with sleeping the hard way");
1796 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1797 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1798 	    SYSCTL_CHILDREN(rack_counters),
1799 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1800 	    &rack_to_alloc_emerg,
1801 	    "Total allocations done from emergency cache");
1802 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1803 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804 	    SYSCTL_CHILDREN(rack_counters),
1805 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1806 	    &rack_to_alloc_limited,
1807 	    "Total allocations dropped due to limit");
1808 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1809 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1810 	    SYSCTL_CHILDREN(rack_counters),
1811 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1812 	    &rack_alloc_limited_conns,
1813 	    "Connections with allocations dropped due to limit");
1814 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1815 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1816 	    SYSCTL_CHILDREN(rack_counters),
1817 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1818 	    &rack_split_limited,
1819 	    "Split allocations dropped due to limit");
1820 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1821 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1822 	    SYSCTL_CHILDREN(rack_counters),
1823 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1824 	    &rack_rxt_clamps_cwnd,
1825 	    "Number of times that excessive rxt clamped the cwnd down");
1826 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1827 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1828 	    SYSCTL_CHILDREN(rack_counters),
1829 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1830 	    &rack_rxt_clamps_cwnd_uniq,
1831 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1832 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1833 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1834 	    SYSCTL_CHILDREN(rack_counters),
1835 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1836 	    &rack_persists_sends,
1837 	    "Number of times we sent a persist probe");
1838 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1839 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1840 	    SYSCTL_CHILDREN(rack_counters),
1841 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1842 	    &rack_persists_acks,
1843 	    "Number of times a persist probe was acked");
1844 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1845 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1846 	    SYSCTL_CHILDREN(rack_counters),
1847 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1848 	    &rack_persists_loss,
1849 	    "Number of times we detected a lost persist probe (no ack)");
1850 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1851 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1852 	    SYSCTL_CHILDREN(rack_counters),
1853 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1854 	    &rack_persists_lost_ends,
1855 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1856 #ifdef INVARIANTS
1857 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1858 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1859 	    SYSCTL_CHILDREN(rack_counters),
1860 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1861 	    &rack_adjust_map_bw,
1862 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1863 #endif
1864 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1865 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1866 	    SYSCTL_CHILDREN(rack_counters),
1867 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1868 	    &rack_multi_single_eq,
1869 	    "Number of compressed acks total represented");
1870 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1871 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1872 	    SYSCTL_CHILDREN(rack_counters),
1873 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1874 	    &rack_proc_non_comp_ack,
1875 	    "Number of non compresseds acks that we processed");
1876 
1877 
1878 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1879 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1880 	    SYSCTL_CHILDREN(rack_counters),
1881 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1882 	    &rack_sack_proc_all,
1883 	    "Total times we had to walk whole list for sack processing");
1884 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1885 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1886 	    SYSCTL_CHILDREN(rack_counters),
1887 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1888 	    &rack_sack_proc_restart,
1889 	    "Total times we had to walk whole list due to a restart");
1890 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1891 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1892 	    SYSCTL_CHILDREN(rack_counters),
1893 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1894 	    &rack_sack_proc_short,
1895 	    "Total times we took shortcut for sack processing");
1896 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1897 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1898 	    SYSCTL_CHILDREN(rack_attack),
1899 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1900 	    &rack_sack_skipped_acked,
1901 	    "Total number of times we skipped previously sacked");
1902 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1903 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1904 	    SYSCTL_CHILDREN(rack_attack),
1905 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1906 	    &rack_sack_splits,
1907 	    "Total number of times we did the old fashion tree split");
1908 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1909 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1910 	    SYSCTL_CHILDREN(rack_counters),
1911 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1912 	    &rack_input_idle_reduces,
1913 	    "Total number of idle reductions on input");
1914 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1915 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1916 	    SYSCTL_CHILDREN(rack_counters),
1917 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1918 	    &rack_collapsed_win_seen,
1919 	    "Total number of collapsed window events seen (where our window shrinks)");
1920 
1921 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1922 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1923 	    SYSCTL_CHILDREN(rack_counters),
1924 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1925 	    &rack_collapsed_win,
1926 	    "Total number of collapsed window events where we mark packets");
1927 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1928 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1929 	    SYSCTL_CHILDREN(rack_counters),
1930 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1931 	    &rack_collapsed_win_rxt,
1932 	    "Total number of packets that were retransmitted");
1933 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1934 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1935 	    SYSCTL_CHILDREN(rack_counters),
1936 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1937 	    &rack_collapsed_win_rxt_bytes,
1938 	    "Total number of bytes that were retransmitted");
1939 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1940 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1941 	    SYSCTL_CHILDREN(rack_counters),
1942 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1943 	    &rack_try_scwnd,
1944 	    "Total number of scwnd attempts");
1945 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1946 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1947 	    OID_AUTO, "outsize", CTLFLAG_RD,
1948 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1949 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1950 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1951 	    OID_AUTO, "opts", CTLFLAG_RD,
1952 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1953 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1954 	    SYSCTL_CHILDREN(rack_sysctl_root),
1955 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1956 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1957 }
1958 
1959 static uint32_t
1960 rc_init_window(struct tcp_rack *rack)
1961 {
1962 	uint32_t win;
1963 
1964 	if (rack->rc_init_win == 0) {
1965 		/*
1966 		 * Nothing set by the user, use the system stack
1967 		 * default.
1968 		 */
1969 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1970 	}
1971 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1972 	return (win);
1973 }
1974 
1975 static uint64_t
1976 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1977 {
1978 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1979 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1980 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1981 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1982 	else
1983 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1984 }
1985 
1986 static void
1987 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1988 	uint64_t data, uint8_t mod, uint16_t aux,
1989 	struct tcp_sendfile_track *cur, int line)
1990 {
1991 #ifdef TCP_REQUEST_TRK
1992 	int do_log = 0;
1993 
1994 	/*
1995 	 * The rate cap one is noisy and only should come out when normal BB logging
1996 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
1997 	 * once per chunk and make up the BBpoint that can be turned on by the client.
1998 	 */
1999 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2000 		/*
2001 		 * The very noisy two need to only come out when
2002 		 * we have verbose logging on.
2003 		 */
2004 		if (rack_verbose_logging != 0)
2005 			do_log = tcp_bblogging_on(rack->rc_tp);
2006 		else
2007 			do_log = 0;
2008 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2009 		/*
2010 		 * All other less noisy logs here except the measure which
2011 		 * also needs to come out on the point and the log.
2012 		 */
2013 		do_log = tcp_bblogging_on(rack->rc_tp);
2014 	} else {
2015 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2016 	}
2017 
2018 	if (do_log) {
2019 		union tcp_log_stackspecific log;
2020 		struct timeval tv;
2021 		uint64_t lt_bw;
2022 
2023 		/* Convert our ms to a microsecond */
2024 		memset(&log, 0, sizeof(log));
2025 
2026 		log.u_bbr.cwnd_gain = line;
2027 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2028 		log.u_bbr.rttProp = tim;
2029 		log.u_bbr.bw_inuse = cbw;
2030 		log.u_bbr.delRate = rack_get_gp_est(rack);
2031 		lt_bw = rack_get_lt_bw(rack);
2032 		log.u_bbr.flex1 = seq;
2033 		log.u_bbr.pacing_gain = aux;
2034 		/* lt_bw = < flex3 | flex2 > */
2035 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2036 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2037 		/* Record the last obtained us rtt in inflight */
2038 		if (cur == NULL) {
2039 			/* Make sure we are looking at the right log if an overide comes in */
2040 			cur = rack->r_ctl.rc_last_sft;
2041 		}
2042 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2043 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2044 		else {
2045 			/* Use the last known rtt i.e. the rack-rtt */
2046 			log.u_bbr.inflight = rack->rc_rack_rtt;
2047 		}
2048 		if (cur != NULL) {
2049 			uint64_t off;
2050 
2051 			log.u_bbr.cur_del_rate = cur->deadline;
2052 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2053 				/* start = < lost | pkt_epoch > */
2054 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2055 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2056 				log.u_bbr.flex6 = cur->start_seq;
2057 				log.u_bbr.pkts_out = cur->end_seq;
2058 			} else {
2059 				/* start = < lost | pkt_epoch > */
2060 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2061 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2062 				/* end = < pkts_out | flex6 > */
2063 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2064 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2065 			}
2066 			/* first_send = <lt_epoch | epoch> */
2067 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2068 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2069 			/* localtime = <delivered | applimited>*/
2070 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2071 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2072 #ifdef TCP_REQUEST_TRK
2073 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2074 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2075 #endif
2076 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2077 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2078 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2079 		} else {
2080 			log.u_bbr.flex7 = 0xffff;
2081 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2082 		}
2083 		/*
2084 		 * Compose bbr_state to be a bit wise 0000ADHF
2085 		 * where A is the always_pace flag
2086 		 * where D is the dgp_on flag
2087 		 * where H is the hybrid_mode on flag
2088 		 * where F is the use_fixed_rate flag.
2089 		 */
2090 		log.u_bbr.bbr_state = rack->rc_always_pace;
2091 		log.u_bbr.bbr_state <<= 1;
2092 		log.u_bbr.bbr_state |= rack->dgp_on;
2093 		log.u_bbr.bbr_state <<= 1;
2094 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2095 		log.u_bbr.bbr_state <<= 1;
2096 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2097 		log.u_bbr.flex8 = mod;
2098 		tcp_log_event(rack->rc_tp, NULL,
2099 		    &rack->rc_inp->inp_socket->so_rcv,
2100 		    &rack->rc_inp->inp_socket->so_snd,
2101 		    TCP_HYBRID_PACING_LOG, 0,
2102 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2103 
2104 	}
2105 #endif
2106 }
2107 
2108 #ifdef TCP_REQUEST_TRK
2109 static void
2110 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2111 {
2112 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2113 		union tcp_log_stackspecific log;
2114 		struct timeval tv;
2115 		uint64_t off;
2116 
2117 		/* Convert our ms to a microsecond */
2118 		memset(&log, 0, sizeof(log));
2119 
2120 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2121 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2122 		log.u_bbr.delRate = cur->sent_at_fs;
2123 		log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2124 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2125 		log.u_bbr.cwnd_gain = line;
2126 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2127 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2128 		/* start = < flex1 | flex2 > */
2129 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2130 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2131 		/* end = < flex3 | flex4 > */
2132 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2133 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2134 
2135 		/* localtime = <delivered | applimited>*/
2136 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2137 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2138 		/* client timestamp = <lt_epoch | epoch>*/
2139 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2140 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2141 		/* now set all the flags in */
2142 		log.u_bbr.pkts_out = cur->hybrid_flags;
2143 		log.u_bbr.flex6 = cur->flags;
2144 		/*
2145 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2146 		 * where a false retransmit occurred so first_send  <-> lastsend may
2147 		 * include longer time then it actually took if we have a false rxt.
2148 		 */
2149 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2150 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2151 
2152 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2153 		tcp_log_event(rack->rc_tp, NULL,
2154 		    &rack->rc_inp->inp_socket->so_rcv,
2155 		    &rack->rc_inp->inp_socket->so_snd,
2156 		    TCP_HYBRID_PACING_LOG, 0,
2157 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2158 	}
2159 }
2160 #endif
2161 
2162 static inline uint64_t
2163 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2164 {
2165 	uint64_t ret_bw, ether;
2166 	uint64_t u_segsiz;
2167 
2168 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2169 	if (rack->r_is_v6){
2170 #ifdef INET6
2171 		ether += sizeof(struct ip6_hdr);
2172 #endif
2173 		ether += 14;	/* eheader size 6+6+2 */
2174 	} else {
2175 #ifdef INET
2176 		ether += sizeof(struct ip);
2177 #endif
2178 		ether += 14;	/* eheader size 6+6+2 */
2179 	}
2180 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2181 	ret_bw = bw;
2182 	ret_bw *= ether;
2183 	ret_bw /= u_segsiz;
2184 	return (ret_bw);
2185 }
2186 
2187 static void
2188 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2189 {
2190 #ifdef TCP_REQUEST_TRK
2191 	struct timeval tv;
2192 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2193 #endif
2194 
2195 	if (rack->r_ctl.bw_rate_cap == 0)
2196 		return;
2197 #ifdef TCP_REQUEST_TRK
2198 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2199 	    (rack->r_ctl.rc_last_sft != NULL)) {
2200 		/*
2201 		 * We have a dynamic cap. The original target
2202 		 * is in bw_rate_cap, but we need to look at
2203 		 * how long it is until we hit the deadline.
2204 		 */
2205 		struct tcp_sendfile_track *ent;
2206 
2207       		ent = rack->r_ctl.rc_last_sft;
2208 		microuptime(&tv);
2209 		timenow = tcp_tv_to_lusectick(&tv);
2210 		if (timenow >= ent->deadline) {
2211 			/* No time left we do DGP only */
2212 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2213 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2214 			rack->r_ctl.bw_rate_cap = 0;
2215 			return;
2216 		}
2217 		/* We have the time */
2218 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2219 		if (timeleft < HPTS_MSEC_IN_SEC) {
2220 			/* If there is less than a ms left just use DGPs rate */
2221 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2222 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2223 			rack->r_ctl.bw_rate_cap = 0;
2224 			return;
2225 		}
2226 		/*
2227 		 * Now lets find the amount of data left to send.
2228 		 *
2229 		 * Now ideally we want to use the end_seq to figure out how much more
2230 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2231 		 */
2232 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2233 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2234 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2235 			else {
2236 				/* TSNH, we should catch it at the send */
2237 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2238 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2239 				rack->r_ctl.bw_rate_cap = 0;
2240 				return;
2241 			}
2242 		} else {
2243 			/*
2244 			 * The hard way, figure out how much is gone and then
2245 			 * take that away from the total the client asked for
2246 			 * (thats off by tls overhead if this is tls).
2247 			 */
2248 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2249 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2250 			else
2251 				lengone = 0;
2252 			if (lengone < (ent->end - ent->start))
2253 				lenleft = (ent->end - ent->start) - lengone;
2254 			else {
2255 				/* TSNH, we should catch it at the send */
2256 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2257 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2258 				rack->r_ctl.bw_rate_cap = 0;
2259 				return;
2260 			}
2261 		}
2262 		if (lenleft == 0) {
2263 			/* We have it all sent */
2264 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2265 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2266 			if (rack->r_ctl.bw_rate_cap)
2267 				goto normal_ratecap;
2268 			else
2269 				return;
2270 		}
2271 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2272 		calcbw /= timeleft;
2273 		/* Now we must compensate for IP/TCP overhead */
2274 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2275 		/* Update the bit rate cap */
2276 		rack->r_ctl.bw_rate_cap = calcbw;
2277 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2278 		    (rack_hybrid_allow_set_maxseg == 1) &&
2279 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2280 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2281 			uint32_t orig_max;
2282 
2283 			orig_max = rack->r_ctl.rc_pace_max_segs;
2284 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2285 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2286 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2287 		}
2288 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2289 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2290 		if ((calcbw > 0) && (*bw > calcbw)) {
2291 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2292 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2293 			*capped = 1;
2294 			*bw = calcbw;
2295 		}
2296 		return;
2297 	}
2298 normal_ratecap:
2299 #endif
2300 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2301 #ifdef TCP_REQUEST_TRK
2302 		if (rack->rc_hybrid_mode &&
2303 		    rack->rc_catch_up &&
2304 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2305 		    (rack_hybrid_allow_set_maxseg == 1) &&
2306 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2307 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2308 			uint32_t orig_max;
2309 
2310 			orig_max = rack->r_ctl.rc_pace_max_segs;
2311 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2312 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2313 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2314 		}
2315 #endif
2316 		*capped = 1;
2317 		*bw = rack->r_ctl.bw_rate_cap;
2318 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2319 				   *bw, 0, 0,
2320 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2321 	}
2322 }
2323 
2324 static uint64_t
2325 rack_get_gp_est(struct tcp_rack *rack)
2326 {
2327 	uint64_t bw, lt_bw, ret_bw;
2328 
2329 	if (rack->rc_gp_filled == 0) {
2330 		/*
2331 		 * We have yet no b/w measurement,
2332 		 * if we have a user set initial bw
2333 		 * return it. If we don't have that and
2334 		 * we have an srtt, use the tcp IW (10) to
2335 		 * calculate a fictional b/w over the SRTT
2336 		 * which is more or less a guess. Note
2337 		 * we don't use our IW from rack on purpose
2338 		 * so if we have like IW=30, we are not
2339 		 * calculating a "huge" b/w.
2340 		 */
2341 		uint64_t srtt;
2342 
2343 		lt_bw = rack_get_lt_bw(rack);
2344 		if (lt_bw) {
2345 			/*
2346 			 * No goodput bw but a long-term b/w does exist
2347 			 * lets use that.
2348 			 */
2349 			ret_bw = lt_bw;
2350 			goto compensate;
2351 		}
2352 		if (rack->r_ctl.init_rate)
2353 			return (rack->r_ctl.init_rate);
2354 
2355 		/* Ok lets come up with the IW guess, if we have a srtt */
2356 		if (rack->rc_tp->t_srtt == 0) {
2357 			/*
2358 			 * Go with old pacing method
2359 			 * i.e. burst mitigation only.
2360 			 */
2361 			return (0);
2362 		}
2363 		/* Ok lets get the initial TCP win (not racks) */
2364 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2365 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2366 		bw *= (uint64_t)USECS_IN_SECOND;
2367 		bw /= srtt;
2368 		ret_bw = bw;
2369 		goto compensate;
2370 
2371 	}
2372 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2373 		/* Averaging is done, we can return the value */
2374 		bw = rack->r_ctl.gp_bw;
2375 	} else {
2376 		/* Still doing initial average must calculate */
2377 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2378 	}
2379 	lt_bw = rack_get_lt_bw(rack);
2380 	if (lt_bw == 0) {
2381 		/* If we don't have one then equate it to the gp_bw */
2382 		lt_bw = rack->r_ctl.gp_bw;
2383 	}
2384 	if ((rack->r_cwnd_was_clamped == 1) && (rack->r_clamped_gets_lower > 0)){
2385 		/*  if clamped take the lowest */
2386 		if (lt_bw < bw)
2387 			ret_bw = lt_bw;
2388 		else
2389 			ret_bw = bw;
2390 	} else {
2391 		/* If not set for clamped to get lowest, take the highest */
2392 		if (lt_bw > bw)
2393 			ret_bw = lt_bw;
2394 		else
2395 			ret_bw = bw;
2396 	}
2397 	/*
2398 	 * Now lets compensate based on the TCP/IP overhead. Our
2399 	 * Goodput estimate does not include this so we must pace out
2400 	 * a bit faster since our pacing calculations do. The pacing
2401 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2402 	 * we are using to do this, so we do that here in the opposite
2403 	 * direction as well. This means that if we are tunneled and the
2404 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2405 	 * compensated for in the pacing time the opposite way.
2406 	 */
2407 compensate:
2408 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2409 	return(ret_bw);
2410 }
2411 
2412 
2413 static uint64_t
2414 rack_get_bw(struct tcp_rack *rack)
2415 {
2416 	uint64_t bw;
2417 
2418 	if (rack->use_fixed_rate) {
2419 		/* Return the fixed pacing rate */
2420 		return (rack_get_fixed_pacing_bw(rack));
2421 	}
2422 	bw = rack_get_gp_est(rack);
2423 	return (bw);
2424 }
2425 
2426 static uint16_t
2427 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2428 {
2429 	if (rack->use_fixed_rate) {
2430 		return (100);
2431 	} else if (rack->in_probe_rtt && (rsm == NULL))
2432 		return (rack->r_ctl.rack_per_of_gp_probertt);
2433 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2434 		  rack->r_ctl.rack_per_of_gp_rec)) {
2435 		if (rsm) {
2436 			/* a retransmission always use the recovery rate */
2437 			return (rack->r_ctl.rack_per_of_gp_rec);
2438 		} else if (rack->rack_rec_nonrxt_use_cr) {
2439 			/* Directed to use the configured rate */
2440 			goto configured_rate;
2441 		} else if (rack->rack_no_prr &&
2442 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2443 			/* No PRR, lets just use the b/w estimate only */
2444 			return (100);
2445 		} else {
2446 			/*
2447 			 * Here we may have a non-retransmit but we
2448 			 * have no overrides, so just use the recovery
2449 			 * rate (prr is in effect).
2450 			 */
2451 			return (rack->r_ctl.rack_per_of_gp_rec);
2452 		}
2453 	}
2454 configured_rate:
2455 	/* For the configured rate we look at our cwnd vs the ssthresh */
2456 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2457 		return (rack->r_ctl.rack_per_of_gp_ss);
2458 	else
2459 		return (rack->r_ctl.rack_per_of_gp_ca);
2460 }
2461 
2462 static void
2463 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2464 {
2465 	/*
2466 	 * Types of logs (mod value)
2467 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2468 	 * 2 = a dsack round begins, persist is reset to 16.
2469 	 * 3 = a dsack round ends
2470 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2471 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2472 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2473 	 */
2474 	if (tcp_bblogging_on(rack->rc_tp)) {
2475 		union tcp_log_stackspecific log;
2476 		struct timeval tv;
2477 
2478 		memset(&log, 0, sizeof(log));
2479 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2480 		log.u_bbr.flex1 <<= 1;
2481 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2482 		log.u_bbr.flex1 <<= 1;
2483 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2484 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2485 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2486 		log.u_bbr.flex4 = flex4;
2487 		log.u_bbr.flex5 = flex5;
2488 		log.u_bbr.flex6 = flex6;
2489 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2490 		log.u_bbr.flex8 = mod;
2491 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2492 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2493 		    &rack->rc_inp->inp_socket->so_rcv,
2494 		    &rack->rc_inp->inp_socket->so_snd,
2495 		    RACK_DSACK_HANDLING, 0,
2496 		    0, &log, false, &tv);
2497 	}
2498 }
2499 
2500 static void
2501 rack_log_hdwr_pacing(struct tcp_rack *rack,
2502 		     uint64_t rate, uint64_t hw_rate, int line,
2503 		     int error, uint16_t mod)
2504 {
2505 	if (tcp_bblogging_on(rack->rc_tp)) {
2506 		union tcp_log_stackspecific log;
2507 		struct timeval tv;
2508 		const struct ifnet *ifp;
2509 
2510 		memset(&log, 0, sizeof(log));
2511 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2512 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2513 		if (rack->r_ctl.crte) {
2514 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2515 		} else if (rack->rc_inp->inp_route.ro_nh &&
2516 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2517 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2518 		} else
2519 			ifp = NULL;
2520 		if (ifp) {
2521 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2522 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2523 		}
2524 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2525 		log.u_bbr.bw_inuse = rate;
2526 		log.u_bbr.flex5 = line;
2527 		log.u_bbr.flex6 = error;
2528 		log.u_bbr.flex7 = mod;
2529 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2530 		log.u_bbr.flex8 = rack->use_fixed_rate;
2531 		log.u_bbr.flex8 <<= 1;
2532 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2533 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2534 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2535 		if (rack->r_ctl.crte)
2536 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2537 		else
2538 			log.u_bbr.cur_del_rate = 0;
2539 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2540 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2541 		    &rack->rc_inp->inp_socket->so_rcv,
2542 		    &rack->rc_inp->inp_socket->so_snd,
2543 		    BBR_LOG_HDWR_PACE, 0,
2544 		    0, &log, false, &tv);
2545 	}
2546 }
2547 
2548 static uint64_t
2549 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2550 {
2551 	/*
2552 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2553 	 */
2554 	uint64_t bw_est, high_rate;
2555 	uint64_t gain;
2556 
2557 	if ((rack->r_pacing_discount == 0) ||
2558 	    (rack_full_buffer_discount == 0)) {
2559 		/*
2560 		 * No buffer level based discount from client buffer
2561 		 * level is enabled or the feature is disabled.
2562 		 */
2563 		gain = (uint64_t)rack_get_output_gain(rack, rsm);
2564 		bw_est = bw * gain;
2565 		bw_est /= (uint64_t)100;
2566 	} else {
2567 		/*
2568 		 * We have a discount in place apply it with
2569 		 * just a 100% gain (we get no boost if the buffer
2570 		 * is full).
2571 		 */
2572 		uint64_t discount;
2573 
2574 		discount = bw * (uint64_t)(rack_full_buffer_discount * rack->r_ctl.pacing_discount_amm);
2575 		discount /= 100;
2576 		/* What %% of the b/w do we discount */
2577 		bw_est = bw - discount;
2578 	}
2579 	/* Never fall below the minimum (def 64kbps) */
2580 	if (bw_est < RACK_MIN_BW)
2581 		bw_est = RACK_MIN_BW;
2582 	if (rack->r_rack_hw_rate_caps) {
2583 		/* Rate caps are in place */
2584 		if (rack->r_ctl.crte != NULL) {
2585 			/* We have a hdwr rate already */
2586 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2587 			if (bw_est >= high_rate) {
2588 				/* We are capping bw at the highest rate table entry */
2589 				if (rack_hw_rate_cap_per &&
2590 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2591 					rack->r_rack_hw_rate_caps = 0;
2592 					goto done;
2593 				}
2594 				rack_log_hdwr_pacing(rack,
2595 						     bw_est, high_rate, __LINE__,
2596 						     0, 3);
2597 				bw_est = high_rate;
2598 				if (capped)
2599 					*capped = 1;
2600 			}
2601 		} else if ((rack->rack_hdrw_pacing == 0) &&
2602 			   (rack->rack_hdw_pace_ena) &&
2603 			   (rack->rack_attempt_hdwr_pace == 0) &&
2604 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2605 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2606 			/*
2607 			 * Special case, we have not yet attempted hardware
2608 			 * pacing, and yet we may, when we do, find out if we are
2609 			 * above the highest rate. We need to know the maxbw for the interface
2610 			 * in question (if it supports ratelimiting). We get back
2611 			 * a 0, if the interface is not found in the RL lists.
2612 			 */
2613 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2614 			if (high_rate) {
2615 				/* Yep, we have a rate is it above this rate? */
2616 				if (bw_est > high_rate) {
2617 					bw_est = high_rate;
2618 					if (capped)
2619 						*capped = 1;
2620 				}
2621 			}
2622 		}
2623 	}
2624 done:
2625 	return (bw_est);
2626 }
2627 
2628 static void
2629 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2630 {
2631 	if (tcp_bblogging_on(rack->rc_tp)) {
2632 		union tcp_log_stackspecific log;
2633 		struct timeval tv;
2634 
2635 		if (rack->sack_attack_disable > 0)
2636 			goto log_anyway;
2637 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2638 			/*
2639 			 * We get 3 values currently for mod
2640 			 * 1 - We are retransmitting and this tells the reason.
2641 			 * 2 - We are clearing a dup-ack count.
2642 			 * 3 - We are incrementing a dup-ack count.
2643 			 *
2644 			 * The clear/increment are only logged
2645 			 * if you have BBverbose on.
2646 			 */
2647 			return;
2648 		}
2649 log_anyway:
2650 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2651 		log.u_bbr.flex1 = tsused;
2652 		log.u_bbr.flex2 = thresh;
2653 		log.u_bbr.flex3 = rsm->r_flags;
2654 		log.u_bbr.flex4 = rsm->r_dupack;
2655 		log.u_bbr.flex5 = rsm->r_start;
2656 		log.u_bbr.flex6 = rsm->r_end;
2657 		log.u_bbr.flex8 = mod;
2658 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2659 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2660 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2661 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2662 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2663 		log.u_bbr.pacing_gain = rack->r_must_retran;
2664 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2665 		    &rack->rc_inp->inp_socket->so_rcv,
2666 		    &rack->rc_inp->inp_socket->so_snd,
2667 		    BBR_LOG_SETTINGS_CHG, 0,
2668 		    0, &log, false, &tv);
2669 	}
2670 }
2671 
2672 static void
2673 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2674 {
2675 	if (tcp_bblogging_on(rack->rc_tp)) {
2676 		union tcp_log_stackspecific log;
2677 		struct timeval tv;
2678 
2679 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2680 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2681 		log.u_bbr.flex2 = to;
2682 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2683 		log.u_bbr.flex4 = slot;
2684 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2685 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2686 		log.u_bbr.flex7 = rack->rc_in_persist;
2687 		log.u_bbr.flex8 = which;
2688 		if (rack->rack_no_prr)
2689 			log.u_bbr.pkts_out = 0;
2690 		else
2691 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2692 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2693 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2694 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2695 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2696 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2697 		log.u_bbr.pacing_gain = rack->r_must_retran;
2698 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2699 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2700 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2701 		log.u_bbr.lost = rack_rto_min;
2702 		log.u_bbr.epoch = rack->r_ctl.roundends;
2703 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2704 		    &rack->rc_inp->inp_socket->so_rcv,
2705 		    &rack->rc_inp->inp_socket->so_snd,
2706 		    BBR_LOG_TIMERSTAR, 0,
2707 		    0, &log, false, &tv);
2708 	}
2709 }
2710 
2711 static void
2712 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2713 {
2714 	if (tcp_bblogging_on(rack->rc_tp)) {
2715 		union tcp_log_stackspecific log;
2716 		struct timeval tv;
2717 
2718 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2719 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2720 		log.u_bbr.flex8 = to_num;
2721 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2722 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2723 		if (rsm == NULL)
2724 			log.u_bbr.flex3 = 0;
2725 		else
2726 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2727 		if (rack->rack_no_prr)
2728 			log.u_bbr.flex5 = 0;
2729 		else
2730 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2731 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2732 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2733 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2734 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2735 		log.u_bbr.pacing_gain = rack->r_must_retran;
2736 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2737 		    &rack->rc_inp->inp_socket->so_rcv,
2738 		    &rack->rc_inp->inp_socket->so_snd,
2739 		    BBR_LOG_RTO, 0,
2740 		    0, &log, false, &tv);
2741 	}
2742 }
2743 
2744 static void
2745 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2746 		 struct rack_sendmap *prev,
2747 		 struct rack_sendmap *rsm,
2748 		 struct rack_sendmap *next,
2749 		 int flag, uint32_t th_ack, int line)
2750 {
2751 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2752 		union tcp_log_stackspecific log;
2753 		struct timeval tv;
2754 
2755 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2756 		log.u_bbr.flex8 = flag;
2757 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2758 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2759 		log.u_bbr.delRate = (uint64_t)rsm;
2760 		log.u_bbr.rttProp = (uint64_t)next;
2761 		log.u_bbr.flex7 = 0;
2762 		if (prev) {
2763 			log.u_bbr.flex1 = prev->r_start;
2764 			log.u_bbr.flex2 = prev->r_end;
2765 			log.u_bbr.flex7 |= 0x4;
2766 		}
2767 		if (rsm) {
2768 			log.u_bbr.flex3 = rsm->r_start;
2769 			log.u_bbr.flex4 = rsm->r_end;
2770 			log.u_bbr.flex7 |= 0x2;
2771 		}
2772 		if (next) {
2773 			log.u_bbr.flex5 = next->r_start;
2774 			log.u_bbr.flex6 = next->r_end;
2775 			log.u_bbr.flex7 |= 0x1;
2776 		}
2777 		log.u_bbr.applimited = line;
2778 		log.u_bbr.pkts_out = th_ack;
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 		if (rack->rack_no_prr)
2782 			log.u_bbr.lost = 0;
2783 		else
2784 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2785 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2786 		    &rack->rc_inp->inp_socket->so_rcv,
2787 		    &rack->rc_inp->inp_socket->so_snd,
2788 		    TCP_LOG_MAPCHG, 0,
2789 		    0, &log, false, &tv);
2790 	}
2791 }
2792 
2793 static void
2794 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2795 		 struct rack_sendmap *rsm, int conf)
2796 {
2797 	if (tcp_bblogging_on(tp)) {
2798 		union tcp_log_stackspecific log;
2799 		struct timeval tv;
2800 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2801 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2802 		log.u_bbr.flex1 = t;
2803 		log.u_bbr.flex2 = len;
2804 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2805 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2806 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2807 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2808 		log.u_bbr.flex7 = conf;
2809 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2810 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2811 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2812 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2813 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2814 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2815 		if (rsm) {
2816 			log.u_bbr.pkt_epoch = rsm->r_start;
2817 			log.u_bbr.lost = rsm->r_end;
2818 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2819 			/* We loose any upper of the 24 bits */
2820 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2821 		} else {
2822 			/* Its a SYN */
2823 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2824 			log.u_bbr.lost = 0;
2825 			log.u_bbr.cwnd_gain = 0;
2826 			log.u_bbr.pacing_gain = 0;
2827 		}
2828 		/* Write out general bits of interest rrs here */
2829 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2830 		log.u_bbr.use_lt_bw <<= 1;
2831 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2832 		log.u_bbr.use_lt_bw <<= 1;
2833 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2834 		log.u_bbr.use_lt_bw <<= 1;
2835 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2836 		log.u_bbr.use_lt_bw <<= 1;
2837 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2838 		log.u_bbr.use_lt_bw <<= 1;
2839 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2840 		log.u_bbr.use_lt_bw <<= 1;
2841 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2842 		log.u_bbr.use_lt_bw <<= 1;
2843 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2844 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2845 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2846 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2847 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2848 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2849 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2850 		log.u_bbr.bw_inuse <<= 32;
2851 		if (rsm)
2852 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2853 		TCP_LOG_EVENTP(tp, NULL,
2854 		    &rack->rc_inp->inp_socket->so_rcv,
2855 		    &rack->rc_inp->inp_socket->so_snd,
2856 		    BBR_LOG_BBRRTT, 0,
2857 		    0, &log, false, &tv);
2858 
2859 
2860 	}
2861 }
2862 
2863 static void
2864 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2865 {
2866 	/*
2867 	 * Log the rtt sample we are
2868 	 * applying to the srtt algorithm in
2869 	 * useconds.
2870 	 */
2871 	if (tcp_bblogging_on(rack->rc_tp)) {
2872 		union tcp_log_stackspecific log;
2873 		struct timeval tv;
2874 
2875 		/* Convert our ms to a microsecond */
2876 		memset(&log, 0, sizeof(log));
2877 		log.u_bbr.flex1 = rtt;
2878 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2879 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2880 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2881 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2882 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2883 		log.u_bbr.flex7 = 1;
2884 		log.u_bbr.flex8 = rack->sack_attack_disable;
2885 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2886 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2887 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2888 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2889 		log.u_bbr.pacing_gain = rack->r_must_retran;
2890 		/*
2891 		 * We capture in delRate the upper 32 bits as
2892 		 * the confidence level we had declared, and the
2893 		 * lower 32 bits as the actual RTT using the arrival
2894 		 * timestamp.
2895 		 */
2896 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2897 		log.u_bbr.delRate <<= 32;
2898 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2899 		/* Lets capture all the things that make up t_rtxcur */
2900 		log.u_bbr.applimited = rack_rto_min;
2901 		log.u_bbr.epoch = rack_rto_max;
2902 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2903 		log.u_bbr.lost = rack_rto_min;
2904 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2905 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2906 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2907 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2908 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2909 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2910 		    &rack->rc_inp->inp_socket->so_rcv,
2911 		    &rack->rc_inp->inp_socket->so_snd,
2912 		    TCP_LOG_RTT, 0,
2913 		    0, &log, false, &tv);
2914 	}
2915 }
2916 
2917 static void
2918 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2919 {
2920 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2921 		union tcp_log_stackspecific log;
2922 		struct timeval tv;
2923 
2924 		/* Convert our ms to a microsecond */
2925 		memset(&log, 0, sizeof(log));
2926 		log.u_bbr.flex1 = rtt;
2927 		log.u_bbr.flex2 = send_time;
2928 		log.u_bbr.flex3 = ack_time;
2929 		log.u_bbr.flex4 = where;
2930 		log.u_bbr.flex7 = 2;
2931 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2932 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2933 		    &rack->rc_inp->inp_socket->so_rcv,
2934 		    &rack->rc_inp->inp_socket->so_snd,
2935 		    TCP_LOG_RTT, 0,
2936 		    0, &log, false, &tv);
2937 	}
2938 }
2939 
2940 
2941 static void
2942 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2943 {
2944 	if (tcp_bblogging_on(rack->rc_tp)) {
2945 		union tcp_log_stackspecific log;
2946 		struct timeval tv;
2947 
2948 		/* Convert our ms to a microsecond */
2949 		memset(&log, 0, sizeof(log));
2950 		log.u_bbr.flex1 = idx;
2951 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2952 		log.u_bbr.flex3 = tsecho;
2953 		log.u_bbr.flex7 = 3;
2954 		log.u_bbr.rttProp = tsv;
2955 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2956 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2957 		    &rack->rc_inp->inp_socket->so_rcv,
2958 		    &rack->rc_inp->inp_socket->so_snd,
2959 		    TCP_LOG_RTT, 0,
2960 		    0, &log, false, &tv);
2961 	}
2962 }
2963 
2964 
2965 static inline void
2966 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2967 {
2968 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2969 		union tcp_log_stackspecific log;
2970 		struct timeval tv;
2971 
2972 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2973 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2974 		log.u_bbr.flex1 = line;
2975 		log.u_bbr.flex2 = tick;
2976 		log.u_bbr.flex3 = tp->t_maxunacktime;
2977 		log.u_bbr.flex4 = tp->t_acktime;
2978 		log.u_bbr.flex8 = event;
2979 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2980 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2981 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2982 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2983 		log.u_bbr.pacing_gain = rack->r_must_retran;
2984 		TCP_LOG_EVENTP(tp, NULL,
2985 		    &rack->rc_inp->inp_socket->so_rcv,
2986 		    &rack->rc_inp->inp_socket->so_snd,
2987 		    BBR_LOG_PROGRESS, 0,
2988 		    0, &log, false, &tv);
2989 	}
2990 }
2991 
2992 static void
2993 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
2994 {
2995 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2996 		union tcp_log_stackspecific log;
2997 
2998 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2999 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3000 		log.u_bbr.flex1 = slot;
3001 		if (rack->rack_no_prr)
3002 			log.u_bbr.flex2 = 0;
3003 		else
3004 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3005 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3006 		log.u_bbr.flex5 = rack->r_ctl.ack_during_sd;
3007 		log.u_bbr.flex6 = line;
3008 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3009 		log.u_bbr.flex8 = rack->rc_in_persist;
3010 		log.u_bbr.timeStamp = cts;
3011 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3012 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3013 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3014 		log.u_bbr.pacing_gain = rack->r_must_retran;
3015 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3016 		    &rack->rc_inp->inp_socket->so_rcv,
3017 		    &rack->rc_inp->inp_socket->so_snd,
3018 		    BBR_LOG_BBRSND, 0,
3019 		    0, &log, false, tv);
3020 	}
3021 }
3022 
3023 static void
3024 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3025 {
3026 	if (tcp_bblogging_on(rack->rc_tp)) {
3027 		union tcp_log_stackspecific log;
3028 		struct timeval tv;
3029 
3030 		memset(&log, 0, sizeof(log));
3031 		log.u_bbr.flex1 = did_out;
3032 		log.u_bbr.flex2 = nxt_pkt;
3033 		log.u_bbr.flex3 = way_out;
3034 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3035 		if (rack->rack_no_prr)
3036 			log.u_bbr.flex5 = 0;
3037 		else
3038 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3039 		log.u_bbr.flex6 = nsegs;
3040 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3041 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3042 		log.u_bbr.flex7 <<= 1;
3043 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3044 		log.u_bbr.flex7 <<= 1;
3045 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3046 		log.u_bbr.flex8 = rack->rc_in_persist;
3047 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3048 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3049 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3050 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3051 		log.u_bbr.use_lt_bw <<= 1;
3052 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3053 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3054 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3055 		log.u_bbr.pacing_gain = rack->r_must_retran;
3056 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3057 		    &rack->rc_inp->inp_socket->so_rcv,
3058 		    &rack->rc_inp->inp_socket->so_snd,
3059 		    BBR_LOG_DOSEG_DONE, 0,
3060 		    0, &log, false, &tv);
3061 	}
3062 }
3063 
3064 static void
3065 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3066 {
3067 	if (tcp_bblogging_on(rack->rc_tp)) {
3068 		union tcp_log_stackspecific log;
3069 		struct timeval tv;
3070 
3071 		memset(&log, 0, sizeof(log));
3072 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3073 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3074 		log.u_bbr.flex4 = arg1;
3075 		log.u_bbr.flex5 = arg2;
3076 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3077 		log.u_bbr.flex6 = arg3;
3078 		log.u_bbr.flex8 = frm;
3079 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3080 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3081 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3082 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3083 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3084 		log.u_bbr.pacing_gain = rack->r_must_retran;
3085 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3086 		    &tptosocket(tp)->so_snd,
3087 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3088 	}
3089 }
3090 
3091 static void
3092 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3093 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3094 {
3095 	if (tcp_bblogging_on(rack->rc_tp)) {
3096 		union tcp_log_stackspecific log;
3097 		struct timeval tv;
3098 
3099 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3100 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3101 		log.u_bbr.flex1 = slot;
3102 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3103 		log.u_bbr.flex4 = reason;
3104 		if (rack->rack_no_prr)
3105 			log.u_bbr.flex5 = 0;
3106 		else
3107 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3108 		log.u_bbr.flex7 = hpts_calling;
3109 		log.u_bbr.flex8 = rack->rc_in_persist;
3110 		log.u_bbr.lt_epoch = cwnd_to_use;
3111 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3112 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3113 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3114 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3115 		log.u_bbr.pacing_gain = rack->r_must_retran;
3116 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3117 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3118 		    &rack->rc_inp->inp_socket->so_rcv,
3119 		    &rack->rc_inp->inp_socket->so_snd,
3120 		    BBR_LOG_JUSTRET, 0,
3121 		    tlen, &log, false, &tv);
3122 	}
3123 }
3124 
3125 static void
3126 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3127 		   struct timeval *tv, uint32_t flags_on_entry)
3128 {
3129 	if (tcp_bblogging_on(rack->rc_tp)) {
3130 		union tcp_log_stackspecific log;
3131 
3132 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3133 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3134 		log.u_bbr.flex1 = line;
3135 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3136 		log.u_bbr.flex3 = flags_on_entry;
3137 		log.u_bbr.flex4 = us_cts;
3138 		if (rack->rack_no_prr)
3139 			log.u_bbr.flex5 = 0;
3140 		else
3141 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3142 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3143 		log.u_bbr.flex7 = hpts_removed;
3144 		log.u_bbr.flex8 = 1;
3145 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3146 		log.u_bbr.timeStamp = us_cts;
3147 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3148 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3149 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3150 		log.u_bbr.pacing_gain = rack->r_must_retran;
3151 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3152 		    &rack->rc_inp->inp_socket->so_rcv,
3153 		    &rack->rc_inp->inp_socket->so_snd,
3154 		    BBR_LOG_TIMERCANC, 0,
3155 		    0, &log, false, tv);
3156 	}
3157 }
3158 
3159 static void
3160 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3161 			  uint32_t flex1, uint32_t flex2,
3162 			  uint32_t flex3, uint32_t flex4,
3163 			  uint32_t flex5, uint32_t flex6,
3164 			  uint16_t flex7, uint8_t mod)
3165 {
3166 	if (tcp_bblogging_on(rack->rc_tp)) {
3167 		union tcp_log_stackspecific log;
3168 		struct timeval tv;
3169 
3170 		if (mod == 1) {
3171 			/* No you can't use 1, its for the real to cancel */
3172 			return;
3173 		}
3174 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3175 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3176 		log.u_bbr.flex1 = flex1;
3177 		log.u_bbr.flex2 = flex2;
3178 		log.u_bbr.flex3 = flex3;
3179 		log.u_bbr.flex4 = flex4;
3180 		log.u_bbr.flex5 = flex5;
3181 		log.u_bbr.flex6 = flex6;
3182 		log.u_bbr.flex7 = flex7;
3183 		log.u_bbr.flex8 = mod;
3184 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3185 		    &rack->rc_inp->inp_socket->so_rcv,
3186 		    &rack->rc_inp->inp_socket->so_snd,
3187 		    BBR_LOG_TIMERCANC, 0,
3188 		    0, &log, false, &tv);
3189 	}
3190 }
3191 
3192 static void
3193 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3194 {
3195 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3196 		union tcp_log_stackspecific log;
3197 		struct timeval tv;
3198 
3199 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3200 		log.u_bbr.flex1 = timers;
3201 		log.u_bbr.flex2 = ret;
3202 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3203 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3204 		log.u_bbr.flex5 = cts;
3205 		if (rack->rack_no_prr)
3206 			log.u_bbr.flex6 = 0;
3207 		else
3208 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3209 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3210 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3211 		log.u_bbr.pacing_gain = rack->r_must_retran;
3212 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3213 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3214 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3215 		    &rack->rc_inp->inp_socket->so_rcv,
3216 		    &rack->rc_inp->inp_socket->so_snd,
3217 		    BBR_LOG_TO_PROCESS, 0,
3218 		    0, &log, false, &tv);
3219 	}
3220 }
3221 
3222 static void
3223 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3224 {
3225 	if (tcp_bblogging_on(rack->rc_tp)) {
3226 		union tcp_log_stackspecific log;
3227 		struct timeval tv;
3228 
3229 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3230 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3231 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3232 		if (rack->rack_no_prr)
3233 			log.u_bbr.flex3 = 0;
3234 		else
3235 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3236 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3237 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3238 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3239 		log.u_bbr.flex7 = line;
3240 		log.u_bbr.flex8 = frm;
3241 		log.u_bbr.pkts_out = orig_cwnd;
3242 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3243 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3244 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3245 		log.u_bbr.use_lt_bw <<= 1;
3246 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3247 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3248 		    &rack->rc_inp->inp_socket->so_rcv,
3249 		    &rack->rc_inp->inp_socket->so_snd,
3250 		    BBR_LOG_BBRUPD, 0,
3251 		    0, &log, false, &tv);
3252 	}
3253 }
3254 
3255 #ifdef TCP_SAD_DETECTION
3256 static void
3257 rack_log_sad(struct tcp_rack *rack, int event)
3258 {
3259 	if (tcp_bblogging_on(rack->rc_tp)) {
3260 		union tcp_log_stackspecific log;
3261 		struct timeval tv;
3262 
3263 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3264 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
3265 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
3266 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
3267 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
3268 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
3269 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
3270 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
3271 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
3272 		log.u_bbr.lt_epoch |= rack->do_detection;
3273 		log.u_bbr.applimited = tcp_map_minimum;
3274 		log.u_bbr.flex7 = rack->sack_attack_disable;
3275 		log.u_bbr.flex8 = event;
3276 		log.u_bbr.bbr_state = rack->rc_suspicious;
3277 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3278 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3279 		log.u_bbr.delivered = tcp_sad_decay_val;
3280 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3281 		    &rack->rc_inp->inp_socket->so_rcv,
3282 		    &rack->rc_inp->inp_socket->so_snd,
3283 		    TCP_SAD_DETECT, 0,
3284 		    0, &log, false, &tv);
3285 	}
3286 }
3287 #endif
3288 
3289 static void
3290 rack_counter_destroy(void)
3291 {
3292 	counter_u64_free(rack_total_bytes);
3293 	counter_u64_free(rack_fto_send);
3294 	counter_u64_free(rack_fto_rsm_send);
3295 	counter_u64_free(rack_nfto_resend);
3296 	counter_u64_free(rack_hw_pace_init_fail);
3297 	counter_u64_free(rack_hw_pace_lost);
3298 	counter_u64_free(rack_non_fto_send);
3299 	counter_u64_free(rack_extended_rfo);
3300 	counter_u64_free(rack_ack_total);
3301 	counter_u64_free(rack_express_sack);
3302 	counter_u64_free(rack_sack_total);
3303 	counter_u64_free(rack_move_none);
3304 	counter_u64_free(rack_move_some);
3305 	counter_u64_free(rack_sack_attacks_detected);
3306 	counter_u64_free(rack_sack_attacks_reversed);
3307 	counter_u64_free(rack_sack_attacks_suspect);
3308 	counter_u64_free(rack_sack_used_next_merge);
3309 	counter_u64_free(rack_sack_used_prev_merge);
3310 	counter_u64_free(rack_tlp_tot);
3311 	counter_u64_free(rack_tlp_newdata);
3312 	counter_u64_free(rack_tlp_retran);
3313 	counter_u64_free(rack_tlp_retran_bytes);
3314 	counter_u64_free(rack_to_tot);
3315 	counter_u64_free(rack_saw_enobuf);
3316 	counter_u64_free(rack_saw_enobuf_hw);
3317 	counter_u64_free(rack_saw_enetunreach);
3318 	counter_u64_free(rack_hot_alloc);
3319 	counter_u64_free(rack_to_alloc);
3320 	counter_u64_free(rack_to_alloc_hard);
3321 	counter_u64_free(rack_to_alloc_emerg);
3322 	counter_u64_free(rack_to_alloc_limited);
3323 	counter_u64_free(rack_alloc_limited_conns);
3324 	counter_u64_free(rack_split_limited);
3325 	counter_u64_free(rack_multi_single_eq);
3326 	counter_u64_free(rack_rxt_clamps_cwnd);
3327 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3328 	counter_u64_free(rack_proc_non_comp_ack);
3329 	counter_u64_free(rack_sack_proc_all);
3330 	counter_u64_free(rack_sack_proc_restart);
3331 	counter_u64_free(rack_sack_proc_short);
3332 	counter_u64_free(rack_sack_skipped_acked);
3333 	counter_u64_free(rack_sack_splits);
3334 	counter_u64_free(rack_input_idle_reduces);
3335 	counter_u64_free(rack_collapsed_win);
3336 	counter_u64_free(rack_collapsed_win_rxt);
3337 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3338 	counter_u64_free(rack_collapsed_win_seen);
3339 	counter_u64_free(rack_try_scwnd);
3340 	counter_u64_free(rack_persists_sends);
3341 	counter_u64_free(rack_persists_acks);
3342 	counter_u64_free(rack_persists_loss);
3343 	counter_u64_free(rack_persists_lost_ends);
3344 #ifdef INVARIANTS
3345 	counter_u64_free(rack_adjust_map_bw);
3346 #endif
3347 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3348 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3349 }
3350 
3351 static struct rack_sendmap *
3352 rack_alloc(struct tcp_rack *rack)
3353 {
3354 	struct rack_sendmap *rsm;
3355 
3356 	/*
3357 	 * First get the top of the list it in
3358 	 * theory is the "hottest" rsm we have,
3359 	 * possibly just freed by ack processing.
3360 	 */
3361 	if (rack->rc_free_cnt > rack_free_cache) {
3362 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3363 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3364 		counter_u64_add(rack_hot_alloc, 1);
3365 		rack->rc_free_cnt--;
3366 		return (rsm);
3367 	}
3368 	/*
3369 	 * Once we get under our free cache we probably
3370 	 * no longer have a "hot" one available. Lets
3371 	 * get one from UMA.
3372 	 */
3373 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3374 	if (rsm) {
3375 		rack->r_ctl.rc_num_maps_alloced++;
3376 		counter_u64_add(rack_to_alloc, 1);
3377 		return (rsm);
3378 	}
3379 	/*
3380 	 * Dig in to our aux rsm's (the last two) since
3381 	 * UMA failed to get us one.
3382 	 */
3383 	if (rack->rc_free_cnt) {
3384 		counter_u64_add(rack_to_alloc_emerg, 1);
3385 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3386 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3387 		rack->rc_free_cnt--;
3388 		return (rsm);
3389 	}
3390 	return (NULL);
3391 }
3392 
3393 static struct rack_sendmap *
3394 rack_alloc_full_limit(struct tcp_rack *rack)
3395 {
3396 	if ((V_tcp_map_entries_limit > 0) &&
3397 	    (rack->do_detection == 0) &&
3398 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3399 		counter_u64_add(rack_to_alloc_limited, 1);
3400 		if (!rack->alloc_limit_reported) {
3401 			rack->alloc_limit_reported = 1;
3402 			counter_u64_add(rack_alloc_limited_conns, 1);
3403 		}
3404 		return (NULL);
3405 	}
3406 	return (rack_alloc(rack));
3407 }
3408 
3409 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3410 static struct rack_sendmap *
3411 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3412 {
3413 	struct rack_sendmap *rsm;
3414 
3415 	if (limit_type) {
3416 		/* currently there is only one limit type */
3417 		if (rack->r_ctl.rc_split_limit > 0 &&
3418 		    (rack->do_detection == 0) &&
3419 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3420 			counter_u64_add(rack_split_limited, 1);
3421 			if (!rack->alloc_limit_reported) {
3422 				rack->alloc_limit_reported = 1;
3423 				counter_u64_add(rack_alloc_limited_conns, 1);
3424 			}
3425 			return (NULL);
3426 #ifdef TCP_SAD_DETECTION
3427 		} else if ((tcp_sad_limit != 0) &&
3428 			   (rack->do_detection == 1) &&
3429 			   (rack->r_ctl.rc_num_split_allocs >= tcp_sad_limit)) {
3430 			counter_u64_add(rack_split_limited, 1);
3431 			if (!rack->alloc_limit_reported) {
3432 				rack->alloc_limit_reported = 1;
3433 				counter_u64_add(rack_alloc_limited_conns, 1);
3434 			}
3435 			return (NULL);
3436 #endif
3437 		}
3438 	}
3439 
3440 	/* allocate and mark in the limit type, if set */
3441 	rsm = rack_alloc(rack);
3442 	if (rsm != NULL && limit_type) {
3443 		rsm->r_limit_type = limit_type;
3444 		rack->r_ctl.rc_num_split_allocs++;
3445 	}
3446 	return (rsm);
3447 }
3448 
3449 static void
3450 rack_free_trim(struct tcp_rack *rack)
3451 {
3452 	struct rack_sendmap *rsm;
3453 
3454 	/*
3455 	 * Free up all the tail entries until
3456 	 * we get our list down to the limit.
3457 	 */
3458 	while (rack->rc_free_cnt > rack_free_cache) {
3459 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3460 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3461 		rack->rc_free_cnt--;
3462 		rack->r_ctl.rc_num_maps_alloced--;
3463 		uma_zfree(rack_zone, rsm);
3464 	}
3465 }
3466 
3467 static void
3468 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3469 {
3470 	if (rsm->r_flags & RACK_APP_LIMITED) {
3471 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3472 			rack->r_ctl.rc_app_limited_cnt--;
3473 		}
3474 	}
3475 	if (rsm->r_limit_type) {
3476 		/* currently there is only one limit type */
3477 		rack->r_ctl.rc_num_split_allocs--;
3478 	}
3479 	if (rsm == rack->r_ctl.rc_first_appl) {
3480 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3481 			rack->r_ctl.rc_first_appl = NULL;
3482 		else
3483 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3484 	}
3485 	if (rsm == rack->r_ctl.rc_resend)
3486 		rack->r_ctl.rc_resend = NULL;
3487 	if (rsm == rack->r_ctl.rc_end_appl)
3488 		rack->r_ctl.rc_end_appl = NULL;
3489 	if (rack->r_ctl.rc_tlpsend == rsm)
3490 		rack->r_ctl.rc_tlpsend = NULL;
3491 	if (rack->r_ctl.rc_sacklast == rsm)
3492 		rack->r_ctl.rc_sacklast = NULL;
3493 	memset(rsm, 0, sizeof(struct rack_sendmap));
3494 	/* Make sure we are not going to overrun our count limit of 0xff */
3495 	if ((rack->rc_free_cnt + 1) > 0xff) {
3496 		rack_free_trim(rack);
3497 	}
3498 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3499 	rack->rc_free_cnt++;
3500 }
3501 
3502 static uint32_t
3503 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3504 {
3505 	uint64_t srtt, bw, len, tim;
3506 	uint32_t segsiz, def_len, minl;
3507 
3508 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3509 	def_len = rack_def_data_window * segsiz;
3510 	if (rack->rc_gp_filled == 0) {
3511 		/*
3512 		 * We have no measurement (IW is in flight?) so
3513 		 * we can only guess using our data_window sysctl
3514 		 * value (usually 20MSS).
3515 		 */
3516 		return (def_len);
3517 	}
3518 	/*
3519 	 * Now we have a number of factors to consider.
3520 	 *
3521 	 * 1) We have a desired BDP which is usually
3522 	 *    at least 2.
3523 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3524 	 *    but we allow it too to be more.
3525 	 * 3) We want to make sure a measurement last N useconds (if
3526 	 *    we have set rack_min_measure_usec.
3527 	 *
3528 	 * We handle the first concern here by trying to create a data
3529 	 * window of max(rack_def_data_window, DesiredBDP). The
3530 	 * second concern we handle in not letting the measurement
3531 	 * window end normally until at least the required SRTT's
3532 	 * have gone by which is done further below in
3533 	 * rack_enough_for_measurement(). Finally the third concern
3534 	 * we also handle here by calculating how long that time
3535 	 * would take at the current BW and then return the
3536 	 * max of our first calculation and that length. Note
3537 	 * that if rack_min_measure_usec is 0, we don't deal
3538 	 * with concern 3. Also for both Concern 1 and 3 an
3539 	 * application limited period could end the measurement
3540 	 * earlier.
3541 	 *
3542 	 * So lets calculate the BDP with the "known" b/w using
3543 	 * the SRTT has our rtt and then multiply it by the
3544 	 * goal.
3545 	 */
3546 	bw = rack_get_bw(rack);
3547 	srtt = (uint64_t)tp->t_srtt;
3548 	len = bw * srtt;
3549 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3550 	len *= max(1, rack_goal_bdp);
3551 	/* Now we need to round up to the nearest MSS */
3552 	len = roundup(len, segsiz);
3553 	if (rack_min_measure_usec) {
3554 		/* Now calculate our min length for this b/w */
3555 		tim = rack_min_measure_usec;
3556 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3557 		if (minl == 0)
3558 			minl = 1;
3559 		minl = roundup(minl, segsiz);
3560 		if (len < minl)
3561 			len = minl;
3562 	}
3563 	/*
3564 	 * Now if we have a very small window we want
3565 	 * to attempt to get the window that is
3566 	 * as small as possible. This happens on
3567 	 * low b/w connections and we don't want to
3568 	 * span huge numbers of rtt's between measurements.
3569 	 *
3570 	 * We basically include 2 over our "MIN window" so
3571 	 * that the measurement can be shortened (possibly) by
3572 	 * an ack'ed packet.
3573 	 */
3574 	if (len < def_len)
3575 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3576 	else
3577 		return (max((uint32_t)len, def_len));
3578 
3579 }
3580 
3581 static int
3582 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3583 {
3584 	uint32_t tim, srtts, segsiz;
3585 
3586 	/*
3587 	 * Has enough time passed for the GP measurement to be valid?
3588 	 */
3589 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3590 		/* Not enough bytes yet */
3591 		return (0);
3592 	}
3593 	if ((tp->snd_max == tp->snd_una) ||
3594 	    (th_ack == tp->snd_max)){
3595 		/*
3596 		 * All is acked quality of all acked is
3597 		 * usually low or medium, but we in theory could split
3598 		 * all acked into two cases, where you got
3599 		 * a signifigant amount of your window and
3600 		 * where you did not. For now we leave it
3601 		 * but it is something to contemplate in the
3602 		 * future. The danger here is that delayed ack
3603 		 * is effecting the last byte (which is a 50:50 chance).
3604 		 */
3605 		*quality = RACK_QUALITY_ALLACKED;
3606 		return (1);
3607 	}
3608 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3609 		/*
3610 		 * We obtained our entire window of data we wanted
3611 		 * no matter if we are in recovery or not then
3612 		 * its ok since expanding the window does not
3613 		 * make things fuzzy (or at least not as much).
3614 		 */
3615 		*quality = RACK_QUALITY_HIGH;
3616 		return (1);
3617 	}
3618 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3619 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3620 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3621 		/* Not enough bytes yet */
3622 		return (0);
3623 	}
3624 	if (rack->r_ctl.rc_first_appl &&
3625 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3626 		/*
3627 		 * We are up to the app limited send point
3628 		 * we have to measure irrespective of the time..
3629 		 */
3630 		*quality = RACK_QUALITY_APPLIMITED;
3631 		return (1);
3632 	}
3633 	/* Now what about time? */
3634 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3635 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3636 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3637 		/*
3638 		 * We do not allow a measurement if we are in recovery
3639 		 * that would shrink the goodput window we wanted.
3640 		 * This is to prevent cloudyness of when the last send
3641 		 * was actually made.
3642 		 */
3643 		*quality = RACK_QUALITY_HIGH;
3644 		return (1);
3645 	}
3646 	/* Nope not even a full SRTT has passed */
3647 	return (0);
3648 }
3649 
3650 static void
3651 rack_log_timely(struct tcp_rack *rack,
3652 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3653 		uint64_t up_bnd, int line, uint8_t method)
3654 {
3655 	if (tcp_bblogging_on(rack->rc_tp)) {
3656 		union tcp_log_stackspecific log;
3657 		struct timeval tv;
3658 
3659 		memset(&log, 0, sizeof(log));
3660 		log.u_bbr.flex1 = logged;
3661 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3662 		log.u_bbr.flex2 <<= 4;
3663 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3664 		log.u_bbr.flex2 <<= 4;
3665 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3666 		log.u_bbr.flex2 <<= 4;
3667 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3668 		log.u_bbr.flex3 = rack->rc_gp_incr;
3669 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3670 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3671 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3672 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3673 		log.u_bbr.flex8 = method;
3674 		log.u_bbr.cur_del_rate = cur_bw;
3675 		log.u_bbr.delRate = low_bnd;
3676 		log.u_bbr.bw_inuse = up_bnd;
3677 		log.u_bbr.rttProp = rack_get_bw(rack);
3678 		log.u_bbr.pkt_epoch = line;
3679 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3680 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3681 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3682 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3683 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3684 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3685 		log.u_bbr.cwnd_gain <<= 1;
3686 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3687 		log.u_bbr.cwnd_gain <<= 1;
3688 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3689 		log.u_bbr.cwnd_gain <<= 1;
3690 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3691 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3692 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3693 		    &rack->rc_inp->inp_socket->so_rcv,
3694 		    &rack->rc_inp->inp_socket->so_snd,
3695 		    TCP_TIMELY_WORK, 0,
3696 		    0, &log, false, &tv);
3697 	}
3698 }
3699 
3700 static int
3701 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3702 {
3703 	/*
3704 	 * Before we increase we need to know if
3705 	 * the estimate just made was less than
3706 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3707 	 *
3708 	 * If we already are pacing at a fast enough
3709 	 * rate to push us faster there is no sense of
3710 	 * increasing.
3711 	 *
3712 	 * We first caculate our actual pacing rate (ss or ca multiplier
3713 	 * times our cur_bw).
3714 	 *
3715 	 * Then we take the last measured rate and multipy by our
3716 	 * maximum pacing overage to give us a max allowable rate.
3717 	 *
3718 	 * If our act_rate is smaller than our max_allowable rate
3719 	 * then we should increase. Else we should hold steady.
3720 	 *
3721 	 */
3722 	uint64_t act_rate, max_allow_rate;
3723 
3724 	if (rack_timely_no_stopping)
3725 		return (1);
3726 
3727 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3728 		/*
3729 		 * Initial startup case or
3730 		 * everything is acked case.
3731 		 */
3732 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3733 				__LINE__, 9);
3734 		return (1);
3735 	}
3736 	if (mult <= 100) {
3737 		/*
3738 		 * We can always pace at or slightly above our rate.
3739 		 */
3740 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3741 				__LINE__, 9);
3742 		return (1);
3743 	}
3744 	act_rate = cur_bw * (uint64_t)mult;
3745 	act_rate /= 100;
3746 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3747 	max_allow_rate /= 100;
3748 	if (act_rate < max_allow_rate) {
3749 		/*
3750 		 * Here the rate we are actually pacing at
3751 		 * is smaller than 10% above our last measurement.
3752 		 * This means we are pacing below what we would
3753 		 * like to try to achieve (plus some wiggle room).
3754 		 */
3755 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3756 				__LINE__, 9);
3757 		return (1);
3758 	} else {
3759 		/*
3760 		 * Here we are already pacing at least rack_max_per_above(10%)
3761 		 * what we are getting back. This indicates most likely
3762 		 * that we are being limited (cwnd/rwnd/app) and can't
3763 		 * get any more b/w. There is no sense of trying to
3764 		 * raise up the pacing rate its not speeding us up
3765 		 * and we already are pacing faster than we are getting.
3766 		 */
3767 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3768 				__LINE__, 8);
3769 		return (0);
3770 	}
3771 }
3772 
3773 static void
3774 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3775 {
3776 	/*
3777 	 * When we drag bottom, we want to assure
3778 	 * that no multiplier is below 1.0, if so
3779 	 * we want to restore it to at least that.
3780 	 */
3781 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3782 		/* This is unlikely we usually do not touch recovery */
3783 		rack->r_ctl.rack_per_of_gp_rec = 100;
3784 	}
3785 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3786 		rack->r_ctl.rack_per_of_gp_ca = 100;
3787 	}
3788 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3789 		rack->r_ctl.rack_per_of_gp_ss = 100;
3790 	}
3791 }
3792 
3793 static void
3794 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3795 {
3796 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3797 		rack->r_ctl.rack_per_of_gp_ca = 100;
3798 	}
3799 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3800 		rack->r_ctl.rack_per_of_gp_ss = 100;
3801 	}
3802 }
3803 
3804 static void
3805 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3806 {
3807 	int32_t  calc, logged, plus;
3808 
3809 	logged = 0;
3810 
3811 	if (override) {
3812 		/*
3813 		 * override is passed when we are
3814 		 * loosing b/w and making one last
3815 		 * gasp at trying to not loose out
3816 		 * to a new-reno flow.
3817 		 */
3818 		goto extra_boost;
3819 	}
3820 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3821 	if (rack->rc_gp_incr &&
3822 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3823 		/*
3824 		 * Reset and get 5 strokes more before the boost. Note
3825 		 * that the count is 0 based so we have to add one.
3826 		 */
3827 extra_boost:
3828 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3829 		rack->rc_gp_timely_inc_cnt = 0;
3830 	} else
3831 		plus = (uint32_t)rack_gp_increase_per;
3832 	/* Must be at least 1% increase for true timely increases */
3833 	if ((plus < 1) &&
3834 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3835 		plus = 1;
3836 	if (rack->rc_gp_saw_rec &&
3837 	    (rack->rc_gp_no_rec_chg == 0) &&
3838 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3839 				  rack->r_ctl.rack_per_of_gp_rec)) {
3840 		/* We have been in recovery ding it too */
3841 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3842 		if (calc > 0xffff)
3843 			calc = 0xffff;
3844 		logged |= 1;
3845 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3846 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3847 		    (rack->rc_dragged_bottom == 0) &&
3848 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3849 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3850 	}
3851 	if (rack->rc_gp_saw_ca &&
3852 	    (rack->rc_gp_saw_ss == 0) &&
3853 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3854 				  rack->r_ctl.rack_per_of_gp_ca)) {
3855 		/* In CA */
3856 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3857 		if (calc > 0xffff)
3858 			calc = 0xffff;
3859 		logged |= 2;
3860 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3861 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3862 		    (rack->rc_dragged_bottom == 0) &&
3863 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3864 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3865 	}
3866 	if (rack->rc_gp_saw_ss &&
3867 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3868 				  rack->r_ctl.rack_per_of_gp_ss)) {
3869 		/* In SS */
3870 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3871 		if (calc > 0xffff)
3872 			calc = 0xffff;
3873 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3874 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3875 		    (rack->rc_dragged_bottom == 0) &&
3876 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3877 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3878 		logged |= 4;
3879 	}
3880 	if (logged &&
3881 	    (rack->rc_gp_incr == 0)){
3882 		/* Go into increment mode */
3883 		rack->rc_gp_incr = 1;
3884 		rack->rc_gp_timely_inc_cnt = 0;
3885 	}
3886 	if (rack->rc_gp_incr &&
3887 	    logged &&
3888 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3889 		rack->rc_gp_timely_inc_cnt++;
3890 	}
3891 	rack_log_timely(rack,  logged, plus, 0, 0,
3892 			__LINE__, 1);
3893 }
3894 
3895 static uint32_t
3896 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3897 {
3898 	/*-
3899 	 * norm_grad = rtt_diff / minrtt;
3900 	 * new_per = curper * (1 - B * norm_grad)
3901 	 *
3902 	 * B = rack_gp_decrease_per (default 80%)
3903 	 * rtt_dif = input var current rtt-diff
3904 	 * curper = input var current percentage
3905 	 * minrtt = from rack filter
3906 	 *
3907 	 * In order to do the floating point calculations above we
3908 	 * do an integer conversion. The code looks confusing so let me
3909 	 * translate it into something that use more variables and
3910 	 * is clearer for us humans :)
3911 	 *
3912 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
3913 	 * uint32_t perf;
3914 	 *
3915 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3916 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3917 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3918 	 * inverse /= 1000000;
3919 	 * reduce_by = (1000000 - inverse);
3920 	 * final_result = (cur_per * reduce_by) / 1000000;
3921 	 * perf = (uint32_t)final_result;
3922 	 */
3923 	uint64_t perf;
3924 
3925 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3926 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3927 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3928 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3929 		     (uint64_t)1000000)) /
3930 		(uint64_t)1000000);
3931 	if (perf > curper) {
3932 		/* TSNH */
3933 		perf = curper - 1;
3934 	}
3935 	return ((uint32_t)perf);
3936 }
3937 
3938 static uint32_t
3939 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3940 {
3941 	/*
3942 	 *                                   highrttthresh
3943 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3944 	 *                                     gp_srtt
3945 	 *
3946 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
3947 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3948 	 */
3949 	uint64_t perf;
3950 	uint32_t highrttthresh;
3951 
3952 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3953 
3954 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3955 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3956 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3957 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3958 	if (tcp_bblogging_on(rack->rc_tp)) {
3959 		uint64_t log1;
3960 
3961 		log1 = rtt;
3962 		log1 <<= 32;
3963 		log1 |= highrttthresh;
3964 		rack_log_timely(rack,
3965 				rack_gp_decrease_per,
3966 				(uint64_t)curper,
3967 				log1,
3968 				perf,
3969 				__LINE__,
3970 				15);
3971 	}
3972 	return (perf);
3973 }
3974 
3975 static void
3976 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3977 {
3978 	uint64_t logvar, logvar2, logvar3;
3979 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3980 
3981 	if (rack->rc_gp_incr) {
3982 		/* Turn off increment counting */
3983 		rack->rc_gp_incr = 0;
3984 		rack->rc_gp_timely_inc_cnt = 0;
3985 	}
3986 	ss_red = ca_red = rec_red = 0;
3987 	logged = 0;
3988 	/* Calculate the reduction value */
3989 	if (rtt_diff < 0) {
3990 		rtt_diff *= -1;
3991 	}
3992 	/* Must be at least 1% reduction */
3993 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3994 		/* We have been in recovery ding it too */
3995 		if (timely_says == 2) {
3996 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3997 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3998 			if (alt < new_per)
3999 				val = alt;
4000 			else
4001 				val = new_per;
4002 		} else
4003 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4004 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4005 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4006 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4007 		} else {
4008 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4009 			rec_red = 0;
4010 		}
4011 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4012 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4013 		logged |= 1;
4014 	}
4015 	if (rack->rc_gp_saw_ss) {
4016 		/* Sent in SS */
4017 		if (timely_says == 2) {
4018 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4019 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4020 			if (alt < new_per)
4021 				val = alt;
4022 			else
4023 				val = new_per;
4024 		} else
4025 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4026 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4027 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4028 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4029 		} else {
4030 			ss_red = new_per;
4031 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4032 			logvar = new_per;
4033 			logvar <<= 32;
4034 			logvar |= alt;
4035 			logvar2 = (uint32_t)rtt;
4036 			logvar2 <<= 32;
4037 			logvar2 |= (uint32_t)rtt_diff;
4038 			logvar3 = rack_gp_rtt_maxmul;
4039 			logvar3 <<= 32;
4040 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4041 			rack_log_timely(rack, timely_says,
4042 					logvar2, logvar3,
4043 					logvar, __LINE__, 10);
4044 		}
4045 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4046 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4047 		logged |= 4;
4048 	} else if (rack->rc_gp_saw_ca) {
4049 		/* Sent in CA */
4050 		if (timely_says == 2) {
4051 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4052 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4053 			if (alt < new_per)
4054 				val = alt;
4055 			else
4056 				val = new_per;
4057 		} else
4058 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4059 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4060 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4061 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4062 		} else {
4063 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4064 			ca_red = 0;
4065 			logvar = new_per;
4066 			logvar <<= 32;
4067 			logvar |= alt;
4068 			logvar2 = (uint32_t)rtt;
4069 			logvar2 <<= 32;
4070 			logvar2 |= (uint32_t)rtt_diff;
4071 			logvar3 = rack_gp_rtt_maxmul;
4072 			logvar3 <<= 32;
4073 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4074 			rack_log_timely(rack, timely_says,
4075 					logvar2, logvar3,
4076 					logvar, __LINE__, 10);
4077 		}
4078 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4079 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4080 		logged |= 2;
4081 	}
4082 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4083 		rack->rc_gp_timely_dec_cnt++;
4084 		if (rack_timely_dec_clear &&
4085 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4086 			rack->rc_gp_timely_dec_cnt = 0;
4087 	}
4088 	logvar = ss_red;
4089 	logvar <<= 32;
4090 	logvar |= ca_red;
4091 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4092 			__LINE__, 2);
4093 }
4094 
4095 static void
4096 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4097 		     uint32_t rtt, uint32_t line, uint8_t reas)
4098 {
4099 	if (tcp_bblogging_on(rack->rc_tp)) {
4100 		union tcp_log_stackspecific log;
4101 		struct timeval tv;
4102 
4103 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4104 		log.u_bbr.flex1 = line;
4105 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4106 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4107 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4108 		log.u_bbr.flex5 = rtt;
4109 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4110 		log.u_bbr.flex6 <<= 1;
4111 		log.u_bbr.flex6 |= rack->forced_ack;
4112 		log.u_bbr.flex6 <<= 1;
4113 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4114 		log.u_bbr.flex6 <<= 1;
4115 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4116 		log.u_bbr.flex6 <<= 1;
4117 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4118 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4119 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4120 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4121 		log.u_bbr.flex8 = reas;
4122 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4123 		log.u_bbr.delRate = rack_get_bw(rack);
4124 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4125 		log.u_bbr.cur_del_rate <<= 32;
4126 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4127 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4128 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4129 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4130 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4131 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4132 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4133 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4134 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4135 		log.u_bbr.rttProp = us_cts;
4136 		log.u_bbr.rttProp <<= 32;
4137 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4138 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4139 		    &rack->rc_inp->inp_socket->so_rcv,
4140 		    &rack->rc_inp->inp_socket->so_snd,
4141 		    BBR_LOG_RTT_SHRINKS, 0,
4142 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4143 	}
4144 }
4145 
4146 static void
4147 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4148 {
4149 	uint64_t bwdp;
4150 
4151 	bwdp = rack_get_bw(rack);
4152 	bwdp *= (uint64_t)rtt;
4153 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4154 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4155 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4156 		/*
4157 		 * A window protocol must be able to have 4 packets
4158 		 * outstanding as the floor in order to function
4159 		 * (especially considering delayed ack :D).
4160 		 */
4161 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4162 	}
4163 }
4164 
4165 static void
4166 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4167 {
4168 	/**
4169 	 * ProbeRTT is a bit different in rack_pacing than in
4170 	 * BBR. It is like BBR in that it uses the lowering of
4171 	 * the RTT as a signal that we saw something new and
4172 	 * counts from there for how long between. But it is
4173 	 * different in that its quite simple. It does not
4174 	 * play with the cwnd and wait until we get down
4175 	 * to N segments outstanding and hold that for
4176 	 * 200ms. Instead it just sets the pacing reduction
4177 	 * rate to a set percentage (70 by default) and hold
4178 	 * that for a number of recent GP Srtt's.
4179 	 */
4180 	uint32_t segsiz;
4181 
4182 	if (rack->rc_gp_dyn_mul == 0)
4183 		return;
4184 
4185 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4186 		/* We are idle */
4187 		return;
4188 	}
4189 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4190 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4191 		/*
4192 		 * Stop the goodput now, the idea here is
4193 		 * that future measurements with in_probe_rtt
4194 		 * won't register if they are not greater so
4195 		 * we want to get what info (if any) is available
4196 		 * now.
4197 		 */
4198 		rack_do_goodput_measurement(rack->rc_tp, rack,
4199 					    rack->rc_tp->snd_una, __LINE__,
4200 					    RACK_QUALITY_PROBERTT);
4201 	}
4202 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4203 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4204 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4205 		     rack->r_ctl.rc_pace_min_segs);
4206 	rack->in_probe_rtt = 1;
4207 	rack->measure_saw_probe_rtt = 1;
4208 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4209 	rack->r_ctl.rc_time_probertt_starts = 0;
4210 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4211 	if (rack_probertt_use_min_rtt_entry)
4212 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4213 	else
4214 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4215 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4216 			     __LINE__, RACK_RTTS_ENTERPROBE);
4217 }
4218 
4219 static void
4220 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4221 {
4222 	struct rack_sendmap *rsm;
4223 	uint32_t segsiz;
4224 
4225 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4226 		     rack->r_ctl.rc_pace_min_segs);
4227 	rack->in_probe_rtt = 0;
4228 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4229 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4230 		/*
4231 		 * Stop the goodput now, the idea here is
4232 		 * that future measurements with in_probe_rtt
4233 		 * won't register if they are not greater so
4234 		 * we want to get what info (if any) is available
4235 		 * now.
4236 		 */
4237 		rack_do_goodput_measurement(rack->rc_tp, rack,
4238 					    rack->rc_tp->snd_una, __LINE__,
4239 					    RACK_QUALITY_PROBERTT);
4240 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4241 		/*
4242 		 * We don't have enough data to make a measurement.
4243 		 * So lets just stop and start here after exiting
4244 		 * probe-rtt. We probably are not interested in
4245 		 * the results anyway.
4246 		 */
4247 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4248 	}
4249 	/*
4250 	 * Measurements through the current snd_max are going
4251 	 * to be limited by the slower pacing rate.
4252 	 *
4253 	 * We need to mark these as app-limited so we
4254 	 * don't collapse the b/w.
4255 	 */
4256 	rsm = tqhash_max(rack->r_ctl.tqh);
4257 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4258 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4259 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4260 		else {
4261 			/*
4262 			 * Go out to the end app limited and mark
4263 			 * this new one as next and move the end_appl up
4264 			 * to this guy.
4265 			 */
4266 			if (rack->r_ctl.rc_end_appl)
4267 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4268 			rack->r_ctl.rc_end_appl = rsm;
4269 		}
4270 		rsm->r_flags |= RACK_APP_LIMITED;
4271 		rack->r_ctl.rc_app_limited_cnt++;
4272 	}
4273 	/*
4274 	 * Now, we need to examine our pacing rate multipliers.
4275 	 * If its under 100%, we need to kick it back up to
4276 	 * 100%. We also don't let it be over our "max" above
4277 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4278 	 * Note setting clamp_atexit_prtt to 0 has the effect
4279 	 * of setting CA/SS to 100% always at exit (which is
4280 	 * the default behavior).
4281 	 */
4282 	if (rack_probertt_clear_is) {
4283 		rack->rc_gp_incr = 0;
4284 		rack->rc_gp_bwred = 0;
4285 		rack->rc_gp_timely_inc_cnt = 0;
4286 		rack->rc_gp_timely_dec_cnt = 0;
4287 	}
4288 	/* Do we do any clamping at exit? */
4289 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4290 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4291 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4292 	}
4293 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4294 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4295 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4296 	}
4297 	/*
4298 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4299 	 * after exiting.
4300 	 */
4301 	rack->r_ctl.rc_rtt_diff = 0;
4302 
4303 	/* Clear all flags so we start fresh */
4304 	rack->rc_tp->t_bytes_acked = 0;
4305 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4306 	/*
4307 	 * If configured to, set the cwnd and ssthresh to
4308 	 * our targets.
4309 	 */
4310 	if (rack_probe_rtt_sets_cwnd) {
4311 		uint64_t ebdp;
4312 		uint32_t setto;
4313 
4314 		/* Set ssthresh so we get into CA once we hit our target */
4315 		if (rack_probertt_use_min_rtt_exit == 1) {
4316 			/* Set to min rtt */
4317 			rack_set_prtt_target(rack, segsiz,
4318 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4319 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4320 			/* Set to current gp rtt */
4321 			rack_set_prtt_target(rack, segsiz,
4322 					     rack->r_ctl.rc_gp_srtt);
4323 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4324 			/* Set to entry gp rtt */
4325 			rack_set_prtt_target(rack, segsiz,
4326 					     rack->r_ctl.rc_entry_gp_rtt);
4327 		} else {
4328 			uint64_t sum;
4329 			uint32_t setval;
4330 
4331 			sum = rack->r_ctl.rc_entry_gp_rtt;
4332 			sum *= 10;
4333 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4334 			if (sum >= 20) {
4335 				/*
4336 				 * A highly buffered path needs
4337 				 * cwnd space for timely to work.
4338 				 * Lets set things up as if
4339 				 * we are heading back here again.
4340 				 */
4341 				setval = rack->r_ctl.rc_entry_gp_rtt;
4342 			} else if (sum >= 15) {
4343 				/*
4344 				 * Lets take the smaller of the
4345 				 * two since we are just somewhat
4346 				 * buffered.
4347 				 */
4348 				setval = rack->r_ctl.rc_gp_srtt;
4349 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4350 					setval = rack->r_ctl.rc_entry_gp_rtt;
4351 			} else {
4352 				/*
4353 				 * Here we are not highly buffered
4354 				 * and should pick the min we can to
4355 				 * keep from causing loss.
4356 				 */
4357 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4358 			}
4359 			rack_set_prtt_target(rack, segsiz,
4360 					     setval);
4361 		}
4362 		if (rack_probe_rtt_sets_cwnd > 1) {
4363 			/* There is a percentage here to boost */
4364 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4365 			ebdp *= rack_probe_rtt_sets_cwnd;
4366 			ebdp /= 100;
4367 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4368 		} else
4369 			setto = rack->r_ctl.rc_target_probertt_flight;
4370 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4371 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4372 			/* Enforce a min */
4373 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4374 		}
4375 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4376 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4377 	}
4378 	rack_log_rtt_shrinks(rack,  us_cts,
4379 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4380 			     __LINE__, RACK_RTTS_EXITPROBE);
4381 	/* Clear times last so log has all the info */
4382 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4383 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4384 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4385 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4386 }
4387 
4388 static void
4389 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4390 {
4391 	/* Check in on probe-rtt */
4392 	if (rack->rc_gp_filled == 0) {
4393 		/* We do not do p-rtt unless we have gp measurements */
4394 		return;
4395 	}
4396 	if (rack->in_probe_rtt) {
4397 		uint64_t no_overflow;
4398 		uint32_t endtime, must_stay;
4399 
4400 		if (rack->r_ctl.rc_went_idle_time &&
4401 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4402 			/*
4403 			 * We went idle during prtt, just exit now.
4404 			 */
4405 			rack_exit_probertt(rack, us_cts);
4406 		} else if (rack_probe_rtt_safety_val &&
4407 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4408 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4409 			/*
4410 			 * Probe RTT safety value triggered!
4411 			 */
4412 			rack_log_rtt_shrinks(rack,  us_cts,
4413 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4414 					     __LINE__, RACK_RTTS_SAFETY);
4415 			rack_exit_probertt(rack, us_cts);
4416 		}
4417 		/* Calculate the max we will wait */
4418 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4419 		if (rack->rc_highly_buffered)
4420 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4421 		/* Calculate the min we must wait */
4422 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4423 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4424 		    TSTMP_LT(us_cts, endtime)) {
4425 			uint32_t calc;
4426 			/* Do we lower more? */
4427 no_exit:
4428 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4429 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4430 			else
4431 				calc = 0;
4432 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4433 			if (calc) {
4434 				/* Maybe */
4435 				calc *= rack_per_of_gp_probertt_reduce;
4436 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4437 				/* Limit it too */
4438 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4439 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4440 			}
4441 			/* We must reach target or the time set */
4442 			return;
4443 		}
4444 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4445 			if ((TSTMP_LT(us_cts, must_stay) &&
4446 			     rack->rc_highly_buffered) ||
4447 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4448 			      rack->r_ctl.rc_target_probertt_flight)) {
4449 				/* We are not past the must_stay time */
4450 				goto no_exit;
4451 			}
4452 			rack_log_rtt_shrinks(rack,  us_cts,
4453 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4454 					     __LINE__, RACK_RTTS_REACHTARGET);
4455 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4456 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4457 				rack->r_ctl.rc_time_probertt_starts = 1;
4458 			/* Restore back to our rate we want to pace at in prtt */
4459 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4460 		}
4461 		/*
4462 		 * Setup our end time, some number of gp_srtts plus 200ms.
4463 		 */
4464 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4465 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4466 		if (rack_probertt_gpsrtt_cnt_div)
4467 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4468 		else
4469 			endtime = 0;
4470 		endtime += rack_min_probertt_hold;
4471 		endtime += rack->r_ctl.rc_time_probertt_starts;
4472 		if (TSTMP_GEQ(us_cts,  endtime)) {
4473 			/* yes, exit probertt */
4474 			rack_exit_probertt(rack, us_cts);
4475 		}
4476 
4477 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
4478 		/* Go into probertt, its been too long since we went lower */
4479 		rack_enter_probertt(rack, us_cts);
4480 	}
4481 }
4482 
4483 static void
4484 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4485 		       uint32_t rtt, int32_t rtt_diff)
4486 {
4487 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4488 	uint32_t losses;
4489 
4490 	if ((rack->rc_gp_dyn_mul == 0) ||
4491 	    (rack->use_fixed_rate) ||
4492 	    (rack->in_probe_rtt) ||
4493 	    (rack->rc_always_pace == 0)) {
4494 		/* No dynamic GP multiplier in play */
4495 		return;
4496 	}
4497 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4498 	cur_bw = rack_get_bw(rack);
4499 	/* Calculate our up and down range */
4500 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4501 	up_bnd /= 100;
4502 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4503 
4504 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4505 	subfr /= 100;
4506 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4507 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4508 		/*
4509 		 * This is the case where our RTT is above
4510 		 * the max target and we have been configured
4511 		 * to just do timely no bonus up stuff in that case.
4512 		 *
4513 		 * There are two configurations, set to 1, and we
4514 		 * just do timely if we are over our max. If its
4515 		 * set above 1 then we slam the multipliers down
4516 		 * to 100 and then decrement per timely.
4517 		 */
4518 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4519 				__LINE__, 3);
4520 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4521 			rack_validate_multipliers_at_or_below_100(rack);
4522 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4523 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4524 		/*
4525 		 * We are decreasing this is a bit complicated this
4526 		 * means we are loosing ground. This could be
4527 		 * because another flow entered and we are competing
4528 		 * for b/w with it. This will push the RTT up which
4529 		 * makes timely unusable unless we want to get shoved
4530 		 * into a corner and just be backed off (the age
4531 		 * old problem with delay based CC).
4532 		 *
4533 		 * On the other hand if it was a route change we
4534 		 * would like to stay somewhat contained and not
4535 		 * blow out the buffers.
4536 		 */
4537 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4538 				__LINE__, 3);
4539 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4540 		if (rack->rc_gp_bwred == 0) {
4541 			/* Go into reduction counting */
4542 			rack->rc_gp_bwred = 1;
4543 			rack->rc_gp_timely_dec_cnt = 0;
4544 		}
4545 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4546 			/*
4547 			 * Push another time with a faster pacing
4548 			 * to try to gain back (we include override to
4549 			 * get a full raise factor).
4550 			 */
4551 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4552 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4553 			    (timely_says == 0) ||
4554 			    (rack_down_raise_thresh == 0)) {
4555 				/*
4556 				 * Do an override up in b/w if we were
4557 				 * below the threshold or if the threshold
4558 				 * is zero we always do the raise.
4559 				 */
4560 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4561 			} else {
4562 				/* Log it stays the same */
4563 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4564 						__LINE__, 11);
4565 			}
4566 			rack->rc_gp_timely_dec_cnt++;
4567 			/* We are not incrementing really no-count */
4568 			rack->rc_gp_incr = 0;
4569 			rack->rc_gp_timely_inc_cnt = 0;
4570 		} else {
4571 			/*
4572 			 * Lets just use the RTT
4573 			 * information and give up
4574 			 * pushing.
4575 			 */
4576 			goto use_timely;
4577 		}
4578 	} else if ((timely_says != 2) &&
4579 		    !losses &&
4580 		    (last_bw_est > up_bnd)) {
4581 		/*
4582 		 * We are increasing b/w lets keep going, updating
4583 		 * our b/w and ignoring any timely input, unless
4584 		 * of course we are at our max raise (if there is one).
4585 		 */
4586 
4587 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4588 				__LINE__, 3);
4589 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4590 		if (rack->rc_gp_saw_ss &&
4591 		    rack->r_ctl.rack_per_upper_bound_ss &&
4592 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4593 			    /*
4594 			     * In cases where we can't go higher
4595 			     * we should just use timely.
4596 			     */
4597 			    goto use_timely;
4598 		}
4599 		if (rack->rc_gp_saw_ca &&
4600 		    rack->r_ctl.rack_per_upper_bound_ca &&
4601 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4602 			    /*
4603 			     * In cases where we can't go higher
4604 			     * we should just use timely.
4605 			     */
4606 			    goto use_timely;
4607 		}
4608 		rack->rc_gp_bwred = 0;
4609 		rack->rc_gp_timely_dec_cnt = 0;
4610 		/* You get a set number of pushes if timely is trying to reduce */
4611 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4612 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4613 		} else {
4614 			/* Log it stays the same */
4615 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4616 			    __LINE__, 12);
4617 		}
4618 		return;
4619 	} else {
4620 		/*
4621 		 * We are staying between the lower and upper range bounds
4622 		 * so use timely to decide.
4623 		 */
4624 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4625 				__LINE__, 3);
4626 use_timely:
4627 		if (timely_says) {
4628 			rack->rc_gp_incr = 0;
4629 			rack->rc_gp_timely_inc_cnt = 0;
4630 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4631 			    !losses &&
4632 			    (last_bw_est < low_bnd)) {
4633 				/* We are loosing ground */
4634 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4635 				rack->rc_gp_timely_dec_cnt++;
4636 				/* We are not incrementing really no-count */
4637 				rack->rc_gp_incr = 0;
4638 				rack->rc_gp_timely_inc_cnt = 0;
4639 			} else
4640 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4641 		} else {
4642 			rack->rc_gp_bwred = 0;
4643 			rack->rc_gp_timely_dec_cnt = 0;
4644 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4645 		}
4646 	}
4647 }
4648 
4649 static int32_t
4650 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4651 {
4652 	int32_t timely_says;
4653 	uint64_t log_mult, log_rtt_a_diff;
4654 
4655 	log_rtt_a_diff = rtt;
4656 	log_rtt_a_diff <<= 32;
4657 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4658 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4659 		    rack_gp_rtt_maxmul)) {
4660 		/* Reduce the b/w multiplier */
4661 		timely_says = 2;
4662 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4663 		log_mult <<= 32;
4664 		log_mult |= prev_rtt;
4665 		rack_log_timely(rack,  timely_says, log_mult,
4666 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4667 				log_rtt_a_diff, __LINE__, 4);
4668 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4669 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4670 			    max(rack_gp_rtt_mindiv , 1)))) {
4671 		/* Increase the b/w multiplier */
4672 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4673 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4674 			 max(rack_gp_rtt_mindiv , 1));
4675 		log_mult <<= 32;
4676 		log_mult |= prev_rtt;
4677 		timely_says = 0;
4678 		rack_log_timely(rack,  timely_says, log_mult ,
4679 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4680 				log_rtt_a_diff, __LINE__, 5);
4681 	} else {
4682 		/*
4683 		 * Use a gradient to find it the timely gradient
4684 		 * is:
4685 		 * grad = rc_rtt_diff / min_rtt;
4686 		 *
4687 		 * anything below or equal to 0 will be
4688 		 * a increase indication. Anything above
4689 		 * zero is a decrease. Note we take care
4690 		 * of the actual gradient calculation
4691 		 * in the reduction (its not needed for
4692 		 * increase).
4693 		 */
4694 		log_mult = prev_rtt;
4695 		if (rtt_diff <= 0) {
4696 			/*
4697 			 * Rttdiff is less than zero, increase the
4698 			 * b/w multiplier (its 0 or negative)
4699 			 */
4700 			timely_says = 0;
4701 			rack_log_timely(rack,  timely_says, log_mult,
4702 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4703 		} else {
4704 			/* Reduce the b/w multiplier */
4705 			timely_says = 1;
4706 			rack_log_timely(rack,  timely_says, log_mult,
4707 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4708 		}
4709 	}
4710 	return (timely_says);
4711 }
4712 
4713 static __inline int
4714 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4715 {
4716 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4717 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4718 		/**
4719 		 * This covers the case that the
4720 		 * resent is completely inside
4721 		 * the gp range or up to it.
4722 		 *      |----------------|
4723 		 *      |-----| <or>
4724 		 *            |----|
4725 		 *            <or>   |---|
4726 		 */
4727 		return (1);
4728 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4729 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4730 		/**
4731 		 * This covers the case of
4732 		 *      |--------------|
4733 		 *  |-------->|
4734 		 */
4735 		return (1);
4736 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4737 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4738 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4739 
4740 		/**
4741 		 * This covers the case of
4742 		 *      |--------------|
4743 		 *              |-------->|
4744 		 */
4745 		return (1);
4746 	}
4747 	return (0);
4748 }
4749 
4750 static __inline void
4751 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4752 {
4753 
4754 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4755 		return;
4756 	/*
4757 	 * We have a Goodput measurement in progress. Mark
4758 	 * the send if its within the window. If its not
4759 	 * in the window make sure it does not have the mark.
4760 	 */
4761 	if (rack_in_gp_window(tp, rsm))
4762 		rsm->r_flags |= RACK_IN_GP_WIN;
4763 	else
4764 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4765 }
4766 
4767 static __inline void
4768 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4769 {
4770 	/* A GP measurement is ending, clear all marks on the send map*/
4771 	struct rack_sendmap *rsm = NULL;
4772 
4773 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4774 	if (rsm == NULL) {
4775 		rsm = tqhash_min(rack->r_ctl.tqh);
4776 	}
4777 	/* Nothing left? */
4778 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4779 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4780 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4781 	}
4782 }
4783 
4784 
4785 static __inline void
4786 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4787 {
4788 	struct rack_sendmap *rsm = NULL;
4789 
4790 	if (tp->snd_una == tp->snd_max) {
4791 		/* Nothing outstanding yet, nothing to do here */
4792 		return;
4793 	}
4794 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4795 		/*
4796 		 * We are measuring ahead of some outstanding
4797 		 * data. We need to walk through up until we get
4798 		 * to gp_seq marking so that no rsm is set incorrectly
4799 		 * with RACK_IN_GP_WIN.
4800 		 */
4801 		rsm = tqhash_min(rack->r_ctl.tqh);
4802 		while (rsm != NULL) {
4803 			rack_mark_in_gp_win(tp, rsm);
4804 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4805 				break;
4806 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4807 		}
4808 	}
4809 	if (rsm == NULL) {
4810 		/*
4811 		 * Need to find the GP seq, if rsm is
4812 		 * set we stopped as we hit it.
4813 		 */
4814 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4815 		if (rsm == NULL)
4816 			return;
4817 		rack_mark_in_gp_win(tp, rsm);
4818 	}
4819 	/*
4820 	 * Now we may need to mark already sent rsm, ahead of
4821 	 * gput_seq in the window since they may have been sent
4822 	 * *before* we started our measurment. The rsm, if non-null
4823 	 * has been marked (note if rsm would have been NULL we would have
4824 	 * returned in the previous block). So we go to the next, and continue
4825 	 * until we run out of entries or we exceed the gp_ack value.
4826 	 */
4827 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4828 	while (rsm) {
4829 		rack_mark_in_gp_win(tp, rsm);
4830 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4831 			break;
4832 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4833 	}
4834 }
4835 
4836 static void
4837 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4838 			    tcp_seq th_ack, int line, uint8_t quality)
4839 {
4840 	uint64_t tim, bytes_ps, stim, utim;
4841 	uint32_t segsiz, bytes, reqbytes, us_cts;
4842 	int32_t gput, new_rtt_diff, timely_says;
4843 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4844 	int did_add = 0;
4845 
4846 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4847 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4848 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4849 		tim = us_cts - tp->gput_ts;
4850 	else
4851 		tim = 0;
4852 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4853 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4854 	else
4855 		stim = 0;
4856 	/*
4857 	 * Use the larger of the send time or ack time. This prevents us
4858 	 * from being influenced by ack artifacts to come up with too
4859 	 * high of measurement. Note that since we are spanning over many more
4860 	 * bytes in most of our measurements hopefully that is less likely to
4861 	 * occur.
4862 	 */
4863 	if (tim > stim)
4864 		utim = max(tim, 1);
4865 	else
4866 		utim = max(stim, 1);
4867 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4868 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4869 	if ((tim == 0) && (stim == 0)) {
4870 		/*
4871 		 * Invalid measurement time, maybe
4872 		 * all on one ack/one send?
4873 		 */
4874 		bytes = 0;
4875 		bytes_ps = 0;
4876 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4877 					   0, 0, 0, 10, __LINE__, NULL, quality);
4878 		goto skip_measurement;
4879 	}
4880 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4881 		/* We never made a us_rtt measurement? */
4882 		bytes = 0;
4883 		bytes_ps = 0;
4884 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4885 					   0, 0, 0, 10, __LINE__, NULL, quality);
4886 		goto skip_measurement;
4887 	}
4888 	/*
4889 	 * Calculate the maximum possible b/w this connection
4890 	 * could have. We base our calculation on the lowest
4891 	 * rtt we have seen during the measurement and the
4892 	 * largest rwnd the client has given us in that time. This
4893 	 * forms a BDP that is the maximum that we could ever
4894 	 * get to the client. Anything larger is not valid.
4895 	 *
4896 	 * I originally had code here that rejected measurements
4897 	 * where the time was less than 1/2 the latest us_rtt.
4898 	 * But after thinking on that I realized its wrong since
4899 	 * say you had a 150Mbps or even 1Gbps link, and you
4900 	 * were a long way away.. example I am in Europe (100ms rtt)
4901 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4902 	 * bytes my time would be 1.2ms, and yet my rtt would say
4903 	 * the measurement was invalid the time was < 50ms. The
4904 	 * same thing is true for 150Mb (8ms of time).
4905 	 *
4906 	 * A better way I realized is to look at what the maximum
4907 	 * the connection could possibly do. This is gated on
4908 	 * the lowest RTT we have seen and the highest rwnd.
4909 	 * We should in theory never exceed that, if we are
4910 	 * then something on the path is storing up packets
4911 	 * and then feeding them all at once to our endpoint
4912 	 * messing up our measurement.
4913 	 */
4914 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4915 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4916 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4917 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4918 		/* No measurement can be made */
4919 		bytes = 0;
4920 		bytes_ps = 0;
4921 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4922 					   0, 0, 0, 10, __LINE__, NULL, quality);
4923 		goto skip_measurement;
4924 	} else
4925 		bytes = (th_ack - tp->gput_seq);
4926 	bytes_ps = (uint64_t)bytes;
4927 	/*
4928 	 * Don't measure a b/w for pacing unless we have gotten at least
4929 	 * an initial windows worth of data in this measurement interval.
4930 	 *
4931 	 * Small numbers of bytes get badly influenced by delayed ack and
4932 	 * other artifacts. Note we take the initial window or our
4933 	 * defined minimum GP (defaulting to 10 which hopefully is the
4934 	 * IW).
4935 	 */
4936 	if (rack->rc_gp_filled == 0) {
4937 		/*
4938 		 * The initial estimate is special. We
4939 		 * have blasted out an IW worth of packets
4940 		 * without a real valid ack ts results. We
4941 		 * then setup the app_limited_needs_set flag,
4942 		 * this should get the first ack in (probably 2
4943 		 * MSS worth) to be recorded as the timestamp.
4944 		 * We thus allow a smaller number of bytes i.e.
4945 		 * IW - 2MSS.
4946 		 */
4947 		reqbytes -= (2 * segsiz);
4948 		/* Also lets fill previous for our first measurement to be neutral */
4949 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4950 	}
4951 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4952 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4953 					   rack->r_ctl.rc_app_limited_cnt,
4954 					   0, 0, 10, __LINE__, NULL, quality);
4955 		goto skip_measurement;
4956 	}
4957 	/*
4958 	 * We now need to calculate the Timely like status so
4959 	 * we can update (possibly) the b/w multipliers.
4960 	 */
4961 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4962 	if (rack->rc_gp_filled == 0) {
4963 		/* No previous reading */
4964 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4965 	} else {
4966 		if (rack->measure_saw_probe_rtt == 0) {
4967 			/*
4968 			 * We don't want a probertt to be counted
4969 			 * since it will be negative incorrectly. We
4970 			 * expect to be reducing the RTT when we
4971 			 * pace at a slower rate.
4972 			 */
4973 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4974 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4975 		}
4976 	}
4977 	timely_says = rack_make_timely_judgement(rack,
4978 	    rack->r_ctl.rc_gp_srtt,
4979 	    rack->r_ctl.rc_rtt_diff,
4980 	    rack->r_ctl.rc_prev_gp_srtt
4981 	);
4982 	bytes_ps *= HPTS_USEC_IN_SEC;
4983 	bytes_ps /= utim;
4984 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4985 		/*
4986 		 * Something is on path playing
4987 		 * since this b/w is not possible based
4988 		 * on our BDP (highest rwnd and lowest rtt
4989 		 * we saw in the measurement window).
4990 		 *
4991 		 * Another option here would be to
4992 		 * instead skip the measurement.
4993 		 */
4994 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4995 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4996 					   11, __LINE__, NULL, quality);
4997 		bytes_ps = rack->r_ctl.last_max_bw;
4998 	}
4999 	/* We store gp for b/w in bytes per second */
5000 	if (rack->rc_gp_filled == 0) {
5001 		/* Initial measurement */
5002 		if (bytes_ps) {
5003 			rack->r_ctl.gp_bw = bytes_ps;
5004 			rack->rc_gp_filled = 1;
5005 			rack->r_ctl.num_measurements = 1;
5006 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5007 		} else {
5008 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5009 						   rack->r_ctl.rc_app_limited_cnt,
5010 						   0, 0, 10, __LINE__, NULL, quality);
5011 		}
5012 		if (tcp_in_hpts(rack->rc_tp) &&
5013 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5014 			/*
5015 			 * Ok we can't trust the pacer in this case
5016 			 * where we transition from un-paced to paced.
5017 			 * Or for that matter when the burst mitigation
5018 			 * was making a wild guess and got it wrong.
5019 			 * Stop the pacer and clear up all the aggregate
5020 			 * delays etc.
5021 			 */
5022 			tcp_hpts_remove(rack->rc_tp);
5023 			rack->r_ctl.rc_hpts_flags = 0;
5024 			rack->r_ctl.rc_last_output_to = 0;
5025 		}
5026 		did_add = 2;
5027 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5028 		/* Still a small number run an average */
5029 		rack->r_ctl.gp_bw += bytes_ps;
5030 		addpart = rack->r_ctl.num_measurements;
5031 		rack->r_ctl.num_measurements++;
5032 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5033 			/* We have collected enough to move forward */
5034 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5035 		}
5036 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5037 		did_add = 3;
5038 	} else {
5039 		/*
5040 		 * We want to take 1/wma of the goodput and add in to 7/8th
5041 		 * of the old value weighted by the srtt. So if your measurement
5042 		 * period is say 2 SRTT's long you would get 1/4 as the
5043 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5044 		 *
5045 		 * But we must be careful not to take too much i.e. if the
5046 		 * srtt is say 20ms and the measurement is taken over
5047 		 * 400ms our weight would be 400/20 i.e. 20. On the
5048 		 * other hand if we get a measurement over 1ms with a
5049 		 * 10ms rtt we only want to take a much smaller portion.
5050 		 */
5051 		if (rack->r_ctl.num_measurements < 0xff) {
5052 			rack->r_ctl.num_measurements++;
5053 		}
5054 		srtt = (uint64_t)tp->t_srtt;
5055 		if (srtt == 0) {
5056 			/*
5057 			 * Strange why did t_srtt go back to zero?
5058 			 */
5059 			if (rack->r_ctl.rc_rack_min_rtt)
5060 				srtt = rack->r_ctl.rc_rack_min_rtt;
5061 			else
5062 				srtt = HPTS_USEC_IN_MSEC;
5063 		}
5064 		/*
5065 		 * XXXrrs: Note for reviewers, in playing with
5066 		 * dynamic pacing I discovered this GP calculation
5067 		 * as done originally leads to some undesired results.
5068 		 * Basically you can get longer measurements contributing
5069 		 * too much to the WMA. Thus I changed it if you are doing
5070 		 * dynamic adjustments to only do the aportioned adjustment
5071 		 * if we have a very small (time wise) measurement. Longer
5072 		 * measurements just get there weight (defaulting to 1/8)
5073 		 * add to the WMA. We may want to think about changing
5074 		 * this to always do that for both sides i.e. dynamic
5075 		 * and non-dynamic... but considering lots of folks
5076 		 * were playing with this I did not want to change the
5077 		 * calculation per.se. without your thoughts.. Lawerence?
5078 		 * Peter??
5079 		 */
5080 		if (rack->rc_gp_dyn_mul == 0) {
5081 			subpart = rack->r_ctl.gp_bw * utim;
5082 			subpart /= (srtt * 8);
5083 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5084 				/*
5085 				 * The b/w update takes no more
5086 				 * away then 1/2 our running total
5087 				 * so factor it in.
5088 				 */
5089 				addpart = bytes_ps * utim;
5090 				addpart /= (srtt * 8);
5091 			} else {
5092 				/*
5093 				 * Don't allow a single measurement
5094 				 * to account for more than 1/2 of the
5095 				 * WMA. This could happen on a retransmission
5096 				 * where utim becomes huge compared to
5097 				 * srtt (multiple retransmissions when using
5098 				 * the sending rate which factors in all the
5099 				 * transmissions from the first one).
5100 				 */
5101 				subpart = rack->r_ctl.gp_bw / 2;
5102 				addpart = bytes_ps / 2;
5103 			}
5104 			resid_bw = rack->r_ctl.gp_bw - subpart;
5105 			rack->r_ctl.gp_bw = resid_bw + addpart;
5106 			did_add = 1;
5107 		} else {
5108 			if ((utim / srtt) <= 1) {
5109 				/*
5110 				 * The b/w update was over a small period
5111 				 * of time. The idea here is to prevent a small
5112 				 * measurement time period from counting
5113 				 * too much. So we scale it based on the
5114 				 * time so it attributes less than 1/rack_wma_divisor
5115 				 * of its measurement.
5116 				 */
5117 				subpart = rack->r_ctl.gp_bw * utim;
5118 				subpart /= (srtt * rack_wma_divisor);
5119 				addpart = bytes_ps * utim;
5120 				addpart /= (srtt * rack_wma_divisor);
5121 			} else {
5122 				/*
5123 				 * The scaled measurement was long
5124 				 * enough so lets just add in the
5125 				 * portion of the measurement i.e. 1/rack_wma_divisor
5126 				 */
5127 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5128 				addpart = bytes_ps / rack_wma_divisor;
5129 			}
5130 			if ((rack->measure_saw_probe_rtt == 0) ||
5131 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5132 				/*
5133 				 * For probe-rtt we only add it in
5134 				 * if its larger, all others we just
5135 				 * add in.
5136 				 */
5137 				did_add = 1;
5138 				resid_bw = rack->r_ctl.gp_bw - subpart;
5139 				rack->r_ctl.gp_bw = resid_bw + addpart;
5140 			}
5141 		}
5142 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5143 	}
5144 	if ((rack->gp_ready == 0) &&
5145 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5146 		/* We have enough measurements now */
5147 		rack->gp_ready = 1;
5148 		if (rack->dgp_on ||
5149 		    rack->rack_hibeta)
5150 			rack_set_cc_pacing(rack);
5151 		if (rack->defer_options)
5152 			rack_apply_deferred_options(rack);
5153 	}
5154 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5155 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5156 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5157 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
5158 		rack_update_multiplier(rack, timely_says, bytes_ps,
5159 				       rack->r_ctl.rc_gp_srtt,
5160 				       rack->r_ctl.rc_rtt_diff);
5161 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5162 				   rack_get_bw(rack), 3, line, NULL, quality);
5163 	rack_log_pacing_delay_calc(rack,
5164 				   bytes, /* flex2 */
5165 				   tim, /* flex1 */
5166 				   bytes_ps, /* bw_inuse */
5167 				   rack->r_ctl.gp_bw, /* delRate */
5168 				   rack_get_lt_bw(rack), /* rttProp */
5169 				   20, line, NULL, 0);
5170 	/* reset the gp srtt and setup the new prev */
5171 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5172 	/* Record the lost count for the next measurement */
5173 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5174 skip_measurement:
5175 	/*
5176 	 * We restart our diffs based on the gpsrtt in the
5177 	 * measurement window.
5178 	 */
5179 	rack->rc_gp_rtt_set = 0;
5180 	rack->rc_gp_saw_rec = 0;
5181 	rack->rc_gp_saw_ca = 0;
5182 	rack->rc_gp_saw_ss = 0;
5183 	rack->rc_dragged_bottom = 0;
5184 
5185 	if (quality == RACK_QUALITY_HIGH) {
5186 		/*
5187 		 * Gput in the stats world is in kbps where bytes_ps is
5188 		 * bytes per second so we do ((x * 8)/ 1000).
5189 		 */
5190 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5191 #ifdef STATS
5192 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5193 					 gput);
5194 		/*
5195 		 * XXXLAS: This is a temporary hack, and should be
5196 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5197 		 * API to deal with chained VOIs.
5198 		 */
5199 		if (tp->t_stats_gput_prev > 0)
5200 			stats_voi_update_abs_s32(tp->t_stats,
5201 						 VOI_TCP_GPUT_ND,
5202 						 ((gput - tp->t_stats_gput_prev) * 100) /
5203 						 tp->t_stats_gput_prev);
5204 #endif
5205 		tp->t_stats_gput_prev = gput;
5206 	}
5207 	tp->t_flags &= ~TF_GPUTINPROG;
5208 	/*
5209 	 * Now are we app limited now and there is space from where we
5210 	 * were to where we want to go?
5211 	 *
5212 	 * We don't do the other case i.e. non-applimited here since
5213 	 * the next send will trigger us picking up the missing data.
5214 	 */
5215 	if (rack->r_ctl.rc_first_appl &&
5216 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5217 	    rack->r_ctl.rc_app_limited_cnt &&
5218 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5219 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5220 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5221 		/*
5222 		 * Yep there is enough outstanding to make a measurement here.
5223 		 */
5224 		struct rack_sendmap *rsm;
5225 
5226 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5227 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5228 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5229 		rack->app_limited_needs_set = 0;
5230 		tp->gput_seq = th_ack;
5231 		if (rack->in_probe_rtt)
5232 			rack->measure_saw_probe_rtt = 1;
5233 		else if ((rack->measure_saw_probe_rtt) &&
5234 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5235 			rack->measure_saw_probe_rtt = 0;
5236 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5237 			/* There is a full window to gain info from */
5238 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5239 		} else {
5240 			/* We can only measure up to the applimited point */
5241 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5242 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5243 				/*
5244 				 * We don't have enough to make a measurement.
5245 				 */
5246 				tp->t_flags &= ~TF_GPUTINPROG;
5247 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5248 							   0, 0, 0, 6, __LINE__, NULL, quality);
5249 				return;
5250 			}
5251 		}
5252 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5253 			/*
5254 			 * We will get no more data into the SB
5255 			 * this means we need to have the data available
5256 			 * before we start a measurement.
5257 			 */
5258 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5259 				/* Nope not enough data. */
5260 				return;
5261 			}
5262 		}
5263 		tp->t_flags |= TF_GPUTINPROG;
5264 		/*
5265 		 * Now we need to find the timestamp of the send at tp->gput_seq
5266 		 * for the send based measurement.
5267 		 */
5268 		rack->r_ctl.rc_gp_cumack_ts = 0;
5269 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5270 		if (rsm) {
5271 			/* Ok send-based limit is set */
5272 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5273 				/*
5274 				 * Move back to include the earlier part
5275 				 * so our ack time lines up right (this may
5276 				 * make an overlapping measurement but thats
5277 				 * ok).
5278 				 */
5279 				tp->gput_seq = rsm->r_start;
5280 			}
5281 			if (rsm->r_flags & RACK_ACKED) {
5282 				struct rack_sendmap *nrsm;
5283 
5284 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5285 				tp->gput_seq = rsm->r_end;
5286 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5287 				if (nrsm)
5288 					rsm = nrsm;
5289 				else {
5290 					rack->app_limited_needs_set = 1;
5291 				}
5292 			} else
5293 				rack->app_limited_needs_set = 1;
5294 			/* We always go from the first send */
5295 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5296 		} else {
5297 			/*
5298 			 * If we don't find the rsm due to some
5299 			 * send-limit set the current time, which
5300 			 * basically disables the send-limit.
5301 			 */
5302 			struct timeval tv;
5303 
5304 			microuptime(&tv);
5305 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5306 		}
5307 		rack_tend_gp_marks(tp, rack);
5308 		rack_log_pacing_delay_calc(rack,
5309 					   tp->gput_seq,
5310 					   tp->gput_ack,
5311 					   (uint64_t)rsm,
5312 					   tp->gput_ts,
5313 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5314 					   9,
5315 					   __LINE__, rsm, quality);
5316 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5317 	} else {
5318 		/*
5319 		 * To make sure proper timestamp merging occurs, we need to clear
5320 		 * all GP marks if we don't start a measurement.
5321 		 */
5322 		rack_clear_gp_marks(tp, rack);
5323 	}
5324 }
5325 
5326 /*
5327  * CC wrapper hook functions
5328  */
5329 static void
5330 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5331     uint16_t type, int32_t recovery)
5332 {
5333 	uint32_t prior_cwnd, acked;
5334 	struct tcp_log_buffer *lgb = NULL;
5335 	uint8_t labc_to_use, quality;
5336 
5337 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5338 	tp->t_ccv.nsegs = nsegs;
5339 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5340 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5341 		uint32_t max;
5342 
5343 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5344 		if (tp->t_ccv.bytes_this_ack > max) {
5345 			tp->t_ccv.bytes_this_ack = max;
5346 		}
5347 	}
5348 #ifdef STATS
5349 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5350 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5351 #endif
5352 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5353 		/* We will ack all, time
5354 		 * to end any lt_bw_up we
5355 		 * have running until something
5356 		 * new is sent.
5357 		 */
5358 		struct timeval tv;
5359 
5360 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5361 		rack->r_ctl.lt_seq = tp->snd_max;
5362 		(void)tcp_get_usecs(&tv);
5363 		rack->r_ctl.lt_bw_time += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
5364 		rack->lt_bw_up = 0;
5365 	}
5366 	quality = RACK_QUALITY_NONE;
5367 	if ((tp->t_flags & TF_GPUTINPROG) &&
5368 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5369 		/* Measure the Goodput */
5370 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5371 	}
5372 	/* Which way our we limited, if not cwnd limited no advance in CA */
5373 	if (tp->snd_cwnd <= tp->snd_wnd)
5374 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5375 	else
5376 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5377 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5378 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5379 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5380 		/* For the setting of a window past use the actual scwnd we are using */
5381 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5382 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5383 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5384 		}
5385 	} else {
5386 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5387 		tp->t_bytes_acked = 0;
5388 	}
5389 	prior_cwnd = tp->snd_cwnd;
5390 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5391 	    (rack_client_low_buf && rack->client_bufferlvl &&
5392 	    (rack->client_bufferlvl < rack_client_low_buf)))
5393 		labc_to_use = rack->rc_labc;
5394 	else
5395 		labc_to_use = rack_max_abc_post_recovery;
5396 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5397 		union tcp_log_stackspecific log;
5398 		struct timeval tv;
5399 
5400 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5401 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5402 		log.u_bbr.flex1 = th_ack;
5403 		log.u_bbr.flex2 = tp->t_ccv.flags;
5404 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5405 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5406 		log.u_bbr.flex5 = labc_to_use;
5407 		log.u_bbr.flex6 = prior_cwnd;
5408 		log.u_bbr.flex7 = V_tcp_do_newsack;
5409 		log.u_bbr.flex8 = 1;
5410 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5411 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5412 	}
5413 	if (CC_ALGO(tp)->ack_received != NULL) {
5414 		/* XXXLAS: Find a way to live without this */
5415 		tp->t_ccv.curack = th_ack;
5416 		tp->t_ccv.labc = labc_to_use;
5417 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5418 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5419 	}
5420 	if (lgb) {
5421 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5422 	}
5423 	if (rack->r_must_retran) {
5424 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5425 			/*
5426 			 * We now are beyond the rxt point so lets disable
5427 			 * the flag.
5428 			 */
5429 			rack->r_ctl.rc_out_at_rto = 0;
5430 			rack->r_must_retran = 0;
5431 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5432 			/*
5433 			 * Only decrement the rc_out_at_rto if the cwnd advances
5434 			 * at least a whole segment. Otherwise next time the peer
5435 			 * acks, we won't be able to send this generaly happens
5436 			 * when we are in Congestion Avoidance.
5437 			 */
5438 			if (acked <= rack->r_ctl.rc_out_at_rto){
5439 				rack->r_ctl.rc_out_at_rto -= acked;
5440 			} else {
5441 				rack->r_ctl.rc_out_at_rto = 0;
5442 			}
5443 		}
5444 	}
5445 #ifdef STATS
5446 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5447 #endif
5448 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5449 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5450 	}
5451 }
5452 
5453 static void
5454 tcp_rack_partialack(struct tcpcb *tp)
5455 {
5456 	struct tcp_rack *rack;
5457 
5458 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5459 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5460 	/*
5461 	 * If we are doing PRR and have enough
5462 	 * room to send <or> we are pacing and prr
5463 	 * is disabled we will want to see if we
5464 	 * can send data (by setting r_wanted_output to
5465 	 * true).
5466 	 */
5467 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5468 	    rack->rack_no_prr)
5469 		rack->r_wanted_output = 1;
5470 }
5471 
5472 static inline void
5473 rack_set_most_aggr(struct tcp_rack *rack)
5474 {
5475 	rack->r_fill_less_agg = 0;
5476 	/* Once the cwnd as been clamped we don't do fill_cw */
5477 	if (rack->r_cwnd_was_clamped == 0)
5478 		rack->rc_pace_to_cwnd = 1;
5479 	rack->r_pacing_discount = 0;
5480 }
5481 
5482 static inline void
5483 rack_limit_fillcw(struct tcp_rack *rack)
5484 {
5485 	rack->r_fill_less_agg = 1;
5486 	/* Once the cwnd as been clamped we don't do fill_cw */
5487 	if (rack->r_cwnd_was_clamped == 0)
5488 		rack->rc_pace_to_cwnd = 1;
5489 	rack->r_pacing_discount = 0;
5490 }
5491 
5492 static inline void
5493 rack_disable_fillcw(struct tcp_rack *rack)
5494 {
5495 	rack->r_fill_less_agg = 1;
5496 	rack->rc_pace_to_cwnd = 0;
5497 	rack->r_pacing_discount = 0;
5498 }
5499 
5500 static void
5501 rack_client_buffer_level_set(struct tcp_rack *rack)
5502 {
5503 	/*
5504 	 * Only if DGP is on do we do anything that
5505 	 * changes stack behavior. If DGP is off all
5506 	 * we will do is issue a BB log (if BB logging is
5507 	 * on) and return.
5508 	 */
5509 	if (rack->dgp_on == 0) {
5510 		rack_log_pacing_delay_calc(rack, 0, rack->client_bufferlvl,
5511 					   0, 0, 0, 30, __LINE__, NULL, 0);
5512 		return;
5513 	}
5514 	if (IN_RECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.full_dgp_in_rec) {
5515 		goto set_most_agg;
5516 	}
5517 	/*
5518 	 * We are in DGP so what setting should we
5519 	 * apply based on where the client is?
5520 	 */
5521 	switch(rack->r_ctl.rc_dgp_bl_agg) {
5522 	default:
5523 	case DGP_LEVEL0:
5524 set_most_agg:
5525 		rack_set_most_aggr(rack);
5526 		break;
5527 	case DGP_LEVEL1:
5528 		if (rack->client_bufferlvl == 4)
5529 			rack_limit_fillcw(rack);
5530 		else if (rack->client_bufferlvl == 5)
5531 			rack_disable_fillcw(rack);
5532 		else
5533 			rack_set_most_aggr(rack);
5534 		break;
5535 	case DGP_LEVEL2:
5536 		if (rack->client_bufferlvl == 3)
5537 			rack_limit_fillcw(rack);
5538 		else if (rack->client_bufferlvl == 4)
5539 			rack_disable_fillcw(rack);
5540 		else if (rack->client_bufferlvl == 5) {
5541 			rack_disable_fillcw(rack);
5542 			rack->r_pacing_discount = 1;
5543 			rack->r_ctl.pacing_discount_amm = 1;
5544 		} else
5545 			rack_set_most_aggr(rack);
5546 		break;
5547 	case DGP_LEVEL3:
5548 		if (rack->client_bufferlvl == 2)
5549 			rack_limit_fillcw(rack);
5550 		else if (rack->client_bufferlvl == 3)
5551 			rack_disable_fillcw(rack);
5552 		else if (rack->client_bufferlvl == 4) {
5553 			rack_disable_fillcw(rack);
5554 			rack->r_pacing_discount = 1;
5555 			rack->r_ctl.pacing_discount_amm = 1;
5556 		} else if (rack->client_bufferlvl == 5) {
5557 			rack_disable_fillcw(rack);
5558 			rack->r_pacing_discount = 1;
5559 			rack->r_ctl.pacing_discount_amm = 2;
5560 		} else
5561 			rack_set_most_aggr(rack);
5562 		break;
5563 	}
5564 	rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_dgp_bl_agg, rack->client_bufferlvl, 0,
5565 				   0, 0, 30, __LINE__, NULL, 0);
5566 }
5567 
5568 static void
5569 do_rack_check_for_unclamp(struct tcpcb *tp, struct tcp_rack *rack)
5570 {
5571 	/*
5572 	 * Can we unclamp. We unclamp if more than
5573 	 * N rounds have transpired with no loss.
5574 	 */
5575 	uint64_t snds, rxts, rxt_per;
5576 	uint32_t rnds;
5577 
5578 	rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5579 	if ((rack_unclamp_round_thresh > 0) &&
5580 	    (rnds >= rack_unclamp_round_thresh)) {
5581 		snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5582 		KASSERT ((snds > 0), ("rack:%p tp:%p snds:%ju is 0", rack, tp,
5583 		    (uintmax_t)snds));
5584 		rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5585 		rxt_per = rxts * 1000;
5586 		rxt_per /= snds;
5587 		if ((uint32_t)rxt_per <= rack_unclamp_rxt_thresh) {
5588 			/* Unclamp */
5589 			if (tcp_bblogging_on(rack->rc_tp)) {
5590 				union tcp_log_stackspecific log;
5591 				struct timeval tv;
5592 
5593 				memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5594 				log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5595 				log.u_bbr.flex3 = rnds;
5596 				log.u_bbr.flex4 = rack_unclamp_round_thresh;
5597 				log.u_bbr.flex5 = (uint32_t)rxt_per;
5598 				log.u_bbr.flex8 = 6;
5599 				log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5600 				log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5601 				log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5602 				log.u_bbr.applimited = rack->r_ctl.max_clamps;
5603 				log.u_bbr.epoch = rack->r_ctl.clamp_options;
5604 				log.u_bbr.cur_del_rate = rxts;
5605 				log.u_bbr.bw_inuse = rack_get_lt_bw(rack);
5606 				log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5607 				log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5608 				log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5609 				tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5610 					      0, &log, false, NULL, NULL, 0, &tv);
5611 			}
5612 			rack->r_ctl.num_of_clamps_applied = 0;
5613 			rack->r_cwnd_was_clamped = 0;
5614 			rack->excess_rxt_on = 1;
5615 			if (rack->r_ctl.clamp_options) {
5616 				/*
5617 				 * We only allow fillcw to be toggled
5618 				 * if you are setting a max seg too.
5619 				 */
5620 				if (rack->r_ctl.clamp_options & 0x1) {
5621 					if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5622 						/* turn on fill cw  for non-dgp*/
5623 						rack->rc_pace_to_cwnd = 0;
5624 					} else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5625 						/* For DGP we want it off */
5626 						rack->rc_pace_to_cwnd = 1;
5627 					}
5628 				}
5629 			}
5630 			if (rack->dgp_on) {
5631 				/* Reset all multipliers to 100.0 so just the measured bw */
5632 				/* Crash any per boosts down to 100% */
5633 				rack->r_ctl.rack_per_of_gp_rec = 100;
5634 				rack->r_ctl.rack_per_of_gp_ss = 100;
5635 				rack->r_ctl.rack_per_of_gp_ca = 100;
5636 				/* Set in an upper bound for ss/ca % increase */
5637 				rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
5638 				rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
5639 			}
5640 		}
5641 	}
5642 }
5643 
5644 static void
5645 do_rack_excess_rxt(struct tcpcb *tp, struct tcp_rack *rack)
5646 {
5647 	/*
5648 	 * Rack excess rxt accounting is turned on. If we
5649 	 * are above a threshold of rxt's in at least N
5650 	 * rounds, then back off the cwnd and ssthresh
5651 	 * to fit into the long-term b/w.
5652 	 */
5653 	uint64_t snds, rxts, rxt_per, lt_bw, bdp;
5654 	uint32_t rnds, new_cwnd, new_ssthresh, rtt, shared_cwnd_was_enabled = 0;
5655 
5656 	/* Is it shut off by 0 rounds? */
5657 	if (rack_rxt_min_rnds == 0)
5658 		return;
5659 	if ((rack->r_ctl.max_clamps > 0) &&
5660 	    (rack->r_ctl.num_of_clamps_applied >= rack->r_ctl.max_clamps)) {
5661 		/*
5662 		 * The idea, if max_clamps is set, is that if clamping it
5663 		 * N times did not work again, then there is no sense
5664 		 * clamping it again. The link is just a lossy link and
5665 		 * our clamps are doing no good. Turn it off so we don't come
5666 		 * back here again.
5667 		 */
5668 		rack->excess_rxt_on = 0;
5669 		rack->r_cwnd_was_clamped = 0;
5670 		rack->r_ctl.num_of_clamps_applied = 0;
5671 		return;
5672 	}
5673 	snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5674 	rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5675 	rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5676 	/* Has enough rounds progressed for us to re-measure? */
5677 	if ((rnds >= rack_rxt_min_rnds) &&
5678 	    (rack->r_ctl.rxt_threshold > 0)){
5679 		rxt_per = rxts * 1000;
5680 		rxt_per /= snds;
5681 		if (rxt_per >= rack->r_ctl.rxt_threshold) {
5682 			/*
5683 			 * Action required:
5684 			 *  We are above our excess retransmit level, lets
5685 			 *  cut down the cwnd and ssthresh to match the long-term
5686 			 *  b/w we are getting.
5687 			 */
5688 			/* First disable scwnd if enabled */
5689 #ifdef NETFLIX_SHARED_CWND
5690 			rack->rack_enable_scwnd = 0;
5691 			if (rack->r_ctl.rc_scw) {
5692 				uint32_t limit;
5693 
5694 				shared_cwnd_was_enabled = 1;
5695 				if (rack->r_limit_scw)
5696 					limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
5697 				else
5698 					limit = 0;
5699 				tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
5700 							  rack->r_ctl.rc_scw_index,
5701 							  limit);
5702 				rack->r_ctl.rc_scw = NULL;
5703 			}
5704 
5705 #endif
5706 			/* Calculate what the cwnd and ssthresh should be */
5707 			tcp_trace_point(rack->rc_tp, TCP_TP_EXCESS_RXT);
5708 			lt_bw = rack_get_lt_bw(rack);
5709 			if (lt_bw == 0) {
5710 				/*
5711 				 * No lt_bw, lets chop things to one MSS
5712 				 * and the ssthresh to the iwnd.
5713 				 */
5714 reset_to_iw:
5715 				new_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5716 				new_ssthresh = tcp_compute_initwnd(tcp_maxseg(tp));
5717 			} else {
5718 				rtt = rack->rc_rack_rtt;
5719 				if (rtt == 0) {
5720 					/* If we have no rack_rtt drop to the IW situation */
5721 					goto reset_to_iw;
5722 				}
5723 				bdp = lt_bw * (uint64_t)rtt;
5724 				bdp /= HPTS_USEC_IN_SEC;
5725 				new_cwnd = (uint32_t)bdp;
5726 				new_ssthresh = new_cwnd - 1;
5727 				if (new_cwnd < ctf_fixed_maxseg(tp)) {
5728 					/* Rock bottom, goto IW settings  */
5729 					goto reset_to_iw;
5730 				}
5731 			}
5732 			rack->r_cwnd_was_clamped = 1;
5733 			rack->r_ctl.num_of_clamps_applied++;
5734 			/* Reset the counter fromn now */
5735 			tp->t_bytes_acked = 0;
5736 			/*
5737 			 * Now what about options?
5738 			 * We look at the bottom  8 bits:
5739 			 * F = fill cw bit (toggle it if set)
5740 			 * S = Segment bits
5741 			 * M = set max segment bit
5742 			 *
5743 			 * SSSS SSMF
5744 			 */
5745 			if (rack->r_ctl.clamp_options) {
5746 				if (rack->r_ctl.clamp_options & 0x1) {
5747 					if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5748 						/* turn on fill cw  for non-dgp*/
5749 						rack->rc_pace_to_cwnd = 1;
5750 					} else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5751 						/* For DGP we want it off */
5752 						rack->rc_pace_to_cwnd = 0;
5753 					}
5754 				}
5755 			}
5756 			if (rack->dgp_on) {
5757 				/* Reset all multipliers to 100.0 so just the measured bw */
5758 				/* Crash any per boosts down to 100% */
5759 				rack->r_ctl.rack_per_of_gp_rec = 100;
5760 				rack->r_ctl.rack_per_of_gp_ss = 100;
5761 				rack->r_ctl.rack_per_of_gp_ca = 100;
5762 				/* Set in an upper bound for ss/ca % increase */
5763 				rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_clamp_ss_upper;
5764 				rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_clamp_ca_upper;
5765 				/* Now move to the lt_bw */
5766 				rack->r_ctl.gp_bw = lt_bw;
5767 				rack->rc_gp_filled = 1;
5768 				rack->r_ctl.num_measurements = RACK_REQ_AVG;
5769 			}
5770 			if (tcp_bblogging_on(rack->rc_tp)) {
5771 				union tcp_log_stackspecific log;
5772 				struct timeval tv;
5773 
5774 				memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5775 				log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5776 				log.u_bbr.flex1 = new_cwnd;
5777 				log.u_bbr.flex2 = new_ssthresh;
5778 				log.u_bbr.flex3 = rnds;
5779 				log.u_bbr.flex4 = rack_rxt_min_rnds;
5780 				log.u_bbr.flex5 = rtt;
5781 				log.u_bbr.flex6 = shared_cwnd_was_enabled;
5782 				log.u_bbr.flex8 = 5;
5783 				log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5784 				log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5785 				log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5786 				log.u_bbr.applimited = rack->r_ctl.max_clamps;
5787 				log.u_bbr.epoch = rack->r_ctl.clamp_options;
5788 				log.u_bbr.cur_del_rate = rxts;
5789 				log.u_bbr.delRate = snds;
5790 				log.u_bbr.rttProp = rack->r_ctl.rxt_threshold;
5791 				log.u_bbr.bw_inuse = lt_bw;
5792 				log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5793 				log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5794 				log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5795 				tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5796 					       0, &log, false, NULL, NULL, 0, &tv);
5797 			}
5798 			/* Update our point where we did it */
5799 			if (rack->r_ctl.already_had_a_excess == 0) {
5800 				rack->r_ctl.already_had_a_excess = 1;
5801 				counter_u64_add(rack_rxt_clamps_cwnd_uniq, 1);
5802 			}
5803 			counter_u64_add(rack_rxt_clamps_cwnd, 1);
5804 			rack->r_ctl.last_sndbytes = tp->t_sndbytes;
5805 			rack->r_ctl.last_snd_rxt_bytes = tp->t_snd_rxt_bytes;
5806 			rack->r_ctl.last_rnd_rxt_clamped = rack->r_ctl.current_round;
5807 			if (new_cwnd < tp->snd_cwnd)
5808 				tp->snd_cwnd = new_cwnd;
5809 			if (new_ssthresh < tp->snd_ssthresh)
5810 				tp->snd_ssthresh = new_ssthresh;
5811 		}
5812 	}
5813 }
5814 
5815 static void
5816 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5817 {
5818 	struct tcp_rack *rack;
5819 	uint32_t orig_cwnd;
5820 
5821 	orig_cwnd = tp->snd_cwnd;
5822 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5823 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5824 	/* only alert CC if we alerted when we entered */
5825 	if (CC_ALGO(tp)->post_recovery != NULL) {
5826 		tp->t_ccv.curack = th_ack;
5827 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5828 		if (tp->snd_cwnd < tp->snd_ssthresh) {
5829 			/*
5830 			 * Rack has burst control and pacing
5831 			 * so lets not set this any lower than
5832 			 * snd_ssthresh per RFC-6582 (option 2).
5833 			 */
5834 			tp->snd_cwnd = tp->snd_ssthresh;
5835 		}
5836 	}
5837 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5838 		union tcp_log_stackspecific log;
5839 		struct timeval tv;
5840 
5841 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5842 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5843 		log.u_bbr.flex1 = th_ack;
5844 		log.u_bbr.flex2 = tp->t_ccv.flags;
5845 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5846 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5847 		log.u_bbr.flex5 = V_tcp_abc_l_var;
5848 		log.u_bbr.flex6 = orig_cwnd;
5849 		log.u_bbr.flex7 = V_tcp_do_newsack;
5850 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5851 		log.u_bbr.flex8 = 2;
5852 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5853 			       0, &log, false, NULL, __func__, __LINE__, &tv);
5854 	}
5855 	if ((rack->rack_no_prr == 0) &&
5856 	    (rack->no_prr_addback == 0) &&
5857 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
5858 		/*
5859 		 * Suck the next prr cnt back into cwnd, but
5860 		 * only do that if we are not application limited.
5861 		 */
5862 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5863 			/*
5864 			 * We are allowed to add back to the cwnd the amount we did
5865 			 * not get out if:
5866 			 * a) no_prr_addback is off.
5867 			 * b) we are not app limited
5868 			 * c) we are doing prr
5869 			 * <and>
5870 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5871 			 */
5872 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5873 					    rack->r_ctl.rc_prr_sndcnt);
5874 		}
5875 		rack->r_ctl.rc_prr_sndcnt = 0;
5876 		rack_log_to_prr(rack, 1, 0, __LINE__);
5877 	}
5878 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5879 	tp->snd_recover = tp->snd_una;
5880 	if (rack->r_ctl.dsack_persist) {
5881 		rack->r_ctl.dsack_persist--;
5882 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5883 			rack->r_ctl.num_dsack = 0;
5884 		}
5885 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5886 	}
5887 	EXIT_RECOVERY(tp->t_flags);
5888 	if (rack->r_ctl.full_dgp_in_rec)
5889 		rack_client_buffer_level_set(rack);
5890 }
5891 
5892 static void
5893 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5894 {
5895 	struct tcp_rack *rack;
5896 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5897 
5898 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5899 #ifdef STATS
5900 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5901 #endif
5902 	if (IN_RECOVERY(tp->t_flags) == 0) {
5903 		in_rec_at_entry = 0;
5904 		ssthresh_enter = tp->snd_ssthresh;
5905 		cwnd_enter = tp->snd_cwnd;
5906 	} else
5907 		in_rec_at_entry = 1;
5908 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5909 	switch (type) {
5910 	case CC_NDUPACK:
5911 		tp->t_flags &= ~TF_WASFRECOVERY;
5912 		tp->t_flags &= ~TF_WASCRECOVERY;
5913 		if (!IN_FASTRECOVERY(tp->t_flags)) {
5914 			if (rack->dgp_on && rack->r_cwnd_was_clamped) {
5915 				/* Reset the gains so that on exit we will be softer longer */
5916 				rack->r_ctl.rack_per_of_gp_rec = 100;
5917 				rack->r_ctl.rack_per_of_gp_ss = 98;
5918 				rack->r_ctl.rack_per_of_gp_ca = 98;
5919 			}
5920 			rack->r_ctl.rc_prr_delivered = 0;
5921 			rack->r_ctl.rc_prr_out = 0;
5922 			rack->r_fast_output = 0;
5923 			if (rack->rack_no_prr == 0) {
5924 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5925 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5926 			}
5927 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5928 			tp->snd_recover = tp->snd_max;
5929 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5930 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5931 		}
5932 		break;
5933 	case CC_ECN:
5934 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5935 		    /*
5936 		     * Allow ECN reaction on ACK to CWR, if
5937 		     * that data segment was also CE marked.
5938 		     */
5939 		    SEQ_GEQ(ack, tp->snd_recover)) {
5940 			EXIT_CONGRECOVERY(tp->t_flags);
5941 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5942 			rack->r_fast_output = 0;
5943 			tp->snd_recover = tp->snd_max + 1;
5944 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5945 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5946 		}
5947 		break;
5948 	case CC_RTO:
5949 		tp->t_dupacks = 0;
5950 		tp->t_bytes_acked = 0;
5951 		rack->r_fast_output = 0;
5952 		EXIT_RECOVERY(tp->t_flags);
5953 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5954 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5955 		orig_cwnd = tp->snd_cwnd;
5956 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
5957 		rack_log_to_prr(rack, 16, orig_cwnd, line);
5958 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5959 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5960 		break;
5961 	case CC_RTO_ERR:
5962 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5963 		/* RTO was unnecessary, so reset everything. */
5964 		tp->snd_cwnd = tp->snd_cwnd_prev;
5965 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5966 		tp->snd_recover = tp->snd_recover_prev;
5967 		if (tp->t_flags & TF_WASFRECOVERY) {
5968 			ENTER_FASTRECOVERY(tp->t_flags);
5969 			tp->t_flags &= ~TF_WASFRECOVERY;
5970 		}
5971 		if (tp->t_flags & TF_WASCRECOVERY) {
5972 			ENTER_CONGRECOVERY(tp->t_flags);
5973 			tp->t_flags &= ~TF_WASCRECOVERY;
5974 		}
5975 		tp->snd_nxt = tp->snd_max;
5976 		tp->t_badrxtwin = 0;
5977 		break;
5978 	}
5979 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5980 	    (type != CC_RTO)){
5981 		tp->t_ccv.curack = ack;
5982 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5983 	}
5984 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5985 		rack_log_to_prr(rack, 15, cwnd_enter, line);
5986 		if (rack->r_ctl.full_dgp_in_rec)
5987 			rack_client_buffer_level_set(rack);
5988 		rack->r_ctl.dsack_byte_cnt = 0;
5989 		rack->r_ctl.retran_during_recovery = 0;
5990 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5991 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5992 		rack->r_ent_rec_ns = 1;
5993 	}
5994 }
5995 
5996 static inline void
5997 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5998 {
5999 	uint32_t i_cwnd;
6000 
6001 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6002 
6003 	if (CC_ALGO(tp)->after_idle != NULL)
6004 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
6005 
6006 	if (tp->snd_cwnd == 1)
6007 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
6008 	else
6009 		i_cwnd = rc_init_window(rack);
6010 
6011 	/*
6012 	 * Being idle is no different than the initial window. If the cc
6013 	 * clamps it down below the initial window raise it to the initial
6014 	 * window.
6015 	 */
6016 	if (tp->snd_cwnd < i_cwnd) {
6017 		tp->snd_cwnd = i_cwnd;
6018 	}
6019 }
6020 
6021 /*
6022  * Indicate whether this ack should be delayed.  We can delay the ack if
6023  * following conditions are met:
6024  *	- There is no delayed ack timer in progress.
6025  *	- Our last ack wasn't a 0-sized window. We never want to delay
6026  *	  the ack that opens up a 0-sized window.
6027  *	- LRO wasn't used for this segment. We make sure by checking that the
6028  *	  segment size is not larger than the MSS.
6029  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
6030  *	  connection.
6031  */
6032 #define DELAY_ACK(tp, tlen)			 \
6033 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
6034 	((tp->t_flags & TF_DELACK) == 0) &&	 \
6035 	(tlen <= tp->t_maxseg) &&		 \
6036 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
6037 
6038 static struct rack_sendmap *
6039 rack_find_lowest_rsm(struct tcp_rack *rack)
6040 {
6041 	struct rack_sendmap *rsm;
6042 
6043 	/*
6044 	 * Walk the time-order transmitted list looking for an rsm that is
6045 	 * not acked. This will be the one that was sent the longest time
6046 	 * ago that is still outstanding.
6047 	 */
6048 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6049 		if (rsm->r_flags & RACK_ACKED) {
6050 			continue;
6051 		}
6052 		goto finish;
6053 	}
6054 finish:
6055 	return (rsm);
6056 }
6057 
6058 static struct rack_sendmap *
6059 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
6060 {
6061 	struct rack_sendmap *prsm;
6062 
6063 	/*
6064 	 * Walk the sequence order list backward until we hit and arrive at
6065 	 * the highest seq not acked. In theory when this is called it
6066 	 * should be the last segment (which it was not).
6067 	 */
6068 	prsm = rsm;
6069 
6070 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
6071 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
6072 			continue;
6073 		}
6074 		return (prsm);
6075 	}
6076 	return (NULL);
6077 }
6078 
6079 static uint32_t
6080 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
6081 {
6082 	int32_t lro;
6083 	uint32_t thresh;
6084 
6085 	/*
6086 	 * lro is the flag we use to determine if we have seen reordering.
6087 	 * If it gets set we have seen reordering. The reorder logic either
6088 	 * works in one of two ways:
6089 	 *
6090 	 * If reorder-fade is configured, then we track the last time we saw
6091 	 * re-ordering occur. If we reach the point where enough time as
6092 	 * passed we no longer consider reordering has occuring.
6093 	 *
6094 	 * Or if reorder-face is 0, then once we see reordering we consider
6095 	 * the connection to alway be subject to reordering and just set lro
6096 	 * to 1.
6097 	 *
6098 	 * In the end if lro is non-zero we add the extra time for
6099 	 * reordering in.
6100 	 */
6101 	if (srtt == 0)
6102 		srtt = 1;
6103 	if (rack->r_ctl.rc_reorder_ts) {
6104 		if (rack->r_ctl.rc_reorder_fade) {
6105 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
6106 				lro = cts - rack->r_ctl.rc_reorder_ts;
6107 				if (lro == 0) {
6108 					/*
6109 					 * No time as passed since the last
6110 					 * reorder, mark it as reordering.
6111 					 */
6112 					lro = 1;
6113 				}
6114 			} else {
6115 				/* Negative time? */
6116 				lro = 0;
6117 			}
6118 			if (lro > rack->r_ctl.rc_reorder_fade) {
6119 				/* Turn off reordering seen too */
6120 				rack->r_ctl.rc_reorder_ts = 0;
6121 				lro = 0;
6122 			}
6123 		} else {
6124 			/* Reodering does not fade */
6125 			lro = 1;
6126 		}
6127 	} else {
6128 		lro = 0;
6129 	}
6130 	if (rack->rc_rack_tmr_std_based == 0) {
6131 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
6132 	} else {
6133 		/* Standards based pkt-delay is 1/4 srtt */
6134 		thresh = srtt +  (srtt >> 2);
6135 	}
6136 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
6137 		/* It must be set, if not you get 1/4 rtt */
6138 		if (rack->r_ctl.rc_reorder_shift)
6139 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
6140 		else
6141 			thresh += (srtt >> 2);
6142 	}
6143 	if (rack->rc_rack_use_dsack &&
6144 	    lro &&
6145 	    (rack->r_ctl.num_dsack > 0)) {
6146 		/*
6147 		 * We only increase the reordering window if we
6148 		 * have seen reordering <and> we have a DSACK count.
6149 		 */
6150 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
6151 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
6152 	}
6153 	/* SRTT * 2 is the ceiling */
6154 	if (thresh > (srtt * 2)) {
6155 		thresh = srtt * 2;
6156 	}
6157 	/* And we don't want it above the RTO max either */
6158 	if (thresh > rack_rto_max) {
6159 		thresh = rack_rto_max;
6160 	}
6161 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
6162 	return (thresh);
6163 }
6164 
6165 static uint32_t
6166 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6167 		     struct rack_sendmap *rsm, uint32_t srtt)
6168 {
6169 	struct rack_sendmap *prsm;
6170 	uint32_t thresh, len;
6171 	int segsiz;
6172 
6173 	if (srtt == 0)
6174 		srtt = 1;
6175 	if (rack->r_ctl.rc_tlp_threshold)
6176 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6177 	else
6178 		thresh = (srtt * 2);
6179 
6180 	/* Get the previous sent packet, if any */
6181 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6182 	len = rsm->r_end - rsm->r_start;
6183 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6184 		/* Exactly like the ID */
6185 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6186 			uint32_t alt_thresh;
6187 			/*
6188 			 * Compensate for delayed-ack with the d-ack time.
6189 			 */
6190 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6191 			if (alt_thresh > thresh)
6192 				thresh = alt_thresh;
6193 		}
6194 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6195 		/* 2.1 behavior */
6196 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6197 		if (prsm && (len <= segsiz)) {
6198 			/*
6199 			 * Two packets outstanding, thresh should be (2*srtt) +
6200 			 * possible inter-packet delay (if any).
6201 			 */
6202 			uint32_t inter_gap = 0;
6203 			int idx, nidx;
6204 
6205 			idx = rsm->r_rtr_cnt - 1;
6206 			nidx = prsm->r_rtr_cnt - 1;
6207 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6208 				/* Yes it was sent later (or at the same time) */
6209 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6210 			}
6211 			thresh += inter_gap;
6212 		} else if (len <= segsiz) {
6213 			/*
6214 			 * Possibly compensate for delayed-ack.
6215 			 */
6216 			uint32_t alt_thresh;
6217 
6218 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6219 			if (alt_thresh > thresh)
6220 				thresh = alt_thresh;
6221 		}
6222 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6223 		/* 2.2 behavior */
6224 		if (len <= segsiz) {
6225 			uint32_t alt_thresh;
6226 			/*
6227 			 * Compensate for delayed-ack with the d-ack time.
6228 			 */
6229 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6230 			if (alt_thresh > thresh)
6231 				thresh = alt_thresh;
6232 		}
6233 	}
6234 	/* Not above an RTO */
6235 	if (thresh > tp->t_rxtcur) {
6236 		thresh = tp->t_rxtcur;
6237 	}
6238 	/* Not above a RTO max */
6239 	if (thresh > rack_rto_max) {
6240 		thresh = rack_rto_max;
6241 	}
6242 	/* Apply user supplied min TLP */
6243 	if (thresh < rack_tlp_min) {
6244 		thresh = rack_tlp_min;
6245 	}
6246 	return (thresh);
6247 }
6248 
6249 static uint32_t
6250 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6251 {
6252 	/*
6253 	 * We want the rack_rtt which is the
6254 	 * last rtt we measured. However if that
6255 	 * does not exist we fallback to the srtt (which
6256 	 * we probably will never do) and then as a last
6257 	 * resort we use RACK_INITIAL_RTO if no srtt is
6258 	 * yet set.
6259 	 */
6260 	if (rack->rc_rack_rtt)
6261 		return (rack->rc_rack_rtt);
6262 	else if (tp->t_srtt == 0)
6263 		return (RACK_INITIAL_RTO);
6264 	return (tp->t_srtt);
6265 }
6266 
6267 static struct rack_sendmap *
6268 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6269 {
6270 	/*
6271 	 * Check to see that we don't need to fall into recovery. We will
6272 	 * need to do so if our oldest transmit is past the time we should
6273 	 * have had an ack.
6274 	 */
6275 	struct tcp_rack *rack;
6276 	struct rack_sendmap *rsm;
6277 	int32_t idx;
6278 	uint32_t srtt, thresh;
6279 
6280 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6281 	if (tqhash_empty(rack->r_ctl.tqh)) {
6282 		return (NULL);
6283 	}
6284 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6285 	if (rsm == NULL)
6286 		return (NULL);
6287 
6288 
6289 	if (rsm->r_flags & RACK_ACKED) {
6290 		rsm = rack_find_lowest_rsm(rack);
6291 		if (rsm == NULL)
6292 			return (NULL);
6293 	}
6294 	idx = rsm->r_rtr_cnt - 1;
6295 	srtt = rack_grab_rtt(tp, rack);
6296 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
6297 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6298 		return (NULL);
6299 	}
6300 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6301 		return (NULL);
6302 	}
6303 	/* Ok if we reach here we are over-due and this guy can be sent */
6304 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6305 	return (rsm);
6306 }
6307 
6308 static uint32_t
6309 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6310 {
6311 	int32_t t;
6312 	int32_t tt;
6313 	uint32_t ret_val;
6314 
6315 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6316 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6317  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6318 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6319 	ret_val = (uint32_t)tt;
6320 	return (ret_val);
6321 }
6322 
6323 static uint32_t
6324 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6325 {
6326 	/*
6327 	 * Start the FR timer, we do this based on getting the first one in
6328 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6329 	 * events we need to stop the running timer (if its running) before
6330 	 * starting the new one.
6331 	 */
6332 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6333 	uint32_t srtt_cur;
6334 	int32_t idx;
6335 	int32_t is_tlp_timer = 0;
6336 	struct rack_sendmap *rsm;
6337 
6338 	if (rack->t_timers_stopped) {
6339 		/* All timers have been stopped none are to run */
6340 		return (0);
6341 	}
6342 	if (rack->rc_in_persist) {
6343 		/* We can't start any timer in persists */
6344 		return (rack_get_persists_timer_val(tp, rack));
6345 	}
6346 	rack->rc_on_min_to = 0;
6347 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6348 	    (rack->sack_attack_disable > 0) ||
6349 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6350 		goto activate_rxt;
6351 	}
6352 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6353 	if ((rsm == NULL) || sup_rack) {
6354 		/* Nothing on the send map or no rack */
6355 activate_rxt:
6356 		time_since_sent = 0;
6357 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6358 		if (rsm) {
6359 			/*
6360 			 * Should we discount the RTX timer any?
6361 			 *
6362 			 * We want to discount it the smallest amount.
6363 			 * If a timer (Rack/TLP or RXT) has gone off more
6364 			 * recently thats the discount we want to use (now - timer time).
6365 			 * If the retransmit of the oldest packet was more recent then
6366 			 * we want to use that (now - oldest-packet-last_transmit_time).
6367 			 *
6368 			 */
6369 			idx = rsm->r_rtr_cnt - 1;
6370 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6371 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6372 			else
6373 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6374 			if (TSTMP_GT(cts, tstmp_touse))
6375 			    time_since_sent = cts - tstmp_touse;
6376 		}
6377 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6378 		    sbavail(&tptosocket(tp)->so_snd)) {
6379 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6380 			to = tp->t_rxtcur;
6381 			if (to > time_since_sent)
6382 				to -= time_since_sent;
6383 			else
6384 				to = rack->r_ctl.rc_min_to;
6385 			if (to == 0)
6386 				to = 1;
6387 			/* Special case for KEEPINIT */
6388 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6389 			    (TP_KEEPINIT(tp) != 0) &&
6390 			    rsm) {
6391 				/*
6392 				 * We have to put a ceiling on the rxt timer
6393 				 * of the keep-init timeout.
6394 				 */
6395 				uint32_t max_time, red;
6396 
6397 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6398 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6399 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6400 					if (red < max_time)
6401 						max_time -= red;
6402 					else
6403 						max_time = 1;
6404 				}
6405 				/* Reduce timeout to the keep value if needed */
6406 				if (max_time < to)
6407 					to = max_time;
6408 			}
6409 			return (to);
6410 		}
6411 		return (0);
6412 	}
6413 	if (rsm->r_flags & RACK_ACKED) {
6414 		rsm = rack_find_lowest_rsm(rack);
6415 		if (rsm == NULL) {
6416 			/* No lowest? */
6417 			goto activate_rxt;
6418 		}
6419 	}
6420 	if (rack->sack_attack_disable) {
6421 		/*
6422 		 * We don't want to do
6423 		 * any TLP's if you are an attacker.
6424 		 * Though if you are doing what
6425 		 * is expected you may still have
6426 		 * SACK-PASSED marks.
6427 		 */
6428 		goto activate_rxt;
6429 	}
6430 	/* Convert from ms to usecs */
6431 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6432 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6433 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6434 		if ((tp->t_flags & TF_SENTFIN) &&
6435 		    ((tp->snd_max - tp->snd_una) == 1) &&
6436 		    (rsm->r_flags & RACK_HAS_FIN)) {
6437 			/*
6438 			 * We don't start a rack timer if all we have is a
6439 			 * FIN outstanding.
6440 			 */
6441 			goto activate_rxt;
6442 		}
6443 		if ((rack->use_rack_rr == 0) &&
6444 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6445 		    (rack->rack_no_prr == 0) &&
6446 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6447 			/*
6448 			 * We are not cheating, in recovery  and
6449 			 * not enough ack's to yet get our next
6450 			 * retransmission out.
6451 			 *
6452 			 * Note that classified attackers do not
6453 			 * get to use the rack-cheat.
6454 			 */
6455 			goto activate_tlp;
6456 		}
6457 		srtt = rack_grab_rtt(tp, rack);
6458 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
6459 		idx = rsm->r_rtr_cnt - 1;
6460 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6461 		if (SEQ_GEQ(exp, cts)) {
6462 			to = exp - cts;
6463 			if (to < rack->r_ctl.rc_min_to) {
6464 				to = rack->r_ctl.rc_min_to;
6465 				if (rack->r_rr_config == 3)
6466 					rack->rc_on_min_to = 1;
6467 			}
6468 		} else {
6469 			to = rack->r_ctl.rc_min_to;
6470 			if (rack->r_rr_config == 3)
6471 				rack->rc_on_min_to = 1;
6472 		}
6473 	} else {
6474 		/* Ok we need to do a TLP not RACK */
6475 activate_tlp:
6476 		if ((rack->rc_tlp_in_progress != 0) &&
6477 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6478 			/*
6479 			 * The previous send was a TLP and we have sent
6480 			 * N TLP's without sending new data.
6481 			 */
6482 			goto activate_rxt;
6483 		}
6484 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6485 		if (rsm == NULL) {
6486 			/* We found no rsm to TLP with. */
6487 			goto activate_rxt;
6488 		}
6489 		if (rsm->r_flags & RACK_HAS_FIN) {
6490 			/* If its a FIN we dont do TLP */
6491 			rsm = NULL;
6492 			goto activate_rxt;
6493 		}
6494 		idx = rsm->r_rtr_cnt - 1;
6495 		time_since_sent = 0;
6496 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6497 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6498 		else
6499 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6500 		if (TSTMP_GT(cts, tstmp_touse))
6501 		    time_since_sent = cts - tstmp_touse;
6502 		is_tlp_timer = 1;
6503 		if (tp->t_srtt) {
6504 			if ((rack->rc_srtt_measure_made == 0) &&
6505 			    (tp->t_srtt == 1)) {
6506 				/*
6507 				 * If another stack as run and set srtt to 1,
6508 				 * then the srtt was 0, so lets use the initial.
6509 				 */
6510 				srtt = RACK_INITIAL_RTO;
6511 			} else {
6512 				srtt_cur = tp->t_srtt;
6513 				srtt = srtt_cur;
6514 			}
6515 		} else
6516 			srtt = RACK_INITIAL_RTO;
6517 		/*
6518 		 * If the SRTT is not keeping up and the
6519 		 * rack RTT has spiked we want to use
6520 		 * the last RTT not the smoothed one.
6521 		 */
6522 		if (rack_tlp_use_greater &&
6523 		    tp->t_srtt &&
6524 		    (srtt < rack_grab_rtt(tp, rack))) {
6525 			srtt = rack_grab_rtt(tp, rack);
6526 		}
6527 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6528 		if (thresh > time_since_sent) {
6529 			to = thresh - time_since_sent;
6530 		} else {
6531 			to = rack->r_ctl.rc_min_to;
6532 			rack_log_alt_to_to_cancel(rack,
6533 						  thresh,		/* flex1 */
6534 						  time_since_sent,	/* flex2 */
6535 						  tstmp_touse,		/* flex3 */
6536 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6537 						  (uint32_t)rsm->r_tim_lastsent[idx],
6538 						  srtt,
6539 						  idx, 99);
6540 		}
6541 		if (to < rack_tlp_min) {
6542 			to = rack_tlp_min;
6543 		}
6544 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6545 			/*
6546 			 * If the TLP time works out to larger than the max
6547 			 * RTO lets not do TLP.. just RTO.
6548 			 */
6549 			goto activate_rxt;
6550 		}
6551 	}
6552 	if (is_tlp_timer == 0) {
6553 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6554 	} else {
6555 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6556 	}
6557 	if (to == 0)
6558 		to = 1;
6559 	return (to);
6560 }
6561 
6562 static void
6563 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6564 {
6565 	struct timeval tv;
6566 
6567 	if (rack->rc_in_persist == 0) {
6568 		if (tp->t_flags & TF_GPUTINPROG) {
6569 			/*
6570 			 * Stop the goodput now, the calling of the
6571 			 * measurement function clears the flag.
6572 			 */
6573 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6574 						    RACK_QUALITY_PERSIST);
6575 		}
6576 #ifdef NETFLIX_SHARED_CWND
6577 		if (rack->r_ctl.rc_scw) {
6578 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6579 			rack->rack_scwnd_is_idle = 1;
6580 		}
6581 #endif
6582 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(&tv);
6583 		if (rack->lt_bw_up) {
6584 			/* Suspend our LT BW measurement */
6585 			uint64_t tmark;
6586 
6587 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6588 			rack->r_ctl.lt_seq = snd_una;
6589 			tmark = tcp_tv_to_lusectick(&tv);
6590 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6591 			rack->r_ctl.lt_timemark = tmark;
6592 			rack->lt_bw_up = 0;
6593 			rack->r_persist_lt_bw_off = 1;
6594 		}
6595 		if (rack->r_ctl.rc_went_idle_time == 0)
6596 			rack->r_ctl.rc_went_idle_time = 1;
6597 		rack_timer_cancel(tp, rack, cts, __LINE__);
6598 		rack->r_ctl.persist_lost_ends = 0;
6599 		rack->probe_not_answered = 0;
6600 		rack->forced_ack = 0;
6601 		tp->t_rxtshift = 0;
6602 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6603 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6604 		rack->rc_in_persist = 1;
6605 	}
6606 }
6607 
6608 static void
6609 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6610 {
6611 	struct timeval tv;
6612 	uint32_t t_time;
6613 
6614 	if (tcp_in_hpts(rack->rc_tp)) {
6615 		tcp_hpts_remove(rack->rc_tp);
6616 		rack->r_ctl.rc_hpts_flags = 0;
6617 	}
6618 #ifdef NETFLIX_SHARED_CWND
6619 	if (rack->r_ctl.rc_scw) {
6620 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6621 		rack->rack_scwnd_is_idle = 0;
6622 	}
6623 #endif
6624 	t_time = tcp_get_usecs(&tv);
6625 	if (rack->rc_gp_dyn_mul &&
6626 	    (rack->use_fixed_rate == 0) &&
6627 	    (rack->rc_always_pace)) {
6628 		/*
6629 		 * Do we count this as if a probe-rtt just
6630 		 * finished?
6631 		 */
6632 		uint32_t time_idle, idle_min;
6633 
6634 		time_idle = t_time - rack->r_ctl.rc_went_idle_time;
6635 		idle_min = rack_min_probertt_hold;
6636 		if (rack_probertt_gpsrtt_cnt_div) {
6637 			uint64_t extra;
6638 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6639 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
6640 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6641 			idle_min += (uint32_t)extra;
6642 		}
6643 		if (time_idle >= idle_min) {
6644 			/* Yes, we count it as a probe-rtt. */
6645 			uint32_t us_cts;
6646 
6647 			us_cts = tcp_get_usecs(NULL);
6648 			if (rack->in_probe_rtt == 0) {
6649 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6650 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6651 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6652 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6653 			} else {
6654 				rack_exit_probertt(rack, us_cts);
6655 			}
6656 		}
6657 	}
6658 	if (rack->r_persist_lt_bw_off) {
6659 		/* Continue where we left off */
6660 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
6661 		rack->lt_bw_up = 1;
6662 		rack->r_persist_lt_bw_off = 0;
6663 	}
6664 	rack->rc_in_persist = 0;
6665 	rack->r_ctl.rc_went_idle_time = 0;
6666 	tp->t_rxtshift = 0;
6667 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6668 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6669 	rack->r_ctl.rc_agg_delayed = 0;
6670 	rack->r_early = 0;
6671 	rack->r_late = 0;
6672 	rack->r_ctl.rc_agg_early = 0;
6673 }
6674 
6675 static void
6676 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6677 		   struct hpts_diag *diag, struct timeval *tv)
6678 {
6679 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6680 		union tcp_log_stackspecific log;
6681 
6682 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6683 		log.u_bbr.flex1 = diag->p_nxt_slot;
6684 		log.u_bbr.flex2 = diag->p_cur_slot;
6685 		log.u_bbr.flex3 = diag->slot_req;
6686 		log.u_bbr.flex4 = diag->inp_hptsslot;
6687 		log.u_bbr.flex5 = diag->slot_remaining;
6688 		log.u_bbr.flex6 = diag->need_new_to;
6689 		log.u_bbr.flex7 = diag->p_hpts_active;
6690 		log.u_bbr.flex8 = diag->p_on_min_sleep;
6691 		/* Hijack other fields as needed */
6692 		log.u_bbr.epoch = diag->have_slept;
6693 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
6694 		log.u_bbr.pkts_out = diag->co_ret;
6695 		log.u_bbr.applimited = diag->hpts_sleep_time;
6696 		log.u_bbr.delivered = diag->p_prev_slot;
6697 		log.u_bbr.inflight = diag->p_runningslot;
6698 		log.u_bbr.bw_inuse = diag->wheel_slot;
6699 		log.u_bbr.rttProp = diag->wheel_cts;
6700 		log.u_bbr.timeStamp = cts;
6701 		log.u_bbr.delRate = diag->maxslots;
6702 		log.u_bbr.cur_del_rate = diag->p_curtick;
6703 		log.u_bbr.cur_del_rate <<= 32;
6704 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
6705 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6706 		    &rack->rc_inp->inp_socket->so_rcv,
6707 		    &rack->rc_inp->inp_socket->so_snd,
6708 		    BBR_LOG_HPTSDIAG, 0,
6709 		    0, &log, false, tv);
6710 	}
6711 
6712 }
6713 
6714 static void
6715 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6716 {
6717 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6718 		union tcp_log_stackspecific log;
6719 		struct timeval tv;
6720 
6721 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6722 		log.u_bbr.flex1 = sb->sb_flags;
6723 		log.u_bbr.flex2 = len;
6724 		log.u_bbr.flex3 = sb->sb_state;
6725 		log.u_bbr.flex8 = type;
6726 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6727 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6728 		    &rack->rc_inp->inp_socket->so_rcv,
6729 		    &rack->rc_inp->inp_socket->so_snd,
6730 		    TCP_LOG_SB_WAKE, 0,
6731 		    len, &log, false, &tv);
6732 	}
6733 }
6734 
6735 static void
6736 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6737       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6738 {
6739 	struct hpts_diag diag;
6740 	struct inpcb *inp = tptoinpcb(tp);
6741 	struct timeval tv;
6742 	uint32_t delayed_ack = 0;
6743 	uint32_t hpts_timeout;
6744 	uint32_t entry_slot = slot;
6745 	uint8_t stopped;
6746 	uint32_t left = 0;
6747 	uint32_t us_cts;
6748 
6749 	if ((tp->t_state == TCPS_CLOSED) ||
6750 	    (tp->t_state == TCPS_LISTEN)) {
6751 		return;
6752 	}
6753 	if (tcp_in_hpts(tp)) {
6754 		/* Already on the pacer */
6755 		return;
6756 	}
6757 	stopped = rack->rc_tmr_stopped;
6758 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6759 		left = rack->r_ctl.rc_timer_exp - cts;
6760 	}
6761 	rack->r_ctl.rc_timer_exp = 0;
6762 	rack->r_ctl.rc_hpts_flags = 0;
6763 	us_cts = tcp_get_usecs(&tv);
6764 	/* Now early/late accounting */
6765 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6766 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6767 		/*
6768 		 * We have a early carry over set,
6769 		 * we can always add more time so we
6770 		 * can always make this compensation.
6771 		 *
6772 		 * Note if ack's are allowed to wake us do not
6773 		 * penalize the next timer for being awoke
6774 		 * by an ack aka the rc_agg_early (non-paced mode).
6775 		 */
6776 		slot += rack->r_ctl.rc_agg_early;
6777 		rack->r_early = 0;
6778 		rack->r_ctl.rc_agg_early = 0;
6779 	}
6780 	if (rack->r_late) {
6781 		/*
6782 		 * This is harder, we can
6783 		 * compensate some but it
6784 		 * really depends on what
6785 		 * the current pacing time is.
6786 		 */
6787 		if (rack->r_ctl.rc_agg_delayed >= slot) {
6788 			/*
6789 			 * We can't compensate for it all.
6790 			 * And we have to have some time
6791 			 * on the clock. We always have a min
6792 			 * 10 slots (10 x 10 i.e. 100 usecs).
6793 			 */
6794 			if (slot <= HPTS_TICKS_PER_SLOT) {
6795 				/* We gain delay */
6796 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6797 				slot = HPTS_TICKS_PER_SLOT;
6798 			} else {
6799 				/* We take off some */
6800 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6801 				slot = HPTS_TICKS_PER_SLOT;
6802 			}
6803 		} else {
6804 			slot -= rack->r_ctl.rc_agg_delayed;
6805 			rack->r_ctl.rc_agg_delayed = 0;
6806 			/* Make sure we have 100 useconds at minimum */
6807 			if (slot < HPTS_TICKS_PER_SLOT) {
6808 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6809 				slot = HPTS_TICKS_PER_SLOT;
6810 			}
6811 			if (rack->r_ctl.rc_agg_delayed == 0)
6812 				rack->r_late = 0;
6813 		}
6814 	}
6815 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6816 #ifdef TCP_SAD_DETECTION
6817 	if (rack->sack_attack_disable &&
6818 	    (rack->r_ctl.ack_during_sd > 0) &&
6819 	    (slot < tcp_sad_pacing_interval)) {
6820 		/*
6821 		 * We have a potential attacker on
6822 		 * the line. We have possibly some
6823 		 * (or now) pacing time set. We want to
6824 		 * slow down the processing of sacks by some
6825 		 * amount (if it is an attacker). Set the default
6826 		 * slot for attackers in place (unless the original
6827 		 * interval is longer). Its stored in
6828 		 * micro-seconds, so lets convert to msecs.
6829 		 */
6830 		slot = tcp_sad_pacing_interval;
6831 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6832 		rack->r_ctl.ack_during_sd = 0;
6833 	}
6834 #endif
6835 	if (tp->t_flags & TF_DELACK) {
6836 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
6837 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6838 	}
6839 	if (delayed_ack && ((hpts_timeout == 0) ||
6840 			    (delayed_ack < hpts_timeout)))
6841 		hpts_timeout = delayed_ack;
6842 	else
6843 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6844 	/*
6845 	 * If no timers are going to run and we will fall off the hptsi
6846 	 * wheel, we resort to a keep-alive timer if its configured.
6847 	 */
6848 	if ((hpts_timeout == 0) &&
6849 	    (slot == 0)) {
6850 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6851 		    (tp->t_state <= TCPS_CLOSING)) {
6852 			/*
6853 			 * Ok we have no timer (persists, rack, tlp, rxt  or
6854 			 * del-ack), we don't have segments being paced. So
6855 			 * all that is left is the keepalive timer.
6856 			 */
6857 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6858 				/* Get the established keep-alive time */
6859 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6860 			} else {
6861 				/*
6862 				 * Get the initial setup keep-alive time,
6863 				 * note that this is probably not going to
6864 				 * happen, since rack will be running a rxt timer
6865 				 * if a SYN of some sort is outstanding. It is
6866 				 * actually handled in rack_timeout_rxt().
6867 				 */
6868 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6869 			}
6870 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6871 			if (rack->in_probe_rtt) {
6872 				/*
6873 				 * We want to instead not wake up a long time from
6874 				 * now but to wake up about the time we would
6875 				 * exit probe-rtt and initiate a keep-alive ack.
6876 				 * This will get us out of probe-rtt and update
6877 				 * our min-rtt.
6878 				 */
6879 				hpts_timeout = rack_min_probertt_hold;
6880 			}
6881 		}
6882 	}
6883 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6884 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6885 		/*
6886 		 * RACK, TLP, persists and RXT timers all are restartable
6887 		 * based on actions input .. i.e we received a packet (ack
6888 		 * or sack) and that changes things (rw, or snd_una etc).
6889 		 * Thus we can restart them with a new value. For
6890 		 * keep-alive, delayed_ack we keep track of what was left
6891 		 * and restart the timer with a smaller value.
6892 		 */
6893 		if (left < hpts_timeout)
6894 			hpts_timeout = left;
6895 	}
6896 	if (hpts_timeout) {
6897 		/*
6898 		 * Hack alert for now we can't time-out over 2,147,483
6899 		 * seconds (a bit more than 596 hours), which is probably ok
6900 		 * :).
6901 		 */
6902 		if (hpts_timeout > 0x7ffffffe)
6903 			hpts_timeout = 0x7ffffffe;
6904 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6905 	}
6906 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6907 	if ((rack->gp_ready == 0) &&
6908 	    (rack->use_fixed_rate == 0) &&
6909 	    (hpts_timeout < slot) &&
6910 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6911 		/*
6912 		 * We have no good estimate yet for the
6913 		 * old clunky burst mitigation or the
6914 		 * real pacing. And the tlp or rxt is smaller
6915 		 * than the pacing calculation. Lets not
6916 		 * pace that long since we know the calculation
6917 		 * so far is not accurate.
6918 		 */
6919 		slot = hpts_timeout;
6920 	}
6921 	/**
6922 	 * Turn off all the flags for queuing by default. The
6923 	 * flags have important meanings to what happens when
6924 	 * LRO interacts with the transport. Most likely (by default now)
6925 	 * mbuf_queueing and ack compression are on. So the transport
6926 	 * has a couple of flags that control what happens (if those
6927 	 * are not on then these flags won't have any effect since it
6928 	 * won't go through the queuing LRO path).
6929 	 *
6930 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6931 	 *                        pacing output, so don't disturb. But
6932 	 *                        it also means LRO can wake me if there
6933 	 *                        is a SACK arrival.
6934 	 *
6935 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6936 	 *                       with the above flag (QUEUE_READY) and
6937 	 *                       when present it says don't even wake me
6938 	 *                       if a SACK arrives.
6939 	 *
6940 	 * The idea behind these flags is that if we are pacing we
6941 	 * set the MBUF_QUEUE_READY and only get woken up if
6942 	 * a SACK arrives (which could change things) or if
6943 	 * our pacing timer expires. If, however, we have a rack
6944 	 * timer running, then we don't even want a sack to wake
6945 	 * us since the rack timer has to expire before we can send.
6946 	 *
6947 	 * Other cases should usually have none of the flags set
6948 	 * so LRO can call into us.
6949 	 */
6950 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6951 	if (slot) {
6952 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6953 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6954 		/*
6955 		 * A pacing timer (slot) is being set, in
6956 		 * such a case we cannot send (we are blocked by
6957 		 * the timer). So lets tell LRO that it should not
6958 		 * wake us unless there is a SACK. Note this only
6959 		 * will be effective if mbuf queueing is on or
6960 		 * compressed acks are being processed.
6961 		 */
6962 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6963 		/*
6964 		 * But wait if we have a Rack timer running
6965 		 * even a SACK should not disturb us (with
6966 		 * the exception of r_rr_config 3).
6967 		 */
6968 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6969 		    (IN_RECOVERY(tp->t_flags))) {
6970 			if (rack->r_rr_config != 3)
6971 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6972 			else if (rack->rc_pace_dnd) {
6973 				/*
6974 				 * When DND is on, we only let a sack
6975 				 * interrupt us if we are not in recovery.
6976 				 *
6977 				 * If DND is off, then we never hit here
6978 				 * and let all sacks wake us up.
6979 				 *
6980 				 */
6981 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6982 			}
6983 		}
6984 		/* For sack attackers we want to ignore sack */
6985 		if (rack->sack_attack_disable == 1) {
6986 			tp->t_flags2 |= (TF2_DONT_SACK_QUEUE |
6987 			    TF2_MBUF_QUEUE_READY);
6988 		} else if (rack->rc_ack_can_sendout_data) {
6989 			/*
6990 			 * Ahh but wait, this is that special case
6991 			 * where the pacing timer can be disturbed
6992 			 * backout the changes (used for non-paced
6993 			 * burst limiting).
6994 			 */
6995 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6996 			    TF2_MBUF_QUEUE_READY);
6997 		}
6998 		if ((rack->use_rack_rr) &&
6999 		    (rack->r_rr_config < 2) &&
7000 		    ((hpts_timeout) && (hpts_timeout < slot))) {
7001 			/*
7002 			 * Arrange for the hpts to kick back in after the
7003 			 * t-o if the t-o does not cause a send.
7004 			 */
7005 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7006 						   __LINE__, &diag);
7007 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7008 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7009 		} else {
7010 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
7011 						   __LINE__, &diag);
7012 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7013 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
7014 		}
7015 	} else if (hpts_timeout) {
7016 		/*
7017 		 * With respect to t_flags2(?) here, lets let any new acks wake
7018 		 * us up here. Since we are not pacing (no pacing timer), output
7019 		 * can happen so we should let it. If its a Rack timer, then any inbound
7020 		 * packet probably won't change the sending (we will be blocked)
7021 		 * but it may change the prr stats so letting it in (the set defaults
7022 		 * at the start of this block) are good enough.
7023 		 */
7024 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7025 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7026 					   __LINE__, &diag);
7027 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7028 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7029 	} else {
7030 		/* No timer starting */
7031 #ifdef INVARIANTS
7032 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
7033 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
7034 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
7035 		}
7036 #endif
7037 	}
7038 	rack->rc_tmr_stopped = 0;
7039 	if (slot)
7040 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
7041 }
7042 
7043 /*
7044  * RACK Timer, here we simply do logging and house keeping.
7045  * the normal rack_output() function will call the
7046  * appropriate thing to check if we need to do a RACK retransmit.
7047  * We return 1, saying don't proceed with rack_output only
7048  * when all timers have been stopped (destroyed PCB?).
7049  */
7050 static int
7051 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7052 {
7053 	/*
7054 	 * This timer simply provides an internal trigger to send out data.
7055 	 * The check_recovery_mode call will see if there are needed
7056 	 * retransmissions, if so we will enter fast-recovery. The output
7057 	 * call may or may not do the same thing depending on sysctl
7058 	 * settings.
7059 	 */
7060 	struct rack_sendmap *rsm;
7061 
7062 	counter_u64_add(rack_to_tot, 1);
7063 	if (rack->r_state && (rack->r_state != tp->t_state))
7064 		rack_set_state(tp, rack);
7065 	rack->rc_on_min_to = 0;
7066 	rsm = rack_check_recovery_mode(tp, cts);
7067 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
7068 	if (rsm) {
7069 		rack->r_ctl.rc_resend = rsm;
7070 		rack->r_timer_override = 1;
7071 		if (rack->use_rack_rr) {
7072 			/*
7073 			 * Don't accumulate extra pacing delay
7074 			 * we are allowing the rack timer to
7075 			 * over-ride pacing i.e. rrr takes precedence
7076 			 * if the pacing interval is longer than the rrr
7077 			 * time (in other words we get the min pacing
7078 			 * time versus rrr pacing time).
7079 			 */
7080 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7081 		}
7082 	}
7083 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
7084 	if (rsm == NULL) {
7085 		/* restart a timer and return 1 */
7086 		rack_start_hpts_timer(rack, tp, cts,
7087 				      0, 0, 0);
7088 		return (1);
7089 	}
7090 	return (0);
7091 }
7092 
7093 
7094 
7095 static void
7096 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
7097 {
7098 
7099 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
7100 		/*
7101 		 * The trailing space changed, mbufs can grow
7102 		 * at the tail but they can't shrink from
7103 		 * it, KASSERT that. Adjust the orig_m_len to
7104 		 * compensate for this change.
7105 		 */
7106 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7107 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7108 			 rsm->m,
7109 			 rsm,
7110 			 (intmax_t)M_TRAILINGROOM(rsm->m),
7111 			 rsm->orig_t_space,
7112 			 rsm->orig_m_len,
7113 			 rsm->m->m_len));
7114 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7115 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7116 	}
7117 	if (rsm->m->m_len < rsm->orig_m_len) {
7118 		/*
7119 		 * Mbuf shrank, trimmed off the top by an ack, our
7120 		 * offset changes.
7121 		 */
7122 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7123 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7124 			 rsm->m, rsm->m->m_len,
7125 			 rsm, rsm->orig_m_len,
7126 			 rsm->soff));
7127 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7128 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7129 		else
7130 			rsm->soff = 0;
7131 		rsm->orig_m_len = rsm->m->m_len;
7132 #ifdef INVARIANTS
7133 	} else if (rsm->m->m_len > rsm->orig_m_len) {
7134 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7135 		      rsm, rsm->m);
7136 #endif
7137 	}
7138 }
7139 
7140 static void
7141 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7142 {
7143 	struct mbuf *m;
7144 	uint32_t soff;
7145 
7146 	if (src_rsm->m &&
7147 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7148 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7149 		/* Fix up the orig_m_len and possibly the mbuf offset */
7150 		rack_adjust_orig_mlen(src_rsm);
7151 	}
7152 	m = src_rsm->m;
7153 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7154 	while (soff >= m->m_len) {
7155 		/* Move out past this mbuf */
7156 		soff -= m->m_len;
7157 		m = m->m_next;
7158 		KASSERT((m != NULL),
7159 			("rsm:%p nrsm:%p hit at soff:%u null m",
7160 			 src_rsm, rsm, soff));
7161 		if (m == NULL) {
7162 			/* This should *not* happen which is why there is a kassert */
7163 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7164 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7165 					       &src_rsm->soff);
7166 			src_rsm->orig_m_len = src_rsm->m->m_len;
7167 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7168 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7169 					   (rsm->r_start - rack->rc_tp->snd_una),
7170 					   &rsm->soff);
7171 			rsm->orig_m_len = rsm->m->m_len;
7172 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7173 			return;
7174 		}
7175 	}
7176 	rsm->m = m;
7177 	rsm->soff = soff;
7178 	rsm->orig_m_len = m->m_len;
7179 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7180 }
7181 
7182 static __inline void
7183 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7184 	       struct rack_sendmap *rsm, uint32_t start)
7185 {
7186 	int idx;
7187 
7188 	nrsm->r_start = start;
7189 	nrsm->r_end = rsm->r_end;
7190 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7191 	nrsm->r_flags = rsm->r_flags;
7192 	nrsm->r_dupack = rsm->r_dupack;
7193 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7194 	nrsm->r_rtr_bytes = 0;
7195 	nrsm->r_fas = rsm->r_fas;
7196 	nrsm->r_bas = rsm->r_bas;
7197 	rsm->r_end = nrsm->r_start;
7198 	nrsm->r_just_ret = rsm->r_just_ret;
7199 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7200 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7201 	}
7202 	/* Now if we have SYN flag we keep it on the left edge */
7203 	if (nrsm->r_flags & RACK_HAS_SYN)
7204 		nrsm->r_flags &= ~RACK_HAS_SYN;
7205 	/* Now if we have a FIN flag we keep it on the right edge */
7206 	if (rsm->r_flags & RACK_HAS_FIN)
7207 		rsm->r_flags &= ~RACK_HAS_FIN;
7208 	/* Push bit must go to the right edge as well */
7209 	if (rsm->r_flags & RACK_HAD_PUSH)
7210 		rsm->r_flags &= ~RACK_HAD_PUSH;
7211 	/* Clone over the state of the hw_tls flag */
7212 	nrsm->r_hw_tls = rsm->r_hw_tls;
7213 	/*
7214 	 * Now we need to find nrsm's new location in the mbuf chain
7215 	 * we basically calculate a new offset, which is soff +
7216 	 * how much is left in original rsm. Then we walk out the mbuf
7217 	 * chain to find the righ position, it may be the same mbuf
7218 	 * or maybe not.
7219 	 */
7220 	KASSERT(((rsm->m != NULL) ||
7221 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7222 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7223 	if (rsm->m)
7224 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7225 }
7226 
7227 static struct rack_sendmap *
7228 rack_merge_rsm(struct tcp_rack *rack,
7229 	       struct rack_sendmap *l_rsm,
7230 	       struct rack_sendmap *r_rsm)
7231 {
7232 	/*
7233 	 * We are merging two ack'd RSM's,
7234 	 * the l_rsm is on the left (lower seq
7235 	 * values) and the r_rsm is on the right
7236 	 * (higher seq value). The simplest way
7237 	 * to merge these is to move the right
7238 	 * one into the left. I don't think there
7239 	 * is any reason we need to try to find
7240 	 * the oldest (or last oldest retransmitted).
7241 	 */
7242 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7243 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7244 	l_rsm->r_end = r_rsm->r_end;
7245 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7246 		l_rsm->r_dupack = r_rsm->r_dupack;
7247 	if (r_rsm->r_rtr_bytes)
7248 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7249 	if (r_rsm->r_in_tmap) {
7250 		/* This really should not happen */
7251 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7252 		r_rsm->r_in_tmap = 0;
7253 	}
7254 
7255 	/* Now the flags */
7256 	if (r_rsm->r_flags & RACK_HAS_FIN)
7257 		l_rsm->r_flags |= RACK_HAS_FIN;
7258 	if (r_rsm->r_flags & RACK_TLP)
7259 		l_rsm->r_flags |= RACK_TLP;
7260 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7261 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7262 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7263 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7264 		/*
7265 		 * If both are app-limited then let the
7266 		 * free lower the count. If right is app
7267 		 * limited and left is not, transfer.
7268 		 */
7269 		l_rsm->r_flags |= RACK_APP_LIMITED;
7270 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7271 		if (r_rsm == rack->r_ctl.rc_first_appl)
7272 			rack->r_ctl.rc_first_appl = l_rsm;
7273 	}
7274 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7275 	/*
7276 	 * We keep the largest value, which is the newest
7277 	 * send. We do this in case a segment that is
7278 	 * joined together and not part of a GP estimate
7279 	 * later gets expanded into the GP estimate.
7280 	 *
7281 	 * We prohibit the merging of unlike kinds i.e.
7282 	 * all pieces that are in the GP estimate can be
7283 	 * merged and all pieces that are not in a GP estimate
7284 	 * can be merged, but not disimilar pieces. Combine
7285 	 * this with taking the highest here and we should
7286 	 * be ok unless of course the client reneges. Then
7287 	 * all bets are off.
7288 	 */
7289 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7290 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7291 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7292 	}
7293 	/*
7294 	 * When merging two RSM's we also need to consider the ack time and keep
7295 	 * newest. If the ack gets merged into a measurement then that is the
7296 	 * one we will want to be using.
7297 	 */
7298 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7299 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7300 
7301 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7302 		/* Transfer the split limit to the map we free */
7303 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7304 		l_rsm->r_limit_type = 0;
7305 	}
7306 	rack_free(rack, r_rsm);
7307 	l_rsm->r_flags |= RACK_MERGED;
7308 	return (l_rsm);
7309 }
7310 
7311 /*
7312  * TLP Timer, here we simply setup what segment we want to
7313  * have the TLP expire on, the normal rack_output() will then
7314  * send it out.
7315  *
7316  * We return 1, saying don't proceed with rack_output only
7317  * when all timers have been stopped (destroyed PCB?).
7318  */
7319 static int
7320 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7321 {
7322 	/*
7323 	 * Tail Loss Probe.
7324 	 */
7325 	struct rack_sendmap *rsm = NULL;
7326 	int insret __diagused;
7327 	struct socket *so = tptosocket(tp);
7328 	uint32_t amm;
7329 	uint32_t out, avail;
7330 	int collapsed_win = 0;
7331 
7332 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7333 		/* Its not time yet */
7334 		return (0);
7335 	}
7336 	if (ctf_progress_timeout_check(tp, true)) {
7337 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7338 		return (-ETIMEDOUT);	/* tcp_drop() */
7339 	}
7340 	/*
7341 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7342 	 * need to figure out how to force a full MSS segment out.
7343 	 */
7344 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7345 	rack->r_ctl.retran_during_recovery = 0;
7346 	rack->r_ctl.dsack_byte_cnt = 0;
7347 	counter_u64_add(rack_tlp_tot, 1);
7348 	if (rack->r_state && (rack->r_state != tp->t_state))
7349 		rack_set_state(tp, rack);
7350 	avail = sbavail(&so->so_snd);
7351 	out = tp->snd_max - tp->snd_una;
7352 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7353 		/* special case, we need a retransmission */
7354 		collapsed_win = 1;
7355 		goto need_retran;
7356 	}
7357 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7358 		rack->r_ctl.dsack_persist--;
7359 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7360 			rack->r_ctl.num_dsack = 0;
7361 		}
7362 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7363 	}
7364 	if ((tp->t_flags & TF_GPUTINPROG) &&
7365 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7366 		/*
7367 		 * If this is the second in a row
7368 		 * TLP and we are doing a measurement
7369 		 * its time to abandon the measurement.
7370 		 * Something is likely broken on
7371 		 * the clients network and measuring a
7372 		 * broken network does us no good.
7373 		 */
7374 		tp->t_flags &= ~TF_GPUTINPROG;
7375 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7376 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7377 					   tp->gput_seq,
7378 					   0, 0, 18, __LINE__, NULL, 0);
7379 	}
7380 	/*
7381 	 * Check our send oldest always settings, and if
7382 	 * there is an oldest to send jump to the need_retran.
7383 	 */
7384 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7385 		goto need_retran;
7386 
7387 	if (avail > out) {
7388 		/* New data is available */
7389 		amm = avail - out;
7390 		if (amm > ctf_fixed_maxseg(tp)) {
7391 			amm = ctf_fixed_maxseg(tp);
7392 			if ((amm + out) > tp->snd_wnd) {
7393 				/* We are rwnd limited */
7394 				goto need_retran;
7395 			}
7396 		} else if (amm < ctf_fixed_maxseg(tp)) {
7397 			/* not enough to fill a MTU */
7398 			goto need_retran;
7399 		}
7400 		if (IN_FASTRECOVERY(tp->t_flags)) {
7401 			/* Unlikely */
7402 			if (rack->rack_no_prr == 0) {
7403 				if (out + amm <= tp->snd_wnd) {
7404 					rack->r_ctl.rc_prr_sndcnt = amm;
7405 					rack->r_ctl.rc_tlp_new_data = amm;
7406 					rack_log_to_prr(rack, 4, 0, __LINE__);
7407 				}
7408 			} else
7409 				goto need_retran;
7410 		} else {
7411 			/* Set the send-new override */
7412 			if (out + amm <= tp->snd_wnd)
7413 				rack->r_ctl.rc_tlp_new_data = amm;
7414 			else
7415 				goto need_retran;
7416 		}
7417 		rack->r_ctl.rc_tlpsend = NULL;
7418 		counter_u64_add(rack_tlp_newdata, 1);
7419 		goto send;
7420 	}
7421 need_retran:
7422 	/*
7423 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7424 	 * optionally the first un-acked segment.
7425 	 */
7426 	if (collapsed_win == 0) {
7427 		if (rack_always_send_oldest)
7428 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7429 		else {
7430 			rsm = tqhash_max(rack->r_ctl.tqh);
7431 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7432 				rsm = rack_find_high_nonack(rack, rsm);
7433 			}
7434 		}
7435 		if (rsm == NULL) {
7436 #ifdef TCP_BLACKBOX
7437 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7438 #endif
7439 			goto out;
7440 		}
7441 	} else {
7442 		/*
7443 		 * We had a collapsed window, lets find
7444 		 * the point before the collapse.
7445 		 */
7446 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7447 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7448 		else {
7449 			rsm = tqhash_min(rack->r_ctl.tqh);
7450 		}
7451 		if (rsm == NULL) {
7452 			/* Huh */
7453 			goto out;
7454 		}
7455 	}
7456 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7457 		/*
7458 		 * We need to split this the last segment in two.
7459 		 */
7460 		struct rack_sendmap *nrsm;
7461 
7462 		nrsm = rack_alloc_full_limit(rack);
7463 		if (nrsm == NULL) {
7464 			/*
7465 			 * No memory to split, we will just exit and punt
7466 			 * off to the RXT timer.
7467 			 */
7468 			goto out;
7469 		}
7470 		rack_clone_rsm(rack, nrsm, rsm,
7471 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7472 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7473 #ifndef INVARIANTS
7474 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7475 #else
7476 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7477 			panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
7478 			      nrsm, insret, rack, rsm);
7479 		}
7480 #endif
7481 		if (rsm->r_in_tmap) {
7482 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7483 			nrsm->r_in_tmap = 1;
7484 		}
7485 		rsm = nrsm;
7486 	}
7487 	rack->r_ctl.rc_tlpsend = rsm;
7488 send:
7489 	/* Make sure output path knows we are doing a TLP */
7490 	*doing_tlp = 1;
7491 	rack->r_timer_override = 1;
7492 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7493 	return (0);
7494 out:
7495 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7496 	return (0);
7497 }
7498 
7499 /*
7500  * Delayed ack Timer, here we simply need to setup the
7501  * ACK_NOW flag and remove the DELACK flag. From there
7502  * the output routine will send the ack out.
7503  *
7504  * We only return 1, saying don't proceed, if all timers
7505  * are stopped (destroyed PCB?).
7506  */
7507 static int
7508 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7509 {
7510 
7511 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7512 	tp->t_flags &= ~TF_DELACK;
7513 	tp->t_flags |= TF_ACKNOW;
7514 	KMOD_TCPSTAT_INC(tcps_delack);
7515 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7516 	return (0);
7517 }
7518 
7519 /*
7520  * Persists timer, here we simply send the
7521  * same thing as a keepalive will.
7522  * the one byte send.
7523  *
7524  * We only return 1, saying don't proceed, if all timers
7525  * are stopped (destroyed PCB?).
7526  */
7527 static int
7528 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7529 {
7530 	struct tcptemp *t_template;
7531 	int32_t retval = 1;
7532 
7533 	if (rack->rc_in_persist == 0)
7534 		return (0);
7535 	if (ctf_progress_timeout_check(tp, false)) {
7536 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7537 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7538 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7539 		return (-ETIMEDOUT);	/* tcp_drop() */
7540 	}
7541 	/*
7542 	 * Persistence timer into zero window. Force a byte to be output, if
7543 	 * possible.
7544 	 */
7545 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
7546 	/*
7547 	 * Hack: if the peer is dead/unreachable, we do not time out if the
7548 	 * window is closed.  After a full backoff, drop the connection if
7549 	 * the idle time (no responses to probes) reaches the maximum
7550 	 * backoff that we would use if retransmitting.
7551 	 */
7552 	if (tp->t_rxtshift >= V_tcp_retries &&
7553 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7554 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7555 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7556 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7557 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7558 		retval = -ETIMEDOUT;	/* tcp_drop() */
7559 		goto out;
7560 	}
7561 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7562 	    tp->snd_una == tp->snd_max)
7563 		rack_exit_persist(tp, rack, cts);
7564 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7565 	/*
7566 	 * If the user has closed the socket then drop a persisting
7567 	 * connection after a much reduced timeout.
7568 	 */
7569 	if (tp->t_state > TCPS_CLOSE_WAIT &&
7570 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7571 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7572 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7573 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7574 		retval = -ETIMEDOUT;	/* tcp_drop() */
7575 		goto out;
7576 	}
7577 	t_template = tcpip_maketemplate(rack->rc_inp);
7578 	if (t_template) {
7579 		/* only set it if we were answered */
7580 		if (rack->forced_ack == 0) {
7581 			rack->forced_ack = 1;
7582 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7583 		} else {
7584 			rack->probe_not_answered = 1;
7585 			counter_u64_add(rack_persists_loss, 1);
7586 			rack->r_ctl.persist_lost_ends++;
7587 		}
7588 		counter_u64_add(rack_persists_sends, 1);
7589 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7590 		tcp_respond(tp, t_template->tt_ipgen,
7591 			    &t_template->tt_t, (struct mbuf *)NULL,
7592 			    tp->rcv_nxt, tp->snd_una - 1, 0);
7593 		/* This sends an ack */
7594 		if (tp->t_flags & TF_DELACK)
7595 			tp->t_flags &= ~TF_DELACK;
7596 		free(t_template, M_TEMP);
7597 	}
7598 	if (tp->t_rxtshift < V_tcp_retries)
7599 		tp->t_rxtshift++;
7600 out:
7601 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7602 	rack_start_hpts_timer(rack, tp, cts,
7603 			      0, 0, 0);
7604 	return (retval);
7605 }
7606 
7607 /*
7608  * If a keepalive goes off, we had no other timers
7609  * happening. We always return 1 here since this
7610  * routine either drops the connection or sends
7611  * out a segment with respond.
7612  */
7613 static int
7614 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7615 {
7616 	struct tcptemp *t_template;
7617 	struct inpcb *inp = tptoinpcb(tp);
7618 
7619 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7620 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7621 	/*
7622 	 * Keep-alive timer went off; send something or drop connection if
7623 	 * idle for too long.
7624 	 */
7625 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
7626 	if (tp->t_state < TCPS_ESTABLISHED)
7627 		goto dropit;
7628 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7629 	    tp->t_state <= TCPS_CLOSING) {
7630 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7631 			goto dropit;
7632 		/*
7633 		 * Send a packet designed to force a response if the peer is
7634 		 * up and reachable: either an ACK if the connection is
7635 		 * still alive, or an RST if the peer has closed the
7636 		 * connection due to timeout or reboot. Using sequence
7637 		 * number tp->snd_una-1 causes the transmitted zero-length
7638 		 * segment to lie outside the receive window; by the
7639 		 * protocol spec, this requires the correspondent TCP to
7640 		 * respond.
7641 		 */
7642 		KMOD_TCPSTAT_INC(tcps_keepprobe);
7643 		t_template = tcpip_maketemplate(inp);
7644 		if (t_template) {
7645 			if (rack->forced_ack == 0) {
7646 				rack->forced_ack = 1;
7647 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7648 			} else {
7649 				rack->probe_not_answered = 1;
7650 			}
7651 			tcp_respond(tp, t_template->tt_ipgen,
7652 			    &t_template->tt_t, (struct mbuf *)NULL,
7653 			    tp->rcv_nxt, tp->snd_una - 1, 0);
7654 			free(t_template, M_TEMP);
7655 		}
7656 	}
7657 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7658 	return (1);
7659 dropit:
7660 	KMOD_TCPSTAT_INC(tcps_keepdrops);
7661 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7662 	return (-ETIMEDOUT);	/* tcp_drop() */
7663 }
7664 
7665 /*
7666  * Retransmit helper function, clear up all the ack
7667  * flags and take care of important book keeping.
7668  */
7669 static void
7670 rack_remxt_tmr(struct tcpcb *tp)
7671 {
7672 	/*
7673 	 * The retransmit timer went off, all sack'd blocks must be
7674 	 * un-acked.
7675 	 */
7676 	struct rack_sendmap *rsm, *trsm = NULL;
7677 	struct tcp_rack *rack;
7678 
7679 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7680 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7681 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7682 	if (rack->r_state && (rack->r_state != tp->t_state))
7683 		rack_set_state(tp, rack);
7684 	/*
7685 	 * Ideally we would like to be able to
7686 	 * mark SACK-PASS on anything not acked here.
7687 	 *
7688 	 * However, if we do that we would burst out
7689 	 * all that data 1ms apart. This would be unwise,
7690 	 * so for now we will just let the normal rxt timer
7691 	 * and tlp timer take care of it.
7692 	 *
7693 	 * Also we really need to stick them back in sequence
7694 	 * order. This way we send in the proper order and any
7695 	 * sacks that come floating in will "re-ack" the data.
7696 	 * To do this we zap the tmap with an INIT and then
7697 	 * walk through and place every rsm in the RB tree
7698 	 * back in its seq ordered place.
7699 	 */
7700 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
7701 
7702 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7703 		rsm->r_dupack = 0;
7704 		if (rack_verbose_logging)
7705 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7706 		/* We must re-add it back to the tlist */
7707 		if (trsm == NULL) {
7708 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7709 		} else {
7710 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7711 		}
7712 		rsm->r_in_tmap = 1;
7713 		trsm = rsm;
7714 		if (rsm->r_flags & RACK_ACKED)
7715 			rsm->r_flags |= RACK_WAS_ACKED;
7716 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
7717 		rsm->r_flags |= RACK_MUST_RXT;
7718 	}
7719 	/* Clear the count (we just un-acked them) */
7720 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7721 	rack->r_ctl.rc_sacked = 0;
7722 	rack->r_ctl.rc_sacklast = NULL;
7723 	rack->r_ctl.rc_agg_delayed = 0;
7724 	rack->r_early = 0;
7725 	rack->r_ctl.rc_agg_early = 0;
7726 	rack->r_late = 0;
7727 	/* Clear the tlp rtx mark */
7728 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7729 	if (rack->r_ctl.rc_resend != NULL)
7730 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7731 	rack->r_ctl.rc_prr_sndcnt = 0;
7732 	rack_log_to_prr(rack, 6, 0, __LINE__);
7733 	rack->r_timer_override = 1;
7734 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
7735 #ifdef TCP_SAD_DETECTION
7736 	    || (rack->sack_attack_disable != 0)
7737 #endif
7738 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
7739 		/*
7740 		 * For non-sack customers new data
7741 		 * needs to go out as retransmits until
7742 		 * we retransmit up to snd_max.
7743 		 */
7744 		rack->r_must_retran = 1;
7745 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7746 						rack->r_ctl.rc_sacked);
7747 	}
7748 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7749 }
7750 
7751 static void
7752 rack_convert_rtts(struct tcpcb *tp)
7753 {
7754 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7755 	tp->t_rxtcur = RACK_REXMTVAL(tp);
7756 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7757 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7758 	}
7759 	if (tp->t_rxtcur > rack_rto_max) {
7760 		tp->t_rxtcur = rack_rto_max;
7761 	}
7762 }
7763 
7764 static void
7765 rack_cc_conn_init(struct tcpcb *tp)
7766 {
7767 	struct tcp_rack *rack;
7768 	uint32_t srtt;
7769 
7770 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7771 	srtt = tp->t_srtt;
7772 	cc_conn_init(tp);
7773 	/*
7774 	 * Now convert to rack's internal format,
7775 	 * if required.
7776 	 */
7777 	if ((srtt == 0) && (tp->t_srtt != 0))
7778 		rack_convert_rtts(tp);
7779 	/*
7780 	 * We want a chance to stay in slowstart as
7781 	 * we create a connection. TCP spec says that
7782 	 * initially ssthresh is infinite. For our
7783 	 * purposes that is the snd_wnd.
7784 	 */
7785 	if (tp->snd_ssthresh < tp->snd_wnd) {
7786 		tp->snd_ssthresh = tp->snd_wnd;
7787 	}
7788 	/*
7789 	 * We also want to assure a IW worth of
7790 	 * data can get inflight.
7791 	 */
7792 	if (rc_init_window(rack) < tp->snd_cwnd)
7793 		tp->snd_cwnd = rc_init_window(rack);
7794 }
7795 
7796 /*
7797  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7798  * we will setup to retransmit the lowest seq number outstanding.
7799  */
7800 static int
7801 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7802 {
7803 	struct inpcb *inp = tptoinpcb(tp);
7804 	int32_t rexmt;
7805 	int32_t retval = 0;
7806 	bool isipv6;
7807 
7808 	if ((tp->t_flags & TF_GPUTINPROG) &&
7809 	    (tp->t_rxtshift)) {
7810 		/*
7811 		 * We have had a second timeout
7812 		 * measurements on successive rxt's are not profitable.
7813 		 * It is unlikely to be of any use (the network is
7814 		 * broken or the client went away).
7815 		 */
7816 		tp->t_flags &= ~TF_GPUTINPROG;
7817 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7818 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7819 					   tp->gput_seq,
7820 					   0, 0, 18, __LINE__, NULL, 0);
7821 	}
7822 	if (ctf_progress_timeout_check(tp, false)) {
7823 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7824 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7825 		return (-ETIMEDOUT);	/* tcp_drop() */
7826 	}
7827 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7828 	rack->r_ctl.retran_during_recovery = 0;
7829 	rack->rc_ack_required = 1;
7830 	rack->r_ctl.dsack_byte_cnt = 0;
7831 	if (IN_FASTRECOVERY(tp->t_flags))
7832 		tp->t_flags |= TF_WASFRECOVERY;
7833 	else
7834 		tp->t_flags &= ~TF_WASFRECOVERY;
7835 	if (IN_CONGRECOVERY(tp->t_flags))
7836 		tp->t_flags |= TF_WASCRECOVERY;
7837 	else
7838 		tp->t_flags &= ~TF_WASCRECOVERY;
7839 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7840 	    (tp->snd_una == tp->snd_max)) {
7841 		/* Nothing outstanding .. nothing to do */
7842 		return (0);
7843 	}
7844 	if (rack->r_ctl.dsack_persist) {
7845 		rack->r_ctl.dsack_persist--;
7846 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7847 			rack->r_ctl.num_dsack = 0;
7848 		}
7849 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7850 	}
7851 	/*
7852 	 * Rack can only run one timer  at a time, so we cannot
7853 	 * run a KEEPINIT (gating SYN sending) and a retransmit
7854 	 * timer for the SYN. So if we are in a front state and
7855 	 * have a KEEPINIT timer we need to check the first transmit
7856 	 * against now to see if we have exceeded the KEEPINIT time
7857 	 * (if one is set).
7858 	 */
7859 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7860 	    (TP_KEEPINIT(tp) != 0)) {
7861 		struct rack_sendmap *rsm;
7862 
7863 		rsm = tqhash_min(rack->r_ctl.tqh);
7864 		if (rsm) {
7865 			/* Ok we have something outstanding to test keepinit with */
7866 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7867 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7868 				/* We have exceeded the KEEPINIT time */
7869 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7870 				goto drop_it;
7871 			}
7872 		}
7873 	}
7874 	/*
7875 	 * Retransmission timer went off.  Message has not been acked within
7876 	 * retransmit interval.  Back off to a longer retransmit interval
7877 	 * and retransmit one segment.
7878 	 */
7879 	rack_remxt_tmr(tp);
7880 	if ((rack->r_ctl.rc_resend == NULL) ||
7881 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7882 		/*
7883 		 * If the rwnd collapsed on
7884 		 * the one we are retransmitting
7885 		 * it does not count against the
7886 		 * rxt count.
7887 		 */
7888 		tp->t_rxtshift++;
7889 	}
7890 	if (tp->t_rxtshift > V_tcp_retries) {
7891 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7892 drop_it:
7893 		tp->t_rxtshift = V_tcp_retries;
7894 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7895 		/* XXXGL: previously t_softerror was casted to uint16_t */
7896 		MPASS(tp->t_softerror >= 0);
7897 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7898 		goto out;	/* tcp_drop() */
7899 	}
7900 	if (tp->t_state == TCPS_SYN_SENT) {
7901 		/*
7902 		 * If the SYN was retransmitted, indicate CWND to be limited
7903 		 * to 1 segment in cc_conn_init().
7904 		 */
7905 		tp->snd_cwnd = 1;
7906 	} else if (tp->t_rxtshift == 1) {
7907 		/*
7908 		 * first retransmit; record ssthresh and cwnd so they can be
7909 		 * recovered if this turns out to be a "bad" retransmit. A
7910 		 * retransmit is considered "bad" if an ACK for this segment
7911 		 * is received within RTT/2 interval; the assumption here is
7912 		 * that the ACK was already in flight.  See "On Estimating
7913 		 * End-to-End Network Path Properties" by Allman and Paxson
7914 		 * for more details.
7915 		 */
7916 		tp->snd_cwnd_prev = tp->snd_cwnd;
7917 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
7918 		tp->snd_recover_prev = tp->snd_recover;
7919 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7920 		tp->t_flags |= TF_PREVVALID;
7921 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7922 		tp->t_flags &= ~TF_PREVVALID;
7923 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7924 	if ((tp->t_state == TCPS_SYN_SENT) ||
7925 	    (tp->t_state == TCPS_SYN_RECEIVED))
7926 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7927 	else
7928 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7929 
7930 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7931 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7932 	/*
7933 	 * We enter the path for PLMTUD if connection is established or, if
7934 	 * connection is FIN_WAIT_1 status, reason for the last is that if
7935 	 * amount of data we send is very small, we could send it in couple
7936 	 * of packets and process straight to FIN. In that case we won't
7937 	 * catch ESTABLISHED state.
7938 	 */
7939 #ifdef INET6
7940 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7941 #else
7942 	isipv6 = false;
7943 #endif
7944 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
7945 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7946 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7947 	    ((tp->t_state == TCPS_ESTABLISHED) ||
7948 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
7949 		/*
7950 		 * Idea here is that at each stage of mtu probe (usually,
7951 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7952 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7953 		 * should take care of that.
7954 		 */
7955 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7956 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7957 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7958 		    tp->t_rxtshift % 2 == 0)) {
7959 			/*
7960 			 * Enter Path MTU Black-hole Detection mechanism: -
7961 			 * Disable Path MTU Discovery (IP "DF" bit). -
7962 			 * Reduce MTU to lower value than what we negotiated
7963 			 * with peer.
7964 			 */
7965 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7966 				/* Record that we may have found a black hole. */
7967 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7968 				/* Keep track of previous MSS. */
7969 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7970 			}
7971 
7972 			/*
7973 			 * Reduce the MSS to blackhole value or to the
7974 			 * default in an attempt to retransmit.
7975 			 */
7976 #ifdef INET6
7977 			if (isipv6 &&
7978 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7979 				/* Use the sysctl tuneable blackhole MSS. */
7980 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7981 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7982 			} else if (isipv6) {
7983 				/* Use the default MSS. */
7984 				tp->t_maxseg = V_tcp_v6mssdflt;
7985 				/*
7986 				 * Disable Path MTU Discovery when we switch
7987 				 * to minmss.
7988 				 */
7989 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7990 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7991 			}
7992 #endif
7993 #if defined(INET6) && defined(INET)
7994 			else
7995 #endif
7996 #ifdef INET
7997 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7998 				/* Use the sysctl tuneable blackhole MSS. */
7999 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
8000 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
8001 			} else {
8002 				/* Use the default MSS. */
8003 				tp->t_maxseg = V_tcp_mssdflt;
8004 				/*
8005 				 * Disable Path MTU Discovery when we switch
8006 				 * to minmss.
8007 				 */
8008 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8009 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8010 			}
8011 #endif
8012 		} else {
8013 			/*
8014 			 * If further retransmissions are still unsuccessful
8015 			 * with a lowered MTU, maybe this isn't a blackhole
8016 			 * and we restore the previous MSS and blackhole
8017 			 * detection flags. The limit '6' is determined by
8018 			 * giving each probe stage (1448, 1188, 524) 2
8019 			 * chances to recover.
8020 			 */
8021 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
8022 			    (tp->t_rxtshift >= 6)) {
8023 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8024 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
8025 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
8026 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
8027 			}
8028 		}
8029 	}
8030 	/*
8031 	 * Disable RFC1323 and SACK if we haven't got any response to
8032 	 * our third SYN to work-around some broken terminal servers
8033 	 * (most of which have hopefully been retired) that have bad VJ
8034 	 * header compression code which trashes TCP segments containing
8035 	 * unknown-to-them TCP options.
8036 	 */
8037 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
8038 	    (tp->t_rxtshift == 3))
8039 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
8040 	/*
8041 	 * If we backed off this far, our srtt estimate is probably bogus.
8042 	 * Clobber it so we'll take the next rtt measurement as our srtt;
8043 	 * move the current srtt into rttvar to keep the current retransmit
8044 	 * times until then.
8045 	 */
8046 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
8047 #ifdef INET6
8048 		if ((inp->inp_vflag & INP_IPV6) != 0)
8049 			in6_losing(inp);
8050 		else
8051 #endif
8052 			in_losing(inp);
8053 		tp->t_rttvar += tp->t_srtt;
8054 		tp->t_srtt = 0;
8055 	}
8056 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8057 	tp->snd_recover = tp->snd_max;
8058 	tp->t_flags |= TF_ACKNOW;
8059 	tp->t_rtttime = 0;
8060 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
8061 out:
8062 	return (retval);
8063 }
8064 
8065 static int
8066 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
8067 {
8068 	int32_t ret = 0;
8069 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
8070 
8071 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8072 	    (tp->t_flags & TF_GPUTINPROG)) {
8073 		/*
8074 		 * We have a goodput in progress
8075 		 * and we have entered a late state.
8076 		 * Do we have enough data in the sb
8077 		 * to handle the GPUT request?
8078 		 */
8079 		uint32_t bytes;
8080 
8081 		bytes = tp->gput_ack - tp->gput_seq;
8082 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
8083 			bytes += tp->gput_seq - tp->snd_una;
8084 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
8085 			/*
8086 			 * There are not enough bytes in the socket
8087 			 * buffer that have been sent to cover this
8088 			 * measurement. Cancel it.
8089 			 */
8090 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8091 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
8092 						   tp->gput_seq,
8093 						   0, 0, 18, __LINE__, NULL, 0);
8094 			tp->t_flags &= ~TF_GPUTINPROG;
8095 		}
8096 	}
8097 	if (timers == 0) {
8098 		return (0);
8099 	}
8100 	if (tp->t_state == TCPS_LISTEN) {
8101 		/* no timers on listen sockets */
8102 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
8103 			return (0);
8104 		return (1);
8105 	}
8106 	if ((timers & PACE_TMR_RACK) &&
8107 	    rack->rc_on_min_to) {
8108 		/*
8109 		 * For the rack timer when we
8110 		 * are on a min-timeout (which means rrr_conf = 3)
8111 		 * we don't want to check the timer. It may
8112 		 * be going off for a pace and thats ok we
8113 		 * want to send the retransmit (if its ready).
8114 		 *
8115 		 * If its on a normal rack timer (non-min) then
8116 		 * we will check if its expired.
8117 		 */
8118 		goto skip_time_check;
8119 	}
8120 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8121 		uint32_t left;
8122 
8123 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8124 			ret = -1;
8125 			rack_log_to_processing(rack, cts, ret, 0);
8126 			return (0);
8127 		}
8128 		if (hpts_calling == 0) {
8129 			/*
8130 			 * A user send or queued mbuf (sack) has called us? We
8131 			 * return 0 and let the pacing guards
8132 			 * deal with it if they should or
8133 			 * should not cause a send.
8134 			 */
8135 			ret = -2;
8136 			rack_log_to_processing(rack, cts, ret, 0);
8137 			return (0);
8138 		}
8139 		/*
8140 		 * Ok our timer went off early and we are not paced false
8141 		 * alarm, go back to sleep. We make sure we don't have
8142 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8143 		 * flag in place.
8144 		 */
8145 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8146 		ret = -3;
8147 		left = rack->r_ctl.rc_timer_exp - cts;
8148 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8149 		rack_log_to_processing(rack, cts, ret, left);
8150 		return (1);
8151 	}
8152 skip_time_check:
8153 	rack->rc_tmr_stopped = 0;
8154 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8155 	if (timers & PACE_TMR_DELACK) {
8156 		ret = rack_timeout_delack(tp, rack, cts);
8157 	} else if (timers & PACE_TMR_RACK) {
8158 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8159 		rack->r_fast_output = 0;
8160 		ret = rack_timeout_rack(tp, rack, cts);
8161 	} else if (timers & PACE_TMR_TLP) {
8162 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8163 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8164 	} else if (timers & PACE_TMR_RXT) {
8165 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8166 		rack->r_fast_output = 0;
8167 		ret = rack_timeout_rxt(tp, rack, cts);
8168 	} else if (timers & PACE_TMR_PERSIT) {
8169 		ret = rack_timeout_persist(tp, rack, cts);
8170 	} else if (timers & PACE_TMR_KEEP) {
8171 		ret = rack_timeout_keepalive(tp, rack, cts);
8172 	}
8173 	rack_log_to_processing(rack, cts, ret, timers);
8174 	return (ret);
8175 }
8176 
8177 static void
8178 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8179 {
8180 	struct timeval tv;
8181 	uint32_t us_cts, flags_on_entry;
8182 	uint8_t hpts_removed = 0;
8183 
8184 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8185 	us_cts = tcp_get_usecs(&tv);
8186 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8187 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8188 	     ((tp->snd_max - tp->snd_una) == 0))) {
8189 		tcp_hpts_remove(rack->rc_tp);
8190 		hpts_removed = 1;
8191 		/* If we were not delayed cancel out the flag. */
8192 		if ((tp->snd_max - tp->snd_una) == 0)
8193 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8194 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8195 	}
8196 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8197 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8198 		if (tcp_in_hpts(rack->rc_tp) &&
8199 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8200 			/*
8201 			 * Canceling timer's when we have no output being
8202 			 * paced. We also must remove ourselves from the
8203 			 * hpts.
8204 			 */
8205 			tcp_hpts_remove(rack->rc_tp);
8206 			hpts_removed = 1;
8207 		}
8208 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8209 	}
8210 	if (hpts_removed == 0)
8211 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8212 }
8213 
8214 static int
8215 rack_stopall(struct tcpcb *tp)
8216 {
8217 	struct tcp_rack *rack;
8218 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8219 	rack->t_timers_stopped = 1;
8220 	return (0);
8221 }
8222 
8223 static void
8224 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8225 {
8226 	/*
8227 	 * Assure no timers are running.
8228 	 */
8229 	if (tcp_timer_active(tp, TT_PERSIST)) {
8230 		/* We enter in persists, set the flag appropriately */
8231 		rack->rc_in_persist = 1;
8232 	}
8233 	if (tcp_in_hpts(rack->rc_tp)) {
8234 		tcp_hpts_remove(rack->rc_tp);
8235 	}
8236 }
8237 
8238 static void
8239 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8240     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz)
8241 {
8242 	int32_t idx;
8243 
8244 	rsm->r_rtr_cnt++;
8245 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8246 	rsm->r_dupack = 0;
8247 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8248 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8249 		rsm->r_flags |= RACK_OVERMAX;
8250 	}
8251 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8252 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8253 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8254 	}
8255 	idx = rsm->r_rtr_cnt - 1;
8256 	rsm->r_tim_lastsent[idx] = ts;
8257 	/*
8258 	 * Here we don't add in the len of send, since its already
8259 	 * in snduna <->snd_max.
8260 	 */
8261 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8262 				     rack->r_ctl.rc_sacked);
8263 	if (rsm->r_flags & RACK_ACKED) {
8264 		/* Problably MTU discovery messing with us */
8265 		rsm->r_flags &= ~RACK_ACKED;
8266 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8267 	}
8268 	if (rsm->r_in_tmap) {
8269 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8270 		rsm->r_in_tmap = 0;
8271 	}
8272 	/* Lets make sure it really is in or not the GP window */
8273 	rack_mark_in_gp_win(tp, rsm);
8274 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8275 	rsm->r_in_tmap = 1;
8276 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8277 	/* Take off the must retransmit flag, if its on */
8278 	if (rsm->r_flags & RACK_MUST_RXT) {
8279 		if (rack->r_must_retran)
8280 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8281 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8282 			/*
8283 			 * We have retransmitted all we need. Clear
8284 			 * any must retransmit flags.
8285 			 */
8286 			rack->r_must_retran = 0;
8287 			rack->r_ctl.rc_out_at_rto = 0;
8288 		}
8289 		rsm->r_flags &= ~RACK_MUST_RXT;
8290 	}
8291 	/* Remove any collapsed flag */
8292 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8293 	if (rsm->r_flags & RACK_SACK_PASSED) {
8294 		/* We have retransmitted due to the SACK pass */
8295 		rsm->r_flags &= ~RACK_SACK_PASSED;
8296 		rsm->r_flags |= RACK_WAS_SACKPASS;
8297 	}
8298 }
8299 
8300 static uint32_t
8301 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8302     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag, int segsiz)
8303 {
8304 	/*
8305 	 * We (re-)transmitted starting at rsm->r_start for some length
8306 	 * (possibly less than r_end.
8307 	 */
8308 	struct rack_sendmap *nrsm;
8309 	int insret __diagused;
8310 	uint32_t c_end;
8311 	int32_t len;
8312 
8313 	len = *lenp;
8314 	c_end = rsm->r_start + len;
8315 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8316 		/*
8317 		 * We retransmitted the whole piece or more than the whole
8318 		 * slopping into the next rsm.
8319 		 */
8320 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8321 		if (c_end == rsm->r_end) {
8322 			*lenp = 0;
8323 			return (0);
8324 		} else {
8325 			int32_t act_len;
8326 
8327 			/* Hangs over the end return whats left */
8328 			act_len = rsm->r_end - rsm->r_start;
8329 			*lenp = (len - act_len);
8330 			return (rsm->r_end);
8331 		}
8332 		/* We don't get out of this block. */
8333 	}
8334 	/*
8335 	 * Here we retransmitted less than the whole thing which means we
8336 	 * have to split this into what was transmitted and what was not.
8337 	 */
8338 	nrsm = rack_alloc_full_limit(rack);
8339 	if (nrsm == NULL) {
8340 		/*
8341 		 * We can't get memory, so lets not proceed.
8342 		 */
8343 		*lenp = 0;
8344 		return (0);
8345 	}
8346 	/*
8347 	 * So here we are going to take the original rsm and make it what we
8348 	 * retransmitted. nrsm will be the tail portion we did not
8349 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8350 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8351 	 * 1, 6 and the new piece will be 6, 11.
8352 	 */
8353 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8354 	nrsm->r_dupack = 0;
8355 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8356 #ifndef INVARIANTS
8357 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8358 #else
8359 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8360 		panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8361 		      nrsm, insret, rack, rsm);
8362 	}
8363 #endif
8364 	if (rsm->r_in_tmap) {
8365 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8366 		nrsm->r_in_tmap = 1;
8367 	}
8368 	rsm->r_flags &= (~RACK_HAS_FIN);
8369 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8370 	/* Log a split of rsm into rsm and nrsm */
8371 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8372 	*lenp = 0;
8373 	return (0);
8374 }
8375 
8376 static void
8377 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8378 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8379 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb,
8380 		uint32_t s_moff, int hw_tls, int segsiz)
8381 {
8382 	struct tcp_rack *rack;
8383 	struct rack_sendmap *rsm, *nrsm;
8384 	int insret __diagused;
8385 
8386 	register uint32_t snd_max, snd_una;
8387 
8388 	/*
8389 	 * Add to the RACK log of packets in flight or retransmitted. If
8390 	 * there is a TS option we will use the TS echoed, if not we will
8391 	 * grab a TS.
8392 	 *
8393 	 * Retransmissions will increment the count and move the ts to its
8394 	 * proper place. Note that if options do not include TS's then we
8395 	 * won't be able to effectively use the ACK for an RTT on a retran.
8396 	 *
8397 	 * Notes about r_start and r_end. Lets consider a send starting at
8398 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8399 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8400 	 * This means that r_end is actually the first sequence for the next
8401 	 * slot (11).
8402 	 *
8403 	 */
8404 	/*
8405 	 * If err is set what do we do XXXrrs? should we not add the thing?
8406 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8407 	 * i.e. proceed with add ** do this for now.
8408 	 */
8409 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8410 	if (err)
8411 		/*
8412 		 * We don't log errors -- we could but snd_max does not
8413 		 * advance in this case either.
8414 		 */
8415 		return;
8416 
8417 	if (th_flags & TH_RST) {
8418 		/*
8419 		 * We don't log resets and we return immediately from
8420 		 * sending
8421 		 */
8422 		return;
8423 	}
8424 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8425 	snd_una = tp->snd_una;
8426 	snd_max = tp->snd_max;
8427 	if (th_flags & (TH_SYN | TH_FIN)) {
8428 		/*
8429 		 * The call to rack_log_output is made before bumping
8430 		 * snd_max. This means we can record one extra byte on a SYN
8431 		 * or FIN if seq_out is adding more on and a FIN is present
8432 		 * (and we are not resending).
8433 		 */
8434 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8435 			len++;
8436 		if (th_flags & TH_FIN)
8437 			len++;
8438 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
8439 			/*
8440 			 * The add/update as not been done for the FIN/SYN
8441 			 * yet.
8442 			 */
8443 			snd_max = tp->snd_nxt;
8444 		}
8445 	}
8446 	if (SEQ_LEQ((seq_out + len), snd_una)) {
8447 		/* Are sending an old segment to induce an ack (keep-alive)? */
8448 		return;
8449 	}
8450 	if (SEQ_LT(seq_out, snd_una)) {
8451 		/* huh? should we panic? */
8452 		uint32_t end;
8453 
8454 		end = seq_out + len;
8455 		seq_out = snd_una;
8456 		if (SEQ_GEQ(end, seq_out))
8457 			len = end - seq_out;
8458 		else
8459 			len = 0;
8460 	}
8461 	if (len == 0) {
8462 		/* We don't log zero window probes */
8463 		return;
8464 	}
8465 	if (IN_FASTRECOVERY(tp->t_flags)) {
8466 		rack->r_ctl.rc_prr_out += len;
8467 	}
8468 	/* First question is it a retransmission or new? */
8469 	if (seq_out == snd_max) {
8470 		/* Its new */
8471 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8472 again:
8473 		rsm = rack_alloc(rack);
8474 		if (rsm == NULL) {
8475 			/*
8476 			 * Hmm out of memory and the tcb got destroyed while
8477 			 * we tried to wait.
8478 			 */
8479 			return;
8480 		}
8481 		if (th_flags & TH_FIN) {
8482 			rsm->r_flags = RACK_HAS_FIN|add_flag;
8483 		} else {
8484 			rsm->r_flags = add_flag;
8485 		}
8486 		if (hw_tls)
8487 			rsm->r_hw_tls = 1;
8488 		rsm->r_tim_lastsent[0] = cts;
8489 		rsm->r_rtr_cnt = 1;
8490 		rsm->r_rtr_bytes = 0;
8491 		if (th_flags & TH_SYN) {
8492 			/* The data space is one beyond snd_una */
8493 			rsm->r_flags |= RACK_HAS_SYN;
8494 		}
8495 		rsm->r_start = seq_out;
8496 		rsm->r_end = rsm->r_start + len;
8497 		rack_mark_in_gp_win(tp, rsm);
8498 		rsm->r_dupack = 0;
8499 		/*
8500 		 * save off the mbuf location that
8501 		 * sndmbuf_noadv returned (which is
8502 		 * where we started copying from)..
8503 		 */
8504 		rsm->m = s_mb;
8505 		rsm->soff = s_moff;
8506 		/*
8507 		 * Here we do add in the len of send, since its not yet
8508 		 * reflected in in snduna <->snd_max
8509 		 */
8510 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8511 					      rack->r_ctl.rc_sacked) +
8512 			      (rsm->r_end - rsm->r_start));
8513 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8514 		if (rsm->m) {
8515 			if (rsm->m->m_len <= rsm->soff) {
8516 				/*
8517 				 * XXXrrs Question, will this happen?
8518 				 *
8519 				 * If sbsndptr is set at the correct place
8520 				 * then s_moff should always be somewhere
8521 				 * within rsm->m. But if the sbsndptr was
8522 				 * off then that won't be true. If it occurs
8523 				 * we need to walkout to the correct location.
8524 				 */
8525 				struct mbuf *lm;
8526 
8527 				lm = rsm->m;
8528 				while (lm->m_len <= rsm->soff) {
8529 					rsm->soff -= lm->m_len;
8530 					lm = lm->m_next;
8531 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8532 							     __func__, rack, s_moff, s_mb, rsm->soff));
8533 				}
8534 				rsm->m = lm;
8535 			}
8536 			rsm->orig_m_len = rsm->m->m_len;
8537 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8538 		} else {
8539 			rsm->orig_m_len = 0;
8540 			rsm->orig_t_space = 0;
8541 		}
8542 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8543 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8544 		/* Log a new rsm */
8545 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8546 #ifndef INVARIANTS
8547 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
8548 #else
8549 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8550 			panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8551 			      nrsm, insret, rack, rsm);
8552 		}
8553 #endif
8554 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8555 		rsm->r_in_tmap = 1;
8556 		/*
8557 		 * Special case detection, is there just a single
8558 		 * packet outstanding when we are not in recovery?
8559 		 *
8560 		 * If this is true mark it so.
8561 		 */
8562 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8563 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8564 			struct rack_sendmap *prsm;
8565 
8566 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8567 			if (prsm)
8568 				prsm->r_one_out_nr = 1;
8569 		}
8570 		return;
8571 	}
8572 	/*
8573 	 * If we reach here its a retransmission and we need to find it.
8574 	 */
8575 more:
8576 	if (hintrsm && (hintrsm->r_start == seq_out)) {
8577 		rsm = hintrsm;
8578 		hintrsm = NULL;
8579 	} else {
8580 		/* No hints sorry */
8581 		rsm = NULL;
8582 	}
8583 	if ((rsm) && (rsm->r_start == seq_out)) {
8584 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8585 		if (len == 0) {
8586 			return;
8587 		} else {
8588 			goto more;
8589 		}
8590 	}
8591 	/* Ok it was not the last pointer go through it the hard way. */
8592 refind:
8593 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8594 	if (rsm) {
8595 		if (rsm->r_start == seq_out) {
8596 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8597 			if (len == 0) {
8598 				return;
8599 			} else {
8600 				goto refind;
8601 			}
8602 		}
8603 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8604 			/* Transmitted within this piece */
8605 			/*
8606 			 * Ok we must split off the front and then let the
8607 			 * update do the rest
8608 			 */
8609 			nrsm = rack_alloc_full_limit(rack);
8610 			if (nrsm == NULL) {
8611 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8612 				return;
8613 			}
8614 			/*
8615 			 * copy rsm to nrsm and then trim the front of rsm
8616 			 * to not include this part.
8617 			 */
8618 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
8619 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8620 #ifndef INVARIANTS
8621 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8622 #else
8623 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8624 				panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
8625 				      nrsm, insret, rack, rsm);
8626 			}
8627 #endif
8628 			if (rsm->r_in_tmap) {
8629 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8630 				nrsm->r_in_tmap = 1;
8631 			}
8632 			rsm->r_flags &= (~RACK_HAS_FIN);
8633 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8634 			if (len == 0) {
8635 				return;
8636 			} else if (len > 0)
8637 				goto refind;
8638 		}
8639 	}
8640 	/*
8641 	 * Hmm not found in map did they retransmit both old and on into the
8642 	 * new?
8643 	 */
8644 	if (seq_out == tp->snd_max) {
8645 		goto again;
8646 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
8647 #ifdef INVARIANTS
8648 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8649 		       seq_out, len, tp->snd_una, tp->snd_max);
8650 		printf("Starting Dump of all rack entries\n");
8651 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8652 			printf("rsm:%p start:%u end:%u\n",
8653 			       rsm, rsm->r_start, rsm->r_end);
8654 		}
8655 		printf("Dump complete\n");
8656 		panic("seq_out not found rack:%p tp:%p",
8657 		      rack, tp);
8658 #endif
8659 	} else {
8660 #ifdef INVARIANTS
8661 		/*
8662 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8663 		 * flag)
8664 		 */
8665 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8666 		      seq_out, len, tp->snd_max, tp);
8667 #endif
8668 	}
8669 }
8670 
8671 /*
8672  * Record one of the RTT updates from an ack into
8673  * our sample structure.
8674  */
8675 
8676 static void
8677 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8678 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8679 {
8680 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8681 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8682 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8683 	}
8684 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8685 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8686 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8687 	}
8688 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8689 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8690 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
8691 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8692 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8693 	}
8694 	if ((confidence == 1) &&
8695 	    ((rsm == NULL) ||
8696 	     (rsm->r_just_ret) ||
8697 	     (rsm->r_one_out_nr &&
8698 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8699 		/*
8700 		 * If the rsm had a just return
8701 		 * hit it then we can't trust the
8702 		 * rtt measurement for buffer deterimination
8703 		 * Note that a confidence of 2, indicates
8704 		 * SACK'd which overrides the r_just_ret or
8705 		 * the r_one_out_nr. If it was a CUM-ACK and
8706 		 * we had only two outstanding, but get an
8707 		 * ack for only 1. Then that also lowers our
8708 		 * confidence.
8709 		 */
8710 		confidence = 0;
8711 	}
8712 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8713 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8714 		if (rack->r_ctl.rack_rs.confidence == 0) {
8715 			/*
8716 			 * We take anything with no current confidence
8717 			 * saved.
8718 			 */
8719 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8720 			rack->r_ctl.rack_rs.confidence = confidence;
8721 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8722 		} else if (confidence != 0) {
8723 			/*
8724 			 * Once we have a confident number,
8725 			 * we can update it with a smaller
8726 			 * value since this confident number
8727 			 * may include the DSACK time until
8728 			 * the next segment (the second one) arrived.
8729 			 */
8730 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8731 			rack->r_ctl.rack_rs.confidence = confidence;
8732 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8733 		}
8734 	}
8735 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8736 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8737 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8738 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
8739 }
8740 
8741 /*
8742  * Collect new round-trip time estimate
8743  * and update averages and current timeout.
8744  */
8745 static void
8746 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8747 {
8748 	int32_t delta;
8749 	int32_t rtt;
8750 
8751 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8752 		/* No valid sample */
8753 		return;
8754 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8755 		/* We are to use the lowest RTT seen in a single ack */
8756 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8757 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8758 		/* We are to use the highest RTT seen in a single ack */
8759 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8760 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8761 		/* We are to use the average RTT seen in a single ack */
8762 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8763 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8764 	} else {
8765 #ifdef INVARIANTS
8766 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8767 #endif
8768 		return;
8769 	}
8770 	if (rtt == 0)
8771 		rtt = 1;
8772 	if (rack->rc_gp_rtt_set == 0) {
8773 		/*
8774 		 * With no RTT we have to accept
8775 		 * even one we are not confident of.
8776 		 */
8777 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8778 		rack->rc_gp_rtt_set = 1;
8779 	} else if (rack->r_ctl.rack_rs.confidence) {
8780 		/* update the running gp srtt */
8781 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8782 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8783 	}
8784 	if (rack->r_ctl.rack_rs.confidence) {
8785 		/*
8786 		 * record the low and high for highly buffered path computation,
8787 		 * we only do this if we are confident (not a retransmission).
8788 		 */
8789 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8790 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8791 		}
8792 		if (rack->rc_highly_buffered == 0) {
8793 			/*
8794 			 * Currently once we declare a path has
8795 			 * highly buffered there is no going
8796 			 * back, which may be a problem...
8797 			 */
8798 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8799 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8800 						     rack->r_ctl.rc_highest_us_rtt,
8801 						     rack->r_ctl.rc_lowest_us_rtt,
8802 						     RACK_RTTS_SEEHBP);
8803 				rack->rc_highly_buffered = 1;
8804 			}
8805 		}
8806 	}
8807 	if ((rack->r_ctl.rack_rs.confidence) ||
8808 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8809 		/*
8810 		 * If we are highly confident of it <or> it was
8811 		 * never retransmitted we accept it as the last us_rtt.
8812 		 */
8813 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8814 		/* The lowest rtt can be set if its was not retransmited */
8815 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8816 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8817 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
8818 				rack->r_ctl.rc_lowest_us_rtt = 1;
8819 		}
8820 	}
8821 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8822 	if (tp->t_srtt != 0) {
8823 		/*
8824 		 * We keep a simple srtt in microseconds, like our rtt
8825 		 * measurement. We don't need to do any tricks with shifting
8826 		 * etc. Instead we just add in 1/8th of the new measurement
8827 		 * and subtract out 1/8 of the old srtt. We do the same with
8828 		 * the variance after finding the absolute value of the
8829 		 * difference between this sample and the current srtt.
8830 		 */
8831 		delta = tp->t_srtt - rtt;
8832 		/* Take off 1/8th of the current sRTT */
8833 		tp->t_srtt -= (tp->t_srtt >> 3);
8834 		/* Add in 1/8th of the new RTT just measured */
8835 		tp->t_srtt += (rtt >> 3);
8836 		if (tp->t_srtt <= 0)
8837 			tp->t_srtt = 1;
8838 		/* Now lets make the absolute value of the variance */
8839 		if (delta < 0)
8840 			delta = -delta;
8841 		/* Subtract out 1/8th */
8842 		tp->t_rttvar -= (tp->t_rttvar >> 3);
8843 		/* Add in 1/8th of the new variance we just saw */
8844 		tp->t_rttvar += (delta >> 3);
8845 		if (tp->t_rttvar <= 0)
8846 			tp->t_rttvar = 1;
8847 	} else {
8848 		/*
8849 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
8850 		 * variance to half the rtt (so our first retransmit happens
8851 		 * at 3*rtt).
8852 		 */
8853 		tp->t_srtt = rtt;
8854 		tp->t_rttvar = rtt >> 1;
8855 	}
8856 	rack->rc_srtt_measure_made = 1;
8857 	KMOD_TCPSTAT_INC(tcps_rttupdated);
8858 	if (tp->t_rttupdated < UCHAR_MAX)
8859 		tp->t_rttupdated++;
8860 #ifdef STATS
8861 	if (rack_stats_gets_ms_rtt == 0) {
8862 		/* Send in the microsecond rtt used for rxt timeout purposes */
8863 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8864 	} else if (rack_stats_gets_ms_rtt == 1) {
8865 		/* Send in the millisecond rtt used for rxt timeout purposes */
8866 		int32_t ms_rtt;
8867 
8868 		/* Round up */
8869 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8870 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8871 	} else if (rack_stats_gets_ms_rtt == 2) {
8872 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
8873 		int32_t ms_rtt;
8874 
8875 		/* Round up */
8876 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8877 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8878 	}  else {
8879 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
8880 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8881 	}
8882 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8883 #endif
8884 	/*
8885 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8886 	 * way we do the smoothing, srtt and rttvar will each average +1/2
8887 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
8888 	 * tick of rounding and 1 extra tick because of +-1/2 tick
8889 	 * uncertainty in the firing of the timer.  The bias will give us
8890 	 * exactly the 1.5 tick we need.  But, because the bias is
8891 	 * statistical, we have to test that we don't drop below the minimum
8892 	 * feasible timer (which is 2 ticks).
8893 	 */
8894 	tp->t_rxtshift = 0;
8895 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8896 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8897 	rack_log_rtt_sample(rack, rtt);
8898 	tp->t_softerror = 0;
8899 }
8900 
8901 
8902 static void
8903 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8904 {
8905 	/*
8906 	 * Apply to filter the inbound us-rtt at us_cts.
8907 	 */
8908 	uint32_t old_rtt;
8909 
8910 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8911 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8912 			       us_rtt, us_cts);
8913 	if (old_rtt > us_rtt) {
8914 		/* We just hit a new lower rtt time */
8915 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8916 				     __LINE__, RACK_RTTS_NEWRTT);
8917 		/*
8918 		 * Only count it if its lower than what we saw within our
8919 		 * calculated range.
8920 		 */
8921 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8922 			if (rack_probertt_lower_within &&
8923 			    rack->rc_gp_dyn_mul &&
8924 			    (rack->use_fixed_rate == 0) &&
8925 			    (rack->rc_always_pace)) {
8926 				/*
8927 				 * We are seeing a new lower rtt very close
8928 				 * to the time that we would have entered probe-rtt.
8929 				 * This is probably due to the fact that a peer flow
8930 				 * has entered probe-rtt. Lets go in now too.
8931 				 */
8932 				uint32_t val;
8933 
8934 				val = rack_probertt_lower_within * rack_time_between_probertt;
8935 				val /= 100;
8936 				if ((rack->in_probe_rtt == 0)  &&
8937 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
8938 					rack_enter_probertt(rack, us_cts);
8939 				}
8940 			}
8941 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8942 		}
8943 	}
8944 }
8945 
8946 static int
8947 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8948     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8949 {
8950 	uint32_t us_rtt;
8951 	int32_t i, all;
8952 	uint32_t t, len_acked;
8953 
8954 	if ((rsm->r_flags & RACK_ACKED) ||
8955 	    (rsm->r_flags & RACK_WAS_ACKED))
8956 		/* Already done */
8957 		return (0);
8958 	if (rsm->r_no_rtt_allowed) {
8959 		/* Not allowed */
8960 		return (0);
8961 	}
8962 	if (ack_type == CUM_ACKED) {
8963 		if (SEQ_GT(th_ack, rsm->r_end)) {
8964 			len_acked = rsm->r_end - rsm->r_start;
8965 			all = 1;
8966 		} else {
8967 			len_acked = th_ack - rsm->r_start;
8968 			all = 0;
8969 		}
8970 	} else {
8971 		len_acked = rsm->r_end - rsm->r_start;
8972 		all = 0;
8973 	}
8974 	if (rsm->r_rtr_cnt == 1) {
8975 
8976 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8977 		if ((int)t <= 0)
8978 			t = 1;
8979 		if (!tp->t_rttlow || tp->t_rttlow > t)
8980 			tp->t_rttlow = t;
8981 		if (!rack->r_ctl.rc_rack_min_rtt ||
8982 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8983 			rack->r_ctl.rc_rack_min_rtt = t;
8984 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8985 				rack->r_ctl.rc_rack_min_rtt = 1;
8986 			}
8987 		}
8988 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8989 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8990 		else
8991 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8992 		if (us_rtt == 0)
8993 			us_rtt = 1;
8994 		if (CC_ALGO(tp)->rttsample != NULL) {
8995 			/* Kick the RTT to the CC */
8996 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8997 		}
8998 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8999 		if (ack_type == SACKED) {
9000 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
9001 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
9002 		} else {
9003 			/*
9004 			 * We need to setup what our confidence
9005 			 * is in this ack.
9006 			 *
9007 			 * If the rsm was app limited and it is
9008 			 * less than a mss in length (the end
9009 			 * of the send) then we have a gap. If we
9010 			 * were app limited but say we were sending
9011 			 * multiple MSS's then we are more confident
9012 			 * int it.
9013 			 *
9014 			 * When we are not app-limited then we see if
9015 			 * the rsm is being included in the current
9016 			 * measurement, we tell this by the app_limited_needs_set
9017 			 * flag.
9018 			 *
9019 			 * Note that being cwnd blocked is not applimited
9020 			 * as well as the pacing delay between packets which
9021 			 * are sending only 1 or 2 MSS's also will show up
9022 			 * in the RTT. We probably need to examine this algorithm
9023 			 * a bit more and enhance it to account for the delay
9024 			 * between rsm's. We could do that by saving off the
9025 			 * pacing delay of each rsm (in an rsm) and then
9026 			 * factoring that in somehow though for now I am
9027 			 * not sure how :)
9028 			 */
9029 			int calc_conf = 0;
9030 
9031 			if (rsm->r_flags & RACK_APP_LIMITED) {
9032 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
9033 					calc_conf = 0;
9034 				else
9035 					calc_conf = 1;
9036 			} else if (rack->app_limited_needs_set == 0) {
9037 				calc_conf = 1;
9038 			} else {
9039 				calc_conf = 0;
9040 			}
9041 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
9042 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
9043 					    calc_conf, rsm, rsm->r_rtr_cnt);
9044 		}
9045 		if ((rsm->r_flags & RACK_TLP) &&
9046 		    (!IN_FASTRECOVERY(tp->t_flags))) {
9047 			/* Segment was a TLP and our retrans matched */
9048 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
9049 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9050 			}
9051 		}
9052 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9053 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9054 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9055 			/* New more recent rack_tmit_time */
9056 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9057 			if (rack->r_ctl.rc_rack_tmit_time == 0)
9058 				rack->r_ctl.rc_rack_tmit_time = 1;
9059 			rack->rc_rack_rtt = t;
9060 		}
9061 		return (1);
9062 	}
9063 	/*
9064 	 * We clear the soft/rxtshift since we got an ack.
9065 	 * There is no assurance we will call the commit() function
9066 	 * so we need to clear these to avoid incorrect handling.
9067 	 */
9068 	tp->t_rxtshift = 0;
9069 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9070 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9071 	tp->t_softerror = 0;
9072 	if (to && (to->to_flags & TOF_TS) &&
9073 	    (ack_type == CUM_ACKED) &&
9074 	    (to->to_tsecr) &&
9075 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
9076 		/*
9077 		 * Now which timestamp does it match? In this block the ACK
9078 		 * must be coming from a previous transmission.
9079 		 */
9080 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9081 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9082 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9083 				if ((int)t <= 0)
9084 					t = 1;
9085 				if (CC_ALGO(tp)->rttsample != NULL) {
9086 					/*
9087 					 * Kick the RTT to the CC, here
9088 					 * we lie a bit in that we know the
9089 					 * retransmission is correct even though
9090 					 * we retransmitted. This is because
9091 					 * we match the timestamps.
9092 					 */
9093 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9094 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9095 					else
9096 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9097 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9098 				}
9099 				if ((i + 1) < rsm->r_rtr_cnt) {
9100 					/*
9101 					 * The peer ack'd from our previous
9102 					 * transmission. We have a spurious
9103 					 * retransmission and thus we dont
9104 					 * want to update our rack_rtt.
9105 					 *
9106 					 * Hmm should there be a CC revert here?
9107 					 *
9108 					 */
9109 					return (0);
9110 				}
9111 				if (!tp->t_rttlow || tp->t_rttlow > t)
9112 					tp->t_rttlow = t;
9113 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9114 					rack->r_ctl.rc_rack_min_rtt = t;
9115 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9116 						rack->r_ctl.rc_rack_min_rtt = 1;
9117 					}
9118 				}
9119 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9120 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9121 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9122 					/* New more recent rack_tmit_time */
9123 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9124 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9125 						rack->r_ctl.rc_rack_tmit_time = 1;
9126 					rack->rc_rack_rtt = t;
9127 				}
9128 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9129 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9130 						    rsm->r_rtr_cnt);
9131 				return (1);
9132 			}
9133 		}
9134 		/* If we are logging log out the sendmap */
9135 		if (tcp_bblogging_on(rack->rc_tp)) {
9136 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9137 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9138 			}
9139 		}
9140 		goto ts_not_found;
9141 	} else {
9142 		/*
9143 		 * Ok its a SACK block that we retransmitted. or a windows
9144 		 * machine without timestamps. We can tell nothing from the
9145 		 * time-stamp since its not there or the time the peer last
9146 		 * recieved a segment that moved forward its cum-ack point.
9147 		 */
9148 ts_not_found:
9149 		i = rsm->r_rtr_cnt - 1;
9150 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9151 		if ((int)t <= 0)
9152 			t = 1;
9153 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9154 			/*
9155 			 * We retransmitted and the ack came back in less
9156 			 * than the smallest rtt we have observed. We most
9157 			 * likely did an improper retransmit as outlined in
9158 			 * 6.2 Step 2 point 2 in the rack-draft so we
9159 			 * don't want to update our rack_rtt. We in
9160 			 * theory (in future) might want to think about reverting our
9161 			 * cwnd state but we won't for now.
9162 			 */
9163 			return (0);
9164 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9165 			/*
9166 			 * We retransmitted it and the retransmit did the
9167 			 * job.
9168 			 */
9169 			if (!rack->r_ctl.rc_rack_min_rtt ||
9170 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9171 				rack->r_ctl.rc_rack_min_rtt = t;
9172 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9173 					rack->r_ctl.rc_rack_min_rtt = 1;
9174 				}
9175 			}
9176 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9177 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9178 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9179 				/* New more recent rack_tmit_time */
9180 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9181 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9182 					rack->r_ctl.rc_rack_tmit_time = 1;
9183 				rack->rc_rack_rtt = t;
9184 			}
9185 			return (1);
9186 		}
9187 	}
9188 	return (0);
9189 }
9190 
9191 /*
9192  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9193  */
9194 static void
9195 rack_log_sack_passed(struct tcpcb *tp,
9196     struct tcp_rack *rack, struct rack_sendmap *rsm)
9197 {
9198 	struct rack_sendmap *nrsm;
9199 
9200 	nrsm = rsm;
9201 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9202 	    rack_head, r_tnext) {
9203 		if (nrsm == rsm) {
9204 			/* Skip original segment he is acked */
9205 			continue;
9206 		}
9207 		if (nrsm->r_flags & RACK_ACKED) {
9208 			/*
9209 			 * Skip ack'd segments, though we
9210 			 * should not see these, since tmap
9211 			 * should not have ack'd segments.
9212 			 */
9213 			continue;
9214 		}
9215 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9216 			/*
9217 			 * If the peer dropped the rwnd on
9218 			 * these then we don't worry about them.
9219 			 */
9220 			continue;
9221 		}
9222 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9223 			/*
9224 			 * We found one that is already marked
9225 			 * passed, we have been here before and
9226 			 * so all others below this are marked.
9227 			 */
9228 			break;
9229 		}
9230 		nrsm->r_flags |= RACK_SACK_PASSED;
9231 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9232 	}
9233 }
9234 
9235 static void
9236 rack_need_set_test(struct tcpcb *tp,
9237 		   struct tcp_rack *rack,
9238 		   struct rack_sendmap *rsm,
9239 		   tcp_seq th_ack,
9240 		   int line,
9241 		   int use_which)
9242 {
9243 	struct rack_sendmap *s_rsm;
9244 
9245 	if ((tp->t_flags & TF_GPUTINPROG) &&
9246 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9247 		/*
9248 		 * We were app limited, and this ack
9249 		 * butts up or goes beyond the point where we want
9250 		 * to start our next measurement. We need
9251 		 * to record the new gput_ts as here and
9252 		 * possibly update the start sequence.
9253 		 */
9254 		uint32_t seq, ts;
9255 
9256 		if (rsm->r_rtr_cnt > 1) {
9257 			/*
9258 			 * This is a retransmit, can we
9259 			 * really make any assessment at this
9260 			 * point?  We are not really sure of
9261 			 * the timestamp, is it this or the
9262 			 * previous transmission?
9263 			 *
9264 			 * Lets wait for something better that
9265 			 * is not retransmitted.
9266 			 */
9267 			return;
9268 		}
9269 		seq = tp->gput_seq;
9270 		ts = tp->gput_ts;
9271 		rack->app_limited_needs_set = 0;
9272 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9273 		/* Do we start at a new end? */
9274 		if ((use_which == RACK_USE_BEG) &&
9275 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9276 			/*
9277 			 * When we get an ACK that just eats
9278 			 * up some of the rsm, we set RACK_USE_BEG
9279 			 * since whats at r_start (i.e. th_ack)
9280 			 * is left unacked and thats where the
9281 			 * measurement now starts.
9282 			 */
9283 			tp->gput_seq = rsm->r_start;
9284 		}
9285 		if ((use_which == RACK_USE_END) &&
9286 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9287 			/*
9288 			 * We use the end when the cumack
9289 			 * is moving forward and completely
9290 			 * deleting the rsm passed so basically
9291 			 * r_end holds th_ack.
9292 			 *
9293 			 * For SACK's we also want to use the end
9294 			 * since this piece just got sacked and
9295 			 * we want to target anything after that
9296 			 * in our measurement.
9297 			 */
9298 			tp->gput_seq = rsm->r_end;
9299 		}
9300 		if (use_which == RACK_USE_END_OR_THACK) {
9301 			/*
9302 			 * special case for ack moving forward,
9303 			 * not a sack, we need to move all the
9304 			 * way up to where this ack cum-ack moves
9305 			 * to.
9306 			 */
9307 			if (SEQ_GT(th_ack, rsm->r_end))
9308 				tp->gput_seq = th_ack;
9309 			else
9310 				tp->gput_seq = rsm->r_end;
9311 		}
9312 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9313 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9314 		else
9315 			s_rsm = NULL;
9316 		/*
9317 		 * Pick up the correct send time if we can the rsm passed in
9318 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9319 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9320 		 * find a different seq i.e. the next send up.
9321 		 *
9322 		 * If that has not been sent, s_rsm will be NULL and we must
9323 		 * arrange it so this function will get called again by setting
9324 		 * app_limited_needs_set.
9325 		 */
9326 		if (s_rsm)
9327 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9328 		else {
9329 			/* If we hit here we have to have *not* sent tp->gput_seq */
9330 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9331 			/* Set it up so we will go through here again */
9332 			rack->app_limited_needs_set = 1;
9333 		}
9334 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9335 			/*
9336 			 * We moved beyond this guy's range, re-calculate
9337 			 * the new end point.
9338 			 */
9339 			if (rack->rc_gp_filled == 0) {
9340 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9341 			} else {
9342 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9343 			}
9344 		}
9345 		/*
9346 		 * We are moving the goal post, we may be able to clear the
9347 		 * measure_saw_probe_rtt flag.
9348 		 */
9349 		if ((rack->in_probe_rtt == 0) &&
9350 		    (rack->measure_saw_probe_rtt) &&
9351 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9352 			rack->measure_saw_probe_rtt = 0;
9353 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9354 					   seq, tp->gput_seq,
9355 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9356 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9357 					   5, line, NULL, 0);
9358 		if (rack->rc_gp_filled &&
9359 		    ((tp->gput_ack - tp->gput_seq) <
9360 		     max(rc_init_window(rack), (MIN_GP_WIN *
9361 						ctf_fixed_maxseg(tp))))) {
9362 			uint32_t ideal_amount;
9363 
9364 			ideal_amount = rack_get_measure_window(tp, rack);
9365 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9366 				/*
9367 				 * There is no sense of continuing this measurement
9368 				 * because its too small to gain us anything we
9369 				 * trust. Skip it and that way we can start a new
9370 				 * measurement quicker.
9371 				 */
9372 				tp->t_flags &= ~TF_GPUTINPROG;
9373 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9374 							   0, 0,
9375 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9376 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9377 							   6, __LINE__, NULL, 0);
9378 			} else {
9379 				/*
9380 				 * Reset the window further out.
9381 				 */
9382 				tp->gput_ack = tp->gput_seq + ideal_amount;
9383 			}
9384 		}
9385 		rack_tend_gp_marks(tp, rack);
9386 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9387 	}
9388 }
9389 
9390 static inline int
9391 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9392 {
9393 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9394 		/* Behind our TLP definition or right at */
9395 		return (0);
9396 	}
9397 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9398 		/* The start is beyond or right at our end of TLP definition */
9399 		return (0);
9400 	}
9401 	/* It has to be a sub-part of the original TLP recorded */
9402 	return (1);
9403 }
9404 
9405 
9406 
9407 static uint32_t
9408 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9409 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9410 		   int *no_extra,
9411 		   int *moved_two, uint32_t segsiz)
9412 {
9413 	uint32_t start, end, changed = 0;
9414 	struct rack_sendmap stack_map;
9415 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
9416 	int insret __diagused;
9417 	int32_t used_ref = 1;
9418 	int moved = 0;
9419 #ifdef TCP_SAD_DETECTION
9420 	int allow_segsiz;
9421 	int first_time_through = 1;
9422 #endif
9423 	int noextra = 0;
9424 	int can_use_hookery = 0;
9425 
9426 	start = sack->start;
9427 	end = sack->end;
9428 	rsm = *prsm;
9429 
9430 #ifdef TCP_SAD_DETECTION
9431 	/*
9432 	 * There are a strange number of proxys and meddle boxes in the world
9433 	 * that seem to cut up segments on different boundaries. This gets us
9434 	 * smaller sacks that are still ok in terms of it being an attacker.
9435 	 * We use the base segsiz to calculate an allowable smallness but
9436 	 * also enforce a min on the segsiz in case it is an attacker playing
9437 	 * games with MSS. So basically if the sack arrives and it is
9438 	 * larger than a worse case 960 bytes, we don't classify the guy
9439 	 * as supicious.
9440 	 */
9441 	allow_segsiz = max(segsiz, 1200) * sad_seg_size_per;
9442 	allow_segsiz /= 1000;
9443 #endif
9444 do_rest_ofb:
9445 	if ((rsm == NULL) ||
9446 	    (SEQ_LT(end, rsm->r_start)) ||
9447 	    (SEQ_GEQ(start, rsm->r_end)) ||
9448 	    (SEQ_LT(start, rsm->r_start))) {
9449 		/*
9450 		 * We are not in the right spot,
9451 		 * find the correct spot in the tree.
9452 		 */
9453 		used_ref = 0;
9454 		rsm = tqhash_find(rack->r_ctl.tqh, start);
9455 		moved++;
9456 	}
9457 	if (rsm == NULL) {
9458 		/* TSNH */
9459 		goto out;
9460 	}
9461 #ifdef TCP_SAD_DETECTION
9462 	/* Now we must check for suspicous activity */
9463 	if ((first_time_through == 1) &&
9464 	    ((end - start) < min((rsm->r_end - rsm->r_start), allow_segsiz)) &&
9465 	    ((rsm->r_flags & RACK_PMTU_CHG) == 0) &&
9466 	    ((rsm->r_flags & RACK_TLP) == 0)) {
9467 		/*
9468 		 * Its less than a full MSS or the segment being acked
9469 		 * this should only happen if the rsm in question had the
9470 		 * r_just_ret flag set <and> the end matches the end of
9471 		 * the rsm block.
9472 		 *
9473 		 * Note we do not look at segments that have had TLP's on
9474 		 * them since we can get un-reported rwnd collapses that
9475 		 * basically we TLP on and then we get back a sack block
9476 		 * that goes from the start to only a small way.
9477 		 *
9478 		 */
9479 		int loss, ok;
9480 
9481 		ok = 0;
9482 		if (SEQ_GEQ(end, rsm->r_end)) {
9483 			if (rsm->r_just_ret == 1) {
9484 				/* This was at the end of a send which is ok */
9485 				ok = 1;
9486 			} else {
9487 				/* A bit harder was it the end of our segment */
9488 				int segs, len;
9489 
9490 				len = (rsm->r_end - rsm->r_start);
9491 				segs = len / segsiz;
9492 				segs *= segsiz;
9493 				if ((segs + (rsm->r_end - start)) == len) {
9494 					/*
9495 					 * So this last bit was the
9496 					 * end of our send if we cut it
9497 					 * up into segsiz pieces so its ok.
9498 					 */
9499 					ok = 1;
9500 				}
9501 			}
9502 		}
9503 		if (ok == 0) {
9504 			/*
9505 			 * This guy is doing something suspicious
9506 			 * lets start detection.
9507 			 */
9508 			if (rack->rc_suspicious == 0) {
9509 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_SUSPECT);
9510 				counter_u64_add(rack_sack_attacks_suspect, 1);
9511 				rack->rc_suspicious = 1;
9512 				rack_log_sad(rack, 4);
9513 				if (tcp_bblogging_on(rack->rc_tp)) {
9514 					union tcp_log_stackspecific log;
9515 					struct timeval tv;
9516 
9517 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
9518 					log.u_bbr.flex1 = end;
9519 					log.u_bbr.flex2 = start;
9520 					log.u_bbr.flex3 = rsm->r_end;
9521 					log.u_bbr.flex4 = rsm->r_start;
9522 					log.u_bbr.flex5 = segsiz;
9523 					log.u_bbr.flex6 = rsm->r_fas;
9524 					log.u_bbr.flex7 = rsm->r_bas;
9525 					log.u_bbr.flex8 = 5;
9526 					log.u_bbr.pkts_out = rsm->r_flags;
9527 					log.u_bbr.bbr_state = rack->rc_suspicious;
9528 					log.u_bbr.bbr_substate = rsm->r_just_ret;
9529 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
9530 					log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9531 					TCP_LOG_EVENTP(rack->rc_tp, NULL,
9532 						       &rack->rc_inp->inp_socket->so_rcv,
9533 						       &rack->rc_inp->inp_socket->so_snd,
9534 						       TCP_SAD_DETECTION, 0,
9535 						       0, &log, false, &tv);
9536 				}
9537 			}
9538 			/* You loose some ack count every time you sack
9539 			 * a small bit that is not butting to the end of
9540 			 * what we have sent. This is because we never
9541 			 * send small bits unless its the end of the sb.
9542 			 * Anyone sending a sack that is not at the end
9543 			 * is thus very very suspicious.
9544 			 */
9545 			loss = (segsiz/2) / (end - start);
9546 			if (loss < rack->r_ctl.ack_count)
9547 				rack->r_ctl.ack_count -= loss;
9548 			else
9549 				rack->r_ctl.ack_count = 0;
9550 		}
9551 	}
9552 	first_time_through = 0;
9553 #endif
9554 	/* Ok we have an ACK for some piece of this rsm */
9555 	if (rsm->r_start != start) {
9556 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9557 			/*
9558 			 * Before any splitting or hookery is
9559 			 * done is it a TLP of interest i.e. rxt?
9560 			 */
9561 			if ((rsm->r_flags & RACK_TLP) &&
9562 			    (rsm->r_rtr_cnt > 1)) {
9563 				/*
9564 				 * We are splitting a rxt TLP, check
9565 				 * if we need to save off the start/end
9566 				 */
9567 				if (rack->rc_last_tlp_acked_set &&
9568 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9569 					/*
9570 					 * We already turned this on since we are inside
9571 					 * the previous one was a partially sack now we
9572 					 * are getting another one (maybe all of it).
9573 					 *
9574 					 */
9575 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9576 					/*
9577 					 * Lets make sure we have all of it though.
9578 					 */
9579 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9580 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9581 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9582 								     rack->r_ctl.last_tlp_acked_end);
9583 					}
9584 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9585 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9586 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9587 								     rack->r_ctl.last_tlp_acked_end);
9588 					}
9589 				} else {
9590 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9591 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9592 					rack->rc_last_tlp_past_cumack = 0;
9593 					rack->rc_last_tlp_acked_set = 1;
9594 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9595 				}
9596 			}
9597 			/**
9598 			 * Need to split this in two pieces the before and after,
9599 			 * the before remains in the map, the after must be
9600 			 * added. In other words we have:
9601 			 * rsm        |--------------|
9602 			 * sackblk        |------->
9603 			 * rsm will become
9604 			 *     rsm    |---|
9605 			 * and nrsm will be  the sacked piece
9606 			 *     nrsm       |----------|
9607 			 *
9608 			 * But before we start down that path lets
9609 			 * see if the sack spans over on top of
9610 			 * the next guy and it is already sacked.
9611 			 *
9612 			 */
9613 			/*
9614 			 * Hookery can only be used if the two entries
9615 			 * are in the same bucket and neither one of
9616 			 * them staddle the bucket line.
9617 			 */
9618 			next = tqhash_next(rack->r_ctl.tqh, rsm);
9619 			if (next &&
9620 			    (rsm->bindex == next->bindex) &&
9621 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9622 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
9623 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
9624 			    (next->r_flags & RACK_IN_GP_WIN))
9625 				can_use_hookery = 1;
9626 			else if (next &&
9627 				 (rsm->bindex == next->bindex) &&
9628 				 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9629 				 ((next->r_flags & RACK_STRADDLE) == 0) &&
9630 				 ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9631 				 ((next->r_flags & RACK_IN_GP_WIN) == 0))
9632 				can_use_hookery = 1;
9633 			else
9634 				can_use_hookery = 0;
9635 			if (next && can_use_hookery &&
9636 			    (next->r_flags & RACK_ACKED) &&
9637 			    SEQ_GEQ(end, next->r_start)) {
9638 				/**
9639 				 * So the next one is already acked, and
9640 				 * we can thus by hookery use our stack_map
9641 				 * to reflect the piece being sacked and
9642 				 * then adjust the two tree entries moving
9643 				 * the start and ends around. So we start like:
9644 				 *  rsm     |------------|             (not-acked)
9645 				 *  next                 |-----------| (acked)
9646 				 *  sackblk        |-------->
9647 				 *  We want to end like so:
9648 				 *  rsm     |------|                   (not-acked)
9649 				 *  next           |-----------------| (acked)
9650 				 *  nrsm           |-----|
9651 				 * Where nrsm is a temporary stack piece we
9652 				 * use to update all the gizmos.
9653 				 */
9654 				/* Copy up our fudge block */
9655 				noextra++;
9656 				nrsm = &stack_map;
9657 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9658 				/* Now adjust our tree blocks */
9659 				rsm->r_end = start;
9660 				next->r_start = start;
9661  				rsm->r_flags |= RACK_SHUFFLED;
9662 				next->r_flags |= RACK_SHUFFLED;
9663 				/* Now we must adjust back where next->m is */
9664 				rack_setup_offset_for_rsm(rack, rsm, next);
9665 				/*
9666 				 * Which timestamp do we keep? It is rather
9667 				 * important in GP measurements to have the
9668 				 * accurate end of the send window.
9669 				 *
9670 				 * We keep the largest value, which is the newest
9671 				 * send. We do this in case a segment that is
9672 				 * joined together and not part of a GP estimate
9673 				 * later gets expanded into the GP estimate.
9674 				 *
9675 				 * We prohibit the merging of unlike kinds i.e.
9676 				 * all pieces that are in the GP estimate can be
9677 				 * merged and all pieces that are not in a GP estimate
9678 				 * can be merged, but not disimilar pieces. Combine
9679 				 * this with taking the highest here and we should
9680 				 * be ok unless of course the client reneges. Then
9681 				 * all bets are off.
9682 				 */
9683 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9684 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9685 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9686 				/*
9687 				 * And we must keep the newest ack arrival time.
9688 				 */
9689 				if (next->r_ack_arrival <
9690 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9691 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9692 
9693 
9694 				/* We don't need to adjust rsm, it did not change */
9695 				/* Clear out the dup ack count of the remainder */
9696 				rsm->r_dupack = 0;
9697 				rsm->r_just_ret = 0;
9698 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9699 				/* Now lets make sure our fudge block is right */
9700 				nrsm->r_start = start;
9701 				/* Now lets update all the stats and such */
9702 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9703 				if (rack->app_limited_needs_set)
9704 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9705 				changed += (nrsm->r_end - nrsm->r_start);
9706 				/* You get a count for acking a whole segment or more */
9707 				if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9708 					rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9709 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9710 				if (nrsm->r_flags & RACK_SACK_PASSED) {
9711 					rack->r_ctl.rc_reorder_ts = cts;
9712 					if (rack->r_ctl.rc_reorder_ts == 0)
9713 						rack->r_ctl.rc_reorder_ts = 1;
9714 				}
9715 				/*
9716 				 * Now we want to go up from rsm (the
9717 				 * one left un-acked) to the next one
9718 				 * in the tmap. We do this so when
9719 				 * we walk backwards we include marking
9720 				 * sack-passed on rsm (The one passed in
9721 				 * is skipped since it is generally called
9722 				 * on something sacked before removing it
9723 				 * from the tmap).
9724 				 */
9725 				if (rsm->r_in_tmap) {
9726 					nrsm = TAILQ_NEXT(rsm, r_tnext);
9727 					/*
9728 					 * Now that we have the next
9729 					 * one walk backwards from there.
9730 					 */
9731 					if (nrsm && nrsm->r_in_tmap)
9732 						rack_log_sack_passed(tp, rack, nrsm);
9733 				}
9734 				/* Now are we done? */
9735 				if (SEQ_LT(end, next->r_end) ||
9736 				    (end == next->r_end)) {
9737 					/* Done with block */
9738 					goto out;
9739 				}
9740 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9741 				counter_u64_add(rack_sack_used_next_merge, 1);
9742 				/* Postion for the next block */
9743 				start = next->r_end;
9744 				rsm = tqhash_next(rack->r_ctl.tqh, next);
9745 				if (rsm == NULL)
9746 					goto out;
9747 			} else {
9748 				/**
9749 				 * We can't use any hookery here, so we
9750 				 * need to split the map. We enter like
9751 				 * so:
9752 				 *  rsm      |--------|
9753 				 *  sackblk       |----->
9754 				 * We will add the new block nrsm and
9755 				 * that will be the new portion, and then
9756 				 * fall through after reseting rsm. So we
9757 				 * split and look like this:
9758 				 *  rsm      |----|
9759 				 *  sackblk       |----->
9760 				 *  nrsm          |---|
9761 				 * We then fall through reseting
9762 				 * rsm to nrsm, so the next block
9763 				 * picks it up.
9764 				 */
9765 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9766 				if (nrsm == NULL) {
9767 					/*
9768 					 * failed XXXrrs what can we do but loose the sack
9769 					 * info?
9770 					 */
9771 					goto out;
9772 				}
9773 				counter_u64_add(rack_sack_splits, 1);
9774 				rack_clone_rsm(rack, nrsm, rsm, start);
9775 				moved++;
9776 				rsm->r_just_ret = 0;
9777 #ifndef INVARIANTS
9778 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9779 #else
9780 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9781 					panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
9782 					      nrsm, insret, rack, rsm);
9783 				}
9784 #endif
9785 				if (rsm->r_in_tmap) {
9786 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9787 					nrsm->r_in_tmap = 1;
9788 				}
9789 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9790 				rsm->r_flags &= (~RACK_HAS_FIN);
9791 				/* Position us to point to the new nrsm that starts the sack blk */
9792 				rsm = nrsm;
9793 			}
9794 		} else {
9795 			/* Already sacked this piece */
9796 			counter_u64_add(rack_sack_skipped_acked, 1);
9797 			moved++;
9798 			if (end == rsm->r_end) {
9799 				/* Done with block */
9800 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9801 				goto out;
9802 			} else if (SEQ_LT(end, rsm->r_end)) {
9803 				/* A partial sack to a already sacked block */
9804 				moved++;
9805 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9806 				goto out;
9807 			} else {
9808 				/*
9809 				 * The end goes beyond this guy
9810 				 * reposition the start to the
9811 				 * next block.
9812 				 */
9813 				start = rsm->r_end;
9814 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9815 				if (rsm == NULL)
9816 					goto out;
9817 			}
9818 		}
9819 	}
9820 	if (SEQ_GEQ(end, rsm->r_end)) {
9821 		/**
9822 		 * The end of this block is either beyond this guy or right
9823 		 * at this guy. I.e.:
9824 		 *  rsm ---                 |-----|
9825 		 *  end                     |-----|
9826 		 *  <or>
9827 		 *  end                     |---------|
9828 		 */
9829 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9830 			/*
9831 			 * Is it a TLP of interest?
9832 			 */
9833 			if ((rsm->r_flags & RACK_TLP) &&
9834 			    (rsm->r_rtr_cnt > 1)) {
9835 				/*
9836 				 * We are splitting a rxt TLP, check
9837 				 * if we need to save off the start/end
9838 				 */
9839 				if (rack->rc_last_tlp_acked_set &&
9840 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9841 					/*
9842 					 * We already turned this on since we are inside
9843 					 * the previous one was a partially sack now we
9844 					 * are getting another one (maybe all of it).
9845 					 */
9846 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9847 					/*
9848 					 * Lets make sure we have all of it though.
9849 					 */
9850 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9851 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9852 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9853 								     rack->r_ctl.last_tlp_acked_end);
9854 					}
9855 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9856 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9857 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9858 								     rack->r_ctl.last_tlp_acked_end);
9859 					}
9860 				} else {
9861 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9862 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9863 					rack->rc_last_tlp_past_cumack = 0;
9864 					rack->rc_last_tlp_acked_set = 1;
9865 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9866 				}
9867 			}
9868 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9869 			changed += (rsm->r_end - rsm->r_start);
9870 			/* You get a count for acking a whole segment or more */
9871 			if ((rsm->r_end - rsm->r_start) >= segsiz)
9872 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
9873 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9874 			if (rsm->r_in_tmap) /* should be true */
9875 				rack_log_sack_passed(tp, rack, rsm);
9876 			/* Is Reordering occuring? */
9877 			if (rsm->r_flags & RACK_SACK_PASSED) {
9878 				rsm->r_flags &= ~RACK_SACK_PASSED;
9879 				rack->r_ctl.rc_reorder_ts = cts;
9880 				if (rack->r_ctl.rc_reorder_ts == 0)
9881 					rack->r_ctl.rc_reorder_ts = 1;
9882 			}
9883 			if (rack->app_limited_needs_set)
9884 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9885 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9886 			rsm->r_flags |= RACK_ACKED;
9887 			if (rsm->r_in_tmap) {
9888 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9889 				rsm->r_in_tmap = 0;
9890 			}
9891 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9892 		} else {
9893 			counter_u64_add(rack_sack_skipped_acked, 1);
9894 			moved++;
9895 		}
9896 		if (end == rsm->r_end) {
9897 			/* This block only - done, setup for next */
9898 			goto out;
9899 		}
9900 		/*
9901 		 * There is more not coverend by this rsm move on
9902 		 * to the next block in the RB tree.
9903 		 */
9904 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9905 		start = rsm->r_end;
9906 		rsm = nrsm;
9907 		if (rsm == NULL)
9908 			goto out;
9909 		goto do_rest_ofb;
9910 	}
9911 	/**
9912 	 * The end of this sack block is smaller than
9913 	 * our rsm i.e.:
9914 	 *  rsm ---                 |-----|
9915 	 *  end                     |--|
9916 	 */
9917 	if ((rsm->r_flags & RACK_ACKED) == 0) {
9918 		/*
9919 		 * Is it a TLP of interest?
9920 		 */
9921 		if ((rsm->r_flags & RACK_TLP) &&
9922 		    (rsm->r_rtr_cnt > 1)) {
9923 			/*
9924 			 * We are splitting a rxt TLP, check
9925 			 * if we need to save off the start/end
9926 			 */
9927 			if (rack->rc_last_tlp_acked_set &&
9928 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9929 				/*
9930 				 * We already turned this on since we are inside
9931 				 * the previous one was a partially sack now we
9932 				 * are getting another one (maybe all of it).
9933 				 */
9934 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9935 				/*
9936 				 * Lets make sure we have all of it though.
9937 				 */
9938 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9939 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9940 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9941 							     rack->r_ctl.last_tlp_acked_end);
9942 				}
9943 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9944 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9945 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9946 							     rack->r_ctl.last_tlp_acked_end);
9947 				}
9948 			} else {
9949 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9950 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9951 				rack->rc_last_tlp_past_cumack = 0;
9952 				rack->rc_last_tlp_acked_set = 1;
9953 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9954 			}
9955 		}
9956 		/*
9957 		 * Hookery can only be used if the two entries
9958 		 * are in the same bucket and neither one of
9959 		 * them staddle the bucket line.
9960 		 */
9961 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9962 		if (prev &&
9963 		    (rsm->bindex == prev->bindex) &&
9964 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9965 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
9966 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
9967 		    (prev->r_flags & RACK_IN_GP_WIN))
9968 			can_use_hookery = 1;
9969 		else if (prev &&
9970 			 (rsm->bindex == prev->bindex) &&
9971 			 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9972 			 ((prev->r_flags & RACK_STRADDLE) == 0) &&
9973 			 ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9974 			 ((prev->r_flags & RACK_IN_GP_WIN) == 0))
9975 			can_use_hookery = 1;
9976 		else
9977 			can_use_hookery = 0;
9978 
9979 		if (prev && can_use_hookery &&
9980 		    (prev->r_flags & RACK_ACKED)) {
9981 			/**
9982 			 * Goal, we want the right remainder of rsm to shrink
9983 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
9984 			 * We want to expand prev to go all the way
9985 			 * to prev->r_end <- end.
9986 			 * so in the tree we have before:
9987 			 *   prev     |--------|         (acked)
9988 			 *   rsm               |-------| (non-acked)
9989 			 *   sackblk           |-|
9990 			 * We churn it so we end up with
9991 			 *   prev     |----------|       (acked)
9992 			 *   rsm                 |-----| (non-acked)
9993 			 *   nrsm              |-| (temporary)
9994 			 *
9995 			 * Note if either prev/rsm is a TLP we don't
9996 			 * do this.
9997 			 */
9998 			noextra++;
9999 			nrsm = &stack_map;
10000 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10001 			prev->r_end = end;
10002 			rsm->r_start = end;
10003 			rsm->r_flags |= RACK_SHUFFLED;
10004 			prev->r_flags |= RACK_SHUFFLED;
10005 			/* Now adjust nrsm (stack copy) to be
10006 			 * the one that is the small
10007 			 * piece that was "sacked".
10008 			 */
10009 			nrsm->r_end = end;
10010 			rsm->r_dupack = 0;
10011 			/*
10012 			 * Which timestamp do we keep? It is rather
10013 			 * important in GP measurements to have the
10014 			 * accurate end of the send window.
10015 			 *
10016 			 * We keep the largest value, which is the newest
10017 			 * send. We do this in case a segment that is
10018 			 * joined together and not part of a GP estimate
10019 			 * later gets expanded into the GP estimate.
10020 			 *
10021 			 * We prohibit the merging of unlike kinds i.e.
10022 			 * all pieces that are in the GP estimate can be
10023 			 * merged and all pieces that are not in a GP estimate
10024 			 * can be merged, but not disimilar pieces. Combine
10025 			 * this with taking the highest here and we should
10026 			 * be ok unless of course the client reneges. Then
10027 			 * all bets are off.
10028 			 */
10029 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
10030 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
10031 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
10032 			}
10033 			/*
10034 			 * And we must keep the newest ack arrival time.
10035 			 */
10036 
10037 			if(prev->r_ack_arrival <
10038 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10039 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10040 
10041 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10042 			/*
10043 			 * Now that the rsm has had its start moved forward
10044 			 * lets go ahead and get its new place in the world.
10045 			 */
10046 			rack_setup_offset_for_rsm(rack, prev, rsm);
10047 			/*
10048 			 * Now nrsm is our new little piece
10049 			 * that is acked (which was merged
10050 			 * to prev). Update the rtt and changed
10051 			 * based on that. Also check for reordering.
10052 			 */
10053 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10054 			if (rack->app_limited_needs_set)
10055 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10056 			changed += (nrsm->r_end - nrsm->r_start);
10057 			/* You get a count for acking a whole segment or more */
10058 			if ((nrsm->r_end - nrsm->r_start) >= segsiz)
10059 				rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
10060 
10061 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10062 			if (nrsm->r_flags & RACK_SACK_PASSED) {
10063 				rack->r_ctl.rc_reorder_ts = cts;
10064 				if (rack->r_ctl.rc_reorder_ts == 0)
10065 					rack->r_ctl.rc_reorder_ts = 1;
10066 			}
10067 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
10068 			rsm = prev;
10069 			counter_u64_add(rack_sack_used_prev_merge, 1);
10070 		} else {
10071 			/**
10072 			 * This is the case where our previous
10073 			 * block is not acked either, so we must
10074 			 * split the block in two.
10075 			 */
10076 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10077 			if (nrsm == NULL) {
10078 				/* failed rrs what can we do but loose the sack info? */
10079 				goto out;
10080 			}
10081 			if ((rsm->r_flags & RACK_TLP) &&
10082 			    (rsm->r_rtr_cnt > 1)) {
10083 				/*
10084 				 * We are splitting a rxt TLP, check
10085 				 * if we need to save off the start/end
10086 				 */
10087 				if (rack->rc_last_tlp_acked_set &&
10088 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10089 					/*
10090 					 * We already turned this on since this block is inside
10091 					 * the previous one was a partially sack now we
10092 					 * are getting another one (maybe all of it).
10093 					 */
10094 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10095 					/*
10096 					 * Lets make sure we have all of it though.
10097 					 */
10098 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10099 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10100 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10101 								     rack->r_ctl.last_tlp_acked_end);
10102 					}
10103 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10104 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10105 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10106 								     rack->r_ctl.last_tlp_acked_end);
10107 					}
10108 				} else {
10109 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10110 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10111 					rack->rc_last_tlp_acked_set = 1;
10112 					rack->rc_last_tlp_past_cumack = 0;
10113 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10114 				}
10115 			}
10116 			/**
10117 			 * In this case nrsm becomes
10118 			 * nrsm->r_start = end;
10119 			 * nrsm->r_end = rsm->r_end;
10120 			 * which is un-acked.
10121 			 * <and>
10122 			 * rsm->r_end = nrsm->r_start;
10123 			 * i.e. the remaining un-acked
10124 			 * piece is left on the left
10125 			 * hand side.
10126 			 *
10127 			 * So we start like this
10128 			 * rsm      |----------| (not acked)
10129 			 * sackblk  |---|
10130 			 * build it so we have
10131 			 * rsm      |---|         (acked)
10132 			 * nrsm         |------|  (not acked)
10133 			 */
10134 			counter_u64_add(rack_sack_splits, 1);
10135 			rack_clone_rsm(rack, nrsm, rsm, end);
10136 			moved++;
10137 			rsm->r_flags &= (~RACK_HAS_FIN);
10138 			rsm->r_just_ret = 0;
10139 #ifndef INVARIANTS
10140 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10141 #else
10142 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10143 				panic("Insert in rb tree of %p fails ret:% rack:%p rsm:%p",
10144 				      nrsm, insret, rack, rsm);
10145 			}
10146 #endif
10147 			if (rsm->r_in_tmap) {
10148 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10149 				nrsm->r_in_tmap = 1;
10150 			}
10151 			nrsm->r_dupack = 0;
10152 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10153 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10154 			changed += (rsm->r_end - rsm->r_start);
10155 			/* You get a count for acking a whole segment or more */
10156 			if ((rsm->r_end - rsm->r_start) >= segsiz)
10157 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10158 
10159 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10160 			if (rsm->r_in_tmap) /* should be true */
10161 				rack_log_sack_passed(tp, rack, rsm);
10162 			/* Is Reordering occuring? */
10163 			if (rsm->r_flags & RACK_SACK_PASSED) {
10164 				rsm->r_flags &= ~RACK_SACK_PASSED;
10165 				rack->r_ctl.rc_reorder_ts = cts;
10166 				if (rack->r_ctl.rc_reorder_ts == 0)
10167 					rack->r_ctl.rc_reorder_ts = 1;
10168 			}
10169 			if (rack->app_limited_needs_set)
10170 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10171 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10172 			rsm->r_flags |= RACK_ACKED;
10173 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10174 			if (rsm->r_in_tmap) {
10175 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10176 				rsm->r_in_tmap = 0;
10177 			}
10178 		}
10179 	} else if (start != end){
10180 		/*
10181 		 * The block was already acked.
10182 		 */
10183 		counter_u64_add(rack_sack_skipped_acked, 1);
10184 		moved++;
10185 	}
10186 out:
10187 	if (rsm &&
10188 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10189 	    (rsm->r_flags & RACK_ACKED)) {
10190 		/*
10191 		 * Now can we merge where we worked
10192 		 * with either the previous or
10193 		 * next block?
10194 		 */
10195 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10196 		while (next) {
10197 			if (next->r_flags & RACK_TLP)
10198 				break;
10199 			/* Only allow merges between ones in or out of GP window */
10200 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10201 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10202 				break;
10203 			}
10204 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10205 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10206 				break;
10207 			}
10208 			if (rsm->bindex != next->bindex)
10209 				break;
10210 			if (rsm->r_flags & RACK_STRADDLE)
10211 				break;
10212 			if (next->r_flags & RACK_STRADDLE)
10213 				break;
10214 			if (next->r_flags & RACK_ACKED) {
10215 				/* yep this and next can be merged */
10216 				rsm = rack_merge_rsm(rack, rsm, next);
10217 				noextra++;
10218 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10219 			} else
10220 				break;
10221 		}
10222 		/* Now what about the previous? */
10223 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10224 		while (prev) {
10225 			if (prev->r_flags & RACK_TLP)
10226 				break;
10227 			/* Only allow merges between ones in or out of GP window */
10228 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10229 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10230 				break;
10231 			}
10232 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10233 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10234 				break;
10235 			}
10236 			if (rsm->bindex != prev->bindex)
10237 				break;
10238 			if (rsm->r_flags & RACK_STRADDLE)
10239 				break;
10240 			if (prev->r_flags & RACK_STRADDLE)
10241 				break;
10242 			if (prev->r_flags & RACK_ACKED) {
10243 				/* yep the previous and this can be merged */
10244 				rsm = rack_merge_rsm(rack, prev, rsm);
10245 				noextra++;
10246 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10247 			} else
10248 				break;
10249 		}
10250 	}
10251 	if (used_ref == 0) {
10252 		counter_u64_add(rack_sack_proc_all, 1);
10253 	} else {
10254 		counter_u64_add(rack_sack_proc_short, 1);
10255 	}
10256 	/* Save off the next one for quick reference. */
10257 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10258 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10259 	/* Pass back the moved. */
10260 	*moved_two = moved;
10261 	*no_extra = noextra;
10262 	return (changed);
10263 }
10264 
10265 static void inline
10266 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10267 {
10268 	struct rack_sendmap *tmap;
10269 
10270 	tmap = NULL;
10271 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10272 		/* Its no longer sacked, mark it so */
10273 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10274 #ifdef INVARIANTS
10275 		if (rsm->r_in_tmap) {
10276 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10277 			      rack, rsm, rsm->r_flags);
10278 		}
10279 #endif
10280 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10281 		/* Rebuild it into our tmap */
10282 		if (tmap == NULL) {
10283 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10284 			tmap = rsm;
10285 		} else {
10286 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10287 			tmap = rsm;
10288 		}
10289 		tmap->r_in_tmap = 1;
10290 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10291 	}
10292 	/*
10293 	 * Now lets possibly clear the sack filter so we start
10294 	 * recognizing sacks that cover this area.
10295 	 */
10296 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10297 
10298 }
10299 
10300 static void
10301 rack_do_decay(struct tcp_rack *rack)
10302 {
10303 	struct timeval res;
10304 
10305 #define	timersub(tvp, uvp, vvp)						\
10306 	do {								\
10307 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
10308 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
10309 		if ((vvp)->tv_usec < 0) {				\
10310 			(vvp)->tv_sec--;				\
10311 			(vvp)->tv_usec += 1000000;			\
10312 		}							\
10313 	} while (0)
10314 
10315 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
10316 #undef timersub
10317 
10318 	rack->r_ctl.input_pkt++;
10319 	if ((rack->rc_in_persist) ||
10320 	    (res.tv_sec >= 1) ||
10321 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
10322 		/*
10323 		 * Check for decay of non-SAD,
10324 		 * we want all SAD detection metrics to
10325 		 * decay 1/4 per second (or more) passed.
10326 		 * Current default is 800 so it decays
10327 		 * 80% every second.
10328 		 */
10329 #ifdef TCP_SAD_DETECTION
10330 		uint32_t pkt_delta;
10331 
10332 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
10333 #endif
10334 		/* Update our saved tracking values */
10335 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
10336 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10337 		/* Now do we escape without decay? */
10338 #ifdef TCP_SAD_DETECTION
10339 		if (rack->rc_in_persist ||
10340 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
10341 		    (pkt_delta < tcp_sad_low_pps)){
10342 			/*
10343 			 * We don't decay idle connections
10344 			 * or ones that have a low input pps.
10345 			 */
10346 			return;
10347 		}
10348 		/* Decay the counters */
10349 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
10350 							tcp_sad_decay_val);
10351 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
10352 							 tcp_sad_decay_val);
10353 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
10354 							       tcp_sad_decay_val);
10355 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
10356 								tcp_sad_decay_val);
10357 #endif
10358 	}
10359 }
10360 
10361 static void inline
10362 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10363 {
10364 	/*
10365 	 * We look at advancing the end send time for our GP
10366 	 * measurement tracking only as the cumulative acknowledgment
10367 	 * moves forward. You might wonder about this, why not
10368 	 * at every transmission or retransmission within the
10369 	 * GP window update the rc_gp_cumack_ts? Well its rather
10370 	 * nuanced but basically the GP window *may* expand (as
10371 	 * it does below) or worse and harder to track it may shrink.
10372 	 *
10373 	 * This last makes it impossible to track at the time of
10374 	 * the send, since you may set forward your rc_gp_cumack_ts
10375 	 * when you send, because that send *is* in your currently
10376 	 * "guessed" window, but then it shrinks. Now which was
10377 	 * the send time of the last bytes in the window, by the
10378 	 * time you ask that question that part of the sendmap
10379 	 * is freed. So you don't know and you will have too
10380 	 * long of send window. Instead by updating the time
10381 	 * marker only when the cumack advances this assures us
10382 	 * that we will have only the sends in the window of our
10383 	 * GP measurement.
10384 	 *
10385 	 * Another complication from this is the
10386 	 * merging of sendmap entries. During SACK processing this
10387 	 * can happen to conserve the sendmap size. That breaks
10388 	 * everything down in tracking the send window of the GP
10389 	 * estimate. So to prevent that and keep it working with
10390 	 * a tiny bit more limited merging, we only allow like
10391 	 * types to be merged. I.e. if two sends are in the GP window
10392 	 * then its ok to merge them together. If two sends are not
10393 	 * in the GP window its ok to merge them together too. Though
10394 	 * one send in and one send out cannot be merged. We combine
10395 	 * this with never allowing the shrinking of the GP window when
10396 	 * we are in recovery so that we can properly calculate the
10397 	 * sending times.
10398 	 *
10399 	 * This all of course seems complicated, because it is.. :)
10400 	 *
10401 	 * The cum-ack is being advanced upon the sendmap.
10402 	 * If we are not doing a GP estimate don't
10403 	 * proceed.
10404 	 */
10405 	uint64_t ts;
10406 
10407 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10408 		return;
10409 	/*
10410 	 * If this sendmap entry is going
10411 	 * beyond the measurement window we had picked,
10412 	 * expand the measurement window by that much.
10413 	 */
10414 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10415 		tp->gput_ack = rsm->r_end;
10416 	}
10417 	/*
10418 	 * If we have not setup a ack, then we
10419 	 * have no idea if the newly acked pieces
10420 	 * will be "in our seq measurement range". If
10421 	 * it is when we clear the app_limited_needs_set
10422 	 * flag the timestamp will be updated.
10423 	 */
10424 	if (rack->app_limited_needs_set)
10425 		return;
10426 	/*
10427 	 * Finally, we grab out the latest timestamp
10428 	 * that this packet was sent and then see
10429 	 * if:
10430 	 *  a) The packet touches are newly defined GP range.
10431 	 *  b) The time is greater than (newer) than the
10432 	 *     one we currently have. If so we update
10433 	 *     our sending end time window.
10434 	 *
10435 	 * Note we *do not* do this at send time. The reason
10436 	 * is that if you do you *may* pick up a newer timestamp
10437 	 * for a range you are not going to measure. We project
10438 	 * out how far and then sometimes modify that to be
10439 	 * smaller. If that occurs then you will have a send
10440 	 * that does not belong to the range included.
10441 	 */
10442 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10443 	    rack->r_ctl.rc_gp_cumack_ts)
10444 		return;
10445 	if (rack_in_gp_window(tp, rsm)) {
10446 		rack->r_ctl.rc_gp_cumack_ts = ts;
10447 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10448 			       __LINE__, from, rsm);
10449 	}
10450 }
10451 
10452 static void
10453 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)
10454 {
10455 	struct rack_sendmap *rsm;
10456 	/*
10457 	 * The ACK point is advancing to th_ack, we must drop off
10458 	 * the packets in the rack log and calculate any eligble
10459 	 * RTT's.
10460 	 */
10461 
10462 	rack->r_wanted_output = 1;
10463 	if (SEQ_GT(th_ack, tp->snd_una))
10464 		rack->r_ctl.last_cumack_advance = acktime;
10465 
10466 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10467 	if ((rack->rc_last_tlp_acked_set == 1)&&
10468 	    (rack->rc_last_tlp_past_cumack == 1) &&
10469 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10470 		/*
10471 		 * We have reached the point where our last rack
10472 		 * tlp retransmit sequence is ahead of the cum-ack.
10473 		 * This can only happen when the cum-ack moves all
10474 		 * the way around (its been a full 2^^31+1 bytes
10475 		 * or more since we sent a retransmitted TLP). Lets
10476 		 * turn off the valid flag since its not really valid.
10477 		 *
10478 		 * Note since sack's also turn on this event we have
10479 		 * a complication, we have to wait to age it out until
10480 		 * the cum-ack is by the TLP before checking which is
10481 		 * what the next else clause does.
10482 		 */
10483 		rack_log_dsack_event(rack, 9, __LINE__,
10484 				     rack->r_ctl.last_tlp_acked_start,
10485 				     rack->r_ctl.last_tlp_acked_end);
10486 		rack->rc_last_tlp_acked_set = 0;
10487 		rack->rc_last_tlp_past_cumack = 0;
10488 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10489 		   (rack->rc_last_tlp_past_cumack == 0) &&
10490 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10491 		/*
10492 		 * It is safe to start aging TLP's out.
10493 		 */
10494 		rack->rc_last_tlp_past_cumack = 1;
10495 	}
10496 	/* We do the same for the tlp send seq as well */
10497 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10498 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10499 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10500 		rack_log_dsack_event(rack, 9, __LINE__,
10501 				     rack->r_ctl.last_sent_tlp_seq,
10502 				     (rack->r_ctl.last_sent_tlp_seq +
10503 				      rack->r_ctl.last_sent_tlp_len));
10504 		rack->rc_last_sent_tlp_seq_valid = 0;
10505 		rack->rc_last_sent_tlp_past_cumack = 0;
10506 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10507 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10508 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10509 		/*
10510 		 * It is safe to start aging TLP's send.
10511 		 */
10512 		rack->rc_last_sent_tlp_past_cumack = 1;
10513 	}
10514 more:
10515 	rsm = tqhash_min(rack->r_ctl.tqh);
10516 	if (rsm == NULL) {
10517 		if ((th_ack - 1) == tp->iss) {
10518 			/*
10519 			 * For the SYN incoming case we will not
10520 			 * have called tcp_output for the sending of
10521 			 * the SYN, so there will be no map. All
10522 			 * other cases should probably be a panic.
10523 			 */
10524 			return;
10525 		}
10526 		if (tp->t_flags & TF_SENTFIN) {
10527 			/* if we sent a FIN we often will not have map */
10528 			return;
10529 		}
10530 #ifdef INVARIANTS
10531 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
10532 		      tp,
10533 		      tp->t_state, th_ack, rack,
10534 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
10535 #endif
10536 		return;
10537 	}
10538 	if (SEQ_LT(th_ack, rsm->r_start)) {
10539 		/* Huh map is missing this */
10540 #ifdef INVARIANTS
10541 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10542 		       rsm->r_start,
10543 		       th_ack, tp->t_state, rack->r_state);
10544 #endif
10545 		return;
10546 	}
10547 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10548 
10549 	/* Now was it a retransmitted TLP? */
10550 	if ((rsm->r_flags & RACK_TLP) &&
10551 	    (rsm->r_rtr_cnt > 1)) {
10552 		/*
10553 		 * Yes, this rsm was a TLP and retransmitted, remember that
10554 		 * since if a DSACK comes back on this we don't want
10555 		 * to think of it as a reordered segment. This may
10556 		 * get updated again with possibly even other TLPs
10557 		 * in flight, but thats ok. Only when we don't send
10558 		 * a retransmitted TLP for 1/2 the sequences space
10559 		 * will it get turned off (above).
10560 		 */
10561 		if (rack->rc_last_tlp_acked_set &&
10562 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10563 			/*
10564 			 * We already turned this on since the end matches,
10565 			 * the previous one was a partially ack now we
10566 			 * are getting another one (maybe all of it).
10567 			 */
10568 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10569 			/*
10570 			 * Lets make sure we have all of it though.
10571 			 */
10572 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10573 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10574 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10575 						     rack->r_ctl.last_tlp_acked_end);
10576 			}
10577 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10578 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10579 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10580 						     rack->r_ctl.last_tlp_acked_end);
10581 			}
10582 		} else {
10583 			rack->rc_last_tlp_past_cumack = 1;
10584 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10585 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10586 			rack->rc_last_tlp_acked_set = 1;
10587 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10588 		}
10589 	}
10590 	/* Now do we consume the whole thing? */
10591 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10592 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
10593 		/* Its all consumed. */
10594 		uint32_t left;
10595 		uint8_t newly_acked;
10596 
10597 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10598 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10599 		rsm->r_rtr_bytes = 0;
10600 		/*
10601 		 * Record the time of highest cumack sent if its in our measurement
10602 		 * window and possibly bump out the end.
10603 		 */
10604 		rack_rsm_sender_update(rack, tp, rsm, 4);
10605 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10606 		if (rsm->r_in_tmap) {
10607 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10608 			rsm->r_in_tmap = 0;
10609 		}
10610 		newly_acked = 1;
10611 		if (rsm->r_flags & RACK_ACKED) {
10612 			/*
10613 			 * It was acked on the scoreboard -- remove
10614 			 * it from total
10615 			 */
10616 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10617 			newly_acked = 0;
10618 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
10619 			/*
10620 			 * There are segments ACKED on the
10621 			 * scoreboard further up. We are seeing
10622 			 * reordering.
10623 			 */
10624 			rsm->r_flags &= ~RACK_SACK_PASSED;
10625 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10626 			rsm->r_flags |= RACK_ACKED;
10627 			rack->r_ctl.rc_reorder_ts = cts;
10628 			if (rack->r_ctl.rc_reorder_ts == 0)
10629 				rack->r_ctl.rc_reorder_ts = 1;
10630 			if (rack->r_ent_rec_ns) {
10631 				/*
10632 				 * We have sent no more, and we saw an sack
10633 				 * then ack arrive.
10634 				 */
10635 				rack->r_might_revert = 1;
10636 			}
10637 		}
10638 		if ((rsm->r_flags & RACK_TO_REXT) &&
10639 		    (tp->t_flags & TF_RCVD_TSTMP) &&
10640 		    (to->to_flags & TOF_TS) &&
10641 		    (to->to_tsecr != 0) &&
10642 		    (tp->t_flags & TF_PREVVALID)) {
10643 			/*
10644 			 * We can use the timestamp to see
10645 			 * if this retransmission was from the
10646 			 * first transmit. If so we made a mistake.
10647 			 */
10648 			tp->t_flags &= ~TF_PREVVALID;
10649 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10650 				/* The first transmit is what this ack is for */
10651 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10652 			}
10653 		}
10654 		left = th_ack - rsm->r_end;
10655 		if (rack->app_limited_needs_set && newly_acked)
10656 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10657 		/* Free back to zone */
10658 		rack_free(rack, rsm);
10659 		if (left) {
10660 			goto more;
10661 		}
10662 		/* Check for reneging */
10663 		rsm = tqhash_min(rack->r_ctl.tqh);
10664 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10665 			/*
10666 			 * The peer has moved snd_una up to
10667 			 * the edge of this send, i.e. one
10668 			 * that it had previously acked. The only
10669 			 * way that can be true if the peer threw
10670 			 * away data (space issues) that it had
10671 			 * previously sacked (else it would have
10672 			 * given us snd_una up to (rsm->r_end).
10673 			 * We need to undo the acked markings here.
10674 			 *
10675 			 * Note we have to look to make sure th_ack is
10676 			 * our rsm->r_start in case we get an old ack
10677 			 * where th_ack is behind snd_una.
10678 			 */
10679 			rack_peer_reneges(rack, rsm, th_ack);
10680 		}
10681 		return;
10682 	}
10683 	if (rsm->r_flags & RACK_ACKED) {
10684 		/*
10685 		 * It was acked on the scoreboard -- remove it from
10686 		 * total for the part being cum-acked.
10687 		 */
10688 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10689 	}
10690 	/*
10691 	 * Clear the dup ack count for
10692 	 * the piece that remains.
10693 	 */
10694 	rsm->r_dupack = 0;
10695 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10696 	if (rsm->r_rtr_bytes) {
10697 		/*
10698 		 * It was retransmitted adjust the
10699 		 * sack holes for what was acked.
10700 		 */
10701 		int ack_am;
10702 
10703 		ack_am = (th_ack - rsm->r_start);
10704 		if (ack_am >= rsm->r_rtr_bytes) {
10705 			rack->r_ctl.rc_holes_rxt -= ack_am;
10706 			rsm->r_rtr_bytes -= ack_am;
10707 		}
10708 	}
10709 	/*
10710 	 * Update where the piece starts and record
10711 	 * the time of send of highest cumack sent if
10712 	 * its in our GP range.
10713 	 */
10714 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10715 	/* Now we need to move our offset forward too */
10716 	if (rsm->m &&
10717 	    ((rsm->orig_m_len != rsm->m->m_len) ||
10718 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10719 		/* Fix up the orig_m_len and possibly the mbuf offset */
10720 		rack_adjust_orig_mlen(rsm);
10721 	}
10722 	rsm->soff += (th_ack - rsm->r_start);
10723 	rack_rsm_sender_update(rack, tp, rsm, 5);
10724 	/* The trim will move th_ack into r_start for us */
10725 	tqhash_trim(rack->r_ctl.tqh, th_ack);
10726 	/* Now do we need to move the mbuf fwd too? */
10727 	{
10728 		struct mbuf *m;
10729 		uint32_t soff;
10730 
10731 		m = rsm->m;
10732 		soff = rsm->soff;
10733 		if (m) {
10734 			while (soff >= m->m_len) {
10735 				soff -= m->m_len;
10736 				KASSERT((m->m_next != NULL),
10737 					(" rsm:%p  off:%u soff:%u m:%p",
10738 					 rsm, rsm->soff, soff, m));
10739 				m = m->m_next;
10740 				if (m == NULL) {
10741 					/*
10742 					 * This is a fall-back that prevents a panic. In reality
10743 					 * we should be able to walk the mbuf's and find our place.
10744 					 * At this point snd_una has not been updated with the sbcut() yet
10745 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
10746 					 * should work fine. This is undesirable since we will take cache
10747 					 * hits to access the socket buffer. And even more puzzling is that
10748 					 * it happens occasionally. It should not :(
10749 					 */
10750 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10751 						      (rsm->r_start - tp->snd_una),
10752 						      &soff);
10753 					break;
10754 				}
10755 			}
10756 			/*
10757 			 * Now save in our updated values.
10758 			 */
10759 			rsm->m = m;
10760 			rsm->soff = soff;
10761 			rsm->orig_m_len = rsm->m->m_len;
10762 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10763 		}
10764 	}
10765 	if (rack->app_limited_needs_set &&
10766 	    SEQ_GEQ(th_ack, tp->gput_seq))
10767 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10768 }
10769 
10770 static void
10771 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10772 {
10773 	struct rack_sendmap *rsm;
10774 	int sack_pass_fnd = 0;
10775 
10776 	if (rack->r_might_revert) {
10777 		/*
10778 		 * Ok we have reordering, have not sent anything, we
10779 		 * might want to revert the congestion state if nothing
10780 		 * further has SACK_PASSED on it. Lets check.
10781 		 *
10782 		 * We also get here when we have DSACKs come in for
10783 		 * all the data that we FR'd. Note that a rxt or tlp
10784 		 * timer clears this from happening.
10785 		 */
10786 
10787 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10788 			if (rsm->r_flags & RACK_SACK_PASSED) {
10789 				sack_pass_fnd = 1;
10790 				break;
10791 			}
10792 		}
10793 		if (sack_pass_fnd == 0) {
10794 			/*
10795 			 * We went into recovery
10796 			 * incorrectly due to reordering!
10797 			 */
10798 			int orig_cwnd;
10799 
10800 			rack->r_ent_rec_ns = 0;
10801 			orig_cwnd = tp->snd_cwnd;
10802 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10803 			tp->snd_recover = tp->snd_una;
10804 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10805 			EXIT_RECOVERY(tp->t_flags);
10806 		}
10807 		rack->r_might_revert = 0;
10808 	}
10809 }
10810 
10811 #ifdef TCP_SAD_DETECTION
10812 
10813 static void
10814 rack_merge_out_sacks(struct tcp_rack *rack)
10815 {
10816 	struct rack_sendmap *cur, *next, *rsm, *trsm = NULL;
10817 
10818 	cur = tqhash_min(rack->r_ctl.tqh);
10819 	while(cur) {
10820 		next = tqhash_next(rack->r_ctl.tqh, cur);
10821 		/*
10822 		 * The idea is to go through all and merge back
10823 		 * together the pieces sent together,
10824 		 */
10825 		if ((next != NULL) &&
10826 		    (cur->r_tim_lastsent[0] == next->r_tim_lastsent[0])) {
10827 			rack_merge_rsm(rack, cur, next);
10828 		} else {
10829 			cur = next;
10830 		}
10831 	}
10832 	/*
10833 	 * now treat it like a rxt event, everything is outstanding
10834 	 * and sent nothing acvked and dupacks are all zero. If this
10835 	 * is not an attacker it will have to dupack its way through
10836 	 * it all.
10837 	 */
10838 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
10839 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
10840 		rsm->r_dupack = 0;
10841 		/* We must re-add it back to the tlist */
10842 		if (trsm == NULL) {
10843 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10844 		} else {
10845 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
10846 		}
10847 		rsm->r_in_tmap = 1;
10848 		trsm = rsm;
10849 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
10850 	}
10851 	sack_filter_clear(&rack->r_ctl.rack_sf, rack->rc_tp->snd_una);
10852 }
10853 
10854 static void
10855 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
10856 {
10857 	int do_detection = 0;
10858 
10859 	if (rack->sack_attack_disable || rack->rc_suspicious) {
10860 		/*
10861 		 * If we have been disabled we must detect
10862 		 * to possibly reverse it. Or if the guy has
10863 		 * sent in suspicious sacks we want to do detection too.
10864 		 */
10865 		do_detection = 1;
10866 
10867 	} else if  ((rack->do_detection || tcp_force_detection) &&
10868 		    (tcp_sack_to_ack_thresh > 0) &&
10869 		    (tcp_sack_to_move_thresh > 0) &&
10870 		    (rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum)) {
10871 		/*
10872 		 * We only detect here if:
10873 		 * 1) System wide forcing is on <or> do_detection is on
10874 		 *   <and>
10875 		 * 2) We have thresholds for move and ack (set one to 0 and we are off)
10876 		 *   <and>
10877 		 * 3) We have maps allocated larger than our min (500).
10878 		 */
10879 		do_detection = 1;
10880 	}
10881 	if (do_detection > 0) {
10882 		/*
10883 		 * We have thresholds set to find
10884 		 * possible attackers and disable sack.
10885 		 * Check them.
10886 		 */
10887 		uint64_t ackratio, moveratio, movetotal;
10888 
10889 		/* Log detecting */
10890 		rack_log_sad(rack, 1);
10891 		/* Do we establish a ack ratio */
10892 		if ((rack->r_ctl.sack_count > tcp_map_minimum)  ||
10893 		    (rack->rc_suspicious == 1) ||
10894 		    (rack->sack_attack_disable > 0)) {
10895 			ackratio = (uint64_t)(rack->r_ctl.sack_count);
10896 			ackratio *= (uint64_t)(1000);
10897 			if (rack->r_ctl.ack_count)
10898 				ackratio /= (uint64_t)(rack->r_ctl.ack_count);
10899 			else {
10900 				/* We can hit this due to ack totals degregation (via small sacks) */
10901 				ackratio = 1000;
10902 			}
10903 		} else {
10904 			/*
10905 			 * No ack ratio needed if we have not
10906 			 * seen more sacks then the number of map entries.
10907 			 * The exception to that is if we have disabled sack then
10908 			 * we need to find a ratio.
10909 			 */
10910 			ackratio = 0;
10911 		}
10912 
10913 		if ((rack->sack_attack_disable == 0) &&
10914 		    (ackratio > rack_highest_sack_thresh_seen))
10915 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
10916 		/* Do we establish a move ratio? */
10917 		if ((rack->r_ctl.sack_moved_extra > tcp_map_minimum) ||
10918 		    (rack->rc_suspicious == 1) ||
10919 		    (rack->sack_attack_disable > 0)) {
10920 			/*
10921 			 * We need to have more sack moves than maps
10922 			 * allocated to have a move ratio considered.
10923 			 */
10924 			movetotal = rack->r_ctl.sack_moved_extra;
10925 			movetotal += rack->r_ctl.sack_noextra_move;
10926 			moveratio = rack->r_ctl.sack_moved_extra;
10927 			moveratio *= (uint64_t)1000;
10928 			if (movetotal)
10929 				moveratio /= movetotal;
10930 			else {
10931 				/* No moves, thats pretty good */
10932 				moveratio = 0;
10933 			}
10934 		} else {
10935 			/*
10936 			 * Not enough moves have occured to consider
10937 			 * if we are out of whack in that ratio.
10938 			 * The exception to that is if we have disabled sack then
10939 			 * we need to find a ratio.
10940 			 */
10941 			moveratio = 0;
10942 		}
10943 		if ((rack->sack_attack_disable == 0) &&
10944 		    (moveratio > rack_highest_move_thresh_seen))
10945 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
10946 		/* Now the tests */
10947 		if (rack->sack_attack_disable == 0) {
10948 			/* Not disabled, do we need to disable? */
10949 			if ((ackratio > tcp_sack_to_ack_thresh) &&
10950 			    (moveratio > tcp_sack_to_move_thresh)) {
10951 				/* Disable sack processing */
10952 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10953 				rack->sack_attack_disable = 1;
10954 				/* set it so we have the built in delay */
10955 				rack->r_ctl.ack_during_sd = 1;
10956 				if (rack_merge_out_sacks_on_attack)
10957 					rack_merge_out_sacks(rack);
10958 				counter_u64_add(rack_sack_attacks_detected, 1);
10959 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10960 				/* Clamp the cwnd at flight size */
10961 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
10962 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10963 				rack_log_sad(rack, 2);
10964 			}
10965 		} else {
10966 			/* We are sack-disabled check for false positives */
10967 			if ((ackratio <= tcp_restoral_thresh) ||
10968 			    ((rack_merge_out_sacks_on_attack == 0) &&
10969 			     (rack->rc_suspicious == 0) &&
10970 			     (rack->r_ctl.rc_num_maps_alloced <= (tcp_map_minimum/2)))) {
10971 				rack->sack_attack_disable = 0;
10972 				rack_log_sad(rack, 3);
10973 				/* Restart counting */
10974 				rack->r_ctl.sack_count = 0;
10975 				rack->r_ctl.sack_moved_extra = 0;
10976 				rack->r_ctl.sack_noextra_move = 1;
10977 				rack->rc_suspicious = 0;
10978 				rack->r_ctl.ack_count = max(1,
10979 							    (bytes_this_ack / segsiz));
10980 
10981 				counter_u64_add(rack_sack_attacks_reversed, 1);
10982 				/* Restore the cwnd */
10983 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
10984 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
10985 			}
10986 		}
10987 	}
10988 }
10989 #endif
10990 
10991 static int
10992 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10993 {
10994 
10995 	uint32_t am, l_end;
10996 	int was_tlp = 0;
10997 
10998 	if (SEQ_GT(end, start))
10999 		am = end - start;
11000 	else
11001 		am = 0;
11002 	if ((rack->rc_last_tlp_acked_set ) &&
11003 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
11004 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
11005 		/*
11006 		 * The DSACK is because of a TLP which we don't
11007 		 * do anything with the reordering window over since
11008 		 * it was not reordering that caused the DSACK but
11009 		 * our previous retransmit TLP.
11010 		 */
11011 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
11012 		was_tlp = 1;
11013 		goto skip_dsack_round;
11014 	}
11015 	if (rack->rc_last_sent_tlp_seq_valid) {
11016 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
11017 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
11018 		    (SEQ_LEQ(end, l_end))) {
11019 			/*
11020 			 * This dsack is from the last sent TLP, ignore it
11021 			 * for reordering purposes.
11022 			 */
11023 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
11024 			was_tlp = 1;
11025 			goto skip_dsack_round;
11026 		}
11027 	}
11028 	if (rack->rc_dsack_round_seen == 0) {
11029 		rack->rc_dsack_round_seen = 1;
11030 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
11031 		rack->r_ctl.num_dsack++;
11032 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
11033 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
11034 	}
11035 skip_dsack_round:
11036 	/*
11037 	 * We keep track of how many DSACK blocks we get
11038 	 * after a recovery incident.
11039 	 */
11040 	rack->r_ctl.dsack_byte_cnt += am;
11041 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
11042 	    rack->r_ctl.retran_during_recovery &&
11043 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
11044 		/*
11045 		 * False recovery most likely culprit is reordering. If
11046 		 * nothing else is missing we need to revert.
11047 		 */
11048 		rack->r_might_revert = 1;
11049 		rack_handle_might_revert(rack->rc_tp, rack);
11050 		rack->r_might_revert = 0;
11051 		rack->r_ctl.retran_during_recovery = 0;
11052 		rack->r_ctl.dsack_byte_cnt = 0;
11053 	}
11054 	return (was_tlp);
11055 }
11056 
11057 static uint32_t
11058 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
11059 {
11060 	return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
11061 }
11062 
11063 static int32_t
11064 rack_compute_pipe(struct tcpcb *tp)
11065 {
11066 	return ((int32_t)do_rack_compute_pipe(tp,
11067 					      (struct tcp_rack *)tp->t_fb_ptr,
11068 					      tp->snd_una));
11069 }
11070 
11071 static void
11072 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
11073 {
11074 	/* Deal with changed and PRR here (in recovery only) */
11075 	uint32_t pipe, snd_una;
11076 
11077 	rack->r_ctl.rc_prr_delivered += changed;
11078 
11079 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
11080 		/*
11081 		 * It is all outstanding, we are application limited
11082 		 * and thus we don't need more room to send anything.
11083 		 * Note we use tp->snd_una here and not th_ack because
11084 		 * the data as yet not been cut from the sb.
11085 		 */
11086 		rack->r_ctl.rc_prr_sndcnt = 0;
11087 		return;
11088 	}
11089 	/* Compute prr_sndcnt */
11090 	if (SEQ_GT(tp->snd_una, th_ack)) {
11091 		snd_una = tp->snd_una;
11092 	} else {
11093 		snd_una = th_ack;
11094 	}
11095 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
11096 	if (pipe > tp->snd_ssthresh) {
11097 		long sndcnt;
11098 
11099 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
11100 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
11101 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
11102 		else {
11103 			rack->r_ctl.rc_prr_sndcnt = 0;
11104 			rack_log_to_prr(rack, 9, 0, __LINE__);
11105 			sndcnt = 0;
11106 		}
11107 		sndcnt++;
11108 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
11109 			sndcnt -= rack->r_ctl.rc_prr_out;
11110 		else
11111 			sndcnt = 0;
11112 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
11113 		rack_log_to_prr(rack, 10, 0, __LINE__);
11114 	} else {
11115 		uint32_t limit;
11116 
11117 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
11118 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
11119 		else
11120 			limit = 0;
11121 		if (changed > limit)
11122 			limit = changed;
11123 		limit += ctf_fixed_maxseg(tp);
11124 		if (tp->snd_ssthresh > pipe) {
11125 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
11126 			rack_log_to_prr(rack, 11, 0, __LINE__);
11127 		} else {
11128 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
11129 			rack_log_to_prr(rack, 12, 0, __LINE__);
11130 		}
11131 	}
11132 }
11133 
11134 static void
11135 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11136 	     int *dsack_seen, int *sacks_seen)
11137 {
11138 	uint32_t changed;
11139 	struct tcp_rack *rack;
11140 	struct rack_sendmap *rsm;
11141 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11142 	register uint32_t th_ack;
11143 	int32_t i, j, k, num_sack_blks = 0;
11144 	uint32_t cts, acked, ack_point;
11145 	int loop_start = 0, moved_two = 0, no_extra = 0;
11146 	uint32_t tsused;
11147 	uint32_t segsiz, o_cnt;
11148 
11149 
11150 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11151 	if (tcp_get_flags(th) & TH_RST) {
11152 		/* We don't log resets */
11153 		return;
11154 	}
11155 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11156 	cts = tcp_get_usecs(NULL);
11157 	rsm = tqhash_min(rack->r_ctl.tqh);
11158 	changed = 0;
11159 	th_ack = th->th_ack;
11160 	if (rack->sack_attack_disable == 0)
11161 		rack_do_decay(rack);
11162 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
11163 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11164 		/*
11165 		 * You only get credit for
11166 		 * MSS and greater (and you get extra
11167 		 * credit for larger cum-ack moves).
11168 		 */
11169 		int ac;
11170 
11171 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11172 		rack->r_ctl.ack_count += ac;
11173 		counter_u64_add(rack_ack_total, ac);
11174 	}
11175 	if (rack->r_ctl.ack_count > 0xfff00000) {
11176 		/*
11177 		 * reduce the number to keep us under
11178 		 * a uint32_t.
11179 		 */
11180 		rack->r_ctl.ack_count /= 2;
11181 		rack->r_ctl.sack_count /= 2;
11182 	}
11183 	if (SEQ_GT(th_ack, tp->snd_una)) {
11184 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
11185 		tp->t_acktime = ticks;
11186 	}
11187 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
11188 		changed = th_ack - rsm->r_start;
11189 	if (changed) {
11190 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
11191 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
11192 	}
11193 	if ((to->to_flags & TOF_SACK) == 0) {
11194 		/* We are done nothing left and no sack. */
11195 		rack_handle_might_revert(tp, rack);
11196 		/*
11197 		 * For cases where we struck a dup-ack
11198 		 * with no SACK, add to the changes so
11199 		 * PRR will work right.
11200 		 */
11201 		if (dup_ack_struck && (changed == 0)) {
11202 			changed += ctf_fixed_maxseg(rack->rc_tp);
11203 		}
11204 		goto out;
11205 	}
11206 	/* Sack block processing */
11207 	if (SEQ_GT(th_ack, tp->snd_una))
11208 		ack_point = th_ack;
11209 	else
11210 		ack_point = tp->snd_una;
11211 	for (i = 0; i < to->to_nsacks; i++) {
11212 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
11213 		      &sack, sizeof(sack));
11214 		sack.start = ntohl(sack.start);
11215 		sack.end = ntohl(sack.end);
11216 		if (SEQ_GT(sack.end, sack.start) &&
11217 		    SEQ_GT(sack.start, ack_point) &&
11218 		    SEQ_LT(sack.start, tp->snd_max) &&
11219 		    SEQ_GT(sack.end, ack_point) &&
11220 		    SEQ_LEQ(sack.end, tp->snd_max)) {
11221 			sack_blocks[num_sack_blks] = sack;
11222 			num_sack_blks++;
11223 		} else if (SEQ_LEQ(sack.start, th_ack) &&
11224 			   SEQ_LEQ(sack.end, th_ack)) {
11225 			int was_tlp;
11226 
11227 			if (dsack_seen != NULL)
11228 				*dsack_seen = 1;
11229 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
11230 			/*
11231 			 * Its a D-SACK block.
11232 			 */
11233 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
11234 		}
11235 	}
11236 	if (rack->rc_dsack_round_seen) {
11237 		/* Is the dsack roound over? */
11238 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
11239 			/* Yes it is */
11240 			rack->rc_dsack_round_seen = 0;
11241 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
11242 		}
11243 	}
11244 	/*
11245 	 * Sort the SACK blocks so we can update the rack scoreboard with
11246 	 * just one pass.
11247 	 */
11248 	o_cnt = num_sack_blks;
11249 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
11250 					 num_sack_blks, th->th_ack);
11251 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
11252 	if (sacks_seen != NULL)
11253 		*sacks_seen = num_sack_blks;
11254 	if (num_sack_blks == 0) {
11255 		/* Nothing to sack, but we need to update counts */
11256 		if ((o_cnt == 1) &&
11257 		    (*dsack_seen != 1))
11258 			rack->r_ctl.sack_count++;
11259 		else if (o_cnt > 1)
11260 			rack->r_ctl.sack_count++;
11261 		goto out_with_totals;
11262 	}
11263 	if (rack->sack_attack_disable) {
11264 		/*
11265 		 * An attacker disablement is in place, for
11266 		 * every sack block that is not at least a full MSS
11267 		 * count up sack_count.
11268 		 */
11269 		for (i = 0; i < num_sack_blks; i++) {
11270 			if ((sack_blocks[i].end - sack_blocks[i].start) < segsiz) {
11271 				rack->r_ctl.sack_count++;
11272 			}
11273 			if (rack->r_ctl.sack_count > 0xfff00000) {
11274 				/*
11275 				 * reduce the number to keep us under
11276 				 * a uint32_t.
11277 				 */
11278 				rack->r_ctl.ack_count /= 2;
11279 				rack->r_ctl.sack_count /= 2;
11280 			}
11281 		}
11282 		goto out;
11283 	}
11284 	/* Its a sack of some sort */
11285 	rack->r_ctl.sack_count += num_sack_blks;
11286 	if (rack->r_ctl.sack_count > 0xfff00000) {
11287 		/*
11288 		 * reduce the number to keep us under
11289 		 * a uint32_t.
11290 		 */
11291 		rack->r_ctl.ack_count /= 2;
11292 		rack->r_ctl.sack_count /= 2;
11293 	}
11294 	if (num_sack_blks < 2) {
11295 		/* Only one, we don't need to sort */
11296 		goto do_sack_work;
11297 	}
11298 	/* Sort the sacks */
11299 	for (i = 0; i < num_sack_blks; i++) {
11300 		for (j = i + 1; j < num_sack_blks; j++) {
11301 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
11302 				sack = sack_blocks[i];
11303 				sack_blocks[i] = sack_blocks[j];
11304 				sack_blocks[j] = sack;
11305 			}
11306 		}
11307 	}
11308 	/*
11309 	 * Now are any of the sack block ends the same (yes some
11310 	 * implementations send these)?
11311 	 */
11312 again:
11313 	if (num_sack_blks == 0)
11314 		goto out_with_totals;
11315 	if (num_sack_blks > 1) {
11316 		for (i = 0; i < num_sack_blks; i++) {
11317 			for (j = i + 1; j < num_sack_blks; j++) {
11318 				if (sack_blocks[i].end == sack_blocks[j].end) {
11319 					/*
11320 					 * Ok these two have the same end we
11321 					 * want the smallest end and then
11322 					 * throw away the larger and start
11323 					 * again.
11324 					 */
11325 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
11326 						/*
11327 						 * The second block covers
11328 						 * more area use that
11329 						 */
11330 						sack_blocks[i].start = sack_blocks[j].start;
11331 					}
11332 					/*
11333 					 * Now collapse out the dup-sack and
11334 					 * lower the count
11335 					 */
11336 					for (k = (j + 1); k < num_sack_blks; k++) {
11337 						sack_blocks[j].start = sack_blocks[k].start;
11338 						sack_blocks[j].end = sack_blocks[k].end;
11339 						j++;
11340 					}
11341 					num_sack_blks--;
11342 					goto again;
11343 				}
11344 			}
11345 		}
11346 	}
11347 do_sack_work:
11348 	/*
11349 	 * First lets look to see if
11350 	 * we have retransmitted and
11351 	 * can use the transmit next?
11352 	 */
11353 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11354 	if (rsm &&
11355 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
11356 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
11357 		/*
11358 		 * We probably did the FR and the next
11359 		 * SACK in continues as we would expect.
11360 		 */
11361 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11362 		if (acked) {
11363 			rack->r_wanted_output = 1;
11364 			changed += acked;
11365 		}
11366 		if (num_sack_blks == 1) {
11367 			/*
11368 			 * This is what we would expect from
11369 			 * a normal implementation to happen
11370 			 * after we have retransmitted the FR,
11371 			 * i.e the sack-filter pushes down
11372 			 * to 1 block and the next to be retransmitted
11373 			 * is the sequence in the sack block (has more
11374 			 * are acked). Count this as ACK'd data to boost
11375 			 * up the chances of recovering any false positives.
11376 			 */
11377 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
11378 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
11379 			counter_u64_add(rack_express_sack, 1);
11380 			if (rack->r_ctl.ack_count > 0xfff00000) {
11381 				/*
11382 				 * reduce the number to keep us under
11383 				 * a uint32_t.
11384 				 */
11385 				rack->r_ctl.ack_count /= 2;
11386 				rack->r_ctl.sack_count /= 2;
11387 			}
11388 			if (moved_two) {
11389 				/*
11390 				 * If we did not get a SACK for at least a MSS and
11391 				 * had to move at all, or if we moved more than our
11392 				 * threshold, it counts against the "extra" move.
11393 				 */
11394 				rack->r_ctl.sack_moved_extra += moved_two;
11395 				rack->r_ctl.sack_noextra_move += no_extra;
11396 				counter_u64_add(rack_move_some, 1);
11397 			} else {
11398 				/*
11399 				 * else we did not have to move
11400 				 * any more than we would expect.
11401 				 */
11402 				rack->r_ctl.sack_noextra_move += no_extra;
11403 				rack->r_ctl.sack_noextra_move++;
11404 				counter_u64_add(rack_move_none, 1);
11405 			}
11406 			if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11407 			    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11408 				rack->r_ctl.sack_moved_extra /= 2;
11409 				rack->r_ctl.sack_noextra_move /= 2;
11410 			}
11411 			goto out_with_totals;
11412 		} else {
11413 			/*
11414 			 * Start the loop through the
11415 			 * rest of blocks, past the first block.
11416 			 */
11417 			loop_start = 1;
11418 		}
11419 	}
11420 	counter_u64_add(rack_sack_total, 1);
11421 	rsm = rack->r_ctl.rc_sacklast;
11422 	for (i = loop_start; i < num_sack_blks; i++) {
11423 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11424 		if (acked) {
11425 			rack->r_wanted_output = 1;
11426 			changed += acked;
11427 		}
11428 		if (moved_two) {
11429 			/*
11430 			 * If we did not get a SACK for at least a MSS and
11431 			 * had to move at all, or if we moved more than our
11432 			 * threshold, it counts against the "extra" move.
11433 			 */
11434 			rack->r_ctl.sack_moved_extra += moved_two;
11435 			rack->r_ctl.sack_noextra_move += no_extra;
11436 			counter_u64_add(rack_move_some, 1);
11437 		} else {
11438 			/*
11439 			 * else we did not have to move
11440 			 * any more than we would expect.
11441 			 */
11442 			rack->r_ctl.sack_noextra_move += no_extra;
11443 			rack->r_ctl.sack_noextra_move++;
11444 			counter_u64_add(rack_move_none, 1);
11445 		}
11446 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11447 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11448 			rack->r_ctl.sack_moved_extra /= 2;
11449 			rack->r_ctl.sack_noextra_move /= 2;
11450 		}
11451 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
11452 			/*
11453 			 * If the SACK was not a full MSS then
11454 			 * we add to sack_count the number of
11455 			 * MSS's (or possibly more than
11456 			 * a MSS if its a TSO send) we had to skip by.
11457 			 */
11458 			rack->r_ctl.sack_count += moved_two;
11459 			if (rack->r_ctl.sack_count > 0xfff00000) {
11460 				rack->r_ctl.ack_count /= 2;
11461 				rack->r_ctl.sack_count /= 2;
11462 			}
11463 			counter_u64_add(rack_sack_total, moved_two);
11464 		}
11465 		/*
11466 		 * Now we need to setup for the next
11467 		 * round. First we make sure we won't
11468 		 * exceed the size of our uint32_t on
11469 		 * the various counts, and then clear out
11470 		 * moved_two.
11471 		 */
11472 		moved_two = 0;
11473 		no_extra = 0;
11474 	}
11475 out_with_totals:
11476 	if (num_sack_blks > 1) {
11477 		/*
11478 		 * You get an extra stroke if
11479 		 * you have more than one sack-blk, this
11480 		 * could be where we are skipping forward
11481 		 * and the sack-filter is still working, or
11482 		 * it could be an attacker constantly
11483 		 * moving us.
11484 		 */
11485 		rack->r_ctl.sack_moved_extra++;
11486 		counter_u64_add(rack_move_some, 1);
11487 	}
11488 out:
11489 #ifdef TCP_SAD_DETECTION
11490 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
11491 #endif
11492 	if (changed) {
11493 		/* Something changed cancel the rack timer */
11494 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11495 	}
11496 	tsused = tcp_get_usecs(NULL);
11497 	rsm = tcp_rack_output(tp, rack, tsused);
11498 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11499 	    rsm &&
11500 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11501 		/* Enter recovery */
11502 		entered_recovery = 1;
11503 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
11504 		/*
11505 		 * When we enter recovery we need to assure we send
11506 		 * one packet.
11507 		 */
11508 		if (rack->rack_no_prr == 0) {
11509 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11510 			rack_log_to_prr(rack, 8, 0, __LINE__);
11511 		}
11512 		rack->r_timer_override = 1;
11513 		rack->r_early = 0;
11514 		rack->r_ctl.rc_agg_early = 0;
11515 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11516 		   rsm &&
11517 		   (rack->r_rr_config == 3)) {
11518 		/*
11519 		 * Assure we can output and we get no
11520 		 * remembered pace time except the retransmit.
11521 		 */
11522 		rack->r_timer_override = 1;
11523 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11524 		rack->r_ctl.rc_resend = rsm;
11525 	}
11526 	if (IN_FASTRECOVERY(tp->t_flags) &&
11527 	    (rack->rack_no_prr == 0) &&
11528 	    (entered_recovery == 0)) {
11529 		rack_update_prr(tp, rack, changed, th_ack);
11530 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11531 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11532 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11533 			/*
11534 			 * If you are pacing output you don't want
11535 			 * to override.
11536 			 */
11537 			rack->r_early = 0;
11538 			rack->r_ctl.rc_agg_early = 0;
11539 			rack->r_timer_override = 1;
11540 		}
11541 	}
11542 }
11543 
11544 static void
11545 rack_strike_dupack(struct tcp_rack *rack)
11546 {
11547 	struct rack_sendmap *rsm;
11548 
11549 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11550 	while (rsm) {
11551 		/*
11552 		 * We need to skip anything already set
11553 		 * to be retransmitted.
11554 		 */
11555 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11556 		    (rsm->r_flags & RACK_MUST_RXT)) {
11557 			rsm = TAILQ_NEXT(rsm, r_tnext);
11558 			continue;
11559 		}
11560 		break;
11561 	}
11562 	if (rsm && (rsm->r_dupack < 0xff)) {
11563 		rsm->r_dupack++;
11564 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11565 			struct timeval tv;
11566 			uint32_t cts;
11567 			/*
11568 			 * Here we see if we need to retransmit. For
11569 			 * a SACK type connection if enough time has passed
11570 			 * we will get a return of the rsm. For a non-sack
11571 			 * connection we will get the rsm returned if the
11572 			 * dupack value is 3 or more.
11573 			 */
11574 			cts = tcp_get_usecs(&tv);
11575 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11576 			if (rack->r_ctl.rc_resend != NULL) {
11577 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11578 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11579 							 rack->rc_tp->snd_una, __LINE__);
11580 				}
11581 				rack->r_wanted_output = 1;
11582 				rack->r_timer_override = 1;
11583 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11584 			}
11585 		} else {
11586 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11587 		}
11588 	}
11589 }
11590 
11591 static void
11592 rack_check_bottom_drag(struct tcpcb *tp,
11593 		       struct tcp_rack *rack,
11594 		       struct socket *so)
11595 {
11596 	uint32_t segsiz, minseg;
11597 
11598 	segsiz = ctf_fixed_maxseg(tp);
11599 	minseg = segsiz;
11600 	if (tp->snd_max == tp->snd_una) {
11601 		/*
11602 		 * We are doing dynamic pacing and we are way
11603 		 * under. Basically everything got acked while
11604 		 * we were still waiting on the pacer to expire.
11605 		 *
11606 		 * This means we need to boost the b/w in
11607 		 * addition to any earlier boosting of
11608 		 * the multiplier.
11609 		 */
11610 		uint64_t lt_bw;
11611 
11612 		lt_bw = rack_get_lt_bw(rack);
11613 		rack->rc_dragged_bottom = 1;
11614 		rack_validate_multipliers_at_or_above100(rack);
11615 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11616 		    (lt_bw > 0)) {
11617 			/*
11618 			 * Lets use the long-term b/w we have
11619 			 * been getting as a base.
11620 			 */
11621 			if (rack->rc_gp_filled == 0) {
11622 				if (lt_bw > ONE_POINT_TWO_MEG) {
11623 					/*
11624 					 * If we have no measurement
11625 					 * don't let us set in more than
11626 					 * 1.2Mbps. If we are still too
11627 					 * low after pacing with this we
11628 					 * will hopefully have a max b/w
11629 					 * available to sanity check things.
11630 					 */
11631 					lt_bw = ONE_POINT_TWO_MEG;
11632 				}
11633 				rack->r_ctl.rc_rtt_diff = 0;
11634 				rack->r_ctl.gp_bw = lt_bw;
11635 				rack->rc_gp_filled = 1;
11636 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11637 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11638 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11639 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11640 				rack->r_ctl.rc_rtt_diff = 0;
11641 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11642 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11643 				rack->r_ctl.gp_bw = lt_bw;
11644 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11645 			} else
11646 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11647 			if ((rack->gp_ready == 0) &&
11648 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11649 				/* We have enough measurements now */
11650 				rack->gp_ready = 1;
11651 				if (rack->dgp_on ||
11652 				    rack->rack_hibeta)
11653 					rack_set_cc_pacing(rack);
11654 				if (rack->defer_options)
11655 					rack_apply_deferred_options(rack);
11656 			}
11657 		} else {
11658 			/*
11659 			 * zero rtt possibly?, settle for just an old increase.
11660 			 */
11661 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11662 		}
11663 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11664 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11665 					       minseg)) &&
11666 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11667 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11668 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11669 		    (segsiz * rack_req_segs))) {
11670 		/*
11671 		 * We are doing dynamic GP pacing and
11672 		 * we have everything except 1MSS or less
11673 		 * bytes left out. We are still pacing away.
11674 		 * And there is data that could be sent, This
11675 		 * means we are inserting delayed ack time in
11676 		 * our measurements because we are pacing too slow.
11677 		 */
11678 		rack_validate_multipliers_at_or_above100(rack);
11679 		rack->rc_dragged_bottom = 1;
11680 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11681 	}
11682 }
11683 
11684 #ifdef TCP_REQUEST_TRK
11685 static void
11686 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11687 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11688 {
11689 	int do_log;
11690 
11691 	do_log = tcp_bblogging_on(rack->rc_tp);
11692 	if (do_log == 0) {
11693 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11694 			return;
11695 		/* We only allow the three below with point logging on */
11696 		if ((mod != HYBRID_LOG_RULES_APP) &&
11697 		    (mod != HYBRID_LOG_RULES_SET) &&
11698 		    (mod != HYBRID_LOG_REQ_COMP))
11699 			return;
11700 
11701 	}
11702 	if (do_log) {
11703 		union tcp_log_stackspecific log;
11704 		struct timeval tv;
11705 
11706 		/* Convert our ms to a microsecond */
11707 		memset(&log, 0, sizeof(log));
11708 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11709 		log.u_bbr.flex1 = seq;
11710 		log.u_bbr.cwnd_gain = line;
11711 		if (cur != NULL) {
11712 			uint64_t off;
11713 
11714 			log.u_bbr.flex2 = cur->start_seq;
11715 			log.u_bbr.flex3 = cur->end_seq;
11716 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11717 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11718 			log.u_bbr.flex6 = cur->flags;
11719 			log.u_bbr.pkts_out = cur->hybrid_flags;
11720 			log.u_bbr.rttProp = cur->timestamp;
11721 			log.u_bbr.cur_del_rate = cur->cspr;
11722 			log.u_bbr.bw_inuse = cur->start;
11723 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11724 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11725 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11726 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11727 			log.u_bbr.bbr_state = 1;
11728 #ifdef TCP_REQUEST_TRK
11729 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11730 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11731 #endif
11732 		} else {
11733 			log.u_bbr.flex2 = err;
11734 		}
11735 		/*
11736 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11737 		 */
11738 		log.u_bbr.flex7 = rack->rc_catch_up;
11739 		log.u_bbr.flex7 <<= 1;
11740 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11741 		log.u_bbr.flex7 <<= 1;
11742 		log.u_bbr.flex7 |= rack->dgp_on;
11743 		log.u_bbr.flex8 = mod;
11744 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11745 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11746 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11747 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11748 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11749 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11750 		tcp_log_event(rack->rc_tp, NULL,
11751 		    &rack->rc_inp->inp_socket->so_rcv,
11752 		    &rack->rc_inp->inp_socket->so_snd,
11753 		    TCP_HYBRID_PACING_LOG, 0,
11754 	            0, &log, false, NULL, __func__, __LINE__, &tv);
11755 	}
11756 }
11757 #endif
11758 
11759 #ifdef TCP_REQUEST_TRK
11760 static void
11761 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len)
11762 {
11763 	struct tcp_sendfile_track *rc_cur;
11764 	struct tcpcb *tp;
11765 	int err = 0;
11766 
11767 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11768 	if (rc_cur == NULL) {
11769 		/* If not in the beginning what about the end piece */
11770 		if (rack->rc_hybrid_mode)
11771 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11772 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11773 	} else {
11774 		err = 12345;
11775 	}
11776 	/* If we find no parameters we are in straight DGP mode */
11777 	if(rc_cur == NULL) {
11778 		/* None found for this seq, just DGP for now */
11779 		rack->r_ctl.client_suggested_maxseg = 0;
11780 		rack->rc_catch_up = 0;
11781 		rack->r_ctl.bw_rate_cap = 0;
11782 		if (rack->rc_hybrid_mode)
11783 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11784 		if (rack->r_ctl.rc_last_sft) {
11785 			rack->r_ctl.rc_last_sft = NULL;
11786 		}
11787 		return;
11788 	}
11789 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11790 		/* This entry was never setup for hybrid pacing on/off etc */
11791 		return;
11792 	}
11793 	/*
11794 	 * Ok if we have a new entry *or* have never
11795 	 * set up an entry we need to proceed. If
11796 	 * we have already set it up this entry we
11797 	 * just continue along with what we already
11798 	 * setup.
11799 	 */
11800 	tp = rack->rc_tp;
11801 	if ((rack->r_ctl.rc_last_sft != NULL) &&
11802 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
11803 		/* Its already in place */
11804 		if (rack->rc_hybrid_mode)
11805 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11806 		return;
11807 	}
11808 	if (rack->rc_hybrid_mode == 0) {
11809 		rack->r_ctl.rc_last_sft = rc_cur;
11810 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11811 		return;
11812 	}
11813 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11814 		/* Compensate for all the header overhead's */
11815 		rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
11816 	} else
11817 		rack->r_ctl.bw_rate_cap = 0;
11818 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11819 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11820 	else
11821 		rack->r_ctl.client_suggested_maxseg = 0;
11822 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11823 	    (rc_cur->cspr > 0)) {
11824 		uint64_t len;
11825 
11826 		rack->rc_catch_up = 1;
11827 		/*
11828 		 * Calculate the deadline time, first set the
11829 		 * time to when the request arrived.
11830 		 */
11831 		rc_cur->deadline = rc_cur->localtime;
11832 		/*
11833 		 * Next calculate the length and compensate for
11834 		 * TLS if need be.
11835 		 */
11836 		len = rc_cur->end - rc_cur->start;
11837 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11838 			/*
11839 			 * This session is doing TLS. Take a swag guess
11840 			 * at the overhead.
11841 			 */
11842 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11843 		}
11844 		/*
11845 		 * Now considering the size, and the cspr, what is the time that
11846 		 * would be required at the cspr rate. Here we use the raw
11847 		 * cspr value since the client only looks at the raw data. We
11848 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
11849 		 * That will get made up for in the CU pacing rate set.
11850 		 */
11851 		len *= HPTS_USEC_IN_SEC;
11852 		len /= rc_cur->cspr;
11853 		rc_cur->deadline += len;
11854 	} else {
11855 		rack->rc_catch_up = 0;
11856 		rc_cur->deadline = 0;
11857 	}
11858 	if (rack->r_ctl.client_suggested_maxseg != 0) {
11859 		/*
11860 		 * We need to reset the max pace segs if we have a
11861 		 * client_suggested_maxseg.
11862 		 */
11863 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
11864 	}
11865 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11866 	/* Remember it for next time and for CU mode */
11867 	rack->r_ctl.rc_last_sft = rc_cur;
11868 }
11869 #endif
11870 
11871 static void
11872 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11873 {
11874 #ifdef TCP_REQUEST_TRK
11875 	struct tcp_sendfile_track *ent;
11876 
11877 	ent = rack->r_ctl.rc_last_sft;
11878 	if ((ent == NULL) ||
11879 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11880 	    (SEQ_GEQ(seq, ent->end_seq))) {
11881 		/* Time to update the track. */
11882 		rack_set_dgp_hybrid_mode(rack, seq, len);
11883 		ent = rack->r_ctl.rc_last_sft;
11884 	}
11885 	/* Out of all */
11886 	if (ent == NULL) {
11887 		return;
11888 	}
11889 	if (SEQ_LT(ent->end_seq, (seq + len))) {
11890 		/*
11891 		 * This is the case where our end_seq guess
11892 		 * was wrong. This is usually due to TLS having
11893 		 * more bytes then our guess. It could also be the
11894 		 * case that the client sent in two requests closely
11895 		 * and the SB is full of both so we are sending part
11896 		 * of each (end|beg). In such a case lets move this
11897 		 * guys end to match the end of this send. That
11898 		 * way it will complete when all of it is acked.
11899 		 */
11900 		ent->end_seq = (seq + len);
11901 		if (rack->rc_hybrid_mode)
11902 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11903 	}
11904 	/* Now validate we have set the send time of this one */
11905 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11906 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11907 		ent->first_send = cts;
11908 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11909 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11910 	}
11911 #endif
11912 }
11913 
11914 static void
11915 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11916 {
11917 	/*
11918 	 * The fast output path is enabled and we
11919 	 * have moved the cumack forward. Lets see if
11920 	 * we can expand forward the fast path length by
11921 	 * that amount. What we would ideally like to
11922 	 * do is increase the number of bytes in the
11923 	 * fast path block (left_to_send) by the
11924 	 * acked amount. However we have to gate that
11925 	 * by two factors:
11926 	 * 1) The amount outstanding and the rwnd of the peer
11927 	 *    (i.e. we don't want to exceed the rwnd of the peer).
11928 	 *    <and>
11929 	 * 2) The amount of data left in the socket buffer (i.e.
11930 	 *    we can't send beyond what is in the buffer).
11931 	 *
11932 	 * Note that this does not take into account any increase
11933 	 * in the cwnd. We will only extend the fast path by
11934 	 * what was acked.
11935 	 */
11936 	uint32_t new_total, gating_val;
11937 
11938 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11939 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11940 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11941 	if (new_total <= gating_val) {
11942 		/* We can increase left_to_send by the acked amount */
11943 		counter_u64_add(rack_extended_rfo, 1);
11944 		rack->r_ctl.fsb.left_to_send = new_total;
11945 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11946 			("rack:%p left_to_send:%u sbavail:%u out:%u",
11947 			 rack, rack->r_ctl.fsb.left_to_send,
11948 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
11949 			 (tp->snd_max - tp->snd_una)));
11950 
11951 	}
11952 }
11953 
11954 static void
11955 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11956 {
11957 	/*
11958 	 * Here any sendmap entry that points to the
11959 	 * beginning mbuf must be adjusted to the correct
11960 	 * offset. This must be called with:
11961 	 * 1) The socket buffer locked
11962 	 * 2) snd_una adjusted to its new position.
11963 	 *
11964 	 * Note that (2) implies rack_ack_received has also
11965 	 * been called and all the sbcut's have been done.
11966 	 *
11967 	 * We grab the first mbuf in the socket buffer and
11968 	 * then go through the front of the sendmap, recalculating
11969 	 * the stored offset for any sendmap entry that has
11970 	 * that mbuf. We must use the sb functions to do this
11971 	 * since its possible an add was done has well as
11972 	 * the subtraction we may have just completed. This should
11973 	 * not be a penalty though, since we just referenced the sb
11974 	 * to go in and trim off the mbufs that we freed (of course
11975 	 * there will be a penalty for the sendmap references though).
11976 	 *
11977 	 * Note also with INVARIANT on, we validate with a KASSERT
11978 	 * that the first sendmap entry has a soff of 0.
11979 	 *
11980 	 */
11981 	struct mbuf *m;
11982 	struct rack_sendmap *rsm;
11983 	tcp_seq snd_una;
11984 #ifdef INVARIANTS
11985 	int first_processed = 0;
11986 #endif
11987 
11988 	snd_una = rack->rc_tp->snd_una;
11989 	SOCKBUF_LOCK_ASSERT(sb);
11990 	m = sb->sb_mb;
11991 	rsm = tqhash_min(rack->r_ctl.tqh);
11992 	if ((rsm == NULL) || (m == NULL)) {
11993 		/* Nothing outstanding */
11994 		return;
11995 	}
11996 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
11997 	KASSERT((rsm->m == m),
11998 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11999 		 rack, sb, rsm));
12000 	while (rsm->m && (rsm->m == m)) {
12001 		/* one to adjust */
12002 #ifdef INVARIANTS
12003 		struct mbuf *tm;
12004 		uint32_t soff;
12005 
12006 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
12007 		if ((rsm->orig_m_len != m->m_len) ||
12008 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
12009 			rack_adjust_orig_mlen(rsm);
12010 		}
12011 		if (first_processed == 0) {
12012 			KASSERT((rsm->soff == 0),
12013 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
12014 				 rack, rsm));
12015 			first_processed = 1;
12016 		}
12017 		if ((rsm->soff != soff) || (rsm->m != tm)) {
12018 			/*
12019 			 * This is not a fatal error, we anticipate it
12020 			 * might happen (the else code), so we count it here
12021 			 * so that under invariant we can see that it really
12022 			 * does happen.
12023 			 */
12024 			counter_u64_add(rack_adjust_map_bw, 1);
12025 		}
12026 		rsm->m = tm;
12027 		rsm->soff = soff;
12028 		if (tm) {
12029 			rsm->orig_m_len = rsm->m->m_len;
12030 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12031 		} else {
12032 			rsm->orig_m_len = 0;
12033 			rsm->orig_t_space = 0;
12034 		}
12035 #else
12036 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
12037 		if (rsm->m) {
12038 			rsm->orig_m_len = rsm->m->m_len;
12039 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12040 		} else {
12041 			rsm->orig_m_len = 0;
12042 			rsm->orig_t_space = 0;
12043 		}
12044 #endif
12045 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
12046 		if (rsm == NULL)
12047 			break;
12048 	}
12049 }
12050 
12051 #ifdef TCP_REQUEST_TRK
12052 static inline void
12053 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
12054 {
12055 	struct tcp_sendfile_track *ent;
12056 	int i;
12057 
12058 	if ((rack->rc_hybrid_mode == 0) &&
12059 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
12060 		/*
12061 		 * Just do normal completions hybrid pacing is not on
12062 		 * and CLDL is off as well.
12063 		 */
12064 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
12065 		return;
12066 	}
12067 	/*
12068 	 * Originally I was just going to find the th_ack associated
12069 	 * with an entry. But then I realized a large strech ack could
12070 	 * in theory ack two or more requests at once. So instead we
12071 	 * need to find all entries that are completed by th_ack not
12072 	 * just a single entry and do our logging.
12073 	 */
12074 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12075 	while (ent != NULL) {
12076 		/*
12077 		 * We may be doing hybrid pacing or CLDL and need more details possibly
12078 		 * so we do it manually instead of calling
12079 		 * tcp_req_check_for_comp()
12080 		 */
12081 		uint64_t laa, tim, data, cbw, ftim;
12082 
12083 		/* Ok this ack frees it */
12084 		rack_log_hybrid(rack, th_ack,
12085 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
12086 		rack_log_hybrid_sends(rack, ent, __LINE__);
12087 		/* calculate the time based on the ack arrival */
12088 		data = ent->end - ent->start;
12089 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
12090 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
12091 			if (ent->first_send > ent->localtime)
12092 				ftim = ent->first_send;
12093 			else
12094 				ftim = ent->localtime;
12095 		} else {
12096 			/* TSNH */
12097 			ftim = ent->localtime;
12098 		}
12099 		if (laa > ent->localtime)
12100 			tim = laa - ftim;
12101 		else
12102 			tim = 0;
12103 		cbw = data * HPTS_USEC_IN_SEC;
12104 		if (tim > 0)
12105 			cbw /= tim;
12106 		else
12107 			cbw = 0;
12108 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
12109 		/*
12110 		 * Check to see if we are freeing what we are pointing to send wise
12111 		 * if so be sure to NULL the pointer so we know we are no longer
12112 		 * set to anything.
12113 		 */
12114 		if (ent == rack->r_ctl.rc_last_sft)
12115 			rack->r_ctl.rc_last_sft = NULL;
12116 		/* Generate the log that the tcp_netflix call would have */
12117 		tcp_req_log_req_info(rack->rc_tp, ent,
12118 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
12119 		/* Free it and see if there is another one */
12120 		tcp_req_free_a_slot(rack->rc_tp, ent);
12121 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12122 	}
12123 }
12124 #endif
12125 
12126 
12127 /*
12128  * Return value of 1, we do not need to call rack_process_data().
12129  * return value of 0, rack_process_data can be called.
12130  * For ret_val if its 0 the TCP is locked, if its non-zero
12131  * its unlocked and probably unsafe to touch the TCB.
12132  */
12133 static int
12134 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12135     struct tcpcb *tp, struct tcpopt *to,
12136     uint32_t tiwin, int32_t tlen,
12137     int32_t * ofia, int32_t thflags, int32_t *ret_val)
12138 {
12139 	int32_t ourfinisacked = 0;
12140 	int32_t nsegs, acked_amount;
12141 	int32_t acked;
12142 	struct mbuf *mfree;
12143 	struct tcp_rack *rack;
12144 	int32_t under_pacing = 0;
12145 	int32_t recovery = 0;
12146 
12147 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12148 
12149 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12150 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
12151 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
12152 				      &rack->r_ctl.challenge_ack_ts,
12153 				      &rack->r_ctl.challenge_ack_cnt);
12154 		rack->r_wanted_output = 1;
12155 		return (1);
12156 	}
12157 	if (rack->gp_ready &&
12158 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12159 		under_pacing = 1;
12160 	}
12161 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
12162 		int in_rec, dup_ack_struck = 0;
12163 		int dsack_seen = 0, sacks_seen = 0;
12164 
12165 		in_rec = IN_FASTRECOVERY(tp->t_flags);
12166 		if (rack->rc_in_persist) {
12167 			tp->t_rxtshift = 0;
12168 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12169 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12170 		}
12171 
12172 		if ((th->th_ack == tp->snd_una) &&
12173 		    (tiwin == tp->snd_wnd) &&
12174 		    ((to->to_flags & TOF_SACK) == 0)) {
12175 			rack_strike_dupack(rack);
12176 			dup_ack_struck = 1;
12177 		}
12178 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
12179 			     dup_ack_struck, &dsack_seen, &sacks_seen);
12180 		if ((rack->sack_attack_disable > 0) &&
12181 		    (th->th_ack == tp->snd_una) &&
12182 		    (tiwin == tp->snd_wnd) &&
12183 		    (dsack_seen == 0) &&
12184 		    (sacks_seen > 0)) {
12185 			/*
12186 			 * If sacks have been disabled we may
12187 			 * want to strike a dup-ack "ignoring" the
12188 			 * sack as long as the sack was not a "dsack". Note
12189 			 * that if no sack is sent (TOF_SACK is off) then the
12190 			 * normal dsack code above rack_log_ack() would have
12191 			 * already struck. So this is just to catch the case
12192 			 * were we are ignoring sacks from this guy due to
12193 			 * it being a suspected attacker.
12194 			 */
12195 			rack_strike_dupack(rack);
12196 		}
12197 
12198 	}
12199 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12200 		/*
12201 		 * Old ack, behind (or duplicate to) the last one rcv'd
12202 		 * Note: We mark reordering is occuring if its
12203 		 * less than and we have not closed our window.
12204 		 */
12205 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
12206 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12207 			if (rack->r_ctl.rc_reorder_ts == 0)
12208 				rack->r_ctl.rc_reorder_ts = 1;
12209 		}
12210 		return (0);
12211 	}
12212 	/*
12213 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
12214 	 * something we sent.
12215 	 */
12216 	if (tp->t_flags & TF_NEEDSYN) {
12217 		/*
12218 		 * T/TCP: Connection was half-synchronized, and our SYN has
12219 		 * been ACK'd (so connection is now fully synchronized).  Go
12220 		 * to non-starred state, increment snd_una for ACK of SYN,
12221 		 * and check if we can do window scaling.
12222 		 */
12223 		tp->t_flags &= ~TF_NEEDSYN;
12224 		tp->snd_una++;
12225 		/* Do window scaling? */
12226 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12227 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12228 			tp->rcv_scale = tp->request_r_scale;
12229 			/* Send window already scaled. */
12230 		}
12231 	}
12232 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12233 
12234 	acked = BYTES_THIS_ACK(tp, th);
12235 	if (acked) {
12236 		/*
12237 		 * Any time we move the cum-ack forward clear
12238 		 * keep-alive tied probe-not-answered. The
12239 		 * persists clears its own on entry.
12240 		 */
12241 		rack->probe_not_answered = 0;
12242 	}
12243 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12244 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12245 	/*
12246 	 * If we just performed our first retransmit, and the ACK arrives
12247 	 * within our recovery window, then it was a mistake to do the
12248 	 * retransmit in the first place.  Recover our original cwnd and
12249 	 * ssthresh, and proceed to transmit where we left off.
12250 	 */
12251 	if ((tp->t_flags & TF_PREVVALID) &&
12252 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12253 		tp->t_flags &= ~TF_PREVVALID;
12254 		if (tp->t_rxtshift == 1 &&
12255 		    (int)(ticks - tp->t_badrxtwin) < 0)
12256 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12257 	}
12258 	if (acked) {
12259 		/* assure we are not backed off */
12260 		tp->t_rxtshift = 0;
12261 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12262 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12263 		rack->rc_tlp_in_progress = 0;
12264 		rack->r_ctl.rc_tlp_cnt_out = 0;
12265 		/*
12266 		 * If it is the RXT timer we want to
12267 		 * stop it, so we can restart a TLP.
12268 		 */
12269 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12270 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12271 #ifdef TCP_REQUEST_TRK
12272 		rack_req_check_for_comp(rack, th->th_ack);
12273 #endif
12274 	}
12275 	/*
12276 	 * If we have a timestamp reply, update smoothed round trip time. If
12277 	 * no timestamp is present but transmit timer is running and timed
12278 	 * sequence number was acked, update smoothed round trip time. Since
12279 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
12280 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
12281 	 * timer.
12282 	 *
12283 	 * Some boxes send broken timestamp replies during the SYN+ACK
12284 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12285 	 * and blow up the retransmit timer.
12286 	 */
12287 	/*
12288 	 * If all outstanding data is acked, stop retransmit timer and
12289 	 * remember to restart (more output or persist). If there is more
12290 	 * data to be acked, restart retransmit timer, using current
12291 	 * (possibly backed-off) value.
12292 	 */
12293 	if (acked == 0) {
12294 		if (ofia)
12295 			*ofia = ourfinisacked;
12296 		return (0);
12297 	}
12298 	if (IN_RECOVERY(tp->t_flags)) {
12299 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
12300 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
12301 			tcp_rack_partialack(tp);
12302 		} else {
12303 			rack_post_recovery(tp, th->th_ack);
12304 			recovery = 1;
12305 		}
12306 	}
12307 	/*
12308 	 * Let the congestion control algorithm update congestion control
12309 	 * related information. This typically means increasing the
12310 	 * congestion window.
12311 	 */
12312 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
12313 	SOCKBUF_LOCK(&so->so_snd);
12314 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
12315 	tp->snd_wnd -= acked_amount;
12316 	mfree = sbcut_locked(&so->so_snd, acked_amount);
12317 	if ((sbused(&so->so_snd) == 0) &&
12318 	    (acked > acked_amount) &&
12319 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
12320 	    (tp->t_flags & TF_SENTFIN)) {
12321 		/*
12322 		 * We must be sure our fin
12323 		 * was sent and acked (we can be
12324 		 * in FIN_WAIT_1 without having
12325 		 * sent the fin).
12326 		 */
12327 		ourfinisacked = 1;
12328 	}
12329 	tp->snd_una = th->th_ack;
12330 	/* wakeups? */
12331 	if (acked_amount && sbavail(&so->so_snd))
12332 		rack_adjust_sendmap_head(rack, &so->so_snd);
12333 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12334 	/* NB: sowwakeup_locked() does an implicit unlock. */
12335 	sowwakeup_locked(so);
12336 	/* now check the rxt clamps */
12337 	if ((recovery == 1) &&
12338 	    (rack->excess_rxt_on) &&
12339 	    (rack->r_cwnd_was_clamped == 0))  {
12340 		do_rack_excess_rxt(tp, rack);
12341 	} else if (rack->r_cwnd_was_clamped)
12342 		do_rack_check_for_unclamp(tp, rack);
12343 	m_freem(mfree);
12344 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
12345 		tp->snd_recover = tp->snd_una;
12346 
12347 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
12348 		tp->snd_nxt = tp->snd_una;
12349 	}
12350 	if (under_pacing &&
12351 	    (rack->use_fixed_rate == 0) &&
12352 	    (rack->in_probe_rtt == 0) &&
12353 	    rack->rc_gp_dyn_mul &&
12354 	    rack->rc_always_pace) {
12355 		/* Check if we are dragging bottom */
12356 		rack_check_bottom_drag(tp, rack, so);
12357 	}
12358 	if (tp->snd_una == tp->snd_max) {
12359 		/* Nothing left outstanding */
12360 		tp->t_flags &= ~TF_PREVVALID;
12361 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12362 		rack->r_ctl.retran_during_recovery = 0;
12363 		rack->r_ctl.dsack_byte_cnt = 0;
12364 		if (rack->r_ctl.rc_went_idle_time == 0)
12365 			rack->r_ctl.rc_went_idle_time = 1;
12366 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12367 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12368 			tp->t_acktime = 0;
12369 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12370 		rack->rc_suspicious = 0;
12371 		/* Set need output so persist might get set */
12372 		rack->r_wanted_output = 1;
12373 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12374 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12375 		    (sbavail(&so->so_snd) == 0) &&
12376 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12377 			/*
12378 			 * The socket was gone and the
12379 			 * peer sent data (now or in the past), time to
12380 			 * reset him.
12381 			 */
12382 			*ret_val = 1;
12383 			/* tcp_close will kill the inp pre-log the Reset */
12384 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12385 			tp = tcp_close(tp);
12386 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12387 			return (1);
12388 		}
12389 	}
12390 	if (ofia)
12391 		*ofia = ourfinisacked;
12392 	return (0);
12393 }
12394 
12395 
12396 static void
12397 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12398 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12399 {
12400 	if (tcp_bblogging_on(rack->rc_tp)) {
12401 		union tcp_log_stackspecific log;
12402 		struct timeval tv;
12403 
12404 		memset(&log, 0, sizeof(log));
12405 		log.u_bbr.flex1 = cnt;
12406 		log.u_bbr.flex2 = split;
12407 		log.u_bbr.flex3 = out;
12408 		log.u_bbr.flex4 = line;
12409 		log.u_bbr.flex5 = rack->r_must_retran;
12410 		log.u_bbr.flex6 = flags;
12411 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12412 		log.u_bbr.flex8 = dir;	/*
12413 					 * 1 is collapsed, 0 is uncollapsed,
12414 					 * 2 is log of a rsm being marked, 3 is a split.
12415 					 */
12416 		if (rsm == NULL)
12417 			log.u_bbr.rttProp = 0;
12418 		else
12419 			log.u_bbr.rttProp = (uint64_t)rsm;
12420 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12421 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12422 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12423 		    &rack->rc_inp->inp_socket->so_rcv,
12424 		    &rack->rc_inp->inp_socket->so_snd,
12425 		    TCP_RACK_LOG_COLLAPSE, 0,
12426 		    0, &log, false, &tv);
12427 	}
12428 }
12429 
12430 static void
12431 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12432 {
12433 	/*
12434 	 * Here all we do is mark the collapsed point and set the flag.
12435 	 * This may happen again and again, but there is no
12436 	 * sense splitting our map until we know where the
12437 	 * peer finally lands in the collapse.
12438 	 */
12439 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12440 	if ((rack->rc_has_collapsed == 0) ||
12441 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12442 		counter_u64_add(rack_collapsed_win_seen, 1);
12443 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12444 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12445 	rack->rc_has_collapsed = 1;
12446 	rack->r_collapse_point_valid = 1;
12447 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12448 }
12449 
12450 static void
12451 rack_un_collapse_window(struct tcp_rack *rack, int line)
12452 {
12453 	struct rack_sendmap *nrsm, *rsm;
12454 	int cnt = 0, split = 0;
12455 	int insret __diagused;
12456 
12457 
12458 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12459 	rack->rc_has_collapsed = 0;
12460 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12461 	if (rsm == NULL) {
12462 		/* Nothing to do maybe the peer ack'ed it all */
12463 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12464 		return;
12465 	}
12466 	/* Now do we need to split this one? */
12467 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12468 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12469 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12470 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12471 		if (nrsm == NULL) {
12472 			/* We can't get a rsm, mark all? */
12473 			nrsm = rsm;
12474 			goto no_split;
12475 		}
12476 		/* Clone it */
12477 		split = 1;
12478 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12479 #ifndef INVARIANTS
12480 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12481 #else
12482 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12483 			panic("Insert in rb tree of %p fails ret:%d rack:%p rsm:%p",
12484 			      nrsm, insret, rack, rsm);
12485 		}
12486 #endif
12487 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12488 				 rack->r_ctl.last_collapse_point, __LINE__);
12489 		if (rsm->r_in_tmap) {
12490 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12491 			nrsm->r_in_tmap = 1;
12492 		}
12493 		/*
12494 		 * Set in the new RSM as the
12495 		 * collapsed starting point
12496 		 */
12497 		rsm = nrsm;
12498 	}
12499 
12500 no_split:
12501 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12502 		cnt++;
12503 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12504 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12505 		cnt++;
12506 	}
12507 	if (cnt) {
12508 		counter_u64_add(rack_collapsed_win, 1);
12509 	}
12510 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12511 }
12512 
12513 static void
12514 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12515 			int32_t tlen, int32_t tfo_syn)
12516 {
12517 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12518 		rack_timer_cancel(tp, rack,
12519 				  rack->r_ctl.rc_rcvtime, __LINE__);
12520 		tp->t_flags |= TF_DELACK;
12521 	} else {
12522 		rack->r_wanted_output = 1;
12523 		tp->t_flags |= TF_ACKNOW;
12524 	}
12525 }
12526 
12527 static void
12528 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12529 {
12530 	/*
12531 	 * If fast output is in progress, lets validate that
12532 	 * the new window did not shrink on us and make it
12533 	 * so fast output should end.
12534 	 */
12535 	if (rack->r_fast_output) {
12536 		uint32_t out;
12537 
12538 		/*
12539 		 * Calculate what we will send if left as is
12540 		 * and compare that to our send window.
12541 		 */
12542 		out = ctf_outstanding(tp);
12543 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12544 			/* ok we have an issue */
12545 			if (out >= tp->snd_wnd) {
12546 				/* Turn off fast output the window is met or collapsed */
12547 				rack->r_fast_output = 0;
12548 			} else {
12549 				/* we have some room left */
12550 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12551 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12552 					/* If not at least 1 full segment never mind */
12553 					rack->r_fast_output = 0;
12554 				}
12555 			}
12556 		}
12557 	}
12558 }
12559 
12560 
12561 /*
12562  * Return value of 1, the TCB is unlocked and most
12563  * likely gone, return value of 0, the TCP is still
12564  * locked.
12565  */
12566 static int
12567 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12568     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12569     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12570 {
12571 	/*
12572 	 * Update window information. Don't look at window if no ACK: TAC's
12573 	 * send garbage on first SYN.
12574 	 */
12575 	int32_t nsegs;
12576 	int32_t tfo_syn;
12577 	struct tcp_rack *rack;
12578 
12579 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12580 
12581 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12582 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12583 	if ((thflags & TH_ACK) &&
12584 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12585 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12586 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12587 		/* keep track of pure window updates */
12588 		if (tlen == 0 &&
12589 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12590 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12591 		tp->snd_wnd = tiwin;
12592 		rack_validate_fo_sendwin_up(tp, rack);
12593 		tp->snd_wl1 = th->th_seq;
12594 		tp->snd_wl2 = th->th_ack;
12595 		if (tp->snd_wnd > tp->max_sndwnd)
12596 			tp->max_sndwnd = tp->snd_wnd;
12597 		rack->r_wanted_output = 1;
12598 	} else if (thflags & TH_ACK) {
12599 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12600 			tp->snd_wnd = tiwin;
12601 			rack_validate_fo_sendwin_up(tp, rack);
12602 			tp->snd_wl1 = th->th_seq;
12603 			tp->snd_wl2 = th->th_ack;
12604 		}
12605 	}
12606 	if (tp->snd_wnd < ctf_outstanding(tp))
12607 		/* The peer collapsed the window */
12608 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12609 	else if (rack->rc_has_collapsed)
12610 		rack_un_collapse_window(rack, __LINE__);
12611 	if ((rack->r_collapse_point_valid) &&
12612 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12613 		rack->r_collapse_point_valid = 0;
12614 	/* Was persist timer active and now we have window space? */
12615 	if ((rack->rc_in_persist != 0) &&
12616 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12617 				rack->r_ctl.rc_pace_min_segs))) {
12618 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12619 		tp->snd_nxt = tp->snd_max;
12620 		/* Make sure we output to start the timer */
12621 		rack->r_wanted_output = 1;
12622 	}
12623 	/* Do we enter persists? */
12624 	if ((rack->rc_in_persist == 0) &&
12625 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12626 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12627 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12628 	    sbavail(&tptosocket(tp)->so_snd) &&
12629 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12630 		/*
12631 		 * Here the rwnd is less than
12632 		 * the pacing size, we are established,
12633 		 * nothing is outstanding, and there is
12634 		 * data to send. Enter persists.
12635 		 */
12636 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12637 	}
12638 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12639 		m_freem(m);
12640 		return (0);
12641 	}
12642 	/*
12643 	 * don't process the URG bit, ignore them drag
12644 	 * along the up.
12645 	 */
12646 	tp->rcv_up = tp->rcv_nxt;
12647 
12648 	/*
12649 	 * Process the segment text, merging it into the TCP sequencing
12650 	 * queue, and arranging for acknowledgment of receipt if necessary.
12651 	 * This process logically involves adjusting tp->rcv_wnd as data is
12652 	 * presented to the user (this happens in tcp_usrreq.c, case
12653 	 * PRU_RCVD).  If a FIN has already been received on this connection
12654 	 * then we just ignore the text.
12655 	 */
12656 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12657 		   IS_FASTOPEN(tp->t_flags));
12658 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12659 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12660 		tcp_seq save_start = th->th_seq;
12661 		tcp_seq save_rnxt  = tp->rcv_nxt;
12662 		int     save_tlen  = tlen;
12663 
12664 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12665 		/*
12666 		 * Insert segment which includes th into TCP reassembly
12667 		 * queue with control block tp.  Set thflags to whether
12668 		 * reassembly now includes a segment with FIN.  This handles
12669 		 * the common case inline (segment is the next to be
12670 		 * received on an established connection, and the queue is
12671 		 * empty), avoiding linkage into and removal from the queue
12672 		 * and repetition of various conversions. Set DELACK for
12673 		 * segments received in order, but ack immediately when
12674 		 * segments are out of order (so fast retransmit can work).
12675 		 */
12676 		if (th->th_seq == tp->rcv_nxt &&
12677 		    SEGQ_EMPTY(tp) &&
12678 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
12679 		    tfo_syn)) {
12680 #ifdef NETFLIX_SB_LIMITS
12681 			u_int mcnt, appended;
12682 
12683 			if (so->so_rcv.sb_shlim) {
12684 				mcnt = m_memcnt(m);
12685 				appended = 0;
12686 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12687 				    CFO_NOSLEEP, NULL) == false) {
12688 					counter_u64_add(tcp_sb_shlim_fails, 1);
12689 					m_freem(m);
12690 					return (0);
12691 				}
12692 			}
12693 #endif
12694 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12695 			tp->rcv_nxt += tlen;
12696 			if (tlen &&
12697 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12698 			    (tp->t_fbyte_in == 0)) {
12699 				tp->t_fbyte_in = ticks;
12700 				if (tp->t_fbyte_in == 0)
12701 					tp->t_fbyte_in = 1;
12702 				if (tp->t_fbyte_out && tp->t_fbyte_in)
12703 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12704 			}
12705 			thflags = tcp_get_flags(th) & TH_FIN;
12706 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12707 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12708 			SOCKBUF_LOCK(&so->so_rcv);
12709 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12710 				m_freem(m);
12711 			} else
12712 #ifdef NETFLIX_SB_LIMITS
12713 				appended =
12714 #endif
12715 					sbappendstream_locked(&so->so_rcv, m, 0);
12716 
12717 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12718 			/* NB: sorwakeup_locked() does an implicit unlock. */
12719 			sorwakeup_locked(so);
12720 #ifdef NETFLIX_SB_LIMITS
12721 			if (so->so_rcv.sb_shlim && appended != mcnt)
12722 				counter_fo_release(so->so_rcv.sb_shlim,
12723 				    mcnt - appended);
12724 #endif
12725 		} else {
12726 			/*
12727 			 * XXX: Due to the header drop above "th" is
12728 			 * theoretically invalid by now.  Fortunately
12729 			 * m_adj() doesn't actually frees any mbufs when
12730 			 * trimming from the head.
12731 			 */
12732 			tcp_seq temp = save_start;
12733 
12734 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
12735 			tp->t_flags |= TF_ACKNOW;
12736 			if (tp->t_flags & TF_WAKESOR) {
12737 				tp->t_flags &= ~TF_WAKESOR;
12738 				/* NB: sorwakeup_locked() does an implicit unlock. */
12739 				sorwakeup_locked(so);
12740 			}
12741 		}
12742 		if ((tp->t_flags & TF_SACK_PERMIT) &&
12743 		    (save_tlen > 0) &&
12744 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
12745 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12746 				/*
12747 				 * DSACK actually handled in the fastpath
12748 				 * above.
12749 				 */
12750 				tcp_update_sack_list(tp, save_start,
12751 				    save_start + save_tlen);
12752 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12753 				if ((tp->rcv_numsacks >= 1) &&
12754 				    (tp->sackblks[0].end == save_start)) {
12755 					/*
12756 					 * Partial overlap, recorded at todrop
12757 					 * above.
12758 					 */
12759 					tcp_update_sack_list(tp,
12760 					    tp->sackblks[0].start,
12761 					    tp->sackblks[0].end);
12762 				} else {
12763 					tcp_update_dsack_list(tp, save_start,
12764 					    save_start + save_tlen);
12765 				}
12766 			} else if (tlen >= save_tlen) {
12767 				/* Update of sackblks. */
12768 				tcp_update_dsack_list(tp, save_start,
12769 				    save_start + save_tlen);
12770 			} else if (tlen > 0) {
12771 				tcp_update_dsack_list(tp, save_start,
12772 				    save_start + tlen);
12773 			}
12774 		}
12775 	} else {
12776 		m_freem(m);
12777 		thflags &= ~TH_FIN;
12778 	}
12779 
12780 	/*
12781 	 * If FIN is received ACK the FIN and let the user know that the
12782 	 * connection is closing.
12783 	 */
12784 	if (thflags & TH_FIN) {
12785 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12786 			/* The socket upcall is handled by socantrcvmore. */
12787 			socantrcvmore(so);
12788 			/*
12789 			 * If connection is half-synchronized (ie NEEDSYN
12790 			 * flag on) then delay ACK, so it may be piggybacked
12791 			 * when SYN is sent. Otherwise, since we received a
12792 			 * FIN then no more input can be expected, send ACK
12793 			 * now.
12794 			 */
12795 			if (tp->t_flags & TF_NEEDSYN) {
12796 				rack_timer_cancel(tp, rack,
12797 				    rack->r_ctl.rc_rcvtime, __LINE__);
12798 				tp->t_flags |= TF_DELACK;
12799 			} else {
12800 				tp->t_flags |= TF_ACKNOW;
12801 			}
12802 			tp->rcv_nxt++;
12803 		}
12804 		switch (tp->t_state) {
12805 			/*
12806 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12807 			 * CLOSE_WAIT state.
12808 			 */
12809 		case TCPS_SYN_RECEIVED:
12810 			tp->t_starttime = ticks;
12811 			/* FALLTHROUGH */
12812 		case TCPS_ESTABLISHED:
12813 			rack_timer_cancel(tp, rack,
12814 			    rack->r_ctl.rc_rcvtime, __LINE__);
12815 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
12816 			break;
12817 
12818 			/*
12819 			 * If still in FIN_WAIT_1 STATE FIN has not been
12820 			 * acked so enter the CLOSING state.
12821 			 */
12822 		case TCPS_FIN_WAIT_1:
12823 			rack_timer_cancel(tp, rack,
12824 			    rack->r_ctl.rc_rcvtime, __LINE__);
12825 			tcp_state_change(tp, TCPS_CLOSING);
12826 			break;
12827 
12828 			/*
12829 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12830 			 * starting the time-wait timer, turning off the
12831 			 * other standard timers.
12832 			 */
12833 		case TCPS_FIN_WAIT_2:
12834 			rack_timer_cancel(tp, rack,
12835 			    rack->r_ctl.rc_rcvtime, __LINE__);
12836 			tcp_twstart(tp);
12837 			return (1);
12838 		}
12839 	}
12840 	/*
12841 	 * Return any desired output.
12842 	 */
12843 	if ((tp->t_flags & TF_ACKNOW) ||
12844 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12845 		rack->r_wanted_output = 1;
12846 	}
12847 	return (0);
12848 }
12849 
12850 /*
12851  * Here nothing is really faster, its just that we
12852  * have broken out the fast-data path also just like
12853  * the fast-ack.
12854  */
12855 static int
12856 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12857     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12858     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12859 {
12860 	int32_t nsegs;
12861 	int32_t newsize = 0;	/* automatic sockbuf scaling */
12862 	struct tcp_rack *rack;
12863 #ifdef NETFLIX_SB_LIMITS
12864 	u_int mcnt, appended;
12865 #endif
12866 
12867 	/*
12868 	 * If last ACK falls within this segment's sequence numbers, record
12869 	 * the timestamp. NOTE that the test is modified according to the
12870 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12871 	 */
12872 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12873 		return (0);
12874 	}
12875 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12876 		return (0);
12877 	}
12878 	if (tiwin && tiwin != tp->snd_wnd) {
12879 		return (0);
12880 	}
12881 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12882 		return (0);
12883 	}
12884 	if (__predict_false((to->to_flags & TOF_TS) &&
12885 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12886 		return (0);
12887 	}
12888 	if (__predict_false((th->th_ack != tp->snd_una))) {
12889 		return (0);
12890 	}
12891 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12892 		return (0);
12893 	}
12894 	if ((to->to_flags & TOF_TS) != 0 &&
12895 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12896 		tp->ts_recent_age = tcp_ts_getticks();
12897 		tp->ts_recent = to->to_tsval;
12898 	}
12899 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12900 	/*
12901 	 * This is a pure, in-sequence data packet with nothing on the
12902 	 * reassembly queue and we have enough buffer space to take it.
12903 	 */
12904 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12905 
12906 #ifdef NETFLIX_SB_LIMITS
12907 	if (so->so_rcv.sb_shlim) {
12908 		mcnt = m_memcnt(m);
12909 		appended = 0;
12910 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12911 		    CFO_NOSLEEP, NULL) == false) {
12912 			counter_u64_add(tcp_sb_shlim_fails, 1);
12913 			m_freem(m);
12914 			return (1);
12915 		}
12916 	}
12917 #endif
12918 	/* Clean receiver SACK report if present */
12919 	if (tp->rcv_numsacks)
12920 		tcp_clean_sackreport(tp);
12921 	KMOD_TCPSTAT_INC(tcps_preddat);
12922 	tp->rcv_nxt += tlen;
12923 	if (tlen &&
12924 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12925 	    (tp->t_fbyte_in == 0)) {
12926 		tp->t_fbyte_in = ticks;
12927 		if (tp->t_fbyte_in == 0)
12928 			tp->t_fbyte_in = 1;
12929 		if (tp->t_fbyte_out && tp->t_fbyte_in)
12930 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12931 	}
12932 	/*
12933 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12934 	 */
12935 	tp->snd_wl1 = th->th_seq;
12936 	/*
12937 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12938 	 */
12939 	tp->rcv_up = tp->rcv_nxt;
12940 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12941 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12942 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12943 
12944 	/* Add data to socket buffer. */
12945 	SOCKBUF_LOCK(&so->so_rcv);
12946 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12947 		m_freem(m);
12948 	} else {
12949 		/*
12950 		 * Set new socket buffer size. Give up when limit is
12951 		 * reached.
12952 		 */
12953 		if (newsize)
12954 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12955 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12956 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12957 #ifdef NETFLIX_SB_LIMITS
12958 		appended =
12959 #endif
12960 			sbappendstream_locked(&so->so_rcv, m, 0);
12961 		ctf_calc_rwin(so, tp);
12962 	}
12963 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12964 	/* NB: sorwakeup_locked() does an implicit unlock. */
12965 	sorwakeup_locked(so);
12966 #ifdef NETFLIX_SB_LIMITS
12967 	if (so->so_rcv.sb_shlim && mcnt != appended)
12968 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12969 #endif
12970 	rack_handle_delayed_ack(tp, rack, tlen, 0);
12971 	if (tp->snd_una == tp->snd_max)
12972 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12973 	return (1);
12974 }
12975 
12976 /*
12977  * This subfunction is used to try to highly optimize the
12978  * fast path. We again allow window updates that are
12979  * in sequence to remain in the fast-path. We also add
12980  * in the __predict's to attempt to help the compiler.
12981  * Note that if we return a 0, then we can *not* process
12982  * it and the caller should push the packet into the
12983  * slow-path.
12984  */
12985 static int
12986 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12987     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12988     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12989 {
12990 	int32_t acked;
12991 	int32_t nsegs;
12992 	int32_t under_pacing = 0;
12993 	struct tcp_rack *rack;
12994 
12995 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12996 		/* Old ack, behind (or duplicate to) the last one rcv'd */
12997 		return (0);
12998 	}
12999 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
13000 		/* Above what we have sent? */
13001 		return (0);
13002 	}
13003 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
13004 		/* We are retransmitting */
13005 		return (0);
13006 	}
13007 	if (__predict_false(tiwin == 0)) {
13008 		/* zero window */
13009 		return (0);
13010 	}
13011 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
13012 		/* We need a SYN or a FIN, unlikely.. */
13013 		return (0);
13014 	}
13015 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
13016 		/* Timestamp is behind .. old ack with seq wrap? */
13017 		return (0);
13018 	}
13019 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
13020 		/* Still recovering */
13021 		return (0);
13022 	}
13023 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13024 	if (rack->r_ctl.rc_sacked) {
13025 		/* We have sack holes on our scoreboard */
13026 		return (0);
13027 	}
13028 	/* Ok if we reach here, we can process a fast-ack */
13029 	if (rack->gp_ready &&
13030 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13031 		under_pacing = 1;
13032 	}
13033 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13034 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
13035 	/* Did the window get updated? */
13036 	if (tiwin != tp->snd_wnd) {
13037 		tp->snd_wnd = tiwin;
13038 		rack_validate_fo_sendwin_up(tp, rack);
13039 		tp->snd_wl1 = th->th_seq;
13040 		if (tp->snd_wnd > tp->max_sndwnd)
13041 			tp->max_sndwnd = tp->snd_wnd;
13042 	}
13043 	/* Do we exit persists? */
13044 	if ((rack->rc_in_persist != 0) &&
13045 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13046 			       rack->r_ctl.rc_pace_min_segs))) {
13047 		rack_exit_persist(tp, rack, cts);
13048 	}
13049 	/* Do we enter persists? */
13050 	if ((rack->rc_in_persist == 0) &&
13051 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13052 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13053 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13054 	    sbavail(&tptosocket(tp)->so_snd) &&
13055 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13056 		/*
13057 		 * Here the rwnd is less than
13058 		 * the pacing size, we are established,
13059 		 * nothing is outstanding, and there is
13060 		 * data to send. Enter persists.
13061 		 */
13062 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
13063 	}
13064 	/*
13065 	 * If last ACK falls within this segment's sequence numbers, record
13066 	 * the timestamp. NOTE that the test is modified according to the
13067 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
13068 	 */
13069 	if ((to->to_flags & TOF_TS) != 0 &&
13070 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
13071 		tp->ts_recent_age = tcp_ts_getticks();
13072 		tp->ts_recent = to->to_tsval;
13073 	}
13074 	/*
13075 	 * This is a pure ack for outstanding data.
13076 	 */
13077 	KMOD_TCPSTAT_INC(tcps_predack);
13078 
13079 	/*
13080 	 * "bad retransmit" recovery.
13081 	 */
13082 	if ((tp->t_flags & TF_PREVVALID) &&
13083 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13084 		tp->t_flags &= ~TF_PREVVALID;
13085 		if (tp->t_rxtshift == 1 &&
13086 		    (int)(ticks - tp->t_badrxtwin) < 0)
13087 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
13088 	}
13089 	/*
13090 	 * Recalculate the transmit timer / rtt.
13091 	 *
13092 	 * Some boxes send broken timestamp replies during the SYN+ACK
13093 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
13094 	 * and blow up the retransmit timer.
13095 	 */
13096 	acked = BYTES_THIS_ACK(tp, th);
13097 
13098 #ifdef TCP_HHOOK
13099 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
13100 	hhook_run_tcp_est_in(tp, th, to);
13101 #endif
13102 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
13103 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13104 	if (acked) {
13105 		struct mbuf *mfree;
13106 
13107 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
13108 		SOCKBUF_LOCK(&so->so_snd);
13109 		mfree = sbcut_locked(&so->so_snd, acked);
13110 		tp->snd_una = th->th_ack;
13111 		/* Note we want to hold the sb lock through the sendmap adjust */
13112 		rack_adjust_sendmap_head(rack, &so->so_snd);
13113 		/* Wake up the socket if we have room to write more */
13114 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13115 		sowwakeup_locked(so);
13116 		m_freem(mfree);
13117 		tp->t_rxtshift = 0;
13118 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13119 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13120 		rack->rc_tlp_in_progress = 0;
13121 		rack->r_ctl.rc_tlp_cnt_out = 0;
13122 		/*
13123 		 * If it is the RXT timer we want to
13124 		 * stop it, so we can restart a TLP.
13125 		 */
13126 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13127 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13128 
13129 #ifdef TCP_REQUEST_TRK
13130 		rack_req_check_for_comp(rack, th->th_ack);
13131 #endif
13132 	}
13133 	/*
13134 	 * Let the congestion control algorithm update congestion control
13135 	 * related information. This typically means increasing the
13136 	 * congestion window.
13137 	 */
13138 	if (tp->snd_wnd < ctf_outstanding(tp)) {
13139 		/* The peer collapsed the window */
13140 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
13141 	} else if (rack->rc_has_collapsed)
13142 		rack_un_collapse_window(rack, __LINE__);
13143 	if ((rack->r_collapse_point_valid) &&
13144 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
13145 		rack->r_collapse_point_valid = 0;
13146 	/*
13147 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
13148 	 */
13149 	tp->snd_wl2 = th->th_ack;
13150 	tp->t_dupacks = 0;
13151 	m_freem(m);
13152 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
13153 
13154 	/*
13155 	 * If all outstanding data are acked, stop retransmit timer,
13156 	 * otherwise restart timer using current (possibly backed-off)
13157 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
13158 	 * If data are ready to send, let tcp_output decide between more
13159 	 * output or persist.
13160 	 */
13161 	if (under_pacing &&
13162 	    (rack->use_fixed_rate == 0) &&
13163 	    (rack->in_probe_rtt == 0) &&
13164 	    rack->rc_gp_dyn_mul &&
13165 	    rack->rc_always_pace) {
13166 		/* Check if we are dragging bottom */
13167 		rack_check_bottom_drag(tp, rack, so);
13168 	}
13169 	if (tp->snd_una == tp->snd_max) {
13170 		tp->t_flags &= ~TF_PREVVALID;
13171 		rack->r_ctl.retran_during_recovery = 0;
13172 		rack->rc_suspicious = 0;
13173 		rack->r_ctl.dsack_byte_cnt = 0;
13174 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13175 		if (rack->r_ctl.rc_went_idle_time == 0)
13176 			rack->r_ctl.rc_went_idle_time = 1;
13177 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13178 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
13179 			tp->t_acktime = 0;
13180 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13181 	}
13182 	if (acked && rack->r_fast_output)
13183 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
13184 	if (sbavail(&so->so_snd)) {
13185 		rack->r_wanted_output = 1;
13186 	}
13187 	return (1);
13188 }
13189 
13190 /*
13191  * Return value of 1, the TCB is unlocked and most
13192  * likely gone, return value of 0, the TCP is still
13193  * locked.
13194  */
13195 static int
13196 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
13197     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13198     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13199 {
13200 	int32_t ret_val = 0;
13201 	int32_t todrop;
13202 	int32_t ourfinisacked = 0;
13203 	struct tcp_rack *rack;
13204 
13205 	INP_WLOCK_ASSERT(tptoinpcb(tp));
13206 
13207 	ctf_calc_rwin(so, tp);
13208 	/*
13209 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
13210 	 * SYN, drop the input. if seg contains a RST, then drop the
13211 	 * connection. if seg does not contain SYN, then drop it. Otherwise
13212 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
13213 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
13214 	 * contains an ECE and ECN support is enabled, the stream is ECN
13215 	 * capable. if SYN has been acked change to ESTABLISHED else
13216 	 * SYN_RCVD state arrange for segment to be acked (eventually)
13217 	 * continue processing rest of data/controls.
13218 	 */
13219 	if ((thflags & TH_ACK) &&
13220 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
13221 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13222 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13223 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13224 		return (1);
13225 	}
13226 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
13227 		TCP_PROBE5(connect__refused, NULL, tp,
13228 		    mtod(m, const char *), tp, th);
13229 		tp = tcp_drop(tp, ECONNREFUSED);
13230 		ctf_do_drop(m, tp);
13231 		return (1);
13232 	}
13233 	if (thflags & TH_RST) {
13234 		ctf_do_drop(m, tp);
13235 		return (1);
13236 	}
13237 	if (!(thflags & TH_SYN)) {
13238 		ctf_do_drop(m, tp);
13239 		return (1);
13240 	}
13241 	tp->irs = th->th_seq;
13242 	tcp_rcvseqinit(tp);
13243 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13244 	if (thflags & TH_ACK) {
13245 		int tfo_partial = 0;
13246 
13247 		KMOD_TCPSTAT_INC(tcps_connects);
13248 		soisconnected(so);
13249 #ifdef MAC
13250 		mac_socketpeer_set_from_mbuf(m, so);
13251 #endif
13252 		/* Do window scaling on this connection? */
13253 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13254 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13255 			tp->rcv_scale = tp->request_r_scale;
13256 		}
13257 		tp->rcv_adv += min(tp->rcv_wnd,
13258 		    TCP_MAXWIN << tp->rcv_scale);
13259 		/*
13260 		 * If not all the data that was sent in the TFO SYN
13261 		 * has been acked, resend the remainder right away.
13262 		 */
13263 		if (IS_FASTOPEN(tp->t_flags) &&
13264 		    (tp->snd_una != tp->snd_max)) {
13265 			tp->snd_nxt = th->th_ack;
13266 			tfo_partial = 1;
13267 		}
13268 		/*
13269 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
13270 		 * will be turned on later.
13271 		 */
13272 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
13273 			rack_timer_cancel(tp, rack,
13274 					  rack->r_ctl.rc_rcvtime, __LINE__);
13275 			tp->t_flags |= TF_DELACK;
13276 		} else {
13277 			rack->r_wanted_output = 1;
13278 			tp->t_flags |= TF_ACKNOW;
13279 		}
13280 
13281 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
13282 
13283 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
13284 			/*
13285 			 * We advance snd_una for the
13286 			 * fast open case. If th_ack is
13287 			 * acknowledging data beyond
13288 			 * snd_una we can't just call
13289 			 * ack-processing since the
13290 			 * data stream in our send-map
13291 			 * will start at snd_una + 1 (one
13292 			 * beyond the SYN). If its just
13293 			 * equal we don't need to do that
13294 			 * and there is no send_map.
13295 			 */
13296 			tp->snd_una++;
13297 		}
13298 		/*
13299 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
13300 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
13301 		 */
13302 		tp->t_starttime = ticks;
13303 		if (tp->t_flags & TF_NEEDFIN) {
13304 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
13305 			tp->t_flags &= ~TF_NEEDFIN;
13306 			thflags &= ~TH_SYN;
13307 		} else {
13308 			tcp_state_change(tp, TCPS_ESTABLISHED);
13309 			TCP_PROBE5(connect__established, NULL, tp,
13310 			    mtod(m, const char *), tp, th);
13311 			rack_cc_conn_init(tp);
13312 		}
13313 	} else {
13314 		/*
13315 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
13316 		 * open.  If segment contains CC option and there is a
13317 		 * cached CC, apply TAO test. If it succeeds, connection is *
13318 		 * half-synchronized. Otherwise, do 3-way handshake:
13319 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13320 		 * there was no CC option, clear cached CC value.
13321 		 */
13322 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13323 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
13324 	}
13325 	/*
13326 	 * Advance th->th_seq to correspond to first data byte. If data,
13327 	 * trim to stay within window, dropping FIN if necessary.
13328 	 */
13329 	th->th_seq++;
13330 	if (tlen > tp->rcv_wnd) {
13331 		todrop = tlen - tp->rcv_wnd;
13332 		m_adj(m, -todrop);
13333 		tlen = tp->rcv_wnd;
13334 		thflags &= ~TH_FIN;
13335 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13336 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13337 	}
13338 	tp->snd_wl1 = th->th_seq - 1;
13339 	tp->rcv_up = th->th_seq;
13340 	/*
13341 	 * Client side of transaction: already sent SYN and data. If the
13342 	 * remote host used T/TCP to validate the SYN, our data will be
13343 	 * ACK'd; if so, enter normal data segment processing in the middle
13344 	 * of step 5, ack processing. Otherwise, goto step 6.
13345 	 */
13346 	if (thflags & TH_ACK) {
13347 		/* For syn-sent we need to possibly update the rtt */
13348 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13349 			uint32_t t, mcts;
13350 
13351 			mcts = tcp_ts_getticks();
13352 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13353 			if (!tp->t_rttlow || tp->t_rttlow > t)
13354 				tp->t_rttlow = t;
13355 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13356 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13357 			tcp_rack_xmit_timer_commit(rack, tp);
13358 		}
13359 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
13360 			return (ret_val);
13361 		/* We may have changed to FIN_WAIT_1 above */
13362 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13363 			/*
13364 			 * In FIN_WAIT_1 STATE in addition to the processing
13365 			 * for the ESTABLISHED state if our FIN is now
13366 			 * acknowledged then enter FIN_WAIT_2.
13367 			 */
13368 			if (ourfinisacked) {
13369 				/*
13370 				 * If we can't receive any more data, then
13371 				 * closing user can proceed. Starting the
13372 				 * timer is contrary to the specification,
13373 				 * but if we don't get a FIN we'll hang
13374 				 * forever.
13375 				 *
13376 				 * XXXjl: we should release the tp also, and
13377 				 * use a compressed state.
13378 				 */
13379 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13380 					soisdisconnected(so);
13381 					tcp_timer_activate(tp, TT_2MSL,
13382 					    (tcp_fast_finwait2_recycle ?
13383 					    tcp_finwait2_timeout :
13384 					    TP_MAXIDLE(tp)));
13385 				}
13386 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13387 			}
13388 		}
13389 	}
13390 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13391 	   tiwin, thflags, nxt_pkt));
13392 }
13393 
13394 /*
13395  * Return value of 1, the TCB is unlocked and most
13396  * likely gone, return value of 0, the TCP is still
13397  * locked.
13398  */
13399 static int
13400 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13401     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13402     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13403 {
13404 	struct tcp_rack *rack;
13405 	int32_t ret_val = 0;
13406 	int32_t ourfinisacked = 0;
13407 
13408 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13409 	ctf_calc_rwin(so, tp);
13410 	if ((thflags & TH_RST) ||
13411 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13412 		return (__ctf_process_rst(m, th, so, tp,
13413 					  &rack->r_ctl.challenge_ack_ts,
13414 					  &rack->r_ctl.challenge_ack_cnt));
13415 	if ((thflags & TH_ACK) &&
13416 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13417 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13418 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13419 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13420 		return (1);
13421 	}
13422 	if (IS_FASTOPEN(tp->t_flags)) {
13423 		/*
13424 		 * When a TFO connection is in SYN_RECEIVED, the
13425 		 * only valid packets are the initial SYN, a
13426 		 * retransmit/copy of the initial SYN (possibly with
13427 		 * a subset of the original data), a valid ACK, a
13428 		 * FIN, or a RST.
13429 		 */
13430 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13431 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13432 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13433 			return (1);
13434 		} else if (thflags & TH_SYN) {
13435 			/* non-initial SYN is ignored */
13436 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13437 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13438 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13439 				ctf_do_drop(m, NULL);
13440 				return (0);
13441 			}
13442 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13443 			ctf_do_drop(m, NULL);
13444 			return (0);
13445 		}
13446 	}
13447 
13448 	/*
13449 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13450 	 * it's less than ts_recent, drop it.
13451 	 */
13452 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13453 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13454 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13455 			return (ret_val);
13456 	}
13457 	/*
13458 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13459 	 * this connection before trimming the data to fit the receive
13460 	 * window.  Check the sequence number versus IRS since we know the
13461 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13462 	 * "LAND" DoS attack.
13463 	 */
13464 	if (SEQ_LT(th->th_seq, tp->irs)) {
13465 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13466 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13467 		return (1);
13468 	}
13469 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13470 			      &rack->r_ctl.challenge_ack_ts,
13471 			      &rack->r_ctl.challenge_ack_cnt)) {
13472 		return (ret_val);
13473 	}
13474 	/*
13475 	 * If last ACK falls within this segment's sequence numbers, record
13476 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13477 	 * from the latest proposal of the tcplw@cray.com list (Braden
13478 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13479 	 * with our earlier PAWS tests, so this check should be solely
13480 	 * predicated on the sequence space of this segment. 3) That we
13481 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13482 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13483 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13484 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13485 	 * p.869. In such cases, we can still calculate the RTT correctly
13486 	 * when RCV.NXT == Last.ACK.Sent.
13487 	 */
13488 	if ((to->to_flags & TOF_TS) != 0 &&
13489 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13490 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13491 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13492 		tp->ts_recent_age = tcp_ts_getticks();
13493 		tp->ts_recent = to->to_tsval;
13494 	}
13495 	tp->snd_wnd = tiwin;
13496 	rack_validate_fo_sendwin_up(tp, rack);
13497 	/*
13498 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13499 	 * is on (half-synchronized state), then queue data for later
13500 	 * processing; else drop segment and return.
13501 	 */
13502 	if ((thflags & TH_ACK) == 0) {
13503 		if (IS_FASTOPEN(tp->t_flags)) {
13504 			rack_cc_conn_init(tp);
13505 		}
13506 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13507 		    tiwin, thflags, nxt_pkt));
13508 	}
13509 	KMOD_TCPSTAT_INC(tcps_connects);
13510 	if (tp->t_flags & TF_SONOTCONN) {
13511 		tp->t_flags &= ~TF_SONOTCONN;
13512 		soisconnected(so);
13513 	}
13514 	/* Do window scaling? */
13515 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13516 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13517 		tp->rcv_scale = tp->request_r_scale;
13518 	}
13519 	/*
13520 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13521 	 * FIN-WAIT-1
13522 	 */
13523 	tp->t_starttime = ticks;
13524 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
13525 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13526 		tp->t_tfo_pending = NULL;
13527 	}
13528 	if (tp->t_flags & TF_NEEDFIN) {
13529 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13530 		tp->t_flags &= ~TF_NEEDFIN;
13531 	} else {
13532 		tcp_state_change(tp, TCPS_ESTABLISHED);
13533 		TCP_PROBE5(accept__established, NULL, tp,
13534 		    mtod(m, const char *), tp, th);
13535 		/*
13536 		 * TFO connections call cc_conn_init() during SYN
13537 		 * processing.  Calling it again here for such connections
13538 		 * is not harmless as it would undo the snd_cwnd reduction
13539 		 * that occurs when a TFO SYN|ACK is retransmitted.
13540 		 */
13541 		if (!IS_FASTOPEN(tp->t_flags))
13542 			rack_cc_conn_init(tp);
13543 	}
13544 	/*
13545 	 * Account for the ACK of our SYN prior to
13546 	 * regular ACK processing below, except for
13547 	 * simultaneous SYN, which is handled later.
13548 	 */
13549 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13550 		tp->snd_una++;
13551 	/*
13552 	 * If segment contains data or ACK, will call tcp_reass() later; if
13553 	 * not, do so now to pass queued data to user.
13554 	 */
13555 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13556 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13557 		    (struct mbuf *)0);
13558 		if (tp->t_flags & TF_WAKESOR) {
13559 			tp->t_flags &= ~TF_WAKESOR;
13560 			/* NB: sorwakeup_locked() does an implicit unlock. */
13561 			sorwakeup_locked(so);
13562 		}
13563 	}
13564 	tp->snd_wl1 = th->th_seq - 1;
13565 	/* For syn-recv we need to possibly update the rtt */
13566 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13567 		uint32_t t, mcts;
13568 
13569 		mcts = tcp_ts_getticks();
13570 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13571 		if (!tp->t_rttlow || tp->t_rttlow > t)
13572 			tp->t_rttlow = t;
13573 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13574 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13575 		tcp_rack_xmit_timer_commit(rack, tp);
13576 	}
13577 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13578 		return (ret_val);
13579 	}
13580 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13581 		/* We could have went to FIN_WAIT_1 (or EST) above */
13582 		/*
13583 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13584 		 * ESTABLISHED state if our FIN is now acknowledged then
13585 		 * enter FIN_WAIT_2.
13586 		 */
13587 		if (ourfinisacked) {
13588 			/*
13589 			 * If we can't receive any more data, then closing
13590 			 * user can proceed. Starting the timer is contrary
13591 			 * to the specification, but if we don't get a FIN
13592 			 * we'll hang forever.
13593 			 *
13594 			 * XXXjl: we should release the tp also, and use a
13595 			 * compressed state.
13596 			 */
13597 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13598 				soisdisconnected(so);
13599 				tcp_timer_activate(tp, TT_2MSL,
13600 				    (tcp_fast_finwait2_recycle ?
13601 				    tcp_finwait2_timeout :
13602 				    TP_MAXIDLE(tp)));
13603 			}
13604 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13605 		}
13606 	}
13607 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13608 	    tiwin, thflags, nxt_pkt));
13609 }
13610 
13611 /*
13612  * Return value of 1, the TCB is unlocked and most
13613  * likely gone, return value of 0, the TCP is still
13614  * locked.
13615  */
13616 static int
13617 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13618     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13619     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13620 {
13621 	int32_t ret_val = 0;
13622 	struct tcp_rack *rack;
13623 
13624 	/*
13625 	 * Header prediction: check for the two common cases of a
13626 	 * uni-directional data xfer.  If the packet has no control flags,
13627 	 * is in-sequence, the window didn't change and we're not
13628 	 * retransmitting, it's a candidate.  If the length is zero and the
13629 	 * ack moved forward, we're the sender side of the xfer.  Just free
13630 	 * the data acked & wake any higher level process that was blocked
13631 	 * waiting for space.  If the length is non-zero and the ack didn't
13632 	 * move, we're the receiver side.  If we're getting packets in-order
13633 	 * (the reassembly queue is empty), add the data toc The socket
13634 	 * buffer and note that we need a delayed ack. Make sure that the
13635 	 * hidden state-flags are also off. Since we check for
13636 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13637 	 */
13638 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13639 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13640 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13641 	    __predict_true(SEGQ_EMPTY(tp)) &&
13642 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13643 		if (tlen == 0) {
13644 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13645 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13646 				return (0);
13647 			}
13648 		} else {
13649 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13650 			    tiwin, nxt_pkt, iptos)) {
13651 				return (0);
13652 			}
13653 		}
13654 	}
13655 	ctf_calc_rwin(so, tp);
13656 
13657 	if ((thflags & TH_RST) ||
13658 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13659 		return (__ctf_process_rst(m, th, so, tp,
13660 					  &rack->r_ctl.challenge_ack_ts,
13661 					  &rack->r_ctl.challenge_ack_cnt));
13662 
13663 	/*
13664 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13665 	 * synchronized state.
13666 	 */
13667 	if (thflags & TH_SYN) {
13668 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13669 		return (ret_val);
13670 	}
13671 	/*
13672 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13673 	 * it's less than ts_recent, drop it.
13674 	 */
13675 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13676 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13677 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13678 			return (ret_val);
13679 	}
13680 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13681 			      &rack->r_ctl.challenge_ack_ts,
13682 			      &rack->r_ctl.challenge_ack_cnt)) {
13683 		return (ret_val);
13684 	}
13685 	/*
13686 	 * If last ACK falls within this segment's sequence numbers, record
13687 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13688 	 * from the latest proposal of the tcplw@cray.com list (Braden
13689 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13690 	 * with our earlier PAWS tests, so this check should be solely
13691 	 * predicated on the sequence space of this segment. 3) That we
13692 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13693 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13694 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13695 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13696 	 * p.869. In such cases, we can still calculate the RTT correctly
13697 	 * when RCV.NXT == Last.ACK.Sent.
13698 	 */
13699 	if ((to->to_flags & TOF_TS) != 0 &&
13700 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13701 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13702 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13703 		tp->ts_recent_age = tcp_ts_getticks();
13704 		tp->ts_recent = to->to_tsval;
13705 	}
13706 	/*
13707 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13708 	 * is on (half-synchronized state), then queue data for later
13709 	 * processing; else drop segment and return.
13710 	 */
13711 	if ((thflags & TH_ACK) == 0) {
13712 		if (tp->t_flags & TF_NEEDSYN) {
13713 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13714 			    tiwin, thflags, nxt_pkt));
13715 
13716 		} else if (tp->t_flags & TF_ACKNOW) {
13717 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13718 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13719 			return (ret_val);
13720 		} else {
13721 			ctf_do_drop(m, NULL);
13722 			return (0);
13723 		}
13724 	}
13725 	/*
13726 	 * Ack processing.
13727 	 */
13728 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13729 		return (ret_val);
13730 	}
13731 	if (sbavail(&so->so_snd)) {
13732 		if (ctf_progress_timeout_check(tp, true)) {
13733 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13734 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13735 			return (1);
13736 		}
13737 	}
13738 	/* State changes only happen in rack_process_data() */
13739 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13740 	    tiwin, thflags, nxt_pkt));
13741 }
13742 
13743 /*
13744  * Return value of 1, the TCB is unlocked and most
13745  * likely gone, return value of 0, the TCP is still
13746  * locked.
13747  */
13748 static int
13749 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13750     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13751     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13752 {
13753 	int32_t ret_val = 0;
13754 	struct tcp_rack *rack;
13755 
13756 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13757 	ctf_calc_rwin(so, tp);
13758 	if ((thflags & TH_RST) ||
13759 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13760 		return (__ctf_process_rst(m, th, so, tp,
13761 					  &rack->r_ctl.challenge_ack_ts,
13762 					  &rack->r_ctl.challenge_ack_cnt));
13763 	/*
13764 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13765 	 * synchronized state.
13766 	 */
13767 	if (thflags & TH_SYN) {
13768 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13769 		return (ret_val);
13770 	}
13771 	/*
13772 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13773 	 * it's less than ts_recent, drop it.
13774 	 */
13775 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13776 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13777 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13778 			return (ret_val);
13779 	}
13780 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13781 			      &rack->r_ctl.challenge_ack_ts,
13782 			      &rack->r_ctl.challenge_ack_cnt)) {
13783 		return (ret_val);
13784 	}
13785 	/*
13786 	 * If last ACK falls within this segment's sequence numbers, record
13787 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13788 	 * from the latest proposal of the tcplw@cray.com list (Braden
13789 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13790 	 * with our earlier PAWS tests, so this check should be solely
13791 	 * predicated on the sequence space of this segment. 3) That we
13792 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13793 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13794 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13795 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13796 	 * p.869. In such cases, we can still calculate the RTT correctly
13797 	 * when RCV.NXT == Last.ACK.Sent.
13798 	 */
13799 	if ((to->to_flags & TOF_TS) != 0 &&
13800 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13801 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13802 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13803 		tp->ts_recent_age = tcp_ts_getticks();
13804 		tp->ts_recent = to->to_tsval;
13805 	}
13806 	/*
13807 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13808 	 * is on (half-synchronized state), then queue data for later
13809 	 * processing; else drop segment and return.
13810 	 */
13811 	if ((thflags & TH_ACK) == 0) {
13812 		if (tp->t_flags & TF_NEEDSYN) {
13813 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13814 			    tiwin, thflags, nxt_pkt));
13815 
13816 		} else if (tp->t_flags & TF_ACKNOW) {
13817 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13818 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13819 			return (ret_val);
13820 		} else {
13821 			ctf_do_drop(m, NULL);
13822 			return (0);
13823 		}
13824 	}
13825 	/*
13826 	 * Ack processing.
13827 	 */
13828 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13829 		return (ret_val);
13830 	}
13831 	if (sbavail(&so->so_snd)) {
13832 		if (ctf_progress_timeout_check(tp, true)) {
13833 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13834 						tp, tick, PROGRESS_DROP, __LINE__);
13835 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13836 			return (1);
13837 		}
13838 	}
13839 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13840 	    tiwin, thflags, nxt_pkt));
13841 }
13842 
13843 static int
13844 rack_check_data_after_close(struct mbuf *m,
13845     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13846 {
13847 	struct tcp_rack *rack;
13848 
13849 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13850 	if (rack->rc_allow_data_af_clo == 0) {
13851 	close_now:
13852 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13853 		/* tcp_close will kill the inp pre-log the Reset */
13854 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13855 		tp = tcp_close(tp);
13856 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13857 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13858 		return (1);
13859 	}
13860 	if (sbavail(&so->so_snd) == 0)
13861 		goto close_now;
13862 	/* Ok we allow data that is ignored and a followup reset */
13863 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13864 	tp->rcv_nxt = th->th_seq + *tlen;
13865 	tp->t_flags2 |= TF2_DROP_AF_DATA;
13866 	rack->r_wanted_output = 1;
13867 	*tlen = 0;
13868 	return (0);
13869 }
13870 
13871 /*
13872  * Return value of 1, the TCB is unlocked and most
13873  * likely gone, return value of 0, the TCP is still
13874  * locked.
13875  */
13876 static int
13877 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13878     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13879     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13880 {
13881 	int32_t ret_val = 0;
13882 	int32_t ourfinisacked = 0;
13883 	struct tcp_rack *rack;
13884 
13885 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13886 	ctf_calc_rwin(so, tp);
13887 
13888 	if ((thflags & TH_RST) ||
13889 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13890 		return (__ctf_process_rst(m, th, so, tp,
13891 					  &rack->r_ctl.challenge_ack_ts,
13892 					  &rack->r_ctl.challenge_ack_cnt));
13893 	/*
13894 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13895 	 * synchronized state.
13896 	 */
13897 	if (thflags & TH_SYN) {
13898 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13899 		return (ret_val);
13900 	}
13901 	/*
13902 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13903 	 * it's less than ts_recent, drop it.
13904 	 */
13905 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13906 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13907 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13908 			return (ret_val);
13909 	}
13910 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13911 			      &rack->r_ctl.challenge_ack_ts,
13912 			      &rack->r_ctl.challenge_ack_cnt)) {
13913 		return (ret_val);
13914 	}
13915 	/*
13916 	 * If new data are received on a connection after the user processes
13917 	 * are gone, then RST the other end.
13918 	 */
13919 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13920 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13921 		return (1);
13922 	/*
13923 	 * If last ACK falls within this segment's sequence numbers, record
13924 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13925 	 * from the latest proposal of the tcplw@cray.com list (Braden
13926 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13927 	 * with our earlier PAWS tests, so this check should be solely
13928 	 * predicated on the sequence space of this segment. 3) That we
13929 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13930 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13931 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13932 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13933 	 * p.869. In such cases, we can still calculate the RTT correctly
13934 	 * when RCV.NXT == Last.ACK.Sent.
13935 	 */
13936 	if ((to->to_flags & TOF_TS) != 0 &&
13937 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13938 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13939 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13940 		tp->ts_recent_age = tcp_ts_getticks();
13941 		tp->ts_recent = to->to_tsval;
13942 	}
13943 	/*
13944 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13945 	 * is on (half-synchronized state), then queue data for later
13946 	 * processing; else drop segment and return.
13947 	 */
13948 	if ((thflags & TH_ACK) == 0) {
13949 		if (tp->t_flags & TF_NEEDSYN) {
13950 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13951 			    tiwin, thflags, nxt_pkt));
13952 		} else if (tp->t_flags & TF_ACKNOW) {
13953 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13954 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13955 			return (ret_val);
13956 		} else {
13957 			ctf_do_drop(m, NULL);
13958 			return (0);
13959 		}
13960 	}
13961 	/*
13962 	 * Ack processing.
13963 	 */
13964 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13965 		return (ret_val);
13966 	}
13967 	if (ourfinisacked) {
13968 		/*
13969 		 * If we can't receive any more data, then closing user can
13970 		 * proceed. Starting the timer is contrary to the
13971 		 * specification, but if we don't get a FIN we'll hang
13972 		 * forever.
13973 		 *
13974 		 * XXXjl: we should release the tp also, and use a
13975 		 * compressed state.
13976 		 */
13977 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13978 			soisdisconnected(so);
13979 			tcp_timer_activate(tp, TT_2MSL,
13980 			    (tcp_fast_finwait2_recycle ?
13981 			    tcp_finwait2_timeout :
13982 			    TP_MAXIDLE(tp)));
13983 		}
13984 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
13985 	}
13986 	if (sbavail(&so->so_snd)) {
13987 		if (ctf_progress_timeout_check(tp, true)) {
13988 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13989 						tp, tick, PROGRESS_DROP, __LINE__);
13990 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13991 			return (1);
13992 		}
13993 	}
13994 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13995 	    tiwin, thflags, nxt_pkt));
13996 }
13997 
13998 /*
13999  * Return value of 1, the TCB is unlocked and most
14000  * likely gone, return value of 0, the TCP is still
14001  * locked.
14002  */
14003 static int
14004 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
14005     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14006     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14007 {
14008 	int32_t ret_val = 0;
14009 	int32_t ourfinisacked = 0;
14010 	struct tcp_rack *rack;
14011 
14012 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14013 	ctf_calc_rwin(so, tp);
14014 
14015 	if ((thflags & TH_RST) ||
14016 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14017 		return (__ctf_process_rst(m, th, so, tp,
14018 					  &rack->r_ctl.challenge_ack_ts,
14019 					  &rack->r_ctl.challenge_ack_cnt));
14020 	/*
14021 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14022 	 * synchronized state.
14023 	 */
14024 	if (thflags & TH_SYN) {
14025 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14026 		return (ret_val);
14027 	}
14028 	/*
14029 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14030 	 * it's less than ts_recent, drop it.
14031 	 */
14032 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14033 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14034 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14035 			return (ret_val);
14036 	}
14037 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14038 			      &rack->r_ctl.challenge_ack_ts,
14039 			      &rack->r_ctl.challenge_ack_cnt)) {
14040 		return (ret_val);
14041 	}
14042 	/*
14043 	 * If new data are received on a connection after the user processes
14044 	 * are gone, then RST the other end.
14045 	 */
14046 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14047 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14048 		return (1);
14049 	/*
14050 	 * If last ACK falls within this segment's sequence numbers, record
14051 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14052 	 * from the latest proposal of the tcplw@cray.com list (Braden
14053 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14054 	 * with our earlier PAWS tests, so this check should be solely
14055 	 * predicated on the sequence space of this segment. 3) That we
14056 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14057 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14058 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14059 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14060 	 * p.869. In such cases, we can still calculate the RTT correctly
14061 	 * when RCV.NXT == Last.ACK.Sent.
14062 	 */
14063 	if ((to->to_flags & TOF_TS) != 0 &&
14064 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14065 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14066 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14067 		tp->ts_recent_age = tcp_ts_getticks();
14068 		tp->ts_recent = to->to_tsval;
14069 	}
14070 	/*
14071 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14072 	 * is on (half-synchronized state), then queue data for later
14073 	 * processing; else drop segment and return.
14074 	 */
14075 	if ((thflags & TH_ACK) == 0) {
14076 		if (tp->t_flags & TF_NEEDSYN) {
14077 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14078 			    tiwin, thflags, nxt_pkt));
14079 		} else if (tp->t_flags & TF_ACKNOW) {
14080 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14081 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14082 			return (ret_val);
14083 		} else {
14084 			ctf_do_drop(m, NULL);
14085 			return (0);
14086 		}
14087 	}
14088 	/*
14089 	 * Ack processing.
14090 	 */
14091 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14092 		return (ret_val);
14093 	}
14094 	if (ourfinisacked) {
14095 		tcp_twstart(tp);
14096 		m_freem(m);
14097 		return (1);
14098 	}
14099 	if (sbavail(&so->so_snd)) {
14100 		if (ctf_progress_timeout_check(tp, true)) {
14101 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14102 						tp, tick, PROGRESS_DROP, __LINE__);
14103 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14104 			return (1);
14105 		}
14106 	}
14107 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14108 	    tiwin, thflags, nxt_pkt));
14109 }
14110 
14111 /*
14112  * Return value of 1, the TCB is unlocked and most
14113  * likely gone, return value of 0, the TCP is still
14114  * locked.
14115  */
14116 static int
14117 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
14118     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14119     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14120 {
14121 	int32_t ret_val = 0;
14122 	int32_t ourfinisacked = 0;
14123 	struct tcp_rack *rack;
14124 
14125 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14126 	ctf_calc_rwin(so, tp);
14127 
14128 	if ((thflags & TH_RST) ||
14129 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14130 		return (__ctf_process_rst(m, th, so, tp,
14131 					  &rack->r_ctl.challenge_ack_ts,
14132 					  &rack->r_ctl.challenge_ack_cnt));
14133 	/*
14134 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14135 	 * synchronized state.
14136 	 */
14137 	if (thflags & TH_SYN) {
14138 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14139 		return (ret_val);
14140 	}
14141 	/*
14142 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14143 	 * it's less than ts_recent, drop it.
14144 	 */
14145 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14146 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14147 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14148 			return (ret_val);
14149 	}
14150 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14151 			      &rack->r_ctl.challenge_ack_ts,
14152 			      &rack->r_ctl.challenge_ack_cnt)) {
14153 		return (ret_val);
14154 	}
14155 	/*
14156 	 * If new data are received on a connection after the user processes
14157 	 * are gone, then RST the other end.
14158 	 */
14159 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14160 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14161 		return (1);
14162 	/*
14163 	 * If last ACK falls within this segment's sequence numbers, record
14164 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14165 	 * from the latest proposal of the tcplw@cray.com list (Braden
14166 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14167 	 * with our earlier PAWS tests, so this check should be solely
14168 	 * predicated on the sequence space of this segment. 3) That we
14169 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14170 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14171 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14172 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14173 	 * p.869. In such cases, we can still calculate the RTT correctly
14174 	 * when RCV.NXT == Last.ACK.Sent.
14175 	 */
14176 	if ((to->to_flags & TOF_TS) != 0 &&
14177 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14178 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14179 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14180 		tp->ts_recent_age = tcp_ts_getticks();
14181 		tp->ts_recent = to->to_tsval;
14182 	}
14183 	/*
14184 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14185 	 * is on (half-synchronized state), then queue data for later
14186 	 * processing; else drop segment and return.
14187 	 */
14188 	if ((thflags & TH_ACK) == 0) {
14189 		if (tp->t_flags & TF_NEEDSYN) {
14190 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14191 			    tiwin, thflags, nxt_pkt));
14192 		} else if (tp->t_flags & TF_ACKNOW) {
14193 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14194 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14195 			return (ret_val);
14196 		} else {
14197 			ctf_do_drop(m, NULL);
14198 			return (0);
14199 		}
14200 	}
14201 	/*
14202 	 * case TCPS_LAST_ACK: Ack processing.
14203 	 */
14204 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14205 		return (ret_val);
14206 	}
14207 	if (ourfinisacked) {
14208 		tp = tcp_close(tp);
14209 		ctf_do_drop(m, tp);
14210 		return (1);
14211 	}
14212 	if (sbavail(&so->so_snd)) {
14213 		if (ctf_progress_timeout_check(tp, true)) {
14214 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14215 						tp, tick, PROGRESS_DROP, __LINE__);
14216 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14217 			return (1);
14218 		}
14219 	}
14220 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14221 	    tiwin, thflags, nxt_pkt));
14222 }
14223 
14224 /*
14225  * Return value of 1, the TCB is unlocked and most
14226  * likely gone, return value of 0, the TCP is still
14227  * locked.
14228  */
14229 static int
14230 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
14231     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14232     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14233 {
14234 	int32_t ret_val = 0;
14235 	int32_t ourfinisacked = 0;
14236 	struct tcp_rack *rack;
14237 
14238 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14239 	ctf_calc_rwin(so, tp);
14240 
14241 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
14242 	if ((thflags & TH_RST) ||
14243 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14244 		return (__ctf_process_rst(m, th, so, tp,
14245 					  &rack->r_ctl.challenge_ack_ts,
14246 					  &rack->r_ctl.challenge_ack_cnt));
14247 	/*
14248 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14249 	 * synchronized state.
14250 	 */
14251 	if (thflags & TH_SYN) {
14252 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14253 		return (ret_val);
14254 	}
14255 	/*
14256 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14257 	 * it's less than ts_recent, drop it.
14258 	 */
14259 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14260 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14261 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14262 			return (ret_val);
14263 	}
14264 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14265 			      &rack->r_ctl.challenge_ack_ts,
14266 			      &rack->r_ctl.challenge_ack_cnt)) {
14267 		return (ret_val);
14268 	}
14269 	/*
14270 	 * If new data are received on a connection after the user processes
14271 	 * are gone, then RST the other end.
14272 	 */
14273 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14274 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14275 		return (1);
14276 	/*
14277 	 * If last ACK falls within this segment's sequence numbers, record
14278 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14279 	 * from the latest proposal of the tcplw@cray.com list (Braden
14280 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14281 	 * with our earlier PAWS tests, so this check should be solely
14282 	 * predicated on the sequence space of this segment. 3) That we
14283 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14284 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14285 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14286 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14287 	 * p.869. In such cases, we can still calculate the RTT correctly
14288 	 * when RCV.NXT == Last.ACK.Sent.
14289 	 */
14290 	if ((to->to_flags & TOF_TS) != 0 &&
14291 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14292 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14293 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14294 		tp->ts_recent_age = tcp_ts_getticks();
14295 		tp->ts_recent = to->to_tsval;
14296 	}
14297 	/*
14298 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14299 	 * is on (half-synchronized state), then queue data for later
14300 	 * processing; else drop segment and return.
14301 	 */
14302 	if ((thflags & TH_ACK) == 0) {
14303 		if (tp->t_flags & TF_NEEDSYN) {
14304 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14305 			    tiwin, thflags, nxt_pkt));
14306 		} else if (tp->t_flags & TF_ACKNOW) {
14307 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14308 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14309 			return (ret_val);
14310 		} else {
14311 			ctf_do_drop(m, NULL);
14312 			return (0);
14313 		}
14314 	}
14315 	/*
14316 	 * Ack processing.
14317 	 */
14318 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14319 		return (ret_val);
14320 	}
14321 	if (sbavail(&so->so_snd)) {
14322 		if (ctf_progress_timeout_check(tp, true)) {
14323 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14324 						tp, tick, PROGRESS_DROP, __LINE__);
14325 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14326 			return (1);
14327 		}
14328 	}
14329 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14330 	    tiwin, thflags, nxt_pkt));
14331 }
14332 
14333 static void inline
14334 rack_clear_rate_sample(struct tcp_rack *rack)
14335 {
14336 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
14337 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
14338 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
14339 }
14340 
14341 static void
14342 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
14343 {
14344 	uint64_t bw_est, rate_wanted;
14345 	int chged = 0;
14346 	uint32_t user_max, orig_min, orig_max;
14347 
14348 #ifdef TCP_REQUEST_TRK
14349 	if (rack->rc_hybrid_mode &&
14350 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
14351 	    (rack_hybrid_allow_set_maxseg == 1) &&
14352 	    (rack->r_ctl.rc_last_sft != NULL)) {
14353 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14354 		return;
14355 	}
14356 #endif
14357 	orig_min = rack->r_ctl.rc_pace_min_segs;
14358 	orig_max = rack->r_ctl.rc_pace_max_segs;
14359 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14360 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14361 		chged = 1;
14362 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14363 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14364 		if (user_max != rack->r_ctl.rc_pace_max_segs)
14365 			chged = 1;
14366 	}
14367 	if (rack->rc_force_max_seg) {
14368 		rack->r_ctl.rc_pace_max_segs = user_max;
14369 	} else if (rack->use_fixed_rate) {
14370 		bw_est = rack_get_bw(rack);
14371 		if ((rack->r_ctl.crte == NULL) ||
14372 		    (bw_est != rack->r_ctl.crte->rate)) {
14373 			rack->r_ctl.rc_pace_max_segs = user_max;
14374 		} else {
14375 			/* We are pacing right at the hardware rate */
14376 			uint32_t segsiz, pace_one;
14377 
14378 			if (rack_pace_one_seg ||
14379 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14380 				pace_one = 1;
14381 			else
14382 				pace_one = 0;
14383 			segsiz = min(ctf_fixed_maxseg(tp),
14384 				     rack->r_ctl.rc_pace_min_segs);
14385 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14386 				tp, bw_est, segsiz, pace_one,
14387 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14388 		}
14389 	} else if (rack->rc_always_pace) {
14390 		if (rack->r_ctl.gp_bw ||
14391 		    rack->r_ctl.init_rate) {
14392 			/* We have a rate of some sort set */
14393 			uint32_t  orig;
14394 
14395 			bw_est = rack_get_bw(rack);
14396 			orig = rack->r_ctl.rc_pace_max_segs;
14397 			if (fill_override)
14398 				rate_wanted = *fill_override;
14399 			else
14400 				rate_wanted = rack_get_gp_est(rack);
14401 			if (rate_wanted) {
14402 				/* We have something */
14403 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14404 										   rate_wanted,
14405 										   ctf_fixed_maxseg(rack->rc_tp));
14406 			} else
14407 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14408 			if (orig != rack->r_ctl.rc_pace_max_segs)
14409 				chged = 1;
14410 		} else if ((rack->r_ctl.gp_bw == 0) &&
14411 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14412 			/*
14413 			 * If we have nothing limit us to bursting
14414 			 * out IW sized pieces.
14415 			 */
14416 			chged = 1;
14417 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14418 		}
14419 	}
14420 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14421 		chged = 1;
14422 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14423 	}
14424 	if (chged)
14425 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14426 }
14427 
14428 
14429 static void
14430 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14431 {
14432 #ifdef INET6
14433 	struct ip6_hdr *ip6 = NULL;
14434 #endif
14435 #ifdef INET
14436 	struct ip *ip = NULL;
14437 #endif
14438 	struct udphdr *udp = NULL;
14439 
14440 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14441 #ifdef INET6
14442 	if (rack->r_is_v6) {
14443 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14444 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14445 		if (tp->t_port) {
14446 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14447 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14448 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14449 			udp->uh_dport = tp->t_port;
14450 			rack->r_ctl.fsb.udp = udp;
14451 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14452 		} else
14453 		{
14454 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14455 			rack->r_ctl.fsb.udp = NULL;
14456 		}
14457 		tcpip_fillheaders(rack->rc_inp,
14458 				  tp->t_port,
14459 				  ip6, rack->r_ctl.fsb.th);
14460 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14461 	} else
14462 #endif				/* INET6 */
14463 #ifdef INET
14464 	{
14465 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14466 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14467 		if (tp->t_port) {
14468 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14469 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14470 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14471 			udp->uh_dport = tp->t_port;
14472 			rack->r_ctl.fsb.udp = udp;
14473 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14474 		} else
14475 		{
14476 			rack->r_ctl.fsb.udp = NULL;
14477 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14478 		}
14479 		tcpip_fillheaders(rack->rc_inp,
14480 				  tp->t_port,
14481 				  ip, rack->r_ctl.fsb.th);
14482 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14483 	}
14484 #endif
14485 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14486 	    (long)TCP_MAXWIN << tp->rcv_scale);
14487 	rack->r_fsb_inited = 1;
14488 }
14489 
14490 static int
14491 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14492 {
14493 	/*
14494 	 * Allocate the larger of spaces V6 if available else just
14495 	 * V4 and include udphdr (overbook)
14496 	 */
14497 #ifdef INET6
14498 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14499 #else
14500 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14501 #endif
14502 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14503 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14504 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14505 		return (ENOMEM);
14506 	}
14507 	rack->r_fsb_inited = 0;
14508 	return (0);
14509 }
14510 
14511 static void
14512 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14513 {
14514 	/*
14515 	 * Types of logs (mod value)
14516 	 * 20 - Initial round setup
14517 	 * 21 - Rack declares a new round.
14518 	 */
14519 	struct tcpcb *tp;
14520 
14521 	tp = rack->rc_tp;
14522 	if (tcp_bblogging_on(tp)) {
14523 		union tcp_log_stackspecific log;
14524 		struct timeval tv;
14525 
14526 		memset(&log, 0, sizeof(log));
14527 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14528 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14529 		log.u_bbr.flex3 = high_seq;
14530 		log.u_bbr.flex4 = tp->snd_max;
14531 		log.u_bbr.flex8 = mod;
14532 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14533 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14534 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14535 		TCP_LOG_EVENTP(tp, NULL,
14536 		    &tptosocket(tp)->so_rcv,
14537 		    &tptosocket(tp)->so_snd,
14538 		    TCP_HYSTART, 0,
14539 		    0, &log, false, &tv);
14540 	}
14541 }
14542 
14543 static void
14544 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14545 {
14546 	rack->rack_deferred_inited = 1;
14547 	rack->r_ctl.roundends = tp->snd_max;
14548 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14549 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14550 }
14551 
14552 static void
14553 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14554 {
14555 	/* Retransmit bit controls.
14556 	 *
14557 	 * The setting of these values control one of
14558 	 * three settings you can have and dictate
14559 	 * how rack does retransmissions. Note this
14560 	 * is in *any* mode i.e. pacing on or off DGP
14561 	 * fixed rate pacing, or just bursting rack.
14562 	 *
14563 	 * 1 - Use full sized retransmits i.e. limit
14564 	 *     the size to whatever the pace_max_segments
14565 	 *     size is.
14566 	 *
14567 	 * 2 - Use pacer min granularity as a guide to
14568 	 *     the size combined with the current calculated
14569 	 *     goodput b/w measurement. So for example if
14570 	 *     the goodput is measured at 20Mbps we would
14571 	 *     calculate 8125 (pacer minimum 250usec in
14572 	 *     that b/w) and then round it up to the next
14573 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14574 	 *
14575 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14576 	 *     fall here too if we are setting via rack_init()).
14577 	 *
14578 	 */
14579 	if (ctl == 1) {
14580 		rack->full_size_rxt = 1;
14581 		rack->shape_rxt_to_pacing_min  = 0;
14582 	} else if (ctl == 2) {
14583 		rack->full_size_rxt = 0;
14584 		rack->shape_rxt_to_pacing_min  = 1;
14585 	} else {
14586 		rack->full_size_rxt = 0;
14587 		rack->shape_rxt_to_pacing_min  = 0;
14588 	}
14589 }
14590 
14591 static void
14592 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14593 		  uint32_t flex1,
14594 		  uint32_t flex2,
14595 		  uint32_t flex3)
14596 {
14597 	if (tcp_bblogging_on(rack->rc_tp)) {
14598 		union tcp_log_stackspecific log;
14599 		struct timeval tv;
14600 
14601 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14602 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14603 		log.u_bbr.flex8 = mod;
14604 		log.u_bbr.flex1 = flex1;
14605 		log.u_bbr.flex2 = flex2;
14606 		log.u_bbr.flex3 = flex3;
14607 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14608 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14609 	}
14610 }
14611 
14612 static int
14613 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14614 {
14615 	struct tcp_rack *rack;
14616 	struct rack_sendmap *rsm;
14617 	int i;
14618 
14619 
14620 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14621 	switch (reqr->req) {
14622 	case TCP_QUERY_SENDMAP:
14623 		if ((reqr->req_param == tp->snd_max) ||
14624 		    (tp->snd_max == tp->snd_una)){
14625 			/* Unlikely */
14626 			return (0);
14627 		}
14628 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14629 		if (rsm == NULL) {
14630 			/* Can't find that seq -- unlikely */
14631 			return (0);
14632 		}
14633 		reqr->sendmap_start = rsm->r_start;
14634 		reqr->sendmap_end = rsm->r_end;
14635 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14636 		reqr->sendmap_fas = rsm->r_fas;
14637 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14638 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14639 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14640 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14641 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14642 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14643 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14644 		reqr->sendmap_dupacks = rsm->r_dupack;
14645 		rack_log_chg_info(tp, rack, 1,
14646 				  rsm->r_start,
14647 				  rsm->r_end,
14648 				  rsm->r_flags);
14649 		return(1);
14650 		break;
14651 	case TCP_QUERY_TIMERS_UP:
14652 		if (rack->r_ctl.rc_hpts_flags == 0) {
14653 			/* no timers up */
14654 			return (0);
14655 		}
14656 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14657 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14658 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14659 		}
14660 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14661 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14662 		}
14663 		rack_log_chg_info(tp, rack, 2,
14664 				  rack->r_ctl.rc_hpts_flags,
14665 				  rack->r_ctl.rc_last_output_to,
14666 				  rack->r_ctl.rc_timer_exp);
14667 		return (1);
14668 		break;
14669 	case TCP_QUERY_RACK_TIMES:
14670 		/* Reordering items */
14671 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14672 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14673 		/* Timerstamps and timers */
14674 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14675 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14676 		reqr->rack_rtt = rack->rc_rack_rtt;
14677 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14678 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14679 		/* PRR data */
14680 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
14681 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14682 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14683 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14684 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14685 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14686 		/* TLP and persists info */
14687 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14688 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14689 		if (rack->rc_in_persist) {
14690 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14691 			reqr->rack_in_persist = 1;
14692 		} else {
14693 			reqr->rack_time_went_idle = 0;
14694 			reqr->rack_in_persist = 0;
14695 		}
14696 		if (rack->r_wanted_output)
14697 			reqr->rack_wanted_output = 1;
14698 		else
14699 			reqr->rack_wanted_output = 0;
14700 		return (1);
14701 		break;
14702 	default:
14703 		return (-EINVAL);
14704 	}
14705 }
14706 
14707 static void
14708 rack_switch_failed(struct tcpcb *tp)
14709 {
14710 	/*
14711 	 * This method gets called if a stack switch was
14712 	 * attempted and it failed. We are left
14713 	 * but our hpts timers were stopped and we
14714 	 * need to validate time units and t_flags2.
14715 	 */
14716 	struct tcp_rack *rack;
14717 	struct timeval tv;
14718 	uint32_t cts;
14719 	uint32_t toval;
14720 	struct hpts_diag diag;
14721 
14722 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14723 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14724 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14725 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14726 	else
14727 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14728 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14729 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14730 	if (tp->t_in_hpts > IHPTS_NONE) {
14731 		/* Strange */
14732 		return;
14733 	}
14734 	cts = tcp_get_usecs(&tv);
14735 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14736 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14737 			toval = rack->r_ctl.rc_last_output_to - cts;
14738 		} else {
14739 			/* one slot please */
14740 			toval = HPTS_TICKS_PER_SLOT;
14741 		}
14742 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14743 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14744 			toval = rack->r_ctl.rc_timer_exp - cts;
14745 		} else {
14746 			/* one slot please */
14747 			toval = HPTS_TICKS_PER_SLOT;
14748 		}
14749 	} else
14750 		toval = HPTS_TICKS_PER_SLOT;
14751 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14752 				   __LINE__, &diag);
14753 	rack_log_hpts_diag(rack, cts, &diag, &tv);
14754 }
14755 
14756 static int
14757 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14758 {
14759 	struct rack_sendmap *rsm, *ersm;
14760 	int insret __diagused;
14761 	/*
14762 	 * When initing outstanding, we must be quite careful
14763 	 * to not refer to tp->t_fb_ptr. This has the old rack
14764 	 * pointer in it, not the "new" one (when we are doing
14765 	 * a stack switch).
14766 	 */
14767 
14768 
14769 	if (tp->t_fb->tfb_chg_query == NULL) {
14770 		/* Create a send map for the current outstanding data */
14771 
14772 		rsm = rack_alloc(rack);
14773 		if (rsm == NULL) {
14774 			uma_zfree(rack_pcb_zone, ptr);
14775 			return (ENOMEM);
14776 		}
14777 		rsm->r_no_rtt_allowed = 1;
14778 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14779 		rsm->r_rtr_cnt = 1;
14780 		rsm->r_rtr_bytes = 0;
14781 		if (tp->t_flags & TF_SENTFIN)
14782 			rsm->r_flags |= RACK_HAS_FIN;
14783 		rsm->r_end = tp->snd_max;
14784 		if (tp->snd_una == tp->iss) {
14785 			/* The data space is one beyond snd_una */
14786 			rsm->r_flags |= RACK_HAS_SYN;
14787 			rsm->r_start = tp->iss;
14788 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14789 		} else
14790 			rsm->r_start = tp->snd_una;
14791 		rsm->r_dupack = 0;
14792 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14793 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14794 			if (rsm->m) {
14795 				rsm->orig_m_len = rsm->m->m_len;
14796 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14797 			} else {
14798 				rsm->orig_m_len = 0;
14799 				rsm->orig_t_space = 0;
14800 			}
14801 		} else {
14802 			/*
14803 			 * This can happen if we have a stand-alone FIN or
14804 			 *  SYN.
14805 			 */
14806 			rsm->m = NULL;
14807 			rsm->orig_m_len = 0;
14808 			rsm->orig_t_space = 0;
14809 			rsm->soff = 0;
14810 		}
14811 #ifdef INVARIANTS
14812 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14813 			panic("Insert in rb tree fails ret:%d rack:%p rsm:%p",
14814 			      insret, rack, rsm);
14815 		}
14816 #else
14817 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14818 #endif
14819 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14820 		rsm->r_in_tmap = 1;
14821 	} else {
14822 		/* We have a query mechanism, lets use it */
14823 		struct tcp_query_resp qr;
14824 		int i;
14825 		tcp_seq at;
14826 
14827 		at = tp->snd_una;
14828 		while (at != tp->snd_max) {
14829 			memset(&qr, 0, sizeof(qr));
14830 			qr.req = TCP_QUERY_SENDMAP;
14831 			qr.req_param = at;
14832 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14833 				break;
14834 			/* Move forward */
14835 			at = qr.sendmap_end;
14836 			/* Now lets build the entry for this one */
14837 			rsm = rack_alloc(rack);
14838 			if (rsm == NULL) {
14839 				uma_zfree(rack_pcb_zone, ptr);
14840 				return (ENOMEM);
14841 			}
14842 			memset(rsm, 0, sizeof(struct rack_sendmap));
14843 			/* Now configure the rsm and insert it */
14844 			rsm->r_dupack = qr.sendmap_dupacks;
14845 			rsm->r_start = qr.sendmap_start;
14846 			rsm->r_end = qr.sendmap_end;
14847 			if (qr.sendmap_fas)
14848 				rsm->r_fas = qr.sendmap_end;
14849 			else
14850 				rsm->r_fas = rsm->r_start - tp->snd_una;
14851 			/*
14852 			 * We have carefully aligned the bits
14853 			 * so that all we have to do is copy over
14854 			 * the bits with the mask.
14855 			 */
14856 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14857 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14858 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14859 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14860 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
14861 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
14862 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14863 					   (rsm->r_start - tp->snd_una), &rsm->soff);
14864 			if (rsm->m) {
14865 				rsm->orig_m_len = rsm->m->m_len;
14866 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14867 			} else {
14868 				rsm->orig_m_len = 0;
14869 				rsm->orig_t_space = 0;
14870 			}
14871 #ifdef INVARIANTS
14872 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14873 				panic("Insert in rb tree fails ret:%d rack:%p rsm:%p",
14874 				      insret, rack, rsm);
14875 			}
14876 #else
14877 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14878 #endif
14879 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
14880 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14881 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14882 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14883 						/*
14884 						 * If the existing ersm was sent at
14885 						 * a later time than the new one, then
14886 						 * the new one should appear ahead of this
14887 						 * ersm.
14888 						 */
14889 						rsm->r_in_tmap = 1;
14890 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14891 						break;
14892 					}
14893 				}
14894 				if (rsm->r_in_tmap == 0) {
14895 					/*
14896 					 * Not found so shove it on the tail.
14897 					 */
14898 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14899 					rsm->r_in_tmap = 1;
14900 				}
14901  			} else {
14902 				if ((rack->r_ctl.rc_sacklast == NULL) ||
14903 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14904 					rack->r_ctl.rc_sacklast = rsm;
14905 				}
14906 			}
14907 			rack_log_chg_info(tp, rack, 3,
14908 					  rsm->r_start,
14909 					  rsm->r_end,
14910 					  rsm->r_flags);
14911 		}
14912 	}
14913 	return (0);
14914 }
14915 
14916 static void
14917 rack_translate_clamp_value(struct tcp_rack *rack, uint32_t optval)
14918 {
14919 	/*
14920 	 * P = percent bits
14921 	 * F = fill cw bit -- Toggle fillcw if this bit is set.
14922 	 * S = Segment bits
14923 	 * M = set max segment bit
14924 	 * U = Unclamined
14925 	 * C = If set to non-zero override the max number of clamps.
14926 	 * L = Bit to indicate if clamped gets lower.
14927 	 *
14928 	 * CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP
14929 	 *
14930 	 * The lowest 3 nibbles is the perentage .1 - 6553.5%
14931 	 * where 10.1 = 101, max 6553.5
14932 	 * The upper 16 bits  holds some options.
14933 	 * The F bit will turn on fill-cw on if you are
14934 	 * not pacing, it will turn it off if dgp is on.
14935 	 * The L bit will change it so when clamped we get
14936 	 * the min(gp, lt-bw) for dgp.
14937 	 */
14938 	uint16_t per;
14939 
14940 	rack->r_ctl.saved_rxt_clamp_val = optval;
14941 	per = optval & 0x0000ffff;
14942 	rack->r_ctl.rxt_threshold = (uint64_t)(per & 0xffff);
14943 	if (optval > 0) {
14944 		uint16_t clamp_opt;
14945 
14946 		rack->excess_rxt_on = 1;
14947 		clamp_opt = ((optval & 0xffff0000) >> 16);
14948 		rack->r_ctl.clamp_options = clamp_opt & 0x00ff;
14949 		if (clamp_opt & 0xff00) {
14950 			/* A max clamps is also present */
14951 			rack->r_ctl.max_clamps = (clamp_opt >> 8);
14952 		} else {
14953 			/* No specified clamps means no limit */
14954 			rack->r_ctl.max_clamps = 0;
14955 		}
14956 		if (rack->r_ctl.clamp_options & 0x0002) {
14957 			rack->r_clamped_gets_lower  = 1;
14958 		} else {
14959 			rack->r_clamped_gets_lower  = 0;
14960 		}
14961 	} else {
14962 		/* Turn it off back to default */
14963 		rack->excess_rxt_on = 0;
14964 		rack->r_clamped_gets_lower  = 0;
14965 	}
14966 
14967 }
14968 
14969 
14970 static int32_t
14971 rack_init(struct tcpcb *tp, void **ptr)
14972 {
14973 	struct inpcb *inp = tptoinpcb(tp);
14974 	struct tcp_rack *rack = NULL;
14975 	uint32_t iwin, snt, us_cts;
14976 	int err, no_query;
14977 
14978 	/*
14979 	 * First are we the initial or are we a switched stack?
14980 	 * If we are initing via tcp_newtcppcb the ptr passed
14981 	 * will be tp->t_fb_ptr. If its a stack switch that
14982 	 * has a previous stack we can query it will be a local
14983 	 * var that will in the end be set into t_fb_ptr.
14984 	 */
14985 	if (ptr == &tp->t_fb_ptr)
14986 		no_query = 1;
14987 	else
14988 		no_query = 0;
14989 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14990 	if (*ptr == NULL) {
14991 		/*
14992 		 * We need to allocate memory but cant. The INP and INP_INFO
14993 		 * locks and they are recursive (happens during setup. So a
14994 		 * scheme to drop the locks fails :(
14995 		 *
14996 		 */
14997 		return(ENOMEM);
14998 	}
14999 	memset(*ptr, 0, sizeof(struct tcp_rack));
15000 	rack = (struct tcp_rack *)*ptr;
15001 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
15002 	if (rack->r_ctl.tqh == NULL) {
15003 		uma_zfree(rack_pcb_zone, rack);
15004 		return(ENOMEM);
15005 	}
15006 	tqhash_init(rack->r_ctl.tqh);
15007 	TAILQ_INIT(&rack->r_ctl.rc_free);
15008 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
15009 	rack->rc_tp = tp;
15010 	rack->rc_inp = inp;
15011 	/* Set the flag */
15012 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
15013 	/* Probably not needed but lets be sure */
15014 	rack_clear_rate_sample(rack);
15015 	/*
15016 	 * Save off the default values, socket options will poke
15017 	 * at these if pacing is not on or we have not yet
15018 	 * reached where pacing is on (gp_ready/fixed enabled).
15019 	 * When they get set into the CC module (when gp_ready
15020 	 * is enabled or we enable fixed) then we will set these
15021 	 * values into the CC and place in here the old values
15022 	 * so we have a restoral. Then we will set the flag
15023 	 * rc_pacing_cc_set. That way whenever we turn off pacing
15024 	 * or switch off this stack, we will know to go restore
15025 	 * the saved values.
15026 	 *
15027 	 * We specifically put into the beta the ecn value for pacing.
15028 	 */
15029 	rack->rc_new_rnd_needed = 1;
15030 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
15031 	/* We want abe like behavior as well */
15032 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
15033 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
15034 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
15035 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
15036 	if (rack_rxt_clamp_thresh) {
15037 		rack_translate_clamp_value(rack, rack_rxt_clamp_thresh);
15038 		rack->excess_rxt_on = 1;
15039 	}
15040 	if (rack_uses_full_dgp_in_rec)
15041 		rack->r_ctl.full_dgp_in_rec = 1;
15042 	if (rack_fill_cw_state)
15043 		rack->rc_pace_to_cwnd = 1;
15044 	if (rack_pacing_min_seg)
15045 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
15046 	if (use_rack_rr)
15047 		rack->use_rack_rr = 1;
15048 	if (rack_dnd_default) {
15049 		rack->rc_pace_dnd = 1;
15050 	}
15051 	if (V_tcp_delack_enabled)
15052 		tp->t_delayed_ack = 1;
15053 	else
15054 		tp->t_delayed_ack = 0;
15055 #ifdef TCP_ACCOUNTING
15056 	if (rack_tcp_accounting) {
15057 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
15058 	}
15059 #endif
15060 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
15061 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
15062 	if (rack_enable_shared_cwnd)
15063 		rack->rack_enable_scwnd = 1;
15064 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
15065 	rack->rc_user_set_max_segs = rack_hptsi_segments;
15066 	rack->rc_force_max_seg = 0;
15067 	TAILQ_INIT(&rack->r_ctl.opt_list);
15068 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
15069 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
15070 	if (rack_hibeta_setting) {
15071 		rack->rack_hibeta = 1;
15072 		if ((rack_hibeta_setting >= 50) &&
15073 		    (rack_hibeta_setting <= 100)) {
15074 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
15075 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
15076 		}
15077 	} else {
15078 		rack->r_ctl.saved_hibeta = 50;
15079 	}
15080 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
15081 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
15082 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
15083 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
15084 	rack->r_ctl.rc_highest_us_rtt = 0;
15085 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
15086 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
15087 	if (rack_use_cmp_acks)
15088 		rack->r_use_cmp_ack = 1;
15089 	if (rack_disable_prr)
15090 		rack->rack_no_prr = 1;
15091 	if (rack_gp_no_rec_chg)
15092 		rack->rc_gp_no_rec_chg = 1;
15093 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
15094 		rack->rc_always_pace = 1;
15095 		if (rack->rack_hibeta)
15096 			rack_set_cc_pacing(rack);
15097 	} else
15098 		rack->rc_always_pace = 0;
15099 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
15100 		rack->r_mbuf_queue = 1;
15101 	else
15102 		rack->r_mbuf_queue = 0;
15103 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
15104 	if (rack_limits_scwnd)
15105 		rack->r_limit_scw = 1;
15106 	else
15107 		rack->r_limit_scw = 0;
15108 	rack_init_retransmit_value(rack, rack_rxt_controls);
15109 	rack->rc_labc = V_tcp_abc_l_var;
15110 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
15111 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
15112 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
15113 	rack->r_ctl.rc_min_to = rack_min_to;
15114 	microuptime(&rack->r_ctl.act_rcv_time);
15115 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
15116 	rack->rc_init_win = rack_default_init_window;
15117 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
15118 	if (rack_hw_up_only)
15119 		rack->r_up_only = 1;
15120 	if (rack_do_dyn_mul) {
15121 		/* When dynamic adjustment is on CA needs to start at 100% */
15122 		rack->rc_gp_dyn_mul = 1;
15123 		if (rack_do_dyn_mul >= 100)
15124 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
15125 	} else
15126 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
15127 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
15128 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
15129 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
15130 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
15131 				rack_probertt_filter_life);
15132 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15133 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
15134 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
15135 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
15136 	rack->r_ctl.rc_time_probertt_starts = 0;
15137 	if (rack_dsack_std_based & 0x1) {
15138 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
15139 		rack->rc_rack_tmr_std_based = 1;
15140 	}
15141 	if (rack_dsack_std_based & 0x2) {
15142 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
15143 		rack->rc_rack_use_dsack = 1;
15144 	}
15145 	/* We require at least one measurement, even if the sysctl is 0 */
15146 	if (rack_req_measurements)
15147 		rack->r_ctl.req_measurements = rack_req_measurements;
15148 	else
15149 		rack->r_ctl.req_measurements = 1;
15150 	if (rack_enable_hw_pacing)
15151 		rack->rack_hdw_pace_ena = 1;
15152 	if (rack_hw_rate_caps)
15153 		rack->r_rack_hw_rate_caps = 1;
15154 #ifdef TCP_SAD_DETECTION
15155 	rack->do_detection = 1;
15156 #else
15157 	rack->do_detection = 0;
15158 #endif
15159 	if (rack_non_rxt_use_cr)
15160 		rack->rack_rec_nonrxt_use_cr = 1;
15161 	/* Lets setup the fsb block */
15162 	err = rack_init_fsb(tp, rack);
15163 	if (err) {
15164 		uma_zfree(rack_pcb_zone, *ptr);
15165 		*ptr = NULL;
15166 		return (err);
15167 	}
15168 	if (rack_do_hystart) {
15169 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
15170 		if (rack_do_hystart > 1)
15171 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
15172 		if (rack_do_hystart > 2)
15173 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
15174 	}
15175 	/* Log what we will do with queries */
15176 	rack_log_chg_info(tp, rack, 7,
15177 			  no_query, 0, 0);
15178 	if (rack_def_profile)
15179 		rack_set_profile(rack, rack_def_profile);
15180 	/* Cancel the GP measurement in progress */
15181 	tp->t_flags &= ~TF_GPUTINPROG;
15182 	if ((tp->t_state != TCPS_CLOSED) &&
15183 	    (tp->t_state != TCPS_TIME_WAIT)) {
15184 		/*
15185 		 * We are already open, we may
15186 		 * need to adjust a few things.
15187 		 */
15188 		if (SEQ_GT(tp->snd_max, tp->iss))
15189 			snt = tp->snd_max - tp->iss;
15190 		else
15191 			snt = 0;
15192 		iwin = rc_init_window(rack);
15193 		if ((snt < iwin) &&
15194 		    (no_query == 1)) {
15195 			/* We are not past the initial window
15196 			 * on the first init (i.e. a stack switch
15197 			 * has not yet occured) so we need to make
15198 			 * sure cwnd and ssthresh is correct.
15199 			 */
15200 			if (tp->snd_cwnd < iwin)
15201 				tp->snd_cwnd = iwin;
15202 			/*
15203 			 * If we are within the initial window
15204 			 * we want ssthresh to be unlimited. Setting
15205 			 * it to the rwnd (which the default stack does
15206 			 * and older racks) is not really a good idea
15207 			 * since we want to be in SS and grow both the
15208 			 * cwnd and the rwnd (via dynamic rwnd growth). If
15209 			 * we set it to the rwnd then as the peer grows its
15210 			 * rwnd we will be stuck in CA and never hit SS.
15211 			 *
15212 			 * Its far better to raise it up high (this takes the
15213 			 * risk that there as been a loss already, probably
15214 			 * we should have an indicator in all stacks of loss
15215 			 * but we don't), but considering the normal use this
15216 			 * is a risk worth taking. The consequences of not
15217 			 * hitting SS are far worse than going one more time
15218 			 * into it early on (before we have sent even a IW).
15219 			 * It is highly unlikely that we will have had a loss
15220 			 * before getting the IW out.
15221 			 */
15222 			tp->snd_ssthresh = 0xffffffff;
15223 		}
15224 		/*
15225 		 * Any init based on sequence numbers
15226 		 * should be done in the deferred init path
15227 		 * since we can be CLOSED and not have them
15228 		 * inited when rack_init() is called. We
15229 		 * are not closed so lets call it.
15230 		 */
15231 		rack_deferred_init(tp, rack);
15232 	}
15233 	if ((tp->t_state != TCPS_CLOSED) &&
15234 	    (tp->t_state != TCPS_TIME_WAIT) &&
15235 	    (no_query == 0) &&
15236 	    (tp->snd_una != tp->snd_max))  {
15237 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
15238 		if (err) {
15239 			*ptr = NULL;
15240 			return(err);
15241 		}
15242 	}
15243 	rack_stop_all_timers(tp, rack);
15244 	/* Setup all the t_flags2 */
15245 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15246 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15247 	else
15248 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15249 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15250 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
15251 	/*
15252 	 * Timers in Rack are kept in microseconds so lets
15253 	 * convert any initial incoming variables
15254 	 * from ticks into usecs. Note that we
15255 	 * also change the values of t_srtt and t_rttvar, if
15256 	 * they are non-zero. They are kept with a 5
15257 	 * bit decimal so we have to carefully convert
15258 	 * these to get the full precision.
15259 	 */
15260 	rack_convert_rtts(tp);
15261 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
15262 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
15263 		/* We do not start any timers on DROPPED connections */
15264 		if (tp->t_fb->tfb_chg_query == NULL) {
15265 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15266 		} else {
15267 			struct tcp_query_resp qr;
15268 			int ret;
15269 
15270 			memset(&qr, 0, sizeof(qr));
15271 
15272 			/* Get the misc time stamps and such for rack */
15273 			qr.req = TCP_QUERY_RACK_TIMES;
15274 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15275 			if (ret == 1) {
15276 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
15277 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
15278 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
15279 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
15280 				rack->rc_rack_rtt = qr.rack_rtt;
15281 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
15282 				rack->r_ctl.rc_sacked = qr.rack_sacked;
15283 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
15284 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
15285 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
15286 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
15287 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
15288 				if (qr.rack_tlp_out) {
15289 					rack->rc_tlp_in_progress = 1;
15290 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
15291 				} else {
15292 					rack->rc_tlp_in_progress = 0;
15293 					rack->r_ctl.rc_tlp_cnt_out = 0;
15294 				}
15295 				if (qr.rack_srtt_measured)
15296 					rack->rc_srtt_measure_made = 1;
15297 				if (qr.rack_in_persist == 1) {
15298 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
15299 #ifdef NETFLIX_SHARED_CWND
15300 					if (rack->r_ctl.rc_scw) {
15301 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
15302 						rack->rack_scwnd_is_idle = 1;
15303 					}
15304 #endif
15305 					rack->r_ctl.persist_lost_ends = 0;
15306 					rack->probe_not_answered = 0;
15307 					rack->forced_ack = 0;
15308 					tp->t_rxtshift = 0;
15309 					rack->rc_in_persist = 1;
15310 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
15311 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
15312 				}
15313 				if (qr.rack_wanted_output)
15314 					rack->r_wanted_output = 1;
15315 				rack_log_chg_info(tp, rack, 6,
15316 						  qr.rack_min_rtt,
15317 						  qr.rack_rtt,
15318 						  qr.rack_reorder_ts);
15319 			}
15320 			/* Get the old stack timers */
15321 			qr.req_param = 0;
15322 			qr.req = TCP_QUERY_TIMERS_UP;
15323 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15324 			if (ret) {
15325 				/*
15326 				 * non-zero return means we have a timer('s)
15327 				 * to start. Zero means no timer (no keepalive
15328 				 * I suppose).
15329 				 */
15330 				uint32_t tov = 0;
15331 
15332 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
15333 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
15334 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
15335 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
15336 						tov = qr.timer_pacing_to - us_cts;
15337 					else
15338 						tov = HPTS_TICKS_PER_SLOT;
15339 				}
15340 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
15341 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
15342 					if (tov == 0) {
15343 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
15344 							tov = qr.timer_timer_exp - us_cts;
15345 						else
15346 							tov = HPTS_TICKS_PER_SLOT;
15347 					}
15348 				}
15349 				rack_log_chg_info(tp, rack, 4,
15350 						  rack->r_ctl.rc_hpts_flags,
15351 						  rack->r_ctl.rc_last_output_to,
15352 						  rack->r_ctl.rc_timer_exp);
15353 				if (tov) {
15354 					struct hpts_diag diag;
15355 
15356 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15357 								   __LINE__, &diag);
15358 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15359 				}
15360 			}
15361 		}
15362 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15363 				     __LINE__, RACK_RTTS_INIT);
15364 	}
15365 	return (0);
15366 }
15367 
15368 static int
15369 rack_handoff_ok(struct tcpcb *tp)
15370 {
15371 	if ((tp->t_state == TCPS_CLOSED) ||
15372 	    (tp->t_state == TCPS_LISTEN)) {
15373 		/* Sure no problem though it may not stick */
15374 		return (0);
15375 	}
15376 	if ((tp->t_state == TCPS_SYN_SENT) ||
15377 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
15378 		/*
15379 		 * We really don't know if you support sack,
15380 		 * you have to get to ESTAB or beyond to tell.
15381 		 */
15382 		return (EAGAIN);
15383 	}
15384 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15385 		/*
15386 		 * Rack will only send a FIN after all data is acknowledged.
15387 		 * So in this case we have more data outstanding. We can't
15388 		 * switch stacks until either all data and only the FIN
15389 		 * is left (in which case rack_init() now knows how
15390 		 * to deal with that) <or> all is acknowledged and we
15391 		 * are only left with incoming data, though why you
15392 		 * would want to switch to rack after all data is acknowledged
15393 		 * I have no idea (rrs)!
15394 		 */
15395 		return (EAGAIN);
15396 	}
15397 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15398 		return (0);
15399 	}
15400 	/*
15401 	 * If we reach here we don't do SACK on this connection so we can
15402 	 * never do rack.
15403 	 */
15404 	return (EINVAL);
15405 }
15406 
15407 static void
15408 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15409 {
15410 
15411 	if (tp->t_fb_ptr) {
15412 		uint32_t cnt_free = 0;
15413 		struct tcp_rack *rack;
15414 		struct rack_sendmap *rsm;
15415 
15416 		tcp_handle_orphaned_packets(tp);
15417 		tp->t_flags &= ~TF_FORCEDATA;
15418 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15419 		rack_log_pacing_delay_calc(rack,
15420 					   0,
15421 					   0,
15422 					   0,
15423 					   rack_get_gp_est(rack), /* delRate */
15424 					   rack_get_lt_bw(rack), /* rttProp */
15425 					   20, __LINE__, NULL, 0);
15426 #ifdef NETFLIX_SHARED_CWND
15427 		if (rack->r_ctl.rc_scw) {
15428 			uint32_t limit;
15429 
15430 			if (rack->r_limit_scw)
15431 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15432 			else
15433 				limit = 0;
15434 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15435 						  rack->r_ctl.rc_scw_index,
15436 						  limit);
15437 			rack->r_ctl.rc_scw = NULL;
15438 		}
15439 #endif
15440 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15441 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15442 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15443 			rack->r_ctl.fsb.th = NULL;
15444 		}
15445 		if (rack->rc_always_pace) {
15446 			tcp_decrement_paced_conn();
15447 			rack_undo_cc_pacing(rack);
15448 			rack->rc_always_pace = 0;
15449 		}
15450 		/* Clean up any options if they were not applied */
15451 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15452 			struct deferred_opt_list *dol;
15453 
15454 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15455 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15456 			free(dol, M_TCPDO);
15457 		}
15458 		/* rack does not use force data but other stacks may clear it */
15459 		if (rack->r_ctl.crte != NULL) {
15460 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15461 			rack->rack_hdrw_pacing = 0;
15462 			rack->r_ctl.crte = NULL;
15463 		}
15464 #ifdef TCP_BLACKBOX
15465 		tcp_log_flowend(tp);
15466 #endif
15467 		/*
15468 		 * Lets take a different approach to purging just
15469 		 * get each one and free it like a cum-ack would and
15470 		 * not use a foreach loop.
15471 		 */
15472 		rsm = tqhash_min(rack->r_ctl.tqh);
15473 		while (rsm) {
15474 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15475 			rack->r_ctl.rc_num_maps_alloced--;
15476 			uma_zfree(rack_zone, rsm);
15477 			rsm = tqhash_min(rack->r_ctl.tqh);
15478 		}
15479 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15480 		while (rsm) {
15481 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15482 			rack->r_ctl.rc_num_maps_alloced--;
15483 			rack->rc_free_cnt--;
15484 			cnt_free++;
15485 			uma_zfree(rack_zone, rsm);
15486 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15487 		}
15488 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15489 		    (tcp_bblogging_on(tp))) {
15490 			union tcp_log_stackspecific log;
15491 			struct timeval tv;
15492 
15493 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15494 			log.u_bbr.flex8 = 10;
15495 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15496 			log.u_bbr.flex2 = rack->rc_free_cnt;
15497 			log.u_bbr.flex3 = cnt_free;
15498 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15499 			rsm = tqhash_min(rack->r_ctl.tqh);
15500 			log.u_bbr.delRate = (uint64_t)rsm;
15501 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15502 			log.u_bbr.cur_del_rate = (uint64_t)rsm;
15503 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15504 			log.u_bbr.pkt_epoch = __LINE__;
15505 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15506 					     0, &log, false, NULL, NULL, 0, &tv);
15507 		}
15508 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15509 			("rack:%p num_aloc:%u after freeing all?",
15510 			 rack,
15511 			 rack->r_ctl.rc_num_maps_alloced));
15512 		rack->rc_free_cnt = 0;
15513 		free(rack->r_ctl.tqh, M_TCPFSB);
15514 		rack->r_ctl.tqh = NULL;
15515 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15516 		tp->t_fb_ptr = NULL;
15517 	}
15518 	/* Make sure snd_nxt is correctly set */
15519 	tp->snd_nxt = tp->snd_max;
15520 }
15521 
15522 static void
15523 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15524 {
15525 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15526 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15527 	}
15528 	switch (tp->t_state) {
15529 	case TCPS_SYN_SENT:
15530 		rack->r_state = TCPS_SYN_SENT;
15531 		rack->r_substate = rack_do_syn_sent;
15532 		break;
15533 	case TCPS_SYN_RECEIVED:
15534 		rack->r_state = TCPS_SYN_RECEIVED;
15535 		rack->r_substate = rack_do_syn_recv;
15536 		break;
15537 	case TCPS_ESTABLISHED:
15538 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15539 		rack->r_state = TCPS_ESTABLISHED;
15540 		rack->r_substate = rack_do_established;
15541 		break;
15542 	case TCPS_CLOSE_WAIT:
15543 		rack->r_state = TCPS_CLOSE_WAIT;
15544 		rack->r_substate = rack_do_close_wait;
15545 		break;
15546 	case TCPS_FIN_WAIT_1:
15547 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15548 		rack->r_state = TCPS_FIN_WAIT_1;
15549 		rack->r_substate = rack_do_fin_wait_1;
15550 		break;
15551 	case TCPS_CLOSING:
15552 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15553 		rack->r_state = TCPS_CLOSING;
15554 		rack->r_substate = rack_do_closing;
15555 		break;
15556 	case TCPS_LAST_ACK:
15557 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15558 		rack->r_state = TCPS_LAST_ACK;
15559 		rack->r_substate = rack_do_lastack;
15560 		break;
15561 	case TCPS_FIN_WAIT_2:
15562 		rack->r_state = TCPS_FIN_WAIT_2;
15563 		rack->r_substate = rack_do_fin_wait_2;
15564 		break;
15565 	case TCPS_LISTEN:
15566 	case TCPS_CLOSED:
15567 	case TCPS_TIME_WAIT:
15568 	default:
15569 		break;
15570 	};
15571 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15572 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15573 
15574 }
15575 
15576 static void
15577 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15578 {
15579 	/*
15580 	 * We received an ack, and then did not
15581 	 * call send or were bounced out due to the
15582 	 * hpts was running. Now a timer is up as well, is
15583 	 * it the right timer?
15584 	 */
15585 	struct rack_sendmap *rsm;
15586 	int tmr_up;
15587 
15588 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15589 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15590 		return;
15591 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15592 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15593 	    (tmr_up == PACE_TMR_RXT)) {
15594 		/* Should be an RXT */
15595 		return;
15596 	}
15597 	if (rsm == NULL) {
15598 		/* Nothing outstanding? */
15599 		if (tp->t_flags & TF_DELACK) {
15600 			if (tmr_up == PACE_TMR_DELACK)
15601 				/* We are supposed to have delayed ack up and we do */
15602 				return;
15603 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15604 			/*
15605 			 * if we hit enobufs then we would expect the possibility
15606 			 * of nothing outstanding and the RXT up (and the hptsi timer).
15607 			 */
15608 			return;
15609 		} else if (((V_tcp_always_keepalive ||
15610 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15611 			    (tp->t_state <= TCPS_CLOSING)) &&
15612 			   (tmr_up == PACE_TMR_KEEP) &&
15613 			   (tp->snd_max == tp->snd_una)) {
15614 			/* We should have keep alive up and we do */
15615 			return;
15616 		}
15617 	}
15618 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15619 		   ((tmr_up == PACE_TMR_TLP) ||
15620 		    (tmr_up == PACE_TMR_RACK) ||
15621 		    (tmr_up == PACE_TMR_RXT))) {
15622 		/*
15623 		 * Either a Rack, TLP or RXT is fine if  we
15624 		 * have outstanding data.
15625 		 */
15626 		return;
15627 	} else if (tmr_up == PACE_TMR_DELACK) {
15628 		/*
15629 		 * If the delayed ack was going to go off
15630 		 * before the rtx/tlp/rack timer were going to
15631 		 * expire, then that would be the timer in control.
15632 		 * Note we don't check the time here trusting the
15633 		 * code is correct.
15634 		 */
15635 		return;
15636 	}
15637 	/*
15638 	 * Ok the timer originally started is not what we want now.
15639 	 * We will force the hpts to be stopped if any, and restart
15640 	 * with the slot set to what was in the saved slot.
15641 	 */
15642 	if (tcp_in_hpts(rack->rc_tp)) {
15643 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15644 			uint32_t us_cts;
15645 
15646 			us_cts = tcp_get_usecs(NULL);
15647 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15648 				rack->r_early = 1;
15649 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15650 			}
15651 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15652 		}
15653 		tcp_hpts_remove(rack->rc_tp);
15654 	}
15655 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15656 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15657 }
15658 
15659 
15660 static void
15661 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15662 {
15663 	if ((SEQ_LT(tp->snd_wl1, seq) ||
15664 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15665 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15666 		/* keep track of pure window updates */
15667 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15668 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15669 		tp->snd_wnd = tiwin;
15670 		rack_validate_fo_sendwin_up(tp, rack);
15671 		tp->snd_wl1 = seq;
15672 		tp->snd_wl2 = ack;
15673 		if (tp->snd_wnd > tp->max_sndwnd)
15674 			tp->max_sndwnd = tp->snd_wnd;
15675 	    rack->r_wanted_output = 1;
15676 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15677 		tp->snd_wnd = tiwin;
15678 		rack_validate_fo_sendwin_up(tp, rack);
15679 		tp->snd_wl1 = seq;
15680 		tp->snd_wl2 = ack;
15681 	} else {
15682 		/* Not a valid win update */
15683 		return;
15684 	}
15685 	if (tp->snd_wnd > tp->max_sndwnd)
15686 		tp->max_sndwnd = tp->snd_wnd;
15687 	/* Do we exit persists? */
15688 	if ((rack->rc_in_persist != 0) &&
15689 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15690 				rack->r_ctl.rc_pace_min_segs))) {
15691 		rack_exit_persist(tp, rack, cts);
15692 	}
15693 	/* Do we enter persists? */
15694 	if ((rack->rc_in_persist == 0) &&
15695 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15696 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
15697 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15698 	    sbavail(&tptosocket(tp)->so_snd) &&
15699 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15700 		/*
15701 		 * Here the rwnd is less than
15702 		 * the pacing size, we are established,
15703 		 * nothing is outstanding, and there is
15704 		 * data to send. Enter persists.
15705 		 */
15706 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15707 	}
15708 }
15709 
15710 static void
15711 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15712 {
15713 
15714 	if (tcp_bblogging_on(rack->rc_tp)) {
15715 		struct inpcb *inp = tptoinpcb(tp);
15716 		union tcp_log_stackspecific log;
15717 		struct timeval ltv;
15718 		char tcp_hdr_buf[60];
15719 		struct tcphdr *th;
15720 		struct timespec ts;
15721 		uint32_t orig_snd_una;
15722 		uint8_t xx = 0;
15723 
15724 #ifdef TCP_REQUEST_TRK
15725 		struct tcp_sendfile_track *tcp_req;
15726 
15727 		if (SEQ_GT(ae->ack, tp->snd_una)) {
15728 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15729 		} else {
15730 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15731 		}
15732 #endif
15733 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15734 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15735 		if (rack->rack_no_prr == 0)
15736 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15737 		else
15738 			log.u_bbr.flex1 = 0;
15739 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15740 		log.u_bbr.use_lt_bw <<= 1;
15741 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
15742 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15743 		log.u_bbr.bbr_state = rack->rc_free_cnt;
15744 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15745 		log.u_bbr.pkts_out = tp->t_maxseg;
15746 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15747 		log.u_bbr.flex7 = 1;
15748 		log.u_bbr.lost = ae->flags;
15749 		log.u_bbr.cwnd_gain = ackval;
15750 		log.u_bbr.pacing_gain = 0x2;
15751 		if (ae->flags & TSTMP_HDWR) {
15752 			/* Record the hardware timestamp if present */
15753 			log.u_bbr.flex3 = M_TSTMP;
15754 			ts.tv_sec = ae->timestamp / 1000000000;
15755 			ts.tv_nsec = ae->timestamp % 1000000000;
15756 			ltv.tv_sec = ts.tv_sec;
15757 			ltv.tv_usec = ts.tv_nsec / 1000;
15758 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15759 		} else if (ae->flags & TSTMP_LRO) {
15760 			/* Record the LRO the arrival timestamp */
15761 			log.u_bbr.flex3 = M_TSTMP_LRO;
15762 			ts.tv_sec = ae->timestamp / 1000000000;
15763 			ts.tv_nsec = ae->timestamp % 1000000000;
15764 			ltv.tv_sec = ts.tv_sec;
15765 			ltv.tv_usec = ts.tv_nsec / 1000;
15766 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15767 		}
15768 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15769 		/* Log the rcv time */
15770 		log.u_bbr.delRate = ae->timestamp;
15771 #ifdef TCP_REQUEST_TRK
15772 		log.u_bbr.applimited = tp->t_tcpreq_closed;
15773 		log.u_bbr.applimited <<= 8;
15774 		log.u_bbr.applimited |= tp->t_tcpreq_open;
15775 		log.u_bbr.applimited <<= 8;
15776 		log.u_bbr.applimited |= tp->t_tcpreq_req;
15777 		if (tcp_req) {
15778 			/* Copy out any client req info */
15779 			/* seconds */
15780 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15781 			/* useconds */
15782 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15783 			log.u_bbr.rttProp = tcp_req->timestamp;
15784 			log.u_bbr.cur_del_rate = tcp_req->start;
15785 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15786 				log.u_bbr.flex8 |= 1;
15787 			} else {
15788 				log.u_bbr.flex8 |= 2;
15789 				log.u_bbr.bw_inuse = tcp_req->end;
15790 			}
15791 			log.u_bbr.flex6 = tcp_req->start_seq;
15792 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15793 				log.u_bbr.flex8 |= 4;
15794 				log.u_bbr.epoch = tcp_req->end_seq;
15795 			}
15796 		}
15797 #endif
15798 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15799 		th = (struct tcphdr *)tcp_hdr_buf;
15800 		th->th_seq = ae->seq;
15801 		th->th_ack = ae->ack;
15802 		th->th_win = ae->win;
15803 		/* Now fill in the ports */
15804 		th->th_sport = inp->inp_fport;
15805 		th->th_dport = inp->inp_lport;
15806 		tcp_set_flags(th, ae->flags);
15807 		/* Now do we have a timestamp option? */
15808 		if (ae->flags & HAS_TSTMP) {
15809 			u_char *cp;
15810 			uint32_t val;
15811 
15812 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15813 			cp = (u_char *)(th + 1);
15814 			*cp = TCPOPT_NOP;
15815 			cp++;
15816 			*cp = TCPOPT_NOP;
15817 			cp++;
15818 			*cp = TCPOPT_TIMESTAMP;
15819 			cp++;
15820 			*cp = TCPOLEN_TIMESTAMP;
15821 			cp++;
15822 			val = htonl(ae->ts_value);
15823 			bcopy((char *)&val,
15824 			      (char *)cp, sizeof(uint32_t));
15825 			val = htonl(ae->ts_echo);
15826 			bcopy((char *)&val,
15827 			      (char *)(cp + 4), sizeof(uint32_t));
15828 		} else
15829 			th->th_off = (sizeof(struct tcphdr) >> 2);
15830 
15831 		/*
15832 		 * For sane logging we need to play a little trick.
15833 		 * If the ack were fully processed we would have moved
15834 		 * snd_una to high_seq, but since compressed acks are
15835 		 * processed in two phases, at this point (logging) snd_una
15836 		 * won't be advanced. So we would see multiple acks showing
15837 		 * the advancement. We can prevent that by "pretending" that
15838 		 * snd_una was advanced and then un-advancing it so that the
15839 		 * logging code has the right value for tlb_snd_una.
15840 		 */
15841 		if (tp->snd_una != high_seq) {
15842 			orig_snd_una = tp->snd_una;
15843 			tp->snd_una = high_seq;
15844 			xx = 1;
15845 		} else
15846 			xx = 0;
15847 		TCP_LOG_EVENTP(tp, th,
15848 			       &tptosocket(tp)->so_rcv,
15849 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15850 			       0, &log, true, &ltv);
15851 		if (xx) {
15852 			tp->snd_una = orig_snd_una;
15853 		}
15854 	}
15855 
15856 }
15857 
15858 static void
15859 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15860 {
15861 	uint32_t us_rtt;
15862 	/*
15863 	 * A persist or keep-alive was forced out, update our
15864 	 * min rtt time. Note now worry about lost responses.
15865 	 * When a subsequent keep-alive or persist times out
15866 	 * and forced_ack is still on, then the last probe
15867 	 * was not responded to. In such cases we have a
15868 	 * sysctl that controls the behavior. Either we apply
15869 	 * the rtt but with reduced confidence (0). Or we just
15870 	 * plain don't apply the rtt estimate. Having data flow
15871 	 * will clear the probe_not_answered flag i.e. cum-ack
15872 	 * move forward <or> exiting and reentering persists.
15873 	 */
15874 
15875 	rack->forced_ack = 0;
15876 	rack->rc_tp->t_rxtshift = 0;
15877 	if ((rack->rc_in_persist &&
15878 	     (tiwin == rack->rc_tp->snd_wnd)) ||
15879 	    (rack->rc_in_persist == 0)) {
15880 		/*
15881 		 * In persists only apply the RTT update if this is
15882 		 * a response to our window probe. And that
15883 		 * means the rwnd sent must match the current
15884 		 * snd_wnd. If it does not, then we got a
15885 		 * window update ack instead. For keepalive
15886 		 * we allow the answer no matter what the window.
15887 		 *
15888 		 * Note that if the probe_not_answered is set then
15889 		 * the forced_ack_ts is the oldest one i.e. the first
15890 		 * probe sent that might have been lost. This assures
15891 		 * us that if we do calculate an RTT it is longer not
15892 		 * some short thing.
15893 		 */
15894 		if (rack->rc_in_persist)
15895 			counter_u64_add(rack_persists_acks, 1);
15896 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15897 		if (us_rtt == 0)
15898 			us_rtt = 1;
15899 		if (rack->probe_not_answered == 0) {
15900 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15901 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15902 		} else {
15903 			/* We have a retransmitted probe here too */
15904 			if (rack_apply_rtt_with_reduced_conf) {
15905 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15906 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15907 			}
15908 		}
15909 	}
15910 }
15911 
15912 static int
15913 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15914 {
15915 	/*
15916 	 * Handle a "special" compressed ack mbuf. Each incoming
15917 	 * ack has only four possible dispositions:
15918 	 *
15919 	 * A) It moves the cum-ack forward
15920 	 * B) It is behind the cum-ack.
15921 	 * C) It is a window-update ack.
15922 	 * D) It is a dup-ack.
15923 	 *
15924 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15925 	 * in the incoming mbuf. We also need to still pay attention
15926 	 * to nxt_pkt since there may be another packet after this
15927 	 * one.
15928 	 */
15929 #ifdef TCP_ACCOUNTING
15930 	uint64_t ts_val;
15931 	uint64_t rdstc;
15932 #endif
15933 	int segsiz;
15934 	struct timespec ts;
15935 	struct tcp_rack *rack;
15936 	struct tcp_ackent *ae;
15937 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15938 	int cnt, i, did_out, ourfinisacked = 0;
15939 	struct tcpopt to_holder, *to = NULL;
15940 #ifdef TCP_ACCOUNTING
15941 	int win_up_req = 0;
15942 #endif
15943 	int nsegs = 0;
15944 	int under_pacing = 0;
15945 	int recovery = 0;
15946 #ifdef TCP_ACCOUNTING
15947 	sched_pin();
15948 #endif
15949 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15950 	if (rack->gp_ready &&
15951 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15952 		under_pacing = 1;
15953 
15954 	if (rack->r_state != tp->t_state)
15955 		rack_set_state(tp, rack);
15956 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15957 	    (tp->t_flags & TF_GPUTINPROG)) {
15958 		/*
15959 		 * We have a goodput in progress
15960 		 * and we have entered a late state.
15961 		 * Do we have enough data in the sb
15962 		 * to handle the GPUT request?
15963 		 */
15964 		uint32_t bytes;
15965 
15966 		bytes = tp->gput_ack - tp->gput_seq;
15967 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
15968 			bytes += tp->gput_seq - tp->snd_una;
15969 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15970 			/*
15971 			 * There are not enough bytes in the socket
15972 			 * buffer that have been sent to cover this
15973 			 * measurement. Cancel it.
15974 			 */
15975 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15976 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
15977 						   tp->gput_seq,
15978 						   0, 0, 18, __LINE__, NULL, 0);
15979 			tp->t_flags &= ~TF_GPUTINPROG;
15980 		}
15981 	}
15982 	to = &to_holder;
15983 	to->to_flags = 0;
15984 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15985 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15986 	cnt = m->m_len / sizeof(struct tcp_ackent);
15987 	counter_u64_add(rack_multi_single_eq, cnt);
15988 	high_seq = tp->snd_una;
15989 	the_win = tp->snd_wnd;
15990 	win_seq = tp->snd_wl1;
15991 	win_upd_ack = tp->snd_wl2;
15992 	cts = tcp_tv_to_usectick(tv);
15993 	ms_cts = tcp_tv_to_mssectick(tv);
15994 	rack->r_ctl.rc_rcvtime = cts;
15995 	segsiz = ctf_fixed_maxseg(tp);
15996 	if ((rack->rc_gp_dyn_mul) &&
15997 	    (rack->use_fixed_rate == 0) &&
15998 	    (rack->rc_always_pace)) {
15999 		/* Check in on probertt */
16000 		rack_check_probe_rtt(rack, cts);
16001 	}
16002 	for (i = 0; i < cnt; i++) {
16003 #ifdef TCP_ACCOUNTING
16004 		ts_val = get_cyclecount();
16005 #endif
16006 		rack_clear_rate_sample(rack);
16007 		ae = ((mtod(m, struct tcp_ackent *)) + i);
16008 		if (ae->flags & TH_FIN)
16009 			rack_log_pacing_delay_calc(rack,
16010 						   0,
16011 						   0,
16012 						   0,
16013 						   rack_get_gp_est(rack), /* delRate */
16014 						   rack_get_lt_bw(rack), /* rttProp */
16015 						   20, __LINE__, NULL, 0);
16016 		/* Setup the window */
16017 		tiwin = ae->win << tp->snd_scale;
16018 		if (tiwin > rack->r_ctl.rc_high_rwnd)
16019 			rack->r_ctl.rc_high_rwnd = tiwin;
16020 		/* figure out the type of ack */
16021 		if (SEQ_LT(ae->ack, high_seq)) {
16022 			/* Case B*/
16023 			ae->ack_val_set = ACK_BEHIND;
16024 		} else if (SEQ_GT(ae->ack, high_seq)) {
16025 			/* Case A */
16026 			ae->ack_val_set = ACK_CUMACK;
16027 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
16028 			/* Case D */
16029 			ae->ack_val_set = ACK_DUPACK;
16030 		} else {
16031 			/* Case C */
16032 			ae->ack_val_set = ACK_RWND;
16033 		}
16034 		if (rack->sack_attack_disable > 0) {
16035 			rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16036 			rack->r_ctl.ack_during_sd++;
16037 		}
16038 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
16039 		/* Validate timestamp */
16040 		if (ae->flags & HAS_TSTMP) {
16041 			/* Setup for a timestamp */
16042 			to->to_flags = TOF_TS;
16043 			ae->ts_echo -= tp->ts_offset;
16044 			to->to_tsecr = ae->ts_echo;
16045 			to->to_tsval = ae->ts_value;
16046 			/*
16047 			 * If echoed timestamp is later than the current time, fall back to
16048 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16049 			 * were used when this connection was established.
16050 			 */
16051 			if (TSTMP_GT(ae->ts_echo, ms_cts))
16052 				to->to_tsecr = 0;
16053 			if (tp->ts_recent &&
16054 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
16055 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
16056 #ifdef TCP_ACCOUNTING
16057 					rdstc = get_cyclecount();
16058 					if (rdstc > ts_val) {
16059 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16060 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16061 						}
16062 					}
16063 #endif
16064 					continue;
16065 				}
16066 			}
16067 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
16068 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
16069 				tp->ts_recent_age = tcp_ts_getticks();
16070 				tp->ts_recent = ae->ts_value;
16071 			}
16072 		} else {
16073 			/* Setup for a no options */
16074 			to->to_flags = 0;
16075 		}
16076 		/* Update the rcv time and perform idle reduction possibly */
16077 		if  (tp->t_idle_reduce &&
16078 		     (tp->snd_max == tp->snd_una) &&
16079 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16080 			counter_u64_add(rack_input_idle_reduces, 1);
16081 			rack_cc_after_idle(rack, tp);
16082 		}
16083 		tp->t_rcvtime = ticks;
16084 		/* Now what about ECN of a chain of pure ACKs? */
16085 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
16086 			tcp_packets_this_ack(tp, ae->ack),
16087 			ae->codepoint))
16088 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
16089 #ifdef TCP_ACCOUNTING
16090 		/* Count for the specific type of ack in */
16091 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16092 			tp->tcp_cnt_counters[ae->ack_val_set]++;
16093 		}
16094 #endif
16095 		/*
16096 		 * Note how we could move up these in the determination
16097 		 * above, but we don't so that way the timestamp checks (and ECN)
16098 		 * is done first before we do any processing on the ACK.
16099 		 * The non-compressed path through the code has this
16100 		 * weakness (noted by @jtl) that it actually does some
16101 		 * processing before verifying the timestamp information.
16102 		 * We don't take that path here which is why we set
16103 		 * the ack_val_set first, do the timestamp and ecn
16104 		 * processing, and then look at what we have setup.
16105 		 */
16106 		if (ae->ack_val_set == ACK_BEHIND) {
16107 			/*
16108 			 * Case B flag reordering, if window is not closed
16109 			 * or it could be a keep-alive or persists
16110 			 */
16111 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
16112 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16113 				if (rack->r_ctl.rc_reorder_ts == 0)
16114 					rack->r_ctl.rc_reorder_ts = 1;
16115 			}
16116 		} else if (ae->ack_val_set == ACK_DUPACK) {
16117 			/* Case D */
16118 			rack_strike_dupack(rack);
16119 		} else if (ae->ack_val_set == ACK_RWND) {
16120 			/* Case C */
16121 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16122 				ts.tv_sec = ae->timestamp / 1000000000;
16123 				ts.tv_nsec = ae->timestamp % 1000000000;
16124 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16125 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16126 			} else {
16127 				rack->r_ctl.act_rcv_time = *tv;
16128 			}
16129 			if (rack->forced_ack) {
16130 				rack_handle_probe_response(rack, tiwin,
16131 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
16132 			}
16133 #ifdef TCP_ACCOUNTING
16134 			win_up_req = 1;
16135 #endif
16136 			win_upd_ack = ae->ack;
16137 			win_seq = ae->seq;
16138 			the_win = tiwin;
16139 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16140 		} else {
16141 			/* Case A */
16142 			if (SEQ_GT(ae->ack, tp->snd_max)) {
16143 				/*
16144 				 * We just send an ack since the incoming
16145 				 * ack is beyond the largest seq we sent.
16146 				 */
16147 				if ((tp->t_flags & TF_ACKNOW) == 0) {
16148 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
16149 					if (tp->t_flags && TF_ACKNOW)
16150 						rack->r_wanted_output = 1;
16151 				}
16152 			} else {
16153 				nsegs++;
16154 				/* If the window changed setup to update */
16155 				if (tiwin != tp->snd_wnd) {
16156 					win_upd_ack = ae->ack;
16157 					win_seq = ae->seq;
16158 					the_win = tiwin;
16159 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16160 				}
16161 #ifdef TCP_ACCOUNTING
16162 				/* Account for the acks */
16163 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16164 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
16165 				}
16166 #endif
16167 				high_seq = ae->ack;
16168 				if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp))
16169 					rack_log_hystart_event(rack, high_seq, 8);
16170 				/* Setup our act_rcv_time */
16171 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16172 					ts.tv_sec = ae->timestamp / 1000000000;
16173 					ts.tv_nsec = ae->timestamp % 1000000000;
16174 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16175 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16176 				} else {
16177 					rack->r_ctl.act_rcv_time = *tv;
16178 				}
16179 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
16180 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
16181 #ifdef TCP_REQUEST_TRK
16182 				rack_req_check_for_comp(rack, high_seq);
16183 #endif
16184 				if (rack->rc_dsack_round_seen) {
16185 					/* Is the dsack round over? */
16186 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
16187 						/* Yes it is */
16188 						rack->rc_dsack_round_seen = 0;
16189 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
16190 					}
16191 				}
16192 			}
16193 		}
16194 		/* And lets be sure to commit the rtt measurements for this ack */
16195 		tcp_rack_xmit_timer_commit(rack, tp);
16196 #ifdef TCP_ACCOUNTING
16197 		rdstc = get_cyclecount();
16198 		if (rdstc > ts_val) {
16199 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16200 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16201 				if (ae->ack_val_set == ACK_CUMACK)
16202 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
16203 			}
16204 		}
16205 #endif
16206 	}
16207 #ifdef TCP_ACCOUNTING
16208 	ts_val = get_cyclecount();
16209 #endif
16210 	/* Tend to any collapsed window */
16211 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
16212 		/* The peer collapsed the window */
16213 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
16214 	} else if (rack->rc_has_collapsed)
16215 		rack_un_collapse_window(rack, __LINE__);
16216 	if ((rack->r_collapse_point_valid) &&
16217 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16218 		rack->r_collapse_point_valid = 0;
16219 	acked_amount = acked = (high_seq - tp->snd_una);
16220 	if (acked) {
16221 		/*
16222 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16223 		 * causes issues when we are just going app limited. Lets
16224 		 * instead use SEQ_GT <or> where its equal but more data
16225 		 * is outstanding.
16226 		 *
16227 		 * Also make sure we are on the last ack of a series. We
16228 		 * have to have all the ack's processed in queue to know
16229 		 * if there is something left outstanding.
16230 		 *
16231 		 */
16232 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16233 		    (rack->rc_new_rnd_needed == 0) &&
16234 		    (nxt_pkt == 0)) {
16235 			rack_log_hystart_event(rack, high_seq, 21);
16236 			rack->r_ctl.current_round++;
16237 			/* Force the next send to setup the next round */
16238 			rack->rc_new_rnd_needed = 1;
16239 			if (CC_ALGO(tp)->newround != NULL) {
16240 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
16241 			}
16242 		}
16243 		/*
16244 		 * Clear the probe not answered flag
16245 		 * since cum-ack moved forward.
16246 		 */
16247 		rack->probe_not_answered = 0;
16248 		if (rack->sack_attack_disable == 0)
16249 			rack_do_decay(rack);
16250 		if (acked >= segsiz) {
16251 			/*
16252 			 * You only get credit for
16253 			 * MSS and greater (and you get extra
16254 			 * credit for larger cum-ack moves).
16255 			 */
16256 			int ac;
16257 
16258 			ac = acked / segsiz;
16259 			rack->r_ctl.ack_count += ac;
16260 			counter_u64_add(rack_ack_total, ac);
16261 		}
16262 		if (rack->r_ctl.ack_count > 0xfff00000) {
16263 			/*
16264 			 * reduce the number to keep us under
16265 			 * a uint32_t.
16266 			 */
16267 			rack->r_ctl.ack_count /= 2;
16268 			rack->r_ctl.sack_count /= 2;
16269 		}
16270 		if (tp->t_flags & TF_NEEDSYN) {
16271 			/*
16272 			 * T/TCP: Connection was half-synchronized, and our SYN has
16273 			 * been ACK'd (so connection is now fully synchronized).  Go
16274 			 * to non-starred state, increment snd_una for ACK of SYN,
16275 			 * and check if we can do window scaling.
16276 			 */
16277 			tp->t_flags &= ~TF_NEEDSYN;
16278 			tp->snd_una++;
16279 			acked_amount = acked = (high_seq - tp->snd_una);
16280 		}
16281 		if (acked > sbavail(&so->so_snd))
16282 			acked_amount = sbavail(&so->so_snd);
16283 #ifdef TCP_SAD_DETECTION
16284 		/*
16285 		 * We only care on a cum-ack move if we are in a sack-disabled
16286 		 * state. We have already added in to the ack_count, and we never
16287 		 * would disable on a cum-ack move, so we only care to do the
16288 		 * detection if it may "undo" it, i.e. we were in disabled already.
16289 		 */
16290 		if (rack->sack_attack_disable)
16291 			rack_do_detection(tp, rack, acked_amount, segsiz);
16292 #endif
16293 		if (IN_FASTRECOVERY(tp->t_flags) &&
16294 		    (rack->rack_no_prr == 0))
16295 			rack_update_prr(tp, rack, acked_amount, high_seq);
16296 		if (IN_RECOVERY(tp->t_flags)) {
16297 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16298 			    (SEQ_LT(high_seq, tp->snd_max))) {
16299 				tcp_rack_partialack(tp);
16300 			} else {
16301 				rack_post_recovery(tp, high_seq);
16302 				recovery = 1;
16303 			}
16304 		}
16305 		/* Handle the rack-log-ack part (sendmap) */
16306 		if ((sbused(&so->so_snd) == 0) &&
16307 		    (acked > acked_amount) &&
16308 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16309 		    (tp->t_flags & TF_SENTFIN)) {
16310 			/*
16311 			 * We must be sure our fin
16312 			 * was sent and acked (we can be
16313 			 * in FIN_WAIT_1 without having
16314 			 * sent the fin).
16315 			 */
16316 			ourfinisacked = 1;
16317 			/*
16318 			 * Lets make sure snd_una is updated
16319 			 * since most likely acked_amount = 0 (it
16320 			 * should be).
16321 			 */
16322 			tp->snd_una = high_seq;
16323 		}
16324 		/* Did we make a RTO error? */
16325 		if ((tp->t_flags & TF_PREVVALID) &&
16326 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16327 			tp->t_flags &= ~TF_PREVVALID;
16328 			if (tp->t_rxtshift == 1 &&
16329 			    (int)(ticks - tp->t_badrxtwin) < 0)
16330 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16331 		}
16332 		/* Handle the data in the socket buffer */
16333 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16334 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16335 		if (acked_amount > 0) {
16336 			struct mbuf *mfree;
16337 
16338 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
16339 			SOCKBUF_LOCK(&so->so_snd);
16340 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16341 			tp->snd_una = high_seq;
16342 			/* Note we want to hold the sb lock through the sendmap adjust */
16343 			rack_adjust_sendmap_head(rack, &so->so_snd);
16344 			/* Wake up the socket if we have room to write more */
16345 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16346 			sowwakeup_locked(so);
16347 			if ((recovery == 1) &&
16348 			    (rack->excess_rxt_on) &&
16349 			    (rack->r_cwnd_was_clamped == 0)) {
16350 				do_rack_excess_rxt(tp, rack);
16351 			} else if (rack->r_cwnd_was_clamped)
16352 				do_rack_check_for_unclamp(tp, rack);
16353 			m_freem(mfree);
16354 		}
16355 		/* update progress */
16356 		tp->t_acktime = ticks;
16357 		rack_log_progress_event(rack, tp, tp->t_acktime,
16358 					PROGRESS_UPDATE, __LINE__);
16359 		/* Clear out shifts and such */
16360 		tp->t_rxtshift = 0;
16361 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16362 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16363 		rack->rc_tlp_in_progress = 0;
16364 		rack->r_ctl.rc_tlp_cnt_out = 0;
16365 		/* Send recover and snd_nxt must be dragged along */
16366 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16367 			tp->snd_recover = tp->snd_una;
16368 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
16369 			tp->snd_nxt = tp->snd_una;
16370 		/*
16371 		 * If the RXT timer is running we want to
16372 		 * stop it, so we can restart a TLP (or new RXT).
16373 		 */
16374 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16375 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16376 		tp->snd_wl2 = high_seq;
16377 		tp->t_dupacks = 0;
16378 		if (under_pacing &&
16379 		    (rack->use_fixed_rate == 0) &&
16380 		    (rack->in_probe_rtt == 0) &&
16381 		    rack->rc_gp_dyn_mul &&
16382 		    rack->rc_always_pace) {
16383 			/* Check if we are dragging bottom */
16384 			rack_check_bottom_drag(tp, rack, so);
16385 		}
16386 		if (tp->snd_una == tp->snd_max) {
16387 			tp->t_flags &= ~TF_PREVVALID;
16388 			rack->r_ctl.retran_during_recovery = 0;
16389 			rack->rc_suspicious = 0;
16390 			rack->r_ctl.dsack_byte_cnt = 0;
16391 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16392 			if (rack->r_ctl.rc_went_idle_time == 0)
16393 				rack->r_ctl.rc_went_idle_time = 1;
16394 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16395 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16396 				tp->t_acktime = 0;
16397 			/* Set so we might enter persists... */
16398 			rack->r_wanted_output = 1;
16399 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16400 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16401 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16402 			    (sbavail(&so->so_snd) == 0) &&
16403 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16404 				/*
16405 				 * The socket was gone and the
16406 				 * peer sent data (not now in the past), time to
16407 				 * reset him.
16408 				 */
16409 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16410 				/* tcp_close will kill the inp pre-log the Reset */
16411 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16412 #ifdef TCP_ACCOUNTING
16413 				rdstc = get_cyclecount();
16414 				if (rdstc > ts_val) {
16415 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16416 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16417 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16418 					}
16419 				}
16420 #endif
16421 				m_freem(m);
16422 				tp = tcp_close(tp);
16423 				if (tp == NULL) {
16424 #ifdef TCP_ACCOUNTING
16425 					sched_unpin();
16426 #endif
16427 					return (1);
16428 				}
16429 				/*
16430 				 * We would normally do drop-with-reset which would
16431 				 * send back a reset. We can't since we don't have
16432 				 * all the needed bits. Instead lets arrange for
16433 				 * a call to tcp_output(). That way since we
16434 				 * are in the closed state we will generate a reset.
16435 				 *
16436 				 * Note if tcp_accounting is on we don't unpin since
16437 				 * we do that after the goto label.
16438 				 */
16439 				goto send_out_a_rst;
16440 			}
16441 			if ((sbused(&so->so_snd) == 0) &&
16442 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16443 			    (tp->t_flags & TF_SENTFIN)) {
16444 				/*
16445 				 * If we can't receive any more data, then closing user can
16446 				 * proceed. Starting the timer is contrary to the
16447 				 * specification, but if we don't get a FIN we'll hang
16448 				 * forever.
16449 				 *
16450 				 */
16451 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16452 					soisdisconnected(so);
16453 					tcp_timer_activate(tp, TT_2MSL,
16454 							   (tcp_fast_finwait2_recycle ?
16455 							    tcp_finwait2_timeout :
16456 							    TP_MAXIDLE(tp)));
16457 				}
16458 				if (ourfinisacked == 0) {
16459 					/*
16460 					 * We don't change to fin-wait-2 if we have our fin acked
16461 					 * which means we are probably in TCPS_CLOSING.
16462 					 */
16463 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16464 				}
16465 			}
16466 		}
16467 		/* Wake up the socket if we have room to write more */
16468 		if (sbavail(&so->so_snd)) {
16469 			rack->r_wanted_output = 1;
16470 			if (ctf_progress_timeout_check(tp, true)) {
16471 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16472 							tp, tick, PROGRESS_DROP, __LINE__);
16473 				/*
16474 				 * We cheat here and don't send a RST, we should send one
16475 				 * when the pacer drops the connection.
16476 				 */
16477 #ifdef TCP_ACCOUNTING
16478 				rdstc = get_cyclecount();
16479 				if (rdstc > ts_val) {
16480 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16481 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16482 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16483 					}
16484 				}
16485 				sched_unpin();
16486 #endif
16487 				(void)tcp_drop(tp, ETIMEDOUT);
16488 				m_freem(m);
16489 				return (1);
16490 			}
16491 		}
16492 		if (ourfinisacked) {
16493 			switch(tp->t_state) {
16494 			case TCPS_CLOSING:
16495 #ifdef TCP_ACCOUNTING
16496 				rdstc = get_cyclecount();
16497 				if (rdstc > ts_val) {
16498 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16499 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16500 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16501 					}
16502 				}
16503 				sched_unpin();
16504 #endif
16505 				tcp_twstart(tp);
16506 				m_freem(m);
16507 				return (1);
16508 				break;
16509 			case TCPS_LAST_ACK:
16510 #ifdef TCP_ACCOUNTING
16511 				rdstc = get_cyclecount();
16512 				if (rdstc > ts_val) {
16513 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16514 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16515 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16516 					}
16517 				}
16518 				sched_unpin();
16519 #endif
16520 				tp = tcp_close(tp);
16521 				ctf_do_drop(m, tp);
16522 				return (1);
16523 				break;
16524 			case TCPS_FIN_WAIT_1:
16525 #ifdef TCP_ACCOUNTING
16526 				rdstc = get_cyclecount();
16527 				if (rdstc > ts_val) {
16528 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16529 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16530 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16531 					}
16532 				}
16533 #endif
16534 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16535 					soisdisconnected(so);
16536 					tcp_timer_activate(tp, TT_2MSL,
16537 							   (tcp_fast_finwait2_recycle ?
16538 							    tcp_finwait2_timeout :
16539 							    TP_MAXIDLE(tp)));
16540 				}
16541 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
16542 				break;
16543 			default:
16544 				break;
16545 			}
16546 		}
16547 		if (rack->r_fast_output) {
16548 			/*
16549 			 * We re doing fast output.. can we expand that?
16550 			 */
16551 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16552 		}
16553 #ifdef TCP_ACCOUNTING
16554 		rdstc = get_cyclecount();
16555 		if (rdstc > ts_val) {
16556 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16557 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16558 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16559 			}
16560 		}
16561 
16562 	} else if (win_up_req) {
16563 		rdstc = get_cyclecount();
16564 		if (rdstc > ts_val) {
16565 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16566 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16567 			}
16568 		}
16569 #endif
16570 	}
16571 	/* Now is there a next packet, if so we are done */
16572 	m_freem(m);
16573 	did_out = 0;
16574 	if (nxt_pkt) {
16575 #ifdef TCP_ACCOUNTING
16576 		sched_unpin();
16577 #endif
16578 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16579 		return (0);
16580 	}
16581 	rack_handle_might_revert(tp, rack);
16582 	ctf_calc_rwin(so, tp);
16583 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16584 	send_out_a_rst:
16585 		if (tcp_output(tp) < 0) {
16586 #ifdef TCP_ACCOUNTING
16587 			sched_unpin();
16588 #endif
16589 			return (1);
16590 		}
16591 		did_out = 1;
16592 	}
16593 	if (tp->t_flags2 & TF2_HPTS_CALLS)
16594 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
16595 	rack_free_trim(rack);
16596 #ifdef TCP_ACCOUNTING
16597 	sched_unpin();
16598 #endif
16599 	rack_timer_audit(tp, rack, &so->so_snd);
16600 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16601 	return (0);
16602 }
16603 
16604 #define	TCP_LRO_TS_OPTION \
16605     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16606 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16607 
16608 static int
16609 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16610     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16611     struct timeval *tv)
16612 {
16613 	struct inpcb *inp = tptoinpcb(tp);
16614 	struct socket *so = tptosocket(tp);
16615 #ifdef TCP_ACCOUNTING
16616 	uint64_t ts_val;
16617 #endif
16618 	int32_t thflags, retval, did_out = 0;
16619 	int32_t way_out = 0;
16620 	/*
16621 	 * cts - is the current time from tv (caller gets ts) in microseconds.
16622 	 * ms_cts - is the current time from tv in milliseconds.
16623 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16624 	 */
16625 	uint32_t cts, us_cts, ms_cts;
16626 	uint32_t tiwin, high_seq;
16627 	struct timespec ts;
16628 	struct tcpopt to;
16629 	struct tcp_rack *rack;
16630 	struct rack_sendmap *rsm;
16631 	int32_t prev_state = 0;
16632 	int no_output = 0;
16633 	int slot_remaining = 0;
16634 #ifdef TCP_ACCOUNTING
16635 	int ack_val_set = 0xf;
16636 #endif
16637 	int nsegs;
16638 
16639 	NET_EPOCH_ASSERT();
16640 	INP_WLOCK_ASSERT(inp);
16641 
16642 	/*
16643 	 * tv passed from common code is from either M_TSTMP_LRO or
16644 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16645 	 */
16646 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16647 	if (rack->rack_deferred_inited == 0) {
16648 		/*
16649 		 * If we are the connecting socket we will
16650 		 * hit rack_init() when no sequence numbers
16651 		 * are setup. This makes it so we must defer
16652 		 * some initialization. Call that now.
16653 		 */
16654 		rack_deferred_init(tp, rack);
16655 	}
16656 	/*
16657 	 * Check to see if we need to skip any output plans. This
16658 	 * can happen in the non-LRO path where we are pacing and
16659 	 * must process the ack coming in but need to defer sending
16660 	 * anything becase a pacing timer is running.
16661 	 */
16662 	us_cts = tcp_tv_to_usectick(tv);
16663 	if (m->m_flags & M_ACKCMP) {
16664 		/*
16665 		 * All compressed ack's are ack's by definition so
16666 		 * remove any ack required flag and then do the processing.
16667 		 */
16668 		rack->rc_ack_required = 0;
16669 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16670 	}
16671 	thflags = tcp_get_flags(th);
16672 	if ((rack->rc_always_pace == 1) &&
16673 	    (rack->rc_ack_can_sendout_data == 0) &&
16674 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16675 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16676 		/*
16677 		 * Ok conditions are right for queuing the packets
16678 		 * but we do have to check the flags in the inp, it
16679 		 * could be, if a sack is present, we want to be awoken and
16680 		 * so should process the packets.
16681 		 */
16682 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16683 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16684 			no_output = 1;
16685 		} else {
16686 			/*
16687 			 * If there is no options, or just a
16688 			 * timestamp option, we will want to queue
16689 			 * the packets. This is the same that LRO does
16690 			 * and will need to change with accurate ECN.
16691 			 */
16692 			uint32_t *ts_ptr;
16693 			int optlen;
16694 
16695 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16696 			ts_ptr = (uint32_t *)(th + 1);
16697 			if ((optlen == 0) ||
16698 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
16699 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
16700 				no_output = 1;
16701 		}
16702 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16703 			/*
16704 			 * It is unrealistic to think we can pace in less than
16705 			 * the minimum granularity of the pacer (def:250usec). So
16706 			 * if we have less than that time remaining we should go
16707 			 * ahead and allow output to be "early". We will attempt to
16708 			 * make up for it in any pacing time we try to apply on
16709 			 * the outbound packet.
16710 			 */
16711 			no_output = 0;
16712 		}
16713 	}
16714 	/*
16715 	 * If there is a RST or FIN lets dump out the bw
16716 	 * with a FIN the connection may go on but we
16717 	 * may not.
16718 	 */
16719 	if ((thflags & TH_FIN) || (thflags & TH_RST))
16720 		rack_log_pacing_delay_calc(rack,
16721 					   rack->r_ctl.gp_bw,
16722 					   0,
16723 					   0,
16724 					   rack_get_gp_est(rack), /* delRate */
16725 					   rack_get_lt_bw(rack), /* rttProp */
16726 					   20, __LINE__, NULL, 0);
16727 	if (m->m_flags & M_ACKCMP) {
16728 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16729 	}
16730 	cts = tcp_tv_to_usectick(tv);
16731 	ms_cts =  tcp_tv_to_mssectick(tv);
16732 	nsegs = m->m_pkthdr.lro_nsegs;
16733 	counter_u64_add(rack_proc_non_comp_ack, 1);
16734 #ifdef TCP_ACCOUNTING
16735 	sched_pin();
16736 	if (thflags & TH_ACK)
16737 		ts_val = get_cyclecount();
16738 #endif
16739 	if ((m->m_flags & M_TSTMP) ||
16740 	    (m->m_flags & M_TSTMP_LRO)) {
16741 		mbuf_tstmp2timespec(m, &ts);
16742 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16743 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16744 	} else
16745 		rack->r_ctl.act_rcv_time = *tv;
16746 	kern_prefetch(rack, &prev_state);
16747 	prev_state = 0;
16748 	/*
16749 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
16750 	 * the scale is zero.
16751 	 */
16752 	tiwin = th->th_win << tp->snd_scale;
16753 #ifdef TCP_ACCOUNTING
16754 	if (thflags & TH_ACK) {
16755 		/*
16756 		 * We have a tradeoff here. We can either do what we are
16757 		 * doing i.e. pinning to this CPU and then doing the accounting
16758 		 * <or> we could do a critical enter, setup the rdtsc and cpu
16759 		 * as in below, and then validate we are on the same CPU on
16760 		 * exit. I have choosen to not do the critical enter since
16761 		 * that often will gain you a context switch, and instead lock
16762 		 * us (line above this if) to the same CPU with sched_pin(). This
16763 		 * means we may be context switched out for a higher priority
16764 		 * interupt but we won't be moved to another CPU.
16765 		 *
16766 		 * If this occurs (which it won't very often since we most likely
16767 		 * are running this code in interupt context and only a higher
16768 		 * priority will bump us ... clock?) we will falsely add in
16769 		 * to the time the interupt processing time plus the ack processing
16770 		 * time. This is ok since its a rare event.
16771 		 */
16772 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16773 						    ctf_fixed_maxseg(tp));
16774 	}
16775 #endif
16776 	/*
16777 	 * Parse options on any incoming segment.
16778 	 */
16779 	memset(&to, 0, sizeof(to));
16780 	tcp_dooptions(&to, (u_char *)(th + 1),
16781 	    (th->th_off << 2) - sizeof(struct tcphdr),
16782 	    (thflags & TH_SYN) ? TO_SYN : 0);
16783 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16784 	    __func__));
16785 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16786 	    __func__));
16787 
16788 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16789 	    (tp->t_flags & TF_GPUTINPROG)) {
16790 		/*
16791 		 * We have a goodput in progress
16792 		 * and we have entered a late state.
16793 		 * Do we have enough data in the sb
16794 		 * to handle the GPUT request?
16795 		 */
16796 		uint32_t bytes;
16797 
16798 		bytes = tp->gput_ack - tp->gput_seq;
16799 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16800 			bytes += tp->gput_seq - tp->snd_una;
16801 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16802 			/*
16803 			 * There are not enough bytes in the socket
16804 			 * buffer that have been sent to cover this
16805 			 * measurement. Cancel it.
16806 			 */
16807 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16808 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16809 						   tp->gput_seq,
16810 						   0, 0, 18, __LINE__, NULL, 0);
16811 			tp->t_flags &= ~TF_GPUTINPROG;
16812 		}
16813 	}
16814 	high_seq = th->th_ack;
16815 	if (tcp_bblogging_on(rack->rc_tp)) {
16816 		union tcp_log_stackspecific log;
16817 		struct timeval ltv;
16818 #ifdef TCP_REQUEST_TRK
16819 		struct tcp_sendfile_track *tcp_req;
16820 
16821 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
16822 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16823 		} else {
16824 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16825 		}
16826 #endif
16827 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16828 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16829 		if (rack->rack_no_prr == 0)
16830 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16831 		else
16832 			log.u_bbr.flex1 = 0;
16833 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16834 		log.u_bbr.use_lt_bw <<= 1;
16835 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16836 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16837 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16838 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16839 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16840 		log.u_bbr.flex3 = m->m_flags;
16841 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16842 		log.u_bbr.lost = thflags;
16843 		log.u_bbr.pacing_gain = 0x1;
16844 #ifdef TCP_ACCOUNTING
16845 		log.u_bbr.cwnd_gain = ack_val_set;
16846 #endif
16847 		log.u_bbr.flex7 = 2;
16848 		if (m->m_flags & M_TSTMP) {
16849 			/* Record the hardware timestamp if present */
16850 			mbuf_tstmp2timespec(m, &ts);
16851 			ltv.tv_sec = ts.tv_sec;
16852 			ltv.tv_usec = ts.tv_nsec / 1000;
16853 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16854 		} else if (m->m_flags & M_TSTMP_LRO) {
16855 			/* Record the LRO the arrival timestamp */
16856 			mbuf_tstmp2timespec(m, &ts);
16857 			ltv.tv_sec = ts.tv_sec;
16858 			ltv.tv_usec = ts.tv_nsec / 1000;
16859 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16860 		}
16861 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16862 		/* Log the rcv time */
16863 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16864 #ifdef TCP_REQUEST_TRK
16865 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16866 		log.u_bbr.applimited <<= 8;
16867 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16868 		log.u_bbr.applimited <<= 8;
16869 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16870 		if (tcp_req) {
16871 			/* Copy out any client req info */
16872 			/* seconds */
16873 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16874 			/* useconds */
16875 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16876 			log.u_bbr.rttProp = tcp_req->timestamp;
16877 			log.u_bbr.cur_del_rate = tcp_req->start;
16878 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16879 				log.u_bbr.flex8 |= 1;
16880 			} else {
16881 				log.u_bbr.flex8 |= 2;
16882 				log.u_bbr.bw_inuse = tcp_req->end;
16883 			}
16884 			log.u_bbr.flex6 = tcp_req->start_seq;
16885 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16886 				log.u_bbr.flex8 |= 4;
16887 				log.u_bbr.epoch = tcp_req->end_seq;
16888 			}
16889 		}
16890 #endif
16891 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16892 		    tlen, &log, true, &ltv);
16893 	}
16894 	/* Remove ack required flag if set, we have one  */
16895 	if (thflags & TH_ACK)
16896 		rack->rc_ack_required = 0;
16897 	if (rack->sack_attack_disable > 0) {
16898 		rack->r_ctl.ack_during_sd++;
16899 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16900 	}
16901 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16902 		way_out = 4;
16903 		retval = 0;
16904 		m_freem(m);
16905 		goto done_with_input;
16906 	}
16907 	/*
16908 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16909 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16910 	 */
16911 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16912 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16913 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16914 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16915 #ifdef TCP_ACCOUNTING
16916 		sched_unpin();
16917 #endif
16918 		return (1);
16919 	}
16920 	/*
16921 	 * If timestamps were negotiated during SYN/ACK and a
16922 	 * segment without a timestamp is received, silently drop
16923 	 * the segment, unless it is a RST segment or missing timestamps are
16924 	 * tolerated.
16925 	 * See section 3.2 of RFC 7323.
16926 	 */
16927 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16928 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16929 		way_out = 5;
16930 		retval = 0;
16931 		m_freem(m);
16932 		goto done_with_input;
16933 	}
16934 
16935 	/*
16936 	 * Segment received on connection. Reset idle time and keep-alive
16937 	 * timer. XXX: This should be done after segment validation to
16938 	 * ignore broken/spoofed segs.
16939 	 */
16940 	if  (tp->t_idle_reduce &&
16941 	     (tp->snd_max == tp->snd_una) &&
16942 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16943 		counter_u64_add(rack_input_idle_reduces, 1);
16944 		rack_cc_after_idle(rack, tp);
16945 	}
16946 	tp->t_rcvtime = ticks;
16947 #ifdef STATS
16948 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16949 #endif
16950 	if (tiwin > rack->r_ctl.rc_high_rwnd)
16951 		rack->r_ctl.rc_high_rwnd = tiwin;
16952 	/*
16953 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16954 	 * this to occur after we've validated the segment.
16955 	 */
16956 	if (tcp_ecn_input_segment(tp, thflags, tlen,
16957 	    tcp_packets_this_ack(tp, th->th_ack),
16958 	    iptos))
16959 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16960 
16961 	/*
16962 	 * If echoed timestamp is later than the current time, fall back to
16963 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16964 	 * were used when this connection was established.
16965 	 */
16966 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16967 		to.to_tsecr -= tp->ts_offset;
16968 		if (TSTMP_GT(to.to_tsecr, ms_cts))
16969 			to.to_tsecr = 0;
16970 	}
16971 
16972 	/*
16973 	 * If its the first time in we need to take care of options and
16974 	 * verify we can do SACK for rack!
16975 	 */
16976 	if (rack->r_state == 0) {
16977 		/* Should be init'd by rack_init() */
16978 		KASSERT(rack->rc_inp != NULL,
16979 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
16980 		if (rack->rc_inp == NULL) {
16981 			rack->rc_inp = inp;
16982 		}
16983 
16984 		/*
16985 		 * Process options only when we get SYN/ACK back. The SYN
16986 		 * case for incoming connections is handled in tcp_syncache.
16987 		 * According to RFC1323 the window field in a SYN (i.e., a
16988 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16989 		 * this is traditional behavior, may need to be cleaned up.
16990 		 */
16991 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16992 			/* Handle parallel SYN for ECN */
16993 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16994 			if ((to.to_flags & TOF_SCALE) &&
16995 			    (tp->t_flags & TF_REQ_SCALE)) {
16996 				tp->t_flags |= TF_RCVD_SCALE;
16997 				tp->snd_scale = to.to_wscale;
16998 			} else
16999 				tp->t_flags &= ~TF_REQ_SCALE;
17000 			/*
17001 			 * Initial send window.  It will be updated with the
17002 			 * next incoming segment to the scaled value.
17003 			 */
17004 			tp->snd_wnd = th->th_win;
17005 			rack_validate_fo_sendwin_up(tp, rack);
17006 			if ((to.to_flags & TOF_TS) &&
17007 			    (tp->t_flags & TF_REQ_TSTMP)) {
17008 				tp->t_flags |= TF_RCVD_TSTMP;
17009 				tp->ts_recent = to.to_tsval;
17010 				tp->ts_recent_age = cts;
17011 			} else
17012 				tp->t_flags &= ~TF_REQ_TSTMP;
17013 			if (to.to_flags & TOF_MSS) {
17014 				tcp_mss(tp, to.to_mss);
17015 			}
17016 			if ((tp->t_flags & TF_SACK_PERMIT) &&
17017 			    (to.to_flags & TOF_SACKPERM) == 0)
17018 				tp->t_flags &= ~TF_SACK_PERMIT;
17019 			if (IS_FASTOPEN(tp->t_flags)) {
17020 				if (to.to_flags & TOF_FASTOPEN) {
17021 					uint16_t mss;
17022 
17023 					if (to.to_flags & TOF_MSS)
17024 						mss = to.to_mss;
17025 					else
17026 						if ((inp->inp_vflag & INP_IPV6) != 0)
17027 							mss = TCP6_MSS;
17028 						else
17029 							mss = TCP_MSS;
17030 					tcp_fastopen_update_cache(tp, mss,
17031 					    to.to_tfo_len, to.to_tfo_cookie);
17032 				} else
17033 					tcp_fastopen_disable_path(tp);
17034 			}
17035 		}
17036 		/*
17037 		 * At this point we are at the initial call. Here we decide
17038 		 * if we are doing RACK or not. We do this by seeing if
17039 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
17040 		 * The code now does do dup-ack counting so if you don't
17041 		 * switch back you won't get rack & TLP, but you will still
17042 		 * get this stack.
17043 		 */
17044 
17045 		if ((rack_sack_not_required == 0) &&
17046 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
17047 			tcp_switch_back_to_default(tp);
17048 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
17049 			    tlen, iptos);
17050 #ifdef TCP_ACCOUNTING
17051 			sched_unpin();
17052 #endif
17053 			return (1);
17054 		}
17055 		tcp_set_hpts(tp);
17056 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
17057 	}
17058 	if (thflags & TH_FIN)
17059 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
17060 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
17061 	if ((rack->rc_gp_dyn_mul) &&
17062 	    (rack->use_fixed_rate == 0) &&
17063 	    (rack->rc_always_pace)) {
17064 		/* Check in on probertt */
17065 		rack_check_probe_rtt(rack, us_cts);
17066 	}
17067 	rack_clear_rate_sample(rack);
17068 	if ((rack->forced_ack) &&
17069 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
17070 		rack_handle_probe_response(rack, tiwin, us_cts);
17071 	}
17072 	/*
17073 	 * This is the one exception case where we set the rack state
17074 	 * always. All other times (timers etc) we must have a rack-state
17075 	 * set (so we assure we have done the checks above for SACK).
17076 	 */
17077 	rack->r_ctl.rc_rcvtime = cts;
17078 	if (rack->r_state != tp->t_state)
17079 		rack_set_state(tp, rack);
17080 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
17081 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
17082 		kern_prefetch(rsm, &prev_state);
17083 	prev_state = rack->r_state;
17084 	if ((thflags & TH_RST) &&
17085 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
17086 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
17087 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
17088 		/* The connection will be killed by a reset check the tracepoint */
17089 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
17090 	}
17091 	retval = (*rack->r_substate) (m, th, so,
17092 	    tp, &to, drop_hdrlen,
17093 	    tlen, tiwin, thflags, nxt_pkt, iptos);
17094 	if (retval == 0) {
17095 		/*
17096 		 * If retval is 1 the tcb is unlocked and most likely the tp
17097 		 * is gone.
17098 		 */
17099 		INP_WLOCK_ASSERT(inp);
17100 		if ((rack->rc_gp_dyn_mul) &&
17101 		    (rack->rc_always_pace) &&
17102 		    (rack->use_fixed_rate == 0) &&
17103 		    rack->in_probe_rtt &&
17104 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
17105 			/*
17106 			 * If we are going for target, lets recheck before
17107 			 * we output.
17108 			 */
17109 			rack_check_probe_rtt(rack, us_cts);
17110 		}
17111 		if (rack->set_pacing_done_a_iw == 0) {
17112 			/* How much has been acked? */
17113 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
17114 				/* We have enough to set in the pacing segment size */
17115 				rack->set_pacing_done_a_iw = 1;
17116 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
17117 			}
17118 		}
17119 		tcp_rack_xmit_timer_commit(rack, tp);
17120 #ifdef TCP_ACCOUNTING
17121 		/*
17122 		 * If we set the ack_val_se to what ack processing we are doing
17123 		 * we also want to track how many cycles we burned. Note
17124 		 * the bits after tcp_output we let be "free". This is because
17125 		 * we are also tracking the tcp_output times as well. Note the
17126 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
17127 		 * 0xf cannot be returned and is what we initialize it too to
17128 		 * indicate we are not doing the tabulations.
17129 		 */
17130 		if (ack_val_set != 0xf) {
17131 			uint64_t crtsc;
17132 
17133 			crtsc = get_cyclecount();
17134 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17135 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
17136 			}
17137 		}
17138 #endif
17139 		if ((nxt_pkt == 0) && (no_output == 0)) {
17140 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
17141 do_output_now:
17142 				if (tcp_output(tp) < 0) {
17143 #ifdef TCP_ACCOUNTING
17144 					sched_unpin();
17145 #endif
17146 					return (1);
17147 				}
17148 				did_out = 1;
17149 			}
17150 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
17151 			rack_free_trim(rack);
17152 		} else if ((no_output == 1) &&
17153 			   (nxt_pkt == 0)  &&
17154 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
17155 			/*
17156 			 * We are not in hpts and we had a pacing timer up. Use
17157 			 * the remaining time (slot_remaining) to restart the timer.
17158 			 */
17159 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
17160 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
17161 			rack_free_trim(rack);
17162 		}
17163 		/* Clear the flag, it may have been cleared by output but we may not have  */
17164 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
17165 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
17166 		/* Update any rounds needed */
17167 		if (rack_verbose_logging &&  tcp_bblogging_on(rack->rc_tp))
17168 			rack_log_hystart_event(rack, high_seq, 8);
17169 		/*
17170 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
17171 		 * causes issues when we are just going app limited. Lets
17172 		 * instead use SEQ_GT <or> where its equal but more data
17173 		 * is outstanding.
17174 		 *
17175 		 * Also make sure we are on the last ack of a series. We
17176 		 * have to have all the ack's processed in queue to know
17177 		 * if there is something left outstanding.
17178 		 */
17179 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
17180 		    (rack->rc_new_rnd_needed == 0) &&
17181 		    (nxt_pkt == 0)) {
17182 			rack_log_hystart_event(rack, tp->snd_una, 21);
17183 			rack->r_ctl.current_round++;
17184 			/* Force the next send to setup the next round */
17185 			rack->rc_new_rnd_needed = 1;
17186 			if (CC_ALGO(tp)->newround != NULL) {
17187 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
17188 			}
17189 		}
17190 		if ((nxt_pkt == 0) &&
17191 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
17192 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
17193 		     (tp->t_flags & TF_DELACK) ||
17194 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
17195 		      (tp->t_state <= TCPS_CLOSING)))) {
17196 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
17197 			if ((tp->snd_max == tp->snd_una) &&
17198 			    ((tp->t_flags & TF_DELACK) == 0) &&
17199 			    (tcp_in_hpts(rack->rc_tp)) &&
17200 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
17201 				/* keep alive not needed if we are hptsi output yet */
17202 				;
17203 			} else {
17204 				int late = 0;
17205 				if (tcp_in_hpts(tp)) {
17206 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
17207 						us_cts = tcp_get_usecs(NULL);
17208 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17209 							rack->r_early = 1;
17210 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17211 						} else
17212 							late = 1;
17213 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17214 					}
17215 					tcp_hpts_remove(tp);
17216 				}
17217 				if (late && (did_out == 0)) {
17218 					/*
17219 					 * We are late in the sending
17220 					 * and we did not call the output
17221 					 * (this probably should not happen).
17222 					 */
17223 					goto do_output_now;
17224 				}
17225 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17226 			}
17227 			way_out = 1;
17228 		} else if (nxt_pkt == 0) {
17229 			/* Do we have the correct timer running? */
17230 			rack_timer_audit(tp, rack, &so->so_snd);
17231 			way_out = 2;
17232 		}
17233 	done_with_input:
17234 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17235 		if (did_out)
17236 			rack->r_wanted_output = 0;
17237 	}
17238 #ifdef TCP_ACCOUNTING
17239 	sched_unpin();
17240 #endif
17241 	return (retval);
17242 }
17243 
17244 static void
17245 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17246     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17247 {
17248 	struct timeval tv;
17249 
17250 	/* First lets see if we have old packets */
17251 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17252 		if (ctf_do_queued_segments(tp, 1)) {
17253 			m_freem(m);
17254 			return;
17255 		}
17256 	}
17257 	if (m->m_flags & M_TSTMP_LRO) {
17258 		mbuf_tstmp2timeval(m, &tv);
17259 	} else {
17260 		/* Should not be should we kassert instead? */
17261 		tcp_get_usecs(&tv);
17262 	}
17263 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17264 	    &tv) == 0) {
17265 		INP_WUNLOCK(tptoinpcb(tp));
17266 	}
17267 }
17268 
17269 struct rack_sendmap *
17270 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17271 {
17272 	struct rack_sendmap *rsm = NULL;
17273 	int32_t idx;
17274 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17275 	int no_sack = 0;
17276 
17277 	/* Return the next guy to be re-transmitted */
17278 	if (tqhash_empty(rack->r_ctl.tqh)) {
17279 		return (NULL);
17280 	}
17281 	if (tp->t_flags & TF_SENTFIN) {
17282 		/* retran the end FIN? */
17283 		return (NULL);
17284 	}
17285 	/* ok lets look at this one */
17286 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17287 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17288 		return (rsm);
17289 	}
17290 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17291 		goto check_it;
17292 	}
17293 	rsm = rack_find_lowest_rsm(rack);
17294 	if (rsm == NULL) {
17295 		return (NULL);
17296 	}
17297 check_it:
17298 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) ||
17299 	    (rack->sack_attack_disable > 0)) {
17300 		no_sack = 1;
17301 	}
17302 	if ((no_sack > 0) &&
17303 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17304 		/*
17305 		 * No sack so we automatically do the 3 strikes and
17306 		 * retransmit (no rack timer would be started).
17307 		 */
17308 		return (rsm);
17309 	}
17310 	if (rsm->r_flags & RACK_ACKED) {
17311 		return (NULL);
17312 	}
17313 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17314 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17315 		/* Its not yet ready */
17316 		return (NULL);
17317 	}
17318 	srtt = rack_grab_rtt(tp, rack);
17319 	idx = rsm->r_rtr_cnt - 1;
17320 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17321 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
17322 	if ((tsused == ts_low) ||
17323 	    (TSTMP_LT(tsused, ts_low))) {
17324 		/* No time since sending */
17325 		return (NULL);
17326 	}
17327 	if ((tsused - ts_low) < thresh) {
17328 		/* It has not been long enough yet */
17329 		return (NULL);
17330 	}
17331 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17332 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
17333 	     (rack->sack_attack_disable == 0))) {
17334 		/*
17335 		 * We have passed the dup-ack threshold <or>
17336 		 * a SACK has indicated this is missing.
17337 		 * Note that if you are a declared attacker
17338 		 * it is only the dup-ack threshold that
17339 		 * will cause retransmits.
17340 		 */
17341 		/* log retransmit reason */
17342 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17343 		rack->r_fast_output = 0;
17344 		return (rsm);
17345 	}
17346 	return (NULL);
17347 }
17348 
17349 static void
17350 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
17351 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17352 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17353 {
17354 	if (tcp_bblogging_on(rack->rc_tp)) {
17355 		union tcp_log_stackspecific log;
17356 		struct timeval tv;
17357 
17358 		if (rack_verbose_logging == 0) {
17359 			/*
17360 			 * We are not verbose screen out all but
17361 			 * ones we always want.
17362 			 */
17363 			if ((method != 2) &&
17364 			    (method != 3) &&
17365 			    (method != 7) &&
17366 			    (method != 14) &&
17367 			    (method != 20)) {
17368 				return;
17369 			}
17370 		}
17371 		memset(&log, 0, sizeof(log));
17372 		log.u_bbr.flex1 = slot;
17373 		log.u_bbr.flex2 = len;
17374 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17375 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17376 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17377 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17378 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17379 		log.u_bbr.use_lt_bw <<= 1;
17380 		log.u_bbr.use_lt_bw |= rack->r_late;
17381 		log.u_bbr.use_lt_bw <<= 1;
17382 		log.u_bbr.use_lt_bw |= rack->r_early;
17383 		log.u_bbr.use_lt_bw <<= 1;
17384 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17385 		log.u_bbr.use_lt_bw <<= 1;
17386 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17387 		log.u_bbr.use_lt_bw <<= 1;
17388 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17389 		log.u_bbr.use_lt_bw <<= 1;
17390 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17391 		log.u_bbr.use_lt_bw <<= 1;
17392 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17393 		log.u_bbr.pkt_epoch = line;
17394 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17395 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17396 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17397 		log.u_bbr.bw_inuse = bw_est;
17398 		log.u_bbr.delRate = bw;
17399 		if (rack->r_ctl.gp_bw == 0)
17400 			log.u_bbr.cur_del_rate = 0;
17401 		else
17402 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17403 		log.u_bbr.rttProp = len_time;
17404 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17405 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17406 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17407 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17408 			/* We are in slow start */
17409 			log.u_bbr.flex7 = 1;
17410 		} else {
17411 			/* we are on congestion avoidance */
17412 			log.u_bbr.flex7 = 0;
17413 		}
17414 		log.u_bbr.flex8 = method;
17415 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17416 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17417 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17418 		log.u_bbr.cwnd_gain <<= 1;
17419 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17420 		log.u_bbr.cwnd_gain <<= 1;
17421 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17422 		log.u_bbr.bbr_substate = quality;
17423 		log.u_bbr.bbr_state = rack->dgp_on;
17424 		log.u_bbr.bbr_state <<= 1;
17425 		log.u_bbr.bbr_state |= rack->r_fill_less_agg;
17426 		log.u_bbr.bbr_state <<= 1;
17427 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17428 		log.u_bbr.bbr_state <<= 2;
17429 		log.u_bbr.bbr_state |= rack->r_pacing_discount;
17430 		log.u_bbr.flex7 = ((rack->r_ctl.pacing_discount_amm << 1) | log.u_bbr.flex7);
17431 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17432 		    &rack->rc_inp->inp_socket->so_rcv,
17433 		    &rack->rc_inp->inp_socket->so_snd,
17434 		    BBR_LOG_HPTSI_CALC, 0,
17435 		    0, &log, false, &tv);
17436 	}
17437 }
17438 
17439 static uint32_t
17440 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17441 {
17442 	uint32_t new_tso, user_max, pace_one;
17443 
17444 	user_max = rack->rc_user_set_max_segs * mss;
17445 	if (rack->rc_force_max_seg) {
17446 		return (user_max);
17447 	}
17448 	if (rack->use_fixed_rate &&
17449 	    ((rack->r_ctl.crte == NULL) ||
17450 	     (bw != rack->r_ctl.crte->rate))) {
17451 		/* Use the user mss since we are not exactly matched */
17452 		return (user_max);
17453 	}
17454 	if (rack_pace_one_seg ||
17455 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17456 		pace_one = 1;
17457 	else
17458 		pace_one = 0;
17459 
17460 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17461 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17462 	if (new_tso > user_max)
17463 		new_tso = user_max;
17464 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17465 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17466 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17467 	}
17468 	if (rack->r_ctl.rc_user_set_min_segs &&
17469 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17470 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17471 	return (new_tso);
17472 }
17473 
17474 static uint64_t
17475 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17476 {
17477 	uint64_t reduced_win;
17478 	uint32_t gain;
17479 
17480 	if (window_input < rc_init_window(rack)) {
17481 		/*
17482 		 * The cwnd is collapsed to
17483 		 * nearly zero, maybe because of a time-out?
17484 		 * Lets drop back to the lt-bw.
17485 		 */
17486 		reduced_win = rack_get_lt_bw(rack);
17487 		/* Set the flag so the caller knows its a rate and not a reduced window */
17488 		*rate_set = 1;
17489 		gain = 100;
17490 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
17491 		/*
17492 		 * If we are in recover our cwnd needs to be less for
17493 		 * our pacing consideration.
17494 		 */
17495 		if (rack->rack_hibeta == 0) {
17496 			reduced_win = window_input / 2;
17497 			gain = 50;
17498 		} else {
17499 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
17500 			reduced_win /= 100;
17501 			gain = rack->r_ctl.saved_hibeta;
17502 		}
17503 	} else {
17504 		/*
17505 		 * Apply Timely factor to increase/decrease the
17506 		 * amount we are pacing at.
17507 		 */
17508 		gain = rack_get_output_gain(rack, NULL);
17509 		if (gain > rack_gain_p5_ub) {
17510 			gain = rack_gain_p5_ub;
17511 		}
17512 		reduced_win = window_input * gain;
17513 		reduced_win /= 100;
17514 	}
17515 	if (gain_b != NULL)
17516 		*gain_b = gain;
17517 	/*
17518 	 * What is being returned here is a trimmed down
17519 	 * window values in all cases where rate_set is left
17520 	 * at 0. In one case we actually return the rate (lt_bw).
17521 	 * the "reduced_win" is returned as a slimmed down cwnd that
17522 	 * is then calculated by the caller into a rate when rate_set
17523 	 * is 0.
17524 	 */
17525 	return (reduced_win);
17526 }
17527 
17528 static int32_t
17529 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)
17530 {
17531 	uint64_t lentim, fill_bw;
17532 
17533 	/* Lets first see if we are full, if so continue with normal rate */
17534 	rack->r_via_fill_cw = 0;
17535 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17536 		return (slot);
17537 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17538 		return (slot);
17539 	if (rack->r_ctl.rc_last_us_rtt == 0)
17540 		return (slot);
17541 	if (rack->rc_pace_fill_if_rttin_range &&
17542 	    (rack->r_ctl.rc_last_us_rtt >=
17543 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17544 		/* The rtt is huge, N * smallest, lets not fill */
17545 		return (slot);
17546 	}
17547 	/*
17548 	 * first lets calculate the b/w based on the last us-rtt
17549 	 * and the the smallest send window.
17550 	 */
17551 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17552 	if (rack->rc_fillcw_apply_discount) {
17553 		uint32_t rate_set = 0;
17554 
17555 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17556 		if (rate_set) {
17557 			goto at_lt_bw;
17558 		}
17559 	}
17560 	/* Take the rwnd if its smaller */
17561 	if (fill_bw > rack->rc_tp->snd_wnd)
17562 		fill_bw = rack->rc_tp->snd_wnd;
17563 	/* Now lets make it into a b/w */
17564 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17565 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17566 at_lt_bw:
17567 	if (rack->r_fill_less_agg) {
17568 		/*
17569 		 * We want the average of the rate_wanted
17570 		 * and our fill-cw calculated bw. We also want
17571 		 * to cap any increase to be no more than
17572 		 * X times the lt_bw (where X is the rack_bw_multipler).
17573 		 */
17574 		uint64_t lt_bw, rate;
17575 
17576 		lt_bw = rack_get_lt_bw(rack);
17577 		if (lt_bw > *rate_wanted)
17578 			rate = lt_bw;
17579 		else
17580 			rate = *rate_wanted;
17581 		fill_bw += rate;
17582 		fill_bw /= 2;
17583 		if (rack_bw_multipler && (fill_bw > (rate * rack_bw_multipler))) {
17584 			fill_bw = rate * rack_bw_multipler;
17585 		}
17586 	}
17587 	/* We are below the min b/w */
17588 	if (non_paced)
17589 		*rate_wanted = fill_bw;
17590 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17591 		return (slot);
17592 	rack->r_via_fill_cw = 1;
17593 	if (rack->r_rack_hw_rate_caps &&
17594 	    (rack->r_ctl.crte != NULL)) {
17595 		uint64_t high_rate;
17596 
17597 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17598 		if (fill_bw > high_rate) {
17599 			/* We are capping bw at the highest rate table entry */
17600 			if (*rate_wanted > high_rate) {
17601 				/* The original rate was also capped */
17602 				rack->r_via_fill_cw = 0;
17603 			}
17604 			rack_log_hdwr_pacing(rack,
17605 					     fill_bw, high_rate, __LINE__,
17606 					     0, 3);
17607 			fill_bw = high_rate;
17608 			if (capped)
17609 				*capped = 1;
17610 		}
17611 	} else if ((rack->r_ctl.crte == NULL) &&
17612 		   (rack->rack_hdrw_pacing == 0) &&
17613 		   (rack->rack_hdw_pace_ena) &&
17614 		   rack->r_rack_hw_rate_caps &&
17615 		   (rack->rack_attempt_hdwr_pace == 0) &&
17616 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
17617 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17618 		/*
17619 		 * Ok we may have a first attempt that is greater than our top rate
17620 		 * lets check.
17621 		 */
17622 		uint64_t high_rate;
17623 
17624 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17625 		if (high_rate) {
17626 			if (fill_bw > high_rate) {
17627 				fill_bw = high_rate;
17628 				if (capped)
17629 					*capped = 1;
17630 			}
17631 		}
17632 	}
17633 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17634 		if (rack->rc_hybrid_mode)
17635 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17636 					   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17637 		fill_bw = rack->r_ctl.bw_rate_cap;
17638 	}
17639 	/*
17640 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
17641 	 * in an rtt (unless it was capped), what does that
17642 	 * time wise equate too?
17643 	 */
17644 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17645 	lentim /= fill_bw;
17646 	*rate_wanted = fill_bw;
17647 	if (non_paced || (lentim < slot)) {
17648 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17649 					   0, lentim, 12, __LINE__, NULL, 0);
17650 		return ((int32_t)lentim);
17651 	} else
17652 		return (slot);
17653 }
17654 
17655 static int32_t
17656 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
17657 {
17658 	uint64_t srtt;
17659 	int32_t slot = 0;
17660 	int32_t minslot = 0;
17661 	int can_start_hw_pacing = 1;
17662 	int err;
17663 	int pace_one;
17664 
17665 	if (rack_pace_one_seg ||
17666 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17667 		pace_one = 1;
17668 	else
17669 		pace_one = 0;
17670 	if (rack->rc_always_pace == 0) {
17671 		/*
17672 		 * We use the most optimistic possible cwnd/srtt for
17673 		 * sending calculations. This will make our
17674 		 * calculation anticipate getting more through
17675 		 * quicker then possible. But thats ok we don't want
17676 		 * the peer to have a gap in data sending.
17677 		 */
17678 		uint64_t cwnd, tr_perms = 0;
17679 		int32_t reduce = 0;
17680 
17681 	old_method:
17682 		/*
17683 		 * We keep no precise pacing with the old method
17684 		 * instead we use the pacer to mitigate bursts.
17685 		 */
17686 		if (rack->r_ctl.rc_rack_min_rtt)
17687 			srtt = rack->r_ctl.rc_rack_min_rtt;
17688 		else
17689 			srtt = max(tp->t_srtt, 1);
17690 		if (rack->r_ctl.rc_rack_largest_cwnd)
17691 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17692 		else
17693 			cwnd = rack->r_ctl.cwnd_to_use;
17694 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
17695 		tr_perms = (cwnd * 1000) / srtt;
17696 		if (tr_perms == 0) {
17697 			tr_perms = ctf_fixed_maxseg(tp);
17698 		}
17699 		/*
17700 		 * Calculate how long this will take to drain, if
17701 		 * the calculation comes out to zero, thats ok we
17702 		 * will use send_a_lot to possibly spin around for
17703 		 * more increasing tot_len_this_send to the point
17704 		 * that its going to require a pace, or we hit the
17705 		 * cwnd. Which in that case we are just waiting for
17706 		 * a ACK.
17707 		 */
17708 		slot = len / tr_perms;
17709 		/* Now do we reduce the time so we don't run dry? */
17710 		if (slot && rack_slot_reduction) {
17711 			reduce = (slot / rack_slot_reduction);
17712 			if (reduce < slot) {
17713 				slot -= reduce;
17714 			} else
17715 				slot = 0;
17716 		}
17717 		slot *= HPTS_USEC_IN_MSEC;
17718 		if (rack->rc_pace_to_cwnd) {
17719 			uint64_t rate_wanted = 0;
17720 
17721 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17722 			rack->rc_ack_can_sendout_data = 1;
17723 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17724 		} else
17725 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17726 		/*******************************************************/
17727 		/* RRS: We insert non-paced call to stats here for len */
17728 		/*******************************************************/
17729 	} else {
17730 		uint64_t bw_est, res, lentim, rate_wanted;
17731 		uint32_t segs, oh;
17732 		int capped = 0;
17733 		int prev_fill;
17734 
17735 		if ((rack->r_rr_config == 1) && rsm) {
17736 			return (rack->r_ctl.rc_min_to);
17737 		}
17738 		if (rack->use_fixed_rate) {
17739 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17740 		} else if ((rack->r_ctl.init_rate == 0) &&
17741 			   (rack->r_ctl.gp_bw == 0)) {
17742 			/* no way to yet do an estimate */
17743 			bw_est = rate_wanted = 0;
17744 		} else if (rack->dgp_on) {
17745 			bw_est = rack_get_bw(rack);
17746 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17747 		} else {
17748 			uint32_t gain, rate_set = 0;
17749 
17750 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17751 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17752 			if (rate_set == 0) {
17753 				if (rate_wanted > rack->rc_tp->snd_wnd)
17754 					rate_wanted = rack->rc_tp->snd_wnd;
17755 				/* Now lets make it into a b/w */
17756 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17757 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17758 			}
17759 			bw_est = rate_wanted;
17760 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17761 						   rack->r_ctl.cwnd_to_use,
17762 						   rate_wanted, bw_est,
17763 						   rack->r_ctl.rc_last_us_rtt,
17764 						   88, __LINE__, NULL, gain);
17765 		}
17766 		if ((bw_est == 0) || (rate_wanted == 0) ||
17767 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17768 			/*
17769 			 * No way yet to make a b/w estimate or
17770 			 * our raise is set incorrectly.
17771 			 */
17772 			goto old_method;
17773 		}
17774 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
17775 		/* We need to account for all the overheads */
17776 		segs = (len + segsiz - 1) / segsiz;
17777 		/*
17778 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17779 		 * and how much data we put in each packet. Yes this
17780 		 * means we may be off if we are larger than 1500 bytes
17781 		 * or smaller. But this just makes us more conservative.
17782 		 */
17783 
17784 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17785 		if (rack->r_is_v6) {
17786 #ifdef INET6
17787 			oh += sizeof(struct ip6_hdr);
17788 #endif
17789 		} else {
17790 #ifdef INET
17791 			oh += sizeof(struct ip);
17792 #endif
17793 		}
17794 		/* We add a fixed 14 for the ethernet header */
17795 		oh += 14;
17796 		segs *= oh;
17797 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17798 		res = lentim / rate_wanted;
17799 		slot = (uint32_t)res;
17800 		if (rack_hw_rate_min &&
17801 		    (rate_wanted < rack_hw_rate_min)) {
17802 			can_start_hw_pacing = 0;
17803 			if (rack->r_ctl.crte) {
17804 				/*
17805 				 * Ok we need to release it, we
17806 				 * have fallen too low.
17807 				 */
17808 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17809 				rack->r_ctl.crte = NULL;
17810 				rack->rack_attempt_hdwr_pace = 0;
17811 				rack->rack_hdrw_pacing = 0;
17812 			}
17813 		}
17814 		if (rack->r_ctl.crte &&
17815 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17816 			/*
17817 			 * We want more than the hardware can give us,
17818 			 * don't start any hw pacing.
17819 			 */
17820 			can_start_hw_pacing = 0;
17821 			if (rack->r_rack_hw_rate_caps == 0) {
17822 				/*
17823 				 * Ok we need to release it, we
17824 				 * want more than the card can give us and
17825 				 * no rate cap is in place. Set it up so
17826 				 * when we want less we can retry.
17827 				 */
17828 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17829 				rack->r_ctl.crte = NULL;
17830 				rack->rack_attempt_hdwr_pace = 0;
17831 				rack->rack_hdrw_pacing = 0;
17832 			}
17833 		}
17834 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17835 			/*
17836 			 * We lost our rate somehow, this can happen
17837 			 * if the interface changed underneath us.
17838 			 */
17839 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17840 			rack->r_ctl.crte = NULL;
17841 			/* Lets re-allow attempting to setup pacing */
17842 			rack->rack_hdrw_pacing = 0;
17843 			rack->rack_attempt_hdwr_pace = 0;
17844 			rack_log_hdwr_pacing(rack,
17845 					     rate_wanted, bw_est, __LINE__,
17846 					     0, 6);
17847 		}
17848 		prev_fill = rack->r_via_fill_cw;
17849 		if ((rack->rc_pace_to_cwnd) &&
17850 		    (capped == 0) &&
17851 		    (rack->dgp_on == 1) &&
17852 		    (rack->use_fixed_rate == 0) &&
17853 		    (rack->in_probe_rtt == 0) &&
17854 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17855 			/*
17856 			 * We want to pace at our rate *or* faster to
17857 			 * fill the cwnd to the max if its not full.
17858 			 */
17859 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17860 			/* Re-check to make sure we are not exceeding our max b/w */
17861 			if ((rack->r_ctl.crte != NULL) &&
17862 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17863 				/*
17864 				 * We want more than the hardware can give us,
17865 				 * don't start any hw pacing.
17866 				 */
17867 				can_start_hw_pacing = 0;
17868 				if (rack->r_rack_hw_rate_caps == 0) {
17869 					/*
17870 					 * Ok we need to release it, we
17871 					 * want more than the card can give us and
17872 					 * no rate cap is in place. Set it up so
17873 					 * when we want less we can retry.
17874 					 */
17875 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17876 					rack->r_ctl.crte = NULL;
17877 					rack->rack_attempt_hdwr_pace = 0;
17878 					rack->rack_hdrw_pacing = 0;
17879 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17880 				}
17881 			}
17882 		}
17883 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17884 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17885 			if ((rack->rack_hdw_pace_ena) &&
17886 			    (can_start_hw_pacing > 0) &&
17887 			    (rack->rack_hdrw_pacing == 0) &&
17888 			    (rack->rack_attempt_hdwr_pace == 0)) {
17889 				/*
17890 				 * Lets attempt to turn on hardware pacing
17891 				 * if we can.
17892 				 */
17893 				rack->rack_attempt_hdwr_pace = 1;
17894 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17895 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
17896 								       rate_wanted,
17897 								       RS_PACING_GEQ,
17898 								       &err, &rack->r_ctl.crte_prev_rate);
17899 				if (rack->r_ctl.crte) {
17900 					rack->rack_hdrw_pacing = 1;
17901 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17902 													   pace_one, rack->r_ctl.crte,
17903 													   NULL, rack->r_ctl.pace_len_divisor);
17904 					rack_log_hdwr_pacing(rack,
17905 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17906 							     err, 0);
17907 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17908 				} else {
17909 					counter_u64_add(rack_hw_pace_init_fail, 1);
17910 				}
17911 			} else if (rack->rack_hdrw_pacing &&
17912 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17913 				/* Do we need to adjust our rate? */
17914 				const struct tcp_hwrate_limit_table *nrte;
17915 
17916 				if (rack->r_up_only &&
17917 				    (rate_wanted < rack->r_ctl.crte->rate)) {
17918 					/**
17919 					 * We have four possible states here
17920 					 * having to do with the previous time
17921 					 * and this time.
17922 					 *   previous  |  this-time
17923 					 * A)     0      |     0   -- fill_cw not in the picture
17924 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
17925 					 * C)     1      |     1   -- all rates from fill_cw
17926 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
17927 					 *
17928 					 * For case A, C and D we don't allow a drop. But for
17929 					 * case B where we now our on our steady rate we do
17930 					 * allow a drop.
17931 					 *
17932 					 */
17933 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17934 						goto done_w_hdwr;
17935 				}
17936 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
17937 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17938 					if (rack_hw_rate_to_low &&
17939 					    (bw_est < rack_hw_rate_to_low)) {
17940 						/*
17941 						 * The pacing rate is too low for hardware, but
17942 						 * do allow hardware pacing to be restarted.
17943 						 */
17944 						rack_log_hdwr_pacing(rack,
17945 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
17946 								     0, 5);
17947 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17948 						rack->r_ctl.crte = NULL;
17949 						rack->rack_attempt_hdwr_pace = 0;
17950 						rack->rack_hdrw_pacing = 0;
17951 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17952 						goto done_w_hdwr;
17953 					}
17954 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17955 								   rack->rc_tp,
17956 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
17957 								   rate_wanted,
17958 								   RS_PACING_GEQ,
17959 								   &err, &rack->r_ctl.crte_prev_rate);
17960 					if (nrte == NULL) {
17961 						/*
17962 						 * Lost the rate, lets drop hardware pacing
17963 						 * period.
17964 						 */
17965 						rack->rack_hdrw_pacing = 0;
17966 						rack->r_ctl.crte = NULL;
17967 						rack_log_hdwr_pacing(rack,
17968 								     rate_wanted, 0, __LINE__,
17969 								     err, 1);
17970 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17971 						counter_u64_add(rack_hw_pace_lost, 1);
17972 					} else if (nrte != rack->r_ctl.crte) {
17973 						rack->r_ctl.crte = nrte;
17974 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17975 														   segsiz, pace_one, rack->r_ctl.crte,
17976 														   NULL, rack->r_ctl.pace_len_divisor);
17977 						rack_log_hdwr_pacing(rack,
17978 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17979 								     err, 2);
17980 						rack->r_ctl.last_hw_bw_req = rate_wanted;
17981 					}
17982 				} else {
17983 					/* We just need to adjust the segment size */
17984 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17985 					rack_log_hdwr_pacing(rack,
17986 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17987 							     0, 4);
17988 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17989 				}
17990 			}
17991 		}
17992 		if (minslot && (minslot > slot)) {
17993 			rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
17994 						   98, __LINE__, NULL, 0);
17995 			slot = minslot;
17996 		}
17997 	done_w_hdwr:
17998 		if (rack_limit_time_with_srtt &&
17999 		    (rack->use_fixed_rate == 0) &&
18000 		    (rack->rack_hdrw_pacing == 0)) {
18001 			/*
18002 			 * Sanity check, we do not allow the pacing delay
18003 			 * to be longer than the SRTT of the path. If it is
18004 			 * a slow path, then adding a packet should increase
18005 			 * the RTT and compensate for this i.e. the srtt will
18006 			 * be greater so the allowed pacing time will be greater.
18007 			 *
18008 			 * Note this restriction is not for where a peak rate
18009 			 * is set, we are doing fixed pacing or hardware pacing.
18010 			 */
18011 			if (rack->rc_tp->t_srtt)
18012 				srtt = rack->rc_tp->t_srtt;
18013 			else
18014 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
18015 			if (srtt < (uint64_t)slot) {
18016 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
18017 				slot = srtt;
18018 			}
18019 		}
18020 		/*******************************************************************/
18021 		/* RRS: We insert paced call to stats here for len and rate_wanted */
18022 		/*******************************************************************/
18023 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
18024 	}
18025 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
18026 		/*
18027 		 * If this rate is seeing enobufs when it
18028 		 * goes to send then either the nic is out
18029 		 * of gas or we are mis-estimating the time
18030 		 * somehow and not letting the queue empty
18031 		 * completely. Lets add to the pacing time.
18032 		 */
18033 		int hw_boost_delay;
18034 
18035 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
18036 		if (hw_boost_delay > rack_enobuf_hw_max)
18037 			hw_boost_delay = rack_enobuf_hw_max;
18038 		else if (hw_boost_delay < rack_enobuf_hw_min)
18039 			hw_boost_delay = rack_enobuf_hw_min;
18040 		slot += hw_boost_delay;
18041 	}
18042 	return (slot);
18043 }
18044 
18045 static void
18046 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
18047     tcp_seq startseq, uint32_t sb_offset)
18048 {
18049 	struct rack_sendmap *my_rsm = NULL;
18050 
18051 	if (tp->t_state < TCPS_ESTABLISHED) {
18052 		/*
18053 		 * We don't start any measurements if we are
18054 		 * not at least established.
18055 		 */
18056 		return;
18057 	}
18058 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
18059 		/*
18060 		 * We will get no more data into the SB
18061 		 * this means we need to have the data available
18062 		 * before we start a measurement.
18063 		 */
18064 
18065 		if (sbavail(&tptosocket(tp)->so_snd) <
18066 		    max(rc_init_window(rack),
18067 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
18068 			/* Nope not enough data */
18069 			return;
18070 		}
18071 	}
18072 	tp->t_flags |= TF_GPUTINPROG;
18073 	rack->r_ctl.rc_gp_cumack_ts = 0;
18074 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
18075 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
18076 	tp->gput_seq = startseq;
18077 	rack->app_limited_needs_set = 0;
18078 	if (rack->in_probe_rtt)
18079 		rack->measure_saw_probe_rtt = 1;
18080 	else if ((rack->measure_saw_probe_rtt) &&
18081 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
18082 		rack->measure_saw_probe_rtt = 0;
18083 	if (rack->rc_gp_filled)
18084 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
18085 	else {
18086 		/* Special case initial measurement */
18087 		struct timeval tv;
18088 
18089 		tp->gput_ts = tcp_get_usecs(&tv);
18090 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18091 	}
18092 	/*
18093 	 * We take a guess out into the future,
18094 	 * if we have no measurement and no
18095 	 * initial rate, we measure the first
18096 	 * initial-windows worth of data to
18097 	 * speed up getting some GP measurement and
18098 	 * thus start pacing.
18099 	 */
18100 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
18101 		rack->app_limited_needs_set = 1;
18102 		tp->gput_ack = startseq + max(rc_init_window(rack),
18103 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
18104 		rack_log_pacing_delay_calc(rack,
18105 					   tp->gput_seq,
18106 					   tp->gput_ack,
18107 					   0,
18108 					   tp->gput_ts,
18109 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18110 					   9,
18111 					   __LINE__, NULL, 0);
18112 		rack_tend_gp_marks(tp, rack);
18113 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18114 		return;
18115 	}
18116 	if (sb_offset) {
18117 		/*
18118 		 * We are out somewhere in the sb
18119 		 * can we use the already outstanding data?
18120 		 */
18121 
18122 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
18123 			/*
18124 			 * Yes first one is good and in this case
18125 			 * the tp->gput_ts is correctly set based on
18126 			 * the last ack that arrived (no need to
18127 			 * set things up when an ack comes in).
18128 			 */
18129 			my_rsm = tqhash_min(rack->r_ctl.tqh);
18130 			if ((my_rsm == NULL) ||
18131 			    (my_rsm->r_rtr_cnt != 1)) {
18132 				/* retransmission? */
18133 				goto use_latest;
18134 			}
18135 		} else {
18136 			if (rack->r_ctl.rc_first_appl == NULL) {
18137 				/*
18138 				 * If rc_first_appl is NULL
18139 				 * then the cnt should be 0.
18140 				 * This is probably an error, maybe
18141 				 * a KASSERT would be approprate.
18142 				 */
18143 				goto use_latest;
18144 			}
18145 			/*
18146 			 * If we have a marker pointer to the last one that is
18147 			 * app limited we can use that, but we need to set
18148 			 * things up so that when it gets ack'ed we record
18149 			 * the ack time (if its not already acked).
18150 			 */
18151 			rack->app_limited_needs_set = 1;
18152 			/*
18153 			 * We want to get to the rsm that is either
18154 			 * next with space i.e. over 1 MSS or the one
18155 			 * after that (after the app-limited).
18156 			 */
18157 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
18158 			if (my_rsm) {
18159 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
18160 					/* Have to use the next one */
18161 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18162 				else {
18163 					/* Use after the first MSS of it is acked */
18164 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
18165 					goto start_set;
18166 				}
18167 			}
18168 			if ((my_rsm == NULL) ||
18169 			    (my_rsm->r_rtr_cnt != 1)) {
18170 				/*
18171 				 * Either its a retransmit or
18172 				 * the last is the app-limited one.
18173 				 */
18174 				goto use_latest;
18175 			}
18176 		}
18177 		tp->gput_seq = my_rsm->r_start;
18178 start_set:
18179 		if (my_rsm->r_flags & RACK_ACKED) {
18180 			/*
18181 			 * This one has been acked use the arrival ack time
18182 			 */
18183 			struct rack_sendmap *nrsm;
18184 
18185 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18186 			rack->app_limited_needs_set = 0;
18187 			/*
18188 			 * Ok in this path we need to use the r_end now
18189 			 * since this guy is the starting ack.
18190 			 */
18191 			tp->gput_seq = my_rsm->r_end;
18192 			/*
18193 			 * We also need to adjust up the sendtime
18194 			 * to the send of the next data after my_rsm.
18195 			 */
18196 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18197 			if (nrsm != NULL)
18198 				my_rsm = nrsm;
18199 			else {
18200 				/*
18201 				 * The next as not been sent, thats the
18202 				 * case for using the latest.
18203 				 */
18204 				goto use_latest;
18205 			}
18206 		}
18207 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18208 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
18209 		rack->r_ctl.rc_gp_cumack_ts = 0;
18210 		rack_log_pacing_delay_calc(rack,
18211 					   tp->gput_seq,
18212 					   tp->gput_ack,
18213 					   (uint64_t)my_rsm,
18214 					   tp->gput_ts,
18215 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18216 					   9,
18217 					   __LINE__, my_rsm, 0);
18218 		/* Now lets make sure all are marked as they should be */
18219 		rack_tend_gp_marks(tp, rack);
18220 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18221 		return;
18222 	}
18223 
18224 use_latest:
18225 	/*
18226 	 * We don't know how long we may have been
18227 	 * idle or if this is the first-send. Lets
18228 	 * setup the flag so we will trim off
18229 	 * the first ack'd data so we get a true
18230 	 * measurement.
18231 	 */
18232 	rack->app_limited_needs_set = 1;
18233 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18234 	rack->r_ctl.rc_gp_cumack_ts = 0;
18235 	/* Find this guy so we can pull the send time */
18236 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18237 	if (my_rsm) {
18238 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18239 		if (my_rsm->r_flags & RACK_ACKED) {
18240 			/*
18241 			 * Unlikely since its probably what was
18242 			 * just transmitted (but I am paranoid).
18243 			 */
18244 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18245 			rack->app_limited_needs_set = 0;
18246 		}
18247 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18248 			/* This also is unlikely */
18249 			tp->gput_seq = my_rsm->r_start;
18250 		}
18251 	} else {
18252 		/*
18253 		 * TSNH unless we have some send-map limit,
18254 		 * and even at that it should not be hitting
18255 		 * that limit (we should have stopped sending).
18256 		 */
18257 		struct timeval tv;
18258 
18259 		microuptime(&tv);
18260 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18261 	}
18262 	rack_tend_gp_marks(tp, rack);
18263 	rack_log_pacing_delay_calc(rack,
18264 				   tp->gput_seq,
18265 				   tp->gput_ack,
18266 				   (uint64_t)my_rsm,
18267 				   tp->gput_ts,
18268 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18269 				   9, __LINE__, NULL, 0);
18270 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18271 }
18272 
18273 static inline uint32_t
18274 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18275     uint32_t avail, int32_t sb_offset)
18276 {
18277 	uint32_t len;
18278 	uint32_t sendwin;
18279 
18280 	if (tp->snd_wnd > cwnd_to_use)
18281 		sendwin = cwnd_to_use;
18282 	else
18283 		sendwin = tp->snd_wnd;
18284 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18285 		/* We never want to go over our peers rcv-window */
18286 		len = 0;
18287 	} else {
18288 		uint32_t flight;
18289 
18290 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18291 		if (flight >= sendwin) {
18292 			/*
18293 			 * We have in flight what we are allowed by cwnd (if
18294 			 * it was rwnd blocking it would have hit above out
18295 			 * >= tp->snd_wnd).
18296 			 */
18297 			return (0);
18298 		}
18299 		len = sendwin - flight;
18300 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18301 			/* We would send too much (beyond the rwnd) */
18302 			len = tp->snd_wnd - ctf_outstanding(tp);
18303 		}
18304 		if ((len + sb_offset) > avail) {
18305 			/*
18306 			 * We don't have that much in the SB, how much is
18307 			 * there?
18308 			 */
18309 			len = avail - sb_offset;
18310 		}
18311 	}
18312 	return (len);
18313 }
18314 
18315 static void
18316 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18317 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18318 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18319 {
18320 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18321 		union tcp_log_stackspecific log;
18322 		struct timeval tv;
18323 
18324 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18325 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18326 		log.u_bbr.flex1 = error;
18327 		log.u_bbr.flex2 = flags;
18328 		log.u_bbr.flex3 = rsm_is_null;
18329 		log.u_bbr.flex4 = ipoptlen;
18330 		log.u_bbr.flex5 = tp->rcv_numsacks;
18331 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18332 		log.u_bbr.flex7 = optlen;
18333 		log.u_bbr.flex8 = rack->r_fsb_inited;
18334 		log.u_bbr.applimited = rack->r_fast_output;
18335 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18336 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18337 		log.u_bbr.cwnd_gain = mode;
18338 		log.u_bbr.pkts_out = orig_len;
18339 		log.u_bbr.lt_epoch = len;
18340 		log.u_bbr.delivered = line;
18341 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18342 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18343 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18344 			       len, &log, false, NULL, __func__, __LINE__, &tv);
18345 	}
18346 }
18347 
18348 
18349 static struct mbuf *
18350 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18351 		   struct rack_fast_send_blk *fsb,
18352 		   int32_t seglimit, int32_t segsize, int hw_tls)
18353 {
18354 #ifdef KERN_TLS
18355 	struct ktls_session *tls, *ntls;
18356 #ifdef INVARIANTS
18357 	struct mbuf *start;
18358 #endif
18359 #endif
18360 	struct mbuf *m, *n, **np, *smb;
18361 	struct mbuf *top;
18362 	int32_t off, soff;
18363 	int32_t len = *plen;
18364 	int32_t fragsize;
18365 	int32_t len_cp = 0;
18366 	uint32_t mlen, frags;
18367 
18368 	soff = off = the_off;
18369 	smb = m = the_m;
18370 	np = &top;
18371 	top = NULL;
18372 #ifdef KERN_TLS
18373 	if (hw_tls && (m->m_flags & M_EXTPG))
18374 		tls = m->m_epg_tls;
18375 	else
18376 		tls = NULL;
18377 #ifdef INVARIANTS
18378 	start = m;
18379 #endif
18380 #endif
18381 	while (len > 0) {
18382 		if (m == NULL) {
18383 			*plen = len_cp;
18384 			break;
18385 		}
18386 #ifdef KERN_TLS
18387 		if (hw_tls) {
18388 			if (m->m_flags & M_EXTPG)
18389 				ntls = m->m_epg_tls;
18390 			else
18391 				ntls = NULL;
18392 
18393 			/*
18394 			 * Avoid mixing TLS records with handshake
18395 			 * data or TLS records from different
18396 			 * sessions.
18397 			 */
18398 			if (tls != ntls) {
18399 				MPASS(m != start);
18400 				*plen = len_cp;
18401 				break;
18402 			}
18403 		}
18404 #endif
18405 		mlen = min(len, m->m_len - off);
18406 		if (seglimit) {
18407 			/*
18408 			 * For M_EXTPG mbufs, add 3 segments
18409 			 * + 1 in case we are crossing page boundaries
18410 			 * + 2 in case the TLS hdr/trailer are used
18411 			 * It is cheaper to just add the segments
18412 			 * than it is to take the cache miss to look
18413 			 * at the mbuf ext_pgs state in detail.
18414 			 */
18415 			if (m->m_flags & M_EXTPG) {
18416 				fragsize = min(segsize, PAGE_SIZE);
18417 				frags = 3;
18418 			} else {
18419 				fragsize = segsize;
18420 				frags = 0;
18421 			}
18422 
18423 			/* Break if we really can't fit anymore. */
18424 			if ((frags + 1) >= seglimit) {
18425 				*plen =	len_cp;
18426 				break;
18427 			}
18428 
18429 			/*
18430 			 * Reduce size if you can't copy the whole
18431 			 * mbuf. If we can't copy the whole mbuf, also
18432 			 * adjust len so the loop will end after this
18433 			 * mbuf.
18434 			 */
18435 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18436 				mlen = (seglimit - frags - 1) * fragsize;
18437 				len = mlen;
18438 				*plen = len_cp + len;
18439 			}
18440 			frags += howmany(mlen, fragsize);
18441 			if (frags == 0)
18442 				frags++;
18443 			seglimit -= frags;
18444 			KASSERT(seglimit > 0,
18445 			    ("%s: seglimit went too low", __func__));
18446 		}
18447 		n = m_get(M_NOWAIT, m->m_type);
18448 		*np = n;
18449 		if (n == NULL)
18450 			goto nospace;
18451 		n->m_len = mlen;
18452 		soff += mlen;
18453 		len_cp += n->m_len;
18454 		if (m->m_flags & (M_EXT|M_EXTPG)) {
18455 			n->m_data = m->m_data + off;
18456 			mb_dupcl(n, m);
18457 		} else {
18458 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18459 			    (u_int)n->m_len);
18460 		}
18461 		len -= n->m_len;
18462 		off = 0;
18463 		m = m->m_next;
18464 		np = &n->m_next;
18465 		if (len || (soff == smb->m_len)) {
18466 			/*
18467 			 * We have more so we move forward  or
18468 			 * we have consumed the entire mbuf and
18469 			 * len has fell to 0.
18470 			 */
18471 			soff = 0;
18472 			smb = m;
18473 		}
18474 
18475 	}
18476 	if (fsb != NULL) {
18477 		fsb->m = smb;
18478 		fsb->off = soff;
18479 		if (smb) {
18480 			/*
18481 			 * Save off the size of the mbuf. We do
18482 			 * this so that we can recognize when it
18483 			 * has been trimmed by sbcut() as acks
18484 			 * come in.
18485 			 */
18486 			fsb->o_m_len = smb->m_len;
18487 			fsb->o_t_len = M_TRAILINGROOM(smb);
18488 		} else {
18489 			/*
18490 			 * This is the case where the next mbuf went to NULL. This
18491 			 * means with this copy we have sent everything in the sb.
18492 			 * In theory we could clear the fast_output flag, but lets
18493 			 * not since its possible that we could get more added
18494 			 * and acks that call the extend function which would let
18495 			 * us send more.
18496 			 */
18497 			fsb->o_m_len = 0;
18498 			fsb->o_t_len = 0;
18499 		}
18500 	}
18501 	return (top);
18502 nospace:
18503 	if (top)
18504 		m_freem(top);
18505 	return (NULL);
18506 
18507 }
18508 
18509 /*
18510  * This is a copy of m_copym(), taking the TSO segment size/limit
18511  * constraints into account, and advancing the sndptr as it goes.
18512  */
18513 static struct mbuf *
18514 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18515 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18516 {
18517 	struct mbuf *m, *n;
18518 	int32_t soff;
18519 
18520 	m = rack->r_ctl.fsb.m;
18521 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18522 		/*
18523 		 * The trailing space changed, mbufs can grow
18524 		 * at the tail but they can't shrink from
18525 		 * it, KASSERT that. Adjust the orig_m_len to
18526 		 * compensate for this change.
18527 		 */
18528 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18529 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18530 			 m,
18531 			 rack,
18532 			 (intmax_t)M_TRAILINGROOM(m),
18533 			 rack->r_ctl.fsb.o_t_len,
18534 			 rack->r_ctl.fsb.o_m_len,
18535 			 m->m_len));
18536 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18537 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18538 	}
18539 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18540 		/*
18541 		 * Mbuf shrank, trimmed off the top by an ack, our
18542 		 * offset changes.
18543 		 */
18544 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18545 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18546 			 m, m->m_len,
18547 			 rack, rack->r_ctl.fsb.o_m_len,
18548 			 rack->r_ctl.fsb.off));
18549 
18550 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18551 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18552 		else
18553 			rack->r_ctl.fsb.off = 0;
18554 		rack->r_ctl.fsb.o_m_len = m->m_len;
18555 #ifdef INVARIANTS
18556 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18557 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
18558 		      rack, m);
18559 #endif
18560 	}
18561 	soff = rack->r_ctl.fsb.off;
18562 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18563 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18564 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18565 				 __FUNCTION__,
18566 				 rack, *plen, m, m->m_len));
18567 	/* Save off the right location before we copy and advance */
18568 	*s_soff = soff;
18569 	*s_mb = rack->r_ctl.fsb.m;
18570 	n = rack_fo_base_copym(m, soff, plen,
18571 			       &rack->r_ctl.fsb,
18572 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18573 	return (n);
18574 }
18575 
18576 /* Log the buffer level */
18577 static void
18578 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18579 		     int len, struct timeval *tv,
18580 		     uint32_t cts)
18581 {
18582 	uint32_t p_rate = 0, p_queue = 0, err = 0;
18583 	union tcp_log_stackspecific log;
18584 
18585 #ifdef RATELIMIT
18586 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18587 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
18588 #endif
18589 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18590 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18591 	log.u_bbr.flex1 = p_rate;
18592 	log.u_bbr.flex2 = p_queue;
18593 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18594 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18595 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18596 	log.u_bbr.flex7 = 99;
18597 	log.u_bbr.flex8 = 0;
18598 	log.u_bbr.pkts_out = err;
18599 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
18600 	log.u_bbr.timeStamp = cts;
18601 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18602 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18603 		       len, &log, false, NULL, __func__, __LINE__, tv);
18604 
18605 }
18606 
18607 static uint32_t
18608 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18609 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18610 {
18611 	uint64_t lentime = 0;
18612 #ifdef RATELIMIT
18613 	uint32_t p_rate = 0, p_queue = 0, err;
18614 	union tcp_log_stackspecific log;
18615 	uint64_t bw;
18616 
18617 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18618 	/* Failed or queue is zero */
18619 	if (err || (p_queue == 0)) {
18620 		lentime = 0;
18621 		goto out;
18622 	}
18623 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18624 	if (err) {
18625 		lentime = 0;
18626 		goto out;
18627 	}
18628 	/*
18629 	 * If we reach here we have some bytes in
18630 	 * the queue. The number returned is a value
18631 	 * between 0 and 0xffff where ffff is full
18632 	 * and 0 is empty. So how best to make this into
18633 	 * something usable?
18634 	 *
18635 	 * The "safer" way is lets take the b/w gotten
18636 	 * from the query (which should be our b/w rate)
18637 	 * and pretend that a full send (our rc_pace_max_segs)
18638 	 * is outstanding. We factor it so its as if a full
18639 	 * number of our MSS segment is terms of full
18640 	 * ethernet segments are outstanding.
18641 	 */
18642 	bw = p_rate / 8;
18643 	if (bw) {
18644 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18645 		lentime *= ETHERNET_SEGMENT_SIZE;
18646 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18647 		lentime /= bw;
18648 	} else {
18649 		/* TSNH -- KASSERT? */
18650 		lentime = 0;
18651 	}
18652 out:
18653 	if (tcp_bblogging_on(tp)) {
18654 		memset(&log, 0, sizeof(log));
18655 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18656 		log.u_bbr.flex1 = p_rate;
18657 		log.u_bbr.flex2 = p_queue;
18658 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18659 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18660 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18661 		log.u_bbr.flex7 = 99;
18662 		log.u_bbr.flex8 = 0;
18663 		log.u_bbr.pkts_out = err;
18664 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
18665 		log.u_bbr.cur_del_rate = lentime;
18666 		log.u_bbr.timeStamp = cts;
18667 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18668 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18669 			       len, &log, false, NULL, __func__, __LINE__,tv);
18670 	}
18671 #endif
18672 	return ((uint32_t)lentime);
18673 }
18674 
18675 static int
18676 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18677 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18678 {
18679 	/*
18680 	 * Enter the fast retransmit path. We are given that a sched_pin is
18681 	 * in place (if accounting is compliled in) and the cycle count taken
18682 	 * at the entry is in the ts_val. The concept her is that the rsm
18683 	 * now holds the mbuf offsets and such so we can directly transmit
18684 	 * without a lot of overhead, the len field is already set for
18685 	 * us to prohibit us from sending too much (usually its 1MSS).
18686 	 */
18687 	struct ip *ip = NULL;
18688 	struct udphdr *udp = NULL;
18689 	struct tcphdr *th = NULL;
18690 	struct mbuf *m = NULL;
18691 	struct inpcb *inp;
18692 	uint8_t *cpto;
18693 	struct tcp_log_buffer *lgb;
18694 #ifdef TCP_ACCOUNTING
18695 	uint64_t crtsc;
18696 	int cnt_thru = 1;
18697 #endif
18698 	struct tcpopt to;
18699 	u_char opt[TCP_MAXOLEN];
18700 	uint32_t hdrlen, optlen;
18701 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18702 	uint16_t flags;
18703 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
18704 	uint32_t if_hw_tsomaxsegsize;
18705 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18706 
18707 #ifdef INET6
18708 	struct ip6_hdr *ip6 = NULL;
18709 
18710 	if (rack->r_is_v6) {
18711 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18712 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18713 	} else
18714 #endif				/* INET6 */
18715 	{
18716 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18717 		hdrlen = sizeof(struct tcpiphdr);
18718 	}
18719 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18720 		goto failed;
18721 	}
18722 	if (doing_tlp) {
18723 		/* Its a TLP add the flag, it may already be there but be sure */
18724 		rsm->r_flags |= RACK_TLP;
18725 	} else {
18726 		/* If it was a TLP it is not not on this retransmit */
18727 		rsm->r_flags &= ~RACK_TLP;
18728 	}
18729 	startseq = rsm->r_start;
18730 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18731 	inp = rack->rc_inp;
18732 	to.to_flags = 0;
18733 	flags = tcp_outflags[tp->t_state];
18734 	if (flags & (TH_SYN|TH_RST)) {
18735 		goto failed;
18736 	}
18737 	if (rsm->r_flags & RACK_HAS_FIN) {
18738 		/* We can't send a FIN here */
18739 		goto failed;
18740 	}
18741 	if (flags & TH_FIN) {
18742 		/* We never send a FIN */
18743 		flags &= ~TH_FIN;
18744 	}
18745 	if (tp->t_flags & TF_RCVD_TSTMP) {
18746 		to.to_tsval = ms_cts + tp->ts_offset;
18747 		to.to_tsecr = tp->ts_recent;
18748 		to.to_flags = TOF_TS;
18749 	}
18750 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18751 	/* TCP-MD5 (RFC2385). */
18752 	if (tp->t_flags & TF_SIGNATURE)
18753 		to.to_flags |= TOF_SIGNATURE;
18754 #endif
18755 	optlen = tcp_addoptions(&to, opt);
18756 	hdrlen += optlen;
18757 	udp = rack->r_ctl.fsb.udp;
18758 	if (udp)
18759 		hdrlen += sizeof(struct udphdr);
18760 	if (rack->r_ctl.rc_pace_max_segs)
18761 		max_val = rack->r_ctl.rc_pace_max_segs;
18762 	else if (rack->rc_user_set_max_segs)
18763 		max_val = rack->rc_user_set_max_segs * segsiz;
18764 	else
18765 		max_val = len;
18766 	if ((tp->t_flags & TF_TSO) &&
18767 	    V_tcp_do_tso &&
18768 	    (len > segsiz) &&
18769 	    (tp->t_port == 0))
18770 		tso = 1;
18771 #ifdef INET6
18772 	if (MHLEN < hdrlen + max_linkhdr)
18773 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18774 	else
18775 #endif
18776 		m = m_gethdr(M_NOWAIT, MT_DATA);
18777 	if (m == NULL)
18778 		goto failed;
18779 	m->m_data += max_linkhdr;
18780 	m->m_len = hdrlen;
18781 	th = rack->r_ctl.fsb.th;
18782 	/* Establish the len to send */
18783 	if (len > max_val)
18784 		len = max_val;
18785 	if ((tso) && (len + optlen > segsiz)) {
18786 		uint32_t if_hw_tsomax;
18787 		int32_t max_len;
18788 
18789 		/* extract TSO information */
18790 		if_hw_tsomax = tp->t_tsomax;
18791 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18792 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18793 		/*
18794 		 * Check if we should limit by maximum payload
18795 		 * length:
18796 		 */
18797 		if (if_hw_tsomax != 0) {
18798 			/* compute maximum TSO length */
18799 			max_len = (if_hw_tsomax - hdrlen -
18800 				   max_linkhdr);
18801 			if (max_len <= 0) {
18802 				goto failed;
18803 			} else if (len > max_len) {
18804 				len = max_len;
18805 			}
18806 		}
18807 		if (len <= segsiz) {
18808 			/*
18809 			 * In case there are too many small fragments don't
18810 			 * use TSO:
18811 			 */
18812 			tso = 0;
18813 		}
18814 	} else {
18815 		tso = 0;
18816 	}
18817 	if ((tso == 0) && (len > segsiz))
18818 		len = segsiz;
18819 	(void)tcp_get_usecs(tv);
18820 	if ((len == 0) ||
18821 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
18822 		goto failed;
18823 	}
18824 	th->th_seq = htonl(rsm->r_start);
18825 	th->th_ack = htonl(tp->rcv_nxt);
18826 	/*
18827 	 * The PUSH bit should only be applied
18828 	 * if the full retransmission is made. If
18829 	 * we are sending less than this is the
18830 	 * left hand edge and should not have
18831 	 * the PUSH bit.
18832 	 */
18833 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
18834 	    (len == (rsm->r_end - rsm->r_start)))
18835 		flags |= TH_PUSH;
18836 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18837 	if (th->th_win == 0) {
18838 		tp->t_sndzerowin++;
18839 		tp->t_flags |= TF_RXWIN0SENT;
18840 	} else
18841 		tp->t_flags &= ~TF_RXWIN0SENT;
18842 	if (rsm->r_flags & RACK_TLP) {
18843 		/*
18844 		 * TLP should not count in retran count, but
18845 		 * in its own bin
18846 		 */
18847 		counter_u64_add(rack_tlp_retran, 1);
18848 		counter_u64_add(rack_tlp_retran_bytes, len);
18849 	} else {
18850 		tp->t_sndrexmitpack++;
18851 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18852 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18853 	}
18854 #ifdef STATS
18855 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18856 				 len);
18857 #endif
18858 	if (rsm->m == NULL)
18859 		goto failed;
18860 	if (rsm->m &&
18861 	    ((rsm->orig_m_len != rsm->m->m_len) ||
18862 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18863 		/* Fix up the orig_m_len and possibly the mbuf offset */
18864 		rack_adjust_orig_mlen(rsm);
18865 	}
18866 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18867 	if (len <= segsiz) {
18868 		/*
18869 		 * Must have ran out of mbufs for the copy
18870 		 * shorten it to no longer need tso. Lets
18871 		 * not put on sendalot since we are low on
18872 		 * mbufs.
18873 		 */
18874 		tso = 0;
18875 	}
18876 	if ((m->m_next == NULL) || (len <= 0)){
18877 		goto failed;
18878 	}
18879 	if (udp) {
18880 		if (rack->r_is_v6)
18881 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
18882 		else
18883 			ulen = hdrlen + len - sizeof(struct ip);
18884 		udp->uh_ulen = htons(ulen);
18885 	}
18886 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18887 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18888 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18889 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
18890 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18891 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18892 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18893 #ifdef INET6
18894 		if (rack->r_is_v6) {
18895 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18896 		    ip6->ip6_flow |= htonl(ect << 20);
18897 		}
18898 		else
18899 #endif
18900 		{
18901 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
18902 		    ip->ip_tos |= ect;
18903 		}
18904 	}
18905 	if (rack->r_ctl.crte != NULL) {
18906 		/* See if we can send via the hw queue */
18907 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18908 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
18909 		if (slot == 0)
18910 			ip_sendflag = 0;
18911 	}
18912 	tcp_set_flags(th, flags);
18913 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18914 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18915 	if (to.to_flags & TOF_SIGNATURE) {
18916 		/*
18917 		 * Calculate MD5 signature and put it into the place
18918 		 * determined before.
18919 		 * NOTE: since TCP options buffer doesn't point into
18920 		 * mbuf's data, calculate offset and use it.
18921 		 */
18922 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18923 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18924 			/*
18925 			 * Do not send segment if the calculation of MD5
18926 			 * digest has failed.
18927 			 */
18928 			goto failed;
18929 		}
18930 	}
18931 #endif
18932 #ifdef INET6
18933 	if (rack->r_is_v6) {
18934 		if (tp->t_port) {
18935 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18936 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18937 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18938 			th->th_sum = htons(0);
18939 			UDPSTAT_INC(udps_opackets);
18940 		} else {
18941 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18942 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18943 			th->th_sum = in6_cksum_pseudo(ip6,
18944 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18945 						      0);
18946 		}
18947 	}
18948 #endif
18949 #if defined(INET6) && defined(INET)
18950 	else
18951 #endif
18952 #ifdef INET
18953 	{
18954 		if (tp->t_port) {
18955 			m->m_pkthdr.csum_flags = CSUM_UDP;
18956 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18957 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18958 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18959 			th->th_sum = htons(0);
18960 			UDPSTAT_INC(udps_opackets);
18961 		} else {
18962 			m->m_pkthdr.csum_flags = CSUM_TCP;
18963 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18964 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18965 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18966 									IPPROTO_TCP + len + optlen));
18967 		}
18968 		/* IP version must be set here for ipv4/ipv6 checking later */
18969 		KASSERT(ip->ip_v == IPVERSION,
18970 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18971 	}
18972 #endif
18973 	if (tso) {
18974 		/*
18975 		 * Here we use segsiz since we have no added options besides
18976 		 * any standard timestamp options (no DSACKs or SACKS are sent
18977 		 * via either fast-path).
18978 		 */
18979 		KASSERT(len > segsiz,
18980 			("%s: len <= tso_segsz tp:%p", __func__, tp));
18981 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18982 		m->m_pkthdr.tso_segsz = segsiz;
18983 	}
18984 #ifdef INET6
18985 	if (rack->r_is_v6) {
18986 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18987 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18988 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18989 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18990 		else
18991 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18992 	}
18993 #endif
18994 #if defined(INET) && defined(INET6)
18995 	else
18996 #endif
18997 #ifdef INET
18998 	{
18999 		ip->ip_len = htons(m->m_pkthdr.len);
19000 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19001 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19002 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19003 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19004 				ip->ip_off |= htons(IP_DF);
19005 			}
19006 		} else {
19007 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19008 		}
19009 	}
19010 #endif
19011 	if (doing_tlp == 0) {
19012 		/* Set we retransmitted */
19013 		rack->rc_gp_saw_rec = 1;
19014 	} else {
19015 		/* Its a TLP set ca or ss */
19016 		if (tp->snd_cwnd > tp->snd_ssthresh) {
19017 			/* Set we sent in CA */
19018 			rack->rc_gp_saw_ca = 1;
19019 		} else {
19020 			/* Set we sent in SS */
19021 			rack->rc_gp_saw_ss = 1;
19022 		}
19023 	}
19024 	/* Time to copy in our header */
19025 	cpto = mtod(m, uint8_t *);
19026 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19027 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19028 	if (optlen) {
19029 		bcopy(opt, th + 1, optlen);
19030 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19031 	} else {
19032 		th->th_off = sizeof(struct tcphdr) >> 2;
19033 	}
19034 	if (tcp_bblogging_on(rack->rc_tp)) {
19035 		union tcp_log_stackspecific log;
19036 
19037 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
19038 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
19039 			counter_u64_add(rack_collapsed_win_rxt, 1);
19040 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
19041 		}
19042 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19043 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19044 		if (rack->rack_no_prr)
19045 			log.u_bbr.flex1 = 0;
19046 		else
19047 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19048 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19049 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19050 		log.u_bbr.flex4 = max_val;
19051 		/* Save off the early/late values */
19052 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19053 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19054 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19055 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19056 		if (doing_tlp == 0)
19057 			log.u_bbr.flex8 = 1;
19058 		else
19059 			log.u_bbr.flex8 = 2;
19060 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19061 		log.u_bbr.flex7 = 55;
19062 		log.u_bbr.pkts_out = tp->t_maxseg;
19063 		log.u_bbr.timeStamp = cts;
19064 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19065 		if (rsm && (rsm->r_rtr_cnt > 0)) {
19066 			/*
19067 			 * When we have a retransmit we want to log the
19068 			 * burst at send and flight at send from before.
19069 			 */
19070 			log.u_bbr.flex5 = rsm->r_fas;
19071 			log.u_bbr.bbr_substate = rsm->r_bas;
19072 		} else {
19073 			/*
19074 			 * This is currently unlikely until we do the
19075 			 * packet pair probes but I will add it for completeness.
19076 			 */
19077 			log.u_bbr.flex5 = log.u_bbr.inflight;
19078 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19079 		}
19080 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19081 		log.u_bbr.delivered = 0;
19082 		log.u_bbr.rttProp = (uint64_t)rsm;
19083 		log.u_bbr.delRate = rsm->r_flags;
19084 		log.u_bbr.delRate <<= 31;
19085 		log.u_bbr.delRate |= rack->r_must_retran;
19086 		log.u_bbr.delRate <<= 1;
19087 		log.u_bbr.delRate |= 1;
19088 		log.u_bbr.pkt_epoch = __LINE__;
19089 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19090 				     len, &log, false, NULL, __func__, __LINE__, tv);
19091 	} else
19092 		lgb = NULL;
19093 	if ((rack->r_ctl.crte != NULL) &&
19094 	    tcp_bblogging_on(tp)) {
19095 		rack_log_queue_level(tp, rack, len, tv, cts);
19096 	}
19097 #ifdef INET6
19098 	if (rack->r_is_v6) {
19099 		error = ip6_output(m, inp->in6p_outputopts,
19100 				   &inp->inp_route6,
19101 				   ip_sendflag, NULL, NULL, inp);
19102 	}
19103 	else
19104 #endif
19105 #ifdef INET
19106 	{
19107 		error = ip_output(m, NULL,
19108 				  &inp->inp_route,
19109 				  ip_sendflag, 0, inp);
19110 	}
19111 #endif
19112 	m = NULL;
19113 	if (lgb) {
19114 		lgb->tlb_errno = error;
19115 		lgb = NULL;
19116 	}
19117 	if (error) {
19118 		goto failed;
19119 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
19120 		rack->rc_hw_nobuf = 0;
19121 		rack->r_ctl.rc_agg_delayed = 0;
19122 		rack->r_early = 0;
19123 		rack->r_late = 0;
19124 		rack->r_ctl.rc_agg_early = 0;
19125 	}
19126 
19127 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
19128 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
19129 	if (doing_tlp) {
19130 		rack->rc_tlp_in_progress = 1;
19131 		rack->r_ctl.rc_tlp_cnt_out++;
19132 	}
19133 	if (error == 0) {
19134 		counter_u64_add(rack_total_bytes, len);
19135 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
19136 		if (doing_tlp) {
19137 			rack->rc_last_sent_tlp_past_cumack = 0;
19138 			rack->rc_last_sent_tlp_seq_valid = 1;
19139 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
19140 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
19141 		}
19142 		if (rack->r_ctl.rc_prr_sndcnt >= len)
19143 			rack->r_ctl.rc_prr_sndcnt -= len;
19144 		else
19145 			rack->r_ctl.rc_prr_sndcnt = 0;
19146 	}
19147 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19148 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19149 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19150 		rack->r_ctl.retran_during_recovery += len;
19151 	{
19152 		int idx;
19153 
19154 		idx = (len / segsiz) + 3;
19155 		if (idx >= TCP_MSS_ACCT_ATIMER)
19156 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19157 		else
19158 			counter_u64_add(rack_out_size[idx], 1);
19159 	}
19160 	if (tp->t_rtttime == 0) {
19161 		tp->t_rtttime = ticks;
19162 		tp->t_rtseq = startseq;
19163 		KMOD_TCPSTAT_INC(tcps_segstimed);
19164 	}
19165 	counter_u64_add(rack_fto_rsm_send, 1);
19166 	if (error && (error == ENOBUFS)) {
19167 		if (rack->r_ctl.crte != NULL) {
19168 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
19169 			if (tcp_bblogging_on(rack->rc_tp))
19170 				rack_log_queue_level(tp, rack, len, tv, cts);
19171 		} else
19172 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
19173 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19174 		if (rack->rc_enobuf < 0x7f)
19175 			rack->rc_enobuf++;
19176 		if (slot < (10 * HPTS_USEC_IN_MSEC))
19177 			slot = 10 * HPTS_USEC_IN_MSEC;
19178 		if (rack->r_ctl.crte != NULL) {
19179 			counter_u64_add(rack_saw_enobuf_hw, 1);
19180 			tcp_rl_log_enobuf(rack->r_ctl.crte);
19181 		}
19182 		counter_u64_add(rack_saw_enobuf, 1);
19183 	} else
19184 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
19185 	if ((slot == 0) ||
19186 	    (rack->rc_always_pace == 0) ||
19187 	    (rack->r_rr_config == 1)) {
19188 		/*
19189 		 * We have no pacing set or we
19190 		 * are using old-style rack or
19191 		 * we are overridden to use the old 1ms pacing.
19192 		 */
19193 		slot = rack->r_ctl.rc_min_to;
19194 	}
19195 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
19196 #ifdef TCP_ACCOUNTING
19197 	crtsc = get_cyclecount();
19198 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19199 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19200 	}
19201 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19202 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19203 	}
19204 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19205 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
19206 	}
19207 	sched_unpin();
19208 #endif
19209 	return (0);
19210 failed:
19211 	if (m)
19212 		m_free(m);
19213 	return (-1);
19214 }
19215 
19216 static void
19217 rack_sndbuf_autoscale(struct tcp_rack *rack)
19218 {
19219 	/*
19220 	 * Automatic sizing of send socket buffer.  Often the send buffer
19221 	 * size is not optimally adjusted to the actual network conditions
19222 	 * at hand (delay bandwidth product).  Setting the buffer size too
19223 	 * small limits throughput on links with high bandwidth and high
19224 	 * delay (eg. trans-continental/oceanic links).  Setting the
19225 	 * buffer size too big consumes too much real kernel memory,
19226 	 * especially with many connections on busy servers.
19227 	 *
19228 	 * The criteria to step up the send buffer one notch are:
19229 	 *  1. receive window of remote host is larger than send buffer
19230 	 *     (with a fudge factor of 5/4th);
19231 	 *  2. send buffer is filled to 7/8th with data (so we actually
19232 	 *     have data to make use of it);
19233 	 *  3. send buffer fill has not hit maximal automatic size;
19234 	 *  4. our send window (slow start and cogestion controlled) is
19235 	 *     larger than sent but unacknowledged data in send buffer.
19236 	 *
19237 	 * Note that the rack version moves things much faster since
19238 	 * we want to avoid hitting cache lines in the rack_fast_output()
19239 	 * path so this is called much less often and thus moves
19240 	 * the SB forward by a percentage.
19241 	 */
19242 	struct socket *so;
19243 	struct tcpcb *tp;
19244 	uint32_t sendwin, scaleup;
19245 
19246 	tp = rack->rc_tp;
19247 	so = rack->rc_inp->inp_socket;
19248 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19249 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19250 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19251 		    sbused(&so->so_snd) >=
19252 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19253 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19254 		    sendwin >= (sbused(&so->so_snd) -
19255 		    (tp->snd_nxt - tp->snd_una))) {
19256 			if (rack_autosndbuf_inc)
19257 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19258 			else
19259 				scaleup = V_tcp_autosndbuf_inc;
19260 			if (scaleup < V_tcp_autosndbuf_inc)
19261 				scaleup = V_tcp_autosndbuf_inc;
19262 			scaleup += so->so_snd.sb_hiwat;
19263 			if (scaleup > V_tcp_autosndbuf_max)
19264 				scaleup = V_tcp_autosndbuf_max;
19265 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19266 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19267 		}
19268 	}
19269 }
19270 
19271 static int
19272 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19273 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19274 {
19275 	/*
19276 	 * Enter to do fast output. We are given that the sched_pin is
19277 	 * in place (if accounting is compiled in) and the cycle count taken
19278 	 * at entry is in place in ts_val. The idea here is that
19279 	 * we know how many more bytes needs to be sent (presumably either
19280 	 * during pacing or to fill the cwnd and that was greater than
19281 	 * the max-burst). We have how much to send and all the info we
19282 	 * need to just send.
19283 	 */
19284 #ifdef INET
19285 	struct ip *ip = NULL;
19286 #endif
19287 	struct udphdr *udp = NULL;
19288 	struct tcphdr *th = NULL;
19289 	struct mbuf *m, *s_mb;
19290 	struct inpcb *inp;
19291 	uint8_t *cpto;
19292 	struct tcp_log_buffer *lgb;
19293 #ifdef TCP_ACCOUNTING
19294 	uint64_t crtsc;
19295 #endif
19296 	struct tcpopt to;
19297 	u_char opt[TCP_MAXOLEN];
19298 	uint32_t hdrlen, optlen;
19299 #ifdef TCP_ACCOUNTING
19300 	int cnt_thru = 1;
19301 #endif
19302 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19303 	uint16_t flags;
19304 	uint32_t s_soff;
19305 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19306 	uint32_t if_hw_tsomaxsegsize;
19307 	uint16_t add_flag = RACK_SENT_FP;
19308 #ifdef INET6
19309 	struct ip6_hdr *ip6 = NULL;
19310 
19311 	if (rack->r_is_v6) {
19312 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19313 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19314 	} else
19315 #endif				/* INET6 */
19316 	{
19317 #ifdef INET
19318 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19319 		hdrlen = sizeof(struct tcpiphdr);
19320 #endif
19321 	}
19322 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19323 		m = NULL;
19324 		goto failed;
19325 	}
19326 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19327 	startseq = tp->snd_max;
19328 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19329 	inp = rack->rc_inp;
19330 	len = rack->r_ctl.fsb.left_to_send;
19331 	to.to_flags = 0;
19332 	flags = rack->r_ctl.fsb.tcp_flags;
19333 	if (tp->t_flags & TF_RCVD_TSTMP) {
19334 		to.to_tsval = ms_cts + tp->ts_offset;
19335 		to.to_tsecr = tp->ts_recent;
19336 		to.to_flags = TOF_TS;
19337 	}
19338 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19339 	/* TCP-MD5 (RFC2385). */
19340 	if (tp->t_flags & TF_SIGNATURE)
19341 		to.to_flags |= TOF_SIGNATURE;
19342 #endif
19343 	optlen = tcp_addoptions(&to, opt);
19344 	hdrlen += optlen;
19345 	udp = rack->r_ctl.fsb.udp;
19346 	if (udp)
19347 		hdrlen += sizeof(struct udphdr);
19348 	if (rack->r_ctl.rc_pace_max_segs)
19349 		max_val = rack->r_ctl.rc_pace_max_segs;
19350 	else if (rack->rc_user_set_max_segs)
19351 		max_val = rack->rc_user_set_max_segs * segsiz;
19352 	else
19353 		max_val = len;
19354 	if ((tp->t_flags & TF_TSO) &&
19355 	    V_tcp_do_tso &&
19356 	    (len > segsiz) &&
19357 	    (tp->t_port == 0))
19358 		tso = 1;
19359 again:
19360 #ifdef INET6
19361 	if (MHLEN < hdrlen + max_linkhdr)
19362 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19363 	else
19364 #endif
19365 		m = m_gethdr(M_NOWAIT, MT_DATA);
19366 	if (m == NULL)
19367 		goto failed;
19368 	m->m_data += max_linkhdr;
19369 	m->m_len = hdrlen;
19370 	th = rack->r_ctl.fsb.th;
19371 	/* Establish the len to send */
19372 	if (len > max_val)
19373 		len = max_val;
19374 	if ((tso) && (len + optlen > segsiz)) {
19375 		uint32_t if_hw_tsomax;
19376 		int32_t max_len;
19377 
19378 		/* extract TSO information */
19379 		if_hw_tsomax = tp->t_tsomax;
19380 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19381 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19382 		/*
19383 		 * Check if we should limit by maximum payload
19384 		 * length:
19385 		 */
19386 		if (if_hw_tsomax != 0) {
19387 			/* compute maximum TSO length */
19388 			max_len = (if_hw_tsomax - hdrlen -
19389 				   max_linkhdr);
19390 			if (max_len <= 0) {
19391 				goto failed;
19392 			} else if (len > max_len) {
19393 				len = max_len;
19394 			}
19395 		}
19396 		if (len <= segsiz) {
19397 			/*
19398 			 * In case there are too many small fragments don't
19399 			 * use TSO:
19400 			 */
19401 			tso = 0;
19402 		}
19403 	} else {
19404 		tso = 0;
19405 	}
19406 	if ((tso == 0) && (len > segsiz))
19407 		len = segsiz;
19408 	(void)tcp_get_usecs(tv);
19409 	if ((len == 0) ||
19410 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19411 		goto failed;
19412 	}
19413 	sb_offset = tp->snd_max - tp->snd_una;
19414 	th->th_seq = htonl(tp->snd_max);
19415 	th->th_ack = htonl(tp->rcv_nxt);
19416 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19417 	if (th->th_win == 0) {
19418 		tp->t_sndzerowin++;
19419 		tp->t_flags |= TF_RXWIN0SENT;
19420 	} else
19421 		tp->t_flags &= ~TF_RXWIN0SENT;
19422 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
19423 	KMOD_TCPSTAT_INC(tcps_sndpack);
19424 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19425 #ifdef STATS
19426 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19427 				 len);
19428 #endif
19429 	if (rack->r_ctl.fsb.m == NULL)
19430 		goto failed;
19431 
19432 	/* s_mb and s_soff are saved for rack_log_output */
19433 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19434 				    &s_mb, &s_soff);
19435 	if (len <= segsiz) {
19436 		/*
19437 		 * Must have ran out of mbufs for the copy
19438 		 * shorten it to no longer need tso. Lets
19439 		 * not put on sendalot since we are low on
19440 		 * mbufs.
19441 		 */
19442 		tso = 0;
19443 	}
19444 	if (rack->r_ctl.fsb.rfo_apply_push &&
19445 	    (len == rack->r_ctl.fsb.left_to_send)) {
19446 		tcp_set_flags(th, flags | TH_PUSH);
19447 		add_flag |= RACK_HAD_PUSH;
19448 	}
19449 	if ((m->m_next == NULL) || (len <= 0)){
19450 		goto failed;
19451 	}
19452 	if (udp) {
19453 		if (rack->r_is_v6)
19454 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19455 		else
19456 			ulen = hdrlen + len - sizeof(struct ip);
19457 		udp->uh_ulen = htons(ulen);
19458 	}
19459 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19460 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19461 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19462 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
19463 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19464 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19465 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19466 #ifdef INET6
19467 		if (rack->r_is_v6) {
19468 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19469 			ip6->ip6_flow |= htonl(ect << 20);
19470 		}
19471 		else
19472 #endif
19473 		{
19474 #ifdef INET
19475 			ip->ip_tos &= ~IPTOS_ECN_MASK;
19476 			ip->ip_tos |= ect;
19477 #endif
19478 		}
19479 	}
19480 	tcp_set_flags(th, flags);
19481 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19482 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19483 	if (to.to_flags & TOF_SIGNATURE) {
19484 		/*
19485 		 * Calculate MD5 signature and put it into the place
19486 		 * determined before.
19487 		 * NOTE: since TCP options buffer doesn't point into
19488 		 * mbuf's data, calculate offset and use it.
19489 		 */
19490 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19491 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19492 			/*
19493 			 * Do not send segment if the calculation of MD5
19494 			 * digest has failed.
19495 			 */
19496 			goto failed;
19497 		}
19498 	}
19499 #endif
19500 #ifdef INET6
19501 	if (rack->r_is_v6) {
19502 		if (tp->t_port) {
19503 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19504 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19505 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19506 			th->th_sum = htons(0);
19507 			UDPSTAT_INC(udps_opackets);
19508 		} else {
19509 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19510 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19511 			th->th_sum = in6_cksum_pseudo(ip6,
19512 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19513 						      0);
19514 		}
19515 	}
19516 #endif
19517 #if defined(INET6) && defined(INET)
19518 	else
19519 #endif
19520 #ifdef INET
19521 	{
19522 		if (tp->t_port) {
19523 			m->m_pkthdr.csum_flags = CSUM_UDP;
19524 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19525 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19526 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19527 			th->th_sum = htons(0);
19528 			UDPSTAT_INC(udps_opackets);
19529 		} else {
19530 			m->m_pkthdr.csum_flags = CSUM_TCP;
19531 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19532 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19533 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19534 									IPPROTO_TCP + len + optlen));
19535 		}
19536 		/* IP version must be set here for ipv4/ipv6 checking later */
19537 		KASSERT(ip->ip_v == IPVERSION,
19538 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19539 	}
19540 #endif
19541 	if (tso) {
19542 		/*
19543 		 * Here we use segsiz since we have no added options besides
19544 		 * any standard timestamp options (no DSACKs or SACKS are sent
19545 		 * via either fast-path).
19546 		 */
19547 		KASSERT(len > segsiz,
19548 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19549 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19550 		m->m_pkthdr.tso_segsz = segsiz;
19551 	}
19552 #ifdef INET6
19553 	if (rack->r_is_v6) {
19554 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19555 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19556 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19557 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19558 		else
19559 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19560 	}
19561 #endif
19562 #if defined(INET) && defined(INET6)
19563 	else
19564 #endif
19565 #ifdef INET
19566 	{
19567 		ip->ip_len = htons(m->m_pkthdr.len);
19568 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19569 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19570 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19571 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19572 				ip->ip_off |= htons(IP_DF);
19573 			}
19574 		} else {
19575 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19576 		}
19577 	}
19578 #endif
19579 	if (tp->snd_cwnd > tp->snd_ssthresh) {
19580 		/* Set we sent in CA */
19581 		rack->rc_gp_saw_ca = 1;
19582 	} else {
19583 		/* Set we sent in SS */
19584 		rack->rc_gp_saw_ss = 1;
19585 	}
19586 	/* Time to copy in our header */
19587 	cpto = mtod(m, uint8_t *);
19588 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19589 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19590 	if (optlen) {
19591 		bcopy(opt, th + 1, optlen);
19592 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19593 	} else {
19594 		th->th_off = sizeof(struct tcphdr) >> 2;
19595 	}
19596 	if ((rack->r_ctl.crte != NULL) &&
19597 	    tcp_bblogging_on(tp)) {
19598 		rack_log_queue_level(tp, rack, len, tv, cts);
19599 	}
19600 	if (tcp_bblogging_on(rack->rc_tp)) {
19601 		union tcp_log_stackspecific log;
19602 
19603 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19604 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19605 		if (rack->rack_no_prr)
19606 			log.u_bbr.flex1 = 0;
19607 		else
19608 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19609 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19610 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19611 		log.u_bbr.flex4 = max_val;
19612 		/* Save off the early/late values */
19613 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19614 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19615 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19616 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19617 		log.u_bbr.flex8 = 0;
19618 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19619 		log.u_bbr.flex7 = 44;
19620 		log.u_bbr.pkts_out = tp->t_maxseg;
19621 		log.u_bbr.timeStamp = cts;
19622 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19623 		log.u_bbr.flex5 = log.u_bbr.inflight;
19624 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19625 		log.u_bbr.delivered = 0;
19626 		log.u_bbr.rttProp = 0;
19627 		log.u_bbr.delRate = rack->r_must_retran;
19628 		log.u_bbr.delRate <<= 1;
19629 		log.u_bbr.pkt_epoch = __LINE__;
19630 		/* For fast output no retrans so just inflight and how many mss we send */
19631 		log.u_bbr.flex5 = log.u_bbr.inflight;
19632 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19633 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19634 				     len, &log, false, NULL, __func__, __LINE__, tv);
19635 	} else
19636 		lgb = NULL;
19637 #ifdef INET6
19638 	if (rack->r_is_v6) {
19639 		error = ip6_output(m, inp->in6p_outputopts,
19640 				   &inp->inp_route6,
19641 				   0, NULL, NULL, inp);
19642 	}
19643 #endif
19644 #if defined(INET) && defined(INET6)
19645 	else
19646 #endif
19647 #ifdef INET
19648 	{
19649 		error = ip_output(m, NULL,
19650 				  &inp->inp_route,
19651 				  0, 0, inp);
19652 	}
19653 #endif
19654 	if (lgb) {
19655 		lgb->tlb_errno = error;
19656 		lgb = NULL;
19657 	}
19658 	if (error) {
19659 		*send_err = error;
19660 		m = NULL;
19661 		goto failed;
19662 	} else if (rack->rc_hw_nobuf) {
19663 		rack->rc_hw_nobuf = 0;
19664 		rack->r_ctl.rc_agg_delayed = 0;
19665 		rack->r_early = 0;
19666 		rack->r_late = 0;
19667 		rack->r_ctl.rc_agg_early = 0;
19668 	}
19669 	if ((error == 0) && (rack->lt_bw_up == 0)) {
19670 		/* Unlikely */
19671 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19672 		rack->r_ctl.lt_seq = tp->snd_una;
19673 		rack->lt_bw_up = 1;
19674 	}
19675 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19676 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19677 	m = NULL;
19678 	if (tp->snd_una == tp->snd_max) {
19679 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19680 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19681 		tp->t_acktime = ticks;
19682 	}
19683 	counter_u64_add(rack_total_bytes, len);
19684 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19685 
19686 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19687 	tot_len += len;
19688 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
19689 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19690 	tp->snd_max += len;
19691 	tp->snd_nxt = tp->snd_max;
19692 	if (rack->rc_new_rnd_needed) {
19693 		/*
19694 		 * Update the rnd to start ticking not
19695 		 * that from a time perspective all of
19696 		 * the preceding idle time is "in the round"
19697 		 */
19698 		rack->rc_new_rnd_needed = 0;
19699 		rack->r_ctl.roundends = tp->snd_max;
19700 	}
19701 	{
19702 		int idx;
19703 
19704 		idx = (len / segsiz) + 3;
19705 		if (idx >= TCP_MSS_ACCT_ATIMER)
19706 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19707 		else
19708 			counter_u64_add(rack_out_size[idx], 1);
19709 	}
19710 	if (len <= rack->r_ctl.fsb.left_to_send)
19711 		rack->r_ctl.fsb.left_to_send -= len;
19712 	else
19713 		rack->r_ctl.fsb.left_to_send = 0;
19714 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
19715 		rack->r_fast_output = 0;
19716 		rack->r_ctl.fsb.left_to_send = 0;
19717 		/* At the end of fast_output scale up the sb */
19718 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
19719 		rack_sndbuf_autoscale(rack);
19720 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
19721 	}
19722 	if (tp->t_rtttime == 0) {
19723 		tp->t_rtttime = ticks;
19724 		tp->t_rtseq = startseq;
19725 		KMOD_TCPSTAT_INC(tcps_segstimed);
19726 	}
19727 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19728 	    (max_val > len) &&
19729 	    (tso == 0)) {
19730 		max_val -= len;
19731 		len = segsiz;
19732 		th = rack->r_ctl.fsb.th;
19733 #ifdef TCP_ACCOUNTING
19734 		cnt_thru++;
19735 #endif
19736 		goto again;
19737 	}
19738 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19739 	counter_u64_add(rack_fto_send, 1);
19740 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
19741 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19742 #ifdef TCP_ACCOUNTING
19743 	crtsc = get_cyclecount();
19744 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19745 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19746 	}
19747 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19748 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19749 	}
19750 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19751 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19752 	}
19753 	sched_unpin();
19754 #endif
19755 	return (0);
19756 failed:
19757 	if (m)
19758 		m_free(m);
19759 	rack->r_fast_output = 0;
19760 	return (-1);
19761 }
19762 
19763 static inline void
19764 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19765 		       struct sockbuf *sb,
19766 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19767 		       bool hw_tls,
19768 		       uint16_t flags)
19769 {
19770 	rack->r_fast_output = 1;
19771 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19772 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19773 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19774 	rack->r_ctl.fsb.tcp_flags = flags;
19775 	rack->r_ctl.fsb.left_to_send = orig_len - len;
19776 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19777 		/* Less than a full sized pace, lets not  */
19778 		rack->r_fast_output = 0;
19779 		return;
19780 	} else {
19781 		/* Round down to the nearest pace_max_seg */
19782 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19783 	}
19784 	if (hw_tls)
19785 		rack->r_ctl.fsb.hw_tls = 1;
19786 	else
19787 		rack->r_ctl.fsb.hw_tls = 0;
19788 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19789 		("rack:%p left_to_send:%u sbavail:%u out:%u",
19790 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19791 		 (tp->snd_max - tp->snd_una)));
19792 	if (rack->r_ctl.fsb.left_to_send < segsiz)
19793 		rack->r_fast_output = 0;
19794 	else {
19795 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19796 			rack->r_ctl.fsb.rfo_apply_push = 1;
19797 		else
19798 			rack->r_ctl.fsb.rfo_apply_push = 0;
19799 	}
19800 }
19801 
19802 static uint32_t
19803 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19804 {
19805 	uint64_t min_time;
19806 	uint32_t maxlen;
19807 
19808 	min_time = (uint64_t)get_hpts_min_sleep_time();
19809 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19810 	maxlen = roundup(maxlen, segsiz);
19811 	return (maxlen);
19812 }
19813 
19814 static struct rack_sendmap *
19815 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19816 {
19817 	struct rack_sendmap *rsm = NULL;
19818 	int thresh;
19819 
19820 restart:
19821 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19822 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19823 		/* Nothing, strange turn off validity  */
19824 		rack->r_collapse_point_valid = 0;
19825 		return (NULL);
19826 	}
19827 	/* Can we send it yet? */
19828 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19829 		/*
19830 		 * Receiver window has not grown enough for
19831 		 * the segment to be put on the wire.
19832 		 */
19833 		return (NULL);
19834 	}
19835 	if (rsm->r_flags & RACK_ACKED) {
19836 		/*
19837 		 * It has been sacked, lets move to the
19838 		 * next one if possible.
19839 		 */
19840 		rack->r_ctl.last_collapse_point = rsm->r_end;
19841 		/* Are we done? */
19842 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19843 			    rack->r_ctl.high_collapse_point)) {
19844 			rack->r_collapse_point_valid = 0;
19845 			return (NULL);
19846 		}
19847 		goto restart;
19848 	}
19849 	/* Now has it been long enough ? */
19850 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
19851 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19852 		rack_log_collapse(rack, rsm->r_start,
19853 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19854 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
19855 		return (rsm);
19856 	}
19857 	/* Not enough time */
19858 	rack_log_collapse(rack, rsm->r_start,
19859 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19860 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
19861 	return (NULL);
19862 }
19863 
19864 static inline void
19865 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19866 {
19867 	if ((rack->full_size_rxt == 0) &&
19868 	    (rack->shape_rxt_to_pacing_min == 0) &&
19869 	    (*len >= segsiz)) {
19870 		*len = segsiz;
19871 	} else if (rack->shape_rxt_to_pacing_min &&
19872 		 rack->gp_ready) {
19873 		/* We use pacing min as shaping len req */
19874 		uint32_t maxlen;
19875 
19876 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19877 		if (*len > maxlen)
19878 			*len = maxlen;
19879 	} else {
19880 		/*
19881 		 * The else is full_size_rxt is on so send it all
19882 		 * note we do need to check this for exceeding
19883 		 * our max segment size due to the fact that
19884 		 * we do sometimes merge chunks together i.e.
19885 		 * we cannot just assume that we will never have
19886 		 * a chunk greater than pace_max_seg
19887 		 */
19888 		if (*len > pace_max_seg)
19889 			*len = pace_max_seg;
19890 	}
19891 }
19892 
19893 static int
19894 rack_output(struct tcpcb *tp)
19895 {
19896 	struct socket *so;
19897 	uint32_t recwin;
19898 	uint32_t sb_offset, s_moff = 0;
19899 	int32_t len, error = 0;
19900 	uint16_t flags;
19901 	struct mbuf *m, *s_mb = NULL;
19902 	struct mbuf *mb;
19903 	uint32_t if_hw_tsomaxsegcount = 0;
19904 	uint32_t if_hw_tsomaxsegsize;
19905 	int32_t segsiz, minseg;
19906 	long tot_len_this_send = 0;
19907 #ifdef INET
19908 	struct ip *ip = NULL;
19909 #endif
19910 	struct udphdr *udp = NULL;
19911 	struct tcp_rack *rack;
19912 	struct tcphdr *th;
19913 	uint8_t pass = 0;
19914 	uint8_t mark = 0;
19915 	uint8_t check_done = 0;
19916 	uint8_t wanted_cookie = 0;
19917 	u_char opt[TCP_MAXOLEN];
19918 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
19919 	uint32_t rack_seq;
19920 
19921 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19922 	unsigned ipsec_optlen = 0;
19923 
19924 #endif
19925 	int32_t idle, sendalot;
19926 	int32_t sub_from_prr = 0;
19927 	volatile int32_t sack_rxmit;
19928 	struct rack_sendmap *rsm = NULL;
19929 	int32_t tso, mtu;
19930 	struct tcpopt to;
19931 	int32_t slot = 0;
19932 	int32_t sup_rack = 0;
19933 	uint32_t cts, ms_cts, delayed, early;
19934 	uint16_t add_flag = RACK_SENT_SP;
19935 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19936 	uint8_t doing_tlp = 0;
19937 	uint32_t cwnd_to_use, pace_max_seg;
19938 	int32_t do_a_prefetch = 0;
19939 	int32_t prefetch_rsm = 0;
19940 	int32_t orig_len = 0;
19941 	struct timeval tv;
19942 	int32_t prefetch_so_done = 0;
19943 	struct tcp_log_buffer *lgb;
19944 	struct inpcb *inp = tptoinpcb(tp);
19945 	struct sockbuf *sb;
19946 	uint64_t ts_val = 0;
19947 #ifdef TCP_ACCOUNTING
19948 	uint64_t crtsc;
19949 #endif
19950 #ifdef INET6
19951 	struct ip6_hdr *ip6 = NULL;
19952 	int32_t isipv6;
19953 #endif
19954 	bool hpts_calling, hw_tls = false;
19955 
19956 	NET_EPOCH_ASSERT();
19957 	INP_WLOCK_ASSERT(inp);
19958 
19959 	/* setup and take the cache hits here */
19960 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19961 #ifdef TCP_ACCOUNTING
19962 	sched_pin();
19963 	ts_val = get_cyclecount();
19964 #endif
19965 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19966 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
19967 #ifdef TCP_OFFLOAD
19968 	if (tp->t_flags & TF_TOE) {
19969 #ifdef TCP_ACCOUNTING
19970 		sched_unpin();
19971 #endif
19972 		return (tcp_offload_output(tp));
19973 	}
19974 #endif
19975 	if (rack->rack_deferred_inited == 0) {
19976 		/*
19977 		 * If we are the connecting socket we will
19978 		 * hit rack_init() when no sequence numbers
19979 		 * are setup. This makes it so we must defer
19980 		 * some initialization. Call that now.
19981 		 */
19982 		rack_deferred_init(tp, rack);
19983 	}
19984 	/*
19985 	 * For TFO connections in SYN_RECEIVED, only allow the initial
19986 	 * SYN|ACK and those sent by the retransmit timer.
19987 	 */
19988 	if (IS_FASTOPEN(tp->t_flags) &&
19989 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
19990 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19991 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19992 #ifdef TCP_ACCOUNTING
19993 		sched_unpin();
19994 #endif
19995 		return (0);
19996 	}
19997 #ifdef INET6
19998 	if (rack->r_state) {
19999 		/* Use the cache line loaded if possible */
20000 		isipv6 = rack->r_is_v6;
20001 	} else {
20002 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
20003 	}
20004 #endif
20005 	early = 0;
20006 	cts = tcp_get_usecs(&tv);
20007 	ms_cts = tcp_tv_to_mssectick(&tv);
20008 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
20009 	    tcp_in_hpts(rack->rc_tp)) {
20010 		/*
20011 		 * We are on the hpts for some timer but not hptsi output.
20012 		 * Remove from the hpts unconditionally.
20013 		 */
20014 		rack_timer_cancel(tp, rack, cts, __LINE__);
20015 	}
20016 	/* Are we pacing and late? */
20017 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20018 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
20019 		/* We are delayed */
20020 		delayed = cts - rack->r_ctl.rc_last_output_to;
20021 	} else {
20022 		delayed = 0;
20023 	}
20024 	/* Do the timers, which may override the pacer */
20025 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
20026 		int retval;
20027 
20028 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
20029 					     &doing_tlp);
20030 		if (retval != 0) {
20031 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
20032 #ifdef TCP_ACCOUNTING
20033 			sched_unpin();
20034 #endif
20035 			/*
20036 			 * If timers want tcp_drop(), then pass error out,
20037 			 * otherwise suppress it.
20038 			 */
20039 			return (retval < 0 ? retval : 0);
20040 		}
20041 	}
20042 	if (rack->rc_in_persist) {
20043 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20044 			/* Timer is not running */
20045 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20046 		}
20047 #ifdef TCP_ACCOUNTING
20048 		sched_unpin();
20049 #endif
20050 		return (0);
20051 	}
20052 	if ((rack->rc_ack_required == 1) &&
20053 	    (rack->r_timer_override == 0)){
20054 		/* A timeout occurred and no ack has arrived */
20055 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20056 			/* Timer is not running */
20057 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20058 		}
20059 #ifdef TCP_ACCOUNTING
20060 		sched_unpin();
20061 #endif
20062 		return (0);
20063 	}
20064 	if ((rack->r_timer_override) ||
20065 	    (rack->rc_ack_can_sendout_data) ||
20066 	    (delayed) ||
20067 	    (tp->t_state < TCPS_ESTABLISHED)) {
20068 		rack->rc_ack_can_sendout_data = 0;
20069 		if (tcp_in_hpts(rack->rc_tp))
20070 			tcp_hpts_remove(rack->rc_tp);
20071 	} else if (tcp_in_hpts(rack->rc_tp)) {
20072 		/*
20073 		 * On the hpts you can't pass even if ACKNOW is on, we will
20074 		 * when the hpts fires.
20075 		 */
20076 #ifdef TCP_ACCOUNTING
20077 		crtsc = get_cyclecount();
20078 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20079 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
20080 		}
20081 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20082 			tp->tcp_cnt_counters[SND_BLOCKED]++;
20083 		}
20084 		sched_unpin();
20085 #endif
20086 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
20087 		return (0);
20088 	}
20089 	/* Finish out both pacing early and late accounting */
20090 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20091 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
20092 		early = rack->r_ctl.rc_last_output_to - cts;
20093 	} else
20094 		early = 0;
20095 	if (delayed) {
20096 		rack->r_ctl.rc_agg_delayed += delayed;
20097 		rack->r_late = 1;
20098 	} else if (early) {
20099 		rack->r_ctl.rc_agg_early += early;
20100 		rack->r_early = 1;
20101 	}
20102 	/* Now that early/late accounting is done turn off the flag */
20103 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
20104 	rack->r_wanted_output = 0;
20105 	rack->r_timer_override = 0;
20106 	if ((tp->t_state != rack->r_state) &&
20107 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
20108 		rack_set_state(tp, rack);
20109 	}
20110 	if ((rack->r_fast_output) &&
20111 	    (doing_tlp == 0) &&
20112 	    (tp->rcv_numsacks == 0)) {
20113 		int ret;
20114 
20115 		error = 0;
20116 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
20117 		if (ret >= 0)
20118 			return(ret);
20119 		else if (error) {
20120 			inp = rack->rc_inp;
20121 			so = inp->inp_socket;
20122 			sb = &so->so_snd;
20123 			goto nomore;
20124 		}
20125 	}
20126 	inp = rack->rc_inp;
20127 	/*
20128 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
20129 	 * only allow the initial SYN or SYN|ACK and those sent
20130 	 * by the retransmit timer.
20131 	 */
20132 	if (IS_FASTOPEN(tp->t_flags) &&
20133 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
20134 	     (tp->t_state == TCPS_SYN_SENT)) &&
20135 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
20136 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
20137 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20138 		so = inp->inp_socket;
20139 		sb = &so->so_snd;
20140 		goto just_return_nolock;
20141 	}
20142 	/*
20143 	 * Determine length of data that should be transmitted, and flags
20144 	 * that will be used. If there is some data or critical controls
20145 	 * (SYN, RST) to send, then transmit; otherwise, investigate
20146 	 * further.
20147 	 */
20148 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
20149 	if (tp->t_idle_reduce) {
20150 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
20151 			rack_cc_after_idle(rack, tp);
20152 	}
20153 	tp->t_flags &= ~TF_LASTIDLE;
20154 	if (idle) {
20155 		if (tp->t_flags & TF_MORETOCOME) {
20156 			tp->t_flags |= TF_LASTIDLE;
20157 			idle = 0;
20158 		}
20159 	}
20160 	if ((tp->snd_una == tp->snd_max) &&
20161 	    rack->r_ctl.rc_went_idle_time &&
20162 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
20163 		idle = cts - rack->r_ctl.rc_went_idle_time;
20164 		if (idle > rack_min_probertt_hold) {
20165 			/* Count as a probe rtt */
20166 			if (rack->in_probe_rtt == 0) {
20167 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
20168 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
20169 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
20170 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
20171 			} else {
20172 				rack_exit_probertt(rack, cts);
20173 			}
20174 		}
20175 		idle = 0;
20176 	}
20177 	if (rack_use_fsb &&
20178 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
20179 	    (rack->r_fsb_inited == 0) &&
20180 	    (rack->r_state != TCPS_CLOSED))
20181 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
20182 again:
20183 	/*
20184 	 * If we've recently taken a timeout, snd_max will be greater than
20185 	 * snd_nxt.  There may be SACK information that allows us to avoid
20186 	 * resending already delivered data.  Adjust snd_nxt accordingly.
20187 	 */
20188 	sendalot = 0;
20189 	cts = tcp_get_usecs(&tv);
20190 	ms_cts = tcp_tv_to_mssectick(&tv);
20191 	tso = 0;
20192 	mtu = 0;
20193 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20194 	minseg = segsiz;
20195 	if (rack->r_ctl.rc_pace_max_segs == 0)
20196 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20197 	else
20198 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20199 	sb_offset = tp->snd_max - tp->snd_una;
20200 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20201 	flags = tcp_outflags[tp->t_state];
20202 	while (rack->rc_free_cnt < rack_free_cache) {
20203 		rsm = rack_alloc(rack);
20204 		if (rsm == NULL) {
20205 			if (hpts_calling)
20206 				/* Retry in a ms */
20207 				slot = (1 * HPTS_USEC_IN_MSEC);
20208 			so = inp->inp_socket;
20209 			sb = &so->so_snd;
20210 			goto just_return_nolock;
20211 		}
20212 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20213 		rack->rc_free_cnt++;
20214 		rsm = NULL;
20215 	}
20216 	sack_rxmit = 0;
20217 	len = 0;
20218 	rsm = NULL;
20219 	if (flags & TH_RST) {
20220 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
20221 		so = inp->inp_socket;
20222 		sb = &so->so_snd;
20223 		goto send;
20224 	}
20225 	if (rack->r_ctl.rc_resend) {
20226 		/* Retransmit timer */
20227 		rsm = rack->r_ctl.rc_resend;
20228 		rack->r_ctl.rc_resend = NULL;
20229 		len = rsm->r_end - rsm->r_start;
20230 		sack_rxmit = 1;
20231 		sendalot = 0;
20232 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20233 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20234 			 __func__, __LINE__,
20235 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20236 		sb_offset = rsm->r_start - tp->snd_una;
20237 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20238 	} else if (rack->r_collapse_point_valid &&
20239 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20240 		/*
20241 		 * If an RSM is returned then enough time has passed
20242 		 * for us to retransmit it. Move up the collapse point,
20243 		 * since this rsm has its chance to retransmit now.
20244 		 */
20245 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20246 		rack->r_ctl.last_collapse_point = rsm->r_end;
20247 		/* Are we done? */
20248 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20249 			    rack->r_ctl.high_collapse_point))
20250 			rack->r_collapse_point_valid = 0;
20251 		sack_rxmit = 1;
20252 		/* We are not doing a TLP */
20253 		doing_tlp = 0;
20254 		len = rsm->r_end - rsm->r_start;
20255 		sb_offset = rsm->r_start - tp->snd_una;
20256 		sendalot = 0;
20257 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20258 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20259 		/* We have a retransmit that takes precedence */
20260 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20261 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20262 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20263 			/* Enter recovery if not induced by a time-out */
20264 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20265 		}
20266 #ifdef INVARIANTS
20267 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20268 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20269 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
20270 		}
20271 #endif
20272 		len = rsm->r_end - rsm->r_start;
20273 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20274 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20275 			 __func__, __LINE__,
20276 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20277 		sb_offset = rsm->r_start - tp->snd_una;
20278 		sendalot = 0;
20279 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20280 		if (len > 0) {
20281 			sack_rxmit = 1;
20282 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20283 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20284 					 min(len, segsiz));
20285 		}
20286 	} else if (rack->r_ctl.rc_tlpsend) {
20287 		/* Tail loss probe */
20288 		long cwin;
20289 		long tlen;
20290 
20291 		/*
20292 		 * Check if we can do a TLP with a RACK'd packet
20293 		 * this can happen if we are not doing the rack
20294 		 * cheat and we skipped to a TLP and it
20295 		 * went off.
20296 		 */
20297 		rsm = rack->r_ctl.rc_tlpsend;
20298 		/* We are doing a TLP make sure the flag is preent */
20299 		rsm->r_flags |= RACK_TLP;
20300 		rack->r_ctl.rc_tlpsend = NULL;
20301 		sack_rxmit = 1;
20302 		tlen = rsm->r_end - rsm->r_start;
20303 		if (tlen > segsiz)
20304 			tlen = segsiz;
20305 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20306 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20307 			 __func__, __LINE__,
20308 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20309 		sb_offset = rsm->r_start - tp->snd_una;
20310 		cwin = min(tp->snd_wnd, tlen);
20311 		len = cwin;
20312 	}
20313 	if (rack->r_must_retran &&
20314 	    (doing_tlp == 0) &&
20315 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20316 	    (rsm == NULL)) {
20317 		/*
20318 		 * There are two different ways that we
20319 		 * can get into this block:
20320 		 * a) This is a non-sack connection, we had a time-out
20321 		 *    and thus r_must_retran was set and everything
20322 		 *    left outstanding as been marked for retransmit.
20323 		 * b) The MTU of the path shrank, so that everything
20324 		 *    was marked to be retransmitted with the smaller
20325 		 *    mtu and r_must_retran was set.
20326 		 *
20327 		 * This means that we expect the sendmap (outstanding)
20328 		 * to all be marked must. We can use the tmap to
20329 		 * look at them.
20330 		 *
20331 		 */
20332 		int sendwin, flight;
20333 
20334 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20335 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20336 		if (flight >= sendwin) {
20337 			/*
20338 			 * We can't send yet.
20339 			 */
20340 			so = inp->inp_socket;
20341 			sb = &so->so_snd;
20342 			goto just_return_nolock;
20343 		}
20344 		/*
20345 		 * This is the case a/b mentioned above. All
20346 		 * outstanding/not-acked should be marked.
20347 		 * We can use the tmap to find them.
20348 		 */
20349 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20350 		if (rsm == NULL) {
20351 			/* TSNH */
20352 			rack->r_must_retran = 0;
20353 			rack->r_ctl.rc_out_at_rto = 0;
20354 			so = inp->inp_socket;
20355 			sb = &so->so_snd;
20356 			goto just_return_nolock;
20357 		}
20358 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20359 			/*
20360 			 * The first one does not have the flag, did we collapse
20361 			 * further up in our list?
20362 			 */
20363 			rack->r_must_retran = 0;
20364 			rack->r_ctl.rc_out_at_rto = 0;
20365 			rsm = NULL;
20366 			sack_rxmit = 0;
20367 		} else {
20368 			sack_rxmit = 1;
20369 			len = rsm->r_end - rsm->r_start;
20370 			sb_offset = rsm->r_start - tp->snd_una;
20371 			sendalot = 0;
20372 			if ((rack->full_size_rxt == 0) &&
20373 			    (rack->shape_rxt_to_pacing_min == 0) &&
20374 			    (len >= segsiz))
20375 				len = segsiz;
20376 			else if (rack->shape_rxt_to_pacing_min &&
20377 				 rack->gp_ready) {
20378 				/* We use pacing min as shaping len req */
20379 				uint32_t maxlen;
20380 
20381 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20382 				if (len > maxlen)
20383 					len = maxlen;
20384 			}
20385 			/*
20386 			 * Delay removing the flag RACK_MUST_RXT so
20387 			 * that the fastpath for retransmit will
20388 			 * work with this rsm.
20389 			 */
20390 		}
20391 	}
20392 	/*
20393 	 * Enforce a connection sendmap count limit if set
20394 	 * as long as we are not retransmiting.
20395 	 */
20396 	if ((rsm == NULL) &&
20397 	    (rack->do_detection == 0) &&
20398 	    (V_tcp_map_entries_limit > 0) &&
20399 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20400 		counter_u64_add(rack_to_alloc_limited, 1);
20401 		if (!rack->alloc_limit_reported) {
20402 			rack->alloc_limit_reported = 1;
20403 			counter_u64_add(rack_alloc_limited_conns, 1);
20404 		}
20405 		so = inp->inp_socket;
20406 		sb = &so->so_snd;
20407 		goto just_return_nolock;
20408 	}
20409 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20410 		/* we are retransmitting the fin */
20411 		len--;
20412 		if (len) {
20413 			/*
20414 			 * When retransmitting data do *not* include the
20415 			 * FIN. This could happen from a TLP probe.
20416 			 */
20417 			flags &= ~TH_FIN;
20418 		}
20419 	}
20420 	if (rsm && rack->r_fsb_inited &&
20421 	    rack_use_rsm_rfo &&
20422 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20423 		int ret;
20424 
20425 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20426 		if (ret == 0)
20427 			return (0);
20428 	}
20429 	so = inp->inp_socket;
20430 	sb = &so->so_snd;
20431 	if (do_a_prefetch == 0) {
20432 		kern_prefetch(sb, &do_a_prefetch);
20433 		do_a_prefetch = 1;
20434 	}
20435 #ifdef NETFLIX_SHARED_CWND
20436 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20437 	    rack->rack_enable_scwnd) {
20438 		/* We are doing cwnd sharing */
20439 		if (rack->gp_ready &&
20440 		    (rack->rack_attempted_scwnd == 0) &&
20441 		    (rack->r_ctl.rc_scw == NULL) &&
20442 		    tp->t_lib) {
20443 			/* The pcbid is in, lets make an attempt */
20444 			counter_u64_add(rack_try_scwnd, 1);
20445 			rack->rack_attempted_scwnd = 1;
20446 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20447 								   &rack->r_ctl.rc_scw_index,
20448 								   segsiz);
20449 		}
20450 		if (rack->r_ctl.rc_scw &&
20451 		    (rack->rack_scwnd_is_idle == 1) &&
20452 		    sbavail(&so->so_snd)) {
20453 			/* we are no longer out of data */
20454 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20455 			rack->rack_scwnd_is_idle = 0;
20456 		}
20457 		if (rack->r_ctl.rc_scw) {
20458 			/* First lets update and get the cwnd */
20459 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20460 										       rack->r_ctl.rc_scw_index,
20461 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
20462 		}
20463 	}
20464 #endif
20465 	/*
20466 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20467 	 * state flags.
20468 	 */
20469 	if (tp->t_flags & TF_NEEDFIN)
20470 		flags |= TH_FIN;
20471 	if (tp->t_flags & TF_NEEDSYN)
20472 		flags |= TH_SYN;
20473 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20474 		void *end_rsm;
20475 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20476 		if (end_rsm)
20477 			kern_prefetch(end_rsm, &prefetch_rsm);
20478 		prefetch_rsm = 1;
20479 	}
20480 	SOCKBUF_LOCK(sb);
20481 	/*
20482 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
20483 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
20484 	 * negative length.  This can also occur when TCP opens up its
20485 	 * congestion window while receiving additional duplicate acks after
20486 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
20487 	 * the fast-retransmit.
20488 	 *
20489 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
20490 	 * set to snd_una, the sb_offset will be 0, and the length may wind
20491 	 * up 0.
20492 	 *
20493 	 * If sack_rxmit is true we are retransmitting from the scoreboard
20494 	 * in which case len is already set.
20495 	 */
20496 	if ((sack_rxmit == 0) &&
20497 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
20498 		uint32_t avail;
20499 
20500 		avail = sbavail(sb);
20501 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
20502 			sb_offset = tp->snd_nxt - tp->snd_una;
20503 		else
20504 			sb_offset = 0;
20505 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20506 			if (rack->r_ctl.rc_tlp_new_data) {
20507 				/* TLP is forcing out new data */
20508 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20509 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20510 				}
20511 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20512 					if (tp->snd_wnd > sb_offset)
20513 						len = tp->snd_wnd - sb_offset;
20514 					else
20515 						len = 0;
20516 				} else {
20517 					len = rack->r_ctl.rc_tlp_new_data;
20518 				}
20519 				rack->r_ctl.rc_tlp_new_data = 0;
20520 			}  else {
20521 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20522 			}
20523 			if ((rack->r_ctl.crte == NULL) &&
20524 			    IN_FASTRECOVERY(tp->t_flags) &&
20525 			    (rack->full_size_rxt == 0) &&
20526 			    (rack->shape_rxt_to_pacing_min == 0) &&
20527 			    (len > segsiz)) {
20528 				/*
20529 				 * For prr=off, we need to send only 1 MSS
20530 				 * at a time. We do this because another sack could
20531 				 * be arriving that causes us to send retransmits and
20532 				 * we don't want to be on a long pace due to a larger send
20533 				 * that keeps us from sending out the retransmit.
20534 				 */
20535 				len = segsiz;
20536 			} else if (rack->shape_rxt_to_pacing_min &&
20537 				   rack->gp_ready) {
20538 				/* We use pacing min as shaping len req */
20539 				uint32_t maxlen;
20540 
20541 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20542 				if (len > maxlen)
20543 					len = maxlen;
20544 			}/* The else is full_size_rxt is on so send it all */
20545 		} else {
20546 			uint32_t outstanding;
20547 			/*
20548 			 * We are inside of a Fast recovery episode, this
20549 			 * is caused by a SACK or 3 dup acks. At this point
20550 			 * we have sent all the retransmissions and we rely
20551 			 * on PRR to dictate what we will send in the form of
20552 			 * new data.
20553 			 */
20554 
20555 			outstanding = tp->snd_max - tp->snd_una;
20556 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20557 				if (tp->snd_wnd > outstanding) {
20558 					len = tp->snd_wnd - outstanding;
20559 					/* Check to see if we have the data */
20560 					if ((sb_offset + len) > avail) {
20561 						/* It does not all fit */
20562 						if (avail > sb_offset)
20563 							len = avail - sb_offset;
20564 						else
20565 							len = 0;
20566 					}
20567 				} else {
20568 					len = 0;
20569 				}
20570 			} else if (avail > sb_offset) {
20571 				len = avail - sb_offset;
20572 			} else {
20573 				len = 0;
20574 			}
20575 			if (len > 0) {
20576 				if (len > rack->r_ctl.rc_prr_sndcnt) {
20577 					len = rack->r_ctl.rc_prr_sndcnt;
20578 				}
20579 				if (len > 0) {
20580 					sub_from_prr = 1;
20581 				}
20582 			}
20583 			if (len > segsiz) {
20584 				/*
20585 				 * We should never send more than a MSS when
20586 				 * retransmitting or sending new data in prr
20587 				 * mode unless the override flag is on. Most
20588 				 * likely the PRR algorithm is not going to
20589 				 * let us send a lot as well :-)
20590 				 */
20591 				if (rack->r_ctl.rc_prr_sendalot == 0) {
20592 					len = segsiz;
20593 				}
20594 			} else if (len < segsiz) {
20595 				/*
20596 				 * Do we send any? The idea here is if the
20597 				 * send empty's the socket buffer we want to
20598 				 * do it. However if not then lets just wait
20599 				 * for our prr_sndcnt to get bigger.
20600 				 */
20601 				long leftinsb;
20602 
20603 				leftinsb = sbavail(sb) - sb_offset;
20604 				if (leftinsb > len) {
20605 					/* This send does not empty the sb */
20606 					len = 0;
20607 				}
20608 			}
20609 		}
20610 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20611 		/*
20612 		 * If you have not established
20613 		 * and are not doing FAST OPEN
20614 		 * no data please.
20615 		 */
20616 		if ((sack_rxmit == 0) &&
20617 		    (!IS_FASTOPEN(tp->t_flags))){
20618 			len = 0;
20619 			sb_offset = 0;
20620 		}
20621 	}
20622 	if (prefetch_so_done == 0) {
20623 		kern_prefetch(so, &prefetch_so_done);
20624 		prefetch_so_done = 1;
20625 	}
20626 	/*
20627 	 * Lop off SYN bit if it has already been sent.  However, if this is
20628 	 * SYN-SENT state and if segment contains data and if we don't know
20629 	 * that foreign host supports TAO, suppress sending segment.
20630 	 */
20631 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
20632 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
20633 		/*
20634 		 * When sending additional segments following a TFO SYN|ACK,
20635 		 * do not include the SYN bit.
20636 		 */
20637 		if (IS_FASTOPEN(tp->t_flags) &&
20638 		    (tp->t_state == TCPS_SYN_RECEIVED))
20639 			flags &= ~TH_SYN;
20640 	}
20641 	/*
20642 	 * Be careful not to send data and/or FIN on SYN segments. This
20643 	 * measure is needed to prevent interoperability problems with not
20644 	 * fully conformant TCP implementations.
20645 	 */
20646 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20647 		len = 0;
20648 		flags &= ~TH_FIN;
20649 	}
20650 	/*
20651 	 * On TFO sockets, ensure no data is sent in the following cases:
20652 	 *
20653 	 *  - When retransmitting SYN|ACK on a passively-created socket
20654 	 *
20655 	 *  - When retransmitting SYN on an actively created socket
20656 	 *
20657 	 *  - When sending a zero-length cookie (cookie request) on an
20658 	 *    actively created socket
20659 	 *
20660 	 *  - When the socket is in the CLOSED state (RST is being sent)
20661 	 */
20662 	if (IS_FASTOPEN(tp->t_flags) &&
20663 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20664 	     ((tp->t_state == TCPS_SYN_SENT) &&
20665 	      (tp->t_tfo_client_cookie_len == 0)) ||
20666 	     (flags & TH_RST))) {
20667 		sack_rxmit = 0;
20668 		len = 0;
20669 	}
20670 	/* Without fast-open there should never be data sent on a SYN */
20671 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
20672 		tp->snd_nxt = tp->iss;
20673 		len = 0;
20674 	}
20675 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20676 		/* We only send 1 MSS if we have a DSACK block */
20677 		add_flag |= RACK_SENT_W_DSACK;
20678 		len = segsiz;
20679 	}
20680 	orig_len = len;
20681 	if (len <= 0) {
20682 		/*
20683 		 * If FIN has been sent but not acked, but we haven't been
20684 		 * called to retransmit, len will be < 0.  Otherwise, window
20685 		 * shrank after we sent into it.  If window shrank to 0,
20686 		 * cancel pending retransmit, pull snd_nxt back to (closed)
20687 		 * window, and set the persist timer if it isn't already
20688 		 * going.  If the window didn't close completely, just wait
20689 		 * for an ACK.
20690 		 *
20691 		 * We also do a general check here to ensure that we will
20692 		 * set the persist timer when we have data to send, but a
20693 		 * 0-byte window. This makes sure the persist timer is set
20694 		 * even if the packet hits one of the "goto send" lines
20695 		 * below.
20696 		 */
20697 		len = 0;
20698 		if ((tp->snd_wnd == 0) &&
20699 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20700 		    (tp->snd_una == tp->snd_max) &&
20701 		    (sb_offset < (int)sbavail(sb))) {
20702 			rack_enter_persist(tp, rack, cts, tp->snd_una);
20703 		}
20704 	} else if ((rsm == NULL) &&
20705 		   (doing_tlp == 0) &&
20706 		   (len < pace_max_seg)) {
20707 		/*
20708 		 * We are not sending a maximum sized segment for
20709 		 * some reason. Should we not send anything (think
20710 		 * sws or persists)?
20711 		 */
20712 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20713 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20714 		    (len < minseg) &&
20715 		    (len < (int)(sbavail(sb) - sb_offset))) {
20716 			/*
20717 			 * Here the rwnd is less than
20718 			 * the minimum pacing size, this is not a retransmit,
20719 			 * we are established and
20720 			 * the send is not the last in the socket buffer
20721 			 * we send nothing, and we may enter persists
20722 			 * if nothing is outstanding.
20723 			 */
20724 			len = 0;
20725 			if (tp->snd_max == tp->snd_una) {
20726 				/*
20727 				 * Nothing out we can
20728 				 * go into persists.
20729 				 */
20730 				rack_enter_persist(tp, rack, cts, tp->snd_una);
20731 			}
20732 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20733 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20734 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20735 			   (len < minseg)) {
20736 			/*
20737 			 * Here we are not retransmitting, and
20738 			 * the cwnd is not so small that we could
20739 			 * not send at least a min size (rxt timer
20740 			 * not having gone off), We have 2 segments or
20741 			 * more already in flight, its not the tail end
20742 			 * of the socket buffer  and the cwnd is blocking
20743 			 * us from sending out a minimum pacing segment size.
20744 			 * Lets not send anything.
20745 			 */
20746 			len = 0;
20747 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20748 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20749 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20750 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20751 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
20752 			/*
20753 			 * Here we have a send window but we have
20754 			 * filled it up and we can't send another pacing segment.
20755 			 * We also have in flight more than 2 segments
20756 			 * and we are not completing the sb i.e. we allow
20757 			 * the last bytes of the sb to go out even if
20758 			 * its not a full pacing segment.
20759 			 */
20760 			len = 0;
20761 		} else if ((rack->r_ctl.crte != NULL) &&
20762 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20763 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20764 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20765 			   (len < (int)(sbavail(sb) - sb_offset))) {
20766 			/*
20767 			 * Here we are doing hardware pacing, this is not a TLP,
20768 			 * we are not sending a pace max segment size, there is rwnd
20769 			 * room to send at least N pace_max_seg, the cwnd is greater
20770 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20771 			 * more segments in flight and its not the tail of the socket buffer.
20772 			 *
20773 			 * We don't want to send instead we need to get more ack's in to
20774 			 * allow us to send a full pacing segment. Normally, if we are pacing
20775 			 * about the right speed, we should have finished our pacing
20776 			 * send as most of the acks have come back if we are at the
20777 			 * right rate. This is a bit fuzzy since return path delay
20778 			 * can delay the acks, which is why we want to make sure we
20779 			 * have cwnd space to have a bit more than a max pace segments in flight.
20780 			 *
20781 			 * If we have not gotten our acks back we are pacing at too high a
20782 			 * rate delaying will not hurt and will bring our GP estimate down by
20783 			 * injecting the delay. If we don't do this we will send
20784 			 * 2 MSS out in response to the acks being clocked in which
20785 			 * defeats the point of hw-pacing (i.e. to help us get
20786 			 * larger TSO's out).
20787 			 */
20788 			len = 0;
20789 		}
20790 
20791 	}
20792 	/* len will be >= 0 after this point. */
20793 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20794 	rack_sndbuf_autoscale(rack);
20795 	/*
20796 	 * Decide if we can use TCP Segmentation Offloading (if supported by
20797 	 * hardware).
20798 	 *
20799 	 * TSO may only be used if we are in a pure bulk sending state.  The
20800 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20801 	 * options prevent using TSO.  With TSO the TCP header is the same
20802 	 * (except for the sequence number) for all generated packets.  This
20803 	 * makes it impossible to transmit any options which vary per
20804 	 * generated segment or packet.
20805 	 *
20806 	 * IPv4 handling has a clear separation of ip options and ip header
20807 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20808 	 * the right thing below to provide length of just ip options and thus
20809 	 * checking for ipoptlen is enough to decide if ip options are present.
20810 	 */
20811 	ipoptlen = 0;
20812 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20813 	/*
20814 	 * Pre-calculate here as we save another lookup into the darknesses
20815 	 * of IPsec that way and can actually decide if TSO is ok.
20816 	 */
20817 #ifdef INET6
20818 	if (isipv6 && IPSEC_ENABLED(ipv6))
20819 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20820 #ifdef INET
20821 	else
20822 #endif
20823 #endif				/* INET6 */
20824 #ifdef INET
20825 		if (IPSEC_ENABLED(ipv4))
20826 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20827 #endif				/* INET */
20828 #endif
20829 
20830 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20831 	ipoptlen += ipsec_optlen;
20832 #endif
20833 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20834 	    (tp->t_port == 0) &&
20835 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
20836 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
20837 	    ipoptlen == 0)
20838 		tso = 1;
20839 	{
20840 		uint32_t outstanding __unused;
20841 
20842 		outstanding = tp->snd_max - tp->snd_una;
20843 		if (tp->t_flags & TF_SENTFIN) {
20844 			/*
20845 			 * If we sent a fin, snd_max is 1 higher than
20846 			 * snd_una
20847 			 */
20848 			outstanding--;
20849 		}
20850 		if (sack_rxmit) {
20851 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20852 				flags &= ~TH_FIN;
20853 		} else {
20854 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
20855 				   sbused(sb)))
20856 				flags &= ~TH_FIN;
20857 		}
20858 	}
20859 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20860 		      (long)TCP_MAXWIN << tp->rcv_scale);
20861 
20862 	/*
20863 	 * Sender silly window avoidance.   We transmit under the following
20864 	 * conditions when len is non-zero:
20865 	 *
20866 	 * - We have a full segment (or more with TSO) - This is the last
20867 	 * buffer in a write()/send() and we are either idle or running
20868 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
20869 	 * then 1/2 the maximum send window's worth of data (receiver may be
20870 	 * limited the window size) - we need to retransmit
20871 	 */
20872 	if (len) {
20873 		if (len >= segsiz) {
20874 			goto send;
20875 		}
20876 		/*
20877 		 * NOTE! on localhost connections an 'ack' from the remote
20878 		 * end may occur synchronously with the output and cause us
20879 		 * to flush a buffer queued with moretocome.  XXX
20880 		 *
20881 		 */
20882 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
20883 		    (idle || (tp->t_flags & TF_NODELAY)) &&
20884 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20885 		    (tp->t_flags & TF_NOPUSH) == 0) {
20886 			pass = 2;
20887 			goto send;
20888 		}
20889 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
20890 			pass = 22;
20891 			goto send;
20892 		}
20893 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20894 			pass = 4;
20895 			goto send;
20896 		}
20897 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
20898 			pass = 5;
20899 			goto send;
20900 		}
20901 		if (sack_rxmit) {
20902 			pass = 6;
20903 			goto send;
20904 		}
20905 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20906 		    (ctf_outstanding(tp) < (segsiz * 2))) {
20907 			/*
20908 			 * We have less than two MSS outstanding (delayed ack)
20909 			 * and our rwnd will not let us send a full sized
20910 			 * MSS. Lets go ahead and let this small segment
20911 			 * out because we want to try to have at least two
20912 			 * packets inflight to not be caught by delayed ack.
20913 			 */
20914 			pass = 12;
20915 			goto send;
20916 		}
20917 	}
20918 	/*
20919 	 * Sending of standalone window updates.
20920 	 *
20921 	 * Window updates are important when we close our window due to a
20922 	 * full socket buffer and are opening it again after the application
20923 	 * reads data from it.  Once the window has opened again and the
20924 	 * remote end starts to send again the ACK clock takes over and
20925 	 * provides the most current window information.
20926 	 *
20927 	 * We must avoid the silly window syndrome whereas every read from
20928 	 * the receive buffer, no matter how small, causes a window update
20929 	 * to be sent.  We also should avoid sending a flurry of window
20930 	 * updates when the socket buffer had queued a lot of data and the
20931 	 * application is doing small reads.
20932 	 *
20933 	 * Prevent a flurry of pointless window updates by only sending an
20934 	 * update when we can increase the advertized window by more than
20935 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
20936 	 * full or is very small be more aggressive and send an update
20937 	 * whenever we can increase by two mss sized segments. In all other
20938 	 * situations the ACK's to new incoming data will carry further
20939 	 * window increases.
20940 	 *
20941 	 * Don't send an independent window update if a delayed ACK is
20942 	 * pending (it will get piggy-backed on it) or the remote side
20943 	 * already has done a half-close and won't send more data.  Skip
20944 	 * this if the connection is in T/TCP half-open state.
20945 	 */
20946 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20947 	    !(tp->t_flags & TF_DELACK) &&
20948 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
20949 		/*
20950 		 * "adv" is the amount we could increase the window, taking
20951 		 * into account that we are limited by TCP_MAXWIN <<
20952 		 * tp->rcv_scale.
20953 		 */
20954 		int32_t adv;
20955 		int oldwin;
20956 
20957 		adv = recwin;
20958 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20959 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
20960 			if (adv > oldwin)
20961 				adv -= oldwin;
20962 			else {
20963 				/* We can't increase the window */
20964 				adv = 0;
20965 			}
20966 		} else
20967 			oldwin = 0;
20968 
20969 		/*
20970 		 * If the new window size ends up being the same as or less
20971 		 * than the old size when it is scaled, then don't force
20972 		 * a window update.
20973 		 */
20974 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20975 			goto dontupdate;
20976 
20977 		if (adv >= (int32_t)(2 * segsiz) &&
20978 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20979 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20980 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20981 			pass = 7;
20982 			goto send;
20983 		}
20984 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20985 			pass = 23;
20986 			goto send;
20987 		}
20988 	}
20989 dontupdate:
20990 
20991 	/*
20992 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20993 	 * is also a catch-all for the retransmit timer timeout case.
20994 	 */
20995 	if (tp->t_flags & TF_ACKNOW) {
20996 		pass = 8;
20997 		goto send;
20998 	}
20999 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
21000 		pass = 9;
21001 		goto send;
21002 	}
21003 	/*
21004 	 * If our state indicates that FIN should be sent and we have not
21005 	 * yet done so, then we need to send.
21006 	 */
21007 	if ((flags & TH_FIN) &&
21008 	    (tp->snd_nxt == tp->snd_una)) {
21009 		pass = 11;
21010 		goto send;
21011 	}
21012 	/*
21013 	 * No reason to send a segment, just return.
21014 	 */
21015 just_return:
21016 	SOCKBUF_UNLOCK(sb);
21017 just_return_nolock:
21018 	{
21019 		int app_limited = CTF_JR_SENT_DATA;
21020 
21021 		if (tot_len_this_send > 0) {
21022 			/* Make sure snd_nxt is up to max */
21023 			rack->r_ctl.fsb.recwin = recwin;
21024 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
21025 			if ((error == 0) &&
21026 			    rack_use_rfo &&
21027 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21028 			    (ipoptlen == 0) &&
21029 			    (tp->snd_nxt == tp->snd_max) &&
21030 			    (tp->rcv_numsacks == 0) &&
21031 			    rack->r_fsb_inited &&
21032 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
21033 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
21034 			    (rack->r_must_retran == 0) &&
21035 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
21036 			    (len > 0) && (orig_len > 0) &&
21037 			    (orig_len > len) &&
21038 			    ((orig_len - len) >= segsiz) &&
21039 			    ((optlen == 0) ||
21040 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
21041 				/* We can send at least one more MSS using our fsb */
21042 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
21043 						       segsiz, pace_max_seg, hw_tls, flags);
21044 			} else
21045 				rack->r_fast_output = 0;
21046 
21047 
21048 			rack_log_fsb(rack, tp, so, flags,
21049 				     ipoptlen, orig_len, len, 0,
21050 				     1, optlen, __LINE__, 1);
21051 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
21052 				tp->snd_nxt = tp->snd_max;
21053 		} else {
21054 			int end_window = 0;
21055 			uint32_t seq = tp->gput_ack;
21056 
21057 			rsm = tqhash_max(rack->r_ctl.tqh);
21058 			if (rsm) {
21059 				/*
21060 				 * Mark the last sent that we just-returned (hinting
21061 				 * that delayed ack may play a role in any rtt measurement).
21062 				 */
21063 				rsm->r_just_ret = 1;
21064 			}
21065 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
21066 			rack->r_ctl.rc_agg_delayed = 0;
21067 			rack->r_early = 0;
21068 			rack->r_late = 0;
21069 			rack->r_ctl.rc_agg_early = 0;
21070 			if ((ctf_outstanding(tp) +
21071 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
21072 				 minseg)) >= tp->snd_wnd) {
21073 				/* We are limited by the rwnd */
21074 				app_limited = CTF_JR_RWND_LIMITED;
21075 				if (IN_FASTRECOVERY(tp->t_flags))
21076 					rack->r_ctl.rc_prr_sndcnt = 0;
21077 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
21078 				/* We are limited by whats available -- app limited */
21079 				app_limited = CTF_JR_APP_LIMITED;
21080 				if (IN_FASTRECOVERY(tp->t_flags))
21081 					rack->r_ctl.rc_prr_sndcnt = 0;
21082 			} else if ((idle == 0) &&
21083 				   ((tp->t_flags & TF_NODELAY) == 0) &&
21084 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
21085 				   (len < segsiz)) {
21086 				/*
21087 				 * No delay is not on and the
21088 				 * user is sending less than 1MSS. This
21089 				 * brings out SWS avoidance so we
21090 				 * don't send. Another app-limited case.
21091 				 */
21092 				app_limited = CTF_JR_APP_LIMITED;
21093 			} else if (tp->t_flags & TF_NOPUSH) {
21094 				/*
21095 				 * The user has requested no push of
21096 				 * the last segment and we are
21097 				 * at the last segment. Another app
21098 				 * limited case.
21099 				 */
21100 				app_limited = CTF_JR_APP_LIMITED;
21101 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
21102 				/* Its the cwnd */
21103 				app_limited = CTF_JR_CWND_LIMITED;
21104 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
21105 				   (rack->rack_no_prr == 0) &&
21106 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
21107 				app_limited = CTF_JR_PRR;
21108 			} else {
21109 				/* Now why here are we not sending? */
21110 #ifdef NOW
21111 #ifdef INVARIANTS
21112 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
21113 #endif
21114 #endif
21115 				app_limited = CTF_JR_ASSESSING;
21116 			}
21117 			/*
21118 			 * App limited in some fashion, for our pacing GP
21119 			 * measurements we don't want any gap (even cwnd).
21120 			 * Close  down the measurement window.
21121 			 */
21122 			if (rack_cwnd_block_ends_measure &&
21123 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
21124 			     (app_limited == CTF_JR_PRR))) {
21125 				/*
21126 				 * The reason we are not sending is
21127 				 * the cwnd (or prr). We have been configured
21128 				 * to end the measurement window in
21129 				 * this case.
21130 				 */
21131 				end_window = 1;
21132 			} else if (rack_rwnd_block_ends_measure &&
21133 				   (app_limited == CTF_JR_RWND_LIMITED)) {
21134 				/*
21135 				 * We are rwnd limited and have been
21136 				 * configured to end the measurement
21137 				 * window in this case.
21138 				 */
21139 				end_window = 1;
21140 			} else if (app_limited == CTF_JR_APP_LIMITED) {
21141 				/*
21142 				 * A true application limited period, we have
21143 				 * ran out of data.
21144 				 */
21145 				end_window = 1;
21146 			} else if (app_limited == CTF_JR_ASSESSING) {
21147 				/*
21148 				 * In the assessing case we hit the end of
21149 				 * the if/else and had no known reason
21150 				 * This will panic us under invariants..
21151 				 *
21152 				 * If we get this out in logs we need to
21153 				 * investagate which reason we missed.
21154 				 */
21155 				end_window = 1;
21156 			}
21157 			if (end_window) {
21158 				uint8_t log = 0;
21159 
21160 				/* Adjust the Gput measurement */
21161 				if ((tp->t_flags & TF_GPUTINPROG) &&
21162 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21163 					tp->gput_ack = tp->snd_max;
21164 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21165 						/*
21166 						 * There is not enough to measure.
21167 						 */
21168 						tp->t_flags &= ~TF_GPUTINPROG;
21169 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21170 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
21171 									   tp->gput_seq,
21172 									   0, 0, 18, __LINE__, NULL, 0);
21173 					} else
21174 						log = 1;
21175 				}
21176 				/* Mark the last packet has app limited */
21177 				rsm = tqhash_max(rack->r_ctl.tqh);
21178 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21179 					if (rack->r_ctl.rc_app_limited_cnt == 0)
21180 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21181 					else {
21182 						/*
21183 						 * Go out to the end app limited and mark
21184 						 * this new one as next and move the end_appl up
21185 						 * to this guy.
21186 						 */
21187 						if (rack->r_ctl.rc_end_appl)
21188 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21189 						rack->r_ctl.rc_end_appl = rsm;
21190 					}
21191 					rsm->r_flags |= RACK_APP_LIMITED;
21192 					rack->r_ctl.rc_app_limited_cnt++;
21193 				}
21194 				if (log)
21195 					rack_log_pacing_delay_calc(rack,
21196 								   rack->r_ctl.rc_app_limited_cnt, seq,
21197 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21198 			}
21199 		}
21200 		/* Check if we need to go into persists or not */
21201 		if ((tp->snd_max == tp->snd_una) &&
21202 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
21203 		    sbavail(sb) &&
21204 		    (sbavail(sb) > tp->snd_wnd) &&
21205 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21206 			/* Yes lets make sure to move to persist before timer-start */
21207 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21208 		}
21209 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21210 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21211 	}
21212 #ifdef NETFLIX_SHARED_CWND
21213 	if ((sbavail(sb) == 0) &&
21214 	    rack->r_ctl.rc_scw) {
21215 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21216 		rack->rack_scwnd_is_idle = 1;
21217 	}
21218 #endif
21219 #ifdef TCP_ACCOUNTING
21220 	if (tot_len_this_send > 0) {
21221 		crtsc = get_cyclecount();
21222 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21223 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
21224 		}
21225 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21226 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21227 		}
21228 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21229 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21230 		}
21231 	} else {
21232 		crtsc = get_cyclecount();
21233 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21234 			tp->tcp_cnt_counters[SND_LIMITED]++;
21235 		}
21236 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21237 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21238 		}
21239 	}
21240 	sched_unpin();
21241 #endif
21242 	return (0);
21243 
21244 send:
21245 	if ((rack->r_ctl.crte != NULL) &&
21246 	    (rsm == NULL) &&
21247 	    ((rack->rc_hw_nobuf == 1) ||
21248 	     (rack_hw_check_queue && (check_done == 0)))) {
21249 		/*
21250 		 * We only want to do this once with the hw_check_queue,
21251 		 * for the enobuf case we would only do it once if
21252 		 * we come around to again, the flag will be clear.
21253 		 */
21254 		check_done = 1;
21255 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21256 		if (slot) {
21257 			rack->r_ctl.rc_agg_delayed = 0;
21258 			rack->r_ctl.rc_agg_early = 0;
21259 			rack->r_early = 0;
21260 			rack->r_late = 0;
21261 			SOCKBUF_UNLOCK(&so->so_snd);
21262 			goto skip_all_send;
21263 		}
21264 	}
21265 	if (rsm || sack_rxmit)
21266 		counter_u64_add(rack_nfto_resend, 1);
21267 	else
21268 		counter_u64_add(rack_non_fto_send, 1);
21269 	if ((flags & TH_FIN) &&
21270 	    sbavail(sb)) {
21271 		/*
21272 		 * We do not transmit a FIN
21273 		 * with data outstanding. We
21274 		 * need to make it so all data
21275 		 * is acked first.
21276 		 */
21277 		flags &= ~TH_FIN;
21278 	}
21279 	/* Enforce stack imposed max seg size if we have one */
21280 	if (rack->r_ctl.rc_pace_max_segs &&
21281 	    (len > rack->r_ctl.rc_pace_max_segs)) {
21282 		mark = 1;
21283 		len = rack->r_ctl.rc_pace_max_segs;
21284 	}
21285 	SOCKBUF_LOCK_ASSERT(sb);
21286 	if (len > 0) {
21287 		if (len >= segsiz)
21288 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21289 		else
21290 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21291 	}
21292 	/*
21293 	 * Before ESTABLISHED, force sending of initial options unless TCP
21294 	 * set not to do any options. NOTE: we assume that the IP/TCP header
21295 	 * plus TCP options always fit in a single mbuf, leaving room for a
21296 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21297 	 * + optlen <= MCLBYTES
21298 	 */
21299 	optlen = 0;
21300 #ifdef INET6
21301 	if (isipv6)
21302 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21303 	else
21304 #endif
21305 		hdrlen = sizeof(struct tcpiphdr);
21306 
21307 	/*
21308 	 * Compute options for segment. We only have to care about SYN and
21309 	 * established connection segments.  Options for SYN-ACK segments
21310 	 * are handled in TCP syncache.
21311 	 */
21312 	to.to_flags = 0;
21313 	if ((tp->t_flags & TF_NOOPT) == 0) {
21314 		/* Maximum segment size. */
21315 		if (flags & TH_SYN) {
21316 			tp->snd_nxt = tp->iss;
21317 			to.to_mss = tcp_mssopt(&inp->inp_inc);
21318 			if (tp->t_port)
21319 				to.to_mss -= V_tcp_udp_tunneling_overhead;
21320 			to.to_flags |= TOF_MSS;
21321 
21322 			/*
21323 			 * On SYN or SYN|ACK transmits on TFO connections,
21324 			 * only include the TFO option if it is not a
21325 			 * retransmit, as the presence of the TFO option may
21326 			 * have caused the original SYN or SYN|ACK to have
21327 			 * been dropped by a middlebox.
21328 			 */
21329 			if (IS_FASTOPEN(tp->t_flags) &&
21330 			    (tp->t_rxtshift == 0)) {
21331 				if (tp->t_state == TCPS_SYN_RECEIVED) {
21332 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21333 					to.to_tfo_cookie =
21334 						(u_int8_t *)&tp->t_tfo_cookie.server;
21335 					to.to_flags |= TOF_FASTOPEN;
21336 					wanted_cookie = 1;
21337 				} else if (tp->t_state == TCPS_SYN_SENT) {
21338 					to.to_tfo_len =
21339 						tp->t_tfo_client_cookie_len;
21340 					to.to_tfo_cookie =
21341 						tp->t_tfo_cookie.client;
21342 					to.to_flags |= TOF_FASTOPEN;
21343 					wanted_cookie = 1;
21344 					/*
21345 					 * If we wind up having more data to
21346 					 * send with the SYN than can fit in
21347 					 * one segment, don't send any more
21348 					 * until the SYN|ACK comes back from
21349 					 * the other end.
21350 					 */
21351 					sendalot = 0;
21352 				}
21353 			}
21354 		}
21355 		/* Window scaling. */
21356 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21357 			to.to_wscale = tp->request_r_scale;
21358 			to.to_flags |= TOF_SCALE;
21359 		}
21360 		/* Timestamps. */
21361 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
21362 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21363 			to.to_tsval = ms_cts + tp->ts_offset;
21364 			to.to_tsecr = tp->ts_recent;
21365 			to.to_flags |= TOF_TS;
21366 		}
21367 		/* Set receive buffer autosizing timestamp. */
21368 		if (tp->rfbuf_ts == 0 &&
21369 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
21370 			tp->rfbuf_ts = tcp_ts_getticks();
21371 		/* Selective ACK's. */
21372 		if (tp->t_flags & TF_SACK_PERMIT) {
21373 			if (flags & TH_SYN)
21374 				to.to_flags |= TOF_SACKPERM;
21375 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21376 				 tp->rcv_numsacks > 0) {
21377 				to.to_flags |= TOF_SACK;
21378 				to.to_nsacks = tp->rcv_numsacks;
21379 				to.to_sacks = (u_char *)tp->sackblks;
21380 			}
21381 		}
21382 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21383 		/* TCP-MD5 (RFC2385). */
21384 		if (tp->t_flags & TF_SIGNATURE)
21385 			to.to_flags |= TOF_SIGNATURE;
21386 #endif
21387 
21388 		/* Processing the options. */
21389 		hdrlen += optlen = tcp_addoptions(&to, opt);
21390 		/*
21391 		 * If we wanted a TFO option to be added, but it was unable
21392 		 * to fit, ensure no data is sent.
21393 		 */
21394 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
21395 		    !(to.to_flags & TOF_FASTOPEN))
21396 			len = 0;
21397 	}
21398 	if (tp->t_port) {
21399 		if (V_tcp_udp_tunneling_port == 0) {
21400 			/* The port was removed?? */
21401 			SOCKBUF_UNLOCK(&so->so_snd);
21402 #ifdef TCP_ACCOUNTING
21403 			crtsc = get_cyclecount();
21404 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21405 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21406 			}
21407 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21408 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21409 			}
21410 			sched_unpin();
21411 #endif
21412 			return (EHOSTUNREACH);
21413 		}
21414 		hdrlen += sizeof(struct udphdr);
21415 	}
21416 #ifdef INET6
21417 	if (isipv6)
21418 		ipoptlen = ip6_optlen(inp);
21419 	else
21420 #endif
21421 		if (inp->inp_options)
21422 			ipoptlen = inp->inp_options->m_len -
21423 				offsetof(struct ipoption, ipopt_list);
21424 		else
21425 			ipoptlen = 0;
21426 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21427 	ipoptlen += ipsec_optlen;
21428 #endif
21429 
21430 	/*
21431 	 * Adjust data length if insertion of options will bump the packet
21432 	 * length beyond the t_maxseg length. Clear the FIN bit because we
21433 	 * cut off the tail of the segment.
21434 	 */
21435 	if (len + optlen + ipoptlen > tp->t_maxseg) {
21436 		if (tso) {
21437 			uint32_t if_hw_tsomax;
21438 			uint32_t moff;
21439 			int32_t max_len;
21440 
21441 			/* extract TSO information */
21442 			if_hw_tsomax = tp->t_tsomax;
21443 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21444 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21445 			KASSERT(ipoptlen == 0,
21446 				("%s: TSO can't do IP options", __func__));
21447 
21448 			/*
21449 			 * Check if we should limit by maximum payload
21450 			 * length:
21451 			 */
21452 			if (if_hw_tsomax != 0) {
21453 				/* compute maximum TSO length */
21454 				max_len = (if_hw_tsomax - hdrlen -
21455 					   max_linkhdr);
21456 				if (max_len <= 0) {
21457 					len = 0;
21458 				} else if (len > max_len) {
21459 					sendalot = 1;
21460 					len = max_len;
21461 					mark = 2;
21462 				}
21463 			}
21464 			/*
21465 			 * Prevent the last segment from being fractional
21466 			 * unless the send sockbuf can be emptied:
21467 			 */
21468 			max_len = (tp->t_maxseg - optlen);
21469 			if ((sb_offset + len) < sbavail(sb)) {
21470 				moff = len % (u_int)max_len;
21471 				if (moff != 0) {
21472 					mark = 3;
21473 					len -= moff;
21474 				}
21475 			}
21476 			/*
21477 			 * In case there are too many small fragments don't
21478 			 * use TSO:
21479 			 */
21480 			if (len <= max_len) {
21481 				mark = 4;
21482 				tso = 0;
21483 			}
21484 			/*
21485 			 * Send the FIN in a separate segment after the bulk
21486 			 * sending is done. We don't trust the TSO
21487 			 * implementations to clear the FIN flag on all but
21488 			 * the last segment.
21489 			 */
21490 			if (tp->t_flags & TF_NEEDFIN) {
21491 				sendalot = 4;
21492 			}
21493 		} else {
21494 			mark = 5;
21495 			if (optlen + ipoptlen >= tp->t_maxseg) {
21496 				/*
21497 				 * Since we don't have enough space to put
21498 				 * the IP header chain and the TCP header in
21499 				 * one packet as required by RFC 7112, don't
21500 				 * send it. Also ensure that at least one
21501 				 * byte of the payload can be put into the
21502 				 * TCP segment.
21503 				 */
21504 				SOCKBUF_UNLOCK(&so->so_snd);
21505 				error = EMSGSIZE;
21506 				sack_rxmit = 0;
21507 				goto out;
21508 			}
21509 			len = tp->t_maxseg - optlen - ipoptlen;
21510 			sendalot = 5;
21511 		}
21512 	} else {
21513 		tso = 0;
21514 		mark = 6;
21515 	}
21516 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21517 		("%s: len > IP_MAXPACKET", __func__));
21518 #ifdef DIAGNOSTIC
21519 #ifdef INET6
21520 	if (max_linkhdr + hdrlen > MCLBYTES)
21521 #else
21522 		if (max_linkhdr + hdrlen > MHLEN)
21523 #endif
21524 			panic("tcphdr too big");
21525 #endif
21526 
21527 	/*
21528 	 * This KASSERT is here to catch edge cases at a well defined place.
21529 	 * Before, those had triggered (random) panic conditions further
21530 	 * down.
21531 	 */
21532 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21533 	if ((len == 0) &&
21534 	    (flags & TH_FIN) &&
21535 	    (sbused(sb))) {
21536 		/*
21537 		 * We have outstanding data, don't send a fin by itself!.
21538 		 */
21539 		goto just_return;
21540 	}
21541 	/*
21542 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
21543 	 * and initialize the header from the template for sends on this
21544 	 * connection.
21545 	 */
21546 	hw_tls = tp->t_nic_ktls_xmit != 0;
21547 	if (len) {
21548 		uint32_t max_val;
21549 		uint32_t moff;
21550 
21551 		if (rack->r_ctl.rc_pace_max_segs)
21552 			max_val = rack->r_ctl.rc_pace_max_segs;
21553 		else if (rack->rc_user_set_max_segs)
21554 			max_val = rack->rc_user_set_max_segs * segsiz;
21555 		else
21556 			max_val = len;
21557 		/*
21558 		 * We allow a limit on sending with hptsi.
21559 		 */
21560 		if (len > max_val) {
21561 			mark = 7;
21562 			len = max_val;
21563 		}
21564 #ifdef INET6
21565 		if (MHLEN < hdrlen + max_linkhdr)
21566 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21567 		else
21568 #endif
21569 			m = m_gethdr(M_NOWAIT, MT_DATA);
21570 
21571 		if (m == NULL) {
21572 			SOCKBUF_UNLOCK(sb);
21573 			error = ENOBUFS;
21574 			sack_rxmit = 0;
21575 			goto out;
21576 		}
21577 		m->m_data += max_linkhdr;
21578 		m->m_len = hdrlen;
21579 
21580 		/*
21581 		 * Start the m_copy functions from the closest mbuf to the
21582 		 * sb_offset in the socket buffer chain.
21583 		 */
21584 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
21585 		s_mb = mb;
21586 		s_moff = moff;
21587 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21588 			m_copydata(mb, moff, (int)len,
21589 				   mtod(m, caddr_t)+hdrlen);
21590 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21591 				sbsndptr_adv(sb, mb, len);
21592 			m->m_len += len;
21593 		} else {
21594 			struct sockbuf *msb;
21595 
21596 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21597 				msb = NULL;
21598 			else
21599 				msb = sb;
21600 			m->m_next = tcp_m_copym(
21601 				mb, moff, &len,
21602 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21603 				((rsm == NULL) ? hw_tls : 0)
21604 #ifdef NETFLIX_COPY_ARGS
21605 				, &s_mb, &s_moff
21606 #endif
21607 				);
21608 			if (len <= (tp->t_maxseg - optlen)) {
21609 				/*
21610 				 * Must have ran out of mbufs for the copy
21611 				 * shorten it to no longer need tso. Lets
21612 				 * not put on sendalot since we are low on
21613 				 * mbufs.
21614 				 */
21615 				tso = 0;
21616 			}
21617 			if (m->m_next == NULL) {
21618 				SOCKBUF_UNLOCK(sb);
21619 				(void)m_free(m);
21620 				error = ENOBUFS;
21621 				sack_rxmit = 0;
21622 				goto out;
21623 			}
21624 		}
21625 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
21626 			if (rsm && (rsm->r_flags & RACK_TLP)) {
21627 				/*
21628 				 * TLP should not count in retran count, but
21629 				 * in its own bin
21630 				 */
21631 				counter_u64_add(rack_tlp_retran, 1);
21632 				counter_u64_add(rack_tlp_retran_bytes, len);
21633 			} else {
21634 				tp->t_sndrexmitpack++;
21635 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21636 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21637 			}
21638 #ifdef STATS
21639 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21640 						 len);
21641 #endif
21642 		} else {
21643 			KMOD_TCPSTAT_INC(tcps_sndpack);
21644 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21645 #ifdef STATS
21646 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21647 						 len);
21648 #endif
21649 		}
21650 		/*
21651 		 * If we're sending everything we've got, set PUSH. (This
21652 		 * will keep happy those implementations which only give
21653 		 * data to the user when a buffer fills or a PUSH comes in.)
21654 		 */
21655 		if (sb_offset + len == sbused(sb) &&
21656 		    sbused(sb) &&
21657 		    !(flags & TH_SYN)) {
21658 			flags |= TH_PUSH;
21659 			add_flag |= RACK_HAD_PUSH;
21660 		}
21661 
21662 		SOCKBUF_UNLOCK(sb);
21663 	} else {
21664 		SOCKBUF_UNLOCK(sb);
21665 		if (tp->t_flags & TF_ACKNOW)
21666 			KMOD_TCPSTAT_INC(tcps_sndacks);
21667 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
21668 			KMOD_TCPSTAT_INC(tcps_sndctrl);
21669 		else
21670 			KMOD_TCPSTAT_INC(tcps_sndwinup);
21671 
21672 		m = m_gethdr(M_NOWAIT, MT_DATA);
21673 		if (m == NULL) {
21674 			error = ENOBUFS;
21675 			sack_rxmit = 0;
21676 			goto out;
21677 		}
21678 #ifdef INET6
21679 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21680 		    MHLEN >= hdrlen) {
21681 			M_ALIGN(m, hdrlen);
21682 		} else
21683 #endif
21684 			m->m_data += max_linkhdr;
21685 		m->m_len = hdrlen;
21686 	}
21687 	SOCKBUF_UNLOCK_ASSERT(sb);
21688 	m->m_pkthdr.rcvif = (struct ifnet *)0;
21689 #ifdef MAC
21690 	mac_inpcb_create_mbuf(inp, m);
21691 #endif
21692 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21693 #ifdef INET6
21694 		if (isipv6)
21695 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21696 		else
21697 #endif				/* INET6 */
21698 #ifdef INET
21699 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21700 #endif
21701 		th = rack->r_ctl.fsb.th;
21702 		udp = rack->r_ctl.fsb.udp;
21703 		if (udp) {
21704 #ifdef INET6
21705 			if (isipv6)
21706 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21707 			else
21708 #endif				/* INET6 */
21709 				ulen = hdrlen + len - sizeof(struct ip);
21710 			udp->uh_ulen = htons(ulen);
21711 		}
21712 	} else {
21713 #ifdef INET6
21714 		if (isipv6) {
21715 			ip6 = mtod(m, struct ip6_hdr *);
21716 			if (tp->t_port) {
21717 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21718 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21719 				udp->uh_dport = tp->t_port;
21720 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21721 				udp->uh_ulen = htons(ulen);
21722 				th = (struct tcphdr *)(udp + 1);
21723 			} else
21724 				th = (struct tcphdr *)(ip6 + 1);
21725 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
21726 		} else
21727 #endif				/* INET6 */
21728 		{
21729 #ifdef INET
21730 			ip = mtod(m, struct ip *);
21731 			if (tp->t_port) {
21732 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21733 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21734 				udp->uh_dport = tp->t_port;
21735 				ulen = hdrlen + len - sizeof(struct ip);
21736 				udp->uh_ulen = htons(ulen);
21737 				th = (struct tcphdr *)(udp + 1);
21738 			} else
21739 				th = (struct tcphdr *)(ip + 1);
21740 			tcpip_fillheaders(inp, tp->t_port, ip, th);
21741 #endif
21742 		}
21743 	}
21744 	/*
21745 	 * Fill in fields, remembering maximum advertised window for use in
21746 	 * delaying messages about window sizes. If resending a FIN, be sure
21747 	 * not to use a new sequence number.
21748 	 */
21749 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
21750 	    tp->snd_nxt == tp->snd_max)
21751 		tp->snd_nxt--;
21752 	/*
21753 	 * If we are starting a connection, send ECN setup SYN packet. If we
21754 	 * are on a retransmit, we may resend those bits a number of times
21755 	 * as per RFC 3168.
21756 	 */
21757 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21758 		flags |= tcp_ecn_output_syn_sent(tp);
21759 	}
21760 	/* Also handle parallel SYN for ECN */
21761 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
21762 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21763 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21764 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21765 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
21766 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21767 #ifdef INET6
21768 		if (isipv6) {
21769 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21770 			ip6->ip6_flow |= htonl(ect << 20);
21771 		}
21772 		else
21773 #endif
21774 		{
21775 #ifdef INET
21776 			ip->ip_tos &= ~IPTOS_ECN_MASK;
21777 			ip->ip_tos |= ect;
21778 #endif
21779 		}
21780 	}
21781 	/*
21782 	 * If we are doing retransmissions, then snd_nxt will not reflect
21783 	 * the first unsent octet.  For ACK only packets, we do not want the
21784 	 * sequence number of the retransmitted packet, we want the sequence
21785 	 * number of the next unsent octet.  So, if there is no data (and no
21786 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
21787 	 * ti_seq.  But if we are in persist state, snd_max might reflect
21788 	 * one byte beyond the right edge of the window, so use snd_nxt in
21789 	 * that case, since we know we aren't doing a retransmission.
21790 	 * (retransmit and persist are mutually exclusive...)
21791 	 */
21792 	if (sack_rxmit == 0) {
21793 		if (len || (flags & (TH_SYN | TH_FIN))) {
21794 			th->th_seq = htonl(tp->snd_nxt);
21795 			rack_seq = tp->snd_nxt;
21796 		} else {
21797 			th->th_seq = htonl(tp->snd_max);
21798 			rack_seq = tp->snd_max;
21799 		}
21800 	} else {
21801 		th->th_seq = htonl(rsm->r_start);
21802 		rack_seq = rsm->r_start;
21803 	}
21804 	th->th_ack = htonl(tp->rcv_nxt);
21805 	tcp_set_flags(th, flags);
21806 	/*
21807 	 * Calculate receive window.  Don't shrink window, but avoid silly
21808 	 * window syndrome.
21809 	 * If a RST segment is sent, advertise a window of zero.
21810 	 */
21811 	if (flags & TH_RST) {
21812 		recwin = 0;
21813 	} else {
21814 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21815 		    recwin < (long)segsiz) {
21816 			recwin = 0;
21817 		}
21818 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21819 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21820 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21821 	}
21822 
21823 	/*
21824 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21825 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21826 	 * handled in syncache.
21827 	 */
21828 	if (flags & TH_SYN)
21829 		th->th_win = htons((u_short)
21830 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21831 	else {
21832 		/* Avoid shrinking window with window scaling. */
21833 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
21834 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21835 	}
21836 	/*
21837 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21838 	 * window.  This may cause the remote transmitter to stall.  This
21839 	 * flag tells soreceive() to disable delayed acknowledgements when
21840 	 * draining the buffer.  This can occur if the receiver is
21841 	 * attempting to read more data than can be buffered prior to
21842 	 * transmitting on the connection.
21843 	 */
21844 	if (th->th_win == 0) {
21845 		tp->t_sndzerowin++;
21846 		tp->t_flags |= TF_RXWIN0SENT;
21847 	} else
21848 		tp->t_flags &= ~TF_RXWIN0SENT;
21849 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
21850 	/* Now are we using fsb?, if so copy the template data to the mbuf */
21851 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21852 		uint8_t *cpto;
21853 
21854 		cpto = mtod(m, uint8_t *);
21855 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21856 		/*
21857 		 * We have just copied in:
21858 		 * IP/IP6
21859 		 * <optional udphdr>
21860 		 * tcphdr (no options)
21861 		 *
21862 		 * We need to grab the correct pointers into the mbuf
21863 		 * for both the tcp header, and possibly the udp header (if tunneling).
21864 		 * We do this by using the offset in the copy buffer and adding it
21865 		 * to the mbuf base pointer (cpto).
21866 		 */
21867 #ifdef INET6
21868 		if (isipv6)
21869 			ip6 = mtod(m, struct ip6_hdr *);
21870 		else
21871 #endif				/* INET6 */
21872 #ifdef INET
21873 			ip = mtod(m, struct ip *);
21874 #endif
21875 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21876 		/* If we have a udp header lets set it into the mbuf as well */
21877 		if (udp)
21878 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21879 	}
21880 	if (optlen) {
21881 		bcopy(opt, th + 1, optlen);
21882 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21883 	}
21884 	/*
21885 	 * Put TCP length in extended header, and then checksum extended
21886 	 * header and data.
21887 	 */
21888 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
21889 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21890 	if (to.to_flags & TOF_SIGNATURE) {
21891 		/*
21892 		 * Calculate MD5 signature and put it into the place
21893 		 * determined before.
21894 		 * NOTE: since TCP options buffer doesn't point into
21895 		 * mbuf's data, calculate offset and use it.
21896 		 */
21897 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21898 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21899 			/*
21900 			 * Do not send segment if the calculation of MD5
21901 			 * digest has failed.
21902 			 */
21903 			goto out;
21904 		}
21905 	}
21906 #endif
21907 #ifdef INET6
21908 	if (isipv6) {
21909 		/*
21910 		 * ip6_plen is not need to be filled now, and will be filled
21911 		 * in ip6_output.
21912 		 */
21913 		if (tp->t_port) {
21914 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21915 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21916 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21917 			th->th_sum = htons(0);
21918 			UDPSTAT_INC(udps_opackets);
21919 		} else {
21920 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21921 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21922 			th->th_sum = in6_cksum_pseudo(ip6,
21923 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21924 						      0);
21925 		}
21926 	}
21927 #endif
21928 #if defined(INET6) && defined(INET)
21929 	else
21930 #endif
21931 #ifdef INET
21932 	{
21933 		if (tp->t_port) {
21934 			m->m_pkthdr.csum_flags = CSUM_UDP;
21935 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21936 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21937 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21938 			th->th_sum = htons(0);
21939 			UDPSTAT_INC(udps_opackets);
21940 		} else {
21941 			m->m_pkthdr.csum_flags = CSUM_TCP;
21942 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21943 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
21944 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21945 									IPPROTO_TCP + len + optlen));
21946 		}
21947 		/* IP version must be set here for ipv4/ipv6 checking later */
21948 		KASSERT(ip->ip_v == IPVERSION,
21949 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
21950 	}
21951 #endif
21952 	/*
21953 	 * Enable TSO and specify the size of the segments. The TCP pseudo
21954 	 * header checksum is always provided. XXX: Fixme: This is currently
21955 	 * not the case for IPv6.
21956 	 */
21957 	if (tso) {
21958 		/*
21959 		 * Here we must use t_maxseg and the optlen since
21960 		 * the optlen may include SACK's (or DSACK).
21961 		 */
21962 		KASSERT(len > tp->t_maxseg - optlen,
21963 			("%s: len <= tso_segsz", __func__));
21964 		m->m_pkthdr.csum_flags |= CSUM_TSO;
21965 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21966 	}
21967 	KASSERT(len + hdrlen == m_length(m, NULL),
21968 		("%s: mbuf chain different than expected: %d + %u != %u",
21969 		 __func__, len, hdrlen, m_length(m, NULL)));
21970 
21971 #ifdef TCP_HHOOK
21972 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21973 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
21974 #endif
21975 	if ((rack->r_ctl.crte != NULL) &&
21976 	    (rack->rc_hw_nobuf == 0) &&
21977 	    tcp_bblogging_on(tp)) {
21978 		rack_log_queue_level(tp, rack, len, &tv, cts);
21979 	}
21980 	/* We're getting ready to send; log now. */
21981 	if (tcp_bblogging_on(rack->rc_tp)) {
21982 		union tcp_log_stackspecific log;
21983 
21984 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21985 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21986 		if (rack->rack_no_prr)
21987 			log.u_bbr.flex1 = 0;
21988 		else
21989 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21990 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21991 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21992 		log.u_bbr.flex4 = orig_len;
21993 		/* Save off the early/late values */
21994 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21995 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21996 		log.u_bbr.bw_inuse = rack_get_bw(rack);
21997 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21998 		log.u_bbr.flex8 = 0;
21999 		if (rsm) {
22000 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
22001 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
22002 				counter_u64_add(rack_collapsed_win_rxt, 1);
22003 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
22004 			}
22005 			if (doing_tlp)
22006 				log.u_bbr.flex8 = 2;
22007 			else
22008 				log.u_bbr.flex8 = 1;
22009 		} else {
22010 			if (doing_tlp)
22011 				log.u_bbr.flex8 = 3;
22012 		}
22013 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
22014 		log.u_bbr.flex7 = mark;
22015 		log.u_bbr.flex7 <<= 8;
22016 		log.u_bbr.flex7 |= pass;
22017 		log.u_bbr.pkts_out = tp->t_maxseg;
22018 		log.u_bbr.timeStamp = cts;
22019 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
22020 		if (rsm && (rsm->r_rtr_cnt > 0)) {
22021 			/*
22022 			 * When we have a retransmit we want to log the
22023 			 * burst at send and flight at send from before.
22024 			 */
22025 			log.u_bbr.flex5 = rsm->r_fas;
22026 			log.u_bbr.bbr_substate = rsm->r_bas;
22027 		} else {
22028 			/*
22029 			 * New transmits we log in flex5 the inflight again as
22030 			 * well as the number of segments in our send in the
22031 			 * substate field.
22032 			 */
22033 			log.u_bbr.flex5 = log.u_bbr.inflight;
22034 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
22035 		}
22036 		log.u_bbr.lt_epoch = cwnd_to_use;
22037 		log.u_bbr.delivered = sendalot;
22038 		log.u_bbr.rttProp = (uint64_t)rsm;
22039 		log.u_bbr.pkt_epoch = __LINE__;
22040 		if (rsm) {
22041 			log.u_bbr.delRate = rsm->r_flags;
22042 			log.u_bbr.delRate <<= 31;
22043 			log.u_bbr.delRate |= rack->r_must_retran;
22044 			log.u_bbr.delRate <<= 1;
22045 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22046 		} else {
22047 			log.u_bbr.delRate = rack->r_must_retran;
22048 			log.u_bbr.delRate <<= 1;
22049 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22050 		}
22051 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
22052 				    len, &log, false, NULL, __func__, __LINE__, &tv);
22053 	} else
22054 		lgb = NULL;
22055 
22056 	/*
22057 	 * Fill in IP length and desired time to live and send to IP level.
22058 	 * There should be a better way to handle ttl and tos; we could keep
22059 	 * them in the template, but need a way to checksum without them.
22060 	 */
22061 	/*
22062 	 * m->m_pkthdr.len should have been set before cksum calcuration,
22063 	 * because in6_cksum() need it.
22064 	 */
22065 #ifdef INET6
22066 	if (isipv6) {
22067 		/*
22068 		 * we separately set hoplimit for every segment, since the
22069 		 * user might want to change the value via setsockopt. Also,
22070 		 * desired default hop limit might be changed via Neighbor
22071 		 * Discovery.
22072 		 */
22073 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
22074 
22075 		/*
22076 		 * Set the packet size here for the benefit of DTrace
22077 		 * probes. ip6_output() will set it properly; it's supposed
22078 		 * to include the option header lengths as well.
22079 		 */
22080 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
22081 
22082 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22083 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22084 		else
22085 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22086 
22087 		if (tp->t_state == TCPS_SYN_SENT)
22088 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22089 
22090 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22091 		/* TODO: IPv6 IP6TOS_ECT bit on */
22092 		error = ip6_output(m,
22093 				   inp->in6p_outputopts,
22094 				   &inp->inp_route6,
22095 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22096 				   NULL, NULL, inp);
22097 
22098 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22099 			mtu = inp->inp_route6.ro_nh->nh_mtu;
22100 	}
22101 #endif				/* INET6 */
22102 #if defined(INET) && defined(INET6)
22103 	else
22104 #endif
22105 #ifdef INET
22106 	{
22107 		ip->ip_len = htons(m->m_pkthdr.len);
22108 #ifdef INET6
22109 		if (inp->inp_vflag & INP_IPV6PROTO)
22110 			ip->ip_ttl = in6_selecthlim(inp, NULL);
22111 #endif				/* INET6 */
22112 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22113 		/*
22114 		 * If we do path MTU discovery, then we set DF on every
22115 		 * packet. This might not be the best thing to do according
22116 		 * to RFC3390 Section 2. However the tcp hostcache migitates
22117 		 * the problem so it affects only the first tcp connection
22118 		 * with a host.
22119 		 *
22120 		 * NB: Don't set DF on small MTU/MSS to have a safe
22121 		 * fallback.
22122 		 */
22123 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22124 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22125 			if (tp->t_port == 0 || len < V_tcp_minmss) {
22126 				ip->ip_off |= htons(IP_DF);
22127 			}
22128 		} else {
22129 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22130 		}
22131 
22132 		if (tp->t_state == TCPS_SYN_SENT)
22133 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22134 
22135 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
22136 
22137 		error = ip_output(m,
22138 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22139 				  inp->inp_options,
22140 #else
22141 				  NULL,
22142 #endif
22143 				  &inp->inp_route,
22144 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22145 				  inp);
22146 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22147 			mtu = inp->inp_route.ro_nh->nh_mtu;
22148 	}
22149 #endif				/* INET */
22150 
22151 out:
22152 	if (lgb) {
22153 		lgb->tlb_errno = error;
22154 		lgb = NULL;
22155 	}
22156 	/*
22157 	 * In transmit state, time the transmission and arrange for the
22158 	 * retransmit.  In persist state, just set snd_max.
22159 	 */
22160 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22161 			rack_to_usec_ts(&tv),
22162 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22163 	if (error == 0) {
22164 		if (rsm == NULL) {
22165 			if (rack->lt_bw_up == 0) {
22166 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
22167 				rack->r_ctl.lt_seq = tp->snd_una;
22168 				rack->lt_bw_up = 1;
22169 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22170 				/*
22171 				 * Need to record what we have since we are
22172 				 * approaching seq wrap.
22173 				 */
22174 				uint64_t tmark;
22175 
22176 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22177 				rack->r_ctl.lt_seq = tp->snd_una;
22178 				tmark = tcp_tv_to_lusectick(&tv);
22179 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22180 				rack->r_ctl.lt_timemark = tmark;
22181 			}
22182 		}
22183 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
22184 		counter_u64_add(rack_total_bytes, len);
22185 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22186 		if (rsm && doing_tlp) {
22187 			rack->rc_last_sent_tlp_past_cumack = 0;
22188 			rack->rc_last_sent_tlp_seq_valid = 1;
22189 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22190 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22191 		}
22192 		if (rack->rc_hw_nobuf) {
22193 			rack->rc_hw_nobuf = 0;
22194 			rack->r_ctl.rc_agg_delayed = 0;
22195 			rack->r_early = 0;
22196 			rack->r_late = 0;
22197 			rack->r_ctl.rc_agg_early = 0;
22198 		}
22199 		if (rsm && (doing_tlp == 0)) {
22200 			/* Set we retransmitted */
22201 			rack->rc_gp_saw_rec = 1;
22202 		} else {
22203 			if (cwnd_to_use > tp->snd_ssthresh) {
22204 				/* Set we sent in CA */
22205 				rack->rc_gp_saw_ca = 1;
22206 			} else {
22207 				/* Set we sent in SS */
22208 				rack->rc_gp_saw_ss = 1;
22209 			}
22210 		}
22211 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22212 		    (tp->t_flags & TF_SACK_PERMIT) &&
22213 		    tp->rcv_numsacks > 0)
22214 			tcp_clean_dsack_blocks(tp);
22215 		tot_len_this_send += len;
22216 		if (len == 0) {
22217 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22218 		} else {
22219 			int idx;
22220 
22221 			idx = (len / segsiz) + 3;
22222 			if (idx >= TCP_MSS_ACCT_ATIMER)
22223 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22224 			else
22225 				counter_u64_add(rack_out_size[idx], 1);
22226 		}
22227 	}
22228 	if ((rack->rack_no_prr == 0) &&
22229 	    sub_from_prr &&
22230 	    (error == 0)) {
22231 		if (rack->r_ctl.rc_prr_sndcnt >= len)
22232 			rack->r_ctl.rc_prr_sndcnt -= len;
22233 		else
22234 			rack->r_ctl.rc_prr_sndcnt = 0;
22235 	}
22236 	sub_from_prr = 0;
22237 	if (doing_tlp) {
22238 		/* Make sure the TLP is added */
22239 		add_flag |= RACK_TLP;
22240 	} else if (rsm) {
22241 		/* If its a resend without TLP then it must not have the flag */
22242 		rsm->r_flags &= ~RACK_TLP;
22243 	}
22244 
22245 
22246 	if ((error == 0) &&
22247 	    (len > 0) &&
22248 	    (tp->snd_una == tp->snd_max))
22249 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
22250 	{
22251 		tcp_seq startseq = tp->snd_nxt;
22252 
22253 		/* Track our lost count */
22254 		if (rsm && (doing_tlp == 0))
22255 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22256 		/*
22257 		 * Advance snd_nxt over sequence space of this segment.
22258 		 */
22259 		if (error)
22260 			/* We don't log or do anything with errors */
22261 			goto nomore;
22262 		if (doing_tlp == 0) {
22263 			if (rsm == NULL) {
22264 				/*
22265 				 * Not a retransmission of some
22266 				 * sort, new data is going out so
22267 				 * clear our TLP count and flag.
22268 				 */
22269 				rack->rc_tlp_in_progress = 0;
22270 				rack->r_ctl.rc_tlp_cnt_out = 0;
22271 			}
22272 		} else {
22273 			/*
22274 			 * We have just sent a TLP, mark that it is true
22275 			 * and make sure our in progress is set so we
22276 			 * continue to check the count.
22277 			 */
22278 			rack->rc_tlp_in_progress = 1;
22279 			rack->r_ctl.rc_tlp_cnt_out++;
22280 		}
22281 		if (flags & (TH_SYN | TH_FIN)) {
22282 			if (flags & TH_SYN)
22283 				tp->snd_nxt++;
22284 			if (flags & TH_FIN) {
22285 				tp->snd_nxt++;
22286 				tp->t_flags |= TF_SENTFIN;
22287 			}
22288 		}
22289 		/* In the ENOBUFS case we do *not* update snd_max */
22290 		if (sack_rxmit)
22291 			goto nomore;
22292 
22293 		tp->snd_nxt += len;
22294 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
22295 			if (tp->snd_una == tp->snd_max) {
22296 				/*
22297 				 * Update the time we just added data since
22298 				 * none was outstanding.
22299 				 */
22300 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22301 				tp->t_acktime = ticks;
22302 			}
22303 			tp->snd_max = tp->snd_nxt;
22304 			if (rack->rc_new_rnd_needed) {
22305 				/*
22306 				 * Update the rnd to start ticking not
22307 				 * that from a time perspective all of
22308 				 * the preceding idle time is "in the round"
22309 				 */
22310 				rack->rc_new_rnd_needed = 0;
22311 				rack->r_ctl.roundends = tp->snd_max;
22312 			}
22313 			/*
22314 			 * Time this transmission if not a retransmission and
22315 			 * not currently timing anything.
22316 			 * This is only relevant in case of switching back to
22317 			 * the base stack.
22318 			 */
22319 			if (tp->t_rtttime == 0) {
22320 				tp->t_rtttime = ticks;
22321 				tp->t_rtseq = startseq;
22322 				KMOD_TCPSTAT_INC(tcps_segstimed);
22323 			}
22324 			if (len &&
22325 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
22326 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22327 		}
22328 		/*
22329 		 * If we are doing FO we need to update the mbuf position and subtract
22330 		 * this happens when the peer sends us duplicate information and
22331 		 * we thus want to send a DSACK.
22332 		 *
22333 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22334 		 * turned off? If not then we are going to echo multiple DSACK blocks
22335 		 * out (with the TSO), which we should not be doing.
22336 		 */
22337 		if (rack->r_fast_output && len) {
22338 			if (rack->r_ctl.fsb.left_to_send > len)
22339 				rack->r_ctl.fsb.left_to_send -= len;
22340 			else
22341 				rack->r_ctl.fsb.left_to_send = 0;
22342 			if (rack->r_ctl.fsb.left_to_send < segsiz)
22343 				rack->r_fast_output = 0;
22344 			if (rack->r_fast_output) {
22345 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22346 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22347 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22348 			}
22349 		}
22350 	}
22351 nomore:
22352 	if (error) {
22353 		rack->r_ctl.rc_agg_delayed = 0;
22354 		rack->r_early = 0;
22355 		rack->r_late = 0;
22356 		rack->r_ctl.rc_agg_early = 0;
22357 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
22358 		/*
22359 		 * Failures do not advance the seq counter above. For the
22360 		 * case of ENOBUFS we will fall out and retry in 1ms with
22361 		 * the hpts. Everything else will just have to retransmit
22362 		 * with the timer.
22363 		 *
22364 		 * In any case, we do not want to loop around for another
22365 		 * send without a good reason.
22366 		 */
22367 		sendalot = 0;
22368 		switch (error) {
22369 		case EPERM:
22370 			tp->t_softerror = error;
22371 #ifdef TCP_ACCOUNTING
22372 			crtsc = get_cyclecount();
22373 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22374 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22375 			}
22376 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22377 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22378 			}
22379 			sched_unpin();
22380 #endif
22381 			return (error);
22382 		case ENOBUFS:
22383 			/*
22384 			 * Pace us right away to retry in a some
22385 			 * time
22386 			 */
22387 			if (rack->r_ctl.crte != NULL) {
22388 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22389 				if (tcp_bblogging_on(rack->rc_tp))
22390 					rack_log_queue_level(tp, rack, len, &tv, cts);
22391 			} else
22392 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22393 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22394 			if (rack->rc_enobuf < 0x7f)
22395 				rack->rc_enobuf++;
22396 			if (slot < (10 * HPTS_USEC_IN_MSEC))
22397 				slot = 10 * HPTS_USEC_IN_MSEC;
22398 			if (rack->r_ctl.crte != NULL) {
22399 				counter_u64_add(rack_saw_enobuf_hw, 1);
22400 				tcp_rl_log_enobuf(rack->r_ctl.crte);
22401 			}
22402 			counter_u64_add(rack_saw_enobuf, 1);
22403 			goto enobufs;
22404 		case EMSGSIZE:
22405 			/*
22406 			 * For some reason the interface we used initially
22407 			 * to send segments changed to another or lowered
22408 			 * its MTU. If TSO was active we either got an
22409 			 * interface without TSO capabilits or TSO was
22410 			 * turned off. If we obtained mtu from ip_output()
22411 			 * then update it and try again.
22412 			 */
22413 			if (tso)
22414 				tp->t_flags &= ~TF_TSO;
22415 			if (mtu != 0) {
22416 				int saved_mtu;
22417 
22418 				saved_mtu = tp->t_maxseg;
22419 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
22420 				if (saved_mtu > tp->t_maxseg) {
22421 					goto again;
22422 				}
22423 			}
22424 			slot = 10 * HPTS_USEC_IN_MSEC;
22425 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22426 #ifdef TCP_ACCOUNTING
22427 			crtsc = get_cyclecount();
22428 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22429 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22430 			}
22431 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22432 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22433 			}
22434 			sched_unpin();
22435 #endif
22436 			return (error);
22437 		case ENETUNREACH:
22438 			counter_u64_add(rack_saw_enetunreach, 1);
22439 		case EHOSTDOWN:
22440 		case EHOSTUNREACH:
22441 		case ENETDOWN:
22442 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
22443 				tp->t_softerror = error;
22444 			}
22445 			/* FALLTHROUGH */
22446 		default:
22447 			slot = 10 * HPTS_USEC_IN_MSEC;
22448 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22449 #ifdef TCP_ACCOUNTING
22450 			crtsc = get_cyclecount();
22451 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22452 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22453 			}
22454 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22455 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22456 			}
22457 			sched_unpin();
22458 #endif
22459 			return (error);
22460 		}
22461 	} else {
22462 		rack->rc_enobuf = 0;
22463 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22464 			rack->r_ctl.retran_during_recovery += len;
22465 	}
22466 	KMOD_TCPSTAT_INC(tcps_sndtotal);
22467 
22468 	/*
22469 	 * Data sent (as far as we can tell). If this advertises a larger
22470 	 * window than any other segment, then remember the size of the
22471 	 * advertised window. Any pending ACK has now been sent.
22472 	 */
22473 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22474 		tp->rcv_adv = tp->rcv_nxt + recwin;
22475 
22476 	tp->last_ack_sent = tp->rcv_nxt;
22477 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22478 enobufs:
22479 	if (sendalot) {
22480 		/* Do we need to turn off sendalot? */
22481 		if (rack->r_ctl.rc_pace_max_segs &&
22482 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
22483 			/* We hit our max. */
22484 			sendalot = 0;
22485 		} else if ((rack->rc_user_set_max_segs) &&
22486 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
22487 			/* We hit the user defined max */
22488 			sendalot = 0;
22489 		}
22490 	}
22491 	if ((error == 0) && (flags & TH_FIN))
22492 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22493 	if (flags & TH_RST) {
22494 		/*
22495 		 * We don't send again after sending a RST.
22496 		 */
22497 		slot = 0;
22498 		sendalot = 0;
22499 		if (error == 0)
22500 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22501 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22502 		/*
22503 		 * Get our pacing rate, if an error
22504 		 * occurred in sending (ENOBUF) we would
22505 		 * hit the else if with slot preset. Other
22506 		 * errors return.
22507 		 */
22508 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
22509 	}
22510 	if (rsm &&
22511 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
22512 	    rack->use_rack_rr) {
22513 		/* Its a retransmit and we use the rack cheat? */
22514 		if ((slot == 0) ||
22515 		    (rack->rc_always_pace == 0) ||
22516 		    (rack->r_rr_config == 1)) {
22517 			/*
22518 			 * We have no pacing set or we
22519 			 * are using old-style rack or
22520 			 * we are overridden to use the old 1ms pacing.
22521 			 */
22522 			slot = rack->r_ctl.rc_min_to;
22523 		}
22524 	}
22525 	/* We have sent clear the flag */
22526 	rack->r_ent_rec_ns = 0;
22527 	if (rack->r_must_retran) {
22528 		if (rsm) {
22529 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22530 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22531 				/*
22532 				 * We have retransmitted all.
22533 				 */
22534 				rack->r_must_retran = 0;
22535 				rack->r_ctl.rc_out_at_rto = 0;
22536 			}
22537 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22538 			/*
22539 			 * Sending new data will also kill
22540 			 * the loop.
22541 			 */
22542 			rack->r_must_retran = 0;
22543 			rack->r_ctl.rc_out_at_rto = 0;
22544 		}
22545 	}
22546 	rack->r_ctl.fsb.recwin = recwin;
22547 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22548 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22549 		/*
22550 		 * We hit an RTO and now have past snd_max at the RTO
22551 		 * clear all the WAS flags.
22552 		 */
22553 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22554 	}
22555 	if (slot) {
22556 		/* set the rack tcb into the slot N */
22557 		if ((error == 0) &&
22558 		    rack_use_rfo &&
22559 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22560 		    (rsm == NULL) &&
22561 		    (tp->snd_nxt == tp->snd_max) &&
22562 		    (ipoptlen == 0) &&
22563 		    (tp->rcv_numsacks == 0) &&
22564 		    rack->r_fsb_inited &&
22565 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22566 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22567 		    (rack->r_must_retran == 0) &&
22568 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22569 		    (len > 0) && (orig_len > 0) &&
22570 		    (orig_len > len) &&
22571 		    ((orig_len - len) >= segsiz) &&
22572 		    ((optlen == 0) ||
22573 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22574 			/* We can send at least one more MSS using our fsb */
22575 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22576 					       segsiz, pace_max_seg, hw_tls, flags);
22577 		} else
22578 			rack->r_fast_output = 0;
22579 		rack_log_fsb(rack, tp, so, flags,
22580 			     ipoptlen, orig_len, len, error,
22581 			     (rsm == NULL), optlen, __LINE__, 2);
22582 	} else if (sendalot) {
22583 		int ret;
22584 
22585 		sack_rxmit = 0;
22586 		if ((error == 0) &&
22587 		    rack_use_rfo &&
22588 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22589 		    (rsm == NULL) &&
22590 		    (ipoptlen == 0) &&
22591 		    (tp->rcv_numsacks == 0) &&
22592 		    (tp->snd_nxt == tp->snd_max) &&
22593 		    (rack->r_must_retran == 0) &&
22594 		    rack->r_fsb_inited &&
22595 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22596 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22597 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22598 		    (len > 0) && (orig_len > 0) &&
22599 		    (orig_len > len) &&
22600 		    ((orig_len - len) >= segsiz) &&
22601 		    ((optlen == 0) ||
22602 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22603 			/* we can use fast_output for more */
22604 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22605 					       segsiz, pace_max_seg, hw_tls, flags);
22606 			if (rack->r_fast_output) {
22607 				error = 0;
22608 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22609 				if (ret >= 0)
22610 					return (ret);
22611 			        else if (error)
22612 					goto nomore;
22613 
22614 			}
22615 		}
22616 		goto again;
22617 	}
22618 	/* Assure when we leave that snd_nxt will point to top */
22619 skip_all_send:
22620 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22621 		tp->snd_nxt = tp->snd_max;
22622 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22623 #ifdef TCP_ACCOUNTING
22624 	crtsc = get_cyclecount() - ts_val;
22625 	if (tot_len_this_send) {
22626 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22627 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22628 		}
22629 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22630 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22631 		}
22632 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22633 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22634 		}
22635 	} else {
22636 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22637 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
22638 		}
22639 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22640 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22641 		}
22642 	}
22643 	sched_unpin();
22644 #endif
22645 	if (error == ENOBUFS)
22646 		error = 0;
22647 	return (error);
22648 }
22649 
22650 static void
22651 rack_update_seg(struct tcp_rack *rack)
22652 {
22653 	uint32_t orig_val;
22654 
22655 	orig_val = rack->r_ctl.rc_pace_max_segs;
22656 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22657 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
22658 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22659 }
22660 
22661 static void
22662 rack_mtu_change(struct tcpcb *tp)
22663 {
22664 	/*
22665 	 * The MSS may have changed
22666 	 */
22667 	struct tcp_rack *rack;
22668 	struct rack_sendmap *rsm;
22669 
22670 	rack = (struct tcp_rack *)tp->t_fb_ptr;
22671 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22672 		/*
22673 		 * The MTU has changed we need to resend everything
22674 		 * since all we have sent is lost. We first fix
22675 		 * up the mtu though.
22676 		 */
22677 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
22678 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
22679 		rack_remxt_tmr(tp);
22680 		rack->r_fast_output = 0;
22681 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22682 						rack->r_ctl.rc_sacked);
22683 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22684 		rack->r_must_retran = 1;
22685 		/* Mark all inflight to needing to be rxt'd */
22686 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22687 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22688 		}
22689 	}
22690 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22691 	/* We don't use snd_nxt to retransmit */
22692 	tp->snd_nxt = tp->snd_max;
22693 }
22694 
22695 static int
22696 rack_set_dgp(struct tcp_rack *rack)
22697 {
22698 	/* pace_always=1 */
22699 	if (rack->rc_always_pace == 0) {
22700 		if (tcp_can_enable_pacing() == 0)
22701 			return (EBUSY);
22702 	}
22703 	rack->rc_fillcw_apply_discount = 0;
22704 	rack->dgp_on = 1;
22705 	rack->rc_always_pace = 1;
22706 	rack->use_fixed_rate = 0;
22707 	if (rack->gp_ready)
22708 		rack_set_cc_pacing(rack);
22709 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22710 	rack->rack_attempt_hdwr_pace = 0;
22711 	/* rxt settings */
22712 	rack->full_size_rxt = 1;
22713 	rack->shape_rxt_to_pacing_min  = 0;
22714 	/* cmpack=1 */
22715 	rack->r_use_cmp_ack = 1;
22716 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22717 	    rack->r_use_cmp_ack)
22718 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22719 	/* scwnd=1 */
22720 	rack->rack_enable_scwnd = 1;
22721 	/* dynamic=100 */
22722 	rack->rc_gp_dyn_mul = 1;
22723 	/* gp_inc_ca */
22724 	rack->r_ctl.rack_per_of_gp_ca = 100;
22725 	/* rrr_conf=3 */
22726 	rack->r_rr_config = 3;
22727 	/* npush=2 */
22728 	rack->r_ctl.rc_no_push_at_mrtt = 2;
22729 	/* fillcw=1 */
22730 	if (rack->r_cwnd_was_clamped == 0) {
22731 		rack->rc_pace_to_cwnd = 1;
22732 	} else {
22733 		rack->rc_pace_to_cwnd = 0;
22734 		/* Reset all multipliers to 100.0 so just the measured bw */
22735 		rack->r_ctl.rack_per_of_gp_ss = 100;
22736 		rack->r_ctl.rack_per_of_gp_ca = 100;
22737 	}
22738 	rack->rc_pace_fill_if_rttin_range = 0;
22739 	rack->rtt_limit_mul = 0;
22740 	/* noprr=1 */
22741 	rack->rack_no_prr = 1;
22742 	/* lscwnd=1 */
22743 	rack->r_limit_scw = 1;
22744 	/* gp_inc_rec */
22745 	rack->r_ctl.rack_per_of_gp_rec = 90;
22746 	rack_client_buffer_level_set(rack);
22747 	return (0);
22748 }
22749 
22750 
22751 
22752 static int
22753 rack_set_profile(struct tcp_rack *rack, int prof)
22754 {
22755 	int err = EINVAL;
22756 	if (prof == 1) {
22757 		/*
22758 		 * Profile 1 is "standard" DGP. It ignores
22759 		 * client buffer level.
22760 		 */
22761 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL0;
22762 		err = rack_set_dgp(rack);
22763 		if (err)
22764 			return (err);
22765 	} else if (prof == 2) {
22766 		/*
22767 		 * Profile 2 is DGP. Less aggressive with
22768 		 * respect to client buffer level.
22769 		 */
22770 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL1;
22771 		err = rack_set_dgp(rack);
22772 		if (err)
22773 			return (err);
22774 	} else if (prof == 3) {
22775 		/*
22776 		 * Profile 3 is DGP. Even Less aggressive with
22777 		 * respect to client buffer level.
22778 		 */
22779 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL2;
22780 		err = rack_set_dgp(rack);
22781 		if (err)
22782 			return (err);
22783 	} else if (prof == 4) {
22784 		/*
22785 		 * Profile 4 is DGP with the most responsiveness
22786 		 * to client buffer level.
22787 		 */
22788 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL3;
22789 		err = rack_set_dgp(rack);
22790 		if (err)
22791 			return (err);
22792 	} else if (prof == 5) {
22793 		err = rack_set_dgp(rack);
22794 		if (err)
22795 			return (err);
22796 		/*
22797 		 * By turning DGP off we change the rate
22798 		 * picked to be only the one the cwnd and rtt
22799 		 * get us.
22800 		 */
22801 		rack->dgp_on = 0;
22802 	} else if (prof == 6) {
22803 		err = rack_set_dgp(rack);
22804 		if (err)
22805 			return (err);
22806 		/*
22807 		 * Profile 6 tweaks DGP so that it will apply to
22808 		 * fill-cw the same settings that profile5 does
22809 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22810 		 */
22811 		rack->rc_fillcw_apply_discount = 1;
22812 	} else if (prof == 0) {
22813 		/* This changes things back to the default settings */
22814 		rack->dgp_on = 0;
22815 		rack->rc_hybrid_mode = 0;
22816 		err = 0;
22817 		if (rack_fill_cw_state)
22818 			rack->rc_pace_to_cwnd = 1;
22819 		else
22820 			rack->rc_pace_to_cwnd = 0;
22821 		if (rack->rc_always_pace) {
22822 			tcp_decrement_paced_conn();
22823 			rack_undo_cc_pacing(rack);
22824 			rack->rc_always_pace = 0;
22825 		}
22826 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22827 			rack->rc_always_pace = 1;
22828 			if (rack->rack_hibeta)
22829 				rack_set_cc_pacing(rack);
22830 		} else
22831 			rack->rc_always_pace = 0;
22832 		if (rack_dsack_std_based & 0x1) {
22833 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22834 			rack->rc_rack_tmr_std_based = 1;
22835 		}
22836 		if (rack_dsack_std_based & 0x2) {
22837 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22838 			rack->rc_rack_use_dsack = 1;
22839 		}
22840 		if (rack_use_cmp_acks)
22841 			rack->r_use_cmp_ack = 1;
22842 		else
22843 			rack->r_use_cmp_ack = 0;
22844 		if (rack_disable_prr)
22845 			rack->rack_no_prr = 1;
22846 		else
22847 			rack->rack_no_prr = 0;
22848 		if (rack_gp_no_rec_chg)
22849 			rack->rc_gp_no_rec_chg = 1;
22850 		else
22851 			rack->rc_gp_no_rec_chg = 0;
22852 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22853 			rack->r_mbuf_queue = 1;
22854 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22855 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22856 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22857 		} else {
22858 			rack->r_mbuf_queue = 0;
22859 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22860 		}
22861 		if (rack_enable_shared_cwnd)
22862 			rack->rack_enable_scwnd = 1;
22863 		else
22864 			rack->rack_enable_scwnd = 0;
22865 		if (rack_do_dyn_mul) {
22866 			/* When dynamic adjustment is on CA needs to start at 100% */
22867 			rack->rc_gp_dyn_mul = 1;
22868 			if (rack_do_dyn_mul >= 100)
22869 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22870 		} else {
22871 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22872 			rack->rc_gp_dyn_mul = 0;
22873 		}
22874 		rack->r_rr_config = 0;
22875 		rack->r_ctl.rc_no_push_at_mrtt = 0;
22876 		rack->rc_pace_to_cwnd = 0;
22877 		rack->rc_pace_fill_if_rttin_range = 0;
22878 		rack->rtt_limit_mul = 0;
22879 
22880 		if (rack_enable_hw_pacing)
22881 			rack->rack_hdw_pace_ena = 1;
22882 		else
22883 			rack->rack_hdw_pace_ena = 0;
22884 		if (rack_disable_prr)
22885 			rack->rack_no_prr = 1;
22886 		else
22887 			rack->rack_no_prr = 0;
22888 		if (rack_limits_scwnd)
22889 			rack->r_limit_scw  = 1;
22890 		else
22891 			rack->r_limit_scw  = 0;
22892 		rack_init_retransmit_value(rack, rack_rxt_controls);
22893 		err = 0;
22894 	}
22895 	return (err);
22896 }
22897 
22898 static int
22899 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22900 {
22901 	struct deferred_opt_list *dol;
22902 
22903 	dol = malloc(sizeof(struct deferred_opt_list),
22904 		     M_TCPFSB, M_NOWAIT|M_ZERO);
22905 	if (dol == NULL) {
22906 		/*
22907 		 * No space yikes -- fail out..
22908 		 */
22909 		return (0);
22910 	}
22911 	dol->optname = sopt_name;
22912 	dol->optval = loptval;
22913 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22914 	return (1);
22915 }
22916 
22917 static int
22918 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22919 {
22920 #ifdef TCP_REQUEST_TRK
22921 	struct tcp_sendfile_track *sft;
22922 	struct timeval tv;
22923 	tcp_seq seq;
22924 	int err;
22925 
22926 	microuptime(&tv);
22927 
22928 	/*
22929 	 * If BB logging is not on we need to look at the DTL flag.
22930 	 * If its on already then those reasons override the DTL input.
22931 	 * We do this with any request, you can turn DTL on, but it does
22932 	 * not turn off at least from hybrid pacing requests.
22933 	 */
22934 	if (tcp_bblogging_on(rack->rc_tp) == 0) {
22935 		if (hybrid->hybrid_flags & TCP_HYBRID_PACING_DTL) {
22936 			/* Turn on BB point logging  */
22937 			tcp_set_bblog_state(rack->rc_tp, TCP_LOG_VIA_BBPOINTS,
22938 					    TCP_BBPOINT_REQ_LEVEL_LOGGING);
22939 		}
22940 	}
22941 	/* Make sure no fixed rate is on */
22942 	rack->use_fixed_rate = 0;
22943 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22944 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22945 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22946 	/* Now allocate or find our entry that will have these settings */
22947 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22948 	if (sft == NULL) {
22949 		rack->rc_tp->tcp_hybrid_error++;
22950 		/* no space, where would it have gone? */
22951 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22952 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22953 		return (ENOSPC);
22954 	}
22955 	/* The seq will be snd_una + everything in the buffer */
22956 	seq = sft->start_seq;
22957 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22958 		/* Disabling hybrid pacing */
22959 		if (rack->rc_hybrid_mode) {
22960 			rack_set_profile(rack, 0);
22961 			rack->rc_tp->tcp_hybrid_stop++;
22962 		}
22963 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22964 		return (0);
22965 	}
22966 	if (rack->dgp_on == 0) {
22967 		/*
22968 		 * If we have not yet turned DGP on, do so
22969 		 * now setting pure DGP mode, no buffer level
22970 		 * response.
22971 		 */
22972 		if ((err = rack_set_profile(rack, 1)) != 0){
22973 			/* Failed to turn pacing on */
22974 			rack->rc_tp->tcp_hybrid_error++;
22975 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22976 			return (err);
22977 		}
22978 	}
22979 	/* Now set in our flags */
22980 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22981 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22982 		sft->cspr = hybrid->cspr;
22983 	else
22984 		sft->cspr = 0;
22985 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22986 		sft->hint_maxseg = hybrid->hint_maxseg;
22987 	else
22988 		sft->hint_maxseg = 0;
22989 	rack->rc_hybrid_mode = 1;
22990 	rack->rc_tp->tcp_hybrid_start++;
22991 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22992 	return (0);
22993 #else
22994 	return (ENOTSUP);
22995 #endif
22996 }
22997 
22998 static int
22999 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
23000 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
23001 
23002 {
23003 	struct epoch_tracker et;
23004 	struct sockopt sopt;
23005 	struct cc_newreno_opts opt;
23006 	uint64_t val;
23007 	int error = 0;
23008 	uint16_t ca, ss;
23009 
23010 	switch (sopt_name) {
23011 	case TCP_RACK_SET_RXT_OPTIONS:
23012 		if ((optval >= 0) && (optval <= 2)) {
23013 			rack_init_retransmit_value(rack, optval);
23014 		} else {
23015 			/*
23016 			 * You must send in 0, 1 or 2 all else is
23017 			 * invalid.
23018 			 */
23019 			error = EINVAL;
23020 		}
23021 		break;
23022 	case TCP_RACK_DSACK_OPT:
23023 		RACK_OPTS_INC(tcp_rack_dsack_opt);
23024 		if (optval & 0x1) {
23025 			rack->rc_rack_tmr_std_based = 1;
23026 		} else {
23027 			rack->rc_rack_tmr_std_based = 0;
23028 		}
23029 		if (optval & 0x2) {
23030 			rack->rc_rack_use_dsack = 1;
23031 		} else {
23032 			rack->rc_rack_use_dsack = 0;
23033 		}
23034 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
23035 		break;
23036 	case TCP_RACK_PACING_DIVISOR:
23037 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
23038 		if (optval == 0) {
23039 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
23040 		} else {
23041 			if (optval < RL_MIN_DIVISOR)
23042 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
23043 			else
23044 				rack->r_ctl.pace_len_divisor = optval;
23045 		}
23046 		break;
23047 	case TCP_RACK_HI_BETA:
23048 		RACK_OPTS_INC(tcp_rack_hi_beta);
23049 		if (optval > 0) {
23050 			rack->rack_hibeta = 1;
23051 			if ((optval >= 50) &&
23052 			    (optval <= 100)) {
23053 				/*
23054 				 * User wants to set a custom beta.
23055 				 */
23056 				rack->r_ctl.saved_hibeta = optval;
23057 				if (rack->rc_pacing_cc_set)
23058 					rack_undo_cc_pacing(rack);
23059 				rack->r_ctl.rc_saved_beta.beta = optval;
23060 			}
23061 			if (rack->rc_pacing_cc_set == 0)
23062 				rack_set_cc_pacing(rack);
23063 		} else {
23064 			rack->rack_hibeta = 0;
23065 			if (rack->rc_pacing_cc_set)
23066 				rack_undo_cc_pacing(rack);
23067 		}
23068 		break;
23069 	case TCP_RACK_PACING_BETA:
23070 		RACK_OPTS_INC(tcp_rack_beta);
23071 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23072 			/* This only works for newreno. */
23073 			error = EINVAL;
23074 			break;
23075 		}
23076 		if (rack->rc_pacing_cc_set) {
23077 			/*
23078 			 * Set them into the real CC module
23079 			 * whats in the rack pcb is the old values
23080 			 * to be used on restoral/
23081 			 */
23082 			sopt.sopt_dir = SOPT_SET;
23083 			opt.name = CC_NEWRENO_BETA;
23084 			opt.val = optval;
23085 			if (CC_ALGO(tp)->ctl_output != NULL)
23086 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23087 			else {
23088 				error = ENOENT;
23089 				break;
23090 			}
23091 		} else {
23092 			/*
23093 			 * Not pacing yet so set it into our local
23094 			 * rack pcb storage.
23095 			 */
23096 			rack->r_ctl.rc_saved_beta.beta = optval;
23097 		}
23098 		break;
23099 	case TCP_RACK_TIMER_SLOP:
23100 		RACK_OPTS_INC(tcp_rack_timer_slop);
23101 		rack->r_ctl.timer_slop = optval;
23102 		if (rack->rc_tp->t_srtt) {
23103 			/*
23104 			 * If we have an SRTT lets update t_rxtcur
23105 			 * to have the new slop.
23106 			 */
23107 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23108 					   rack_rto_min, rack_rto_max,
23109 					   rack->r_ctl.timer_slop);
23110 		}
23111 		break;
23112 	case TCP_RACK_PACING_BETA_ECN:
23113 		RACK_OPTS_INC(tcp_rack_beta_ecn);
23114 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23115 			/* This only works for newreno. */
23116 			error = EINVAL;
23117 			break;
23118 		}
23119 		if (rack->rc_pacing_cc_set) {
23120 			/*
23121 			 * Set them into the real CC module
23122 			 * whats in the rack pcb is the old values
23123 			 * to be used on restoral/
23124 			 */
23125 			sopt.sopt_dir = SOPT_SET;
23126 			opt.name = CC_NEWRENO_BETA_ECN;
23127 			opt.val = optval;
23128 			if (CC_ALGO(tp)->ctl_output != NULL)
23129 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23130 			else
23131 				error = ENOENT;
23132 		} else {
23133 			/*
23134 			 * Not pacing yet so set it into our local
23135 			 * rack pcb storage.
23136 			 */
23137 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
23138 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
23139 		}
23140 		break;
23141 	case TCP_DEFER_OPTIONS:
23142 		RACK_OPTS_INC(tcp_defer_opt);
23143 		if (optval) {
23144 			if (rack->gp_ready) {
23145 				/* Too late */
23146 				error = EINVAL;
23147 				break;
23148 			}
23149 			rack->defer_options = 1;
23150 		} else
23151 			rack->defer_options = 0;
23152 		break;
23153 	case TCP_RACK_MEASURE_CNT:
23154 		RACK_OPTS_INC(tcp_rack_measure_cnt);
23155 		if (optval && (optval <= 0xff)) {
23156 			rack->r_ctl.req_measurements = optval;
23157 		} else
23158 			error = EINVAL;
23159 		break;
23160 	case TCP_REC_ABC_VAL:
23161 		RACK_OPTS_INC(tcp_rec_abc_val);
23162 		if (optval > 0)
23163 			rack->r_use_labc_for_rec = 1;
23164 		else
23165 			rack->r_use_labc_for_rec = 0;
23166 		break;
23167 	case TCP_RACK_ABC_VAL:
23168 		RACK_OPTS_INC(tcp_rack_abc_val);
23169 		if ((optval > 0) && (optval < 255))
23170 			rack->rc_labc = optval;
23171 		else
23172 			error = EINVAL;
23173 		break;
23174 	case TCP_HDWR_UP_ONLY:
23175 		RACK_OPTS_INC(tcp_pacing_up_only);
23176 		if (optval)
23177 			rack->r_up_only = 1;
23178 		else
23179 			rack->r_up_only = 0;
23180 		break;
23181 	case TCP_PACING_RATE_CAP:
23182 		RACK_OPTS_INC(tcp_pacing_rate_cap);
23183 		rack->r_ctl.bw_rate_cap = loptval;
23184 		break;
23185 	case TCP_HYBRID_PACING:
23186 		if (hybrid == NULL) {
23187 			error = EINVAL;
23188 			break;
23189 		}
23190 		error = process_hybrid_pacing(rack, hybrid);
23191 		break;
23192 	case TCP_RACK_PROFILE:
23193 		RACK_OPTS_INC(tcp_profile);
23194 		error = rack_set_profile(rack, optval);
23195 		break;
23196 	case TCP_USE_CMP_ACKS:
23197 		RACK_OPTS_INC(tcp_use_cmp_acks);
23198 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23199 			/* You can't turn it off once its on! */
23200 			error = EINVAL;
23201 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23202 			rack->r_use_cmp_ack = 1;
23203 			rack->r_mbuf_queue = 1;
23204 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23205 		}
23206 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23207 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
23208 		break;
23209 	case TCP_SHARED_CWND_TIME_LIMIT:
23210 		RACK_OPTS_INC(tcp_lscwnd);
23211 		if (optval)
23212 			rack->r_limit_scw = 1;
23213 		else
23214 			rack->r_limit_scw = 0;
23215 		break;
23216 	case TCP_RACK_DGP_IN_REC:
23217 		RACK_OPTS_INC(tcp_dgp_in_rec);
23218 		if (optval)
23219 			rack->r_ctl.full_dgp_in_rec = 1;
23220 		else
23221 			rack->r_ctl.full_dgp_in_rec = 0;
23222 		break;
23223 	case TCP_RXT_CLAMP:
23224 		RACK_OPTS_INC(tcp_rxt_clamp);
23225 		rack_translate_clamp_value(rack, optval);
23226 		break;
23227  	case TCP_RACK_PACE_TO_FILL:
23228 		RACK_OPTS_INC(tcp_fillcw);
23229 		if (optval == 0)
23230 			rack->rc_pace_to_cwnd = 0;
23231 		else {
23232 			rack->rc_pace_to_cwnd = 1;
23233 			if (optval > 1)
23234 				rack->r_fill_less_agg = 1;
23235 		}
23236 		if ((optval >= rack_gp_rtt_maxmul) &&
23237 		    rack_gp_rtt_maxmul &&
23238 		    (optval < 0xf)) {
23239 			rack->rc_pace_fill_if_rttin_range = 1;
23240 			rack->rtt_limit_mul = optval;
23241 		} else {
23242 			rack->rc_pace_fill_if_rttin_range = 0;
23243 			rack->rtt_limit_mul = 0;
23244 		}
23245 		break;
23246 	case TCP_RACK_NO_PUSH_AT_MAX:
23247 		RACK_OPTS_INC(tcp_npush);
23248 		if (optval == 0)
23249 			rack->r_ctl.rc_no_push_at_mrtt = 0;
23250 		else if (optval < 0xff)
23251 			rack->r_ctl.rc_no_push_at_mrtt = optval;
23252 		else
23253 			error = EINVAL;
23254 		break;
23255 	case TCP_SHARED_CWND_ENABLE:
23256 		RACK_OPTS_INC(tcp_rack_scwnd);
23257 		if (optval == 0)
23258 			rack->rack_enable_scwnd = 0;
23259 		else
23260 			rack->rack_enable_scwnd = 1;
23261 		break;
23262 	case TCP_RACK_MBUF_QUEUE:
23263 		/* Now do we use the LRO mbuf-queue feature */
23264 		RACK_OPTS_INC(tcp_rack_mbufq);
23265 		if (optval || rack->r_use_cmp_ack)
23266 			rack->r_mbuf_queue = 1;
23267 		else
23268 			rack->r_mbuf_queue = 0;
23269 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23270 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23271 		else
23272 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23273 		break;
23274 	case TCP_RACK_NONRXT_CFG_RATE:
23275 		RACK_OPTS_INC(tcp_rack_cfg_rate);
23276 		if (optval == 0)
23277 			rack->rack_rec_nonrxt_use_cr = 0;
23278 		else
23279 			rack->rack_rec_nonrxt_use_cr = 1;
23280 		break;
23281 	case TCP_NO_PRR:
23282 		RACK_OPTS_INC(tcp_rack_noprr);
23283 		if (optval == 0)
23284 			rack->rack_no_prr = 0;
23285 		else if (optval == 1)
23286 			rack->rack_no_prr = 1;
23287 		else if (optval == 2)
23288 			rack->no_prr_addback = 1;
23289 		else
23290 			error = EINVAL;
23291 		break;
23292 	case TCP_TIMELY_DYN_ADJ:
23293 		RACK_OPTS_INC(tcp_timely_dyn);
23294 		if (optval == 0)
23295 			rack->rc_gp_dyn_mul = 0;
23296 		else {
23297 			rack->rc_gp_dyn_mul = 1;
23298 			if (optval >= 100) {
23299 				/*
23300 				 * If the user sets something 100 or more
23301 				 * its the gp_ca value.
23302 				 */
23303 				rack->r_ctl.rack_per_of_gp_ca  = optval;
23304 			}
23305 		}
23306 		break;
23307 	case TCP_RACK_DO_DETECTION:
23308 		RACK_OPTS_INC(tcp_rack_do_detection);
23309 		if (optval == 0)
23310 			rack->do_detection = 0;
23311 		else
23312 			rack->do_detection = 1;
23313 		break;
23314 	case TCP_RACK_TLP_USE:
23315 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23316 			error = EINVAL;
23317 			break;
23318 		}
23319 		RACK_OPTS_INC(tcp_tlp_use);
23320 		rack->rack_tlp_threshold_use = optval;
23321 		break;
23322 	case TCP_RACK_TLP_REDUCE:
23323 		/* RACK TLP cwnd reduction (bool) */
23324 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
23325 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23326 		break;
23327 		/*  Pacing related ones */
23328 	case TCP_RACK_PACE_ALWAYS:
23329 		/*
23330 		 * zero is old rack method, 1 is new
23331 		 * method using a pacing rate.
23332 		 */
23333 		RACK_OPTS_INC(tcp_rack_pace_always);
23334 		if (optval > 0) {
23335 			if (rack->rc_always_pace) {
23336 				error = EALREADY;
23337 				break;
23338 			} else if (tcp_can_enable_pacing()) {
23339 				rack->rc_always_pace = 1;
23340 				if (rack->rack_hibeta)
23341 					rack_set_cc_pacing(rack);
23342 			}
23343 			else {
23344 				error = ENOSPC;
23345 				break;
23346 			}
23347 		} else {
23348 			if (rack->rc_always_pace) {
23349 				tcp_decrement_paced_conn();
23350 				rack->rc_always_pace = 0;
23351 				rack_undo_cc_pacing(rack);
23352 			}
23353 		}
23354 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23355 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23356 		else
23357 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23358 		/* A rate may be set irate or other, if so set seg size */
23359 		rack_update_seg(rack);
23360 		break;
23361 	case TCP_BBR_RACK_INIT_RATE:
23362 		RACK_OPTS_INC(tcp_initial_rate);
23363 		val = optval;
23364 		/* Change from kbits per second to bytes per second */
23365 		val *= 1000;
23366 		val /= 8;
23367 		rack->r_ctl.init_rate = val;
23368 		if (rack->rc_init_win != rack_default_init_window) {
23369 			uint32_t win, snt;
23370 
23371 			/*
23372 			 * Options don't always get applied
23373 			 * in the order you think. So in order
23374 			 * to assure we update a cwnd we need
23375 			 * to check and see if we are still
23376 			 * where we should raise the cwnd.
23377 			 */
23378 			win = rc_init_window(rack);
23379 			if (SEQ_GT(tp->snd_max, tp->iss))
23380 				snt = tp->snd_max - tp->iss;
23381 			else
23382 				snt = 0;
23383 			if ((snt < win) &&
23384 			    (tp->snd_cwnd < win))
23385 				tp->snd_cwnd = win;
23386 		}
23387 		if (rack->rc_always_pace)
23388 			rack_update_seg(rack);
23389 		break;
23390 	case TCP_BBR_IWINTSO:
23391 		RACK_OPTS_INC(tcp_initial_win);
23392 		if (optval && (optval <= 0xff)) {
23393 			uint32_t win, snt;
23394 
23395 			rack->rc_init_win = optval;
23396 			win = rc_init_window(rack);
23397 			if (SEQ_GT(tp->snd_max, tp->iss))
23398 				snt = tp->snd_max - tp->iss;
23399 			else
23400 				snt = 0;
23401 			if ((snt < win) &&
23402 			    (tp->t_srtt |
23403 			     rack->r_ctl.init_rate)) {
23404 				/*
23405 				 * We are not past the initial window
23406 				 * and we have some bases for pacing,
23407 				 * so we need to possibly adjust up
23408 				 * the cwnd. Note even if we don't set
23409 				 * the cwnd, its still ok to raise the rc_init_win
23410 				 * which can be used coming out of idle when we
23411 				 * would have a rate.
23412 				 */
23413 				if (tp->snd_cwnd < win)
23414 					tp->snd_cwnd = win;
23415 			}
23416 			if (rack->rc_always_pace)
23417 				rack_update_seg(rack);
23418 		} else
23419 			error = EINVAL;
23420 		break;
23421 	case TCP_RACK_FORCE_MSEG:
23422 		RACK_OPTS_INC(tcp_rack_force_max_seg);
23423 		if (optval)
23424 			rack->rc_force_max_seg = 1;
23425 		else
23426 			rack->rc_force_max_seg = 0;
23427 		break;
23428 	case TCP_RACK_PACE_MIN_SEG:
23429 		RACK_OPTS_INC(tcp_rack_min_seg);
23430 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23431 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23432 		break;
23433 	case TCP_RACK_PACE_MAX_SEG:
23434 		/* Max segments size in a pace in bytes */
23435 		RACK_OPTS_INC(tcp_rack_max_seg);
23436 		if (optval <= MAX_USER_SET_SEG)
23437 			rack->rc_user_set_max_segs = optval;
23438 		else
23439 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23440 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23441 		break;
23442 	case TCP_RACK_PACE_RATE_REC:
23443 		/* Set the fixed pacing rate in Bytes per second ca */
23444 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23445 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23446 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23447 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23448 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23449 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23450 		rack->use_fixed_rate = 1;
23451 		if (rack->rack_hibeta)
23452 			rack_set_cc_pacing(rack);
23453 		rack_log_pacing_delay_calc(rack,
23454 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23455 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23456 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23457 					   __LINE__, NULL,0);
23458 		break;
23459 
23460 	case TCP_RACK_PACE_RATE_SS:
23461 		/* Set the fixed pacing rate in Bytes per second ca */
23462 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23463 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23464 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23465 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23466 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23467 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23468 		rack->use_fixed_rate = 1;
23469 		if (rack->rack_hibeta)
23470 			rack_set_cc_pacing(rack);
23471 		rack_log_pacing_delay_calc(rack,
23472 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23473 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23474 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23475 					   __LINE__, NULL, 0);
23476 		break;
23477 
23478 	case TCP_RACK_PACE_RATE_CA:
23479 		/* Set the fixed pacing rate in Bytes per second ca */
23480 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23481 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23482 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23483 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23484 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23485 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23486 		rack->use_fixed_rate = 1;
23487 		if (rack->rack_hibeta)
23488 			rack_set_cc_pacing(rack);
23489 		rack_log_pacing_delay_calc(rack,
23490 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23491 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23492 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23493 					   __LINE__, NULL, 0);
23494 		break;
23495 	case TCP_RACK_GP_INCREASE_REC:
23496 		RACK_OPTS_INC(tcp_gp_inc_rec);
23497 		rack->r_ctl.rack_per_of_gp_rec = optval;
23498 		rack_log_pacing_delay_calc(rack,
23499 					   rack->r_ctl.rack_per_of_gp_ss,
23500 					   rack->r_ctl.rack_per_of_gp_ca,
23501 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23502 					   __LINE__, NULL, 0);
23503 		break;
23504 	case TCP_RACK_GP_INCREASE_CA:
23505 		RACK_OPTS_INC(tcp_gp_inc_ca);
23506 		ca = optval;
23507 		if (ca < 100) {
23508 			/*
23509 			 * We don't allow any reduction
23510 			 * over the GP b/w.
23511 			 */
23512 			error = EINVAL;
23513 			break;
23514 		}
23515 		rack->r_ctl.rack_per_of_gp_ca = ca;
23516 		rack_log_pacing_delay_calc(rack,
23517 					   rack->r_ctl.rack_per_of_gp_ss,
23518 					   rack->r_ctl.rack_per_of_gp_ca,
23519 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23520 					   __LINE__, NULL, 0);
23521 		break;
23522 	case TCP_RACK_GP_INCREASE_SS:
23523 		RACK_OPTS_INC(tcp_gp_inc_ss);
23524 		ss = optval;
23525 		if (ss < 100) {
23526 			/*
23527 			 * We don't allow any reduction
23528 			 * over the GP b/w.
23529 			 */
23530 			error = EINVAL;
23531 			break;
23532 		}
23533 		rack->r_ctl.rack_per_of_gp_ss = ss;
23534 		rack_log_pacing_delay_calc(rack,
23535 					   rack->r_ctl.rack_per_of_gp_ss,
23536 					   rack->r_ctl.rack_per_of_gp_ca,
23537 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23538 					   __LINE__, NULL, 0);
23539 		break;
23540 	case TCP_RACK_RR_CONF:
23541 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23542 		if (optval && optval <= 3)
23543 			rack->r_rr_config = optval;
23544 		else
23545 			rack->r_rr_config = 0;
23546 		break;
23547 	case TCP_PACING_DND:			/*  URL:dnd */
23548 		if (optval > 0)
23549 			rack->rc_pace_dnd = 1;
23550 		else
23551 			rack->rc_pace_dnd = 0;
23552 		break;
23553 	case TCP_HDWR_RATE_CAP:
23554 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
23555 		if (optval) {
23556 			if (rack->r_rack_hw_rate_caps == 0)
23557 				rack->r_rack_hw_rate_caps = 1;
23558 			else
23559 				error = EALREADY;
23560 		} else {
23561 			rack->r_rack_hw_rate_caps = 0;
23562 		}
23563 		break;
23564 	case TCP_RACK_SPLIT_LIMIT:
23565 		RACK_OPTS_INC(tcp_split_limit);
23566 		rack->r_ctl.rc_split_limit = optval;
23567 		break;
23568 	case TCP_BBR_HDWR_PACE:
23569 		RACK_OPTS_INC(tcp_hdwr_pacing);
23570 		if (optval){
23571 			if (rack->rack_hdrw_pacing == 0) {
23572 				rack->rack_hdw_pace_ena = 1;
23573 				rack->rack_attempt_hdwr_pace = 0;
23574 			} else
23575 				error = EALREADY;
23576 		} else {
23577 			rack->rack_hdw_pace_ena = 0;
23578 #ifdef RATELIMIT
23579 			if (rack->r_ctl.crte != NULL) {
23580 				rack->rack_hdrw_pacing = 0;
23581 				rack->rack_attempt_hdwr_pace = 0;
23582 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23583 				rack->r_ctl.crte = NULL;
23584 			}
23585 #endif
23586 		}
23587 		break;
23588 		/*  End Pacing related ones */
23589 	case TCP_RACK_PRR_SENDALOT:
23590 		/* Allow PRR to send more than one seg */
23591 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
23592 		rack->r_ctl.rc_prr_sendalot = optval;
23593 		break;
23594 	case TCP_RACK_MIN_TO:
23595 		/* Minimum time between rack t-o's in ms */
23596 		RACK_OPTS_INC(tcp_rack_min_to);
23597 		rack->r_ctl.rc_min_to = optval;
23598 		break;
23599 	case TCP_RACK_EARLY_SEG:
23600 		/* If early recovery max segments */
23601 		RACK_OPTS_INC(tcp_rack_early_seg);
23602 		rack->r_ctl.rc_early_recovery_segs = optval;
23603 		break;
23604 	case TCP_RACK_ENABLE_HYSTART:
23605 	{
23606 		if (optval) {
23607 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23608 			if (rack_do_hystart > RACK_HYSTART_ON)
23609 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23610 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23611 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23612 		} else {
23613 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23614 		}
23615 	}
23616 	break;
23617 	case TCP_RACK_REORD_THRESH:
23618 		/* RACK reorder threshold (shift amount) */
23619 		RACK_OPTS_INC(tcp_rack_reord_thresh);
23620 		if ((optval > 0) && (optval < 31))
23621 			rack->r_ctl.rc_reorder_shift = optval;
23622 		else
23623 			error = EINVAL;
23624 		break;
23625 	case TCP_RACK_REORD_FADE:
23626 		/* Does reordering fade after ms time */
23627 		RACK_OPTS_INC(tcp_rack_reord_fade);
23628 		rack->r_ctl.rc_reorder_fade = optval;
23629 		break;
23630 	case TCP_RACK_TLP_THRESH:
23631 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
23632 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
23633 		if (optval)
23634 			rack->r_ctl.rc_tlp_threshold = optval;
23635 		else
23636 			error = EINVAL;
23637 		break;
23638 	case TCP_BBR_USE_RACK_RR:
23639 		RACK_OPTS_INC(tcp_rack_rr);
23640 		if (optval)
23641 			rack->use_rack_rr = 1;
23642 		else
23643 			rack->use_rack_rr = 0;
23644 		break;
23645 	case TCP_RACK_PKT_DELAY:
23646 		/* RACK added ms i.e. rack-rtt + reord + N */
23647 		RACK_OPTS_INC(tcp_rack_pkt_delay);
23648 		rack->r_ctl.rc_pkt_delay = optval;
23649 		break;
23650 	case TCP_DELACK:
23651 		RACK_OPTS_INC(tcp_rack_delayed_ack);
23652 		if (optval == 0)
23653 			tp->t_delayed_ack = 0;
23654 		else
23655 			tp->t_delayed_ack = 1;
23656 		if (tp->t_flags & TF_DELACK) {
23657 			tp->t_flags &= ~TF_DELACK;
23658 			tp->t_flags |= TF_ACKNOW;
23659 			NET_EPOCH_ENTER(et);
23660 			rack_output(tp);
23661 			NET_EPOCH_EXIT(et);
23662 		}
23663 		break;
23664 
23665 	case TCP_BBR_RACK_RTT_USE:
23666 		RACK_OPTS_INC(tcp_rack_rtt_use);
23667 		if ((optval != USE_RTT_HIGH) &&
23668 		    (optval != USE_RTT_LOW) &&
23669 		    (optval != USE_RTT_AVG))
23670 			error = EINVAL;
23671 		else
23672 			rack->r_ctl.rc_rate_sample_method = optval;
23673 		break;
23674 	case TCP_DATA_AFTER_CLOSE:
23675 		RACK_OPTS_INC(tcp_data_after_close);
23676 		if (optval)
23677 			rack->rc_allow_data_af_clo = 1;
23678 		else
23679 			rack->rc_allow_data_af_clo = 0;
23680 		break;
23681 	default:
23682 		break;
23683 	}
23684 	tcp_log_socket_option(tp, sopt_name, optval, error);
23685 	return (error);
23686 }
23687 
23688 
23689 static void
23690 rack_apply_deferred_options(struct tcp_rack *rack)
23691 {
23692 	struct deferred_opt_list *dol, *sdol;
23693 	uint32_t s_optval;
23694 
23695 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
23696 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
23697 		/* Disadvantage of deferal is you loose the error return */
23698 		s_optval = (uint32_t)dol->optval;
23699 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
23700 		free(dol, M_TCPDO);
23701 	}
23702 }
23703 
23704 static void
23705 rack_hw_tls_change(struct tcpcb *tp, int chg)
23706 {
23707 	/* Update HW tls state */
23708 	struct tcp_rack *rack;
23709 
23710 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23711 	if (chg)
23712 		rack->r_ctl.fsb.hw_tls = 1;
23713 	else
23714 		rack->r_ctl.fsb.hw_tls = 0;
23715 }
23716 
23717 static int
23718 rack_pru_options(struct tcpcb *tp, int flags)
23719 {
23720 	if (flags & PRUS_OOB)
23721 		return (EOPNOTSUPP);
23722 	return (0);
23723 }
23724 
23725 static bool
23726 rack_wake_check(struct tcpcb *tp)
23727 {
23728 	struct tcp_rack *rack;
23729 	struct timeval tv;
23730 	uint32_t cts;
23731 
23732 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23733 	if (rack->r_ctl.rc_hpts_flags) {
23734 		cts = tcp_get_usecs(&tv);
23735 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
23736 			/*
23737 			 * Pacing timer is up, check if we are ready.
23738 			 */
23739 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
23740 				return (true);
23741 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
23742 			/*
23743 			 * A timer is up, check if we are ready.
23744 			 */
23745 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
23746 				return (true);
23747 		}
23748 	}
23749 	return (false);
23750 }
23751 
23752 static struct tcp_function_block __tcp_rack = {
23753 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
23754 	.tfb_tcp_output = rack_output,
23755 	.tfb_do_queued_segments = ctf_do_queued_segments,
23756 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
23757 	.tfb_tcp_do_segment = rack_do_segment,
23758 	.tfb_tcp_ctloutput = rack_ctloutput,
23759 	.tfb_tcp_fb_init = rack_init,
23760 	.tfb_tcp_fb_fini = rack_fini,
23761 	.tfb_tcp_timer_stop_all = rack_stopall,
23762 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
23763 	.tfb_tcp_handoff_ok = rack_handoff_ok,
23764 	.tfb_tcp_mtu_chg = rack_mtu_change,
23765 	.tfb_pru_options = rack_pru_options,
23766 	.tfb_hwtls_change = rack_hw_tls_change,
23767 	.tfb_chg_query = rack_chg_query,
23768 	.tfb_switch_failed = rack_switch_failed,
23769 	.tfb_early_wake_check = rack_wake_check,
23770 	.tfb_compute_pipe = rack_compute_pipe,
23771 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
23772 };
23773 
23774 /*
23775  * rack_ctloutput() must drop the inpcb lock before performing copyin on
23776  * socket option arguments.  When it re-acquires the lock after the copy, it
23777  * has to revalidate that the connection is still valid for the socket
23778  * option.
23779  */
23780 static int
23781 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23782 {
23783 	struct inpcb *inp = tptoinpcb(tp);
23784 #ifdef INET
23785 	struct ip *ip;
23786 #endif
23787 	struct tcp_rack *rack;
23788 	struct tcp_hybrid_req hybrid;
23789 	uint64_t loptval;
23790 	int32_t error = 0, optval;
23791 
23792 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23793 	if (rack == NULL) {
23794 		INP_WUNLOCK(inp);
23795 		return (EINVAL);
23796 	}
23797 #ifdef INET
23798 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
23799 #endif
23800 
23801 	switch (sopt->sopt_level) {
23802 #ifdef INET6
23803 	case IPPROTO_IPV6:
23804 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
23805 		switch (sopt->sopt_name) {
23806 		case IPV6_USE_MIN_MTU:
23807 			tcp6_use_min_mtu(tp);
23808 			break;
23809 		}
23810 		INP_WUNLOCK(inp);
23811 		return (0);
23812 #endif
23813 #ifdef INET
23814 	case IPPROTO_IP:
23815 		switch (sopt->sopt_name) {
23816 		case IP_TOS:
23817 			/*
23818 			 * The DSCP codepoint has changed, update the fsb.
23819 			 */
23820 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
23821 			break;
23822 		case IP_TTL:
23823 			/*
23824 			 * The TTL has changed, update the fsb.
23825 			 */
23826 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
23827 			break;
23828 		}
23829 		INP_WUNLOCK(inp);
23830 		return (0);
23831 #endif
23832 #ifdef SO_PEERPRIO
23833 	case SOL_SOCKET:
23834 		switch (sopt->sopt_name) {
23835 		case SO_PEERPRIO:			/*  SC-URL:bs */
23836 			/* Already read in and sanity checked in sosetopt(). */
23837 			if (inp->inp_socket) {
23838 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
23839 				rack_client_buffer_level_set(rack);
23840 			}
23841 			break;
23842 		}
23843 		INP_WUNLOCK(inp);
23844 		return (0);
23845 #endif
23846 	case IPPROTO_TCP:
23847 		switch (sopt->sopt_name) {
23848 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
23849 		/*  Pacing related ones */
23850 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
23851 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
23852 		case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
23853 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
23854 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
23855 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
23856 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
23857 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
23858 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
23859 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
23860 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
23861 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
23862 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
23863 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
23864 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
23865 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
23866 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
23867 		case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
23868 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
23869 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
23870 		case TCP_RACK_DGP_IN_REC:		/*  URL:dgpinrec */
23871 			/* End pacing related */
23872 		case TCP_RXT_CLAMP:			/*  URL:rxtclamp */
23873 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
23874 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
23875 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
23876 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
23877 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
23878 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
23879 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
23880 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
23881 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
23882 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
23883 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
23884 		case TCP_RACK_DO_DETECTION:		/*  URL:detect */
23885 		case TCP_NO_PRR:			/*  URL:noprr */
23886 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
23887 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
23888 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
23889 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
23890 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
23891 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
23892 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
23893 		case TCP_RACK_PROFILE:			/*  URL:profile */
23894 		case TCP_HYBRID_PACING:			/*  URL:hybrid */
23895 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
23896 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
23897 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
23898 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
23899 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
23900 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
23901 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
23902 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
23903 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
23904 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
23905 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
23906 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
23907 		case TCP_PACING_DND:			/*  URL:dnd */
23908 			goto process_opt;
23909 			break;
23910 		default:
23911 			/* Filter off all unknown options to the base stack */
23912 			return (tcp_default_ctloutput(tp, sopt));
23913 			break;
23914 		}
23915 
23916 	default:
23917 		INP_WUNLOCK(inp);
23918 		return (0);
23919 	}
23920 process_opt:
23921 	INP_WUNLOCK(inp);
23922 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
23923 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
23924 		/*
23925 		 * We truncate it down to 32 bits for the socket-option trace this
23926 		 * means rates > 34Gbps won't show right, but thats probably ok.
23927 		 */
23928 		optval = (uint32_t)loptval;
23929 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
23930 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
23931 	} else {
23932 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
23933 		/* Save it in 64 bit form too */
23934 		loptval = optval;
23935 	}
23936 	if (error)
23937 		return (error);
23938 	INP_WLOCK(inp);
23939 	if (tp->t_fb != &__tcp_rack) {
23940 		INP_WUNLOCK(inp);
23941 		return (ENOPROTOOPT);
23942 	}
23943 	if (rack->defer_options && (rack->gp_ready == 0) &&
23944 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
23945 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
23946 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
23947 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
23948 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
23949 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
23950 		/* Options are beind deferred */
23951 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
23952 			INP_WUNLOCK(inp);
23953 			return (0);
23954 		} else {
23955 			/* No memory to defer, fail */
23956 			INP_WUNLOCK(inp);
23957 			return (ENOMEM);
23958 		}
23959 	}
23960 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
23961 	INP_WUNLOCK(inp);
23962 	return (error);
23963 }
23964 
23965 static void
23966 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
23967 {
23968 
23969 	INP_WLOCK_ASSERT(tptoinpcb(tp));
23970 	bzero(ti, sizeof(*ti));
23971 
23972 	ti->tcpi_state = tp->t_state;
23973 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
23974 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
23975 	if (tp->t_flags & TF_SACK_PERMIT)
23976 		ti->tcpi_options |= TCPI_OPT_SACK;
23977 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
23978 		ti->tcpi_options |= TCPI_OPT_WSCALE;
23979 		ti->tcpi_snd_wscale = tp->snd_scale;
23980 		ti->tcpi_rcv_wscale = tp->rcv_scale;
23981 	}
23982 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
23983 		ti->tcpi_options |= TCPI_OPT_ECN;
23984 	if (tp->t_flags & TF_FASTOPEN)
23985 		ti->tcpi_options |= TCPI_OPT_TFO;
23986 	/* still kept in ticks is t_rcvtime */
23987 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
23988 	/* Since we hold everything in precise useconds this is easy */
23989 	ti->tcpi_rtt = tp->t_srtt;
23990 	ti->tcpi_rttvar = tp->t_rttvar;
23991 	ti->tcpi_rto = tp->t_rxtcur;
23992 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
23993 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
23994 	/*
23995 	 * FreeBSD-specific extension fields for tcp_info.
23996 	 */
23997 	ti->tcpi_rcv_space = tp->rcv_wnd;
23998 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
23999 	ti->tcpi_snd_wnd = tp->snd_wnd;
24000 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
24001 	ti->tcpi_snd_nxt = tp->snd_nxt;
24002 	ti->tcpi_snd_mss = tp->t_maxseg;
24003 	ti->tcpi_rcv_mss = tp->t_maxseg;
24004 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24005 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24006 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24007 	ti->tcpi_total_tlp = tp->t_sndtlppack;
24008 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24009 #ifdef NETFLIX_STATS
24010 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24011 #endif
24012 #ifdef TCP_OFFLOAD
24013 	if (tp->t_flags & TF_TOE) {
24014 		ti->tcpi_options |= TCPI_OPT_TOE;
24015 		tcp_offload_tcp_info(tp, ti);
24016 	}
24017 #endif
24018 }
24019 
24020 static int
24021 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24022 {
24023 	struct inpcb *inp = tptoinpcb(tp);
24024 	struct tcp_rack *rack;
24025 	int32_t error, optval;
24026 	uint64_t val, loptval;
24027 	struct	tcp_info ti;
24028 	/*
24029 	 * Because all our options are either boolean or an int, we can just
24030 	 * pull everything into optval and then unlock and copy. If we ever
24031 	 * add a option that is not a int, then this will have quite an
24032 	 * impact to this routine.
24033 	 */
24034 	error = 0;
24035 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24036 	if (rack == NULL) {
24037 		INP_WUNLOCK(inp);
24038 		return (EINVAL);
24039 	}
24040 	switch (sopt->sopt_name) {
24041 	case TCP_INFO:
24042 		/* First get the info filled */
24043 		rack_fill_info(tp, &ti);
24044 		/* Fix up the rtt related fields if needed */
24045 		INP_WUNLOCK(inp);
24046 		error = sooptcopyout(sopt, &ti, sizeof ti);
24047 		return (error);
24048 	/*
24049 	 * Beta is the congestion control value for NewReno that influences how
24050 	 * much of a backoff happens when loss is detected. It is normally set
24051 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24052 	 * when you exit recovery.
24053 	 */
24054 	case TCP_RACK_PACING_BETA:
24055 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24056 			error = EINVAL;
24057 		else if (rack->rc_pacing_cc_set == 0)
24058 			optval = rack->r_ctl.rc_saved_beta.beta;
24059 		else {
24060 			/*
24061 			 * Reach out into the CC data and report back what
24062 			 * I have previously set. Yeah it looks hackish but
24063 			 * we don't want to report the saved values.
24064 			 */
24065 			if (tp->t_ccv.cc_data)
24066 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24067 			else
24068 				error = EINVAL;
24069 		}
24070 		break;
24071 		/*
24072 		 * Beta_ecn is the congestion control value for NewReno that influences how
24073 		 * much of a backoff happens when a ECN mark is detected. It is normally set
24074 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24075 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24076 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24077 		 */
24078 
24079 	case TCP_RACK_PACING_BETA_ECN:
24080 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24081 			error = EINVAL;
24082 		else if (rack->rc_pacing_cc_set == 0)
24083 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
24084 		else {
24085 			/*
24086 			 * Reach out into the CC data and report back what
24087 			 * I have previously set. Yeah it looks hackish but
24088 			 * we don't want to report the saved values.
24089 			 */
24090 			if (tp->t_ccv.cc_data)
24091 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24092 			else
24093 				error = EINVAL;
24094 		}
24095 		break;
24096 	case TCP_RACK_DSACK_OPT:
24097 		optval = 0;
24098 		if (rack->rc_rack_tmr_std_based) {
24099 			optval |= 1;
24100 		}
24101 		if (rack->rc_rack_use_dsack) {
24102 			optval |= 2;
24103 		}
24104 		break;
24105  	case TCP_RACK_ENABLE_HYSTART:
24106 	{
24107 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24108 			optval = RACK_HYSTART_ON;
24109 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24110 				optval = RACK_HYSTART_ON_W_SC;
24111 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24112 				optval = RACK_HYSTART_ON_W_SC_C;
24113 		} else {
24114 			optval = RACK_HYSTART_OFF;
24115 		}
24116 	}
24117 	break;
24118 	case TCP_RACK_DGP_IN_REC:
24119 		optval = rack->r_ctl.full_dgp_in_rec;
24120 		break;
24121 	case TCP_RACK_HI_BETA:
24122 		optval = rack->rack_hibeta;
24123 		break;
24124 	case TCP_RXT_CLAMP:
24125 		optval = rack->r_ctl.saved_rxt_clamp_val;
24126 		break;
24127 	case TCP_DEFER_OPTIONS:
24128 		optval = rack->defer_options;
24129 		break;
24130 	case TCP_RACK_MEASURE_CNT:
24131 		optval = rack->r_ctl.req_measurements;
24132 		break;
24133 	case TCP_REC_ABC_VAL:
24134 		optval = rack->r_use_labc_for_rec;
24135 		break;
24136 	case TCP_RACK_ABC_VAL:
24137 		optval = rack->rc_labc;
24138 		break;
24139 	case TCP_HDWR_UP_ONLY:
24140 		optval= rack->r_up_only;
24141 		break;
24142 	case TCP_PACING_RATE_CAP:
24143 		loptval = rack->r_ctl.bw_rate_cap;
24144 		break;
24145 	case TCP_RACK_PROFILE:
24146 		/* You cannot retrieve a profile, its write only */
24147 		error = EINVAL;
24148 		break;
24149 	case TCP_HYBRID_PACING:
24150 		/* You cannot retrieve hybrid pacing information, its write only */
24151 		error = EINVAL;
24152 		break;
24153 	case TCP_USE_CMP_ACKS:
24154 		optval = rack->r_use_cmp_ack;
24155 		break;
24156 	case TCP_RACK_PACE_TO_FILL:
24157 		optval = rack->rc_pace_to_cwnd;
24158 		if (optval && rack->r_fill_less_agg)
24159 			optval++;
24160 		break;
24161 	case TCP_RACK_NO_PUSH_AT_MAX:
24162 		optval = rack->r_ctl.rc_no_push_at_mrtt;
24163 		break;
24164 	case TCP_SHARED_CWND_ENABLE:
24165 		optval = rack->rack_enable_scwnd;
24166 		break;
24167 	case TCP_RACK_NONRXT_CFG_RATE:
24168 		optval = rack->rack_rec_nonrxt_use_cr;
24169 		break;
24170 	case TCP_NO_PRR:
24171 		if (rack->rack_no_prr  == 1)
24172 			optval = 1;
24173 		else if (rack->no_prr_addback == 1)
24174 			optval = 2;
24175 		else
24176 			optval = 0;
24177 		break;
24178 	case TCP_RACK_DO_DETECTION:
24179 		optval = rack->do_detection;
24180 		break;
24181 	case TCP_RACK_MBUF_QUEUE:
24182 		/* Now do we use the LRO mbuf-queue feature */
24183 		optval = rack->r_mbuf_queue;
24184 		break;
24185 	case TCP_TIMELY_DYN_ADJ:
24186 		optval = rack->rc_gp_dyn_mul;
24187 		break;
24188 	case TCP_BBR_IWINTSO:
24189 		optval = rack->rc_init_win;
24190 		break;
24191 	case TCP_RACK_TLP_REDUCE:
24192 		/* RACK TLP cwnd reduction (bool) */
24193 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24194 		break;
24195 	case TCP_BBR_RACK_INIT_RATE:
24196 		val = rack->r_ctl.init_rate;
24197 		/* convert to kbits per sec */
24198 		val *= 8;
24199 		val /= 1000;
24200 		optval = (uint32_t)val;
24201 		break;
24202 	case TCP_RACK_FORCE_MSEG:
24203 		optval = rack->rc_force_max_seg;
24204 		break;
24205 	case TCP_RACK_PACE_MIN_SEG:
24206 		optval = rack->r_ctl.rc_user_set_min_segs;
24207 		break;
24208 	case TCP_RACK_PACE_MAX_SEG:
24209 		/* Max segments in a pace */
24210 		optval = rack->rc_user_set_max_segs;
24211 		break;
24212 	case TCP_RACK_PACE_ALWAYS:
24213 		/* Use the always pace method */
24214 		optval = rack->rc_always_pace;
24215 		break;
24216 	case TCP_RACK_PRR_SENDALOT:
24217 		/* Allow PRR to send more than one seg */
24218 		optval = rack->r_ctl.rc_prr_sendalot;
24219 		break;
24220 	case TCP_RACK_MIN_TO:
24221 		/* Minimum time between rack t-o's in ms */
24222 		optval = rack->r_ctl.rc_min_to;
24223 		break;
24224 	case TCP_RACK_SPLIT_LIMIT:
24225 		optval = rack->r_ctl.rc_split_limit;
24226 		break;
24227 	case TCP_RACK_EARLY_SEG:
24228 		/* If early recovery max segments */
24229 		optval = rack->r_ctl.rc_early_recovery_segs;
24230 		break;
24231 	case TCP_RACK_REORD_THRESH:
24232 		/* RACK reorder threshold (shift amount) */
24233 		optval = rack->r_ctl.rc_reorder_shift;
24234 		break;
24235 	case TCP_RACK_REORD_FADE:
24236 		/* Does reordering fade after ms time */
24237 		optval = rack->r_ctl.rc_reorder_fade;
24238 		break;
24239 	case TCP_BBR_USE_RACK_RR:
24240 		/* Do we use the rack cheat for rxt */
24241 		optval = rack->use_rack_rr;
24242 		break;
24243 	case TCP_RACK_RR_CONF:
24244 		optval = rack->r_rr_config;
24245 		break;
24246 	case TCP_HDWR_RATE_CAP:
24247 		optval = rack->r_rack_hw_rate_caps;
24248 		break;
24249 	case TCP_BBR_HDWR_PACE:
24250 		optval = rack->rack_hdw_pace_ena;
24251 		break;
24252 	case TCP_RACK_TLP_THRESH:
24253 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24254 		optval = rack->r_ctl.rc_tlp_threshold;
24255 		break;
24256 	case TCP_RACK_PKT_DELAY:
24257 		/* RACK added ms i.e. rack-rtt + reord + N */
24258 		optval = rack->r_ctl.rc_pkt_delay;
24259 		break;
24260 	case TCP_RACK_TLP_USE:
24261 		optval = rack->rack_tlp_threshold_use;
24262 		break;
24263 	case TCP_PACING_DND:
24264 		optval = rack->rc_pace_dnd;
24265 		break;
24266 	case TCP_RACK_PACE_RATE_CA:
24267 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24268 		break;
24269 	case TCP_RACK_PACE_RATE_SS:
24270 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24271 		break;
24272 	case TCP_RACK_PACE_RATE_REC:
24273 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24274 		break;
24275 	case TCP_RACK_GP_INCREASE_SS:
24276 		optval = rack->r_ctl.rack_per_of_gp_ca;
24277 		break;
24278 	case TCP_RACK_GP_INCREASE_CA:
24279 		optval = rack->r_ctl.rack_per_of_gp_ss;
24280 		break;
24281 	case TCP_RACK_PACING_DIVISOR:
24282 		optval = rack->r_ctl.pace_len_divisor;
24283 		break;
24284 	case TCP_BBR_RACK_RTT_USE:
24285 		optval = rack->r_ctl.rc_rate_sample_method;
24286 		break;
24287 	case TCP_DELACK:
24288 		optval = tp->t_delayed_ack;
24289 		break;
24290 	case TCP_DATA_AFTER_CLOSE:
24291 		optval = rack->rc_allow_data_af_clo;
24292 		break;
24293 	case TCP_SHARED_CWND_TIME_LIMIT:
24294 		optval = rack->r_limit_scw;
24295 		break;
24296 	case TCP_RACK_TIMER_SLOP:
24297 		optval = rack->r_ctl.timer_slop;
24298 		break;
24299 	default:
24300 		return (tcp_default_ctloutput(tp, sopt));
24301 		break;
24302 	}
24303 	INP_WUNLOCK(inp);
24304 	if (error == 0) {
24305 		if (TCP_PACING_RATE_CAP)
24306 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
24307 		else
24308 			error = sooptcopyout(sopt, &optval, sizeof optval);
24309 	}
24310 	return (error);
24311 }
24312 
24313 static int
24314 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24315 {
24316 	if (sopt->sopt_dir == SOPT_SET) {
24317 		return (rack_set_sockopt(tp, sopt));
24318 	} else if (sopt->sopt_dir == SOPT_GET) {
24319 		return (rack_get_sockopt(tp, sopt));
24320 	} else {
24321 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24322 	}
24323 }
24324 
24325 static const char *rack_stack_names[] = {
24326 	__XSTRING(STACKNAME),
24327 #ifdef STACKALIAS
24328 	__XSTRING(STACKALIAS),
24329 #endif
24330 };
24331 
24332 static int
24333 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24334 {
24335 	memset(mem, 0, size);
24336 	return (0);
24337 }
24338 
24339 static void
24340 rack_dtor(void *mem, int32_t size, void *arg)
24341 {
24342 
24343 }
24344 
24345 static bool rack_mod_inited = false;
24346 
24347 static int
24348 tcp_addrack(module_t mod, int32_t type, void *data)
24349 {
24350 	int32_t err = 0;
24351 	int num_stacks;
24352 
24353 	switch (type) {
24354 	case MOD_LOAD:
24355 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24356 		    sizeof(struct rack_sendmap),
24357 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24358 
24359 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24360 		    sizeof(struct tcp_rack),
24361 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24362 
24363 		sysctl_ctx_init(&rack_sysctl_ctx);
24364 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24365 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24366 		    OID_AUTO,
24367 #ifdef STACKALIAS
24368 		    __XSTRING(STACKALIAS),
24369 #else
24370 		    __XSTRING(STACKNAME),
24371 #endif
24372 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24373 		    "");
24374 		if (rack_sysctl_root == NULL) {
24375 			printf("Failed to add sysctl node\n");
24376 			err = EFAULT;
24377 			goto free_uma;
24378 		}
24379 		rack_init_sysctls();
24380 		num_stacks = nitems(rack_stack_names);
24381 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24382 		    rack_stack_names, &num_stacks);
24383 		if (err) {
24384 			printf("Failed to register %s stack name for "
24385 			    "%s module\n", rack_stack_names[num_stacks],
24386 			    __XSTRING(MODNAME));
24387 			sysctl_ctx_free(&rack_sysctl_ctx);
24388 free_uma:
24389 			uma_zdestroy(rack_zone);
24390 			uma_zdestroy(rack_pcb_zone);
24391 			rack_counter_destroy();
24392 			printf("Failed to register rack module -- err:%d\n", err);
24393 			return (err);
24394 		}
24395 		tcp_lro_reg_mbufq();
24396 		rack_mod_inited = true;
24397 		break;
24398 	case MOD_QUIESCE:
24399 		err = deregister_tcp_functions(&__tcp_rack, true, false);
24400 		break;
24401 	case MOD_UNLOAD:
24402 		err = deregister_tcp_functions(&__tcp_rack, false, true);
24403 		if (err == EBUSY)
24404 			break;
24405 		if (rack_mod_inited) {
24406 			uma_zdestroy(rack_zone);
24407 			uma_zdestroy(rack_pcb_zone);
24408 			sysctl_ctx_free(&rack_sysctl_ctx);
24409 			rack_counter_destroy();
24410 			rack_mod_inited = false;
24411 		}
24412 		tcp_lro_dereg_mbufq();
24413 		err = 0;
24414 		break;
24415 	default:
24416 		return (EOPNOTSUPP);
24417 	}
24418 	return (err);
24419 }
24420 
24421 static moduledata_t tcp_rack = {
24422 	.name = __XSTRING(MODNAME),
24423 	.evhand = tcp_addrack,
24424 	.priv = 0
24425 };
24426 
24427 MODULE_VERSION(MODNAME, 1);
24428 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24429 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24430 
24431 #endif /* #if !defined(INET) && !defined(INET6) */
24432