xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision e0c4386e) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30 #include "opt_ipsec.h"
31 #include "opt_ratelimit.h"
32 #include "opt_kern_tls.h"
33 #if defined(INET) || defined(INET6)
34 #include <sys/param.h>
35 #include <sys/arb.h>
36 #include <sys/module.h>
37 #include <sys/kernel.h>
38 #ifdef TCP_HHOOK
39 #include <sys/hhook.h>
40 #endif
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/mbuf.h>
46 #include <sys/proc.h>		/* for proc0 declaration */
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 #ifdef STATS
52 #include <sys/qmath.h>
53 #include <sys/tree.h>
54 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
55 #else
56 #include <sys/tree.h>
57 #endif
58 #include <sys/refcount.h>
59 #include <sys/queue.h>
60 #include <sys/tim_filter.h>
61 #include <sys/smp.h>
62 #include <sys/kthread.h>
63 #include <sys/kern_prefetch.h>
64 #include <sys/protosw.h>
65 #ifdef TCP_ACCOUNTING
66 #include <sys/sched.h>
67 #include <machine/cpu.h>
68 #endif
69 #include <vm/uma.h>
70 
71 #include <net/route.h>
72 #include <net/route/nhop.h>
73 #include <net/vnet.h>
74 
75 #define TCPSTATES		/* for logging */
76 
77 #include <netinet/in.h>
78 #include <netinet/in_kdtrace.h>
79 #include <netinet/in_pcb.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
82 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
83 #include <netinet/ip_var.h>
84 #include <netinet/ip6.h>
85 #include <netinet6/in6_pcb.h>
86 #include <netinet6/ip6_var.h>
87 #include <netinet/tcp.h>
88 #define	TCPOUTFLAGS
89 #include <netinet/tcp_fsm.h>
90 #include <netinet/tcp_seq.h>
91 #include <netinet/tcp_timer.h>
92 #include <netinet/tcp_var.h>
93 #include <netinet/tcp_log_buf.h>
94 #include <netinet/tcp_syncache.h>
95 #include <netinet/tcp_hpts.h>
96 #include <netinet/tcp_ratelimit.h>
97 #include <netinet/tcp_accounting.h>
98 #include <netinet/tcpip.h>
99 #include <netinet/cc/cc.h>
100 #include <netinet/cc/cc_newreno.h>
101 #include <netinet/tcp_fastopen.h>
102 #include <netinet/tcp_lro.h>
103 #ifdef NETFLIX_SHARED_CWND
104 #include <netinet/tcp_shared_cwnd.h>
105 #endif
106 #ifdef TCP_OFFLOAD
107 #include <netinet/tcp_offload.h>
108 #endif
109 #ifdef INET6
110 #include <netinet6/tcp6_var.h>
111 #endif
112 #include <netinet/tcp_ecn.h>
113 
114 #include <netipsec/ipsec_support.h>
115 
116 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
117 #include <netipsec/ipsec.h>
118 #include <netipsec/ipsec6.h>
119 #endif				/* IPSEC */
120 
121 #include <netinet/udp.h>
122 #include <netinet/udp_var.h>
123 #include <machine/in_cksum.h>
124 
125 #ifdef MAC
126 #include <security/mac/mac_framework.h>
127 #endif
128 #include "sack_filter.h"
129 #include "tcp_rack.h"
130 #include "tailq_hash.h"
131 #include "rack_bbr_common.h"
132 
133 uma_zone_t rack_zone;
134 uma_zone_t rack_pcb_zone;
135 
136 #ifndef TICKS2SBT
137 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
138 #endif
139 
140 VNET_DECLARE(uint32_t, newreno_beta);
141 VNET_DECLARE(uint32_t, newreno_beta_ecn);
142 #define V_newreno_beta VNET(newreno_beta)
143 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
144 
145 
146 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
147 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
148 
149 struct sysctl_ctx_list rack_sysctl_ctx;
150 struct sysctl_oid *rack_sysctl_root;
151 
152 #define CUM_ACKED 1
153 #define SACKED 2
154 
155 /*
156  * The RACK module incorporates a number of
157  * TCP ideas that have been put out into the IETF
158  * over the last few years:
159  * - Matt Mathis's Rate Halving which slowly drops
160  *    the congestion window so that the ack clock can
161  *    be maintained during a recovery.
162  * - Yuchung Cheng's RACK TCP (for which its named) that
163  *    will stop us using the number of dup acks and instead
164  *    use time as the gage of when we retransmit.
165  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
166  *    of Dukkipati et.al.
167  * RACK depends on SACK, so if an endpoint arrives that
168  * cannot do SACK the state machine below will shuttle the
169  * connection back to using the "default" TCP stack that is
170  * in FreeBSD.
171  *
172  * To implement RACK the original TCP stack was first decomposed
173  * into a functional state machine with individual states
174  * for each of the possible TCP connection states. The do_segment
175  * functions role in life is to mandate the connection supports SACK
176  * initially and then assure that the RACK state matches the conenction
177  * state before calling the states do_segment function. Each
178  * state is simplified due to the fact that the original do_segment
179  * has been decomposed and we *know* what state we are in (no
180  * switches on the state) and all tests for SACK are gone. This
181  * greatly simplifies what each state does.
182  *
183  * TCP output is also over-written with a new version since it
184  * must maintain the new rack scoreboard.
185  *
186  */
187 static int32_t rack_tlp_thresh = 1;
188 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
189 static int32_t rack_tlp_use_greater = 1;
190 static int32_t rack_reorder_thresh = 2;
191 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
192 						 * - 60 seconds */
193 static uint32_t rack_clamp_ss_upper = 110;
194 static uint32_t rack_clamp_ca_upper = 105;
195 static uint32_t rack_rxt_min_rnds = 10;	/* Min rounds if drastic rxt clamp is in place */
196 static uint32_t rack_unclamp_round_thresh = 100;	/* number of perfect rounds before we unclamp */
197 static uint32_t rack_unclamp_rxt_thresh = 5;	/* .5%  and under */
198 static uint64_t rack_rxt_clamp_thresh = 0;	/* Do we do the rxt clamp thing */
199 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
200 static int32_t rack_rxt_controls = 0;
201 static int32_t rack_fill_cw_state = 0;
202 static uint8_t rack_req_measurements = 1;
203 /* Attack threshold detections */
204 static uint32_t rack_highest_sack_thresh_seen = 0;
205 static uint32_t rack_highest_move_thresh_seen = 0;
206 static uint32_t rack_merge_out_sacks_on_attack = 0;
207 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
208 static int32_t rack_hw_pace_extra_slots = 0;	/* 2 extra MSS time betweens */
209 static int32_t rack_hw_rate_caps = 0; /* 1; */
210 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
211 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
212 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
213 static int32_t rack_hw_up_only = 0;
214 static int32_t rack_stats_gets_ms_rtt = 1;
215 static int32_t rack_prr_addbackmax = 2;
216 static int32_t rack_do_hystart = 0;
217 static int32_t rack_apply_rtt_with_reduced_conf = 0;
218 static int32_t rack_hibeta_setting = 0;
219 static int32_t rack_default_pacing_divisor = 250;
220 static int32_t rack_uses_full_dgp_in_rec = 1;
221 static uint16_t rack_pacing_min_seg = 0;
222 
223 
224 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
225 static int32_t rack_pkt_delay = 1000;
226 static int32_t rack_send_a_lot_in_prr = 1;
227 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
228 static int32_t rack_verbose_logging = 0;
229 static int32_t rack_ignore_data_after_close = 1;
230 static int32_t rack_enable_shared_cwnd = 1;
231 static int32_t rack_use_cmp_acks = 1;
232 static int32_t rack_use_fsb = 1;
233 static int32_t rack_use_rfo = 1;
234 static int32_t rack_use_rsm_rfo = 1;
235 static int32_t rack_max_abc_post_recovery = 2;
236 static int32_t rack_client_low_buf = 0;
237 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 */
238 static int32_t rack_bw_multipler = 2;		/* Limit on fill cw's jump up to be this x gp_est */
239 #ifdef TCP_ACCOUNTING
240 static int32_t rack_tcp_accounting = 0;
241 #endif
242 static int32_t rack_limits_scwnd = 1;
243 static int32_t rack_enable_mqueue_for_nonpaced = 0;
244 static int32_t rack_hybrid_allow_set_maxseg = 0;
245 static int32_t rack_disable_prr = 0;
246 static int32_t use_rack_rr = 1;
247 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
248 static int32_t rack_persist_min = 250000;	/* 250usec */
249 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
250 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
251 static int32_t rack_default_init_window = 0;	/* Use system default */
252 static int32_t rack_limit_time_with_srtt = 0;
253 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
254 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
255 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
256 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
257 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
258 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
259 static int32_t rack_full_buffer_discount = 10;
260 /*
261  * Currently regular tcp has a rto_min of 30ms
262  * the backoff goes 12 times so that ends up
263  * being a total of 122.850 seconds before a
264  * connection is killed.
265  */
266 static uint32_t rack_def_data_window = 20;
267 static uint32_t rack_goal_bdp = 2;
268 static uint32_t rack_min_srtts = 1;
269 static uint32_t rack_min_measure_usec = 0;
270 static int32_t rack_tlp_min = 10000;	/* 10ms */
271 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
272 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
273 static const int32_t rack_free_cache = 2;
274 static int32_t rack_hptsi_segments = 40;
275 static int32_t rack_rate_sample_method = USE_RTT_LOW;
276 static int32_t rack_pace_every_seg = 0;
277 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
278 static int32_t rack_slot_reduction = 4;
279 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
280 static int32_t rack_cwnd_block_ends_measure = 0;
281 static int32_t rack_rwnd_block_ends_measure = 0;
282 static int32_t rack_def_profile = 0;
283 
284 static int32_t rack_lower_cwnd_at_tlp = 0;
285 static int32_t rack_limited_retran = 0;
286 static int32_t rack_always_send_oldest = 0;
287 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
288 
289 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
290 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
291 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
292 
293 /* Probertt */
294 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
295 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
296 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
297 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
298 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
299 
300 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
301 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
302 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
303 static uint32_t rack_probertt_use_min_rtt_exit = 0;
304 static uint32_t rack_probe_rtt_sets_cwnd = 0;
305 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
306 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
307 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
308 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
309 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
310 static uint32_t rack_probertt_filter_life = 10000000;
311 static uint32_t rack_probertt_lower_within = 10;
312 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
313 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
314 static int32_t rack_probertt_clear_is = 1;
315 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
316 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
317 
318 /* Part of pacing */
319 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
320 
321 /* Timely information:
322  *
323  * Here we have various control parameters on how
324  * timely may change the multiplier. rack_gain_p5_ub
325  * is associated with timely but not directly influencing
326  * the rate decision like the other variables. It controls
327  * the way fill-cw interacts with timely and caps how much
328  * timely can boost the fill-cw b/w.
329  *
330  * The other values are various boost/shrink numbers as well
331  * as potential caps when adjustments are made to the timely
332  * gain (returned by rack_get_output_gain(). Remember too that
333  * the gain returned can be overriden by other factors such as
334  * probeRTT as well as fixed-rate-pacing.
335  */
336 static int32_t rack_gain_p5_ub = 250;
337 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
338 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
339 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
340 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
341 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
342 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
343 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
344 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
345 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
346 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
347 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
348 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
349 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
350 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
351 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
352 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
353 static int32_t rack_use_max_for_nobackoff = 0;
354 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
355 static int32_t rack_timely_no_stopping = 0;
356 static int32_t rack_down_raise_thresh = 100;
357 static int32_t rack_req_segs = 1;
358 static uint64_t rack_bw_rate_cap = 0;
359 
360 
361 /* Rack specific counters */
362 counter_u64_t rack_saw_enobuf;
363 counter_u64_t rack_saw_enobuf_hw;
364 counter_u64_t rack_saw_enetunreach;
365 counter_u64_t rack_persists_sends;
366 counter_u64_t rack_persists_acks;
367 counter_u64_t rack_persists_loss;
368 counter_u64_t rack_persists_lost_ends;
369 counter_u64_t rack_total_bytes;
370 #ifdef INVARIANTS
371 counter_u64_t rack_adjust_map_bw;
372 #endif
373 /* Tail loss probe counters */
374 counter_u64_t rack_tlp_tot;
375 counter_u64_t rack_tlp_newdata;
376 counter_u64_t rack_tlp_retran;
377 counter_u64_t rack_tlp_retran_bytes;
378 counter_u64_t rack_to_tot;
379 counter_u64_t rack_hot_alloc;
380 counter_u64_t rack_to_alloc;
381 counter_u64_t rack_to_alloc_hard;
382 counter_u64_t rack_to_alloc_emerg;
383 counter_u64_t rack_to_alloc_limited;
384 counter_u64_t rack_alloc_limited_conns;
385 counter_u64_t rack_split_limited;
386 counter_u64_t rack_rxt_clamps_cwnd;
387 counter_u64_t rack_rxt_clamps_cwnd_uniq;
388 
389 counter_u64_t rack_multi_single_eq;
390 counter_u64_t rack_proc_non_comp_ack;
391 
392 counter_u64_t rack_fto_send;
393 counter_u64_t rack_fto_rsm_send;
394 counter_u64_t rack_nfto_resend;
395 counter_u64_t rack_non_fto_send;
396 counter_u64_t rack_extended_rfo;
397 
398 counter_u64_t rack_sack_proc_all;
399 counter_u64_t rack_sack_proc_short;
400 counter_u64_t rack_sack_proc_restart;
401 counter_u64_t rack_sack_attacks_detected;
402 counter_u64_t rack_sack_attacks_reversed;
403 counter_u64_t rack_sack_attacks_suspect;
404 counter_u64_t rack_sack_used_next_merge;
405 counter_u64_t rack_sack_splits;
406 counter_u64_t rack_sack_used_prev_merge;
407 counter_u64_t rack_sack_skipped_acked;
408 counter_u64_t rack_ack_total;
409 counter_u64_t rack_express_sack;
410 counter_u64_t rack_sack_total;
411 counter_u64_t rack_move_none;
412 counter_u64_t rack_move_some;
413 
414 counter_u64_t rack_input_idle_reduces;
415 counter_u64_t rack_collapsed_win;
416 counter_u64_t rack_collapsed_win_seen;
417 counter_u64_t rack_collapsed_win_rxt;
418 counter_u64_t rack_collapsed_win_rxt_bytes;
419 counter_u64_t rack_try_scwnd;
420 counter_u64_t rack_hw_pace_init_fail;
421 counter_u64_t rack_hw_pace_lost;
422 
423 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
424 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
425 
426 
427 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
428 
429 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
430 	(tv) = (value) + slop;	 \
431 	if ((u_long)(tv) < (u_long)(tvmin)) \
432 		(tv) = (tvmin); \
433 	if ((u_long)(tv) > (u_long)(tvmax)) \
434 		(tv) = (tvmax); \
435 } while (0)
436 
437 static void
438 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
439 
440 static int
441 rack_process_ack(struct mbuf *m, struct tcphdr *th,
442     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
443     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
444 static int
445 rack_process_data(struct mbuf *m, struct tcphdr *th,
446     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
447     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
448 static void
449 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
450    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
451 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
452 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
453     uint8_t limit_type);
454 static struct rack_sendmap *
455 rack_check_recovery_mode(struct tcpcb *tp,
456     uint32_t tsused);
457 static void
458 rack_cong_signal(struct tcpcb *tp,
459 		 uint32_t type, uint32_t ack, int );
460 static void rack_counter_destroy(void);
461 static int
462 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
463 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
464 static void
465 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
466 static void
467 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
468     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
469 static void rack_dtor(void *mem, int32_t size, void *arg);
470 static void
471 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
472     uint32_t flex1, uint32_t flex2,
473     uint32_t flex3, uint32_t flex4,
474     uint32_t flex5, uint32_t flex6,
475     uint16_t flex7, uint8_t mod);
476 
477 static void
478 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
479    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
480    struct rack_sendmap *rsm, uint8_t quality);
481 static struct rack_sendmap *
482 rack_find_high_nonack(struct tcp_rack *rack,
483     struct rack_sendmap *rsm);
484 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
485 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
486 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
487 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
488 static void
489 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
490 			    tcp_seq th_ack, int line, uint8_t quality);
491 static void
492 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
493 
494 static uint32_t
495 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
496 static int32_t rack_handoff_ok(struct tcpcb *tp);
497 static int32_t rack_init(struct tcpcb *tp, void **ptr);
498 static void rack_init_sysctls(void);
499 
500 static void
501 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
502     struct tcphdr *th, int entered_rec, int dup_ack_struck,
503     int *dsack_seen, int *sacks_seen);
504 static void
505 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
506     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
507     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
508 
509 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
510 
511 static void
512 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
513     struct rack_sendmap *rsm);
514 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
515 static int32_t rack_output(struct tcpcb *tp);
516 
517 static uint32_t
518 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
519     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
520     uint32_t cts, int *no_extra, int *moved_two, uint32_t segsiz);
521 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
522 static void rack_remxt_tmr(struct tcpcb *tp);
523 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
524 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
525 static int32_t rack_stopall(struct tcpcb *tp);
526 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
527 static uint32_t
528 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
529     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag, int segsiz);
530 static void
531 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
532     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz);
533 static int
534 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
535     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
536 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
537 static int
538 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
539     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
540     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
541 static int
542 rack_do_closing(struct mbuf *m, struct tcphdr *th,
543     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
544     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
545 static int
546 rack_do_established(struct mbuf *m, struct tcphdr *th,
547     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
548     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
549 static int
550 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
551     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
552     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
553 static int
554 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
555     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
556     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
557 static int
558 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
559     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
560     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
561 static int
562 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
563     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
564     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
565 static int
566 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
567     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
568     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
569 static int
570 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
571     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
572     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
573 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
574 struct rack_sendmap *
575 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
576     uint32_t tsused);
577 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
578     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
579 static void
580      tcp_rack_partialack(struct tcpcb *tp);
581 static int
582 rack_set_profile(struct tcp_rack *rack, int prof);
583 static void
584 rack_apply_deferred_options(struct tcp_rack *rack);
585 
586 int32_t rack_clear_counter=0;
587 
588 static uint64_t
589 rack_get_lt_bw(struct tcp_rack *rack)
590 {
591 	struct timeval tv;
592 	uint64_t tim, bytes;
593 
594 	tim = rack->r_ctl.lt_bw_time;
595 	bytes = rack->r_ctl.lt_bw_bytes;
596 	if (rack->lt_bw_up) {
597 		/* Include all the current bytes too */
598 		microuptime(&tv);
599 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
600 		tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
601 	}
602 	if ((bytes != 0) && (tim != 0))
603 		return ((bytes * (uint64_t)1000000) / tim);
604 	else
605 		return (0);
606 }
607 
608 static void
609 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
610 {
611 	struct sockopt sopt;
612 	struct cc_newreno_opts opt;
613 	struct newreno old;
614 	struct tcpcb *tp;
615 	int error, failed = 0;
616 
617 	tp = rack->rc_tp;
618 	if (tp->t_cc == NULL) {
619 		/* Tcb is leaving */
620 		return;
621 	}
622 	rack->rc_pacing_cc_set = 1;
623 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
624 		/* Not new-reno we can't play games with beta! */
625 		failed = 1;
626 		goto out;
627 
628 	}
629 	if (CC_ALGO(tp)->ctl_output == NULL)  {
630 		/* Huh, not using new-reno so no swaps.? */
631 		failed = 2;
632 		goto out;
633 	}
634 	/* Get the current values out */
635 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
636 	sopt.sopt_dir = SOPT_GET;
637 	opt.name = CC_NEWRENO_BETA;
638 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
639 	if (error)  {
640 		failed = 3;
641 		goto out;
642 	}
643 	old.beta = opt.val;
644 	opt.name = CC_NEWRENO_BETA_ECN;
645 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
646 	if (error)  {
647 		failed = 4;
648 		goto out;
649 	}
650 	old.beta_ecn = opt.val;
651 
652 	/* Now lets set in the values we have stored */
653 	sopt.sopt_dir = SOPT_SET;
654 	opt.name = CC_NEWRENO_BETA;
655 	opt.val = rack->r_ctl.rc_saved_beta.beta;
656 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
657 	if (error)  {
658 		failed = 5;
659 		goto out;
660 	}
661 	opt.name = CC_NEWRENO_BETA_ECN;
662 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
663 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
664 	if (error) {
665 		failed = 6;
666 		goto out;
667 	}
668 	/* Save off the values for restoral */
669 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
670 out:
671 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
672 		union tcp_log_stackspecific log;
673 		struct timeval tv;
674 		struct newreno *ptr;
675 
676 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
677 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
678 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
679 		log.u_bbr.flex1 = ptr->beta;
680 		log.u_bbr.flex2 = ptr->beta_ecn;
681 		log.u_bbr.flex3 = ptr->newreno_flags;
682 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
683 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
684 		log.u_bbr.flex6 = failed;
685 		log.u_bbr.flex7 = rack->gp_ready;
686 		log.u_bbr.flex7 <<= 1;
687 		log.u_bbr.flex7 |= rack->use_fixed_rate;
688 		log.u_bbr.flex7 <<= 1;
689 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
690 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
691 		log.u_bbr.flex8 = flex8;
692 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
693 			       0, &log, false, NULL, NULL, 0, &tv);
694 	}
695 }
696 
697 static void
698 rack_set_cc_pacing(struct tcp_rack *rack)
699 {
700 	if (rack->rc_pacing_cc_set)
701 		return;
702 	/*
703 	 * Use the swap utility placing in 3 for flex8 to id a
704 	 * set of a new set of values.
705 	 */
706 	rack->rc_pacing_cc_set = 1;
707 	rack_swap_beta_values(rack, 3);
708 }
709 
710 static void
711 rack_undo_cc_pacing(struct tcp_rack *rack)
712 {
713 	if (rack->rc_pacing_cc_set == 0)
714 		return;
715 	/*
716 	 * Use the swap utility placing in 4 for flex8 to id a
717 	 * restoral of the old values.
718 	 */
719 	rack->rc_pacing_cc_set = 0;
720 	rack_swap_beta_values(rack, 4);
721 }
722 
723 static void
724 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
725 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
726 {
727 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
728 		union tcp_log_stackspecific log;
729 		struct timeval tv;
730 
731 		memset(&log, 0, sizeof(log));
732 		log.u_bbr.flex1 = seq_end;
733 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
734 		log.u_bbr.flex3 = ack_end_t;
735 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
736 		log.u_bbr.flex5 = send_end_t;
737 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
738 		log.u_bbr.flex7 = mode;
739 		log.u_bbr.flex8 = 69;
740 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
741 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
742 		log.u_bbr.pkts_out = line;
743 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
744 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
745 		if (rsm != NULL) {
746 			log.u_bbr.applimited = rsm->r_start;
747 			log.u_bbr.delivered = rsm->r_end;
748 			log.u_bbr.epoch = rsm->r_flags;
749 		}
750 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
751 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
752 		    &rack->rc_inp->inp_socket->so_rcv,
753 		    &rack->rc_inp->inp_socket->so_snd,
754 		    BBR_LOG_HPTSI_CALC, 0,
755 		    0, &log, false, &tv);
756 	}
757 }
758 
759 static int
760 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
761 {
762 	uint32_t stat;
763 	int32_t error;
764 
765 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
766 	if (error || req->newptr == NULL)
767 		return error;
768 
769 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
770 	if (error)
771 		return (error);
772 	if (stat == 1) {
773 #ifdef INVARIANTS
774 		printf("Clearing RACK counters\n");
775 #endif
776 		counter_u64_zero(rack_tlp_tot);
777 		counter_u64_zero(rack_tlp_newdata);
778 		counter_u64_zero(rack_tlp_retran);
779 		counter_u64_zero(rack_tlp_retran_bytes);
780 		counter_u64_zero(rack_to_tot);
781 		counter_u64_zero(rack_saw_enobuf);
782 		counter_u64_zero(rack_saw_enobuf_hw);
783 		counter_u64_zero(rack_saw_enetunreach);
784 		counter_u64_zero(rack_persists_sends);
785 		counter_u64_zero(rack_total_bytes);
786 		counter_u64_zero(rack_persists_acks);
787 		counter_u64_zero(rack_persists_loss);
788 		counter_u64_zero(rack_persists_lost_ends);
789 #ifdef INVARIANTS
790 		counter_u64_zero(rack_adjust_map_bw);
791 #endif
792 		counter_u64_zero(rack_to_alloc_hard);
793 		counter_u64_zero(rack_to_alloc_emerg);
794 		counter_u64_zero(rack_sack_proc_all);
795 		counter_u64_zero(rack_fto_send);
796 		counter_u64_zero(rack_fto_rsm_send);
797 		counter_u64_zero(rack_extended_rfo);
798 		counter_u64_zero(rack_hw_pace_init_fail);
799 		counter_u64_zero(rack_hw_pace_lost);
800 		counter_u64_zero(rack_non_fto_send);
801 		counter_u64_zero(rack_nfto_resend);
802 		counter_u64_zero(rack_sack_proc_short);
803 		counter_u64_zero(rack_sack_proc_restart);
804 		counter_u64_zero(rack_to_alloc);
805 		counter_u64_zero(rack_to_alloc_limited);
806 		counter_u64_zero(rack_alloc_limited_conns);
807 		counter_u64_zero(rack_split_limited);
808 		counter_u64_zero(rack_rxt_clamps_cwnd);
809 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
810 		counter_u64_zero(rack_multi_single_eq);
811 		counter_u64_zero(rack_proc_non_comp_ack);
812 		counter_u64_zero(rack_sack_attacks_detected);
813 		counter_u64_zero(rack_sack_attacks_reversed);
814 		counter_u64_zero(rack_sack_attacks_suspect);
815 		counter_u64_zero(rack_sack_used_next_merge);
816 		counter_u64_zero(rack_sack_used_prev_merge);
817 		counter_u64_zero(rack_sack_splits);
818 		counter_u64_zero(rack_sack_skipped_acked);
819 		counter_u64_zero(rack_ack_total);
820 		counter_u64_zero(rack_express_sack);
821 		counter_u64_zero(rack_sack_total);
822 		counter_u64_zero(rack_move_none);
823 		counter_u64_zero(rack_move_some);
824 		counter_u64_zero(rack_try_scwnd);
825 		counter_u64_zero(rack_collapsed_win);
826 		counter_u64_zero(rack_collapsed_win_rxt);
827 		counter_u64_zero(rack_collapsed_win_seen);
828 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
829 	} else if (stat == 2) {
830 #ifdef INVARIANTS
831 		printf("Clearing RACK option array\n");
832 #endif
833 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
834 	} else if (stat == 3) {
835 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
836 	} else if (stat == 4) {
837 #ifdef INVARIANTS
838 		printf("Clearing RACK out size array\n");
839 #endif
840 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
841 	}
842 	rack_clear_counter = 0;
843 	return (0);
844 }
845 
846 static void
847 rack_init_sysctls(void)
848 {
849 	struct sysctl_oid *rack_counters;
850 	struct sysctl_oid *rack_attack;
851 	struct sysctl_oid *rack_pacing;
852 	struct sysctl_oid *rack_timely;
853 	struct sysctl_oid *rack_timers;
854 	struct sysctl_oid *rack_tlp;
855 	struct sysctl_oid *rack_misc;
856 	struct sysctl_oid *rack_features;
857 	struct sysctl_oid *rack_measure;
858 	struct sysctl_oid *rack_probertt;
859 	struct sysctl_oid *rack_hw_pacing;
860 
861 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
862 	    SYSCTL_CHILDREN(rack_sysctl_root),
863 	    OID_AUTO,
864 	    "sack_attack",
865 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
866 	    "Rack Sack Attack Counters and Controls");
867 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
868 	    SYSCTL_CHILDREN(rack_sysctl_root),
869 	    OID_AUTO,
870 	    "stats",
871 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
872 	    "Rack Counters");
873 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
874 	    SYSCTL_CHILDREN(rack_sysctl_root),
875 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
876 	    &rack_rate_sample_method , USE_RTT_LOW,
877 	    "What method should we use for rate sampling 0=high, 1=low ");
878 	/* Probe rtt related controls */
879 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
880 	    SYSCTL_CHILDREN(rack_sysctl_root),
881 	    OID_AUTO,
882 	    "probertt",
883 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
884 	    "ProbeRTT related Controls");
885 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
886 	    SYSCTL_CHILDREN(rack_probertt),
887 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
888 	    &rack_atexit_prtt_hbp, 130,
889 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
890 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
891 	    SYSCTL_CHILDREN(rack_probertt),
892 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
893 	    &rack_atexit_prtt, 130,
894 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
895 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
896 	    SYSCTL_CHILDREN(rack_probertt),
897 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
898 	    &rack_per_of_gp_probertt, 60,
899 	    "What percentage of goodput do we pace at in probertt");
900 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
901 	    SYSCTL_CHILDREN(rack_probertt),
902 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
903 	    &rack_per_of_gp_probertt_reduce, 10,
904 	    "What percentage of goodput do we reduce every gp_srtt");
905 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
906 	    SYSCTL_CHILDREN(rack_probertt),
907 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
908 	    &rack_per_of_gp_lowthresh, 40,
909 	    "What percentage of goodput do we allow the multiplier to fall to");
910 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
911 	    SYSCTL_CHILDREN(rack_probertt),
912 	    OID_AUTO, "time_between", CTLFLAG_RW,
913 	    & rack_time_between_probertt, 96000000,
914 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
915 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_probertt),
917 	    OID_AUTO, "safety", CTLFLAG_RW,
918 	    &rack_probe_rtt_safety_val, 2000000,
919 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
920 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
921 	    SYSCTL_CHILDREN(rack_probertt),
922 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
923 	    &rack_probe_rtt_sets_cwnd, 0,
924 	    "Do we set the cwnd too (if always_lower is on)");
925 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
926 	    SYSCTL_CHILDREN(rack_probertt),
927 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
928 	    &rack_max_drain_wait, 2,
929 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
930 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
931 	    SYSCTL_CHILDREN(rack_probertt),
932 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
933 	    &rack_must_drain, 1,
934 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
935 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_probertt),
937 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
938 	    &rack_probertt_use_min_rtt_entry, 1,
939 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
940 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_probertt),
942 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
943 	    &rack_probertt_use_min_rtt_exit, 0,
944 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
945 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_probertt),
947 	    OID_AUTO, "length_div", CTLFLAG_RW,
948 	    &rack_probertt_gpsrtt_cnt_div, 0,
949 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
950 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_probertt),
952 	    OID_AUTO, "length_mul", CTLFLAG_RW,
953 	    &rack_probertt_gpsrtt_cnt_mul, 0,
954 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
955 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_probertt),
957 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
958 	    &rack_min_probertt_hold, 200000,
959 	    "What is the minimum time we hold probertt at target");
960 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_probertt),
962 	    OID_AUTO, "filter_life", CTLFLAG_RW,
963 	    &rack_probertt_filter_life, 10000000,
964 	    "What is the time for the filters life in useconds");
965 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_probertt),
967 	    OID_AUTO, "lower_within", CTLFLAG_RW,
968 	    &rack_probertt_lower_within, 10,
969 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
970 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
971 	    SYSCTL_CHILDREN(rack_probertt),
972 	    OID_AUTO, "must_move", CTLFLAG_RW,
973 	    &rack_min_rtt_movement, 250,
974 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
975 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
976 	    SYSCTL_CHILDREN(rack_probertt),
977 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
978 	    &rack_probertt_clear_is, 1,
979 	    "Do we clear I/S counts on exiting probe-rtt");
980 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_probertt),
982 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
983 	    &rack_max_drain_hbp, 1,
984 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
985 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
986 	    SYSCTL_CHILDREN(rack_probertt),
987 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
988 	    &rack_hbp_thresh, 3,
989 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
990 	/* Pacing related sysctls */
991 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
992 	    SYSCTL_CHILDREN(rack_sysctl_root),
993 	    OID_AUTO,
994 	    "pacing",
995 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
996 	    "Pacing related Controls");
997 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_pacing),
999 	    OID_AUTO, "fulldgpinrec", CTLFLAG_RW,
1000 	    &rack_uses_full_dgp_in_rec, 1,
1001 	    "Do we use all DGP features in recovery (fillcw, timely et.al.)?");
1002 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_pacing),
1004 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1005 	    &rack_full_buffer_discount, 10,
1006 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1007 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_pacing),
1009 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1010 	    &rack_fill_cw_state, 0,
1011 	    "Enable fillcw on new connections (default=0 off)?");
1012 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_pacing),
1014 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1015 	    &rack_pacing_min_seg, 0,
1016 	    "What is the min burst size for pacing (0 disables)?");
1017 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_pacing),
1019 	    OID_AUTO, "divisor", CTLFLAG_RW,
1020 	    &rack_default_pacing_divisor, 4,
1021 	    "What is the default divisor given to the rl code?");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_pacing),
1024 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1025 	    &rack_bw_multipler, 2,
1026 	    "What is the multiplier of the current gp_est that fillcw can increase the b/w too?");
1027 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_pacing),
1029 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1030 	    &rack_max_per_above, 30,
1031 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1032 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_pacing),
1034 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1035 	    &rack_pace_one_seg, 0,
1036 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1037 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1038 	    SYSCTL_CHILDREN(rack_pacing),
1039 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1040 	    &rack_limit_time_with_srtt, 0,
1041 	    "Do we limit pacing time based on srtt");
1042 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1043 	    SYSCTL_CHILDREN(rack_pacing),
1044 	    OID_AUTO, "init_win", CTLFLAG_RW,
1045 	    &rack_default_init_window, 0,
1046 	    "Do we have a rack initial window 0 = system default");
1047 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1048 	    SYSCTL_CHILDREN(rack_pacing),
1049 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1050 	    &rack_per_of_gp_ss, 250,
1051 	    "If non zero, what percentage of goodput to pace at in slow start");
1052 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1053 	    SYSCTL_CHILDREN(rack_pacing),
1054 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1055 	    &rack_per_of_gp_ca, 150,
1056 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1057 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1058 	    SYSCTL_CHILDREN(rack_pacing),
1059 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1060 	    &rack_per_of_gp_rec, 200,
1061 	    "If non zero, what percentage of goodput to pace at in recovery");
1062 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1063 	    SYSCTL_CHILDREN(rack_pacing),
1064 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1065 	    &rack_hptsi_segments, 40,
1066 	    "What size is the max for TSO segments in pacing and burst mitigation");
1067 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1068 	    SYSCTL_CHILDREN(rack_pacing),
1069 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1070 	    &rack_slot_reduction, 4,
1071 	    "When doing only burst mitigation what is the reduce divisor");
1072 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1073 	    SYSCTL_CHILDREN(rack_sysctl_root),
1074 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1075 	    &rack_pace_every_seg, 0,
1076 	    "If set we use pacing, if clear we use only the original burst mitigation");
1077 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1078 	    SYSCTL_CHILDREN(rack_pacing),
1079 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1080 	    &rack_bw_rate_cap, 0,
1081 	    "If set we apply this value to the absolute rate cap used by pacing");
1082 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1083 	    SYSCTL_CHILDREN(rack_sysctl_root),
1084 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1085 	    &rack_req_measurements, 1,
1086 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1087 	/* Hardware pacing */
1088 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1089 	    SYSCTL_CHILDREN(rack_sysctl_root),
1090 	    OID_AUTO,
1091 	    "hdwr_pacing",
1092 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1093 	    "Pacing related Controls");
1094 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_hw_pacing),
1096 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1097 	    &rack_hw_rwnd_factor, 2,
1098 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1099 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100 	    SYSCTL_CHILDREN(rack_hw_pacing),
1101 	    OID_AUTO, "precheck", CTLFLAG_RW,
1102 	    &rack_hw_check_queue, 0,
1103 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1104 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 	    SYSCTL_CHILDREN(rack_hw_pacing),
1106 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1107 	    &rack_enobuf_hw_boost_mult, 0,
1108 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1109 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1110 	    SYSCTL_CHILDREN(rack_hw_pacing),
1111 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1112 	    &rack_enobuf_hw_max, 2,
1113 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1114 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1115 	    SYSCTL_CHILDREN(rack_hw_pacing),
1116 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1117 	    &rack_enobuf_hw_min, 2,
1118 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1119 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1120 	    SYSCTL_CHILDREN(rack_hw_pacing),
1121 	    OID_AUTO, "enable", CTLFLAG_RW,
1122 	    &rack_enable_hw_pacing, 0,
1123 	    "Should RACK attempt to use hw pacing?");
1124 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125 	    SYSCTL_CHILDREN(rack_hw_pacing),
1126 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1127 	    &rack_hw_rate_caps, 0,
1128 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1129 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130 	    SYSCTL_CHILDREN(rack_hw_pacing),
1131 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1132 	    &rack_hw_rate_cap_per, 0,
1133 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1134 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135 	    SYSCTL_CHILDREN(rack_hw_pacing),
1136 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1137 	    &rack_hw_rate_min, 0,
1138 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1139 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_hw_pacing),
1141 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1142 	    &rack_hw_rate_to_low, 0,
1143 	    "If we fall below this rate, dis-engage hw pacing?");
1144 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_hw_pacing),
1146 	    OID_AUTO, "up_only", CTLFLAG_RW,
1147 	    &rack_hw_up_only, 0,
1148 	    "Do we allow hw pacing to lower the rate selected?");
1149 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_hw_pacing),
1151 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1152 	    &rack_hw_pace_extra_slots, 0,
1153 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1154 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1155 	    SYSCTL_CHILDREN(rack_sysctl_root),
1156 	    OID_AUTO,
1157 	    "timely",
1158 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1159 	    "Rack Timely RTT Controls");
1160 	/* Timely based GP dynmics */
1161 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162 	    SYSCTL_CHILDREN(rack_timely),
1163 	    OID_AUTO, "upper", CTLFLAG_RW,
1164 	    &rack_gp_per_bw_mul_up, 2,
1165 	    "Rack timely upper range for equal b/w (in percentage)");
1166 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167 	    SYSCTL_CHILDREN(rack_timely),
1168 	    OID_AUTO, "lower", CTLFLAG_RW,
1169 	    &rack_gp_per_bw_mul_down, 4,
1170 	    "Rack timely lower range for equal b/w (in percentage)");
1171 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(rack_timely),
1173 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1174 	    &rack_gp_rtt_maxmul, 3,
1175 	    "Rack timely multiplier of lowest rtt for rtt_max");
1176 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177 	    SYSCTL_CHILDREN(rack_timely),
1178 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1179 	    &rack_gp_rtt_mindiv, 4,
1180 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1181 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_timely),
1183 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1184 	    &rack_gp_rtt_minmul, 1,
1185 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1186 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187 	    SYSCTL_CHILDREN(rack_timely),
1188 	    OID_AUTO, "decrease", CTLFLAG_RW,
1189 	    &rack_gp_decrease_per, 80,
1190 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1191 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192 	    SYSCTL_CHILDREN(rack_timely),
1193 	    OID_AUTO, "increase", CTLFLAG_RW,
1194 	    &rack_gp_increase_per, 2,
1195 	    "Rack timely increase perentage of our GP multiplication factor");
1196 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1197 	    SYSCTL_CHILDREN(rack_timely),
1198 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1199 	    &rack_per_lower_bound, 50,
1200 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1201 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(rack_timely),
1203 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1204 	    &rack_gain_p5_ub, 250,
1205 	    "Profile 5 upper bound to timely gain");
1206 
1207 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1208 	    SYSCTL_CHILDREN(rack_timely),
1209 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1210 	    &rack_per_upper_bound_ss, 0,
1211 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1212 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1213 	    SYSCTL_CHILDREN(rack_timely),
1214 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1215 	    &rack_per_upper_bound_ca, 0,
1216 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1217 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1218 	    SYSCTL_CHILDREN(rack_timely),
1219 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1220 	    &rack_do_dyn_mul, 0,
1221 	    "Rack timely do we enable dynmaic timely goodput by default");
1222 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1223 	    SYSCTL_CHILDREN(rack_timely),
1224 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1225 	    &rack_gp_no_rec_chg, 1,
1226 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1227 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1228 	    SYSCTL_CHILDREN(rack_timely),
1229 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1230 	    &rack_timely_dec_clear, 6,
1231 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1232 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1233 	    SYSCTL_CHILDREN(rack_timely),
1234 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1235 	    &rack_timely_max_push_rise, 3,
1236 	    "Rack timely how many times do we push up with b/w increase");
1237 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1238 	    SYSCTL_CHILDREN(rack_timely),
1239 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1240 	    &rack_timely_max_push_drop, 3,
1241 	    "Rack timely how many times do we push back on b/w decent");
1242 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1243 	    SYSCTL_CHILDREN(rack_timely),
1244 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1245 	    &rack_timely_min_segs, 4,
1246 	    "Rack timely when setting the cwnd what is the min num segments");
1247 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1248 	    SYSCTL_CHILDREN(rack_timely),
1249 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1250 	    &rack_use_max_for_nobackoff, 0,
1251 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1252 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1253 	    SYSCTL_CHILDREN(rack_timely),
1254 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1255 	    &rack_timely_int_timely_only, 0,
1256 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1257 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1258 	    SYSCTL_CHILDREN(rack_timely),
1259 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1260 	    &rack_timely_no_stopping, 0,
1261 	    "Rack timely don't stop increase");
1262 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1263 	    SYSCTL_CHILDREN(rack_timely),
1264 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1265 	    &rack_down_raise_thresh, 100,
1266 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1267 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1268 	    SYSCTL_CHILDREN(rack_timely),
1269 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1270 	    &rack_req_segs, 1,
1271 	    "Bottom dragging if not these many segments outstanding and room");
1272 
1273 	/* TLP and Rack related parameters */
1274 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1275 	    SYSCTL_CHILDREN(rack_sysctl_root),
1276 	    OID_AUTO,
1277 	    "tlp",
1278 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1279 	    "TLP and Rack related Controls");
1280 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1281 	    SYSCTL_CHILDREN(rack_tlp),
1282 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1283 	    &use_rack_rr, 1,
1284 	    "Do we use Rack Rapid Recovery");
1285 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1286 	    SYSCTL_CHILDREN(rack_tlp),
1287 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1288 	    &rack_max_abc_post_recovery, 2,
1289 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1290 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1291 	    SYSCTL_CHILDREN(rack_tlp),
1292 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1293 	    &rack_non_rxt_use_cr, 0,
1294 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1295 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1296 	    SYSCTL_CHILDREN(rack_tlp),
1297 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1298 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1299 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1300 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1301 	    SYSCTL_CHILDREN(rack_tlp),
1302 	    OID_AUTO, "limit", CTLFLAG_RW,
1303 	    &rack_tlp_limit, 2,
1304 	    "How many TLP's can be sent without sending new data");
1305 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1306 	    SYSCTL_CHILDREN(rack_tlp),
1307 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1308 	    &rack_tlp_use_greater, 1,
1309 	    "Should we use the rack_rtt time if its greater than srtt");
1310 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1311 	    SYSCTL_CHILDREN(rack_tlp),
1312 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1313 	    &rack_tlp_min, 10000,
1314 	    "TLP minimum timeout per the specification (in microseconds)");
1315 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1316 	    SYSCTL_CHILDREN(rack_tlp),
1317 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1318 	    &rack_always_send_oldest, 0,
1319 	    "Should we always send the oldest TLP and RACK-TLP");
1320 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1321 	    SYSCTL_CHILDREN(rack_tlp),
1322 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1323 	    &rack_limited_retran, 0,
1324 	    "How many times can a rack timeout drive out sends");
1325 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1326 	    SYSCTL_CHILDREN(rack_tlp),
1327 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1328 	    &rack_lower_cwnd_at_tlp, 0,
1329 	    "When a TLP completes a retran should we enter recovery");
1330 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_tlp),
1332 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1333 	    &rack_reorder_thresh, 2,
1334 	    "What factor for rack will be added when seeing reordering (shift right)");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_tlp),
1337 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1338 	    &rack_tlp_thresh, 1,
1339 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_tlp),
1342 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1343 	    &rack_reorder_fade, 60000000,
1344 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1345 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_tlp),
1347 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1348 	    &rack_pkt_delay, 1000,
1349 	    "Extra RACK time (in microseconds) besides reordering thresh");
1350 
1351 	/* Timer related controls */
1352 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1353 	    SYSCTL_CHILDREN(rack_sysctl_root),
1354 	    OID_AUTO,
1355 	    "timers",
1356 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1357 	    "Timer related controls");
1358 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1359 	    SYSCTL_CHILDREN(rack_timers),
1360 	    OID_AUTO, "persmin", CTLFLAG_RW,
1361 	    &rack_persist_min, 250000,
1362 	    "What is the minimum time in microseconds between persists");
1363 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1364 	    SYSCTL_CHILDREN(rack_timers),
1365 	    OID_AUTO, "persmax", CTLFLAG_RW,
1366 	    &rack_persist_max, 2000000,
1367 	    "What is the largest delay in microseconds between persists");
1368 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1369 	    SYSCTL_CHILDREN(rack_timers),
1370 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1371 	    &rack_delayed_ack_time, 40000,
1372 	    "Delayed ack time (40ms in microseconds)");
1373 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374 	    SYSCTL_CHILDREN(rack_timers),
1375 	    OID_AUTO, "minrto", CTLFLAG_RW,
1376 	    &rack_rto_min, 30000,
1377 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1378 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379 	    SYSCTL_CHILDREN(rack_timers),
1380 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1381 	    &rack_rto_max, 4000000,
1382 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1383 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384 	    SYSCTL_CHILDREN(rack_timers),
1385 	    OID_AUTO, "minto", CTLFLAG_RW,
1386 	    &rack_min_to, 1000,
1387 	    "Minimum rack timeout in microseconds");
1388 	/* Measure controls */
1389 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1390 	    SYSCTL_CHILDREN(rack_sysctl_root),
1391 	    OID_AUTO,
1392 	    "measure",
1393 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1394 	    "Measure related controls");
1395 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1396 	    SYSCTL_CHILDREN(rack_measure),
1397 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1398 	    &rack_wma_divisor, 8,
1399 	    "When doing b/w calculation what is the  divisor for the WMA");
1400 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1401 	    SYSCTL_CHILDREN(rack_measure),
1402 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1403 	    &rack_cwnd_block_ends_measure, 0,
1404 	    "Does a cwnd just-return end the measurement window (app limited)");
1405 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1406 	    SYSCTL_CHILDREN(rack_measure),
1407 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1408 	    &rack_rwnd_block_ends_measure, 0,
1409 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1410 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1411 	    SYSCTL_CHILDREN(rack_measure),
1412 	    OID_AUTO, "min_target", CTLFLAG_RW,
1413 	    &rack_def_data_window, 20,
1414 	    "What is the minimum target window (in mss) for a GP measurements");
1415 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1416 	    SYSCTL_CHILDREN(rack_measure),
1417 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1418 	    &rack_goal_bdp, 2,
1419 	    "What is the goal BDP to measure");
1420 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_measure),
1422 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1423 	    &rack_min_srtts, 1,
1424 	    "What is the goal BDP to measure");
1425 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1426 	    SYSCTL_CHILDREN(rack_measure),
1427 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1428 	    &rack_min_measure_usec, 0,
1429 	    "What is the Minimum time time for a measurement if 0, this is off");
1430 	/* Features */
1431 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1432 	    SYSCTL_CHILDREN(rack_sysctl_root),
1433 	    OID_AUTO,
1434 	    "features",
1435 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1436 	    "Feature controls");
1437 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1438 	    SYSCTL_CHILDREN(rack_features),
1439 	    OID_AUTO, "rxt_clamp_thresh", CTLFLAG_RW,
1440 	    &rack_rxt_clamp_thresh, 0,
1441 	    "Bit encoded clamping setup bits CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP");
1442 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443 	    SYSCTL_CHILDREN(rack_features),
1444 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1445 	    &rack_hybrid_allow_set_maxseg, 0,
1446 	    "Should hybrid pacing allow the setmss command");
1447 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448 	    SYSCTL_CHILDREN(rack_features),
1449 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1450 	    &rack_use_cmp_acks, 1,
1451 	    "Should RACK have LRO send compressed acks");
1452 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453 	    SYSCTL_CHILDREN(rack_features),
1454 	    OID_AUTO, "fsb", CTLFLAG_RW,
1455 	    &rack_use_fsb, 1,
1456 	    "Should RACK use the fast send block?");
1457 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1458 	    SYSCTL_CHILDREN(rack_features),
1459 	    OID_AUTO, "rfo", CTLFLAG_RW,
1460 	    &rack_use_rfo, 1,
1461 	    "Should RACK use rack_fast_output()?");
1462 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1463 	    SYSCTL_CHILDREN(rack_features),
1464 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1465 	    &rack_use_rsm_rfo, 1,
1466 	    "Should RACK use rack_fast_rsm_output()?");
1467 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1468 	    SYSCTL_CHILDREN(rack_features),
1469 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1470 	    &rack_enable_mqueue_for_nonpaced, 0,
1471 	    "Should RACK use mbuf queuing for non-paced connections");
1472 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1473 	    SYSCTL_CHILDREN(rack_features),
1474 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1475 	    &rack_do_hystart, 0,
1476 	    "Should RACK enable HyStart++ on connections?");
1477 	/* Misc rack controls */
1478 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1479 	    SYSCTL_CHILDREN(rack_sysctl_root),
1480 	    OID_AUTO,
1481 	    "misc",
1482 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1483 	    "Misc related controls");
1484 #ifdef TCP_ACCOUNTING
1485 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1486 	    SYSCTL_CHILDREN(rack_misc),
1487 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1488 	    &rack_tcp_accounting, 0,
1489 	    "Should we turn on TCP accounting for all rack sessions?");
1490 #endif
1491 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1492 	    SYSCTL_CHILDREN(rack_misc),
1493 	    OID_AUTO, "dnd", CTLFLAG_RW,
1494 	    &rack_dnd_default, 0,
1495 	    "Do not disturb default for rack_rrr = 3");
1496 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1497 	    SYSCTL_CHILDREN(rack_misc),
1498 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1499 	    &sad_seg_size_per, 800,
1500 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1501 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1502 	    SYSCTL_CHILDREN(rack_misc),
1503 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1504 	    &rack_rxt_controls, 0,
1505 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1506 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1507 	    SYSCTL_CHILDREN(rack_misc),
1508 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1509 	    &rack_hibeta_setting, 0,
1510 	    "Do we ue a high beta (80 instead of 50)?");
1511 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1512 	    SYSCTL_CHILDREN(rack_misc),
1513 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1514 	    &rack_apply_rtt_with_reduced_conf, 0,
1515 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1516 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1517 	    SYSCTL_CHILDREN(rack_misc),
1518 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1519 	    &rack_dsack_std_based, 3,
1520 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1521 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_misc),
1523 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1524 	    &rack_prr_addbackmax, 2,
1525 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1526 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1527 	    SYSCTL_CHILDREN(rack_misc),
1528 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1529 	    &rack_stats_gets_ms_rtt, 1,
1530 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1531 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(rack_misc),
1533 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1534 	    &rack_client_low_buf, 0,
1535 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1536 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1537 	    SYSCTL_CHILDREN(rack_misc),
1538 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1539 	    &rack_def_profile, 0,
1540 	    "Should RACK use a default profile (0=no, num == profile num)?");
1541 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1542 	    SYSCTL_CHILDREN(rack_misc),
1543 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1544 	    &rack_enable_shared_cwnd, 1,
1545 	    "Should RACK try to use the shared cwnd on connections where allowed");
1546 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1547 	    SYSCTL_CHILDREN(rack_misc),
1548 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1549 	    &rack_limits_scwnd, 1,
1550 	    "Should RACK place low end time limits on the shared cwnd feature");
1551 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1552 	    SYSCTL_CHILDREN(rack_misc),
1553 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1554 	    &rack_disable_prr, 0,
1555 	    "Should RACK not use prr and only pace (must have pacing on)");
1556 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1557 	    SYSCTL_CHILDREN(rack_misc),
1558 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1559 	    &rack_verbose_logging, 0,
1560 	    "Should RACK black box logging be verbose");
1561 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_misc),
1563 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1564 	    &rack_ignore_data_after_close, 1,
1565 	    "Do we hold off sending a RST until all pending data is ack'd");
1566 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_misc),
1568 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1569 	    &rack_sack_not_required, 1,
1570 	    "Do we allow rack to run on connections not supporting SACK");
1571 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_misc),
1573 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1574 	    &rack_send_a_lot_in_prr, 1,
1575 	    "Send a lot in prr");
1576 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_misc),
1578 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1579 	    &rack_autosndbuf_inc, 20,
1580 	    "What percentage should rack scale up its snd buffer by?");
1581 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_misc),
1583 	    OID_AUTO, "rnds_for_rxt_clamp", CTLFLAG_RW,
1584 	    &rack_rxt_min_rnds, 10,
1585 	    "Number of rounds needed between RTT clamps due to high loss rates");
1586 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_misc),
1588 	    OID_AUTO, "rnds_for_unclamp", CTLFLAG_RW,
1589 	    &rack_unclamp_round_thresh, 100,
1590 	    "Number of rounds needed with no loss to unclamp");
1591 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_misc),
1593 	    OID_AUTO, "rxt_threshs_for_unclamp", CTLFLAG_RW,
1594 	    &rack_unclamp_rxt_thresh, 5,
1595 	   "Percentage of retransmits we need to be under to unclamp (5 = .5 percent)\n");
1596 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_misc),
1598 	    OID_AUTO, "clamp_ss_upper", CTLFLAG_RW,
1599 	    &rack_clamp_ss_upper, 110,
1600 	    "Clamp percentage ceiling in SS?");
1601 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1602 	    SYSCTL_CHILDREN(rack_misc),
1603 	    OID_AUTO, "clamp_ca_upper", CTLFLAG_RW,
1604 	    &rack_clamp_ca_upper, 110,
1605 	    "Clamp percentage ceiling in CA?");
1606 	/* Sack Attacker detection stuff */
1607 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1608 	    SYSCTL_CHILDREN(rack_attack),
1609 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1610 	    &rack_merge_out_sacks_on_attack, 0,
1611 	    "Do we merge the sendmap when we decide we are being attacked?");
1612 
1613 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1614 	    SYSCTL_CHILDREN(rack_attack),
1615 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1616 	    &rack_highest_sack_thresh_seen, 0,
1617 	    "Highest sack to ack ratio seen");
1618 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1619 	    SYSCTL_CHILDREN(rack_attack),
1620 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1621 	    &rack_highest_move_thresh_seen, 0,
1622 	    "Highest move to non-move ratio seen");
1623 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1624 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1625 	    SYSCTL_CHILDREN(rack_attack),
1626 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1627 	    &rack_ack_total,
1628 	    "Total number of Ack's");
1629 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1630 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_attack),
1632 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1633 	    &rack_express_sack,
1634 	    "Total expresss number of Sack's");
1635 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1636 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1637 	    SYSCTL_CHILDREN(rack_attack),
1638 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1639 	    &rack_sack_total,
1640 	    "Total number of SACKs");
1641 	rack_move_none = counter_u64_alloc(M_WAITOK);
1642 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1643 	    SYSCTL_CHILDREN(rack_attack),
1644 	    OID_AUTO, "move_none", CTLFLAG_RD,
1645 	    &rack_move_none,
1646 	    "Total number of SACK index reuse of positions under threshold");
1647 	rack_move_some = counter_u64_alloc(M_WAITOK);
1648 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1649 	    SYSCTL_CHILDREN(rack_attack),
1650 	    OID_AUTO, "move_some", CTLFLAG_RD,
1651 	    &rack_move_some,
1652 	    "Total number of SACK index reuse of positions over threshold");
1653 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1654 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1655 	    SYSCTL_CHILDREN(rack_attack),
1656 	    OID_AUTO, "attacks", CTLFLAG_RD,
1657 	    &rack_sack_attacks_detected,
1658 	    "Total number of SACK attackers that had sack disabled");
1659 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1660 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(rack_attack),
1662 	    OID_AUTO, "reversed", CTLFLAG_RD,
1663 	    &rack_sack_attacks_reversed,
1664 	    "Total number of SACK attackers that were later determined false positive");
1665 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1666 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667 	    SYSCTL_CHILDREN(rack_attack),
1668 	    OID_AUTO, "suspect", CTLFLAG_RD,
1669 	    &rack_sack_attacks_suspect,
1670 	    "Total number of SACKs that triggered early detection");
1671 
1672 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1673 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674 	    SYSCTL_CHILDREN(rack_attack),
1675 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1676 	    &rack_sack_used_next_merge,
1677 	    "Total number of times we used the next merge");
1678 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1679 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(rack_attack),
1681 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1682 	    &rack_sack_used_prev_merge,
1683 	    "Total number of times we used the prev merge");
1684 	/* Counters */
1685 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1686 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687 	    SYSCTL_CHILDREN(rack_counters),
1688 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1689 	    &rack_total_bytes,
1690 	    "Total number of bytes sent");
1691 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1692 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693 	    SYSCTL_CHILDREN(rack_counters),
1694 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1695 	    &rack_fto_send, "Total number of rack_fast_output sends");
1696 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1697 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698 	    SYSCTL_CHILDREN(rack_counters),
1699 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1700 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1701 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1702 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1703 	    SYSCTL_CHILDREN(rack_counters),
1704 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1705 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1706 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1707 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1708 	    SYSCTL_CHILDREN(rack_counters),
1709 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1710 	    &rack_non_fto_send, "Total number of rack_output first sends");
1711 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1712 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1713 	    SYSCTL_CHILDREN(rack_counters),
1714 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1715 	    &rack_extended_rfo, "Total number of times we extended rfo");
1716 
1717 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1718 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1719 	    SYSCTL_CHILDREN(rack_counters),
1720 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1721 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1722 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1723 
1724 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1725 	    SYSCTL_CHILDREN(rack_counters),
1726 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1727 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1728 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1729 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1730 	    SYSCTL_CHILDREN(rack_counters),
1731 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1732 	    &rack_tlp_tot,
1733 	    "Total number of tail loss probe expirations");
1734 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1735 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1736 	    SYSCTL_CHILDREN(rack_counters),
1737 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1738 	    &rack_tlp_newdata,
1739 	    "Total number of tail loss probe sending new data");
1740 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1741 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1742 	    SYSCTL_CHILDREN(rack_counters),
1743 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1744 	    &rack_tlp_retran,
1745 	    "Total number of tail loss probe sending retransmitted data");
1746 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1747 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1748 	    SYSCTL_CHILDREN(rack_counters),
1749 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1750 	    &rack_tlp_retran_bytes,
1751 	    "Total bytes of tail loss probe sending retransmitted data");
1752 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1753 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1754 	    SYSCTL_CHILDREN(rack_counters),
1755 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1756 	    &rack_to_tot,
1757 	    "Total number of times the rack to expired");
1758 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1759 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1760 	    SYSCTL_CHILDREN(rack_counters),
1761 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1762 	    &rack_saw_enobuf,
1763 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1764 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1765 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1766 	    SYSCTL_CHILDREN(rack_counters),
1767 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1768 	    &rack_saw_enobuf_hw,
1769 	    "Total number of times a send returned enobuf for hdwr paced connections");
1770 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1771 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1772 	    SYSCTL_CHILDREN(rack_counters),
1773 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1774 	    &rack_saw_enetunreach,
1775 	    "Total number of times a send received a enetunreachable");
1776 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1777 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1778 	    SYSCTL_CHILDREN(rack_counters),
1779 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1780 	    &rack_hot_alloc,
1781 	    "Total allocations from the top of our list");
1782 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1783 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1784 	    SYSCTL_CHILDREN(rack_counters),
1785 	    OID_AUTO, "allocs", CTLFLAG_RD,
1786 	    &rack_to_alloc,
1787 	    "Total allocations of tracking structures");
1788 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1789 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1790 	    SYSCTL_CHILDREN(rack_counters),
1791 	    OID_AUTO, "allochard", CTLFLAG_RD,
1792 	    &rack_to_alloc_hard,
1793 	    "Total allocations done with sleeping the hard way");
1794 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1795 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1796 	    SYSCTL_CHILDREN(rack_counters),
1797 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1798 	    &rack_to_alloc_emerg,
1799 	    "Total allocations done from emergency cache");
1800 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1801 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1802 	    SYSCTL_CHILDREN(rack_counters),
1803 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1804 	    &rack_to_alloc_limited,
1805 	    "Total allocations dropped due to limit");
1806 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1807 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1808 	    SYSCTL_CHILDREN(rack_counters),
1809 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1810 	    &rack_alloc_limited_conns,
1811 	    "Connections with allocations dropped due to limit");
1812 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1813 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814 	    SYSCTL_CHILDREN(rack_counters),
1815 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1816 	    &rack_split_limited,
1817 	    "Split allocations dropped due to limit");
1818 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1819 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820 	    SYSCTL_CHILDREN(rack_counters),
1821 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1822 	    &rack_rxt_clamps_cwnd,
1823 	    "Number of times that excessive rxt clamped the cwnd down");
1824 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1825 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826 	    SYSCTL_CHILDREN(rack_counters),
1827 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1828 	    &rack_rxt_clamps_cwnd_uniq,
1829 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1830 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1831 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1832 	    SYSCTL_CHILDREN(rack_counters),
1833 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1834 	    &rack_persists_sends,
1835 	    "Number of times we sent a persist probe");
1836 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1837 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1838 	    SYSCTL_CHILDREN(rack_counters),
1839 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1840 	    &rack_persists_acks,
1841 	    "Number of times a persist probe was acked");
1842 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1843 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1844 	    SYSCTL_CHILDREN(rack_counters),
1845 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1846 	    &rack_persists_loss,
1847 	    "Number of times we detected a lost persist probe (no ack)");
1848 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1849 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1850 	    SYSCTL_CHILDREN(rack_counters),
1851 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1852 	    &rack_persists_lost_ends,
1853 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1854 #ifdef INVARIANTS
1855 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1856 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1857 	    SYSCTL_CHILDREN(rack_counters),
1858 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1859 	    &rack_adjust_map_bw,
1860 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1861 #endif
1862 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1863 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1864 	    SYSCTL_CHILDREN(rack_counters),
1865 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1866 	    &rack_multi_single_eq,
1867 	    "Number of compressed acks total represented");
1868 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1869 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1870 	    SYSCTL_CHILDREN(rack_counters),
1871 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1872 	    &rack_proc_non_comp_ack,
1873 	    "Number of non compresseds acks that we processed");
1874 
1875 
1876 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1877 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1878 	    SYSCTL_CHILDREN(rack_counters),
1879 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1880 	    &rack_sack_proc_all,
1881 	    "Total times we had to walk whole list for sack processing");
1882 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1883 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1884 	    SYSCTL_CHILDREN(rack_counters),
1885 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1886 	    &rack_sack_proc_restart,
1887 	    "Total times we had to walk whole list due to a restart");
1888 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1889 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1890 	    SYSCTL_CHILDREN(rack_counters),
1891 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1892 	    &rack_sack_proc_short,
1893 	    "Total times we took shortcut for sack processing");
1894 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1895 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1896 	    SYSCTL_CHILDREN(rack_attack),
1897 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1898 	    &rack_sack_skipped_acked,
1899 	    "Total number of times we skipped previously sacked");
1900 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1901 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1902 	    SYSCTL_CHILDREN(rack_attack),
1903 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1904 	    &rack_sack_splits,
1905 	    "Total number of times we did the old fashion tree split");
1906 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1907 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1908 	    SYSCTL_CHILDREN(rack_counters),
1909 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1910 	    &rack_input_idle_reduces,
1911 	    "Total number of idle reductions on input");
1912 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1913 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1914 	    SYSCTL_CHILDREN(rack_counters),
1915 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1916 	    &rack_collapsed_win_seen,
1917 	    "Total number of collapsed window events seen (where our window shrinks)");
1918 
1919 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1920 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1921 	    SYSCTL_CHILDREN(rack_counters),
1922 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1923 	    &rack_collapsed_win,
1924 	    "Total number of collapsed window events where we mark packets");
1925 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1926 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1927 	    SYSCTL_CHILDREN(rack_counters),
1928 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1929 	    &rack_collapsed_win_rxt,
1930 	    "Total number of packets that were retransmitted");
1931 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1932 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1933 	    SYSCTL_CHILDREN(rack_counters),
1934 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1935 	    &rack_collapsed_win_rxt_bytes,
1936 	    "Total number of bytes that were retransmitted");
1937 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1938 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1939 	    SYSCTL_CHILDREN(rack_counters),
1940 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1941 	    &rack_try_scwnd,
1942 	    "Total number of scwnd attempts");
1943 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1944 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1945 	    OID_AUTO, "outsize", CTLFLAG_RD,
1946 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1947 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1948 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1949 	    OID_AUTO, "opts", CTLFLAG_RD,
1950 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1951 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1952 	    SYSCTL_CHILDREN(rack_sysctl_root),
1953 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1954 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1955 }
1956 
1957 static uint32_t
1958 rc_init_window(struct tcp_rack *rack)
1959 {
1960 	uint32_t win;
1961 
1962 	if (rack->rc_init_win == 0) {
1963 		/*
1964 		 * Nothing set by the user, use the system stack
1965 		 * default.
1966 		 */
1967 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1968 	}
1969 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1970 	return (win);
1971 }
1972 
1973 static uint64_t
1974 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1975 {
1976 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1977 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1978 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1979 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1980 	else
1981 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1982 }
1983 
1984 static void
1985 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1986 	uint64_t data, uint8_t mod, uint16_t aux,
1987 	struct tcp_sendfile_track *cur, int line)
1988 {
1989 #ifdef TCP_REQUEST_TRK
1990 	int do_log = 0;
1991 
1992 	/*
1993 	 * The rate cap one is noisy and only should come out when normal BB logging
1994 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
1995 	 * once per chunk and make up the BBpoint that can be turned on by the client.
1996 	 */
1997 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
1998 		/*
1999 		 * The very noisy two need to only come out when
2000 		 * we have verbose logging on.
2001 		 */
2002 		if (rack_verbose_logging != 0)
2003 			do_log = tcp_bblogging_on(rack->rc_tp);
2004 		else
2005 			do_log = 0;
2006 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2007 		/*
2008 		 * All other less noisy logs here except the measure which
2009 		 * also needs to come out on the point and the log.
2010 		 */
2011 		do_log = tcp_bblogging_on(rack->rc_tp);
2012 	} else {
2013 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2014 	}
2015 
2016 	if (do_log) {
2017 		union tcp_log_stackspecific log;
2018 		struct timeval tv;
2019 		uint64_t lt_bw;
2020 
2021 		/* Convert our ms to a microsecond */
2022 		memset(&log, 0, sizeof(log));
2023 
2024 		log.u_bbr.cwnd_gain = line;
2025 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2026 		log.u_bbr.rttProp = tim;
2027 		log.u_bbr.bw_inuse = cbw;
2028 		log.u_bbr.delRate = rack_get_gp_est(rack);
2029 		lt_bw = rack_get_lt_bw(rack);
2030 		log.u_bbr.flex1 = seq;
2031 		log.u_bbr.pacing_gain = aux;
2032 		/* lt_bw = < flex3 | flex2 > */
2033 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2034 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2035 		/* Record the last obtained us rtt in inflight */
2036 		if (cur == NULL) {
2037 			/* Make sure we are looking at the right log if an overide comes in */
2038 			cur = rack->r_ctl.rc_last_sft;
2039 		}
2040 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2041 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2042 		else {
2043 			/* Use the last known rtt i.e. the rack-rtt */
2044 			log.u_bbr.inflight = rack->rc_rack_rtt;
2045 		}
2046 		if (cur != NULL) {
2047 			uint64_t off;
2048 
2049 			log.u_bbr.cur_del_rate = cur->deadline;
2050 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2051 				/* start = < lost | pkt_epoch > */
2052 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2053 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2054 				log.u_bbr.flex6 = cur->start_seq;
2055 				log.u_bbr.pkts_out = cur->end_seq;
2056 			} else {
2057 				/* start = < lost | pkt_epoch > */
2058 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2059 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2060 				/* end = < pkts_out | flex6 > */
2061 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2062 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2063 			}
2064 			/* first_send = <lt_epoch | epoch> */
2065 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2066 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2067 			/* localtime = <delivered | applimited>*/
2068 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2069 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2070 #ifdef TCP_REQUEST_TRK
2071 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2072 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2073 #endif
2074 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2075 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2076 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2077 		} else {
2078 			log.u_bbr.flex7 = 0xffff;
2079 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2080 		}
2081 		/*
2082 		 * Compose bbr_state to be a bit wise 0000ADHF
2083 		 * where A is the always_pace flag
2084 		 * where D is the dgp_on flag
2085 		 * where H is the hybrid_mode on flag
2086 		 * where F is the use_fixed_rate flag.
2087 		 */
2088 		log.u_bbr.bbr_state = rack->rc_always_pace;
2089 		log.u_bbr.bbr_state <<= 1;
2090 		log.u_bbr.bbr_state |= rack->dgp_on;
2091 		log.u_bbr.bbr_state <<= 1;
2092 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2093 		log.u_bbr.bbr_state <<= 1;
2094 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2095 		log.u_bbr.flex8 = mod;
2096 		tcp_log_event(rack->rc_tp, NULL,
2097 		    &rack->rc_inp->inp_socket->so_rcv,
2098 		    &rack->rc_inp->inp_socket->so_snd,
2099 		    TCP_HYBRID_PACING_LOG, 0,
2100 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2101 
2102 	}
2103 #endif
2104 }
2105 
2106 #ifdef TCP_REQUEST_TRK
2107 static void
2108 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2109 {
2110 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2111 		union tcp_log_stackspecific log;
2112 		struct timeval tv;
2113 		uint64_t off;
2114 
2115 		/* Convert our ms to a microsecond */
2116 		memset(&log, 0, sizeof(log));
2117 
2118 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2119 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2120 		log.u_bbr.delRate = cur->sent_at_fs;
2121 		log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2122 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2123 		log.u_bbr.cwnd_gain = line;
2124 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2125 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2126 		/* start = < flex1 | flex2 > */
2127 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2128 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2129 		/* end = < flex3 | flex4 > */
2130 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2131 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2132 
2133 		/* localtime = <delivered | applimited>*/
2134 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2135 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2136 		/* client timestamp = <lt_epoch | epoch>*/
2137 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2138 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2139 		/* now set all the flags in */
2140 		log.u_bbr.pkts_out = cur->hybrid_flags;
2141 		log.u_bbr.flex6 = cur->flags;
2142 		/*
2143 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2144 		 * where a false retransmit occurred so first_send  <-> lastsend may
2145 		 * include longer time then it actually took if we have a false rxt.
2146 		 */
2147 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2148 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2149 
2150 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2151 		tcp_log_event(rack->rc_tp, NULL,
2152 		    &rack->rc_inp->inp_socket->so_rcv,
2153 		    &rack->rc_inp->inp_socket->so_snd,
2154 		    TCP_HYBRID_PACING_LOG, 0,
2155 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2156 	}
2157 }
2158 #endif
2159 
2160 static inline uint64_t
2161 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2162 {
2163 	uint64_t ret_bw, ether;
2164 	uint64_t u_segsiz;
2165 
2166 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2167 	if (rack->r_is_v6){
2168 #ifdef INET6
2169 		ether += sizeof(struct ip6_hdr);
2170 #endif
2171 		ether += 14;	/* eheader size 6+6+2 */
2172 	} else {
2173 #ifdef INET
2174 		ether += sizeof(struct ip);
2175 #endif
2176 		ether += 14;	/* eheader size 6+6+2 */
2177 	}
2178 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2179 	ret_bw = bw;
2180 	ret_bw *= ether;
2181 	ret_bw /= u_segsiz;
2182 	return (ret_bw);
2183 }
2184 
2185 static void
2186 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2187 {
2188 #ifdef TCP_REQUEST_TRK
2189 	struct timeval tv;
2190 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2191 #endif
2192 
2193 	if (rack->r_ctl.bw_rate_cap == 0)
2194 		return;
2195 #ifdef TCP_REQUEST_TRK
2196 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2197 	    (rack->r_ctl.rc_last_sft != NULL)) {
2198 		/*
2199 		 * We have a dynamic cap. The original target
2200 		 * is in bw_rate_cap, but we need to look at
2201 		 * how long it is until we hit the deadline.
2202 		 */
2203 		struct tcp_sendfile_track *ent;
2204 
2205       		ent = rack->r_ctl.rc_last_sft;
2206 		microuptime(&tv);
2207 		timenow = tcp_tv_to_lusectick(&tv);
2208 		if (timenow >= ent->deadline) {
2209 			/* No time left we do DGP only */
2210 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2211 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2212 			rack->r_ctl.bw_rate_cap = 0;
2213 			return;
2214 		}
2215 		/* We have the time */
2216 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2217 		if (timeleft < HPTS_MSEC_IN_SEC) {
2218 			/* If there is less than a ms left just use DGPs rate */
2219 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2220 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2221 			rack->r_ctl.bw_rate_cap = 0;
2222 			return;
2223 		}
2224 		/*
2225 		 * Now lets find the amount of data left to send.
2226 		 *
2227 		 * Now ideally we want to use the end_seq to figure out how much more
2228 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2229 		 */
2230 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2231 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2232 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2233 			else {
2234 				/* TSNH, we should catch it at the send */
2235 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2236 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2237 				rack->r_ctl.bw_rate_cap = 0;
2238 				return;
2239 			}
2240 		} else {
2241 			/*
2242 			 * The hard way, figure out how much is gone and then
2243 			 * take that away from the total the client asked for
2244 			 * (thats off by tls overhead if this is tls).
2245 			 */
2246 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2247 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2248 			else
2249 				lengone = 0;
2250 			if (lengone < (ent->end - ent->start))
2251 				lenleft = (ent->end - ent->start) - lengone;
2252 			else {
2253 				/* TSNH, we should catch it at the send */
2254 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2255 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2256 				rack->r_ctl.bw_rate_cap = 0;
2257 				return;
2258 			}
2259 		}
2260 		if (lenleft == 0) {
2261 			/* We have it all sent */
2262 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2263 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2264 			if (rack->r_ctl.bw_rate_cap)
2265 				goto normal_ratecap;
2266 			else
2267 				return;
2268 		}
2269 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2270 		calcbw /= timeleft;
2271 		/* Now we must compensate for IP/TCP overhead */
2272 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2273 		/* Update the bit rate cap */
2274 		rack->r_ctl.bw_rate_cap = calcbw;
2275 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2276 		    (rack_hybrid_allow_set_maxseg == 1) &&
2277 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2278 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2279 			uint32_t orig_max;
2280 
2281 			orig_max = rack->r_ctl.rc_pace_max_segs;
2282 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2283 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2284 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2285 		}
2286 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2287 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2288 		if ((calcbw > 0) && (*bw > calcbw)) {
2289 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2290 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2291 			*capped = 1;
2292 			*bw = calcbw;
2293 		}
2294 		return;
2295 	}
2296 normal_ratecap:
2297 #endif
2298 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2299 #ifdef TCP_REQUEST_TRK
2300 		if (rack->rc_hybrid_mode &&
2301 		    rack->rc_catch_up &&
2302 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2303 		    (rack_hybrid_allow_set_maxseg == 1) &&
2304 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2305 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2306 			uint32_t orig_max;
2307 
2308 			orig_max = rack->r_ctl.rc_pace_max_segs;
2309 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2310 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2311 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2312 		}
2313 #endif
2314 		*capped = 1;
2315 		*bw = rack->r_ctl.bw_rate_cap;
2316 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2317 				   *bw, 0, 0,
2318 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2319 	}
2320 }
2321 
2322 static uint64_t
2323 rack_get_gp_est(struct tcp_rack *rack)
2324 {
2325 	uint64_t bw, lt_bw, ret_bw;
2326 
2327 	if (rack->rc_gp_filled == 0) {
2328 		/*
2329 		 * We have yet no b/w measurement,
2330 		 * if we have a user set initial bw
2331 		 * return it. If we don't have that and
2332 		 * we have an srtt, use the tcp IW (10) to
2333 		 * calculate a fictional b/w over the SRTT
2334 		 * which is more or less a guess. Note
2335 		 * we don't use our IW from rack on purpose
2336 		 * so if we have like IW=30, we are not
2337 		 * calculating a "huge" b/w.
2338 		 */
2339 		uint64_t srtt;
2340 
2341 		lt_bw = rack_get_lt_bw(rack);
2342 		if (lt_bw) {
2343 			/*
2344 			 * No goodput bw but a long-term b/w does exist
2345 			 * lets use that.
2346 			 */
2347 			ret_bw = lt_bw;
2348 			goto compensate;
2349 		}
2350 		if (rack->r_ctl.init_rate)
2351 			return (rack->r_ctl.init_rate);
2352 
2353 		/* Ok lets come up with the IW guess, if we have a srtt */
2354 		if (rack->rc_tp->t_srtt == 0) {
2355 			/*
2356 			 * Go with old pacing method
2357 			 * i.e. burst mitigation only.
2358 			 */
2359 			return (0);
2360 		}
2361 		/* Ok lets get the initial TCP win (not racks) */
2362 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2363 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2364 		bw *= (uint64_t)USECS_IN_SECOND;
2365 		bw /= srtt;
2366 		ret_bw = bw;
2367 		goto compensate;
2368 
2369 	}
2370 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2371 		/* Averaging is done, we can return the value */
2372 		bw = rack->r_ctl.gp_bw;
2373 	} else {
2374 		/* Still doing initial average must calculate */
2375 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2376 	}
2377 	lt_bw = rack_get_lt_bw(rack);
2378 	if (lt_bw == 0) {
2379 		/* If we don't have one then equate it to the gp_bw */
2380 		lt_bw = rack->r_ctl.gp_bw;
2381 	}
2382 	if ((rack->r_cwnd_was_clamped == 1) && (rack->r_clamped_gets_lower > 0)){
2383 		/*  if clamped take the lowest */
2384 		if (lt_bw < bw)
2385 			ret_bw = lt_bw;
2386 		else
2387 			ret_bw = bw;
2388 	} else {
2389 		/* If not set for clamped to get lowest, take the highest */
2390 		if (lt_bw > bw)
2391 			ret_bw = lt_bw;
2392 		else
2393 			ret_bw = bw;
2394 	}
2395 	/*
2396 	 * Now lets compensate based on the TCP/IP overhead. Our
2397 	 * Goodput estimate does not include this so we must pace out
2398 	 * a bit faster since our pacing calculations do. The pacing
2399 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2400 	 * we are using to do this, so we do that here in the opposite
2401 	 * direction as well. This means that if we are tunneled and the
2402 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2403 	 * compensated for in the pacing time the opposite way.
2404 	 */
2405 compensate:
2406 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2407 	return(ret_bw);
2408 }
2409 
2410 
2411 static uint64_t
2412 rack_get_bw(struct tcp_rack *rack)
2413 {
2414 	uint64_t bw;
2415 
2416 	if (rack->use_fixed_rate) {
2417 		/* Return the fixed pacing rate */
2418 		return (rack_get_fixed_pacing_bw(rack));
2419 	}
2420 	bw = rack_get_gp_est(rack);
2421 	return (bw);
2422 }
2423 
2424 static uint16_t
2425 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2426 {
2427 	if (rack->use_fixed_rate) {
2428 		return (100);
2429 	} else if (rack->in_probe_rtt && (rsm == NULL))
2430 		return (rack->r_ctl.rack_per_of_gp_probertt);
2431 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2432 		  rack->r_ctl.rack_per_of_gp_rec)) {
2433 		if (rsm) {
2434 			/* a retransmission always use the recovery rate */
2435 			return (rack->r_ctl.rack_per_of_gp_rec);
2436 		} else if (rack->rack_rec_nonrxt_use_cr) {
2437 			/* Directed to use the configured rate */
2438 			goto configured_rate;
2439 		} else if (rack->rack_no_prr &&
2440 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2441 			/* No PRR, lets just use the b/w estimate only */
2442 			return (100);
2443 		} else {
2444 			/*
2445 			 * Here we may have a non-retransmit but we
2446 			 * have no overrides, so just use the recovery
2447 			 * rate (prr is in effect).
2448 			 */
2449 			return (rack->r_ctl.rack_per_of_gp_rec);
2450 		}
2451 	}
2452 configured_rate:
2453 	/* For the configured rate we look at our cwnd vs the ssthresh */
2454 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2455 		return (rack->r_ctl.rack_per_of_gp_ss);
2456 	else
2457 		return (rack->r_ctl.rack_per_of_gp_ca);
2458 }
2459 
2460 static void
2461 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2462 {
2463 	/*
2464 	 * Types of logs (mod value)
2465 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2466 	 * 2 = a dsack round begins, persist is reset to 16.
2467 	 * 3 = a dsack round ends
2468 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2469 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2470 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2471 	 */
2472 	if (tcp_bblogging_on(rack->rc_tp)) {
2473 		union tcp_log_stackspecific log;
2474 		struct timeval tv;
2475 
2476 		memset(&log, 0, sizeof(log));
2477 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2478 		log.u_bbr.flex1 <<= 1;
2479 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2480 		log.u_bbr.flex1 <<= 1;
2481 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2482 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2483 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2484 		log.u_bbr.flex4 = flex4;
2485 		log.u_bbr.flex5 = flex5;
2486 		log.u_bbr.flex6 = flex6;
2487 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2488 		log.u_bbr.flex8 = mod;
2489 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2490 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2491 		    &rack->rc_inp->inp_socket->so_rcv,
2492 		    &rack->rc_inp->inp_socket->so_snd,
2493 		    RACK_DSACK_HANDLING, 0,
2494 		    0, &log, false, &tv);
2495 	}
2496 }
2497 
2498 static void
2499 rack_log_hdwr_pacing(struct tcp_rack *rack,
2500 		     uint64_t rate, uint64_t hw_rate, int line,
2501 		     int error, uint16_t mod)
2502 {
2503 	if (tcp_bblogging_on(rack->rc_tp)) {
2504 		union tcp_log_stackspecific log;
2505 		struct timeval tv;
2506 		const struct ifnet *ifp;
2507 
2508 		memset(&log, 0, sizeof(log));
2509 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2510 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2511 		if (rack->r_ctl.crte) {
2512 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2513 		} else if (rack->rc_inp->inp_route.ro_nh &&
2514 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2515 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2516 		} else
2517 			ifp = NULL;
2518 		if (ifp) {
2519 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2520 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2521 		}
2522 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2523 		log.u_bbr.bw_inuse = rate;
2524 		log.u_bbr.flex5 = line;
2525 		log.u_bbr.flex6 = error;
2526 		log.u_bbr.flex7 = mod;
2527 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2528 		log.u_bbr.flex8 = rack->use_fixed_rate;
2529 		log.u_bbr.flex8 <<= 1;
2530 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2531 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2532 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2533 		if (rack->r_ctl.crte)
2534 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2535 		else
2536 			log.u_bbr.cur_del_rate = 0;
2537 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2538 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2539 		    &rack->rc_inp->inp_socket->so_rcv,
2540 		    &rack->rc_inp->inp_socket->so_snd,
2541 		    BBR_LOG_HDWR_PACE, 0,
2542 		    0, &log, false, &tv);
2543 	}
2544 }
2545 
2546 static uint64_t
2547 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2548 {
2549 	/*
2550 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2551 	 */
2552 	uint64_t bw_est, high_rate;
2553 	uint64_t gain;
2554 
2555 	if ((rack->r_pacing_discount == 0) ||
2556 	    (rack_full_buffer_discount == 0)) {
2557 		/*
2558 		 * No buffer level based discount from client buffer
2559 		 * level is enabled or the feature is disabled.
2560 		 */
2561 		gain = (uint64_t)rack_get_output_gain(rack, rsm);
2562 		bw_est = bw * gain;
2563 		bw_est /= (uint64_t)100;
2564 	} else {
2565 		/*
2566 		 * We have a discount in place apply it with
2567 		 * just a 100% gain (we get no boost if the buffer
2568 		 * is full).
2569 		 */
2570 		uint64_t discount;
2571 
2572 		discount = bw * (uint64_t)(rack_full_buffer_discount * rack->r_ctl.pacing_discount_amm);
2573 		discount /= 100;
2574 		/* What %% of the b/w do we discount */
2575 		bw_est = bw - discount;
2576 	}
2577 	/* Never fall below the minimum (def 64kbps) */
2578 	if (bw_est < RACK_MIN_BW)
2579 		bw_est = RACK_MIN_BW;
2580 	if (rack->r_rack_hw_rate_caps) {
2581 		/* Rate caps are in place */
2582 		if (rack->r_ctl.crte != NULL) {
2583 			/* We have a hdwr rate already */
2584 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2585 			if (bw_est >= high_rate) {
2586 				/* We are capping bw at the highest rate table entry */
2587 				if (rack_hw_rate_cap_per &&
2588 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2589 					rack->r_rack_hw_rate_caps = 0;
2590 					goto done;
2591 				}
2592 				rack_log_hdwr_pacing(rack,
2593 						     bw_est, high_rate, __LINE__,
2594 						     0, 3);
2595 				bw_est = high_rate;
2596 				if (capped)
2597 					*capped = 1;
2598 			}
2599 		} else if ((rack->rack_hdrw_pacing == 0) &&
2600 			   (rack->rack_hdw_pace_ena) &&
2601 			   (rack->rack_attempt_hdwr_pace == 0) &&
2602 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2603 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2604 			/*
2605 			 * Special case, we have not yet attempted hardware
2606 			 * pacing, and yet we may, when we do, find out if we are
2607 			 * above the highest rate. We need to know the maxbw for the interface
2608 			 * in question (if it supports ratelimiting). We get back
2609 			 * a 0, if the interface is not found in the RL lists.
2610 			 */
2611 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2612 			if (high_rate) {
2613 				/* Yep, we have a rate is it above this rate? */
2614 				if (bw_est > high_rate) {
2615 					bw_est = high_rate;
2616 					if (capped)
2617 						*capped = 1;
2618 				}
2619 			}
2620 		}
2621 	}
2622 done:
2623 	return (bw_est);
2624 }
2625 
2626 static void
2627 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2628 {
2629 	if (tcp_bblogging_on(rack->rc_tp)) {
2630 		union tcp_log_stackspecific log;
2631 		struct timeval tv;
2632 
2633 		if (rack->sack_attack_disable > 0)
2634 			goto log_anyway;
2635 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2636 			/*
2637 			 * We get 3 values currently for mod
2638 			 * 1 - We are retransmitting and this tells the reason.
2639 			 * 2 - We are clearing a dup-ack count.
2640 			 * 3 - We are incrementing a dup-ack count.
2641 			 *
2642 			 * The clear/increment are only logged
2643 			 * if you have BBverbose on.
2644 			 */
2645 			return;
2646 		}
2647 log_anyway:
2648 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2649 		log.u_bbr.flex1 = tsused;
2650 		log.u_bbr.flex2 = thresh;
2651 		log.u_bbr.flex3 = rsm->r_flags;
2652 		log.u_bbr.flex4 = rsm->r_dupack;
2653 		log.u_bbr.flex5 = rsm->r_start;
2654 		log.u_bbr.flex6 = rsm->r_end;
2655 		log.u_bbr.flex8 = mod;
2656 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2657 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2658 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2659 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2660 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2661 		log.u_bbr.pacing_gain = rack->r_must_retran;
2662 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2663 		    &rack->rc_inp->inp_socket->so_rcv,
2664 		    &rack->rc_inp->inp_socket->so_snd,
2665 		    BBR_LOG_SETTINGS_CHG, 0,
2666 		    0, &log, false, &tv);
2667 	}
2668 }
2669 
2670 static void
2671 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2672 {
2673 	if (tcp_bblogging_on(rack->rc_tp)) {
2674 		union tcp_log_stackspecific log;
2675 		struct timeval tv;
2676 
2677 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2678 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2679 		log.u_bbr.flex2 = to;
2680 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2681 		log.u_bbr.flex4 = slot;
2682 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2683 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2684 		log.u_bbr.flex7 = rack->rc_in_persist;
2685 		log.u_bbr.flex8 = which;
2686 		if (rack->rack_no_prr)
2687 			log.u_bbr.pkts_out = 0;
2688 		else
2689 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2690 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2691 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2692 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2693 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2694 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2695 		log.u_bbr.pacing_gain = rack->r_must_retran;
2696 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2697 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2698 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2699 		log.u_bbr.lost = rack_rto_min;
2700 		log.u_bbr.epoch = rack->r_ctl.roundends;
2701 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2702 		    &rack->rc_inp->inp_socket->so_rcv,
2703 		    &rack->rc_inp->inp_socket->so_snd,
2704 		    BBR_LOG_TIMERSTAR, 0,
2705 		    0, &log, false, &tv);
2706 	}
2707 }
2708 
2709 static void
2710 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2711 {
2712 	if (tcp_bblogging_on(rack->rc_tp)) {
2713 		union tcp_log_stackspecific log;
2714 		struct timeval tv;
2715 
2716 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2717 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2718 		log.u_bbr.flex8 = to_num;
2719 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2720 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2721 		if (rsm == NULL)
2722 			log.u_bbr.flex3 = 0;
2723 		else
2724 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2725 		if (rack->rack_no_prr)
2726 			log.u_bbr.flex5 = 0;
2727 		else
2728 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2729 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2730 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2731 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2732 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2733 		log.u_bbr.pacing_gain = rack->r_must_retran;
2734 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2735 		    &rack->rc_inp->inp_socket->so_rcv,
2736 		    &rack->rc_inp->inp_socket->so_snd,
2737 		    BBR_LOG_RTO, 0,
2738 		    0, &log, false, &tv);
2739 	}
2740 }
2741 
2742 static void
2743 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2744 		 struct rack_sendmap *prev,
2745 		 struct rack_sendmap *rsm,
2746 		 struct rack_sendmap *next,
2747 		 int flag, uint32_t th_ack, int line)
2748 {
2749 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2750 		union tcp_log_stackspecific log;
2751 		struct timeval tv;
2752 
2753 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2754 		log.u_bbr.flex8 = flag;
2755 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2756 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2757 		log.u_bbr.delRate = (uint64_t)rsm;
2758 		log.u_bbr.rttProp = (uint64_t)next;
2759 		log.u_bbr.flex7 = 0;
2760 		if (prev) {
2761 			log.u_bbr.flex1 = prev->r_start;
2762 			log.u_bbr.flex2 = prev->r_end;
2763 			log.u_bbr.flex7 |= 0x4;
2764 		}
2765 		if (rsm) {
2766 			log.u_bbr.flex3 = rsm->r_start;
2767 			log.u_bbr.flex4 = rsm->r_end;
2768 			log.u_bbr.flex7 |= 0x2;
2769 		}
2770 		if (next) {
2771 			log.u_bbr.flex5 = next->r_start;
2772 			log.u_bbr.flex6 = next->r_end;
2773 			log.u_bbr.flex7 |= 0x1;
2774 		}
2775 		log.u_bbr.applimited = line;
2776 		log.u_bbr.pkts_out = th_ack;
2777 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2778 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2779 		if (rack->rack_no_prr)
2780 			log.u_bbr.lost = 0;
2781 		else
2782 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2783 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2784 		    &rack->rc_inp->inp_socket->so_rcv,
2785 		    &rack->rc_inp->inp_socket->so_snd,
2786 		    TCP_LOG_MAPCHG, 0,
2787 		    0, &log, false, &tv);
2788 	}
2789 }
2790 
2791 static void
2792 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2793 		 struct rack_sendmap *rsm, int conf)
2794 {
2795 	if (tcp_bblogging_on(tp)) {
2796 		union tcp_log_stackspecific log;
2797 		struct timeval tv;
2798 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2799 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2800 		log.u_bbr.flex1 = t;
2801 		log.u_bbr.flex2 = len;
2802 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2803 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2804 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2805 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2806 		log.u_bbr.flex7 = conf;
2807 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2808 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2809 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2810 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2811 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2812 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2813 		if (rsm) {
2814 			log.u_bbr.pkt_epoch = rsm->r_start;
2815 			log.u_bbr.lost = rsm->r_end;
2816 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2817 			/* We loose any upper of the 24 bits */
2818 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2819 		} else {
2820 			/* Its a SYN */
2821 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2822 			log.u_bbr.lost = 0;
2823 			log.u_bbr.cwnd_gain = 0;
2824 			log.u_bbr.pacing_gain = 0;
2825 		}
2826 		/* Write out general bits of interest rrs here */
2827 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2828 		log.u_bbr.use_lt_bw <<= 1;
2829 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2830 		log.u_bbr.use_lt_bw <<= 1;
2831 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2832 		log.u_bbr.use_lt_bw <<= 1;
2833 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2834 		log.u_bbr.use_lt_bw <<= 1;
2835 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2836 		log.u_bbr.use_lt_bw <<= 1;
2837 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2838 		log.u_bbr.use_lt_bw <<= 1;
2839 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2840 		log.u_bbr.use_lt_bw <<= 1;
2841 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2842 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2843 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2844 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2845 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2846 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2847 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2848 		log.u_bbr.bw_inuse <<= 32;
2849 		if (rsm)
2850 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2851 		TCP_LOG_EVENTP(tp, NULL,
2852 		    &rack->rc_inp->inp_socket->so_rcv,
2853 		    &rack->rc_inp->inp_socket->so_snd,
2854 		    BBR_LOG_BBRRTT, 0,
2855 		    0, &log, false, &tv);
2856 
2857 
2858 	}
2859 }
2860 
2861 static void
2862 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2863 {
2864 	/*
2865 	 * Log the rtt sample we are
2866 	 * applying to the srtt algorithm in
2867 	 * useconds.
2868 	 */
2869 	if (tcp_bblogging_on(rack->rc_tp)) {
2870 		union tcp_log_stackspecific log;
2871 		struct timeval tv;
2872 
2873 		/* Convert our ms to a microsecond */
2874 		memset(&log, 0, sizeof(log));
2875 		log.u_bbr.flex1 = rtt;
2876 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2877 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2878 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2879 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2880 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2881 		log.u_bbr.flex7 = 1;
2882 		log.u_bbr.flex8 = rack->sack_attack_disable;
2883 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2884 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2885 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2886 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2887 		log.u_bbr.pacing_gain = rack->r_must_retran;
2888 		/*
2889 		 * We capture in delRate the upper 32 bits as
2890 		 * the confidence level we had declared, and the
2891 		 * lower 32 bits as the actual RTT using the arrival
2892 		 * timestamp.
2893 		 */
2894 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2895 		log.u_bbr.delRate <<= 32;
2896 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2897 		/* Lets capture all the things that make up t_rtxcur */
2898 		log.u_bbr.applimited = rack_rto_min;
2899 		log.u_bbr.epoch = rack_rto_max;
2900 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2901 		log.u_bbr.lost = rack_rto_min;
2902 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2903 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2904 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2905 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2906 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2907 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2908 		    &rack->rc_inp->inp_socket->so_rcv,
2909 		    &rack->rc_inp->inp_socket->so_snd,
2910 		    TCP_LOG_RTT, 0,
2911 		    0, &log, false, &tv);
2912 	}
2913 }
2914 
2915 static void
2916 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2917 {
2918 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2919 		union tcp_log_stackspecific log;
2920 		struct timeval tv;
2921 
2922 		/* Convert our ms to a microsecond */
2923 		memset(&log, 0, sizeof(log));
2924 		log.u_bbr.flex1 = rtt;
2925 		log.u_bbr.flex2 = send_time;
2926 		log.u_bbr.flex3 = ack_time;
2927 		log.u_bbr.flex4 = where;
2928 		log.u_bbr.flex7 = 2;
2929 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2930 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2931 		    &rack->rc_inp->inp_socket->so_rcv,
2932 		    &rack->rc_inp->inp_socket->so_snd,
2933 		    TCP_LOG_RTT, 0,
2934 		    0, &log, false, &tv);
2935 	}
2936 }
2937 
2938 
2939 static void
2940 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2941 {
2942 	if (tcp_bblogging_on(rack->rc_tp)) {
2943 		union tcp_log_stackspecific log;
2944 		struct timeval tv;
2945 
2946 		/* Convert our ms to a microsecond */
2947 		memset(&log, 0, sizeof(log));
2948 		log.u_bbr.flex1 = idx;
2949 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2950 		log.u_bbr.flex3 = tsecho;
2951 		log.u_bbr.flex7 = 3;
2952 		log.u_bbr.rttProp = tsv;
2953 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2954 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2955 		    &rack->rc_inp->inp_socket->so_rcv,
2956 		    &rack->rc_inp->inp_socket->so_snd,
2957 		    TCP_LOG_RTT, 0,
2958 		    0, &log, false, &tv);
2959 	}
2960 }
2961 
2962 
2963 static inline void
2964 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2965 {
2966 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2967 		union tcp_log_stackspecific log;
2968 		struct timeval tv;
2969 
2970 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2971 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2972 		log.u_bbr.flex1 = line;
2973 		log.u_bbr.flex2 = tick;
2974 		log.u_bbr.flex3 = tp->t_maxunacktime;
2975 		log.u_bbr.flex4 = tp->t_acktime;
2976 		log.u_bbr.flex8 = event;
2977 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2978 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2979 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2980 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2981 		log.u_bbr.pacing_gain = rack->r_must_retran;
2982 		TCP_LOG_EVENTP(tp, NULL,
2983 		    &rack->rc_inp->inp_socket->so_rcv,
2984 		    &rack->rc_inp->inp_socket->so_snd,
2985 		    BBR_LOG_PROGRESS, 0,
2986 		    0, &log, false, &tv);
2987 	}
2988 }
2989 
2990 static void
2991 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
2992 {
2993 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2994 		union tcp_log_stackspecific log;
2995 
2996 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2997 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2998 		log.u_bbr.flex1 = slot;
2999 		if (rack->rack_no_prr)
3000 			log.u_bbr.flex2 = 0;
3001 		else
3002 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3003 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3004 		log.u_bbr.flex5 = rack->r_ctl.ack_during_sd;
3005 		log.u_bbr.flex6 = line;
3006 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3007 		log.u_bbr.flex8 = rack->rc_in_persist;
3008 		log.u_bbr.timeStamp = cts;
3009 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3010 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3011 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3012 		log.u_bbr.pacing_gain = rack->r_must_retran;
3013 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3014 		    &rack->rc_inp->inp_socket->so_rcv,
3015 		    &rack->rc_inp->inp_socket->so_snd,
3016 		    BBR_LOG_BBRSND, 0,
3017 		    0, &log, false, tv);
3018 	}
3019 }
3020 
3021 static void
3022 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3023 {
3024 	if (tcp_bblogging_on(rack->rc_tp)) {
3025 		union tcp_log_stackspecific log;
3026 		struct timeval tv;
3027 
3028 		memset(&log, 0, sizeof(log));
3029 		log.u_bbr.flex1 = did_out;
3030 		log.u_bbr.flex2 = nxt_pkt;
3031 		log.u_bbr.flex3 = way_out;
3032 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3033 		if (rack->rack_no_prr)
3034 			log.u_bbr.flex5 = 0;
3035 		else
3036 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3037 		log.u_bbr.flex6 = nsegs;
3038 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3039 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3040 		log.u_bbr.flex7 <<= 1;
3041 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3042 		log.u_bbr.flex7 <<= 1;
3043 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3044 		log.u_bbr.flex8 = rack->rc_in_persist;
3045 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3046 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3047 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3048 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3049 		log.u_bbr.use_lt_bw <<= 1;
3050 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3051 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3052 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3053 		log.u_bbr.pacing_gain = rack->r_must_retran;
3054 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3055 		    &rack->rc_inp->inp_socket->so_rcv,
3056 		    &rack->rc_inp->inp_socket->so_snd,
3057 		    BBR_LOG_DOSEG_DONE, 0,
3058 		    0, &log, false, &tv);
3059 	}
3060 }
3061 
3062 static void
3063 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3064 {
3065 	if (tcp_bblogging_on(rack->rc_tp)) {
3066 		union tcp_log_stackspecific log;
3067 		struct timeval tv;
3068 
3069 		memset(&log, 0, sizeof(log));
3070 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3071 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3072 		log.u_bbr.flex4 = arg1;
3073 		log.u_bbr.flex5 = arg2;
3074 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3075 		log.u_bbr.flex6 = arg3;
3076 		log.u_bbr.flex8 = frm;
3077 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3078 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3079 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3080 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3081 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3082 		log.u_bbr.pacing_gain = rack->r_must_retran;
3083 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3084 		    &tptosocket(tp)->so_snd,
3085 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3086 	}
3087 }
3088 
3089 static void
3090 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3091 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3092 {
3093 	if (tcp_bblogging_on(rack->rc_tp)) {
3094 		union tcp_log_stackspecific log;
3095 		struct timeval tv;
3096 
3097 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3098 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3099 		log.u_bbr.flex1 = slot;
3100 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3101 		log.u_bbr.flex4 = reason;
3102 		if (rack->rack_no_prr)
3103 			log.u_bbr.flex5 = 0;
3104 		else
3105 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3106 		log.u_bbr.flex7 = hpts_calling;
3107 		log.u_bbr.flex8 = rack->rc_in_persist;
3108 		log.u_bbr.lt_epoch = cwnd_to_use;
3109 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3110 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3111 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3112 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3113 		log.u_bbr.pacing_gain = rack->r_must_retran;
3114 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3115 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3116 		    &rack->rc_inp->inp_socket->so_rcv,
3117 		    &rack->rc_inp->inp_socket->so_snd,
3118 		    BBR_LOG_JUSTRET, 0,
3119 		    tlen, &log, false, &tv);
3120 	}
3121 }
3122 
3123 static void
3124 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3125 		   struct timeval *tv, uint32_t flags_on_entry)
3126 {
3127 	if (tcp_bblogging_on(rack->rc_tp)) {
3128 		union tcp_log_stackspecific log;
3129 
3130 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3131 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3132 		log.u_bbr.flex1 = line;
3133 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3134 		log.u_bbr.flex3 = flags_on_entry;
3135 		log.u_bbr.flex4 = us_cts;
3136 		if (rack->rack_no_prr)
3137 			log.u_bbr.flex5 = 0;
3138 		else
3139 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3140 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3141 		log.u_bbr.flex7 = hpts_removed;
3142 		log.u_bbr.flex8 = 1;
3143 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3144 		log.u_bbr.timeStamp = us_cts;
3145 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3146 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3147 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3148 		log.u_bbr.pacing_gain = rack->r_must_retran;
3149 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3150 		    &rack->rc_inp->inp_socket->so_rcv,
3151 		    &rack->rc_inp->inp_socket->so_snd,
3152 		    BBR_LOG_TIMERCANC, 0,
3153 		    0, &log, false, tv);
3154 	}
3155 }
3156 
3157 static void
3158 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3159 			  uint32_t flex1, uint32_t flex2,
3160 			  uint32_t flex3, uint32_t flex4,
3161 			  uint32_t flex5, uint32_t flex6,
3162 			  uint16_t flex7, uint8_t mod)
3163 {
3164 	if (tcp_bblogging_on(rack->rc_tp)) {
3165 		union tcp_log_stackspecific log;
3166 		struct timeval tv;
3167 
3168 		if (mod == 1) {
3169 			/* No you can't use 1, its for the real to cancel */
3170 			return;
3171 		}
3172 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3173 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3174 		log.u_bbr.flex1 = flex1;
3175 		log.u_bbr.flex2 = flex2;
3176 		log.u_bbr.flex3 = flex3;
3177 		log.u_bbr.flex4 = flex4;
3178 		log.u_bbr.flex5 = flex5;
3179 		log.u_bbr.flex6 = flex6;
3180 		log.u_bbr.flex7 = flex7;
3181 		log.u_bbr.flex8 = mod;
3182 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3183 		    &rack->rc_inp->inp_socket->so_rcv,
3184 		    &rack->rc_inp->inp_socket->so_snd,
3185 		    BBR_LOG_TIMERCANC, 0,
3186 		    0, &log, false, &tv);
3187 	}
3188 }
3189 
3190 static void
3191 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3192 {
3193 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3194 		union tcp_log_stackspecific log;
3195 		struct timeval tv;
3196 
3197 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3198 		log.u_bbr.flex1 = timers;
3199 		log.u_bbr.flex2 = ret;
3200 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3201 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3202 		log.u_bbr.flex5 = cts;
3203 		if (rack->rack_no_prr)
3204 			log.u_bbr.flex6 = 0;
3205 		else
3206 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3207 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3208 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3209 		log.u_bbr.pacing_gain = rack->r_must_retran;
3210 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3211 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3212 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3213 		    &rack->rc_inp->inp_socket->so_rcv,
3214 		    &rack->rc_inp->inp_socket->so_snd,
3215 		    BBR_LOG_TO_PROCESS, 0,
3216 		    0, &log, false, &tv);
3217 	}
3218 }
3219 
3220 static void
3221 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3222 {
3223 	if (tcp_bblogging_on(rack->rc_tp)) {
3224 		union tcp_log_stackspecific log;
3225 		struct timeval tv;
3226 
3227 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3228 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3229 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3230 		if (rack->rack_no_prr)
3231 			log.u_bbr.flex3 = 0;
3232 		else
3233 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3234 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3235 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3236 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3237 		log.u_bbr.flex7 = line;
3238 		log.u_bbr.flex8 = frm;
3239 		log.u_bbr.pkts_out = orig_cwnd;
3240 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3241 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3242 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3243 		log.u_bbr.use_lt_bw <<= 1;
3244 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3245 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3246 		    &rack->rc_inp->inp_socket->so_rcv,
3247 		    &rack->rc_inp->inp_socket->so_snd,
3248 		    BBR_LOG_BBRUPD, 0,
3249 		    0, &log, false, &tv);
3250 	}
3251 }
3252 
3253 #ifdef TCP_SAD_DETECTION
3254 static void
3255 rack_log_sad(struct tcp_rack *rack, int event)
3256 {
3257 	if (tcp_bblogging_on(rack->rc_tp)) {
3258 		union tcp_log_stackspecific log;
3259 		struct timeval tv;
3260 
3261 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3262 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
3263 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
3264 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
3265 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
3266 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
3267 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
3268 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
3269 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
3270 		log.u_bbr.lt_epoch |= rack->do_detection;
3271 		log.u_bbr.applimited = tcp_map_minimum;
3272 		log.u_bbr.flex7 = rack->sack_attack_disable;
3273 		log.u_bbr.flex8 = event;
3274 		log.u_bbr.bbr_state = rack->rc_suspicious;
3275 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3276 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3277 		log.u_bbr.delivered = tcp_sad_decay_val;
3278 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3279 		    &rack->rc_inp->inp_socket->so_rcv,
3280 		    &rack->rc_inp->inp_socket->so_snd,
3281 		    TCP_SAD_DETECT, 0,
3282 		    0, &log, false, &tv);
3283 	}
3284 }
3285 #endif
3286 
3287 static void
3288 rack_counter_destroy(void)
3289 {
3290 	counter_u64_free(rack_total_bytes);
3291 	counter_u64_free(rack_fto_send);
3292 	counter_u64_free(rack_fto_rsm_send);
3293 	counter_u64_free(rack_nfto_resend);
3294 	counter_u64_free(rack_hw_pace_init_fail);
3295 	counter_u64_free(rack_hw_pace_lost);
3296 	counter_u64_free(rack_non_fto_send);
3297 	counter_u64_free(rack_extended_rfo);
3298 	counter_u64_free(rack_ack_total);
3299 	counter_u64_free(rack_express_sack);
3300 	counter_u64_free(rack_sack_total);
3301 	counter_u64_free(rack_move_none);
3302 	counter_u64_free(rack_move_some);
3303 	counter_u64_free(rack_sack_attacks_detected);
3304 	counter_u64_free(rack_sack_attacks_reversed);
3305 	counter_u64_free(rack_sack_attacks_suspect);
3306 	counter_u64_free(rack_sack_used_next_merge);
3307 	counter_u64_free(rack_sack_used_prev_merge);
3308 	counter_u64_free(rack_tlp_tot);
3309 	counter_u64_free(rack_tlp_newdata);
3310 	counter_u64_free(rack_tlp_retran);
3311 	counter_u64_free(rack_tlp_retran_bytes);
3312 	counter_u64_free(rack_to_tot);
3313 	counter_u64_free(rack_saw_enobuf);
3314 	counter_u64_free(rack_saw_enobuf_hw);
3315 	counter_u64_free(rack_saw_enetunreach);
3316 	counter_u64_free(rack_hot_alloc);
3317 	counter_u64_free(rack_to_alloc);
3318 	counter_u64_free(rack_to_alloc_hard);
3319 	counter_u64_free(rack_to_alloc_emerg);
3320 	counter_u64_free(rack_to_alloc_limited);
3321 	counter_u64_free(rack_alloc_limited_conns);
3322 	counter_u64_free(rack_split_limited);
3323 	counter_u64_free(rack_multi_single_eq);
3324 	counter_u64_free(rack_rxt_clamps_cwnd);
3325 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3326 	counter_u64_free(rack_proc_non_comp_ack);
3327 	counter_u64_free(rack_sack_proc_all);
3328 	counter_u64_free(rack_sack_proc_restart);
3329 	counter_u64_free(rack_sack_proc_short);
3330 	counter_u64_free(rack_sack_skipped_acked);
3331 	counter_u64_free(rack_sack_splits);
3332 	counter_u64_free(rack_input_idle_reduces);
3333 	counter_u64_free(rack_collapsed_win);
3334 	counter_u64_free(rack_collapsed_win_rxt);
3335 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3336 	counter_u64_free(rack_collapsed_win_seen);
3337 	counter_u64_free(rack_try_scwnd);
3338 	counter_u64_free(rack_persists_sends);
3339 	counter_u64_free(rack_persists_acks);
3340 	counter_u64_free(rack_persists_loss);
3341 	counter_u64_free(rack_persists_lost_ends);
3342 #ifdef INVARIANTS
3343 	counter_u64_free(rack_adjust_map_bw);
3344 #endif
3345 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3346 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3347 }
3348 
3349 static struct rack_sendmap *
3350 rack_alloc(struct tcp_rack *rack)
3351 {
3352 	struct rack_sendmap *rsm;
3353 
3354 	/*
3355 	 * First get the top of the list it in
3356 	 * theory is the "hottest" rsm we have,
3357 	 * possibly just freed by ack processing.
3358 	 */
3359 	if (rack->rc_free_cnt > rack_free_cache) {
3360 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3361 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3362 		counter_u64_add(rack_hot_alloc, 1);
3363 		rack->rc_free_cnt--;
3364 		return (rsm);
3365 	}
3366 	/*
3367 	 * Once we get under our free cache we probably
3368 	 * no longer have a "hot" one available. Lets
3369 	 * get one from UMA.
3370 	 */
3371 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3372 	if (rsm) {
3373 		rack->r_ctl.rc_num_maps_alloced++;
3374 		counter_u64_add(rack_to_alloc, 1);
3375 		return (rsm);
3376 	}
3377 	/*
3378 	 * Dig in to our aux rsm's (the last two) since
3379 	 * UMA failed to get us one.
3380 	 */
3381 	if (rack->rc_free_cnt) {
3382 		counter_u64_add(rack_to_alloc_emerg, 1);
3383 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3384 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3385 		rack->rc_free_cnt--;
3386 		return (rsm);
3387 	}
3388 	return (NULL);
3389 }
3390 
3391 static struct rack_sendmap *
3392 rack_alloc_full_limit(struct tcp_rack *rack)
3393 {
3394 	if ((V_tcp_map_entries_limit > 0) &&
3395 	    (rack->do_detection == 0) &&
3396 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3397 		counter_u64_add(rack_to_alloc_limited, 1);
3398 		if (!rack->alloc_limit_reported) {
3399 			rack->alloc_limit_reported = 1;
3400 			counter_u64_add(rack_alloc_limited_conns, 1);
3401 		}
3402 		return (NULL);
3403 	}
3404 	return (rack_alloc(rack));
3405 }
3406 
3407 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3408 static struct rack_sendmap *
3409 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3410 {
3411 	struct rack_sendmap *rsm;
3412 
3413 	if (limit_type) {
3414 		/* currently there is only one limit type */
3415 		if (rack->r_ctl.rc_split_limit > 0 &&
3416 		    (rack->do_detection == 0) &&
3417 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3418 			counter_u64_add(rack_split_limited, 1);
3419 			if (!rack->alloc_limit_reported) {
3420 				rack->alloc_limit_reported = 1;
3421 				counter_u64_add(rack_alloc_limited_conns, 1);
3422 			}
3423 			return (NULL);
3424 #ifdef TCP_SAD_DETECTION
3425 		} else if ((tcp_sad_limit != 0) &&
3426 			   (rack->do_detection == 1) &&
3427 			   (rack->r_ctl.rc_num_split_allocs >= tcp_sad_limit)) {
3428 			counter_u64_add(rack_split_limited, 1);
3429 			if (!rack->alloc_limit_reported) {
3430 				rack->alloc_limit_reported = 1;
3431 				counter_u64_add(rack_alloc_limited_conns, 1);
3432 			}
3433 			return (NULL);
3434 #endif
3435 		}
3436 	}
3437 
3438 	/* allocate and mark in the limit type, if set */
3439 	rsm = rack_alloc(rack);
3440 	if (rsm != NULL && limit_type) {
3441 		rsm->r_limit_type = limit_type;
3442 		rack->r_ctl.rc_num_split_allocs++;
3443 	}
3444 	return (rsm);
3445 }
3446 
3447 static void
3448 rack_free_trim(struct tcp_rack *rack)
3449 {
3450 	struct rack_sendmap *rsm;
3451 
3452 	/*
3453 	 * Free up all the tail entries until
3454 	 * we get our list down to the limit.
3455 	 */
3456 	while (rack->rc_free_cnt > rack_free_cache) {
3457 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3458 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3459 		rack->rc_free_cnt--;
3460 		rack->r_ctl.rc_num_maps_alloced--;
3461 		uma_zfree(rack_zone, rsm);
3462 	}
3463 }
3464 
3465 static void
3466 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3467 {
3468 	if (rsm->r_flags & RACK_APP_LIMITED) {
3469 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3470 			rack->r_ctl.rc_app_limited_cnt--;
3471 		}
3472 	}
3473 	if (rsm->r_limit_type) {
3474 		/* currently there is only one limit type */
3475 		rack->r_ctl.rc_num_split_allocs--;
3476 	}
3477 	if (rsm == rack->r_ctl.rc_first_appl) {
3478 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3479 			rack->r_ctl.rc_first_appl = NULL;
3480 		else
3481 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3482 	}
3483 	if (rsm == rack->r_ctl.rc_resend)
3484 		rack->r_ctl.rc_resend = NULL;
3485 	if (rsm == rack->r_ctl.rc_end_appl)
3486 		rack->r_ctl.rc_end_appl = NULL;
3487 	if (rack->r_ctl.rc_tlpsend == rsm)
3488 		rack->r_ctl.rc_tlpsend = NULL;
3489 	if (rack->r_ctl.rc_sacklast == rsm)
3490 		rack->r_ctl.rc_sacklast = NULL;
3491 	memset(rsm, 0, sizeof(struct rack_sendmap));
3492 	/* Make sure we are not going to overrun our count limit of 0xff */
3493 	if ((rack->rc_free_cnt + 1) > 0xff) {
3494 		rack_free_trim(rack);
3495 	}
3496 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3497 	rack->rc_free_cnt++;
3498 }
3499 
3500 static uint32_t
3501 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3502 {
3503 	uint64_t srtt, bw, len, tim;
3504 	uint32_t segsiz, def_len, minl;
3505 
3506 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3507 	def_len = rack_def_data_window * segsiz;
3508 	if (rack->rc_gp_filled == 0) {
3509 		/*
3510 		 * We have no measurement (IW is in flight?) so
3511 		 * we can only guess using our data_window sysctl
3512 		 * value (usually 20MSS).
3513 		 */
3514 		return (def_len);
3515 	}
3516 	/*
3517 	 * Now we have a number of factors to consider.
3518 	 *
3519 	 * 1) We have a desired BDP which is usually
3520 	 *    at least 2.
3521 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3522 	 *    but we allow it too to be more.
3523 	 * 3) We want to make sure a measurement last N useconds (if
3524 	 *    we have set rack_min_measure_usec.
3525 	 *
3526 	 * We handle the first concern here by trying to create a data
3527 	 * window of max(rack_def_data_window, DesiredBDP). The
3528 	 * second concern we handle in not letting the measurement
3529 	 * window end normally until at least the required SRTT's
3530 	 * have gone by which is done further below in
3531 	 * rack_enough_for_measurement(). Finally the third concern
3532 	 * we also handle here by calculating how long that time
3533 	 * would take at the current BW and then return the
3534 	 * max of our first calculation and that length. Note
3535 	 * that if rack_min_measure_usec is 0, we don't deal
3536 	 * with concern 3. Also for both Concern 1 and 3 an
3537 	 * application limited period could end the measurement
3538 	 * earlier.
3539 	 *
3540 	 * So lets calculate the BDP with the "known" b/w using
3541 	 * the SRTT has our rtt and then multiply it by the
3542 	 * goal.
3543 	 */
3544 	bw = rack_get_bw(rack);
3545 	srtt = (uint64_t)tp->t_srtt;
3546 	len = bw * srtt;
3547 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3548 	len *= max(1, rack_goal_bdp);
3549 	/* Now we need to round up to the nearest MSS */
3550 	len = roundup(len, segsiz);
3551 	if (rack_min_measure_usec) {
3552 		/* Now calculate our min length for this b/w */
3553 		tim = rack_min_measure_usec;
3554 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3555 		if (minl == 0)
3556 			minl = 1;
3557 		minl = roundup(minl, segsiz);
3558 		if (len < minl)
3559 			len = minl;
3560 	}
3561 	/*
3562 	 * Now if we have a very small window we want
3563 	 * to attempt to get the window that is
3564 	 * as small as possible. This happens on
3565 	 * low b/w connections and we don't want to
3566 	 * span huge numbers of rtt's between measurements.
3567 	 *
3568 	 * We basically include 2 over our "MIN window" so
3569 	 * that the measurement can be shortened (possibly) by
3570 	 * an ack'ed packet.
3571 	 */
3572 	if (len < def_len)
3573 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3574 	else
3575 		return (max((uint32_t)len, def_len));
3576 
3577 }
3578 
3579 static int
3580 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3581 {
3582 	uint32_t tim, srtts, segsiz;
3583 
3584 	/*
3585 	 * Has enough time passed for the GP measurement to be valid?
3586 	 */
3587 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3588 		/* Not enough bytes yet */
3589 		return (0);
3590 	}
3591 	if ((tp->snd_max == tp->snd_una) ||
3592 	    (th_ack == tp->snd_max)){
3593 		/*
3594 		 * All is acked quality of all acked is
3595 		 * usually low or medium, but we in theory could split
3596 		 * all acked into two cases, where you got
3597 		 * a signifigant amount of your window and
3598 		 * where you did not. For now we leave it
3599 		 * but it is something to contemplate in the
3600 		 * future. The danger here is that delayed ack
3601 		 * is effecting the last byte (which is a 50:50 chance).
3602 		 */
3603 		*quality = RACK_QUALITY_ALLACKED;
3604 		return (1);
3605 	}
3606 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3607 		/*
3608 		 * We obtained our entire window of data we wanted
3609 		 * no matter if we are in recovery or not then
3610 		 * its ok since expanding the window does not
3611 		 * make things fuzzy (or at least not as much).
3612 		 */
3613 		*quality = RACK_QUALITY_HIGH;
3614 		return (1);
3615 	}
3616 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3617 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3618 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3619 		/* Not enough bytes yet */
3620 		return (0);
3621 	}
3622 	if (rack->r_ctl.rc_first_appl &&
3623 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3624 		/*
3625 		 * We are up to the app limited send point
3626 		 * we have to measure irrespective of the time..
3627 		 */
3628 		*quality = RACK_QUALITY_APPLIMITED;
3629 		return (1);
3630 	}
3631 	/* Now what about time? */
3632 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3633 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3634 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3635 		/*
3636 		 * We do not allow a measurement if we are in recovery
3637 		 * that would shrink the goodput window we wanted.
3638 		 * This is to prevent cloudyness of when the last send
3639 		 * was actually made.
3640 		 */
3641 		*quality = RACK_QUALITY_HIGH;
3642 		return (1);
3643 	}
3644 	/* Nope not even a full SRTT has passed */
3645 	return (0);
3646 }
3647 
3648 static void
3649 rack_log_timely(struct tcp_rack *rack,
3650 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3651 		uint64_t up_bnd, int line, uint8_t method)
3652 {
3653 	if (tcp_bblogging_on(rack->rc_tp)) {
3654 		union tcp_log_stackspecific log;
3655 		struct timeval tv;
3656 
3657 		memset(&log, 0, sizeof(log));
3658 		log.u_bbr.flex1 = logged;
3659 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3660 		log.u_bbr.flex2 <<= 4;
3661 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3662 		log.u_bbr.flex2 <<= 4;
3663 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3664 		log.u_bbr.flex2 <<= 4;
3665 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3666 		log.u_bbr.flex3 = rack->rc_gp_incr;
3667 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3668 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3669 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3670 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3671 		log.u_bbr.flex8 = method;
3672 		log.u_bbr.cur_del_rate = cur_bw;
3673 		log.u_bbr.delRate = low_bnd;
3674 		log.u_bbr.bw_inuse = up_bnd;
3675 		log.u_bbr.rttProp = rack_get_bw(rack);
3676 		log.u_bbr.pkt_epoch = line;
3677 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3678 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3679 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3680 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3681 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3682 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3683 		log.u_bbr.cwnd_gain <<= 1;
3684 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3685 		log.u_bbr.cwnd_gain <<= 1;
3686 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3687 		log.u_bbr.cwnd_gain <<= 1;
3688 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3689 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3690 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3691 		    &rack->rc_inp->inp_socket->so_rcv,
3692 		    &rack->rc_inp->inp_socket->so_snd,
3693 		    TCP_TIMELY_WORK, 0,
3694 		    0, &log, false, &tv);
3695 	}
3696 }
3697 
3698 static int
3699 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3700 {
3701 	/*
3702 	 * Before we increase we need to know if
3703 	 * the estimate just made was less than
3704 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3705 	 *
3706 	 * If we already are pacing at a fast enough
3707 	 * rate to push us faster there is no sense of
3708 	 * increasing.
3709 	 *
3710 	 * We first caculate our actual pacing rate (ss or ca multiplier
3711 	 * times our cur_bw).
3712 	 *
3713 	 * Then we take the last measured rate and multipy by our
3714 	 * maximum pacing overage to give us a max allowable rate.
3715 	 *
3716 	 * If our act_rate is smaller than our max_allowable rate
3717 	 * then we should increase. Else we should hold steady.
3718 	 *
3719 	 */
3720 	uint64_t act_rate, max_allow_rate;
3721 
3722 	if (rack_timely_no_stopping)
3723 		return (1);
3724 
3725 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3726 		/*
3727 		 * Initial startup case or
3728 		 * everything is acked case.
3729 		 */
3730 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3731 				__LINE__, 9);
3732 		return (1);
3733 	}
3734 	if (mult <= 100) {
3735 		/*
3736 		 * We can always pace at or slightly above our rate.
3737 		 */
3738 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3739 				__LINE__, 9);
3740 		return (1);
3741 	}
3742 	act_rate = cur_bw * (uint64_t)mult;
3743 	act_rate /= 100;
3744 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3745 	max_allow_rate /= 100;
3746 	if (act_rate < max_allow_rate) {
3747 		/*
3748 		 * Here the rate we are actually pacing at
3749 		 * is smaller than 10% above our last measurement.
3750 		 * This means we are pacing below what we would
3751 		 * like to try to achieve (plus some wiggle room).
3752 		 */
3753 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3754 				__LINE__, 9);
3755 		return (1);
3756 	} else {
3757 		/*
3758 		 * Here we are already pacing at least rack_max_per_above(10%)
3759 		 * what we are getting back. This indicates most likely
3760 		 * that we are being limited (cwnd/rwnd/app) and can't
3761 		 * get any more b/w. There is no sense of trying to
3762 		 * raise up the pacing rate its not speeding us up
3763 		 * and we already are pacing faster than we are getting.
3764 		 */
3765 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3766 				__LINE__, 8);
3767 		return (0);
3768 	}
3769 }
3770 
3771 static void
3772 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3773 {
3774 	/*
3775 	 * When we drag bottom, we want to assure
3776 	 * that no multiplier is below 1.0, if so
3777 	 * we want to restore it to at least that.
3778 	 */
3779 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3780 		/* This is unlikely we usually do not touch recovery */
3781 		rack->r_ctl.rack_per_of_gp_rec = 100;
3782 	}
3783 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3784 		rack->r_ctl.rack_per_of_gp_ca = 100;
3785 	}
3786 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3787 		rack->r_ctl.rack_per_of_gp_ss = 100;
3788 	}
3789 }
3790 
3791 static void
3792 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3793 {
3794 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3795 		rack->r_ctl.rack_per_of_gp_ca = 100;
3796 	}
3797 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3798 		rack->r_ctl.rack_per_of_gp_ss = 100;
3799 	}
3800 }
3801 
3802 static void
3803 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3804 {
3805 	int32_t  calc, logged, plus;
3806 
3807 	logged = 0;
3808 
3809 	if (override) {
3810 		/*
3811 		 * override is passed when we are
3812 		 * loosing b/w and making one last
3813 		 * gasp at trying to not loose out
3814 		 * to a new-reno flow.
3815 		 */
3816 		goto extra_boost;
3817 	}
3818 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3819 	if (rack->rc_gp_incr &&
3820 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3821 		/*
3822 		 * Reset and get 5 strokes more before the boost. Note
3823 		 * that the count is 0 based so we have to add one.
3824 		 */
3825 extra_boost:
3826 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3827 		rack->rc_gp_timely_inc_cnt = 0;
3828 	} else
3829 		plus = (uint32_t)rack_gp_increase_per;
3830 	/* Must be at least 1% increase for true timely increases */
3831 	if ((plus < 1) &&
3832 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3833 		plus = 1;
3834 	if (rack->rc_gp_saw_rec &&
3835 	    (rack->rc_gp_no_rec_chg == 0) &&
3836 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3837 				  rack->r_ctl.rack_per_of_gp_rec)) {
3838 		/* We have been in recovery ding it too */
3839 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3840 		if (calc > 0xffff)
3841 			calc = 0xffff;
3842 		logged |= 1;
3843 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3844 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3845 		    (rack->rc_dragged_bottom == 0) &&
3846 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3847 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3848 	}
3849 	if (rack->rc_gp_saw_ca &&
3850 	    (rack->rc_gp_saw_ss == 0) &&
3851 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3852 				  rack->r_ctl.rack_per_of_gp_ca)) {
3853 		/* In CA */
3854 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3855 		if (calc > 0xffff)
3856 			calc = 0xffff;
3857 		logged |= 2;
3858 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3859 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3860 		    (rack->rc_dragged_bottom == 0) &&
3861 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3862 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3863 	}
3864 	if (rack->rc_gp_saw_ss &&
3865 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3866 				  rack->r_ctl.rack_per_of_gp_ss)) {
3867 		/* In SS */
3868 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3869 		if (calc > 0xffff)
3870 			calc = 0xffff;
3871 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3872 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3873 		    (rack->rc_dragged_bottom == 0) &&
3874 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3875 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3876 		logged |= 4;
3877 	}
3878 	if (logged &&
3879 	    (rack->rc_gp_incr == 0)){
3880 		/* Go into increment mode */
3881 		rack->rc_gp_incr = 1;
3882 		rack->rc_gp_timely_inc_cnt = 0;
3883 	}
3884 	if (rack->rc_gp_incr &&
3885 	    logged &&
3886 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3887 		rack->rc_gp_timely_inc_cnt++;
3888 	}
3889 	rack_log_timely(rack,  logged, plus, 0, 0,
3890 			__LINE__, 1);
3891 }
3892 
3893 static uint32_t
3894 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3895 {
3896 	/*-
3897 	 * norm_grad = rtt_diff / minrtt;
3898 	 * new_per = curper * (1 - B * norm_grad)
3899 	 *
3900 	 * B = rack_gp_decrease_per (default 80%)
3901 	 * rtt_dif = input var current rtt-diff
3902 	 * curper = input var current percentage
3903 	 * minrtt = from rack filter
3904 	 *
3905 	 * In order to do the floating point calculations above we
3906 	 * do an integer conversion. The code looks confusing so let me
3907 	 * translate it into something that use more variables and
3908 	 * is clearer for us humans :)
3909 	 *
3910 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
3911 	 * uint32_t perf;
3912 	 *
3913 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3914 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3915 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3916 	 * inverse /= 1000000;
3917 	 * reduce_by = (1000000 - inverse);
3918 	 * final_result = (cur_per * reduce_by) / 1000000;
3919 	 * perf = (uint32_t)final_result;
3920 	 */
3921 	uint64_t perf;
3922 
3923 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3924 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3925 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3926 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3927 		     (uint64_t)1000000)) /
3928 		(uint64_t)1000000);
3929 	if (perf > curper) {
3930 		/* TSNH */
3931 		perf = curper - 1;
3932 	}
3933 	return ((uint32_t)perf);
3934 }
3935 
3936 static uint32_t
3937 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3938 {
3939 	/*
3940 	 *                                   highrttthresh
3941 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3942 	 *                                     gp_srtt
3943 	 *
3944 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
3945 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3946 	 */
3947 	uint64_t perf;
3948 	uint32_t highrttthresh;
3949 
3950 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3951 
3952 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3953 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3954 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3955 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3956 	if (tcp_bblogging_on(rack->rc_tp)) {
3957 		uint64_t log1;
3958 
3959 		log1 = rtt;
3960 		log1 <<= 32;
3961 		log1 |= highrttthresh;
3962 		rack_log_timely(rack,
3963 				rack_gp_decrease_per,
3964 				(uint64_t)curper,
3965 				log1,
3966 				perf,
3967 				__LINE__,
3968 				15);
3969 	}
3970 	return (perf);
3971 }
3972 
3973 static void
3974 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3975 {
3976 	uint64_t logvar, logvar2, logvar3;
3977 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3978 
3979 	if (rack->rc_gp_incr) {
3980 		/* Turn off increment counting */
3981 		rack->rc_gp_incr = 0;
3982 		rack->rc_gp_timely_inc_cnt = 0;
3983 	}
3984 	ss_red = ca_red = rec_red = 0;
3985 	logged = 0;
3986 	/* Calculate the reduction value */
3987 	if (rtt_diff < 0) {
3988 		rtt_diff *= -1;
3989 	}
3990 	/* Must be at least 1% reduction */
3991 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3992 		/* We have been in recovery ding it too */
3993 		if (timely_says == 2) {
3994 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3995 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3996 			if (alt < new_per)
3997 				val = alt;
3998 			else
3999 				val = new_per;
4000 		} else
4001 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4002 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4003 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4004 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4005 		} else {
4006 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4007 			rec_red = 0;
4008 		}
4009 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4010 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4011 		logged |= 1;
4012 	}
4013 	if (rack->rc_gp_saw_ss) {
4014 		/* Sent in SS */
4015 		if (timely_says == 2) {
4016 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4017 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4018 			if (alt < new_per)
4019 				val = alt;
4020 			else
4021 				val = new_per;
4022 		} else
4023 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4024 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4025 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4026 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4027 		} else {
4028 			ss_red = new_per;
4029 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4030 			logvar = new_per;
4031 			logvar <<= 32;
4032 			logvar |= alt;
4033 			logvar2 = (uint32_t)rtt;
4034 			logvar2 <<= 32;
4035 			logvar2 |= (uint32_t)rtt_diff;
4036 			logvar3 = rack_gp_rtt_maxmul;
4037 			logvar3 <<= 32;
4038 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4039 			rack_log_timely(rack, timely_says,
4040 					logvar2, logvar3,
4041 					logvar, __LINE__, 10);
4042 		}
4043 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4044 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4045 		logged |= 4;
4046 	} else if (rack->rc_gp_saw_ca) {
4047 		/* Sent in CA */
4048 		if (timely_says == 2) {
4049 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4050 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4051 			if (alt < new_per)
4052 				val = alt;
4053 			else
4054 				val = new_per;
4055 		} else
4056 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4057 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4058 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4059 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4060 		} else {
4061 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4062 			ca_red = 0;
4063 			logvar = new_per;
4064 			logvar <<= 32;
4065 			logvar |= alt;
4066 			logvar2 = (uint32_t)rtt;
4067 			logvar2 <<= 32;
4068 			logvar2 |= (uint32_t)rtt_diff;
4069 			logvar3 = rack_gp_rtt_maxmul;
4070 			logvar3 <<= 32;
4071 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4072 			rack_log_timely(rack, timely_says,
4073 					logvar2, logvar3,
4074 					logvar, __LINE__, 10);
4075 		}
4076 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4077 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4078 		logged |= 2;
4079 	}
4080 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4081 		rack->rc_gp_timely_dec_cnt++;
4082 		if (rack_timely_dec_clear &&
4083 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4084 			rack->rc_gp_timely_dec_cnt = 0;
4085 	}
4086 	logvar = ss_red;
4087 	logvar <<= 32;
4088 	logvar |= ca_red;
4089 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4090 			__LINE__, 2);
4091 }
4092 
4093 static void
4094 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4095 		     uint32_t rtt, uint32_t line, uint8_t reas)
4096 {
4097 	if (tcp_bblogging_on(rack->rc_tp)) {
4098 		union tcp_log_stackspecific log;
4099 		struct timeval tv;
4100 
4101 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4102 		log.u_bbr.flex1 = line;
4103 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4104 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4105 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4106 		log.u_bbr.flex5 = rtt;
4107 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4108 		log.u_bbr.flex6 <<= 1;
4109 		log.u_bbr.flex6 |= rack->forced_ack;
4110 		log.u_bbr.flex6 <<= 1;
4111 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4112 		log.u_bbr.flex6 <<= 1;
4113 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4114 		log.u_bbr.flex6 <<= 1;
4115 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4116 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4117 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4118 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4119 		log.u_bbr.flex8 = reas;
4120 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4121 		log.u_bbr.delRate = rack_get_bw(rack);
4122 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4123 		log.u_bbr.cur_del_rate <<= 32;
4124 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4125 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4126 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4127 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4128 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4129 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4130 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4131 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4132 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4133 		log.u_bbr.rttProp = us_cts;
4134 		log.u_bbr.rttProp <<= 32;
4135 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4136 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4137 		    &rack->rc_inp->inp_socket->so_rcv,
4138 		    &rack->rc_inp->inp_socket->so_snd,
4139 		    BBR_LOG_RTT_SHRINKS, 0,
4140 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4141 	}
4142 }
4143 
4144 static void
4145 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4146 {
4147 	uint64_t bwdp;
4148 
4149 	bwdp = rack_get_bw(rack);
4150 	bwdp *= (uint64_t)rtt;
4151 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4152 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4153 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4154 		/*
4155 		 * A window protocol must be able to have 4 packets
4156 		 * outstanding as the floor in order to function
4157 		 * (especially considering delayed ack :D).
4158 		 */
4159 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4160 	}
4161 }
4162 
4163 static void
4164 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4165 {
4166 	/**
4167 	 * ProbeRTT is a bit different in rack_pacing than in
4168 	 * BBR. It is like BBR in that it uses the lowering of
4169 	 * the RTT as a signal that we saw something new and
4170 	 * counts from there for how long between. But it is
4171 	 * different in that its quite simple. It does not
4172 	 * play with the cwnd and wait until we get down
4173 	 * to N segments outstanding and hold that for
4174 	 * 200ms. Instead it just sets the pacing reduction
4175 	 * rate to a set percentage (70 by default) and hold
4176 	 * that for a number of recent GP Srtt's.
4177 	 */
4178 	uint32_t segsiz;
4179 
4180 	if (rack->rc_gp_dyn_mul == 0)
4181 		return;
4182 
4183 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4184 		/* We are idle */
4185 		return;
4186 	}
4187 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4188 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4189 		/*
4190 		 * Stop the goodput now, the idea here is
4191 		 * that future measurements with in_probe_rtt
4192 		 * won't register if they are not greater so
4193 		 * we want to get what info (if any) is available
4194 		 * now.
4195 		 */
4196 		rack_do_goodput_measurement(rack->rc_tp, rack,
4197 					    rack->rc_tp->snd_una, __LINE__,
4198 					    RACK_QUALITY_PROBERTT);
4199 	}
4200 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4201 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4202 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4203 		     rack->r_ctl.rc_pace_min_segs);
4204 	rack->in_probe_rtt = 1;
4205 	rack->measure_saw_probe_rtt = 1;
4206 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4207 	rack->r_ctl.rc_time_probertt_starts = 0;
4208 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4209 	if (rack_probertt_use_min_rtt_entry)
4210 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4211 	else
4212 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4213 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4214 			     __LINE__, RACK_RTTS_ENTERPROBE);
4215 }
4216 
4217 static void
4218 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4219 {
4220 	struct rack_sendmap *rsm;
4221 	uint32_t segsiz;
4222 
4223 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4224 		     rack->r_ctl.rc_pace_min_segs);
4225 	rack->in_probe_rtt = 0;
4226 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4227 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4228 		/*
4229 		 * Stop the goodput now, the idea here is
4230 		 * that future measurements with in_probe_rtt
4231 		 * won't register if they are not greater so
4232 		 * we want to get what info (if any) is available
4233 		 * now.
4234 		 */
4235 		rack_do_goodput_measurement(rack->rc_tp, rack,
4236 					    rack->rc_tp->snd_una, __LINE__,
4237 					    RACK_QUALITY_PROBERTT);
4238 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4239 		/*
4240 		 * We don't have enough data to make a measurement.
4241 		 * So lets just stop and start here after exiting
4242 		 * probe-rtt. We probably are not interested in
4243 		 * the results anyway.
4244 		 */
4245 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4246 	}
4247 	/*
4248 	 * Measurements through the current snd_max are going
4249 	 * to be limited by the slower pacing rate.
4250 	 *
4251 	 * We need to mark these as app-limited so we
4252 	 * don't collapse the b/w.
4253 	 */
4254 	rsm = tqhash_max(rack->r_ctl.tqh);
4255 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4256 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4257 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4258 		else {
4259 			/*
4260 			 * Go out to the end app limited and mark
4261 			 * this new one as next and move the end_appl up
4262 			 * to this guy.
4263 			 */
4264 			if (rack->r_ctl.rc_end_appl)
4265 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4266 			rack->r_ctl.rc_end_appl = rsm;
4267 		}
4268 		rsm->r_flags |= RACK_APP_LIMITED;
4269 		rack->r_ctl.rc_app_limited_cnt++;
4270 	}
4271 	/*
4272 	 * Now, we need to examine our pacing rate multipliers.
4273 	 * If its under 100%, we need to kick it back up to
4274 	 * 100%. We also don't let it be over our "max" above
4275 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4276 	 * Note setting clamp_atexit_prtt to 0 has the effect
4277 	 * of setting CA/SS to 100% always at exit (which is
4278 	 * the default behavior).
4279 	 */
4280 	if (rack_probertt_clear_is) {
4281 		rack->rc_gp_incr = 0;
4282 		rack->rc_gp_bwred = 0;
4283 		rack->rc_gp_timely_inc_cnt = 0;
4284 		rack->rc_gp_timely_dec_cnt = 0;
4285 	}
4286 	/* Do we do any clamping at exit? */
4287 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4288 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4289 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4290 	}
4291 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4292 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4293 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4294 	}
4295 	/*
4296 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4297 	 * after exiting.
4298 	 */
4299 	rack->r_ctl.rc_rtt_diff = 0;
4300 
4301 	/* Clear all flags so we start fresh */
4302 	rack->rc_tp->t_bytes_acked = 0;
4303 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4304 	/*
4305 	 * If configured to, set the cwnd and ssthresh to
4306 	 * our targets.
4307 	 */
4308 	if (rack_probe_rtt_sets_cwnd) {
4309 		uint64_t ebdp;
4310 		uint32_t setto;
4311 
4312 		/* Set ssthresh so we get into CA once we hit our target */
4313 		if (rack_probertt_use_min_rtt_exit == 1) {
4314 			/* Set to min rtt */
4315 			rack_set_prtt_target(rack, segsiz,
4316 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4317 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4318 			/* Set to current gp rtt */
4319 			rack_set_prtt_target(rack, segsiz,
4320 					     rack->r_ctl.rc_gp_srtt);
4321 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4322 			/* Set to entry gp rtt */
4323 			rack_set_prtt_target(rack, segsiz,
4324 					     rack->r_ctl.rc_entry_gp_rtt);
4325 		} else {
4326 			uint64_t sum;
4327 			uint32_t setval;
4328 
4329 			sum = rack->r_ctl.rc_entry_gp_rtt;
4330 			sum *= 10;
4331 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4332 			if (sum >= 20) {
4333 				/*
4334 				 * A highly buffered path needs
4335 				 * cwnd space for timely to work.
4336 				 * Lets set things up as if
4337 				 * we are heading back here again.
4338 				 */
4339 				setval = rack->r_ctl.rc_entry_gp_rtt;
4340 			} else if (sum >= 15) {
4341 				/*
4342 				 * Lets take the smaller of the
4343 				 * two since we are just somewhat
4344 				 * buffered.
4345 				 */
4346 				setval = rack->r_ctl.rc_gp_srtt;
4347 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4348 					setval = rack->r_ctl.rc_entry_gp_rtt;
4349 			} else {
4350 				/*
4351 				 * Here we are not highly buffered
4352 				 * and should pick the min we can to
4353 				 * keep from causing loss.
4354 				 */
4355 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4356 			}
4357 			rack_set_prtt_target(rack, segsiz,
4358 					     setval);
4359 		}
4360 		if (rack_probe_rtt_sets_cwnd > 1) {
4361 			/* There is a percentage here to boost */
4362 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4363 			ebdp *= rack_probe_rtt_sets_cwnd;
4364 			ebdp /= 100;
4365 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4366 		} else
4367 			setto = rack->r_ctl.rc_target_probertt_flight;
4368 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4369 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4370 			/* Enforce a min */
4371 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4372 		}
4373 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4374 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4375 	}
4376 	rack_log_rtt_shrinks(rack,  us_cts,
4377 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4378 			     __LINE__, RACK_RTTS_EXITPROBE);
4379 	/* Clear times last so log has all the info */
4380 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4381 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4382 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4383 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4384 }
4385 
4386 static void
4387 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4388 {
4389 	/* Check in on probe-rtt */
4390 	if (rack->rc_gp_filled == 0) {
4391 		/* We do not do p-rtt unless we have gp measurements */
4392 		return;
4393 	}
4394 	if (rack->in_probe_rtt) {
4395 		uint64_t no_overflow;
4396 		uint32_t endtime, must_stay;
4397 
4398 		if (rack->r_ctl.rc_went_idle_time &&
4399 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4400 			/*
4401 			 * We went idle during prtt, just exit now.
4402 			 */
4403 			rack_exit_probertt(rack, us_cts);
4404 		} else if (rack_probe_rtt_safety_val &&
4405 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4406 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4407 			/*
4408 			 * Probe RTT safety value triggered!
4409 			 */
4410 			rack_log_rtt_shrinks(rack,  us_cts,
4411 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4412 					     __LINE__, RACK_RTTS_SAFETY);
4413 			rack_exit_probertt(rack, us_cts);
4414 		}
4415 		/* Calculate the max we will wait */
4416 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4417 		if (rack->rc_highly_buffered)
4418 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4419 		/* Calculate the min we must wait */
4420 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4421 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4422 		    TSTMP_LT(us_cts, endtime)) {
4423 			uint32_t calc;
4424 			/* Do we lower more? */
4425 no_exit:
4426 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4427 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4428 			else
4429 				calc = 0;
4430 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4431 			if (calc) {
4432 				/* Maybe */
4433 				calc *= rack_per_of_gp_probertt_reduce;
4434 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4435 				/* Limit it too */
4436 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4437 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4438 			}
4439 			/* We must reach target or the time set */
4440 			return;
4441 		}
4442 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4443 			if ((TSTMP_LT(us_cts, must_stay) &&
4444 			     rack->rc_highly_buffered) ||
4445 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4446 			      rack->r_ctl.rc_target_probertt_flight)) {
4447 				/* We are not past the must_stay time */
4448 				goto no_exit;
4449 			}
4450 			rack_log_rtt_shrinks(rack,  us_cts,
4451 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4452 					     __LINE__, RACK_RTTS_REACHTARGET);
4453 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4454 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4455 				rack->r_ctl.rc_time_probertt_starts = 1;
4456 			/* Restore back to our rate we want to pace at in prtt */
4457 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4458 		}
4459 		/*
4460 		 * Setup our end time, some number of gp_srtts plus 200ms.
4461 		 */
4462 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4463 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4464 		if (rack_probertt_gpsrtt_cnt_div)
4465 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4466 		else
4467 			endtime = 0;
4468 		endtime += rack_min_probertt_hold;
4469 		endtime += rack->r_ctl.rc_time_probertt_starts;
4470 		if (TSTMP_GEQ(us_cts,  endtime)) {
4471 			/* yes, exit probertt */
4472 			rack_exit_probertt(rack, us_cts);
4473 		}
4474 
4475 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
4476 		/* Go into probertt, its been too long since we went lower */
4477 		rack_enter_probertt(rack, us_cts);
4478 	}
4479 }
4480 
4481 static void
4482 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4483 		       uint32_t rtt, int32_t rtt_diff)
4484 {
4485 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4486 	uint32_t losses;
4487 
4488 	if ((rack->rc_gp_dyn_mul == 0) ||
4489 	    (rack->use_fixed_rate) ||
4490 	    (rack->in_probe_rtt) ||
4491 	    (rack->rc_always_pace == 0)) {
4492 		/* No dynamic GP multiplier in play */
4493 		return;
4494 	}
4495 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4496 	cur_bw = rack_get_bw(rack);
4497 	/* Calculate our up and down range */
4498 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4499 	up_bnd /= 100;
4500 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4501 
4502 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4503 	subfr /= 100;
4504 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4505 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4506 		/*
4507 		 * This is the case where our RTT is above
4508 		 * the max target and we have been configured
4509 		 * to just do timely no bonus up stuff in that case.
4510 		 *
4511 		 * There are two configurations, set to 1, and we
4512 		 * just do timely if we are over our max. If its
4513 		 * set above 1 then we slam the multipliers down
4514 		 * to 100 and then decrement per timely.
4515 		 */
4516 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4517 				__LINE__, 3);
4518 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4519 			rack_validate_multipliers_at_or_below_100(rack);
4520 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4521 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4522 		/*
4523 		 * We are decreasing this is a bit complicated this
4524 		 * means we are loosing ground. This could be
4525 		 * because another flow entered and we are competing
4526 		 * for b/w with it. This will push the RTT up which
4527 		 * makes timely unusable unless we want to get shoved
4528 		 * into a corner and just be backed off (the age
4529 		 * old problem with delay based CC).
4530 		 *
4531 		 * On the other hand if it was a route change we
4532 		 * would like to stay somewhat contained and not
4533 		 * blow out the buffers.
4534 		 */
4535 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4536 				__LINE__, 3);
4537 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4538 		if (rack->rc_gp_bwred == 0) {
4539 			/* Go into reduction counting */
4540 			rack->rc_gp_bwred = 1;
4541 			rack->rc_gp_timely_dec_cnt = 0;
4542 		}
4543 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4544 			/*
4545 			 * Push another time with a faster pacing
4546 			 * to try to gain back (we include override to
4547 			 * get a full raise factor).
4548 			 */
4549 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4550 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4551 			    (timely_says == 0) ||
4552 			    (rack_down_raise_thresh == 0)) {
4553 				/*
4554 				 * Do an override up in b/w if we were
4555 				 * below the threshold or if the threshold
4556 				 * is zero we always do the raise.
4557 				 */
4558 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4559 			} else {
4560 				/* Log it stays the same */
4561 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4562 						__LINE__, 11);
4563 			}
4564 			rack->rc_gp_timely_dec_cnt++;
4565 			/* We are not incrementing really no-count */
4566 			rack->rc_gp_incr = 0;
4567 			rack->rc_gp_timely_inc_cnt = 0;
4568 		} else {
4569 			/*
4570 			 * Lets just use the RTT
4571 			 * information and give up
4572 			 * pushing.
4573 			 */
4574 			goto use_timely;
4575 		}
4576 	} else if ((timely_says != 2) &&
4577 		    !losses &&
4578 		    (last_bw_est > up_bnd)) {
4579 		/*
4580 		 * We are increasing b/w lets keep going, updating
4581 		 * our b/w and ignoring any timely input, unless
4582 		 * of course we are at our max raise (if there is one).
4583 		 */
4584 
4585 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4586 				__LINE__, 3);
4587 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4588 		if (rack->rc_gp_saw_ss &&
4589 		    rack->r_ctl.rack_per_upper_bound_ss &&
4590 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4591 			    /*
4592 			     * In cases where we can't go higher
4593 			     * we should just use timely.
4594 			     */
4595 			    goto use_timely;
4596 		}
4597 		if (rack->rc_gp_saw_ca &&
4598 		    rack->r_ctl.rack_per_upper_bound_ca &&
4599 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4600 			    /*
4601 			     * In cases where we can't go higher
4602 			     * we should just use timely.
4603 			     */
4604 			    goto use_timely;
4605 		}
4606 		rack->rc_gp_bwred = 0;
4607 		rack->rc_gp_timely_dec_cnt = 0;
4608 		/* You get a set number of pushes if timely is trying to reduce */
4609 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4610 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4611 		} else {
4612 			/* Log it stays the same */
4613 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4614 			    __LINE__, 12);
4615 		}
4616 		return;
4617 	} else {
4618 		/*
4619 		 * We are staying between the lower and upper range bounds
4620 		 * so use timely to decide.
4621 		 */
4622 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4623 				__LINE__, 3);
4624 use_timely:
4625 		if (timely_says) {
4626 			rack->rc_gp_incr = 0;
4627 			rack->rc_gp_timely_inc_cnt = 0;
4628 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4629 			    !losses &&
4630 			    (last_bw_est < low_bnd)) {
4631 				/* We are loosing ground */
4632 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4633 				rack->rc_gp_timely_dec_cnt++;
4634 				/* We are not incrementing really no-count */
4635 				rack->rc_gp_incr = 0;
4636 				rack->rc_gp_timely_inc_cnt = 0;
4637 			} else
4638 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4639 		} else {
4640 			rack->rc_gp_bwred = 0;
4641 			rack->rc_gp_timely_dec_cnt = 0;
4642 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4643 		}
4644 	}
4645 }
4646 
4647 static int32_t
4648 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4649 {
4650 	int32_t timely_says;
4651 	uint64_t log_mult, log_rtt_a_diff;
4652 
4653 	log_rtt_a_diff = rtt;
4654 	log_rtt_a_diff <<= 32;
4655 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4656 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4657 		    rack_gp_rtt_maxmul)) {
4658 		/* Reduce the b/w multiplier */
4659 		timely_says = 2;
4660 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4661 		log_mult <<= 32;
4662 		log_mult |= prev_rtt;
4663 		rack_log_timely(rack,  timely_says, log_mult,
4664 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4665 				log_rtt_a_diff, __LINE__, 4);
4666 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4667 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4668 			    max(rack_gp_rtt_mindiv , 1)))) {
4669 		/* Increase the b/w multiplier */
4670 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4671 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4672 			 max(rack_gp_rtt_mindiv , 1));
4673 		log_mult <<= 32;
4674 		log_mult |= prev_rtt;
4675 		timely_says = 0;
4676 		rack_log_timely(rack,  timely_says, log_mult ,
4677 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4678 				log_rtt_a_diff, __LINE__, 5);
4679 	} else {
4680 		/*
4681 		 * Use a gradient to find it the timely gradient
4682 		 * is:
4683 		 * grad = rc_rtt_diff / min_rtt;
4684 		 *
4685 		 * anything below or equal to 0 will be
4686 		 * a increase indication. Anything above
4687 		 * zero is a decrease. Note we take care
4688 		 * of the actual gradient calculation
4689 		 * in the reduction (its not needed for
4690 		 * increase).
4691 		 */
4692 		log_mult = prev_rtt;
4693 		if (rtt_diff <= 0) {
4694 			/*
4695 			 * Rttdiff is less than zero, increase the
4696 			 * b/w multiplier (its 0 or negative)
4697 			 */
4698 			timely_says = 0;
4699 			rack_log_timely(rack,  timely_says, log_mult,
4700 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4701 		} else {
4702 			/* Reduce the b/w multiplier */
4703 			timely_says = 1;
4704 			rack_log_timely(rack,  timely_says, log_mult,
4705 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4706 		}
4707 	}
4708 	return (timely_says);
4709 }
4710 
4711 static __inline int
4712 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4713 {
4714 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4715 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4716 		/**
4717 		 * This covers the case that the
4718 		 * resent is completely inside
4719 		 * the gp range or up to it.
4720 		 *      |----------------|
4721 		 *      |-----| <or>
4722 		 *            |----|
4723 		 *            <or>   |---|
4724 		 */
4725 		return (1);
4726 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4727 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4728 		/**
4729 		 * This covers the case of
4730 		 *      |--------------|
4731 		 *  |-------->|
4732 		 */
4733 		return (1);
4734 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4735 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4736 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4737 
4738 		/**
4739 		 * This covers the case of
4740 		 *      |--------------|
4741 		 *              |-------->|
4742 		 */
4743 		return (1);
4744 	}
4745 	return (0);
4746 }
4747 
4748 static __inline void
4749 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4750 {
4751 
4752 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4753 		return;
4754 	/*
4755 	 * We have a Goodput measurement in progress. Mark
4756 	 * the send if its within the window. If its not
4757 	 * in the window make sure it does not have the mark.
4758 	 */
4759 	if (rack_in_gp_window(tp, rsm))
4760 		rsm->r_flags |= RACK_IN_GP_WIN;
4761 	else
4762 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4763 }
4764 
4765 static __inline void
4766 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4767 {
4768 	/* A GP measurement is ending, clear all marks on the send map*/
4769 	struct rack_sendmap *rsm = NULL;
4770 
4771 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4772 	if (rsm == NULL) {
4773 		rsm = tqhash_min(rack->r_ctl.tqh);
4774 	}
4775 	/* Nothing left? */
4776 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4777 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4778 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4779 	}
4780 }
4781 
4782 
4783 static __inline void
4784 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4785 {
4786 	struct rack_sendmap *rsm = NULL;
4787 
4788 	if (tp->snd_una == tp->snd_max) {
4789 		/* Nothing outstanding yet, nothing to do here */
4790 		return;
4791 	}
4792 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4793 		/*
4794 		 * We are measuring ahead of some outstanding
4795 		 * data. We need to walk through up until we get
4796 		 * to gp_seq marking so that no rsm is set incorrectly
4797 		 * with RACK_IN_GP_WIN.
4798 		 */
4799 		rsm = tqhash_min(rack->r_ctl.tqh);
4800 		while (rsm != NULL) {
4801 			rack_mark_in_gp_win(tp, rsm);
4802 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4803 				break;
4804 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4805 		}
4806 	}
4807 	if (rsm == NULL) {
4808 		/*
4809 		 * Need to find the GP seq, if rsm is
4810 		 * set we stopped as we hit it.
4811 		 */
4812 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4813 		if (rsm == NULL)
4814 			return;
4815 		rack_mark_in_gp_win(tp, rsm);
4816 	}
4817 	/*
4818 	 * Now we may need to mark already sent rsm, ahead of
4819 	 * gput_seq in the window since they may have been sent
4820 	 * *before* we started our measurment. The rsm, if non-null
4821 	 * has been marked (note if rsm would have been NULL we would have
4822 	 * returned in the previous block). So we go to the next, and continue
4823 	 * until we run out of entries or we exceed the gp_ack value.
4824 	 */
4825 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4826 	while (rsm) {
4827 		rack_mark_in_gp_win(tp, rsm);
4828 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4829 			break;
4830 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4831 	}
4832 }
4833 
4834 static void
4835 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4836 			    tcp_seq th_ack, int line, uint8_t quality)
4837 {
4838 	uint64_t tim, bytes_ps, stim, utim;
4839 	uint32_t segsiz, bytes, reqbytes, us_cts;
4840 	int32_t gput, new_rtt_diff, timely_says;
4841 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4842 	int did_add = 0;
4843 
4844 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4845 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4846 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4847 		tim = us_cts - tp->gput_ts;
4848 	else
4849 		tim = 0;
4850 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4851 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4852 	else
4853 		stim = 0;
4854 	/*
4855 	 * Use the larger of the send time or ack time. This prevents us
4856 	 * from being influenced by ack artifacts to come up with too
4857 	 * high of measurement. Note that since we are spanning over many more
4858 	 * bytes in most of our measurements hopefully that is less likely to
4859 	 * occur.
4860 	 */
4861 	if (tim > stim)
4862 		utim = max(tim, 1);
4863 	else
4864 		utim = max(stim, 1);
4865 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4866 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4867 	if ((tim == 0) && (stim == 0)) {
4868 		/*
4869 		 * Invalid measurement time, maybe
4870 		 * all on one ack/one send?
4871 		 */
4872 		bytes = 0;
4873 		bytes_ps = 0;
4874 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4875 					   0, 0, 0, 10, __LINE__, NULL, quality);
4876 		goto skip_measurement;
4877 	}
4878 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4879 		/* We never made a us_rtt measurement? */
4880 		bytes = 0;
4881 		bytes_ps = 0;
4882 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4883 					   0, 0, 0, 10, __LINE__, NULL, quality);
4884 		goto skip_measurement;
4885 	}
4886 	/*
4887 	 * Calculate the maximum possible b/w this connection
4888 	 * could have. We base our calculation on the lowest
4889 	 * rtt we have seen during the measurement and the
4890 	 * largest rwnd the client has given us in that time. This
4891 	 * forms a BDP that is the maximum that we could ever
4892 	 * get to the client. Anything larger is not valid.
4893 	 *
4894 	 * I originally had code here that rejected measurements
4895 	 * where the time was less than 1/2 the latest us_rtt.
4896 	 * But after thinking on that I realized its wrong since
4897 	 * say you had a 150Mbps or even 1Gbps link, and you
4898 	 * were a long way away.. example I am in Europe (100ms rtt)
4899 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4900 	 * bytes my time would be 1.2ms, and yet my rtt would say
4901 	 * the measurement was invalid the time was < 50ms. The
4902 	 * same thing is true for 150Mb (8ms of time).
4903 	 *
4904 	 * A better way I realized is to look at what the maximum
4905 	 * the connection could possibly do. This is gated on
4906 	 * the lowest RTT we have seen and the highest rwnd.
4907 	 * We should in theory never exceed that, if we are
4908 	 * then something on the path is storing up packets
4909 	 * and then feeding them all at once to our endpoint
4910 	 * messing up our measurement.
4911 	 */
4912 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4913 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4914 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4915 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4916 		/* No measurement can be made */
4917 		bytes = 0;
4918 		bytes_ps = 0;
4919 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4920 					   0, 0, 0, 10, __LINE__, NULL, quality);
4921 		goto skip_measurement;
4922 	} else
4923 		bytes = (th_ack - tp->gput_seq);
4924 	bytes_ps = (uint64_t)bytes;
4925 	/*
4926 	 * Don't measure a b/w for pacing unless we have gotten at least
4927 	 * an initial windows worth of data in this measurement interval.
4928 	 *
4929 	 * Small numbers of bytes get badly influenced by delayed ack and
4930 	 * other artifacts. Note we take the initial window or our
4931 	 * defined minimum GP (defaulting to 10 which hopefully is the
4932 	 * IW).
4933 	 */
4934 	if (rack->rc_gp_filled == 0) {
4935 		/*
4936 		 * The initial estimate is special. We
4937 		 * have blasted out an IW worth of packets
4938 		 * without a real valid ack ts results. We
4939 		 * then setup the app_limited_needs_set flag,
4940 		 * this should get the first ack in (probably 2
4941 		 * MSS worth) to be recorded as the timestamp.
4942 		 * We thus allow a smaller number of bytes i.e.
4943 		 * IW - 2MSS.
4944 		 */
4945 		reqbytes -= (2 * segsiz);
4946 		/* Also lets fill previous for our first measurement to be neutral */
4947 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4948 	}
4949 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4950 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4951 					   rack->r_ctl.rc_app_limited_cnt,
4952 					   0, 0, 10, __LINE__, NULL, quality);
4953 		goto skip_measurement;
4954 	}
4955 	/*
4956 	 * We now need to calculate the Timely like status so
4957 	 * we can update (possibly) the b/w multipliers.
4958 	 */
4959 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4960 	if (rack->rc_gp_filled == 0) {
4961 		/* No previous reading */
4962 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4963 	} else {
4964 		if (rack->measure_saw_probe_rtt == 0) {
4965 			/*
4966 			 * We don't want a probertt to be counted
4967 			 * since it will be negative incorrectly. We
4968 			 * expect to be reducing the RTT when we
4969 			 * pace at a slower rate.
4970 			 */
4971 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4972 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4973 		}
4974 	}
4975 	timely_says = rack_make_timely_judgement(rack,
4976 	    rack->r_ctl.rc_gp_srtt,
4977 	    rack->r_ctl.rc_rtt_diff,
4978 	    rack->r_ctl.rc_prev_gp_srtt
4979 	);
4980 	bytes_ps *= HPTS_USEC_IN_SEC;
4981 	bytes_ps /= utim;
4982 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4983 		/*
4984 		 * Something is on path playing
4985 		 * since this b/w is not possible based
4986 		 * on our BDP (highest rwnd and lowest rtt
4987 		 * we saw in the measurement window).
4988 		 *
4989 		 * Another option here would be to
4990 		 * instead skip the measurement.
4991 		 */
4992 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4993 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4994 					   11, __LINE__, NULL, quality);
4995 		bytes_ps = rack->r_ctl.last_max_bw;
4996 	}
4997 	/* We store gp for b/w in bytes per second */
4998 	if (rack->rc_gp_filled == 0) {
4999 		/* Initial measurement */
5000 		if (bytes_ps) {
5001 			rack->r_ctl.gp_bw = bytes_ps;
5002 			rack->rc_gp_filled = 1;
5003 			rack->r_ctl.num_measurements = 1;
5004 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5005 		} else {
5006 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5007 						   rack->r_ctl.rc_app_limited_cnt,
5008 						   0, 0, 10, __LINE__, NULL, quality);
5009 		}
5010 		if (tcp_in_hpts(rack->rc_tp) &&
5011 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5012 			/*
5013 			 * Ok we can't trust the pacer in this case
5014 			 * where we transition from un-paced to paced.
5015 			 * Or for that matter when the burst mitigation
5016 			 * was making a wild guess and got it wrong.
5017 			 * Stop the pacer and clear up all the aggregate
5018 			 * delays etc.
5019 			 */
5020 			tcp_hpts_remove(rack->rc_tp);
5021 			rack->r_ctl.rc_hpts_flags = 0;
5022 			rack->r_ctl.rc_last_output_to = 0;
5023 		}
5024 		did_add = 2;
5025 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5026 		/* Still a small number run an average */
5027 		rack->r_ctl.gp_bw += bytes_ps;
5028 		addpart = rack->r_ctl.num_measurements;
5029 		rack->r_ctl.num_measurements++;
5030 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5031 			/* We have collected enough to move forward */
5032 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5033 		}
5034 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5035 		did_add = 3;
5036 	} else {
5037 		/*
5038 		 * We want to take 1/wma of the goodput and add in to 7/8th
5039 		 * of the old value weighted by the srtt. So if your measurement
5040 		 * period is say 2 SRTT's long you would get 1/4 as the
5041 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5042 		 *
5043 		 * But we must be careful not to take too much i.e. if the
5044 		 * srtt is say 20ms and the measurement is taken over
5045 		 * 400ms our weight would be 400/20 i.e. 20. On the
5046 		 * other hand if we get a measurement over 1ms with a
5047 		 * 10ms rtt we only want to take a much smaller portion.
5048 		 */
5049 		if (rack->r_ctl.num_measurements < 0xff) {
5050 			rack->r_ctl.num_measurements++;
5051 		}
5052 		srtt = (uint64_t)tp->t_srtt;
5053 		if (srtt == 0) {
5054 			/*
5055 			 * Strange why did t_srtt go back to zero?
5056 			 */
5057 			if (rack->r_ctl.rc_rack_min_rtt)
5058 				srtt = rack->r_ctl.rc_rack_min_rtt;
5059 			else
5060 				srtt = HPTS_USEC_IN_MSEC;
5061 		}
5062 		/*
5063 		 * XXXrrs: Note for reviewers, in playing with
5064 		 * dynamic pacing I discovered this GP calculation
5065 		 * as done originally leads to some undesired results.
5066 		 * Basically you can get longer measurements contributing
5067 		 * too much to the WMA. Thus I changed it if you are doing
5068 		 * dynamic adjustments to only do the aportioned adjustment
5069 		 * if we have a very small (time wise) measurement. Longer
5070 		 * measurements just get there weight (defaulting to 1/8)
5071 		 * add to the WMA. We may want to think about changing
5072 		 * this to always do that for both sides i.e. dynamic
5073 		 * and non-dynamic... but considering lots of folks
5074 		 * were playing with this I did not want to change the
5075 		 * calculation per.se. without your thoughts.. Lawerence?
5076 		 * Peter??
5077 		 */
5078 		if (rack->rc_gp_dyn_mul == 0) {
5079 			subpart = rack->r_ctl.gp_bw * utim;
5080 			subpart /= (srtt * 8);
5081 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5082 				/*
5083 				 * The b/w update takes no more
5084 				 * away then 1/2 our running total
5085 				 * so factor it in.
5086 				 */
5087 				addpart = bytes_ps * utim;
5088 				addpart /= (srtt * 8);
5089 			} else {
5090 				/*
5091 				 * Don't allow a single measurement
5092 				 * to account for more than 1/2 of the
5093 				 * WMA. This could happen on a retransmission
5094 				 * where utim becomes huge compared to
5095 				 * srtt (multiple retransmissions when using
5096 				 * the sending rate which factors in all the
5097 				 * transmissions from the first one).
5098 				 */
5099 				subpart = rack->r_ctl.gp_bw / 2;
5100 				addpart = bytes_ps / 2;
5101 			}
5102 			resid_bw = rack->r_ctl.gp_bw - subpart;
5103 			rack->r_ctl.gp_bw = resid_bw + addpart;
5104 			did_add = 1;
5105 		} else {
5106 			if ((utim / srtt) <= 1) {
5107 				/*
5108 				 * The b/w update was over a small period
5109 				 * of time. The idea here is to prevent a small
5110 				 * measurement time period from counting
5111 				 * too much. So we scale it based on the
5112 				 * time so it attributes less than 1/rack_wma_divisor
5113 				 * of its measurement.
5114 				 */
5115 				subpart = rack->r_ctl.gp_bw * utim;
5116 				subpart /= (srtt * rack_wma_divisor);
5117 				addpart = bytes_ps * utim;
5118 				addpart /= (srtt * rack_wma_divisor);
5119 			} else {
5120 				/*
5121 				 * The scaled measurement was long
5122 				 * enough so lets just add in the
5123 				 * portion of the measurement i.e. 1/rack_wma_divisor
5124 				 */
5125 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5126 				addpart = bytes_ps / rack_wma_divisor;
5127 			}
5128 			if ((rack->measure_saw_probe_rtt == 0) ||
5129 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5130 				/*
5131 				 * For probe-rtt we only add it in
5132 				 * if its larger, all others we just
5133 				 * add in.
5134 				 */
5135 				did_add = 1;
5136 				resid_bw = rack->r_ctl.gp_bw - subpart;
5137 				rack->r_ctl.gp_bw = resid_bw + addpart;
5138 			}
5139 		}
5140 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5141 	}
5142 	if ((rack->gp_ready == 0) &&
5143 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5144 		/* We have enough measurements now */
5145 		rack->gp_ready = 1;
5146 		if (rack->dgp_on ||
5147 		    rack->rack_hibeta)
5148 			rack_set_cc_pacing(rack);
5149 		if (rack->defer_options)
5150 			rack_apply_deferred_options(rack);
5151 	}
5152 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5153 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5154 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5155 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
5156 		rack_update_multiplier(rack, timely_says, bytes_ps,
5157 				       rack->r_ctl.rc_gp_srtt,
5158 				       rack->r_ctl.rc_rtt_diff);
5159 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5160 				   rack_get_bw(rack), 3, line, NULL, quality);
5161 	rack_log_pacing_delay_calc(rack,
5162 				   bytes, /* flex2 */
5163 				   tim, /* flex1 */
5164 				   bytes_ps, /* bw_inuse */
5165 				   rack->r_ctl.gp_bw, /* delRate */
5166 				   rack_get_lt_bw(rack), /* rttProp */
5167 				   20, line, NULL, 0);
5168 	/* reset the gp srtt and setup the new prev */
5169 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5170 	/* Record the lost count for the next measurement */
5171 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5172 skip_measurement:
5173 	/*
5174 	 * We restart our diffs based on the gpsrtt in the
5175 	 * measurement window.
5176 	 */
5177 	rack->rc_gp_rtt_set = 0;
5178 	rack->rc_gp_saw_rec = 0;
5179 	rack->rc_gp_saw_ca = 0;
5180 	rack->rc_gp_saw_ss = 0;
5181 	rack->rc_dragged_bottom = 0;
5182 
5183 	if (quality == RACK_QUALITY_HIGH) {
5184 		/*
5185 		 * Gput in the stats world is in kbps where bytes_ps is
5186 		 * bytes per second so we do ((x * 8)/ 1000).
5187 		 */
5188 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5189 #ifdef STATS
5190 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5191 					 gput);
5192 		/*
5193 		 * XXXLAS: This is a temporary hack, and should be
5194 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5195 		 * API to deal with chained VOIs.
5196 		 */
5197 		if (tp->t_stats_gput_prev > 0)
5198 			stats_voi_update_abs_s32(tp->t_stats,
5199 						 VOI_TCP_GPUT_ND,
5200 						 ((gput - tp->t_stats_gput_prev) * 100) /
5201 						 tp->t_stats_gput_prev);
5202 #endif
5203 		tp->t_stats_gput_prev = gput;
5204 	}
5205 	tp->t_flags &= ~TF_GPUTINPROG;
5206 	/*
5207 	 * Now are we app limited now and there is space from where we
5208 	 * were to where we want to go?
5209 	 *
5210 	 * We don't do the other case i.e. non-applimited here since
5211 	 * the next send will trigger us picking up the missing data.
5212 	 */
5213 	if (rack->r_ctl.rc_first_appl &&
5214 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5215 	    rack->r_ctl.rc_app_limited_cnt &&
5216 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5217 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5218 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5219 		/*
5220 		 * Yep there is enough outstanding to make a measurement here.
5221 		 */
5222 		struct rack_sendmap *rsm;
5223 
5224 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5225 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5226 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5227 		rack->app_limited_needs_set = 0;
5228 		tp->gput_seq = th_ack;
5229 		if (rack->in_probe_rtt)
5230 			rack->measure_saw_probe_rtt = 1;
5231 		else if ((rack->measure_saw_probe_rtt) &&
5232 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5233 			rack->measure_saw_probe_rtt = 0;
5234 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5235 			/* There is a full window to gain info from */
5236 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5237 		} else {
5238 			/* We can only measure up to the applimited point */
5239 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5240 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5241 				/*
5242 				 * We don't have enough to make a measurement.
5243 				 */
5244 				tp->t_flags &= ~TF_GPUTINPROG;
5245 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5246 							   0, 0, 0, 6, __LINE__, NULL, quality);
5247 				return;
5248 			}
5249 		}
5250 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5251 			/*
5252 			 * We will get no more data into the SB
5253 			 * this means we need to have the data available
5254 			 * before we start a measurement.
5255 			 */
5256 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5257 				/* Nope not enough data. */
5258 				return;
5259 			}
5260 		}
5261 		tp->t_flags |= TF_GPUTINPROG;
5262 		/*
5263 		 * Now we need to find the timestamp of the send at tp->gput_seq
5264 		 * for the send based measurement.
5265 		 */
5266 		rack->r_ctl.rc_gp_cumack_ts = 0;
5267 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5268 		if (rsm) {
5269 			/* Ok send-based limit is set */
5270 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5271 				/*
5272 				 * Move back to include the earlier part
5273 				 * so our ack time lines up right (this may
5274 				 * make an overlapping measurement but thats
5275 				 * ok).
5276 				 */
5277 				tp->gput_seq = rsm->r_start;
5278 			}
5279 			if (rsm->r_flags & RACK_ACKED) {
5280 				struct rack_sendmap *nrsm;
5281 
5282 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5283 				tp->gput_seq = rsm->r_end;
5284 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5285 				if (nrsm)
5286 					rsm = nrsm;
5287 				else {
5288 					rack->app_limited_needs_set = 1;
5289 				}
5290 			} else
5291 				rack->app_limited_needs_set = 1;
5292 			/* We always go from the first send */
5293 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5294 		} else {
5295 			/*
5296 			 * If we don't find the rsm due to some
5297 			 * send-limit set the current time, which
5298 			 * basically disables the send-limit.
5299 			 */
5300 			struct timeval tv;
5301 
5302 			microuptime(&tv);
5303 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5304 		}
5305 		rack_tend_gp_marks(tp, rack);
5306 		rack_log_pacing_delay_calc(rack,
5307 					   tp->gput_seq,
5308 					   tp->gput_ack,
5309 					   (uint64_t)rsm,
5310 					   tp->gput_ts,
5311 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5312 					   9,
5313 					   __LINE__, rsm, quality);
5314 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5315 	} else {
5316 		/*
5317 		 * To make sure proper timestamp merging occurs, we need to clear
5318 		 * all GP marks if we don't start a measurement.
5319 		 */
5320 		rack_clear_gp_marks(tp, rack);
5321 	}
5322 }
5323 
5324 /*
5325  * CC wrapper hook functions
5326  */
5327 static void
5328 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5329     uint16_t type, int32_t recovery)
5330 {
5331 	uint32_t prior_cwnd, acked;
5332 	struct tcp_log_buffer *lgb = NULL;
5333 	uint8_t labc_to_use, quality;
5334 
5335 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5336 	tp->t_ccv.nsegs = nsegs;
5337 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5338 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5339 		uint32_t max;
5340 
5341 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5342 		if (tp->t_ccv.bytes_this_ack > max) {
5343 			tp->t_ccv.bytes_this_ack = max;
5344 		}
5345 	}
5346 #ifdef STATS
5347 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5348 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5349 #endif
5350 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5351 		/* We will ack all, time
5352 		 * to end any lt_bw_up we
5353 		 * have running until something
5354 		 * new is sent.
5355 		 */
5356 		struct timeval tv;
5357 
5358 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5359 		rack->r_ctl.lt_seq = tp->snd_max;
5360 		(void)tcp_get_usecs(&tv);
5361 		rack->r_ctl.lt_bw_time += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
5362 		rack->lt_bw_up = 0;
5363 	}
5364 	quality = RACK_QUALITY_NONE;
5365 	if ((tp->t_flags & TF_GPUTINPROG) &&
5366 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5367 		/* Measure the Goodput */
5368 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5369 	}
5370 	/* Which way our we limited, if not cwnd limited no advance in CA */
5371 	if (tp->snd_cwnd <= tp->snd_wnd)
5372 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5373 	else
5374 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5375 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5376 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5377 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5378 		/* For the setting of a window past use the actual scwnd we are using */
5379 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5380 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5381 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5382 		}
5383 	} else {
5384 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5385 		tp->t_bytes_acked = 0;
5386 	}
5387 	prior_cwnd = tp->snd_cwnd;
5388 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5389 	    (rack_client_low_buf && rack->client_bufferlvl &&
5390 	    (rack->client_bufferlvl < rack_client_low_buf)))
5391 		labc_to_use = rack->rc_labc;
5392 	else
5393 		labc_to_use = rack_max_abc_post_recovery;
5394 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5395 		union tcp_log_stackspecific log;
5396 		struct timeval tv;
5397 
5398 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5399 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5400 		log.u_bbr.flex1 = th_ack;
5401 		log.u_bbr.flex2 = tp->t_ccv.flags;
5402 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5403 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5404 		log.u_bbr.flex5 = labc_to_use;
5405 		log.u_bbr.flex6 = prior_cwnd;
5406 		log.u_bbr.flex7 = V_tcp_do_newsack;
5407 		log.u_bbr.flex8 = 1;
5408 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5409 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5410 	}
5411 	if (CC_ALGO(tp)->ack_received != NULL) {
5412 		/* XXXLAS: Find a way to live without this */
5413 		tp->t_ccv.curack = th_ack;
5414 		tp->t_ccv.labc = labc_to_use;
5415 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5416 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5417 	}
5418 	if (lgb) {
5419 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5420 	}
5421 	if (rack->r_must_retran) {
5422 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5423 			/*
5424 			 * We now are beyond the rxt point so lets disable
5425 			 * the flag.
5426 			 */
5427 			rack->r_ctl.rc_out_at_rto = 0;
5428 			rack->r_must_retran = 0;
5429 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5430 			/*
5431 			 * Only decrement the rc_out_at_rto if the cwnd advances
5432 			 * at least a whole segment. Otherwise next time the peer
5433 			 * acks, we won't be able to send this generaly happens
5434 			 * when we are in Congestion Avoidance.
5435 			 */
5436 			if (acked <= rack->r_ctl.rc_out_at_rto){
5437 				rack->r_ctl.rc_out_at_rto -= acked;
5438 			} else {
5439 				rack->r_ctl.rc_out_at_rto = 0;
5440 			}
5441 		}
5442 	}
5443 #ifdef STATS
5444 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5445 #endif
5446 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5447 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5448 	}
5449 }
5450 
5451 static void
5452 tcp_rack_partialack(struct tcpcb *tp)
5453 {
5454 	struct tcp_rack *rack;
5455 
5456 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5457 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5458 	/*
5459 	 * If we are doing PRR and have enough
5460 	 * room to send <or> we are pacing and prr
5461 	 * is disabled we will want to see if we
5462 	 * can send data (by setting r_wanted_output to
5463 	 * true).
5464 	 */
5465 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5466 	    rack->rack_no_prr)
5467 		rack->r_wanted_output = 1;
5468 }
5469 
5470 static inline void
5471 rack_set_most_aggr(struct tcp_rack *rack)
5472 {
5473 	rack->r_fill_less_agg = 0;
5474 	/* Once the cwnd as been clamped we don't do fill_cw */
5475 	if (rack->r_cwnd_was_clamped == 0)
5476 		rack->rc_pace_to_cwnd = 1;
5477 	rack->r_pacing_discount = 0;
5478 }
5479 
5480 static inline void
5481 rack_limit_fillcw(struct tcp_rack *rack)
5482 {
5483 	rack->r_fill_less_agg = 1;
5484 	/* Once the cwnd as been clamped we don't do fill_cw */
5485 	if (rack->r_cwnd_was_clamped == 0)
5486 		rack->rc_pace_to_cwnd = 1;
5487 	rack->r_pacing_discount = 0;
5488 }
5489 
5490 static inline void
5491 rack_disable_fillcw(struct tcp_rack *rack)
5492 {
5493 	rack->r_fill_less_agg = 1;
5494 	rack->rc_pace_to_cwnd = 0;
5495 	rack->r_pacing_discount = 0;
5496 }
5497 
5498 static void
5499 rack_client_buffer_level_set(struct tcp_rack *rack)
5500 {
5501 	/*
5502 	 * Only if DGP is on do we do anything that
5503 	 * changes stack behavior. If DGP is off all
5504 	 * we will do is issue a BB log (if BB logging is
5505 	 * on) and return.
5506 	 */
5507 	if (rack->dgp_on == 0) {
5508 		rack_log_pacing_delay_calc(rack, 0, rack->client_bufferlvl,
5509 					   0, 0, 0, 30, __LINE__, NULL, 0);
5510 		return;
5511 	}
5512 	if (IN_RECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.full_dgp_in_rec) {
5513 		goto set_most_agg;
5514 	}
5515 	/*
5516 	 * We are in DGP so what setting should we
5517 	 * apply based on where the client is?
5518 	 */
5519 	switch(rack->r_ctl.rc_dgp_bl_agg) {
5520 	default:
5521 	case DGP_LEVEL0:
5522 set_most_agg:
5523 		rack_set_most_aggr(rack);
5524 		break;
5525 	case DGP_LEVEL1:
5526 		if (rack->client_bufferlvl == 4)
5527 			rack_limit_fillcw(rack);
5528 		else if (rack->client_bufferlvl == 5)
5529 			rack_disable_fillcw(rack);
5530 		else
5531 			rack_set_most_aggr(rack);
5532 		break;
5533 	case DGP_LEVEL2:
5534 		if (rack->client_bufferlvl == 3)
5535 			rack_limit_fillcw(rack);
5536 		else if (rack->client_bufferlvl == 4)
5537 			rack_disable_fillcw(rack);
5538 		else if (rack->client_bufferlvl == 5) {
5539 			rack_disable_fillcw(rack);
5540 			rack->r_pacing_discount = 1;
5541 			rack->r_ctl.pacing_discount_amm = 1;
5542 		} else
5543 			rack_set_most_aggr(rack);
5544 		break;
5545 	case DGP_LEVEL3:
5546 		if (rack->client_bufferlvl == 2)
5547 			rack_limit_fillcw(rack);
5548 		else if (rack->client_bufferlvl == 3)
5549 			rack_disable_fillcw(rack);
5550 		else if (rack->client_bufferlvl == 4) {
5551 			rack_disable_fillcw(rack);
5552 			rack->r_pacing_discount = 1;
5553 			rack->r_ctl.pacing_discount_amm = 1;
5554 		} else if (rack->client_bufferlvl == 5) {
5555 			rack_disable_fillcw(rack);
5556 			rack->r_pacing_discount = 1;
5557 			rack->r_ctl.pacing_discount_amm = 2;
5558 		} else
5559 			rack_set_most_aggr(rack);
5560 		break;
5561 	}
5562 	rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_dgp_bl_agg, rack->client_bufferlvl, 0,
5563 				   0, 0, 30, __LINE__, NULL, 0);
5564 }
5565 
5566 static void
5567 do_rack_check_for_unclamp(struct tcpcb *tp, struct tcp_rack *rack)
5568 {
5569 	/*
5570 	 * Can we unclamp. We unclamp if more than
5571 	 * N rounds have transpired with no loss.
5572 	 */
5573 	uint64_t snds, rxts, rxt_per;
5574 	uint32_t rnds;
5575 
5576 	rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5577 	if ((rack_unclamp_round_thresh > 0) &&
5578 	    (rnds >= rack_unclamp_round_thresh)) {
5579 		snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5580 		KASSERT ((snds > 0), ("rack:%p tp:%p snds:%ju is 0", rack, tp,
5581 		    (uintmax_t)snds));
5582 		rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5583 		rxt_per = rxts * 1000;
5584 		rxt_per /= snds;
5585 		if ((uint32_t)rxt_per <= rack_unclamp_rxt_thresh) {
5586 			/* Unclamp */
5587 			if (tcp_bblogging_on(rack->rc_tp)) {
5588 				union tcp_log_stackspecific log;
5589 				struct timeval tv;
5590 
5591 				memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5592 				log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5593 				log.u_bbr.flex3 = rnds;
5594 				log.u_bbr.flex4 = rack_unclamp_round_thresh;
5595 				log.u_bbr.flex5 = (uint32_t)rxt_per;
5596 				log.u_bbr.flex8 = 6;
5597 				log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5598 				log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5599 				log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5600 				log.u_bbr.applimited = rack->r_ctl.max_clamps;
5601 				log.u_bbr.epoch = rack->r_ctl.clamp_options;
5602 				log.u_bbr.cur_del_rate = rxts;
5603 				log.u_bbr.bw_inuse = rack_get_lt_bw(rack);
5604 				log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5605 				log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5606 				log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5607 				tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5608 					      0, &log, false, NULL, NULL, 0, &tv);
5609 			}
5610 			rack->r_ctl.num_of_clamps_applied = 0;
5611 			rack->r_cwnd_was_clamped = 0;
5612 			rack->excess_rxt_on = 1;
5613 			if (rack->r_ctl.clamp_options) {
5614 				/*
5615 				 * We only allow fillcw to be toggled
5616 				 * if you are setting a max seg too.
5617 				 */
5618 				if (rack->r_ctl.clamp_options & 0x1) {
5619 					if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5620 						/* turn on fill cw  for non-dgp*/
5621 						rack->rc_pace_to_cwnd = 0;
5622 					} else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5623 						/* For DGP we want it off */
5624 						rack->rc_pace_to_cwnd = 1;
5625 					}
5626 				}
5627 			}
5628 			if (rack->dgp_on) {
5629 				/* Reset all multipliers to 100.0 so just the measured bw */
5630 				/* Crash any per boosts down to 100% */
5631 				rack->r_ctl.rack_per_of_gp_rec = 100;
5632 				rack->r_ctl.rack_per_of_gp_ss = 100;
5633 				rack->r_ctl.rack_per_of_gp_ca = 100;
5634 				/* Set in an upper bound for ss/ca % increase */
5635 				rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
5636 				rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
5637 			}
5638 		}
5639 	}
5640 }
5641 
5642 static void
5643 do_rack_excess_rxt(struct tcpcb *tp, struct tcp_rack *rack)
5644 {
5645 	/*
5646 	 * Rack excess rxt accounting is turned on. If we
5647 	 * are above a threshold of rxt's in at least N
5648 	 * rounds, then back off the cwnd and ssthresh
5649 	 * to fit into the long-term b/w.
5650 	 */
5651 	uint64_t snds, rxts, rxt_per, lt_bw, bdp;
5652 	uint32_t rnds, new_cwnd, new_ssthresh, rtt, shared_cwnd_was_enabled = 0;
5653 
5654 	/* Is it shut off by 0 rounds? */
5655 	if (rack_rxt_min_rnds == 0)
5656 		return;
5657 	if ((rack->r_ctl.max_clamps > 0) &&
5658 	    (rack->r_ctl.num_of_clamps_applied >= rack->r_ctl.max_clamps)) {
5659 		/*
5660 		 * The idea, if max_clamps is set, is that if clamping it
5661 		 * N times did not work again, then there is no sense
5662 		 * clamping it again. The link is just a lossy link and
5663 		 * our clamps are doing no good. Turn it off so we don't come
5664 		 * back here again.
5665 		 */
5666 		rack->excess_rxt_on = 0;
5667 		rack->r_cwnd_was_clamped = 0;
5668 		rack->r_ctl.num_of_clamps_applied = 0;
5669 		return;
5670 	}
5671 	snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5672 	rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5673 	rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5674 	/* Has enough rounds progressed for us to re-measure? */
5675 	if ((rnds >= rack_rxt_min_rnds) &&
5676 	    (rack->r_ctl.rxt_threshold > 0)){
5677 		rxt_per = rxts * 1000;
5678 		rxt_per /= snds;
5679 		if (rxt_per >= rack->r_ctl.rxt_threshold) {
5680 			/*
5681 			 * Action required:
5682 			 *  We are above our excess retransmit level, lets
5683 			 *  cut down the cwnd and ssthresh to match the long-term
5684 			 *  b/w we are getting.
5685 			 */
5686 			/* First disable scwnd if enabled */
5687 #ifdef NETFLIX_SHARED_CWND
5688 			rack->rack_enable_scwnd = 0;
5689 			if (rack->r_ctl.rc_scw) {
5690 				uint32_t limit;
5691 
5692 				shared_cwnd_was_enabled = 1;
5693 				if (rack->r_limit_scw)
5694 					limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
5695 				else
5696 					limit = 0;
5697 				tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
5698 							  rack->r_ctl.rc_scw_index,
5699 							  limit);
5700 				rack->r_ctl.rc_scw = NULL;
5701 			}
5702 
5703 #endif
5704 			/* Calculate what the cwnd and ssthresh should be */
5705 			tcp_trace_point(rack->rc_tp, TCP_TP_EXCESS_RXT);
5706 			lt_bw = rack_get_lt_bw(rack);
5707 			if (lt_bw == 0) {
5708 				/*
5709 				 * No lt_bw, lets chop things to one MSS
5710 				 * and the ssthresh to the iwnd.
5711 				 */
5712 reset_to_iw:
5713 				new_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5714 				new_ssthresh = tcp_compute_initwnd(tcp_maxseg(tp));
5715 			} else {
5716 				rtt = rack->rc_rack_rtt;
5717 				if (rtt == 0) {
5718 					/* If we have no rack_rtt drop to the IW situation */
5719 					goto reset_to_iw;
5720 				}
5721 				bdp = lt_bw * (uint64_t)rtt;
5722 				bdp /= HPTS_USEC_IN_SEC;
5723 				new_cwnd = (uint32_t)bdp;
5724 				new_ssthresh = new_cwnd - 1;
5725 				if (new_cwnd < ctf_fixed_maxseg(tp)) {
5726 					/* Rock bottom, goto IW settings  */
5727 					goto reset_to_iw;
5728 				}
5729 			}
5730 			rack->r_cwnd_was_clamped = 1;
5731 			rack->r_ctl.num_of_clamps_applied++;
5732 			/* Reset the counter fromn now */
5733 			tp->t_bytes_acked = 0;
5734 			/*
5735 			 * Now what about options?
5736 			 * We look at the bottom  8 bits:
5737 			 * F = fill cw bit (toggle it if set)
5738 			 * S = Segment bits
5739 			 * M = set max segment bit
5740 			 *
5741 			 * SSSS SSMF
5742 			 */
5743 			if (rack->r_ctl.clamp_options) {
5744 				if (rack->r_ctl.clamp_options & 0x1) {
5745 					if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5746 						/* turn on fill cw  for non-dgp*/
5747 						rack->rc_pace_to_cwnd = 1;
5748 					} else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5749 						/* For DGP we want it off */
5750 						rack->rc_pace_to_cwnd = 0;
5751 					}
5752 				}
5753 			}
5754 			if (rack->dgp_on) {
5755 				/* Reset all multipliers to 100.0 so just the measured bw */
5756 				/* Crash any per boosts down to 100% */
5757 				rack->r_ctl.rack_per_of_gp_rec = 100;
5758 				rack->r_ctl.rack_per_of_gp_ss = 100;
5759 				rack->r_ctl.rack_per_of_gp_ca = 100;
5760 				/* Set in an upper bound for ss/ca % increase */
5761 				rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_clamp_ss_upper;
5762 				rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_clamp_ca_upper;
5763 				/* Now move to the lt_bw */
5764 				rack->r_ctl.gp_bw = lt_bw;
5765 				rack->rc_gp_filled = 1;
5766 				rack->r_ctl.num_measurements = RACK_REQ_AVG;
5767 			}
5768 			if (tcp_bblogging_on(rack->rc_tp)) {
5769 				union tcp_log_stackspecific log;
5770 				struct timeval tv;
5771 
5772 				memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5773 				log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5774 				log.u_bbr.flex1 = new_cwnd;
5775 				log.u_bbr.flex2 = new_ssthresh;
5776 				log.u_bbr.flex3 = rnds;
5777 				log.u_bbr.flex4 = rack_rxt_min_rnds;
5778 				log.u_bbr.flex5 = rtt;
5779 				log.u_bbr.flex6 = shared_cwnd_was_enabled;
5780 				log.u_bbr.flex8 = 5;
5781 				log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5782 				log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5783 				log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5784 				log.u_bbr.applimited = rack->r_ctl.max_clamps;
5785 				log.u_bbr.epoch = rack->r_ctl.clamp_options;
5786 				log.u_bbr.cur_del_rate = rxts;
5787 				log.u_bbr.delRate = snds;
5788 				log.u_bbr.rttProp = rack->r_ctl.rxt_threshold;
5789 				log.u_bbr.bw_inuse = lt_bw;
5790 				log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5791 				log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5792 				log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5793 				tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5794 					       0, &log, false, NULL, NULL, 0, &tv);
5795 			}
5796 			/* Update our point where we did it */
5797 			if (rack->r_ctl.already_had_a_excess == 0) {
5798 				rack->r_ctl.already_had_a_excess = 1;
5799 				counter_u64_add(rack_rxt_clamps_cwnd_uniq, 1);
5800 			}
5801 			counter_u64_add(rack_rxt_clamps_cwnd, 1);
5802 			rack->r_ctl.last_sndbytes = tp->t_sndbytes;
5803 			rack->r_ctl.last_snd_rxt_bytes = tp->t_snd_rxt_bytes;
5804 			rack->r_ctl.last_rnd_rxt_clamped = rack->r_ctl.current_round;
5805 			if (new_cwnd < tp->snd_cwnd)
5806 				tp->snd_cwnd = new_cwnd;
5807 			if (new_ssthresh < tp->snd_ssthresh)
5808 				tp->snd_ssthresh = new_ssthresh;
5809 		}
5810 	}
5811 }
5812 
5813 static void
5814 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5815 {
5816 	struct tcp_rack *rack;
5817 	uint32_t orig_cwnd;
5818 
5819 	orig_cwnd = tp->snd_cwnd;
5820 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5821 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5822 	/* only alert CC if we alerted when we entered */
5823 	if (CC_ALGO(tp)->post_recovery != NULL) {
5824 		tp->t_ccv.curack = th_ack;
5825 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5826 		if (tp->snd_cwnd < tp->snd_ssthresh) {
5827 			/*
5828 			 * Rack has burst control and pacing
5829 			 * so lets not set this any lower than
5830 			 * snd_ssthresh per RFC-6582 (option 2).
5831 			 */
5832 			tp->snd_cwnd = tp->snd_ssthresh;
5833 		}
5834 	}
5835 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5836 		union tcp_log_stackspecific log;
5837 		struct timeval tv;
5838 
5839 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5840 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5841 		log.u_bbr.flex1 = th_ack;
5842 		log.u_bbr.flex2 = tp->t_ccv.flags;
5843 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5844 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5845 		log.u_bbr.flex5 = V_tcp_abc_l_var;
5846 		log.u_bbr.flex6 = orig_cwnd;
5847 		log.u_bbr.flex7 = V_tcp_do_newsack;
5848 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5849 		log.u_bbr.flex8 = 2;
5850 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5851 			       0, &log, false, NULL, __func__, __LINE__, &tv);
5852 	}
5853 	if ((rack->rack_no_prr == 0) &&
5854 	    (rack->no_prr_addback == 0) &&
5855 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
5856 		/*
5857 		 * Suck the next prr cnt back into cwnd, but
5858 		 * only do that if we are not application limited.
5859 		 */
5860 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5861 			/*
5862 			 * We are allowed to add back to the cwnd the amount we did
5863 			 * not get out if:
5864 			 * a) no_prr_addback is off.
5865 			 * b) we are not app limited
5866 			 * c) we are doing prr
5867 			 * <and>
5868 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5869 			 */
5870 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5871 					    rack->r_ctl.rc_prr_sndcnt);
5872 		}
5873 		rack->r_ctl.rc_prr_sndcnt = 0;
5874 		rack_log_to_prr(rack, 1, 0, __LINE__);
5875 	}
5876 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5877 	tp->snd_recover = tp->snd_una;
5878 	if (rack->r_ctl.dsack_persist) {
5879 		rack->r_ctl.dsack_persist--;
5880 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5881 			rack->r_ctl.num_dsack = 0;
5882 		}
5883 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5884 	}
5885 	EXIT_RECOVERY(tp->t_flags);
5886 	if (rack->r_ctl.full_dgp_in_rec)
5887 		rack_client_buffer_level_set(rack);
5888 }
5889 
5890 static void
5891 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5892 {
5893 	struct tcp_rack *rack;
5894 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5895 
5896 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5897 #ifdef STATS
5898 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5899 #endif
5900 	if (IN_RECOVERY(tp->t_flags) == 0) {
5901 		in_rec_at_entry = 0;
5902 		ssthresh_enter = tp->snd_ssthresh;
5903 		cwnd_enter = tp->snd_cwnd;
5904 	} else
5905 		in_rec_at_entry = 1;
5906 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5907 	switch (type) {
5908 	case CC_NDUPACK:
5909 		tp->t_flags &= ~TF_WASFRECOVERY;
5910 		tp->t_flags &= ~TF_WASCRECOVERY;
5911 		if (!IN_FASTRECOVERY(tp->t_flags)) {
5912 			if (rack->dgp_on && rack->r_cwnd_was_clamped) {
5913 				/* Reset the gains so that on exit we will be softer longer */
5914 				rack->r_ctl.rack_per_of_gp_rec = 100;
5915 				rack->r_ctl.rack_per_of_gp_ss = 98;
5916 				rack->r_ctl.rack_per_of_gp_ca = 98;
5917 			}
5918 			rack->r_ctl.rc_prr_delivered = 0;
5919 			rack->r_ctl.rc_prr_out = 0;
5920 			rack->r_fast_output = 0;
5921 			if (rack->rack_no_prr == 0) {
5922 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5923 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5924 			}
5925 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5926 			tp->snd_recover = tp->snd_max;
5927 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5928 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5929 		}
5930 		break;
5931 	case CC_ECN:
5932 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5933 		    /*
5934 		     * Allow ECN reaction on ACK to CWR, if
5935 		     * that data segment was also CE marked.
5936 		     */
5937 		    SEQ_GEQ(ack, tp->snd_recover)) {
5938 			EXIT_CONGRECOVERY(tp->t_flags);
5939 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5940 			rack->r_fast_output = 0;
5941 			tp->snd_recover = tp->snd_max + 1;
5942 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5943 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5944 		}
5945 		break;
5946 	case CC_RTO:
5947 		tp->t_dupacks = 0;
5948 		tp->t_bytes_acked = 0;
5949 		rack->r_fast_output = 0;
5950 		EXIT_RECOVERY(tp->t_flags);
5951 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5952 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5953 		orig_cwnd = tp->snd_cwnd;
5954 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
5955 		rack_log_to_prr(rack, 16, orig_cwnd, line);
5956 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5957 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5958 		break;
5959 	case CC_RTO_ERR:
5960 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5961 		/* RTO was unnecessary, so reset everything. */
5962 		tp->snd_cwnd = tp->snd_cwnd_prev;
5963 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5964 		tp->snd_recover = tp->snd_recover_prev;
5965 		if (tp->t_flags & TF_WASFRECOVERY) {
5966 			ENTER_FASTRECOVERY(tp->t_flags);
5967 			tp->t_flags &= ~TF_WASFRECOVERY;
5968 		}
5969 		if (tp->t_flags & TF_WASCRECOVERY) {
5970 			ENTER_CONGRECOVERY(tp->t_flags);
5971 			tp->t_flags &= ~TF_WASCRECOVERY;
5972 		}
5973 		tp->snd_nxt = tp->snd_max;
5974 		tp->t_badrxtwin = 0;
5975 		break;
5976 	}
5977 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5978 	    (type != CC_RTO)){
5979 		tp->t_ccv.curack = ack;
5980 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5981 	}
5982 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5983 		rack_log_to_prr(rack, 15, cwnd_enter, line);
5984 		if (rack->r_ctl.full_dgp_in_rec)
5985 			rack_client_buffer_level_set(rack);
5986 		rack->r_ctl.dsack_byte_cnt = 0;
5987 		rack->r_ctl.retran_during_recovery = 0;
5988 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5989 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5990 		rack->r_ent_rec_ns = 1;
5991 	}
5992 }
5993 
5994 static inline void
5995 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5996 {
5997 	uint32_t i_cwnd;
5998 
5999 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6000 
6001 	if (CC_ALGO(tp)->after_idle != NULL)
6002 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
6003 
6004 	if (tp->snd_cwnd == 1)
6005 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
6006 	else
6007 		i_cwnd = rc_init_window(rack);
6008 
6009 	/*
6010 	 * Being idle is no different than the initial window. If the cc
6011 	 * clamps it down below the initial window raise it to the initial
6012 	 * window.
6013 	 */
6014 	if (tp->snd_cwnd < i_cwnd) {
6015 		tp->snd_cwnd = i_cwnd;
6016 	}
6017 }
6018 
6019 /*
6020  * Indicate whether this ack should be delayed.  We can delay the ack if
6021  * following conditions are met:
6022  *	- There is no delayed ack timer in progress.
6023  *	- Our last ack wasn't a 0-sized window. We never want to delay
6024  *	  the ack that opens up a 0-sized window.
6025  *	- LRO wasn't used for this segment. We make sure by checking that the
6026  *	  segment size is not larger than the MSS.
6027  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
6028  *	  connection.
6029  */
6030 #define DELAY_ACK(tp, tlen)			 \
6031 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
6032 	((tp->t_flags & TF_DELACK) == 0) &&	 \
6033 	(tlen <= tp->t_maxseg) &&		 \
6034 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
6035 
6036 static struct rack_sendmap *
6037 rack_find_lowest_rsm(struct tcp_rack *rack)
6038 {
6039 	struct rack_sendmap *rsm;
6040 
6041 	/*
6042 	 * Walk the time-order transmitted list looking for an rsm that is
6043 	 * not acked. This will be the one that was sent the longest time
6044 	 * ago that is still outstanding.
6045 	 */
6046 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6047 		if (rsm->r_flags & RACK_ACKED) {
6048 			continue;
6049 		}
6050 		goto finish;
6051 	}
6052 finish:
6053 	return (rsm);
6054 }
6055 
6056 static struct rack_sendmap *
6057 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
6058 {
6059 	struct rack_sendmap *prsm;
6060 
6061 	/*
6062 	 * Walk the sequence order list backward until we hit and arrive at
6063 	 * the highest seq not acked. In theory when this is called it
6064 	 * should be the last segment (which it was not).
6065 	 */
6066 	prsm = rsm;
6067 
6068 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
6069 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
6070 			continue;
6071 		}
6072 		return (prsm);
6073 	}
6074 	return (NULL);
6075 }
6076 
6077 static uint32_t
6078 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
6079 {
6080 	int32_t lro;
6081 	uint32_t thresh;
6082 
6083 	/*
6084 	 * lro is the flag we use to determine if we have seen reordering.
6085 	 * If it gets set we have seen reordering. The reorder logic either
6086 	 * works in one of two ways:
6087 	 *
6088 	 * If reorder-fade is configured, then we track the last time we saw
6089 	 * re-ordering occur. If we reach the point where enough time as
6090 	 * passed we no longer consider reordering has occuring.
6091 	 *
6092 	 * Or if reorder-face is 0, then once we see reordering we consider
6093 	 * the connection to alway be subject to reordering and just set lro
6094 	 * to 1.
6095 	 *
6096 	 * In the end if lro is non-zero we add the extra time for
6097 	 * reordering in.
6098 	 */
6099 	if (srtt == 0)
6100 		srtt = 1;
6101 	if (rack->r_ctl.rc_reorder_ts) {
6102 		if (rack->r_ctl.rc_reorder_fade) {
6103 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
6104 				lro = cts - rack->r_ctl.rc_reorder_ts;
6105 				if (lro == 0) {
6106 					/*
6107 					 * No time as passed since the last
6108 					 * reorder, mark it as reordering.
6109 					 */
6110 					lro = 1;
6111 				}
6112 			} else {
6113 				/* Negative time? */
6114 				lro = 0;
6115 			}
6116 			if (lro > rack->r_ctl.rc_reorder_fade) {
6117 				/* Turn off reordering seen too */
6118 				rack->r_ctl.rc_reorder_ts = 0;
6119 				lro = 0;
6120 			}
6121 		} else {
6122 			/* Reodering does not fade */
6123 			lro = 1;
6124 		}
6125 	} else {
6126 		lro = 0;
6127 	}
6128 	if (rack->rc_rack_tmr_std_based == 0) {
6129 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
6130 	} else {
6131 		/* Standards based pkt-delay is 1/4 srtt */
6132 		thresh = srtt +  (srtt >> 2);
6133 	}
6134 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
6135 		/* It must be set, if not you get 1/4 rtt */
6136 		if (rack->r_ctl.rc_reorder_shift)
6137 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
6138 		else
6139 			thresh += (srtt >> 2);
6140 	}
6141 	if (rack->rc_rack_use_dsack &&
6142 	    lro &&
6143 	    (rack->r_ctl.num_dsack > 0)) {
6144 		/*
6145 		 * We only increase the reordering window if we
6146 		 * have seen reordering <and> we have a DSACK count.
6147 		 */
6148 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
6149 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
6150 	}
6151 	/* SRTT * 2 is the ceiling */
6152 	if (thresh > (srtt * 2)) {
6153 		thresh = srtt * 2;
6154 	}
6155 	/* And we don't want it above the RTO max either */
6156 	if (thresh > rack_rto_max) {
6157 		thresh = rack_rto_max;
6158 	}
6159 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
6160 	return (thresh);
6161 }
6162 
6163 static uint32_t
6164 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6165 		     struct rack_sendmap *rsm, uint32_t srtt)
6166 {
6167 	struct rack_sendmap *prsm;
6168 	uint32_t thresh, len;
6169 	int segsiz;
6170 
6171 	if (srtt == 0)
6172 		srtt = 1;
6173 	if (rack->r_ctl.rc_tlp_threshold)
6174 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6175 	else
6176 		thresh = (srtt * 2);
6177 
6178 	/* Get the previous sent packet, if any */
6179 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6180 	len = rsm->r_end - rsm->r_start;
6181 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6182 		/* Exactly like the ID */
6183 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6184 			uint32_t alt_thresh;
6185 			/*
6186 			 * Compensate for delayed-ack with the d-ack time.
6187 			 */
6188 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6189 			if (alt_thresh > thresh)
6190 				thresh = alt_thresh;
6191 		}
6192 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6193 		/* 2.1 behavior */
6194 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6195 		if (prsm && (len <= segsiz)) {
6196 			/*
6197 			 * Two packets outstanding, thresh should be (2*srtt) +
6198 			 * possible inter-packet delay (if any).
6199 			 */
6200 			uint32_t inter_gap = 0;
6201 			int idx, nidx;
6202 
6203 			idx = rsm->r_rtr_cnt - 1;
6204 			nidx = prsm->r_rtr_cnt - 1;
6205 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6206 				/* Yes it was sent later (or at the same time) */
6207 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6208 			}
6209 			thresh += inter_gap;
6210 		} else if (len <= segsiz) {
6211 			/*
6212 			 * Possibly compensate for delayed-ack.
6213 			 */
6214 			uint32_t alt_thresh;
6215 
6216 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6217 			if (alt_thresh > thresh)
6218 				thresh = alt_thresh;
6219 		}
6220 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6221 		/* 2.2 behavior */
6222 		if (len <= segsiz) {
6223 			uint32_t alt_thresh;
6224 			/*
6225 			 * Compensate for delayed-ack with the d-ack time.
6226 			 */
6227 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6228 			if (alt_thresh > thresh)
6229 				thresh = alt_thresh;
6230 		}
6231 	}
6232 	/* Not above an RTO */
6233 	if (thresh > tp->t_rxtcur) {
6234 		thresh = tp->t_rxtcur;
6235 	}
6236 	/* Not above a RTO max */
6237 	if (thresh > rack_rto_max) {
6238 		thresh = rack_rto_max;
6239 	}
6240 	/* Apply user supplied min TLP */
6241 	if (thresh < rack_tlp_min) {
6242 		thresh = rack_tlp_min;
6243 	}
6244 	return (thresh);
6245 }
6246 
6247 static uint32_t
6248 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6249 {
6250 	/*
6251 	 * We want the rack_rtt which is the
6252 	 * last rtt we measured. However if that
6253 	 * does not exist we fallback to the srtt (which
6254 	 * we probably will never do) and then as a last
6255 	 * resort we use RACK_INITIAL_RTO if no srtt is
6256 	 * yet set.
6257 	 */
6258 	if (rack->rc_rack_rtt)
6259 		return (rack->rc_rack_rtt);
6260 	else if (tp->t_srtt == 0)
6261 		return (RACK_INITIAL_RTO);
6262 	return (tp->t_srtt);
6263 }
6264 
6265 static struct rack_sendmap *
6266 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6267 {
6268 	/*
6269 	 * Check to see that we don't need to fall into recovery. We will
6270 	 * need to do so if our oldest transmit is past the time we should
6271 	 * have had an ack.
6272 	 */
6273 	struct tcp_rack *rack;
6274 	struct rack_sendmap *rsm;
6275 	int32_t idx;
6276 	uint32_t srtt, thresh;
6277 
6278 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6279 	if (tqhash_empty(rack->r_ctl.tqh)) {
6280 		return (NULL);
6281 	}
6282 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6283 	if (rsm == NULL)
6284 		return (NULL);
6285 
6286 
6287 	if (rsm->r_flags & RACK_ACKED) {
6288 		rsm = rack_find_lowest_rsm(rack);
6289 		if (rsm == NULL)
6290 			return (NULL);
6291 	}
6292 	idx = rsm->r_rtr_cnt - 1;
6293 	srtt = rack_grab_rtt(tp, rack);
6294 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
6295 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6296 		return (NULL);
6297 	}
6298 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6299 		return (NULL);
6300 	}
6301 	/* Ok if we reach here we are over-due and this guy can be sent */
6302 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6303 	return (rsm);
6304 }
6305 
6306 static uint32_t
6307 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6308 {
6309 	int32_t t;
6310 	int32_t tt;
6311 	uint32_t ret_val;
6312 
6313 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6314 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6315  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6316 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6317 	ret_val = (uint32_t)tt;
6318 	return (ret_val);
6319 }
6320 
6321 static uint32_t
6322 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6323 {
6324 	/*
6325 	 * Start the FR timer, we do this based on getting the first one in
6326 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6327 	 * events we need to stop the running timer (if its running) before
6328 	 * starting the new one.
6329 	 */
6330 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6331 	uint32_t srtt_cur;
6332 	int32_t idx;
6333 	int32_t is_tlp_timer = 0;
6334 	struct rack_sendmap *rsm;
6335 
6336 	if (rack->t_timers_stopped) {
6337 		/* All timers have been stopped none are to run */
6338 		return (0);
6339 	}
6340 	if (rack->rc_in_persist) {
6341 		/* We can't start any timer in persists */
6342 		return (rack_get_persists_timer_val(tp, rack));
6343 	}
6344 	rack->rc_on_min_to = 0;
6345 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6346 	    (rack->sack_attack_disable > 0) ||
6347 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6348 		goto activate_rxt;
6349 	}
6350 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6351 	if ((rsm == NULL) || sup_rack) {
6352 		/* Nothing on the send map or no rack */
6353 activate_rxt:
6354 		time_since_sent = 0;
6355 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6356 		if (rsm) {
6357 			/*
6358 			 * Should we discount the RTX timer any?
6359 			 *
6360 			 * We want to discount it the smallest amount.
6361 			 * If a timer (Rack/TLP or RXT) has gone off more
6362 			 * recently thats the discount we want to use (now - timer time).
6363 			 * If the retransmit of the oldest packet was more recent then
6364 			 * we want to use that (now - oldest-packet-last_transmit_time).
6365 			 *
6366 			 */
6367 			idx = rsm->r_rtr_cnt - 1;
6368 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6369 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6370 			else
6371 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6372 			if (TSTMP_GT(cts, tstmp_touse))
6373 			    time_since_sent = cts - tstmp_touse;
6374 		}
6375 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6376 		    sbavail(&tptosocket(tp)->so_snd)) {
6377 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6378 			to = tp->t_rxtcur;
6379 			if (to > time_since_sent)
6380 				to -= time_since_sent;
6381 			else
6382 				to = rack->r_ctl.rc_min_to;
6383 			if (to == 0)
6384 				to = 1;
6385 			/* Special case for KEEPINIT */
6386 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6387 			    (TP_KEEPINIT(tp) != 0) &&
6388 			    rsm) {
6389 				/*
6390 				 * We have to put a ceiling on the rxt timer
6391 				 * of the keep-init timeout.
6392 				 */
6393 				uint32_t max_time, red;
6394 
6395 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6396 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6397 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6398 					if (red < max_time)
6399 						max_time -= red;
6400 					else
6401 						max_time = 1;
6402 				}
6403 				/* Reduce timeout to the keep value if needed */
6404 				if (max_time < to)
6405 					to = max_time;
6406 			}
6407 			return (to);
6408 		}
6409 		return (0);
6410 	}
6411 	if (rsm->r_flags & RACK_ACKED) {
6412 		rsm = rack_find_lowest_rsm(rack);
6413 		if (rsm == NULL) {
6414 			/* No lowest? */
6415 			goto activate_rxt;
6416 		}
6417 	}
6418 	if (rack->sack_attack_disable) {
6419 		/*
6420 		 * We don't want to do
6421 		 * any TLP's if you are an attacker.
6422 		 * Though if you are doing what
6423 		 * is expected you may still have
6424 		 * SACK-PASSED marks.
6425 		 */
6426 		goto activate_rxt;
6427 	}
6428 	/* Convert from ms to usecs */
6429 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6430 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6431 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6432 		if ((tp->t_flags & TF_SENTFIN) &&
6433 		    ((tp->snd_max - tp->snd_una) == 1) &&
6434 		    (rsm->r_flags & RACK_HAS_FIN)) {
6435 			/*
6436 			 * We don't start a rack timer if all we have is a
6437 			 * FIN outstanding.
6438 			 */
6439 			goto activate_rxt;
6440 		}
6441 		if ((rack->use_rack_rr == 0) &&
6442 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6443 		    (rack->rack_no_prr == 0) &&
6444 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6445 			/*
6446 			 * We are not cheating, in recovery  and
6447 			 * not enough ack's to yet get our next
6448 			 * retransmission out.
6449 			 *
6450 			 * Note that classified attackers do not
6451 			 * get to use the rack-cheat.
6452 			 */
6453 			goto activate_tlp;
6454 		}
6455 		srtt = rack_grab_rtt(tp, rack);
6456 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
6457 		idx = rsm->r_rtr_cnt - 1;
6458 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6459 		if (SEQ_GEQ(exp, cts)) {
6460 			to = exp - cts;
6461 			if (to < rack->r_ctl.rc_min_to) {
6462 				to = rack->r_ctl.rc_min_to;
6463 				if (rack->r_rr_config == 3)
6464 					rack->rc_on_min_to = 1;
6465 			}
6466 		} else {
6467 			to = rack->r_ctl.rc_min_to;
6468 			if (rack->r_rr_config == 3)
6469 				rack->rc_on_min_to = 1;
6470 		}
6471 	} else {
6472 		/* Ok we need to do a TLP not RACK */
6473 activate_tlp:
6474 		if ((rack->rc_tlp_in_progress != 0) &&
6475 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6476 			/*
6477 			 * The previous send was a TLP and we have sent
6478 			 * N TLP's without sending new data.
6479 			 */
6480 			goto activate_rxt;
6481 		}
6482 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6483 		if (rsm == NULL) {
6484 			/* We found no rsm to TLP with. */
6485 			goto activate_rxt;
6486 		}
6487 		if (rsm->r_flags & RACK_HAS_FIN) {
6488 			/* If its a FIN we dont do TLP */
6489 			rsm = NULL;
6490 			goto activate_rxt;
6491 		}
6492 		idx = rsm->r_rtr_cnt - 1;
6493 		time_since_sent = 0;
6494 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6495 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6496 		else
6497 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6498 		if (TSTMP_GT(cts, tstmp_touse))
6499 		    time_since_sent = cts - tstmp_touse;
6500 		is_tlp_timer = 1;
6501 		if (tp->t_srtt) {
6502 			if ((rack->rc_srtt_measure_made == 0) &&
6503 			    (tp->t_srtt == 1)) {
6504 				/*
6505 				 * If another stack as run and set srtt to 1,
6506 				 * then the srtt was 0, so lets use the initial.
6507 				 */
6508 				srtt = RACK_INITIAL_RTO;
6509 			} else {
6510 				srtt_cur = tp->t_srtt;
6511 				srtt = srtt_cur;
6512 			}
6513 		} else
6514 			srtt = RACK_INITIAL_RTO;
6515 		/*
6516 		 * If the SRTT is not keeping up and the
6517 		 * rack RTT has spiked we want to use
6518 		 * the last RTT not the smoothed one.
6519 		 */
6520 		if (rack_tlp_use_greater &&
6521 		    tp->t_srtt &&
6522 		    (srtt < rack_grab_rtt(tp, rack))) {
6523 			srtt = rack_grab_rtt(tp, rack);
6524 		}
6525 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6526 		if (thresh > time_since_sent) {
6527 			to = thresh - time_since_sent;
6528 		} else {
6529 			to = rack->r_ctl.rc_min_to;
6530 			rack_log_alt_to_to_cancel(rack,
6531 						  thresh,		/* flex1 */
6532 						  time_since_sent,	/* flex2 */
6533 						  tstmp_touse,		/* flex3 */
6534 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6535 						  (uint32_t)rsm->r_tim_lastsent[idx],
6536 						  srtt,
6537 						  idx, 99);
6538 		}
6539 		if (to < rack_tlp_min) {
6540 			to = rack_tlp_min;
6541 		}
6542 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6543 			/*
6544 			 * If the TLP time works out to larger than the max
6545 			 * RTO lets not do TLP.. just RTO.
6546 			 */
6547 			goto activate_rxt;
6548 		}
6549 	}
6550 	if (is_tlp_timer == 0) {
6551 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6552 	} else {
6553 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6554 	}
6555 	if (to == 0)
6556 		to = 1;
6557 	return (to);
6558 }
6559 
6560 static void
6561 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6562 {
6563 	struct timeval tv;
6564 
6565 	if (rack->rc_in_persist == 0) {
6566 		if (tp->t_flags & TF_GPUTINPROG) {
6567 			/*
6568 			 * Stop the goodput now, the calling of the
6569 			 * measurement function clears the flag.
6570 			 */
6571 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6572 						    RACK_QUALITY_PERSIST);
6573 		}
6574 #ifdef NETFLIX_SHARED_CWND
6575 		if (rack->r_ctl.rc_scw) {
6576 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6577 			rack->rack_scwnd_is_idle = 1;
6578 		}
6579 #endif
6580 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(&tv);
6581 		if (rack->lt_bw_up) {
6582 			/* Suspend our LT BW measurement */
6583 			uint64_t tmark;
6584 
6585 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6586 			rack->r_ctl.lt_seq = snd_una;
6587 			tmark = tcp_tv_to_lusectick(&tv);
6588 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6589 			rack->r_ctl.lt_timemark = tmark;
6590 			rack->lt_bw_up = 0;
6591 			rack->r_persist_lt_bw_off = 1;
6592 		}
6593 		if (rack->r_ctl.rc_went_idle_time == 0)
6594 			rack->r_ctl.rc_went_idle_time = 1;
6595 		rack_timer_cancel(tp, rack, cts, __LINE__);
6596 		rack->r_ctl.persist_lost_ends = 0;
6597 		rack->probe_not_answered = 0;
6598 		rack->forced_ack = 0;
6599 		tp->t_rxtshift = 0;
6600 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6601 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6602 		rack->rc_in_persist = 1;
6603 	}
6604 }
6605 
6606 static void
6607 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6608 {
6609 	struct timeval tv;
6610 	uint32_t t_time;
6611 
6612 	if (tcp_in_hpts(rack->rc_tp)) {
6613 		tcp_hpts_remove(rack->rc_tp);
6614 		rack->r_ctl.rc_hpts_flags = 0;
6615 	}
6616 #ifdef NETFLIX_SHARED_CWND
6617 	if (rack->r_ctl.rc_scw) {
6618 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6619 		rack->rack_scwnd_is_idle = 0;
6620 	}
6621 #endif
6622 	t_time = tcp_get_usecs(&tv);
6623 	if (rack->rc_gp_dyn_mul &&
6624 	    (rack->use_fixed_rate == 0) &&
6625 	    (rack->rc_always_pace)) {
6626 		/*
6627 		 * Do we count this as if a probe-rtt just
6628 		 * finished?
6629 		 */
6630 		uint32_t time_idle, idle_min;
6631 
6632 		time_idle = t_time - rack->r_ctl.rc_went_idle_time;
6633 		idle_min = rack_min_probertt_hold;
6634 		if (rack_probertt_gpsrtt_cnt_div) {
6635 			uint64_t extra;
6636 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6637 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
6638 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6639 			idle_min += (uint32_t)extra;
6640 		}
6641 		if (time_idle >= idle_min) {
6642 			/* Yes, we count it as a probe-rtt. */
6643 			uint32_t us_cts;
6644 
6645 			us_cts = tcp_get_usecs(NULL);
6646 			if (rack->in_probe_rtt == 0) {
6647 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6648 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6649 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6650 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6651 			} else {
6652 				rack_exit_probertt(rack, us_cts);
6653 			}
6654 		}
6655 	}
6656 	if (rack->r_persist_lt_bw_off) {
6657 		/* Continue where we left off */
6658 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
6659 		rack->lt_bw_up = 1;
6660 		rack->r_persist_lt_bw_off = 0;
6661 	}
6662 	rack->rc_in_persist = 0;
6663 	rack->r_ctl.rc_went_idle_time = 0;
6664 	tp->t_rxtshift = 0;
6665 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6666 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6667 	rack->r_ctl.rc_agg_delayed = 0;
6668 	rack->r_early = 0;
6669 	rack->r_late = 0;
6670 	rack->r_ctl.rc_agg_early = 0;
6671 }
6672 
6673 static void
6674 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6675 		   struct hpts_diag *diag, struct timeval *tv)
6676 {
6677 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6678 		union tcp_log_stackspecific log;
6679 
6680 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6681 		log.u_bbr.flex1 = diag->p_nxt_slot;
6682 		log.u_bbr.flex2 = diag->p_cur_slot;
6683 		log.u_bbr.flex3 = diag->slot_req;
6684 		log.u_bbr.flex4 = diag->inp_hptsslot;
6685 		log.u_bbr.flex5 = diag->slot_remaining;
6686 		log.u_bbr.flex6 = diag->need_new_to;
6687 		log.u_bbr.flex7 = diag->p_hpts_active;
6688 		log.u_bbr.flex8 = diag->p_on_min_sleep;
6689 		/* Hijack other fields as needed */
6690 		log.u_bbr.epoch = diag->have_slept;
6691 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
6692 		log.u_bbr.pkts_out = diag->co_ret;
6693 		log.u_bbr.applimited = diag->hpts_sleep_time;
6694 		log.u_bbr.delivered = diag->p_prev_slot;
6695 		log.u_bbr.inflight = diag->p_runningslot;
6696 		log.u_bbr.bw_inuse = diag->wheel_slot;
6697 		log.u_bbr.rttProp = diag->wheel_cts;
6698 		log.u_bbr.timeStamp = cts;
6699 		log.u_bbr.delRate = diag->maxslots;
6700 		log.u_bbr.cur_del_rate = diag->p_curtick;
6701 		log.u_bbr.cur_del_rate <<= 32;
6702 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
6703 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6704 		    &rack->rc_inp->inp_socket->so_rcv,
6705 		    &rack->rc_inp->inp_socket->so_snd,
6706 		    BBR_LOG_HPTSDIAG, 0,
6707 		    0, &log, false, tv);
6708 	}
6709 
6710 }
6711 
6712 static void
6713 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6714 {
6715 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6716 		union tcp_log_stackspecific log;
6717 		struct timeval tv;
6718 
6719 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6720 		log.u_bbr.flex1 = sb->sb_flags;
6721 		log.u_bbr.flex2 = len;
6722 		log.u_bbr.flex3 = sb->sb_state;
6723 		log.u_bbr.flex8 = type;
6724 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6725 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6726 		    &rack->rc_inp->inp_socket->so_rcv,
6727 		    &rack->rc_inp->inp_socket->so_snd,
6728 		    TCP_LOG_SB_WAKE, 0,
6729 		    len, &log, false, &tv);
6730 	}
6731 }
6732 
6733 static void
6734 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6735       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6736 {
6737 	struct hpts_diag diag;
6738 	struct inpcb *inp = tptoinpcb(tp);
6739 	struct timeval tv;
6740 	uint32_t delayed_ack = 0;
6741 	uint32_t hpts_timeout;
6742 	uint32_t entry_slot = slot;
6743 	uint8_t stopped;
6744 	uint32_t left = 0;
6745 	uint32_t us_cts;
6746 
6747 	if ((tp->t_state == TCPS_CLOSED) ||
6748 	    (tp->t_state == TCPS_LISTEN)) {
6749 		return;
6750 	}
6751 	if (tcp_in_hpts(tp)) {
6752 		/* Already on the pacer */
6753 		return;
6754 	}
6755 	stopped = rack->rc_tmr_stopped;
6756 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6757 		left = rack->r_ctl.rc_timer_exp - cts;
6758 	}
6759 	rack->r_ctl.rc_timer_exp = 0;
6760 	rack->r_ctl.rc_hpts_flags = 0;
6761 	us_cts = tcp_get_usecs(&tv);
6762 	/* Now early/late accounting */
6763 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6764 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6765 		/*
6766 		 * We have a early carry over set,
6767 		 * we can always add more time so we
6768 		 * can always make this compensation.
6769 		 *
6770 		 * Note if ack's are allowed to wake us do not
6771 		 * penalize the next timer for being awoke
6772 		 * by an ack aka the rc_agg_early (non-paced mode).
6773 		 */
6774 		slot += rack->r_ctl.rc_agg_early;
6775 		rack->r_early = 0;
6776 		rack->r_ctl.rc_agg_early = 0;
6777 	}
6778 	if (rack->r_late) {
6779 		/*
6780 		 * This is harder, we can
6781 		 * compensate some but it
6782 		 * really depends on what
6783 		 * the current pacing time is.
6784 		 */
6785 		if (rack->r_ctl.rc_agg_delayed >= slot) {
6786 			/*
6787 			 * We can't compensate for it all.
6788 			 * And we have to have some time
6789 			 * on the clock. We always have a min
6790 			 * 10 slots (10 x 10 i.e. 100 usecs).
6791 			 */
6792 			if (slot <= HPTS_TICKS_PER_SLOT) {
6793 				/* We gain delay */
6794 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6795 				slot = HPTS_TICKS_PER_SLOT;
6796 			} else {
6797 				/* We take off some */
6798 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6799 				slot = HPTS_TICKS_PER_SLOT;
6800 			}
6801 		} else {
6802 			slot -= rack->r_ctl.rc_agg_delayed;
6803 			rack->r_ctl.rc_agg_delayed = 0;
6804 			/* Make sure we have 100 useconds at minimum */
6805 			if (slot < HPTS_TICKS_PER_SLOT) {
6806 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6807 				slot = HPTS_TICKS_PER_SLOT;
6808 			}
6809 			if (rack->r_ctl.rc_agg_delayed == 0)
6810 				rack->r_late = 0;
6811 		}
6812 	}
6813 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6814 #ifdef TCP_SAD_DETECTION
6815 	if (rack->sack_attack_disable &&
6816 	    (rack->r_ctl.ack_during_sd > 0) &&
6817 	    (slot < tcp_sad_pacing_interval)) {
6818 		/*
6819 		 * We have a potential attacker on
6820 		 * the line. We have possibly some
6821 		 * (or now) pacing time set. We want to
6822 		 * slow down the processing of sacks by some
6823 		 * amount (if it is an attacker). Set the default
6824 		 * slot for attackers in place (unless the original
6825 		 * interval is longer). Its stored in
6826 		 * micro-seconds, so lets convert to msecs.
6827 		 */
6828 		slot = tcp_sad_pacing_interval;
6829 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6830 		rack->r_ctl.ack_during_sd = 0;
6831 	}
6832 #endif
6833 	if (tp->t_flags & TF_DELACK) {
6834 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
6835 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6836 	}
6837 	if (delayed_ack && ((hpts_timeout == 0) ||
6838 			    (delayed_ack < hpts_timeout)))
6839 		hpts_timeout = delayed_ack;
6840 	else
6841 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6842 	/*
6843 	 * If no timers are going to run and we will fall off the hptsi
6844 	 * wheel, we resort to a keep-alive timer if its configured.
6845 	 */
6846 	if ((hpts_timeout == 0) &&
6847 	    (slot == 0)) {
6848 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6849 		    (tp->t_state <= TCPS_CLOSING)) {
6850 			/*
6851 			 * Ok we have no timer (persists, rack, tlp, rxt  or
6852 			 * del-ack), we don't have segments being paced. So
6853 			 * all that is left is the keepalive timer.
6854 			 */
6855 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6856 				/* Get the established keep-alive time */
6857 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6858 			} else {
6859 				/*
6860 				 * Get the initial setup keep-alive time,
6861 				 * note that this is probably not going to
6862 				 * happen, since rack will be running a rxt timer
6863 				 * if a SYN of some sort is outstanding. It is
6864 				 * actually handled in rack_timeout_rxt().
6865 				 */
6866 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6867 			}
6868 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6869 			if (rack->in_probe_rtt) {
6870 				/*
6871 				 * We want to instead not wake up a long time from
6872 				 * now but to wake up about the time we would
6873 				 * exit probe-rtt and initiate a keep-alive ack.
6874 				 * This will get us out of probe-rtt and update
6875 				 * our min-rtt.
6876 				 */
6877 				hpts_timeout = rack_min_probertt_hold;
6878 			}
6879 		}
6880 	}
6881 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6882 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6883 		/*
6884 		 * RACK, TLP, persists and RXT timers all are restartable
6885 		 * based on actions input .. i.e we received a packet (ack
6886 		 * or sack) and that changes things (rw, or snd_una etc).
6887 		 * Thus we can restart them with a new value. For
6888 		 * keep-alive, delayed_ack we keep track of what was left
6889 		 * and restart the timer with a smaller value.
6890 		 */
6891 		if (left < hpts_timeout)
6892 			hpts_timeout = left;
6893 	}
6894 	if (hpts_timeout) {
6895 		/*
6896 		 * Hack alert for now we can't time-out over 2,147,483
6897 		 * seconds (a bit more than 596 hours), which is probably ok
6898 		 * :).
6899 		 */
6900 		if (hpts_timeout > 0x7ffffffe)
6901 			hpts_timeout = 0x7ffffffe;
6902 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6903 	}
6904 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6905 	if ((rack->gp_ready == 0) &&
6906 	    (rack->use_fixed_rate == 0) &&
6907 	    (hpts_timeout < slot) &&
6908 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6909 		/*
6910 		 * We have no good estimate yet for the
6911 		 * old clunky burst mitigation or the
6912 		 * real pacing. And the tlp or rxt is smaller
6913 		 * than the pacing calculation. Lets not
6914 		 * pace that long since we know the calculation
6915 		 * so far is not accurate.
6916 		 */
6917 		slot = hpts_timeout;
6918 	}
6919 	/**
6920 	 * Turn off all the flags for queuing by default. The
6921 	 * flags have important meanings to what happens when
6922 	 * LRO interacts with the transport. Most likely (by default now)
6923 	 * mbuf_queueing and ack compression are on. So the transport
6924 	 * has a couple of flags that control what happens (if those
6925 	 * are not on then these flags won't have any effect since it
6926 	 * won't go through the queuing LRO path).
6927 	 *
6928 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6929 	 *                        pacing output, so don't disturb. But
6930 	 *                        it also means LRO can wake me if there
6931 	 *                        is a SACK arrival.
6932 	 *
6933 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6934 	 *                       with the above flag (QUEUE_READY) and
6935 	 *                       when present it says don't even wake me
6936 	 *                       if a SACK arrives.
6937 	 *
6938 	 * The idea behind these flags is that if we are pacing we
6939 	 * set the MBUF_QUEUE_READY and only get woken up if
6940 	 * a SACK arrives (which could change things) or if
6941 	 * our pacing timer expires. If, however, we have a rack
6942 	 * timer running, then we don't even want a sack to wake
6943 	 * us since the rack timer has to expire before we can send.
6944 	 *
6945 	 * Other cases should usually have none of the flags set
6946 	 * so LRO can call into us.
6947 	 */
6948 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6949 	if (slot) {
6950 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6951 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6952 		/*
6953 		 * A pacing timer (slot) is being set, in
6954 		 * such a case we cannot send (we are blocked by
6955 		 * the timer). So lets tell LRO that it should not
6956 		 * wake us unless there is a SACK. Note this only
6957 		 * will be effective if mbuf queueing is on or
6958 		 * compressed acks are being processed.
6959 		 */
6960 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6961 		/*
6962 		 * But wait if we have a Rack timer running
6963 		 * even a SACK should not disturb us (with
6964 		 * the exception of r_rr_config 3).
6965 		 */
6966 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6967 		    (IN_RECOVERY(tp->t_flags))) {
6968 			if (rack->r_rr_config != 3)
6969 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6970 			else if (rack->rc_pace_dnd) {
6971 				/*
6972 				 * When DND is on, we only let a sack
6973 				 * interrupt us if we are not in recovery.
6974 				 *
6975 				 * If DND is off, then we never hit here
6976 				 * and let all sacks wake us up.
6977 				 *
6978 				 */
6979 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6980 			}
6981 		}
6982 		/* For sack attackers we want to ignore sack */
6983 		if (rack->sack_attack_disable == 1) {
6984 			tp->t_flags2 |= (TF2_DONT_SACK_QUEUE |
6985 			    TF2_MBUF_QUEUE_READY);
6986 		} else if (rack->rc_ack_can_sendout_data) {
6987 			/*
6988 			 * Ahh but wait, this is that special case
6989 			 * where the pacing timer can be disturbed
6990 			 * backout the changes (used for non-paced
6991 			 * burst limiting).
6992 			 */
6993 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6994 			    TF2_MBUF_QUEUE_READY);
6995 		}
6996 		if ((rack->use_rack_rr) &&
6997 		    (rack->r_rr_config < 2) &&
6998 		    ((hpts_timeout) && (hpts_timeout < slot))) {
6999 			/*
7000 			 * Arrange for the hpts to kick back in after the
7001 			 * t-o if the t-o does not cause a send.
7002 			 */
7003 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7004 						   __LINE__, &diag);
7005 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7006 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7007 		} else {
7008 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
7009 						   __LINE__, &diag);
7010 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7011 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
7012 		}
7013 	} else if (hpts_timeout) {
7014 		/*
7015 		 * With respect to t_flags2(?) here, lets let any new acks wake
7016 		 * us up here. Since we are not pacing (no pacing timer), output
7017 		 * can happen so we should let it. If its a Rack timer, then any inbound
7018 		 * packet probably won't change the sending (we will be blocked)
7019 		 * but it may change the prr stats so letting it in (the set defaults
7020 		 * at the start of this block) are good enough.
7021 		 */
7022 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7023 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7024 					   __LINE__, &diag);
7025 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7026 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7027 	} else {
7028 		/* No timer starting */
7029 #ifdef INVARIANTS
7030 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
7031 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
7032 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
7033 		}
7034 #endif
7035 	}
7036 	rack->rc_tmr_stopped = 0;
7037 	if (slot)
7038 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
7039 }
7040 
7041 /*
7042  * RACK Timer, here we simply do logging and house keeping.
7043  * the normal rack_output() function will call the
7044  * appropriate thing to check if we need to do a RACK retransmit.
7045  * We return 1, saying don't proceed with rack_output only
7046  * when all timers have been stopped (destroyed PCB?).
7047  */
7048 static int
7049 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7050 {
7051 	/*
7052 	 * This timer simply provides an internal trigger to send out data.
7053 	 * The check_recovery_mode call will see if there are needed
7054 	 * retransmissions, if so we will enter fast-recovery. The output
7055 	 * call may or may not do the same thing depending on sysctl
7056 	 * settings.
7057 	 */
7058 	struct rack_sendmap *rsm;
7059 
7060 	counter_u64_add(rack_to_tot, 1);
7061 	if (rack->r_state && (rack->r_state != tp->t_state))
7062 		rack_set_state(tp, rack);
7063 	rack->rc_on_min_to = 0;
7064 	rsm = rack_check_recovery_mode(tp, cts);
7065 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
7066 	if (rsm) {
7067 		rack->r_ctl.rc_resend = rsm;
7068 		rack->r_timer_override = 1;
7069 		if (rack->use_rack_rr) {
7070 			/*
7071 			 * Don't accumulate extra pacing delay
7072 			 * we are allowing the rack timer to
7073 			 * over-ride pacing i.e. rrr takes precedence
7074 			 * if the pacing interval is longer than the rrr
7075 			 * time (in other words we get the min pacing
7076 			 * time versus rrr pacing time).
7077 			 */
7078 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7079 		}
7080 	}
7081 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
7082 	if (rsm == NULL) {
7083 		/* restart a timer and return 1 */
7084 		rack_start_hpts_timer(rack, tp, cts,
7085 				      0, 0, 0);
7086 		return (1);
7087 	}
7088 	return (0);
7089 }
7090 
7091 
7092 
7093 static void
7094 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
7095 {
7096 
7097 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
7098 		/*
7099 		 * The trailing space changed, mbufs can grow
7100 		 * at the tail but they can't shrink from
7101 		 * it, KASSERT that. Adjust the orig_m_len to
7102 		 * compensate for this change.
7103 		 */
7104 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7105 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7106 			 rsm->m,
7107 			 rsm,
7108 			 (intmax_t)M_TRAILINGROOM(rsm->m),
7109 			 rsm->orig_t_space,
7110 			 rsm->orig_m_len,
7111 			 rsm->m->m_len));
7112 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7113 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7114 	}
7115 	if (rsm->m->m_len < rsm->orig_m_len) {
7116 		/*
7117 		 * Mbuf shrank, trimmed off the top by an ack, our
7118 		 * offset changes.
7119 		 */
7120 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7121 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7122 			 rsm->m, rsm->m->m_len,
7123 			 rsm, rsm->orig_m_len,
7124 			 rsm->soff));
7125 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7126 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7127 		else
7128 			rsm->soff = 0;
7129 		rsm->orig_m_len = rsm->m->m_len;
7130 #ifdef INVARIANTS
7131 	} else if (rsm->m->m_len > rsm->orig_m_len) {
7132 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7133 		      rsm, rsm->m);
7134 #endif
7135 	}
7136 }
7137 
7138 static void
7139 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7140 {
7141 	struct mbuf *m;
7142 	uint32_t soff;
7143 
7144 	if (src_rsm->m &&
7145 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7146 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7147 		/* Fix up the orig_m_len and possibly the mbuf offset */
7148 		rack_adjust_orig_mlen(src_rsm);
7149 	}
7150 	m = src_rsm->m;
7151 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7152 	while (soff >= m->m_len) {
7153 		/* Move out past this mbuf */
7154 		soff -= m->m_len;
7155 		m = m->m_next;
7156 		KASSERT((m != NULL),
7157 			("rsm:%p nrsm:%p hit at soff:%u null m",
7158 			 src_rsm, rsm, soff));
7159 		if (m == NULL) {
7160 			/* This should *not* happen which is why there is a kassert */
7161 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7162 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7163 					       &src_rsm->soff);
7164 			src_rsm->orig_m_len = src_rsm->m->m_len;
7165 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7166 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7167 					   (rsm->r_start - rack->rc_tp->snd_una),
7168 					   &rsm->soff);
7169 			rsm->orig_m_len = rsm->m->m_len;
7170 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7171 			return;
7172 		}
7173 	}
7174 	rsm->m = m;
7175 	rsm->soff = soff;
7176 	rsm->orig_m_len = m->m_len;
7177 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7178 }
7179 
7180 static __inline void
7181 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7182 	       struct rack_sendmap *rsm, uint32_t start)
7183 {
7184 	int idx;
7185 
7186 	nrsm->r_start = start;
7187 	nrsm->r_end = rsm->r_end;
7188 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7189 	nrsm->r_flags = rsm->r_flags;
7190 	nrsm->r_dupack = rsm->r_dupack;
7191 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7192 	nrsm->r_rtr_bytes = 0;
7193 	nrsm->r_fas = rsm->r_fas;
7194 	nrsm->r_bas = rsm->r_bas;
7195 	rsm->r_end = nrsm->r_start;
7196 	nrsm->r_just_ret = rsm->r_just_ret;
7197 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7198 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7199 	}
7200 	/* Now if we have SYN flag we keep it on the left edge */
7201 	if (nrsm->r_flags & RACK_HAS_SYN)
7202 		nrsm->r_flags &= ~RACK_HAS_SYN;
7203 	/* Now if we have a FIN flag we keep it on the right edge */
7204 	if (rsm->r_flags & RACK_HAS_FIN)
7205 		rsm->r_flags &= ~RACK_HAS_FIN;
7206 	/* Push bit must go to the right edge as well */
7207 	if (rsm->r_flags & RACK_HAD_PUSH)
7208 		rsm->r_flags &= ~RACK_HAD_PUSH;
7209 	/* Clone over the state of the hw_tls flag */
7210 	nrsm->r_hw_tls = rsm->r_hw_tls;
7211 	/*
7212 	 * Now we need to find nrsm's new location in the mbuf chain
7213 	 * we basically calculate a new offset, which is soff +
7214 	 * how much is left in original rsm. Then we walk out the mbuf
7215 	 * chain to find the righ position, it may be the same mbuf
7216 	 * or maybe not.
7217 	 */
7218 	KASSERT(((rsm->m != NULL) ||
7219 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7220 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7221 	if (rsm->m)
7222 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7223 }
7224 
7225 static struct rack_sendmap *
7226 rack_merge_rsm(struct tcp_rack *rack,
7227 	       struct rack_sendmap *l_rsm,
7228 	       struct rack_sendmap *r_rsm)
7229 {
7230 	/*
7231 	 * We are merging two ack'd RSM's,
7232 	 * the l_rsm is on the left (lower seq
7233 	 * values) and the r_rsm is on the right
7234 	 * (higher seq value). The simplest way
7235 	 * to merge these is to move the right
7236 	 * one into the left. I don't think there
7237 	 * is any reason we need to try to find
7238 	 * the oldest (or last oldest retransmitted).
7239 	 */
7240 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7241 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7242 	l_rsm->r_end = r_rsm->r_end;
7243 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7244 		l_rsm->r_dupack = r_rsm->r_dupack;
7245 	if (r_rsm->r_rtr_bytes)
7246 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7247 	if (r_rsm->r_in_tmap) {
7248 		/* This really should not happen */
7249 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7250 		r_rsm->r_in_tmap = 0;
7251 	}
7252 
7253 	/* Now the flags */
7254 	if (r_rsm->r_flags & RACK_HAS_FIN)
7255 		l_rsm->r_flags |= RACK_HAS_FIN;
7256 	if (r_rsm->r_flags & RACK_TLP)
7257 		l_rsm->r_flags |= RACK_TLP;
7258 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7259 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7260 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7261 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7262 		/*
7263 		 * If both are app-limited then let the
7264 		 * free lower the count. If right is app
7265 		 * limited and left is not, transfer.
7266 		 */
7267 		l_rsm->r_flags |= RACK_APP_LIMITED;
7268 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7269 		if (r_rsm == rack->r_ctl.rc_first_appl)
7270 			rack->r_ctl.rc_first_appl = l_rsm;
7271 	}
7272 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7273 	/*
7274 	 * We keep the largest value, which is the newest
7275 	 * send. We do this in case a segment that is
7276 	 * joined together and not part of a GP estimate
7277 	 * later gets expanded into the GP estimate.
7278 	 *
7279 	 * We prohibit the merging of unlike kinds i.e.
7280 	 * all pieces that are in the GP estimate can be
7281 	 * merged and all pieces that are not in a GP estimate
7282 	 * can be merged, but not disimilar pieces. Combine
7283 	 * this with taking the highest here and we should
7284 	 * be ok unless of course the client reneges. Then
7285 	 * all bets are off.
7286 	 */
7287 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7288 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7289 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7290 	}
7291 	/*
7292 	 * When merging two RSM's we also need to consider the ack time and keep
7293 	 * newest. If the ack gets merged into a measurement then that is the
7294 	 * one we will want to be using.
7295 	 */
7296 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7297 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7298 
7299 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7300 		/* Transfer the split limit to the map we free */
7301 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7302 		l_rsm->r_limit_type = 0;
7303 	}
7304 	rack_free(rack, r_rsm);
7305 	l_rsm->r_flags |= RACK_MERGED;
7306 	return (l_rsm);
7307 }
7308 
7309 /*
7310  * TLP Timer, here we simply setup what segment we want to
7311  * have the TLP expire on, the normal rack_output() will then
7312  * send it out.
7313  *
7314  * We return 1, saying don't proceed with rack_output only
7315  * when all timers have been stopped (destroyed PCB?).
7316  */
7317 static int
7318 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7319 {
7320 	/*
7321 	 * Tail Loss Probe.
7322 	 */
7323 	struct rack_sendmap *rsm = NULL;
7324 	int insret __diagused;
7325 	struct socket *so = tptosocket(tp);
7326 	uint32_t amm;
7327 	uint32_t out, avail;
7328 	int collapsed_win = 0;
7329 
7330 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7331 		/* Its not time yet */
7332 		return (0);
7333 	}
7334 	if (ctf_progress_timeout_check(tp, true)) {
7335 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7336 		return (-ETIMEDOUT);	/* tcp_drop() */
7337 	}
7338 	/*
7339 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7340 	 * need to figure out how to force a full MSS segment out.
7341 	 */
7342 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7343 	rack->r_ctl.retran_during_recovery = 0;
7344 	rack->r_ctl.dsack_byte_cnt = 0;
7345 	counter_u64_add(rack_tlp_tot, 1);
7346 	if (rack->r_state && (rack->r_state != tp->t_state))
7347 		rack_set_state(tp, rack);
7348 	avail = sbavail(&so->so_snd);
7349 	out = tp->snd_max - tp->snd_una;
7350 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7351 		/* special case, we need a retransmission */
7352 		collapsed_win = 1;
7353 		goto need_retran;
7354 	}
7355 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7356 		rack->r_ctl.dsack_persist--;
7357 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7358 			rack->r_ctl.num_dsack = 0;
7359 		}
7360 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7361 	}
7362 	if ((tp->t_flags & TF_GPUTINPROG) &&
7363 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7364 		/*
7365 		 * If this is the second in a row
7366 		 * TLP and we are doing a measurement
7367 		 * its time to abandon the measurement.
7368 		 * Something is likely broken on
7369 		 * the clients network and measuring a
7370 		 * broken network does us no good.
7371 		 */
7372 		tp->t_flags &= ~TF_GPUTINPROG;
7373 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7374 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7375 					   tp->gput_seq,
7376 					   0, 0, 18, __LINE__, NULL, 0);
7377 	}
7378 	/*
7379 	 * Check our send oldest always settings, and if
7380 	 * there is an oldest to send jump to the need_retran.
7381 	 */
7382 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7383 		goto need_retran;
7384 
7385 	if (avail > out) {
7386 		/* New data is available */
7387 		amm = avail - out;
7388 		if (amm > ctf_fixed_maxseg(tp)) {
7389 			amm = ctf_fixed_maxseg(tp);
7390 			if ((amm + out) > tp->snd_wnd) {
7391 				/* We are rwnd limited */
7392 				goto need_retran;
7393 			}
7394 		} else if (amm < ctf_fixed_maxseg(tp)) {
7395 			/* not enough to fill a MTU */
7396 			goto need_retran;
7397 		}
7398 		if (IN_FASTRECOVERY(tp->t_flags)) {
7399 			/* Unlikely */
7400 			if (rack->rack_no_prr == 0) {
7401 				if (out + amm <= tp->snd_wnd) {
7402 					rack->r_ctl.rc_prr_sndcnt = amm;
7403 					rack->r_ctl.rc_tlp_new_data = amm;
7404 					rack_log_to_prr(rack, 4, 0, __LINE__);
7405 				}
7406 			} else
7407 				goto need_retran;
7408 		} else {
7409 			/* Set the send-new override */
7410 			if (out + amm <= tp->snd_wnd)
7411 				rack->r_ctl.rc_tlp_new_data = amm;
7412 			else
7413 				goto need_retran;
7414 		}
7415 		rack->r_ctl.rc_tlpsend = NULL;
7416 		counter_u64_add(rack_tlp_newdata, 1);
7417 		goto send;
7418 	}
7419 need_retran:
7420 	/*
7421 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7422 	 * optionally the first un-acked segment.
7423 	 */
7424 	if (collapsed_win == 0) {
7425 		if (rack_always_send_oldest)
7426 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7427 		else {
7428 			rsm = tqhash_max(rack->r_ctl.tqh);
7429 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7430 				rsm = rack_find_high_nonack(rack, rsm);
7431 			}
7432 		}
7433 		if (rsm == NULL) {
7434 #ifdef TCP_BLACKBOX
7435 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7436 #endif
7437 			goto out;
7438 		}
7439 	} else {
7440 		/*
7441 		 * We had a collapsed window, lets find
7442 		 * the point before the collapse.
7443 		 */
7444 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7445 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7446 		else {
7447 			rsm = tqhash_min(rack->r_ctl.tqh);
7448 		}
7449 		if (rsm == NULL) {
7450 			/* Huh */
7451 			goto out;
7452 		}
7453 	}
7454 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7455 		/*
7456 		 * We need to split this the last segment in two.
7457 		 */
7458 		struct rack_sendmap *nrsm;
7459 
7460 		nrsm = rack_alloc_full_limit(rack);
7461 		if (nrsm == NULL) {
7462 			/*
7463 			 * No memory to split, we will just exit and punt
7464 			 * off to the RXT timer.
7465 			 */
7466 			goto out;
7467 		}
7468 		rack_clone_rsm(rack, nrsm, rsm,
7469 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7470 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7471 #ifndef INVARIANTS
7472 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7473 #else
7474 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7475 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7476 			      nrsm, insret, rack, rsm);
7477 		}
7478 #endif
7479 		if (rsm->r_in_tmap) {
7480 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7481 			nrsm->r_in_tmap = 1;
7482 		}
7483 		rsm = nrsm;
7484 	}
7485 	rack->r_ctl.rc_tlpsend = rsm;
7486 send:
7487 	/* Make sure output path knows we are doing a TLP */
7488 	*doing_tlp = 1;
7489 	rack->r_timer_override = 1;
7490 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7491 	return (0);
7492 out:
7493 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7494 	return (0);
7495 }
7496 
7497 /*
7498  * Delayed ack Timer, here we simply need to setup the
7499  * ACK_NOW flag and remove the DELACK flag. From there
7500  * the output routine will send the ack out.
7501  *
7502  * We only return 1, saying don't proceed, if all timers
7503  * are stopped (destroyed PCB?).
7504  */
7505 static int
7506 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7507 {
7508 
7509 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7510 	tp->t_flags &= ~TF_DELACK;
7511 	tp->t_flags |= TF_ACKNOW;
7512 	KMOD_TCPSTAT_INC(tcps_delack);
7513 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7514 	return (0);
7515 }
7516 
7517 /*
7518  * Persists timer, here we simply send the
7519  * same thing as a keepalive will.
7520  * the one byte send.
7521  *
7522  * We only return 1, saying don't proceed, if all timers
7523  * are stopped (destroyed PCB?).
7524  */
7525 static int
7526 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7527 {
7528 	struct tcptemp *t_template;
7529 	int32_t retval = 1;
7530 
7531 	if (rack->rc_in_persist == 0)
7532 		return (0);
7533 	if (ctf_progress_timeout_check(tp, false)) {
7534 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7535 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7536 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7537 		return (-ETIMEDOUT);	/* tcp_drop() */
7538 	}
7539 	/*
7540 	 * Persistence timer into zero window. Force a byte to be output, if
7541 	 * possible.
7542 	 */
7543 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
7544 	/*
7545 	 * Hack: if the peer is dead/unreachable, we do not time out if the
7546 	 * window is closed.  After a full backoff, drop the connection if
7547 	 * the idle time (no responses to probes) reaches the maximum
7548 	 * backoff that we would use if retransmitting.
7549 	 */
7550 	if (tp->t_rxtshift >= V_tcp_retries &&
7551 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7552 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7553 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7554 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7555 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7556 		retval = -ETIMEDOUT;	/* tcp_drop() */
7557 		goto out;
7558 	}
7559 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7560 	    tp->snd_una == tp->snd_max)
7561 		rack_exit_persist(tp, rack, cts);
7562 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7563 	/*
7564 	 * If the user has closed the socket then drop a persisting
7565 	 * connection after a much reduced timeout.
7566 	 */
7567 	if (tp->t_state > TCPS_CLOSE_WAIT &&
7568 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7569 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7570 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7571 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7572 		retval = -ETIMEDOUT;	/* tcp_drop() */
7573 		goto out;
7574 	}
7575 	t_template = tcpip_maketemplate(rack->rc_inp);
7576 	if (t_template) {
7577 		/* only set it if we were answered */
7578 		if (rack->forced_ack == 0) {
7579 			rack->forced_ack = 1;
7580 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7581 		} else {
7582 			rack->probe_not_answered = 1;
7583 			counter_u64_add(rack_persists_loss, 1);
7584 			rack->r_ctl.persist_lost_ends++;
7585 		}
7586 		counter_u64_add(rack_persists_sends, 1);
7587 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7588 		tcp_respond(tp, t_template->tt_ipgen,
7589 			    &t_template->tt_t, (struct mbuf *)NULL,
7590 			    tp->rcv_nxt, tp->snd_una - 1, 0);
7591 		/* This sends an ack */
7592 		if (tp->t_flags & TF_DELACK)
7593 			tp->t_flags &= ~TF_DELACK;
7594 		free(t_template, M_TEMP);
7595 	}
7596 	if (tp->t_rxtshift < V_tcp_retries)
7597 		tp->t_rxtshift++;
7598 out:
7599 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7600 	rack_start_hpts_timer(rack, tp, cts,
7601 			      0, 0, 0);
7602 	return (retval);
7603 }
7604 
7605 /*
7606  * If a keepalive goes off, we had no other timers
7607  * happening. We always return 1 here since this
7608  * routine either drops the connection or sends
7609  * out a segment with respond.
7610  */
7611 static int
7612 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7613 {
7614 	struct tcptemp *t_template;
7615 	struct inpcb *inp = tptoinpcb(tp);
7616 
7617 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7618 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7619 	/*
7620 	 * Keep-alive timer went off; send something or drop connection if
7621 	 * idle for too long.
7622 	 */
7623 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
7624 	if (tp->t_state < TCPS_ESTABLISHED)
7625 		goto dropit;
7626 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7627 	    tp->t_state <= TCPS_CLOSING) {
7628 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7629 			goto dropit;
7630 		/*
7631 		 * Send a packet designed to force a response if the peer is
7632 		 * up and reachable: either an ACK if the connection is
7633 		 * still alive, or an RST if the peer has closed the
7634 		 * connection due to timeout or reboot. Using sequence
7635 		 * number tp->snd_una-1 causes the transmitted zero-length
7636 		 * segment to lie outside the receive window; by the
7637 		 * protocol spec, this requires the correspondent TCP to
7638 		 * respond.
7639 		 */
7640 		KMOD_TCPSTAT_INC(tcps_keepprobe);
7641 		t_template = tcpip_maketemplate(inp);
7642 		if (t_template) {
7643 			if (rack->forced_ack == 0) {
7644 				rack->forced_ack = 1;
7645 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7646 			} else {
7647 				rack->probe_not_answered = 1;
7648 			}
7649 			tcp_respond(tp, t_template->tt_ipgen,
7650 			    &t_template->tt_t, (struct mbuf *)NULL,
7651 			    tp->rcv_nxt, tp->snd_una - 1, 0);
7652 			free(t_template, M_TEMP);
7653 		}
7654 	}
7655 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7656 	return (1);
7657 dropit:
7658 	KMOD_TCPSTAT_INC(tcps_keepdrops);
7659 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7660 	return (-ETIMEDOUT);	/* tcp_drop() */
7661 }
7662 
7663 /*
7664  * Retransmit helper function, clear up all the ack
7665  * flags and take care of important book keeping.
7666  */
7667 static void
7668 rack_remxt_tmr(struct tcpcb *tp)
7669 {
7670 	/*
7671 	 * The retransmit timer went off, all sack'd blocks must be
7672 	 * un-acked.
7673 	 */
7674 	struct rack_sendmap *rsm, *trsm = NULL;
7675 	struct tcp_rack *rack;
7676 
7677 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7678 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7679 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7680 	if (rack->r_state && (rack->r_state != tp->t_state))
7681 		rack_set_state(tp, rack);
7682 	/*
7683 	 * Ideally we would like to be able to
7684 	 * mark SACK-PASS on anything not acked here.
7685 	 *
7686 	 * However, if we do that we would burst out
7687 	 * all that data 1ms apart. This would be unwise,
7688 	 * so for now we will just let the normal rxt timer
7689 	 * and tlp timer take care of it.
7690 	 *
7691 	 * Also we really need to stick them back in sequence
7692 	 * order. This way we send in the proper order and any
7693 	 * sacks that come floating in will "re-ack" the data.
7694 	 * To do this we zap the tmap with an INIT and then
7695 	 * walk through and place every rsm in the tail queue
7696 	 * hash table back in its seq ordered place.
7697 	 */
7698 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
7699 
7700 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7701 		rsm->r_dupack = 0;
7702 		if (rack_verbose_logging)
7703 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7704 		/* We must re-add it back to the tlist */
7705 		if (trsm == NULL) {
7706 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7707 		} else {
7708 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7709 		}
7710 		rsm->r_in_tmap = 1;
7711 		trsm = rsm;
7712 		if (rsm->r_flags & RACK_ACKED)
7713 			rsm->r_flags |= RACK_WAS_ACKED;
7714 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
7715 		rsm->r_flags |= RACK_MUST_RXT;
7716 	}
7717 	/* Clear the count (we just un-acked them) */
7718 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7719 	rack->r_ctl.rc_sacked = 0;
7720 	rack->r_ctl.rc_sacklast = NULL;
7721 	rack->r_ctl.rc_agg_delayed = 0;
7722 	rack->r_early = 0;
7723 	rack->r_ctl.rc_agg_early = 0;
7724 	rack->r_late = 0;
7725 	/* Clear the tlp rtx mark */
7726 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7727 	if (rack->r_ctl.rc_resend != NULL)
7728 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7729 	rack->r_ctl.rc_prr_sndcnt = 0;
7730 	rack_log_to_prr(rack, 6, 0, __LINE__);
7731 	rack->r_timer_override = 1;
7732 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
7733 #ifdef TCP_SAD_DETECTION
7734 	    || (rack->sack_attack_disable != 0)
7735 #endif
7736 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
7737 		/*
7738 		 * For non-sack customers new data
7739 		 * needs to go out as retransmits until
7740 		 * we retransmit up to snd_max.
7741 		 */
7742 		rack->r_must_retran = 1;
7743 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7744 						rack->r_ctl.rc_sacked);
7745 	}
7746 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7747 }
7748 
7749 static void
7750 rack_convert_rtts(struct tcpcb *tp)
7751 {
7752 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7753 	tp->t_rxtcur = RACK_REXMTVAL(tp);
7754 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7755 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7756 	}
7757 	if (tp->t_rxtcur > rack_rto_max) {
7758 		tp->t_rxtcur = rack_rto_max;
7759 	}
7760 }
7761 
7762 static void
7763 rack_cc_conn_init(struct tcpcb *tp)
7764 {
7765 	struct tcp_rack *rack;
7766 	uint32_t srtt;
7767 
7768 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7769 	srtt = tp->t_srtt;
7770 	cc_conn_init(tp);
7771 	/*
7772 	 * Now convert to rack's internal format,
7773 	 * if required.
7774 	 */
7775 	if ((srtt == 0) && (tp->t_srtt != 0))
7776 		rack_convert_rtts(tp);
7777 	/*
7778 	 * We want a chance to stay in slowstart as
7779 	 * we create a connection. TCP spec says that
7780 	 * initially ssthresh is infinite. For our
7781 	 * purposes that is the snd_wnd.
7782 	 */
7783 	if (tp->snd_ssthresh < tp->snd_wnd) {
7784 		tp->snd_ssthresh = tp->snd_wnd;
7785 	}
7786 	/*
7787 	 * We also want to assure a IW worth of
7788 	 * data can get inflight.
7789 	 */
7790 	if (rc_init_window(rack) < tp->snd_cwnd)
7791 		tp->snd_cwnd = rc_init_window(rack);
7792 }
7793 
7794 /*
7795  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7796  * we will setup to retransmit the lowest seq number outstanding.
7797  */
7798 static int
7799 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7800 {
7801 	struct inpcb *inp = tptoinpcb(tp);
7802 	int32_t rexmt;
7803 	int32_t retval = 0;
7804 	bool isipv6;
7805 
7806 	if ((tp->t_flags & TF_GPUTINPROG) &&
7807 	    (tp->t_rxtshift)) {
7808 		/*
7809 		 * We have had a second timeout
7810 		 * measurements on successive rxt's are not profitable.
7811 		 * It is unlikely to be of any use (the network is
7812 		 * broken or the client went away).
7813 		 */
7814 		tp->t_flags &= ~TF_GPUTINPROG;
7815 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7816 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7817 					   tp->gput_seq,
7818 					   0, 0, 18, __LINE__, NULL, 0);
7819 	}
7820 	if (ctf_progress_timeout_check(tp, false)) {
7821 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7822 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7823 		return (-ETIMEDOUT);	/* tcp_drop() */
7824 	}
7825 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7826 	rack->r_ctl.retran_during_recovery = 0;
7827 	rack->rc_ack_required = 1;
7828 	rack->r_ctl.dsack_byte_cnt = 0;
7829 	if (IN_FASTRECOVERY(tp->t_flags))
7830 		tp->t_flags |= TF_WASFRECOVERY;
7831 	else
7832 		tp->t_flags &= ~TF_WASFRECOVERY;
7833 	if (IN_CONGRECOVERY(tp->t_flags))
7834 		tp->t_flags |= TF_WASCRECOVERY;
7835 	else
7836 		tp->t_flags &= ~TF_WASCRECOVERY;
7837 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7838 	    (tp->snd_una == tp->snd_max)) {
7839 		/* Nothing outstanding .. nothing to do */
7840 		return (0);
7841 	}
7842 	if (rack->r_ctl.dsack_persist) {
7843 		rack->r_ctl.dsack_persist--;
7844 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7845 			rack->r_ctl.num_dsack = 0;
7846 		}
7847 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7848 	}
7849 	/*
7850 	 * Rack can only run one timer  at a time, so we cannot
7851 	 * run a KEEPINIT (gating SYN sending) and a retransmit
7852 	 * timer for the SYN. So if we are in a front state and
7853 	 * have a KEEPINIT timer we need to check the first transmit
7854 	 * against now to see if we have exceeded the KEEPINIT time
7855 	 * (if one is set).
7856 	 */
7857 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7858 	    (TP_KEEPINIT(tp) != 0)) {
7859 		struct rack_sendmap *rsm;
7860 
7861 		rsm = tqhash_min(rack->r_ctl.tqh);
7862 		if (rsm) {
7863 			/* Ok we have something outstanding to test keepinit with */
7864 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7865 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7866 				/* We have exceeded the KEEPINIT time */
7867 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7868 				goto drop_it;
7869 			}
7870 		}
7871 	}
7872 	/*
7873 	 * Retransmission timer went off.  Message has not been acked within
7874 	 * retransmit interval.  Back off to a longer retransmit interval
7875 	 * and retransmit one segment.
7876 	 */
7877 	rack_remxt_tmr(tp);
7878 	if ((rack->r_ctl.rc_resend == NULL) ||
7879 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7880 		/*
7881 		 * If the rwnd collapsed on
7882 		 * the one we are retransmitting
7883 		 * it does not count against the
7884 		 * rxt count.
7885 		 */
7886 		tp->t_rxtshift++;
7887 	}
7888 	if (tp->t_rxtshift > V_tcp_retries) {
7889 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7890 drop_it:
7891 		tp->t_rxtshift = V_tcp_retries;
7892 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7893 		/* XXXGL: previously t_softerror was casted to uint16_t */
7894 		MPASS(tp->t_softerror >= 0);
7895 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7896 		goto out;	/* tcp_drop() */
7897 	}
7898 	if (tp->t_state == TCPS_SYN_SENT) {
7899 		/*
7900 		 * If the SYN was retransmitted, indicate CWND to be limited
7901 		 * to 1 segment in cc_conn_init().
7902 		 */
7903 		tp->snd_cwnd = 1;
7904 	} else if (tp->t_rxtshift == 1) {
7905 		/*
7906 		 * first retransmit; record ssthresh and cwnd so they can be
7907 		 * recovered if this turns out to be a "bad" retransmit. A
7908 		 * retransmit is considered "bad" if an ACK for this segment
7909 		 * is received within RTT/2 interval; the assumption here is
7910 		 * that the ACK was already in flight.  See "On Estimating
7911 		 * End-to-End Network Path Properties" by Allman and Paxson
7912 		 * for more details.
7913 		 */
7914 		tp->snd_cwnd_prev = tp->snd_cwnd;
7915 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
7916 		tp->snd_recover_prev = tp->snd_recover;
7917 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7918 		tp->t_flags |= TF_PREVVALID;
7919 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7920 		tp->t_flags &= ~TF_PREVVALID;
7921 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7922 	if ((tp->t_state == TCPS_SYN_SENT) ||
7923 	    (tp->t_state == TCPS_SYN_RECEIVED))
7924 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7925 	else
7926 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7927 
7928 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7929 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7930 	/*
7931 	 * We enter the path for PLMTUD if connection is established or, if
7932 	 * connection is FIN_WAIT_1 status, reason for the last is that if
7933 	 * amount of data we send is very small, we could send it in couple
7934 	 * of packets and process straight to FIN. In that case we won't
7935 	 * catch ESTABLISHED state.
7936 	 */
7937 #ifdef INET6
7938 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7939 #else
7940 	isipv6 = false;
7941 #endif
7942 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
7943 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7944 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7945 	    ((tp->t_state == TCPS_ESTABLISHED) ||
7946 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
7947 		/*
7948 		 * Idea here is that at each stage of mtu probe (usually,
7949 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7950 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7951 		 * should take care of that.
7952 		 */
7953 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7954 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7955 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7956 		    tp->t_rxtshift % 2 == 0)) {
7957 			/*
7958 			 * Enter Path MTU Black-hole Detection mechanism: -
7959 			 * Disable Path MTU Discovery (IP "DF" bit). -
7960 			 * Reduce MTU to lower value than what we negotiated
7961 			 * with peer.
7962 			 */
7963 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7964 				/* Record that we may have found a black hole. */
7965 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7966 				/* Keep track of previous MSS. */
7967 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7968 			}
7969 
7970 			/*
7971 			 * Reduce the MSS to blackhole value or to the
7972 			 * default in an attempt to retransmit.
7973 			 */
7974 #ifdef INET6
7975 			if (isipv6 &&
7976 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7977 				/* Use the sysctl tuneable blackhole MSS. */
7978 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7979 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7980 			} else if (isipv6) {
7981 				/* Use the default MSS. */
7982 				tp->t_maxseg = V_tcp_v6mssdflt;
7983 				/*
7984 				 * Disable Path MTU Discovery when we switch
7985 				 * to minmss.
7986 				 */
7987 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7988 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7989 			}
7990 #endif
7991 #if defined(INET6) && defined(INET)
7992 			else
7993 #endif
7994 #ifdef INET
7995 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7996 				/* Use the sysctl tuneable blackhole MSS. */
7997 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7998 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7999 			} else {
8000 				/* Use the default MSS. */
8001 				tp->t_maxseg = V_tcp_mssdflt;
8002 				/*
8003 				 * Disable Path MTU Discovery when we switch
8004 				 * to minmss.
8005 				 */
8006 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8007 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8008 			}
8009 #endif
8010 		} else {
8011 			/*
8012 			 * If further retransmissions are still unsuccessful
8013 			 * with a lowered MTU, maybe this isn't a blackhole
8014 			 * and we restore the previous MSS and blackhole
8015 			 * detection flags. The limit '6' is determined by
8016 			 * giving each probe stage (1448, 1188, 524) 2
8017 			 * chances to recover.
8018 			 */
8019 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
8020 			    (tp->t_rxtshift >= 6)) {
8021 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8022 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
8023 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
8024 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
8025 			}
8026 		}
8027 	}
8028 	/*
8029 	 * Disable RFC1323 and SACK if we haven't got any response to
8030 	 * our third SYN to work-around some broken terminal servers
8031 	 * (most of which have hopefully been retired) that have bad VJ
8032 	 * header compression code which trashes TCP segments containing
8033 	 * unknown-to-them TCP options.
8034 	 */
8035 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
8036 	    (tp->t_rxtshift == 3))
8037 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
8038 	/*
8039 	 * If we backed off this far, our srtt estimate is probably bogus.
8040 	 * Clobber it so we'll take the next rtt measurement as our srtt;
8041 	 * move the current srtt into rttvar to keep the current retransmit
8042 	 * times until then.
8043 	 */
8044 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
8045 #ifdef INET6
8046 		if ((inp->inp_vflag & INP_IPV6) != 0)
8047 			in6_losing(inp);
8048 		else
8049 #endif
8050 			in_losing(inp);
8051 		tp->t_rttvar += tp->t_srtt;
8052 		tp->t_srtt = 0;
8053 	}
8054 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8055 	tp->snd_recover = tp->snd_max;
8056 	tp->t_flags |= TF_ACKNOW;
8057 	tp->t_rtttime = 0;
8058 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
8059 out:
8060 	return (retval);
8061 }
8062 
8063 static int
8064 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
8065 {
8066 	int32_t ret = 0;
8067 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
8068 
8069 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8070 	    (tp->t_flags & TF_GPUTINPROG)) {
8071 		/*
8072 		 * We have a goodput in progress
8073 		 * and we have entered a late state.
8074 		 * Do we have enough data in the sb
8075 		 * to handle the GPUT request?
8076 		 */
8077 		uint32_t bytes;
8078 
8079 		bytes = tp->gput_ack - tp->gput_seq;
8080 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
8081 			bytes += tp->gput_seq - tp->snd_una;
8082 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
8083 			/*
8084 			 * There are not enough bytes in the socket
8085 			 * buffer that have been sent to cover this
8086 			 * measurement. Cancel it.
8087 			 */
8088 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8089 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
8090 						   tp->gput_seq,
8091 						   0, 0, 18, __LINE__, NULL, 0);
8092 			tp->t_flags &= ~TF_GPUTINPROG;
8093 		}
8094 	}
8095 	if (timers == 0) {
8096 		return (0);
8097 	}
8098 	if (tp->t_state == TCPS_LISTEN) {
8099 		/* no timers on listen sockets */
8100 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
8101 			return (0);
8102 		return (1);
8103 	}
8104 	if ((timers & PACE_TMR_RACK) &&
8105 	    rack->rc_on_min_to) {
8106 		/*
8107 		 * For the rack timer when we
8108 		 * are on a min-timeout (which means rrr_conf = 3)
8109 		 * we don't want to check the timer. It may
8110 		 * be going off for a pace and thats ok we
8111 		 * want to send the retransmit (if its ready).
8112 		 *
8113 		 * If its on a normal rack timer (non-min) then
8114 		 * we will check if its expired.
8115 		 */
8116 		goto skip_time_check;
8117 	}
8118 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8119 		uint32_t left;
8120 
8121 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8122 			ret = -1;
8123 			rack_log_to_processing(rack, cts, ret, 0);
8124 			return (0);
8125 		}
8126 		if (hpts_calling == 0) {
8127 			/*
8128 			 * A user send or queued mbuf (sack) has called us? We
8129 			 * return 0 and let the pacing guards
8130 			 * deal with it if they should or
8131 			 * should not cause a send.
8132 			 */
8133 			ret = -2;
8134 			rack_log_to_processing(rack, cts, ret, 0);
8135 			return (0);
8136 		}
8137 		/*
8138 		 * Ok our timer went off early and we are not paced false
8139 		 * alarm, go back to sleep. We make sure we don't have
8140 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8141 		 * flag in place.
8142 		 */
8143 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8144 		ret = -3;
8145 		left = rack->r_ctl.rc_timer_exp - cts;
8146 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8147 		rack_log_to_processing(rack, cts, ret, left);
8148 		return (1);
8149 	}
8150 skip_time_check:
8151 	rack->rc_tmr_stopped = 0;
8152 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8153 	if (timers & PACE_TMR_DELACK) {
8154 		ret = rack_timeout_delack(tp, rack, cts);
8155 	} else if (timers & PACE_TMR_RACK) {
8156 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8157 		rack->r_fast_output = 0;
8158 		ret = rack_timeout_rack(tp, rack, cts);
8159 	} else if (timers & PACE_TMR_TLP) {
8160 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8161 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8162 	} else if (timers & PACE_TMR_RXT) {
8163 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8164 		rack->r_fast_output = 0;
8165 		ret = rack_timeout_rxt(tp, rack, cts);
8166 	} else if (timers & PACE_TMR_PERSIT) {
8167 		ret = rack_timeout_persist(tp, rack, cts);
8168 	} else if (timers & PACE_TMR_KEEP) {
8169 		ret = rack_timeout_keepalive(tp, rack, cts);
8170 	}
8171 	rack_log_to_processing(rack, cts, ret, timers);
8172 	return (ret);
8173 }
8174 
8175 static void
8176 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8177 {
8178 	struct timeval tv;
8179 	uint32_t us_cts, flags_on_entry;
8180 	uint8_t hpts_removed = 0;
8181 
8182 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8183 	us_cts = tcp_get_usecs(&tv);
8184 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8185 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8186 	     ((tp->snd_max - tp->snd_una) == 0))) {
8187 		tcp_hpts_remove(rack->rc_tp);
8188 		hpts_removed = 1;
8189 		/* If we were not delayed cancel out the flag. */
8190 		if ((tp->snd_max - tp->snd_una) == 0)
8191 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8192 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8193 	}
8194 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8195 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8196 		if (tcp_in_hpts(rack->rc_tp) &&
8197 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8198 			/*
8199 			 * Canceling timer's when we have no output being
8200 			 * paced. We also must remove ourselves from the
8201 			 * hpts.
8202 			 */
8203 			tcp_hpts_remove(rack->rc_tp);
8204 			hpts_removed = 1;
8205 		}
8206 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8207 	}
8208 	if (hpts_removed == 0)
8209 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8210 }
8211 
8212 static int
8213 rack_stopall(struct tcpcb *tp)
8214 {
8215 	struct tcp_rack *rack;
8216 
8217 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8218 	rack->t_timers_stopped = 1;
8219 
8220 	tcp_hpts_remove(tp);
8221 
8222 	return (0);
8223 }
8224 
8225 static void
8226 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8227 {
8228 	/*
8229 	 * Assure no timers are running.
8230 	 */
8231 	if (tcp_timer_active(tp, TT_PERSIST)) {
8232 		/* We enter in persists, set the flag appropriately */
8233 		rack->rc_in_persist = 1;
8234 	}
8235 	if (tcp_in_hpts(rack->rc_tp)) {
8236 		tcp_hpts_remove(rack->rc_tp);
8237 	}
8238 }
8239 
8240 static void
8241 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8242     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz)
8243 {
8244 	int32_t idx;
8245 
8246 	rsm->r_rtr_cnt++;
8247 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8248 	rsm->r_dupack = 0;
8249 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8250 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8251 		rsm->r_flags |= RACK_OVERMAX;
8252 	}
8253 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8254 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8255 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8256 	}
8257 	idx = rsm->r_rtr_cnt - 1;
8258 	rsm->r_tim_lastsent[idx] = ts;
8259 	/*
8260 	 * Here we don't add in the len of send, since its already
8261 	 * in snduna <->snd_max.
8262 	 */
8263 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8264 				     rack->r_ctl.rc_sacked);
8265 	if (rsm->r_flags & RACK_ACKED) {
8266 		/* Problably MTU discovery messing with us */
8267 		rsm->r_flags &= ~RACK_ACKED;
8268 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8269 	}
8270 	if (rsm->r_in_tmap) {
8271 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8272 		rsm->r_in_tmap = 0;
8273 	}
8274 	/* Lets make sure it really is in or not the GP window */
8275 	rack_mark_in_gp_win(tp, rsm);
8276 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8277 	rsm->r_in_tmap = 1;
8278 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8279 	/* Take off the must retransmit flag, if its on */
8280 	if (rsm->r_flags & RACK_MUST_RXT) {
8281 		if (rack->r_must_retran)
8282 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8283 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8284 			/*
8285 			 * We have retransmitted all we need. Clear
8286 			 * any must retransmit flags.
8287 			 */
8288 			rack->r_must_retran = 0;
8289 			rack->r_ctl.rc_out_at_rto = 0;
8290 		}
8291 		rsm->r_flags &= ~RACK_MUST_RXT;
8292 	}
8293 	/* Remove any collapsed flag */
8294 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8295 	if (rsm->r_flags & RACK_SACK_PASSED) {
8296 		/* We have retransmitted due to the SACK pass */
8297 		rsm->r_flags &= ~RACK_SACK_PASSED;
8298 		rsm->r_flags |= RACK_WAS_SACKPASS;
8299 	}
8300 }
8301 
8302 static uint32_t
8303 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8304     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag, int segsiz)
8305 {
8306 	/*
8307 	 * We (re-)transmitted starting at rsm->r_start for some length
8308 	 * (possibly less than r_end.
8309 	 */
8310 	struct rack_sendmap *nrsm;
8311 	int insret __diagused;
8312 	uint32_t c_end;
8313 	int32_t len;
8314 
8315 	len = *lenp;
8316 	c_end = rsm->r_start + len;
8317 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8318 		/*
8319 		 * We retransmitted the whole piece or more than the whole
8320 		 * slopping into the next rsm.
8321 		 */
8322 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8323 		if (c_end == rsm->r_end) {
8324 			*lenp = 0;
8325 			return (0);
8326 		} else {
8327 			int32_t act_len;
8328 
8329 			/* Hangs over the end return whats left */
8330 			act_len = rsm->r_end - rsm->r_start;
8331 			*lenp = (len - act_len);
8332 			return (rsm->r_end);
8333 		}
8334 		/* We don't get out of this block. */
8335 	}
8336 	/*
8337 	 * Here we retransmitted less than the whole thing which means we
8338 	 * have to split this into what was transmitted and what was not.
8339 	 */
8340 	nrsm = rack_alloc_full_limit(rack);
8341 	if (nrsm == NULL) {
8342 		/*
8343 		 * We can't get memory, so lets not proceed.
8344 		 */
8345 		*lenp = 0;
8346 		return (0);
8347 	}
8348 	/*
8349 	 * So here we are going to take the original rsm and make it what we
8350 	 * retransmitted. nrsm will be the tail portion we did not
8351 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8352 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8353 	 * 1, 6 and the new piece will be 6, 11.
8354 	 */
8355 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8356 	nrsm->r_dupack = 0;
8357 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8358 #ifndef INVARIANTS
8359 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8360 #else
8361 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8362 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8363 		      nrsm, insret, rack, rsm);
8364 	}
8365 #endif
8366 	if (rsm->r_in_tmap) {
8367 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8368 		nrsm->r_in_tmap = 1;
8369 	}
8370 	rsm->r_flags &= (~RACK_HAS_FIN);
8371 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8372 	/* Log a split of rsm into rsm and nrsm */
8373 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8374 	*lenp = 0;
8375 	return (0);
8376 }
8377 
8378 static void
8379 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8380 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8381 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb,
8382 		uint32_t s_moff, int hw_tls, int segsiz)
8383 {
8384 	struct tcp_rack *rack;
8385 	struct rack_sendmap *rsm, *nrsm;
8386 	int insret __diagused;
8387 
8388 	register uint32_t snd_max, snd_una;
8389 
8390 	/*
8391 	 * Add to the RACK log of packets in flight or retransmitted. If
8392 	 * there is a TS option we will use the TS echoed, if not we will
8393 	 * grab a TS.
8394 	 *
8395 	 * Retransmissions will increment the count and move the ts to its
8396 	 * proper place. Note that if options do not include TS's then we
8397 	 * won't be able to effectively use the ACK for an RTT on a retran.
8398 	 *
8399 	 * Notes about r_start and r_end. Lets consider a send starting at
8400 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8401 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8402 	 * This means that r_end is actually the first sequence for the next
8403 	 * slot (11).
8404 	 *
8405 	 */
8406 	/*
8407 	 * If err is set what do we do XXXrrs? should we not add the thing?
8408 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8409 	 * i.e. proceed with add ** do this for now.
8410 	 */
8411 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8412 	if (err)
8413 		/*
8414 		 * We don't log errors -- we could but snd_max does not
8415 		 * advance in this case either.
8416 		 */
8417 		return;
8418 
8419 	if (th_flags & TH_RST) {
8420 		/*
8421 		 * We don't log resets and we return immediately from
8422 		 * sending
8423 		 */
8424 		return;
8425 	}
8426 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8427 	snd_una = tp->snd_una;
8428 	snd_max = tp->snd_max;
8429 	if (th_flags & (TH_SYN | TH_FIN)) {
8430 		/*
8431 		 * The call to rack_log_output is made before bumping
8432 		 * snd_max. This means we can record one extra byte on a SYN
8433 		 * or FIN if seq_out is adding more on and a FIN is present
8434 		 * (and we are not resending).
8435 		 */
8436 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8437 			len++;
8438 		if (th_flags & TH_FIN)
8439 			len++;
8440 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
8441 			/*
8442 			 * The add/update as not been done for the FIN/SYN
8443 			 * yet.
8444 			 */
8445 			snd_max = tp->snd_nxt;
8446 		}
8447 	}
8448 	if (SEQ_LEQ((seq_out + len), snd_una)) {
8449 		/* Are sending an old segment to induce an ack (keep-alive)? */
8450 		return;
8451 	}
8452 	if (SEQ_LT(seq_out, snd_una)) {
8453 		/* huh? should we panic? */
8454 		uint32_t end;
8455 
8456 		end = seq_out + len;
8457 		seq_out = snd_una;
8458 		if (SEQ_GEQ(end, seq_out))
8459 			len = end - seq_out;
8460 		else
8461 			len = 0;
8462 	}
8463 	if (len == 0) {
8464 		/* We don't log zero window probes */
8465 		return;
8466 	}
8467 	if (IN_FASTRECOVERY(tp->t_flags)) {
8468 		rack->r_ctl.rc_prr_out += len;
8469 	}
8470 	/* First question is it a retransmission or new? */
8471 	if (seq_out == snd_max) {
8472 		/* Its new */
8473 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8474 again:
8475 		rsm = rack_alloc(rack);
8476 		if (rsm == NULL) {
8477 			/*
8478 			 * Hmm out of memory and the tcb got destroyed while
8479 			 * we tried to wait.
8480 			 */
8481 			return;
8482 		}
8483 		if (th_flags & TH_FIN) {
8484 			rsm->r_flags = RACK_HAS_FIN|add_flag;
8485 		} else {
8486 			rsm->r_flags = add_flag;
8487 		}
8488 		if (hw_tls)
8489 			rsm->r_hw_tls = 1;
8490 		rsm->r_tim_lastsent[0] = cts;
8491 		rsm->r_rtr_cnt = 1;
8492 		rsm->r_rtr_bytes = 0;
8493 		if (th_flags & TH_SYN) {
8494 			/* The data space is one beyond snd_una */
8495 			rsm->r_flags |= RACK_HAS_SYN;
8496 		}
8497 		rsm->r_start = seq_out;
8498 		rsm->r_end = rsm->r_start + len;
8499 		rack_mark_in_gp_win(tp, rsm);
8500 		rsm->r_dupack = 0;
8501 		/*
8502 		 * save off the mbuf location that
8503 		 * sndmbuf_noadv returned (which is
8504 		 * where we started copying from)..
8505 		 */
8506 		rsm->m = s_mb;
8507 		rsm->soff = s_moff;
8508 		/*
8509 		 * Here we do add in the len of send, since its not yet
8510 		 * reflected in in snduna <->snd_max
8511 		 */
8512 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8513 					      rack->r_ctl.rc_sacked) +
8514 			      (rsm->r_end - rsm->r_start));
8515 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8516 		if (rsm->m) {
8517 			if (rsm->m->m_len <= rsm->soff) {
8518 				/*
8519 				 * XXXrrs Question, will this happen?
8520 				 *
8521 				 * If sbsndptr is set at the correct place
8522 				 * then s_moff should always be somewhere
8523 				 * within rsm->m. But if the sbsndptr was
8524 				 * off then that won't be true. If it occurs
8525 				 * we need to walkout to the correct location.
8526 				 */
8527 				struct mbuf *lm;
8528 
8529 				lm = rsm->m;
8530 				while (lm->m_len <= rsm->soff) {
8531 					rsm->soff -= lm->m_len;
8532 					lm = lm->m_next;
8533 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8534 							     __func__, rack, s_moff, s_mb, rsm->soff));
8535 				}
8536 				rsm->m = lm;
8537 			}
8538 			rsm->orig_m_len = rsm->m->m_len;
8539 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8540 		} else {
8541 			rsm->orig_m_len = 0;
8542 			rsm->orig_t_space = 0;
8543 		}
8544 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8545 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8546 		/* Log a new rsm */
8547 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8548 #ifndef INVARIANTS
8549 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
8550 #else
8551 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8552 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8553 			      nrsm, insret, rack, rsm);
8554 		}
8555 #endif
8556 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8557 		rsm->r_in_tmap = 1;
8558 		/*
8559 		 * Special case detection, is there just a single
8560 		 * packet outstanding when we are not in recovery?
8561 		 *
8562 		 * If this is true mark it so.
8563 		 */
8564 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8565 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8566 			struct rack_sendmap *prsm;
8567 
8568 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8569 			if (prsm)
8570 				prsm->r_one_out_nr = 1;
8571 		}
8572 		return;
8573 	}
8574 	/*
8575 	 * If we reach here its a retransmission and we need to find it.
8576 	 */
8577 more:
8578 	if (hintrsm && (hintrsm->r_start == seq_out)) {
8579 		rsm = hintrsm;
8580 		hintrsm = NULL;
8581 	} else {
8582 		/* No hints sorry */
8583 		rsm = NULL;
8584 	}
8585 	if ((rsm) && (rsm->r_start == seq_out)) {
8586 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8587 		if (len == 0) {
8588 			return;
8589 		} else {
8590 			goto more;
8591 		}
8592 	}
8593 	/* Ok it was not the last pointer go through it the hard way. */
8594 refind:
8595 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8596 	if (rsm) {
8597 		if (rsm->r_start == seq_out) {
8598 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8599 			if (len == 0) {
8600 				return;
8601 			} else {
8602 				goto refind;
8603 			}
8604 		}
8605 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8606 			/* Transmitted within this piece */
8607 			/*
8608 			 * Ok we must split off the front and then let the
8609 			 * update do the rest
8610 			 */
8611 			nrsm = rack_alloc_full_limit(rack);
8612 			if (nrsm == NULL) {
8613 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8614 				return;
8615 			}
8616 			/*
8617 			 * copy rsm to nrsm and then trim the front of rsm
8618 			 * to not include this part.
8619 			 */
8620 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
8621 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8622 #ifndef INVARIANTS
8623 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8624 #else
8625 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8626 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8627 				      nrsm, insret, rack, rsm);
8628 			}
8629 #endif
8630 			if (rsm->r_in_tmap) {
8631 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8632 				nrsm->r_in_tmap = 1;
8633 			}
8634 			rsm->r_flags &= (~RACK_HAS_FIN);
8635 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8636 			if (len == 0) {
8637 				return;
8638 			} else if (len > 0)
8639 				goto refind;
8640 		}
8641 	}
8642 	/*
8643 	 * Hmm not found in map did they retransmit both old and on into the
8644 	 * new?
8645 	 */
8646 	if (seq_out == tp->snd_max) {
8647 		goto again;
8648 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
8649 #ifdef INVARIANTS
8650 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8651 		       seq_out, len, tp->snd_una, tp->snd_max);
8652 		printf("Starting Dump of all rack entries\n");
8653 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8654 			printf("rsm:%p start:%u end:%u\n",
8655 			       rsm, rsm->r_start, rsm->r_end);
8656 		}
8657 		printf("Dump complete\n");
8658 		panic("seq_out not found rack:%p tp:%p",
8659 		      rack, tp);
8660 #endif
8661 	} else {
8662 #ifdef INVARIANTS
8663 		/*
8664 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8665 		 * flag)
8666 		 */
8667 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8668 		      seq_out, len, tp->snd_max, tp);
8669 #endif
8670 	}
8671 }
8672 
8673 /*
8674  * Record one of the RTT updates from an ack into
8675  * our sample structure.
8676  */
8677 
8678 static void
8679 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8680 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8681 {
8682 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8683 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8684 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8685 	}
8686 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8687 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8688 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8689 	}
8690 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8691 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8692 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
8693 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8694 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8695 	}
8696 	if ((confidence == 1) &&
8697 	    ((rsm == NULL) ||
8698 	     (rsm->r_just_ret) ||
8699 	     (rsm->r_one_out_nr &&
8700 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8701 		/*
8702 		 * If the rsm had a just return
8703 		 * hit it then we can't trust the
8704 		 * rtt measurement for buffer deterimination
8705 		 * Note that a confidence of 2, indicates
8706 		 * SACK'd which overrides the r_just_ret or
8707 		 * the r_one_out_nr. If it was a CUM-ACK and
8708 		 * we had only two outstanding, but get an
8709 		 * ack for only 1. Then that also lowers our
8710 		 * confidence.
8711 		 */
8712 		confidence = 0;
8713 	}
8714 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8715 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8716 		if (rack->r_ctl.rack_rs.confidence == 0) {
8717 			/*
8718 			 * We take anything with no current confidence
8719 			 * saved.
8720 			 */
8721 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8722 			rack->r_ctl.rack_rs.confidence = confidence;
8723 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8724 		} else if (confidence != 0) {
8725 			/*
8726 			 * Once we have a confident number,
8727 			 * we can update it with a smaller
8728 			 * value since this confident number
8729 			 * may include the DSACK time until
8730 			 * the next segment (the second one) arrived.
8731 			 */
8732 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8733 			rack->r_ctl.rack_rs.confidence = confidence;
8734 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8735 		}
8736 	}
8737 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8738 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8739 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8740 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
8741 }
8742 
8743 /*
8744  * Collect new round-trip time estimate
8745  * and update averages and current timeout.
8746  */
8747 static void
8748 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8749 {
8750 	int32_t delta;
8751 	int32_t rtt;
8752 
8753 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8754 		/* No valid sample */
8755 		return;
8756 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8757 		/* We are to use the lowest RTT seen in a single ack */
8758 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8759 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8760 		/* We are to use the highest RTT seen in a single ack */
8761 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8762 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8763 		/* We are to use the average RTT seen in a single ack */
8764 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8765 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8766 	} else {
8767 #ifdef INVARIANTS
8768 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8769 #endif
8770 		return;
8771 	}
8772 	if (rtt == 0)
8773 		rtt = 1;
8774 	if (rack->rc_gp_rtt_set == 0) {
8775 		/*
8776 		 * With no RTT we have to accept
8777 		 * even one we are not confident of.
8778 		 */
8779 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8780 		rack->rc_gp_rtt_set = 1;
8781 	} else if (rack->r_ctl.rack_rs.confidence) {
8782 		/* update the running gp srtt */
8783 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8784 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8785 	}
8786 	if (rack->r_ctl.rack_rs.confidence) {
8787 		/*
8788 		 * record the low and high for highly buffered path computation,
8789 		 * we only do this if we are confident (not a retransmission).
8790 		 */
8791 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8792 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8793 		}
8794 		if (rack->rc_highly_buffered == 0) {
8795 			/*
8796 			 * Currently once we declare a path has
8797 			 * highly buffered there is no going
8798 			 * back, which may be a problem...
8799 			 */
8800 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8801 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8802 						     rack->r_ctl.rc_highest_us_rtt,
8803 						     rack->r_ctl.rc_lowest_us_rtt,
8804 						     RACK_RTTS_SEEHBP);
8805 				rack->rc_highly_buffered = 1;
8806 			}
8807 		}
8808 	}
8809 	if ((rack->r_ctl.rack_rs.confidence) ||
8810 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8811 		/*
8812 		 * If we are highly confident of it <or> it was
8813 		 * never retransmitted we accept it as the last us_rtt.
8814 		 */
8815 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8816 		/* The lowest rtt can be set if its was not retransmited */
8817 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8818 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8819 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
8820 				rack->r_ctl.rc_lowest_us_rtt = 1;
8821 		}
8822 	}
8823 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8824 	if (tp->t_srtt != 0) {
8825 		/*
8826 		 * We keep a simple srtt in microseconds, like our rtt
8827 		 * measurement. We don't need to do any tricks with shifting
8828 		 * etc. Instead we just add in 1/8th of the new measurement
8829 		 * and subtract out 1/8 of the old srtt. We do the same with
8830 		 * the variance after finding the absolute value of the
8831 		 * difference between this sample and the current srtt.
8832 		 */
8833 		delta = tp->t_srtt - rtt;
8834 		/* Take off 1/8th of the current sRTT */
8835 		tp->t_srtt -= (tp->t_srtt >> 3);
8836 		/* Add in 1/8th of the new RTT just measured */
8837 		tp->t_srtt += (rtt >> 3);
8838 		if (tp->t_srtt <= 0)
8839 			tp->t_srtt = 1;
8840 		/* Now lets make the absolute value of the variance */
8841 		if (delta < 0)
8842 			delta = -delta;
8843 		/* Subtract out 1/8th */
8844 		tp->t_rttvar -= (tp->t_rttvar >> 3);
8845 		/* Add in 1/8th of the new variance we just saw */
8846 		tp->t_rttvar += (delta >> 3);
8847 		if (tp->t_rttvar <= 0)
8848 			tp->t_rttvar = 1;
8849 	} else {
8850 		/*
8851 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
8852 		 * variance to half the rtt (so our first retransmit happens
8853 		 * at 3*rtt).
8854 		 */
8855 		tp->t_srtt = rtt;
8856 		tp->t_rttvar = rtt >> 1;
8857 	}
8858 	rack->rc_srtt_measure_made = 1;
8859 	KMOD_TCPSTAT_INC(tcps_rttupdated);
8860 	if (tp->t_rttupdated < UCHAR_MAX)
8861 		tp->t_rttupdated++;
8862 #ifdef STATS
8863 	if (rack_stats_gets_ms_rtt == 0) {
8864 		/* Send in the microsecond rtt used for rxt timeout purposes */
8865 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8866 	} else if (rack_stats_gets_ms_rtt == 1) {
8867 		/* Send in the millisecond rtt used for rxt timeout purposes */
8868 		int32_t ms_rtt;
8869 
8870 		/* Round up */
8871 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8872 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8873 	} else if (rack_stats_gets_ms_rtt == 2) {
8874 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
8875 		int32_t ms_rtt;
8876 
8877 		/* Round up */
8878 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8879 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8880 	}  else {
8881 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
8882 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8883 	}
8884 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8885 #endif
8886 	/*
8887 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8888 	 * way we do the smoothing, srtt and rttvar will each average +1/2
8889 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
8890 	 * tick of rounding and 1 extra tick because of +-1/2 tick
8891 	 * uncertainty in the firing of the timer.  The bias will give us
8892 	 * exactly the 1.5 tick we need.  But, because the bias is
8893 	 * statistical, we have to test that we don't drop below the minimum
8894 	 * feasible timer (which is 2 ticks).
8895 	 */
8896 	tp->t_rxtshift = 0;
8897 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8898 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8899 	rack_log_rtt_sample(rack, rtt);
8900 	tp->t_softerror = 0;
8901 }
8902 
8903 
8904 static void
8905 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8906 {
8907 	/*
8908 	 * Apply to filter the inbound us-rtt at us_cts.
8909 	 */
8910 	uint32_t old_rtt;
8911 
8912 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8913 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8914 			       us_rtt, us_cts);
8915 	if (old_rtt > us_rtt) {
8916 		/* We just hit a new lower rtt time */
8917 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8918 				     __LINE__, RACK_RTTS_NEWRTT);
8919 		/*
8920 		 * Only count it if its lower than what we saw within our
8921 		 * calculated range.
8922 		 */
8923 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8924 			if (rack_probertt_lower_within &&
8925 			    rack->rc_gp_dyn_mul &&
8926 			    (rack->use_fixed_rate == 0) &&
8927 			    (rack->rc_always_pace)) {
8928 				/*
8929 				 * We are seeing a new lower rtt very close
8930 				 * to the time that we would have entered probe-rtt.
8931 				 * This is probably due to the fact that a peer flow
8932 				 * has entered probe-rtt. Lets go in now too.
8933 				 */
8934 				uint32_t val;
8935 
8936 				val = rack_probertt_lower_within * rack_time_between_probertt;
8937 				val /= 100;
8938 				if ((rack->in_probe_rtt == 0)  &&
8939 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
8940 					rack_enter_probertt(rack, us_cts);
8941 				}
8942 			}
8943 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8944 		}
8945 	}
8946 }
8947 
8948 static int
8949 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8950     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8951 {
8952 	uint32_t us_rtt;
8953 	int32_t i, all;
8954 	uint32_t t, len_acked;
8955 
8956 	if ((rsm->r_flags & RACK_ACKED) ||
8957 	    (rsm->r_flags & RACK_WAS_ACKED))
8958 		/* Already done */
8959 		return (0);
8960 	if (rsm->r_no_rtt_allowed) {
8961 		/* Not allowed */
8962 		return (0);
8963 	}
8964 	if (ack_type == CUM_ACKED) {
8965 		if (SEQ_GT(th_ack, rsm->r_end)) {
8966 			len_acked = rsm->r_end - rsm->r_start;
8967 			all = 1;
8968 		} else {
8969 			len_acked = th_ack - rsm->r_start;
8970 			all = 0;
8971 		}
8972 	} else {
8973 		len_acked = rsm->r_end - rsm->r_start;
8974 		all = 0;
8975 	}
8976 	if (rsm->r_rtr_cnt == 1) {
8977 
8978 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8979 		if ((int)t <= 0)
8980 			t = 1;
8981 		if (!tp->t_rttlow || tp->t_rttlow > t)
8982 			tp->t_rttlow = t;
8983 		if (!rack->r_ctl.rc_rack_min_rtt ||
8984 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8985 			rack->r_ctl.rc_rack_min_rtt = t;
8986 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8987 				rack->r_ctl.rc_rack_min_rtt = 1;
8988 			}
8989 		}
8990 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8991 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8992 		else
8993 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8994 		if (us_rtt == 0)
8995 			us_rtt = 1;
8996 		if (CC_ALGO(tp)->rttsample != NULL) {
8997 			/* Kick the RTT to the CC */
8998 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8999 		}
9000 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
9001 		if (ack_type == SACKED) {
9002 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
9003 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
9004 		} else {
9005 			/*
9006 			 * We need to setup what our confidence
9007 			 * is in this ack.
9008 			 *
9009 			 * If the rsm was app limited and it is
9010 			 * less than a mss in length (the end
9011 			 * of the send) then we have a gap. If we
9012 			 * were app limited but say we were sending
9013 			 * multiple MSS's then we are more confident
9014 			 * int it.
9015 			 *
9016 			 * When we are not app-limited then we see if
9017 			 * the rsm is being included in the current
9018 			 * measurement, we tell this by the app_limited_needs_set
9019 			 * flag.
9020 			 *
9021 			 * Note that being cwnd blocked is not applimited
9022 			 * as well as the pacing delay between packets which
9023 			 * are sending only 1 or 2 MSS's also will show up
9024 			 * in the RTT. We probably need to examine this algorithm
9025 			 * a bit more and enhance it to account for the delay
9026 			 * between rsm's. We could do that by saving off the
9027 			 * pacing delay of each rsm (in an rsm) and then
9028 			 * factoring that in somehow though for now I am
9029 			 * not sure how :)
9030 			 */
9031 			int calc_conf = 0;
9032 
9033 			if (rsm->r_flags & RACK_APP_LIMITED) {
9034 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
9035 					calc_conf = 0;
9036 				else
9037 					calc_conf = 1;
9038 			} else if (rack->app_limited_needs_set == 0) {
9039 				calc_conf = 1;
9040 			} else {
9041 				calc_conf = 0;
9042 			}
9043 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
9044 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
9045 					    calc_conf, rsm, rsm->r_rtr_cnt);
9046 		}
9047 		if ((rsm->r_flags & RACK_TLP) &&
9048 		    (!IN_FASTRECOVERY(tp->t_flags))) {
9049 			/* Segment was a TLP and our retrans matched */
9050 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
9051 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9052 			}
9053 		}
9054 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9055 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9056 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9057 			/* New more recent rack_tmit_time */
9058 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9059 			if (rack->r_ctl.rc_rack_tmit_time == 0)
9060 				rack->r_ctl.rc_rack_tmit_time = 1;
9061 			rack->rc_rack_rtt = t;
9062 		}
9063 		return (1);
9064 	}
9065 	/*
9066 	 * We clear the soft/rxtshift since we got an ack.
9067 	 * There is no assurance we will call the commit() function
9068 	 * so we need to clear these to avoid incorrect handling.
9069 	 */
9070 	tp->t_rxtshift = 0;
9071 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9072 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9073 	tp->t_softerror = 0;
9074 	if (to && (to->to_flags & TOF_TS) &&
9075 	    (ack_type == CUM_ACKED) &&
9076 	    (to->to_tsecr) &&
9077 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
9078 		/*
9079 		 * Now which timestamp does it match? In this block the ACK
9080 		 * must be coming from a previous transmission.
9081 		 */
9082 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9083 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9084 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9085 				if ((int)t <= 0)
9086 					t = 1;
9087 				if (CC_ALGO(tp)->rttsample != NULL) {
9088 					/*
9089 					 * Kick the RTT to the CC, here
9090 					 * we lie a bit in that we know the
9091 					 * retransmission is correct even though
9092 					 * we retransmitted. This is because
9093 					 * we match the timestamps.
9094 					 */
9095 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9096 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9097 					else
9098 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9099 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9100 				}
9101 				if ((i + 1) < rsm->r_rtr_cnt) {
9102 					/*
9103 					 * The peer ack'd from our previous
9104 					 * transmission. We have a spurious
9105 					 * retransmission and thus we dont
9106 					 * want to update our rack_rtt.
9107 					 *
9108 					 * Hmm should there be a CC revert here?
9109 					 *
9110 					 */
9111 					return (0);
9112 				}
9113 				if (!tp->t_rttlow || tp->t_rttlow > t)
9114 					tp->t_rttlow = t;
9115 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9116 					rack->r_ctl.rc_rack_min_rtt = t;
9117 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9118 						rack->r_ctl.rc_rack_min_rtt = 1;
9119 					}
9120 				}
9121 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9122 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9123 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9124 					/* New more recent rack_tmit_time */
9125 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9126 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9127 						rack->r_ctl.rc_rack_tmit_time = 1;
9128 					rack->rc_rack_rtt = t;
9129 				}
9130 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9131 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9132 						    rsm->r_rtr_cnt);
9133 				return (1);
9134 			}
9135 		}
9136 		/* If we are logging log out the sendmap */
9137 		if (tcp_bblogging_on(rack->rc_tp)) {
9138 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9139 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9140 			}
9141 		}
9142 		goto ts_not_found;
9143 	} else {
9144 		/*
9145 		 * Ok its a SACK block that we retransmitted. or a windows
9146 		 * machine without timestamps. We can tell nothing from the
9147 		 * time-stamp since its not there or the time the peer last
9148 		 * received a segment that moved forward its cum-ack point.
9149 		 */
9150 ts_not_found:
9151 		i = rsm->r_rtr_cnt - 1;
9152 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9153 		if ((int)t <= 0)
9154 			t = 1;
9155 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9156 			/*
9157 			 * We retransmitted and the ack came back in less
9158 			 * than the smallest rtt we have observed. We most
9159 			 * likely did an improper retransmit as outlined in
9160 			 * 6.2 Step 2 point 2 in the rack-draft so we
9161 			 * don't want to update our rack_rtt. We in
9162 			 * theory (in future) might want to think about reverting our
9163 			 * cwnd state but we won't for now.
9164 			 */
9165 			return (0);
9166 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9167 			/*
9168 			 * We retransmitted it and the retransmit did the
9169 			 * job.
9170 			 */
9171 			if (!rack->r_ctl.rc_rack_min_rtt ||
9172 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9173 				rack->r_ctl.rc_rack_min_rtt = t;
9174 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9175 					rack->r_ctl.rc_rack_min_rtt = 1;
9176 				}
9177 			}
9178 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9179 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9180 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9181 				/* New more recent rack_tmit_time */
9182 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9183 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9184 					rack->r_ctl.rc_rack_tmit_time = 1;
9185 				rack->rc_rack_rtt = t;
9186 			}
9187 			return (1);
9188 		}
9189 	}
9190 	return (0);
9191 }
9192 
9193 /*
9194  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9195  */
9196 static void
9197 rack_log_sack_passed(struct tcpcb *tp,
9198     struct tcp_rack *rack, struct rack_sendmap *rsm)
9199 {
9200 	struct rack_sendmap *nrsm;
9201 
9202 	nrsm = rsm;
9203 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9204 	    rack_head, r_tnext) {
9205 		if (nrsm == rsm) {
9206 			/* Skip original segment he is acked */
9207 			continue;
9208 		}
9209 		if (nrsm->r_flags & RACK_ACKED) {
9210 			/*
9211 			 * Skip ack'd segments, though we
9212 			 * should not see these, since tmap
9213 			 * should not have ack'd segments.
9214 			 */
9215 			continue;
9216 		}
9217 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9218 			/*
9219 			 * If the peer dropped the rwnd on
9220 			 * these then we don't worry about them.
9221 			 */
9222 			continue;
9223 		}
9224 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9225 			/*
9226 			 * We found one that is already marked
9227 			 * passed, we have been here before and
9228 			 * so all others below this are marked.
9229 			 */
9230 			break;
9231 		}
9232 		nrsm->r_flags |= RACK_SACK_PASSED;
9233 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9234 	}
9235 }
9236 
9237 static void
9238 rack_need_set_test(struct tcpcb *tp,
9239 		   struct tcp_rack *rack,
9240 		   struct rack_sendmap *rsm,
9241 		   tcp_seq th_ack,
9242 		   int line,
9243 		   int use_which)
9244 {
9245 	struct rack_sendmap *s_rsm;
9246 
9247 	if ((tp->t_flags & TF_GPUTINPROG) &&
9248 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9249 		/*
9250 		 * We were app limited, and this ack
9251 		 * butts up or goes beyond the point where we want
9252 		 * to start our next measurement. We need
9253 		 * to record the new gput_ts as here and
9254 		 * possibly update the start sequence.
9255 		 */
9256 		uint32_t seq, ts;
9257 
9258 		if (rsm->r_rtr_cnt > 1) {
9259 			/*
9260 			 * This is a retransmit, can we
9261 			 * really make any assessment at this
9262 			 * point?  We are not really sure of
9263 			 * the timestamp, is it this or the
9264 			 * previous transmission?
9265 			 *
9266 			 * Lets wait for something better that
9267 			 * is not retransmitted.
9268 			 */
9269 			return;
9270 		}
9271 		seq = tp->gput_seq;
9272 		ts = tp->gput_ts;
9273 		rack->app_limited_needs_set = 0;
9274 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9275 		/* Do we start at a new end? */
9276 		if ((use_which == RACK_USE_BEG) &&
9277 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9278 			/*
9279 			 * When we get an ACK that just eats
9280 			 * up some of the rsm, we set RACK_USE_BEG
9281 			 * since whats at r_start (i.e. th_ack)
9282 			 * is left unacked and thats where the
9283 			 * measurement now starts.
9284 			 */
9285 			tp->gput_seq = rsm->r_start;
9286 		}
9287 		if ((use_which == RACK_USE_END) &&
9288 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9289 			/*
9290 			 * We use the end when the cumack
9291 			 * is moving forward and completely
9292 			 * deleting the rsm passed so basically
9293 			 * r_end holds th_ack.
9294 			 *
9295 			 * For SACK's we also want to use the end
9296 			 * since this piece just got sacked and
9297 			 * we want to target anything after that
9298 			 * in our measurement.
9299 			 */
9300 			tp->gput_seq = rsm->r_end;
9301 		}
9302 		if (use_which == RACK_USE_END_OR_THACK) {
9303 			/*
9304 			 * special case for ack moving forward,
9305 			 * not a sack, we need to move all the
9306 			 * way up to where this ack cum-ack moves
9307 			 * to.
9308 			 */
9309 			if (SEQ_GT(th_ack, rsm->r_end))
9310 				tp->gput_seq = th_ack;
9311 			else
9312 				tp->gput_seq = rsm->r_end;
9313 		}
9314 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9315 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9316 		else
9317 			s_rsm = NULL;
9318 		/*
9319 		 * Pick up the correct send time if we can the rsm passed in
9320 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9321 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9322 		 * find a different seq i.e. the next send up.
9323 		 *
9324 		 * If that has not been sent, s_rsm will be NULL and we must
9325 		 * arrange it so this function will get called again by setting
9326 		 * app_limited_needs_set.
9327 		 */
9328 		if (s_rsm)
9329 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9330 		else {
9331 			/* If we hit here we have to have *not* sent tp->gput_seq */
9332 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9333 			/* Set it up so we will go through here again */
9334 			rack->app_limited_needs_set = 1;
9335 		}
9336 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9337 			/*
9338 			 * We moved beyond this guy's range, re-calculate
9339 			 * the new end point.
9340 			 */
9341 			if (rack->rc_gp_filled == 0) {
9342 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9343 			} else {
9344 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9345 			}
9346 		}
9347 		/*
9348 		 * We are moving the goal post, we may be able to clear the
9349 		 * measure_saw_probe_rtt flag.
9350 		 */
9351 		if ((rack->in_probe_rtt == 0) &&
9352 		    (rack->measure_saw_probe_rtt) &&
9353 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9354 			rack->measure_saw_probe_rtt = 0;
9355 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9356 					   seq, tp->gput_seq,
9357 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9358 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9359 					   5, line, NULL, 0);
9360 		if (rack->rc_gp_filled &&
9361 		    ((tp->gput_ack - tp->gput_seq) <
9362 		     max(rc_init_window(rack), (MIN_GP_WIN *
9363 						ctf_fixed_maxseg(tp))))) {
9364 			uint32_t ideal_amount;
9365 
9366 			ideal_amount = rack_get_measure_window(tp, rack);
9367 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9368 				/*
9369 				 * There is no sense of continuing this measurement
9370 				 * because its too small to gain us anything we
9371 				 * trust. Skip it and that way we can start a new
9372 				 * measurement quicker.
9373 				 */
9374 				tp->t_flags &= ~TF_GPUTINPROG;
9375 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9376 							   0, 0,
9377 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9378 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9379 							   6, __LINE__, NULL, 0);
9380 			} else {
9381 				/*
9382 				 * Reset the window further out.
9383 				 */
9384 				tp->gput_ack = tp->gput_seq + ideal_amount;
9385 			}
9386 		}
9387 		rack_tend_gp_marks(tp, rack);
9388 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9389 	}
9390 }
9391 
9392 static inline int
9393 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9394 {
9395 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9396 		/* Behind our TLP definition or right at */
9397 		return (0);
9398 	}
9399 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9400 		/* The start is beyond or right at our end of TLP definition */
9401 		return (0);
9402 	}
9403 	/* It has to be a sub-part of the original TLP recorded */
9404 	return (1);
9405 }
9406 
9407 
9408 
9409 static uint32_t
9410 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9411 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9412 		   int *no_extra,
9413 		   int *moved_two, uint32_t segsiz)
9414 {
9415 	uint32_t start, end, changed = 0;
9416 	struct rack_sendmap stack_map;
9417 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
9418 	int insret __diagused;
9419 	int32_t used_ref = 1;
9420 	int moved = 0;
9421 #ifdef TCP_SAD_DETECTION
9422 	int allow_segsiz;
9423 	int first_time_through = 1;
9424 #endif
9425 	int noextra = 0;
9426 	int can_use_hookery = 0;
9427 
9428 	start = sack->start;
9429 	end = sack->end;
9430 	rsm = *prsm;
9431 
9432 #ifdef TCP_SAD_DETECTION
9433 	/*
9434 	 * There are a strange number of proxys and meddle boxes in the world
9435 	 * that seem to cut up segments on different boundaries. This gets us
9436 	 * smaller sacks that are still ok in terms of it being an attacker.
9437 	 * We use the base segsiz to calculate an allowable smallness but
9438 	 * also enforce a min on the segsiz in case it is an attacker playing
9439 	 * games with MSS. So basically if the sack arrives and it is
9440 	 * larger than a worse case 960 bytes, we don't classify the guy
9441 	 * as supicious.
9442 	 */
9443 	allow_segsiz = max(segsiz, 1200) * sad_seg_size_per;
9444 	allow_segsiz /= 1000;
9445 #endif
9446 do_rest_ofb:
9447 	if ((rsm == NULL) ||
9448 	    (SEQ_LT(end, rsm->r_start)) ||
9449 	    (SEQ_GEQ(start, rsm->r_end)) ||
9450 	    (SEQ_LT(start, rsm->r_start))) {
9451 		/*
9452 		 * We are not in the right spot,
9453 		 * find the correct spot in the tree.
9454 		 */
9455 		used_ref = 0;
9456 		rsm = tqhash_find(rack->r_ctl.tqh, start);
9457 		moved++;
9458 	}
9459 	if (rsm == NULL) {
9460 		/* TSNH */
9461 		goto out;
9462 	}
9463 #ifdef TCP_SAD_DETECTION
9464 	/* Now we must check for suspicous activity */
9465 	if ((first_time_through == 1) &&
9466 	    ((end - start) < min((rsm->r_end - rsm->r_start), allow_segsiz)) &&
9467 	    ((rsm->r_flags & RACK_PMTU_CHG) == 0) &&
9468 	    ((rsm->r_flags & RACK_TLP) == 0)) {
9469 		/*
9470 		 * Its less than a full MSS or the segment being acked
9471 		 * this should only happen if the rsm in question had the
9472 		 * r_just_ret flag set <and> the end matches the end of
9473 		 * the rsm block.
9474 		 *
9475 		 * Note we do not look at segments that have had TLP's on
9476 		 * them since we can get un-reported rwnd collapses that
9477 		 * basically we TLP on and then we get back a sack block
9478 		 * that goes from the start to only a small way.
9479 		 *
9480 		 */
9481 		int loss, ok;
9482 
9483 		ok = 0;
9484 		if (SEQ_GEQ(end, rsm->r_end)) {
9485 			if (rsm->r_just_ret == 1) {
9486 				/* This was at the end of a send which is ok */
9487 				ok = 1;
9488 			} else {
9489 				/* A bit harder was it the end of our segment */
9490 				int segs, len;
9491 
9492 				len = (rsm->r_end - rsm->r_start);
9493 				segs = len / segsiz;
9494 				segs *= segsiz;
9495 				if ((segs + (rsm->r_end - start)) == len) {
9496 					/*
9497 					 * So this last bit was the
9498 					 * end of our send if we cut it
9499 					 * up into segsiz pieces so its ok.
9500 					 */
9501 					ok = 1;
9502 				}
9503 			}
9504 		}
9505 		if (ok == 0) {
9506 			/*
9507 			 * This guy is doing something suspicious
9508 			 * lets start detection.
9509 			 */
9510 			if (rack->rc_suspicious == 0) {
9511 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_SUSPECT);
9512 				counter_u64_add(rack_sack_attacks_suspect, 1);
9513 				rack->rc_suspicious = 1;
9514 				rack_log_sad(rack, 4);
9515 				if (tcp_bblogging_on(rack->rc_tp)) {
9516 					union tcp_log_stackspecific log;
9517 					struct timeval tv;
9518 
9519 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
9520 					log.u_bbr.flex1 = end;
9521 					log.u_bbr.flex2 = start;
9522 					log.u_bbr.flex3 = rsm->r_end;
9523 					log.u_bbr.flex4 = rsm->r_start;
9524 					log.u_bbr.flex5 = segsiz;
9525 					log.u_bbr.flex6 = rsm->r_fas;
9526 					log.u_bbr.flex7 = rsm->r_bas;
9527 					log.u_bbr.flex8 = 5;
9528 					log.u_bbr.pkts_out = rsm->r_flags;
9529 					log.u_bbr.bbr_state = rack->rc_suspicious;
9530 					log.u_bbr.bbr_substate = rsm->r_just_ret;
9531 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
9532 					log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9533 					TCP_LOG_EVENTP(rack->rc_tp, NULL,
9534 						       &rack->rc_inp->inp_socket->so_rcv,
9535 						       &rack->rc_inp->inp_socket->so_snd,
9536 						       TCP_SAD_DETECTION, 0,
9537 						       0, &log, false, &tv);
9538 				}
9539 			}
9540 			/* You loose some ack count every time you sack
9541 			 * a small bit that is not butting to the end of
9542 			 * what we have sent. This is because we never
9543 			 * send small bits unless its the end of the sb.
9544 			 * Anyone sending a sack that is not at the end
9545 			 * is thus very very suspicious.
9546 			 */
9547 			loss = (segsiz/2) / (end - start);
9548 			if (loss < rack->r_ctl.ack_count)
9549 				rack->r_ctl.ack_count -= loss;
9550 			else
9551 				rack->r_ctl.ack_count = 0;
9552 		}
9553 	}
9554 	first_time_through = 0;
9555 #endif
9556 	/* Ok we have an ACK for some piece of this rsm */
9557 	if (rsm->r_start != start) {
9558 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9559 			/*
9560 			 * Before any splitting or hookery is
9561 			 * done is it a TLP of interest i.e. rxt?
9562 			 */
9563 			if ((rsm->r_flags & RACK_TLP) &&
9564 			    (rsm->r_rtr_cnt > 1)) {
9565 				/*
9566 				 * We are splitting a rxt TLP, check
9567 				 * if we need to save off the start/end
9568 				 */
9569 				if (rack->rc_last_tlp_acked_set &&
9570 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9571 					/*
9572 					 * We already turned this on since we are inside
9573 					 * the previous one was a partially sack now we
9574 					 * are getting another one (maybe all of it).
9575 					 *
9576 					 */
9577 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9578 					/*
9579 					 * Lets make sure we have all of it though.
9580 					 */
9581 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9582 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9583 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9584 								     rack->r_ctl.last_tlp_acked_end);
9585 					}
9586 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9587 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9588 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9589 								     rack->r_ctl.last_tlp_acked_end);
9590 					}
9591 				} else {
9592 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9593 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9594 					rack->rc_last_tlp_past_cumack = 0;
9595 					rack->rc_last_tlp_acked_set = 1;
9596 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9597 				}
9598 			}
9599 			/**
9600 			 * Need to split this in two pieces the before and after,
9601 			 * the before remains in the map, the after must be
9602 			 * added. In other words we have:
9603 			 * rsm        |--------------|
9604 			 * sackblk        |------->
9605 			 * rsm will become
9606 			 *     rsm    |---|
9607 			 * and nrsm will be  the sacked piece
9608 			 *     nrsm       |----------|
9609 			 *
9610 			 * But before we start down that path lets
9611 			 * see if the sack spans over on top of
9612 			 * the next guy and it is already sacked.
9613 			 *
9614 			 */
9615 			/*
9616 			 * Hookery can only be used if the two entries
9617 			 * are in the same bucket and neither one of
9618 			 * them staddle the bucket line.
9619 			 */
9620 			next = tqhash_next(rack->r_ctl.tqh, rsm);
9621 			if (next &&
9622 			    (rsm->bindex == next->bindex) &&
9623 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9624 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
9625 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
9626 			    (next->r_flags & RACK_IN_GP_WIN))
9627 				can_use_hookery = 1;
9628 			else if (next &&
9629 				 (rsm->bindex == next->bindex) &&
9630 				 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9631 				 ((next->r_flags & RACK_STRADDLE) == 0) &&
9632 				 ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9633 				 ((next->r_flags & RACK_IN_GP_WIN) == 0))
9634 				can_use_hookery = 1;
9635 			else
9636 				can_use_hookery = 0;
9637 			if (next && can_use_hookery &&
9638 			    (next->r_flags & RACK_ACKED) &&
9639 			    SEQ_GEQ(end, next->r_start)) {
9640 				/**
9641 				 * So the next one is already acked, and
9642 				 * we can thus by hookery use our stack_map
9643 				 * to reflect the piece being sacked and
9644 				 * then adjust the two tree entries moving
9645 				 * the start and ends around. So we start like:
9646 				 *  rsm     |------------|             (not-acked)
9647 				 *  next                 |-----------| (acked)
9648 				 *  sackblk        |-------->
9649 				 *  We want to end like so:
9650 				 *  rsm     |------|                   (not-acked)
9651 				 *  next           |-----------------| (acked)
9652 				 *  nrsm           |-----|
9653 				 * Where nrsm is a temporary stack piece we
9654 				 * use to update all the gizmos.
9655 				 */
9656 				/* Copy up our fudge block */
9657 				noextra++;
9658 				nrsm = &stack_map;
9659 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9660 				/* Now adjust our tree blocks */
9661 				rsm->r_end = start;
9662 				next->r_start = start;
9663  				rsm->r_flags |= RACK_SHUFFLED;
9664 				next->r_flags |= RACK_SHUFFLED;
9665 				/* Now we must adjust back where next->m is */
9666 				rack_setup_offset_for_rsm(rack, rsm, next);
9667 				/*
9668 				 * Which timestamp do we keep? It is rather
9669 				 * important in GP measurements to have the
9670 				 * accurate end of the send window.
9671 				 *
9672 				 * We keep the largest value, which is the newest
9673 				 * send. We do this in case a segment that is
9674 				 * joined together and not part of a GP estimate
9675 				 * later gets expanded into the GP estimate.
9676 				 *
9677 				 * We prohibit the merging of unlike kinds i.e.
9678 				 * all pieces that are in the GP estimate can be
9679 				 * merged and all pieces that are not in a GP estimate
9680 				 * can be merged, but not disimilar pieces. Combine
9681 				 * this with taking the highest here and we should
9682 				 * be ok unless of course the client reneges. Then
9683 				 * all bets are off.
9684 				 */
9685 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9686 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9687 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9688 				/*
9689 				 * And we must keep the newest ack arrival time.
9690 				 */
9691 				if (next->r_ack_arrival <
9692 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9693 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9694 
9695 
9696 				/* We don't need to adjust rsm, it did not change */
9697 				/* Clear out the dup ack count of the remainder */
9698 				rsm->r_dupack = 0;
9699 				rsm->r_just_ret = 0;
9700 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9701 				/* Now lets make sure our fudge block is right */
9702 				nrsm->r_start = start;
9703 				/* Now lets update all the stats and such */
9704 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9705 				if (rack->app_limited_needs_set)
9706 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9707 				changed += (nrsm->r_end - nrsm->r_start);
9708 				/* You get a count for acking a whole segment or more */
9709 				if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9710 					rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9711 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9712 				if (nrsm->r_flags & RACK_SACK_PASSED) {
9713 					rack->r_ctl.rc_reorder_ts = cts;
9714 					if (rack->r_ctl.rc_reorder_ts == 0)
9715 						rack->r_ctl.rc_reorder_ts = 1;
9716 				}
9717 				/*
9718 				 * Now we want to go up from rsm (the
9719 				 * one left un-acked) to the next one
9720 				 * in the tmap. We do this so when
9721 				 * we walk backwards we include marking
9722 				 * sack-passed on rsm (The one passed in
9723 				 * is skipped since it is generally called
9724 				 * on something sacked before removing it
9725 				 * from the tmap).
9726 				 */
9727 				if (rsm->r_in_tmap) {
9728 					nrsm = TAILQ_NEXT(rsm, r_tnext);
9729 					/*
9730 					 * Now that we have the next
9731 					 * one walk backwards from there.
9732 					 */
9733 					if (nrsm && nrsm->r_in_tmap)
9734 						rack_log_sack_passed(tp, rack, nrsm);
9735 				}
9736 				/* Now are we done? */
9737 				if (SEQ_LT(end, next->r_end) ||
9738 				    (end == next->r_end)) {
9739 					/* Done with block */
9740 					goto out;
9741 				}
9742 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9743 				counter_u64_add(rack_sack_used_next_merge, 1);
9744 				/* Postion for the next block */
9745 				start = next->r_end;
9746 				rsm = tqhash_next(rack->r_ctl.tqh, next);
9747 				if (rsm == NULL)
9748 					goto out;
9749 			} else {
9750 				/**
9751 				 * We can't use any hookery here, so we
9752 				 * need to split the map. We enter like
9753 				 * so:
9754 				 *  rsm      |--------|
9755 				 *  sackblk       |----->
9756 				 * We will add the new block nrsm and
9757 				 * that will be the new portion, and then
9758 				 * fall through after reseting rsm. So we
9759 				 * split and look like this:
9760 				 *  rsm      |----|
9761 				 *  sackblk       |----->
9762 				 *  nrsm          |---|
9763 				 * We then fall through reseting
9764 				 * rsm to nrsm, so the next block
9765 				 * picks it up.
9766 				 */
9767 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9768 				if (nrsm == NULL) {
9769 					/*
9770 					 * failed XXXrrs what can we do but loose the sack
9771 					 * info?
9772 					 */
9773 					goto out;
9774 				}
9775 				counter_u64_add(rack_sack_splits, 1);
9776 				rack_clone_rsm(rack, nrsm, rsm, start);
9777 				moved++;
9778 				rsm->r_just_ret = 0;
9779 #ifndef INVARIANTS
9780 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9781 #else
9782 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9783 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9784 					      nrsm, insret, rack, rsm);
9785 				}
9786 #endif
9787 				if (rsm->r_in_tmap) {
9788 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9789 					nrsm->r_in_tmap = 1;
9790 				}
9791 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9792 				rsm->r_flags &= (~RACK_HAS_FIN);
9793 				/* Position us to point to the new nrsm that starts the sack blk */
9794 				rsm = nrsm;
9795 			}
9796 		} else {
9797 			/* Already sacked this piece */
9798 			counter_u64_add(rack_sack_skipped_acked, 1);
9799 			moved++;
9800 			if (end == rsm->r_end) {
9801 				/* Done with block */
9802 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9803 				goto out;
9804 			} else if (SEQ_LT(end, rsm->r_end)) {
9805 				/* A partial sack to a already sacked block */
9806 				moved++;
9807 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9808 				goto out;
9809 			} else {
9810 				/*
9811 				 * The end goes beyond this guy
9812 				 * reposition the start to the
9813 				 * next block.
9814 				 */
9815 				start = rsm->r_end;
9816 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9817 				if (rsm == NULL)
9818 					goto out;
9819 			}
9820 		}
9821 	}
9822 	if (SEQ_GEQ(end, rsm->r_end)) {
9823 		/**
9824 		 * The end of this block is either beyond this guy or right
9825 		 * at this guy. I.e.:
9826 		 *  rsm ---                 |-----|
9827 		 *  end                     |-----|
9828 		 *  <or>
9829 		 *  end                     |---------|
9830 		 */
9831 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9832 			/*
9833 			 * Is it a TLP of interest?
9834 			 */
9835 			if ((rsm->r_flags & RACK_TLP) &&
9836 			    (rsm->r_rtr_cnt > 1)) {
9837 				/*
9838 				 * We are splitting a rxt TLP, check
9839 				 * if we need to save off the start/end
9840 				 */
9841 				if (rack->rc_last_tlp_acked_set &&
9842 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9843 					/*
9844 					 * We already turned this on since we are inside
9845 					 * the previous one was a partially sack now we
9846 					 * are getting another one (maybe all of it).
9847 					 */
9848 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9849 					/*
9850 					 * Lets make sure we have all of it though.
9851 					 */
9852 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9853 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9854 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9855 								     rack->r_ctl.last_tlp_acked_end);
9856 					}
9857 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9858 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9859 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9860 								     rack->r_ctl.last_tlp_acked_end);
9861 					}
9862 				} else {
9863 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9864 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9865 					rack->rc_last_tlp_past_cumack = 0;
9866 					rack->rc_last_tlp_acked_set = 1;
9867 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9868 				}
9869 			}
9870 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9871 			changed += (rsm->r_end - rsm->r_start);
9872 			/* You get a count for acking a whole segment or more */
9873 			if ((rsm->r_end - rsm->r_start) >= segsiz)
9874 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
9875 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9876 			if (rsm->r_in_tmap) /* should be true */
9877 				rack_log_sack_passed(tp, rack, rsm);
9878 			/* Is Reordering occuring? */
9879 			if (rsm->r_flags & RACK_SACK_PASSED) {
9880 				rsm->r_flags &= ~RACK_SACK_PASSED;
9881 				rack->r_ctl.rc_reorder_ts = cts;
9882 				if (rack->r_ctl.rc_reorder_ts == 0)
9883 					rack->r_ctl.rc_reorder_ts = 1;
9884 			}
9885 			if (rack->app_limited_needs_set)
9886 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9887 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9888 			rsm->r_flags |= RACK_ACKED;
9889 			if (rsm->r_in_tmap) {
9890 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9891 				rsm->r_in_tmap = 0;
9892 			}
9893 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9894 		} else {
9895 			counter_u64_add(rack_sack_skipped_acked, 1);
9896 			moved++;
9897 		}
9898 		if (end == rsm->r_end) {
9899 			/* This block only - done, setup for next */
9900 			goto out;
9901 		}
9902 		/*
9903 		 * There is more not coverend by this rsm move on
9904 		 * to the next block in the tail queue hash table.
9905 		 */
9906 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9907 		start = rsm->r_end;
9908 		rsm = nrsm;
9909 		if (rsm == NULL)
9910 			goto out;
9911 		goto do_rest_ofb;
9912 	}
9913 	/**
9914 	 * The end of this sack block is smaller than
9915 	 * our rsm i.e.:
9916 	 *  rsm ---                 |-----|
9917 	 *  end                     |--|
9918 	 */
9919 	if ((rsm->r_flags & RACK_ACKED) == 0) {
9920 		/*
9921 		 * Is it a TLP of interest?
9922 		 */
9923 		if ((rsm->r_flags & RACK_TLP) &&
9924 		    (rsm->r_rtr_cnt > 1)) {
9925 			/*
9926 			 * We are splitting a rxt TLP, check
9927 			 * if we need to save off the start/end
9928 			 */
9929 			if (rack->rc_last_tlp_acked_set &&
9930 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9931 				/*
9932 				 * We already turned this on since we are inside
9933 				 * the previous one was a partially sack now we
9934 				 * are getting another one (maybe all of it).
9935 				 */
9936 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9937 				/*
9938 				 * Lets make sure we have all of it though.
9939 				 */
9940 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9941 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9942 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9943 							     rack->r_ctl.last_tlp_acked_end);
9944 				}
9945 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9946 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9947 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9948 							     rack->r_ctl.last_tlp_acked_end);
9949 				}
9950 			} else {
9951 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9952 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9953 				rack->rc_last_tlp_past_cumack = 0;
9954 				rack->rc_last_tlp_acked_set = 1;
9955 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9956 			}
9957 		}
9958 		/*
9959 		 * Hookery can only be used if the two entries
9960 		 * are in the same bucket and neither one of
9961 		 * them staddle the bucket line.
9962 		 */
9963 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9964 		if (prev &&
9965 		    (rsm->bindex == prev->bindex) &&
9966 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9967 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
9968 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
9969 		    (prev->r_flags & RACK_IN_GP_WIN))
9970 			can_use_hookery = 1;
9971 		else if (prev &&
9972 			 (rsm->bindex == prev->bindex) &&
9973 			 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9974 			 ((prev->r_flags & RACK_STRADDLE) == 0) &&
9975 			 ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9976 			 ((prev->r_flags & RACK_IN_GP_WIN) == 0))
9977 			can_use_hookery = 1;
9978 		else
9979 			can_use_hookery = 0;
9980 
9981 		if (prev && can_use_hookery &&
9982 		    (prev->r_flags & RACK_ACKED)) {
9983 			/**
9984 			 * Goal, we want the right remainder of rsm to shrink
9985 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
9986 			 * We want to expand prev to go all the way
9987 			 * to prev->r_end <- end.
9988 			 * so in the tree we have before:
9989 			 *   prev     |--------|         (acked)
9990 			 *   rsm               |-------| (non-acked)
9991 			 *   sackblk           |-|
9992 			 * We churn it so we end up with
9993 			 *   prev     |----------|       (acked)
9994 			 *   rsm                 |-----| (non-acked)
9995 			 *   nrsm              |-| (temporary)
9996 			 *
9997 			 * Note if either prev/rsm is a TLP we don't
9998 			 * do this.
9999 			 */
10000 			noextra++;
10001 			nrsm = &stack_map;
10002 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10003 			prev->r_end = end;
10004 			rsm->r_start = end;
10005 			rsm->r_flags |= RACK_SHUFFLED;
10006 			prev->r_flags |= RACK_SHUFFLED;
10007 			/* Now adjust nrsm (stack copy) to be
10008 			 * the one that is the small
10009 			 * piece that was "sacked".
10010 			 */
10011 			nrsm->r_end = end;
10012 			rsm->r_dupack = 0;
10013 			/*
10014 			 * Which timestamp do we keep? It is rather
10015 			 * important in GP measurements to have the
10016 			 * accurate end of the send window.
10017 			 *
10018 			 * We keep the largest value, which is the newest
10019 			 * send. We do this in case a segment that is
10020 			 * joined together and not part of a GP estimate
10021 			 * later gets expanded into the GP estimate.
10022 			 *
10023 			 * We prohibit the merging of unlike kinds i.e.
10024 			 * all pieces that are in the GP estimate can be
10025 			 * merged and all pieces that are not in a GP estimate
10026 			 * can be merged, but not disimilar pieces. Combine
10027 			 * this with taking the highest here and we should
10028 			 * be ok unless of course the client reneges. Then
10029 			 * all bets are off.
10030 			 */
10031 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
10032 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
10033 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
10034 			}
10035 			/*
10036 			 * And we must keep the newest ack arrival time.
10037 			 */
10038 
10039 			if(prev->r_ack_arrival <
10040 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10041 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10042 
10043 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10044 			/*
10045 			 * Now that the rsm has had its start moved forward
10046 			 * lets go ahead and get its new place in the world.
10047 			 */
10048 			rack_setup_offset_for_rsm(rack, prev, rsm);
10049 			/*
10050 			 * Now nrsm is our new little piece
10051 			 * that is acked (which was merged
10052 			 * to prev). Update the rtt and changed
10053 			 * based on that. Also check for reordering.
10054 			 */
10055 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10056 			if (rack->app_limited_needs_set)
10057 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10058 			changed += (nrsm->r_end - nrsm->r_start);
10059 			/* You get a count for acking a whole segment or more */
10060 			if ((nrsm->r_end - nrsm->r_start) >= segsiz)
10061 				rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
10062 
10063 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10064 			if (nrsm->r_flags & RACK_SACK_PASSED) {
10065 				rack->r_ctl.rc_reorder_ts = cts;
10066 				if (rack->r_ctl.rc_reorder_ts == 0)
10067 					rack->r_ctl.rc_reorder_ts = 1;
10068 			}
10069 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
10070 			rsm = prev;
10071 			counter_u64_add(rack_sack_used_prev_merge, 1);
10072 		} else {
10073 			/**
10074 			 * This is the case where our previous
10075 			 * block is not acked either, so we must
10076 			 * split the block in two.
10077 			 */
10078 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10079 			if (nrsm == NULL) {
10080 				/* failed rrs what can we do but loose the sack info? */
10081 				goto out;
10082 			}
10083 			if ((rsm->r_flags & RACK_TLP) &&
10084 			    (rsm->r_rtr_cnt > 1)) {
10085 				/*
10086 				 * We are splitting a rxt TLP, check
10087 				 * if we need to save off the start/end
10088 				 */
10089 				if (rack->rc_last_tlp_acked_set &&
10090 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10091 					/*
10092 					 * We already turned this on since this block is inside
10093 					 * the previous one was a partially sack now we
10094 					 * are getting another one (maybe all of it).
10095 					 */
10096 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10097 					/*
10098 					 * Lets make sure we have all of it though.
10099 					 */
10100 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10101 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10102 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10103 								     rack->r_ctl.last_tlp_acked_end);
10104 					}
10105 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10106 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10107 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10108 								     rack->r_ctl.last_tlp_acked_end);
10109 					}
10110 				} else {
10111 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10112 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10113 					rack->rc_last_tlp_acked_set = 1;
10114 					rack->rc_last_tlp_past_cumack = 0;
10115 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10116 				}
10117 			}
10118 			/**
10119 			 * In this case nrsm becomes
10120 			 * nrsm->r_start = end;
10121 			 * nrsm->r_end = rsm->r_end;
10122 			 * which is un-acked.
10123 			 * <and>
10124 			 * rsm->r_end = nrsm->r_start;
10125 			 * i.e. the remaining un-acked
10126 			 * piece is left on the left
10127 			 * hand side.
10128 			 *
10129 			 * So we start like this
10130 			 * rsm      |----------| (not acked)
10131 			 * sackblk  |---|
10132 			 * build it so we have
10133 			 * rsm      |---|         (acked)
10134 			 * nrsm         |------|  (not acked)
10135 			 */
10136 			counter_u64_add(rack_sack_splits, 1);
10137 			rack_clone_rsm(rack, nrsm, rsm, end);
10138 			moved++;
10139 			rsm->r_flags &= (~RACK_HAS_FIN);
10140 			rsm->r_just_ret = 0;
10141 #ifndef INVARIANTS
10142 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10143 #else
10144 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10145 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
10146 				      nrsm, insret, rack, rsm);
10147 			}
10148 #endif
10149 			if (rsm->r_in_tmap) {
10150 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10151 				nrsm->r_in_tmap = 1;
10152 			}
10153 			nrsm->r_dupack = 0;
10154 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10155 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10156 			changed += (rsm->r_end - rsm->r_start);
10157 			/* You get a count for acking a whole segment or more */
10158 			if ((rsm->r_end - rsm->r_start) >= segsiz)
10159 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10160 
10161 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10162 			if (rsm->r_in_tmap) /* should be true */
10163 				rack_log_sack_passed(tp, rack, rsm);
10164 			/* Is Reordering occuring? */
10165 			if (rsm->r_flags & RACK_SACK_PASSED) {
10166 				rsm->r_flags &= ~RACK_SACK_PASSED;
10167 				rack->r_ctl.rc_reorder_ts = cts;
10168 				if (rack->r_ctl.rc_reorder_ts == 0)
10169 					rack->r_ctl.rc_reorder_ts = 1;
10170 			}
10171 			if (rack->app_limited_needs_set)
10172 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10173 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10174 			rsm->r_flags |= RACK_ACKED;
10175 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10176 			if (rsm->r_in_tmap) {
10177 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10178 				rsm->r_in_tmap = 0;
10179 			}
10180 		}
10181 	} else if (start != end){
10182 		/*
10183 		 * The block was already acked.
10184 		 */
10185 		counter_u64_add(rack_sack_skipped_acked, 1);
10186 		moved++;
10187 	}
10188 out:
10189 	if (rsm &&
10190 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10191 	    (rsm->r_flags & RACK_ACKED)) {
10192 		/*
10193 		 * Now can we merge where we worked
10194 		 * with either the previous or
10195 		 * next block?
10196 		 */
10197 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10198 		while (next) {
10199 			if (next->r_flags & RACK_TLP)
10200 				break;
10201 			/* Only allow merges between ones in or out of GP window */
10202 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10203 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10204 				break;
10205 			}
10206 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10207 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10208 				break;
10209 			}
10210 			if (rsm->bindex != next->bindex)
10211 				break;
10212 			if (rsm->r_flags & RACK_STRADDLE)
10213 				break;
10214 			if (next->r_flags & RACK_STRADDLE)
10215 				break;
10216 			if (next->r_flags & RACK_ACKED) {
10217 				/* yep this and next can be merged */
10218 				rsm = rack_merge_rsm(rack, rsm, next);
10219 				noextra++;
10220 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10221 			} else
10222 				break;
10223 		}
10224 		/* Now what about the previous? */
10225 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10226 		while (prev) {
10227 			if (prev->r_flags & RACK_TLP)
10228 				break;
10229 			/* Only allow merges between ones in or out of GP window */
10230 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10231 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10232 				break;
10233 			}
10234 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10235 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10236 				break;
10237 			}
10238 			if (rsm->bindex != prev->bindex)
10239 				break;
10240 			if (rsm->r_flags & RACK_STRADDLE)
10241 				break;
10242 			if (prev->r_flags & RACK_STRADDLE)
10243 				break;
10244 			if (prev->r_flags & RACK_ACKED) {
10245 				/* yep the previous and this can be merged */
10246 				rsm = rack_merge_rsm(rack, prev, rsm);
10247 				noextra++;
10248 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10249 			} else
10250 				break;
10251 		}
10252 	}
10253 	if (used_ref == 0) {
10254 		counter_u64_add(rack_sack_proc_all, 1);
10255 	} else {
10256 		counter_u64_add(rack_sack_proc_short, 1);
10257 	}
10258 	/* Save off the next one for quick reference. */
10259 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10260 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10261 	/* Pass back the moved. */
10262 	*moved_two = moved;
10263 	*no_extra = noextra;
10264 	return (changed);
10265 }
10266 
10267 static void inline
10268 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10269 {
10270 	struct rack_sendmap *tmap;
10271 
10272 	tmap = NULL;
10273 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10274 		/* Its no longer sacked, mark it so */
10275 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10276 #ifdef INVARIANTS
10277 		if (rsm->r_in_tmap) {
10278 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10279 			      rack, rsm, rsm->r_flags);
10280 		}
10281 #endif
10282 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10283 		/* Rebuild it into our tmap */
10284 		if (tmap == NULL) {
10285 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10286 			tmap = rsm;
10287 		} else {
10288 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10289 			tmap = rsm;
10290 		}
10291 		tmap->r_in_tmap = 1;
10292 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10293 	}
10294 	/*
10295 	 * Now lets possibly clear the sack filter so we start
10296 	 * recognizing sacks that cover this area.
10297 	 */
10298 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10299 
10300 }
10301 
10302 static void
10303 rack_do_decay(struct tcp_rack *rack)
10304 {
10305 	struct timeval res;
10306 
10307 #define	timersub(tvp, uvp, vvp)						\
10308 	do {								\
10309 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
10310 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
10311 		if ((vvp)->tv_usec < 0) {				\
10312 			(vvp)->tv_sec--;				\
10313 			(vvp)->tv_usec += 1000000;			\
10314 		}							\
10315 	} while (0)
10316 
10317 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
10318 #undef timersub
10319 
10320 	rack->r_ctl.input_pkt++;
10321 	if ((rack->rc_in_persist) ||
10322 	    (res.tv_sec >= 1) ||
10323 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
10324 		/*
10325 		 * Check for decay of non-SAD,
10326 		 * we want all SAD detection metrics to
10327 		 * decay 1/4 per second (or more) passed.
10328 		 * Current default is 800 so it decays
10329 		 * 80% every second.
10330 		 */
10331 #ifdef TCP_SAD_DETECTION
10332 		uint32_t pkt_delta;
10333 
10334 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
10335 #endif
10336 		/* Update our saved tracking values */
10337 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
10338 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10339 		/* Now do we escape without decay? */
10340 #ifdef TCP_SAD_DETECTION
10341 		if (rack->rc_in_persist ||
10342 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
10343 		    (pkt_delta < tcp_sad_low_pps)){
10344 			/*
10345 			 * We don't decay idle connections
10346 			 * or ones that have a low input pps.
10347 			 */
10348 			return;
10349 		}
10350 		/* Decay the counters */
10351 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
10352 							tcp_sad_decay_val);
10353 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
10354 							 tcp_sad_decay_val);
10355 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
10356 							       tcp_sad_decay_val);
10357 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
10358 								tcp_sad_decay_val);
10359 #endif
10360 	}
10361 }
10362 
10363 static void inline
10364 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10365 {
10366 	/*
10367 	 * We look at advancing the end send time for our GP
10368 	 * measurement tracking only as the cumulative acknowledgment
10369 	 * moves forward. You might wonder about this, why not
10370 	 * at every transmission or retransmission within the
10371 	 * GP window update the rc_gp_cumack_ts? Well its rather
10372 	 * nuanced but basically the GP window *may* expand (as
10373 	 * it does below) or worse and harder to track it may shrink.
10374 	 *
10375 	 * This last makes it impossible to track at the time of
10376 	 * the send, since you may set forward your rc_gp_cumack_ts
10377 	 * when you send, because that send *is* in your currently
10378 	 * "guessed" window, but then it shrinks. Now which was
10379 	 * the send time of the last bytes in the window, by the
10380 	 * time you ask that question that part of the sendmap
10381 	 * is freed. So you don't know and you will have too
10382 	 * long of send window. Instead by updating the time
10383 	 * marker only when the cumack advances this assures us
10384 	 * that we will have only the sends in the window of our
10385 	 * GP measurement.
10386 	 *
10387 	 * Another complication from this is the
10388 	 * merging of sendmap entries. During SACK processing this
10389 	 * can happen to conserve the sendmap size. That breaks
10390 	 * everything down in tracking the send window of the GP
10391 	 * estimate. So to prevent that and keep it working with
10392 	 * a tiny bit more limited merging, we only allow like
10393 	 * types to be merged. I.e. if two sends are in the GP window
10394 	 * then its ok to merge them together. If two sends are not
10395 	 * in the GP window its ok to merge them together too. Though
10396 	 * one send in and one send out cannot be merged. We combine
10397 	 * this with never allowing the shrinking of the GP window when
10398 	 * we are in recovery so that we can properly calculate the
10399 	 * sending times.
10400 	 *
10401 	 * This all of course seems complicated, because it is.. :)
10402 	 *
10403 	 * The cum-ack is being advanced upon the sendmap.
10404 	 * If we are not doing a GP estimate don't
10405 	 * proceed.
10406 	 */
10407 	uint64_t ts;
10408 
10409 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10410 		return;
10411 	/*
10412 	 * If this sendmap entry is going
10413 	 * beyond the measurement window we had picked,
10414 	 * expand the measurement window by that much.
10415 	 */
10416 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10417 		tp->gput_ack = rsm->r_end;
10418 	}
10419 	/*
10420 	 * If we have not setup a ack, then we
10421 	 * have no idea if the newly acked pieces
10422 	 * will be "in our seq measurement range". If
10423 	 * it is when we clear the app_limited_needs_set
10424 	 * flag the timestamp will be updated.
10425 	 */
10426 	if (rack->app_limited_needs_set)
10427 		return;
10428 	/*
10429 	 * Finally, we grab out the latest timestamp
10430 	 * that this packet was sent and then see
10431 	 * if:
10432 	 *  a) The packet touches are newly defined GP range.
10433 	 *  b) The time is greater than (newer) than the
10434 	 *     one we currently have. If so we update
10435 	 *     our sending end time window.
10436 	 *
10437 	 * Note we *do not* do this at send time. The reason
10438 	 * is that if you do you *may* pick up a newer timestamp
10439 	 * for a range you are not going to measure. We project
10440 	 * out how far and then sometimes modify that to be
10441 	 * smaller. If that occurs then you will have a send
10442 	 * that does not belong to the range included.
10443 	 */
10444 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10445 	    rack->r_ctl.rc_gp_cumack_ts)
10446 		return;
10447 	if (rack_in_gp_window(tp, rsm)) {
10448 		rack->r_ctl.rc_gp_cumack_ts = ts;
10449 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10450 			       __LINE__, from, rsm);
10451 	}
10452 }
10453 
10454 static void
10455 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)
10456 {
10457 	struct rack_sendmap *rsm;
10458 	/*
10459 	 * The ACK point is advancing to th_ack, we must drop off
10460 	 * the packets in the rack log and calculate any eligble
10461 	 * RTT's.
10462 	 */
10463 
10464 	rack->r_wanted_output = 1;
10465 	if (SEQ_GT(th_ack, tp->snd_una))
10466 		rack->r_ctl.last_cumack_advance = acktime;
10467 
10468 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10469 	if ((rack->rc_last_tlp_acked_set == 1)&&
10470 	    (rack->rc_last_tlp_past_cumack == 1) &&
10471 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10472 		/*
10473 		 * We have reached the point where our last rack
10474 		 * tlp retransmit sequence is ahead of the cum-ack.
10475 		 * This can only happen when the cum-ack moves all
10476 		 * the way around (its been a full 2^^31+1 bytes
10477 		 * or more since we sent a retransmitted TLP). Lets
10478 		 * turn off the valid flag since its not really valid.
10479 		 *
10480 		 * Note since sack's also turn on this event we have
10481 		 * a complication, we have to wait to age it out until
10482 		 * the cum-ack is by the TLP before checking which is
10483 		 * what the next else clause does.
10484 		 */
10485 		rack_log_dsack_event(rack, 9, __LINE__,
10486 				     rack->r_ctl.last_tlp_acked_start,
10487 				     rack->r_ctl.last_tlp_acked_end);
10488 		rack->rc_last_tlp_acked_set = 0;
10489 		rack->rc_last_tlp_past_cumack = 0;
10490 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10491 		   (rack->rc_last_tlp_past_cumack == 0) &&
10492 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10493 		/*
10494 		 * It is safe to start aging TLP's out.
10495 		 */
10496 		rack->rc_last_tlp_past_cumack = 1;
10497 	}
10498 	/* We do the same for the tlp send seq as well */
10499 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10500 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10501 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10502 		rack_log_dsack_event(rack, 9, __LINE__,
10503 				     rack->r_ctl.last_sent_tlp_seq,
10504 				     (rack->r_ctl.last_sent_tlp_seq +
10505 				      rack->r_ctl.last_sent_tlp_len));
10506 		rack->rc_last_sent_tlp_seq_valid = 0;
10507 		rack->rc_last_sent_tlp_past_cumack = 0;
10508 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10509 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10510 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10511 		/*
10512 		 * It is safe to start aging TLP's send.
10513 		 */
10514 		rack->rc_last_sent_tlp_past_cumack = 1;
10515 	}
10516 more:
10517 	rsm = tqhash_min(rack->r_ctl.tqh);
10518 	if (rsm == NULL) {
10519 		if ((th_ack - 1) == tp->iss) {
10520 			/*
10521 			 * For the SYN incoming case we will not
10522 			 * have called tcp_output for the sending of
10523 			 * the SYN, so there will be no map. All
10524 			 * other cases should probably be a panic.
10525 			 */
10526 			return;
10527 		}
10528 		if (tp->t_flags & TF_SENTFIN) {
10529 			/* if we sent a FIN we often will not have map */
10530 			return;
10531 		}
10532 #ifdef INVARIANTS
10533 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
10534 		      tp,
10535 		      tp->t_state, th_ack, rack,
10536 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
10537 #endif
10538 		return;
10539 	}
10540 	if (SEQ_LT(th_ack, rsm->r_start)) {
10541 		/* Huh map is missing this */
10542 #ifdef INVARIANTS
10543 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10544 		       rsm->r_start,
10545 		       th_ack, tp->t_state, rack->r_state);
10546 #endif
10547 		return;
10548 	}
10549 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10550 
10551 	/* Now was it a retransmitted TLP? */
10552 	if ((rsm->r_flags & RACK_TLP) &&
10553 	    (rsm->r_rtr_cnt > 1)) {
10554 		/*
10555 		 * Yes, this rsm was a TLP and retransmitted, remember that
10556 		 * since if a DSACK comes back on this we don't want
10557 		 * to think of it as a reordered segment. This may
10558 		 * get updated again with possibly even other TLPs
10559 		 * in flight, but thats ok. Only when we don't send
10560 		 * a retransmitted TLP for 1/2 the sequences space
10561 		 * will it get turned off (above).
10562 		 */
10563 		if (rack->rc_last_tlp_acked_set &&
10564 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10565 			/*
10566 			 * We already turned this on since the end matches,
10567 			 * the previous one was a partially ack now we
10568 			 * are getting another one (maybe all of it).
10569 			 */
10570 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10571 			/*
10572 			 * Lets make sure we have all of it though.
10573 			 */
10574 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10575 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10576 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10577 						     rack->r_ctl.last_tlp_acked_end);
10578 			}
10579 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10580 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10581 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10582 						     rack->r_ctl.last_tlp_acked_end);
10583 			}
10584 		} else {
10585 			rack->rc_last_tlp_past_cumack = 1;
10586 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10587 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10588 			rack->rc_last_tlp_acked_set = 1;
10589 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10590 		}
10591 	}
10592 	/* Now do we consume the whole thing? */
10593 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10594 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
10595 		/* Its all consumed. */
10596 		uint32_t left;
10597 		uint8_t newly_acked;
10598 
10599 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10600 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10601 		rsm->r_rtr_bytes = 0;
10602 		/*
10603 		 * Record the time of highest cumack sent if its in our measurement
10604 		 * window and possibly bump out the end.
10605 		 */
10606 		rack_rsm_sender_update(rack, tp, rsm, 4);
10607 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10608 		if (rsm->r_in_tmap) {
10609 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10610 			rsm->r_in_tmap = 0;
10611 		}
10612 		newly_acked = 1;
10613 		if (rsm->r_flags & RACK_ACKED) {
10614 			/*
10615 			 * It was acked on the scoreboard -- remove
10616 			 * it from total
10617 			 */
10618 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10619 			newly_acked = 0;
10620 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
10621 			/*
10622 			 * There are segments ACKED on the
10623 			 * scoreboard further up. We are seeing
10624 			 * reordering.
10625 			 */
10626 			rsm->r_flags &= ~RACK_SACK_PASSED;
10627 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10628 			rsm->r_flags |= RACK_ACKED;
10629 			rack->r_ctl.rc_reorder_ts = cts;
10630 			if (rack->r_ctl.rc_reorder_ts == 0)
10631 				rack->r_ctl.rc_reorder_ts = 1;
10632 			if (rack->r_ent_rec_ns) {
10633 				/*
10634 				 * We have sent no more, and we saw an sack
10635 				 * then ack arrive.
10636 				 */
10637 				rack->r_might_revert = 1;
10638 			}
10639 		}
10640 		if ((rsm->r_flags & RACK_TO_REXT) &&
10641 		    (tp->t_flags & TF_RCVD_TSTMP) &&
10642 		    (to->to_flags & TOF_TS) &&
10643 		    (to->to_tsecr != 0) &&
10644 		    (tp->t_flags & TF_PREVVALID)) {
10645 			/*
10646 			 * We can use the timestamp to see
10647 			 * if this retransmission was from the
10648 			 * first transmit. If so we made a mistake.
10649 			 */
10650 			tp->t_flags &= ~TF_PREVVALID;
10651 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10652 				/* The first transmit is what this ack is for */
10653 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10654 			}
10655 		}
10656 		left = th_ack - rsm->r_end;
10657 		if (rack->app_limited_needs_set && newly_acked)
10658 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10659 		/* Free back to zone */
10660 		rack_free(rack, rsm);
10661 		if (left) {
10662 			goto more;
10663 		}
10664 		/* Check for reneging */
10665 		rsm = tqhash_min(rack->r_ctl.tqh);
10666 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10667 			/*
10668 			 * The peer has moved snd_una up to
10669 			 * the edge of this send, i.e. one
10670 			 * that it had previously acked. The only
10671 			 * way that can be true if the peer threw
10672 			 * away data (space issues) that it had
10673 			 * previously sacked (else it would have
10674 			 * given us snd_una up to (rsm->r_end).
10675 			 * We need to undo the acked markings here.
10676 			 *
10677 			 * Note we have to look to make sure th_ack is
10678 			 * our rsm->r_start in case we get an old ack
10679 			 * where th_ack is behind snd_una.
10680 			 */
10681 			rack_peer_reneges(rack, rsm, th_ack);
10682 		}
10683 		return;
10684 	}
10685 	if (rsm->r_flags & RACK_ACKED) {
10686 		/*
10687 		 * It was acked on the scoreboard -- remove it from
10688 		 * total for the part being cum-acked.
10689 		 */
10690 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10691 	}
10692 	/*
10693 	 * Clear the dup ack count for
10694 	 * the piece that remains.
10695 	 */
10696 	rsm->r_dupack = 0;
10697 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10698 	if (rsm->r_rtr_bytes) {
10699 		/*
10700 		 * It was retransmitted adjust the
10701 		 * sack holes for what was acked.
10702 		 */
10703 		int ack_am;
10704 
10705 		ack_am = (th_ack - rsm->r_start);
10706 		if (ack_am >= rsm->r_rtr_bytes) {
10707 			rack->r_ctl.rc_holes_rxt -= ack_am;
10708 			rsm->r_rtr_bytes -= ack_am;
10709 		}
10710 	}
10711 	/*
10712 	 * Update where the piece starts and record
10713 	 * the time of send of highest cumack sent if
10714 	 * its in our GP range.
10715 	 */
10716 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10717 	/* Now we need to move our offset forward too */
10718 	if (rsm->m &&
10719 	    ((rsm->orig_m_len != rsm->m->m_len) ||
10720 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10721 		/* Fix up the orig_m_len and possibly the mbuf offset */
10722 		rack_adjust_orig_mlen(rsm);
10723 	}
10724 	rsm->soff += (th_ack - rsm->r_start);
10725 	rack_rsm_sender_update(rack, tp, rsm, 5);
10726 	/* The trim will move th_ack into r_start for us */
10727 	tqhash_trim(rack->r_ctl.tqh, th_ack);
10728 	/* Now do we need to move the mbuf fwd too? */
10729 	{
10730 		struct mbuf *m;
10731 		uint32_t soff;
10732 
10733 		m = rsm->m;
10734 		soff = rsm->soff;
10735 		if (m) {
10736 			while (soff >= m->m_len) {
10737 				soff -= m->m_len;
10738 				KASSERT((m->m_next != NULL),
10739 					(" rsm:%p  off:%u soff:%u m:%p",
10740 					 rsm, rsm->soff, soff, m));
10741 				m = m->m_next;
10742 				if (m == NULL) {
10743 					/*
10744 					 * This is a fall-back that prevents a panic. In reality
10745 					 * we should be able to walk the mbuf's and find our place.
10746 					 * At this point snd_una has not been updated with the sbcut() yet
10747 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
10748 					 * should work fine. This is undesirable since we will take cache
10749 					 * hits to access the socket buffer. And even more puzzling is that
10750 					 * it happens occasionally. It should not :(
10751 					 */
10752 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10753 						      (rsm->r_start - tp->snd_una),
10754 						      &soff);
10755 					break;
10756 				}
10757 			}
10758 			/*
10759 			 * Now save in our updated values.
10760 			 */
10761 			rsm->m = m;
10762 			rsm->soff = soff;
10763 			rsm->orig_m_len = rsm->m->m_len;
10764 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10765 		}
10766 	}
10767 	if (rack->app_limited_needs_set &&
10768 	    SEQ_GEQ(th_ack, tp->gput_seq))
10769 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10770 }
10771 
10772 static void
10773 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10774 {
10775 	struct rack_sendmap *rsm;
10776 	int sack_pass_fnd = 0;
10777 
10778 	if (rack->r_might_revert) {
10779 		/*
10780 		 * Ok we have reordering, have not sent anything, we
10781 		 * might want to revert the congestion state if nothing
10782 		 * further has SACK_PASSED on it. Lets check.
10783 		 *
10784 		 * We also get here when we have DSACKs come in for
10785 		 * all the data that we FR'd. Note that a rxt or tlp
10786 		 * timer clears this from happening.
10787 		 */
10788 
10789 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10790 			if (rsm->r_flags & RACK_SACK_PASSED) {
10791 				sack_pass_fnd = 1;
10792 				break;
10793 			}
10794 		}
10795 		if (sack_pass_fnd == 0) {
10796 			/*
10797 			 * We went into recovery
10798 			 * incorrectly due to reordering!
10799 			 */
10800 			int orig_cwnd;
10801 
10802 			rack->r_ent_rec_ns = 0;
10803 			orig_cwnd = tp->snd_cwnd;
10804 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10805 			tp->snd_recover = tp->snd_una;
10806 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10807 			EXIT_RECOVERY(tp->t_flags);
10808 		}
10809 		rack->r_might_revert = 0;
10810 	}
10811 }
10812 
10813 #ifdef TCP_SAD_DETECTION
10814 
10815 static void
10816 rack_merge_out_sacks(struct tcp_rack *rack)
10817 {
10818 	struct rack_sendmap *cur, *next, *rsm, *trsm = NULL;
10819 
10820 	cur = tqhash_min(rack->r_ctl.tqh);
10821 	while(cur) {
10822 		next = tqhash_next(rack->r_ctl.tqh, cur);
10823 		/*
10824 		 * The idea is to go through all and merge back
10825 		 * together the pieces sent together,
10826 		 */
10827 		if ((next != NULL) &&
10828 		    (cur->r_tim_lastsent[0] == next->r_tim_lastsent[0])) {
10829 			rack_merge_rsm(rack, cur, next);
10830 		} else {
10831 			cur = next;
10832 		}
10833 	}
10834 	/*
10835 	 * now treat it like a rxt event, everything is outstanding
10836 	 * and sent nothing acvked and dupacks are all zero. If this
10837 	 * is not an attacker it will have to dupack its way through
10838 	 * it all.
10839 	 */
10840 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
10841 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
10842 		rsm->r_dupack = 0;
10843 		/* We must re-add it back to the tlist */
10844 		if (trsm == NULL) {
10845 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10846 		} else {
10847 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
10848 		}
10849 		rsm->r_in_tmap = 1;
10850 		trsm = rsm;
10851 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
10852 	}
10853 	sack_filter_clear(&rack->r_ctl.rack_sf, rack->rc_tp->snd_una);
10854 }
10855 
10856 static void
10857 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
10858 {
10859 	int do_detection = 0;
10860 
10861 	if (rack->sack_attack_disable || rack->rc_suspicious) {
10862 		/*
10863 		 * If we have been disabled we must detect
10864 		 * to possibly reverse it. Or if the guy has
10865 		 * sent in suspicious sacks we want to do detection too.
10866 		 */
10867 		do_detection = 1;
10868 
10869 	} else if  ((rack->do_detection || tcp_force_detection) &&
10870 		    (tcp_sack_to_ack_thresh > 0) &&
10871 		    (tcp_sack_to_move_thresh > 0) &&
10872 		    (rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum)) {
10873 		/*
10874 		 * We only detect here if:
10875 		 * 1) System wide forcing is on <or> do_detection is on
10876 		 *   <and>
10877 		 * 2) We have thresholds for move and ack (set one to 0 and we are off)
10878 		 *   <and>
10879 		 * 3) We have maps allocated larger than our min (500).
10880 		 */
10881 		do_detection = 1;
10882 	}
10883 	if (do_detection > 0) {
10884 		/*
10885 		 * We have thresholds set to find
10886 		 * possible attackers and disable sack.
10887 		 * Check them.
10888 		 */
10889 		uint64_t ackratio, moveratio, movetotal;
10890 
10891 		/* Log detecting */
10892 		rack_log_sad(rack, 1);
10893 		/* Do we establish a ack ratio */
10894 		if ((rack->r_ctl.sack_count > tcp_map_minimum)  ||
10895 		    (rack->rc_suspicious == 1) ||
10896 		    (rack->sack_attack_disable > 0)) {
10897 			ackratio = (uint64_t)(rack->r_ctl.sack_count);
10898 			ackratio *= (uint64_t)(1000);
10899 			if (rack->r_ctl.ack_count)
10900 				ackratio /= (uint64_t)(rack->r_ctl.ack_count);
10901 			else {
10902 				/* We can hit this due to ack totals degregation (via small sacks) */
10903 				ackratio = 1000;
10904 			}
10905 		} else {
10906 			/*
10907 			 * No ack ratio needed if we have not
10908 			 * seen more sacks then the number of map entries.
10909 			 * The exception to that is if we have disabled sack then
10910 			 * we need to find a ratio.
10911 			 */
10912 			ackratio = 0;
10913 		}
10914 
10915 		if ((rack->sack_attack_disable == 0) &&
10916 		    (ackratio > rack_highest_sack_thresh_seen))
10917 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
10918 		/* Do we establish a move ratio? */
10919 		if ((rack->r_ctl.sack_moved_extra > tcp_map_minimum) ||
10920 		    (rack->rc_suspicious == 1) ||
10921 		    (rack->sack_attack_disable > 0)) {
10922 			/*
10923 			 * We need to have more sack moves than maps
10924 			 * allocated to have a move ratio considered.
10925 			 */
10926 			movetotal = rack->r_ctl.sack_moved_extra;
10927 			movetotal += rack->r_ctl.sack_noextra_move;
10928 			moveratio = rack->r_ctl.sack_moved_extra;
10929 			moveratio *= (uint64_t)1000;
10930 			if (movetotal)
10931 				moveratio /= movetotal;
10932 			else {
10933 				/* No moves, thats pretty good */
10934 				moveratio = 0;
10935 			}
10936 		} else {
10937 			/*
10938 			 * Not enough moves have occured to consider
10939 			 * if we are out of whack in that ratio.
10940 			 * The exception to that is if we have disabled sack then
10941 			 * we need to find a ratio.
10942 			 */
10943 			moveratio = 0;
10944 		}
10945 		if ((rack->sack_attack_disable == 0) &&
10946 		    (moveratio > rack_highest_move_thresh_seen))
10947 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
10948 		/* Now the tests */
10949 		if (rack->sack_attack_disable == 0) {
10950 			/* Not disabled, do we need to disable? */
10951 			if ((ackratio > tcp_sack_to_ack_thresh) &&
10952 			    (moveratio > tcp_sack_to_move_thresh)) {
10953 				/* Disable sack processing */
10954 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10955 				rack->sack_attack_disable = 1;
10956 				/* set it so we have the built in delay */
10957 				rack->r_ctl.ack_during_sd = 1;
10958 				if (rack_merge_out_sacks_on_attack)
10959 					rack_merge_out_sacks(rack);
10960 				counter_u64_add(rack_sack_attacks_detected, 1);
10961 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10962 				/* Clamp the cwnd at flight size */
10963 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
10964 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10965 				rack_log_sad(rack, 2);
10966 			}
10967 		} else {
10968 			/* We are sack-disabled check for false positives */
10969 			if ((ackratio <= tcp_restoral_thresh) ||
10970 			    ((rack_merge_out_sacks_on_attack == 0) &&
10971 			     (rack->rc_suspicious == 0) &&
10972 			     (rack->r_ctl.rc_num_maps_alloced <= (tcp_map_minimum/2)))) {
10973 				rack->sack_attack_disable = 0;
10974 				rack_log_sad(rack, 3);
10975 				/* Restart counting */
10976 				rack->r_ctl.sack_count = 0;
10977 				rack->r_ctl.sack_moved_extra = 0;
10978 				rack->r_ctl.sack_noextra_move = 1;
10979 				rack->rc_suspicious = 0;
10980 				rack->r_ctl.ack_count = max(1,
10981 							    (bytes_this_ack / segsiz));
10982 
10983 				counter_u64_add(rack_sack_attacks_reversed, 1);
10984 				/* Restore the cwnd */
10985 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
10986 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
10987 			}
10988 		}
10989 	}
10990 }
10991 #endif
10992 
10993 static int
10994 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10995 {
10996 
10997 	uint32_t am, l_end;
10998 	int was_tlp = 0;
10999 
11000 	if (SEQ_GT(end, start))
11001 		am = end - start;
11002 	else
11003 		am = 0;
11004 	if ((rack->rc_last_tlp_acked_set ) &&
11005 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
11006 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
11007 		/*
11008 		 * The DSACK is because of a TLP which we don't
11009 		 * do anything with the reordering window over since
11010 		 * it was not reordering that caused the DSACK but
11011 		 * our previous retransmit TLP.
11012 		 */
11013 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
11014 		was_tlp = 1;
11015 		goto skip_dsack_round;
11016 	}
11017 	if (rack->rc_last_sent_tlp_seq_valid) {
11018 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
11019 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
11020 		    (SEQ_LEQ(end, l_end))) {
11021 			/*
11022 			 * This dsack is from the last sent TLP, ignore it
11023 			 * for reordering purposes.
11024 			 */
11025 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
11026 			was_tlp = 1;
11027 			goto skip_dsack_round;
11028 		}
11029 	}
11030 	if (rack->rc_dsack_round_seen == 0) {
11031 		rack->rc_dsack_round_seen = 1;
11032 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
11033 		rack->r_ctl.num_dsack++;
11034 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
11035 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
11036 	}
11037 skip_dsack_round:
11038 	/*
11039 	 * We keep track of how many DSACK blocks we get
11040 	 * after a recovery incident.
11041 	 */
11042 	rack->r_ctl.dsack_byte_cnt += am;
11043 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
11044 	    rack->r_ctl.retran_during_recovery &&
11045 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
11046 		/*
11047 		 * False recovery most likely culprit is reordering. If
11048 		 * nothing else is missing we need to revert.
11049 		 */
11050 		rack->r_might_revert = 1;
11051 		rack_handle_might_revert(rack->rc_tp, rack);
11052 		rack->r_might_revert = 0;
11053 		rack->r_ctl.retran_during_recovery = 0;
11054 		rack->r_ctl.dsack_byte_cnt = 0;
11055 	}
11056 	return (was_tlp);
11057 }
11058 
11059 static uint32_t
11060 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
11061 {
11062 	return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
11063 }
11064 
11065 static int32_t
11066 rack_compute_pipe(struct tcpcb *tp)
11067 {
11068 	return ((int32_t)do_rack_compute_pipe(tp,
11069 					      (struct tcp_rack *)tp->t_fb_ptr,
11070 					      tp->snd_una));
11071 }
11072 
11073 static void
11074 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
11075 {
11076 	/* Deal with changed and PRR here (in recovery only) */
11077 	uint32_t pipe, snd_una;
11078 
11079 	rack->r_ctl.rc_prr_delivered += changed;
11080 
11081 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
11082 		/*
11083 		 * It is all outstanding, we are application limited
11084 		 * and thus we don't need more room to send anything.
11085 		 * Note we use tp->snd_una here and not th_ack because
11086 		 * the data as yet not been cut from the sb.
11087 		 */
11088 		rack->r_ctl.rc_prr_sndcnt = 0;
11089 		return;
11090 	}
11091 	/* Compute prr_sndcnt */
11092 	if (SEQ_GT(tp->snd_una, th_ack)) {
11093 		snd_una = tp->snd_una;
11094 	} else {
11095 		snd_una = th_ack;
11096 	}
11097 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
11098 	if (pipe > tp->snd_ssthresh) {
11099 		long sndcnt;
11100 
11101 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
11102 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
11103 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
11104 		else {
11105 			rack->r_ctl.rc_prr_sndcnt = 0;
11106 			rack_log_to_prr(rack, 9, 0, __LINE__);
11107 			sndcnt = 0;
11108 		}
11109 		sndcnt++;
11110 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
11111 			sndcnt -= rack->r_ctl.rc_prr_out;
11112 		else
11113 			sndcnt = 0;
11114 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
11115 		rack_log_to_prr(rack, 10, 0, __LINE__);
11116 	} else {
11117 		uint32_t limit;
11118 
11119 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
11120 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
11121 		else
11122 			limit = 0;
11123 		if (changed > limit)
11124 			limit = changed;
11125 		limit += ctf_fixed_maxseg(tp);
11126 		if (tp->snd_ssthresh > pipe) {
11127 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
11128 			rack_log_to_prr(rack, 11, 0, __LINE__);
11129 		} else {
11130 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
11131 			rack_log_to_prr(rack, 12, 0, __LINE__);
11132 		}
11133 	}
11134 }
11135 
11136 static void
11137 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11138 	     int *dsack_seen, int *sacks_seen)
11139 {
11140 	uint32_t changed;
11141 	struct tcp_rack *rack;
11142 	struct rack_sendmap *rsm;
11143 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11144 	register uint32_t th_ack;
11145 	int32_t i, j, k, num_sack_blks = 0;
11146 	uint32_t cts, acked, ack_point;
11147 	int loop_start = 0, moved_two = 0, no_extra = 0;
11148 	uint32_t tsused;
11149 	uint32_t segsiz, o_cnt;
11150 
11151 
11152 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11153 	if (tcp_get_flags(th) & TH_RST) {
11154 		/* We don't log resets */
11155 		return;
11156 	}
11157 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11158 	cts = tcp_get_usecs(NULL);
11159 	rsm = tqhash_min(rack->r_ctl.tqh);
11160 	changed = 0;
11161 	th_ack = th->th_ack;
11162 	if (rack->sack_attack_disable == 0)
11163 		rack_do_decay(rack);
11164 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
11165 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11166 		/*
11167 		 * You only get credit for
11168 		 * MSS and greater (and you get extra
11169 		 * credit for larger cum-ack moves).
11170 		 */
11171 		int ac;
11172 
11173 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11174 		rack->r_ctl.ack_count += ac;
11175 		counter_u64_add(rack_ack_total, ac);
11176 	}
11177 	if (rack->r_ctl.ack_count > 0xfff00000) {
11178 		/*
11179 		 * reduce the number to keep us under
11180 		 * a uint32_t.
11181 		 */
11182 		rack->r_ctl.ack_count /= 2;
11183 		rack->r_ctl.sack_count /= 2;
11184 	}
11185 	if (SEQ_GT(th_ack, tp->snd_una)) {
11186 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
11187 		tp->t_acktime = ticks;
11188 	}
11189 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
11190 		changed = th_ack - rsm->r_start;
11191 	if (changed) {
11192 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
11193 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
11194 	}
11195 	if ((to->to_flags & TOF_SACK) == 0) {
11196 		/* We are done nothing left and no sack. */
11197 		rack_handle_might_revert(tp, rack);
11198 		/*
11199 		 * For cases where we struck a dup-ack
11200 		 * with no SACK, add to the changes so
11201 		 * PRR will work right.
11202 		 */
11203 		if (dup_ack_struck && (changed == 0)) {
11204 			changed += ctf_fixed_maxseg(rack->rc_tp);
11205 		}
11206 		goto out;
11207 	}
11208 	/* Sack block processing */
11209 	if (SEQ_GT(th_ack, tp->snd_una))
11210 		ack_point = th_ack;
11211 	else
11212 		ack_point = tp->snd_una;
11213 	for (i = 0; i < to->to_nsacks; i++) {
11214 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
11215 		      &sack, sizeof(sack));
11216 		sack.start = ntohl(sack.start);
11217 		sack.end = ntohl(sack.end);
11218 		if (SEQ_GT(sack.end, sack.start) &&
11219 		    SEQ_GT(sack.start, ack_point) &&
11220 		    SEQ_LT(sack.start, tp->snd_max) &&
11221 		    SEQ_GT(sack.end, ack_point) &&
11222 		    SEQ_LEQ(sack.end, tp->snd_max)) {
11223 			sack_blocks[num_sack_blks] = sack;
11224 			num_sack_blks++;
11225 		} else if (SEQ_LEQ(sack.start, th_ack) &&
11226 			   SEQ_LEQ(sack.end, th_ack)) {
11227 			int was_tlp;
11228 
11229 			if (dsack_seen != NULL)
11230 				*dsack_seen = 1;
11231 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
11232 			/*
11233 			 * Its a D-SACK block.
11234 			 */
11235 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
11236 		}
11237 	}
11238 	if (rack->rc_dsack_round_seen) {
11239 		/* Is the dsack roound over? */
11240 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
11241 			/* Yes it is */
11242 			rack->rc_dsack_round_seen = 0;
11243 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
11244 		}
11245 	}
11246 	/*
11247 	 * Sort the SACK blocks so we can update the rack scoreboard with
11248 	 * just one pass.
11249 	 */
11250 	o_cnt = num_sack_blks;
11251 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
11252 					 num_sack_blks, th->th_ack);
11253 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
11254 	if (sacks_seen != NULL)
11255 		*sacks_seen = num_sack_blks;
11256 	if (num_sack_blks == 0) {
11257 		/* Nothing to sack, but we need to update counts */
11258 		if ((o_cnt == 1) &&
11259 		    (*dsack_seen != 1))
11260 			rack->r_ctl.sack_count++;
11261 		else if (o_cnt > 1)
11262 			rack->r_ctl.sack_count++;
11263 		goto out_with_totals;
11264 	}
11265 	if (rack->sack_attack_disable) {
11266 		/*
11267 		 * An attacker disablement is in place, for
11268 		 * every sack block that is not at least a full MSS
11269 		 * count up sack_count.
11270 		 */
11271 		for (i = 0; i < num_sack_blks; i++) {
11272 			if ((sack_blocks[i].end - sack_blocks[i].start) < segsiz) {
11273 				rack->r_ctl.sack_count++;
11274 			}
11275 			if (rack->r_ctl.sack_count > 0xfff00000) {
11276 				/*
11277 				 * reduce the number to keep us under
11278 				 * a uint32_t.
11279 				 */
11280 				rack->r_ctl.ack_count /= 2;
11281 				rack->r_ctl.sack_count /= 2;
11282 			}
11283 		}
11284 		goto out;
11285 	}
11286 	/* Its a sack of some sort */
11287 	rack->r_ctl.sack_count += num_sack_blks;
11288 	if (rack->r_ctl.sack_count > 0xfff00000) {
11289 		/*
11290 		 * reduce the number to keep us under
11291 		 * a uint32_t.
11292 		 */
11293 		rack->r_ctl.ack_count /= 2;
11294 		rack->r_ctl.sack_count /= 2;
11295 	}
11296 	if (num_sack_blks < 2) {
11297 		/* Only one, we don't need to sort */
11298 		goto do_sack_work;
11299 	}
11300 	/* Sort the sacks */
11301 	for (i = 0; i < num_sack_blks; i++) {
11302 		for (j = i + 1; j < num_sack_blks; j++) {
11303 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
11304 				sack = sack_blocks[i];
11305 				sack_blocks[i] = sack_blocks[j];
11306 				sack_blocks[j] = sack;
11307 			}
11308 		}
11309 	}
11310 	/*
11311 	 * Now are any of the sack block ends the same (yes some
11312 	 * implementations send these)?
11313 	 */
11314 again:
11315 	if (num_sack_blks == 0)
11316 		goto out_with_totals;
11317 	if (num_sack_blks > 1) {
11318 		for (i = 0; i < num_sack_blks; i++) {
11319 			for (j = i + 1; j < num_sack_blks; j++) {
11320 				if (sack_blocks[i].end == sack_blocks[j].end) {
11321 					/*
11322 					 * Ok these two have the same end we
11323 					 * want the smallest end and then
11324 					 * throw away the larger and start
11325 					 * again.
11326 					 */
11327 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
11328 						/*
11329 						 * The second block covers
11330 						 * more area use that
11331 						 */
11332 						sack_blocks[i].start = sack_blocks[j].start;
11333 					}
11334 					/*
11335 					 * Now collapse out the dup-sack and
11336 					 * lower the count
11337 					 */
11338 					for (k = (j + 1); k < num_sack_blks; k++) {
11339 						sack_blocks[j].start = sack_blocks[k].start;
11340 						sack_blocks[j].end = sack_blocks[k].end;
11341 						j++;
11342 					}
11343 					num_sack_blks--;
11344 					goto again;
11345 				}
11346 			}
11347 		}
11348 	}
11349 do_sack_work:
11350 	/*
11351 	 * First lets look to see if
11352 	 * we have retransmitted and
11353 	 * can use the transmit next?
11354 	 */
11355 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11356 	if (rsm &&
11357 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
11358 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
11359 		/*
11360 		 * We probably did the FR and the next
11361 		 * SACK in continues as we would expect.
11362 		 */
11363 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11364 		if (acked) {
11365 			rack->r_wanted_output = 1;
11366 			changed += acked;
11367 		}
11368 		if (num_sack_blks == 1) {
11369 			/*
11370 			 * This is what we would expect from
11371 			 * a normal implementation to happen
11372 			 * after we have retransmitted the FR,
11373 			 * i.e the sack-filter pushes down
11374 			 * to 1 block and the next to be retransmitted
11375 			 * is the sequence in the sack block (has more
11376 			 * are acked). Count this as ACK'd data to boost
11377 			 * up the chances of recovering any false positives.
11378 			 */
11379 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
11380 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
11381 			counter_u64_add(rack_express_sack, 1);
11382 			if (rack->r_ctl.ack_count > 0xfff00000) {
11383 				/*
11384 				 * reduce the number to keep us under
11385 				 * a uint32_t.
11386 				 */
11387 				rack->r_ctl.ack_count /= 2;
11388 				rack->r_ctl.sack_count /= 2;
11389 			}
11390 			if (moved_two) {
11391 				/*
11392 				 * If we did not get a SACK for at least a MSS and
11393 				 * had to move at all, or if we moved more than our
11394 				 * threshold, it counts against the "extra" move.
11395 				 */
11396 				rack->r_ctl.sack_moved_extra += moved_two;
11397 				rack->r_ctl.sack_noextra_move += no_extra;
11398 				counter_u64_add(rack_move_some, 1);
11399 			} else {
11400 				/*
11401 				 * else we did not have to move
11402 				 * any more than we would expect.
11403 				 */
11404 				rack->r_ctl.sack_noextra_move += no_extra;
11405 				rack->r_ctl.sack_noextra_move++;
11406 				counter_u64_add(rack_move_none, 1);
11407 			}
11408 			if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11409 			    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11410 				rack->r_ctl.sack_moved_extra /= 2;
11411 				rack->r_ctl.sack_noextra_move /= 2;
11412 			}
11413 			goto out_with_totals;
11414 		} else {
11415 			/*
11416 			 * Start the loop through the
11417 			 * rest of blocks, past the first block.
11418 			 */
11419 			loop_start = 1;
11420 		}
11421 	}
11422 	counter_u64_add(rack_sack_total, 1);
11423 	rsm = rack->r_ctl.rc_sacklast;
11424 	for (i = loop_start; i < num_sack_blks; i++) {
11425 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11426 		if (acked) {
11427 			rack->r_wanted_output = 1;
11428 			changed += acked;
11429 		}
11430 		if (moved_two) {
11431 			/*
11432 			 * If we did not get a SACK for at least a MSS and
11433 			 * had to move at all, or if we moved more than our
11434 			 * threshold, it counts against the "extra" move.
11435 			 */
11436 			rack->r_ctl.sack_moved_extra += moved_two;
11437 			rack->r_ctl.sack_noextra_move += no_extra;
11438 			counter_u64_add(rack_move_some, 1);
11439 		} else {
11440 			/*
11441 			 * else we did not have to move
11442 			 * any more than we would expect.
11443 			 */
11444 			rack->r_ctl.sack_noextra_move += no_extra;
11445 			rack->r_ctl.sack_noextra_move++;
11446 			counter_u64_add(rack_move_none, 1);
11447 		}
11448 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11449 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11450 			rack->r_ctl.sack_moved_extra /= 2;
11451 			rack->r_ctl.sack_noextra_move /= 2;
11452 		}
11453 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
11454 			/*
11455 			 * If the SACK was not a full MSS then
11456 			 * we add to sack_count the number of
11457 			 * MSS's (or possibly more than
11458 			 * a MSS if its a TSO send) we had to skip by.
11459 			 */
11460 			rack->r_ctl.sack_count += moved_two;
11461 			if (rack->r_ctl.sack_count > 0xfff00000) {
11462 				rack->r_ctl.ack_count /= 2;
11463 				rack->r_ctl.sack_count /= 2;
11464 			}
11465 			counter_u64_add(rack_sack_total, moved_two);
11466 		}
11467 		/*
11468 		 * Now we need to setup for the next
11469 		 * round. First we make sure we won't
11470 		 * exceed the size of our uint32_t on
11471 		 * the various counts, and then clear out
11472 		 * moved_two.
11473 		 */
11474 		moved_two = 0;
11475 		no_extra = 0;
11476 	}
11477 out_with_totals:
11478 	if (num_sack_blks > 1) {
11479 		/*
11480 		 * You get an extra stroke if
11481 		 * you have more than one sack-blk, this
11482 		 * could be where we are skipping forward
11483 		 * and the sack-filter is still working, or
11484 		 * it could be an attacker constantly
11485 		 * moving us.
11486 		 */
11487 		rack->r_ctl.sack_moved_extra++;
11488 		counter_u64_add(rack_move_some, 1);
11489 	}
11490 out:
11491 #ifdef TCP_SAD_DETECTION
11492 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
11493 #endif
11494 	if (changed) {
11495 		/* Something changed cancel the rack timer */
11496 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11497 	}
11498 	tsused = tcp_get_usecs(NULL);
11499 	rsm = tcp_rack_output(tp, rack, tsused);
11500 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11501 	    rsm &&
11502 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11503 		/* Enter recovery */
11504 		entered_recovery = 1;
11505 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
11506 		/*
11507 		 * When we enter recovery we need to assure we send
11508 		 * one packet.
11509 		 */
11510 		if (rack->rack_no_prr == 0) {
11511 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11512 			rack_log_to_prr(rack, 8, 0, __LINE__);
11513 		}
11514 		rack->r_timer_override = 1;
11515 		rack->r_early = 0;
11516 		rack->r_ctl.rc_agg_early = 0;
11517 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11518 		   rsm &&
11519 		   (rack->r_rr_config == 3)) {
11520 		/*
11521 		 * Assure we can output and we get no
11522 		 * remembered pace time except the retransmit.
11523 		 */
11524 		rack->r_timer_override = 1;
11525 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11526 		rack->r_ctl.rc_resend = rsm;
11527 	}
11528 	if (IN_FASTRECOVERY(tp->t_flags) &&
11529 	    (rack->rack_no_prr == 0) &&
11530 	    (entered_recovery == 0)) {
11531 		rack_update_prr(tp, rack, changed, th_ack);
11532 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11533 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11534 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11535 			/*
11536 			 * If you are pacing output you don't want
11537 			 * to override.
11538 			 */
11539 			rack->r_early = 0;
11540 			rack->r_ctl.rc_agg_early = 0;
11541 			rack->r_timer_override = 1;
11542 		}
11543 	}
11544 }
11545 
11546 static void
11547 rack_strike_dupack(struct tcp_rack *rack)
11548 {
11549 	struct rack_sendmap *rsm;
11550 
11551 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11552 	while (rsm) {
11553 		/*
11554 		 * We need to skip anything already set
11555 		 * to be retransmitted.
11556 		 */
11557 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11558 		    (rsm->r_flags & RACK_MUST_RXT)) {
11559 			rsm = TAILQ_NEXT(rsm, r_tnext);
11560 			continue;
11561 		}
11562 		break;
11563 	}
11564 	if (rsm && (rsm->r_dupack < 0xff)) {
11565 		rsm->r_dupack++;
11566 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11567 			struct timeval tv;
11568 			uint32_t cts;
11569 			/*
11570 			 * Here we see if we need to retransmit. For
11571 			 * a SACK type connection if enough time has passed
11572 			 * we will get a return of the rsm. For a non-sack
11573 			 * connection we will get the rsm returned if the
11574 			 * dupack value is 3 or more.
11575 			 */
11576 			cts = tcp_get_usecs(&tv);
11577 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11578 			if (rack->r_ctl.rc_resend != NULL) {
11579 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11580 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11581 							 rack->rc_tp->snd_una, __LINE__);
11582 				}
11583 				rack->r_wanted_output = 1;
11584 				rack->r_timer_override = 1;
11585 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11586 			}
11587 		} else {
11588 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11589 		}
11590 	}
11591 }
11592 
11593 static void
11594 rack_check_bottom_drag(struct tcpcb *tp,
11595 		       struct tcp_rack *rack,
11596 		       struct socket *so)
11597 {
11598 	uint32_t segsiz, minseg;
11599 
11600 	segsiz = ctf_fixed_maxseg(tp);
11601 	minseg = segsiz;
11602 	if (tp->snd_max == tp->snd_una) {
11603 		/*
11604 		 * We are doing dynamic pacing and we are way
11605 		 * under. Basically everything got acked while
11606 		 * we were still waiting on the pacer to expire.
11607 		 *
11608 		 * This means we need to boost the b/w in
11609 		 * addition to any earlier boosting of
11610 		 * the multiplier.
11611 		 */
11612 		uint64_t lt_bw;
11613 
11614 		lt_bw = rack_get_lt_bw(rack);
11615 		rack->rc_dragged_bottom = 1;
11616 		rack_validate_multipliers_at_or_above100(rack);
11617 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11618 		    (lt_bw > 0)) {
11619 			/*
11620 			 * Lets use the long-term b/w we have
11621 			 * been getting as a base.
11622 			 */
11623 			if (rack->rc_gp_filled == 0) {
11624 				if (lt_bw > ONE_POINT_TWO_MEG) {
11625 					/*
11626 					 * If we have no measurement
11627 					 * don't let us set in more than
11628 					 * 1.2Mbps. If we are still too
11629 					 * low after pacing with this we
11630 					 * will hopefully have a max b/w
11631 					 * available to sanity check things.
11632 					 */
11633 					lt_bw = ONE_POINT_TWO_MEG;
11634 				}
11635 				rack->r_ctl.rc_rtt_diff = 0;
11636 				rack->r_ctl.gp_bw = lt_bw;
11637 				rack->rc_gp_filled = 1;
11638 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11639 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11640 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11641 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11642 				rack->r_ctl.rc_rtt_diff = 0;
11643 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11644 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11645 				rack->r_ctl.gp_bw = lt_bw;
11646 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11647 			} else
11648 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11649 			if ((rack->gp_ready == 0) &&
11650 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11651 				/* We have enough measurements now */
11652 				rack->gp_ready = 1;
11653 				if (rack->dgp_on ||
11654 				    rack->rack_hibeta)
11655 					rack_set_cc_pacing(rack);
11656 				if (rack->defer_options)
11657 					rack_apply_deferred_options(rack);
11658 			}
11659 		} else {
11660 			/*
11661 			 * zero rtt possibly?, settle for just an old increase.
11662 			 */
11663 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11664 		}
11665 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11666 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11667 					       minseg)) &&
11668 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11669 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11670 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11671 		    (segsiz * rack_req_segs))) {
11672 		/*
11673 		 * We are doing dynamic GP pacing and
11674 		 * we have everything except 1MSS or less
11675 		 * bytes left out. We are still pacing away.
11676 		 * And there is data that could be sent, This
11677 		 * means we are inserting delayed ack time in
11678 		 * our measurements because we are pacing too slow.
11679 		 */
11680 		rack_validate_multipliers_at_or_above100(rack);
11681 		rack->rc_dragged_bottom = 1;
11682 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11683 	}
11684 }
11685 
11686 #ifdef TCP_REQUEST_TRK
11687 static void
11688 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11689 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11690 {
11691 	int do_log;
11692 
11693 	do_log = tcp_bblogging_on(rack->rc_tp);
11694 	if (do_log == 0) {
11695 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11696 			return;
11697 		/* We only allow the three below with point logging on */
11698 		if ((mod != HYBRID_LOG_RULES_APP) &&
11699 		    (mod != HYBRID_LOG_RULES_SET) &&
11700 		    (mod != HYBRID_LOG_REQ_COMP))
11701 			return;
11702 
11703 	}
11704 	if (do_log) {
11705 		union tcp_log_stackspecific log;
11706 		struct timeval tv;
11707 
11708 		/* Convert our ms to a microsecond */
11709 		memset(&log, 0, sizeof(log));
11710 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11711 		log.u_bbr.flex1 = seq;
11712 		log.u_bbr.cwnd_gain = line;
11713 		if (cur != NULL) {
11714 			uint64_t off;
11715 
11716 			log.u_bbr.flex2 = cur->start_seq;
11717 			log.u_bbr.flex3 = cur->end_seq;
11718 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11719 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11720 			log.u_bbr.flex6 = cur->flags;
11721 			log.u_bbr.pkts_out = cur->hybrid_flags;
11722 			log.u_bbr.rttProp = cur->timestamp;
11723 			log.u_bbr.cur_del_rate = cur->cspr;
11724 			log.u_bbr.bw_inuse = cur->start;
11725 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11726 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11727 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11728 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11729 			log.u_bbr.bbr_state = 1;
11730 #ifdef TCP_REQUEST_TRK
11731 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11732 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11733 #endif
11734 		} else {
11735 			log.u_bbr.flex2 = err;
11736 		}
11737 		/*
11738 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11739 		 */
11740 		log.u_bbr.flex7 = rack->rc_catch_up;
11741 		log.u_bbr.flex7 <<= 1;
11742 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11743 		log.u_bbr.flex7 <<= 1;
11744 		log.u_bbr.flex7 |= rack->dgp_on;
11745 		log.u_bbr.flex8 = mod;
11746 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11747 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11748 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11749 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11750 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11751 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11752 		tcp_log_event(rack->rc_tp, NULL,
11753 		    &rack->rc_inp->inp_socket->so_rcv,
11754 		    &rack->rc_inp->inp_socket->so_snd,
11755 		    TCP_HYBRID_PACING_LOG, 0,
11756 	            0, &log, false, NULL, __func__, __LINE__, &tv);
11757 	}
11758 }
11759 #endif
11760 
11761 #ifdef TCP_REQUEST_TRK
11762 static void
11763 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len)
11764 {
11765 	struct tcp_sendfile_track *rc_cur;
11766 	struct tcpcb *tp;
11767 	int err = 0;
11768 
11769 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11770 	if (rc_cur == NULL) {
11771 		/* If not in the beginning what about the end piece */
11772 		if (rack->rc_hybrid_mode)
11773 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11774 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11775 	} else {
11776 		err = 12345;
11777 	}
11778 	/* If we find no parameters we are in straight DGP mode */
11779 	if(rc_cur == NULL) {
11780 		/* None found for this seq, just DGP for now */
11781 		rack->r_ctl.client_suggested_maxseg = 0;
11782 		rack->rc_catch_up = 0;
11783 		rack->r_ctl.bw_rate_cap = 0;
11784 		if (rack->rc_hybrid_mode)
11785 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11786 		if (rack->r_ctl.rc_last_sft) {
11787 			rack->r_ctl.rc_last_sft = NULL;
11788 		}
11789 		return;
11790 	}
11791 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11792 		/* This entry was never setup for hybrid pacing on/off etc */
11793 		return;
11794 	}
11795 	/*
11796 	 * Ok if we have a new entry *or* have never
11797 	 * set up an entry we need to proceed. If
11798 	 * we have already set it up this entry we
11799 	 * just continue along with what we already
11800 	 * setup.
11801 	 */
11802 	tp = rack->rc_tp;
11803 	if ((rack->r_ctl.rc_last_sft != NULL) &&
11804 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
11805 		/* Its already in place */
11806 		if (rack->rc_hybrid_mode)
11807 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11808 		return;
11809 	}
11810 	if (rack->rc_hybrid_mode == 0) {
11811 		rack->r_ctl.rc_last_sft = rc_cur;
11812 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11813 		return;
11814 	}
11815 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11816 		/* Compensate for all the header overhead's */
11817 		rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
11818 	} else
11819 		rack->r_ctl.bw_rate_cap = 0;
11820 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11821 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11822 	else
11823 		rack->r_ctl.client_suggested_maxseg = 0;
11824 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11825 	    (rc_cur->cspr > 0)) {
11826 		uint64_t len;
11827 
11828 		rack->rc_catch_up = 1;
11829 		/*
11830 		 * Calculate the deadline time, first set the
11831 		 * time to when the request arrived.
11832 		 */
11833 		rc_cur->deadline = rc_cur->localtime;
11834 		/*
11835 		 * Next calculate the length and compensate for
11836 		 * TLS if need be.
11837 		 */
11838 		len = rc_cur->end - rc_cur->start;
11839 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11840 			/*
11841 			 * This session is doing TLS. Take a swag guess
11842 			 * at the overhead.
11843 			 */
11844 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11845 		}
11846 		/*
11847 		 * Now considering the size, and the cspr, what is the time that
11848 		 * would be required at the cspr rate. Here we use the raw
11849 		 * cspr value since the client only looks at the raw data. We
11850 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
11851 		 * That will get made up for in the CU pacing rate set.
11852 		 */
11853 		len *= HPTS_USEC_IN_SEC;
11854 		len /= rc_cur->cspr;
11855 		rc_cur->deadline += len;
11856 	} else {
11857 		rack->rc_catch_up = 0;
11858 		rc_cur->deadline = 0;
11859 	}
11860 	if (rack->r_ctl.client_suggested_maxseg != 0) {
11861 		/*
11862 		 * We need to reset the max pace segs if we have a
11863 		 * client_suggested_maxseg.
11864 		 */
11865 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
11866 	}
11867 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11868 	/* Remember it for next time and for CU mode */
11869 	rack->r_ctl.rc_last_sft = rc_cur;
11870 }
11871 #endif
11872 
11873 static void
11874 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11875 {
11876 #ifdef TCP_REQUEST_TRK
11877 	struct tcp_sendfile_track *ent;
11878 
11879 	ent = rack->r_ctl.rc_last_sft;
11880 	if ((ent == NULL) ||
11881 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11882 	    (SEQ_GEQ(seq, ent->end_seq))) {
11883 		/* Time to update the track. */
11884 		rack_set_dgp_hybrid_mode(rack, seq, len);
11885 		ent = rack->r_ctl.rc_last_sft;
11886 	}
11887 	/* Out of all */
11888 	if (ent == NULL) {
11889 		return;
11890 	}
11891 	if (SEQ_LT(ent->end_seq, (seq + len))) {
11892 		/*
11893 		 * This is the case where our end_seq guess
11894 		 * was wrong. This is usually due to TLS having
11895 		 * more bytes then our guess. It could also be the
11896 		 * case that the client sent in two requests closely
11897 		 * and the SB is full of both so we are sending part
11898 		 * of each (end|beg). In such a case lets move this
11899 		 * guys end to match the end of this send. That
11900 		 * way it will complete when all of it is acked.
11901 		 */
11902 		ent->end_seq = (seq + len);
11903 		if (rack->rc_hybrid_mode)
11904 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11905 	}
11906 	/* Now validate we have set the send time of this one */
11907 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11908 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11909 		ent->first_send = cts;
11910 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11911 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11912 	}
11913 #endif
11914 }
11915 
11916 static void
11917 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11918 {
11919 	/*
11920 	 * The fast output path is enabled and we
11921 	 * have moved the cumack forward. Lets see if
11922 	 * we can expand forward the fast path length by
11923 	 * that amount. What we would ideally like to
11924 	 * do is increase the number of bytes in the
11925 	 * fast path block (left_to_send) by the
11926 	 * acked amount. However we have to gate that
11927 	 * by two factors:
11928 	 * 1) The amount outstanding and the rwnd of the peer
11929 	 *    (i.e. we don't want to exceed the rwnd of the peer).
11930 	 *    <and>
11931 	 * 2) The amount of data left in the socket buffer (i.e.
11932 	 *    we can't send beyond what is in the buffer).
11933 	 *
11934 	 * Note that this does not take into account any increase
11935 	 * in the cwnd. We will only extend the fast path by
11936 	 * what was acked.
11937 	 */
11938 	uint32_t new_total, gating_val;
11939 
11940 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11941 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11942 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11943 	if (new_total <= gating_val) {
11944 		/* We can increase left_to_send by the acked amount */
11945 		counter_u64_add(rack_extended_rfo, 1);
11946 		rack->r_ctl.fsb.left_to_send = new_total;
11947 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11948 			("rack:%p left_to_send:%u sbavail:%u out:%u",
11949 			 rack, rack->r_ctl.fsb.left_to_send,
11950 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
11951 			 (tp->snd_max - tp->snd_una)));
11952 
11953 	}
11954 }
11955 
11956 static void
11957 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11958 {
11959 	/*
11960 	 * Here any sendmap entry that points to the
11961 	 * beginning mbuf must be adjusted to the correct
11962 	 * offset. This must be called with:
11963 	 * 1) The socket buffer locked
11964 	 * 2) snd_una adjusted to its new position.
11965 	 *
11966 	 * Note that (2) implies rack_ack_received has also
11967 	 * been called and all the sbcut's have been done.
11968 	 *
11969 	 * We grab the first mbuf in the socket buffer and
11970 	 * then go through the front of the sendmap, recalculating
11971 	 * the stored offset for any sendmap entry that has
11972 	 * that mbuf. We must use the sb functions to do this
11973 	 * since its possible an add was done has well as
11974 	 * the subtraction we may have just completed. This should
11975 	 * not be a penalty though, since we just referenced the sb
11976 	 * to go in and trim off the mbufs that we freed (of course
11977 	 * there will be a penalty for the sendmap references though).
11978 	 *
11979 	 * Note also with INVARIANT on, we validate with a KASSERT
11980 	 * that the first sendmap entry has a soff of 0.
11981 	 *
11982 	 */
11983 	struct mbuf *m;
11984 	struct rack_sendmap *rsm;
11985 	tcp_seq snd_una;
11986 #ifdef INVARIANTS
11987 	int first_processed = 0;
11988 #endif
11989 
11990 	snd_una = rack->rc_tp->snd_una;
11991 	SOCKBUF_LOCK_ASSERT(sb);
11992 	m = sb->sb_mb;
11993 	rsm = tqhash_min(rack->r_ctl.tqh);
11994 	if ((rsm == NULL) || (m == NULL)) {
11995 		/* Nothing outstanding */
11996 		return;
11997 	}
11998 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
11999 	KASSERT((rsm->m == m),
12000 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
12001 		 rack, sb, rsm));
12002 	while (rsm->m && (rsm->m == m)) {
12003 		/* one to adjust */
12004 #ifdef INVARIANTS
12005 		struct mbuf *tm;
12006 		uint32_t soff;
12007 
12008 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
12009 		if ((rsm->orig_m_len != m->m_len) ||
12010 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
12011 			rack_adjust_orig_mlen(rsm);
12012 		}
12013 		if (first_processed == 0) {
12014 			KASSERT((rsm->soff == 0),
12015 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
12016 				 rack, rsm));
12017 			first_processed = 1;
12018 		}
12019 		if ((rsm->soff != soff) || (rsm->m != tm)) {
12020 			/*
12021 			 * This is not a fatal error, we anticipate it
12022 			 * might happen (the else code), so we count it here
12023 			 * so that under invariant we can see that it really
12024 			 * does happen.
12025 			 */
12026 			counter_u64_add(rack_adjust_map_bw, 1);
12027 		}
12028 		rsm->m = tm;
12029 		rsm->soff = soff;
12030 		if (tm) {
12031 			rsm->orig_m_len = rsm->m->m_len;
12032 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12033 		} else {
12034 			rsm->orig_m_len = 0;
12035 			rsm->orig_t_space = 0;
12036 		}
12037 #else
12038 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
12039 		if (rsm->m) {
12040 			rsm->orig_m_len = rsm->m->m_len;
12041 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12042 		} else {
12043 			rsm->orig_m_len = 0;
12044 			rsm->orig_t_space = 0;
12045 		}
12046 #endif
12047 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
12048 		if (rsm == NULL)
12049 			break;
12050 	}
12051 }
12052 
12053 #ifdef TCP_REQUEST_TRK
12054 static inline void
12055 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
12056 {
12057 	struct tcp_sendfile_track *ent;
12058 	int i;
12059 
12060 	if ((rack->rc_hybrid_mode == 0) &&
12061 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
12062 		/*
12063 		 * Just do normal completions hybrid pacing is not on
12064 		 * and CLDL is off as well.
12065 		 */
12066 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
12067 		return;
12068 	}
12069 	/*
12070 	 * Originally I was just going to find the th_ack associated
12071 	 * with an entry. But then I realized a large strech ack could
12072 	 * in theory ack two or more requests at once. So instead we
12073 	 * need to find all entries that are completed by th_ack not
12074 	 * just a single entry and do our logging.
12075 	 */
12076 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12077 	while (ent != NULL) {
12078 		/*
12079 		 * We may be doing hybrid pacing or CLDL and need more details possibly
12080 		 * so we do it manually instead of calling
12081 		 * tcp_req_check_for_comp()
12082 		 */
12083 		uint64_t laa, tim, data, cbw, ftim;
12084 
12085 		/* Ok this ack frees it */
12086 		rack_log_hybrid(rack, th_ack,
12087 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
12088 		rack_log_hybrid_sends(rack, ent, __LINE__);
12089 		/* calculate the time based on the ack arrival */
12090 		data = ent->end - ent->start;
12091 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
12092 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
12093 			if (ent->first_send > ent->localtime)
12094 				ftim = ent->first_send;
12095 			else
12096 				ftim = ent->localtime;
12097 		} else {
12098 			/* TSNH */
12099 			ftim = ent->localtime;
12100 		}
12101 		if (laa > ent->localtime)
12102 			tim = laa - ftim;
12103 		else
12104 			tim = 0;
12105 		cbw = data * HPTS_USEC_IN_SEC;
12106 		if (tim > 0)
12107 			cbw /= tim;
12108 		else
12109 			cbw = 0;
12110 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
12111 		/*
12112 		 * Check to see if we are freeing what we are pointing to send wise
12113 		 * if so be sure to NULL the pointer so we know we are no longer
12114 		 * set to anything.
12115 		 */
12116 		if (ent == rack->r_ctl.rc_last_sft)
12117 			rack->r_ctl.rc_last_sft = NULL;
12118 		/* Generate the log that the tcp_netflix call would have */
12119 		tcp_req_log_req_info(rack->rc_tp, ent,
12120 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
12121 		/* Free it and see if there is another one */
12122 		tcp_req_free_a_slot(rack->rc_tp, ent);
12123 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12124 	}
12125 }
12126 #endif
12127 
12128 
12129 /*
12130  * Return value of 1, we do not need to call rack_process_data().
12131  * return value of 0, rack_process_data can be called.
12132  * For ret_val if its 0 the TCP is locked, if its non-zero
12133  * its unlocked and probably unsafe to touch the TCB.
12134  */
12135 static int
12136 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12137     struct tcpcb *tp, struct tcpopt *to,
12138     uint32_t tiwin, int32_t tlen,
12139     int32_t * ofia, int32_t thflags, int32_t *ret_val)
12140 {
12141 	int32_t ourfinisacked = 0;
12142 	int32_t nsegs, acked_amount;
12143 	int32_t acked;
12144 	struct mbuf *mfree;
12145 	struct tcp_rack *rack;
12146 	int32_t under_pacing = 0;
12147 	int32_t recovery = 0;
12148 
12149 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12150 
12151 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12152 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
12153 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
12154 				      &rack->r_ctl.challenge_ack_ts,
12155 				      &rack->r_ctl.challenge_ack_cnt);
12156 		rack->r_wanted_output = 1;
12157 		return (1);
12158 	}
12159 	if (rack->gp_ready &&
12160 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12161 		under_pacing = 1;
12162 	}
12163 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
12164 		int in_rec, dup_ack_struck = 0;
12165 		int dsack_seen = 0, sacks_seen = 0;
12166 
12167 		in_rec = IN_FASTRECOVERY(tp->t_flags);
12168 		if (rack->rc_in_persist) {
12169 			tp->t_rxtshift = 0;
12170 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12171 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12172 		}
12173 
12174 		if ((th->th_ack == tp->snd_una) &&
12175 		    (tiwin == tp->snd_wnd) &&
12176 		    ((to->to_flags & TOF_SACK) == 0)) {
12177 			rack_strike_dupack(rack);
12178 			dup_ack_struck = 1;
12179 		}
12180 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
12181 			     dup_ack_struck, &dsack_seen, &sacks_seen);
12182 		if ((rack->sack_attack_disable > 0) &&
12183 		    (th->th_ack == tp->snd_una) &&
12184 		    (tiwin == tp->snd_wnd) &&
12185 		    (dsack_seen == 0) &&
12186 		    (sacks_seen > 0)) {
12187 			/*
12188 			 * If sacks have been disabled we may
12189 			 * want to strike a dup-ack "ignoring" the
12190 			 * sack as long as the sack was not a "dsack". Note
12191 			 * that if no sack is sent (TOF_SACK is off) then the
12192 			 * normal dsack code above rack_log_ack() would have
12193 			 * already struck. So this is just to catch the case
12194 			 * were we are ignoring sacks from this guy due to
12195 			 * it being a suspected attacker.
12196 			 */
12197 			rack_strike_dupack(rack);
12198 		}
12199 
12200 	}
12201 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12202 		/*
12203 		 * Old ack, behind (or duplicate to) the last one rcv'd
12204 		 * Note: We mark reordering is occuring if its
12205 		 * less than and we have not closed our window.
12206 		 */
12207 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
12208 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12209 			if (rack->r_ctl.rc_reorder_ts == 0)
12210 				rack->r_ctl.rc_reorder_ts = 1;
12211 		}
12212 		return (0);
12213 	}
12214 	/*
12215 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
12216 	 * something we sent.
12217 	 */
12218 	if (tp->t_flags & TF_NEEDSYN) {
12219 		/*
12220 		 * T/TCP: Connection was half-synchronized, and our SYN has
12221 		 * been ACK'd (so connection is now fully synchronized).  Go
12222 		 * to non-starred state, increment snd_una for ACK of SYN,
12223 		 * and check if we can do window scaling.
12224 		 */
12225 		tp->t_flags &= ~TF_NEEDSYN;
12226 		tp->snd_una++;
12227 		/* Do window scaling? */
12228 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12229 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12230 			tp->rcv_scale = tp->request_r_scale;
12231 			/* Send window already scaled. */
12232 		}
12233 	}
12234 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12235 
12236 	acked = BYTES_THIS_ACK(tp, th);
12237 	if (acked) {
12238 		/*
12239 		 * Any time we move the cum-ack forward clear
12240 		 * keep-alive tied probe-not-answered. The
12241 		 * persists clears its own on entry.
12242 		 */
12243 		rack->probe_not_answered = 0;
12244 	}
12245 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12246 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12247 	/*
12248 	 * If we just performed our first retransmit, and the ACK arrives
12249 	 * within our recovery window, then it was a mistake to do the
12250 	 * retransmit in the first place.  Recover our original cwnd and
12251 	 * ssthresh, and proceed to transmit where we left off.
12252 	 */
12253 	if ((tp->t_flags & TF_PREVVALID) &&
12254 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12255 		tp->t_flags &= ~TF_PREVVALID;
12256 		if (tp->t_rxtshift == 1 &&
12257 		    (int)(ticks - tp->t_badrxtwin) < 0)
12258 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12259 	}
12260 	if (acked) {
12261 		/* assure we are not backed off */
12262 		tp->t_rxtshift = 0;
12263 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12264 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12265 		rack->rc_tlp_in_progress = 0;
12266 		rack->r_ctl.rc_tlp_cnt_out = 0;
12267 		/*
12268 		 * If it is the RXT timer we want to
12269 		 * stop it, so we can restart a TLP.
12270 		 */
12271 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12272 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12273 #ifdef TCP_REQUEST_TRK
12274 		rack_req_check_for_comp(rack, th->th_ack);
12275 #endif
12276 	}
12277 	/*
12278 	 * If we have a timestamp reply, update smoothed round trip time. If
12279 	 * no timestamp is present but transmit timer is running and timed
12280 	 * sequence number was acked, update smoothed round trip time. Since
12281 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
12282 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
12283 	 * timer.
12284 	 *
12285 	 * Some boxes send broken timestamp replies during the SYN+ACK
12286 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12287 	 * and blow up the retransmit timer.
12288 	 */
12289 	/*
12290 	 * If all outstanding data is acked, stop retransmit timer and
12291 	 * remember to restart (more output or persist). If there is more
12292 	 * data to be acked, restart retransmit timer, using current
12293 	 * (possibly backed-off) value.
12294 	 */
12295 	if (acked == 0) {
12296 		if (ofia)
12297 			*ofia = ourfinisacked;
12298 		return (0);
12299 	}
12300 	if (IN_RECOVERY(tp->t_flags)) {
12301 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
12302 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
12303 			tcp_rack_partialack(tp);
12304 		} else {
12305 			rack_post_recovery(tp, th->th_ack);
12306 			recovery = 1;
12307 		}
12308 	}
12309 	/*
12310 	 * Let the congestion control algorithm update congestion control
12311 	 * related information. This typically means increasing the
12312 	 * congestion window.
12313 	 */
12314 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
12315 	SOCKBUF_LOCK(&so->so_snd);
12316 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
12317 	tp->snd_wnd -= acked_amount;
12318 	mfree = sbcut_locked(&so->so_snd, acked_amount);
12319 	if ((sbused(&so->so_snd) == 0) &&
12320 	    (acked > acked_amount) &&
12321 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
12322 	    (tp->t_flags & TF_SENTFIN)) {
12323 		/*
12324 		 * We must be sure our fin
12325 		 * was sent and acked (we can be
12326 		 * in FIN_WAIT_1 without having
12327 		 * sent the fin).
12328 		 */
12329 		ourfinisacked = 1;
12330 	}
12331 	tp->snd_una = th->th_ack;
12332 	/* wakeups? */
12333 	if (acked_amount && sbavail(&so->so_snd))
12334 		rack_adjust_sendmap_head(rack, &so->so_snd);
12335 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12336 	/* NB: sowwakeup_locked() does an implicit unlock. */
12337 	sowwakeup_locked(so);
12338 	/* now check the rxt clamps */
12339 	if ((recovery == 1) &&
12340 	    (rack->excess_rxt_on) &&
12341 	    (rack->r_cwnd_was_clamped == 0))  {
12342 		do_rack_excess_rxt(tp, rack);
12343 	} else if (rack->r_cwnd_was_clamped)
12344 		do_rack_check_for_unclamp(tp, rack);
12345 	m_freem(mfree);
12346 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
12347 		tp->snd_recover = tp->snd_una;
12348 
12349 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12350 		tp->snd_nxt = tp->snd_max;
12351 	}
12352 	if (under_pacing &&
12353 	    (rack->use_fixed_rate == 0) &&
12354 	    (rack->in_probe_rtt == 0) &&
12355 	    rack->rc_gp_dyn_mul &&
12356 	    rack->rc_always_pace) {
12357 		/* Check if we are dragging bottom */
12358 		rack_check_bottom_drag(tp, rack, so);
12359 	}
12360 	if (tp->snd_una == tp->snd_max) {
12361 		/* Nothing left outstanding */
12362 		tp->t_flags &= ~TF_PREVVALID;
12363 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12364 		rack->r_ctl.retran_during_recovery = 0;
12365 		rack->r_ctl.dsack_byte_cnt = 0;
12366 		if (rack->r_ctl.rc_went_idle_time == 0)
12367 			rack->r_ctl.rc_went_idle_time = 1;
12368 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12369 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12370 			tp->t_acktime = 0;
12371 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12372 		rack->rc_suspicious = 0;
12373 		/* Set need output so persist might get set */
12374 		rack->r_wanted_output = 1;
12375 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12376 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12377 		    (sbavail(&so->so_snd) == 0) &&
12378 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12379 			/*
12380 			 * The socket was gone and the
12381 			 * peer sent data (now or in the past), time to
12382 			 * reset him.
12383 			 */
12384 			*ret_val = 1;
12385 			/* tcp_close will kill the inp pre-log the Reset */
12386 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12387 			tp = tcp_close(tp);
12388 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12389 			return (1);
12390 		}
12391 	}
12392 	if (ofia)
12393 		*ofia = ourfinisacked;
12394 	return (0);
12395 }
12396 
12397 
12398 static void
12399 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12400 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12401 {
12402 	if (tcp_bblogging_on(rack->rc_tp)) {
12403 		union tcp_log_stackspecific log;
12404 		struct timeval tv;
12405 
12406 		memset(&log, 0, sizeof(log));
12407 		log.u_bbr.flex1 = cnt;
12408 		log.u_bbr.flex2 = split;
12409 		log.u_bbr.flex3 = out;
12410 		log.u_bbr.flex4 = line;
12411 		log.u_bbr.flex5 = rack->r_must_retran;
12412 		log.u_bbr.flex6 = flags;
12413 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12414 		log.u_bbr.flex8 = dir;	/*
12415 					 * 1 is collapsed, 0 is uncollapsed,
12416 					 * 2 is log of a rsm being marked, 3 is a split.
12417 					 */
12418 		if (rsm == NULL)
12419 			log.u_bbr.rttProp = 0;
12420 		else
12421 			log.u_bbr.rttProp = (uint64_t)rsm;
12422 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12423 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12424 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12425 		    &rack->rc_inp->inp_socket->so_rcv,
12426 		    &rack->rc_inp->inp_socket->so_snd,
12427 		    TCP_RACK_LOG_COLLAPSE, 0,
12428 		    0, &log, false, &tv);
12429 	}
12430 }
12431 
12432 static void
12433 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12434 {
12435 	/*
12436 	 * Here all we do is mark the collapsed point and set the flag.
12437 	 * This may happen again and again, but there is no
12438 	 * sense splitting our map until we know where the
12439 	 * peer finally lands in the collapse.
12440 	 */
12441 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12442 	if ((rack->rc_has_collapsed == 0) ||
12443 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12444 		counter_u64_add(rack_collapsed_win_seen, 1);
12445 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12446 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12447 	rack->rc_has_collapsed = 1;
12448 	rack->r_collapse_point_valid = 1;
12449 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12450 }
12451 
12452 static void
12453 rack_un_collapse_window(struct tcp_rack *rack, int line)
12454 {
12455 	struct rack_sendmap *nrsm, *rsm;
12456 	int cnt = 0, split = 0;
12457 	int insret __diagused;
12458 
12459 
12460 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12461 	rack->rc_has_collapsed = 0;
12462 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12463 	if (rsm == NULL) {
12464 		/* Nothing to do maybe the peer ack'ed it all */
12465 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12466 		return;
12467 	}
12468 	/* Now do we need to split this one? */
12469 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12470 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12471 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12472 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12473 		if (nrsm == NULL) {
12474 			/* We can't get a rsm, mark all? */
12475 			nrsm = rsm;
12476 			goto no_split;
12477 		}
12478 		/* Clone it */
12479 		split = 1;
12480 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12481 #ifndef INVARIANTS
12482 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12483 #else
12484 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12485 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12486 			      nrsm, insret, rack, rsm);
12487 		}
12488 #endif
12489 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12490 				 rack->r_ctl.last_collapse_point, __LINE__);
12491 		if (rsm->r_in_tmap) {
12492 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12493 			nrsm->r_in_tmap = 1;
12494 		}
12495 		/*
12496 		 * Set in the new RSM as the
12497 		 * collapsed starting point
12498 		 */
12499 		rsm = nrsm;
12500 	}
12501 
12502 no_split:
12503 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12504 		cnt++;
12505 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12506 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12507 		cnt++;
12508 	}
12509 	if (cnt) {
12510 		counter_u64_add(rack_collapsed_win, 1);
12511 	}
12512 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12513 }
12514 
12515 static void
12516 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12517 			int32_t tlen, int32_t tfo_syn)
12518 {
12519 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12520 		rack_timer_cancel(tp, rack,
12521 				  rack->r_ctl.rc_rcvtime, __LINE__);
12522 		tp->t_flags |= TF_DELACK;
12523 	} else {
12524 		rack->r_wanted_output = 1;
12525 		tp->t_flags |= TF_ACKNOW;
12526 	}
12527 }
12528 
12529 static void
12530 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12531 {
12532 	/*
12533 	 * If fast output is in progress, lets validate that
12534 	 * the new window did not shrink on us and make it
12535 	 * so fast output should end.
12536 	 */
12537 	if (rack->r_fast_output) {
12538 		uint32_t out;
12539 
12540 		/*
12541 		 * Calculate what we will send if left as is
12542 		 * and compare that to our send window.
12543 		 */
12544 		out = ctf_outstanding(tp);
12545 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12546 			/* ok we have an issue */
12547 			if (out >= tp->snd_wnd) {
12548 				/* Turn off fast output the window is met or collapsed */
12549 				rack->r_fast_output = 0;
12550 			} else {
12551 				/* we have some room left */
12552 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12553 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12554 					/* If not at least 1 full segment never mind */
12555 					rack->r_fast_output = 0;
12556 				}
12557 			}
12558 		}
12559 	}
12560 }
12561 
12562 
12563 /*
12564  * Return value of 1, the TCB is unlocked and most
12565  * likely gone, return value of 0, the TCP is still
12566  * locked.
12567  */
12568 static int
12569 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12570     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12571     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12572 {
12573 	/*
12574 	 * Update window information. Don't look at window if no ACK: TAC's
12575 	 * send garbage on first SYN.
12576 	 */
12577 	int32_t nsegs;
12578 	int32_t tfo_syn;
12579 	struct tcp_rack *rack;
12580 
12581 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12582 
12583 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12584 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12585 	if ((thflags & TH_ACK) &&
12586 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12587 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12588 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12589 		/* keep track of pure window updates */
12590 		if (tlen == 0 &&
12591 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12592 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12593 		tp->snd_wnd = tiwin;
12594 		rack_validate_fo_sendwin_up(tp, rack);
12595 		tp->snd_wl1 = th->th_seq;
12596 		tp->snd_wl2 = th->th_ack;
12597 		if (tp->snd_wnd > tp->max_sndwnd)
12598 			tp->max_sndwnd = tp->snd_wnd;
12599 		rack->r_wanted_output = 1;
12600 	} else if (thflags & TH_ACK) {
12601 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12602 			tp->snd_wnd = tiwin;
12603 			rack_validate_fo_sendwin_up(tp, rack);
12604 			tp->snd_wl1 = th->th_seq;
12605 			tp->snd_wl2 = th->th_ack;
12606 		}
12607 	}
12608 	if (tp->snd_wnd < ctf_outstanding(tp))
12609 		/* The peer collapsed the window */
12610 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12611 	else if (rack->rc_has_collapsed)
12612 		rack_un_collapse_window(rack, __LINE__);
12613 	if ((rack->r_collapse_point_valid) &&
12614 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12615 		rack->r_collapse_point_valid = 0;
12616 	/* Was persist timer active and now we have window space? */
12617 	if ((rack->rc_in_persist != 0) &&
12618 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12619 				rack->r_ctl.rc_pace_min_segs))) {
12620 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12621 		tp->snd_nxt = tp->snd_max;
12622 		/* Make sure we output to start the timer */
12623 		rack->r_wanted_output = 1;
12624 	}
12625 	/* Do we enter persists? */
12626 	if ((rack->rc_in_persist == 0) &&
12627 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12628 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12629 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12630 	    sbavail(&tptosocket(tp)->so_snd) &&
12631 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12632 		/*
12633 		 * Here the rwnd is less than
12634 		 * the pacing size, we are established,
12635 		 * nothing is outstanding, and there is
12636 		 * data to send. Enter persists.
12637 		 */
12638 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12639 	}
12640 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12641 		m_freem(m);
12642 		return (0);
12643 	}
12644 	/*
12645 	 * don't process the URG bit, ignore them drag
12646 	 * along the up.
12647 	 */
12648 	tp->rcv_up = tp->rcv_nxt;
12649 
12650 	/*
12651 	 * Process the segment text, merging it into the TCP sequencing
12652 	 * queue, and arranging for acknowledgment of receipt if necessary.
12653 	 * This process logically involves adjusting tp->rcv_wnd as data is
12654 	 * presented to the user (this happens in tcp_usrreq.c, case
12655 	 * PRU_RCVD).  If a FIN has already been received on this connection
12656 	 * then we just ignore the text.
12657 	 */
12658 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12659 		   IS_FASTOPEN(tp->t_flags));
12660 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12661 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12662 		tcp_seq save_start = th->th_seq;
12663 		tcp_seq save_rnxt  = tp->rcv_nxt;
12664 		int     save_tlen  = tlen;
12665 
12666 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12667 		/*
12668 		 * Insert segment which includes th into TCP reassembly
12669 		 * queue with control block tp.  Set thflags to whether
12670 		 * reassembly now includes a segment with FIN.  This handles
12671 		 * the common case inline (segment is the next to be
12672 		 * received on an established connection, and the queue is
12673 		 * empty), avoiding linkage into and removal from the queue
12674 		 * and repetition of various conversions. Set DELACK for
12675 		 * segments received in order, but ack immediately when
12676 		 * segments are out of order (so fast retransmit can work).
12677 		 */
12678 		if (th->th_seq == tp->rcv_nxt &&
12679 		    SEGQ_EMPTY(tp) &&
12680 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
12681 		    tfo_syn)) {
12682 #ifdef NETFLIX_SB_LIMITS
12683 			u_int mcnt, appended;
12684 
12685 			if (so->so_rcv.sb_shlim) {
12686 				mcnt = m_memcnt(m);
12687 				appended = 0;
12688 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12689 				    CFO_NOSLEEP, NULL) == false) {
12690 					counter_u64_add(tcp_sb_shlim_fails, 1);
12691 					m_freem(m);
12692 					return (0);
12693 				}
12694 			}
12695 #endif
12696 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12697 			tp->rcv_nxt += tlen;
12698 			if (tlen &&
12699 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12700 			    (tp->t_fbyte_in == 0)) {
12701 				tp->t_fbyte_in = ticks;
12702 				if (tp->t_fbyte_in == 0)
12703 					tp->t_fbyte_in = 1;
12704 				if (tp->t_fbyte_out && tp->t_fbyte_in)
12705 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12706 			}
12707 			thflags = tcp_get_flags(th) & TH_FIN;
12708 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12709 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12710 			SOCKBUF_LOCK(&so->so_rcv);
12711 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12712 				m_freem(m);
12713 			} else
12714 #ifdef NETFLIX_SB_LIMITS
12715 				appended =
12716 #endif
12717 					sbappendstream_locked(&so->so_rcv, m, 0);
12718 
12719 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12720 			/* NB: sorwakeup_locked() does an implicit unlock. */
12721 			sorwakeup_locked(so);
12722 #ifdef NETFLIX_SB_LIMITS
12723 			if (so->so_rcv.sb_shlim && appended != mcnt)
12724 				counter_fo_release(so->so_rcv.sb_shlim,
12725 				    mcnt - appended);
12726 #endif
12727 		} else {
12728 			/*
12729 			 * XXX: Due to the header drop above "th" is
12730 			 * theoretically invalid by now.  Fortunately
12731 			 * m_adj() doesn't actually frees any mbufs when
12732 			 * trimming from the head.
12733 			 */
12734 			tcp_seq temp = save_start;
12735 
12736 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
12737 			tp->t_flags |= TF_ACKNOW;
12738 			if (tp->t_flags & TF_WAKESOR) {
12739 				tp->t_flags &= ~TF_WAKESOR;
12740 				/* NB: sorwakeup_locked() does an implicit unlock. */
12741 				sorwakeup_locked(so);
12742 			}
12743 		}
12744 		if ((tp->t_flags & TF_SACK_PERMIT) &&
12745 		    (save_tlen > 0) &&
12746 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
12747 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12748 				/*
12749 				 * DSACK actually handled in the fastpath
12750 				 * above.
12751 				 */
12752 				tcp_update_sack_list(tp, save_start,
12753 				    save_start + save_tlen);
12754 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12755 				if ((tp->rcv_numsacks >= 1) &&
12756 				    (tp->sackblks[0].end == save_start)) {
12757 					/*
12758 					 * Partial overlap, recorded at todrop
12759 					 * above.
12760 					 */
12761 					tcp_update_sack_list(tp,
12762 					    tp->sackblks[0].start,
12763 					    tp->sackblks[0].end);
12764 				} else {
12765 					tcp_update_dsack_list(tp, save_start,
12766 					    save_start + save_tlen);
12767 				}
12768 			} else if (tlen >= save_tlen) {
12769 				/* Update of sackblks. */
12770 				tcp_update_dsack_list(tp, save_start,
12771 				    save_start + save_tlen);
12772 			} else if (tlen > 0) {
12773 				tcp_update_dsack_list(tp, save_start,
12774 				    save_start + tlen);
12775 			}
12776 		}
12777 	} else {
12778 		m_freem(m);
12779 		thflags &= ~TH_FIN;
12780 	}
12781 
12782 	/*
12783 	 * If FIN is received ACK the FIN and let the user know that the
12784 	 * connection is closing.
12785 	 */
12786 	if (thflags & TH_FIN) {
12787 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12788 			/* The socket upcall is handled by socantrcvmore. */
12789 			socantrcvmore(so);
12790 			/*
12791 			 * If connection is half-synchronized (ie NEEDSYN
12792 			 * flag on) then delay ACK, so it may be piggybacked
12793 			 * when SYN is sent. Otherwise, since we received a
12794 			 * FIN then no more input can be expected, send ACK
12795 			 * now.
12796 			 */
12797 			if (tp->t_flags & TF_NEEDSYN) {
12798 				rack_timer_cancel(tp, rack,
12799 				    rack->r_ctl.rc_rcvtime, __LINE__);
12800 				tp->t_flags |= TF_DELACK;
12801 			} else {
12802 				tp->t_flags |= TF_ACKNOW;
12803 			}
12804 			tp->rcv_nxt++;
12805 		}
12806 		switch (tp->t_state) {
12807 			/*
12808 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12809 			 * CLOSE_WAIT state.
12810 			 */
12811 		case TCPS_SYN_RECEIVED:
12812 			tp->t_starttime = ticks;
12813 			/* FALLTHROUGH */
12814 		case TCPS_ESTABLISHED:
12815 			rack_timer_cancel(tp, rack,
12816 			    rack->r_ctl.rc_rcvtime, __LINE__);
12817 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
12818 			break;
12819 
12820 			/*
12821 			 * If still in FIN_WAIT_1 STATE FIN has not been
12822 			 * acked so enter the CLOSING state.
12823 			 */
12824 		case TCPS_FIN_WAIT_1:
12825 			rack_timer_cancel(tp, rack,
12826 			    rack->r_ctl.rc_rcvtime, __LINE__);
12827 			tcp_state_change(tp, TCPS_CLOSING);
12828 			break;
12829 
12830 			/*
12831 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12832 			 * starting the time-wait timer, turning off the
12833 			 * other standard timers.
12834 			 */
12835 		case TCPS_FIN_WAIT_2:
12836 			rack_timer_cancel(tp, rack,
12837 			    rack->r_ctl.rc_rcvtime, __LINE__);
12838 			tcp_twstart(tp);
12839 			return (1);
12840 		}
12841 	}
12842 	/*
12843 	 * Return any desired output.
12844 	 */
12845 	if ((tp->t_flags & TF_ACKNOW) ||
12846 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12847 		rack->r_wanted_output = 1;
12848 	}
12849 	return (0);
12850 }
12851 
12852 /*
12853  * Here nothing is really faster, its just that we
12854  * have broken out the fast-data path also just like
12855  * the fast-ack.
12856  */
12857 static int
12858 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12859     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12860     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12861 {
12862 	int32_t nsegs;
12863 	int32_t newsize = 0;	/* automatic sockbuf scaling */
12864 	struct tcp_rack *rack;
12865 #ifdef NETFLIX_SB_LIMITS
12866 	u_int mcnt, appended;
12867 #endif
12868 
12869 	/*
12870 	 * If last ACK falls within this segment's sequence numbers, record
12871 	 * the timestamp. NOTE that the test is modified according to the
12872 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12873 	 */
12874 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12875 		return (0);
12876 	}
12877 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12878 		return (0);
12879 	}
12880 	if (tiwin && tiwin != tp->snd_wnd) {
12881 		return (0);
12882 	}
12883 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12884 		return (0);
12885 	}
12886 	if (__predict_false((to->to_flags & TOF_TS) &&
12887 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12888 		return (0);
12889 	}
12890 	if (__predict_false((th->th_ack != tp->snd_una))) {
12891 		return (0);
12892 	}
12893 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12894 		return (0);
12895 	}
12896 	if ((to->to_flags & TOF_TS) != 0 &&
12897 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12898 		tp->ts_recent_age = tcp_ts_getticks();
12899 		tp->ts_recent = to->to_tsval;
12900 	}
12901 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12902 	/*
12903 	 * This is a pure, in-sequence data packet with nothing on the
12904 	 * reassembly queue and we have enough buffer space to take it.
12905 	 */
12906 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12907 
12908 #ifdef NETFLIX_SB_LIMITS
12909 	if (so->so_rcv.sb_shlim) {
12910 		mcnt = m_memcnt(m);
12911 		appended = 0;
12912 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12913 		    CFO_NOSLEEP, NULL) == false) {
12914 			counter_u64_add(tcp_sb_shlim_fails, 1);
12915 			m_freem(m);
12916 			return (1);
12917 		}
12918 	}
12919 #endif
12920 	/* Clean receiver SACK report if present */
12921 	if (tp->rcv_numsacks)
12922 		tcp_clean_sackreport(tp);
12923 	KMOD_TCPSTAT_INC(tcps_preddat);
12924 	tp->rcv_nxt += tlen;
12925 	if (tlen &&
12926 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12927 	    (tp->t_fbyte_in == 0)) {
12928 		tp->t_fbyte_in = ticks;
12929 		if (tp->t_fbyte_in == 0)
12930 			tp->t_fbyte_in = 1;
12931 		if (tp->t_fbyte_out && tp->t_fbyte_in)
12932 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12933 	}
12934 	/*
12935 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12936 	 */
12937 	tp->snd_wl1 = th->th_seq;
12938 	/*
12939 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12940 	 */
12941 	tp->rcv_up = tp->rcv_nxt;
12942 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12943 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12944 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12945 
12946 	/* Add data to socket buffer. */
12947 	SOCKBUF_LOCK(&so->so_rcv);
12948 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12949 		m_freem(m);
12950 	} else {
12951 		/*
12952 		 * Set new socket buffer size. Give up when limit is
12953 		 * reached.
12954 		 */
12955 		if (newsize)
12956 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12957 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12958 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12959 #ifdef NETFLIX_SB_LIMITS
12960 		appended =
12961 #endif
12962 			sbappendstream_locked(&so->so_rcv, m, 0);
12963 		ctf_calc_rwin(so, tp);
12964 	}
12965 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12966 	/* NB: sorwakeup_locked() does an implicit unlock. */
12967 	sorwakeup_locked(so);
12968 #ifdef NETFLIX_SB_LIMITS
12969 	if (so->so_rcv.sb_shlim && mcnt != appended)
12970 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12971 #endif
12972 	rack_handle_delayed_ack(tp, rack, tlen, 0);
12973 	if (tp->snd_una == tp->snd_max)
12974 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12975 	return (1);
12976 }
12977 
12978 /*
12979  * This subfunction is used to try to highly optimize the
12980  * fast path. We again allow window updates that are
12981  * in sequence to remain in the fast-path. We also add
12982  * in the __predict's to attempt to help the compiler.
12983  * Note that if we return a 0, then we can *not* process
12984  * it and the caller should push the packet into the
12985  * slow-path.
12986  */
12987 static int
12988 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12989     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12990     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12991 {
12992 	int32_t acked;
12993 	int32_t nsegs;
12994 	int32_t under_pacing = 0;
12995 	struct tcp_rack *rack;
12996 
12997 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12998 		/* Old ack, behind (or duplicate to) the last one rcv'd */
12999 		return (0);
13000 	}
13001 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
13002 		/* Above what we have sent? */
13003 		return (0);
13004 	}
13005 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
13006 		/* We are retransmitting */
13007 		return (0);
13008 	}
13009 	if (__predict_false(tiwin == 0)) {
13010 		/* zero window */
13011 		return (0);
13012 	}
13013 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
13014 		/* We need a SYN or a FIN, unlikely.. */
13015 		return (0);
13016 	}
13017 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
13018 		/* Timestamp is behind .. old ack with seq wrap? */
13019 		return (0);
13020 	}
13021 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
13022 		/* Still recovering */
13023 		return (0);
13024 	}
13025 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13026 	if (rack->r_ctl.rc_sacked) {
13027 		/* We have sack holes on our scoreboard */
13028 		return (0);
13029 	}
13030 	/* Ok if we reach here, we can process a fast-ack */
13031 	if (rack->gp_ready &&
13032 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13033 		under_pacing = 1;
13034 	}
13035 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13036 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
13037 	/* Did the window get updated? */
13038 	if (tiwin != tp->snd_wnd) {
13039 		tp->snd_wnd = tiwin;
13040 		rack_validate_fo_sendwin_up(tp, rack);
13041 		tp->snd_wl1 = th->th_seq;
13042 		if (tp->snd_wnd > tp->max_sndwnd)
13043 			tp->max_sndwnd = tp->snd_wnd;
13044 	}
13045 	/* Do we exit persists? */
13046 	if ((rack->rc_in_persist != 0) &&
13047 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13048 			       rack->r_ctl.rc_pace_min_segs))) {
13049 		rack_exit_persist(tp, rack, cts);
13050 	}
13051 	/* Do we enter persists? */
13052 	if ((rack->rc_in_persist == 0) &&
13053 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13054 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13055 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13056 	    sbavail(&tptosocket(tp)->so_snd) &&
13057 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13058 		/*
13059 		 * Here the rwnd is less than
13060 		 * the pacing size, we are established,
13061 		 * nothing is outstanding, and there is
13062 		 * data to send. Enter persists.
13063 		 */
13064 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
13065 	}
13066 	/*
13067 	 * If last ACK falls within this segment's sequence numbers, record
13068 	 * the timestamp. NOTE that the test is modified according to the
13069 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
13070 	 */
13071 	if ((to->to_flags & TOF_TS) != 0 &&
13072 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
13073 		tp->ts_recent_age = tcp_ts_getticks();
13074 		tp->ts_recent = to->to_tsval;
13075 	}
13076 	/*
13077 	 * This is a pure ack for outstanding data.
13078 	 */
13079 	KMOD_TCPSTAT_INC(tcps_predack);
13080 
13081 	/*
13082 	 * "bad retransmit" recovery.
13083 	 */
13084 	if ((tp->t_flags & TF_PREVVALID) &&
13085 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13086 		tp->t_flags &= ~TF_PREVVALID;
13087 		if (tp->t_rxtshift == 1 &&
13088 		    (int)(ticks - tp->t_badrxtwin) < 0)
13089 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
13090 	}
13091 	/*
13092 	 * Recalculate the transmit timer / rtt.
13093 	 *
13094 	 * Some boxes send broken timestamp replies during the SYN+ACK
13095 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
13096 	 * and blow up the retransmit timer.
13097 	 */
13098 	acked = BYTES_THIS_ACK(tp, th);
13099 
13100 #ifdef TCP_HHOOK
13101 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
13102 	hhook_run_tcp_est_in(tp, th, to);
13103 #endif
13104 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
13105 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13106 	if (acked) {
13107 		struct mbuf *mfree;
13108 
13109 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
13110 		SOCKBUF_LOCK(&so->so_snd);
13111 		mfree = sbcut_locked(&so->so_snd, acked);
13112 		tp->snd_una = th->th_ack;
13113 		/* Note we want to hold the sb lock through the sendmap adjust */
13114 		rack_adjust_sendmap_head(rack, &so->so_snd);
13115 		/* Wake up the socket if we have room to write more */
13116 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13117 		sowwakeup_locked(so);
13118 		m_freem(mfree);
13119 		tp->t_rxtshift = 0;
13120 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13121 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13122 		rack->rc_tlp_in_progress = 0;
13123 		rack->r_ctl.rc_tlp_cnt_out = 0;
13124 		/*
13125 		 * If it is the RXT timer we want to
13126 		 * stop it, so we can restart a TLP.
13127 		 */
13128 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13129 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13130 
13131 #ifdef TCP_REQUEST_TRK
13132 		rack_req_check_for_comp(rack, th->th_ack);
13133 #endif
13134 	}
13135 	/*
13136 	 * Let the congestion control algorithm update congestion control
13137 	 * related information. This typically means increasing the
13138 	 * congestion window.
13139 	 */
13140 	if (tp->snd_wnd < ctf_outstanding(tp)) {
13141 		/* The peer collapsed the window */
13142 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
13143 	} else if (rack->rc_has_collapsed)
13144 		rack_un_collapse_window(rack, __LINE__);
13145 	if ((rack->r_collapse_point_valid) &&
13146 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
13147 		rack->r_collapse_point_valid = 0;
13148 	/*
13149 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
13150 	 */
13151 	tp->snd_wl2 = th->th_ack;
13152 	tp->t_dupacks = 0;
13153 	m_freem(m);
13154 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
13155 
13156 	/*
13157 	 * If all outstanding data are acked, stop retransmit timer,
13158 	 * otherwise restart timer using current (possibly backed-off)
13159 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
13160 	 * If data are ready to send, let tcp_output decide between more
13161 	 * output or persist.
13162 	 */
13163 	if (under_pacing &&
13164 	    (rack->use_fixed_rate == 0) &&
13165 	    (rack->in_probe_rtt == 0) &&
13166 	    rack->rc_gp_dyn_mul &&
13167 	    rack->rc_always_pace) {
13168 		/* Check if we are dragging bottom */
13169 		rack_check_bottom_drag(tp, rack, so);
13170 	}
13171 	if (tp->snd_una == tp->snd_max) {
13172 		tp->t_flags &= ~TF_PREVVALID;
13173 		rack->r_ctl.retran_during_recovery = 0;
13174 		rack->rc_suspicious = 0;
13175 		rack->r_ctl.dsack_byte_cnt = 0;
13176 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13177 		if (rack->r_ctl.rc_went_idle_time == 0)
13178 			rack->r_ctl.rc_went_idle_time = 1;
13179 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13180 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
13181 			tp->t_acktime = 0;
13182 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13183 	}
13184 	if (acked && rack->r_fast_output)
13185 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
13186 	if (sbavail(&so->so_snd)) {
13187 		rack->r_wanted_output = 1;
13188 	}
13189 	return (1);
13190 }
13191 
13192 /*
13193  * Return value of 1, the TCB is unlocked and most
13194  * likely gone, return value of 0, the TCP is still
13195  * locked.
13196  */
13197 static int
13198 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
13199     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13200     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13201 {
13202 	int32_t ret_val = 0;
13203 	int32_t todrop;
13204 	int32_t ourfinisacked = 0;
13205 	struct tcp_rack *rack;
13206 
13207 	INP_WLOCK_ASSERT(tptoinpcb(tp));
13208 
13209 	ctf_calc_rwin(so, tp);
13210 	/*
13211 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
13212 	 * SYN, drop the input. if seg contains a RST, then drop the
13213 	 * connection. if seg does not contain SYN, then drop it. Otherwise
13214 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
13215 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
13216 	 * contains an ECE and ECN support is enabled, the stream is ECN
13217 	 * capable. if SYN has been acked change to ESTABLISHED else
13218 	 * SYN_RCVD state arrange for segment to be acked (eventually)
13219 	 * continue processing rest of data/controls.
13220 	 */
13221 	if ((thflags & TH_ACK) &&
13222 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
13223 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13224 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13225 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13226 		return (1);
13227 	}
13228 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
13229 		TCP_PROBE5(connect__refused, NULL, tp,
13230 		    mtod(m, const char *), tp, th);
13231 		tp = tcp_drop(tp, ECONNREFUSED);
13232 		ctf_do_drop(m, tp);
13233 		return (1);
13234 	}
13235 	if (thflags & TH_RST) {
13236 		ctf_do_drop(m, tp);
13237 		return (1);
13238 	}
13239 	if (!(thflags & TH_SYN)) {
13240 		ctf_do_drop(m, tp);
13241 		return (1);
13242 	}
13243 	tp->irs = th->th_seq;
13244 	tcp_rcvseqinit(tp);
13245 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13246 	if (thflags & TH_ACK) {
13247 		int tfo_partial = 0;
13248 
13249 		KMOD_TCPSTAT_INC(tcps_connects);
13250 		soisconnected(so);
13251 #ifdef MAC
13252 		mac_socketpeer_set_from_mbuf(m, so);
13253 #endif
13254 		/* Do window scaling on this connection? */
13255 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13256 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13257 			tp->rcv_scale = tp->request_r_scale;
13258 		}
13259 		tp->rcv_adv += min(tp->rcv_wnd,
13260 		    TCP_MAXWIN << tp->rcv_scale);
13261 		/*
13262 		 * If not all the data that was sent in the TFO SYN
13263 		 * has been acked, resend the remainder right away.
13264 		 */
13265 		if (IS_FASTOPEN(tp->t_flags) &&
13266 		    (tp->snd_una != tp->snd_max)) {
13267 			tp->snd_nxt = th->th_ack;
13268 			tfo_partial = 1;
13269 		}
13270 		/*
13271 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
13272 		 * will be turned on later.
13273 		 */
13274 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
13275 			rack_timer_cancel(tp, rack,
13276 					  rack->r_ctl.rc_rcvtime, __LINE__);
13277 			tp->t_flags |= TF_DELACK;
13278 		} else {
13279 			rack->r_wanted_output = 1;
13280 			tp->t_flags |= TF_ACKNOW;
13281 		}
13282 
13283 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
13284 
13285 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
13286 			/*
13287 			 * We advance snd_una for the
13288 			 * fast open case. If th_ack is
13289 			 * acknowledging data beyond
13290 			 * snd_una we can't just call
13291 			 * ack-processing since the
13292 			 * data stream in our send-map
13293 			 * will start at snd_una + 1 (one
13294 			 * beyond the SYN). If its just
13295 			 * equal we don't need to do that
13296 			 * and there is no send_map.
13297 			 */
13298 			tp->snd_una++;
13299 		}
13300 		/*
13301 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
13302 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
13303 		 */
13304 		tp->t_starttime = ticks;
13305 		if (tp->t_flags & TF_NEEDFIN) {
13306 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
13307 			tp->t_flags &= ~TF_NEEDFIN;
13308 			thflags &= ~TH_SYN;
13309 		} else {
13310 			tcp_state_change(tp, TCPS_ESTABLISHED);
13311 			TCP_PROBE5(connect__established, NULL, tp,
13312 			    mtod(m, const char *), tp, th);
13313 			rack_cc_conn_init(tp);
13314 		}
13315 	} else {
13316 		/*
13317 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
13318 		 * open.  If segment contains CC option and there is a
13319 		 * cached CC, apply TAO test. If it succeeds, connection is *
13320 		 * half-synchronized. Otherwise, do 3-way handshake:
13321 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13322 		 * there was no CC option, clear cached CC value.
13323 		 */
13324 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13325 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
13326 	}
13327 	/*
13328 	 * Advance th->th_seq to correspond to first data byte. If data,
13329 	 * trim to stay within window, dropping FIN if necessary.
13330 	 */
13331 	th->th_seq++;
13332 	if (tlen > tp->rcv_wnd) {
13333 		todrop = tlen - tp->rcv_wnd;
13334 		m_adj(m, -todrop);
13335 		tlen = tp->rcv_wnd;
13336 		thflags &= ~TH_FIN;
13337 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13338 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13339 	}
13340 	tp->snd_wl1 = th->th_seq - 1;
13341 	tp->rcv_up = th->th_seq;
13342 	/*
13343 	 * Client side of transaction: already sent SYN and data. If the
13344 	 * remote host used T/TCP to validate the SYN, our data will be
13345 	 * ACK'd; if so, enter normal data segment processing in the middle
13346 	 * of step 5, ack processing. Otherwise, goto step 6.
13347 	 */
13348 	if (thflags & TH_ACK) {
13349 		/* For syn-sent we need to possibly update the rtt */
13350 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13351 			uint32_t t, mcts;
13352 
13353 			mcts = tcp_ts_getticks();
13354 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13355 			if (!tp->t_rttlow || tp->t_rttlow > t)
13356 				tp->t_rttlow = t;
13357 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13358 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13359 			tcp_rack_xmit_timer_commit(rack, tp);
13360 		}
13361 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
13362 			return (ret_val);
13363 		/* We may have changed to FIN_WAIT_1 above */
13364 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13365 			/*
13366 			 * In FIN_WAIT_1 STATE in addition to the processing
13367 			 * for the ESTABLISHED state if our FIN is now
13368 			 * acknowledged then enter FIN_WAIT_2.
13369 			 */
13370 			if (ourfinisacked) {
13371 				/*
13372 				 * If we can't receive any more data, then
13373 				 * closing user can proceed. Starting the
13374 				 * timer is contrary to the specification,
13375 				 * but if we don't get a FIN we'll hang
13376 				 * forever.
13377 				 *
13378 				 * XXXjl: we should release the tp also, and
13379 				 * use a compressed state.
13380 				 */
13381 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13382 					soisdisconnected(so);
13383 					tcp_timer_activate(tp, TT_2MSL,
13384 					    (tcp_fast_finwait2_recycle ?
13385 					    tcp_finwait2_timeout :
13386 					    TP_MAXIDLE(tp)));
13387 				}
13388 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13389 			}
13390 		}
13391 	}
13392 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13393 	   tiwin, thflags, nxt_pkt));
13394 }
13395 
13396 /*
13397  * Return value of 1, the TCB is unlocked and most
13398  * likely gone, return value of 0, the TCP is still
13399  * locked.
13400  */
13401 static int
13402 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13403     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13404     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13405 {
13406 	struct tcp_rack *rack;
13407 	int32_t ret_val = 0;
13408 	int32_t ourfinisacked = 0;
13409 
13410 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13411 	ctf_calc_rwin(so, tp);
13412 	if ((thflags & TH_RST) ||
13413 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13414 		return (__ctf_process_rst(m, th, so, tp,
13415 					  &rack->r_ctl.challenge_ack_ts,
13416 					  &rack->r_ctl.challenge_ack_cnt));
13417 	if ((thflags & TH_ACK) &&
13418 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13419 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13420 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13421 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13422 		return (1);
13423 	}
13424 	if (IS_FASTOPEN(tp->t_flags)) {
13425 		/*
13426 		 * When a TFO connection is in SYN_RECEIVED, the
13427 		 * only valid packets are the initial SYN, a
13428 		 * retransmit/copy of the initial SYN (possibly with
13429 		 * a subset of the original data), a valid ACK, a
13430 		 * FIN, or a RST.
13431 		 */
13432 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13433 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13434 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13435 			return (1);
13436 		} else if (thflags & TH_SYN) {
13437 			/* non-initial SYN is ignored */
13438 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13439 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13440 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13441 				ctf_do_drop(m, NULL);
13442 				return (0);
13443 			}
13444 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13445 			ctf_do_drop(m, NULL);
13446 			return (0);
13447 		}
13448 	}
13449 
13450 	/*
13451 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13452 	 * it's less than ts_recent, drop it.
13453 	 */
13454 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13455 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13456 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13457 			return (ret_val);
13458 	}
13459 	/*
13460 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13461 	 * this connection before trimming the data to fit the receive
13462 	 * window.  Check the sequence number versus IRS since we know the
13463 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13464 	 * "LAND" DoS attack.
13465 	 */
13466 	if (SEQ_LT(th->th_seq, tp->irs)) {
13467 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13468 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13469 		return (1);
13470 	}
13471 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13472 			      &rack->r_ctl.challenge_ack_ts,
13473 			      &rack->r_ctl.challenge_ack_cnt)) {
13474 		return (ret_val);
13475 	}
13476 	/*
13477 	 * If last ACK falls within this segment's sequence numbers, record
13478 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13479 	 * from the latest proposal of the tcplw@cray.com list (Braden
13480 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13481 	 * with our earlier PAWS tests, so this check should be solely
13482 	 * predicated on the sequence space of this segment. 3) That we
13483 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13484 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13485 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13486 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13487 	 * p.869. In such cases, we can still calculate the RTT correctly
13488 	 * when RCV.NXT == Last.ACK.Sent.
13489 	 */
13490 	if ((to->to_flags & TOF_TS) != 0 &&
13491 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13492 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13493 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13494 		tp->ts_recent_age = tcp_ts_getticks();
13495 		tp->ts_recent = to->to_tsval;
13496 	}
13497 	tp->snd_wnd = tiwin;
13498 	rack_validate_fo_sendwin_up(tp, rack);
13499 	/*
13500 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13501 	 * is on (half-synchronized state), then queue data for later
13502 	 * processing; else drop segment and return.
13503 	 */
13504 	if ((thflags & TH_ACK) == 0) {
13505 		if (IS_FASTOPEN(tp->t_flags)) {
13506 			rack_cc_conn_init(tp);
13507 		}
13508 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13509 		    tiwin, thflags, nxt_pkt));
13510 	}
13511 	KMOD_TCPSTAT_INC(tcps_connects);
13512 	if (tp->t_flags & TF_SONOTCONN) {
13513 		tp->t_flags &= ~TF_SONOTCONN;
13514 		soisconnected(so);
13515 	}
13516 	/* Do window scaling? */
13517 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13518 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13519 		tp->rcv_scale = tp->request_r_scale;
13520 	}
13521 	/*
13522 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13523 	 * FIN-WAIT-1
13524 	 */
13525 	tp->t_starttime = ticks;
13526 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
13527 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13528 		tp->t_tfo_pending = NULL;
13529 	}
13530 	if (tp->t_flags & TF_NEEDFIN) {
13531 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13532 		tp->t_flags &= ~TF_NEEDFIN;
13533 	} else {
13534 		tcp_state_change(tp, TCPS_ESTABLISHED);
13535 		TCP_PROBE5(accept__established, NULL, tp,
13536 		    mtod(m, const char *), tp, th);
13537 		/*
13538 		 * TFO connections call cc_conn_init() during SYN
13539 		 * processing.  Calling it again here for such connections
13540 		 * is not harmless as it would undo the snd_cwnd reduction
13541 		 * that occurs when a TFO SYN|ACK is retransmitted.
13542 		 */
13543 		if (!IS_FASTOPEN(tp->t_flags))
13544 			rack_cc_conn_init(tp);
13545 	}
13546 	/*
13547 	 * Account for the ACK of our SYN prior to
13548 	 * regular ACK processing below, except for
13549 	 * simultaneous SYN, which is handled later.
13550 	 */
13551 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13552 		tp->snd_una++;
13553 	/*
13554 	 * If segment contains data or ACK, will call tcp_reass() later; if
13555 	 * not, do so now to pass queued data to user.
13556 	 */
13557 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13558 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13559 		    (struct mbuf *)0);
13560 		if (tp->t_flags & TF_WAKESOR) {
13561 			tp->t_flags &= ~TF_WAKESOR;
13562 			/* NB: sorwakeup_locked() does an implicit unlock. */
13563 			sorwakeup_locked(so);
13564 		}
13565 	}
13566 	tp->snd_wl1 = th->th_seq - 1;
13567 	/* For syn-recv we need to possibly update the rtt */
13568 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13569 		uint32_t t, mcts;
13570 
13571 		mcts = tcp_ts_getticks();
13572 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13573 		if (!tp->t_rttlow || tp->t_rttlow > t)
13574 			tp->t_rttlow = t;
13575 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13576 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13577 		tcp_rack_xmit_timer_commit(rack, tp);
13578 	}
13579 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13580 		return (ret_val);
13581 	}
13582 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13583 		/* We could have went to FIN_WAIT_1 (or EST) above */
13584 		/*
13585 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13586 		 * ESTABLISHED state if our FIN is now acknowledged then
13587 		 * enter FIN_WAIT_2.
13588 		 */
13589 		if (ourfinisacked) {
13590 			/*
13591 			 * If we can't receive any more data, then closing
13592 			 * user can proceed. Starting the timer is contrary
13593 			 * to the specification, but if we don't get a FIN
13594 			 * we'll hang forever.
13595 			 *
13596 			 * XXXjl: we should release the tp also, and use a
13597 			 * compressed state.
13598 			 */
13599 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13600 				soisdisconnected(so);
13601 				tcp_timer_activate(tp, TT_2MSL,
13602 				    (tcp_fast_finwait2_recycle ?
13603 				    tcp_finwait2_timeout :
13604 				    TP_MAXIDLE(tp)));
13605 			}
13606 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13607 		}
13608 	}
13609 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13610 	    tiwin, thflags, nxt_pkt));
13611 }
13612 
13613 /*
13614  * Return value of 1, the TCB is unlocked and most
13615  * likely gone, return value of 0, the TCP is still
13616  * locked.
13617  */
13618 static int
13619 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13620     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13621     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13622 {
13623 	int32_t ret_val = 0;
13624 	struct tcp_rack *rack;
13625 
13626 	/*
13627 	 * Header prediction: check for the two common cases of a
13628 	 * uni-directional data xfer.  If the packet has no control flags,
13629 	 * is in-sequence, the window didn't change and we're not
13630 	 * retransmitting, it's a candidate.  If the length is zero and the
13631 	 * ack moved forward, we're the sender side of the xfer.  Just free
13632 	 * the data acked & wake any higher level process that was blocked
13633 	 * waiting for space.  If the length is non-zero and the ack didn't
13634 	 * move, we're the receiver side.  If we're getting packets in-order
13635 	 * (the reassembly queue is empty), add the data toc The socket
13636 	 * buffer and note that we need a delayed ack. Make sure that the
13637 	 * hidden state-flags are also off. Since we check for
13638 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13639 	 */
13640 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13641 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13642 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13643 	    __predict_true(SEGQ_EMPTY(tp)) &&
13644 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13645 		if (tlen == 0) {
13646 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13647 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13648 				return (0);
13649 			}
13650 		} else {
13651 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13652 			    tiwin, nxt_pkt, iptos)) {
13653 				return (0);
13654 			}
13655 		}
13656 	}
13657 	ctf_calc_rwin(so, tp);
13658 
13659 	if ((thflags & TH_RST) ||
13660 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13661 		return (__ctf_process_rst(m, th, so, tp,
13662 					  &rack->r_ctl.challenge_ack_ts,
13663 					  &rack->r_ctl.challenge_ack_cnt));
13664 
13665 	/*
13666 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13667 	 * synchronized state.
13668 	 */
13669 	if (thflags & TH_SYN) {
13670 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13671 		return (ret_val);
13672 	}
13673 	/*
13674 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13675 	 * it's less than ts_recent, drop it.
13676 	 */
13677 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13678 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13679 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13680 			return (ret_val);
13681 	}
13682 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13683 			      &rack->r_ctl.challenge_ack_ts,
13684 			      &rack->r_ctl.challenge_ack_cnt)) {
13685 		return (ret_val);
13686 	}
13687 	/*
13688 	 * If last ACK falls within this segment's sequence numbers, record
13689 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13690 	 * from the latest proposal of the tcplw@cray.com list (Braden
13691 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13692 	 * with our earlier PAWS tests, so this check should be solely
13693 	 * predicated on the sequence space of this segment. 3) That we
13694 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13695 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13696 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13697 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13698 	 * p.869. In such cases, we can still calculate the RTT correctly
13699 	 * when RCV.NXT == Last.ACK.Sent.
13700 	 */
13701 	if ((to->to_flags & TOF_TS) != 0 &&
13702 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13703 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13704 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13705 		tp->ts_recent_age = tcp_ts_getticks();
13706 		tp->ts_recent = to->to_tsval;
13707 	}
13708 	/*
13709 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13710 	 * is on (half-synchronized state), then queue data for later
13711 	 * processing; else drop segment and return.
13712 	 */
13713 	if ((thflags & TH_ACK) == 0) {
13714 		if (tp->t_flags & TF_NEEDSYN) {
13715 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13716 			    tiwin, thflags, nxt_pkt));
13717 
13718 		} else if (tp->t_flags & TF_ACKNOW) {
13719 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13720 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13721 			return (ret_val);
13722 		} else {
13723 			ctf_do_drop(m, NULL);
13724 			return (0);
13725 		}
13726 	}
13727 	/*
13728 	 * Ack processing.
13729 	 */
13730 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13731 		return (ret_val);
13732 	}
13733 	if (sbavail(&so->so_snd)) {
13734 		if (ctf_progress_timeout_check(tp, true)) {
13735 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13736 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13737 			return (1);
13738 		}
13739 	}
13740 	/* State changes only happen in rack_process_data() */
13741 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13742 	    tiwin, thflags, nxt_pkt));
13743 }
13744 
13745 /*
13746  * Return value of 1, the TCB is unlocked and most
13747  * likely gone, return value of 0, the TCP is still
13748  * locked.
13749  */
13750 static int
13751 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13752     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13753     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13754 {
13755 	int32_t ret_val = 0;
13756 	struct tcp_rack *rack;
13757 
13758 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13759 	ctf_calc_rwin(so, tp);
13760 	if ((thflags & TH_RST) ||
13761 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13762 		return (__ctf_process_rst(m, th, so, tp,
13763 					  &rack->r_ctl.challenge_ack_ts,
13764 					  &rack->r_ctl.challenge_ack_cnt));
13765 	/*
13766 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13767 	 * synchronized state.
13768 	 */
13769 	if (thflags & TH_SYN) {
13770 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13771 		return (ret_val);
13772 	}
13773 	/*
13774 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13775 	 * it's less than ts_recent, drop it.
13776 	 */
13777 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13778 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13779 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13780 			return (ret_val);
13781 	}
13782 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13783 			      &rack->r_ctl.challenge_ack_ts,
13784 			      &rack->r_ctl.challenge_ack_cnt)) {
13785 		return (ret_val);
13786 	}
13787 	/*
13788 	 * If last ACK falls within this segment's sequence numbers, record
13789 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13790 	 * from the latest proposal of the tcplw@cray.com list (Braden
13791 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13792 	 * with our earlier PAWS tests, so this check should be solely
13793 	 * predicated on the sequence space of this segment. 3) That we
13794 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13795 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13796 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13797 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13798 	 * p.869. In such cases, we can still calculate the RTT correctly
13799 	 * when RCV.NXT == Last.ACK.Sent.
13800 	 */
13801 	if ((to->to_flags & TOF_TS) != 0 &&
13802 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13803 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13804 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13805 		tp->ts_recent_age = tcp_ts_getticks();
13806 		tp->ts_recent = to->to_tsval;
13807 	}
13808 	/*
13809 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13810 	 * is on (half-synchronized state), then queue data for later
13811 	 * processing; else drop segment and return.
13812 	 */
13813 	if ((thflags & TH_ACK) == 0) {
13814 		if (tp->t_flags & TF_NEEDSYN) {
13815 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13816 			    tiwin, thflags, nxt_pkt));
13817 
13818 		} else if (tp->t_flags & TF_ACKNOW) {
13819 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13820 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13821 			return (ret_val);
13822 		} else {
13823 			ctf_do_drop(m, NULL);
13824 			return (0);
13825 		}
13826 	}
13827 	/*
13828 	 * Ack processing.
13829 	 */
13830 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13831 		return (ret_val);
13832 	}
13833 	if (sbavail(&so->so_snd)) {
13834 		if (ctf_progress_timeout_check(tp, true)) {
13835 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13836 						tp, tick, PROGRESS_DROP, __LINE__);
13837 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13838 			return (1);
13839 		}
13840 	}
13841 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13842 	    tiwin, thflags, nxt_pkt));
13843 }
13844 
13845 static int
13846 rack_check_data_after_close(struct mbuf *m,
13847     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13848 {
13849 	struct tcp_rack *rack;
13850 
13851 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13852 	if (rack->rc_allow_data_af_clo == 0) {
13853 	close_now:
13854 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13855 		/* tcp_close will kill the inp pre-log the Reset */
13856 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13857 		tp = tcp_close(tp);
13858 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13859 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13860 		return (1);
13861 	}
13862 	if (sbavail(&so->so_snd) == 0)
13863 		goto close_now;
13864 	/* Ok we allow data that is ignored and a followup reset */
13865 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13866 	tp->rcv_nxt = th->th_seq + *tlen;
13867 	tp->t_flags2 |= TF2_DROP_AF_DATA;
13868 	rack->r_wanted_output = 1;
13869 	*tlen = 0;
13870 	return (0);
13871 }
13872 
13873 /*
13874  * Return value of 1, the TCB is unlocked and most
13875  * likely gone, return value of 0, the TCP is still
13876  * locked.
13877  */
13878 static int
13879 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13880     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13881     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13882 {
13883 	int32_t ret_val = 0;
13884 	int32_t ourfinisacked = 0;
13885 	struct tcp_rack *rack;
13886 
13887 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13888 	ctf_calc_rwin(so, tp);
13889 
13890 	if ((thflags & TH_RST) ||
13891 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13892 		return (__ctf_process_rst(m, th, so, tp,
13893 					  &rack->r_ctl.challenge_ack_ts,
13894 					  &rack->r_ctl.challenge_ack_cnt));
13895 	/*
13896 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13897 	 * synchronized state.
13898 	 */
13899 	if (thflags & TH_SYN) {
13900 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13901 		return (ret_val);
13902 	}
13903 	/*
13904 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13905 	 * it's less than ts_recent, drop it.
13906 	 */
13907 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13908 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13909 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13910 			return (ret_val);
13911 	}
13912 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13913 			      &rack->r_ctl.challenge_ack_ts,
13914 			      &rack->r_ctl.challenge_ack_cnt)) {
13915 		return (ret_val);
13916 	}
13917 	/*
13918 	 * If new data are received on a connection after the user processes
13919 	 * are gone, then RST the other end.
13920 	 */
13921 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13922 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13923 		return (1);
13924 	/*
13925 	 * If last ACK falls within this segment's sequence numbers, record
13926 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13927 	 * from the latest proposal of the tcplw@cray.com list (Braden
13928 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13929 	 * with our earlier PAWS tests, so this check should be solely
13930 	 * predicated on the sequence space of this segment. 3) That we
13931 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13932 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13933 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13934 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13935 	 * p.869. In such cases, we can still calculate the RTT correctly
13936 	 * when RCV.NXT == Last.ACK.Sent.
13937 	 */
13938 	if ((to->to_flags & TOF_TS) != 0 &&
13939 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13940 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13941 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13942 		tp->ts_recent_age = tcp_ts_getticks();
13943 		tp->ts_recent = to->to_tsval;
13944 	}
13945 	/*
13946 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13947 	 * is on (half-synchronized state), then queue data for later
13948 	 * processing; else drop segment and return.
13949 	 */
13950 	if ((thflags & TH_ACK) == 0) {
13951 		if (tp->t_flags & TF_NEEDSYN) {
13952 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13953 			    tiwin, thflags, nxt_pkt));
13954 		} else if (tp->t_flags & TF_ACKNOW) {
13955 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13956 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13957 			return (ret_val);
13958 		} else {
13959 			ctf_do_drop(m, NULL);
13960 			return (0);
13961 		}
13962 	}
13963 	/*
13964 	 * Ack processing.
13965 	 */
13966 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13967 		return (ret_val);
13968 	}
13969 	if (ourfinisacked) {
13970 		/*
13971 		 * If we can't receive any more data, then closing user can
13972 		 * proceed. Starting the timer is contrary to the
13973 		 * specification, but if we don't get a FIN we'll hang
13974 		 * forever.
13975 		 *
13976 		 * XXXjl: we should release the tp also, and use a
13977 		 * compressed state.
13978 		 */
13979 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13980 			soisdisconnected(so);
13981 			tcp_timer_activate(tp, TT_2MSL,
13982 			    (tcp_fast_finwait2_recycle ?
13983 			    tcp_finwait2_timeout :
13984 			    TP_MAXIDLE(tp)));
13985 		}
13986 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
13987 	}
13988 	if (sbavail(&so->so_snd)) {
13989 		if (ctf_progress_timeout_check(tp, true)) {
13990 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13991 						tp, tick, PROGRESS_DROP, __LINE__);
13992 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13993 			return (1);
13994 		}
13995 	}
13996 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13997 	    tiwin, thflags, nxt_pkt));
13998 }
13999 
14000 /*
14001  * Return value of 1, the TCB is unlocked and most
14002  * likely gone, return value of 0, the TCP is still
14003  * locked.
14004  */
14005 static int
14006 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
14007     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14008     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14009 {
14010 	int32_t ret_val = 0;
14011 	int32_t ourfinisacked = 0;
14012 	struct tcp_rack *rack;
14013 
14014 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14015 	ctf_calc_rwin(so, tp);
14016 
14017 	if ((thflags & TH_RST) ||
14018 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14019 		return (__ctf_process_rst(m, th, so, tp,
14020 					  &rack->r_ctl.challenge_ack_ts,
14021 					  &rack->r_ctl.challenge_ack_cnt));
14022 	/*
14023 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14024 	 * synchronized state.
14025 	 */
14026 	if (thflags & TH_SYN) {
14027 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14028 		return (ret_val);
14029 	}
14030 	/*
14031 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14032 	 * it's less than ts_recent, drop it.
14033 	 */
14034 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14035 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14036 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14037 			return (ret_val);
14038 	}
14039 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14040 			      &rack->r_ctl.challenge_ack_ts,
14041 			      &rack->r_ctl.challenge_ack_cnt)) {
14042 		return (ret_val);
14043 	}
14044 	/*
14045 	 * If new data are received on a connection after the user processes
14046 	 * are gone, then RST the other end.
14047 	 */
14048 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14049 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14050 		return (1);
14051 	/*
14052 	 * If last ACK falls within this segment's sequence numbers, record
14053 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14054 	 * from the latest proposal of the tcplw@cray.com list (Braden
14055 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14056 	 * with our earlier PAWS tests, so this check should be solely
14057 	 * predicated on the sequence space of this segment. 3) That we
14058 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14059 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14060 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14061 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14062 	 * p.869. In such cases, we can still calculate the RTT correctly
14063 	 * when RCV.NXT == Last.ACK.Sent.
14064 	 */
14065 	if ((to->to_flags & TOF_TS) != 0 &&
14066 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14067 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14068 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14069 		tp->ts_recent_age = tcp_ts_getticks();
14070 		tp->ts_recent = to->to_tsval;
14071 	}
14072 	/*
14073 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14074 	 * is on (half-synchronized state), then queue data for later
14075 	 * processing; else drop segment and return.
14076 	 */
14077 	if ((thflags & TH_ACK) == 0) {
14078 		if (tp->t_flags & TF_NEEDSYN) {
14079 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14080 			    tiwin, thflags, nxt_pkt));
14081 		} else if (tp->t_flags & TF_ACKNOW) {
14082 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14083 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14084 			return (ret_val);
14085 		} else {
14086 			ctf_do_drop(m, NULL);
14087 			return (0);
14088 		}
14089 	}
14090 	/*
14091 	 * Ack processing.
14092 	 */
14093 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14094 		return (ret_val);
14095 	}
14096 	if (ourfinisacked) {
14097 		tcp_twstart(tp);
14098 		m_freem(m);
14099 		return (1);
14100 	}
14101 	if (sbavail(&so->so_snd)) {
14102 		if (ctf_progress_timeout_check(tp, true)) {
14103 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14104 						tp, tick, PROGRESS_DROP, __LINE__);
14105 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14106 			return (1);
14107 		}
14108 	}
14109 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14110 	    tiwin, thflags, nxt_pkt));
14111 }
14112 
14113 /*
14114  * Return value of 1, the TCB is unlocked and most
14115  * likely gone, return value of 0, the TCP is still
14116  * locked.
14117  */
14118 static int
14119 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
14120     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14121     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14122 {
14123 	int32_t ret_val = 0;
14124 	int32_t ourfinisacked = 0;
14125 	struct tcp_rack *rack;
14126 
14127 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14128 	ctf_calc_rwin(so, tp);
14129 
14130 	if ((thflags & TH_RST) ||
14131 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14132 		return (__ctf_process_rst(m, th, so, tp,
14133 					  &rack->r_ctl.challenge_ack_ts,
14134 					  &rack->r_ctl.challenge_ack_cnt));
14135 	/*
14136 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14137 	 * synchronized state.
14138 	 */
14139 	if (thflags & TH_SYN) {
14140 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14141 		return (ret_val);
14142 	}
14143 	/*
14144 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14145 	 * it's less than ts_recent, drop it.
14146 	 */
14147 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14148 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14149 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14150 			return (ret_val);
14151 	}
14152 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14153 			      &rack->r_ctl.challenge_ack_ts,
14154 			      &rack->r_ctl.challenge_ack_cnt)) {
14155 		return (ret_val);
14156 	}
14157 	/*
14158 	 * If new data are received on a connection after the user processes
14159 	 * are gone, then RST the other end.
14160 	 */
14161 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14162 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14163 		return (1);
14164 	/*
14165 	 * If last ACK falls within this segment's sequence numbers, record
14166 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14167 	 * from the latest proposal of the tcplw@cray.com list (Braden
14168 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14169 	 * with our earlier PAWS tests, so this check should be solely
14170 	 * predicated on the sequence space of this segment. 3) That we
14171 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14172 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14173 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14174 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14175 	 * p.869. In such cases, we can still calculate the RTT correctly
14176 	 * when RCV.NXT == Last.ACK.Sent.
14177 	 */
14178 	if ((to->to_flags & TOF_TS) != 0 &&
14179 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14180 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14181 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14182 		tp->ts_recent_age = tcp_ts_getticks();
14183 		tp->ts_recent = to->to_tsval;
14184 	}
14185 	/*
14186 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14187 	 * is on (half-synchronized state), then queue data for later
14188 	 * processing; else drop segment and return.
14189 	 */
14190 	if ((thflags & TH_ACK) == 0) {
14191 		if (tp->t_flags & TF_NEEDSYN) {
14192 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14193 			    tiwin, thflags, nxt_pkt));
14194 		} else if (tp->t_flags & TF_ACKNOW) {
14195 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14196 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14197 			return (ret_val);
14198 		} else {
14199 			ctf_do_drop(m, NULL);
14200 			return (0);
14201 		}
14202 	}
14203 	/*
14204 	 * case TCPS_LAST_ACK: Ack processing.
14205 	 */
14206 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14207 		return (ret_val);
14208 	}
14209 	if (ourfinisacked) {
14210 		tp = tcp_close(tp);
14211 		ctf_do_drop(m, tp);
14212 		return (1);
14213 	}
14214 	if (sbavail(&so->so_snd)) {
14215 		if (ctf_progress_timeout_check(tp, true)) {
14216 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14217 						tp, tick, PROGRESS_DROP, __LINE__);
14218 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14219 			return (1);
14220 		}
14221 	}
14222 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14223 	    tiwin, thflags, nxt_pkt));
14224 }
14225 
14226 /*
14227  * Return value of 1, the TCB is unlocked and most
14228  * likely gone, return value of 0, the TCP is still
14229  * locked.
14230  */
14231 static int
14232 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
14233     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14234     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14235 {
14236 	int32_t ret_val = 0;
14237 	int32_t ourfinisacked = 0;
14238 	struct tcp_rack *rack;
14239 
14240 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14241 	ctf_calc_rwin(so, tp);
14242 
14243 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
14244 	if ((thflags & TH_RST) ||
14245 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14246 		return (__ctf_process_rst(m, th, so, tp,
14247 					  &rack->r_ctl.challenge_ack_ts,
14248 					  &rack->r_ctl.challenge_ack_cnt));
14249 	/*
14250 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14251 	 * synchronized state.
14252 	 */
14253 	if (thflags & TH_SYN) {
14254 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14255 		return (ret_val);
14256 	}
14257 	/*
14258 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14259 	 * it's less than ts_recent, drop it.
14260 	 */
14261 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14262 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14263 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14264 			return (ret_val);
14265 	}
14266 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14267 			      &rack->r_ctl.challenge_ack_ts,
14268 			      &rack->r_ctl.challenge_ack_cnt)) {
14269 		return (ret_val);
14270 	}
14271 	/*
14272 	 * If new data are received on a connection after the user processes
14273 	 * are gone, then RST the other end.
14274 	 */
14275 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14276 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14277 		return (1);
14278 	/*
14279 	 * If last ACK falls within this segment's sequence numbers, record
14280 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14281 	 * from the latest proposal of the tcplw@cray.com list (Braden
14282 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14283 	 * with our earlier PAWS tests, so this check should be solely
14284 	 * predicated on the sequence space of this segment. 3) That we
14285 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14286 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14287 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14288 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14289 	 * p.869. In such cases, we can still calculate the RTT correctly
14290 	 * when RCV.NXT == Last.ACK.Sent.
14291 	 */
14292 	if ((to->to_flags & TOF_TS) != 0 &&
14293 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14294 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14295 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14296 		tp->ts_recent_age = tcp_ts_getticks();
14297 		tp->ts_recent = to->to_tsval;
14298 	}
14299 	/*
14300 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14301 	 * is on (half-synchronized state), then queue data for later
14302 	 * processing; else drop segment and return.
14303 	 */
14304 	if ((thflags & TH_ACK) == 0) {
14305 		if (tp->t_flags & TF_NEEDSYN) {
14306 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14307 			    tiwin, thflags, nxt_pkt));
14308 		} else if (tp->t_flags & TF_ACKNOW) {
14309 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14310 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14311 			return (ret_val);
14312 		} else {
14313 			ctf_do_drop(m, NULL);
14314 			return (0);
14315 		}
14316 	}
14317 	/*
14318 	 * Ack processing.
14319 	 */
14320 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14321 		return (ret_val);
14322 	}
14323 	if (sbavail(&so->so_snd)) {
14324 		if (ctf_progress_timeout_check(tp, true)) {
14325 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14326 						tp, tick, PROGRESS_DROP, __LINE__);
14327 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14328 			return (1);
14329 		}
14330 	}
14331 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14332 	    tiwin, thflags, nxt_pkt));
14333 }
14334 
14335 static void inline
14336 rack_clear_rate_sample(struct tcp_rack *rack)
14337 {
14338 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
14339 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
14340 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
14341 }
14342 
14343 static void
14344 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
14345 {
14346 	uint64_t bw_est, rate_wanted;
14347 	int chged = 0;
14348 	uint32_t user_max, orig_min, orig_max;
14349 
14350 #ifdef TCP_REQUEST_TRK
14351 	if (rack->rc_hybrid_mode &&
14352 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
14353 	    (rack_hybrid_allow_set_maxseg == 1) &&
14354 	    (rack->r_ctl.rc_last_sft != NULL)) {
14355 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14356 		return;
14357 	}
14358 #endif
14359 	orig_min = rack->r_ctl.rc_pace_min_segs;
14360 	orig_max = rack->r_ctl.rc_pace_max_segs;
14361 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14362 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14363 		chged = 1;
14364 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14365 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14366 		if (user_max != rack->r_ctl.rc_pace_max_segs)
14367 			chged = 1;
14368 	}
14369 	if (rack->rc_force_max_seg) {
14370 		rack->r_ctl.rc_pace_max_segs = user_max;
14371 	} else if (rack->use_fixed_rate) {
14372 		bw_est = rack_get_bw(rack);
14373 		if ((rack->r_ctl.crte == NULL) ||
14374 		    (bw_est != rack->r_ctl.crte->rate)) {
14375 			rack->r_ctl.rc_pace_max_segs = user_max;
14376 		} else {
14377 			/* We are pacing right at the hardware rate */
14378 			uint32_t segsiz, pace_one;
14379 
14380 			if (rack_pace_one_seg ||
14381 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14382 				pace_one = 1;
14383 			else
14384 				pace_one = 0;
14385 			segsiz = min(ctf_fixed_maxseg(tp),
14386 				     rack->r_ctl.rc_pace_min_segs);
14387 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14388 				tp, bw_est, segsiz, pace_one,
14389 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14390 		}
14391 	} else if (rack->rc_always_pace) {
14392 		if (rack->r_ctl.gp_bw ||
14393 		    rack->r_ctl.init_rate) {
14394 			/* We have a rate of some sort set */
14395 			uint32_t  orig;
14396 
14397 			bw_est = rack_get_bw(rack);
14398 			orig = rack->r_ctl.rc_pace_max_segs;
14399 			if (fill_override)
14400 				rate_wanted = *fill_override;
14401 			else
14402 				rate_wanted = rack_get_gp_est(rack);
14403 			if (rate_wanted) {
14404 				/* We have something */
14405 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14406 										   rate_wanted,
14407 										   ctf_fixed_maxseg(rack->rc_tp));
14408 			} else
14409 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14410 			if (orig != rack->r_ctl.rc_pace_max_segs)
14411 				chged = 1;
14412 		} else if ((rack->r_ctl.gp_bw == 0) &&
14413 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14414 			/*
14415 			 * If we have nothing limit us to bursting
14416 			 * out IW sized pieces.
14417 			 */
14418 			chged = 1;
14419 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14420 		}
14421 	}
14422 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14423 		chged = 1;
14424 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14425 	}
14426 	if (chged)
14427 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14428 }
14429 
14430 
14431 static void
14432 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14433 {
14434 #ifdef INET6
14435 	struct ip6_hdr *ip6 = NULL;
14436 #endif
14437 #ifdef INET
14438 	struct ip *ip = NULL;
14439 #endif
14440 	struct udphdr *udp = NULL;
14441 
14442 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14443 #ifdef INET6
14444 	if (rack->r_is_v6) {
14445 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14446 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14447 		if (tp->t_port) {
14448 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14449 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14450 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14451 			udp->uh_dport = tp->t_port;
14452 			rack->r_ctl.fsb.udp = udp;
14453 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14454 		} else
14455 		{
14456 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14457 			rack->r_ctl.fsb.udp = NULL;
14458 		}
14459 		tcpip_fillheaders(rack->rc_inp,
14460 				  tp->t_port,
14461 				  ip6, rack->r_ctl.fsb.th);
14462 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14463 	} else
14464 #endif				/* INET6 */
14465 #ifdef INET
14466 	{
14467 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14468 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14469 		if (tp->t_port) {
14470 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14471 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14472 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14473 			udp->uh_dport = tp->t_port;
14474 			rack->r_ctl.fsb.udp = udp;
14475 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14476 		} else
14477 		{
14478 			rack->r_ctl.fsb.udp = NULL;
14479 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14480 		}
14481 		tcpip_fillheaders(rack->rc_inp,
14482 				  tp->t_port,
14483 				  ip, rack->r_ctl.fsb.th);
14484 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14485 	}
14486 #endif
14487 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14488 	    (long)TCP_MAXWIN << tp->rcv_scale);
14489 	rack->r_fsb_inited = 1;
14490 }
14491 
14492 static int
14493 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14494 {
14495 	/*
14496 	 * Allocate the larger of spaces V6 if available else just
14497 	 * V4 and include udphdr (overbook)
14498 	 */
14499 #ifdef INET6
14500 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14501 #else
14502 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14503 #endif
14504 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14505 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14506 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14507 		return (ENOMEM);
14508 	}
14509 	rack->r_fsb_inited = 0;
14510 	return (0);
14511 }
14512 
14513 static void
14514 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14515 {
14516 	/*
14517 	 * Types of logs (mod value)
14518 	 * 20 - Initial round setup
14519 	 * 21 - Rack declares a new round.
14520 	 */
14521 	struct tcpcb *tp;
14522 
14523 	tp = rack->rc_tp;
14524 	if (tcp_bblogging_on(tp)) {
14525 		union tcp_log_stackspecific log;
14526 		struct timeval tv;
14527 
14528 		memset(&log, 0, sizeof(log));
14529 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14530 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14531 		log.u_bbr.flex3 = high_seq;
14532 		log.u_bbr.flex4 = tp->snd_max;
14533 		log.u_bbr.flex8 = mod;
14534 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14535 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14536 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14537 		TCP_LOG_EVENTP(tp, NULL,
14538 		    &tptosocket(tp)->so_rcv,
14539 		    &tptosocket(tp)->so_snd,
14540 		    TCP_HYSTART, 0,
14541 		    0, &log, false, &tv);
14542 	}
14543 }
14544 
14545 static void
14546 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14547 {
14548 	rack->rack_deferred_inited = 1;
14549 	rack->r_ctl.roundends = tp->snd_max;
14550 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14551 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14552 }
14553 
14554 static void
14555 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14556 {
14557 	/* Retransmit bit controls.
14558 	 *
14559 	 * The setting of these values control one of
14560 	 * three settings you can have and dictate
14561 	 * how rack does retransmissions. Note this
14562 	 * is in *any* mode i.e. pacing on or off DGP
14563 	 * fixed rate pacing, or just bursting rack.
14564 	 *
14565 	 * 1 - Use full sized retransmits i.e. limit
14566 	 *     the size to whatever the pace_max_segments
14567 	 *     size is.
14568 	 *
14569 	 * 2 - Use pacer min granularity as a guide to
14570 	 *     the size combined with the current calculated
14571 	 *     goodput b/w measurement. So for example if
14572 	 *     the goodput is measured at 20Mbps we would
14573 	 *     calculate 8125 (pacer minimum 250usec in
14574 	 *     that b/w) and then round it up to the next
14575 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14576 	 *
14577 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14578 	 *     fall here too if we are setting via rack_init()).
14579 	 *
14580 	 */
14581 	if (ctl == 1) {
14582 		rack->full_size_rxt = 1;
14583 		rack->shape_rxt_to_pacing_min  = 0;
14584 	} else if (ctl == 2) {
14585 		rack->full_size_rxt = 0;
14586 		rack->shape_rxt_to_pacing_min  = 1;
14587 	} else {
14588 		rack->full_size_rxt = 0;
14589 		rack->shape_rxt_to_pacing_min  = 0;
14590 	}
14591 }
14592 
14593 static void
14594 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14595 		  uint32_t flex1,
14596 		  uint32_t flex2,
14597 		  uint32_t flex3)
14598 {
14599 	if (tcp_bblogging_on(rack->rc_tp)) {
14600 		union tcp_log_stackspecific log;
14601 		struct timeval tv;
14602 
14603 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14604 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14605 		log.u_bbr.flex8 = mod;
14606 		log.u_bbr.flex1 = flex1;
14607 		log.u_bbr.flex2 = flex2;
14608 		log.u_bbr.flex3 = flex3;
14609 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14610 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14611 	}
14612 }
14613 
14614 static int
14615 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14616 {
14617 	struct tcp_rack *rack;
14618 	struct rack_sendmap *rsm;
14619 	int i;
14620 
14621 
14622 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14623 	switch (reqr->req) {
14624 	case TCP_QUERY_SENDMAP:
14625 		if ((reqr->req_param == tp->snd_max) ||
14626 		    (tp->snd_max == tp->snd_una)){
14627 			/* Unlikely */
14628 			return (0);
14629 		}
14630 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14631 		if (rsm == NULL) {
14632 			/* Can't find that seq -- unlikely */
14633 			return (0);
14634 		}
14635 		reqr->sendmap_start = rsm->r_start;
14636 		reqr->sendmap_end = rsm->r_end;
14637 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14638 		reqr->sendmap_fas = rsm->r_fas;
14639 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14640 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14641 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14642 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14643 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14644 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14645 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14646 		reqr->sendmap_dupacks = rsm->r_dupack;
14647 		rack_log_chg_info(tp, rack, 1,
14648 				  rsm->r_start,
14649 				  rsm->r_end,
14650 				  rsm->r_flags);
14651 		return(1);
14652 		break;
14653 	case TCP_QUERY_TIMERS_UP:
14654 		if (rack->r_ctl.rc_hpts_flags == 0) {
14655 			/* no timers up */
14656 			return (0);
14657 		}
14658 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14659 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14660 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14661 		}
14662 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14663 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14664 		}
14665 		rack_log_chg_info(tp, rack, 2,
14666 				  rack->r_ctl.rc_hpts_flags,
14667 				  rack->r_ctl.rc_last_output_to,
14668 				  rack->r_ctl.rc_timer_exp);
14669 		return (1);
14670 		break;
14671 	case TCP_QUERY_RACK_TIMES:
14672 		/* Reordering items */
14673 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14674 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14675 		/* Timerstamps and timers */
14676 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14677 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14678 		reqr->rack_rtt = rack->rc_rack_rtt;
14679 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14680 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14681 		/* PRR data */
14682 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
14683 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14684 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14685 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14686 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14687 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14688 		/* TLP and persists info */
14689 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14690 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14691 		if (rack->rc_in_persist) {
14692 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14693 			reqr->rack_in_persist = 1;
14694 		} else {
14695 			reqr->rack_time_went_idle = 0;
14696 			reqr->rack_in_persist = 0;
14697 		}
14698 		if (rack->r_wanted_output)
14699 			reqr->rack_wanted_output = 1;
14700 		else
14701 			reqr->rack_wanted_output = 0;
14702 		return (1);
14703 		break;
14704 	default:
14705 		return (-EINVAL);
14706 	}
14707 }
14708 
14709 static void
14710 rack_switch_failed(struct tcpcb *tp)
14711 {
14712 	/*
14713 	 * This method gets called if a stack switch was
14714 	 * attempted and it failed. We are left
14715 	 * but our hpts timers were stopped and we
14716 	 * need to validate time units and t_flags2.
14717 	 */
14718 	struct tcp_rack *rack;
14719 	struct timeval tv;
14720 	uint32_t cts;
14721 	uint32_t toval;
14722 	struct hpts_diag diag;
14723 
14724 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14725 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14726 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14727 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14728 	else
14729 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14730 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14731 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14732 	if (tp->t_in_hpts > IHPTS_NONE) {
14733 		/* Strange */
14734 		return;
14735 	}
14736 	cts = tcp_get_usecs(&tv);
14737 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14738 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14739 			toval = rack->r_ctl.rc_last_output_to - cts;
14740 		} else {
14741 			/* one slot please */
14742 			toval = HPTS_TICKS_PER_SLOT;
14743 		}
14744 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14745 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14746 			toval = rack->r_ctl.rc_timer_exp - cts;
14747 		} else {
14748 			/* one slot please */
14749 			toval = HPTS_TICKS_PER_SLOT;
14750 		}
14751 	} else
14752 		toval = HPTS_TICKS_PER_SLOT;
14753 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14754 				   __LINE__, &diag);
14755 	rack_log_hpts_diag(rack, cts, &diag, &tv);
14756 }
14757 
14758 static int
14759 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14760 {
14761 	struct rack_sendmap *rsm, *ersm;
14762 	int insret __diagused;
14763 	/*
14764 	 * When initing outstanding, we must be quite careful
14765 	 * to not refer to tp->t_fb_ptr. This has the old rack
14766 	 * pointer in it, not the "new" one (when we are doing
14767 	 * a stack switch).
14768 	 */
14769 
14770 
14771 	if (tp->t_fb->tfb_chg_query == NULL) {
14772 		/* Create a send map for the current outstanding data */
14773 
14774 		rsm = rack_alloc(rack);
14775 		if (rsm == NULL) {
14776 			uma_zfree(rack_pcb_zone, ptr);
14777 			return (ENOMEM);
14778 		}
14779 		rsm->r_no_rtt_allowed = 1;
14780 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14781 		rsm->r_rtr_cnt = 1;
14782 		rsm->r_rtr_bytes = 0;
14783 		if (tp->t_flags & TF_SENTFIN)
14784 			rsm->r_flags |= RACK_HAS_FIN;
14785 		rsm->r_end = tp->snd_max;
14786 		if (tp->snd_una == tp->iss) {
14787 			/* The data space is one beyond snd_una */
14788 			rsm->r_flags |= RACK_HAS_SYN;
14789 			rsm->r_start = tp->iss;
14790 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14791 		} else
14792 			rsm->r_start = tp->snd_una;
14793 		rsm->r_dupack = 0;
14794 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14795 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14796 			if (rsm->m) {
14797 				rsm->orig_m_len = rsm->m->m_len;
14798 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14799 			} else {
14800 				rsm->orig_m_len = 0;
14801 				rsm->orig_t_space = 0;
14802 			}
14803 		} else {
14804 			/*
14805 			 * This can happen if we have a stand-alone FIN or
14806 			 *  SYN.
14807 			 */
14808 			rsm->m = NULL;
14809 			rsm->orig_m_len = 0;
14810 			rsm->orig_t_space = 0;
14811 			rsm->soff = 0;
14812 		}
14813 #ifdef INVARIANTS
14814 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14815 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14816 			      insret, rack, rsm);
14817 		}
14818 #else
14819 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14820 #endif
14821 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14822 		rsm->r_in_tmap = 1;
14823 	} else {
14824 		/* We have a query mechanism, lets use it */
14825 		struct tcp_query_resp qr;
14826 		int i;
14827 		tcp_seq at;
14828 
14829 		at = tp->snd_una;
14830 		while (at != tp->snd_max) {
14831 			memset(&qr, 0, sizeof(qr));
14832 			qr.req = TCP_QUERY_SENDMAP;
14833 			qr.req_param = at;
14834 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14835 				break;
14836 			/* Move forward */
14837 			at = qr.sendmap_end;
14838 			/* Now lets build the entry for this one */
14839 			rsm = rack_alloc(rack);
14840 			if (rsm == NULL) {
14841 				uma_zfree(rack_pcb_zone, ptr);
14842 				return (ENOMEM);
14843 			}
14844 			memset(rsm, 0, sizeof(struct rack_sendmap));
14845 			/* Now configure the rsm and insert it */
14846 			rsm->r_dupack = qr.sendmap_dupacks;
14847 			rsm->r_start = qr.sendmap_start;
14848 			rsm->r_end = qr.sendmap_end;
14849 			if (qr.sendmap_fas)
14850 				rsm->r_fas = qr.sendmap_end;
14851 			else
14852 				rsm->r_fas = rsm->r_start - tp->snd_una;
14853 			/*
14854 			 * We have carefully aligned the bits
14855 			 * so that all we have to do is copy over
14856 			 * the bits with the mask.
14857 			 */
14858 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14859 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14860 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14861 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14862 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
14863 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
14864 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14865 					   (rsm->r_start - tp->snd_una), &rsm->soff);
14866 			if (rsm->m) {
14867 				rsm->orig_m_len = rsm->m->m_len;
14868 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14869 			} else {
14870 				rsm->orig_m_len = 0;
14871 				rsm->orig_t_space = 0;
14872 			}
14873 #ifdef INVARIANTS
14874 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14875 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14876 				      insret, rack, rsm);
14877 			}
14878 #else
14879 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14880 #endif
14881 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
14882 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14883 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14884 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14885 						/*
14886 						 * If the existing ersm was sent at
14887 						 * a later time than the new one, then
14888 						 * the new one should appear ahead of this
14889 						 * ersm.
14890 						 */
14891 						rsm->r_in_tmap = 1;
14892 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14893 						break;
14894 					}
14895 				}
14896 				if (rsm->r_in_tmap == 0) {
14897 					/*
14898 					 * Not found so shove it on the tail.
14899 					 */
14900 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14901 					rsm->r_in_tmap = 1;
14902 				}
14903  			} else {
14904 				if ((rack->r_ctl.rc_sacklast == NULL) ||
14905 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14906 					rack->r_ctl.rc_sacklast = rsm;
14907 				}
14908 			}
14909 			rack_log_chg_info(tp, rack, 3,
14910 					  rsm->r_start,
14911 					  rsm->r_end,
14912 					  rsm->r_flags);
14913 		}
14914 	}
14915 	return (0);
14916 }
14917 
14918 static void
14919 rack_translate_clamp_value(struct tcp_rack *rack, uint32_t optval)
14920 {
14921 	/*
14922 	 * P = percent bits
14923 	 * F = fill cw bit -- Toggle fillcw if this bit is set.
14924 	 * S = Segment bits
14925 	 * M = set max segment bit
14926 	 * U = Unclamined
14927 	 * C = If set to non-zero override the max number of clamps.
14928 	 * L = Bit to indicate if clamped gets lower.
14929 	 *
14930 	 * CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP
14931 	 *
14932 	 * The lowest 3 nibbles is the perentage .1 - 6553.5%
14933 	 * where 10.1 = 101, max 6553.5
14934 	 * The upper 16 bits  holds some options.
14935 	 * The F bit will turn on fill-cw on if you are
14936 	 * not pacing, it will turn it off if dgp is on.
14937 	 * The L bit will change it so when clamped we get
14938 	 * the min(gp, lt-bw) for dgp.
14939 	 */
14940 	uint16_t per;
14941 
14942 	rack->r_ctl.saved_rxt_clamp_val = optval;
14943 	per = optval & 0x0000ffff;
14944 	rack->r_ctl.rxt_threshold = (uint64_t)(per & 0xffff);
14945 	if (optval > 0) {
14946 		uint16_t clamp_opt;
14947 
14948 		rack->excess_rxt_on = 1;
14949 		clamp_opt = ((optval & 0xffff0000) >> 16);
14950 		rack->r_ctl.clamp_options = clamp_opt & 0x00ff;
14951 		if (clamp_opt & 0xff00) {
14952 			/* A max clamps is also present */
14953 			rack->r_ctl.max_clamps = (clamp_opt >> 8);
14954 		} else {
14955 			/* No specified clamps means no limit */
14956 			rack->r_ctl.max_clamps = 0;
14957 		}
14958 		if (rack->r_ctl.clamp_options & 0x0002) {
14959 			rack->r_clamped_gets_lower  = 1;
14960 		} else {
14961 			rack->r_clamped_gets_lower  = 0;
14962 		}
14963 	} else {
14964 		/* Turn it off back to default */
14965 		rack->excess_rxt_on = 0;
14966 		rack->r_clamped_gets_lower  = 0;
14967 	}
14968 
14969 }
14970 
14971 
14972 static int32_t
14973 rack_init(struct tcpcb *tp, void **ptr)
14974 {
14975 	struct inpcb *inp = tptoinpcb(tp);
14976 	struct tcp_rack *rack = NULL;
14977 	uint32_t iwin, snt, us_cts;
14978 	int err, no_query;
14979 
14980 	tcp_hpts_init(tp);
14981 
14982 	/*
14983 	 * First are we the initial or are we a switched stack?
14984 	 * If we are initing via tcp_newtcppcb the ptr passed
14985 	 * will be tp->t_fb_ptr. If its a stack switch that
14986 	 * has a previous stack we can query it will be a local
14987 	 * var that will in the end be set into t_fb_ptr.
14988 	 */
14989 	if (ptr == &tp->t_fb_ptr)
14990 		no_query = 1;
14991 	else
14992 		no_query = 0;
14993 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14994 	if (*ptr == NULL) {
14995 		/*
14996 		 * We need to allocate memory but cant. The INP and INP_INFO
14997 		 * locks and they are recursive (happens during setup. So a
14998 		 * scheme to drop the locks fails :(
14999 		 *
15000 		 */
15001 		return(ENOMEM);
15002 	}
15003 	memset(*ptr, 0, sizeof(struct tcp_rack));
15004 	rack = (struct tcp_rack *)*ptr;
15005 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
15006 	if (rack->r_ctl.tqh == NULL) {
15007 		uma_zfree(rack_pcb_zone, rack);
15008 		return(ENOMEM);
15009 	}
15010 	tqhash_init(rack->r_ctl.tqh);
15011 	TAILQ_INIT(&rack->r_ctl.rc_free);
15012 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
15013 	rack->rc_tp = tp;
15014 	rack->rc_inp = inp;
15015 	/* Set the flag */
15016 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
15017 	/* Probably not needed but lets be sure */
15018 	rack_clear_rate_sample(rack);
15019 	/*
15020 	 * Save off the default values, socket options will poke
15021 	 * at these if pacing is not on or we have not yet
15022 	 * reached where pacing is on (gp_ready/fixed enabled).
15023 	 * When they get set into the CC module (when gp_ready
15024 	 * is enabled or we enable fixed) then we will set these
15025 	 * values into the CC and place in here the old values
15026 	 * so we have a restoral. Then we will set the flag
15027 	 * rc_pacing_cc_set. That way whenever we turn off pacing
15028 	 * or switch off this stack, we will know to go restore
15029 	 * the saved values.
15030 	 *
15031 	 * We specifically put into the beta the ecn value for pacing.
15032 	 */
15033 	rack->rc_new_rnd_needed = 1;
15034 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
15035 	/* We want abe like behavior as well */
15036 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
15037 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
15038 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
15039 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
15040 	if (rack_rxt_clamp_thresh) {
15041 		rack_translate_clamp_value(rack, rack_rxt_clamp_thresh);
15042 		rack->excess_rxt_on = 1;
15043 	}
15044 	if (rack_uses_full_dgp_in_rec)
15045 		rack->r_ctl.full_dgp_in_rec = 1;
15046 	if (rack_fill_cw_state)
15047 		rack->rc_pace_to_cwnd = 1;
15048 	if (rack_pacing_min_seg)
15049 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
15050 	if (use_rack_rr)
15051 		rack->use_rack_rr = 1;
15052 	if (rack_dnd_default) {
15053 		rack->rc_pace_dnd = 1;
15054 	}
15055 	if (V_tcp_delack_enabled)
15056 		tp->t_delayed_ack = 1;
15057 	else
15058 		tp->t_delayed_ack = 0;
15059 #ifdef TCP_ACCOUNTING
15060 	if (rack_tcp_accounting) {
15061 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
15062 	}
15063 #endif
15064 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
15065 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
15066 	if (rack_enable_shared_cwnd)
15067 		rack->rack_enable_scwnd = 1;
15068 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
15069 	rack->rc_user_set_max_segs = rack_hptsi_segments;
15070 	rack->rc_force_max_seg = 0;
15071 	TAILQ_INIT(&rack->r_ctl.opt_list);
15072 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
15073 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
15074 	if (rack_hibeta_setting) {
15075 		rack->rack_hibeta = 1;
15076 		if ((rack_hibeta_setting >= 50) &&
15077 		    (rack_hibeta_setting <= 100)) {
15078 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
15079 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
15080 		}
15081 	} else {
15082 		rack->r_ctl.saved_hibeta = 50;
15083 	}
15084 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
15085 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
15086 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
15087 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
15088 	rack->r_ctl.rc_highest_us_rtt = 0;
15089 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
15090 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
15091 	if (rack_use_cmp_acks)
15092 		rack->r_use_cmp_ack = 1;
15093 	if (rack_disable_prr)
15094 		rack->rack_no_prr = 1;
15095 	if (rack_gp_no_rec_chg)
15096 		rack->rc_gp_no_rec_chg = 1;
15097 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
15098 		rack->rc_always_pace = 1;
15099 		if (rack->rack_hibeta)
15100 			rack_set_cc_pacing(rack);
15101 	} else
15102 		rack->rc_always_pace = 0;
15103 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
15104 		rack->r_mbuf_queue = 1;
15105 	else
15106 		rack->r_mbuf_queue = 0;
15107 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
15108 	if (rack_limits_scwnd)
15109 		rack->r_limit_scw = 1;
15110 	else
15111 		rack->r_limit_scw = 0;
15112 	rack_init_retransmit_value(rack, rack_rxt_controls);
15113 	rack->rc_labc = V_tcp_abc_l_var;
15114 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
15115 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
15116 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
15117 	rack->r_ctl.rc_min_to = rack_min_to;
15118 	microuptime(&rack->r_ctl.act_rcv_time);
15119 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
15120 	rack->rc_init_win = rack_default_init_window;
15121 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
15122 	if (rack_hw_up_only)
15123 		rack->r_up_only = 1;
15124 	if (rack_do_dyn_mul) {
15125 		/* When dynamic adjustment is on CA needs to start at 100% */
15126 		rack->rc_gp_dyn_mul = 1;
15127 		if (rack_do_dyn_mul >= 100)
15128 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
15129 	} else
15130 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
15131 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
15132 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
15133 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
15134 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
15135 				rack_probertt_filter_life);
15136 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15137 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
15138 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
15139 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
15140 	rack->r_ctl.rc_time_probertt_starts = 0;
15141 	if (rack_dsack_std_based & 0x1) {
15142 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
15143 		rack->rc_rack_tmr_std_based = 1;
15144 	}
15145 	if (rack_dsack_std_based & 0x2) {
15146 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
15147 		rack->rc_rack_use_dsack = 1;
15148 	}
15149 	/* We require at least one measurement, even if the sysctl is 0 */
15150 	if (rack_req_measurements)
15151 		rack->r_ctl.req_measurements = rack_req_measurements;
15152 	else
15153 		rack->r_ctl.req_measurements = 1;
15154 	if (rack_enable_hw_pacing)
15155 		rack->rack_hdw_pace_ena = 1;
15156 	if (rack_hw_rate_caps)
15157 		rack->r_rack_hw_rate_caps = 1;
15158 #ifdef TCP_SAD_DETECTION
15159 	rack->do_detection = 1;
15160 #else
15161 	rack->do_detection = 0;
15162 #endif
15163 	if (rack_non_rxt_use_cr)
15164 		rack->rack_rec_nonrxt_use_cr = 1;
15165 	/* Lets setup the fsb block */
15166 	err = rack_init_fsb(tp, rack);
15167 	if (err) {
15168 		uma_zfree(rack_pcb_zone, *ptr);
15169 		*ptr = NULL;
15170 		return (err);
15171 	}
15172 	if (rack_do_hystart) {
15173 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
15174 		if (rack_do_hystart > 1)
15175 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
15176 		if (rack_do_hystart > 2)
15177 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
15178 	}
15179 	/* Log what we will do with queries */
15180 	rack_log_chg_info(tp, rack, 7,
15181 			  no_query, 0, 0);
15182 	if (rack_def_profile)
15183 		rack_set_profile(rack, rack_def_profile);
15184 	/* Cancel the GP measurement in progress */
15185 	tp->t_flags &= ~TF_GPUTINPROG;
15186 	if ((tp->t_state != TCPS_CLOSED) &&
15187 	    (tp->t_state != TCPS_TIME_WAIT)) {
15188 		/*
15189 		 * We are already open, we may
15190 		 * need to adjust a few things.
15191 		 */
15192 		if (SEQ_GT(tp->snd_max, tp->iss))
15193 			snt = tp->snd_max - tp->iss;
15194 		else
15195 			snt = 0;
15196 		iwin = rc_init_window(rack);
15197 		if ((snt < iwin) &&
15198 		    (no_query == 1)) {
15199 			/* We are not past the initial window
15200 			 * on the first init (i.e. a stack switch
15201 			 * has not yet occured) so we need to make
15202 			 * sure cwnd and ssthresh is correct.
15203 			 */
15204 			if (tp->snd_cwnd < iwin)
15205 				tp->snd_cwnd = iwin;
15206 			/*
15207 			 * If we are within the initial window
15208 			 * we want ssthresh to be unlimited. Setting
15209 			 * it to the rwnd (which the default stack does
15210 			 * and older racks) is not really a good idea
15211 			 * since we want to be in SS and grow both the
15212 			 * cwnd and the rwnd (via dynamic rwnd growth). If
15213 			 * we set it to the rwnd then as the peer grows its
15214 			 * rwnd we will be stuck in CA and never hit SS.
15215 			 *
15216 			 * Its far better to raise it up high (this takes the
15217 			 * risk that there as been a loss already, probably
15218 			 * we should have an indicator in all stacks of loss
15219 			 * but we don't), but considering the normal use this
15220 			 * is a risk worth taking. The consequences of not
15221 			 * hitting SS are far worse than going one more time
15222 			 * into it early on (before we have sent even a IW).
15223 			 * It is highly unlikely that we will have had a loss
15224 			 * before getting the IW out.
15225 			 */
15226 			tp->snd_ssthresh = 0xffffffff;
15227 		}
15228 		/*
15229 		 * Any init based on sequence numbers
15230 		 * should be done in the deferred init path
15231 		 * since we can be CLOSED and not have them
15232 		 * inited when rack_init() is called. We
15233 		 * are not closed so lets call it.
15234 		 */
15235 		rack_deferred_init(tp, rack);
15236 	}
15237 	if ((tp->t_state != TCPS_CLOSED) &&
15238 	    (tp->t_state != TCPS_TIME_WAIT) &&
15239 	    (no_query == 0) &&
15240 	    (tp->snd_una != tp->snd_max))  {
15241 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
15242 		if (err) {
15243 			*ptr = NULL;
15244 			return(err);
15245 		}
15246 	}
15247 	rack_stop_all_timers(tp, rack);
15248 	/* Setup all the t_flags2 */
15249 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15250 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15251 	else
15252 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15253 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15254 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
15255 	/*
15256 	 * Timers in Rack are kept in microseconds so lets
15257 	 * convert any initial incoming variables
15258 	 * from ticks into usecs. Note that we
15259 	 * also change the values of t_srtt and t_rttvar, if
15260 	 * they are non-zero. They are kept with a 5
15261 	 * bit decimal so we have to carefully convert
15262 	 * these to get the full precision.
15263 	 */
15264 	rack_convert_rtts(tp);
15265 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
15266 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
15267 		/* We do not start any timers on DROPPED connections */
15268 		if (tp->t_fb->tfb_chg_query == NULL) {
15269 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15270 		} else {
15271 			struct tcp_query_resp qr;
15272 			int ret;
15273 
15274 			memset(&qr, 0, sizeof(qr));
15275 
15276 			/* Get the misc time stamps and such for rack */
15277 			qr.req = TCP_QUERY_RACK_TIMES;
15278 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15279 			if (ret == 1) {
15280 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
15281 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
15282 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
15283 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
15284 				rack->rc_rack_rtt = qr.rack_rtt;
15285 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
15286 				rack->r_ctl.rc_sacked = qr.rack_sacked;
15287 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
15288 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
15289 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
15290 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
15291 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
15292 				if (qr.rack_tlp_out) {
15293 					rack->rc_tlp_in_progress = 1;
15294 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
15295 				} else {
15296 					rack->rc_tlp_in_progress = 0;
15297 					rack->r_ctl.rc_tlp_cnt_out = 0;
15298 				}
15299 				if (qr.rack_srtt_measured)
15300 					rack->rc_srtt_measure_made = 1;
15301 				if (qr.rack_in_persist == 1) {
15302 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
15303 #ifdef NETFLIX_SHARED_CWND
15304 					if (rack->r_ctl.rc_scw) {
15305 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
15306 						rack->rack_scwnd_is_idle = 1;
15307 					}
15308 #endif
15309 					rack->r_ctl.persist_lost_ends = 0;
15310 					rack->probe_not_answered = 0;
15311 					rack->forced_ack = 0;
15312 					tp->t_rxtshift = 0;
15313 					rack->rc_in_persist = 1;
15314 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
15315 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
15316 				}
15317 				if (qr.rack_wanted_output)
15318 					rack->r_wanted_output = 1;
15319 				rack_log_chg_info(tp, rack, 6,
15320 						  qr.rack_min_rtt,
15321 						  qr.rack_rtt,
15322 						  qr.rack_reorder_ts);
15323 			}
15324 			/* Get the old stack timers */
15325 			qr.req_param = 0;
15326 			qr.req = TCP_QUERY_TIMERS_UP;
15327 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15328 			if (ret) {
15329 				/*
15330 				 * non-zero return means we have a timer('s)
15331 				 * to start. Zero means no timer (no keepalive
15332 				 * I suppose).
15333 				 */
15334 				uint32_t tov = 0;
15335 
15336 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
15337 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
15338 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
15339 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
15340 						tov = qr.timer_pacing_to - us_cts;
15341 					else
15342 						tov = HPTS_TICKS_PER_SLOT;
15343 				}
15344 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
15345 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
15346 					if (tov == 0) {
15347 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
15348 							tov = qr.timer_timer_exp - us_cts;
15349 						else
15350 							tov = HPTS_TICKS_PER_SLOT;
15351 					}
15352 				}
15353 				rack_log_chg_info(tp, rack, 4,
15354 						  rack->r_ctl.rc_hpts_flags,
15355 						  rack->r_ctl.rc_last_output_to,
15356 						  rack->r_ctl.rc_timer_exp);
15357 				if (tov) {
15358 					struct hpts_diag diag;
15359 
15360 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15361 								   __LINE__, &diag);
15362 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15363 				}
15364 			}
15365 		}
15366 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15367 				     __LINE__, RACK_RTTS_INIT);
15368 	}
15369 	return (0);
15370 }
15371 
15372 static int
15373 rack_handoff_ok(struct tcpcb *tp)
15374 {
15375 	if ((tp->t_state == TCPS_CLOSED) ||
15376 	    (tp->t_state == TCPS_LISTEN)) {
15377 		/* Sure no problem though it may not stick */
15378 		return (0);
15379 	}
15380 	if ((tp->t_state == TCPS_SYN_SENT) ||
15381 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
15382 		/*
15383 		 * We really don't know if you support sack,
15384 		 * you have to get to ESTAB or beyond to tell.
15385 		 */
15386 		return (EAGAIN);
15387 	}
15388 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15389 		/*
15390 		 * Rack will only send a FIN after all data is acknowledged.
15391 		 * So in this case we have more data outstanding. We can't
15392 		 * switch stacks until either all data and only the FIN
15393 		 * is left (in which case rack_init() now knows how
15394 		 * to deal with that) <or> all is acknowledged and we
15395 		 * are only left with incoming data, though why you
15396 		 * would want to switch to rack after all data is acknowledged
15397 		 * I have no idea (rrs)!
15398 		 */
15399 		return (EAGAIN);
15400 	}
15401 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15402 		return (0);
15403 	}
15404 	/*
15405 	 * If we reach here we don't do SACK on this connection so we can
15406 	 * never do rack.
15407 	 */
15408 	return (EINVAL);
15409 }
15410 
15411 static void
15412 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15413 {
15414 
15415 	if (tp->t_fb_ptr) {
15416 		uint32_t cnt_free = 0;
15417 		struct tcp_rack *rack;
15418 		struct rack_sendmap *rsm;
15419 
15420 		tcp_handle_orphaned_packets(tp);
15421 		tp->t_flags &= ~TF_FORCEDATA;
15422 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15423 		rack_log_pacing_delay_calc(rack,
15424 					   0,
15425 					   0,
15426 					   0,
15427 					   rack_get_gp_est(rack), /* delRate */
15428 					   rack_get_lt_bw(rack), /* rttProp */
15429 					   20, __LINE__, NULL, 0);
15430 #ifdef NETFLIX_SHARED_CWND
15431 		if (rack->r_ctl.rc_scw) {
15432 			uint32_t limit;
15433 
15434 			if (rack->r_limit_scw)
15435 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15436 			else
15437 				limit = 0;
15438 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15439 						  rack->r_ctl.rc_scw_index,
15440 						  limit);
15441 			rack->r_ctl.rc_scw = NULL;
15442 		}
15443 #endif
15444 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15445 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15446 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15447 			rack->r_ctl.fsb.th = NULL;
15448 		}
15449 		if (rack->rc_always_pace) {
15450 			tcp_decrement_paced_conn();
15451 			rack_undo_cc_pacing(rack);
15452 			rack->rc_always_pace = 0;
15453 		}
15454 		/* Clean up any options if they were not applied */
15455 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15456 			struct deferred_opt_list *dol;
15457 
15458 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15459 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15460 			free(dol, M_TCPDO);
15461 		}
15462 		/* rack does not use force data but other stacks may clear it */
15463 		if (rack->r_ctl.crte != NULL) {
15464 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15465 			rack->rack_hdrw_pacing = 0;
15466 			rack->r_ctl.crte = NULL;
15467 		}
15468 #ifdef TCP_BLACKBOX
15469 		tcp_log_flowend(tp);
15470 #endif
15471 		/*
15472 		 * Lets take a different approach to purging just
15473 		 * get each one and free it like a cum-ack would and
15474 		 * not use a foreach loop.
15475 		 */
15476 		rsm = tqhash_min(rack->r_ctl.tqh);
15477 		while (rsm) {
15478 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15479 			rack->r_ctl.rc_num_maps_alloced--;
15480 			uma_zfree(rack_zone, rsm);
15481 			rsm = tqhash_min(rack->r_ctl.tqh);
15482 		}
15483 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15484 		while (rsm) {
15485 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15486 			rack->r_ctl.rc_num_maps_alloced--;
15487 			rack->rc_free_cnt--;
15488 			cnt_free++;
15489 			uma_zfree(rack_zone, rsm);
15490 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15491 		}
15492 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15493 		    (tcp_bblogging_on(tp))) {
15494 			union tcp_log_stackspecific log;
15495 			struct timeval tv;
15496 
15497 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15498 			log.u_bbr.flex8 = 10;
15499 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15500 			log.u_bbr.flex2 = rack->rc_free_cnt;
15501 			log.u_bbr.flex3 = cnt_free;
15502 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15503 			rsm = tqhash_min(rack->r_ctl.tqh);
15504 			log.u_bbr.delRate = (uint64_t)rsm;
15505 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15506 			log.u_bbr.cur_del_rate = (uint64_t)rsm;
15507 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15508 			log.u_bbr.pkt_epoch = __LINE__;
15509 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15510 					     0, &log, false, NULL, NULL, 0, &tv);
15511 		}
15512 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15513 			("rack:%p num_aloc:%u after freeing all?",
15514 			 rack,
15515 			 rack->r_ctl.rc_num_maps_alloced));
15516 		rack->rc_free_cnt = 0;
15517 		free(rack->r_ctl.tqh, M_TCPFSB);
15518 		rack->r_ctl.tqh = NULL;
15519 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15520 		tp->t_fb_ptr = NULL;
15521 	}
15522 	/* Make sure snd_nxt is correctly set */
15523 	tp->snd_nxt = tp->snd_max;
15524 }
15525 
15526 static void
15527 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15528 {
15529 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15530 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15531 	}
15532 	switch (tp->t_state) {
15533 	case TCPS_SYN_SENT:
15534 		rack->r_state = TCPS_SYN_SENT;
15535 		rack->r_substate = rack_do_syn_sent;
15536 		break;
15537 	case TCPS_SYN_RECEIVED:
15538 		rack->r_state = TCPS_SYN_RECEIVED;
15539 		rack->r_substate = rack_do_syn_recv;
15540 		break;
15541 	case TCPS_ESTABLISHED:
15542 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15543 		rack->r_state = TCPS_ESTABLISHED;
15544 		rack->r_substate = rack_do_established;
15545 		break;
15546 	case TCPS_CLOSE_WAIT:
15547 		rack->r_state = TCPS_CLOSE_WAIT;
15548 		rack->r_substate = rack_do_close_wait;
15549 		break;
15550 	case TCPS_FIN_WAIT_1:
15551 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15552 		rack->r_state = TCPS_FIN_WAIT_1;
15553 		rack->r_substate = rack_do_fin_wait_1;
15554 		break;
15555 	case TCPS_CLOSING:
15556 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15557 		rack->r_state = TCPS_CLOSING;
15558 		rack->r_substate = rack_do_closing;
15559 		break;
15560 	case TCPS_LAST_ACK:
15561 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15562 		rack->r_state = TCPS_LAST_ACK;
15563 		rack->r_substate = rack_do_lastack;
15564 		break;
15565 	case TCPS_FIN_WAIT_2:
15566 		rack->r_state = TCPS_FIN_WAIT_2;
15567 		rack->r_substate = rack_do_fin_wait_2;
15568 		break;
15569 	case TCPS_LISTEN:
15570 	case TCPS_CLOSED:
15571 	case TCPS_TIME_WAIT:
15572 	default:
15573 		break;
15574 	};
15575 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15576 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15577 
15578 }
15579 
15580 static void
15581 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15582 {
15583 	/*
15584 	 * We received an ack, and then did not
15585 	 * call send or were bounced out due to the
15586 	 * hpts was running. Now a timer is up as well, is
15587 	 * it the right timer?
15588 	 */
15589 	struct rack_sendmap *rsm;
15590 	int tmr_up;
15591 
15592 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15593 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15594 		return;
15595 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15596 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15597 	    (tmr_up == PACE_TMR_RXT)) {
15598 		/* Should be an RXT */
15599 		return;
15600 	}
15601 	if (rsm == NULL) {
15602 		/* Nothing outstanding? */
15603 		if (tp->t_flags & TF_DELACK) {
15604 			if (tmr_up == PACE_TMR_DELACK)
15605 				/* We are supposed to have delayed ack up and we do */
15606 				return;
15607 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15608 			/*
15609 			 * if we hit enobufs then we would expect the possibility
15610 			 * of nothing outstanding and the RXT up (and the hptsi timer).
15611 			 */
15612 			return;
15613 		} else if (((V_tcp_always_keepalive ||
15614 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15615 			    (tp->t_state <= TCPS_CLOSING)) &&
15616 			   (tmr_up == PACE_TMR_KEEP) &&
15617 			   (tp->snd_max == tp->snd_una)) {
15618 			/* We should have keep alive up and we do */
15619 			return;
15620 		}
15621 	}
15622 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15623 		   ((tmr_up == PACE_TMR_TLP) ||
15624 		    (tmr_up == PACE_TMR_RACK) ||
15625 		    (tmr_up == PACE_TMR_RXT))) {
15626 		/*
15627 		 * Either a Rack, TLP or RXT is fine if  we
15628 		 * have outstanding data.
15629 		 */
15630 		return;
15631 	} else if (tmr_up == PACE_TMR_DELACK) {
15632 		/*
15633 		 * If the delayed ack was going to go off
15634 		 * before the rtx/tlp/rack timer were going to
15635 		 * expire, then that would be the timer in control.
15636 		 * Note we don't check the time here trusting the
15637 		 * code is correct.
15638 		 */
15639 		return;
15640 	}
15641 	/*
15642 	 * Ok the timer originally started is not what we want now.
15643 	 * We will force the hpts to be stopped if any, and restart
15644 	 * with the slot set to what was in the saved slot.
15645 	 */
15646 	if (tcp_in_hpts(rack->rc_tp)) {
15647 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15648 			uint32_t us_cts;
15649 
15650 			us_cts = tcp_get_usecs(NULL);
15651 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15652 				rack->r_early = 1;
15653 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15654 			}
15655 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15656 		}
15657 		tcp_hpts_remove(rack->rc_tp);
15658 	}
15659 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15660 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15661 }
15662 
15663 
15664 static void
15665 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15666 {
15667 	if ((SEQ_LT(tp->snd_wl1, seq) ||
15668 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15669 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15670 		/* keep track of pure window updates */
15671 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15672 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15673 		tp->snd_wnd = tiwin;
15674 		rack_validate_fo_sendwin_up(tp, rack);
15675 		tp->snd_wl1 = seq;
15676 		tp->snd_wl2 = ack;
15677 		if (tp->snd_wnd > tp->max_sndwnd)
15678 			tp->max_sndwnd = tp->snd_wnd;
15679 	    rack->r_wanted_output = 1;
15680 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15681 		tp->snd_wnd = tiwin;
15682 		rack_validate_fo_sendwin_up(tp, rack);
15683 		tp->snd_wl1 = seq;
15684 		tp->snd_wl2 = ack;
15685 	} else {
15686 		/* Not a valid win update */
15687 		return;
15688 	}
15689 	if (tp->snd_wnd > tp->max_sndwnd)
15690 		tp->max_sndwnd = tp->snd_wnd;
15691 	/* Do we exit persists? */
15692 	if ((rack->rc_in_persist != 0) &&
15693 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15694 				rack->r_ctl.rc_pace_min_segs))) {
15695 		rack_exit_persist(tp, rack, cts);
15696 	}
15697 	/* Do we enter persists? */
15698 	if ((rack->rc_in_persist == 0) &&
15699 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15700 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
15701 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15702 	    sbavail(&tptosocket(tp)->so_snd) &&
15703 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15704 		/*
15705 		 * Here the rwnd is less than
15706 		 * the pacing size, we are established,
15707 		 * nothing is outstanding, and there is
15708 		 * data to send. Enter persists.
15709 		 */
15710 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15711 	}
15712 }
15713 
15714 static void
15715 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15716 {
15717 
15718 	if (tcp_bblogging_on(rack->rc_tp)) {
15719 		struct inpcb *inp = tptoinpcb(tp);
15720 		union tcp_log_stackspecific log;
15721 		struct timeval ltv;
15722 		char tcp_hdr_buf[60];
15723 		struct tcphdr *th;
15724 		struct timespec ts;
15725 		uint32_t orig_snd_una;
15726 		uint8_t xx = 0;
15727 
15728 #ifdef TCP_REQUEST_TRK
15729 		struct tcp_sendfile_track *tcp_req;
15730 
15731 		if (SEQ_GT(ae->ack, tp->snd_una)) {
15732 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15733 		} else {
15734 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15735 		}
15736 #endif
15737 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15738 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15739 		if (rack->rack_no_prr == 0)
15740 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15741 		else
15742 			log.u_bbr.flex1 = 0;
15743 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15744 		log.u_bbr.use_lt_bw <<= 1;
15745 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
15746 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15747 		log.u_bbr.bbr_state = rack->rc_free_cnt;
15748 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15749 		log.u_bbr.pkts_out = tp->t_maxseg;
15750 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15751 		log.u_bbr.flex7 = 1;
15752 		log.u_bbr.lost = ae->flags;
15753 		log.u_bbr.cwnd_gain = ackval;
15754 		log.u_bbr.pacing_gain = 0x2;
15755 		if (ae->flags & TSTMP_HDWR) {
15756 			/* Record the hardware timestamp if present */
15757 			log.u_bbr.flex3 = M_TSTMP;
15758 			ts.tv_sec = ae->timestamp / 1000000000;
15759 			ts.tv_nsec = ae->timestamp % 1000000000;
15760 			ltv.tv_sec = ts.tv_sec;
15761 			ltv.tv_usec = ts.tv_nsec / 1000;
15762 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15763 		} else if (ae->flags & TSTMP_LRO) {
15764 			/* Record the LRO the arrival timestamp */
15765 			log.u_bbr.flex3 = M_TSTMP_LRO;
15766 			ts.tv_sec = ae->timestamp / 1000000000;
15767 			ts.tv_nsec = ae->timestamp % 1000000000;
15768 			ltv.tv_sec = ts.tv_sec;
15769 			ltv.tv_usec = ts.tv_nsec / 1000;
15770 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15771 		}
15772 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15773 		/* Log the rcv time */
15774 		log.u_bbr.delRate = ae->timestamp;
15775 #ifdef TCP_REQUEST_TRK
15776 		log.u_bbr.applimited = tp->t_tcpreq_closed;
15777 		log.u_bbr.applimited <<= 8;
15778 		log.u_bbr.applimited |= tp->t_tcpreq_open;
15779 		log.u_bbr.applimited <<= 8;
15780 		log.u_bbr.applimited |= tp->t_tcpreq_req;
15781 		if (tcp_req) {
15782 			/* Copy out any client req info */
15783 			/* seconds */
15784 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15785 			/* useconds */
15786 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15787 			log.u_bbr.rttProp = tcp_req->timestamp;
15788 			log.u_bbr.cur_del_rate = tcp_req->start;
15789 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15790 				log.u_bbr.flex8 |= 1;
15791 			} else {
15792 				log.u_bbr.flex8 |= 2;
15793 				log.u_bbr.bw_inuse = tcp_req->end;
15794 			}
15795 			log.u_bbr.flex6 = tcp_req->start_seq;
15796 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15797 				log.u_bbr.flex8 |= 4;
15798 				log.u_bbr.epoch = tcp_req->end_seq;
15799 			}
15800 		}
15801 #endif
15802 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15803 		th = (struct tcphdr *)tcp_hdr_buf;
15804 		th->th_seq = ae->seq;
15805 		th->th_ack = ae->ack;
15806 		th->th_win = ae->win;
15807 		/* Now fill in the ports */
15808 		th->th_sport = inp->inp_fport;
15809 		th->th_dport = inp->inp_lport;
15810 		tcp_set_flags(th, ae->flags);
15811 		/* Now do we have a timestamp option? */
15812 		if (ae->flags & HAS_TSTMP) {
15813 			u_char *cp;
15814 			uint32_t val;
15815 
15816 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15817 			cp = (u_char *)(th + 1);
15818 			*cp = TCPOPT_NOP;
15819 			cp++;
15820 			*cp = TCPOPT_NOP;
15821 			cp++;
15822 			*cp = TCPOPT_TIMESTAMP;
15823 			cp++;
15824 			*cp = TCPOLEN_TIMESTAMP;
15825 			cp++;
15826 			val = htonl(ae->ts_value);
15827 			bcopy((char *)&val,
15828 			      (char *)cp, sizeof(uint32_t));
15829 			val = htonl(ae->ts_echo);
15830 			bcopy((char *)&val,
15831 			      (char *)(cp + 4), sizeof(uint32_t));
15832 		} else
15833 			th->th_off = (sizeof(struct tcphdr) >> 2);
15834 
15835 		/*
15836 		 * For sane logging we need to play a little trick.
15837 		 * If the ack were fully processed we would have moved
15838 		 * snd_una to high_seq, but since compressed acks are
15839 		 * processed in two phases, at this point (logging) snd_una
15840 		 * won't be advanced. So we would see multiple acks showing
15841 		 * the advancement. We can prevent that by "pretending" that
15842 		 * snd_una was advanced and then un-advancing it so that the
15843 		 * logging code has the right value for tlb_snd_una.
15844 		 */
15845 		if (tp->snd_una != high_seq) {
15846 			orig_snd_una = tp->snd_una;
15847 			tp->snd_una = high_seq;
15848 			xx = 1;
15849 		} else
15850 			xx = 0;
15851 		TCP_LOG_EVENTP(tp, th,
15852 			       &tptosocket(tp)->so_rcv,
15853 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15854 			       0, &log, true, &ltv);
15855 		if (xx) {
15856 			tp->snd_una = orig_snd_una;
15857 		}
15858 	}
15859 
15860 }
15861 
15862 static void
15863 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15864 {
15865 	uint32_t us_rtt;
15866 	/*
15867 	 * A persist or keep-alive was forced out, update our
15868 	 * min rtt time. Note now worry about lost responses.
15869 	 * When a subsequent keep-alive or persist times out
15870 	 * and forced_ack is still on, then the last probe
15871 	 * was not responded to. In such cases we have a
15872 	 * sysctl that controls the behavior. Either we apply
15873 	 * the rtt but with reduced confidence (0). Or we just
15874 	 * plain don't apply the rtt estimate. Having data flow
15875 	 * will clear the probe_not_answered flag i.e. cum-ack
15876 	 * move forward <or> exiting and reentering persists.
15877 	 */
15878 
15879 	rack->forced_ack = 0;
15880 	rack->rc_tp->t_rxtshift = 0;
15881 	if ((rack->rc_in_persist &&
15882 	     (tiwin == rack->rc_tp->snd_wnd)) ||
15883 	    (rack->rc_in_persist == 0)) {
15884 		/*
15885 		 * In persists only apply the RTT update if this is
15886 		 * a response to our window probe. And that
15887 		 * means the rwnd sent must match the current
15888 		 * snd_wnd. If it does not, then we got a
15889 		 * window update ack instead. For keepalive
15890 		 * we allow the answer no matter what the window.
15891 		 *
15892 		 * Note that if the probe_not_answered is set then
15893 		 * the forced_ack_ts is the oldest one i.e. the first
15894 		 * probe sent that might have been lost. This assures
15895 		 * us that if we do calculate an RTT it is longer not
15896 		 * some short thing.
15897 		 */
15898 		if (rack->rc_in_persist)
15899 			counter_u64_add(rack_persists_acks, 1);
15900 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15901 		if (us_rtt == 0)
15902 			us_rtt = 1;
15903 		if (rack->probe_not_answered == 0) {
15904 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15905 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15906 		} else {
15907 			/* We have a retransmitted probe here too */
15908 			if (rack_apply_rtt_with_reduced_conf) {
15909 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15910 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15911 			}
15912 		}
15913 	}
15914 }
15915 
15916 static int
15917 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15918 {
15919 	/*
15920 	 * Handle a "special" compressed ack mbuf. Each incoming
15921 	 * ack has only four possible dispositions:
15922 	 *
15923 	 * A) It moves the cum-ack forward
15924 	 * B) It is behind the cum-ack.
15925 	 * C) It is a window-update ack.
15926 	 * D) It is a dup-ack.
15927 	 *
15928 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15929 	 * in the incoming mbuf. We also need to still pay attention
15930 	 * to nxt_pkt since there may be another packet after this
15931 	 * one.
15932 	 */
15933 #ifdef TCP_ACCOUNTING
15934 	uint64_t ts_val;
15935 	uint64_t rdstc;
15936 #endif
15937 	int segsiz;
15938 	struct timespec ts;
15939 	struct tcp_rack *rack;
15940 	struct tcp_ackent *ae;
15941 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15942 	int cnt, i, did_out, ourfinisacked = 0;
15943 	struct tcpopt to_holder, *to = NULL;
15944 #ifdef TCP_ACCOUNTING
15945 	int win_up_req = 0;
15946 #endif
15947 	int nsegs = 0;
15948 	int under_pacing = 0;
15949 	int recovery = 0;
15950 #ifdef TCP_ACCOUNTING
15951 	sched_pin();
15952 #endif
15953 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15954 	if (rack->gp_ready &&
15955 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15956 		under_pacing = 1;
15957 
15958 	if (rack->r_state != tp->t_state)
15959 		rack_set_state(tp, rack);
15960 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15961 	    (tp->t_flags & TF_GPUTINPROG)) {
15962 		/*
15963 		 * We have a goodput in progress
15964 		 * and we have entered a late state.
15965 		 * Do we have enough data in the sb
15966 		 * to handle the GPUT request?
15967 		 */
15968 		uint32_t bytes;
15969 
15970 		bytes = tp->gput_ack - tp->gput_seq;
15971 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
15972 			bytes += tp->gput_seq - tp->snd_una;
15973 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15974 			/*
15975 			 * There are not enough bytes in the socket
15976 			 * buffer that have been sent to cover this
15977 			 * measurement. Cancel it.
15978 			 */
15979 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15980 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
15981 						   tp->gput_seq,
15982 						   0, 0, 18, __LINE__, NULL, 0);
15983 			tp->t_flags &= ~TF_GPUTINPROG;
15984 		}
15985 	}
15986 	to = &to_holder;
15987 	to->to_flags = 0;
15988 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15989 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15990 	cnt = m->m_len / sizeof(struct tcp_ackent);
15991 	counter_u64_add(rack_multi_single_eq, cnt);
15992 	high_seq = tp->snd_una;
15993 	the_win = tp->snd_wnd;
15994 	win_seq = tp->snd_wl1;
15995 	win_upd_ack = tp->snd_wl2;
15996 	cts = tcp_tv_to_usectick(tv);
15997 	ms_cts = tcp_tv_to_mssectick(tv);
15998 	rack->r_ctl.rc_rcvtime = cts;
15999 	segsiz = ctf_fixed_maxseg(tp);
16000 	if ((rack->rc_gp_dyn_mul) &&
16001 	    (rack->use_fixed_rate == 0) &&
16002 	    (rack->rc_always_pace)) {
16003 		/* Check in on probertt */
16004 		rack_check_probe_rtt(rack, cts);
16005 	}
16006 	for (i = 0; i < cnt; i++) {
16007 #ifdef TCP_ACCOUNTING
16008 		ts_val = get_cyclecount();
16009 #endif
16010 		rack_clear_rate_sample(rack);
16011 		ae = ((mtod(m, struct tcp_ackent *)) + i);
16012 		if (ae->flags & TH_FIN)
16013 			rack_log_pacing_delay_calc(rack,
16014 						   0,
16015 						   0,
16016 						   0,
16017 						   rack_get_gp_est(rack), /* delRate */
16018 						   rack_get_lt_bw(rack), /* rttProp */
16019 						   20, __LINE__, NULL, 0);
16020 		/* Setup the window */
16021 		tiwin = ae->win << tp->snd_scale;
16022 		if (tiwin > rack->r_ctl.rc_high_rwnd)
16023 			rack->r_ctl.rc_high_rwnd = tiwin;
16024 		/* figure out the type of ack */
16025 		if (SEQ_LT(ae->ack, high_seq)) {
16026 			/* Case B*/
16027 			ae->ack_val_set = ACK_BEHIND;
16028 		} else if (SEQ_GT(ae->ack, high_seq)) {
16029 			/* Case A */
16030 			ae->ack_val_set = ACK_CUMACK;
16031 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
16032 			/* Case D */
16033 			ae->ack_val_set = ACK_DUPACK;
16034 		} else {
16035 			/* Case C */
16036 			ae->ack_val_set = ACK_RWND;
16037 		}
16038 		if (rack->sack_attack_disable > 0) {
16039 			rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16040 			rack->r_ctl.ack_during_sd++;
16041 		}
16042 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
16043 		/* Validate timestamp */
16044 		if (ae->flags & HAS_TSTMP) {
16045 			/* Setup for a timestamp */
16046 			to->to_flags = TOF_TS;
16047 			ae->ts_echo -= tp->ts_offset;
16048 			to->to_tsecr = ae->ts_echo;
16049 			to->to_tsval = ae->ts_value;
16050 			/*
16051 			 * If echoed timestamp is later than the current time, fall back to
16052 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16053 			 * were used when this connection was established.
16054 			 */
16055 			if (TSTMP_GT(ae->ts_echo, ms_cts))
16056 				to->to_tsecr = 0;
16057 			if (tp->ts_recent &&
16058 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
16059 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
16060 #ifdef TCP_ACCOUNTING
16061 					rdstc = get_cyclecount();
16062 					if (rdstc > ts_val) {
16063 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16064 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16065 						}
16066 					}
16067 #endif
16068 					continue;
16069 				}
16070 			}
16071 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
16072 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
16073 				tp->ts_recent_age = tcp_ts_getticks();
16074 				tp->ts_recent = ae->ts_value;
16075 			}
16076 		} else {
16077 			/* Setup for a no options */
16078 			to->to_flags = 0;
16079 		}
16080 		/* Update the rcv time and perform idle reduction possibly */
16081 		if  (tp->t_idle_reduce &&
16082 		     (tp->snd_max == tp->snd_una) &&
16083 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16084 			counter_u64_add(rack_input_idle_reduces, 1);
16085 			rack_cc_after_idle(rack, tp);
16086 		}
16087 		tp->t_rcvtime = ticks;
16088 		/* Now what about ECN of a chain of pure ACKs? */
16089 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
16090 			tcp_packets_this_ack(tp, ae->ack),
16091 			ae->codepoint))
16092 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
16093 #ifdef TCP_ACCOUNTING
16094 		/* Count for the specific type of ack in */
16095 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16096 			tp->tcp_cnt_counters[ae->ack_val_set]++;
16097 		}
16098 #endif
16099 		/*
16100 		 * Note how we could move up these in the determination
16101 		 * above, but we don't so that way the timestamp checks (and ECN)
16102 		 * is done first before we do any processing on the ACK.
16103 		 * The non-compressed path through the code has this
16104 		 * weakness (noted by @jtl) that it actually does some
16105 		 * processing before verifying the timestamp information.
16106 		 * We don't take that path here which is why we set
16107 		 * the ack_val_set first, do the timestamp and ecn
16108 		 * processing, and then look at what we have setup.
16109 		 */
16110 		if (ae->ack_val_set == ACK_BEHIND) {
16111 			/*
16112 			 * Case B flag reordering, if window is not closed
16113 			 * or it could be a keep-alive or persists
16114 			 */
16115 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
16116 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16117 				if (rack->r_ctl.rc_reorder_ts == 0)
16118 					rack->r_ctl.rc_reorder_ts = 1;
16119 			}
16120 		} else if (ae->ack_val_set == ACK_DUPACK) {
16121 			/* Case D */
16122 			rack_strike_dupack(rack);
16123 		} else if (ae->ack_val_set == ACK_RWND) {
16124 			/* Case C */
16125 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16126 				ts.tv_sec = ae->timestamp / 1000000000;
16127 				ts.tv_nsec = ae->timestamp % 1000000000;
16128 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16129 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16130 			} else {
16131 				rack->r_ctl.act_rcv_time = *tv;
16132 			}
16133 			if (rack->forced_ack) {
16134 				rack_handle_probe_response(rack, tiwin,
16135 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
16136 			}
16137 #ifdef TCP_ACCOUNTING
16138 			win_up_req = 1;
16139 #endif
16140 			win_upd_ack = ae->ack;
16141 			win_seq = ae->seq;
16142 			the_win = tiwin;
16143 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16144 		} else {
16145 			/* Case A */
16146 			if (SEQ_GT(ae->ack, tp->snd_max)) {
16147 				/*
16148 				 * We just send an ack since the incoming
16149 				 * ack is beyond the largest seq we sent.
16150 				 */
16151 				if ((tp->t_flags & TF_ACKNOW) == 0) {
16152 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
16153 					if (tp->t_flags && TF_ACKNOW)
16154 						rack->r_wanted_output = 1;
16155 				}
16156 			} else {
16157 				nsegs++;
16158 				/* If the window changed setup to update */
16159 				if (tiwin != tp->snd_wnd) {
16160 					win_upd_ack = ae->ack;
16161 					win_seq = ae->seq;
16162 					the_win = tiwin;
16163 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16164 				}
16165 #ifdef TCP_ACCOUNTING
16166 				/* Account for the acks */
16167 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16168 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
16169 				}
16170 #endif
16171 				high_seq = ae->ack;
16172 				if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp))
16173 					rack_log_hystart_event(rack, high_seq, 8);
16174 				/* Setup our act_rcv_time */
16175 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16176 					ts.tv_sec = ae->timestamp / 1000000000;
16177 					ts.tv_nsec = ae->timestamp % 1000000000;
16178 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16179 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16180 				} else {
16181 					rack->r_ctl.act_rcv_time = *tv;
16182 				}
16183 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
16184 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
16185 #ifdef TCP_REQUEST_TRK
16186 				rack_req_check_for_comp(rack, high_seq);
16187 #endif
16188 				if (rack->rc_dsack_round_seen) {
16189 					/* Is the dsack round over? */
16190 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
16191 						/* Yes it is */
16192 						rack->rc_dsack_round_seen = 0;
16193 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
16194 					}
16195 				}
16196 			}
16197 		}
16198 		/* And lets be sure to commit the rtt measurements for this ack */
16199 		tcp_rack_xmit_timer_commit(rack, tp);
16200 #ifdef TCP_ACCOUNTING
16201 		rdstc = get_cyclecount();
16202 		if (rdstc > ts_val) {
16203 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16204 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16205 				if (ae->ack_val_set == ACK_CUMACK)
16206 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
16207 			}
16208 		}
16209 #endif
16210 	}
16211 #ifdef TCP_ACCOUNTING
16212 	ts_val = get_cyclecount();
16213 #endif
16214 	/* Tend to any collapsed window */
16215 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
16216 		/* The peer collapsed the window */
16217 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
16218 	} else if (rack->rc_has_collapsed)
16219 		rack_un_collapse_window(rack, __LINE__);
16220 	if ((rack->r_collapse_point_valid) &&
16221 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16222 		rack->r_collapse_point_valid = 0;
16223 	acked_amount = acked = (high_seq - tp->snd_una);
16224 	if (acked) {
16225 		/*
16226 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16227 		 * causes issues when we are just going app limited. Lets
16228 		 * instead use SEQ_GT <or> where its equal but more data
16229 		 * is outstanding.
16230 		 *
16231 		 * Also make sure we are on the last ack of a series. We
16232 		 * have to have all the ack's processed in queue to know
16233 		 * if there is something left outstanding.
16234 		 *
16235 		 */
16236 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16237 		    (rack->rc_new_rnd_needed == 0) &&
16238 		    (nxt_pkt == 0)) {
16239 			rack_log_hystart_event(rack, high_seq, 21);
16240 			rack->r_ctl.current_round++;
16241 			/* Force the next send to setup the next round */
16242 			rack->rc_new_rnd_needed = 1;
16243 			if (CC_ALGO(tp)->newround != NULL) {
16244 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
16245 			}
16246 		}
16247 		/*
16248 		 * Clear the probe not answered flag
16249 		 * since cum-ack moved forward.
16250 		 */
16251 		rack->probe_not_answered = 0;
16252 		if (rack->sack_attack_disable == 0)
16253 			rack_do_decay(rack);
16254 		if (acked >= segsiz) {
16255 			/*
16256 			 * You only get credit for
16257 			 * MSS and greater (and you get extra
16258 			 * credit for larger cum-ack moves).
16259 			 */
16260 			int ac;
16261 
16262 			ac = acked / segsiz;
16263 			rack->r_ctl.ack_count += ac;
16264 			counter_u64_add(rack_ack_total, ac);
16265 		}
16266 		if (rack->r_ctl.ack_count > 0xfff00000) {
16267 			/*
16268 			 * reduce the number to keep us under
16269 			 * a uint32_t.
16270 			 */
16271 			rack->r_ctl.ack_count /= 2;
16272 			rack->r_ctl.sack_count /= 2;
16273 		}
16274 		if (tp->t_flags & TF_NEEDSYN) {
16275 			/*
16276 			 * T/TCP: Connection was half-synchronized, and our SYN has
16277 			 * been ACK'd (so connection is now fully synchronized).  Go
16278 			 * to non-starred state, increment snd_una for ACK of SYN,
16279 			 * and check if we can do window scaling.
16280 			 */
16281 			tp->t_flags &= ~TF_NEEDSYN;
16282 			tp->snd_una++;
16283 			acked_amount = acked = (high_seq - tp->snd_una);
16284 		}
16285 		if (acked > sbavail(&so->so_snd))
16286 			acked_amount = sbavail(&so->so_snd);
16287 #ifdef TCP_SAD_DETECTION
16288 		/*
16289 		 * We only care on a cum-ack move if we are in a sack-disabled
16290 		 * state. We have already added in to the ack_count, and we never
16291 		 * would disable on a cum-ack move, so we only care to do the
16292 		 * detection if it may "undo" it, i.e. we were in disabled already.
16293 		 */
16294 		if (rack->sack_attack_disable)
16295 			rack_do_detection(tp, rack, acked_amount, segsiz);
16296 #endif
16297 		if (IN_FASTRECOVERY(tp->t_flags) &&
16298 		    (rack->rack_no_prr == 0))
16299 			rack_update_prr(tp, rack, acked_amount, high_seq);
16300 		if (IN_RECOVERY(tp->t_flags)) {
16301 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16302 			    (SEQ_LT(high_seq, tp->snd_max))) {
16303 				tcp_rack_partialack(tp);
16304 			} else {
16305 				rack_post_recovery(tp, high_seq);
16306 				recovery = 1;
16307 			}
16308 		}
16309 		/* Handle the rack-log-ack part (sendmap) */
16310 		if ((sbused(&so->so_snd) == 0) &&
16311 		    (acked > acked_amount) &&
16312 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16313 		    (tp->t_flags & TF_SENTFIN)) {
16314 			/*
16315 			 * We must be sure our fin
16316 			 * was sent and acked (we can be
16317 			 * in FIN_WAIT_1 without having
16318 			 * sent the fin).
16319 			 */
16320 			ourfinisacked = 1;
16321 			/*
16322 			 * Lets make sure snd_una is updated
16323 			 * since most likely acked_amount = 0 (it
16324 			 * should be).
16325 			 */
16326 			tp->snd_una = high_seq;
16327 		}
16328 		/* Did we make a RTO error? */
16329 		if ((tp->t_flags & TF_PREVVALID) &&
16330 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16331 			tp->t_flags &= ~TF_PREVVALID;
16332 			if (tp->t_rxtshift == 1 &&
16333 			    (int)(ticks - tp->t_badrxtwin) < 0)
16334 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16335 		}
16336 		/* Handle the data in the socket buffer */
16337 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16338 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16339 		if (acked_amount > 0) {
16340 			struct mbuf *mfree;
16341 
16342 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
16343 			SOCKBUF_LOCK(&so->so_snd);
16344 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16345 			tp->snd_una = high_seq;
16346 			/* Note we want to hold the sb lock through the sendmap adjust */
16347 			rack_adjust_sendmap_head(rack, &so->so_snd);
16348 			/* Wake up the socket if we have room to write more */
16349 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16350 			sowwakeup_locked(so);
16351 			if ((recovery == 1) &&
16352 			    (rack->excess_rxt_on) &&
16353 			    (rack->r_cwnd_was_clamped == 0)) {
16354 				do_rack_excess_rxt(tp, rack);
16355 			} else if (rack->r_cwnd_was_clamped)
16356 				do_rack_check_for_unclamp(tp, rack);
16357 			m_freem(mfree);
16358 		}
16359 		/* update progress */
16360 		tp->t_acktime = ticks;
16361 		rack_log_progress_event(rack, tp, tp->t_acktime,
16362 					PROGRESS_UPDATE, __LINE__);
16363 		/* Clear out shifts and such */
16364 		tp->t_rxtshift = 0;
16365 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16366 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16367 		rack->rc_tlp_in_progress = 0;
16368 		rack->r_ctl.rc_tlp_cnt_out = 0;
16369 		/* Send recover and snd_nxt must be dragged along */
16370 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16371 			tp->snd_recover = tp->snd_una;
16372 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16373 			tp->snd_nxt = tp->snd_max;
16374 		/*
16375 		 * If the RXT timer is running we want to
16376 		 * stop it, so we can restart a TLP (or new RXT).
16377 		 */
16378 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16379 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16380 		tp->snd_wl2 = high_seq;
16381 		tp->t_dupacks = 0;
16382 		if (under_pacing &&
16383 		    (rack->use_fixed_rate == 0) &&
16384 		    (rack->in_probe_rtt == 0) &&
16385 		    rack->rc_gp_dyn_mul &&
16386 		    rack->rc_always_pace) {
16387 			/* Check if we are dragging bottom */
16388 			rack_check_bottom_drag(tp, rack, so);
16389 		}
16390 		if (tp->snd_una == tp->snd_max) {
16391 			tp->t_flags &= ~TF_PREVVALID;
16392 			rack->r_ctl.retran_during_recovery = 0;
16393 			rack->rc_suspicious = 0;
16394 			rack->r_ctl.dsack_byte_cnt = 0;
16395 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16396 			if (rack->r_ctl.rc_went_idle_time == 0)
16397 				rack->r_ctl.rc_went_idle_time = 1;
16398 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16399 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16400 				tp->t_acktime = 0;
16401 			/* Set so we might enter persists... */
16402 			rack->r_wanted_output = 1;
16403 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16404 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16405 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16406 			    (sbavail(&so->so_snd) == 0) &&
16407 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16408 				/*
16409 				 * The socket was gone and the
16410 				 * peer sent data (not now in the past), time to
16411 				 * reset him.
16412 				 */
16413 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16414 				/* tcp_close will kill the inp pre-log the Reset */
16415 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16416 #ifdef TCP_ACCOUNTING
16417 				rdstc = get_cyclecount();
16418 				if (rdstc > ts_val) {
16419 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16420 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16421 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16422 					}
16423 				}
16424 #endif
16425 				m_freem(m);
16426 				tp = tcp_close(tp);
16427 				if (tp == NULL) {
16428 #ifdef TCP_ACCOUNTING
16429 					sched_unpin();
16430 #endif
16431 					return (1);
16432 				}
16433 				/*
16434 				 * We would normally do drop-with-reset which would
16435 				 * send back a reset. We can't since we don't have
16436 				 * all the needed bits. Instead lets arrange for
16437 				 * a call to tcp_output(). That way since we
16438 				 * are in the closed state we will generate a reset.
16439 				 *
16440 				 * Note if tcp_accounting is on we don't unpin since
16441 				 * we do that after the goto label.
16442 				 */
16443 				goto send_out_a_rst;
16444 			}
16445 			if ((sbused(&so->so_snd) == 0) &&
16446 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16447 			    (tp->t_flags & TF_SENTFIN)) {
16448 				/*
16449 				 * If we can't receive any more data, then closing user can
16450 				 * proceed. Starting the timer is contrary to the
16451 				 * specification, but if we don't get a FIN we'll hang
16452 				 * forever.
16453 				 *
16454 				 */
16455 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16456 					soisdisconnected(so);
16457 					tcp_timer_activate(tp, TT_2MSL,
16458 							   (tcp_fast_finwait2_recycle ?
16459 							    tcp_finwait2_timeout :
16460 							    TP_MAXIDLE(tp)));
16461 				}
16462 				if (ourfinisacked == 0) {
16463 					/*
16464 					 * We don't change to fin-wait-2 if we have our fin acked
16465 					 * which means we are probably in TCPS_CLOSING.
16466 					 */
16467 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16468 				}
16469 			}
16470 		}
16471 		/* Wake up the socket if we have room to write more */
16472 		if (sbavail(&so->so_snd)) {
16473 			rack->r_wanted_output = 1;
16474 			if (ctf_progress_timeout_check(tp, true)) {
16475 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16476 							tp, tick, PROGRESS_DROP, __LINE__);
16477 				/*
16478 				 * We cheat here and don't send a RST, we should send one
16479 				 * when the pacer drops the connection.
16480 				 */
16481 #ifdef TCP_ACCOUNTING
16482 				rdstc = get_cyclecount();
16483 				if (rdstc > ts_val) {
16484 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16485 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16486 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16487 					}
16488 				}
16489 				sched_unpin();
16490 #endif
16491 				(void)tcp_drop(tp, ETIMEDOUT);
16492 				m_freem(m);
16493 				return (1);
16494 			}
16495 		}
16496 		if (ourfinisacked) {
16497 			switch(tp->t_state) {
16498 			case TCPS_CLOSING:
16499 #ifdef TCP_ACCOUNTING
16500 				rdstc = get_cyclecount();
16501 				if (rdstc > ts_val) {
16502 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16503 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16504 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16505 					}
16506 				}
16507 				sched_unpin();
16508 #endif
16509 				tcp_twstart(tp);
16510 				m_freem(m);
16511 				return (1);
16512 				break;
16513 			case TCPS_LAST_ACK:
16514 #ifdef TCP_ACCOUNTING
16515 				rdstc = get_cyclecount();
16516 				if (rdstc > ts_val) {
16517 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16518 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16519 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16520 					}
16521 				}
16522 				sched_unpin();
16523 #endif
16524 				tp = tcp_close(tp);
16525 				ctf_do_drop(m, tp);
16526 				return (1);
16527 				break;
16528 			case TCPS_FIN_WAIT_1:
16529 #ifdef TCP_ACCOUNTING
16530 				rdstc = get_cyclecount();
16531 				if (rdstc > ts_val) {
16532 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16533 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16534 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16535 					}
16536 				}
16537 #endif
16538 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16539 					soisdisconnected(so);
16540 					tcp_timer_activate(tp, TT_2MSL,
16541 							   (tcp_fast_finwait2_recycle ?
16542 							    tcp_finwait2_timeout :
16543 							    TP_MAXIDLE(tp)));
16544 				}
16545 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
16546 				break;
16547 			default:
16548 				break;
16549 			}
16550 		}
16551 		if (rack->r_fast_output) {
16552 			/*
16553 			 * We re doing fast output.. can we expand that?
16554 			 */
16555 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16556 		}
16557 #ifdef TCP_ACCOUNTING
16558 		rdstc = get_cyclecount();
16559 		if (rdstc > ts_val) {
16560 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16561 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16562 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16563 			}
16564 		}
16565 
16566 	} else if (win_up_req) {
16567 		rdstc = get_cyclecount();
16568 		if (rdstc > ts_val) {
16569 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16570 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16571 			}
16572 		}
16573 #endif
16574 	}
16575 	/* Now is there a next packet, if so we are done */
16576 	m_freem(m);
16577 	did_out = 0;
16578 	if (nxt_pkt) {
16579 #ifdef TCP_ACCOUNTING
16580 		sched_unpin();
16581 #endif
16582 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16583 		return (0);
16584 	}
16585 	rack_handle_might_revert(tp, rack);
16586 	ctf_calc_rwin(so, tp);
16587 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16588 	send_out_a_rst:
16589 		if (tcp_output(tp) < 0) {
16590 #ifdef TCP_ACCOUNTING
16591 			sched_unpin();
16592 #endif
16593 			return (1);
16594 		}
16595 		did_out = 1;
16596 	}
16597 	if (tp->t_flags2 & TF2_HPTS_CALLS)
16598 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
16599 	rack_free_trim(rack);
16600 #ifdef TCP_ACCOUNTING
16601 	sched_unpin();
16602 #endif
16603 	rack_timer_audit(tp, rack, &so->so_snd);
16604 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16605 	return (0);
16606 }
16607 
16608 #define	TCP_LRO_TS_OPTION \
16609     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16610 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16611 
16612 static int
16613 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16614     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16615     struct timeval *tv)
16616 {
16617 	struct inpcb *inp = tptoinpcb(tp);
16618 	struct socket *so = tptosocket(tp);
16619 #ifdef TCP_ACCOUNTING
16620 	uint64_t ts_val;
16621 #endif
16622 	int32_t thflags, retval, did_out = 0;
16623 	int32_t way_out = 0;
16624 	/*
16625 	 * cts - is the current time from tv (caller gets ts) in microseconds.
16626 	 * ms_cts - is the current time from tv in milliseconds.
16627 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16628 	 */
16629 	uint32_t cts, us_cts, ms_cts;
16630 	uint32_t tiwin, high_seq;
16631 	struct timespec ts;
16632 	struct tcpopt to;
16633 	struct tcp_rack *rack;
16634 	struct rack_sendmap *rsm;
16635 	int32_t prev_state = 0;
16636 	int no_output = 0;
16637 	int slot_remaining = 0;
16638 #ifdef TCP_ACCOUNTING
16639 	int ack_val_set = 0xf;
16640 #endif
16641 	int nsegs;
16642 
16643 	NET_EPOCH_ASSERT();
16644 	INP_WLOCK_ASSERT(inp);
16645 
16646 	/*
16647 	 * tv passed from common code is from either M_TSTMP_LRO or
16648 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16649 	 */
16650 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16651 	if (rack->rack_deferred_inited == 0) {
16652 		/*
16653 		 * If we are the connecting socket we will
16654 		 * hit rack_init() when no sequence numbers
16655 		 * are setup. This makes it so we must defer
16656 		 * some initialization. Call that now.
16657 		 */
16658 		rack_deferred_init(tp, rack);
16659 	}
16660 	/*
16661 	 * Check to see if we need to skip any output plans. This
16662 	 * can happen in the non-LRO path where we are pacing and
16663 	 * must process the ack coming in but need to defer sending
16664 	 * anything becase a pacing timer is running.
16665 	 */
16666 	us_cts = tcp_tv_to_usectick(tv);
16667 	if (m->m_flags & M_ACKCMP) {
16668 		/*
16669 		 * All compressed ack's are ack's by definition so
16670 		 * remove any ack required flag and then do the processing.
16671 		 */
16672 		rack->rc_ack_required = 0;
16673 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16674 	}
16675 	thflags = tcp_get_flags(th);
16676 	if ((rack->rc_always_pace == 1) &&
16677 	    (rack->rc_ack_can_sendout_data == 0) &&
16678 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16679 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16680 		/*
16681 		 * Ok conditions are right for queuing the packets
16682 		 * but we do have to check the flags in the inp, it
16683 		 * could be, if a sack is present, we want to be awoken and
16684 		 * so should process the packets.
16685 		 */
16686 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16687 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16688 			no_output = 1;
16689 		} else {
16690 			/*
16691 			 * If there is no options, or just a
16692 			 * timestamp option, we will want to queue
16693 			 * the packets. This is the same that LRO does
16694 			 * and will need to change with accurate ECN.
16695 			 */
16696 			uint32_t *ts_ptr;
16697 			int optlen;
16698 
16699 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16700 			ts_ptr = (uint32_t *)(th + 1);
16701 			if ((optlen == 0) ||
16702 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
16703 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
16704 				no_output = 1;
16705 		}
16706 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16707 			/*
16708 			 * It is unrealistic to think we can pace in less than
16709 			 * the minimum granularity of the pacer (def:250usec). So
16710 			 * if we have less than that time remaining we should go
16711 			 * ahead and allow output to be "early". We will attempt to
16712 			 * make up for it in any pacing time we try to apply on
16713 			 * the outbound packet.
16714 			 */
16715 			no_output = 0;
16716 		}
16717 	}
16718 	/*
16719 	 * If there is a RST or FIN lets dump out the bw
16720 	 * with a FIN the connection may go on but we
16721 	 * may not.
16722 	 */
16723 	if ((thflags & TH_FIN) || (thflags & TH_RST))
16724 		rack_log_pacing_delay_calc(rack,
16725 					   rack->r_ctl.gp_bw,
16726 					   0,
16727 					   0,
16728 					   rack_get_gp_est(rack), /* delRate */
16729 					   rack_get_lt_bw(rack), /* rttProp */
16730 					   20, __LINE__, NULL, 0);
16731 	if (m->m_flags & M_ACKCMP) {
16732 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16733 	}
16734 	cts = tcp_tv_to_usectick(tv);
16735 	ms_cts =  tcp_tv_to_mssectick(tv);
16736 	nsegs = m->m_pkthdr.lro_nsegs;
16737 	counter_u64_add(rack_proc_non_comp_ack, 1);
16738 #ifdef TCP_ACCOUNTING
16739 	sched_pin();
16740 	if (thflags & TH_ACK)
16741 		ts_val = get_cyclecount();
16742 #endif
16743 	if ((m->m_flags & M_TSTMP) ||
16744 	    (m->m_flags & M_TSTMP_LRO)) {
16745 		mbuf_tstmp2timespec(m, &ts);
16746 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16747 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16748 	} else
16749 		rack->r_ctl.act_rcv_time = *tv;
16750 	kern_prefetch(rack, &prev_state);
16751 	prev_state = 0;
16752 	/*
16753 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
16754 	 * the scale is zero.
16755 	 */
16756 	tiwin = th->th_win << tp->snd_scale;
16757 #ifdef TCP_ACCOUNTING
16758 	if (thflags & TH_ACK) {
16759 		/*
16760 		 * We have a tradeoff here. We can either do what we are
16761 		 * doing i.e. pinning to this CPU and then doing the accounting
16762 		 * <or> we could do a critical enter, setup the rdtsc and cpu
16763 		 * as in below, and then validate we are on the same CPU on
16764 		 * exit. I have choosen to not do the critical enter since
16765 		 * that often will gain you a context switch, and instead lock
16766 		 * us (line above this if) to the same CPU with sched_pin(). This
16767 		 * means we may be context switched out for a higher priority
16768 		 * interupt but we won't be moved to another CPU.
16769 		 *
16770 		 * If this occurs (which it won't very often since we most likely
16771 		 * are running this code in interupt context and only a higher
16772 		 * priority will bump us ... clock?) we will falsely add in
16773 		 * to the time the interupt processing time plus the ack processing
16774 		 * time. This is ok since its a rare event.
16775 		 */
16776 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16777 						    ctf_fixed_maxseg(tp));
16778 	}
16779 #endif
16780 	/*
16781 	 * Parse options on any incoming segment.
16782 	 */
16783 	memset(&to, 0, sizeof(to));
16784 	tcp_dooptions(&to, (u_char *)(th + 1),
16785 	    (th->th_off << 2) - sizeof(struct tcphdr),
16786 	    (thflags & TH_SYN) ? TO_SYN : 0);
16787 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16788 	    __func__));
16789 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16790 	    __func__));
16791 
16792 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16793 	    (tp->t_flags & TF_GPUTINPROG)) {
16794 		/*
16795 		 * We have a goodput in progress
16796 		 * and we have entered a late state.
16797 		 * Do we have enough data in the sb
16798 		 * to handle the GPUT request?
16799 		 */
16800 		uint32_t bytes;
16801 
16802 		bytes = tp->gput_ack - tp->gput_seq;
16803 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16804 			bytes += tp->gput_seq - tp->snd_una;
16805 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16806 			/*
16807 			 * There are not enough bytes in the socket
16808 			 * buffer that have been sent to cover this
16809 			 * measurement. Cancel it.
16810 			 */
16811 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16812 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16813 						   tp->gput_seq,
16814 						   0, 0, 18, __LINE__, NULL, 0);
16815 			tp->t_flags &= ~TF_GPUTINPROG;
16816 		}
16817 	}
16818 	high_seq = th->th_ack;
16819 	if (tcp_bblogging_on(rack->rc_tp)) {
16820 		union tcp_log_stackspecific log;
16821 		struct timeval ltv;
16822 #ifdef TCP_REQUEST_TRK
16823 		struct tcp_sendfile_track *tcp_req;
16824 
16825 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
16826 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16827 		} else {
16828 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16829 		}
16830 #endif
16831 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16832 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16833 		if (rack->rack_no_prr == 0)
16834 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16835 		else
16836 			log.u_bbr.flex1 = 0;
16837 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16838 		log.u_bbr.use_lt_bw <<= 1;
16839 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16840 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16841 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16842 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16843 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16844 		log.u_bbr.flex3 = m->m_flags;
16845 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16846 		log.u_bbr.lost = thflags;
16847 		log.u_bbr.pacing_gain = 0x1;
16848 #ifdef TCP_ACCOUNTING
16849 		log.u_bbr.cwnd_gain = ack_val_set;
16850 #endif
16851 		log.u_bbr.flex7 = 2;
16852 		if (m->m_flags & M_TSTMP) {
16853 			/* Record the hardware timestamp if present */
16854 			mbuf_tstmp2timespec(m, &ts);
16855 			ltv.tv_sec = ts.tv_sec;
16856 			ltv.tv_usec = ts.tv_nsec / 1000;
16857 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16858 		} else if (m->m_flags & M_TSTMP_LRO) {
16859 			/* Record the LRO the arrival timestamp */
16860 			mbuf_tstmp2timespec(m, &ts);
16861 			ltv.tv_sec = ts.tv_sec;
16862 			ltv.tv_usec = ts.tv_nsec / 1000;
16863 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16864 		}
16865 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16866 		/* Log the rcv time */
16867 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16868 #ifdef TCP_REQUEST_TRK
16869 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16870 		log.u_bbr.applimited <<= 8;
16871 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16872 		log.u_bbr.applimited <<= 8;
16873 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16874 		if (tcp_req) {
16875 			/* Copy out any client req info */
16876 			/* seconds */
16877 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16878 			/* useconds */
16879 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16880 			log.u_bbr.rttProp = tcp_req->timestamp;
16881 			log.u_bbr.cur_del_rate = tcp_req->start;
16882 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16883 				log.u_bbr.flex8 |= 1;
16884 			} else {
16885 				log.u_bbr.flex8 |= 2;
16886 				log.u_bbr.bw_inuse = tcp_req->end;
16887 			}
16888 			log.u_bbr.flex6 = tcp_req->start_seq;
16889 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16890 				log.u_bbr.flex8 |= 4;
16891 				log.u_bbr.epoch = tcp_req->end_seq;
16892 			}
16893 		}
16894 #endif
16895 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16896 		    tlen, &log, true, &ltv);
16897 	}
16898 	/* Remove ack required flag if set, we have one  */
16899 	if (thflags & TH_ACK)
16900 		rack->rc_ack_required = 0;
16901 	if (rack->sack_attack_disable > 0) {
16902 		rack->r_ctl.ack_during_sd++;
16903 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16904 	}
16905 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16906 		way_out = 4;
16907 		retval = 0;
16908 		m_freem(m);
16909 		goto done_with_input;
16910 	}
16911 	/*
16912 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16913 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16914 	 */
16915 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16916 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16917 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16918 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16919 #ifdef TCP_ACCOUNTING
16920 		sched_unpin();
16921 #endif
16922 		return (1);
16923 	}
16924 	/*
16925 	 * If timestamps were negotiated during SYN/ACK and a
16926 	 * segment without a timestamp is received, silently drop
16927 	 * the segment, unless it is a RST segment or missing timestamps are
16928 	 * tolerated.
16929 	 * See section 3.2 of RFC 7323.
16930 	 */
16931 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16932 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16933 		way_out = 5;
16934 		retval = 0;
16935 		m_freem(m);
16936 		goto done_with_input;
16937 	}
16938 
16939 	/*
16940 	 * Segment received on connection. Reset idle time and keep-alive
16941 	 * timer. XXX: This should be done after segment validation to
16942 	 * ignore broken/spoofed segs.
16943 	 */
16944 	if  (tp->t_idle_reduce &&
16945 	     (tp->snd_max == tp->snd_una) &&
16946 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16947 		counter_u64_add(rack_input_idle_reduces, 1);
16948 		rack_cc_after_idle(rack, tp);
16949 	}
16950 	tp->t_rcvtime = ticks;
16951 #ifdef STATS
16952 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16953 #endif
16954 	if (tiwin > rack->r_ctl.rc_high_rwnd)
16955 		rack->r_ctl.rc_high_rwnd = tiwin;
16956 	/*
16957 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16958 	 * this to occur after we've validated the segment.
16959 	 */
16960 	if (tcp_ecn_input_segment(tp, thflags, tlen,
16961 	    tcp_packets_this_ack(tp, th->th_ack),
16962 	    iptos))
16963 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16964 
16965 	/*
16966 	 * If echoed timestamp is later than the current time, fall back to
16967 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16968 	 * were used when this connection was established.
16969 	 */
16970 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16971 		to.to_tsecr -= tp->ts_offset;
16972 		if (TSTMP_GT(to.to_tsecr, ms_cts))
16973 			to.to_tsecr = 0;
16974 	}
16975 
16976 	/*
16977 	 * If its the first time in we need to take care of options and
16978 	 * verify we can do SACK for rack!
16979 	 */
16980 	if (rack->r_state == 0) {
16981 		/* Should be init'd by rack_init() */
16982 		KASSERT(rack->rc_inp != NULL,
16983 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
16984 		if (rack->rc_inp == NULL) {
16985 			rack->rc_inp = inp;
16986 		}
16987 
16988 		/*
16989 		 * Process options only when we get SYN/ACK back. The SYN
16990 		 * case for incoming connections is handled in tcp_syncache.
16991 		 * According to RFC1323 the window field in a SYN (i.e., a
16992 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16993 		 * this is traditional behavior, may need to be cleaned up.
16994 		 */
16995 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16996 			/* Handle parallel SYN for ECN */
16997 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16998 			if ((to.to_flags & TOF_SCALE) &&
16999 			    (tp->t_flags & TF_REQ_SCALE)) {
17000 				tp->t_flags |= TF_RCVD_SCALE;
17001 				tp->snd_scale = to.to_wscale;
17002 			} else
17003 				tp->t_flags &= ~TF_REQ_SCALE;
17004 			/*
17005 			 * Initial send window.  It will be updated with the
17006 			 * next incoming segment to the scaled value.
17007 			 */
17008 			tp->snd_wnd = th->th_win;
17009 			rack_validate_fo_sendwin_up(tp, rack);
17010 			if ((to.to_flags & TOF_TS) &&
17011 			    (tp->t_flags & TF_REQ_TSTMP)) {
17012 				tp->t_flags |= TF_RCVD_TSTMP;
17013 				tp->ts_recent = to.to_tsval;
17014 				tp->ts_recent_age = cts;
17015 			} else
17016 				tp->t_flags &= ~TF_REQ_TSTMP;
17017 			if (to.to_flags & TOF_MSS) {
17018 				tcp_mss(tp, to.to_mss);
17019 			}
17020 			if ((tp->t_flags & TF_SACK_PERMIT) &&
17021 			    (to.to_flags & TOF_SACKPERM) == 0)
17022 				tp->t_flags &= ~TF_SACK_PERMIT;
17023 			if (IS_FASTOPEN(tp->t_flags)) {
17024 				if (to.to_flags & TOF_FASTOPEN) {
17025 					uint16_t mss;
17026 
17027 					if (to.to_flags & TOF_MSS)
17028 						mss = to.to_mss;
17029 					else
17030 						if ((inp->inp_vflag & INP_IPV6) != 0)
17031 							mss = TCP6_MSS;
17032 						else
17033 							mss = TCP_MSS;
17034 					tcp_fastopen_update_cache(tp, mss,
17035 					    to.to_tfo_len, to.to_tfo_cookie);
17036 				} else
17037 					tcp_fastopen_disable_path(tp);
17038 			}
17039 		}
17040 		/*
17041 		 * At this point we are at the initial call. Here we decide
17042 		 * if we are doing RACK or not. We do this by seeing if
17043 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
17044 		 * The code now does do dup-ack counting so if you don't
17045 		 * switch back you won't get rack & TLP, but you will still
17046 		 * get this stack.
17047 		 */
17048 
17049 		if ((rack_sack_not_required == 0) &&
17050 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
17051 			tcp_switch_back_to_default(tp);
17052 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
17053 			    tlen, iptos);
17054 #ifdef TCP_ACCOUNTING
17055 			sched_unpin();
17056 #endif
17057 			return (1);
17058 		}
17059 		tcp_set_hpts(tp);
17060 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
17061 	}
17062 	if (thflags & TH_FIN)
17063 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
17064 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
17065 	if ((rack->rc_gp_dyn_mul) &&
17066 	    (rack->use_fixed_rate == 0) &&
17067 	    (rack->rc_always_pace)) {
17068 		/* Check in on probertt */
17069 		rack_check_probe_rtt(rack, us_cts);
17070 	}
17071 	rack_clear_rate_sample(rack);
17072 	if ((rack->forced_ack) &&
17073 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
17074 		rack_handle_probe_response(rack, tiwin, us_cts);
17075 	}
17076 	/*
17077 	 * This is the one exception case where we set the rack state
17078 	 * always. All other times (timers etc) we must have a rack-state
17079 	 * set (so we assure we have done the checks above for SACK).
17080 	 */
17081 	rack->r_ctl.rc_rcvtime = cts;
17082 	if (rack->r_state != tp->t_state)
17083 		rack_set_state(tp, rack);
17084 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
17085 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
17086 		kern_prefetch(rsm, &prev_state);
17087 	prev_state = rack->r_state;
17088 	if ((thflags & TH_RST) &&
17089 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
17090 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
17091 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
17092 		/* The connection will be killed by a reset check the tracepoint */
17093 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
17094 	}
17095 	retval = (*rack->r_substate) (m, th, so,
17096 	    tp, &to, drop_hdrlen,
17097 	    tlen, tiwin, thflags, nxt_pkt, iptos);
17098 	if (retval == 0) {
17099 		/*
17100 		 * If retval is 1 the tcb is unlocked and most likely the tp
17101 		 * is gone.
17102 		 */
17103 		INP_WLOCK_ASSERT(inp);
17104 		if ((rack->rc_gp_dyn_mul) &&
17105 		    (rack->rc_always_pace) &&
17106 		    (rack->use_fixed_rate == 0) &&
17107 		    rack->in_probe_rtt &&
17108 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
17109 			/*
17110 			 * If we are going for target, lets recheck before
17111 			 * we output.
17112 			 */
17113 			rack_check_probe_rtt(rack, us_cts);
17114 		}
17115 		if (rack->set_pacing_done_a_iw == 0) {
17116 			/* How much has been acked? */
17117 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
17118 				/* We have enough to set in the pacing segment size */
17119 				rack->set_pacing_done_a_iw = 1;
17120 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
17121 			}
17122 		}
17123 		tcp_rack_xmit_timer_commit(rack, tp);
17124 #ifdef TCP_ACCOUNTING
17125 		/*
17126 		 * If we set the ack_val_se to what ack processing we are doing
17127 		 * we also want to track how many cycles we burned. Note
17128 		 * the bits after tcp_output we let be "free". This is because
17129 		 * we are also tracking the tcp_output times as well. Note the
17130 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
17131 		 * 0xf cannot be returned and is what we initialize it too to
17132 		 * indicate we are not doing the tabulations.
17133 		 */
17134 		if (ack_val_set != 0xf) {
17135 			uint64_t crtsc;
17136 
17137 			crtsc = get_cyclecount();
17138 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17139 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
17140 			}
17141 		}
17142 #endif
17143 		if ((nxt_pkt == 0) && (no_output == 0)) {
17144 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
17145 do_output_now:
17146 				if (tcp_output(tp) < 0) {
17147 #ifdef TCP_ACCOUNTING
17148 					sched_unpin();
17149 #endif
17150 					return (1);
17151 				}
17152 				did_out = 1;
17153 			}
17154 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
17155 			rack_free_trim(rack);
17156 		} else if ((no_output == 1) &&
17157 			   (nxt_pkt == 0)  &&
17158 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
17159 			/*
17160 			 * We are not in hpts and we had a pacing timer up. Use
17161 			 * the remaining time (slot_remaining) to restart the timer.
17162 			 */
17163 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
17164 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
17165 			rack_free_trim(rack);
17166 		}
17167 		/* Clear the flag, it may have been cleared by output but we may not have  */
17168 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
17169 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
17170 		/* Update any rounds needed */
17171 		if (rack_verbose_logging &&  tcp_bblogging_on(rack->rc_tp))
17172 			rack_log_hystart_event(rack, high_seq, 8);
17173 		/*
17174 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
17175 		 * causes issues when we are just going app limited. Lets
17176 		 * instead use SEQ_GT <or> where its equal but more data
17177 		 * is outstanding.
17178 		 *
17179 		 * Also make sure we are on the last ack of a series. We
17180 		 * have to have all the ack's processed in queue to know
17181 		 * if there is something left outstanding.
17182 		 */
17183 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
17184 		    (rack->rc_new_rnd_needed == 0) &&
17185 		    (nxt_pkt == 0)) {
17186 			rack_log_hystart_event(rack, tp->snd_una, 21);
17187 			rack->r_ctl.current_round++;
17188 			/* Force the next send to setup the next round */
17189 			rack->rc_new_rnd_needed = 1;
17190 			if (CC_ALGO(tp)->newround != NULL) {
17191 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
17192 			}
17193 		}
17194 		if ((nxt_pkt == 0) &&
17195 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
17196 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
17197 		     (tp->t_flags & TF_DELACK) ||
17198 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
17199 		      (tp->t_state <= TCPS_CLOSING)))) {
17200 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
17201 			if ((tp->snd_max == tp->snd_una) &&
17202 			    ((tp->t_flags & TF_DELACK) == 0) &&
17203 			    (tcp_in_hpts(rack->rc_tp)) &&
17204 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
17205 				/* keep alive not needed if we are hptsi output yet */
17206 				;
17207 			} else {
17208 				int late = 0;
17209 				if (tcp_in_hpts(tp)) {
17210 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
17211 						us_cts = tcp_get_usecs(NULL);
17212 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17213 							rack->r_early = 1;
17214 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17215 						} else
17216 							late = 1;
17217 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17218 					}
17219 					tcp_hpts_remove(tp);
17220 				}
17221 				if (late && (did_out == 0)) {
17222 					/*
17223 					 * We are late in the sending
17224 					 * and we did not call the output
17225 					 * (this probably should not happen).
17226 					 */
17227 					goto do_output_now;
17228 				}
17229 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17230 			}
17231 			way_out = 1;
17232 		} else if (nxt_pkt == 0) {
17233 			/* Do we have the correct timer running? */
17234 			rack_timer_audit(tp, rack, &so->so_snd);
17235 			way_out = 2;
17236 		}
17237 	done_with_input:
17238 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17239 		if (did_out)
17240 			rack->r_wanted_output = 0;
17241 	}
17242 #ifdef TCP_ACCOUNTING
17243 	sched_unpin();
17244 #endif
17245 	return (retval);
17246 }
17247 
17248 static void
17249 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17250     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17251 {
17252 	struct timeval tv;
17253 
17254 	/* First lets see if we have old packets */
17255 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17256 		if (ctf_do_queued_segments(tp, 1)) {
17257 			m_freem(m);
17258 			return;
17259 		}
17260 	}
17261 	if (m->m_flags & M_TSTMP_LRO) {
17262 		mbuf_tstmp2timeval(m, &tv);
17263 	} else {
17264 		/* Should not be should we kassert instead? */
17265 		tcp_get_usecs(&tv);
17266 	}
17267 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17268 	    &tv) == 0) {
17269 		INP_WUNLOCK(tptoinpcb(tp));
17270 	}
17271 }
17272 
17273 struct rack_sendmap *
17274 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17275 {
17276 	struct rack_sendmap *rsm = NULL;
17277 	int32_t idx;
17278 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17279 	int no_sack = 0;
17280 
17281 	/* Return the next guy to be re-transmitted */
17282 	if (tqhash_empty(rack->r_ctl.tqh)) {
17283 		return (NULL);
17284 	}
17285 	if (tp->t_flags & TF_SENTFIN) {
17286 		/* retran the end FIN? */
17287 		return (NULL);
17288 	}
17289 	/* ok lets look at this one */
17290 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17291 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17292 		return (rsm);
17293 	}
17294 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17295 		goto check_it;
17296 	}
17297 	rsm = rack_find_lowest_rsm(rack);
17298 	if (rsm == NULL) {
17299 		return (NULL);
17300 	}
17301 check_it:
17302 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) ||
17303 	    (rack->sack_attack_disable > 0)) {
17304 		no_sack = 1;
17305 	}
17306 	if ((no_sack > 0) &&
17307 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17308 		/*
17309 		 * No sack so we automatically do the 3 strikes and
17310 		 * retransmit (no rack timer would be started).
17311 		 */
17312 		return (rsm);
17313 	}
17314 	if (rsm->r_flags & RACK_ACKED) {
17315 		return (NULL);
17316 	}
17317 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17318 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17319 		/* Its not yet ready */
17320 		return (NULL);
17321 	}
17322 	srtt = rack_grab_rtt(tp, rack);
17323 	idx = rsm->r_rtr_cnt - 1;
17324 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17325 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
17326 	if ((tsused == ts_low) ||
17327 	    (TSTMP_LT(tsused, ts_low))) {
17328 		/* No time since sending */
17329 		return (NULL);
17330 	}
17331 	if ((tsused - ts_low) < thresh) {
17332 		/* It has not been long enough yet */
17333 		return (NULL);
17334 	}
17335 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17336 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
17337 	     (rack->sack_attack_disable == 0))) {
17338 		/*
17339 		 * We have passed the dup-ack threshold <or>
17340 		 * a SACK has indicated this is missing.
17341 		 * Note that if you are a declared attacker
17342 		 * it is only the dup-ack threshold that
17343 		 * will cause retransmits.
17344 		 */
17345 		/* log retransmit reason */
17346 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17347 		rack->r_fast_output = 0;
17348 		return (rsm);
17349 	}
17350 	return (NULL);
17351 }
17352 
17353 static void
17354 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
17355 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17356 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17357 {
17358 	if (tcp_bblogging_on(rack->rc_tp)) {
17359 		union tcp_log_stackspecific log;
17360 		struct timeval tv;
17361 
17362 		if (rack_verbose_logging == 0) {
17363 			/*
17364 			 * We are not verbose screen out all but
17365 			 * ones we always want.
17366 			 */
17367 			if ((method != 2) &&
17368 			    (method != 3) &&
17369 			    (method != 7) &&
17370 			    (method != 14) &&
17371 			    (method != 20)) {
17372 				return;
17373 			}
17374 		}
17375 		memset(&log, 0, sizeof(log));
17376 		log.u_bbr.flex1 = slot;
17377 		log.u_bbr.flex2 = len;
17378 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17379 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17380 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17381 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17382 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17383 		log.u_bbr.use_lt_bw <<= 1;
17384 		log.u_bbr.use_lt_bw |= rack->r_late;
17385 		log.u_bbr.use_lt_bw <<= 1;
17386 		log.u_bbr.use_lt_bw |= rack->r_early;
17387 		log.u_bbr.use_lt_bw <<= 1;
17388 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17389 		log.u_bbr.use_lt_bw <<= 1;
17390 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17391 		log.u_bbr.use_lt_bw <<= 1;
17392 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17393 		log.u_bbr.use_lt_bw <<= 1;
17394 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17395 		log.u_bbr.use_lt_bw <<= 1;
17396 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17397 		log.u_bbr.pkt_epoch = line;
17398 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17399 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17400 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17401 		log.u_bbr.bw_inuse = bw_est;
17402 		log.u_bbr.delRate = bw;
17403 		if (rack->r_ctl.gp_bw == 0)
17404 			log.u_bbr.cur_del_rate = 0;
17405 		else
17406 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17407 		log.u_bbr.rttProp = len_time;
17408 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17409 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17410 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17411 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17412 			/* We are in slow start */
17413 			log.u_bbr.flex7 = 1;
17414 		} else {
17415 			/* we are on congestion avoidance */
17416 			log.u_bbr.flex7 = 0;
17417 		}
17418 		log.u_bbr.flex8 = method;
17419 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17420 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17421 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17422 		log.u_bbr.cwnd_gain <<= 1;
17423 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17424 		log.u_bbr.cwnd_gain <<= 1;
17425 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17426 		log.u_bbr.bbr_substate = quality;
17427 		log.u_bbr.bbr_state = rack->dgp_on;
17428 		log.u_bbr.bbr_state <<= 1;
17429 		log.u_bbr.bbr_state |= rack->r_fill_less_agg;
17430 		log.u_bbr.bbr_state <<= 1;
17431 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17432 		log.u_bbr.bbr_state <<= 2;
17433 		log.u_bbr.bbr_state |= rack->r_pacing_discount;
17434 		log.u_bbr.flex7 = ((rack->r_ctl.pacing_discount_amm << 1) | log.u_bbr.flex7);
17435 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17436 		    &rack->rc_inp->inp_socket->so_rcv,
17437 		    &rack->rc_inp->inp_socket->so_snd,
17438 		    BBR_LOG_HPTSI_CALC, 0,
17439 		    0, &log, false, &tv);
17440 	}
17441 }
17442 
17443 static uint32_t
17444 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17445 {
17446 	uint32_t new_tso, user_max, pace_one;
17447 
17448 	user_max = rack->rc_user_set_max_segs * mss;
17449 	if (rack->rc_force_max_seg) {
17450 		return (user_max);
17451 	}
17452 	if (rack->use_fixed_rate &&
17453 	    ((rack->r_ctl.crte == NULL) ||
17454 	     (bw != rack->r_ctl.crte->rate))) {
17455 		/* Use the user mss since we are not exactly matched */
17456 		return (user_max);
17457 	}
17458 	if (rack_pace_one_seg ||
17459 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17460 		pace_one = 1;
17461 	else
17462 		pace_one = 0;
17463 
17464 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17465 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17466 	if (new_tso > user_max)
17467 		new_tso = user_max;
17468 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17469 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17470 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17471 	}
17472 	if (rack->r_ctl.rc_user_set_min_segs &&
17473 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17474 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17475 	return (new_tso);
17476 }
17477 
17478 static uint64_t
17479 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17480 {
17481 	uint64_t reduced_win;
17482 	uint32_t gain;
17483 
17484 	if (window_input < rc_init_window(rack)) {
17485 		/*
17486 		 * The cwnd is collapsed to
17487 		 * nearly zero, maybe because of a time-out?
17488 		 * Lets drop back to the lt-bw.
17489 		 */
17490 		reduced_win = rack_get_lt_bw(rack);
17491 		/* Set the flag so the caller knows its a rate and not a reduced window */
17492 		*rate_set = 1;
17493 		gain = 100;
17494 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
17495 		/*
17496 		 * If we are in recover our cwnd needs to be less for
17497 		 * our pacing consideration.
17498 		 */
17499 		if (rack->rack_hibeta == 0) {
17500 			reduced_win = window_input / 2;
17501 			gain = 50;
17502 		} else {
17503 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
17504 			reduced_win /= 100;
17505 			gain = rack->r_ctl.saved_hibeta;
17506 		}
17507 	} else {
17508 		/*
17509 		 * Apply Timely factor to increase/decrease the
17510 		 * amount we are pacing at.
17511 		 */
17512 		gain = rack_get_output_gain(rack, NULL);
17513 		if (gain > rack_gain_p5_ub) {
17514 			gain = rack_gain_p5_ub;
17515 		}
17516 		reduced_win = window_input * gain;
17517 		reduced_win /= 100;
17518 	}
17519 	if (gain_b != NULL)
17520 		*gain_b = gain;
17521 	/*
17522 	 * What is being returned here is a trimmed down
17523 	 * window values in all cases where rate_set is left
17524 	 * at 0. In one case we actually return the rate (lt_bw).
17525 	 * the "reduced_win" is returned as a slimmed down cwnd that
17526 	 * is then calculated by the caller into a rate when rate_set
17527 	 * is 0.
17528 	 */
17529 	return (reduced_win);
17530 }
17531 
17532 static int32_t
17533 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)
17534 {
17535 	uint64_t lentim, fill_bw;
17536 
17537 	/* Lets first see if we are full, if so continue with normal rate */
17538 	rack->r_via_fill_cw = 0;
17539 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17540 		return (slot);
17541 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17542 		return (slot);
17543 	if (rack->r_ctl.rc_last_us_rtt == 0)
17544 		return (slot);
17545 	if (rack->rc_pace_fill_if_rttin_range &&
17546 	    (rack->r_ctl.rc_last_us_rtt >=
17547 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17548 		/* The rtt is huge, N * smallest, lets not fill */
17549 		return (slot);
17550 	}
17551 	/*
17552 	 * first lets calculate the b/w based on the last us-rtt
17553 	 * and the the smallest send window.
17554 	 */
17555 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17556 	if (rack->rc_fillcw_apply_discount) {
17557 		uint32_t rate_set = 0;
17558 
17559 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17560 		if (rate_set) {
17561 			goto at_lt_bw;
17562 		}
17563 	}
17564 	/* Take the rwnd if its smaller */
17565 	if (fill_bw > rack->rc_tp->snd_wnd)
17566 		fill_bw = rack->rc_tp->snd_wnd;
17567 	/* Now lets make it into a b/w */
17568 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17569 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17570 at_lt_bw:
17571 	if (rack->r_fill_less_agg) {
17572 		/*
17573 		 * We want the average of the rate_wanted
17574 		 * and our fill-cw calculated bw. We also want
17575 		 * to cap any increase to be no more than
17576 		 * X times the lt_bw (where X is the rack_bw_multipler).
17577 		 */
17578 		uint64_t lt_bw, rate;
17579 
17580 		lt_bw = rack_get_lt_bw(rack);
17581 		if (lt_bw > *rate_wanted)
17582 			rate = lt_bw;
17583 		else
17584 			rate = *rate_wanted;
17585 		fill_bw += rate;
17586 		fill_bw /= 2;
17587 		if (rack_bw_multipler && (fill_bw > (rate * rack_bw_multipler))) {
17588 			fill_bw = rate * rack_bw_multipler;
17589 		}
17590 	}
17591 	/* We are below the min b/w */
17592 	if (non_paced)
17593 		*rate_wanted = fill_bw;
17594 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17595 		return (slot);
17596 	rack->r_via_fill_cw = 1;
17597 	if (rack->r_rack_hw_rate_caps &&
17598 	    (rack->r_ctl.crte != NULL)) {
17599 		uint64_t high_rate;
17600 
17601 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17602 		if (fill_bw > high_rate) {
17603 			/* We are capping bw at the highest rate table entry */
17604 			if (*rate_wanted > high_rate) {
17605 				/* The original rate was also capped */
17606 				rack->r_via_fill_cw = 0;
17607 			}
17608 			rack_log_hdwr_pacing(rack,
17609 					     fill_bw, high_rate, __LINE__,
17610 					     0, 3);
17611 			fill_bw = high_rate;
17612 			if (capped)
17613 				*capped = 1;
17614 		}
17615 	} else if ((rack->r_ctl.crte == NULL) &&
17616 		   (rack->rack_hdrw_pacing == 0) &&
17617 		   (rack->rack_hdw_pace_ena) &&
17618 		   rack->r_rack_hw_rate_caps &&
17619 		   (rack->rack_attempt_hdwr_pace == 0) &&
17620 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
17621 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17622 		/*
17623 		 * Ok we may have a first attempt that is greater than our top rate
17624 		 * lets check.
17625 		 */
17626 		uint64_t high_rate;
17627 
17628 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17629 		if (high_rate) {
17630 			if (fill_bw > high_rate) {
17631 				fill_bw = high_rate;
17632 				if (capped)
17633 					*capped = 1;
17634 			}
17635 		}
17636 	}
17637 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17638 		if (rack->rc_hybrid_mode)
17639 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17640 					   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17641 		fill_bw = rack->r_ctl.bw_rate_cap;
17642 	}
17643 	/*
17644 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
17645 	 * in an rtt (unless it was capped), what does that
17646 	 * time wise equate too?
17647 	 */
17648 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17649 	lentim /= fill_bw;
17650 	*rate_wanted = fill_bw;
17651 	if (non_paced || (lentim < slot)) {
17652 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17653 					   0, lentim, 12, __LINE__, NULL, 0);
17654 		return ((int32_t)lentim);
17655 	} else
17656 		return (slot);
17657 }
17658 
17659 static int32_t
17660 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
17661 {
17662 	uint64_t srtt;
17663 	int32_t slot = 0;
17664 	int32_t minslot = 0;
17665 	int can_start_hw_pacing = 1;
17666 	int err;
17667 	int pace_one;
17668 
17669 	if (rack_pace_one_seg ||
17670 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17671 		pace_one = 1;
17672 	else
17673 		pace_one = 0;
17674 	if (rack->rc_always_pace == 0) {
17675 		/*
17676 		 * We use the most optimistic possible cwnd/srtt for
17677 		 * sending calculations. This will make our
17678 		 * calculation anticipate getting more through
17679 		 * quicker then possible. But thats ok we don't want
17680 		 * the peer to have a gap in data sending.
17681 		 */
17682 		uint64_t cwnd, tr_perms = 0;
17683 		int32_t reduce = 0;
17684 
17685 	old_method:
17686 		/*
17687 		 * We keep no precise pacing with the old method
17688 		 * instead we use the pacer to mitigate bursts.
17689 		 */
17690 		if (rack->r_ctl.rc_rack_min_rtt)
17691 			srtt = rack->r_ctl.rc_rack_min_rtt;
17692 		else
17693 			srtt = max(tp->t_srtt, 1);
17694 		if (rack->r_ctl.rc_rack_largest_cwnd)
17695 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17696 		else
17697 			cwnd = rack->r_ctl.cwnd_to_use;
17698 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
17699 		tr_perms = (cwnd * 1000) / srtt;
17700 		if (tr_perms == 0) {
17701 			tr_perms = ctf_fixed_maxseg(tp);
17702 		}
17703 		/*
17704 		 * Calculate how long this will take to drain, if
17705 		 * the calculation comes out to zero, thats ok we
17706 		 * will use send_a_lot to possibly spin around for
17707 		 * more increasing tot_len_this_send to the point
17708 		 * that its going to require a pace, or we hit the
17709 		 * cwnd. Which in that case we are just waiting for
17710 		 * a ACK.
17711 		 */
17712 		slot = len / tr_perms;
17713 		/* Now do we reduce the time so we don't run dry? */
17714 		if (slot && rack_slot_reduction) {
17715 			reduce = (slot / rack_slot_reduction);
17716 			if (reduce < slot) {
17717 				slot -= reduce;
17718 			} else
17719 				slot = 0;
17720 		}
17721 		slot *= HPTS_USEC_IN_MSEC;
17722 		if (rack->rc_pace_to_cwnd) {
17723 			uint64_t rate_wanted = 0;
17724 
17725 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17726 			rack->rc_ack_can_sendout_data = 1;
17727 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17728 		} else
17729 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17730 		/*******************************************************/
17731 		/* RRS: We insert non-paced call to stats here for len */
17732 		/*******************************************************/
17733 	} else {
17734 		uint64_t bw_est, res, lentim, rate_wanted;
17735 		uint32_t segs, oh;
17736 		int capped = 0;
17737 		int prev_fill;
17738 
17739 		if ((rack->r_rr_config == 1) && rsm) {
17740 			return (rack->r_ctl.rc_min_to);
17741 		}
17742 		if (rack->use_fixed_rate) {
17743 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17744 		} else if ((rack->r_ctl.init_rate == 0) &&
17745 			   (rack->r_ctl.gp_bw == 0)) {
17746 			/* no way to yet do an estimate */
17747 			bw_est = rate_wanted = 0;
17748 		} else if (rack->dgp_on) {
17749 			bw_est = rack_get_bw(rack);
17750 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17751 		} else {
17752 			uint32_t gain, rate_set = 0;
17753 
17754 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17755 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17756 			if (rate_set == 0) {
17757 				if (rate_wanted > rack->rc_tp->snd_wnd)
17758 					rate_wanted = rack->rc_tp->snd_wnd;
17759 				/* Now lets make it into a b/w */
17760 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17761 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17762 			}
17763 			bw_est = rate_wanted;
17764 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17765 						   rack->r_ctl.cwnd_to_use,
17766 						   rate_wanted, bw_est,
17767 						   rack->r_ctl.rc_last_us_rtt,
17768 						   88, __LINE__, NULL, gain);
17769 		}
17770 		if ((bw_est == 0) || (rate_wanted == 0) ||
17771 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17772 			/*
17773 			 * No way yet to make a b/w estimate or
17774 			 * our raise is set incorrectly.
17775 			 */
17776 			goto old_method;
17777 		}
17778 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
17779 		/* We need to account for all the overheads */
17780 		segs = (len + segsiz - 1) / segsiz;
17781 		/*
17782 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17783 		 * and how much data we put in each packet. Yes this
17784 		 * means we may be off if we are larger than 1500 bytes
17785 		 * or smaller. But this just makes us more conservative.
17786 		 */
17787 
17788 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17789 		if (rack->r_is_v6) {
17790 #ifdef INET6
17791 			oh += sizeof(struct ip6_hdr);
17792 #endif
17793 		} else {
17794 #ifdef INET
17795 			oh += sizeof(struct ip);
17796 #endif
17797 		}
17798 		/* We add a fixed 14 for the ethernet header */
17799 		oh += 14;
17800 		segs *= oh;
17801 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17802 		res = lentim / rate_wanted;
17803 		slot = (uint32_t)res;
17804 		if (rack_hw_rate_min &&
17805 		    (rate_wanted < rack_hw_rate_min)) {
17806 			can_start_hw_pacing = 0;
17807 			if (rack->r_ctl.crte) {
17808 				/*
17809 				 * Ok we need to release it, we
17810 				 * have fallen too low.
17811 				 */
17812 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17813 				rack->r_ctl.crte = NULL;
17814 				rack->rack_attempt_hdwr_pace = 0;
17815 				rack->rack_hdrw_pacing = 0;
17816 			}
17817 		}
17818 		if (rack->r_ctl.crte &&
17819 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17820 			/*
17821 			 * We want more than the hardware can give us,
17822 			 * don't start any hw pacing.
17823 			 */
17824 			can_start_hw_pacing = 0;
17825 			if (rack->r_rack_hw_rate_caps == 0) {
17826 				/*
17827 				 * Ok we need to release it, we
17828 				 * want more than the card can give us and
17829 				 * no rate cap is in place. Set it up so
17830 				 * when we want less we can retry.
17831 				 */
17832 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17833 				rack->r_ctl.crte = NULL;
17834 				rack->rack_attempt_hdwr_pace = 0;
17835 				rack->rack_hdrw_pacing = 0;
17836 			}
17837 		}
17838 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17839 			/*
17840 			 * We lost our rate somehow, this can happen
17841 			 * if the interface changed underneath us.
17842 			 */
17843 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17844 			rack->r_ctl.crte = NULL;
17845 			/* Lets re-allow attempting to setup pacing */
17846 			rack->rack_hdrw_pacing = 0;
17847 			rack->rack_attempt_hdwr_pace = 0;
17848 			rack_log_hdwr_pacing(rack,
17849 					     rate_wanted, bw_est, __LINE__,
17850 					     0, 6);
17851 		}
17852 		prev_fill = rack->r_via_fill_cw;
17853 		if ((rack->rc_pace_to_cwnd) &&
17854 		    (capped == 0) &&
17855 		    (rack->dgp_on == 1) &&
17856 		    (rack->use_fixed_rate == 0) &&
17857 		    (rack->in_probe_rtt == 0) &&
17858 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17859 			/*
17860 			 * We want to pace at our rate *or* faster to
17861 			 * fill the cwnd to the max if its not full.
17862 			 */
17863 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17864 			/* Re-check to make sure we are not exceeding our max b/w */
17865 			if ((rack->r_ctl.crte != NULL) &&
17866 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17867 				/*
17868 				 * We want more than the hardware can give us,
17869 				 * don't start any hw pacing.
17870 				 */
17871 				can_start_hw_pacing = 0;
17872 				if (rack->r_rack_hw_rate_caps == 0) {
17873 					/*
17874 					 * Ok we need to release it, we
17875 					 * want more than the card can give us and
17876 					 * no rate cap is in place. Set it up so
17877 					 * when we want less we can retry.
17878 					 */
17879 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17880 					rack->r_ctl.crte = NULL;
17881 					rack->rack_attempt_hdwr_pace = 0;
17882 					rack->rack_hdrw_pacing = 0;
17883 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17884 				}
17885 			}
17886 		}
17887 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17888 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17889 			if ((rack->rack_hdw_pace_ena) &&
17890 			    (can_start_hw_pacing > 0) &&
17891 			    (rack->rack_hdrw_pacing == 0) &&
17892 			    (rack->rack_attempt_hdwr_pace == 0)) {
17893 				/*
17894 				 * Lets attempt to turn on hardware pacing
17895 				 * if we can.
17896 				 */
17897 				rack->rack_attempt_hdwr_pace = 1;
17898 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17899 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
17900 								       rate_wanted,
17901 								       RS_PACING_GEQ,
17902 								       &err, &rack->r_ctl.crte_prev_rate);
17903 				if (rack->r_ctl.crte) {
17904 					rack->rack_hdrw_pacing = 1;
17905 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17906 													   pace_one, rack->r_ctl.crte,
17907 													   NULL, rack->r_ctl.pace_len_divisor);
17908 					rack_log_hdwr_pacing(rack,
17909 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17910 							     err, 0);
17911 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17912 				} else {
17913 					counter_u64_add(rack_hw_pace_init_fail, 1);
17914 				}
17915 			} else if (rack->rack_hdrw_pacing &&
17916 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17917 				/* Do we need to adjust our rate? */
17918 				const struct tcp_hwrate_limit_table *nrte;
17919 
17920 				if (rack->r_up_only &&
17921 				    (rate_wanted < rack->r_ctl.crte->rate)) {
17922 					/**
17923 					 * We have four possible states here
17924 					 * having to do with the previous time
17925 					 * and this time.
17926 					 *   previous  |  this-time
17927 					 * A)     0      |     0   -- fill_cw not in the picture
17928 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
17929 					 * C)     1      |     1   -- all rates from fill_cw
17930 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
17931 					 *
17932 					 * For case A, C and D we don't allow a drop. But for
17933 					 * case B where we now our on our steady rate we do
17934 					 * allow a drop.
17935 					 *
17936 					 */
17937 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17938 						goto done_w_hdwr;
17939 				}
17940 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
17941 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17942 					if (rack_hw_rate_to_low &&
17943 					    (bw_est < rack_hw_rate_to_low)) {
17944 						/*
17945 						 * The pacing rate is too low for hardware, but
17946 						 * do allow hardware pacing to be restarted.
17947 						 */
17948 						rack_log_hdwr_pacing(rack,
17949 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
17950 								     0, 5);
17951 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17952 						rack->r_ctl.crte = NULL;
17953 						rack->rack_attempt_hdwr_pace = 0;
17954 						rack->rack_hdrw_pacing = 0;
17955 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17956 						goto done_w_hdwr;
17957 					}
17958 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17959 								   rack->rc_tp,
17960 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
17961 								   rate_wanted,
17962 								   RS_PACING_GEQ,
17963 								   &err, &rack->r_ctl.crte_prev_rate);
17964 					if (nrte == NULL) {
17965 						/*
17966 						 * Lost the rate, lets drop hardware pacing
17967 						 * period.
17968 						 */
17969 						rack->rack_hdrw_pacing = 0;
17970 						rack->r_ctl.crte = NULL;
17971 						rack_log_hdwr_pacing(rack,
17972 								     rate_wanted, 0, __LINE__,
17973 								     err, 1);
17974 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17975 						counter_u64_add(rack_hw_pace_lost, 1);
17976 					} else if (nrte != rack->r_ctl.crte) {
17977 						rack->r_ctl.crte = nrte;
17978 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17979 														   segsiz, pace_one, rack->r_ctl.crte,
17980 														   NULL, rack->r_ctl.pace_len_divisor);
17981 						rack_log_hdwr_pacing(rack,
17982 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17983 								     err, 2);
17984 						rack->r_ctl.last_hw_bw_req = rate_wanted;
17985 					}
17986 				} else {
17987 					/* We just need to adjust the segment size */
17988 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17989 					rack_log_hdwr_pacing(rack,
17990 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17991 							     0, 4);
17992 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17993 				}
17994 			}
17995 		}
17996 		if (minslot && (minslot > slot)) {
17997 			rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
17998 						   98, __LINE__, NULL, 0);
17999 			slot = minslot;
18000 		}
18001 	done_w_hdwr:
18002 		if (rack_limit_time_with_srtt &&
18003 		    (rack->use_fixed_rate == 0) &&
18004 		    (rack->rack_hdrw_pacing == 0)) {
18005 			/*
18006 			 * Sanity check, we do not allow the pacing delay
18007 			 * to be longer than the SRTT of the path. If it is
18008 			 * a slow path, then adding a packet should increase
18009 			 * the RTT and compensate for this i.e. the srtt will
18010 			 * be greater so the allowed pacing time will be greater.
18011 			 *
18012 			 * Note this restriction is not for where a peak rate
18013 			 * is set, we are doing fixed pacing or hardware pacing.
18014 			 */
18015 			if (rack->rc_tp->t_srtt)
18016 				srtt = rack->rc_tp->t_srtt;
18017 			else
18018 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
18019 			if (srtt < (uint64_t)slot) {
18020 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
18021 				slot = srtt;
18022 			}
18023 		}
18024 		/*******************************************************************/
18025 		/* RRS: We insert paced call to stats here for len and rate_wanted */
18026 		/*******************************************************************/
18027 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
18028 	}
18029 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
18030 		/*
18031 		 * If this rate is seeing enobufs when it
18032 		 * goes to send then either the nic is out
18033 		 * of gas or we are mis-estimating the time
18034 		 * somehow and not letting the queue empty
18035 		 * completely. Lets add to the pacing time.
18036 		 */
18037 		int hw_boost_delay;
18038 
18039 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
18040 		if (hw_boost_delay > rack_enobuf_hw_max)
18041 			hw_boost_delay = rack_enobuf_hw_max;
18042 		else if (hw_boost_delay < rack_enobuf_hw_min)
18043 			hw_boost_delay = rack_enobuf_hw_min;
18044 		slot += hw_boost_delay;
18045 	}
18046 	return (slot);
18047 }
18048 
18049 static void
18050 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
18051     tcp_seq startseq, uint32_t sb_offset)
18052 {
18053 	struct rack_sendmap *my_rsm = NULL;
18054 
18055 	if (tp->t_state < TCPS_ESTABLISHED) {
18056 		/*
18057 		 * We don't start any measurements if we are
18058 		 * not at least established.
18059 		 */
18060 		return;
18061 	}
18062 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
18063 		/*
18064 		 * We will get no more data into the SB
18065 		 * this means we need to have the data available
18066 		 * before we start a measurement.
18067 		 */
18068 
18069 		if (sbavail(&tptosocket(tp)->so_snd) <
18070 		    max(rc_init_window(rack),
18071 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
18072 			/* Nope not enough data */
18073 			return;
18074 		}
18075 	}
18076 	tp->t_flags |= TF_GPUTINPROG;
18077 	rack->r_ctl.rc_gp_cumack_ts = 0;
18078 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
18079 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
18080 	tp->gput_seq = startseq;
18081 	rack->app_limited_needs_set = 0;
18082 	if (rack->in_probe_rtt)
18083 		rack->measure_saw_probe_rtt = 1;
18084 	else if ((rack->measure_saw_probe_rtt) &&
18085 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
18086 		rack->measure_saw_probe_rtt = 0;
18087 	if (rack->rc_gp_filled)
18088 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
18089 	else {
18090 		/* Special case initial measurement */
18091 		struct timeval tv;
18092 
18093 		tp->gput_ts = tcp_get_usecs(&tv);
18094 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18095 	}
18096 	/*
18097 	 * We take a guess out into the future,
18098 	 * if we have no measurement and no
18099 	 * initial rate, we measure the first
18100 	 * initial-windows worth of data to
18101 	 * speed up getting some GP measurement and
18102 	 * thus start pacing.
18103 	 */
18104 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
18105 		rack->app_limited_needs_set = 1;
18106 		tp->gput_ack = startseq + max(rc_init_window(rack),
18107 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
18108 		rack_log_pacing_delay_calc(rack,
18109 					   tp->gput_seq,
18110 					   tp->gput_ack,
18111 					   0,
18112 					   tp->gput_ts,
18113 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18114 					   9,
18115 					   __LINE__, NULL, 0);
18116 		rack_tend_gp_marks(tp, rack);
18117 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18118 		return;
18119 	}
18120 	if (sb_offset) {
18121 		/*
18122 		 * We are out somewhere in the sb
18123 		 * can we use the already outstanding data?
18124 		 */
18125 
18126 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
18127 			/*
18128 			 * Yes first one is good and in this case
18129 			 * the tp->gput_ts is correctly set based on
18130 			 * the last ack that arrived (no need to
18131 			 * set things up when an ack comes in).
18132 			 */
18133 			my_rsm = tqhash_min(rack->r_ctl.tqh);
18134 			if ((my_rsm == NULL) ||
18135 			    (my_rsm->r_rtr_cnt != 1)) {
18136 				/* retransmission? */
18137 				goto use_latest;
18138 			}
18139 		} else {
18140 			if (rack->r_ctl.rc_first_appl == NULL) {
18141 				/*
18142 				 * If rc_first_appl is NULL
18143 				 * then the cnt should be 0.
18144 				 * This is probably an error, maybe
18145 				 * a KASSERT would be approprate.
18146 				 */
18147 				goto use_latest;
18148 			}
18149 			/*
18150 			 * If we have a marker pointer to the last one that is
18151 			 * app limited we can use that, but we need to set
18152 			 * things up so that when it gets ack'ed we record
18153 			 * the ack time (if its not already acked).
18154 			 */
18155 			rack->app_limited_needs_set = 1;
18156 			/*
18157 			 * We want to get to the rsm that is either
18158 			 * next with space i.e. over 1 MSS or the one
18159 			 * after that (after the app-limited).
18160 			 */
18161 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
18162 			if (my_rsm) {
18163 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
18164 					/* Have to use the next one */
18165 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18166 				else {
18167 					/* Use after the first MSS of it is acked */
18168 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
18169 					goto start_set;
18170 				}
18171 			}
18172 			if ((my_rsm == NULL) ||
18173 			    (my_rsm->r_rtr_cnt != 1)) {
18174 				/*
18175 				 * Either its a retransmit or
18176 				 * the last is the app-limited one.
18177 				 */
18178 				goto use_latest;
18179 			}
18180 		}
18181 		tp->gput_seq = my_rsm->r_start;
18182 start_set:
18183 		if (my_rsm->r_flags & RACK_ACKED) {
18184 			/*
18185 			 * This one has been acked use the arrival ack time
18186 			 */
18187 			struct rack_sendmap *nrsm;
18188 
18189 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18190 			rack->app_limited_needs_set = 0;
18191 			/*
18192 			 * Ok in this path we need to use the r_end now
18193 			 * since this guy is the starting ack.
18194 			 */
18195 			tp->gput_seq = my_rsm->r_end;
18196 			/*
18197 			 * We also need to adjust up the sendtime
18198 			 * to the send of the next data after my_rsm.
18199 			 */
18200 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18201 			if (nrsm != NULL)
18202 				my_rsm = nrsm;
18203 			else {
18204 				/*
18205 				 * The next as not been sent, thats the
18206 				 * case for using the latest.
18207 				 */
18208 				goto use_latest;
18209 			}
18210 		}
18211 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18212 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
18213 		rack->r_ctl.rc_gp_cumack_ts = 0;
18214 		rack_log_pacing_delay_calc(rack,
18215 					   tp->gput_seq,
18216 					   tp->gput_ack,
18217 					   (uint64_t)my_rsm,
18218 					   tp->gput_ts,
18219 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18220 					   9,
18221 					   __LINE__, my_rsm, 0);
18222 		/* Now lets make sure all are marked as they should be */
18223 		rack_tend_gp_marks(tp, rack);
18224 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18225 		return;
18226 	}
18227 
18228 use_latest:
18229 	/*
18230 	 * We don't know how long we may have been
18231 	 * idle or if this is the first-send. Lets
18232 	 * setup the flag so we will trim off
18233 	 * the first ack'd data so we get a true
18234 	 * measurement.
18235 	 */
18236 	rack->app_limited_needs_set = 1;
18237 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18238 	rack->r_ctl.rc_gp_cumack_ts = 0;
18239 	/* Find this guy so we can pull the send time */
18240 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18241 	if (my_rsm) {
18242 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18243 		if (my_rsm->r_flags & RACK_ACKED) {
18244 			/*
18245 			 * Unlikely since its probably what was
18246 			 * just transmitted (but I am paranoid).
18247 			 */
18248 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18249 			rack->app_limited_needs_set = 0;
18250 		}
18251 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18252 			/* This also is unlikely */
18253 			tp->gput_seq = my_rsm->r_start;
18254 		}
18255 	} else {
18256 		/*
18257 		 * TSNH unless we have some send-map limit,
18258 		 * and even at that it should not be hitting
18259 		 * that limit (we should have stopped sending).
18260 		 */
18261 		struct timeval tv;
18262 
18263 		microuptime(&tv);
18264 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18265 	}
18266 	rack_tend_gp_marks(tp, rack);
18267 	rack_log_pacing_delay_calc(rack,
18268 				   tp->gput_seq,
18269 				   tp->gput_ack,
18270 				   (uint64_t)my_rsm,
18271 				   tp->gput_ts,
18272 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18273 				   9, __LINE__, NULL, 0);
18274 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18275 }
18276 
18277 static inline uint32_t
18278 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18279     uint32_t avail, int32_t sb_offset)
18280 {
18281 	uint32_t len;
18282 	uint32_t sendwin;
18283 
18284 	if (tp->snd_wnd > cwnd_to_use)
18285 		sendwin = cwnd_to_use;
18286 	else
18287 		sendwin = tp->snd_wnd;
18288 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18289 		/* We never want to go over our peers rcv-window */
18290 		len = 0;
18291 	} else {
18292 		uint32_t flight;
18293 
18294 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18295 		if (flight >= sendwin) {
18296 			/*
18297 			 * We have in flight what we are allowed by cwnd (if
18298 			 * it was rwnd blocking it would have hit above out
18299 			 * >= tp->snd_wnd).
18300 			 */
18301 			return (0);
18302 		}
18303 		len = sendwin - flight;
18304 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18305 			/* We would send too much (beyond the rwnd) */
18306 			len = tp->snd_wnd - ctf_outstanding(tp);
18307 		}
18308 		if ((len + sb_offset) > avail) {
18309 			/*
18310 			 * We don't have that much in the SB, how much is
18311 			 * there?
18312 			 */
18313 			len = avail - sb_offset;
18314 		}
18315 	}
18316 	return (len);
18317 }
18318 
18319 static void
18320 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18321 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18322 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18323 {
18324 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18325 		union tcp_log_stackspecific log;
18326 		struct timeval tv;
18327 
18328 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18329 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18330 		log.u_bbr.flex1 = error;
18331 		log.u_bbr.flex2 = flags;
18332 		log.u_bbr.flex3 = rsm_is_null;
18333 		log.u_bbr.flex4 = ipoptlen;
18334 		log.u_bbr.flex5 = tp->rcv_numsacks;
18335 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18336 		log.u_bbr.flex7 = optlen;
18337 		log.u_bbr.flex8 = rack->r_fsb_inited;
18338 		log.u_bbr.applimited = rack->r_fast_output;
18339 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18340 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18341 		log.u_bbr.cwnd_gain = mode;
18342 		log.u_bbr.pkts_out = orig_len;
18343 		log.u_bbr.lt_epoch = len;
18344 		log.u_bbr.delivered = line;
18345 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18346 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18347 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18348 			       len, &log, false, NULL, __func__, __LINE__, &tv);
18349 	}
18350 }
18351 
18352 
18353 static struct mbuf *
18354 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18355 		   struct rack_fast_send_blk *fsb,
18356 		   int32_t seglimit, int32_t segsize, int hw_tls)
18357 {
18358 #ifdef KERN_TLS
18359 	struct ktls_session *tls, *ntls;
18360 #ifdef INVARIANTS
18361 	struct mbuf *start;
18362 #endif
18363 #endif
18364 	struct mbuf *m, *n, **np, *smb;
18365 	struct mbuf *top;
18366 	int32_t off, soff;
18367 	int32_t len = *plen;
18368 	int32_t fragsize;
18369 	int32_t len_cp = 0;
18370 	uint32_t mlen, frags;
18371 
18372 	soff = off = the_off;
18373 	smb = m = the_m;
18374 	np = &top;
18375 	top = NULL;
18376 #ifdef KERN_TLS
18377 	if (hw_tls && (m->m_flags & M_EXTPG))
18378 		tls = m->m_epg_tls;
18379 	else
18380 		tls = NULL;
18381 #ifdef INVARIANTS
18382 	start = m;
18383 #endif
18384 #endif
18385 	while (len > 0) {
18386 		if (m == NULL) {
18387 			*plen = len_cp;
18388 			break;
18389 		}
18390 #ifdef KERN_TLS
18391 		if (hw_tls) {
18392 			if (m->m_flags & M_EXTPG)
18393 				ntls = m->m_epg_tls;
18394 			else
18395 				ntls = NULL;
18396 
18397 			/*
18398 			 * Avoid mixing TLS records with handshake
18399 			 * data or TLS records from different
18400 			 * sessions.
18401 			 */
18402 			if (tls != ntls) {
18403 				MPASS(m != start);
18404 				*plen = len_cp;
18405 				break;
18406 			}
18407 		}
18408 #endif
18409 		mlen = min(len, m->m_len - off);
18410 		if (seglimit) {
18411 			/*
18412 			 * For M_EXTPG mbufs, add 3 segments
18413 			 * + 1 in case we are crossing page boundaries
18414 			 * + 2 in case the TLS hdr/trailer are used
18415 			 * It is cheaper to just add the segments
18416 			 * than it is to take the cache miss to look
18417 			 * at the mbuf ext_pgs state in detail.
18418 			 */
18419 			if (m->m_flags & M_EXTPG) {
18420 				fragsize = min(segsize, PAGE_SIZE);
18421 				frags = 3;
18422 			} else {
18423 				fragsize = segsize;
18424 				frags = 0;
18425 			}
18426 
18427 			/* Break if we really can't fit anymore. */
18428 			if ((frags + 1) >= seglimit) {
18429 				*plen =	len_cp;
18430 				break;
18431 			}
18432 
18433 			/*
18434 			 * Reduce size if you can't copy the whole
18435 			 * mbuf. If we can't copy the whole mbuf, also
18436 			 * adjust len so the loop will end after this
18437 			 * mbuf.
18438 			 */
18439 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18440 				mlen = (seglimit - frags - 1) * fragsize;
18441 				len = mlen;
18442 				*plen = len_cp + len;
18443 			}
18444 			frags += howmany(mlen, fragsize);
18445 			if (frags == 0)
18446 				frags++;
18447 			seglimit -= frags;
18448 			KASSERT(seglimit > 0,
18449 			    ("%s: seglimit went too low", __func__));
18450 		}
18451 		n = m_get(M_NOWAIT, m->m_type);
18452 		*np = n;
18453 		if (n == NULL)
18454 			goto nospace;
18455 		n->m_len = mlen;
18456 		soff += mlen;
18457 		len_cp += n->m_len;
18458 		if (m->m_flags & (M_EXT | M_EXTPG)) {
18459 			n->m_data = m->m_data + off;
18460 			mb_dupcl(n, m);
18461 		} else {
18462 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18463 			    (u_int)n->m_len);
18464 		}
18465 		len -= n->m_len;
18466 		off = 0;
18467 		m = m->m_next;
18468 		np = &n->m_next;
18469 		if (len || (soff == smb->m_len)) {
18470 			/*
18471 			 * We have more so we move forward  or
18472 			 * we have consumed the entire mbuf and
18473 			 * len has fell to 0.
18474 			 */
18475 			soff = 0;
18476 			smb = m;
18477 		}
18478 
18479 	}
18480 	if (fsb != NULL) {
18481 		fsb->m = smb;
18482 		fsb->off = soff;
18483 		if (smb) {
18484 			/*
18485 			 * Save off the size of the mbuf. We do
18486 			 * this so that we can recognize when it
18487 			 * has been trimmed by sbcut() as acks
18488 			 * come in.
18489 			 */
18490 			fsb->o_m_len = smb->m_len;
18491 			fsb->o_t_len = M_TRAILINGROOM(smb);
18492 		} else {
18493 			/*
18494 			 * This is the case where the next mbuf went to NULL. This
18495 			 * means with this copy we have sent everything in the sb.
18496 			 * In theory we could clear the fast_output flag, but lets
18497 			 * not since its possible that we could get more added
18498 			 * and acks that call the extend function which would let
18499 			 * us send more.
18500 			 */
18501 			fsb->o_m_len = 0;
18502 			fsb->o_t_len = 0;
18503 		}
18504 	}
18505 	return (top);
18506 nospace:
18507 	if (top)
18508 		m_freem(top);
18509 	return (NULL);
18510 
18511 }
18512 
18513 /*
18514  * This is a copy of m_copym(), taking the TSO segment size/limit
18515  * constraints into account, and advancing the sndptr as it goes.
18516  */
18517 static struct mbuf *
18518 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18519 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18520 {
18521 	struct mbuf *m, *n;
18522 	int32_t soff;
18523 
18524 	m = rack->r_ctl.fsb.m;
18525 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18526 		/*
18527 		 * The trailing space changed, mbufs can grow
18528 		 * at the tail but they can't shrink from
18529 		 * it, KASSERT that. Adjust the orig_m_len to
18530 		 * compensate for this change.
18531 		 */
18532 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18533 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18534 			 m,
18535 			 rack,
18536 			 (intmax_t)M_TRAILINGROOM(m),
18537 			 rack->r_ctl.fsb.o_t_len,
18538 			 rack->r_ctl.fsb.o_m_len,
18539 			 m->m_len));
18540 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18541 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18542 	}
18543 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18544 		/*
18545 		 * Mbuf shrank, trimmed off the top by an ack, our
18546 		 * offset changes.
18547 		 */
18548 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18549 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18550 			 m, m->m_len,
18551 			 rack, rack->r_ctl.fsb.o_m_len,
18552 			 rack->r_ctl.fsb.off));
18553 
18554 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18555 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18556 		else
18557 			rack->r_ctl.fsb.off = 0;
18558 		rack->r_ctl.fsb.o_m_len = m->m_len;
18559 #ifdef INVARIANTS
18560 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18561 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
18562 		      rack, m);
18563 #endif
18564 	}
18565 	soff = rack->r_ctl.fsb.off;
18566 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18567 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18568 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18569 				 __FUNCTION__,
18570 				 rack, *plen, m, m->m_len));
18571 	/* Save off the right location before we copy and advance */
18572 	*s_soff = soff;
18573 	*s_mb = rack->r_ctl.fsb.m;
18574 	n = rack_fo_base_copym(m, soff, plen,
18575 			       &rack->r_ctl.fsb,
18576 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18577 	return (n);
18578 }
18579 
18580 /* Log the buffer level */
18581 static void
18582 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18583 		     int len, struct timeval *tv,
18584 		     uint32_t cts)
18585 {
18586 	uint32_t p_rate = 0, p_queue = 0, err = 0;
18587 	union tcp_log_stackspecific log;
18588 
18589 #ifdef RATELIMIT
18590 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18591 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
18592 #endif
18593 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18594 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18595 	log.u_bbr.flex1 = p_rate;
18596 	log.u_bbr.flex2 = p_queue;
18597 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18598 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18599 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18600 	log.u_bbr.flex7 = 99;
18601 	log.u_bbr.flex8 = 0;
18602 	log.u_bbr.pkts_out = err;
18603 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
18604 	log.u_bbr.timeStamp = cts;
18605 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18606 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18607 		       len, &log, false, NULL, __func__, __LINE__, tv);
18608 
18609 }
18610 
18611 static uint32_t
18612 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18613 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18614 {
18615 	uint64_t lentime = 0;
18616 #ifdef RATELIMIT
18617 	uint32_t p_rate = 0, p_queue = 0, err;
18618 	union tcp_log_stackspecific log;
18619 	uint64_t bw;
18620 
18621 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18622 	/* Failed or queue is zero */
18623 	if (err || (p_queue == 0)) {
18624 		lentime = 0;
18625 		goto out;
18626 	}
18627 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18628 	if (err) {
18629 		lentime = 0;
18630 		goto out;
18631 	}
18632 	/*
18633 	 * If we reach here we have some bytes in
18634 	 * the queue. The number returned is a value
18635 	 * between 0 and 0xffff where ffff is full
18636 	 * and 0 is empty. So how best to make this into
18637 	 * something usable?
18638 	 *
18639 	 * The "safer" way is lets take the b/w gotten
18640 	 * from the query (which should be our b/w rate)
18641 	 * and pretend that a full send (our rc_pace_max_segs)
18642 	 * is outstanding. We factor it so its as if a full
18643 	 * number of our MSS segment is terms of full
18644 	 * ethernet segments are outstanding.
18645 	 */
18646 	bw = p_rate / 8;
18647 	if (bw) {
18648 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18649 		lentime *= ETHERNET_SEGMENT_SIZE;
18650 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18651 		lentime /= bw;
18652 	} else {
18653 		/* TSNH -- KASSERT? */
18654 		lentime = 0;
18655 	}
18656 out:
18657 	if (tcp_bblogging_on(tp)) {
18658 		memset(&log, 0, sizeof(log));
18659 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18660 		log.u_bbr.flex1 = p_rate;
18661 		log.u_bbr.flex2 = p_queue;
18662 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18663 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18664 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18665 		log.u_bbr.flex7 = 99;
18666 		log.u_bbr.flex8 = 0;
18667 		log.u_bbr.pkts_out = err;
18668 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
18669 		log.u_bbr.cur_del_rate = lentime;
18670 		log.u_bbr.timeStamp = cts;
18671 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18672 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18673 			       len, &log, false, NULL, __func__, __LINE__,tv);
18674 	}
18675 #endif
18676 	return ((uint32_t)lentime);
18677 }
18678 
18679 static int
18680 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18681 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18682 {
18683 	/*
18684 	 * Enter the fast retransmit path. We are given that a sched_pin is
18685 	 * in place (if accounting is compliled in) and the cycle count taken
18686 	 * at the entry is in the ts_val. The concept her is that the rsm
18687 	 * now holds the mbuf offsets and such so we can directly transmit
18688 	 * without a lot of overhead, the len field is already set for
18689 	 * us to prohibit us from sending too much (usually its 1MSS).
18690 	 */
18691 	struct ip *ip = NULL;
18692 	struct udphdr *udp = NULL;
18693 	struct tcphdr *th = NULL;
18694 	struct mbuf *m = NULL;
18695 	struct inpcb *inp;
18696 	uint8_t *cpto;
18697 	struct tcp_log_buffer *lgb;
18698 #ifdef TCP_ACCOUNTING
18699 	uint64_t crtsc;
18700 	int cnt_thru = 1;
18701 #endif
18702 	struct tcpopt to;
18703 	u_char opt[TCP_MAXOLEN];
18704 	uint32_t hdrlen, optlen;
18705 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18706 	uint16_t flags;
18707 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
18708 	uint32_t if_hw_tsomaxsegsize;
18709 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18710 
18711 #ifdef INET6
18712 	struct ip6_hdr *ip6 = NULL;
18713 
18714 	if (rack->r_is_v6) {
18715 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18716 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18717 	} else
18718 #endif				/* INET6 */
18719 	{
18720 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18721 		hdrlen = sizeof(struct tcpiphdr);
18722 	}
18723 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18724 		goto failed;
18725 	}
18726 	if (doing_tlp) {
18727 		/* Its a TLP add the flag, it may already be there but be sure */
18728 		rsm->r_flags |= RACK_TLP;
18729 	} else {
18730 		/* If it was a TLP it is not not on this retransmit */
18731 		rsm->r_flags &= ~RACK_TLP;
18732 	}
18733 	startseq = rsm->r_start;
18734 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18735 	inp = rack->rc_inp;
18736 	to.to_flags = 0;
18737 	flags = tcp_outflags[tp->t_state];
18738 	if (flags & (TH_SYN|TH_RST)) {
18739 		goto failed;
18740 	}
18741 	if (rsm->r_flags & RACK_HAS_FIN) {
18742 		/* We can't send a FIN here */
18743 		goto failed;
18744 	}
18745 	if (flags & TH_FIN) {
18746 		/* We never send a FIN */
18747 		flags &= ~TH_FIN;
18748 	}
18749 	if (tp->t_flags & TF_RCVD_TSTMP) {
18750 		to.to_tsval = ms_cts + tp->ts_offset;
18751 		to.to_tsecr = tp->ts_recent;
18752 		to.to_flags = TOF_TS;
18753 	}
18754 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18755 	/* TCP-MD5 (RFC2385). */
18756 	if (tp->t_flags & TF_SIGNATURE)
18757 		to.to_flags |= TOF_SIGNATURE;
18758 #endif
18759 	optlen = tcp_addoptions(&to, opt);
18760 	hdrlen += optlen;
18761 	udp = rack->r_ctl.fsb.udp;
18762 	if (udp)
18763 		hdrlen += sizeof(struct udphdr);
18764 	if (rack->r_ctl.rc_pace_max_segs)
18765 		max_val = rack->r_ctl.rc_pace_max_segs;
18766 	else if (rack->rc_user_set_max_segs)
18767 		max_val = rack->rc_user_set_max_segs * segsiz;
18768 	else
18769 		max_val = len;
18770 	if ((tp->t_flags & TF_TSO) &&
18771 	    V_tcp_do_tso &&
18772 	    (len > segsiz) &&
18773 	    (tp->t_port == 0))
18774 		tso = 1;
18775 #ifdef INET6
18776 	if (MHLEN < hdrlen + max_linkhdr)
18777 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18778 	else
18779 #endif
18780 		m = m_gethdr(M_NOWAIT, MT_DATA);
18781 	if (m == NULL)
18782 		goto failed;
18783 	m->m_data += max_linkhdr;
18784 	m->m_len = hdrlen;
18785 	th = rack->r_ctl.fsb.th;
18786 	/* Establish the len to send */
18787 	if (len > max_val)
18788 		len = max_val;
18789 	if ((tso) && (len + optlen > segsiz)) {
18790 		uint32_t if_hw_tsomax;
18791 		int32_t max_len;
18792 
18793 		/* extract TSO information */
18794 		if_hw_tsomax = tp->t_tsomax;
18795 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18796 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18797 		/*
18798 		 * Check if we should limit by maximum payload
18799 		 * length:
18800 		 */
18801 		if (if_hw_tsomax != 0) {
18802 			/* compute maximum TSO length */
18803 			max_len = (if_hw_tsomax - hdrlen -
18804 				   max_linkhdr);
18805 			if (max_len <= 0) {
18806 				goto failed;
18807 			} else if (len > max_len) {
18808 				len = max_len;
18809 			}
18810 		}
18811 		if (len <= segsiz) {
18812 			/*
18813 			 * In case there are too many small fragments don't
18814 			 * use TSO:
18815 			 */
18816 			tso = 0;
18817 		}
18818 	} else {
18819 		tso = 0;
18820 	}
18821 	if ((tso == 0) && (len > segsiz))
18822 		len = segsiz;
18823 	(void)tcp_get_usecs(tv);
18824 	if ((len == 0) ||
18825 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
18826 		goto failed;
18827 	}
18828 	th->th_seq = htonl(rsm->r_start);
18829 	th->th_ack = htonl(tp->rcv_nxt);
18830 	/*
18831 	 * The PUSH bit should only be applied
18832 	 * if the full retransmission is made. If
18833 	 * we are sending less than this is the
18834 	 * left hand edge and should not have
18835 	 * the PUSH bit.
18836 	 */
18837 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
18838 	    (len == (rsm->r_end - rsm->r_start)))
18839 		flags |= TH_PUSH;
18840 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18841 	if (th->th_win == 0) {
18842 		tp->t_sndzerowin++;
18843 		tp->t_flags |= TF_RXWIN0SENT;
18844 	} else
18845 		tp->t_flags &= ~TF_RXWIN0SENT;
18846 	if (rsm->r_flags & RACK_TLP) {
18847 		/*
18848 		 * TLP should not count in retran count, but
18849 		 * in its own bin
18850 		 */
18851 		counter_u64_add(rack_tlp_retran, 1);
18852 		counter_u64_add(rack_tlp_retran_bytes, len);
18853 	} else {
18854 		tp->t_sndrexmitpack++;
18855 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18856 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18857 	}
18858 #ifdef STATS
18859 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18860 				 len);
18861 #endif
18862 	if (rsm->m == NULL)
18863 		goto failed;
18864 	if (rsm->m &&
18865 	    ((rsm->orig_m_len != rsm->m->m_len) ||
18866 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18867 		/* Fix up the orig_m_len and possibly the mbuf offset */
18868 		rack_adjust_orig_mlen(rsm);
18869 	}
18870 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18871 	if (len <= segsiz) {
18872 		/*
18873 		 * Must have ran out of mbufs for the copy
18874 		 * shorten it to no longer need tso. Lets
18875 		 * not put on sendalot since we are low on
18876 		 * mbufs.
18877 		 */
18878 		tso = 0;
18879 	}
18880 	if ((m->m_next == NULL) || (len <= 0)){
18881 		goto failed;
18882 	}
18883 	if (udp) {
18884 		if (rack->r_is_v6)
18885 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
18886 		else
18887 			ulen = hdrlen + len - sizeof(struct ip);
18888 		udp->uh_ulen = htons(ulen);
18889 	}
18890 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18891 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18892 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18893 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
18894 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18895 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18896 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18897 #ifdef INET6
18898 		if (rack->r_is_v6) {
18899 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18900 		    ip6->ip6_flow |= htonl(ect << 20);
18901 		}
18902 		else
18903 #endif
18904 		{
18905 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
18906 		    ip->ip_tos |= ect;
18907 		}
18908 	}
18909 	if (rack->r_ctl.crte != NULL) {
18910 		/* See if we can send via the hw queue */
18911 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18912 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
18913 		if (slot == 0)
18914 			ip_sendflag = 0;
18915 	}
18916 	tcp_set_flags(th, flags);
18917 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18918 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18919 	if (to.to_flags & TOF_SIGNATURE) {
18920 		/*
18921 		 * Calculate MD5 signature and put it into the place
18922 		 * determined before.
18923 		 * NOTE: since TCP options buffer doesn't point into
18924 		 * mbuf's data, calculate offset and use it.
18925 		 */
18926 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18927 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18928 			/*
18929 			 * Do not send segment if the calculation of MD5
18930 			 * digest has failed.
18931 			 */
18932 			goto failed;
18933 		}
18934 	}
18935 #endif
18936 #ifdef INET6
18937 	if (rack->r_is_v6) {
18938 		if (tp->t_port) {
18939 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18940 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18941 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18942 			th->th_sum = htons(0);
18943 			UDPSTAT_INC(udps_opackets);
18944 		} else {
18945 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18946 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18947 			th->th_sum = in6_cksum_pseudo(ip6,
18948 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18949 						      0);
18950 		}
18951 	}
18952 #endif
18953 #if defined(INET6) && defined(INET)
18954 	else
18955 #endif
18956 #ifdef INET
18957 	{
18958 		if (tp->t_port) {
18959 			m->m_pkthdr.csum_flags = CSUM_UDP;
18960 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18961 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18962 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18963 			th->th_sum = htons(0);
18964 			UDPSTAT_INC(udps_opackets);
18965 		} else {
18966 			m->m_pkthdr.csum_flags = CSUM_TCP;
18967 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18968 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18969 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18970 									IPPROTO_TCP + len + optlen));
18971 		}
18972 		/* IP version must be set here for ipv4/ipv6 checking later */
18973 		KASSERT(ip->ip_v == IPVERSION,
18974 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18975 	}
18976 #endif
18977 	if (tso) {
18978 		/*
18979 		 * Here we use segsiz since we have no added options besides
18980 		 * any standard timestamp options (no DSACKs or SACKS are sent
18981 		 * via either fast-path).
18982 		 */
18983 		KASSERT(len > segsiz,
18984 			("%s: len <= tso_segsz tp:%p", __func__, tp));
18985 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18986 		m->m_pkthdr.tso_segsz = segsiz;
18987 	}
18988 #ifdef INET6
18989 	if (rack->r_is_v6) {
18990 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18991 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18992 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18993 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18994 		else
18995 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18996 	}
18997 #endif
18998 #if defined(INET) && defined(INET6)
18999 	else
19000 #endif
19001 #ifdef INET
19002 	{
19003 		ip->ip_len = htons(m->m_pkthdr.len);
19004 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19005 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19006 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19007 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19008 				ip->ip_off |= htons(IP_DF);
19009 			}
19010 		} else {
19011 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19012 		}
19013 	}
19014 #endif
19015 	if (doing_tlp == 0) {
19016 		/* Set we retransmitted */
19017 		rack->rc_gp_saw_rec = 1;
19018 	} else {
19019 		/* Its a TLP set ca or ss */
19020 		if (tp->snd_cwnd > tp->snd_ssthresh) {
19021 			/* Set we sent in CA */
19022 			rack->rc_gp_saw_ca = 1;
19023 		} else {
19024 			/* Set we sent in SS */
19025 			rack->rc_gp_saw_ss = 1;
19026 		}
19027 	}
19028 	/* Time to copy in our header */
19029 	cpto = mtod(m, uint8_t *);
19030 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19031 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19032 	if (optlen) {
19033 		bcopy(opt, th + 1, optlen);
19034 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19035 	} else {
19036 		th->th_off = sizeof(struct tcphdr) >> 2;
19037 	}
19038 	if (tcp_bblogging_on(rack->rc_tp)) {
19039 		union tcp_log_stackspecific log;
19040 
19041 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
19042 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
19043 			counter_u64_add(rack_collapsed_win_rxt, 1);
19044 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
19045 		}
19046 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19047 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19048 		if (rack->rack_no_prr)
19049 			log.u_bbr.flex1 = 0;
19050 		else
19051 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19052 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19053 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19054 		log.u_bbr.flex4 = max_val;
19055 		/* Save off the early/late values */
19056 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19057 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19058 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19059 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19060 		if (doing_tlp == 0)
19061 			log.u_bbr.flex8 = 1;
19062 		else
19063 			log.u_bbr.flex8 = 2;
19064 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19065 		log.u_bbr.flex7 = 55;
19066 		log.u_bbr.pkts_out = tp->t_maxseg;
19067 		log.u_bbr.timeStamp = cts;
19068 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19069 		if (rsm && (rsm->r_rtr_cnt > 0)) {
19070 			/*
19071 			 * When we have a retransmit we want to log the
19072 			 * burst at send and flight at send from before.
19073 			 */
19074 			log.u_bbr.flex5 = rsm->r_fas;
19075 			log.u_bbr.bbr_substate = rsm->r_bas;
19076 		} else {
19077 			/*
19078 			 * This is currently unlikely until we do the
19079 			 * packet pair probes but I will add it for completeness.
19080 			 */
19081 			log.u_bbr.flex5 = log.u_bbr.inflight;
19082 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19083 		}
19084 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19085 		log.u_bbr.delivered = 0;
19086 		log.u_bbr.rttProp = (uint64_t)rsm;
19087 		log.u_bbr.delRate = rsm->r_flags;
19088 		log.u_bbr.delRate <<= 31;
19089 		log.u_bbr.delRate |= rack->r_must_retran;
19090 		log.u_bbr.delRate <<= 1;
19091 		log.u_bbr.delRate |= 1;
19092 		log.u_bbr.pkt_epoch = __LINE__;
19093 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19094 				     len, &log, false, NULL, __func__, __LINE__, tv);
19095 	} else
19096 		lgb = NULL;
19097 	if ((rack->r_ctl.crte != NULL) &&
19098 	    tcp_bblogging_on(tp)) {
19099 		rack_log_queue_level(tp, rack, len, tv, cts);
19100 	}
19101 #ifdef INET6
19102 	if (rack->r_is_v6) {
19103 		error = ip6_output(m, inp->in6p_outputopts,
19104 				   &inp->inp_route6,
19105 				   ip_sendflag, NULL, NULL, inp);
19106 	}
19107 	else
19108 #endif
19109 #ifdef INET
19110 	{
19111 		error = ip_output(m, NULL,
19112 				  &inp->inp_route,
19113 				  ip_sendflag, 0, inp);
19114 	}
19115 #endif
19116 	m = NULL;
19117 	if (lgb) {
19118 		lgb->tlb_errno = error;
19119 		lgb = NULL;
19120 	}
19121 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
19122 	tp->snd_nxt = tp->snd_max;
19123 	if (error) {
19124 		goto failed;
19125 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
19126 		rack->rc_hw_nobuf = 0;
19127 		rack->r_ctl.rc_agg_delayed = 0;
19128 		rack->r_early = 0;
19129 		rack->r_late = 0;
19130 		rack->r_ctl.rc_agg_early = 0;
19131 	}
19132 
19133 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
19134 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
19135 	if (doing_tlp) {
19136 		rack->rc_tlp_in_progress = 1;
19137 		rack->r_ctl.rc_tlp_cnt_out++;
19138 	}
19139 	if (error == 0) {
19140 		counter_u64_add(rack_total_bytes, len);
19141 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
19142 		if (doing_tlp) {
19143 			rack->rc_last_sent_tlp_past_cumack = 0;
19144 			rack->rc_last_sent_tlp_seq_valid = 1;
19145 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
19146 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
19147 		}
19148 		if (rack->r_ctl.rc_prr_sndcnt >= len)
19149 			rack->r_ctl.rc_prr_sndcnt -= len;
19150 		else
19151 			rack->r_ctl.rc_prr_sndcnt = 0;
19152 	}
19153 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19154 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19155 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19156 		rack->r_ctl.retran_during_recovery += len;
19157 	{
19158 		int idx;
19159 
19160 		idx = (len / segsiz) + 3;
19161 		if (idx >= TCP_MSS_ACCT_ATIMER)
19162 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19163 		else
19164 			counter_u64_add(rack_out_size[idx], 1);
19165 	}
19166 	if (tp->t_rtttime == 0) {
19167 		tp->t_rtttime = ticks;
19168 		tp->t_rtseq = startseq;
19169 		KMOD_TCPSTAT_INC(tcps_segstimed);
19170 	}
19171 	counter_u64_add(rack_fto_rsm_send, 1);
19172 	if (error && (error == ENOBUFS)) {
19173 		if (rack->r_ctl.crte != NULL) {
19174 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
19175 			if (tcp_bblogging_on(rack->rc_tp))
19176 				rack_log_queue_level(tp, rack, len, tv, cts);
19177 		} else
19178 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
19179 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19180 		if (rack->rc_enobuf < 0x7f)
19181 			rack->rc_enobuf++;
19182 		if (slot < (10 * HPTS_USEC_IN_MSEC))
19183 			slot = 10 * HPTS_USEC_IN_MSEC;
19184 		if (rack->r_ctl.crte != NULL) {
19185 			counter_u64_add(rack_saw_enobuf_hw, 1);
19186 			tcp_rl_log_enobuf(rack->r_ctl.crte);
19187 		}
19188 		counter_u64_add(rack_saw_enobuf, 1);
19189 	} else
19190 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
19191 	if ((slot == 0) ||
19192 	    (rack->rc_always_pace == 0) ||
19193 	    (rack->r_rr_config == 1)) {
19194 		/*
19195 		 * We have no pacing set or we
19196 		 * are using old-style rack or
19197 		 * we are overridden to use the old 1ms pacing.
19198 		 */
19199 		slot = rack->r_ctl.rc_min_to;
19200 	}
19201 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
19202 #ifdef TCP_ACCOUNTING
19203 	crtsc = get_cyclecount();
19204 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19205 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19206 	}
19207 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19208 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19209 	}
19210 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19211 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
19212 	}
19213 	sched_unpin();
19214 #endif
19215 	return (0);
19216 failed:
19217 	if (m)
19218 		m_free(m);
19219 	return (-1);
19220 }
19221 
19222 static void
19223 rack_sndbuf_autoscale(struct tcp_rack *rack)
19224 {
19225 	/*
19226 	 * Automatic sizing of send socket buffer.  Often the send buffer
19227 	 * size is not optimally adjusted to the actual network conditions
19228 	 * at hand (delay bandwidth product).  Setting the buffer size too
19229 	 * small limits throughput on links with high bandwidth and high
19230 	 * delay (eg. trans-continental/oceanic links).  Setting the
19231 	 * buffer size too big consumes too much real kernel memory,
19232 	 * especially with many connections on busy servers.
19233 	 *
19234 	 * The criteria to step up the send buffer one notch are:
19235 	 *  1. receive window of remote host is larger than send buffer
19236 	 *     (with a fudge factor of 5/4th);
19237 	 *  2. send buffer is filled to 7/8th with data (so we actually
19238 	 *     have data to make use of it);
19239 	 *  3. send buffer fill has not hit maximal automatic size;
19240 	 *  4. our send window (slow start and cogestion controlled) is
19241 	 *     larger than sent but unacknowledged data in send buffer.
19242 	 *
19243 	 * Note that the rack version moves things much faster since
19244 	 * we want to avoid hitting cache lines in the rack_fast_output()
19245 	 * path so this is called much less often and thus moves
19246 	 * the SB forward by a percentage.
19247 	 */
19248 	struct socket *so;
19249 	struct tcpcb *tp;
19250 	uint32_t sendwin, scaleup;
19251 
19252 	tp = rack->rc_tp;
19253 	so = rack->rc_inp->inp_socket;
19254 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19255 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19256 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19257 		    sbused(&so->so_snd) >=
19258 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19259 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19260 		    sendwin >= (sbused(&so->so_snd) -
19261 		    (tp->snd_nxt - tp->snd_una))) {
19262 			if (rack_autosndbuf_inc)
19263 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19264 			else
19265 				scaleup = V_tcp_autosndbuf_inc;
19266 			if (scaleup < V_tcp_autosndbuf_inc)
19267 				scaleup = V_tcp_autosndbuf_inc;
19268 			scaleup += so->so_snd.sb_hiwat;
19269 			if (scaleup > V_tcp_autosndbuf_max)
19270 				scaleup = V_tcp_autosndbuf_max;
19271 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19272 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19273 		}
19274 	}
19275 }
19276 
19277 static int
19278 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19279 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19280 {
19281 	/*
19282 	 * Enter to do fast output. We are given that the sched_pin is
19283 	 * in place (if accounting is compiled in) and the cycle count taken
19284 	 * at entry is in place in ts_val. The idea here is that
19285 	 * we know how many more bytes needs to be sent (presumably either
19286 	 * during pacing or to fill the cwnd and that was greater than
19287 	 * the max-burst). We have how much to send and all the info we
19288 	 * need to just send.
19289 	 */
19290 #ifdef INET
19291 	struct ip *ip = NULL;
19292 #endif
19293 	struct udphdr *udp = NULL;
19294 	struct tcphdr *th = NULL;
19295 	struct mbuf *m, *s_mb;
19296 	struct inpcb *inp;
19297 	uint8_t *cpto;
19298 	struct tcp_log_buffer *lgb;
19299 #ifdef TCP_ACCOUNTING
19300 	uint64_t crtsc;
19301 #endif
19302 	struct tcpopt to;
19303 	u_char opt[TCP_MAXOLEN];
19304 	uint32_t hdrlen, optlen;
19305 #ifdef TCP_ACCOUNTING
19306 	int cnt_thru = 1;
19307 #endif
19308 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19309 	uint16_t flags;
19310 	uint32_t s_soff;
19311 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19312 	uint32_t if_hw_tsomaxsegsize;
19313 	uint16_t add_flag = RACK_SENT_FP;
19314 #ifdef INET6
19315 	struct ip6_hdr *ip6 = NULL;
19316 
19317 	if (rack->r_is_v6) {
19318 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19319 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19320 	} else
19321 #endif				/* INET6 */
19322 	{
19323 #ifdef INET
19324 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19325 		hdrlen = sizeof(struct tcpiphdr);
19326 #endif
19327 	}
19328 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19329 		m = NULL;
19330 		goto failed;
19331 	}
19332 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19333 	startseq = tp->snd_max;
19334 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19335 	inp = rack->rc_inp;
19336 	len = rack->r_ctl.fsb.left_to_send;
19337 	to.to_flags = 0;
19338 	flags = rack->r_ctl.fsb.tcp_flags;
19339 	if (tp->t_flags & TF_RCVD_TSTMP) {
19340 		to.to_tsval = ms_cts + tp->ts_offset;
19341 		to.to_tsecr = tp->ts_recent;
19342 		to.to_flags = TOF_TS;
19343 	}
19344 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19345 	/* TCP-MD5 (RFC2385). */
19346 	if (tp->t_flags & TF_SIGNATURE)
19347 		to.to_flags |= TOF_SIGNATURE;
19348 #endif
19349 	optlen = tcp_addoptions(&to, opt);
19350 	hdrlen += optlen;
19351 	udp = rack->r_ctl.fsb.udp;
19352 	if (udp)
19353 		hdrlen += sizeof(struct udphdr);
19354 	if (rack->r_ctl.rc_pace_max_segs)
19355 		max_val = rack->r_ctl.rc_pace_max_segs;
19356 	else if (rack->rc_user_set_max_segs)
19357 		max_val = rack->rc_user_set_max_segs * segsiz;
19358 	else
19359 		max_val = len;
19360 	if ((tp->t_flags & TF_TSO) &&
19361 	    V_tcp_do_tso &&
19362 	    (len > segsiz) &&
19363 	    (tp->t_port == 0))
19364 		tso = 1;
19365 again:
19366 #ifdef INET6
19367 	if (MHLEN < hdrlen + max_linkhdr)
19368 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19369 	else
19370 #endif
19371 		m = m_gethdr(M_NOWAIT, MT_DATA);
19372 	if (m == NULL)
19373 		goto failed;
19374 	m->m_data += max_linkhdr;
19375 	m->m_len = hdrlen;
19376 	th = rack->r_ctl.fsb.th;
19377 	/* Establish the len to send */
19378 	if (len > max_val)
19379 		len = max_val;
19380 	if ((tso) && (len + optlen > segsiz)) {
19381 		uint32_t if_hw_tsomax;
19382 		int32_t max_len;
19383 
19384 		/* extract TSO information */
19385 		if_hw_tsomax = tp->t_tsomax;
19386 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19387 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19388 		/*
19389 		 * Check if we should limit by maximum payload
19390 		 * length:
19391 		 */
19392 		if (if_hw_tsomax != 0) {
19393 			/* compute maximum TSO length */
19394 			max_len = (if_hw_tsomax - hdrlen -
19395 				   max_linkhdr);
19396 			if (max_len <= 0) {
19397 				goto failed;
19398 			} else if (len > max_len) {
19399 				len = max_len;
19400 			}
19401 		}
19402 		if (len <= segsiz) {
19403 			/*
19404 			 * In case there are too many small fragments don't
19405 			 * use TSO:
19406 			 */
19407 			tso = 0;
19408 		}
19409 	} else {
19410 		tso = 0;
19411 	}
19412 	if ((tso == 0) && (len > segsiz))
19413 		len = segsiz;
19414 	(void)tcp_get_usecs(tv);
19415 	if ((len == 0) ||
19416 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19417 		goto failed;
19418 	}
19419 	sb_offset = tp->snd_max - tp->snd_una;
19420 	th->th_seq = htonl(tp->snd_max);
19421 	th->th_ack = htonl(tp->rcv_nxt);
19422 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19423 	if (th->th_win == 0) {
19424 		tp->t_sndzerowin++;
19425 		tp->t_flags |= TF_RXWIN0SENT;
19426 	} else
19427 		tp->t_flags &= ~TF_RXWIN0SENT;
19428 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
19429 	KMOD_TCPSTAT_INC(tcps_sndpack);
19430 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19431 #ifdef STATS
19432 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19433 				 len);
19434 #endif
19435 	if (rack->r_ctl.fsb.m == NULL)
19436 		goto failed;
19437 
19438 	/* s_mb and s_soff are saved for rack_log_output */
19439 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19440 				    &s_mb, &s_soff);
19441 	if (len <= segsiz) {
19442 		/*
19443 		 * Must have ran out of mbufs for the copy
19444 		 * shorten it to no longer need tso. Lets
19445 		 * not put on sendalot since we are low on
19446 		 * mbufs.
19447 		 */
19448 		tso = 0;
19449 	}
19450 	if (rack->r_ctl.fsb.rfo_apply_push &&
19451 	    (len == rack->r_ctl.fsb.left_to_send)) {
19452 		tcp_set_flags(th, flags | TH_PUSH);
19453 		add_flag |= RACK_HAD_PUSH;
19454 	}
19455 	if ((m->m_next == NULL) || (len <= 0)){
19456 		goto failed;
19457 	}
19458 	if (udp) {
19459 		if (rack->r_is_v6)
19460 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19461 		else
19462 			ulen = hdrlen + len - sizeof(struct ip);
19463 		udp->uh_ulen = htons(ulen);
19464 	}
19465 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19466 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19467 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19468 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
19469 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19470 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19471 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19472 #ifdef INET6
19473 		if (rack->r_is_v6) {
19474 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19475 			ip6->ip6_flow |= htonl(ect << 20);
19476 		}
19477 		else
19478 #endif
19479 		{
19480 #ifdef INET
19481 			ip->ip_tos &= ~IPTOS_ECN_MASK;
19482 			ip->ip_tos |= ect;
19483 #endif
19484 		}
19485 	}
19486 	tcp_set_flags(th, flags);
19487 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19488 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19489 	if (to.to_flags & TOF_SIGNATURE) {
19490 		/*
19491 		 * Calculate MD5 signature and put it into the place
19492 		 * determined before.
19493 		 * NOTE: since TCP options buffer doesn't point into
19494 		 * mbuf's data, calculate offset and use it.
19495 		 */
19496 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19497 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19498 			/*
19499 			 * Do not send segment if the calculation of MD5
19500 			 * digest has failed.
19501 			 */
19502 			goto failed;
19503 		}
19504 	}
19505 #endif
19506 #ifdef INET6
19507 	if (rack->r_is_v6) {
19508 		if (tp->t_port) {
19509 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19510 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19511 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19512 			th->th_sum = htons(0);
19513 			UDPSTAT_INC(udps_opackets);
19514 		} else {
19515 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19516 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19517 			th->th_sum = in6_cksum_pseudo(ip6,
19518 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19519 						      0);
19520 		}
19521 	}
19522 #endif
19523 #if defined(INET6) && defined(INET)
19524 	else
19525 #endif
19526 #ifdef INET
19527 	{
19528 		if (tp->t_port) {
19529 			m->m_pkthdr.csum_flags = CSUM_UDP;
19530 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19531 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19532 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19533 			th->th_sum = htons(0);
19534 			UDPSTAT_INC(udps_opackets);
19535 		} else {
19536 			m->m_pkthdr.csum_flags = CSUM_TCP;
19537 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19538 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19539 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19540 									IPPROTO_TCP + len + optlen));
19541 		}
19542 		/* IP version must be set here for ipv4/ipv6 checking later */
19543 		KASSERT(ip->ip_v == IPVERSION,
19544 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19545 	}
19546 #endif
19547 	if (tso) {
19548 		/*
19549 		 * Here we use segsiz since we have no added options besides
19550 		 * any standard timestamp options (no DSACKs or SACKS are sent
19551 		 * via either fast-path).
19552 		 */
19553 		KASSERT(len > segsiz,
19554 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19555 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19556 		m->m_pkthdr.tso_segsz = segsiz;
19557 	}
19558 #ifdef INET6
19559 	if (rack->r_is_v6) {
19560 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19561 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19562 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19563 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19564 		else
19565 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19566 	}
19567 #endif
19568 #if defined(INET) && defined(INET6)
19569 	else
19570 #endif
19571 #ifdef INET
19572 	{
19573 		ip->ip_len = htons(m->m_pkthdr.len);
19574 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19575 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19576 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19577 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19578 				ip->ip_off |= htons(IP_DF);
19579 			}
19580 		} else {
19581 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19582 		}
19583 	}
19584 #endif
19585 	if (tp->snd_cwnd > tp->snd_ssthresh) {
19586 		/* Set we sent in CA */
19587 		rack->rc_gp_saw_ca = 1;
19588 	} else {
19589 		/* Set we sent in SS */
19590 		rack->rc_gp_saw_ss = 1;
19591 	}
19592 	/* Time to copy in our header */
19593 	cpto = mtod(m, uint8_t *);
19594 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19595 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19596 	if (optlen) {
19597 		bcopy(opt, th + 1, optlen);
19598 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19599 	} else {
19600 		th->th_off = sizeof(struct tcphdr) >> 2;
19601 	}
19602 	if ((rack->r_ctl.crte != NULL) &&
19603 	    tcp_bblogging_on(tp)) {
19604 		rack_log_queue_level(tp, rack, len, tv, cts);
19605 	}
19606 	if (tcp_bblogging_on(rack->rc_tp)) {
19607 		union tcp_log_stackspecific log;
19608 
19609 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19610 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19611 		if (rack->rack_no_prr)
19612 			log.u_bbr.flex1 = 0;
19613 		else
19614 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19615 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19616 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19617 		log.u_bbr.flex4 = max_val;
19618 		/* Save off the early/late values */
19619 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19620 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19621 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19622 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19623 		log.u_bbr.flex8 = 0;
19624 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19625 		log.u_bbr.flex7 = 44;
19626 		log.u_bbr.pkts_out = tp->t_maxseg;
19627 		log.u_bbr.timeStamp = cts;
19628 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19629 		log.u_bbr.flex5 = log.u_bbr.inflight;
19630 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19631 		log.u_bbr.delivered = 0;
19632 		log.u_bbr.rttProp = 0;
19633 		log.u_bbr.delRate = rack->r_must_retran;
19634 		log.u_bbr.delRate <<= 1;
19635 		log.u_bbr.pkt_epoch = __LINE__;
19636 		/* For fast output no retrans so just inflight and how many mss we send */
19637 		log.u_bbr.flex5 = log.u_bbr.inflight;
19638 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19639 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19640 				     len, &log, false, NULL, __func__, __LINE__, tv);
19641 	} else
19642 		lgb = NULL;
19643 #ifdef INET6
19644 	if (rack->r_is_v6) {
19645 		error = ip6_output(m, inp->in6p_outputopts,
19646 				   &inp->inp_route6,
19647 				   0, NULL, NULL, inp);
19648 	}
19649 #endif
19650 #if defined(INET) && defined(INET6)
19651 	else
19652 #endif
19653 #ifdef INET
19654 	{
19655 		error = ip_output(m, NULL,
19656 				  &inp->inp_route,
19657 				  0, 0, inp);
19658 	}
19659 #endif
19660 	if (lgb) {
19661 		lgb->tlb_errno = error;
19662 		lgb = NULL;
19663 	}
19664 	if (error) {
19665 		*send_err = error;
19666 		m = NULL;
19667 		goto failed;
19668 	} else if (rack->rc_hw_nobuf) {
19669 		rack->rc_hw_nobuf = 0;
19670 		rack->r_ctl.rc_agg_delayed = 0;
19671 		rack->r_early = 0;
19672 		rack->r_late = 0;
19673 		rack->r_ctl.rc_agg_early = 0;
19674 	}
19675 	if ((error == 0) && (rack->lt_bw_up == 0)) {
19676 		/* Unlikely */
19677 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19678 		rack->r_ctl.lt_seq = tp->snd_una;
19679 		rack->lt_bw_up = 1;
19680 	}
19681 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19682 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19683 	m = NULL;
19684 	if (tp->snd_una == tp->snd_max) {
19685 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19686 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19687 		tp->t_acktime = ticks;
19688 	}
19689 	counter_u64_add(rack_total_bytes, len);
19690 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19691 
19692 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19693 	tot_len += len;
19694 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
19695 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19696 	tp->snd_max += len;
19697 	tp->snd_nxt = tp->snd_max;
19698 	if (rack->rc_new_rnd_needed) {
19699 		/*
19700 		 * Update the rnd to start ticking not
19701 		 * that from a time perspective all of
19702 		 * the preceding idle time is "in the round"
19703 		 */
19704 		rack->rc_new_rnd_needed = 0;
19705 		rack->r_ctl.roundends = tp->snd_max;
19706 	}
19707 	{
19708 		int idx;
19709 
19710 		idx = (len / segsiz) + 3;
19711 		if (idx >= TCP_MSS_ACCT_ATIMER)
19712 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19713 		else
19714 			counter_u64_add(rack_out_size[idx], 1);
19715 	}
19716 	if (len <= rack->r_ctl.fsb.left_to_send)
19717 		rack->r_ctl.fsb.left_to_send -= len;
19718 	else
19719 		rack->r_ctl.fsb.left_to_send = 0;
19720 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
19721 		rack->r_fast_output = 0;
19722 		rack->r_ctl.fsb.left_to_send = 0;
19723 		/* At the end of fast_output scale up the sb */
19724 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
19725 		rack_sndbuf_autoscale(rack);
19726 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
19727 	}
19728 	if (tp->t_rtttime == 0) {
19729 		tp->t_rtttime = ticks;
19730 		tp->t_rtseq = startseq;
19731 		KMOD_TCPSTAT_INC(tcps_segstimed);
19732 	}
19733 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19734 	    (max_val > len) &&
19735 	    (tso == 0)) {
19736 		max_val -= len;
19737 		len = segsiz;
19738 		th = rack->r_ctl.fsb.th;
19739 #ifdef TCP_ACCOUNTING
19740 		cnt_thru++;
19741 #endif
19742 		goto again;
19743 	}
19744 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19745 	counter_u64_add(rack_fto_send, 1);
19746 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
19747 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19748 #ifdef TCP_ACCOUNTING
19749 	crtsc = get_cyclecount();
19750 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19751 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19752 	}
19753 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19754 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19755 	}
19756 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19757 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19758 	}
19759 	sched_unpin();
19760 #endif
19761 	return (0);
19762 failed:
19763 	if (m)
19764 		m_free(m);
19765 	rack->r_fast_output = 0;
19766 	return (-1);
19767 }
19768 
19769 static inline void
19770 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19771 		       struct sockbuf *sb,
19772 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19773 		       bool hw_tls,
19774 		       uint16_t flags)
19775 {
19776 	rack->r_fast_output = 1;
19777 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19778 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19779 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19780 	rack->r_ctl.fsb.tcp_flags = flags;
19781 	rack->r_ctl.fsb.left_to_send = orig_len - len;
19782 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19783 		/* Less than a full sized pace, lets not  */
19784 		rack->r_fast_output = 0;
19785 		return;
19786 	} else {
19787 		/* Round down to the nearest pace_max_seg */
19788 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19789 	}
19790 	if (hw_tls)
19791 		rack->r_ctl.fsb.hw_tls = 1;
19792 	else
19793 		rack->r_ctl.fsb.hw_tls = 0;
19794 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19795 		("rack:%p left_to_send:%u sbavail:%u out:%u",
19796 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19797 		 (tp->snd_max - tp->snd_una)));
19798 	if (rack->r_ctl.fsb.left_to_send < segsiz)
19799 		rack->r_fast_output = 0;
19800 	else {
19801 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19802 			rack->r_ctl.fsb.rfo_apply_push = 1;
19803 		else
19804 			rack->r_ctl.fsb.rfo_apply_push = 0;
19805 	}
19806 }
19807 
19808 static uint32_t
19809 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19810 {
19811 	uint64_t min_time;
19812 	uint32_t maxlen;
19813 
19814 	min_time = (uint64_t)get_hpts_min_sleep_time();
19815 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19816 	maxlen = roundup(maxlen, segsiz);
19817 	return (maxlen);
19818 }
19819 
19820 static struct rack_sendmap *
19821 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19822 {
19823 	struct rack_sendmap *rsm = NULL;
19824 	int thresh;
19825 
19826 restart:
19827 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19828 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19829 		/* Nothing, strange turn off validity  */
19830 		rack->r_collapse_point_valid = 0;
19831 		return (NULL);
19832 	}
19833 	/* Can we send it yet? */
19834 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19835 		/*
19836 		 * Receiver window has not grown enough for
19837 		 * the segment to be put on the wire.
19838 		 */
19839 		return (NULL);
19840 	}
19841 	if (rsm->r_flags & RACK_ACKED) {
19842 		/*
19843 		 * It has been sacked, lets move to the
19844 		 * next one if possible.
19845 		 */
19846 		rack->r_ctl.last_collapse_point = rsm->r_end;
19847 		/* Are we done? */
19848 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19849 			    rack->r_ctl.high_collapse_point)) {
19850 			rack->r_collapse_point_valid = 0;
19851 			return (NULL);
19852 		}
19853 		goto restart;
19854 	}
19855 	/* Now has it been long enough ? */
19856 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
19857 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19858 		rack_log_collapse(rack, rsm->r_start,
19859 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19860 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
19861 		return (rsm);
19862 	}
19863 	/* Not enough time */
19864 	rack_log_collapse(rack, rsm->r_start,
19865 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19866 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
19867 	return (NULL);
19868 }
19869 
19870 static inline void
19871 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19872 {
19873 	if ((rack->full_size_rxt == 0) &&
19874 	    (rack->shape_rxt_to_pacing_min == 0) &&
19875 	    (*len >= segsiz)) {
19876 		*len = segsiz;
19877 	} else if (rack->shape_rxt_to_pacing_min &&
19878 		 rack->gp_ready) {
19879 		/* We use pacing min as shaping len req */
19880 		uint32_t maxlen;
19881 
19882 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19883 		if (*len > maxlen)
19884 			*len = maxlen;
19885 	} else {
19886 		/*
19887 		 * The else is full_size_rxt is on so send it all
19888 		 * note we do need to check this for exceeding
19889 		 * our max segment size due to the fact that
19890 		 * we do sometimes merge chunks together i.e.
19891 		 * we cannot just assume that we will never have
19892 		 * a chunk greater than pace_max_seg
19893 		 */
19894 		if (*len > pace_max_seg)
19895 			*len = pace_max_seg;
19896 	}
19897 }
19898 
19899 static int
19900 rack_output(struct tcpcb *tp)
19901 {
19902 	struct socket *so;
19903 	uint32_t recwin;
19904 	uint32_t sb_offset, s_moff = 0;
19905 	int32_t len, error = 0;
19906 	uint16_t flags;
19907 	struct mbuf *m, *s_mb = NULL;
19908 	struct mbuf *mb;
19909 	uint32_t if_hw_tsomaxsegcount = 0;
19910 	uint32_t if_hw_tsomaxsegsize;
19911 	int32_t segsiz, minseg;
19912 	long tot_len_this_send = 0;
19913 #ifdef INET
19914 	struct ip *ip = NULL;
19915 #endif
19916 	struct udphdr *udp = NULL;
19917 	struct tcp_rack *rack;
19918 	struct tcphdr *th;
19919 	uint8_t pass = 0;
19920 	uint8_t mark = 0;
19921 	uint8_t check_done = 0;
19922 	uint8_t wanted_cookie = 0;
19923 	u_char opt[TCP_MAXOLEN];
19924 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
19925 	uint32_t rack_seq;
19926 
19927 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19928 	unsigned ipsec_optlen = 0;
19929 
19930 #endif
19931 	int32_t idle, sendalot;
19932 	int32_t sub_from_prr = 0;
19933 	volatile int32_t sack_rxmit;
19934 	struct rack_sendmap *rsm = NULL;
19935 	int32_t tso, mtu;
19936 	struct tcpopt to;
19937 	int32_t slot = 0;
19938 	int32_t sup_rack = 0;
19939 	uint32_t cts, ms_cts, delayed, early;
19940 	uint16_t add_flag = RACK_SENT_SP;
19941 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19942 	uint8_t doing_tlp = 0;
19943 	uint32_t cwnd_to_use, pace_max_seg;
19944 	int32_t do_a_prefetch = 0;
19945 	int32_t prefetch_rsm = 0;
19946 	int32_t orig_len = 0;
19947 	struct timeval tv;
19948 	int32_t prefetch_so_done = 0;
19949 	struct tcp_log_buffer *lgb;
19950 	struct inpcb *inp = tptoinpcb(tp);
19951 	struct sockbuf *sb;
19952 	uint64_t ts_val = 0;
19953 #ifdef TCP_ACCOUNTING
19954 	uint64_t crtsc;
19955 #endif
19956 #ifdef INET6
19957 	struct ip6_hdr *ip6 = NULL;
19958 	int32_t isipv6;
19959 #endif
19960 	bool hpts_calling, hw_tls = false;
19961 
19962 	NET_EPOCH_ASSERT();
19963 	INP_WLOCK_ASSERT(inp);
19964 
19965 	/* setup and take the cache hits here */
19966 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19967 #ifdef TCP_ACCOUNTING
19968 	sched_pin();
19969 	ts_val = get_cyclecount();
19970 #endif
19971 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19972 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
19973 #ifdef TCP_OFFLOAD
19974 	if (tp->t_flags & TF_TOE) {
19975 #ifdef TCP_ACCOUNTING
19976 		sched_unpin();
19977 #endif
19978 		return (tcp_offload_output(tp));
19979 	}
19980 #endif
19981 	if (rack->rack_deferred_inited == 0) {
19982 		/*
19983 		 * If we are the connecting socket we will
19984 		 * hit rack_init() when no sequence numbers
19985 		 * are setup. This makes it so we must defer
19986 		 * some initialization. Call that now.
19987 		 */
19988 		rack_deferred_init(tp, rack);
19989 	}
19990 	/*
19991 	 * For TFO connections in SYN_RECEIVED, only allow the initial
19992 	 * SYN|ACK and those sent by the retransmit timer.
19993 	 */
19994 	if (IS_FASTOPEN(tp->t_flags) &&
19995 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
19996 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19997 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19998 #ifdef TCP_ACCOUNTING
19999 		sched_unpin();
20000 #endif
20001 		return (0);
20002 	}
20003 #ifdef INET6
20004 	if (rack->r_state) {
20005 		/* Use the cache line loaded if possible */
20006 		isipv6 = rack->r_is_v6;
20007 	} else {
20008 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
20009 	}
20010 #endif
20011 	early = 0;
20012 	cts = tcp_get_usecs(&tv);
20013 	ms_cts = tcp_tv_to_mssectick(&tv);
20014 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
20015 	    tcp_in_hpts(rack->rc_tp)) {
20016 		/*
20017 		 * We are on the hpts for some timer but not hptsi output.
20018 		 * Remove from the hpts unconditionally.
20019 		 */
20020 		rack_timer_cancel(tp, rack, cts, __LINE__);
20021 	}
20022 	/* Are we pacing and late? */
20023 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20024 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
20025 		/* We are delayed */
20026 		delayed = cts - rack->r_ctl.rc_last_output_to;
20027 	} else {
20028 		delayed = 0;
20029 	}
20030 	/* Do the timers, which may override the pacer */
20031 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
20032 		int retval;
20033 
20034 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
20035 					     &doing_tlp);
20036 		if (retval != 0) {
20037 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
20038 #ifdef TCP_ACCOUNTING
20039 			sched_unpin();
20040 #endif
20041 			/*
20042 			 * If timers want tcp_drop(), then pass error out,
20043 			 * otherwise suppress it.
20044 			 */
20045 			return (retval < 0 ? retval : 0);
20046 		}
20047 	}
20048 	if (rack->rc_in_persist) {
20049 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20050 			/* Timer is not running */
20051 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20052 		}
20053 #ifdef TCP_ACCOUNTING
20054 		sched_unpin();
20055 #endif
20056 		return (0);
20057 	}
20058 	if ((rack->rc_ack_required == 1) &&
20059 	    (rack->r_timer_override == 0)){
20060 		/* A timeout occurred and no ack has arrived */
20061 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20062 			/* Timer is not running */
20063 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20064 		}
20065 #ifdef TCP_ACCOUNTING
20066 		sched_unpin();
20067 #endif
20068 		return (0);
20069 	}
20070 	if ((rack->r_timer_override) ||
20071 	    (rack->rc_ack_can_sendout_data) ||
20072 	    (delayed) ||
20073 	    (tp->t_state < TCPS_ESTABLISHED)) {
20074 		rack->rc_ack_can_sendout_data = 0;
20075 		if (tcp_in_hpts(rack->rc_tp))
20076 			tcp_hpts_remove(rack->rc_tp);
20077 	} else if (tcp_in_hpts(rack->rc_tp)) {
20078 		/*
20079 		 * On the hpts you can't pass even if ACKNOW is on, we will
20080 		 * when the hpts fires.
20081 		 */
20082 #ifdef TCP_ACCOUNTING
20083 		crtsc = get_cyclecount();
20084 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20085 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
20086 		}
20087 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20088 			tp->tcp_cnt_counters[SND_BLOCKED]++;
20089 		}
20090 		sched_unpin();
20091 #endif
20092 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
20093 		return (0);
20094 	}
20095 	/* Finish out both pacing early and late accounting */
20096 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20097 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
20098 		early = rack->r_ctl.rc_last_output_to - cts;
20099 	} else
20100 		early = 0;
20101 	if (delayed) {
20102 		rack->r_ctl.rc_agg_delayed += delayed;
20103 		rack->r_late = 1;
20104 	} else if (early) {
20105 		rack->r_ctl.rc_agg_early += early;
20106 		rack->r_early = 1;
20107 	}
20108 	/* Now that early/late accounting is done turn off the flag */
20109 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
20110 	rack->r_wanted_output = 0;
20111 	rack->r_timer_override = 0;
20112 	if ((tp->t_state != rack->r_state) &&
20113 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
20114 		rack_set_state(tp, rack);
20115 	}
20116 	if ((rack->r_fast_output) &&
20117 	    (doing_tlp == 0) &&
20118 	    (tp->rcv_numsacks == 0)) {
20119 		int ret;
20120 
20121 		error = 0;
20122 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
20123 		if (ret >= 0)
20124 			return(ret);
20125 		else if (error) {
20126 			inp = rack->rc_inp;
20127 			so = inp->inp_socket;
20128 			sb = &so->so_snd;
20129 			goto nomore;
20130 		}
20131 	}
20132 	inp = rack->rc_inp;
20133 	/*
20134 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
20135 	 * only allow the initial SYN or SYN|ACK and those sent
20136 	 * by the retransmit timer.
20137 	 */
20138 	if (IS_FASTOPEN(tp->t_flags) &&
20139 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
20140 	     (tp->t_state == TCPS_SYN_SENT)) &&
20141 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
20142 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
20143 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20144 		so = inp->inp_socket;
20145 		sb = &so->so_snd;
20146 		goto just_return_nolock;
20147 	}
20148 	/*
20149 	 * Determine length of data that should be transmitted, and flags
20150 	 * that will be used. If there is some data or critical controls
20151 	 * (SYN, RST) to send, then transmit; otherwise, investigate
20152 	 * further.
20153 	 */
20154 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
20155 	if (tp->t_idle_reduce) {
20156 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
20157 			rack_cc_after_idle(rack, tp);
20158 	}
20159 	tp->t_flags &= ~TF_LASTIDLE;
20160 	if (idle) {
20161 		if (tp->t_flags & TF_MORETOCOME) {
20162 			tp->t_flags |= TF_LASTIDLE;
20163 			idle = 0;
20164 		}
20165 	}
20166 	if ((tp->snd_una == tp->snd_max) &&
20167 	    rack->r_ctl.rc_went_idle_time &&
20168 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
20169 		idle = cts - rack->r_ctl.rc_went_idle_time;
20170 		if (idle > rack_min_probertt_hold) {
20171 			/* Count as a probe rtt */
20172 			if (rack->in_probe_rtt == 0) {
20173 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
20174 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
20175 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
20176 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
20177 			} else {
20178 				rack_exit_probertt(rack, cts);
20179 			}
20180 		}
20181 		idle = 0;
20182 	}
20183 	if (rack_use_fsb &&
20184 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
20185 	    (rack->r_fsb_inited == 0) &&
20186 	    (rack->r_state != TCPS_CLOSED))
20187 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
20188 again:
20189 	/*
20190 	 * If we've recently taken a timeout, snd_max will be greater than
20191 	 * snd_nxt.  There may be SACK information that allows us to avoid
20192 	 * resending already delivered data.  Adjust snd_nxt accordingly.
20193 	 */
20194 	sendalot = 0;
20195 	cts = tcp_get_usecs(&tv);
20196 	ms_cts = tcp_tv_to_mssectick(&tv);
20197 	tso = 0;
20198 	mtu = 0;
20199 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20200 	minseg = segsiz;
20201 	if (rack->r_ctl.rc_pace_max_segs == 0)
20202 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20203 	else
20204 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20205 	sb_offset = tp->snd_max - tp->snd_una;
20206 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20207 	flags = tcp_outflags[tp->t_state];
20208 	while (rack->rc_free_cnt < rack_free_cache) {
20209 		rsm = rack_alloc(rack);
20210 		if (rsm == NULL) {
20211 			if (hpts_calling)
20212 				/* Retry in a ms */
20213 				slot = (1 * HPTS_USEC_IN_MSEC);
20214 			so = inp->inp_socket;
20215 			sb = &so->so_snd;
20216 			goto just_return_nolock;
20217 		}
20218 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20219 		rack->rc_free_cnt++;
20220 		rsm = NULL;
20221 	}
20222 	sack_rxmit = 0;
20223 	len = 0;
20224 	rsm = NULL;
20225 	if (flags & TH_RST) {
20226 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
20227 		so = inp->inp_socket;
20228 		sb = &so->so_snd;
20229 		goto send;
20230 	}
20231 	if (rack->r_ctl.rc_resend) {
20232 		/* Retransmit timer */
20233 		rsm = rack->r_ctl.rc_resend;
20234 		rack->r_ctl.rc_resend = NULL;
20235 		len = rsm->r_end - rsm->r_start;
20236 		sack_rxmit = 1;
20237 		sendalot = 0;
20238 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20239 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20240 			 __func__, __LINE__,
20241 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20242 		sb_offset = rsm->r_start - tp->snd_una;
20243 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20244 	} else if (rack->r_collapse_point_valid &&
20245 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20246 		/*
20247 		 * If an RSM is returned then enough time has passed
20248 		 * for us to retransmit it. Move up the collapse point,
20249 		 * since this rsm has its chance to retransmit now.
20250 		 */
20251 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20252 		rack->r_ctl.last_collapse_point = rsm->r_end;
20253 		/* Are we done? */
20254 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20255 			    rack->r_ctl.high_collapse_point))
20256 			rack->r_collapse_point_valid = 0;
20257 		sack_rxmit = 1;
20258 		/* We are not doing a TLP */
20259 		doing_tlp = 0;
20260 		len = rsm->r_end - rsm->r_start;
20261 		sb_offset = rsm->r_start - tp->snd_una;
20262 		sendalot = 0;
20263 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20264 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20265 		/* We have a retransmit that takes precedence */
20266 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20267 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20268 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20269 			/* Enter recovery if not induced by a time-out */
20270 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20271 		}
20272 #ifdef INVARIANTS
20273 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20274 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20275 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
20276 		}
20277 #endif
20278 		len = rsm->r_end - rsm->r_start;
20279 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20280 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20281 			 __func__, __LINE__,
20282 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20283 		sb_offset = rsm->r_start - tp->snd_una;
20284 		sendalot = 0;
20285 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20286 		if (len > 0) {
20287 			sack_rxmit = 1;
20288 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20289 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20290 					 min(len, segsiz));
20291 		}
20292 	} else if (rack->r_ctl.rc_tlpsend) {
20293 		/* Tail loss probe */
20294 		long cwin;
20295 		long tlen;
20296 
20297 		/*
20298 		 * Check if we can do a TLP with a RACK'd packet
20299 		 * this can happen if we are not doing the rack
20300 		 * cheat and we skipped to a TLP and it
20301 		 * went off.
20302 		 */
20303 		rsm = rack->r_ctl.rc_tlpsend;
20304 		/* We are doing a TLP make sure the flag is preent */
20305 		rsm->r_flags |= RACK_TLP;
20306 		rack->r_ctl.rc_tlpsend = NULL;
20307 		sack_rxmit = 1;
20308 		tlen = rsm->r_end - rsm->r_start;
20309 		if (tlen > segsiz)
20310 			tlen = segsiz;
20311 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20312 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20313 			 __func__, __LINE__,
20314 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20315 		sb_offset = rsm->r_start - tp->snd_una;
20316 		cwin = min(tp->snd_wnd, tlen);
20317 		len = cwin;
20318 	}
20319 	if (rack->r_must_retran &&
20320 	    (doing_tlp == 0) &&
20321 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20322 	    (rsm == NULL)) {
20323 		/*
20324 		 * There are two different ways that we
20325 		 * can get into this block:
20326 		 * a) This is a non-sack connection, we had a time-out
20327 		 *    and thus r_must_retran was set and everything
20328 		 *    left outstanding as been marked for retransmit.
20329 		 * b) The MTU of the path shrank, so that everything
20330 		 *    was marked to be retransmitted with the smaller
20331 		 *    mtu and r_must_retran was set.
20332 		 *
20333 		 * This means that we expect the sendmap (outstanding)
20334 		 * to all be marked must. We can use the tmap to
20335 		 * look at them.
20336 		 *
20337 		 */
20338 		int sendwin, flight;
20339 
20340 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20341 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20342 		if (flight >= sendwin) {
20343 			/*
20344 			 * We can't send yet.
20345 			 */
20346 			so = inp->inp_socket;
20347 			sb = &so->so_snd;
20348 			goto just_return_nolock;
20349 		}
20350 		/*
20351 		 * This is the case a/b mentioned above. All
20352 		 * outstanding/not-acked should be marked.
20353 		 * We can use the tmap to find them.
20354 		 */
20355 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20356 		if (rsm == NULL) {
20357 			/* TSNH */
20358 			rack->r_must_retran = 0;
20359 			rack->r_ctl.rc_out_at_rto = 0;
20360 			so = inp->inp_socket;
20361 			sb = &so->so_snd;
20362 			goto just_return_nolock;
20363 		}
20364 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20365 			/*
20366 			 * The first one does not have the flag, did we collapse
20367 			 * further up in our list?
20368 			 */
20369 			rack->r_must_retran = 0;
20370 			rack->r_ctl.rc_out_at_rto = 0;
20371 			rsm = NULL;
20372 			sack_rxmit = 0;
20373 		} else {
20374 			sack_rxmit = 1;
20375 			len = rsm->r_end - rsm->r_start;
20376 			sb_offset = rsm->r_start - tp->snd_una;
20377 			sendalot = 0;
20378 			if ((rack->full_size_rxt == 0) &&
20379 			    (rack->shape_rxt_to_pacing_min == 0) &&
20380 			    (len >= segsiz))
20381 				len = segsiz;
20382 			else if (rack->shape_rxt_to_pacing_min &&
20383 				 rack->gp_ready) {
20384 				/* We use pacing min as shaping len req */
20385 				uint32_t maxlen;
20386 
20387 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20388 				if (len > maxlen)
20389 					len = maxlen;
20390 			}
20391 			/*
20392 			 * Delay removing the flag RACK_MUST_RXT so
20393 			 * that the fastpath for retransmit will
20394 			 * work with this rsm.
20395 			 */
20396 		}
20397 	}
20398 	/*
20399 	 * Enforce a connection sendmap count limit if set
20400 	 * as long as we are not retransmiting.
20401 	 */
20402 	if ((rsm == NULL) &&
20403 	    (rack->do_detection == 0) &&
20404 	    (V_tcp_map_entries_limit > 0) &&
20405 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20406 		counter_u64_add(rack_to_alloc_limited, 1);
20407 		if (!rack->alloc_limit_reported) {
20408 			rack->alloc_limit_reported = 1;
20409 			counter_u64_add(rack_alloc_limited_conns, 1);
20410 		}
20411 		so = inp->inp_socket;
20412 		sb = &so->so_snd;
20413 		goto just_return_nolock;
20414 	}
20415 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20416 		/* we are retransmitting the fin */
20417 		len--;
20418 		if (len) {
20419 			/*
20420 			 * When retransmitting data do *not* include the
20421 			 * FIN. This could happen from a TLP probe.
20422 			 */
20423 			flags &= ~TH_FIN;
20424 		}
20425 	}
20426 	if (rsm && rack->r_fsb_inited &&
20427 	    rack_use_rsm_rfo &&
20428 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20429 		int ret;
20430 
20431 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20432 		if (ret == 0)
20433 			return (0);
20434 	}
20435 	so = inp->inp_socket;
20436 	sb = &so->so_snd;
20437 	if (do_a_prefetch == 0) {
20438 		kern_prefetch(sb, &do_a_prefetch);
20439 		do_a_prefetch = 1;
20440 	}
20441 #ifdef NETFLIX_SHARED_CWND
20442 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20443 	    rack->rack_enable_scwnd) {
20444 		/* We are doing cwnd sharing */
20445 		if (rack->gp_ready &&
20446 		    (rack->rack_attempted_scwnd == 0) &&
20447 		    (rack->r_ctl.rc_scw == NULL) &&
20448 		    tp->t_lib) {
20449 			/* The pcbid is in, lets make an attempt */
20450 			counter_u64_add(rack_try_scwnd, 1);
20451 			rack->rack_attempted_scwnd = 1;
20452 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20453 								   &rack->r_ctl.rc_scw_index,
20454 								   segsiz);
20455 		}
20456 		if (rack->r_ctl.rc_scw &&
20457 		    (rack->rack_scwnd_is_idle == 1) &&
20458 		    sbavail(&so->so_snd)) {
20459 			/* we are no longer out of data */
20460 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20461 			rack->rack_scwnd_is_idle = 0;
20462 		}
20463 		if (rack->r_ctl.rc_scw) {
20464 			/* First lets update and get the cwnd */
20465 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20466 										       rack->r_ctl.rc_scw_index,
20467 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
20468 		}
20469 	}
20470 #endif
20471 	/*
20472 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20473 	 * state flags.
20474 	 */
20475 	if (tp->t_flags & TF_NEEDFIN)
20476 		flags |= TH_FIN;
20477 	if (tp->t_flags & TF_NEEDSYN)
20478 		flags |= TH_SYN;
20479 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20480 		void *end_rsm;
20481 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20482 		if (end_rsm)
20483 			kern_prefetch(end_rsm, &prefetch_rsm);
20484 		prefetch_rsm = 1;
20485 	}
20486 	SOCKBUF_LOCK(sb);
20487 	/*
20488 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
20489 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
20490 	 * negative length.  This can also occur when TCP opens up its
20491 	 * congestion window while receiving additional duplicate acks after
20492 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
20493 	 * the fast-retransmit.
20494 	 *
20495 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
20496 	 * set to snd_una, the sb_offset will be 0, and the length may wind
20497 	 * up 0.
20498 	 *
20499 	 * If sack_rxmit is true we are retransmitting from the scoreboard
20500 	 * in which case len is already set.
20501 	 */
20502 	if ((sack_rxmit == 0) &&
20503 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
20504 		uint32_t avail;
20505 
20506 		avail = sbavail(sb);
20507 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
20508 			sb_offset = tp->snd_nxt - tp->snd_una;
20509 		else
20510 			sb_offset = 0;
20511 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20512 			if (rack->r_ctl.rc_tlp_new_data) {
20513 				/* TLP is forcing out new data */
20514 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20515 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20516 				}
20517 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20518 					if (tp->snd_wnd > sb_offset)
20519 						len = tp->snd_wnd - sb_offset;
20520 					else
20521 						len = 0;
20522 				} else {
20523 					len = rack->r_ctl.rc_tlp_new_data;
20524 				}
20525 				rack->r_ctl.rc_tlp_new_data = 0;
20526 			}  else {
20527 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20528 			}
20529 			if ((rack->r_ctl.crte == NULL) &&
20530 			    IN_FASTRECOVERY(tp->t_flags) &&
20531 			    (rack->full_size_rxt == 0) &&
20532 			    (rack->shape_rxt_to_pacing_min == 0) &&
20533 			    (len > segsiz)) {
20534 				/*
20535 				 * For prr=off, we need to send only 1 MSS
20536 				 * at a time. We do this because another sack could
20537 				 * be arriving that causes us to send retransmits and
20538 				 * we don't want to be on a long pace due to a larger send
20539 				 * that keeps us from sending out the retransmit.
20540 				 */
20541 				len = segsiz;
20542 			} else if (rack->shape_rxt_to_pacing_min &&
20543 				   rack->gp_ready) {
20544 				/* We use pacing min as shaping len req */
20545 				uint32_t maxlen;
20546 
20547 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20548 				if (len > maxlen)
20549 					len = maxlen;
20550 			}/* The else is full_size_rxt is on so send it all */
20551 		} else {
20552 			uint32_t outstanding;
20553 			/*
20554 			 * We are inside of a Fast recovery episode, this
20555 			 * is caused by a SACK or 3 dup acks. At this point
20556 			 * we have sent all the retransmissions and we rely
20557 			 * on PRR to dictate what we will send in the form of
20558 			 * new data.
20559 			 */
20560 
20561 			outstanding = tp->snd_max - tp->snd_una;
20562 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20563 				if (tp->snd_wnd > outstanding) {
20564 					len = tp->snd_wnd - outstanding;
20565 					/* Check to see if we have the data */
20566 					if ((sb_offset + len) > avail) {
20567 						/* It does not all fit */
20568 						if (avail > sb_offset)
20569 							len = avail - sb_offset;
20570 						else
20571 							len = 0;
20572 					}
20573 				} else {
20574 					len = 0;
20575 				}
20576 			} else if (avail > sb_offset) {
20577 				len = avail - sb_offset;
20578 			} else {
20579 				len = 0;
20580 			}
20581 			if (len > 0) {
20582 				if (len > rack->r_ctl.rc_prr_sndcnt) {
20583 					len = rack->r_ctl.rc_prr_sndcnt;
20584 				}
20585 				if (len > 0) {
20586 					sub_from_prr = 1;
20587 				}
20588 			}
20589 			if (len > segsiz) {
20590 				/*
20591 				 * We should never send more than a MSS when
20592 				 * retransmitting or sending new data in prr
20593 				 * mode unless the override flag is on. Most
20594 				 * likely the PRR algorithm is not going to
20595 				 * let us send a lot as well :-)
20596 				 */
20597 				if (rack->r_ctl.rc_prr_sendalot == 0) {
20598 					len = segsiz;
20599 				}
20600 			} else if (len < segsiz) {
20601 				/*
20602 				 * Do we send any? The idea here is if the
20603 				 * send empty's the socket buffer we want to
20604 				 * do it. However if not then lets just wait
20605 				 * for our prr_sndcnt to get bigger.
20606 				 */
20607 				long leftinsb;
20608 
20609 				leftinsb = sbavail(sb) - sb_offset;
20610 				if (leftinsb > len) {
20611 					/* This send does not empty the sb */
20612 					len = 0;
20613 				}
20614 			}
20615 		}
20616 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20617 		/*
20618 		 * If you have not established
20619 		 * and are not doing FAST OPEN
20620 		 * no data please.
20621 		 */
20622 		if ((sack_rxmit == 0) &&
20623 		    (!IS_FASTOPEN(tp->t_flags))){
20624 			len = 0;
20625 			sb_offset = 0;
20626 		}
20627 	}
20628 	if (prefetch_so_done == 0) {
20629 		kern_prefetch(so, &prefetch_so_done);
20630 		prefetch_so_done = 1;
20631 	}
20632 	/*
20633 	 * Lop off SYN bit if it has already been sent.  However, if this is
20634 	 * SYN-SENT state and if segment contains data and if we don't know
20635 	 * that foreign host supports TAO, suppress sending segment.
20636 	 */
20637 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
20638 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
20639 		/*
20640 		 * When sending additional segments following a TFO SYN|ACK,
20641 		 * do not include the SYN bit.
20642 		 */
20643 		if (IS_FASTOPEN(tp->t_flags) &&
20644 		    (tp->t_state == TCPS_SYN_RECEIVED))
20645 			flags &= ~TH_SYN;
20646 	}
20647 	/*
20648 	 * Be careful not to send data and/or FIN on SYN segments. This
20649 	 * measure is needed to prevent interoperability problems with not
20650 	 * fully conformant TCP implementations.
20651 	 */
20652 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20653 		len = 0;
20654 		flags &= ~TH_FIN;
20655 	}
20656 	/*
20657 	 * On TFO sockets, ensure no data is sent in the following cases:
20658 	 *
20659 	 *  - When retransmitting SYN|ACK on a passively-created socket
20660 	 *
20661 	 *  - When retransmitting SYN on an actively created socket
20662 	 *
20663 	 *  - When sending a zero-length cookie (cookie request) on an
20664 	 *    actively created socket
20665 	 *
20666 	 *  - When the socket is in the CLOSED state (RST is being sent)
20667 	 */
20668 	if (IS_FASTOPEN(tp->t_flags) &&
20669 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20670 	     ((tp->t_state == TCPS_SYN_SENT) &&
20671 	      (tp->t_tfo_client_cookie_len == 0)) ||
20672 	     (flags & TH_RST))) {
20673 		sack_rxmit = 0;
20674 		len = 0;
20675 	}
20676 	/* Without fast-open there should never be data sent on a SYN */
20677 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
20678 		tp->snd_nxt = tp->iss;
20679 		len = 0;
20680 	}
20681 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20682 		/* We only send 1 MSS if we have a DSACK block */
20683 		add_flag |= RACK_SENT_W_DSACK;
20684 		len = segsiz;
20685 	}
20686 	orig_len = len;
20687 	if (len <= 0) {
20688 		/*
20689 		 * If FIN has been sent but not acked, but we haven't been
20690 		 * called to retransmit, len will be < 0.  Otherwise, window
20691 		 * shrank after we sent into it.  If window shrank to 0,
20692 		 * cancel pending retransmit, pull snd_nxt back to (closed)
20693 		 * window, and set the persist timer if it isn't already
20694 		 * going.  If the window didn't close completely, just wait
20695 		 * for an ACK.
20696 		 *
20697 		 * We also do a general check here to ensure that we will
20698 		 * set the persist timer when we have data to send, but a
20699 		 * 0-byte window. This makes sure the persist timer is set
20700 		 * even if the packet hits one of the "goto send" lines
20701 		 * below.
20702 		 */
20703 		len = 0;
20704 		if ((tp->snd_wnd == 0) &&
20705 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20706 		    (tp->snd_una == tp->snd_max) &&
20707 		    (sb_offset < (int)sbavail(sb))) {
20708 			rack_enter_persist(tp, rack, cts, tp->snd_una);
20709 		}
20710 	} else if ((rsm == NULL) &&
20711 		   (doing_tlp == 0) &&
20712 		   (len < pace_max_seg)) {
20713 		/*
20714 		 * We are not sending a maximum sized segment for
20715 		 * some reason. Should we not send anything (think
20716 		 * sws or persists)?
20717 		 */
20718 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20719 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20720 		    (len < minseg) &&
20721 		    (len < (int)(sbavail(sb) - sb_offset))) {
20722 			/*
20723 			 * Here the rwnd is less than
20724 			 * the minimum pacing size, this is not a retransmit,
20725 			 * we are established and
20726 			 * the send is not the last in the socket buffer
20727 			 * we send nothing, and we may enter persists
20728 			 * if nothing is outstanding.
20729 			 */
20730 			len = 0;
20731 			if (tp->snd_max == tp->snd_una) {
20732 				/*
20733 				 * Nothing out we can
20734 				 * go into persists.
20735 				 */
20736 				rack_enter_persist(tp, rack, cts, tp->snd_una);
20737 			}
20738 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20739 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20740 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20741 			   (len < minseg)) {
20742 			/*
20743 			 * Here we are not retransmitting, and
20744 			 * the cwnd is not so small that we could
20745 			 * not send at least a min size (rxt timer
20746 			 * not having gone off), We have 2 segments or
20747 			 * more already in flight, its not the tail end
20748 			 * of the socket buffer  and the cwnd is blocking
20749 			 * us from sending out a minimum pacing segment size.
20750 			 * Lets not send anything.
20751 			 */
20752 			len = 0;
20753 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20754 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20755 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20756 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20757 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
20758 			/*
20759 			 * Here we have a send window but we have
20760 			 * filled it up and we can't send another pacing segment.
20761 			 * We also have in flight more than 2 segments
20762 			 * and we are not completing the sb i.e. we allow
20763 			 * the last bytes of the sb to go out even if
20764 			 * its not a full pacing segment.
20765 			 */
20766 			len = 0;
20767 		} else if ((rack->r_ctl.crte != NULL) &&
20768 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20769 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20770 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20771 			   (len < (int)(sbavail(sb) - sb_offset))) {
20772 			/*
20773 			 * Here we are doing hardware pacing, this is not a TLP,
20774 			 * we are not sending a pace max segment size, there is rwnd
20775 			 * room to send at least N pace_max_seg, the cwnd is greater
20776 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20777 			 * more segments in flight and its not the tail of the socket buffer.
20778 			 *
20779 			 * We don't want to send instead we need to get more ack's in to
20780 			 * allow us to send a full pacing segment. Normally, if we are pacing
20781 			 * about the right speed, we should have finished our pacing
20782 			 * send as most of the acks have come back if we are at the
20783 			 * right rate. This is a bit fuzzy since return path delay
20784 			 * can delay the acks, which is why we want to make sure we
20785 			 * have cwnd space to have a bit more than a max pace segments in flight.
20786 			 *
20787 			 * If we have not gotten our acks back we are pacing at too high a
20788 			 * rate delaying will not hurt and will bring our GP estimate down by
20789 			 * injecting the delay. If we don't do this we will send
20790 			 * 2 MSS out in response to the acks being clocked in which
20791 			 * defeats the point of hw-pacing (i.e. to help us get
20792 			 * larger TSO's out).
20793 			 */
20794 			len = 0;
20795 		}
20796 
20797 	}
20798 	/* len will be >= 0 after this point. */
20799 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20800 	rack_sndbuf_autoscale(rack);
20801 	/*
20802 	 * Decide if we can use TCP Segmentation Offloading (if supported by
20803 	 * hardware).
20804 	 *
20805 	 * TSO may only be used if we are in a pure bulk sending state.  The
20806 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20807 	 * options prevent using TSO.  With TSO the TCP header is the same
20808 	 * (except for the sequence number) for all generated packets.  This
20809 	 * makes it impossible to transmit any options which vary per
20810 	 * generated segment or packet.
20811 	 *
20812 	 * IPv4 handling has a clear separation of ip options and ip header
20813 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20814 	 * the right thing below to provide length of just ip options and thus
20815 	 * checking for ipoptlen is enough to decide if ip options are present.
20816 	 */
20817 	ipoptlen = 0;
20818 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20819 	/*
20820 	 * Pre-calculate here as we save another lookup into the darknesses
20821 	 * of IPsec that way and can actually decide if TSO is ok.
20822 	 */
20823 #ifdef INET6
20824 	if (isipv6 && IPSEC_ENABLED(ipv6))
20825 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20826 #ifdef INET
20827 	else
20828 #endif
20829 #endif				/* INET6 */
20830 #ifdef INET
20831 		if (IPSEC_ENABLED(ipv4))
20832 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20833 #endif				/* INET */
20834 #endif
20835 
20836 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20837 	ipoptlen += ipsec_optlen;
20838 #endif
20839 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20840 	    (tp->t_port == 0) &&
20841 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
20842 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
20843 	    ipoptlen == 0)
20844 		tso = 1;
20845 	{
20846 		uint32_t outstanding __unused;
20847 
20848 		outstanding = tp->snd_max - tp->snd_una;
20849 		if (tp->t_flags & TF_SENTFIN) {
20850 			/*
20851 			 * If we sent a fin, snd_max is 1 higher than
20852 			 * snd_una
20853 			 */
20854 			outstanding--;
20855 		}
20856 		if (sack_rxmit) {
20857 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20858 				flags &= ~TH_FIN;
20859 		} else {
20860 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
20861 				   sbused(sb)))
20862 				flags &= ~TH_FIN;
20863 		}
20864 	}
20865 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20866 		      (long)TCP_MAXWIN << tp->rcv_scale);
20867 
20868 	/*
20869 	 * Sender silly window avoidance.   We transmit under the following
20870 	 * conditions when len is non-zero:
20871 	 *
20872 	 * - We have a full segment (or more with TSO) - This is the last
20873 	 * buffer in a write()/send() and we are either idle or running
20874 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
20875 	 * then 1/2 the maximum send window's worth of data (receiver may be
20876 	 * limited the window size) - we need to retransmit
20877 	 */
20878 	if (len) {
20879 		if (len >= segsiz) {
20880 			goto send;
20881 		}
20882 		/*
20883 		 * NOTE! on localhost connections an 'ack' from the remote
20884 		 * end may occur synchronously with the output and cause us
20885 		 * to flush a buffer queued with moretocome.  XXX
20886 		 *
20887 		 */
20888 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
20889 		    (idle || (tp->t_flags & TF_NODELAY)) &&
20890 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20891 		    (tp->t_flags & TF_NOPUSH) == 0) {
20892 			pass = 2;
20893 			goto send;
20894 		}
20895 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
20896 			pass = 22;
20897 			goto send;
20898 		}
20899 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20900 			pass = 4;
20901 			goto send;
20902 		}
20903 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
20904 			pass = 5;
20905 			goto send;
20906 		}
20907 		if (sack_rxmit) {
20908 			pass = 6;
20909 			goto send;
20910 		}
20911 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20912 		    (ctf_outstanding(tp) < (segsiz * 2))) {
20913 			/*
20914 			 * We have less than two MSS outstanding (delayed ack)
20915 			 * and our rwnd will not let us send a full sized
20916 			 * MSS. Lets go ahead and let this small segment
20917 			 * out because we want to try to have at least two
20918 			 * packets inflight to not be caught by delayed ack.
20919 			 */
20920 			pass = 12;
20921 			goto send;
20922 		}
20923 	}
20924 	/*
20925 	 * Sending of standalone window updates.
20926 	 *
20927 	 * Window updates are important when we close our window due to a
20928 	 * full socket buffer and are opening it again after the application
20929 	 * reads data from it.  Once the window has opened again and the
20930 	 * remote end starts to send again the ACK clock takes over and
20931 	 * provides the most current window information.
20932 	 *
20933 	 * We must avoid the silly window syndrome whereas every read from
20934 	 * the receive buffer, no matter how small, causes a window update
20935 	 * to be sent.  We also should avoid sending a flurry of window
20936 	 * updates when the socket buffer had queued a lot of data and the
20937 	 * application is doing small reads.
20938 	 *
20939 	 * Prevent a flurry of pointless window updates by only sending an
20940 	 * update when we can increase the advertized window by more than
20941 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
20942 	 * full or is very small be more aggressive and send an update
20943 	 * whenever we can increase by two mss sized segments. In all other
20944 	 * situations the ACK's to new incoming data will carry further
20945 	 * window increases.
20946 	 *
20947 	 * Don't send an independent window update if a delayed ACK is
20948 	 * pending (it will get piggy-backed on it) or the remote side
20949 	 * already has done a half-close and won't send more data.  Skip
20950 	 * this if the connection is in T/TCP half-open state.
20951 	 */
20952 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20953 	    !(tp->t_flags & TF_DELACK) &&
20954 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
20955 		/*
20956 		 * "adv" is the amount we could increase the window, taking
20957 		 * into account that we are limited by TCP_MAXWIN <<
20958 		 * tp->rcv_scale.
20959 		 */
20960 		int32_t adv;
20961 		int oldwin;
20962 
20963 		adv = recwin;
20964 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20965 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
20966 			if (adv > oldwin)
20967 				adv -= oldwin;
20968 			else {
20969 				/* We can't increase the window */
20970 				adv = 0;
20971 			}
20972 		} else
20973 			oldwin = 0;
20974 
20975 		/*
20976 		 * If the new window size ends up being the same as or less
20977 		 * than the old size when it is scaled, then don't force
20978 		 * a window update.
20979 		 */
20980 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20981 			goto dontupdate;
20982 
20983 		if (adv >= (int32_t)(2 * segsiz) &&
20984 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20985 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20986 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20987 			pass = 7;
20988 			goto send;
20989 		}
20990 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20991 			pass = 23;
20992 			goto send;
20993 		}
20994 	}
20995 dontupdate:
20996 
20997 	/*
20998 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20999 	 * is also a catch-all for the retransmit timer timeout case.
21000 	 */
21001 	if (tp->t_flags & TF_ACKNOW) {
21002 		pass = 8;
21003 		goto send;
21004 	}
21005 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
21006 		pass = 9;
21007 		goto send;
21008 	}
21009 	/*
21010 	 * If our state indicates that FIN should be sent and we have not
21011 	 * yet done so, then we need to send.
21012 	 */
21013 	if ((flags & TH_FIN) &&
21014 	    (tp->snd_nxt == tp->snd_una)) {
21015 		pass = 11;
21016 		goto send;
21017 	}
21018 	/*
21019 	 * No reason to send a segment, just return.
21020 	 */
21021 just_return:
21022 	SOCKBUF_UNLOCK(sb);
21023 just_return_nolock:
21024 	{
21025 		int app_limited = CTF_JR_SENT_DATA;
21026 
21027 		if (tot_len_this_send > 0) {
21028 			/* Make sure snd_nxt is up to max */
21029 			rack->r_ctl.fsb.recwin = recwin;
21030 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
21031 			if ((error == 0) &&
21032 			    rack_use_rfo &&
21033 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21034 			    (ipoptlen == 0) &&
21035 			    (tp->snd_nxt == tp->snd_max) &&
21036 			    (tp->rcv_numsacks == 0) &&
21037 			    rack->r_fsb_inited &&
21038 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
21039 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
21040 			    (rack->r_must_retran == 0) &&
21041 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
21042 			    (len > 0) && (orig_len > 0) &&
21043 			    (orig_len > len) &&
21044 			    ((orig_len - len) >= segsiz) &&
21045 			    ((optlen == 0) ||
21046 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
21047 				/* We can send at least one more MSS using our fsb */
21048 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
21049 						       segsiz, pace_max_seg, hw_tls, flags);
21050 			} else
21051 				rack->r_fast_output = 0;
21052 
21053 
21054 			rack_log_fsb(rack, tp, so, flags,
21055 				     ipoptlen, orig_len, len, 0,
21056 				     1, optlen, __LINE__, 1);
21057 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
21058 				tp->snd_nxt = tp->snd_max;
21059 		} else {
21060 			int end_window = 0;
21061 			uint32_t seq = tp->gput_ack;
21062 
21063 			rsm = tqhash_max(rack->r_ctl.tqh);
21064 			if (rsm) {
21065 				/*
21066 				 * Mark the last sent that we just-returned (hinting
21067 				 * that delayed ack may play a role in any rtt measurement).
21068 				 */
21069 				rsm->r_just_ret = 1;
21070 			}
21071 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
21072 			rack->r_ctl.rc_agg_delayed = 0;
21073 			rack->r_early = 0;
21074 			rack->r_late = 0;
21075 			rack->r_ctl.rc_agg_early = 0;
21076 			if ((ctf_outstanding(tp) +
21077 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
21078 				 minseg)) >= tp->snd_wnd) {
21079 				/* We are limited by the rwnd */
21080 				app_limited = CTF_JR_RWND_LIMITED;
21081 				if (IN_FASTRECOVERY(tp->t_flags))
21082 					rack->r_ctl.rc_prr_sndcnt = 0;
21083 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
21084 				/* We are limited by whats available -- app limited */
21085 				app_limited = CTF_JR_APP_LIMITED;
21086 				if (IN_FASTRECOVERY(tp->t_flags))
21087 					rack->r_ctl.rc_prr_sndcnt = 0;
21088 			} else if ((idle == 0) &&
21089 				   ((tp->t_flags & TF_NODELAY) == 0) &&
21090 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
21091 				   (len < segsiz)) {
21092 				/*
21093 				 * No delay is not on and the
21094 				 * user is sending less than 1MSS. This
21095 				 * brings out SWS avoidance so we
21096 				 * don't send. Another app-limited case.
21097 				 */
21098 				app_limited = CTF_JR_APP_LIMITED;
21099 			} else if (tp->t_flags & TF_NOPUSH) {
21100 				/*
21101 				 * The user has requested no push of
21102 				 * the last segment and we are
21103 				 * at the last segment. Another app
21104 				 * limited case.
21105 				 */
21106 				app_limited = CTF_JR_APP_LIMITED;
21107 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
21108 				/* Its the cwnd */
21109 				app_limited = CTF_JR_CWND_LIMITED;
21110 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
21111 				   (rack->rack_no_prr == 0) &&
21112 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
21113 				app_limited = CTF_JR_PRR;
21114 			} else {
21115 				/* Now why here are we not sending? */
21116 #ifdef NOW
21117 #ifdef INVARIANTS
21118 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
21119 #endif
21120 #endif
21121 				app_limited = CTF_JR_ASSESSING;
21122 			}
21123 			/*
21124 			 * App limited in some fashion, for our pacing GP
21125 			 * measurements we don't want any gap (even cwnd).
21126 			 * Close  down the measurement window.
21127 			 */
21128 			if (rack_cwnd_block_ends_measure &&
21129 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
21130 			     (app_limited == CTF_JR_PRR))) {
21131 				/*
21132 				 * The reason we are not sending is
21133 				 * the cwnd (or prr). We have been configured
21134 				 * to end the measurement window in
21135 				 * this case.
21136 				 */
21137 				end_window = 1;
21138 			} else if (rack_rwnd_block_ends_measure &&
21139 				   (app_limited == CTF_JR_RWND_LIMITED)) {
21140 				/*
21141 				 * We are rwnd limited and have been
21142 				 * configured to end the measurement
21143 				 * window in this case.
21144 				 */
21145 				end_window = 1;
21146 			} else if (app_limited == CTF_JR_APP_LIMITED) {
21147 				/*
21148 				 * A true application limited period, we have
21149 				 * ran out of data.
21150 				 */
21151 				end_window = 1;
21152 			} else if (app_limited == CTF_JR_ASSESSING) {
21153 				/*
21154 				 * In the assessing case we hit the end of
21155 				 * the if/else and had no known reason
21156 				 * This will panic us under invariants..
21157 				 *
21158 				 * If we get this out in logs we need to
21159 				 * investagate which reason we missed.
21160 				 */
21161 				end_window = 1;
21162 			}
21163 			if (end_window) {
21164 				uint8_t log = 0;
21165 
21166 				/* Adjust the Gput measurement */
21167 				if ((tp->t_flags & TF_GPUTINPROG) &&
21168 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21169 					tp->gput_ack = tp->snd_max;
21170 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21171 						/*
21172 						 * There is not enough to measure.
21173 						 */
21174 						tp->t_flags &= ~TF_GPUTINPROG;
21175 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21176 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
21177 									   tp->gput_seq,
21178 									   0, 0, 18, __LINE__, NULL, 0);
21179 					} else
21180 						log = 1;
21181 				}
21182 				/* Mark the last packet has app limited */
21183 				rsm = tqhash_max(rack->r_ctl.tqh);
21184 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21185 					if (rack->r_ctl.rc_app_limited_cnt == 0)
21186 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21187 					else {
21188 						/*
21189 						 * Go out to the end app limited and mark
21190 						 * this new one as next and move the end_appl up
21191 						 * to this guy.
21192 						 */
21193 						if (rack->r_ctl.rc_end_appl)
21194 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21195 						rack->r_ctl.rc_end_appl = rsm;
21196 					}
21197 					rsm->r_flags |= RACK_APP_LIMITED;
21198 					rack->r_ctl.rc_app_limited_cnt++;
21199 				}
21200 				if (log)
21201 					rack_log_pacing_delay_calc(rack,
21202 								   rack->r_ctl.rc_app_limited_cnt, seq,
21203 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21204 			}
21205 		}
21206 		/* Check if we need to go into persists or not */
21207 		if ((tp->snd_max == tp->snd_una) &&
21208 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
21209 		    sbavail(sb) &&
21210 		    (sbavail(sb) > tp->snd_wnd) &&
21211 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21212 			/* Yes lets make sure to move to persist before timer-start */
21213 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21214 		}
21215 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21216 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21217 	}
21218 #ifdef NETFLIX_SHARED_CWND
21219 	if ((sbavail(sb) == 0) &&
21220 	    rack->r_ctl.rc_scw) {
21221 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21222 		rack->rack_scwnd_is_idle = 1;
21223 	}
21224 #endif
21225 #ifdef TCP_ACCOUNTING
21226 	if (tot_len_this_send > 0) {
21227 		crtsc = get_cyclecount();
21228 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21229 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
21230 		}
21231 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21232 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21233 		}
21234 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21235 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21236 		}
21237 	} else {
21238 		crtsc = get_cyclecount();
21239 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21240 			tp->tcp_cnt_counters[SND_LIMITED]++;
21241 		}
21242 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21243 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21244 		}
21245 	}
21246 	sched_unpin();
21247 #endif
21248 	return (0);
21249 
21250 send:
21251 	if ((rack->r_ctl.crte != NULL) &&
21252 	    (rsm == NULL) &&
21253 	    ((rack->rc_hw_nobuf == 1) ||
21254 	     (rack_hw_check_queue && (check_done == 0)))) {
21255 		/*
21256 		 * We only want to do this once with the hw_check_queue,
21257 		 * for the enobuf case we would only do it once if
21258 		 * we come around to again, the flag will be clear.
21259 		 */
21260 		check_done = 1;
21261 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21262 		if (slot) {
21263 			rack->r_ctl.rc_agg_delayed = 0;
21264 			rack->r_ctl.rc_agg_early = 0;
21265 			rack->r_early = 0;
21266 			rack->r_late = 0;
21267 			SOCKBUF_UNLOCK(&so->so_snd);
21268 			goto skip_all_send;
21269 		}
21270 	}
21271 	if (rsm || sack_rxmit)
21272 		counter_u64_add(rack_nfto_resend, 1);
21273 	else
21274 		counter_u64_add(rack_non_fto_send, 1);
21275 	if ((flags & TH_FIN) &&
21276 	    sbavail(sb)) {
21277 		/*
21278 		 * We do not transmit a FIN
21279 		 * with data outstanding. We
21280 		 * need to make it so all data
21281 		 * is acked first.
21282 		 */
21283 		flags &= ~TH_FIN;
21284 	}
21285 	/* Enforce stack imposed max seg size if we have one */
21286 	if (rack->r_ctl.rc_pace_max_segs &&
21287 	    (len > rack->r_ctl.rc_pace_max_segs)) {
21288 		mark = 1;
21289 		len = rack->r_ctl.rc_pace_max_segs;
21290 	}
21291 	SOCKBUF_LOCK_ASSERT(sb);
21292 	if (len > 0) {
21293 		if (len >= segsiz)
21294 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21295 		else
21296 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21297 	}
21298 	/*
21299 	 * Before ESTABLISHED, force sending of initial options unless TCP
21300 	 * set not to do any options. NOTE: we assume that the IP/TCP header
21301 	 * plus TCP options always fit in a single mbuf, leaving room for a
21302 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21303 	 * + optlen <= MCLBYTES
21304 	 */
21305 	optlen = 0;
21306 #ifdef INET6
21307 	if (isipv6)
21308 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21309 	else
21310 #endif
21311 		hdrlen = sizeof(struct tcpiphdr);
21312 
21313 	/*
21314 	 * Compute options for segment. We only have to care about SYN and
21315 	 * established connection segments.  Options for SYN-ACK segments
21316 	 * are handled in TCP syncache.
21317 	 */
21318 	to.to_flags = 0;
21319 	if ((tp->t_flags & TF_NOOPT) == 0) {
21320 		/* Maximum segment size. */
21321 		if (flags & TH_SYN) {
21322 			tp->snd_nxt = tp->iss;
21323 			to.to_mss = tcp_mssopt(&inp->inp_inc);
21324 			if (tp->t_port)
21325 				to.to_mss -= V_tcp_udp_tunneling_overhead;
21326 			to.to_flags |= TOF_MSS;
21327 
21328 			/*
21329 			 * On SYN or SYN|ACK transmits on TFO connections,
21330 			 * only include the TFO option if it is not a
21331 			 * retransmit, as the presence of the TFO option may
21332 			 * have caused the original SYN or SYN|ACK to have
21333 			 * been dropped by a middlebox.
21334 			 */
21335 			if (IS_FASTOPEN(tp->t_flags) &&
21336 			    (tp->t_rxtshift == 0)) {
21337 				if (tp->t_state == TCPS_SYN_RECEIVED) {
21338 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21339 					to.to_tfo_cookie =
21340 						(u_int8_t *)&tp->t_tfo_cookie.server;
21341 					to.to_flags |= TOF_FASTOPEN;
21342 					wanted_cookie = 1;
21343 				} else if (tp->t_state == TCPS_SYN_SENT) {
21344 					to.to_tfo_len =
21345 						tp->t_tfo_client_cookie_len;
21346 					to.to_tfo_cookie =
21347 						tp->t_tfo_cookie.client;
21348 					to.to_flags |= TOF_FASTOPEN;
21349 					wanted_cookie = 1;
21350 					/*
21351 					 * If we wind up having more data to
21352 					 * send with the SYN than can fit in
21353 					 * one segment, don't send any more
21354 					 * until the SYN|ACK comes back from
21355 					 * the other end.
21356 					 */
21357 					sendalot = 0;
21358 				}
21359 			}
21360 		}
21361 		/* Window scaling. */
21362 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21363 			to.to_wscale = tp->request_r_scale;
21364 			to.to_flags |= TOF_SCALE;
21365 		}
21366 		/* Timestamps. */
21367 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
21368 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21369 			to.to_tsval = ms_cts + tp->ts_offset;
21370 			to.to_tsecr = tp->ts_recent;
21371 			to.to_flags |= TOF_TS;
21372 		}
21373 		/* Set receive buffer autosizing timestamp. */
21374 		if (tp->rfbuf_ts == 0 &&
21375 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
21376 			tp->rfbuf_ts = tcp_ts_getticks();
21377 		/* Selective ACK's. */
21378 		if (tp->t_flags & TF_SACK_PERMIT) {
21379 			if (flags & TH_SYN)
21380 				to.to_flags |= TOF_SACKPERM;
21381 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21382 				 tp->rcv_numsacks > 0) {
21383 				to.to_flags |= TOF_SACK;
21384 				to.to_nsacks = tp->rcv_numsacks;
21385 				to.to_sacks = (u_char *)tp->sackblks;
21386 			}
21387 		}
21388 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21389 		/* TCP-MD5 (RFC2385). */
21390 		if (tp->t_flags & TF_SIGNATURE)
21391 			to.to_flags |= TOF_SIGNATURE;
21392 #endif
21393 
21394 		/* Processing the options. */
21395 		hdrlen += optlen = tcp_addoptions(&to, opt);
21396 		/*
21397 		 * If we wanted a TFO option to be added, but it was unable
21398 		 * to fit, ensure no data is sent.
21399 		 */
21400 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
21401 		    !(to.to_flags & TOF_FASTOPEN))
21402 			len = 0;
21403 	}
21404 	if (tp->t_port) {
21405 		if (V_tcp_udp_tunneling_port == 0) {
21406 			/* The port was removed?? */
21407 			SOCKBUF_UNLOCK(&so->so_snd);
21408 #ifdef TCP_ACCOUNTING
21409 			crtsc = get_cyclecount();
21410 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21411 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21412 			}
21413 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21414 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21415 			}
21416 			sched_unpin();
21417 #endif
21418 			return (EHOSTUNREACH);
21419 		}
21420 		hdrlen += sizeof(struct udphdr);
21421 	}
21422 #ifdef INET6
21423 	if (isipv6)
21424 		ipoptlen = ip6_optlen(inp);
21425 	else
21426 #endif
21427 		if (inp->inp_options)
21428 			ipoptlen = inp->inp_options->m_len -
21429 				offsetof(struct ipoption, ipopt_list);
21430 		else
21431 			ipoptlen = 0;
21432 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21433 	ipoptlen += ipsec_optlen;
21434 #endif
21435 
21436 	/*
21437 	 * Adjust data length if insertion of options will bump the packet
21438 	 * length beyond the t_maxseg length. Clear the FIN bit because we
21439 	 * cut off the tail of the segment.
21440 	 */
21441 	if (len + optlen + ipoptlen > tp->t_maxseg) {
21442 		if (tso) {
21443 			uint32_t if_hw_tsomax;
21444 			uint32_t moff;
21445 			int32_t max_len;
21446 
21447 			/* extract TSO information */
21448 			if_hw_tsomax = tp->t_tsomax;
21449 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21450 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21451 			KASSERT(ipoptlen == 0,
21452 				("%s: TSO can't do IP options", __func__));
21453 
21454 			/*
21455 			 * Check if we should limit by maximum payload
21456 			 * length:
21457 			 */
21458 			if (if_hw_tsomax != 0) {
21459 				/* compute maximum TSO length */
21460 				max_len = (if_hw_tsomax - hdrlen -
21461 					   max_linkhdr);
21462 				if (max_len <= 0) {
21463 					len = 0;
21464 				} else if (len > max_len) {
21465 					sendalot = 1;
21466 					len = max_len;
21467 					mark = 2;
21468 				}
21469 			}
21470 			/*
21471 			 * Prevent the last segment from being fractional
21472 			 * unless the send sockbuf can be emptied:
21473 			 */
21474 			max_len = (tp->t_maxseg - optlen);
21475 			if ((sb_offset + len) < sbavail(sb)) {
21476 				moff = len % (u_int)max_len;
21477 				if (moff != 0) {
21478 					mark = 3;
21479 					len -= moff;
21480 				}
21481 			}
21482 			/*
21483 			 * In case there are too many small fragments don't
21484 			 * use TSO:
21485 			 */
21486 			if (len <= max_len) {
21487 				mark = 4;
21488 				tso = 0;
21489 			}
21490 			/*
21491 			 * Send the FIN in a separate segment after the bulk
21492 			 * sending is done. We don't trust the TSO
21493 			 * implementations to clear the FIN flag on all but
21494 			 * the last segment.
21495 			 */
21496 			if (tp->t_flags & TF_NEEDFIN) {
21497 				sendalot = 4;
21498 			}
21499 		} else {
21500 			mark = 5;
21501 			if (optlen + ipoptlen >= tp->t_maxseg) {
21502 				/*
21503 				 * Since we don't have enough space to put
21504 				 * the IP header chain and the TCP header in
21505 				 * one packet as required by RFC 7112, don't
21506 				 * send it. Also ensure that at least one
21507 				 * byte of the payload can be put into the
21508 				 * TCP segment.
21509 				 */
21510 				SOCKBUF_UNLOCK(&so->so_snd);
21511 				error = EMSGSIZE;
21512 				sack_rxmit = 0;
21513 				goto out;
21514 			}
21515 			len = tp->t_maxseg - optlen - ipoptlen;
21516 			sendalot = 5;
21517 		}
21518 	} else {
21519 		tso = 0;
21520 		mark = 6;
21521 	}
21522 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21523 		("%s: len > IP_MAXPACKET", __func__));
21524 #ifdef DIAGNOSTIC
21525 #ifdef INET6
21526 	if (max_linkhdr + hdrlen > MCLBYTES)
21527 #else
21528 		if (max_linkhdr + hdrlen > MHLEN)
21529 #endif
21530 			panic("tcphdr too big");
21531 #endif
21532 
21533 	/*
21534 	 * This KASSERT is here to catch edge cases at a well defined place.
21535 	 * Before, those had triggered (random) panic conditions further
21536 	 * down.
21537 	 */
21538 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21539 	if ((len == 0) &&
21540 	    (flags & TH_FIN) &&
21541 	    (sbused(sb))) {
21542 		/*
21543 		 * We have outstanding data, don't send a fin by itself!.
21544 		 */
21545 		goto just_return;
21546 	}
21547 	/*
21548 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
21549 	 * and initialize the header from the template for sends on this
21550 	 * connection.
21551 	 */
21552 	hw_tls = tp->t_nic_ktls_xmit != 0;
21553 	if (len) {
21554 		uint32_t max_val;
21555 		uint32_t moff;
21556 
21557 		if (rack->r_ctl.rc_pace_max_segs)
21558 			max_val = rack->r_ctl.rc_pace_max_segs;
21559 		else if (rack->rc_user_set_max_segs)
21560 			max_val = rack->rc_user_set_max_segs * segsiz;
21561 		else
21562 			max_val = len;
21563 		/*
21564 		 * We allow a limit on sending with hptsi.
21565 		 */
21566 		if (len > max_val) {
21567 			mark = 7;
21568 			len = max_val;
21569 		}
21570 #ifdef INET6
21571 		if (MHLEN < hdrlen + max_linkhdr)
21572 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21573 		else
21574 #endif
21575 			m = m_gethdr(M_NOWAIT, MT_DATA);
21576 
21577 		if (m == NULL) {
21578 			SOCKBUF_UNLOCK(sb);
21579 			error = ENOBUFS;
21580 			sack_rxmit = 0;
21581 			goto out;
21582 		}
21583 		m->m_data += max_linkhdr;
21584 		m->m_len = hdrlen;
21585 
21586 		/*
21587 		 * Start the m_copy functions from the closest mbuf to the
21588 		 * sb_offset in the socket buffer chain.
21589 		 */
21590 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
21591 		s_mb = mb;
21592 		s_moff = moff;
21593 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21594 			m_copydata(mb, moff, (int)len,
21595 				   mtod(m, caddr_t)+hdrlen);
21596 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21597 				sbsndptr_adv(sb, mb, len);
21598 			m->m_len += len;
21599 		} else {
21600 			struct sockbuf *msb;
21601 
21602 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21603 				msb = NULL;
21604 			else
21605 				msb = sb;
21606 			m->m_next = tcp_m_copym(
21607 				mb, moff, &len,
21608 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21609 				((rsm == NULL) ? hw_tls : 0)
21610 #ifdef NETFLIX_COPY_ARGS
21611 				, &s_mb, &s_moff
21612 #endif
21613 				);
21614 			if (len <= (tp->t_maxseg - optlen)) {
21615 				/*
21616 				 * Must have ran out of mbufs for the copy
21617 				 * shorten it to no longer need tso. Lets
21618 				 * not put on sendalot since we are low on
21619 				 * mbufs.
21620 				 */
21621 				tso = 0;
21622 			}
21623 			if (m->m_next == NULL) {
21624 				SOCKBUF_UNLOCK(sb);
21625 				(void)m_free(m);
21626 				error = ENOBUFS;
21627 				sack_rxmit = 0;
21628 				goto out;
21629 			}
21630 		}
21631 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
21632 			if (rsm && (rsm->r_flags & RACK_TLP)) {
21633 				/*
21634 				 * TLP should not count in retran count, but
21635 				 * in its own bin
21636 				 */
21637 				counter_u64_add(rack_tlp_retran, 1);
21638 				counter_u64_add(rack_tlp_retran_bytes, len);
21639 			} else {
21640 				tp->t_sndrexmitpack++;
21641 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21642 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21643 			}
21644 #ifdef STATS
21645 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21646 						 len);
21647 #endif
21648 		} else {
21649 			KMOD_TCPSTAT_INC(tcps_sndpack);
21650 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21651 #ifdef STATS
21652 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21653 						 len);
21654 #endif
21655 		}
21656 		/*
21657 		 * If we're sending everything we've got, set PUSH. (This
21658 		 * will keep happy those implementations which only give
21659 		 * data to the user when a buffer fills or a PUSH comes in.)
21660 		 */
21661 		if (sb_offset + len == sbused(sb) &&
21662 		    sbused(sb) &&
21663 		    !(flags & TH_SYN)) {
21664 			flags |= TH_PUSH;
21665 			add_flag |= RACK_HAD_PUSH;
21666 		}
21667 
21668 		SOCKBUF_UNLOCK(sb);
21669 	} else {
21670 		SOCKBUF_UNLOCK(sb);
21671 		if (tp->t_flags & TF_ACKNOW)
21672 			KMOD_TCPSTAT_INC(tcps_sndacks);
21673 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
21674 			KMOD_TCPSTAT_INC(tcps_sndctrl);
21675 		else
21676 			KMOD_TCPSTAT_INC(tcps_sndwinup);
21677 
21678 		m = m_gethdr(M_NOWAIT, MT_DATA);
21679 		if (m == NULL) {
21680 			error = ENOBUFS;
21681 			sack_rxmit = 0;
21682 			goto out;
21683 		}
21684 #ifdef INET6
21685 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21686 		    MHLEN >= hdrlen) {
21687 			M_ALIGN(m, hdrlen);
21688 		} else
21689 #endif
21690 			m->m_data += max_linkhdr;
21691 		m->m_len = hdrlen;
21692 	}
21693 	SOCKBUF_UNLOCK_ASSERT(sb);
21694 	m->m_pkthdr.rcvif = (struct ifnet *)0;
21695 #ifdef MAC
21696 	mac_inpcb_create_mbuf(inp, m);
21697 #endif
21698 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21699 #ifdef INET6
21700 		if (isipv6)
21701 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21702 		else
21703 #endif				/* INET6 */
21704 #ifdef INET
21705 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21706 #endif
21707 		th = rack->r_ctl.fsb.th;
21708 		udp = rack->r_ctl.fsb.udp;
21709 		if (udp) {
21710 #ifdef INET6
21711 			if (isipv6)
21712 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21713 			else
21714 #endif				/* INET6 */
21715 				ulen = hdrlen + len - sizeof(struct ip);
21716 			udp->uh_ulen = htons(ulen);
21717 		}
21718 	} else {
21719 #ifdef INET6
21720 		if (isipv6) {
21721 			ip6 = mtod(m, struct ip6_hdr *);
21722 			if (tp->t_port) {
21723 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21724 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21725 				udp->uh_dport = tp->t_port;
21726 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21727 				udp->uh_ulen = htons(ulen);
21728 				th = (struct tcphdr *)(udp + 1);
21729 			} else
21730 				th = (struct tcphdr *)(ip6 + 1);
21731 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
21732 		} else
21733 #endif				/* INET6 */
21734 		{
21735 #ifdef INET
21736 			ip = mtod(m, struct ip *);
21737 			if (tp->t_port) {
21738 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21739 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21740 				udp->uh_dport = tp->t_port;
21741 				ulen = hdrlen + len - sizeof(struct ip);
21742 				udp->uh_ulen = htons(ulen);
21743 				th = (struct tcphdr *)(udp + 1);
21744 			} else
21745 				th = (struct tcphdr *)(ip + 1);
21746 			tcpip_fillheaders(inp, tp->t_port, ip, th);
21747 #endif
21748 		}
21749 	}
21750 	/*
21751 	 * Fill in fields, remembering maximum advertised window for use in
21752 	 * delaying messages about window sizes. If resending a FIN, be sure
21753 	 * not to use a new sequence number.
21754 	 */
21755 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
21756 	    tp->snd_nxt == tp->snd_max)
21757 		tp->snd_nxt--;
21758 	/*
21759 	 * If we are starting a connection, send ECN setup SYN packet. If we
21760 	 * are on a retransmit, we may resend those bits a number of times
21761 	 * as per RFC 3168.
21762 	 */
21763 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21764 		flags |= tcp_ecn_output_syn_sent(tp);
21765 	}
21766 	/* Also handle parallel SYN for ECN */
21767 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
21768 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21769 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21770 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21771 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
21772 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21773 #ifdef INET6
21774 		if (isipv6) {
21775 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21776 			ip6->ip6_flow |= htonl(ect << 20);
21777 		}
21778 		else
21779 #endif
21780 		{
21781 #ifdef INET
21782 			ip->ip_tos &= ~IPTOS_ECN_MASK;
21783 			ip->ip_tos |= ect;
21784 #endif
21785 		}
21786 	}
21787 	/*
21788 	 * If we are doing retransmissions, then snd_nxt will not reflect
21789 	 * the first unsent octet.  For ACK only packets, we do not want the
21790 	 * sequence number of the retransmitted packet, we want the sequence
21791 	 * number of the next unsent octet.  So, if there is no data (and no
21792 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
21793 	 * ti_seq.  But if we are in persist state, snd_max might reflect
21794 	 * one byte beyond the right edge of the window, so use snd_nxt in
21795 	 * that case, since we know we aren't doing a retransmission.
21796 	 * (retransmit and persist are mutually exclusive...)
21797 	 */
21798 	if (sack_rxmit == 0) {
21799 		if (len || (flags & (TH_SYN | TH_FIN))) {
21800 			th->th_seq = htonl(tp->snd_nxt);
21801 			rack_seq = tp->snd_nxt;
21802 		} else {
21803 			th->th_seq = htonl(tp->snd_max);
21804 			rack_seq = tp->snd_max;
21805 		}
21806 	} else {
21807 		th->th_seq = htonl(rsm->r_start);
21808 		rack_seq = rsm->r_start;
21809 	}
21810 	th->th_ack = htonl(tp->rcv_nxt);
21811 	tcp_set_flags(th, flags);
21812 	/*
21813 	 * Calculate receive window.  Don't shrink window, but avoid silly
21814 	 * window syndrome.
21815 	 * If a RST segment is sent, advertise a window of zero.
21816 	 */
21817 	if (flags & TH_RST) {
21818 		recwin = 0;
21819 	} else {
21820 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21821 		    recwin < (long)segsiz) {
21822 			recwin = 0;
21823 		}
21824 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21825 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21826 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21827 	}
21828 
21829 	/*
21830 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21831 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21832 	 * handled in syncache.
21833 	 */
21834 	if (flags & TH_SYN)
21835 		th->th_win = htons((u_short)
21836 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21837 	else {
21838 		/* Avoid shrinking window with window scaling. */
21839 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
21840 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21841 	}
21842 	/*
21843 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21844 	 * window.  This may cause the remote transmitter to stall.  This
21845 	 * flag tells soreceive() to disable delayed acknowledgements when
21846 	 * draining the buffer.  This can occur if the receiver is
21847 	 * attempting to read more data than can be buffered prior to
21848 	 * transmitting on the connection.
21849 	 */
21850 	if (th->th_win == 0) {
21851 		tp->t_sndzerowin++;
21852 		tp->t_flags |= TF_RXWIN0SENT;
21853 	} else
21854 		tp->t_flags &= ~TF_RXWIN0SENT;
21855 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
21856 	/* Now are we using fsb?, if so copy the template data to the mbuf */
21857 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21858 		uint8_t *cpto;
21859 
21860 		cpto = mtod(m, uint8_t *);
21861 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21862 		/*
21863 		 * We have just copied in:
21864 		 * IP/IP6
21865 		 * <optional udphdr>
21866 		 * tcphdr (no options)
21867 		 *
21868 		 * We need to grab the correct pointers into the mbuf
21869 		 * for both the tcp header, and possibly the udp header (if tunneling).
21870 		 * We do this by using the offset in the copy buffer and adding it
21871 		 * to the mbuf base pointer (cpto).
21872 		 */
21873 #ifdef INET6
21874 		if (isipv6)
21875 			ip6 = mtod(m, struct ip6_hdr *);
21876 		else
21877 #endif				/* INET6 */
21878 #ifdef INET
21879 			ip = mtod(m, struct ip *);
21880 #endif
21881 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21882 		/* If we have a udp header lets set it into the mbuf as well */
21883 		if (udp)
21884 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21885 	}
21886 	if (optlen) {
21887 		bcopy(opt, th + 1, optlen);
21888 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21889 	}
21890 	/*
21891 	 * Put TCP length in extended header, and then checksum extended
21892 	 * header and data.
21893 	 */
21894 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
21895 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21896 	if (to.to_flags & TOF_SIGNATURE) {
21897 		/*
21898 		 * Calculate MD5 signature and put it into the place
21899 		 * determined before.
21900 		 * NOTE: since TCP options buffer doesn't point into
21901 		 * mbuf's data, calculate offset and use it.
21902 		 */
21903 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21904 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21905 			/*
21906 			 * Do not send segment if the calculation of MD5
21907 			 * digest has failed.
21908 			 */
21909 			goto out;
21910 		}
21911 	}
21912 #endif
21913 #ifdef INET6
21914 	if (isipv6) {
21915 		/*
21916 		 * ip6_plen is not need to be filled now, and will be filled
21917 		 * in ip6_output.
21918 		 */
21919 		if (tp->t_port) {
21920 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21921 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21922 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21923 			th->th_sum = htons(0);
21924 			UDPSTAT_INC(udps_opackets);
21925 		} else {
21926 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21927 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21928 			th->th_sum = in6_cksum_pseudo(ip6,
21929 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21930 						      0);
21931 		}
21932 	}
21933 #endif
21934 #if defined(INET6) && defined(INET)
21935 	else
21936 #endif
21937 #ifdef INET
21938 	{
21939 		if (tp->t_port) {
21940 			m->m_pkthdr.csum_flags = CSUM_UDP;
21941 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21942 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21943 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21944 			th->th_sum = htons(0);
21945 			UDPSTAT_INC(udps_opackets);
21946 		} else {
21947 			m->m_pkthdr.csum_flags = CSUM_TCP;
21948 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21949 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
21950 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21951 									IPPROTO_TCP + len + optlen));
21952 		}
21953 		/* IP version must be set here for ipv4/ipv6 checking later */
21954 		KASSERT(ip->ip_v == IPVERSION,
21955 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
21956 	}
21957 #endif
21958 	/*
21959 	 * Enable TSO and specify the size of the segments. The TCP pseudo
21960 	 * header checksum is always provided. XXX: Fixme: This is currently
21961 	 * not the case for IPv6.
21962 	 */
21963 	if (tso) {
21964 		/*
21965 		 * Here we must use t_maxseg and the optlen since
21966 		 * the optlen may include SACK's (or DSACK).
21967 		 */
21968 		KASSERT(len > tp->t_maxseg - optlen,
21969 			("%s: len <= tso_segsz", __func__));
21970 		m->m_pkthdr.csum_flags |= CSUM_TSO;
21971 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21972 	}
21973 	KASSERT(len + hdrlen == m_length(m, NULL),
21974 		("%s: mbuf chain different than expected: %d + %u != %u",
21975 		 __func__, len, hdrlen, m_length(m, NULL)));
21976 
21977 #ifdef TCP_HHOOK
21978 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21979 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
21980 #endif
21981 	if ((rack->r_ctl.crte != NULL) &&
21982 	    (rack->rc_hw_nobuf == 0) &&
21983 	    tcp_bblogging_on(tp)) {
21984 		rack_log_queue_level(tp, rack, len, &tv, cts);
21985 	}
21986 	/* We're getting ready to send; log now. */
21987 	if (tcp_bblogging_on(rack->rc_tp)) {
21988 		union tcp_log_stackspecific log;
21989 
21990 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21991 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21992 		if (rack->rack_no_prr)
21993 			log.u_bbr.flex1 = 0;
21994 		else
21995 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21996 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21997 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21998 		log.u_bbr.flex4 = orig_len;
21999 		/* Save off the early/late values */
22000 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
22001 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
22002 		log.u_bbr.bw_inuse = rack_get_bw(rack);
22003 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
22004 		log.u_bbr.flex8 = 0;
22005 		if (rsm) {
22006 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
22007 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
22008 				counter_u64_add(rack_collapsed_win_rxt, 1);
22009 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
22010 			}
22011 			if (doing_tlp)
22012 				log.u_bbr.flex8 = 2;
22013 			else
22014 				log.u_bbr.flex8 = 1;
22015 		} else {
22016 			if (doing_tlp)
22017 				log.u_bbr.flex8 = 3;
22018 		}
22019 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
22020 		log.u_bbr.flex7 = mark;
22021 		log.u_bbr.flex7 <<= 8;
22022 		log.u_bbr.flex7 |= pass;
22023 		log.u_bbr.pkts_out = tp->t_maxseg;
22024 		log.u_bbr.timeStamp = cts;
22025 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
22026 		if (rsm && (rsm->r_rtr_cnt > 0)) {
22027 			/*
22028 			 * When we have a retransmit we want to log the
22029 			 * burst at send and flight at send from before.
22030 			 */
22031 			log.u_bbr.flex5 = rsm->r_fas;
22032 			log.u_bbr.bbr_substate = rsm->r_bas;
22033 		} else {
22034 			/*
22035 			 * New transmits we log in flex5 the inflight again as
22036 			 * well as the number of segments in our send in the
22037 			 * substate field.
22038 			 */
22039 			log.u_bbr.flex5 = log.u_bbr.inflight;
22040 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
22041 		}
22042 		log.u_bbr.lt_epoch = cwnd_to_use;
22043 		log.u_bbr.delivered = sendalot;
22044 		log.u_bbr.rttProp = (uint64_t)rsm;
22045 		log.u_bbr.pkt_epoch = __LINE__;
22046 		if (rsm) {
22047 			log.u_bbr.delRate = rsm->r_flags;
22048 			log.u_bbr.delRate <<= 31;
22049 			log.u_bbr.delRate |= rack->r_must_retran;
22050 			log.u_bbr.delRate <<= 1;
22051 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22052 		} else {
22053 			log.u_bbr.delRate = rack->r_must_retran;
22054 			log.u_bbr.delRate <<= 1;
22055 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22056 		}
22057 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
22058 				    len, &log, false, NULL, __func__, __LINE__, &tv);
22059 	} else
22060 		lgb = NULL;
22061 
22062 	/*
22063 	 * Fill in IP length and desired time to live and send to IP level.
22064 	 * There should be a better way to handle ttl and tos; we could keep
22065 	 * them in the template, but need a way to checksum without them.
22066 	 */
22067 	/*
22068 	 * m->m_pkthdr.len should have been set before cksum calcuration,
22069 	 * because in6_cksum() need it.
22070 	 */
22071 #ifdef INET6
22072 	if (isipv6) {
22073 		/*
22074 		 * we separately set hoplimit for every segment, since the
22075 		 * user might want to change the value via setsockopt. Also,
22076 		 * desired default hop limit might be changed via Neighbor
22077 		 * Discovery.
22078 		 */
22079 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
22080 
22081 		/*
22082 		 * Set the packet size here for the benefit of DTrace
22083 		 * probes. ip6_output() will set it properly; it's supposed
22084 		 * to include the option header lengths as well.
22085 		 */
22086 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
22087 
22088 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22089 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22090 		else
22091 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22092 
22093 		if (tp->t_state == TCPS_SYN_SENT)
22094 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22095 
22096 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22097 		/* TODO: IPv6 IP6TOS_ECT bit on */
22098 		error = ip6_output(m,
22099 				   inp->in6p_outputopts,
22100 				   &inp->inp_route6,
22101 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22102 				   NULL, NULL, inp);
22103 
22104 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22105 			mtu = inp->inp_route6.ro_nh->nh_mtu;
22106 	}
22107 #endif				/* INET6 */
22108 #if defined(INET) && defined(INET6)
22109 	else
22110 #endif
22111 #ifdef INET
22112 	{
22113 		ip->ip_len = htons(m->m_pkthdr.len);
22114 #ifdef INET6
22115 		if (inp->inp_vflag & INP_IPV6PROTO)
22116 			ip->ip_ttl = in6_selecthlim(inp, NULL);
22117 #endif				/* INET6 */
22118 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22119 		/*
22120 		 * If we do path MTU discovery, then we set DF on every
22121 		 * packet. This might not be the best thing to do according
22122 		 * to RFC3390 Section 2. However the tcp hostcache migitates
22123 		 * the problem so it affects only the first tcp connection
22124 		 * with a host.
22125 		 *
22126 		 * NB: Don't set DF on small MTU/MSS to have a safe
22127 		 * fallback.
22128 		 */
22129 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22130 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22131 			if (tp->t_port == 0 || len < V_tcp_minmss) {
22132 				ip->ip_off |= htons(IP_DF);
22133 			}
22134 		} else {
22135 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22136 		}
22137 
22138 		if (tp->t_state == TCPS_SYN_SENT)
22139 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22140 
22141 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
22142 
22143 		error = ip_output(m,
22144 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22145 				  inp->inp_options,
22146 #else
22147 				  NULL,
22148 #endif
22149 				  &inp->inp_route,
22150 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22151 				  inp);
22152 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22153 			mtu = inp->inp_route.ro_nh->nh_mtu;
22154 	}
22155 #endif				/* INET */
22156 
22157 out:
22158 	if (lgb) {
22159 		lgb->tlb_errno = error;
22160 		lgb = NULL;
22161 	}
22162 	/*
22163 	 * In transmit state, time the transmission and arrange for the
22164 	 * retransmit.  In persist state, just set snd_max.
22165 	 */
22166 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22167 			rack_to_usec_ts(&tv),
22168 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22169 	if (error == 0) {
22170 		if (rsm == NULL) {
22171 			if (rack->lt_bw_up == 0) {
22172 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
22173 				rack->r_ctl.lt_seq = tp->snd_una;
22174 				rack->lt_bw_up = 1;
22175 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22176 				/*
22177 				 * Need to record what we have since we are
22178 				 * approaching seq wrap.
22179 				 */
22180 				uint64_t tmark;
22181 
22182 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22183 				rack->r_ctl.lt_seq = tp->snd_una;
22184 				tmark = tcp_tv_to_lusectick(&tv);
22185 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22186 				rack->r_ctl.lt_timemark = tmark;
22187 			}
22188 		}
22189 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
22190 		counter_u64_add(rack_total_bytes, len);
22191 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22192 		if (rsm && doing_tlp) {
22193 			rack->rc_last_sent_tlp_past_cumack = 0;
22194 			rack->rc_last_sent_tlp_seq_valid = 1;
22195 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22196 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22197 		}
22198 		if (rack->rc_hw_nobuf) {
22199 			rack->rc_hw_nobuf = 0;
22200 			rack->r_ctl.rc_agg_delayed = 0;
22201 			rack->r_early = 0;
22202 			rack->r_late = 0;
22203 			rack->r_ctl.rc_agg_early = 0;
22204 		}
22205 		if (rsm && (doing_tlp == 0)) {
22206 			/* Set we retransmitted */
22207 			rack->rc_gp_saw_rec = 1;
22208 		} else {
22209 			if (cwnd_to_use > tp->snd_ssthresh) {
22210 				/* Set we sent in CA */
22211 				rack->rc_gp_saw_ca = 1;
22212 			} else {
22213 				/* Set we sent in SS */
22214 				rack->rc_gp_saw_ss = 1;
22215 			}
22216 		}
22217 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22218 		    (tp->t_flags & TF_SACK_PERMIT) &&
22219 		    tp->rcv_numsacks > 0)
22220 			tcp_clean_dsack_blocks(tp);
22221 		tot_len_this_send += len;
22222 		if (len == 0) {
22223 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22224 		} else {
22225 			int idx;
22226 
22227 			idx = (len / segsiz) + 3;
22228 			if (idx >= TCP_MSS_ACCT_ATIMER)
22229 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22230 			else
22231 				counter_u64_add(rack_out_size[idx], 1);
22232 		}
22233 	}
22234 	if ((rack->rack_no_prr == 0) &&
22235 	    sub_from_prr &&
22236 	    (error == 0)) {
22237 		if (rack->r_ctl.rc_prr_sndcnt >= len)
22238 			rack->r_ctl.rc_prr_sndcnt -= len;
22239 		else
22240 			rack->r_ctl.rc_prr_sndcnt = 0;
22241 	}
22242 	sub_from_prr = 0;
22243 	if (doing_tlp) {
22244 		/* Make sure the TLP is added */
22245 		add_flag |= RACK_TLP;
22246 	} else if (rsm) {
22247 		/* If its a resend without TLP then it must not have the flag */
22248 		rsm->r_flags &= ~RACK_TLP;
22249 	}
22250 
22251 
22252 	if ((error == 0) &&
22253 	    (len > 0) &&
22254 	    (tp->snd_una == tp->snd_max))
22255 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
22256 	{
22257 		tcp_seq startseq = tp->snd_nxt;
22258 
22259 		/* Track our lost count */
22260 		if (rsm && (doing_tlp == 0))
22261 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22262 		/*
22263 		 * Advance snd_nxt over sequence space of this segment.
22264 		 */
22265 		if (error)
22266 			/* We don't log or do anything with errors */
22267 			goto nomore;
22268 		if (doing_tlp == 0) {
22269 			if (rsm == NULL) {
22270 				/*
22271 				 * Not a retransmission of some
22272 				 * sort, new data is going out so
22273 				 * clear our TLP count and flag.
22274 				 */
22275 				rack->rc_tlp_in_progress = 0;
22276 				rack->r_ctl.rc_tlp_cnt_out = 0;
22277 			}
22278 		} else {
22279 			/*
22280 			 * We have just sent a TLP, mark that it is true
22281 			 * and make sure our in progress is set so we
22282 			 * continue to check the count.
22283 			 */
22284 			rack->rc_tlp_in_progress = 1;
22285 			rack->r_ctl.rc_tlp_cnt_out++;
22286 		}
22287 		if (flags & (TH_SYN | TH_FIN)) {
22288 			if (flags & TH_SYN)
22289 				tp->snd_nxt++;
22290 			if (flags & TH_FIN) {
22291 				tp->snd_nxt++;
22292 				tp->t_flags |= TF_SENTFIN;
22293 			}
22294 		}
22295 		/* In the ENOBUFS case we do *not* update snd_max */
22296 		if (sack_rxmit)
22297 			goto nomore;
22298 
22299 		tp->snd_nxt += len;
22300 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
22301 			if (tp->snd_una == tp->snd_max) {
22302 				/*
22303 				 * Update the time we just added data since
22304 				 * none was outstanding.
22305 				 */
22306 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22307 				tp->t_acktime = ticks;
22308 			}
22309 			tp->snd_max = tp->snd_nxt;
22310 			if (rack->rc_new_rnd_needed) {
22311 				/*
22312 				 * Update the rnd to start ticking not
22313 				 * that from a time perspective all of
22314 				 * the preceding idle time is "in the round"
22315 				 */
22316 				rack->rc_new_rnd_needed = 0;
22317 				rack->r_ctl.roundends = tp->snd_max;
22318 			}
22319 			/*
22320 			 * Time this transmission if not a retransmission and
22321 			 * not currently timing anything.
22322 			 * This is only relevant in case of switching back to
22323 			 * the base stack.
22324 			 */
22325 			if (tp->t_rtttime == 0) {
22326 				tp->t_rtttime = ticks;
22327 				tp->t_rtseq = startseq;
22328 				KMOD_TCPSTAT_INC(tcps_segstimed);
22329 			}
22330 			if (len &&
22331 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
22332 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22333 		}
22334 		/*
22335 		 * If we are doing FO we need to update the mbuf position and subtract
22336 		 * this happens when the peer sends us duplicate information and
22337 		 * we thus want to send a DSACK.
22338 		 *
22339 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22340 		 * turned off? If not then we are going to echo multiple DSACK blocks
22341 		 * out (with the TSO), which we should not be doing.
22342 		 */
22343 		if (rack->r_fast_output && len) {
22344 			if (rack->r_ctl.fsb.left_to_send > len)
22345 				rack->r_ctl.fsb.left_to_send -= len;
22346 			else
22347 				rack->r_ctl.fsb.left_to_send = 0;
22348 			if (rack->r_ctl.fsb.left_to_send < segsiz)
22349 				rack->r_fast_output = 0;
22350 			if (rack->r_fast_output) {
22351 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22352 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22353 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22354 			}
22355 		}
22356 	}
22357 nomore:
22358 	if (error) {
22359 		rack->r_ctl.rc_agg_delayed = 0;
22360 		rack->r_early = 0;
22361 		rack->r_late = 0;
22362 		rack->r_ctl.rc_agg_early = 0;
22363 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
22364 		/*
22365 		 * Failures do not advance the seq counter above. For the
22366 		 * case of ENOBUFS we will fall out and retry in 1ms with
22367 		 * the hpts. Everything else will just have to retransmit
22368 		 * with the timer.
22369 		 *
22370 		 * In any case, we do not want to loop around for another
22371 		 * send without a good reason.
22372 		 */
22373 		sendalot = 0;
22374 		switch (error) {
22375 		case EPERM:
22376 			tp->t_softerror = error;
22377 #ifdef TCP_ACCOUNTING
22378 			crtsc = get_cyclecount();
22379 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22380 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22381 			}
22382 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22383 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22384 			}
22385 			sched_unpin();
22386 #endif
22387 			return (error);
22388 		case ENOBUFS:
22389 			/*
22390 			 * Pace us right away to retry in a some
22391 			 * time
22392 			 */
22393 			if (rack->r_ctl.crte != NULL) {
22394 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22395 				if (tcp_bblogging_on(rack->rc_tp))
22396 					rack_log_queue_level(tp, rack, len, &tv, cts);
22397 			} else
22398 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22399 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22400 			if (rack->rc_enobuf < 0x7f)
22401 				rack->rc_enobuf++;
22402 			if (slot < (10 * HPTS_USEC_IN_MSEC))
22403 				slot = 10 * HPTS_USEC_IN_MSEC;
22404 			if (rack->r_ctl.crte != NULL) {
22405 				counter_u64_add(rack_saw_enobuf_hw, 1);
22406 				tcp_rl_log_enobuf(rack->r_ctl.crte);
22407 			}
22408 			counter_u64_add(rack_saw_enobuf, 1);
22409 			goto enobufs;
22410 		case EMSGSIZE:
22411 			/*
22412 			 * For some reason the interface we used initially
22413 			 * to send segments changed to another or lowered
22414 			 * its MTU. If TSO was active we either got an
22415 			 * interface without TSO capabilits or TSO was
22416 			 * turned off. If we obtained mtu from ip_output()
22417 			 * then update it and try again.
22418 			 */
22419 			if (tso)
22420 				tp->t_flags &= ~TF_TSO;
22421 			if (mtu != 0) {
22422 				int saved_mtu;
22423 
22424 				saved_mtu = tp->t_maxseg;
22425 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
22426 				if (saved_mtu > tp->t_maxseg) {
22427 					goto again;
22428 				}
22429 			}
22430 			slot = 10 * HPTS_USEC_IN_MSEC;
22431 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22432 #ifdef TCP_ACCOUNTING
22433 			crtsc = get_cyclecount();
22434 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22435 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22436 			}
22437 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22438 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22439 			}
22440 			sched_unpin();
22441 #endif
22442 			return (error);
22443 		case ENETUNREACH:
22444 			counter_u64_add(rack_saw_enetunreach, 1);
22445 		case EHOSTDOWN:
22446 		case EHOSTUNREACH:
22447 		case ENETDOWN:
22448 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
22449 				tp->t_softerror = error;
22450 			}
22451 			/* FALLTHROUGH */
22452 		default:
22453 			slot = 10 * HPTS_USEC_IN_MSEC;
22454 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22455 #ifdef TCP_ACCOUNTING
22456 			crtsc = get_cyclecount();
22457 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22458 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22459 			}
22460 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22461 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22462 			}
22463 			sched_unpin();
22464 #endif
22465 			return (error);
22466 		}
22467 	} else {
22468 		rack->rc_enobuf = 0;
22469 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22470 			rack->r_ctl.retran_during_recovery += len;
22471 	}
22472 	KMOD_TCPSTAT_INC(tcps_sndtotal);
22473 
22474 	/*
22475 	 * Data sent (as far as we can tell). If this advertises a larger
22476 	 * window than any other segment, then remember the size of the
22477 	 * advertised window. Any pending ACK has now been sent.
22478 	 */
22479 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22480 		tp->rcv_adv = tp->rcv_nxt + recwin;
22481 
22482 	tp->last_ack_sent = tp->rcv_nxt;
22483 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22484 enobufs:
22485 	if (sendalot) {
22486 		/* Do we need to turn off sendalot? */
22487 		if (rack->r_ctl.rc_pace_max_segs &&
22488 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
22489 			/* We hit our max. */
22490 			sendalot = 0;
22491 		} else if ((rack->rc_user_set_max_segs) &&
22492 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
22493 			/* We hit the user defined max */
22494 			sendalot = 0;
22495 		}
22496 	}
22497 	if ((error == 0) && (flags & TH_FIN))
22498 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22499 	if (flags & TH_RST) {
22500 		/*
22501 		 * We don't send again after sending a RST.
22502 		 */
22503 		slot = 0;
22504 		sendalot = 0;
22505 		if (error == 0)
22506 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22507 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22508 		/*
22509 		 * Get our pacing rate, if an error
22510 		 * occurred in sending (ENOBUF) we would
22511 		 * hit the else if with slot preset. Other
22512 		 * errors return.
22513 		 */
22514 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
22515 	}
22516 	if (rsm &&
22517 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
22518 	    rack->use_rack_rr) {
22519 		/* Its a retransmit and we use the rack cheat? */
22520 		if ((slot == 0) ||
22521 		    (rack->rc_always_pace == 0) ||
22522 		    (rack->r_rr_config == 1)) {
22523 			/*
22524 			 * We have no pacing set or we
22525 			 * are using old-style rack or
22526 			 * we are overridden to use the old 1ms pacing.
22527 			 */
22528 			slot = rack->r_ctl.rc_min_to;
22529 		}
22530 	}
22531 	/* We have sent clear the flag */
22532 	rack->r_ent_rec_ns = 0;
22533 	if (rack->r_must_retran) {
22534 		if (rsm) {
22535 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22536 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22537 				/*
22538 				 * We have retransmitted all.
22539 				 */
22540 				rack->r_must_retran = 0;
22541 				rack->r_ctl.rc_out_at_rto = 0;
22542 			}
22543 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22544 			/*
22545 			 * Sending new data will also kill
22546 			 * the loop.
22547 			 */
22548 			rack->r_must_retran = 0;
22549 			rack->r_ctl.rc_out_at_rto = 0;
22550 		}
22551 	}
22552 	rack->r_ctl.fsb.recwin = recwin;
22553 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22554 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22555 		/*
22556 		 * We hit an RTO and now have past snd_max at the RTO
22557 		 * clear all the WAS flags.
22558 		 */
22559 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22560 	}
22561 	if (slot) {
22562 		/* set the rack tcb into the slot N */
22563 		if ((error == 0) &&
22564 		    rack_use_rfo &&
22565 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22566 		    (rsm == NULL) &&
22567 		    (tp->snd_nxt == tp->snd_max) &&
22568 		    (ipoptlen == 0) &&
22569 		    (tp->rcv_numsacks == 0) &&
22570 		    rack->r_fsb_inited &&
22571 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22572 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22573 		    (rack->r_must_retran == 0) &&
22574 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22575 		    (len > 0) && (orig_len > 0) &&
22576 		    (orig_len > len) &&
22577 		    ((orig_len - len) >= segsiz) &&
22578 		    ((optlen == 0) ||
22579 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22580 			/* We can send at least one more MSS using our fsb */
22581 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22582 					       segsiz, pace_max_seg, hw_tls, flags);
22583 		} else
22584 			rack->r_fast_output = 0;
22585 		rack_log_fsb(rack, tp, so, flags,
22586 			     ipoptlen, orig_len, len, error,
22587 			     (rsm == NULL), optlen, __LINE__, 2);
22588 	} else if (sendalot) {
22589 		int ret;
22590 
22591 		sack_rxmit = 0;
22592 		if ((error == 0) &&
22593 		    rack_use_rfo &&
22594 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22595 		    (rsm == NULL) &&
22596 		    (ipoptlen == 0) &&
22597 		    (tp->rcv_numsacks == 0) &&
22598 		    (tp->snd_nxt == tp->snd_max) &&
22599 		    (rack->r_must_retran == 0) &&
22600 		    rack->r_fsb_inited &&
22601 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22602 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22603 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22604 		    (len > 0) && (orig_len > 0) &&
22605 		    (orig_len > len) &&
22606 		    ((orig_len - len) >= segsiz) &&
22607 		    ((optlen == 0) ||
22608 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22609 			/* we can use fast_output for more */
22610 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22611 					       segsiz, pace_max_seg, hw_tls, flags);
22612 			if (rack->r_fast_output) {
22613 				error = 0;
22614 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22615 				if (ret >= 0)
22616 					return (ret);
22617 			        else if (error)
22618 					goto nomore;
22619 
22620 			}
22621 		}
22622 		goto again;
22623 	}
22624 	/* Assure when we leave that snd_nxt will point to top */
22625 skip_all_send:
22626 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22627 		tp->snd_nxt = tp->snd_max;
22628 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22629 #ifdef TCP_ACCOUNTING
22630 	crtsc = get_cyclecount() - ts_val;
22631 	if (tot_len_this_send) {
22632 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22633 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22634 		}
22635 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22636 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22637 		}
22638 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22639 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22640 		}
22641 	} else {
22642 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22643 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
22644 		}
22645 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22646 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22647 		}
22648 	}
22649 	sched_unpin();
22650 #endif
22651 	if (error == ENOBUFS)
22652 		error = 0;
22653 	return (error);
22654 }
22655 
22656 static void
22657 rack_update_seg(struct tcp_rack *rack)
22658 {
22659 	uint32_t orig_val;
22660 
22661 	orig_val = rack->r_ctl.rc_pace_max_segs;
22662 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22663 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
22664 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22665 }
22666 
22667 static void
22668 rack_mtu_change(struct tcpcb *tp)
22669 {
22670 	/*
22671 	 * The MSS may have changed
22672 	 */
22673 	struct tcp_rack *rack;
22674 	struct rack_sendmap *rsm;
22675 
22676 	rack = (struct tcp_rack *)tp->t_fb_ptr;
22677 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22678 		/*
22679 		 * The MTU has changed we need to resend everything
22680 		 * since all we have sent is lost. We first fix
22681 		 * up the mtu though.
22682 		 */
22683 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
22684 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
22685 		rack_remxt_tmr(tp);
22686 		rack->r_fast_output = 0;
22687 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22688 						rack->r_ctl.rc_sacked);
22689 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22690 		rack->r_must_retran = 1;
22691 		/* Mark all inflight to needing to be rxt'd */
22692 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22693 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22694 		}
22695 	}
22696 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22697 	/* We don't use snd_nxt to retransmit */
22698 	tp->snd_nxt = tp->snd_max;
22699 }
22700 
22701 static int
22702 rack_set_dgp(struct tcp_rack *rack)
22703 {
22704 	/* pace_always=1 */
22705 	if (rack->rc_always_pace == 0) {
22706 		if (tcp_can_enable_pacing() == 0)
22707 			return (EBUSY);
22708 	}
22709 	rack->rc_fillcw_apply_discount = 0;
22710 	rack->dgp_on = 1;
22711 	rack->rc_always_pace = 1;
22712 	rack->use_fixed_rate = 0;
22713 	if (rack->gp_ready)
22714 		rack_set_cc_pacing(rack);
22715 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22716 	rack->rack_attempt_hdwr_pace = 0;
22717 	/* rxt settings */
22718 	rack->full_size_rxt = 1;
22719 	rack->shape_rxt_to_pacing_min  = 0;
22720 	/* cmpack=1 */
22721 	rack->r_use_cmp_ack = 1;
22722 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22723 	    rack->r_use_cmp_ack)
22724 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22725 	/* scwnd=1 */
22726 	rack->rack_enable_scwnd = 1;
22727 	/* dynamic=100 */
22728 	rack->rc_gp_dyn_mul = 1;
22729 	/* gp_inc_ca */
22730 	rack->r_ctl.rack_per_of_gp_ca = 100;
22731 	/* rrr_conf=3 */
22732 	rack->r_rr_config = 3;
22733 	/* npush=2 */
22734 	rack->r_ctl.rc_no_push_at_mrtt = 2;
22735 	/* fillcw=1 */
22736 	if (rack->r_cwnd_was_clamped == 0) {
22737 		rack->rc_pace_to_cwnd = 1;
22738 	} else {
22739 		rack->rc_pace_to_cwnd = 0;
22740 		/* Reset all multipliers to 100.0 so just the measured bw */
22741 		rack->r_ctl.rack_per_of_gp_ss = 100;
22742 		rack->r_ctl.rack_per_of_gp_ca = 100;
22743 	}
22744 	rack->rc_pace_fill_if_rttin_range = 0;
22745 	rack->rtt_limit_mul = 0;
22746 	/* noprr=1 */
22747 	rack->rack_no_prr = 1;
22748 	/* lscwnd=1 */
22749 	rack->r_limit_scw = 1;
22750 	/* gp_inc_rec */
22751 	rack->r_ctl.rack_per_of_gp_rec = 90;
22752 	rack_client_buffer_level_set(rack);
22753 	return (0);
22754 }
22755 
22756 
22757 
22758 static int
22759 rack_set_profile(struct tcp_rack *rack, int prof)
22760 {
22761 	int err = EINVAL;
22762 	if (prof == 1) {
22763 		/*
22764 		 * Profile 1 is "standard" DGP. It ignores
22765 		 * client buffer level.
22766 		 */
22767 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL0;
22768 		err = rack_set_dgp(rack);
22769 		if (err)
22770 			return (err);
22771 	} else if (prof == 2) {
22772 		/*
22773 		 * Profile 2 is DGP. Less aggressive with
22774 		 * respect to client buffer level.
22775 		 */
22776 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL1;
22777 		err = rack_set_dgp(rack);
22778 		if (err)
22779 			return (err);
22780 	} else if (prof == 3) {
22781 		/*
22782 		 * Profile 3 is DGP. Even Less aggressive with
22783 		 * respect to client buffer level.
22784 		 */
22785 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL2;
22786 		err = rack_set_dgp(rack);
22787 		if (err)
22788 			return (err);
22789 	} else if (prof == 4) {
22790 		/*
22791 		 * Profile 4 is DGP with the most responsiveness
22792 		 * to client buffer level.
22793 		 */
22794 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL3;
22795 		err = rack_set_dgp(rack);
22796 		if (err)
22797 			return (err);
22798 	} else if (prof == 5) {
22799 		err = rack_set_dgp(rack);
22800 		if (err)
22801 			return (err);
22802 		/*
22803 		 * By turning DGP off we change the rate
22804 		 * picked to be only the one the cwnd and rtt
22805 		 * get us.
22806 		 */
22807 		rack->dgp_on = 0;
22808 	} else if (prof == 6) {
22809 		err = rack_set_dgp(rack);
22810 		if (err)
22811 			return (err);
22812 		/*
22813 		 * Profile 6 tweaks DGP so that it will apply to
22814 		 * fill-cw the same settings that profile5 does
22815 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22816 		 */
22817 		rack->rc_fillcw_apply_discount = 1;
22818 	} else if (prof == 0) {
22819 		/* This changes things back to the default settings */
22820 		rack->dgp_on = 0;
22821 		rack->rc_hybrid_mode = 0;
22822 		err = 0;
22823 		if (rack_fill_cw_state)
22824 			rack->rc_pace_to_cwnd = 1;
22825 		else
22826 			rack->rc_pace_to_cwnd = 0;
22827 		if (rack->rc_always_pace) {
22828 			tcp_decrement_paced_conn();
22829 			rack_undo_cc_pacing(rack);
22830 			rack->rc_always_pace = 0;
22831 		}
22832 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22833 			rack->rc_always_pace = 1;
22834 			if (rack->rack_hibeta)
22835 				rack_set_cc_pacing(rack);
22836 		} else
22837 			rack->rc_always_pace = 0;
22838 		if (rack_dsack_std_based & 0x1) {
22839 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22840 			rack->rc_rack_tmr_std_based = 1;
22841 		}
22842 		if (rack_dsack_std_based & 0x2) {
22843 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22844 			rack->rc_rack_use_dsack = 1;
22845 		}
22846 		if (rack_use_cmp_acks)
22847 			rack->r_use_cmp_ack = 1;
22848 		else
22849 			rack->r_use_cmp_ack = 0;
22850 		if (rack_disable_prr)
22851 			rack->rack_no_prr = 1;
22852 		else
22853 			rack->rack_no_prr = 0;
22854 		if (rack_gp_no_rec_chg)
22855 			rack->rc_gp_no_rec_chg = 1;
22856 		else
22857 			rack->rc_gp_no_rec_chg = 0;
22858 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22859 			rack->r_mbuf_queue = 1;
22860 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22861 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22862 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22863 		} else {
22864 			rack->r_mbuf_queue = 0;
22865 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22866 		}
22867 		if (rack_enable_shared_cwnd)
22868 			rack->rack_enable_scwnd = 1;
22869 		else
22870 			rack->rack_enable_scwnd = 0;
22871 		if (rack_do_dyn_mul) {
22872 			/* When dynamic adjustment is on CA needs to start at 100% */
22873 			rack->rc_gp_dyn_mul = 1;
22874 			if (rack_do_dyn_mul >= 100)
22875 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22876 		} else {
22877 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22878 			rack->rc_gp_dyn_mul = 0;
22879 		}
22880 		rack->r_rr_config = 0;
22881 		rack->r_ctl.rc_no_push_at_mrtt = 0;
22882 		rack->rc_pace_to_cwnd = 0;
22883 		rack->rc_pace_fill_if_rttin_range = 0;
22884 		rack->rtt_limit_mul = 0;
22885 
22886 		if (rack_enable_hw_pacing)
22887 			rack->rack_hdw_pace_ena = 1;
22888 		else
22889 			rack->rack_hdw_pace_ena = 0;
22890 		if (rack_disable_prr)
22891 			rack->rack_no_prr = 1;
22892 		else
22893 			rack->rack_no_prr = 0;
22894 		if (rack_limits_scwnd)
22895 			rack->r_limit_scw  = 1;
22896 		else
22897 			rack->r_limit_scw  = 0;
22898 		rack_init_retransmit_value(rack, rack_rxt_controls);
22899 		err = 0;
22900 	}
22901 	return (err);
22902 }
22903 
22904 static int
22905 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22906 {
22907 	struct deferred_opt_list *dol;
22908 
22909 	dol = malloc(sizeof(struct deferred_opt_list),
22910 		     M_TCPFSB, M_NOWAIT|M_ZERO);
22911 	if (dol == NULL) {
22912 		/*
22913 		 * No space yikes -- fail out..
22914 		 */
22915 		return (0);
22916 	}
22917 	dol->optname = sopt_name;
22918 	dol->optval = loptval;
22919 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22920 	return (1);
22921 }
22922 
22923 static int
22924 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22925 {
22926 #ifdef TCP_REQUEST_TRK
22927 	struct tcp_sendfile_track *sft;
22928 	struct timeval tv;
22929 	tcp_seq seq;
22930 	int err;
22931 
22932 	microuptime(&tv);
22933 
22934 	/*
22935 	 * If BB logging is not on we need to look at the DTL flag.
22936 	 * If its on already then those reasons override the DTL input.
22937 	 * We do this with any request, you can turn DTL on, but it does
22938 	 * not turn off at least from hybrid pacing requests.
22939 	 */
22940 	if (tcp_bblogging_on(rack->rc_tp) == 0) {
22941 		if (hybrid->hybrid_flags & TCP_HYBRID_PACING_DTL) {
22942 			/* Turn on BB point logging  */
22943 			tcp_set_bblog_state(rack->rc_tp, TCP_LOG_VIA_BBPOINTS,
22944 					    TCP_BBPOINT_REQ_LEVEL_LOGGING);
22945 		}
22946 	}
22947 	/* Make sure no fixed rate is on */
22948 	rack->use_fixed_rate = 0;
22949 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22950 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22951 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22952 	/* Now allocate or find our entry that will have these settings */
22953 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22954 	if (sft == NULL) {
22955 		rack->rc_tp->tcp_hybrid_error++;
22956 		/* no space, where would it have gone? */
22957 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22958 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22959 		return (ENOSPC);
22960 	}
22961 	/* The seq will be snd_una + everything in the buffer */
22962 	seq = sft->start_seq;
22963 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22964 		/* Disabling hybrid pacing */
22965 		if (rack->rc_hybrid_mode) {
22966 			rack_set_profile(rack, 0);
22967 			rack->rc_tp->tcp_hybrid_stop++;
22968 		}
22969 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22970 		return (0);
22971 	}
22972 	if (rack->dgp_on == 0) {
22973 		/*
22974 		 * If we have not yet turned DGP on, do so
22975 		 * now setting pure DGP mode, no buffer level
22976 		 * response.
22977 		 */
22978 		if ((err = rack_set_profile(rack, 1)) != 0){
22979 			/* Failed to turn pacing on */
22980 			rack->rc_tp->tcp_hybrid_error++;
22981 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22982 			return (err);
22983 		}
22984 	}
22985 	/* Now set in our flags */
22986 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22987 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22988 		sft->cspr = hybrid->cspr;
22989 	else
22990 		sft->cspr = 0;
22991 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22992 		sft->hint_maxseg = hybrid->hint_maxseg;
22993 	else
22994 		sft->hint_maxseg = 0;
22995 	rack->rc_hybrid_mode = 1;
22996 	rack->rc_tp->tcp_hybrid_start++;
22997 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22998 	return (0);
22999 #else
23000 	return (ENOTSUP);
23001 #endif
23002 }
23003 
23004 static int
23005 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
23006 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
23007 
23008 {
23009 	struct epoch_tracker et;
23010 	struct sockopt sopt;
23011 	struct cc_newreno_opts opt;
23012 	uint64_t val;
23013 	int error = 0;
23014 	uint16_t ca, ss;
23015 
23016 	switch (sopt_name) {
23017 	case TCP_RACK_SET_RXT_OPTIONS:
23018 		if ((optval >= 0) && (optval <= 2)) {
23019 			rack_init_retransmit_value(rack, optval);
23020 		} else {
23021 			/*
23022 			 * You must send in 0, 1 or 2 all else is
23023 			 * invalid.
23024 			 */
23025 			error = EINVAL;
23026 		}
23027 		break;
23028 	case TCP_RACK_DSACK_OPT:
23029 		RACK_OPTS_INC(tcp_rack_dsack_opt);
23030 		if (optval & 0x1) {
23031 			rack->rc_rack_tmr_std_based = 1;
23032 		} else {
23033 			rack->rc_rack_tmr_std_based = 0;
23034 		}
23035 		if (optval & 0x2) {
23036 			rack->rc_rack_use_dsack = 1;
23037 		} else {
23038 			rack->rc_rack_use_dsack = 0;
23039 		}
23040 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
23041 		break;
23042 	case TCP_RACK_PACING_DIVISOR:
23043 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
23044 		if (optval == 0) {
23045 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
23046 		} else {
23047 			if (optval < RL_MIN_DIVISOR)
23048 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
23049 			else
23050 				rack->r_ctl.pace_len_divisor = optval;
23051 		}
23052 		break;
23053 	case TCP_RACK_HI_BETA:
23054 		RACK_OPTS_INC(tcp_rack_hi_beta);
23055 		if (optval > 0) {
23056 			rack->rack_hibeta = 1;
23057 			if ((optval >= 50) &&
23058 			    (optval <= 100)) {
23059 				/*
23060 				 * User wants to set a custom beta.
23061 				 */
23062 				rack->r_ctl.saved_hibeta = optval;
23063 				if (rack->rc_pacing_cc_set)
23064 					rack_undo_cc_pacing(rack);
23065 				rack->r_ctl.rc_saved_beta.beta = optval;
23066 			}
23067 			if (rack->rc_pacing_cc_set == 0)
23068 				rack_set_cc_pacing(rack);
23069 		} else {
23070 			rack->rack_hibeta = 0;
23071 			if (rack->rc_pacing_cc_set)
23072 				rack_undo_cc_pacing(rack);
23073 		}
23074 		break;
23075 	case TCP_RACK_PACING_BETA:
23076 		RACK_OPTS_INC(tcp_rack_beta);
23077 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23078 			/* This only works for newreno. */
23079 			error = EINVAL;
23080 			break;
23081 		}
23082 		if (rack->rc_pacing_cc_set) {
23083 			/*
23084 			 * Set them into the real CC module
23085 			 * whats in the rack pcb is the old values
23086 			 * to be used on restoral/
23087 			 */
23088 			sopt.sopt_dir = SOPT_SET;
23089 			opt.name = CC_NEWRENO_BETA;
23090 			opt.val = optval;
23091 			if (CC_ALGO(tp)->ctl_output != NULL)
23092 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23093 			else {
23094 				error = ENOENT;
23095 				break;
23096 			}
23097 		} else {
23098 			/*
23099 			 * Not pacing yet so set it into our local
23100 			 * rack pcb storage.
23101 			 */
23102 			rack->r_ctl.rc_saved_beta.beta = optval;
23103 		}
23104 		break;
23105 	case TCP_RACK_TIMER_SLOP:
23106 		RACK_OPTS_INC(tcp_rack_timer_slop);
23107 		rack->r_ctl.timer_slop = optval;
23108 		if (rack->rc_tp->t_srtt) {
23109 			/*
23110 			 * If we have an SRTT lets update t_rxtcur
23111 			 * to have the new slop.
23112 			 */
23113 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23114 					   rack_rto_min, rack_rto_max,
23115 					   rack->r_ctl.timer_slop);
23116 		}
23117 		break;
23118 	case TCP_RACK_PACING_BETA_ECN:
23119 		RACK_OPTS_INC(tcp_rack_beta_ecn);
23120 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23121 			/* This only works for newreno. */
23122 			error = EINVAL;
23123 			break;
23124 		}
23125 		if (rack->rc_pacing_cc_set) {
23126 			/*
23127 			 * Set them into the real CC module
23128 			 * whats in the rack pcb is the old values
23129 			 * to be used on restoral/
23130 			 */
23131 			sopt.sopt_dir = SOPT_SET;
23132 			opt.name = CC_NEWRENO_BETA_ECN;
23133 			opt.val = optval;
23134 			if (CC_ALGO(tp)->ctl_output != NULL)
23135 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23136 			else
23137 				error = ENOENT;
23138 		} else {
23139 			/*
23140 			 * Not pacing yet so set it into our local
23141 			 * rack pcb storage.
23142 			 */
23143 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
23144 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
23145 		}
23146 		break;
23147 	case TCP_DEFER_OPTIONS:
23148 		RACK_OPTS_INC(tcp_defer_opt);
23149 		if (optval) {
23150 			if (rack->gp_ready) {
23151 				/* Too late */
23152 				error = EINVAL;
23153 				break;
23154 			}
23155 			rack->defer_options = 1;
23156 		} else
23157 			rack->defer_options = 0;
23158 		break;
23159 	case TCP_RACK_MEASURE_CNT:
23160 		RACK_OPTS_INC(tcp_rack_measure_cnt);
23161 		if (optval && (optval <= 0xff)) {
23162 			rack->r_ctl.req_measurements = optval;
23163 		} else
23164 			error = EINVAL;
23165 		break;
23166 	case TCP_REC_ABC_VAL:
23167 		RACK_OPTS_INC(tcp_rec_abc_val);
23168 		if (optval > 0)
23169 			rack->r_use_labc_for_rec = 1;
23170 		else
23171 			rack->r_use_labc_for_rec = 0;
23172 		break;
23173 	case TCP_RACK_ABC_VAL:
23174 		RACK_OPTS_INC(tcp_rack_abc_val);
23175 		if ((optval > 0) && (optval < 255))
23176 			rack->rc_labc = optval;
23177 		else
23178 			error = EINVAL;
23179 		break;
23180 	case TCP_HDWR_UP_ONLY:
23181 		RACK_OPTS_INC(tcp_pacing_up_only);
23182 		if (optval)
23183 			rack->r_up_only = 1;
23184 		else
23185 			rack->r_up_only = 0;
23186 		break;
23187 	case TCP_PACING_RATE_CAP:
23188 		RACK_OPTS_INC(tcp_pacing_rate_cap);
23189 		rack->r_ctl.bw_rate_cap = loptval;
23190 		break;
23191 	case TCP_HYBRID_PACING:
23192 		if (hybrid == NULL) {
23193 			error = EINVAL;
23194 			break;
23195 		}
23196 		error = process_hybrid_pacing(rack, hybrid);
23197 		break;
23198 	case TCP_RACK_PROFILE:
23199 		RACK_OPTS_INC(tcp_profile);
23200 		error = rack_set_profile(rack, optval);
23201 		break;
23202 	case TCP_USE_CMP_ACKS:
23203 		RACK_OPTS_INC(tcp_use_cmp_acks);
23204 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23205 			/* You can't turn it off once its on! */
23206 			error = EINVAL;
23207 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23208 			rack->r_use_cmp_ack = 1;
23209 			rack->r_mbuf_queue = 1;
23210 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23211 		}
23212 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23213 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
23214 		break;
23215 	case TCP_SHARED_CWND_TIME_LIMIT:
23216 		RACK_OPTS_INC(tcp_lscwnd);
23217 		if (optval)
23218 			rack->r_limit_scw = 1;
23219 		else
23220 			rack->r_limit_scw = 0;
23221 		break;
23222 	case TCP_RACK_DGP_IN_REC:
23223 		RACK_OPTS_INC(tcp_dgp_in_rec);
23224 		if (optval)
23225 			rack->r_ctl.full_dgp_in_rec = 1;
23226 		else
23227 			rack->r_ctl.full_dgp_in_rec = 0;
23228 		break;
23229 	case TCP_RXT_CLAMP:
23230 		RACK_OPTS_INC(tcp_rxt_clamp);
23231 		rack_translate_clamp_value(rack, optval);
23232 		break;
23233  	case TCP_RACK_PACE_TO_FILL:
23234 		RACK_OPTS_INC(tcp_fillcw);
23235 		if (optval == 0)
23236 			rack->rc_pace_to_cwnd = 0;
23237 		else {
23238 			rack->rc_pace_to_cwnd = 1;
23239 			if (optval > 1)
23240 				rack->r_fill_less_agg = 1;
23241 		}
23242 		if ((optval >= rack_gp_rtt_maxmul) &&
23243 		    rack_gp_rtt_maxmul &&
23244 		    (optval < 0xf)) {
23245 			rack->rc_pace_fill_if_rttin_range = 1;
23246 			rack->rtt_limit_mul = optval;
23247 		} else {
23248 			rack->rc_pace_fill_if_rttin_range = 0;
23249 			rack->rtt_limit_mul = 0;
23250 		}
23251 		break;
23252 	case TCP_RACK_NO_PUSH_AT_MAX:
23253 		RACK_OPTS_INC(tcp_npush);
23254 		if (optval == 0)
23255 			rack->r_ctl.rc_no_push_at_mrtt = 0;
23256 		else if (optval < 0xff)
23257 			rack->r_ctl.rc_no_push_at_mrtt = optval;
23258 		else
23259 			error = EINVAL;
23260 		break;
23261 	case TCP_SHARED_CWND_ENABLE:
23262 		RACK_OPTS_INC(tcp_rack_scwnd);
23263 		if (optval == 0)
23264 			rack->rack_enable_scwnd = 0;
23265 		else
23266 			rack->rack_enable_scwnd = 1;
23267 		break;
23268 	case TCP_RACK_MBUF_QUEUE:
23269 		/* Now do we use the LRO mbuf-queue feature */
23270 		RACK_OPTS_INC(tcp_rack_mbufq);
23271 		if (optval || rack->r_use_cmp_ack)
23272 			rack->r_mbuf_queue = 1;
23273 		else
23274 			rack->r_mbuf_queue = 0;
23275 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23276 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23277 		else
23278 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23279 		break;
23280 	case TCP_RACK_NONRXT_CFG_RATE:
23281 		RACK_OPTS_INC(tcp_rack_cfg_rate);
23282 		if (optval == 0)
23283 			rack->rack_rec_nonrxt_use_cr = 0;
23284 		else
23285 			rack->rack_rec_nonrxt_use_cr = 1;
23286 		break;
23287 	case TCP_NO_PRR:
23288 		RACK_OPTS_INC(tcp_rack_noprr);
23289 		if (optval == 0)
23290 			rack->rack_no_prr = 0;
23291 		else if (optval == 1)
23292 			rack->rack_no_prr = 1;
23293 		else if (optval == 2)
23294 			rack->no_prr_addback = 1;
23295 		else
23296 			error = EINVAL;
23297 		break;
23298 	case TCP_TIMELY_DYN_ADJ:
23299 		RACK_OPTS_INC(tcp_timely_dyn);
23300 		if (optval == 0)
23301 			rack->rc_gp_dyn_mul = 0;
23302 		else {
23303 			rack->rc_gp_dyn_mul = 1;
23304 			if (optval >= 100) {
23305 				/*
23306 				 * If the user sets something 100 or more
23307 				 * its the gp_ca value.
23308 				 */
23309 				rack->r_ctl.rack_per_of_gp_ca  = optval;
23310 			}
23311 		}
23312 		break;
23313 	case TCP_RACK_DO_DETECTION:
23314 		RACK_OPTS_INC(tcp_rack_do_detection);
23315 		if (optval == 0)
23316 			rack->do_detection = 0;
23317 		else
23318 			rack->do_detection = 1;
23319 		break;
23320 	case TCP_RACK_TLP_USE:
23321 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23322 			error = EINVAL;
23323 			break;
23324 		}
23325 		RACK_OPTS_INC(tcp_tlp_use);
23326 		rack->rack_tlp_threshold_use = optval;
23327 		break;
23328 	case TCP_RACK_TLP_REDUCE:
23329 		/* RACK TLP cwnd reduction (bool) */
23330 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
23331 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23332 		break;
23333 		/*  Pacing related ones */
23334 	case TCP_RACK_PACE_ALWAYS:
23335 		/*
23336 		 * zero is old rack method, 1 is new
23337 		 * method using a pacing rate.
23338 		 */
23339 		RACK_OPTS_INC(tcp_rack_pace_always);
23340 		if (optval > 0) {
23341 			if (rack->rc_always_pace) {
23342 				error = EALREADY;
23343 				break;
23344 			} else if (tcp_can_enable_pacing()) {
23345 				rack->rc_always_pace = 1;
23346 				if (rack->rack_hibeta)
23347 					rack_set_cc_pacing(rack);
23348 			}
23349 			else {
23350 				error = ENOSPC;
23351 				break;
23352 			}
23353 		} else {
23354 			if (rack->rc_always_pace) {
23355 				tcp_decrement_paced_conn();
23356 				rack->rc_always_pace = 0;
23357 				rack_undo_cc_pacing(rack);
23358 			}
23359 		}
23360 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23361 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23362 		else
23363 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23364 		/* A rate may be set irate or other, if so set seg size */
23365 		rack_update_seg(rack);
23366 		break;
23367 	case TCP_BBR_RACK_INIT_RATE:
23368 		RACK_OPTS_INC(tcp_initial_rate);
23369 		val = optval;
23370 		/* Change from kbits per second to bytes per second */
23371 		val *= 1000;
23372 		val /= 8;
23373 		rack->r_ctl.init_rate = val;
23374 		if (rack->rc_init_win != rack_default_init_window) {
23375 			uint32_t win, snt;
23376 
23377 			/*
23378 			 * Options don't always get applied
23379 			 * in the order you think. So in order
23380 			 * to assure we update a cwnd we need
23381 			 * to check and see if we are still
23382 			 * where we should raise the cwnd.
23383 			 */
23384 			win = rc_init_window(rack);
23385 			if (SEQ_GT(tp->snd_max, tp->iss))
23386 				snt = tp->snd_max - tp->iss;
23387 			else
23388 				snt = 0;
23389 			if ((snt < win) &&
23390 			    (tp->snd_cwnd < win))
23391 				tp->snd_cwnd = win;
23392 		}
23393 		if (rack->rc_always_pace)
23394 			rack_update_seg(rack);
23395 		break;
23396 	case TCP_BBR_IWINTSO:
23397 		RACK_OPTS_INC(tcp_initial_win);
23398 		if (optval && (optval <= 0xff)) {
23399 			uint32_t win, snt;
23400 
23401 			rack->rc_init_win = optval;
23402 			win = rc_init_window(rack);
23403 			if (SEQ_GT(tp->snd_max, tp->iss))
23404 				snt = tp->snd_max - tp->iss;
23405 			else
23406 				snt = 0;
23407 			if ((snt < win) &&
23408 			    (tp->t_srtt |
23409 			     rack->r_ctl.init_rate)) {
23410 				/*
23411 				 * We are not past the initial window
23412 				 * and we have some bases for pacing,
23413 				 * so we need to possibly adjust up
23414 				 * the cwnd. Note even if we don't set
23415 				 * the cwnd, its still ok to raise the rc_init_win
23416 				 * which can be used coming out of idle when we
23417 				 * would have a rate.
23418 				 */
23419 				if (tp->snd_cwnd < win)
23420 					tp->snd_cwnd = win;
23421 			}
23422 			if (rack->rc_always_pace)
23423 				rack_update_seg(rack);
23424 		} else
23425 			error = EINVAL;
23426 		break;
23427 	case TCP_RACK_FORCE_MSEG:
23428 		RACK_OPTS_INC(tcp_rack_force_max_seg);
23429 		if (optval)
23430 			rack->rc_force_max_seg = 1;
23431 		else
23432 			rack->rc_force_max_seg = 0;
23433 		break;
23434 	case TCP_RACK_PACE_MIN_SEG:
23435 		RACK_OPTS_INC(tcp_rack_min_seg);
23436 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23437 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23438 		break;
23439 	case TCP_RACK_PACE_MAX_SEG:
23440 		/* Max segments size in a pace in bytes */
23441 		RACK_OPTS_INC(tcp_rack_max_seg);
23442 		if (optval <= MAX_USER_SET_SEG)
23443 			rack->rc_user_set_max_segs = optval;
23444 		else
23445 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23446 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23447 		break;
23448 	case TCP_RACK_PACE_RATE_REC:
23449 		/* Set the fixed pacing rate in Bytes per second ca */
23450 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23451 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23452 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23453 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23454 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23455 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23456 		rack->use_fixed_rate = 1;
23457 		if (rack->rack_hibeta)
23458 			rack_set_cc_pacing(rack);
23459 		rack_log_pacing_delay_calc(rack,
23460 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23461 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23462 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23463 					   __LINE__, NULL,0);
23464 		break;
23465 
23466 	case TCP_RACK_PACE_RATE_SS:
23467 		/* Set the fixed pacing rate in Bytes per second ca */
23468 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23469 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23470 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23471 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23472 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23473 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23474 		rack->use_fixed_rate = 1;
23475 		if (rack->rack_hibeta)
23476 			rack_set_cc_pacing(rack);
23477 		rack_log_pacing_delay_calc(rack,
23478 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23479 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23480 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23481 					   __LINE__, NULL, 0);
23482 		break;
23483 
23484 	case TCP_RACK_PACE_RATE_CA:
23485 		/* Set the fixed pacing rate in Bytes per second ca */
23486 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23487 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23488 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23489 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23490 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23491 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23492 		rack->use_fixed_rate = 1;
23493 		if (rack->rack_hibeta)
23494 			rack_set_cc_pacing(rack);
23495 		rack_log_pacing_delay_calc(rack,
23496 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23497 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23498 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23499 					   __LINE__, NULL, 0);
23500 		break;
23501 	case TCP_RACK_GP_INCREASE_REC:
23502 		RACK_OPTS_INC(tcp_gp_inc_rec);
23503 		rack->r_ctl.rack_per_of_gp_rec = optval;
23504 		rack_log_pacing_delay_calc(rack,
23505 					   rack->r_ctl.rack_per_of_gp_ss,
23506 					   rack->r_ctl.rack_per_of_gp_ca,
23507 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23508 					   __LINE__, NULL, 0);
23509 		break;
23510 	case TCP_RACK_GP_INCREASE_CA:
23511 		RACK_OPTS_INC(tcp_gp_inc_ca);
23512 		ca = optval;
23513 		if (ca < 100) {
23514 			/*
23515 			 * We don't allow any reduction
23516 			 * over the GP b/w.
23517 			 */
23518 			error = EINVAL;
23519 			break;
23520 		}
23521 		rack->r_ctl.rack_per_of_gp_ca = ca;
23522 		rack_log_pacing_delay_calc(rack,
23523 					   rack->r_ctl.rack_per_of_gp_ss,
23524 					   rack->r_ctl.rack_per_of_gp_ca,
23525 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23526 					   __LINE__, NULL, 0);
23527 		break;
23528 	case TCP_RACK_GP_INCREASE_SS:
23529 		RACK_OPTS_INC(tcp_gp_inc_ss);
23530 		ss = optval;
23531 		if (ss < 100) {
23532 			/*
23533 			 * We don't allow any reduction
23534 			 * over the GP b/w.
23535 			 */
23536 			error = EINVAL;
23537 			break;
23538 		}
23539 		rack->r_ctl.rack_per_of_gp_ss = ss;
23540 		rack_log_pacing_delay_calc(rack,
23541 					   rack->r_ctl.rack_per_of_gp_ss,
23542 					   rack->r_ctl.rack_per_of_gp_ca,
23543 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23544 					   __LINE__, NULL, 0);
23545 		break;
23546 	case TCP_RACK_RR_CONF:
23547 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23548 		if (optval && optval <= 3)
23549 			rack->r_rr_config = optval;
23550 		else
23551 			rack->r_rr_config = 0;
23552 		break;
23553 	case TCP_PACING_DND:			/*  URL:dnd */
23554 		if (optval > 0)
23555 			rack->rc_pace_dnd = 1;
23556 		else
23557 			rack->rc_pace_dnd = 0;
23558 		break;
23559 	case TCP_HDWR_RATE_CAP:
23560 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
23561 		if (optval) {
23562 			if (rack->r_rack_hw_rate_caps == 0)
23563 				rack->r_rack_hw_rate_caps = 1;
23564 			else
23565 				error = EALREADY;
23566 		} else {
23567 			rack->r_rack_hw_rate_caps = 0;
23568 		}
23569 		break;
23570 	case TCP_RACK_SPLIT_LIMIT:
23571 		RACK_OPTS_INC(tcp_split_limit);
23572 		rack->r_ctl.rc_split_limit = optval;
23573 		break;
23574 	case TCP_BBR_HDWR_PACE:
23575 		RACK_OPTS_INC(tcp_hdwr_pacing);
23576 		if (optval){
23577 			if (rack->rack_hdrw_pacing == 0) {
23578 				rack->rack_hdw_pace_ena = 1;
23579 				rack->rack_attempt_hdwr_pace = 0;
23580 			} else
23581 				error = EALREADY;
23582 		} else {
23583 			rack->rack_hdw_pace_ena = 0;
23584 #ifdef RATELIMIT
23585 			if (rack->r_ctl.crte != NULL) {
23586 				rack->rack_hdrw_pacing = 0;
23587 				rack->rack_attempt_hdwr_pace = 0;
23588 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23589 				rack->r_ctl.crte = NULL;
23590 			}
23591 #endif
23592 		}
23593 		break;
23594 		/*  End Pacing related ones */
23595 	case TCP_RACK_PRR_SENDALOT:
23596 		/* Allow PRR to send more than one seg */
23597 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
23598 		rack->r_ctl.rc_prr_sendalot = optval;
23599 		break;
23600 	case TCP_RACK_MIN_TO:
23601 		/* Minimum time between rack t-o's in ms */
23602 		RACK_OPTS_INC(tcp_rack_min_to);
23603 		rack->r_ctl.rc_min_to = optval;
23604 		break;
23605 	case TCP_RACK_EARLY_SEG:
23606 		/* If early recovery max segments */
23607 		RACK_OPTS_INC(tcp_rack_early_seg);
23608 		rack->r_ctl.rc_early_recovery_segs = optval;
23609 		break;
23610 	case TCP_RACK_ENABLE_HYSTART:
23611 	{
23612 		if (optval) {
23613 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23614 			if (rack_do_hystart > RACK_HYSTART_ON)
23615 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23616 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23617 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23618 		} else {
23619 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23620 		}
23621 	}
23622 	break;
23623 	case TCP_RACK_REORD_THRESH:
23624 		/* RACK reorder threshold (shift amount) */
23625 		RACK_OPTS_INC(tcp_rack_reord_thresh);
23626 		if ((optval > 0) && (optval < 31))
23627 			rack->r_ctl.rc_reorder_shift = optval;
23628 		else
23629 			error = EINVAL;
23630 		break;
23631 	case TCP_RACK_REORD_FADE:
23632 		/* Does reordering fade after ms time */
23633 		RACK_OPTS_INC(tcp_rack_reord_fade);
23634 		rack->r_ctl.rc_reorder_fade = optval;
23635 		break;
23636 	case TCP_RACK_TLP_THRESH:
23637 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
23638 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
23639 		if (optval)
23640 			rack->r_ctl.rc_tlp_threshold = optval;
23641 		else
23642 			error = EINVAL;
23643 		break;
23644 	case TCP_BBR_USE_RACK_RR:
23645 		RACK_OPTS_INC(tcp_rack_rr);
23646 		if (optval)
23647 			rack->use_rack_rr = 1;
23648 		else
23649 			rack->use_rack_rr = 0;
23650 		break;
23651 	case TCP_RACK_PKT_DELAY:
23652 		/* RACK added ms i.e. rack-rtt + reord + N */
23653 		RACK_OPTS_INC(tcp_rack_pkt_delay);
23654 		rack->r_ctl.rc_pkt_delay = optval;
23655 		break;
23656 	case TCP_DELACK:
23657 		RACK_OPTS_INC(tcp_rack_delayed_ack);
23658 		if (optval == 0)
23659 			tp->t_delayed_ack = 0;
23660 		else
23661 			tp->t_delayed_ack = 1;
23662 		if (tp->t_flags & TF_DELACK) {
23663 			tp->t_flags &= ~TF_DELACK;
23664 			tp->t_flags |= TF_ACKNOW;
23665 			NET_EPOCH_ENTER(et);
23666 			rack_output(tp);
23667 			NET_EPOCH_EXIT(et);
23668 		}
23669 		break;
23670 
23671 	case TCP_BBR_RACK_RTT_USE:
23672 		RACK_OPTS_INC(tcp_rack_rtt_use);
23673 		if ((optval != USE_RTT_HIGH) &&
23674 		    (optval != USE_RTT_LOW) &&
23675 		    (optval != USE_RTT_AVG))
23676 			error = EINVAL;
23677 		else
23678 			rack->r_ctl.rc_rate_sample_method = optval;
23679 		break;
23680 	case TCP_DATA_AFTER_CLOSE:
23681 		RACK_OPTS_INC(tcp_data_after_close);
23682 		if (optval)
23683 			rack->rc_allow_data_af_clo = 1;
23684 		else
23685 			rack->rc_allow_data_af_clo = 0;
23686 		break;
23687 	default:
23688 		break;
23689 	}
23690 	tcp_log_socket_option(tp, sopt_name, optval, error);
23691 	return (error);
23692 }
23693 
23694 
23695 static void
23696 rack_apply_deferred_options(struct tcp_rack *rack)
23697 {
23698 	struct deferred_opt_list *dol, *sdol;
23699 	uint32_t s_optval;
23700 
23701 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
23702 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
23703 		/* Disadvantage of deferal is you loose the error return */
23704 		s_optval = (uint32_t)dol->optval;
23705 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
23706 		free(dol, M_TCPDO);
23707 	}
23708 }
23709 
23710 static void
23711 rack_hw_tls_change(struct tcpcb *tp, int chg)
23712 {
23713 	/* Update HW tls state */
23714 	struct tcp_rack *rack;
23715 
23716 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23717 	if (chg)
23718 		rack->r_ctl.fsb.hw_tls = 1;
23719 	else
23720 		rack->r_ctl.fsb.hw_tls = 0;
23721 }
23722 
23723 static int
23724 rack_pru_options(struct tcpcb *tp, int flags)
23725 {
23726 	if (flags & PRUS_OOB)
23727 		return (EOPNOTSUPP);
23728 	return (0);
23729 }
23730 
23731 static bool
23732 rack_wake_check(struct tcpcb *tp)
23733 {
23734 	struct tcp_rack *rack;
23735 	struct timeval tv;
23736 	uint32_t cts;
23737 
23738 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23739 	if (rack->r_ctl.rc_hpts_flags) {
23740 		cts = tcp_get_usecs(&tv);
23741 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
23742 			/*
23743 			 * Pacing timer is up, check if we are ready.
23744 			 */
23745 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
23746 				return (true);
23747 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
23748 			/*
23749 			 * A timer is up, check if we are ready.
23750 			 */
23751 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
23752 				return (true);
23753 		}
23754 	}
23755 	return (false);
23756 }
23757 
23758 static struct tcp_function_block __tcp_rack = {
23759 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
23760 	.tfb_tcp_output = rack_output,
23761 	.tfb_do_queued_segments = ctf_do_queued_segments,
23762 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
23763 	.tfb_tcp_do_segment = rack_do_segment,
23764 	.tfb_tcp_ctloutput = rack_ctloutput,
23765 	.tfb_tcp_fb_init = rack_init,
23766 	.tfb_tcp_fb_fini = rack_fini,
23767 	.tfb_tcp_timer_stop_all = rack_stopall,
23768 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
23769 	.tfb_tcp_handoff_ok = rack_handoff_ok,
23770 	.tfb_tcp_mtu_chg = rack_mtu_change,
23771 	.tfb_pru_options = rack_pru_options,
23772 	.tfb_hwtls_change = rack_hw_tls_change,
23773 	.tfb_chg_query = rack_chg_query,
23774 	.tfb_switch_failed = rack_switch_failed,
23775 	.tfb_early_wake_check = rack_wake_check,
23776 	.tfb_compute_pipe = rack_compute_pipe,
23777 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
23778 };
23779 
23780 /*
23781  * rack_ctloutput() must drop the inpcb lock before performing copyin on
23782  * socket option arguments.  When it re-acquires the lock after the copy, it
23783  * has to revalidate that the connection is still valid for the socket
23784  * option.
23785  */
23786 static int
23787 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23788 {
23789 	struct inpcb *inp = tptoinpcb(tp);
23790 #ifdef INET
23791 	struct ip *ip;
23792 #endif
23793 	struct tcp_rack *rack;
23794 	struct tcp_hybrid_req hybrid;
23795 	uint64_t loptval;
23796 	int32_t error = 0, optval;
23797 
23798 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23799 	if (rack == NULL) {
23800 		INP_WUNLOCK(inp);
23801 		return (EINVAL);
23802 	}
23803 #ifdef INET
23804 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
23805 #endif
23806 
23807 	switch (sopt->sopt_level) {
23808 #ifdef INET6
23809 	case IPPROTO_IPV6:
23810 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
23811 		switch (sopt->sopt_name) {
23812 		case IPV6_USE_MIN_MTU:
23813 			tcp6_use_min_mtu(tp);
23814 			break;
23815 		}
23816 		INP_WUNLOCK(inp);
23817 		return (0);
23818 #endif
23819 #ifdef INET
23820 	case IPPROTO_IP:
23821 		switch (sopt->sopt_name) {
23822 		case IP_TOS:
23823 			/*
23824 			 * The DSCP codepoint has changed, update the fsb.
23825 			 */
23826 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
23827 			break;
23828 		case IP_TTL:
23829 			/*
23830 			 * The TTL has changed, update the fsb.
23831 			 */
23832 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
23833 			break;
23834 		}
23835 		INP_WUNLOCK(inp);
23836 		return (0);
23837 #endif
23838 #ifdef SO_PEERPRIO
23839 	case SOL_SOCKET:
23840 		switch (sopt->sopt_name) {
23841 		case SO_PEERPRIO:			/*  SC-URL:bs */
23842 			/* Already read in and sanity checked in sosetopt(). */
23843 			if (inp->inp_socket) {
23844 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
23845 				rack_client_buffer_level_set(rack);
23846 			}
23847 			break;
23848 		}
23849 		INP_WUNLOCK(inp);
23850 		return (0);
23851 #endif
23852 	case IPPROTO_TCP:
23853 		switch (sopt->sopt_name) {
23854 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
23855 		/*  Pacing related ones */
23856 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
23857 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
23858 		case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
23859 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
23860 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
23861 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
23862 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
23863 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
23864 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
23865 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
23866 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
23867 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
23868 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
23869 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
23870 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
23871 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
23872 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
23873 		case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
23874 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
23875 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
23876 		case TCP_RACK_DGP_IN_REC:		/*  URL:dgpinrec */
23877 			/* End pacing related */
23878 		case TCP_RXT_CLAMP:			/*  URL:rxtclamp */
23879 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
23880 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
23881 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
23882 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
23883 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
23884 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
23885 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
23886 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
23887 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
23888 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
23889 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
23890 		case TCP_RACK_DO_DETECTION:		/*  URL:detect */
23891 		case TCP_NO_PRR:			/*  URL:noprr */
23892 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
23893 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
23894 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
23895 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
23896 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
23897 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
23898 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
23899 		case TCP_RACK_PROFILE:			/*  URL:profile */
23900 		case TCP_HYBRID_PACING:			/*  URL:hybrid */
23901 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
23902 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
23903 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
23904 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
23905 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
23906 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
23907 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
23908 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
23909 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
23910 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
23911 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
23912 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
23913 		case TCP_PACING_DND:			/*  URL:dnd */
23914 			goto process_opt;
23915 			break;
23916 		default:
23917 			/* Filter off all unknown options to the base stack */
23918 			return (tcp_default_ctloutput(tp, sopt));
23919 			break;
23920 		}
23921 
23922 	default:
23923 		INP_WUNLOCK(inp);
23924 		return (0);
23925 	}
23926 process_opt:
23927 	INP_WUNLOCK(inp);
23928 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
23929 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
23930 		/*
23931 		 * We truncate it down to 32 bits for the socket-option trace this
23932 		 * means rates > 34Gbps won't show right, but thats probably ok.
23933 		 */
23934 		optval = (uint32_t)loptval;
23935 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
23936 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
23937 	} else {
23938 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
23939 		/* Save it in 64 bit form too */
23940 		loptval = optval;
23941 	}
23942 	if (error)
23943 		return (error);
23944 	INP_WLOCK(inp);
23945 	if (tp->t_fb != &__tcp_rack) {
23946 		INP_WUNLOCK(inp);
23947 		return (ENOPROTOOPT);
23948 	}
23949 	if (rack->defer_options && (rack->gp_ready == 0) &&
23950 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
23951 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
23952 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
23953 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
23954 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
23955 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
23956 		/* Options are beind deferred */
23957 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
23958 			INP_WUNLOCK(inp);
23959 			return (0);
23960 		} else {
23961 			/* No memory to defer, fail */
23962 			INP_WUNLOCK(inp);
23963 			return (ENOMEM);
23964 		}
23965 	}
23966 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
23967 	INP_WUNLOCK(inp);
23968 	return (error);
23969 }
23970 
23971 static void
23972 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
23973 {
23974 
23975 	INP_WLOCK_ASSERT(tptoinpcb(tp));
23976 	bzero(ti, sizeof(*ti));
23977 
23978 	ti->tcpi_state = tp->t_state;
23979 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
23980 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
23981 	if (tp->t_flags & TF_SACK_PERMIT)
23982 		ti->tcpi_options |= TCPI_OPT_SACK;
23983 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
23984 		ti->tcpi_options |= TCPI_OPT_WSCALE;
23985 		ti->tcpi_snd_wscale = tp->snd_scale;
23986 		ti->tcpi_rcv_wscale = tp->rcv_scale;
23987 	}
23988 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
23989 		ti->tcpi_options |= TCPI_OPT_ECN;
23990 	if (tp->t_flags & TF_FASTOPEN)
23991 		ti->tcpi_options |= TCPI_OPT_TFO;
23992 	/* still kept in ticks is t_rcvtime */
23993 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
23994 	/* Since we hold everything in precise useconds this is easy */
23995 	ti->tcpi_rtt = tp->t_srtt;
23996 	ti->tcpi_rttvar = tp->t_rttvar;
23997 	ti->tcpi_rto = tp->t_rxtcur;
23998 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
23999 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
24000 	/*
24001 	 * FreeBSD-specific extension fields for tcp_info.
24002 	 */
24003 	ti->tcpi_rcv_space = tp->rcv_wnd;
24004 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
24005 	ti->tcpi_snd_wnd = tp->snd_wnd;
24006 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
24007 	ti->tcpi_snd_nxt = tp->snd_nxt;
24008 	ti->tcpi_snd_mss = tp->t_maxseg;
24009 	ti->tcpi_rcv_mss = tp->t_maxseg;
24010 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24011 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24012 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24013 	ti->tcpi_total_tlp = tp->t_sndtlppack;
24014 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24015 #ifdef NETFLIX_STATS
24016 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24017 #endif
24018 #ifdef TCP_OFFLOAD
24019 	if (tp->t_flags & TF_TOE) {
24020 		ti->tcpi_options |= TCPI_OPT_TOE;
24021 		tcp_offload_tcp_info(tp, ti);
24022 	}
24023 #endif
24024 }
24025 
24026 static int
24027 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24028 {
24029 	struct inpcb *inp = tptoinpcb(tp);
24030 	struct tcp_rack *rack;
24031 	int32_t error, optval;
24032 	uint64_t val, loptval;
24033 	struct	tcp_info ti;
24034 	/*
24035 	 * Because all our options are either boolean or an int, we can just
24036 	 * pull everything into optval and then unlock and copy. If we ever
24037 	 * add a option that is not a int, then this will have quite an
24038 	 * impact to this routine.
24039 	 */
24040 	error = 0;
24041 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24042 	if (rack == NULL) {
24043 		INP_WUNLOCK(inp);
24044 		return (EINVAL);
24045 	}
24046 	switch (sopt->sopt_name) {
24047 	case TCP_INFO:
24048 		/* First get the info filled */
24049 		rack_fill_info(tp, &ti);
24050 		/* Fix up the rtt related fields if needed */
24051 		INP_WUNLOCK(inp);
24052 		error = sooptcopyout(sopt, &ti, sizeof ti);
24053 		return (error);
24054 	/*
24055 	 * Beta is the congestion control value for NewReno that influences how
24056 	 * much of a backoff happens when loss is detected. It is normally set
24057 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24058 	 * when you exit recovery.
24059 	 */
24060 	case TCP_RACK_PACING_BETA:
24061 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24062 			error = EINVAL;
24063 		else if (rack->rc_pacing_cc_set == 0)
24064 			optval = rack->r_ctl.rc_saved_beta.beta;
24065 		else {
24066 			/*
24067 			 * Reach out into the CC data and report back what
24068 			 * I have previously set. Yeah it looks hackish but
24069 			 * we don't want to report the saved values.
24070 			 */
24071 			if (tp->t_ccv.cc_data)
24072 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24073 			else
24074 				error = EINVAL;
24075 		}
24076 		break;
24077 		/*
24078 		 * Beta_ecn is the congestion control value for NewReno that influences how
24079 		 * much of a backoff happens when a ECN mark is detected. It is normally set
24080 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24081 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24082 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24083 		 */
24084 
24085 	case TCP_RACK_PACING_BETA_ECN:
24086 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24087 			error = EINVAL;
24088 		else if (rack->rc_pacing_cc_set == 0)
24089 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
24090 		else {
24091 			/*
24092 			 * Reach out into the CC data and report back what
24093 			 * I have previously set. Yeah it looks hackish but
24094 			 * we don't want to report the saved values.
24095 			 */
24096 			if (tp->t_ccv.cc_data)
24097 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24098 			else
24099 				error = EINVAL;
24100 		}
24101 		break;
24102 	case TCP_RACK_DSACK_OPT:
24103 		optval = 0;
24104 		if (rack->rc_rack_tmr_std_based) {
24105 			optval |= 1;
24106 		}
24107 		if (rack->rc_rack_use_dsack) {
24108 			optval |= 2;
24109 		}
24110 		break;
24111  	case TCP_RACK_ENABLE_HYSTART:
24112 	{
24113 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24114 			optval = RACK_HYSTART_ON;
24115 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24116 				optval = RACK_HYSTART_ON_W_SC;
24117 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24118 				optval = RACK_HYSTART_ON_W_SC_C;
24119 		} else {
24120 			optval = RACK_HYSTART_OFF;
24121 		}
24122 	}
24123 	break;
24124 	case TCP_RACK_DGP_IN_REC:
24125 		optval = rack->r_ctl.full_dgp_in_rec;
24126 		break;
24127 	case TCP_RACK_HI_BETA:
24128 		optval = rack->rack_hibeta;
24129 		break;
24130 	case TCP_RXT_CLAMP:
24131 		optval = rack->r_ctl.saved_rxt_clamp_val;
24132 		break;
24133 	case TCP_DEFER_OPTIONS:
24134 		optval = rack->defer_options;
24135 		break;
24136 	case TCP_RACK_MEASURE_CNT:
24137 		optval = rack->r_ctl.req_measurements;
24138 		break;
24139 	case TCP_REC_ABC_VAL:
24140 		optval = rack->r_use_labc_for_rec;
24141 		break;
24142 	case TCP_RACK_ABC_VAL:
24143 		optval = rack->rc_labc;
24144 		break;
24145 	case TCP_HDWR_UP_ONLY:
24146 		optval= rack->r_up_only;
24147 		break;
24148 	case TCP_PACING_RATE_CAP:
24149 		loptval = rack->r_ctl.bw_rate_cap;
24150 		break;
24151 	case TCP_RACK_PROFILE:
24152 		/* You cannot retrieve a profile, its write only */
24153 		error = EINVAL;
24154 		break;
24155 	case TCP_HYBRID_PACING:
24156 		/* You cannot retrieve hybrid pacing information, its write only */
24157 		error = EINVAL;
24158 		break;
24159 	case TCP_USE_CMP_ACKS:
24160 		optval = rack->r_use_cmp_ack;
24161 		break;
24162 	case TCP_RACK_PACE_TO_FILL:
24163 		optval = rack->rc_pace_to_cwnd;
24164 		if (optval && rack->r_fill_less_agg)
24165 			optval++;
24166 		break;
24167 	case TCP_RACK_NO_PUSH_AT_MAX:
24168 		optval = rack->r_ctl.rc_no_push_at_mrtt;
24169 		break;
24170 	case TCP_SHARED_CWND_ENABLE:
24171 		optval = rack->rack_enable_scwnd;
24172 		break;
24173 	case TCP_RACK_NONRXT_CFG_RATE:
24174 		optval = rack->rack_rec_nonrxt_use_cr;
24175 		break;
24176 	case TCP_NO_PRR:
24177 		if (rack->rack_no_prr  == 1)
24178 			optval = 1;
24179 		else if (rack->no_prr_addback == 1)
24180 			optval = 2;
24181 		else
24182 			optval = 0;
24183 		break;
24184 	case TCP_RACK_DO_DETECTION:
24185 		optval = rack->do_detection;
24186 		break;
24187 	case TCP_RACK_MBUF_QUEUE:
24188 		/* Now do we use the LRO mbuf-queue feature */
24189 		optval = rack->r_mbuf_queue;
24190 		break;
24191 	case TCP_TIMELY_DYN_ADJ:
24192 		optval = rack->rc_gp_dyn_mul;
24193 		break;
24194 	case TCP_BBR_IWINTSO:
24195 		optval = rack->rc_init_win;
24196 		break;
24197 	case TCP_RACK_TLP_REDUCE:
24198 		/* RACK TLP cwnd reduction (bool) */
24199 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24200 		break;
24201 	case TCP_BBR_RACK_INIT_RATE:
24202 		val = rack->r_ctl.init_rate;
24203 		/* convert to kbits per sec */
24204 		val *= 8;
24205 		val /= 1000;
24206 		optval = (uint32_t)val;
24207 		break;
24208 	case TCP_RACK_FORCE_MSEG:
24209 		optval = rack->rc_force_max_seg;
24210 		break;
24211 	case TCP_RACK_PACE_MIN_SEG:
24212 		optval = rack->r_ctl.rc_user_set_min_segs;
24213 		break;
24214 	case TCP_RACK_PACE_MAX_SEG:
24215 		/* Max segments in a pace */
24216 		optval = rack->rc_user_set_max_segs;
24217 		break;
24218 	case TCP_RACK_PACE_ALWAYS:
24219 		/* Use the always pace method */
24220 		optval = rack->rc_always_pace;
24221 		break;
24222 	case TCP_RACK_PRR_SENDALOT:
24223 		/* Allow PRR to send more than one seg */
24224 		optval = rack->r_ctl.rc_prr_sendalot;
24225 		break;
24226 	case TCP_RACK_MIN_TO:
24227 		/* Minimum time between rack t-o's in ms */
24228 		optval = rack->r_ctl.rc_min_to;
24229 		break;
24230 	case TCP_RACK_SPLIT_LIMIT:
24231 		optval = rack->r_ctl.rc_split_limit;
24232 		break;
24233 	case TCP_RACK_EARLY_SEG:
24234 		/* If early recovery max segments */
24235 		optval = rack->r_ctl.rc_early_recovery_segs;
24236 		break;
24237 	case TCP_RACK_REORD_THRESH:
24238 		/* RACK reorder threshold (shift amount) */
24239 		optval = rack->r_ctl.rc_reorder_shift;
24240 		break;
24241 	case TCP_RACK_REORD_FADE:
24242 		/* Does reordering fade after ms time */
24243 		optval = rack->r_ctl.rc_reorder_fade;
24244 		break;
24245 	case TCP_BBR_USE_RACK_RR:
24246 		/* Do we use the rack cheat for rxt */
24247 		optval = rack->use_rack_rr;
24248 		break;
24249 	case TCP_RACK_RR_CONF:
24250 		optval = rack->r_rr_config;
24251 		break;
24252 	case TCP_HDWR_RATE_CAP:
24253 		optval = rack->r_rack_hw_rate_caps;
24254 		break;
24255 	case TCP_BBR_HDWR_PACE:
24256 		optval = rack->rack_hdw_pace_ena;
24257 		break;
24258 	case TCP_RACK_TLP_THRESH:
24259 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24260 		optval = rack->r_ctl.rc_tlp_threshold;
24261 		break;
24262 	case TCP_RACK_PKT_DELAY:
24263 		/* RACK added ms i.e. rack-rtt + reord + N */
24264 		optval = rack->r_ctl.rc_pkt_delay;
24265 		break;
24266 	case TCP_RACK_TLP_USE:
24267 		optval = rack->rack_tlp_threshold_use;
24268 		break;
24269 	case TCP_PACING_DND:
24270 		optval = rack->rc_pace_dnd;
24271 		break;
24272 	case TCP_RACK_PACE_RATE_CA:
24273 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24274 		break;
24275 	case TCP_RACK_PACE_RATE_SS:
24276 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24277 		break;
24278 	case TCP_RACK_PACE_RATE_REC:
24279 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24280 		break;
24281 	case TCP_RACK_GP_INCREASE_SS:
24282 		optval = rack->r_ctl.rack_per_of_gp_ca;
24283 		break;
24284 	case TCP_RACK_GP_INCREASE_CA:
24285 		optval = rack->r_ctl.rack_per_of_gp_ss;
24286 		break;
24287 	case TCP_RACK_PACING_DIVISOR:
24288 		optval = rack->r_ctl.pace_len_divisor;
24289 		break;
24290 	case TCP_BBR_RACK_RTT_USE:
24291 		optval = rack->r_ctl.rc_rate_sample_method;
24292 		break;
24293 	case TCP_DELACK:
24294 		optval = tp->t_delayed_ack;
24295 		break;
24296 	case TCP_DATA_AFTER_CLOSE:
24297 		optval = rack->rc_allow_data_af_clo;
24298 		break;
24299 	case TCP_SHARED_CWND_TIME_LIMIT:
24300 		optval = rack->r_limit_scw;
24301 		break;
24302 	case TCP_RACK_TIMER_SLOP:
24303 		optval = rack->r_ctl.timer_slop;
24304 		break;
24305 	default:
24306 		return (tcp_default_ctloutput(tp, sopt));
24307 		break;
24308 	}
24309 	INP_WUNLOCK(inp);
24310 	if (error == 0) {
24311 		if (TCP_PACING_RATE_CAP)
24312 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
24313 		else
24314 			error = sooptcopyout(sopt, &optval, sizeof optval);
24315 	}
24316 	return (error);
24317 }
24318 
24319 static int
24320 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24321 {
24322 	if (sopt->sopt_dir == SOPT_SET) {
24323 		return (rack_set_sockopt(tp, sopt));
24324 	} else if (sopt->sopt_dir == SOPT_GET) {
24325 		return (rack_get_sockopt(tp, sopt));
24326 	} else {
24327 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24328 	}
24329 }
24330 
24331 static const char *rack_stack_names[] = {
24332 	__XSTRING(STACKNAME),
24333 #ifdef STACKALIAS
24334 	__XSTRING(STACKALIAS),
24335 #endif
24336 };
24337 
24338 static int
24339 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24340 {
24341 	memset(mem, 0, size);
24342 	return (0);
24343 }
24344 
24345 static void
24346 rack_dtor(void *mem, int32_t size, void *arg)
24347 {
24348 
24349 }
24350 
24351 static bool rack_mod_inited = false;
24352 
24353 static int
24354 tcp_addrack(module_t mod, int32_t type, void *data)
24355 {
24356 	int32_t err = 0;
24357 	int num_stacks;
24358 
24359 	switch (type) {
24360 	case MOD_LOAD:
24361 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24362 		    sizeof(struct rack_sendmap),
24363 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24364 
24365 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24366 		    sizeof(struct tcp_rack),
24367 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24368 
24369 		sysctl_ctx_init(&rack_sysctl_ctx);
24370 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24371 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24372 		    OID_AUTO,
24373 #ifdef STACKALIAS
24374 		    __XSTRING(STACKALIAS),
24375 #else
24376 		    __XSTRING(STACKNAME),
24377 #endif
24378 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24379 		    "");
24380 		if (rack_sysctl_root == NULL) {
24381 			printf("Failed to add sysctl node\n");
24382 			err = EFAULT;
24383 			goto free_uma;
24384 		}
24385 		rack_init_sysctls();
24386 		num_stacks = nitems(rack_stack_names);
24387 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24388 		    rack_stack_names, &num_stacks);
24389 		if (err) {
24390 			printf("Failed to register %s stack name for "
24391 			    "%s module\n", rack_stack_names[num_stacks],
24392 			    __XSTRING(MODNAME));
24393 			sysctl_ctx_free(&rack_sysctl_ctx);
24394 free_uma:
24395 			uma_zdestroy(rack_zone);
24396 			uma_zdestroy(rack_pcb_zone);
24397 			rack_counter_destroy();
24398 			printf("Failed to register rack module -- err:%d\n", err);
24399 			return (err);
24400 		}
24401 		tcp_lro_reg_mbufq();
24402 		rack_mod_inited = true;
24403 		break;
24404 	case MOD_QUIESCE:
24405 		err = deregister_tcp_functions(&__tcp_rack, true, false);
24406 		break;
24407 	case MOD_UNLOAD:
24408 		err = deregister_tcp_functions(&__tcp_rack, false, true);
24409 		if (err == EBUSY)
24410 			break;
24411 		if (rack_mod_inited) {
24412 			uma_zdestroy(rack_zone);
24413 			uma_zdestroy(rack_pcb_zone);
24414 			sysctl_ctx_free(&rack_sysctl_ctx);
24415 			rack_counter_destroy();
24416 			rack_mod_inited = false;
24417 		}
24418 		tcp_lro_dereg_mbufq();
24419 		err = 0;
24420 		break;
24421 	default:
24422 		return (EOPNOTSUPP);
24423 	}
24424 	return (err);
24425 }
24426 
24427 static moduledata_t tcp_rack = {
24428 	.name = __XSTRING(MODNAME),
24429 	.evhand = tcp_addrack,
24430 	.priv = 0
24431 };
24432 
24433 MODULE_VERSION(MODNAME, 1);
24434 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24435 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24436 
24437 #endif /* #if !defined(INET) && !defined(INET6) */
24438