xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 4ba91e076ee84101112d8296785098ae31dac35e) 
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/mutex.h>
44 #include <sys/mbuf.h>
45 #include <sys/proc.h>		/* for proc0 declaration */
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 #ifdef STATS
51 #include <sys/qmath.h>
52 #include <sys/tree.h>
53 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
54 #else
55 #include <sys/tree.h>
56 #endif
57 #include <sys/refcount.h>
58 #include <sys/queue.h>
59 #include <sys/tim_filter.h>
60 #include <sys/smp.h>
61 #include <sys/kthread.h>
62 #include <sys/kern_prefetch.h>
63 #include <sys/protosw.h>
64 #ifdef TCP_ACCOUNTING
65 #include <sys/sched.h>
66 #include <machine/cpu.h>
67 #endif
68 #include <vm/uma.h>
69 
70 #include <net/route.h>
71 #include <net/route/nhop.h>
72 #include <net/vnet.h>
73 
74 #define TCPSTATES		/* for logging */
75 
76 #include <netinet/in.h>
77 #include <netinet/in_kdtrace.h>
78 #include <netinet/in_pcb.h>
79 #include <netinet/ip.h>
80 #include <netinet/ip_var.h>
81 #include <netinet/ip6.h>
82 #include <netinet6/in6_pcb.h>
83 #include <netinet6/ip6_var.h>
84 #include <netinet/tcp.h>
85 #define	TCPOUTFLAGS
86 #include <netinet/tcp_fsm.h>
87 #include <netinet/tcp_seq.h>
88 #include <netinet/tcp_timer.h>
89 #include <netinet/tcp_var.h>
90 #include <netinet/tcp_log_buf.h>
91 #include <netinet/tcp_syncache.h>
92 #include <netinet/tcp_hpts.h>
93 #include <netinet/tcp_ratelimit.h>
94 #include <netinet/tcp_accounting.h>
95 #include <netinet/tcpip.h>
96 #include <netinet/cc/cc.h>
97 #include <netinet/cc/cc_newreno.h>
98 #include <netinet/tcp_fastopen.h>
99 #include <netinet/tcp_lro.h>
100 #ifdef NETFLIX_SHARED_CWND
101 #include <netinet/tcp_shared_cwnd.h>
102 #endif
103 #ifdef TCP_OFFLOAD
104 #include <netinet/tcp_offload.h>
105 #endif
106 #ifdef INET6
107 #include <netinet6/tcp6_var.h>
108 #endif
109 #include <netinet/tcp_ecn.h>
110 
111 #include <netipsec/ipsec_support.h>
112 
113 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
114 #include <netipsec/ipsec.h>
115 #include <netipsec/ipsec6.h>
116 #endif				/* IPSEC */
117 
118 #include <netinet/udp.h>
119 #include <netinet/udp_var.h>
120 #include <machine/in_cksum.h>
121 
122 #ifdef MAC
123 #include <security/mac/mac_framework.h>
124 #endif
125 #include "sack_filter.h"
126 #include "tcp_rack.h"
127 #include "tailq_hash.h"
128 #include "rack_bbr_common.h"
129 
130 uma_zone_t rack_zone;
131 uma_zone_t rack_pcb_zone;
132 
133 #ifndef TICKS2SBT
134 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
135 #endif
136 
137 VNET_DECLARE(uint32_t, newreno_beta);
138 VNET_DECLARE(uint32_t, newreno_beta_ecn);
139 #define V_newreno_beta VNET(newreno_beta)
140 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
141 
142 #define	M_TCPFSB	__CONCAT(M_TCPFSB, STACKNAME)
143 #define	M_TCPDO		__CONCAT(M_TCPDO, STACKNAME)
144 
145 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb_" __XSTRING(STACKNAME), "TCP fast send block");
146 MALLOC_DEFINE(M_TCPDO, "tcp_do_" __XSTRING(STACKNAME), "TCP deferred options");
147 MALLOC_DEFINE(M_TCPPCM, "tcp_pcm_" __XSTRING(STACKNAME), "TCP PCM measurement information");
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_pcm_every_n_rounds = 100;
194 static uint32_t rack_pcm_blast = 0;
195 static uint32_t rack_pcm_is_enabled = 1;
196 static uint8_t rack_ssthresh_rest_rto_rec = 0; /* Do we restore ssthresh when we have rec -> rto -> rec */
197 
198 static uint32_t rack_gp_gain_req = 1200;		/* Amount percent wise required to gain to record a round as "gaining" */
199 static uint32_t rack_rnd_cnt_req = 0x10005;		/* Default number of rounds if we are below rack_gp_gain_req where we exit ss */
200 
201 
202 static int32_t rack_rxt_scoreboard_clear_thresh = 2;
203 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
204 static int32_t rack_rxt_controls = 0;
205 static int32_t rack_fill_cw_state = 0;
206 static uint8_t rack_req_measurements = 1;
207 /* Attack threshold detections */
208 static uint32_t rack_highest_sack_thresh_seen = 0;
209 static uint32_t rack_highest_move_thresh_seen = 0;
210 static uint32_t rack_merge_out_sacks_on_attack = 0;
211 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
212 static int32_t rack_hw_rate_caps = 0; /* 1; */
213 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
214 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
215 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
216 static int32_t rack_hw_up_only = 0;
217 static int32_t rack_stats_gets_ms_rtt = 1;
218 static int32_t rack_prr_addbackmax = 2;
219 static int32_t rack_do_hystart = 0;
220 static int32_t rack_apply_rtt_with_reduced_conf = 0;
221 static int32_t rack_hibeta_setting = 0;
222 static int32_t rack_default_pacing_divisor = 250;
223 static uint16_t rack_pacing_min_seg = 0;
224 static int32_t rack_timely_off = 0;
225 
226 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
227 static int32_t rack_pkt_delay = 1000;
228 static int32_t rack_send_a_lot_in_prr = 1;
229 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
230 static int32_t rack_verbose_logging = 0;
231 static int32_t rack_ignore_data_after_close = 1;
232 static int32_t rack_enable_shared_cwnd = 1;
233 static int32_t rack_use_cmp_acks = 1;
234 static int32_t rack_use_fsb = 1;
235 static int32_t rack_use_rfo = 1;
236 static int32_t rack_use_rsm_rfo = 1;
237 static int32_t rack_max_abc_post_recovery = 2;
238 static int32_t rack_client_low_buf = 0;
239 static int32_t rack_dsack_std_based = 0x3;	/* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
240 static int32_t rack_bw_multipler = 0;		/* Limit on fill cw's jump up to be this x gp_est */
241 #ifdef TCP_ACCOUNTING
242 static int32_t rack_tcp_accounting = 0;
243 #endif
244 static int32_t rack_limits_scwnd = 1;
245 static int32_t rack_enable_mqueue_for_nonpaced = 0;
246 static int32_t rack_hybrid_allow_set_maxseg = 0;
247 static int32_t rack_disable_prr = 0;
248 static int32_t use_rack_rr = 1;
249 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
250 static int32_t rack_persist_min = 250000;	/* 250usec */
251 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
252 static int32_t rack_honors_hpts_min_to =  1;	/* Do we honor the hpts minimum time out for pacing timers */
253 static uint32_t rack_max_reduce = 10;		/* Percent we can reduce slot by */
254 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
255 static int32_t rack_limit_time_with_srtt = 0;
256 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
257 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
258 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
259 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
260 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
261 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
262 
263 /*
264  * Currently regular tcp has a rto_min of 30ms
265  * the backoff goes 12 times so that ends up
266  * being a total of 122.850 seconds before a
267  * connection is killed.
268  */
269 static uint32_t rack_def_data_window = 20;
270 static uint32_t rack_goal_bdp = 2;
271 static uint32_t rack_min_srtts = 1;
272 static uint32_t rack_min_measure_usec = 0;
273 static int32_t rack_tlp_min = 10000;	/* 10ms */
274 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
275 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
276 static const int32_t rack_free_cache = 2;
277 static int32_t rack_hptsi_segments = 40;
278 static int32_t rack_rate_sample_method = USE_RTT_LOW;
279 static int32_t rack_pace_every_seg = 0;
280 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
281 static int32_t rack_slot_reduction = 4;
282 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
283 static int32_t rack_cwnd_block_ends_measure = 0;
284 static int32_t rack_rwnd_block_ends_measure = 0;
285 static int32_t rack_def_profile = 0;
286 
287 static int32_t rack_lower_cwnd_at_tlp = 0;
288 static int32_t rack_always_send_oldest = 0;
289 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
290 
291 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
292 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
293 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
294 
295 /* Probertt */
296 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
297 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
298 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
299 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
300 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
301 
302 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
303 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
304 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
305 static uint32_t rack_probertt_use_min_rtt_exit = 0;
306 static uint32_t rack_probe_rtt_sets_cwnd = 0;
307 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
308 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
309 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
310 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
311 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
312 static uint32_t rack_probertt_filter_life = 10000000;
313 static uint32_t rack_probertt_lower_within = 10;
314 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
315 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
316 static int32_t rack_probertt_clear_is = 1;
317 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
318 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
319 
320 /* Part of pacing */
321 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
322 
323 /* Timely information:
324  *
325  * Here we have various control parameters on how
326  * timely may change the multiplier. rack_gain_p5_ub
327  * is associated with timely but not directly influencing
328  * the rate decision like the other variables. It controls
329  * the way fill-cw interacts with timely and caps how much
330  * timely can boost the fill-cw b/w.
331  *
332  * The other values are various boost/shrink numbers as well
333  * as potential caps when adjustments are made to the timely
334  * gain (returned by rack_get_output_gain(). Remember too that
335  * the gain returned can be overriden by other factors such as
336  * probeRTT as well as fixed-rate-pacing.
337  */
338 static int32_t rack_gain_p5_ub = 250;
339 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
340 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
341 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
342 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
343 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
344 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
345 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
346 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
347 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
348 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
349 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
350 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
351 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
352 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
353 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
354 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
355 static int32_t rack_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 static uint64_t rack_fillcw_bw_cap = 3750000;	/* Cap fillcw at 30Mbps */
360 
361 
362 /* Rack specific counters */
363 counter_u64_t rack_saw_enobuf;
364 counter_u64_t rack_saw_enobuf_hw;
365 counter_u64_t rack_saw_enetunreach;
366 counter_u64_t rack_persists_sends;
367 counter_u64_t rack_persists_acks;
368 counter_u64_t rack_persists_loss;
369 counter_u64_t rack_persists_lost_ends;
370 counter_u64_t rack_total_bytes;
371 #ifdef INVARIANTS
372 counter_u64_t rack_adjust_map_bw;
373 #endif
374 /* Tail loss probe counters */
375 counter_u64_t rack_tlp_tot;
376 counter_u64_t rack_tlp_newdata;
377 counter_u64_t rack_tlp_retran;
378 counter_u64_t rack_tlp_retran_bytes;
379 counter_u64_t rack_to_tot;
380 counter_u64_t rack_hot_alloc;
381 counter_u64_t rack_to_alloc;
382 counter_u64_t rack_to_alloc_hard;
383 counter_u64_t rack_to_alloc_emerg;
384 counter_u64_t rack_to_alloc_limited;
385 counter_u64_t rack_alloc_limited_conns;
386 counter_u64_t rack_split_limited;
387 counter_u64_t rack_rxt_clamps_cwnd;
388 counter_u64_t rack_rxt_clamps_cwnd_uniq;
389 
390 counter_u64_t rack_multi_single_eq;
391 counter_u64_t rack_proc_non_comp_ack;
392 
393 counter_u64_t rack_fto_send;
394 counter_u64_t rack_fto_rsm_send;
395 counter_u64_t rack_nfto_resend;
396 counter_u64_t rack_non_fto_send;
397 counter_u64_t rack_extended_rfo;
398 
399 counter_u64_t rack_sack_proc_all;
400 counter_u64_t rack_sack_proc_short;
401 counter_u64_t rack_sack_proc_restart;
402 counter_u64_t rack_sack_attacks_detected;
403 counter_u64_t rack_sack_attacks_reversed;
404 counter_u64_t rack_sack_attacks_suspect;
405 counter_u64_t rack_sack_used_next_merge;
406 counter_u64_t rack_sack_splits;
407 counter_u64_t rack_sack_used_prev_merge;
408 counter_u64_t rack_sack_skipped_acked;
409 counter_u64_t rack_ack_total;
410 counter_u64_t rack_express_sack;
411 counter_u64_t rack_sack_total;
412 counter_u64_t rack_move_none;
413 counter_u64_t rack_move_some;
414 
415 counter_u64_t rack_input_idle_reduces;
416 counter_u64_t rack_collapsed_win;
417 counter_u64_t rack_collapsed_win_seen;
418 counter_u64_t rack_collapsed_win_rxt;
419 counter_u64_t rack_collapsed_win_rxt_bytes;
420 counter_u64_t rack_try_scwnd;
421 counter_u64_t rack_hw_pace_init_fail;
422 counter_u64_t rack_hw_pace_lost;
423 
424 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
425 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
426 
427 
428 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
429 
430 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
431 	(tv) = (value) + slop;	 \
432 	if ((u_long)(tv) < (u_long)(tvmin)) \
433 		(tv) = (tvmin); \
434 	if ((u_long)(tv) > (u_long)(tvmax)) \
435 		(tv) = (tvmax); \
436 } while (0)
437 
438 static void
439 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
440 
441 static int
442 rack_process_ack(struct mbuf *m, struct tcphdr *th,
443     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
444     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val, int32_t orig_tlen);
445 static int
446 rack_process_data(struct mbuf *m, struct tcphdr *th,
447     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
448     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
449 static void
450 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
451    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
452 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
453 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
454     uint8_t limit_type);
455 static struct rack_sendmap *
456 rack_check_recovery_mode(struct tcpcb *tp,
457     uint32_t tsused);
458 static uint32_t
459 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack);
460 static void
461 rack_cong_signal(struct tcpcb *tp,
462 		 uint32_t type, uint32_t ack, int );
463 static void rack_counter_destroy(void);
464 static int
465 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
466 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
467 static void
468 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
469 static void
470 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
471     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
472 static void rack_dtor(void *mem, int32_t size, void *arg);
473 static void
474 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
475     uint32_t flex1, uint32_t flex2,
476     uint32_t flex3, uint32_t flex4,
477     uint32_t flex5, uint32_t flex6,
478     uint16_t flex7, uint8_t mod);
479 
480 static void
481 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
482    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
483    struct rack_sendmap *rsm, uint8_t quality);
484 static struct rack_sendmap *
485 rack_find_high_nonack(struct tcp_rack *rack,
486     struct rack_sendmap *rsm);
487 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
488 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
489 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
490 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
491 static void
492 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
493 			    tcp_seq th_ack, int line, uint8_t quality);
494 static void
495 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
496 
497 static uint32_t
498 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
499 static int32_t rack_handoff_ok(struct tcpcb *tp);
500 static int32_t rack_init(struct tcpcb *tp, void **ptr);
501 static void rack_init_sysctls(void);
502 
503 static void
504 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
505     struct tcphdr *th, int entered_rec, int dup_ack_struck,
506     int *dsack_seen, int *sacks_seen);
507 static void
508 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
509     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
510     struct rack_sendmap *hintrsm, uint32_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
511 
512 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
513 
514 
515 static void
516 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
517     struct rack_sendmap *rsm, uint32_t cts);
518 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
519 static int32_t rack_output(struct tcpcb *tp);
520 
521 static uint32_t
522 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
523     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
524     uint32_t cts, uint32_t segsiz);
525 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
526 static void rack_remxt_tmr(struct tcpcb *tp);
527 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
528 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
529 static int32_t rack_stopall(struct tcpcb *tp);
530 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
531 static uint32_t
532 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
533     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint32_t add_flag, int segsiz);
534 static void
535 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
536     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz);
537 static int
538 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
539     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
540 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
541 static int
542 rack_do_close_wait(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 
546 static int
547 rack_do_closing(struct mbuf *m, struct tcphdr *th,
548     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
549     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
550 static int
551 rack_do_established(struct mbuf *m, struct tcphdr *th,
552     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
553     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
554 static int
555 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
556     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
557     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
558 static int
559 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
560     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
561     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
562 static int
563 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
564     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
565     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
566 static int
567 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
568     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
569     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
570 static int
571 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
572     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
573     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
574 static int
575 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
576     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
577     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
578 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
579 struct rack_sendmap *
580 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
581     uint32_t tsused);
582 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
583     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
584 static void
585      tcp_rack_partialack(struct tcpcb *tp);
586 static int
587 rack_set_profile(struct tcp_rack *rack, int prof);
588 static void
589 rack_apply_deferred_options(struct tcp_rack *rack);
590 
591 int32_t rack_clear_counter=0;
592 
593 static uint64_t
594 rack_get_lt_bw(struct tcp_rack *rack)
595 {
596 	struct timeval tv;
597 	uint64_t tim, bytes;
598 
599 	tim = rack->r_ctl.lt_bw_time;
600 	bytes = rack->r_ctl.lt_bw_bytes;
601 	if (rack->lt_bw_up) {
602 		/* Include all the current bytes too */
603 		microuptime(&tv);
604 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
605 		tim += (tcp_tv_to_lusec(&tv) - rack->r_ctl.lt_timemark);
606 	}
607 	if ((bytes != 0) && (tim != 0))
608 		return ((bytes * (uint64_t)1000000) / tim);
609 	else
610 		return (0);
611 }
612 
613 static void
614 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
615 {
616 	struct sockopt sopt;
617 	struct cc_newreno_opts opt;
618 	struct tcpcb *tp;
619 	uint32_t old_beta;
620 	uint32_t old_beta_ecn;
621 	int error = 0, failed = 0;
622 
623 	tp = rack->rc_tp;
624 	if (tp->t_cc == NULL) {
625 		/* Tcb is leaving */
626 		return;
627 	}
628 	rack->rc_pacing_cc_set = 1;
629 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
630 		/* Not new-reno we can't play games with beta! */
631 		failed = 1;
632 		goto out;
633 
634 	}
635 	if (CC_ALGO(tp)->ctl_output == NULL)  {
636 		/* Huh, not using new-reno so no swaps.? */
637 		failed = 2;
638 		goto out;
639 	}
640 	/* Get the current values out */
641 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
642 	sopt.sopt_dir = SOPT_GET;
643 	opt.name = CC_NEWRENO_BETA;
644 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
645 	if (error)  {
646 		failed = 3;
647 		goto out;
648 	}
649 	old_beta = opt.val;
650 	opt.name = CC_NEWRENO_BETA_ECN;
651 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
652 	if (error)  {
653 		failed = 4;
654 		goto out;
655 	}
656 	old_beta_ecn = opt.val;
657 
658 	/* Now lets set in the values we have stored */
659 	sopt.sopt_dir = SOPT_SET;
660 	opt.name = CC_NEWRENO_BETA;
661 	opt.val = rack->r_ctl.rc_saved_beta;
662 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
663 	if (error)  {
664 		failed = 5;
665 		goto out;
666 	}
667 	opt.name = CC_NEWRENO_BETA_ECN;
668 	opt.val = rack->r_ctl.rc_saved_beta_ecn;
669 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
670 	if (error) {
671 		failed = 6;
672 		goto out;
673 	}
674 	/* Save off the values for restoral */
675 	rack->r_ctl.rc_saved_beta = old_beta;
676 	rack->r_ctl.rc_saved_beta_ecn = old_beta_ecn;
677 out:
678 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
679 		union tcp_log_stackspecific log;
680 		struct timeval tv;
681 		struct newreno *ptr;
682 
683 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
684 		memset(&log, 0, sizeof(log));
685 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
686 		log.u_bbr.flex1 = ptr->beta;
687 		log.u_bbr.flex2 = ptr->beta_ecn;
688 		log.u_bbr.flex3 = ptr->newreno_flags;
689 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta;
690 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta_ecn;
691 		log.u_bbr.flex6 = failed;
692 		log.u_bbr.flex7 = rack->gp_ready;
693 		log.u_bbr.flex7 <<= 1;
694 		log.u_bbr.flex7 |= rack->use_fixed_rate;
695 		log.u_bbr.flex7 <<= 1;
696 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
697 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
698 		log.u_bbr.flex8 = flex8;
699 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
700 			       0, &log, false, NULL, NULL, 0, &tv);
701 	}
702 }
703 
704 static void
705 rack_set_cc_pacing(struct tcp_rack *rack)
706 {
707 	if (rack->rc_pacing_cc_set)
708 		return;
709 	/*
710 	 * Use the swap utility placing in 3 for flex8 to id a
711 	 * set of a new set of values.
712 	 */
713 	rack->rc_pacing_cc_set = 1;
714 	rack_swap_beta_values(rack, 3);
715 }
716 
717 static void
718 rack_undo_cc_pacing(struct tcp_rack *rack)
719 {
720 	if (rack->rc_pacing_cc_set == 0)
721 		return;
722 	/*
723 	 * Use the swap utility placing in 4 for flex8 to id a
724 	 * restoral of the old values.
725 	 */
726 	rack->rc_pacing_cc_set = 0;
727 	rack_swap_beta_values(rack, 4);
728 }
729 
730 static void
731 rack_remove_pacing(struct tcp_rack *rack)
732 {
733 	if (rack->rc_pacing_cc_set)
734 		rack_undo_cc_pacing(rack);
735 	if (rack->r_ctl.pacing_method & RACK_REG_PACING)
736 		tcp_decrement_paced_conn();
737 	if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
738 		tcp_dec_dgp_pacing_cnt();
739 	rack->rc_always_pace = 0;
740 	rack->r_ctl.pacing_method = RACK_PACING_NONE;
741 	rack->dgp_on = 0;
742 	rack->rc_hybrid_mode = 0;
743 	rack->use_fixed_rate = 0;
744 }
745 
746 static void
747 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
748 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
749 {
750 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
751 		union tcp_log_stackspecific log;
752 		struct timeval tv;
753 
754 		memset(&log, 0, sizeof(log));
755 		log.u_bbr.flex1 = seq_end;
756 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
757 		log.u_bbr.flex3 = ack_end_t;
758 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
759 		log.u_bbr.flex5 = send_end_t;
760 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
761 		log.u_bbr.flex7 = mode;
762 		log.u_bbr.flex8 = 69;
763 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
764 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
765 		log.u_bbr.pkts_out = line;
766 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
767 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
768 		log.u_bbr.epoch = rack->r_ctl.current_round;
769 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
770 		if (rsm != NULL) {
771 			log.u_bbr.applimited = rsm->r_start;
772 			log.u_bbr.delivered = rsm->r_end;
773 			log.u_bbr.epoch = rsm->r_flags;
774 		}
775 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
776 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
777 		    &rack->rc_inp->inp_socket->so_rcv,
778 		    &rack->rc_inp->inp_socket->so_snd,
779 		    BBR_LOG_HPTSI_CALC, 0,
780 		    0, &log, false, &tv);
781 	}
782 }
783 
784 static int
785 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
786 {
787 	uint32_t stat;
788 	int32_t error;
789 
790 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
791 	if (error || req->newptr == NULL)
792 		return error;
793 
794 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
795 	if (error)
796 		return (error);
797 	if (stat == 1) {
798 #ifdef INVARIANTS
799 		printf("Clearing RACK counters\n");
800 #endif
801 		counter_u64_zero(rack_tlp_tot);
802 		counter_u64_zero(rack_tlp_newdata);
803 		counter_u64_zero(rack_tlp_retran);
804 		counter_u64_zero(rack_tlp_retran_bytes);
805 		counter_u64_zero(rack_to_tot);
806 		counter_u64_zero(rack_saw_enobuf);
807 		counter_u64_zero(rack_saw_enobuf_hw);
808 		counter_u64_zero(rack_saw_enetunreach);
809 		counter_u64_zero(rack_persists_sends);
810 		counter_u64_zero(rack_total_bytes);
811 		counter_u64_zero(rack_persists_acks);
812 		counter_u64_zero(rack_persists_loss);
813 		counter_u64_zero(rack_persists_lost_ends);
814 #ifdef INVARIANTS
815 		counter_u64_zero(rack_adjust_map_bw);
816 #endif
817 		counter_u64_zero(rack_to_alloc_hard);
818 		counter_u64_zero(rack_to_alloc_emerg);
819 		counter_u64_zero(rack_sack_proc_all);
820 		counter_u64_zero(rack_fto_send);
821 		counter_u64_zero(rack_fto_rsm_send);
822 		counter_u64_zero(rack_extended_rfo);
823 		counter_u64_zero(rack_hw_pace_init_fail);
824 		counter_u64_zero(rack_hw_pace_lost);
825 		counter_u64_zero(rack_non_fto_send);
826 		counter_u64_zero(rack_nfto_resend);
827 		counter_u64_zero(rack_sack_proc_short);
828 		counter_u64_zero(rack_sack_proc_restart);
829 		counter_u64_zero(rack_to_alloc);
830 		counter_u64_zero(rack_to_alloc_limited);
831 		counter_u64_zero(rack_alloc_limited_conns);
832 		counter_u64_zero(rack_split_limited);
833 		counter_u64_zero(rack_rxt_clamps_cwnd);
834 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
835 		counter_u64_zero(rack_multi_single_eq);
836 		counter_u64_zero(rack_proc_non_comp_ack);
837 		counter_u64_zero(rack_sack_attacks_detected);
838 		counter_u64_zero(rack_sack_attacks_reversed);
839 		counter_u64_zero(rack_sack_attacks_suspect);
840 		counter_u64_zero(rack_sack_used_next_merge);
841 		counter_u64_zero(rack_sack_used_prev_merge);
842 		counter_u64_zero(rack_sack_splits);
843 		counter_u64_zero(rack_sack_skipped_acked);
844 		counter_u64_zero(rack_ack_total);
845 		counter_u64_zero(rack_express_sack);
846 		counter_u64_zero(rack_sack_total);
847 		counter_u64_zero(rack_move_none);
848 		counter_u64_zero(rack_move_some);
849 		counter_u64_zero(rack_try_scwnd);
850 		counter_u64_zero(rack_collapsed_win);
851 		counter_u64_zero(rack_collapsed_win_rxt);
852 		counter_u64_zero(rack_collapsed_win_seen);
853 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
854 	} else if (stat == 2) {
855 #ifdef INVARIANTS
856 		printf("Clearing RACK option array\n");
857 #endif
858 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
859 	} else if (stat == 3) {
860 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
861 	} else if (stat == 4) {
862 #ifdef INVARIANTS
863 		printf("Clearing RACK out size array\n");
864 #endif
865 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
866 	}
867 	rack_clear_counter = 0;
868 	return (0);
869 }
870 
871 static void
872 rack_init_sysctls(void)
873 {
874 	struct sysctl_oid *rack_counters;
875 	struct sysctl_oid *rack_attack;
876 	struct sysctl_oid *rack_pacing;
877 	struct sysctl_oid *rack_timely;
878 	struct sysctl_oid *rack_timers;
879 	struct sysctl_oid *rack_tlp;
880 	struct sysctl_oid *rack_misc;
881 	struct sysctl_oid *rack_features;
882 	struct sysctl_oid *rack_measure;
883 	struct sysctl_oid *rack_probertt;
884 	struct sysctl_oid *rack_hw_pacing;
885 
886 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
887 	    SYSCTL_CHILDREN(rack_sysctl_root),
888 	    OID_AUTO,
889 	    "sack_attack",
890 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
891 	    "Rack Sack Attack Counters and Controls");
892 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
893 	    SYSCTL_CHILDREN(rack_sysctl_root),
894 	    OID_AUTO,
895 	    "stats",
896 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
897 	    "Rack Counters");
898 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
899 	    SYSCTL_CHILDREN(rack_sysctl_root),
900 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
901 	    &rack_rate_sample_method , USE_RTT_LOW,
902 	    "What method should we use for rate sampling 0=high, 1=low ");
903 	/* Probe rtt related controls */
904 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
905 	    SYSCTL_CHILDREN(rack_sysctl_root),
906 	    OID_AUTO,
907 	    "probertt",
908 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
909 	    "ProbeRTT related Controls");
910 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
911 	    SYSCTL_CHILDREN(rack_probertt),
912 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
913 	    &rack_atexit_prtt_hbp, 130,
914 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
915 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_probertt),
917 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
918 	    &rack_atexit_prtt, 130,
919 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
920 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
921 	    SYSCTL_CHILDREN(rack_probertt),
922 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
923 	    &rack_per_of_gp_probertt, 60,
924 	    "What percentage of goodput do we pace at in probertt");
925 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
926 	    SYSCTL_CHILDREN(rack_probertt),
927 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
928 	    &rack_per_of_gp_probertt_reduce, 10,
929 	    "What percentage of goodput do we reduce every gp_srtt");
930 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
931 	    SYSCTL_CHILDREN(rack_probertt),
932 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
933 	    &rack_per_of_gp_lowthresh, 40,
934 	    "What percentage of goodput do we allow the multiplier to fall to");
935 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_probertt),
937 	    OID_AUTO, "time_between", CTLFLAG_RW,
938 	    &rack_time_between_probertt, 96000000,
939 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
940 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_probertt),
942 	    OID_AUTO, "safety", CTLFLAG_RW,
943 	    &rack_probe_rtt_safety_val, 2000000,
944 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
945 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_probertt),
947 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
948 	    &rack_probe_rtt_sets_cwnd, 0,
949 	    "Do we set the cwnd too (if always_lower is on)");
950 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_probertt),
952 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
953 	    &rack_max_drain_wait, 2,
954 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
955 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_probertt),
957 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
958 	    &rack_must_drain, 1,
959 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
960 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_probertt),
962 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
963 	    &rack_probertt_use_min_rtt_entry, 1,
964 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
965 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_probertt),
967 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
968 	    &rack_probertt_use_min_rtt_exit, 0,
969 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
970 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
971 	    SYSCTL_CHILDREN(rack_probertt),
972 	    OID_AUTO, "length_div", CTLFLAG_RW,
973 	    &rack_probertt_gpsrtt_cnt_div, 0,
974 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
975 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
976 	    SYSCTL_CHILDREN(rack_probertt),
977 	    OID_AUTO, "length_mul", CTLFLAG_RW,
978 	    &rack_probertt_gpsrtt_cnt_mul, 0,
979 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
980 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_probertt),
982 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
983 	    &rack_min_probertt_hold, 200000,
984 	    "What is the minimum time we hold probertt at target");
985 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
986 	    SYSCTL_CHILDREN(rack_probertt),
987 	    OID_AUTO, "filter_life", CTLFLAG_RW,
988 	    &rack_probertt_filter_life, 10000000,
989 	    "What is the time for the filters life in useconds");
990 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
991 	    SYSCTL_CHILDREN(rack_probertt),
992 	    OID_AUTO, "lower_within", CTLFLAG_RW,
993 	    &rack_probertt_lower_within, 10,
994 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
995 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
996 	    SYSCTL_CHILDREN(rack_probertt),
997 	    OID_AUTO, "must_move", CTLFLAG_RW,
998 	    &rack_min_rtt_movement, 250,
999 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1000 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1001 	    SYSCTL_CHILDREN(rack_probertt),
1002 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1003 	    &rack_probertt_clear_is, 1,
1004 	    "Do we clear I/S counts on exiting probe-rtt");
1005 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1006 	    SYSCTL_CHILDREN(rack_probertt),
1007 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1008 	    &rack_max_drain_hbp, 1,
1009 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
1010 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1011 	    SYSCTL_CHILDREN(rack_probertt),
1012 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1013 	    &rack_hbp_thresh, 3,
1014 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1015 	/* Pacing related sysctls */
1016 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1017 	    SYSCTL_CHILDREN(rack_sysctl_root),
1018 	    OID_AUTO,
1019 	    "pacing",
1020 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1021 	    "Pacing related Controls");
1022 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_pacing),
1024 	    OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1025 	    &rack_pcm_is_enabled, 1,
1026 	    "Do we by default do PCM measurements?");
1027 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_pacing),
1029 	    OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1030 	    &rack_pcm_every_n_rounds, 100,
1031 	    "How many rounds before we need to do a PCM measurement");
1032 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_pacing),
1034 	    OID_AUTO, "pcm_blast", CTLFLAG_RW,
1035 	    &rack_pcm_blast, 0,
1036 	    "Blast out the full cwnd/rwnd when doing a PCM measurement");
1037 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1038 	    SYSCTL_CHILDREN(rack_pacing),
1039 	    OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1040 	    &rack_gp_gain_req, 1200,
1041 	    "How much do we have to increase the GP to record the round 1200 = 120.0");
1042 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1043 	    SYSCTL_CHILDREN(rack_pacing),
1044 	    OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1045 	    &rack_rnd_cnt_req, 0x10005,
1046 	    "How many rounds less than rnd_gp_gain will drop us out of SS");
1047 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1048 	    SYSCTL_CHILDREN(rack_pacing),
1049 	    OID_AUTO, "no_timely", CTLFLAG_RW,
1050 	    &rack_timely_off, 0,
1051 	    "Do we not use timely in DGP?");
1052 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1053 	    SYSCTL_CHILDREN(rack_pacing),
1054 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1055 	    &rack_fill_cw_state, 0,
1056 	    "Enable fillcw on new connections (default=0 off)?");
1057 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1058 	    SYSCTL_CHILDREN(rack_pacing),
1059 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1060 	    &rack_pacing_min_seg, 0,
1061 	    "What is the min burst size for pacing (0 disables)?");
1062 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1063 	    SYSCTL_CHILDREN(rack_pacing),
1064 	    OID_AUTO, "divisor", CTLFLAG_RW,
1065 	    &rack_default_pacing_divisor, 250,
1066 	    "What is the default divisor given to the rl code?");
1067 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1068 	    SYSCTL_CHILDREN(rack_pacing),
1069 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1070 	    &rack_bw_multipler, 0,
1071 	    "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1072 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1073 	    SYSCTL_CHILDREN(rack_pacing),
1074 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1075 	    &rack_max_per_above, 30,
1076 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1077 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1078 	    SYSCTL_CHILDREN(rack_pacing),
1079 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1080 	    &rack_pace_one_seg, 0,
1081 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1082 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1083 	    SYSCTL_CHILDREN(rack_pacing),
1084 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1085 	    &rack_limit_time_with_srtt, 0,
1086 	    "Do we limit pacing time based on srtt");
1087 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1088 	    SYSCTL_CHILDREN(rack_pacing),
1089 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1090 	    &rack_per_of_gp_ss, 250,
1091 	    "If non zero, what percentage of goodput to pace at in slow start");
1092 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1093 	    SYSCTL_CHILDREN(rack_pacing),
1094 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1095 	    &rack_per_of_gp_ca, 150,
1096 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1097 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1098 	    SYSCTL_CHILDREN(rack_pacing),
1099 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1100 	    &rack_per_of_gp_rec, 200,
1101 	    "If non zero, what percentage of goodput to pace at in recovery");
1102 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1103 	    SYSCTL_CHILDREN(rack_pacing),
1104 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1105 	    &rack_hptsi_segments, 40,
1106 	    "What size is the max for TSO segments in pacing and burst mitigation");
1107 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1108 	    SYSCTL_CHILDREN(rack_pacing),
1109 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1110 	    &rack_slot_reduction, 4,
1111 	    "When doing only burst mitigation what is the reduce divisor");
1112 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1113 	    SYSCTL_CHILDREN(rack_sysctl_root),
1114 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1115 	    &rack_pace_every_seg, 0,
1116 	    "If set we use pacing, if clear we use only the original burst mitigation");
1117 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1118 	    SYSCTL_CHILDREN(rack_pacing),
1119 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1120 	    &rack_bw_rate_cap, 0,
1121 	    "If set we apply this value to the absolute rate cap used by pacing");
1122 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1123 	    SYSCTL_CHILDREN(rack_pacing),
1124 	    OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1125 	    &rack_fillcw_bw_cap, 3750000,
1126 	    "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1127 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1128 	    SYSCTL_CHILDREN(rack_sysctl_root),
1129 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1130 	    &rack_req_measurements, 1,
1131 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1132 	/* Hardware pacing */
1133 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1134 	    SYSCTL_CHILDREN(rack_sysctl_root),
1135 	    OID_AUTO,
1136 	    "hdwr_pacing",
1137 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1138 	    "Pacing related Controls");
1139 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_hw_pacing),
1141 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1142 	    &rack_hw_rwnd_factor, 2,
1143 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1144 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_hw_pacing),
1146 	    OID_AUTO, "precheck", CTLFLAG_RW,
1147 	    &rack_hw_check_queue, 0,
1148 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1149 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_hw_pacing),
1151 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1152 	    &rack_enobuf_hw_boost_mult, 0,
1153 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1154 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1155 	    SYSCTL_CHILDREN(rack_hw_pacing),
1156 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1157 	    &rack_enobuf_hw_max, 2,
1158 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1159 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1160 	    SYSCTL_CHILDREN(rack_hw_pacing),
1161 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1162 	    &rack_enobuf_hw_min, 2,
1163 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1164 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1165 	    SYSCTL_CHILDREN(rack_hw_pacing),
1166 	    OID_AUTO, "enable", CTLFLAG_RW,
1167 	    &rack_enable_hw_pacing, 0,
1168 	    "Should RACK attempt to use hw pacing?");
1169 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1170 	    SYSCTL_CHILDREN(rack_hw_pacing),
1171 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1172 	    &rack_hw_rate_caps, 0,
1173 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1174 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1175 	    SYSCTL_CHILDREN(rack_hw_pacing),
1176 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1177 	    &rack_hw_rate_cap_per, 0,
1178 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1179 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1180 	    SYSCTL_CHILDREN(rack_hw_pacing),
1181 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1182 	    &rack_hw_rate_min, 0,
1183 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1184 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1185 	    SYSCTL_CHILDREN(rack_hw_pacing),
1186 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1187 	    &rack_hw_rate_to_low, 0,
1188 	    "If we fall below this rate, dis-engage hw pacing?");
1189 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1190 	    SYSCTL_CHILDREN(rack_hw_pacing),
1191 	    OID_AUTO, "up_only", CTLFLAG_RW,
1192 	    &rack_hw_up_only, 0,
1193 	    "Do we allow hw pacing to lower the rate selected?");
1194 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1195 	    SYSCTL_CHILDREN(rack_sysctl_root),
1196 	    OID_AUTO,
1197 	    "timely",
1198 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1199 	    "Rack Timely RTT Controls");
1200 	/* Timely based GP dynmics */
1201 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(rack_timely),
1203 	    OID_AUTO, "upper", CTLFLAG_RW,
1204 	    &rack_gp_per_bw_mul_up, 2,
1205 	    "Rack timely upper range for equal b/w (in percentage)");
1206 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1207 	    SYSCTL_CHILDREN(rack_timely),
1208 	    OID_AUTO, "lower", CTLFLAG_RW,
1209 	    &rack_gp_per_bw_mul_down, 4,
1210 	    "Rack timely lower range for equal b/w (in percentage)");
1211 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1212 	    SYSCTL_CHILDREN(rack_timely),
1213 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1214 	    &rack_gp_rtt_maxmul, 3,
1215 	    "Rack timely multiplier of lowest rtt for rtt_max");
1216 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1217 	    SYSCTL_CHILDREN(rack_timely),
1218 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1219 	    &rack_gp_rtt_mindiv, 4,
1220 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1221 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1222 	    SYSCTL_CHILDREN(rack_timely),
1223 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1224 	    &rack_gp_rtt_minmul, 1,
1225 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1226 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1227 	    SYSCTL_CHILDREN(rack_timely),
1228 	    OID_AUTO, "decrease", CTLFLAG_RW,
1229 	    &rack_gp_decrease_per, 80,
1230 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1231 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1232 	    SYSCTL_CHILDREN(rack_timely),
1233 	    OID_AUTO, "increase", CTLFLAG_RW,
1234 	    &rack_gp_increase_per, 2,
1235 	    "Rack timely increase perentage of our GP multiplication factor");
1236 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1237 	    SYSCTL_CHILDREN(rack_timely),
1238 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1239 	    &rack_per_lower_bound, 50,
1240 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1241 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1242 	    SYSCTL_CHILDREN(rack_timely),
1243 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1244 	    &rack_gain_p5_ub, 250,
1245 	    "Profile 5 upper bound to timely gain");
1246 
1247 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1248 	    SYSCTL_CHILDREN(rack_timely),
1249 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1250 	    &rack_per_upper_bound_ss, 0,
1251 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1252 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1253 	    SYSCTL_CHILDREN(rack_timely),
1254 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1255 	    &rack_per_upper_bound_ca, 0,
1256 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1257 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1258 	    SYSCTL_CHILDREN(rack_timely),
1259 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1260 	    &rack_do_dyn_mul, 0,
1261 	    "Rack timely do we enable dynmaic timely goodput by default");
1262 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1263 	    SYSCTL_CHILDREN(rack_timely),
1264 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1265 	    &rack_gp_no_rec_chg, 1,
1266 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1267 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1268 	    SYSCTL_CHILDREN(rack_timely),
1269 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1270 	    &rack_timely_dec_clear, 6,
1271 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1272 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1273 	    SYSCTL_CHILDREN(rack_timely),
1274 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1275 	    &rack_timely_max_push_rise, 3,
1276 	    "Rack timely how many times do we push up with b/w increase");
1277 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1278 	    SYSCTL_CHILDREN(rack_timely),
1279 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1280 	    &rack_timely_max_push_drop, 3,
1281 	    "Rack timely how many times do we push back on b/w decent");
1282 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1283 	    SYSCTL_CHILDREN(rack_timely),
1284 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1285 	    &rack_timely_min_segs, 4,
1286 	    "Rack timely when setting the cwnd what is the min num segments");
1287 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(rack_timely),
1289 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1290 	    &rack_timely_no_stopping, 0,
1291 	    "Rack timely don't stop increase");
1292 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(rack_timely),
1294 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1295 	    &rack_down_raise_thresh, 100,
1296 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1297 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(rack_timely),
1299 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1300 	    &rack_req_segs, 1,
1301 	    "Bottom dragging if not these many segments outstanding and room");
1302 
1303 	/* TLP and Rack related parameters */
1304 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1305 	    SYSCTL_CHILDREN(rack_sysctl_root),
1306 	    OID_AUTO,
1307 	    "tlp",
1308 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1309 	    "TLP and Rack related Controls");
1310 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1311 	    SYSCTL_CHILDREN(rack_tlp),
1312 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1313 	    &use_rack_rr, 1,
1314 	    "Do we use Rack Rapid Recovery");
1315 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1316 	    SYSCTL_CHILDREN(rack_tlp),
1317 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1318 	    &rack_max_abc_post_recovery, 2,
1319 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1320 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1321 	    SYSCTL_CHILDREN(rack_tlp),
1322 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1323 	    &rack_non_rxt_use_cr, 0,
1324 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1325 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1326 	    SYSCTL_CHILDREN(rack_tlp),
1327 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1328 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1329 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1330 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_tlp),
1332 	    OID_AUTO, "limit", CTLFLAG_RW,
1333 	    &rack_tlp_limit, 2,
1334 	    "How many TLP's can be sent without sending new data");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_tlp),
1337 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1338 	    &rack_tlp_use_greater, 1,
1339 	    "Should we use the rack_rtt time if its greater than srtt");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_tlp),
1342 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1343 	    &rack_tlp_min, 10000,
1344 	    "TLP minimum timeout per the specification (in microseconds)");
1345 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_tlp),
1347 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1348 	    &rack_always_send_oldest, 0,
1349 	    "Should we always send the oldest TLP and RACK-TLP");
1350 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1351 	    SYSCTL_CHILDREN(rack_tlp),
1352 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1353 	    &rack_lower_cwnd_at_tlp, 0,
1354 	    "When a TLP completes a retran should we enter recovery");
1355 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1356 	    SYSCTL_CHILDREN(rack_tlp),
1357 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1358 	    &rack_reorder_thresh, 2,
1359 	    "What factor for rack will be added when seeing reordering (shift right)");
1360 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_tlp),
1362 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1363 	    &rack_tlp_thresh, 1,
1364 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1365 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1366 	    SYSCTL_CHILDREN(rack_tlp),
1367 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1368 	    &rack_reorder_fade, 60000000,
1369 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1370 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1371 	    SYSCTL_CHILDREN(rack_tlp),
1372 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1373 	    &rack_pkt_delay, 1000,
1374 	    "Extra RACK time (in microseconds) besides reordering thresh");
1375 
1376 	/* Timer related controls */
1377 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1378 	    SYSCTL_CHILDREN(rack_sysctl_root),
1379 	    OID_AUTO,
1380 	    "timers",
1381 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1382 	    "Timer related controls");
1383 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1384 	    SYSCTL_CHILDREN(rack_timers),
1385 	    OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1386 	    &rack_ssthresh_rest_rto_rec, 0,
1387 	    "When doing recovery -> rto -> recovery do we reset SSthresh?");
1388 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1389 	    SYSCTL_CHILDREN(rack_timers),
1390 	    OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1391 	    &rack_rxt_scoreboard_clear_thresh, 2,
1392 	    "How many RTO's are allowed before we clear the scoreboard");
1393 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1394 	    SYSCTL_CHILDREN(rack_timers),
1395 	    OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1396 	    &rack_honors_hpts_min_to, 1,
1397 	    "Do rack pacing timers honor hpts min timeout");
1398 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1399 	    SYSCTL_CHILDREN(rack_timers),
1400 	    OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1401 	    &rack_max_reduce, 10,
1402 	    "Max percentage we will reduce slot by for pacing when we are behind");
1403 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1404 	    SYSCTL_CHILDREN(rack_timers),
1405 	    OID_AUTO, "persmin", CTLFLAG_RW,
1406 	    &rack_persist_min, 250000,
1407 	    "What is the minimum time in microseconds between persists");
1408 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(rack_timers),
1410 	    OID_AUTO, "persmax", CTLFLAG_RW,
1411 	    &rack_persist_max, 2000000,
1412 	    "What is the largest delay in microseconds between persists");
1413 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1414 	    SYSCTL_CHILDREN(rack_timers),
1415 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1416 	    &rack_delayed_ack_time, 40000,
1417 	    "Delayed ack time (40ms in microseconds)");
1418 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1419 	    SYSCTL_CHILDREN(rack_timers),
1420 	    OID_AUTO, "minrto", CTLFLAG_RW,
1421 	    &rack_rto_min, 30000,
1422 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1423 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1424 	    SYSCTL_CHILDREN(rack_timers),
1425 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1426 	    &rack_rto_max, 4000000,
1427 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1428 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1429 	    SYSCTL_CHILDREN(rack_timers),
1430 	    OID_AUTO, "minto", CTLFLAG_RW,
1431 	    &rack_min_to, 1000,
1432 	    "Minimum rack timeout in microseconds");
1433 	/* Measure controls */
1434 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1435 	    SYSCTL_CHILDREN(rack_sysctl_root),
1436 	    OID_AUTO,
1437 	    "measure",
1438 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1439 	    "Measure related controls");
1440 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_measure),
1442 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1443 	    &rack_wma_divisor, 8,
1444 	    "When doing b/w calculation what is the  divisor for the WMA");
1445 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1446 	    SYSCTL_CHILDREN(rack_measure),
1447 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1448 	    &rack_cwnd_block_ends_measure, 0,
1449 	    "Does a cwnd just-return end the measurement window (app limited)");
1450 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1451 	    SYSCTL_CHILDREN(rack_measure),
1452 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1453 	    &rack_rwnd_block_ends_measure, 0,
1454 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1455 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1456 	    SYSCTL_CHILDREN(rack_measure),
1457 	    OID_AUTO, "min_target", CTLFLAG_RW,
1458 	    &rack_def_data_window, 20,
1459 	    "What is the minimum target window (in mss) for a GP measurements");
1460 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1461 	    SYSCTL_CHILDREN(rack_measure),
1462 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1463 	    &rack_goal_bdp, 2,
1464 	    "What is the goal BDP to measure");
1465 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1466 	    SYSCTL_CHILDREN(rack_measure),
1467 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1468 	    &rack_min_srtts, 1,
1469 	    "What is the goal BDP to measure");
1470 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1471 	    SYSCTL_CHILDREN(rack_measure),
1472 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1473 	    &rack_min_measure_usec, 0,
1474 	    "What is the Minimum time time for a measurement if 0, this is off");
1475 	/* Features */
1476 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1477 	    SYSCTL_CHILDREN(rack_sysctl_root),
1478 	    OID_AUTO,
1479 	    "features",
1480 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1481 	    "Feature controls");
1482 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1483 	    SYSCTL_CHILDREN(rack_features),
1484 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1485 	    &rack_hybrid_allow_set_maxseg, 0,
1486 	    "Should hybrid pacing allow the setmss command");
1487 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1488 	    SYSCTL_CHILDREN(rack_features),
1489 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1490 	    &rack_use_cmp_acks, 1,
1491 	    "Should RACK have LRO send compressed acks");
1492 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1493 	    SYSCTL_CHILDREN(rack_features),
1494 	    OID_AUTO, "fsb", CTLFLAG_RW,
1495 	    &rack_use_fsb, 1,
1496 	    "Should RACK use the fast send block?");
1497 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1498 	    SYSCTL_CHILDREN(rack_features),
1499 	    OID_AUTO, "rfo", CTLFLAG_RW,
1500 	    &rack_use_rfo, 1,
1501 	    "Should RACK use rack_fast_output()?");
1502 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1503 	    SYSCTL_CHILDREN(rack_features),
1504 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1505 	    &rack_use_rsm_rfo, 1,
1506 	    "Should RACK use rack_fast_rsm_output()?");
1507 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(rack_features),
1509 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1510 	    &rack_enable_mqueue_for_nonpaced, 0,
1511 	    "Should RACK use mbuf queuing for non-paced connections");
1512 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1513 	    SYSCTL_CHILDREN(rack_features),
1514 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1515 	    &rack_do_hystart, 0,
1516 	    "Should RACK enable HyStart++ on connections?");
1517 	/* Misc rack controls */
1518 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1519 	    SYSCTL_CHILDREN(rack_sysctl_root),
1520 	    OID_AUTO,
1521 	    "misc",
1522 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1523 	    "Misc related controls");
1524 #ifdef TCP_ACCOUNTING
1525 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1526 	    SYSCTL_CHILDREN(rack_misc),
1527 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1528 	    &rack_tcp_accounting, 0,
1529 	    "Should we turn on TCP accounting for all rack sessions?");
1530 #endif
1531 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(rack_misc),
1533 	    OID_AUTO, "dnd", CTLFLAG_RW,
1534 	    &rack_dnd_default, 0,
1535 	    "Do not disturb default for rack_rrr = 3");
1536 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1537 	    SYSCTL_CHILDREN(rack_misc),
1538 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1539 	    &sad_seg_size_per, 800,
1540 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1541 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1542 	    SYSCTL_CHILDREN(rack_misc),
1543 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1544 	    &rack_rxt_controls, 0,
1545 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1546 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1547 	    SYSCTL_CHILDREN(rack_misc),
1548 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1549 	    &rack_hibeta_setting, 0,
1550 	    "Do we ue a high beta (80 instead of 50)?");
1551 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1552 	    SYSCTL_CHILDREN(rack_misc),
1553 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1554 	    &rack_apply_rtt_with_reduced_conf, 0,
1555 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1556 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1557 	    SYSCTL_CHILDREN(rack_misc),
1558 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1559 	    &rack_dsack_std_based, 3,
1560 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1561 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_misc),
1563 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1564 	    &rack_prr_addbackmax, 2,
1565 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1566 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_misc),
1568 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1569 	    &rack_stats_gets_ms_rtt, 1,
1570 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1571 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_misc),
1573 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1574 	    &rack_client_low_buf, 0,
1575 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1576 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_misc),
1578 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1579 	    &rack_def_profile, 0,
1580 	    "Should RACK use a default profile (0=no, num == profile num)?");
1581 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_misc),
1583 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1584 	    &rack_enable_shared_cwnd, 1,
1585 	    "Should RACK try to use the shared cwnd on connections where allowed");
1586 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_misc),
1588 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1589 	    &rack_limits_scwnd, 1,
1590 	    "Should RACK place low end time limits on the shared cwnd feature");
1591 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_misc),
1593 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1594 	    &rack_disable_prr, 0,
1595 	    "Should RACK not use prr and only pace (must have pacing on)");
1596 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_misc),
1598 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1599 	    &rack_verbose_logging, 0,
1600 	    "Should RACK black box logging be verbose");
1601 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1602 	    SYSCTL_CHILDREN(rack_misc),
1603 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1604 	    &rack_ignore_data_after_close, 1,
1605 	    "Do we hold off sending a RST until all pending data is ack'd");
1606 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1607 	    SYSCTL_CHILDREN(rack_misc),
1608 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1609 	    &rack_sack_not_required, 1,
1610 	    "Do we allow rack to run on connections not supporting SACK");
1611 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1612 	    SYSCTL_CHILDREN(rack_misc),
1613 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1614 	    &rack_send_a_lot_in_prr, 1,
1615 	    "Send a lot in prr");
1616 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1617 	    SYSCTL_CHILDREN(rack_misc),
1618 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1619 	    &rack_autosndbuf_inc, 20,
1620 	    "What percentage should rack scale up its snd buffer by?");
1621 
1622 
1623 	/* Sack Attacker detection stuff */
1624 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1625 	    SYSCTL_CHILDREN(rack_attack),
1626 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1627 	    &rack_merge_out_sacks_on_attack, 0,
1628 	    "Do we merge the sendmap when we decide we are being attacked?");
1629 
1630 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_attack),
1632 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1633 	    &rack_highest_sack_thresh_seen, 0,
1634 	    "Highest sack to ack ratio seen");
1635 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(rack_attack),
1637 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1638 	    &rack_highest_move_thresh_seen, 0,
1639 	    "Highest move to non-move ratio seen");
1640 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1641 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1642 	    SYSCTL_CHILDREN(rack_attack),
1643 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1644 	    &rack_ack_total,
1645 	    "Total number of Ack's");
1646 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1647 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1648 	    SYSCTL_CHILDREN(rack_attack),
1649 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1650 	    &rack_express_sack,
1651 	    "Total expresss number of Sack's");
1652 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1653 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1654 	    SYSCTL_CHILDREN(rack_attack),
1655 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1656 	    &rack_sack_total,
1657 	    "Total number of SACKs");
1658 	rack_move_none = counter_u64_alloc(M_WAITOK);
1659 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1660 	    SYSCTL_CHILDREN(rack_attack),
1661 	    OID_AUTO, "move_none", CTLFLAG_RD,
1662 	    &rack_move_none,
1663 	    "Total number of SACK index reuse of positions under threshold");
1664 	rack_move_some = counter_u64_alloc(M_WAITOK);
1665 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_attack),
1667 	    OID_AUTO, "move_some", CTLFLAG_RD,
1668 	    &rack_move_some,
1669 	    "Total number of SACK index reuse of positions over threshold");
1670 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1671 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1672 	    SYSCTL_CHILDREN(rack_attack),
1673 	    OID_AUTO, "attacks", CTLFLAG_RD,
1674 	    &rack_sack_attacks_detected,
1675 	    "Total number of SACK attackers that had sack disabled");
1676 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1677 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1678 	    SYSCTL_CHILDREN(rack_attack),
1679 	    OID_AUTO, "reversed", CTLFLAG_RD,
1680 	    &rack_sack_attacks_reversed,
1681 	    "Total number of SACK attackers that were later determined false positive");
1682 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1683 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1684 	    SYSCTL_CHILDREN(rack_attack),
1685 	    OID_AUTO, "suspect", CTLFLAG_RD,
1686 	    &rack_sack_attacks_suspect,
1687 	    "Total number of SACKs that triggered early detection");
1688 
1689 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1690 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1691 	    SYSCTL_CHILDREN(rack_attack),
1692 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1693 	    &rack_sack_used_next_merge,
1694 	    "Total number of times we used the next merge");
1695 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1696 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1697 	    SYSCTL_CHILDREN(rack_attack),
1698 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1699 	    &rack_sack_used_prev_merge,
1700 	    "Total number of times we used the prev merge");
1701 	/* Counters */
1702 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1703 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1704 	    SYSCTL_CHILDREN(rack_counters),
1705 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1706 	    &rack_total_bytes,
1707 	    "Total number of bytes sent");
1708 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1709 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1710 	    SYSCTL_CHILDREN(rack_counters),
1711 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1712 	    &rack_fto_send, "Total number of rack_fast_output sends");
1713 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1714 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1715 	    SYSCTL_CHILDREN(rack_counters),
1716 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1717 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1718 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1719 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1720 	    SYSCTL_CHILDREN(rack_counters),
1721 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1722 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1723 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1724 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1725 	    SYSCTL_CHILDREN(rack_counters),
1726 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1727 	    &rack_non_fto_send, "Total number of rack_output first sends");
1728 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1729 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1730 	    SYSCTL_CHILDREN(rack_counters),
1731 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1732 	    &rack_extended_rfo, "Total number of times we extended rfo");
1733 
1734 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1735 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1736 	    SYSCTL_CHILDREN(rack_counters),
1737 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1738 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1739 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1740 
1741 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1742 	    SYSCTL_CHILDREN(rack_counters),
1743 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1744 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1745 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1746 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1747 	    SYSCTL_CHILDREN(rack_counters),
1748 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1749 	    &rack_tlp_tot,
1750 	    "Total number of tail loss probe expirations");
1751 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1752 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1753 	    SYSCTL_CHILDREN(rack_counters),
1754 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1755 	    &rack_tlp_newdata,
1756 	    "Total number of tail loss probe sending new data");
1757 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1758 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1759 	    SYSCTL_CHILDREN(rack_counters),
1760 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1761 	    &rack_tlp_retran,
1762 	    "Total number of tail loss probe sending retransmitted data");
1763 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1764 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1765 	    SYSCTL_CHILDREN(rack_counters),
1766 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1767 	    &rack_tlp_retran_bytes,
1768 	    "Total bytes of tail loss probe sending retransmitted data");
1769 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1770 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1771 	    SYSCTL_CHILDREN(rack_counters),
1772 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1773 	    &rack_to_tot,
1774 	    "Total number of times the rack to expired");
1775 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1776 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1777 	    SYSCTL_CHILDREN(rack_counters),
1778 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1779 	    &rack_saw_enobuf,
1780 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1781 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1782 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1783 	    SYSCTL_CHILDREN(rack_counters),
1784 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1785 	    &rack_saw_enobuf_hw,
1786 	    "Total number of times a send returned enobuf for hdwr paced connections");
1787 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1788 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1789 	    SYSCTL_CHILDREN(rack_counters),
1790 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1791 	    &rack_saw_enetunreach,
1792 	    "Total number of times a send received a enetunreachable");
1793 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1794 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1795 	    SYSCTL_CHILDREN(rack_counters),
1796 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1797 	    &rack_hot_alloc,
1798 	    "Total allocations from the top of our list");
1799 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1800 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1801 	    SYSCTL_CHILDREN(rack_counters),
1802 	    OID_AUTO, "allocs", CTLFLAG_RD,
1803 	    &rack_to_alloc,
1804 	    "Total allocations of tracking structures");
1805 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1806 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1807 	    SYSCTL_CHILDREN(rack_counters),
1808 	    OID_AUTO, "allochard", CTLFLAG_RD,
1809 	    &rack_to_alloc_hard,
1810 	    "Total allocations done with sleeping the hard way");
1811 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1812 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1813 	    SYSCTL_CHILDREN(rack_counters),
1814 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1815 	    &rack_to_alloc_emerg,
1816 	    "Total allocations done from emergency cache");
1817 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1818 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1819 	    SYSCTL_CHILDREN(rack_counters),
1820 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1821 	    &rack_to_alloc_limited,
1822 	    "Total allocations dropped due to limit");
1823 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1824 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1825 	    SYSCTL_CHILDREN(rack_counters),
1826 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1827 	    &rack_alloc_limited_conns,
1828 	    "Connections with allocations dropped due to limit");
1829 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1831 	    SYSCTL_CHILDREN(rack_counters),
1832 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1833 	    &rack_split_limited,
1834 	    "Split allocations dropped due to limit");
1835 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1836 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1837 	    SYSCTL_CHILDREN(rack_counters),
1838 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1839 	    &rack_rxt_clamps_cwnd,
1840 	    "Number of times that excessive rxt clamped the cwnd down");
1841 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1842 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1843 	    SYSCTL_CHILDREN(rack_counters),
1844 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1845 	    &rack_rxt_clamps_cwnd_uniq,
1846 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1847 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1848 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1849 	    SYSCTL_CHILDREN(rack_counters),
1850 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1851 	    &rack_persists_sends,
1852 	    "Number of times we sent a persist probe");
1853 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1854 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1855 	    SYSCTL_CHILDREN(rack_counters),
1856 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1857 	    &rack_persists_acks,
1858 	    "Number of times a persist probe was acked");
1859 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1860 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1861 	    SYSCTL_CHILDREN(rack_counters),
1862 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1863 	    &rack_persists_loss,
1864 	    "Number of times we detected a lost persist probe (no ack)");
1865 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1866 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1867 	    SYSCTL_CHILDREN(rack_counters),
1868 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1869 	    &rack_persists_lost_ends,
1870 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1871 #ifdef INVARIANTS
1872 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1873 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1874 	    SYSCTL_CHILDREN(rack_counters),
1875 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1876 	    &rack_adjust_map_bw,
1877 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1878 #endif
1879 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1880 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1881 	    SYSCTL_CHILDREN(rack_counters),
1882 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1883 	    &rack_multi_single_eq,
1884 	    "Number of compressed acks total represented");
1885 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1886 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1887 	    SYSCTL_CHILDREN(rack_counters),
1888 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1889 	    &rack_proc_non_comp_ack,
1890 	    "Number of non compresseds acks that we processed");
1891 
1892 
1893 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1894 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1895 	    SYSCTL_CHILDREN(rack_counters),
1896 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1897 	    &rack_sack_proc_all,
1898 	    "Total times we had to walk whole list for sack processing");
1899 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1900 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1901 	    SYSCTL_CHILDREN(rack_counters),
1902 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1903 	    &rack_sack_proc_restart,
1904 	    "Total times we had to walk whole list due to a restart");
1905 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1906 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1907 	    SYSCTL_CHILDREN(rack_counters),
1908 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1909 	    &rack_sack_proc_short,
1910 	    "Total times we took shortcut for sack processing");
1911 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1912 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1913 	    SYSCTL_CHILDREN(rack_attack),
1914 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1915 	    &rack_sack_skipped_acked,
1916 	    "Total number of times we skipped previously sacked");
1917 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1918 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1919 	    SYSCTL_CHILDREN(rack_attack),
1920 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1921 	    &rack_sack_splits,
1922 	    "Total number of times we did the old fashion tree split");
1923 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1924 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1925 	    SYSCTL_CHILDREN(rack_counters),
1926 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1927 	    &rack_input_idle_reduces,
1928 	    "Total number of idle reductions on input");
1929 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1930 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1931 	    SYSCTL_CHILDREN(rack_counters),
1932 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1933 	    &rack_collapsed_win_seen,
1934 	    "Total number of collapsed window events seen (where our window shrinks)");
1935 
1936 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1937 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1938 	    SYSCTL_CHILDREN(rack_counters),
1939 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1940 	    &rack_collapsed_win,
1941 	    "Total number of collapsed window events where we mark packets");
1942 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1943 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1944 	    SYSCTL_CHILDREN(rack_counters),
1945 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1946 	    &rack_collapsed_win_rxt,
1947 	    "Total number of packets that were retransmitted");
1948 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1949 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1950 	    SYSCTL_CHILDREN(rack_counters),
1951 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1952 	    &rack_collapsed_win_rxt_bytes,
1953 	    "Total number of bytes that were retransmitted");
1954 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1955 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1956 	    SYSCTL_CHILDREN(rack_counters),
1957 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1958 	    &rack_try_scwnd,
1959 	    "Total number of scwnd attempts");
1960 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1961 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1962 	    OID_AUTO, "outsize", CTLFLAG_RD,
1963 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1964 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1965 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1966 	    OID_AUTO, "opts", CTLFLAG_RD,
1967 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1968 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1969 	    SYSCTL_CHILDREN(rack_sysctl_root),
1970 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1971 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1972 }
1973 
1974 static uint32_t
1975 rc_init_window(struct tcp_rack *rack)
1976 {
1977 	return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1978 
1979 }
1980 
1981 static uint64_t
1982 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1983 {
1984 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1985 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1986 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1987 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1988 	else
1989 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1990 }
1991 
1992 static void
1993 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1994 	uint64_t data, uint8_t mod, uint16_t aux,
1995 	struct tcp_sendfile_track *cur, int line)
1996 {
1997 #ifdef TCP_REQUEST_TRK
1998 	int do_log = 0;
1999 
2000 	/*
2001 	 * The rate cap one is noisy and only should come out when normal BB logging
2002 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2003 	 * once per chunk and make up the BBpoint that can be turned on by the client.
2004 	 */
2005 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2006 		/*
2007 		 * The very noisy two need to only come out when
2008 		 * we have verbose logging on.
2009 		 */
2010 		if (rack_verbose_logging != 0)
2011 			do_log = tcp_bblogging_on(rack->rc_tp);
2012 		else
2013 			do_log = 0;
2014 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2015 		/*
2016 		 * All other less noisy logs here except the measure which
2017 		 * also needs to come out on the point and the log.
2018 		 */
2019 		do_log = tcp_bblogging_on(rack->rc_tp);
2020 	} else {
2021 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2022 	}
2023 
2024 	if (do_log) {
2025 		union tcp_log_stackspecific log;
2026 		struct timeval tv;
2027 		uint64_t lt_bw;
2028 
2029 		/* Convert our ms to a microsecond */
2030 		memset(&log, 0, sizeof(log));
2031 
2032 		log.u_bbr.cwnd_gain = line;
2033 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2034 		log.u_bbr.rttProp = tim;
2035 		log.u_bbr.bw_inuse = cbw;
2036 		log.u_bbr.delRate = rack_get_gp_est(rack);
2037 		lt_bw = rack_get_lt_bw(rack);
2038 		log.u_bbr.flex1 = seq;
2039 		log.u_bbr.pacing_gain = aux;
2040 		/* lt_bw = < flex3 | flex2 > */
2041 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2042 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2043 		/* Record the last obtained us rtt in inflight */
2044 		if (cur == NULL) {
2045 			/* Make sure we are looking at the right log if an overide comes in */
2046 			cur = rack->r_ctl.rc_last_sft;
2047 		}
2048 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2049 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2050 		else {
2051 			/* Use the last known rtt i.e. the rack-rtt */
2052 			log.u_bbr.inflight = rack->rc_rack_rtt;
2053 		}
2054 		if (cur != NULL) {
2055 			uint64_t off;
2056 
2057 			log.u_bbr.cur_del_rate = cur->deadline;
2058 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2059 				/* start = < lost | pkt_epoch > */
2060 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2061 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2062 				log.u_bbr.flex6 = cur->start_seq;
2063 				log.u_bbr.pkts_out = cur->end_seq;
2064 			} else {
2065 				/* start = < lost | pkt_epoch > */
2066 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2067 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2068 				/* end = < pkts_out | flex6 > */
2069 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2070 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2071 			}
2072 			/* first_send = <lt_epoch | epoch> */
2073 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2074 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2075 			/* localtime = <delivered | applimited>*/
2076 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2077 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2078 #ifdef TCP_REQUEST_TRK
2079 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2080 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2081 #endif
2082 			log.u_bbr.inhpts = 1;
2083 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2084 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2085 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2086 		} else {
2087 			log.u_bbr.flex7 = 0xffff;
2088 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2089 		}
2090 		/*
2091 		 * Compose bbr_state to be a bit wise 0000ADHF
2092 		 * where A is the always_pace flag
2093 		 * where D is the dgp_on flag
2094 		 * where H is the hybrid_mode on flag
2095 		 * where F is the use_fixed_rate flag.
2096 		 */
2097 		log.u_bbr.bbr_state = rack->rc_always_pace;
2098 		log.u_bbr.bbr_state <<= 1;
2099 		log.u_bbr.bbr_state |= rack->dgp_on;
2100 		log.u_bbr.bbr_state <<= 1;
2101 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2102 		log.u_bbr.bbr_state <<= 1;
2103 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2104 		log.u_bbr.flex8 = mod;
2105 		tcp_log_event(rack->rc_tp, NULL,
2106 		    &rack->rc_inp->inp_socket->so_rcv,
2107 		    &rack->rc_inp->inp_socket->so_snd,
2108 		    TCP_HYBRID_PACING_LOG, 0,
2109 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2110 
2111 	}
2112 #endif
2113 }
2114 
2115 #ifdef TCP_REQUEST_TRK
2116 static void
2117 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2118 {
2119 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2120 		union tcp_log_stackspecific log;
2121 		struct timeval tv;
2122 		uint64_t off;
2123 
2124 		/* Convert our ms to a microsecond */
2125 		memset(&log, 0, sizeof(log));
2126 
2127 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2128 		log.u_bbr.delRate = cur->sent_at_fs;
2129 
2130 		if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2131 			/*
2132 			 * We did not get a new Rules Applied to set so
2133 			 * no overlapping send occured, this means the
2134 			 * current byte counts are correct.
2135 			 */
2136 			log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2137 			log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2138 		} else {
2139 			/*
2140 			 * Overlapping send case, we switched to a new
2141 			 * send and did a rules applied.
2142 			 */
2143 			log.u_bbr.cur_del_rate = cur->sent_at_ls;
2144 			log.u_bbr.rttProp = cur->rxt_at_ls;
2145 		}
2146 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2147 		log.u_bbr.cwnd_gain = line;
2148 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2149 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2150 		/* start = < flex1 | flex2 > */
2151 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2152 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2153 		/* end = < flex3 | flex4 > */
2154 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2155 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2156 
2157 		/* localtime = <delivered | applimited>*/
2158 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2159 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2160 		/* client timestamp = <lt_epoch | epoch>*/
2161 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2162 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2163 		/* now set all the flags in */
2164 		log.u_bbr.pkts_out = cur->hybrid_flags;
2165 		log.u_bbr.lost = cur->playout_ms;
2166 		log.u_bbr.flex6 = cur->flags;
2167 		/*
2168 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2169 		 * where a false retransmit occurred so first_send  <-> lastsend may
2170 		 * include longer time then it actually took if we have a false rxt.
2171 		 */
2172 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2173 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2174 		/*
2175 		 * Compose bbr_state to be a bit wise 0000ADHF
2176 		 * where A is the always_pace flag
2177 		 * where D is the dgp_on flag
2178 		 * where H is the hybrid_mode on flag
2179 		 * where F is the use_fixed_rate flag.
2180 		 */
2181 		log.u_bbr.bbr_state = rack->rc_always_pace;
2182 		log.u_bbr.bbr_state <<= 1;
2183 		log.u_bbr.bbr_state |= rack->dgp_on;
2184 		log.u_bbr.bbr_state <<= 1;
2185 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2186 		log.u_bbr.bbr_state <<= 1;
2187 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2188 
2189 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2190 		tcp_log_event(rack->rc_tp, NULL,
2191 		    &rack->rc_inp->inp_socket->so_rcv,
2192 		    &rack->rc_inp->inp_socket->so_snd,
2193 		    TCP_HYBRID_PACING_LOG, 0,
2194 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2195 	}
2196 }
2197 #endif
2198 
2199 static inline uint64_t
2200 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2201 {
2202 	uint64_t ret_bw, ether;
2203 	uint64_t u_segsiz;
2204 
2205 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2206 	if (rack->r_is_v6){
2207 #ifdef INET6
2208 		ether += sizeof(struct ip6_hdr);
2209 #endif
2210 		ether += 14;	/* eheader size 6+6+2 */
2211 	} else {
2212 #ifdef INET
2213 		ether += sizeof(struct ip);
2214 #endif
2215 		ether += 14;	/* eheader size 6+6+2 */
2216 	}
2217 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2218 	ret_bw = bw;
2219 	ret_bw *= ether;
2220 	ret_bw /= u_segsiz;
2221 	return (ret_bw);
2222 }
2223 
2224 static void
2225 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2226 {
2227 #ifdef TCP_REQUEST_TRK
2228 	struct timeval tv;
2229 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2230 #endif
2231 
2232 	if (rack->r_ctl.bw_rate_cap == 0)
2233 		return;
2234 #ifdef TCP_REQUEST_TRK
2235 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2236 	    (rack->r_ctl.rc_last_sft != NULL)) {
2237 		/*
2238 		 * We have a dynamic cap. The original target
2239 		 * is in bw_rate_cap, but we need to look at
2240 		 * how long it is until we hit the deadline.
2241 		 */
2242 		struct tcp_sendfile_track *ent;
2243 
2244       		ent = rack->r_ctl.rc_last_sft;
2245 		microuptime(&tv);
2246 		timenow = tcp_tv_to_lusec(&tv);
2247 		if (timenow >= ent->deadline) {
2248 			/* No time left we do DGP only */
2249 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2250 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2251 			rack->r_ctl.bw_rate_cap = 0;
2252 			return;
2253 		}
2254 		/* We have the time */
2255 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2256 		if (timeleft < HPTS_MSEC_IN_SEC) {
2257 			/* If there is less than a ms left just use DGPs rate */
2258 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2259 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2260 			rack->r_ctl.bw_rate_cap = 0;
2261 			return;
2262 		}
2263 		/*
2264 		 * Now lets find the amount of data left to send.
2265 		 *
2266 		 * Now ideally we want to use the end_seq to figure out how much more
2267 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2268 		 */
2269 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2270 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2271 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2272 			else {
2273 				/* TSNH, we should catch it at the send */
2274 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2275 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2276 				rack->r_ctl.bw_rate_cap = 0;
2277 				return;
2278 			}
2279 		} else {
2280 			/*
2281 			 * The hard way, figure out how much is gone and then
2282 			 * take that away from the total the client asked for
2283 			 * (thats off by tls overhead if this is tls).
2284 			 */
2285 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2286 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2287 			else
2288 				lengone = 0;
2289 			if (lengone < (ent->end - ent->start))
2290 				lenleft = (ent->end - ent->start) - lengone;
2291 			else {
2292 				/* TSNH, we should catch it at the send */
2293 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2294 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2295 				rack->r_ctl.bw_rate_cap = 0;
2296 				return;
2297 			}
2298 		}
2299 		if (lenleft == 0) {
2300 			/* We have it all sent */
2301 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2302 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2303 			if (rack->r_ctl.bw_rate_cap)
2304 				goto normal_ratecap;
2305 			else
2306 				return;
2307 		}
2308 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2309 		calcbw /= timeleft;
2310 		/* Now we must compensate for IP/TCP overhead */
2311 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2312 		/* Update the bit rate cap */
2313 		rack->r_ctl.bw_rate_cap = calcbw;
2314 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2315 		    (rack_hybrid_allow_set_maxseg == 1) &&
2316 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2317 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2318 			uint32_t orig_max;
2319 
2320 			orig_max = rack->r_ctl.rc_pace_max_segs;
2321 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2322 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2323 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2324 		}
2325 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2326 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2327 		if ((calcbw > 0) && (*bw > calcbw)) {
2328 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2329 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2330 			*capped = 1;
2331 			*bw = calcbw;
2332 		}
2333 		return;
2334 	}
2335 normal_ratecap:
2336 #endif
2337 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2338 #ifdef TCP_REQUEST_TRK
2339 		if (rack->rc_hybrid_mode &&
2340 		    rack->rc_catch_up &&
2341 		    (rack->r_ctl.rc_last_sft != NULL) &&
2342 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2343 		    (rack_hybrid_allow_set_maxseg == 1) &&
2344 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2345 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2346 			uint32_t orig_max;
2347 
2348 			orig_max = rack->r_ctl.rc_pace_max_segs;
2349 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2350 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2351 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2352 		}
2353 #endif
2354 		*capped = 1;
2355 		*bw = rack->r_ctl.bw_rate_cap;
2356 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2357 				   *bw, 0, 0,
2358 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2359 	}
2360 }
2361 
2362 static uint64_t
2363 rack_get_gp_est(struct tcp_rack *rack)
2364 {
2365 	uint64_t bw, lt_bw, ret_bw;
2366 
2367 	if (rack->rc_gp_filled == 0) {
2368 		/*
2369 		 * We have yet no b/w measurement,
2370 		 * if we have a user set initial bw
2371 		 * return it. If we don't have that and
2372 		 * we have an srtt, use the tcp IW (10) to
2373 		 * calculate a fictional b/w over the SRTT
2374 		 * which is more or less a guess. Note
2375 		 * we don't use our IW from rack on purpose
2376 		 * so if we have like IW=30, we are not
2377 		 * calculating a "huge" b/w.
2378 		 */
2379 		uint64_t srtt;
2380 
2381 		if (rack->dis_lt_bw == 1)
2382 			lt_bw = 0;
2383 		else
2384 			lt_bw = rack_get_lt_bw(rack);
2385 		if (lt_bw) {
2386 			/*
2387 			 * No goodput bw but a long-term b/w does exist
2388 			 * lets use that.
2389 			 */
2390 			ret_bw = lt_bw;
2391 			goto compensate;
2392 		}
2393 		if (rack->r_ctl.init_rate)
2394 			return (rack->r_ctl.init_rate);
2395 
2396 		/* Ok lets come up with the IW guess, if we have a srtt */
2397 		if (rack->rc_tp->t_srtt == 0) {
2398 			/*
2399 			 * Go with old pacing method
2400 			 * i.e. burst mitigation only.
2401 			 */
2402 			return (0);
2403 		}
2404 		/* Ok lets get the initial TCP win (not racks) */
2405 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2406 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2407 		bw *= (uint64_t)USECS_IN_SECOND;
2408 		bw /= srtt;
2409 		ret_bw = bw;
2410 		goto compensate;
2411 
2412 	}
2413 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2414 		/* Averaging is done, we can return the value */
2415 		bw = rack->r_ctl.gp_bw;
2416 	} else {
2417 		/* Still doing initial average must calculate */
2418 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2419 	}
2420 	if (rack->dis_lt_bw) {
2421 		/* We are not using lt-bw */
2422 		ret_bw = bw;
2423 		goto compensate;
2424 	}
2425 	lt_bw = rack_get_lt_bw(rack);
2426 	if (lt_bw == 0) {
2427 		/* If we don't have one then equate it to the gp_bw */
2428 		lt_bw = rack->r_ctl.gp_bw;
2429 	}
2430 	if (rack->use_lesser_lt_bw) {
2431 		if (lt_bw < bw)
2432 			ret_bw = lt_bw;
2433 		else
2434 			ret_bw = bw;
2435 	} else {
2436 		if (lt_bw > bw)
2437 			ret_bw = lt_bw;
2438 		else
2439 			ret_bw = bw;
2440 	}
2441 	/*
2442 	 * Now lets compensate based on the TCP/IP overhead. Our
2443 	 * Goodput estimate does not include this so we must pace out
2444 	 * a bit faster since our pacing calculations do. The pacing
2445 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2446 	 * we are using to do this, so we do that here in the opposite
2447 	 * direction as well. This means that if we are tunneled and the
2448 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2449 	 * compensated for in the pacing time the opposite way.
2450 	 */
2451 compensate:
2452 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2453 	return(ret_bw);
2454 }
2455 
2456 
2457 static uint64_t
2458 rack_get_bw(struct tcp_rack *rack)
2459 {
2460 	uint64_t bw;
2461 
2462 	if (rack->use_fixed_rate) {
2463 		/* Return the fixed pacing rate */
2464 		return (rack_get_fixed_pacing_bw(rack));
2465 	}
2466 	bw = rack_get_gp_est(rack);
2467 	return (bw);
2468 }
2469 
2470 static uint16_t
2471 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2472 {
2473 	if (rack->use_fixed_rate) {
2474 		return (100);
2475 	} else if (rack->in_probe_rtt && (rsm == NULL))
2476 		return (rack->r_ctl.rack_per_of_gp_probertt);
2477 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2478 		  rack->r_ctl.rack_per_of_gp_rec)) {
2479 		if (rsm) {
2480 			/* a retransmission always use the recovery rate */
2481 			return (rack->r_ctl.rack_per_of_gp_rec);
2482 		} else if (rack->rack_rec_nonrxt_use_cr) {
2483 			/* Directed to use the configured rate */
2484 			goto configured_rate;
2485 		} else if (rack->rack_no_prr &&
2486 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2487 			/* No PRR, lets just use the b/w estimate only */
2488 			return (100);
2489 		} else {
2490 			/*
2491 			 * Here we may have a non-retransmit but we
2492 			 * have no overrides, so just use the recovery
2493 			 * rate (prr is in effect).
2494 			 */
2495 			return (rack->r_ctl.rack_per_of_gp_rec);
2496 		}
2497 	}
2498 configured_rate:
2499 	/* For the configured rate we look at our cwnd vs the ssthresh */
2500 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2501 		return (rack->r_ctl.rack_per_of_gp_ss);
2502 	else
2503 		return (rack->r_ctl.rack_per_of_gp_ca);
2504 }
2505 
2506 static void
2507 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2508 {
2509 	/*
2510 	 * Types of logs (mod value)
2511 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2512 	 * 2 = a dsack round begins, persist is reset to 16.
2513 	 * 3 = a dsack round ends
2514 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2515 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2516 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2517 	 */
2518 	if (tcp_bblogging_on(rack->rc_tp)) {
2519 		union tcp_log_stackspecific log;
2520 		struct timeval tv;
2521 
2522 		memset(&log, 0, sizeof(log));
2523 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2524 		log.u_bbr.flex1 <<= 1;
2525 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2526 		log.u_bbr.flex1 <<= 1;
2527 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2528 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2529 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2530 		log.u_bbr.flex4 = flex4;
2531 		log.u_bbr.flex5 = flex5;
2532 		log.u_bbr.flex6 = flex6;
2533 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2534 		log.u_bbr.flex8 = mod;
2535 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2536 		log.u_bbr.epoch = rack->r_ctl.current_round;
2537 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
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 		    RACK_DSACK_HANDLING, 0,
2542 		    0, &log, false, &tv);
2543 	}
2544 }
2545 
2546 static void
2547 rack_log_hdwr_pacing(struct tcp_rack *rack,
2548 		     uint64_t rate, uint64_t hw_rate, int line,
2549 		     int error, uint16_t mod)
2550 {
2551 	if (tcp_bblogging_on(rack->rc_tp)) {
2552 		union tcp_log_stackspecific log;
2553 		struct timeval tv;
2554 		const struct ifnet *ifp;
2555 		uint64_t ifp64;
2556 
2557 		memset(&log, 0, sizeof(log));
2558 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2559 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2560 		if (rack->r_ctl.crte) {
2561 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2562 		} else if (rack->rc_inp->inp_route.ro_nh &&
2563 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2564 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2565 		} else
2566 			ifp = NULL;
2567 		if (ifp) {
2568 			ifp64 = (uintptr_t)ifp;
2569 			log.u_bbr.flex3 = ((ifp64  >> 32) & 0x00000000ffffffff);
2570 			log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2571 		}
2572 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2573 		log.u_bbr.bw_inuse = rate;
2574 		log.u_bbr.flex5 = line;
2575 		log.u_bbr.flex6 = error;
2576 		log.u_bbr.flex7 = mod;
2577 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2578 		log.u_bbr.flex8 = rack->use_fixed_rate;
2579 		log.u_bbr.flex8 <<= 1;
2580 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2581 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2582 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2583 		if (rack->r_ctl.crte)
2584 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2585 		else
2586 			log.u_bbr.cur_del_rate = 0;
2587 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2588 		log.u_bbr.epoch = rack->r_ctl.current_round;
2589 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2590 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2591 		    &rack->rc_inp->inp_socket->so_rcv,
2592 		    &rack->rc_inp->inp_socket->so_snd,
2593 		    BBR_LOG_HDWR_PACE, 0,
2594 		    0, &log, false, &tv);
2595 	}
2596 }
2597 
2598 static uint64_t
2599 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2600 {
2601 	/*
2602 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2603 	 */
2604 	uint64_t bw_est, high_rate;
2605 	uint64_t gain;
2606 
2607 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2608 	bw_est = bw * gain;
2609 	bw_est /= (uint64_t)100;
2610 	/* Never fall below the minimum (def 64kbps) */
2611 	if (bw_est < RACK_MIN_BW)
2612 		bw_est = RACK_MIN_BW;
2613 	if (rack->r_rack_hw_rate_caps) {
2614 		/* Rate caps are in place */
2615 		if (rack->r_ctl.crte != NULL) {
2616 			/* We have a hdwr rate already */
2617 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2618 			if (bw_est >= high_rate) {
2619 				/* We are capping bw at the highest rate table entry */
2620 				if (rack_hw_rate_cap_per &&
2621 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2622 					rack->r_rack_hw_rate_caps = 0;
2623 					goto done;
2624 				}
2625 				rack_log_hdwr_pacing(rack,
2626 						     bw_est, high_rate, __LINE__,
2627 						     0, 3);
2628 				bw_est = high_rate;
2629 				if (capped)
2630 					*capped = 1;
2631 			}
2632 		} else if ((rack->rack_hdrw_pacing == 0) &&
2633 			   (rack->rack_hdw_pace_ena) &&
2634 			   (rack->rack_attempt_hdwr_pace == 0) &&
2635 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2636 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2637 			/*
2638 			 * Special case, we have not yet attempted hardware
2639 			 * pacing, and yet we may, when we do, find out if we are
2640 			 * above the highest rate. We need to know the maxbw for the interface
2641 			 * in question (if it supports ratelimiting). We get back
2642 			 * a 0, if the interface is not found in the RL lists.
2643 			 */
2644 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2645 			if (high_rate) {
2646 				/* Yep, we have a rate is it above this rate? */
2647 				if (bw_est > high_rate) {
2648 					bw_est = high_rate;
2649 					if (capped)
2650 						*capped = 1;
2651 				}
2652 			}
2653 		}
2654 	}
2655 done:
2656 	return (bw_est);
2657 }
2658 
2659 static void
2660 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2661 {
2662 	if (tcp_bblogging_on(rack->rc_tp)) {
2663 		union tcp_log_stackspecific log;
2664 		struct timeval tv;
2665 
2666 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2667 			/*
2668 			 * We get 3 values currently for mod
2669 			 * 1 - We are retransmitting and this tells the reason.
2670 			 * 2 - We are clearing a dup-ack count.
2671 			 * 3 - We are incrementing a dup-ack count.
2672 			 *
2673 			 * The clear/increment are only logged
2674 			 * if you have BBverbose on.
2675 			 */
2676 			return;
2677 		}
2678 		memset(&log, 0, sizeof(log));
2679 		log.u_bbr.flex1 = tsused;
2680 		log.u_bbr.flex2 = thresh;
2681 		log.u_bbr.flex3 = rsm->r_flags;
2682 		log.u_bbr.flex4 = rsm->r_dupack;
2683 		log.u_bbr.flex5 = rsm->r_start;
2684 		log.u_bbr.flex6 = rsm->r_end;
2685 		log.u_bbr.flex8 = mod;
2686 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2687 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2688 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2689 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2690 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2691 		log.u_bbr.pacing_gain = rack->r_must_retran;
2692 		log.u_bbr.epoch = rack->r_ctl.current_round;
2693 		log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2694 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2695 		    &rack->rc_inp->inp_socket->so_rcv,
2696 		    &rack->rc_inp->inp_socket->so_snd,
2697 		    BBR_LOG_SETTINGS_CHG, 0,
2698 		    0, &log, false, &tv);
2699 	}
2700 }
2701 
2702 static void
2703 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2704 {
2705 	if (tcp_bblogging_on(rack->rc_tp)) {
2706 		union tcp_log_stackspecific log;
2707 		struct timeval tv;
2708 
2709 		memset(&log, 0, sizeof(log));
2710 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2711 		log.u_bbr.flex2 = to;
2712 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2713 		log.u_bbr.flex4 = slot;
2714 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2715 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2716 		log.u_bbr.flex7 = rack->rc_in_persist;
2717 		log.u_bbr.flex8 = which;
2718 		if (rack->rack_no_prr)
2719 			log.u_bbr.pkts_out = 0;
2720 		else
2721 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2722 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2723 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2724 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2725 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2726 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2727 		log.u_bbr.pacing_gain = rack->r_must_retran;
2728 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2729 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2730 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2731 		log.u_bbr.lost = rack_rto_min;
2732 		log.u_bbr.epoch = rack->r_ctl.roundends;
2733 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2734 		log.u_bbr.bw_inuse <<= 32;
2735 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2736 		log.u_bbr.applimited = rack->rc_tp->t_flags2;
2737 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2738 		    &rack->rc_inp->inp_socket->so_rcv,
2739 		    &rack->rc_inp->inp_socket->so_snd,
2740 		    BBR_LOG_TIMERSTAR, 0,
2741 		    0, &log, false, &tv);
2742 	}
2743 }
2744 
2745 static void
2746 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2747 {
2748 	if (tcp_bblogging_on(rack->rc_tp)) {
2749 		union tcp_log_stackspecific log;
2750 		struct timeval tv;
2751 
2752 		memset(&log, 0, sizeof(log));
2753 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2754 		log.u_bbr.flex8 = to_num;
2755 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2756 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2757 		if (rsm == NULL)
2758 			log.u_bbr.flex3 = 0;
2759 		else
2760 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2761 		if (rack->rack_no_prr)
2762 			log.u_bbr.flex5 = 0;
2763 		else
2764 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2765 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2766 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2767 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2768 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2769 		log.u_bbr.pacing_gain = rack->r_must_retran;
2770 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2771 		log.u_bbr.bw_inuse <<= 32;
2772 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2773 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2774 		    &rack->rc_inp->inp_socket->so_rcv,
2775 		    &rack->rc_inp->inp_socket->so_snd,
2776 		    BBR_LOG_RTO, 0,
2777 		    0, &log, false, &tv);
2778 	}
2779 }
2780 
2781 static void
2782 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2783 		 struct rack_sendmap *prev,
2784 		 struct rack_sendmap *rsm,
2785 		 struct rack_sendmap *next,
2786 		 int flag, uint32_t th_ack, int line)
2787 {
2788 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2789 		union tcp_log_stackspecific log;
2790 		struct timeval tv;
2791 
2792 		memset(&log, 0, sizeof(log));
2793 		log.u_bbr.flex8 = flag;
2794 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2795 		log.u_bbr.cur_del_rate = (uintptr_t)prev;
2796 		log.u_bbr.delRate = (uintptr_t)rsm;
2797 		log.u_bbr.rttProp = (uintptr_t)next;
2798 		log.u_bbr.flex7 = 0;
2799 		if (prev) {
2800 			log.u_bbr.flex1 = prev->r_start;
2801 			log.u_bbr.flex2 = prev->r_end;
2802 			log.u_bbr.flex7 |= 0x4;
2803 		}
2804 		if (rsm) {
2805 			log.u_bbr.flex3 = rsm->r_start;
2806 			log.u_bbr.flex4 = rsm->r_end;
2807 			log.u_bbr.flex7 |= 0x2;
2808 		}
2809 		if (next) {
2810 			log.u_bbr.flex5 = next->r_start;
2811 			log.u_bbr.flex6 = next->r_end;
2812 			log.u_bbr.flex7 |= 0x1;
2813 		}
2814 		log.u_bbr.applimited = line;
2815 		log.u_bbr.pkts_out = th_ack;
2816 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2817 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2818 		if (rack->rack_no_prr)
2819 			log.u_bbr.lost = 0;
2820 		else
2821 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2822 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2823 		log.u_bbr.bw_inuse <<= 32;
2824 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2825 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2826 		    &rack->rc_inp->inp_socket->so_rcv,
2827 		    &rack->rc_inp->inp_socket->so_snd,
2828 		    TCP_LOG_MAPCHG, 0,
2829 		    0, &log, false, &tv);
2830 	}
2831 }
2832 
2833 static void
2834 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2835 		 struct rack_sendmap *rsm, int conf)
2836 {
2837 	if (tcp_bblogging_on(tp)) {
2838 		union tcp_log_stackspecific log;
2839 		struct timeval tv;
2840 		memset(&log, 0, sizeof(log));
2841 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2842 		log.u_bbr.flex1 = t;
2843 		log.u_bbr.flex2 = len;
2844 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2845 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2846 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2847 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2848 		log.u_bbr.flex7 = conf;
2849 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2850 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2851 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2852 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2853 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2854 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2855 		if (rsm) {
2856 			log.u_bbr.pkt_epoch = rsm->r_start;
2857 			log.u_bbr.lost = rsm->r_end;
2858 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2859 			/* We loose any upper of the 24 bits */
2860 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2861 		} else {
2862 			/* Its a SYN */
2863 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2864 			log.u_bbr.lost = 0;
2865 			log.u_bbr.cwnd_gain = 0;
2866 			log.u_bbr.pacing_gain = 0;
2867 		}
2868 		/* Write out general bits of interest rrs here */
2869 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2870 		log.u_bbr.use_lt_bw <<= 1;
2871 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2872 		log.u_bbr.use_lt_bw <<= 1;
2873 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2874 		log.u_bbr.use_lt_bw <<= 1;
2875 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2876 		log.u_bbr.use_lt_bw <<= 1;
2877 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2878 		log.u_bbr.use_lt_bw <<= 1;
2879 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2880 		log.u_bbr.use_lt_bw <<= 1;
2881 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2882 		log.u_bbr.use_lt_bw <<= 1;
2883 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2884 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2885 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2886 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2887 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2888 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2889 		log.u_bbr.bw_inuse = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
2890 		log.u_bbr.bw_inuse <<= 32;
2891 		if (rsm)
2892 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2893 		TCP_LOG_EVENTP(tp, NULL,
2894 		    &rack->rc_inp->inp_socket->so_rcv,
2895 		    &rack->rc_inp->inp_socket->so_snd,
2896 		    BBR_LOG_BBRRTT, 0,
2897 		    0, &log, false, &tv);
2898 
2899 
2900 	}
2901 }
2902 
2903 static void
2904 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2905 {
2906 	/*
2907 	 * Log the rtt sample we are
2908 	 * applying to the srtt algorithm in
2909 	 * useconds.
2910 	 */
2911 	if (tcp_bblogging_on(rack->rc_tp)) {
2912 		union tcp_log_stackspecific log;
2913 		struct timeval tv;
2914 
2915 		/* Convert our ms to a microsecond */
2916 		memset(&log, 0, sizeof(log));
2917 		log.u_bbr.flex1 = rtt;
2918 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2919 		log.u_bbr.flex7 = 1;
2920 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2921 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2922 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2923 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2924 		log.u_bbr.pacing_gain = rack->r_must_retran;
2925 		/*
2926 		 * We capture in delRate the upper 32 bits as
2927 		 * the confidence level we had declared, and the
2928 		 * lower 32 bits as the actual RTT using the arrival
2929 		 * timestamp.
2930 		 */
2931 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2932 		log.u_bbr.delRate <<= 32;
2933 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2934 		/* Lets capture all the things that make up t_rtxcur */
2935 		log.u_bbr.applimited = rack_rto_min;
2936 		log.u_bbr.epoch = rack_rto_max;
2937 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2938 		log.u_bbr.lost = rack_rto_min;
2939 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2940 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2941 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2942 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2943 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2944 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2945 		    &rack->rc_inp->inp_socket->so_rcv,
2946 		    &rack->rc_inp->inp_socket->so_snd,
2947 		    TCP_LOG_RTT, 0,
2948 		    0, &log, false, &tv);
2949 	}
2950 }
2951 
2952 static void
2953 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2954 {
2955 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2956 		union tcp_log_stackspecific log;
2957 		struct timeval tv;
2958 
2959 		/* Convert our ms to a microsecond */
2960 		memset(&log, 0, sizeof(log));
2961 		log.u_bbr.flex1 = rtt;
2962 		log.u_bbr.flex2 = send_time;
2963 		log.u_bbr.flex3 = ack_time;
2964 		log.u_bbr.flex4 = where;
2965 		log.u_bbr.flex7 = 2;
2966 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2967 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2968 		log.u_bbr.bw_inuse <<= 32;
2969 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2970 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2971 		    &rack->rc_inp->inp_socket->so_rcv,
2972 		    &rack->rc_inp->inp_socket->so_snd,
2973 		    TCP_LOG_RTT, 0,
2974 		    0, &log, false, &tv);
2975 	}
2976 }
2977 
2978 
2979 static void
2980 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2981 {
2982 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2983 		union tcp_log_stackspecific log;
2984 		struct timeval tv;
2985 
2986 		/* Convert our ms to a microsecond */
2987 		memset(&log, 0, sizeof(log));
2988 		log.u_bbr.flex1 = idx;
2989 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2990 		log.u_bbr.flex3 = tsecho;
2991 		log.u_bbr.flex7 = 3;
2992 		log.u_bbr.rttProp = tsv;
2993 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2994 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2995 		log.u_bbr.bw_inuse <<= 32;
2996 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2997 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2998 		    &rack->rc_inp->inp_socket->so_rcv,
2999 		    &rack->rc_inp->inp_socket->so_snd,
3000 		    TCP_LOG_RTT, 0,
3001 		    0, &log, false, &tv);
3002 	}
3003 }
3004 
3005 
3006 static inline void
3007 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
3008 {
3009 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3010 		union tcp_log_stackspecific log;
3011 		struct timeval tv;
3012 
3013 		memset(&log, 0, sizeof(log));
3014 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3015 		log.u_bbr.flex1 = line;
3016 		log.u_bbr.flex2 = tick;
3017 		log.u_bbr.flex3 = tp->t_maxunacktime;
3018 		log.u_bbr.flex4 = tp->t_acktime;
3019 		log.u_bbr.flex8 = event;
3020 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3021 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3022 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3023 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3024 		log.u_bbr.pacing_gain = rack->r_must_retran;
3025 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3026 		log.u_bbr.bw_inuse <<= 32;
3027 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3028 		TCP_LOG_EVENTP(tp, NULL,
3029 		    &rack->rc_inp->inp_socket->so_rcv,
3030 		    &rack->rc_inp->inp_socket->so_snd,
3031 		    BBR_LOG_PROGRESS, 0,
3032 		    0, &log, false, &tv);
3033 	}
3034 }
3035 
3036 static void
3037 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3038 {
3039 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3040 		union tcp_log_stackspecific log;
3041 
3042 		memset(&log, 0, sizeof(log));
3043 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3044 		log.u_bbr.flex1 = slot;
3045 		if (rack->rack_no_prr)
3046 			log.u_bbr.flex2 = 0;
3047 		else
3048 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3049 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3050 		log.u_bbr.flex6 = line;
3051 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3052 		log.u_bbr.flex8 = rack->rc_in_persist;
3053 		log.u_bbr.timeStamp = cts;
3054 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3055 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3056 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3057 		log.u_bbr.pacing_gain = rack->r_must_retran;
3058 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3059 		    &rack->rc_inp->inp_socket->so_rcv,
3060 		    &rack->rc_inp->inp_socket->so_snd,
3061 		    BBR_LOG_BBRSND, 0,
3062 		    0, &log, false, tv);
3063 	}
3064 }
3065 
3066 static void
3067 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3068 {
3069 	if (tcp_bblogging_on(rack->rc_tp)) {
3070 		union tcp_log_stackspecific log;
3071 		struct timeval tv;
3072 
3073 		memset(&log, 0, sizeof(log));
3074 		log.u_bbr.flex1 = did_out;
3075 		log.u_bbr.flex2 = nxt_pkt;
3076 		log.u_bbr.flex3 = way_out;
3077 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3078 		if (rack->rack_no_prr)
3079 			log.u_bbr.flex5 = 0;
3080 		else
3081 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3082 		log.u_bbr.flex6 = nsegs;
3083 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3084 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3085 		log.u_bbr.flex7 <<= 1;
3086 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3087 		log.u_bbr.flex7 <<= 1;
3088 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3089 		log.u_bbr.flex8 = rack->rc_in_persist;
3090 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3091 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3092 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3093 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3094 		log.u_bbr.use_lt_bw <<= 1;
3095 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3096 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3097 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3098 		log.u_bbr.pacing_gain = rack->r_must_retran;
3099 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3100 		log.u_bbr.bw_inuse <<= 32;
3101 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3102 		log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3103 		log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3104 		log.u_bbr.lost = rack->rc_tp->t_srtt;
3105 		log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3106 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3107 		    &rack->rc_inp->inp_socket->so_rcv,
3108 		    &rack->rc_inp->inp_socket->so_snd,
3109 		    BBR_LOG_DOSEG_DONE, 0,
3110 		    0, &log, false, &tv);
3111 	}
3112 }
3113 
3114 static void
3115 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3116 {
3117 	if (tcp_bblogging_on(rack->rc_tp)) {
3118 		union tcp_log_stackspecific log;
3119 		struct timeval tv;
3120 
3121 		memset(&log, 0, sizeof(log));
3122 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3123 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3124 		log.u_bbr.flex4 = arg1;
3125 		log.u_bbr.flex5 = arg2;
3126 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3127 		log.u_bbr.flex6 = arg3;
3128 		log.u_bbr.flex8 = frm;
3129 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3130 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3131 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3132 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3133 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3134 		log.u_bbr.pacing_gain = rack->r_must_retran;
3135 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3136 		    &tptosocket(tp)->so_snd,
3137 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3138 	}
3139 }
3140 
3141 static void
3142 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3143 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3144 {
3145 	if (tcp_bblogging_on(rack->rc_tp)) {
3146 		union tcp_log_stackspecific log;
3147 		struct timeval tv;
3148 
3149 		memset(&log, 0, sizeof(log));
3150 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3151 		log.u_bbr.flex1 = slot;
3152 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3153 		log.u_bbr.flex4 = reason;
3154 		if (rack->rack_no_prr)
3155 			log.u_bbr.flex5 = 0;
3156 		else
3157 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3158 		log.u_bbr.flex7 = hpts_calling;
3159 		log.u_bbr.flex8 = rack->rc_in_persist;
3160 		log.u_bbr.lt_epoch = cwnd_to_use;
3161 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3162 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3163 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3164 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3165 		log.u_bbr.pacing_gain = rack->r_must_retran;
3166 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3167 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3168 		log.u_bbr.bw_inuse <<= 32;
3169 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3170 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3171 		    &rack->rc_inp->inp_socket->so_rcv,
3172 		    &rack->rc_inp->inp_socket->so_snd,
3173 		    BBR_LOG_JUSTRET, 0,
3174 		    tlen, &log, false, &tv);
3175 	}
3176 }
3177 
3178 static void
3179 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3180 		   struct timeval *tv, uint32_t flags_on_entry)
3181 {
3182 	if (tcp_bblogging_on(rack->rc_tp)) {
3183 		union tcp_log_stackspecific log;
3184 
3185 		memset(&log, 0, sizeof(log));
3186 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3187 		log.u_bbr.flex1 = line;
3188 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3189 		log.u_bbr.flex3 = flags_on_entry;
3190 		log.u_bbr.flex4 = us_cts;
3191 		if (rack->rack_no_prr)
3192 			log.u_bbr.flex5 = 0;
3193 		else
3194 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3195 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3196 		log.u_bbr.flex7 = hpts_removed;
3197 		log.u_bbr.flex8 = 1;
3198 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3199 		log.u_bbr.timeStamp = us_cts;
3200 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3201 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3202 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3203 		log.u_bbr.pacing_gain = rack->r_must_retran;
3204 		log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3205 		log.u_bbr.bw_inuse <<= 32;
3206 		log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3207 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3208 		    &rack->rc_inp->inp_socket->so_rcv,
3209 		    &rack->rc_inp->inp_socket->so_snd,
3210 		    BBR_LOG_TIMERCANC, 0,
3211 		    0, &log, false, tv);
3212 	}
3213 }
3214 
3215 static void
3216 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3217 			  uint32_t flex1, uint32_t flex2,
3218 			  uint32_t flex3, uint32_t flex4,
3219 			  uint32_t flex5, uint32_t flex6,
3220 			  uint16_t flex7, uint8_t mod)
3221 {
3222 	if (tcp_bblogging_on(rack->rc_tp)) {
3223 		union tcp_log_stackspecific log;
3224 		struct timeval tv;
3225 
3226 		if (mod == 1) {
3227 			/* No you can't use 1, its for the real to cancel */
3228 			return;
3229 		}
3230 		memset(&log, 0, sizeof(log));
3231 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3232 		log.u_bbr.flex1 = flex1;
3233 		log.u_bbr.flex2 = flex2;
3234 		log.u_bbr.flex3 = flex3;
3235 		log.u_bbr.flex4 = flex4;
3236 		log.u_bbr.flex5 = flex5;
3237 		log.u_bbr.flex6 = flex6;
3238 		log.u_bbr.flex7 = flex7;
3239 		log.u_bbr.flex8 = mod;
3240 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3241 		    &rack->rc_inp->inp_socket->so_rcv,
3242 		    &rack->rc_inp->inp_socket->so_snd,
3243 		    BBR_LOG_TIMERCANC, 0,
3244 		    0, &log, false, &tv);
3245 	}
3246 }
3247 
3248 static void
3249 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3250 {
3251 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3252 		union tcp_log_stackspecific log;
3253 		struct timeval tv;
3254 
3255 		memset(&log, 0, sizeof(log));
3256 		log.u_bbr.flex1 = timers;
3257 		log.u_bbr.flex2 = ret;
3258 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3259 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3260 		log.u_bbr.flex5 = cts;
3261 		if (rack->rack_no_prr)
3262 			log.u_bbr.flex6 = 0;
3263 		else
3264 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3265 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3266 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3267 		log.u_bbr.pacing_gain = rack->r_must_retran;
3268 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3269 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3270 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3271 		    &rack->rc_inp->inp_socket->so_rcv,
3272 		    &rack->rc_inp->inp_socket->so_snd,
3273 		    BBR_LOG_TO_PROCESS, 0,
3274 		    0, &log, false, &tv);
3275 	}
3276 }
3277 
3278 static void
3279 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3280 {
3281 	if (tcp_bblogging_on(rack->rc_tp)) {
3282 		union tcp_log_stackspecific log;
3283 		struct timeval tv;
3284 
3285 		memset(&log, 0, sizeof(log));
3286 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3287 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3288 		if (rack->rack_no_prr)
3289 			log.u_bbr.flex3 = 0;
3290 		else
3291 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3292 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3293 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3294 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3295 		log.u_bbr.flex7 = line;
3296 		log.u_bbr.flex8 = frm;
3297 		log.u_bbr.pkts_out = orig_cwnd;
3298 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3299 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3300 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3301 		log.u_bbr.use_lt_bw <<= 1;
3302 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3303 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3304 		    &rack->rc_inp->inp_socket->so_rcv,
3305 		    &rack->rc_inp->inp_socket->so_snd,
3306 		    BBR_LOG_BBRUPD, 0,
3307 		    0, &log, false, &tv);
3308 	}
3309 }
3310 
3311 static void
3312 rack_counter_destroy(void)
3313 {
3314 	counter_u64_free(rack_total_bytes);
3315 	counter_u64_free(rack_fto_send);
3316 	counter_u64_free(rack_fto_rsm_send);
3317 	counter_u64_free(rack_nfto_resend);
3318 	counter_u64_free(rack_hw_pace_init_fail);
3319 	counter_u64_free(rack_hw_pace_lost);
3320 	counter_u64_free(rack_non_fto_send);
3321 	counter_u64_free(rack_extended_rfo);
3322 	counter_u64_free(rack_ack_total);
3323 	counter_u64_free(rack_express_sack);
3324 	counter_u64_free(rack_sack_total);
3325 	counter_u64_free(rack_move_none);
3326 	counter_u64_free(rack_move_some);
3327 	counter_u64_free(rack_sack_attacks_detected);
3328 	counter_u64_free(rack_sack_attacks_reversed);
3329 	counter_u64_free(rack_sack_attacks_suspect);
3330 	counter_u64_free(rack_sack_used_next_merge);
3331 	counter_u64_free(rack_sack_used_prev_merge);
3332 	counter_u64_free(rack_tlp_tot);
3333 	counter_u64_free(rack_tlp_newdata);
3334 	counter_u64_free(rack_tlp_retran);
3335 	counter_u64_free(rack_tlp_retran_bytes);
3336 	counter_u64_free(rack_to_tot);
3337 	counter_u64_free(rack_saw_enobuf);
3338 	counter_u64_free(rack_saw_enobuf_hw);
3339 	counter_u64_free(rack_saw_enetunreach);
3340 	counter_u64_free(rack_hot_alloc);
3341 	counter_u64_free(rack_to_alloc);
3342 	counter_u64_free(rack_to_alloc_hard);
3343 	counter_u64_free(rack_to_alloc_emerg);
3344 	counter_u64_free(rack_to_alloc_limited);
3345 	counter_u64_free(rack_alloc_limited_conns);
3346 	counter_u64_free(rack_split_limited);
3347 	counter_u64_free(rack_multi_single_eq);
3348 	counter_u64_free(rack_rxt_clamps_cwnd);
3349 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3350 	counter_u64_free(rack_proc_non_comp_ack);
3351 	counter_u64_free(rack_sack_proc_all);
3352 	counter_u64_free(rack_sack_proc_restart);
3353 	counter_u64_free(rack_sack_proc_short);
3354 	counter_u64_free(rack_sack_skipped_acked);
3355 	counter_u64_free(rack_sack_splits);
3356 	counter_u64_free(rack_input_idle_reduces);
3357 	counter_u64_free(rack_collapsed_win);
3358 	counter_u64_free(rack_collapsed_win_rxt);
3359 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3360 	counter_u64_free(rack_collapsed_win_seen);
3361 	counter_u64_free(rack_try_scwnd);
3362 	counter_u64_free(rack_persists_sends);
3363 	counter_u64_free(rack_persists_acks);
3364 	counter_u64_free(rack_persists_loss);
3365 	counter_u64_free(rack_persists_lost_ends);
3366 #ifdef INVARIANTS
3367 	counter_u64_free(rack_adjust_map_bw);
3368 #endif
3369 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3370 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3371 }
3372 
3373 static struct rack_sendmap *
3374 rack_alloc(struct tcp_rack *rack)
3375 {
3376 	struct rack_sendmap *rsm;
3377 
3378 	/*
3379 	 * First get the top of the list it in
3380 	 * theory is the "hottest" rsm we have,
3381 	 * possibly just freed by ack processing.
3382 	 */
3383 	if (rack->rc_free_cnt > rack_free_cache) {
3384 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3385 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3386 		counter_u64_add(rack_hot_alloc, 1);
3387 		rack->rc_free_cnt--;
3388 		return (rsm);
3389 	}
3390 	/*
3391 	 * Once we get under our free cache we probably
3392 	 * no longer have a "hot" one available. Lets
3393 	 * get one from UMA.
3394 	 */
3395 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3396 	if (rsm) {
3397 		rack->r_ctl.rc_num_maps_alloced++;
3398 		counter_u64_add(rack_to_alloc, 1);
3399 		return (rsm);
3400 	}
3401 	/*
3402 	 * Dig in to our aux rsm's (the last two) since
3403 	 * UMA failed to get us one.
3404 	 */
3405 	if (rack->rc_free_cnt) {
3406 		counter_u64_add(rack_to_alloc_emerg, 1);
3407 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3408 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3409 		rack->rc_free_cnt--;
3410 		return (rsm);
3411 	}
3412 	return (NULL);
3413 }
3414 
3415 static struct rack_sendmap *
3416 rack_alloc_full_limit(struct tcp_rack *rack)
3417 {
3418 	if ((V_tcp_map_entries_limit > 0) &&
3419 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3420 		counter_u64_add(rack_to_alloc_limited, 1);
3421 		if (!rack->alloc_limit_reported) {
3422 			rack->alloc_limit_reported = 1;
3423 			counter_u64_add(rack_alloc_limited_conns, 1);
3424 		}
3425 		return (NULL);
3426 	}
3427 	return (rack_alloc(rack));
3428 }
3429 
3430 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3431 static struct rack_sendmap *
3432 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3433 {
3434 	struct rack_sendmap *rsm;
3435 
3436 	if (limit_type) {
3437 		/* currently there is only one limit type */
3438 		if (rack->r_ctl.rc_split_limit > 0 &&
3439 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3440 			counter_u64_add(rack_split_limited, 1);
3441 			if (!rack->alloc_limit_reported) {
3442 				rack->alloc_limit_reported = 1;
3443 				counter_u64_add(rack_alloc_limited_conns, 1);
3444 			}
3445 			return (NULL);
3446 		}
3447 	}
3448 
3449 	/* allocate and mark in the limit type, if set */
3450 	rsm = rack_alloc(rack);
3451 	if (rsm != NULL && limit_type) {
3452 		rsm->r_limit_type = limit_type;
3453 		rack->r_ctl.rc_num_split_allocs++;
3454 	}
3455 	return (rsm);
3456 }
3457 
3458 static void
3459 rack_free_trim(struct tcp_rack *rack)
3460 {
3461 	struct rack_sendmap *rsm;
3462 
3463 	/*
3464 	 * Free up all the tail entries until
3465 	 * we get our list down to the limit.
3466 	 */
3467 	while (rack->rc_free_cnt > rack_free_cache) {
3468 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3469 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3470 		rack->rc_free_cnt--;
3471 		rack->r_ctl.rc_num_maps_alloced--;
3472 		uma_zfree(rack_zone, rsm);
3473 	}
3474 }
3475 
3476 static void
3477 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3478 {
3479 	if (rsm->r_flags & RACK_APP_LIMITED) {
3480 		KASSERT((rack->r_ctl.rc_app_limited_cnt > 0),
3481 		    ("app_cnt %u, rsm %p", rack->r_ctl.rc_app_limited_cnt, rsm));
3482 		rack->r_ctl.rc_app_limited_cnt--;
3483 	}
3484 	if (rsm->r_limit_type) {
3485 		/* currently there is only one limit type */
3486 		rack->r_ctl.rc_num_split_allocs--;
3487 	}
3488 	if (rsm == rack->r_ctl.rc_first_appl) {
3489 		rack->r_ctl.cleared_app_ack_seq = rsm->r_end;
3490 		rack->r_ctl.cleared_app_ack = 1;
3491 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3492 			rack->r_ctl.rc_first_appl = NULL;
3493 		else
3494 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3495 	}
3496 	if (rsm == rack->r_ctl.rc_resend)
3497 		rack->r_ctl.rc_resend = NULL;
3498 	if (rsm == rack->r_ctl.rc_end_appl)
3499 		rack->r_ctl.rc_end_appl = NULL;
3500 	if (rack->r_ctl.rc_tlpsend == rsm)
3501 		rack->r_ctl.rc_tlpsend = NULL;
3502 	if (rack->r_ctl.rc_sacklast == rsm)
3503 		rack->r_ctl.rc_sacklast = NULL;
3504 	memset(rsm, 0, sizeof(struct rack_sendmap));
3505 	/* Make sure we are not going to overrun our count limit of 0xff */
3506 	if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3507 		rack_free_trim(rack);
3508 	}
3509 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3510 	rack->rc_free_cnt++;
3511 }
3512 
3513 static uint32_t
3514 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3515 {
3516 	uint64_t srtt, bw, len, tim;
3517 	uint32_t segsiz, def_len, minl;
3518 
3519 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3520 	def_len = rack_def_data_window * segsiz;
3521 	if (rack->rc_gp_filled == 0) {
3522 		/*
3523 		 * We have no measurement (IW is in flight?) so
3524 		 * we can only guess using our data_window sysctl
3525 		 * value (usually 20MSS).
3526 		 */
3527 		return (def_len);
3528 	}
3529 	/*
3530 	 * Now we have a number of factors to consider.
3531 	 *
3532 	 * 1) We have a desired BDP which is usually
3533 	 *    at least 2.
3534 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3535 	 *    but we allow it too to be more.
3536 	 * 3) We want to make sure a measurement last N useconds (if
3537 	 *    we have set rack_min_measure_usec.
3538 	 *
3539 	 * We handle the first concern here by trying to create a data
3540 	 * window of max(rack_def_data_window, DesiredBDP). The
3541 	 * second concern we handle in not letting the measurement
3542 	 * window end normally until at least the required SRTT's
3543 	 * have gone by which is done further below in
3544 	 * rack_enough_for_measurement(). Finally the third concern
3545 	 * we also handle here by calculating how long that time
3546 	 * would take at the current BW and then return the
3547 	 * max of our first calculation and that length. Note
3548 	 * that if rack_min_measure_usec is 0, we don't deal
3549 	 * with concern 3. Also for both Concern 1 and 3 an
3550 	 * application limited period could end the measurement
3551 	 * earlier.
3552 	 *
3553 	 * So lets calculate the BDP with the "known" b/w using
3554 	 * the SRTT as our rtt and then multiply it by the goal.
3555 	 */
3556 	bw = rack_get_bw(rack);
3557 	srtt = (uint64_t)tp->t_srtt;
3558 	len = bw * srtt;
3559 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3560 	len *= max(1, rack_goal_bdp);
3561 	/* Now we need to round up to the nearest MSS */
3562 	len = roundup(len, segsiz);
3563 	if (rack_min_measure_usec) {
3564 		/* Now calculate our min length for this b/w */
3565 		tim = rack_min_measure_usec;
3566 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3567 		if (minl == 0)
3568 			minl = 1;
3569 		minl = roundup(minl, segsiz);
3570 		if (len < minl)
3571 			len = minl;
3572 	}
3573 	/*
3574 	 * Now if we have a very small window we want
3575 	 * to attempt to get the window that is
3576 	 * as small as possible. This happens on
3577 	 * low b/w connections and we don't want to
3578 	 * span huge numbers of rtt's between measurements.
3579 	 *
3580 	 * We basically include 2 over our "MIN window" so
3581 	 * that the measurement can be shortened (possibly) by
3582 	 * an ack'ed packet.
3583 	 */
3584 	if (len < def_len)
3585 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3586 	else
3587 		return (max((uint32_t)len, def_len));
3588 
3589 }
3590 
3591 static int
3592 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3593 {
3594 	uint32_t tim, srtts, segsiz;
3595 
3596 	/*
3597 	 * Has enough time passed for the GP measurement to be valid?
3598 	 */
3599 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3600 		/* Not enough bytes yet */
3601 		return (0);
3602 	}
3603 	if ((tp->snd_max == tp->snd_una) ||
3604 	    (th_ack == tp->snd_max)){
3605 		/*
3606 		 * All is acked quality of all acked is
3607 		 * usually low or medium, but we in theory could split
3608 		 * all acked into two cases, where you got
3609 		 * a signifigant amount of your window and
3610 		 * where you did not. For now we leave it
3611 		 * but it is something to contemplate in the
3612 		 * future. The danger here is that delayed ack
3613 		 * is effecting the last byte (which is a 50:50 chance).
3614 		 */
3615 		*quality = RACK_QUALITY_ALLACKED;
3616 		return (1);
3617 	}
3618 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3619 		/*
3620 		 * We obtained our entire window of data we wanted
3621 		 * no matter if we are in recovery or not then
3622 		 * its ok since expanding the window does not
3623 		 * make things fuzzy (or at least not as much).
3624 		 */
3625 		*quality = RACK_QUALITY_HIGH;
3626 		return (1);
3627 	}
3628 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3629 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3630 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3631 		/* Not enough bytes yet */
3632 		return (0);
3633 	}
3634 	if (rack->r_ctl.rc_first_appl &&
3635 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3636 		/*
3637 		 * We are up to the app limited send point
3638 		 * we have to measure irrespective of the time..
3639 		 */
3640 		*quality = RACK_QUALITY_APPLIMITED;
3641 		return (1);
3642 	}
3643 	/* Now what about time? */
3644 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3645 	tim = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3646 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3647 		/*
3648 		 * We do not allow a measurement if we are in recovery
3649 		 * that would shrink the goodput window we wanted.
3650 		 * This is to prevent cloudyness of when the last send
3651 		 * was actually made.
3652 		 */
3653 		*quality = RACK_QUALITY_HIGH;
3654 		return (1);
3655 	}
3656 	/* Nope not even a full SRTT has passed */
3657 	return (0);
3658 }
3659 
3660 static void
3661 rack_log_timely(struct tcp_rack *rack,
3662 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3663 		uint64_t up_bnd, int line, uint8_t method)
3664 {
3665 	if (tcp_bblogging_on(rack->rc_tp)) {
3666 		union tcp_log_stackspecific log;
3667 		struct timeval tv;
3668 
3669 		memset(&log, 0, sizeof(log));
3670 		log.u_bbr.flex1 = logged;
3671 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3672 		log.u_bbr.flex2 <<= 4;
3673 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3674 		log.u_bbr.flex2 <<= 4;
3675 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3676 		log.u_bbr.flex2 <<= 4;
3677 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3678 		log.u_bbr.flex3 = rack->rc_gp_incr;
3679 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3680 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3681 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3682 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3683 		log.u_bbr.flex8 = method;
3684 		log.u_bbr.cur_del_rate = cur_bw;
3685 		log.u_bbr.delRate = low_bnd;
3686 		log.u_bbr.bw_inuse = up_bnd;
3687 		log.u_bbr.rttProp = rack_get_bw(rack);
3688 		log.u_bbr.pkt_epoch = line;
3689 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3690 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3691 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3692 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3693 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3694 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3695 		log.u_bbr.cwnd_gain <<= 1;
3696 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3697 		log.u_bbr.cwnd_gain <<= 1;
3698 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3699 		log.u_bbr.cwnd_gain <<= 1;
3700 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3701 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3702 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3703 		    &rack->rc_inp->inp_socket->so_rcv,
3704 		    &rack->rc_inp->inp_socket->so_snd,
3705 		    TCP_TIMELY_WORK, 0,
3706 		    0, &log, false, &tv);
3707 	}
3708 }
3709 
3710 static int
3711 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3712 {
3713 	/*
3714 	 * Before we increase we need to know if
3715 	 * the estimate just made was less than
3716 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3717 	 *
3718 	 * If we already are pacing at a fast enough
3719 	 * rate to push us faster there is no sense of
3720 	 * increasing.
3721 	 *
3722 	 * We first caculate our actual pacing rate (ss or ca multiplier
3723 	 * times our cur_bw).
3724 	 *
3725 	 * Then we take the last measured rate and multipy by our
3726 	 * maximum pacing overage to give us a max allowable rate.
3727 	 *
3728 	 * If our act_rate is smaller than our max_allowable rate
3729 	 * then we should increase. Else we should hold steady.
3730 	 *
3731 	 */
3732 	uint64_t act_rate, max_allow_rate;
3733 
3734 	if (rack_timely_no_stopping)
3735 		return (1);
3736 
3737 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3738 		/*
3739 		 * Initial startup case or
3740 		 * everything is acked case.
3741 		 */
3742 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3743 				__LINE__, 9);
3744 		return (1);
3745 	}
3746 	if (mult <= 100) {
3747 		/*
3748 		 * We can always pace at or slightly above our rate.
3749 		 */
3750 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3751 				__LINE__, 9);
3752 		return (1);
3753 	}
3754 	act_rate = cur_bw * (uint64_t)mult;
3755 	act_rate /= 100;
3756 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3757 	max_allow_rate /= 100;
3758 	if (act_rate < max_allow_rate) {
3759 		/*
3760 		 * Here the rate we are actually pacing at
3761 		 * is smaller than 10% above our last measurement.
3762 		 * This means we are pacing below what we would
3763 		 * like to try to achieve (plus some wiggle room).
3764 		 */
3765 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3766 				__LINE__, 9);
3767 		return (1);
3768 	} else {
3769 		/*
3770 		 * Here we are already pacing at least rack_max_per_above(10%)
3771 		 * what we are getting back. This indicates most likely
3772 		 * that we are being limited (cwnd/rwnd/app) and can't
3773 		 * get any more b/w. There is no sense of trying to
3774 		 * raise up the pacing rate its not speeding us up
3775 		 * and we already are pacing faster than we are getting.
3776 		 */
3777 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3778 				__LINE__, 8);
3779 		return (0);
3780 	}
3781 }
3782 
3783 static void
3784 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3785 {
3786 	/*
3787 	 * When we drag bottom, we want to assure
3788 	 * that no multiplier is below 1.0, if so
3789 	 * we want to restore it to at least that.
3790 	 */
3791 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3792 		/* This is unlikely we usually do not touch recovery */
3793 		rack->r_ctl.rack_per_of_gp_rec = 100;
3794 	}
3795 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3796 		rack->r_ctl.rack_per_of_gp_ca = 100;
3797 	}
3798 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3799 		rack->r_ctl.rack_per_of_gp_ss = 100;
3800 	}
3801 }
3802 
3803 static void
3804 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3805 {
3806 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3807 		rack->r_ctl.rack_per_of_gp_ca = 100;
3808 	}
3809 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3810 		rack->r_ctl.rack_per_of_gp_ss = 100;
3811 	}
3812 }
3813 
3814 static void
3815 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3816 {
3817 	int32_t  calc, logged, plus;
3818 
3819 	logged = 0;
3820 
3821 	if (rack->rc_skip_timely)
3822 		return;
3823 	if (override) {
3824 		/*
3825 		 * override is passed when we are
3826 		 * loosing b/w and making one last
3827 		 * gasp at trying to not loose out
3828 		 * to a new-reno flow.
3829 		 */
3830 		goto extra_boost;
3831 	}
3832 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3833 	if (rack->rc_gp_incr &&
3834 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3835 		/*
3836 		 * Reset and get 5 strokes more before the boost. Note
3837 		 * that the count is 0 based so we have to add one.
3838 		 */
3839 extra_boost:
3840 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3841 		rack->rc_gp_timely_inc_cnt = 0;
3842 	} else
3843 		plus = (uint32_t)rack_gp_increase_per;
3844 	/* Must be at least 1% increase for true timely increases */
3845 	if ((plus < 1) &&
3846 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3847 		plus = 1;
3848 	if (rack->rc_gp_saw_rec &&
3849 	    (rack->rc_gp_no_rec_chg == 0) &&
3850 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3851 				  rack->r_ctl.rack_per_of_gp_rec)) {
3852 		/* We have been in recovery ding it too */
3853 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3854 		if (calc > 0xffff)
3855 			calc = 0xffff;
3856 		logged |= 1;
3857 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3858 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3859 		    (rack->rc_dragged_bottom == 0) &&
3860 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3861 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3862 	}
3863 	if (rack->rc_gp_saw_ca &&
3864 	    (rack->rc_gp_saw_ss == 0) &&
3865 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3866 				  rack->r_ctl.rack_per_of_gp_ca)) {
3867 		/* In CA */
3868 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3869 		if (calc > 0xffff)
3870 			calc = 0xffff;
3871 		logged |= 2;
3872 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3873 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3874 		    (rack->rc_dragged_bottom == 0) &&
3875 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3876 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3877 	}
3878 	if (rack->rc_gp_saw_ss &&
3879 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3880 				  rack->r_ctl.rack_per_of_gp_ss)) {
3881 		/* In SS */
3882 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3883 		if (calc > 0xffff)
3884 			calc = 0xffff;
3885 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3886 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3887 		    (rack->rc_dragged_bottom == 0) &&
3888 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3889 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3890 		logged |= 4;
3891 	}
3892 	if (logged &&
3893 	    (rack->rc_gp_incr == 0)){
3894 		/* Go into increment mode */
3895 		rack->rc_gp_incr = 1;
3896 		rack->rc_gp_timely_inc_cnt = 0;
3897 	}
3898 	if (rack->rc_gp_incr &&
3899 	    logged &&
3900 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3901 		rack->rc_gp_timely_inc_cnt++;
3902 	}
3903 	rack_log_timely(rack,  logged, plus, 0, 0,
3904 			__LINE__, 1);
3905 }
3906 
3907 static uint32_t
3908 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3909 {
3910 	/*-
3911 	 * norm_grad = rtt_diff / minrtt;
3912 	 * new_per = curper * (1 - B * norm_grad)
3913 	 *
3914 	 * B = rack_gp_decrease_per (default 80%)
3915 	 * rtt_dif = input var current rtt-diff
3916 	 * curper = input var current percentage
3917 	 * minrtt = from rack filter
3918 	 *
3919 	 * In order to do the floating point calculations above we
3920 	 * do an integer conversion. The code looks confusing so let me
3921 	 * translate it into something that use more variables and
3922 	 * is clearer for us humans :)
3923 	 *
3924 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
3925 	 * uint32_t perf;
3926 	 *
3927 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3928 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3929 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3930 	 * inverse /= 1000000;
3931 	 * reduce_by = (1000000 - inverse);
3932 	 * final_result = (cur_per * reduce_by) / 1000000;
3933 	 * perf = (uint32_t)final_result;
3934 	 */
3935 	uint64_t perf;
3936 
3937 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3938 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3939 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3940 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3941 		     (uint64_t)1000000)) /
3942 		(uint64_t)1000000);
3943 	if (perf > curper) {
3944 		/* TSNH */
3945 		perf = curper - 1;
3946 	}
3947 	return ((uint32_t)perf);
3948 }
3949 
3950 static uint32_t
3951 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3952 {
3953 	/*
3954 	 *                                   highrttthresh
3955 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3956 	 *                                     gp_srtt
3957 	 *
3958 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
3959 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3960 	 */
3961 	uint64_t perf;
3962 	uint32_t highrttthresh;
3963 
3964 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3965 
3966 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3967 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3968 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3969 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3970 	if (tcp_bblogging_on(rack->rc_tp)) {
3971 		uint64_t log1;
3972 
3973 		log1 = rtt;
3974 		log1 <<= 32;
3975 		log1 |= highrttthresh;
3976 		rack_log_timely(rack,
3977 				rack_gp_decrease_per,
3978 				(uint64_t)curper,
3979 				log1,
3980 				perf,
3981 				__LINE__,
3982 				15);
3983 	}
3984 	return (perf);
3985 }
3986 
3987 static void
3988 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3989 {
3990 	uint64_t logvar, logvar2, logvar3;
3991 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3992 
3993 	if (rack->rc_skip_timely)
3994 		return;
3995 	if (rack->rc_gp_incr) {
3996 		/* Turn off increment counting */
3997 		rack->rc_gp_incr = 0;
3998 		rack->rc_gp_timely_inc_cnt = 0;
3999 	}
4000 	ss_red = ca_red = rec_red = 0;
4001 	logged = 0;
4002 	/* Calculate the reduction value */
4003 	if (rtt_diff < 0) {
4004 		rtt_diff *= -1;
4005 	}
4006 	/* Must be at least 1% reduction */
4007 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4008 		/* We have been in recovery ding it too */
4009 		if (timely_says == 2) {
4010 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4011 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4012 			if (alt < new_per)
4013 				val = alt;
4014 			else
4015 				val = new_per;
4016 		} else
4017 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4018 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4019 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4020 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4021 		} else {
4022 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4023 			rec_red = 0;
4024 		}
4025 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4026 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4027 		logged |= 1;
4028 	}
4029 	if (rack->rc_gp_saw_ss) {
4030 		/* Sent in SS */
4031 		if (timely_says == 2) {
4032 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4033 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4034 			if (alt < new_per)
4035 				val = alt;
4036 			else
4037 				val = new_per;
4038 		} else
4039 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4040 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4041 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4042 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4043 		} else {
4044 			ss_red = new_per;
4045 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4046 			logvar = new_per;
4047 			logvar <<= 32;
4048 			logvar |= alt;
4049 			logvar2 = (uint32_t)rtt;
4050 			logvar2 <<= 32;
4051 			logvar2 |= (uint32_t)rtt_diff;
4052 			logvar3 = rack_gp_rtt_maxmul;
4053 			logvar3 <<= 32;
4054 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4055 			rack_log_timely(rack, timely_says,
4056 					logvar2, logvar3,
4057 					logvar, __LINE__, 10);
4058 		}
4059 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4060 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4061 		logged |= 4;
4062 	} else if (rack->rc_gp_saw_ca) {
4063 		/* Sent in CA */
4064 		if (timely_says == 2) {
4065 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4066 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4067 			if (alt < new_per)
4068 				val = alt;
4069 			else
4070 				val = new_per;
4071 		} else
4072 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4073 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4074 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4075 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4076 		} else {
4077 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4078 			ca_red = 0;
4079 			logvar = new_per;
4080 			logvar <<= 32;
4081 			logvar |= alt;
4082 			logvar2 = (uint32_t)rtt;
4083 			logvar2 <<= 32;
4084 			logvar2 |= (uint32_t)rtt_diff;
4085 			logvar3 = rack_gp_rtt_maxmul;
4086 			logvar3 <<= 32;
4087 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4088 			rack_log_timely(rack, timely_says,
4089 					logvar2, logvar3,
4090 					logvar, __LINE__, 10);
4091 		}
4092 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4093 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4094 		logged |= 2;
4095 	}
4096 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4097 		rack->rc_gp_timely_dec_cnt++;
4098 		if (rack_timely_dec_clear &&
4099 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4100 			rack->rc_gp_timely_dec_cnt = 0;
4101 	}
4102 	logvar = ss_red;
4103 	logvar <<= 32;
4104 	logvar |= ca_red;
4105 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4106 			__LINE__, 2);
4107 }
4108 
4109 static void
4110 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4111 		     uint32_t rtt, uint32_t line, uint8_t reas)
4112 {
4113 	if (tcp_bblogging_on(rack->rc_tp)) {
4114 		union tcp_log_stackspecific log;
4115 		struct timeval tv;
4116 
4117 		memset(&log, 0, sizeof(log));
4118 		log.u_bbr.flex1 = line;
4119 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4120 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4121 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4122 		log.u_bbr.flex5 = rtt;
4123 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4124 		log.u_bbr.flex6 <<= 1;
4125 		log.u_bbr.flex6 |= rack->forced_ack;
4126 		log.u_bbr.flex6 <<= 1;
4127 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4128 		log.u_bbr.flex6 <<= 1;
4129 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4130 		log.u_bbr.flex6 <<= 1;
4131 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4132 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4133 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4134 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4135 		log.u_bbr.flex8 = reas;
4136 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4137 		log.u_bbr.delRate = rack_get_bw(rack);
4138 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4139 		log.u_bbr.cur_del_rate <<= 32;
4140 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4141 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4142 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4143 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4144 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4145 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4146 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4147 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4148 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4149 		log.u_bbr.rttProp = us_cts;
4150 		log.u_bbr.rttProp <<= 32;
4151 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4152 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4153 		    &rack->rc_inp->inp_socket->so_rcv,
4154 		    &rack->rc_inp->inp_socket->so_snd,
4155 		    BBR_LOG_RTT_SHRINKS, 0,
4156 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4157 	}
4158 }
4159 
4160 static void
4161 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4162 {
4163 	uint64_t bwdp;
4164 
4165 	bwdp = rack_get_bw(rack);
4166 	bwdp *= (uint64_t)rtt;
4167 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4168 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4169 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4170 		/*
4171 		 * A window protocol must be able to have 4 packets
4172 		 * outstanding as the floor in order to function
4173 		 * (especially considering delayed ack :D).
4174 		 */
4175 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4176 	}
4177 }
4178 
4179 static void
4180 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4181 {
4182 	/**
4183 	 * ProbeRTT is a bit different in rack_pacing than in
4184 	 * BBR. It is like BBR in that it uses the lowering of
4185 	 * the RTT as a signal that we saw something new and
4186 	 * counts from there for how long between. But it is
4187 	 * different in that its quite simple. It does not
4188 	 * play with the cwnd and wait until we get down
4189 	 * to N segments outstanding and hold that for
4190 	 * 200ms. Instead it just sets the pacing reduction
4191 	 * rate to a set percentage (70 by default) and hold
4192 	 * that for a number of recent GP Srtt's.
4193 	 */
4194 	uint32_t segsiz;
4195 
4196 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4197 	if (rack->rc_gp_dyn_mul == 0)
4198 		return;
4199 
4200 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4201 		/* We are idle */
4202 		return;
4203 	}
4204 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4205 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4206 		/*
4207 		 * Stop the goodput now, the idea here is
4208 		 * that future measurements with in_probe_rtt
4209 		 * won't register if they are not greater so
4210 		 * we want to get what info (if any) is available
4211 		 * now.
4212 		 */
4213 		rack_do_goodput_measurement(rack->rc_tp, rack,
4214 					    rack->rc_tp->snd_una, __LINE__,
4215 					    RACK_QUALITY_PROBERTT);
4216 	}
4217 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4218 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4219 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4220 		     rack->r_ctl.rc_pace_min_segs);
4221 	rack->in_probe_rtt = 1;
4222 	rack->measure_saw_probe_rtt = 1;
4223 	rack->r_ctl.rc_time_probertt_starts = 0;
4224 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4225 	if (rack_probertt_use_min_rtt_entry)
4226 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4227 	else
4228 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4229 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4230 			     __LINE__, RACK_RTTS_ENTERPROBE);
4231 }
4232 
4233 static void
4234 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4235 {
4236 	struct rack_sendmap *rsm;
4237 	uint32_t segsiz;
4238 
4239 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4240 		     rack->r_ctl.rc_pace_min_segs);
4241 	rack->in_probe_rtt = 0;
4242 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4243 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4244 		/*
4245 		 * Stop the goodput now, the idea here is
4246 		 * that future measurements with in_probe_rtt
4247 		 * won't register if they are not greater so
4248 		 * we want to get what info (if any) is available
4249 		 * now.
4250 		 */
4251 		rack_do_goodput_measurement(rack->rc_tp, rack,
4252 					    rack->rc_tp->snd_una, __LINE__,
4253 					    RACK_QUALITY_PROBERTT);
4254 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4255 		/*
4256 		 * We don't have enough data to make a measurement.
4257 		 * So lets just stop and start here after exiting
4258 		 * probe-rtt. We probably are not interested in
4259 		 * the results anyway.
4260 		 */
4261 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4262 	}
4263 	/*
4264 	 * Measurements through the current snd_max are going
4265 	 * to be limited by the slower pacing rate.
4266 	 *
4267 	 * We need to mark these as app-limited so we
4268 	 * don't collapse the b/w.
4269 	 */
4270 	rsm = tqhash_max(rack->r_ctl.tqh);
4271 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4272 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4273 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4274 		else {
4275 			/*
4276 			 * Go out to the end app limited and mark
4277 			 * this new one as next and move the end_appl up
4278 			 * to this guy.
4279 			 */
4280 			if (rack->r_ctl.rc_end_appl)
4281 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4282 			rack->r_ctl.rc_end_appl = rsm;
4283 		}
4284 		rsm->r_flags |= RACK_APP_LIMITED;
4285 		rack->r_ctl.rc_app_limited_cnt++;
4286 	}
4287 	/*
4288 	 * Now, we need to examine our pacing rate multipliers.
4289 	 * If its under 100%, we need to kick it back up to
4290 	 * 100%. We also don't let it be over our "max" above
4291 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4292 	 * Note setting clamp_atexit_prtt to 0 has the effect
4293 	 * of setting CA/SS to 100% always at exit (which is
4294 	 * the default behavior).
4295 	 */
4296 	if (rack_probertt_clear_is) {
4297 		rack->rc_gp_incr = 0;
4298 		rack->rc_gp_bwred = 0;
4299 		rack->rc_gp_timely_inc_cnt = 0;
4300 		rack->rc_gp_timely_dec_cnt = 0;
4301 	}
4302 	/* Do we do any clamping at exit? */
4303 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4304 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4305 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4306 	}
4307 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4308 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4309 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4310 	}
4311 	/*
4312 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4313 	 * after exiting.
4314 	 */
4315 	rack->r_ctl.rc_rtt_diff = 0;
4316 
4317 	/* Clear all flags so we start fresh */
4318 	rack->rc_tp->t_bytes_acked = 0;
4319 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4320 	/*
4321 	 * If configured to, set the cwnd and ssthresh to
4322 	 * our targets.
4323 	 */
4324 	if (rack_probe_rtt_sets_cwnd) {
4325 		uint64_t ebdp;
4326 		uint32_t setto;
4327 
4328 		/* Set ssthresh so we get into CA once we hit our target */
4329 		if (rack_probertt_use_min_rtt_exit == 1) {
4330 			/* Set to min rtt */
4331 			rack_set_prtt_target(rack, segsiz,
4332 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4333 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4334 			/* Set to current gp rtt */
4335 			rack_set_prtt_target(rack, segsiz,
4336 					     rack->r_ctl.rc_gp_srtt);
4337 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4338 			/* Set to entry gp rtt */
4339 			rack_set_prtt_target(rack, segsiz,
4340 					     rack->r_ctl.rc_entry_gp_rtt);
4341 		} else {
4342 			uint64_t sum;
4343 			uint32_t setval;
4344 
4345 			sum = rack->r_ctl.rc_entry_gp_rtt;
4346 			sum *= 10;
4347 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4348 			if (sum >= 20) {
4349 				/*
4350 				 * A highly buffered path needs
4351 				 * cwnd space for timely to work.
4352 				 * Lets set things up as if
4353 				 * we are heading back here again.
4354 				 */
4355 				setval = rack->r_ctl.rc_entry_gp_rtt;
4356 			} else if (sum >= 15) {
4357 				/*
4358 				 * Lets take the smaller of the
4359 				 * two since we are just somewhat
4360 				 * buffered.
4361 				 */
4362 				setval = rack->r_ctl.rc_gp_srtt;
4363 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4364 					setval = rack->r_ctl.rc_entry_gp_rtt;
4365 			} else {
4366 				/*
4367 				 * Here we are not highly buffered
4368 				 * and should pick the min we can to
4369 				 * keep from causing loss.
4370 				 */
4371 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4372 			}
4373 			rack_set_prtt_target(rack, segsiz,
4374 					     setval);
4375 		}
4376 		if (rack_probe_rtt_sets_cwnd > 1) {
4377 			/* There is a percentage here to boost */
4378 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4379 			ebdp *= rack_probe_rtt_sets_cwnd;
4380 			ebdp /= 100;
4381 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4382 		} else
4383 			setto = rack->r_ctl.rc_target_probertt_flight;
4384 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4385 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4386 			/* Enforce a min */
4387 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4388 		}
4389 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4390 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4391 	}
4392 	rack_log_rtt_shrinks(rack,  us_cts,
4393 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4394 			     __LINE__, RACK_RTTS_EXITPROBE);
4395 	/* Clear times last so log has all the info */
4396 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4397 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4398 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4399 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4400 }
4401 
4402 static void
4403 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4404 {
4405 	/* Check in on probe-rtt */
4406 
4407 	if (rack->rc_gp_filled == 0) {
4408 		/* We do not do p-rtt unless we have gp measurements */
4409 		return;
4410 	}
4411 	if (rack->in_probe_rtt) {
4412 		uint64_t no_overflow;
4413 		uint32_t endtime, must_stay;
4414 
4415 		if (rack->r_ctl.rc_went_idle_time &&
4416 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4417 			/*
4418 			 * We went idle during prtt, just exit now.
4419 			 */
4420 			rack_exit_probertt(rack, us_cts);
4421 		} else if (rack_probe_rtt_safety_val &&
4422 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4423 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4424 			/*
4425 			 * Probe RTT safety value triggered!
4426 			 */
4427 			rack_log_rtt_shrinks(rack,  us_cts,
4428 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4429 					     __LINE__, RACK_RTTS_SAFETY);
4430 			rack_exit_probertt(rack, us_cts);
4431 		}
4432 		/* Calculate the max we will wait */
4433 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4434 		if (rack->rc_highly_buffered)
4435 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4436 		/* Calculate the min we must wait */
4437 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4438 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4439 		    TSTMP_LT(us_cts, endtime)) {
4440 			uint32_t calc;
4441 			/* Do we lower more? */
4442 no_exit:
4443 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4444 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4445 			else
4446 				calc = 0;
4447 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4448 			if (calc) {
4449 				/* Maybe */
4450 				calc *= rack_per_of_gp_probertt_reduce;
4451 				if (calc > rack_per_of_gp_probertt)
4452 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4453 				else
4454 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4455 				/* Limit it too */
4456 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4457 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4458 			}
4459 			/* We must reach target or the time set */
4460 			return;
4461 		}
4462 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4463 			if ((TSTMP_LT(us_cts, must_stay) &&
4464 			     rack->rc_highly_buffered) ||
4465 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4466 			      rack->r_ctl.rc_target_probertt_flight)) {
4467 				/* We are not past the must_stay time */
4468 				goto no_exit;
4469 			}
4470 			rack_log_rtt_shrinks(rack,  us_cts,
4471 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4472 					     __LINE__, RACK_RTTS_REACHTARGET);
4473 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4474 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4475 				rack->r_ctl.rc_time_probertt_starts = 1;
4476 			/* Restore back to our rate we want to pace at in prtt */
4477 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4478 		}
4479 		/*
4480 		 * Setup our end time, some number of gp_srtts plus 200ms.
4481 		 */
4482 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4483 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4484 		if (rack_probertt_gpsrtt_cnt_div)
4485 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4486 		else
4487 			endtime = 0;
4488 		endtime += rack_min_probertt_hold;
4489 		endtime += rack->r_ctl.rc_time_probertt_starts;
4490 		if (TSTMP_GEQ(us_cts,  endtime)) {
4491 			/* yes, exit probertt */
4492 			rack_exit_probertt(rack, us_cts);
4493 		}
4494 
4495 	} else if ((rack->rc_skip_timely == 0) &&
4496 		   (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4497 		   ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4498 		/* Go into probertt, its been too long since we went lower */
4499 		rack_enter_probertt(rack, us_cts);
4500 	}
4501 }
4502 
4503 static void
4504 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4505 		       uint32_t rtt, int32_t rtt_diff)
4506 {
4507 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4508 	uint32_t losses;
4509 
4510 	if ((rack->rc_gp_dyn_mul == 0) ||
4511 	    (rack->use_fixed_rate) ||
4512 	    (rack->in_probe_rtt) ||
4513 	    (rack->rc_always_pace == 0)) {
4514 		/* No dynamic GP multiplier in play */
4515 		return;
4516 	}
4517 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4518 	cur_bw = rack_get_bw(rack);
4519 	/* Calculate our up and down range */
4520 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4521 	up_bnd /= 100;
4522 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4523 
4524 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4525 	subfr /= 100;
4526 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4527 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4528 		/*
4529 		 * This is the case where our RTT is above
4530 		 * the max target and we have been configured
4531 		 * to just do timely no bonus up stuff in that case.
4532 		 *
4533 		 * There are two configurations, set to 1, and we
4534 		 * just do timely if we are over our max. If its
4535 		 * set above 1 then we slam the multipliers down
4536 		 * to 100 and then decrement per timely.
4537 		 */
4538 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4539 				__LINE__, 3);
4540 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4541 			rack_validate_multipliers_at_or_below_100(rack);
4542 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4543 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4544 		/*
4545 		 * We are decreasing this is a bit complicated this
4546 		 * means we are loosing ground. This could be
4547 		 * because another flow entered and we are competing
4548 		 * for b/w with it. This will push the RTT up which
4549 		 * makes timely unusable unless we want to get shoved
4550 		 * into a corner and just be backed off (the age
4551 		 * old problem with delay based CC).
4552 		 *
4553 		 * On the other hand if it was a route change we
4554 		 * would like to stay somewhat contained and not
4555 		 * blow out the buffers.
4556 		 */
4557 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4558 				__LINE__, 3);
4559 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4560 		if (rack->rc_gp_bwred == 0) {
4561 			/* Go into reduction counting */
4562 			rack->rc_gp_bwred = 1;
4563 			rack->rc_gp_timely_dec_cnt = 0;
4564 		}
4565 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4566 			/*
4567 			 * Push another time with a faster pacing
4568 			 * to try to gain back (we include override to
4569 			 * get a full raise factor).
4570 			 */
4571 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4572 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4573 			    (timely_says == 0) ||
4574 			    (rack_down_raise_thresh == 0)) {
4575 				/*
4576 				 * Do an override up in b/w if we were
4577 				 * below the threshold or if the threshold
4578 				 * is zero we always do the raise.
4579 				 */
4580 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4581 			} else {
4582 				/* Log it stays the same */
4583 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4584 						__LINE__, 11);
4585 			}
4586 			rack->rc_gp_timely_dec_cnt++;
4587 			/* We are not incrementing really no-count */
4588 			rack->rc_gp_incr = 0;
4589 			rack->rc_gp_timely_inc_cnt = 0;
4590 		} else {
4591 			/*
4592 			 * Lets just use the RTT
4593 			 * information and give up
4594 			 * pushing.
4595 			 */
4596 			goto use_timely;
4597 		}
4598 	} else if ((timely_says != 2) &&
4599 		    !losses &&
4600 		    (last_bw_est > up_bnd)) {
4601 		/*
4602 		 * We are increasing b/w lets keep going, updating
4603 		 * our b/w and ignoring any timely input, unless
4604 		 * of course we are at our max raise (if there is one).
4605 		 */
4606 
4607 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4608 				__LINE__, 3);
4609 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4610 		if (rack->rc_gp_saw_ss &&
4611 		    rack->r_ctl.rack_per_upper_bound_ss &&
4612 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4613 			    /*
4614 			     * In cases where we can't go higher
4615 			     * we should just use timely.
4616 			     */
4617 			    goto use_timely;
4618 		}
4619 		if (rack->rc_gp_saw_ca &&
4620 		    rack->r_ctl.rack_per_upper_bound_ca &&
4621 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4622 			    /*
4623 			     * In cases where we can't go higher
4624 			     * we should just use timely.
4625 			     */
4626 			    goto use_timely;
4627 		}
4628 		rack->rc_gp_bwred = 0;
4629 		rack->rc_gp_timely_dec_cnt = 0;
4630 		/* You get a set number of pushes if timely is trying to reduce */
4631 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4632 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4633 		} else {
4634 			/* Log it stays the same */
4635 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4636 			    __LINE__, 12);
4637 		}
4638 		return;
4639 	} else {
4640 		/*
4641 		 * We are staying between the lower and upper range bounds
4642 		 * so use timely to decide.
4643 		 */
4644 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4645 				__LINE__, 3);
4646 use_timely:
4647 		if (timely_says) {
4648 			rack->rc_gp_incr = 0;
4649 			rack->rc_gp_timely_inc_cnt = 0;
4650 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4651 			    !losses &&
4652 			    (last_bw_est < low_bnd)) {
4653 				/* We are loosing ground */
4654 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4655 				rack->rc_gp_timely_dec_cnt++;
4656 				/* We are not incrementing really no-count */
4657 				rack->rc_gp_incr = 0;
4658 				rack->rc_gp_timely_inc_cnt = 0;
4659 			} else
4660 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4661 		} else {
4662 			rack->rc_gp_bwred = 0;
4663 			rack->rc_gp_timely_dec_cnt = 0;
4664 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4665 		}
4666 	}
4667 }
4668 
4669 static int32_t
4670 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4671 {
4672 	int32_t timely_says;
4673 	uint64_t log_mult, log_rtt_a_diff;
4674 
4675 	log_rtt_a_diff = rtt;
4676 	log_rtt_a_diff <<= 32;
4677 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4678 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4679 		    rack_gp_rtt_maxmul)) {
4680 		/* Reduce the b/w multiplier */
4681 		timely_says = 2;
4682 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4683 		log_mult <<= 32;
4684 		log_mult |= prev_rtt;
4685 		rack_log_timely(rack,  timely_says, log_mult,
4686 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4687 				log_rtt_a_diff, __LINE__, 4);
4688 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4689 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4690 			    max(rack_gp_rtt_mindiv , 1)))) {
4691 		/* Increase the b/w multiplier */
4692 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4693 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4694 			 max(rack_gp_rtt_mindiv , 1));
4695 		log_mult <<= 32;
4696 		log_mult |= prev_rtt;
4697 		timely_says = 0;
4698 		rack_log_timely(rack,  timely_says, log_mult ,
4699 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4700 				log_rtt_a_diff, __LINE__, 5);
4701 	} else {
4702 		/*
4703 		 * Use a gradient to find it the timely gradient
4704 		 * is:
4705 		 * grad = rc_rtt_diff / min_rtt;
4706 		 *
4707 		 * anything below or equal to 0 will be
4708 		 * a increase indication. Anything above
4709 		 * zero is a decrease. Note we take care
4710 		 * of the actual gradient calculation
4711 		 * in the reduction (its not needed for
4712 		 * increase).
4713 		 */
4714 		log_mult = prev_rtt;
4715 		if (rtt_diff <= 0) {
4716 			/*
4717 			 * Rttdiff is less than zero, increase the
4718 			 * b/w multiplier (its 0 or negative)
4719 			 */
4720 			timely_says = 0;
4721 			rack_log_timely(rack,  timely_says, log_mult,
4722 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4723 		} else {
4724 			/* Reduce the b/w multiplier */
4725 			timely_says = 1;
4726 			rack_log_timely(rack,  timely_says, log_mult,
4727 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4728 		}
4729 	}
4730 	return (timely_says);
4731 }
4732 
4733 static __inline int
4734 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4735 {
4736 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4737 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4738 		/**
4739 		 * This covers the case that the
4740 		 * resent is completely inside
4741 		 * the gp range or up to it.
4742 		 *      |----------------|
4743 		 *      |-----| <or>
4744 		 *            |----|
4745 		 *            <or>   |---|
4746 		 */
4747 		return (1);
4748 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4749 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4750 		/**
4751 		 * This covers the case of
4752 		 *      |--------------|
4753 		 *  |-------->|
4754 		 */
4755 		return (1);
4756 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4757 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4758 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4759 
4760 		/**
4761 		 * This covers the case of
4762 		 *      |--------------|
4763 		 *              |-------->|
4764 		 */
4765 		return (1);
4766 	}
4767 	return (0);
4768 }
4769 
4770 static __inline void
4771 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4772 {
4773 
4774 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4775 		return;
4776 	/*
4777 	 * We have a Goodput measurement in progress. Mark
4778 	 * the send if its within the window. If its not
4779 	 * in the window make sure it does not have the mark.
4780 	 */
4781 	if (rack_in_gp_window(tp, rsm))
4782 		rsm->r_flags |= RACK_IN_GP_WIN;
4783 	else
4784 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4785 }
4786 
4787 static __inline void
4788 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4789 {
4790 	/* A GP measurement is ending, clear all marks on the send map*/
4791 	struct rack_sendmap *rsm = NULL;
4792 
4793 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4794 	if (rsm == NULL) {
4795 		rsm = tqhash_min(rack->r_ctl.tqh);
4796 	}
4797 	/* Nothing left? */
4798 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4799 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4800 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4801 	}
4802 }
4803 
4804 
4805 static __inline void
4806 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4807 {
4808 	struct rack_sendmap *rsm = NULL;
4809 
4810 	if (tp->snd_una == tp->snd_max) {
4811 		/* Nothing outstanding yet, nothing to do here */
4812 		return;
4813 	}
4814 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4815 		/*
4816 		 * We are measuring ahead of some outstanding
4817 		 * data. We need to walk through up until we get
4818 		 * to gp_seq marking so that no rsm is set incorrectly
4819 		 * with RACK_IN_GP_WIN.
4820 		 */
4821 		rsm = tqhash_min(rack->r_ctl.tqh);
4822 		while (rsm != NULL) {
4823 			rack_mark_in_gp_win(tp, rsm);
4824 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4825 				break;
4826 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4827 		}
4828 	}
4829 	if (rsm == NULL) {
4830 		/*
4831 		 * Need to find the GP seq, if rsm is
4832 		 * set we stopped as we hit it.
4833 		 */
4834 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4835 		if (rsm == NULL)
4836 			return;
4837 		rack_mark_in_gp_win(tp, rsm);
4838 	}
4839 	/*
4840 	 * Now we may need to mark already sent rsm, ahead of
4841 	 * gput_seq in the window since they may have been sent
4842 	 * *before* we started our measurment. The rsm, if non-null
4843 	 * has been marked (note if rsm would have been NULL we would have
4844 	 * returned in the previous block). So we go to the next, and continue
4845 	 * until we run out of entries or we exceed the gp_ack value.
4846 	 */
4847 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4848 	while (rsm) {
4849 		rack_mark_in_gp_win(tp, rsm);
4850 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4851 			break;
4852 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4853 	}
4854 }
4855 
4856 static void
4857 rack_log_gp_calc(struct tcp_rack *rack, uint32_t add_part, uint32_t sub_part, uint32_t srtt, uint64_t meas_bw, uint64_t utim, uint8_t meth, uint32_t line)
4858 {
4859 	if (tcp_bblogging_on(rack->rc_tp)) {
4860 		union tcp_log_stackspecific log;
4861 		struct timeval tv;
4862 
4863 		memset(&log, 0, sizeof(log));
4864 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4865 		log.u_bbr.flex1 = add_part;
4866 		log.u_bbr.flex2 = sub_part;
4867 		log.u_bbr.flex3 = rack_wma_divisor;
4868 		log.u_bbr.flex4 = srtt;
4869 		log.u_bbr.flex7 = (uint16_t)line;
4870 		log.u_bbr.flex8 = meth;
4871 		log.u_bbr.delRate = rack->r_ctl.gp_bw;
4872 		log.u_bbr.cur_del_rate = meas_bw;
4873 		log.u_bbr.rttProp = utim;
4874 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4875 		    &rack->rc_inp->inp_socket->so_rcv,
4876 		    &rack->rc_inp->inp_socket->so_snd,
4877 		    BBR_LOG_THRESH_CALC, 0,
4878 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4879 	}
4880 }
4881 
4882 static void
4883 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4884 			    tcp_seq th_ack, int line, uint8_t quality)
4885 {
4886 	uint64_t tim, bytes_ps, stim, utim;
4887 	uint32_t segsiz, bytes, reqbytes, us_cts;
4888 	int32_t gput, new_rtt_diff, timely_says;
4889 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4890 	int did_add = 0;
4891 
4892 	us_cts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
4893 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4894 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4895 		tim = us_cts - tp->gput_ts;
4896 	else
4897 		tim = 0;
4898 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4899 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4900 	else
4901 		stim = 0;
4902 	/*
4903 	 * Use the larger of the send time or ack time. This prevents us
4904 	 * from being influenced by ack artifacts to come up with too
4905 	 * high of measurement. Note that since we are spanning over many more
4906 	 * bytes in most of our measurements hopefully that is less likely to
4907 	 * occur.
4908 	 */
4909 	if (tim > stim)
4910 		utim = max(tim, 1);
4911 	else
4912 		utim = max(stim, 1);
4913 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4914 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4915 	if ((tim == 0) && (stim == 0)) {
4916 		/*
4917 		 * Invalid measurement time, maybe
4918 		 * all on one ack/one send?
4919 		 */
4920 		bytes = 0;
4921 		bytes_ps = 0;
4922 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4923 					   0, 0, 0, 10, __LINE__, NULL, quality);
4924 		goto skip_measurement;
4925 	}
4926 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4927 		/* We never made a us_rtt measurement? */
4928 		bytes = 0;
4929 		bytes_ps = 0;
4930 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4931 					   0, 0, 0, 10, __LINE__, NULL, quality);
4932 		goto skip_measurement;
4933 	}
4934 	/*
4935 	 * Calculate the maximum possible b/w this connection
4936 	 * could have. We base our calculation on the lowest
4937 	 * rtt we have seen during the measurement and the
4938 	 * largest rwnd the client has given us in that time. This
4939 	 * forms a BDP that is the maximum that we could ever
4940 	 * get to the client. Anything larger is not valid.
4941 	 *
4942 	 * I originally had code here that rejected measurements
4943 	 * where the time was less than 1/2 the latest us_rtt.
4944 	 * But after thinking on that I realized its wrong since
4945 	 * say you had a 150Mbps or even 1Gbps link, and you
4946 	 * were a long way away.. example I am in Europe (100ms rtt)
4947 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4948 	 * bytes my time would be 1.2ms, and yet my rtt would say
4949 	 * the measurement was invalid the time was < 50ms. The
4950 	 * same thing is true for 150Mb (8ms of time).
4951 	 *
4952 	 * A better way I realized is to look at what the maximum
4953 	 * the connection could possibly do. This is gated on
4954 	 * the lowest RTT we have seen and the highest rwnd.
4955 	 * We should in theory never exceed that, if we are
4956 	 * then something on the path is storing up packets
4957 	 * and then feeding them all at once to our endpoint
4958 	 * messing up our measurement.
4959 	 */
4960 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4961 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4962 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4963 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4964 		/* No measurement can be made */
4965 		bytes = 0;
4966 		bytes_ps = 0;
4967 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4968 					   0, 0, 0, 10, __LINE__, NULL, quality);
4969 		goto skip_measurement;
4970 	} else
4971 		bytes = (th_ack - tp->gput_seq);
4972 	bytes_ps = (uint64_t)bytes;
4973 	/*
4974 	 * Don't measure a b/w for pacing unless we have gotten at least
4975 	 * an initial windows worth of data in this measurement interval.
4976 	 *
4977 	 * Small numbers of bytes get badly influenced by delayed ack and
4978 	 * other artifacts. Note we take the initial window or our
4979 	 * defined minimum GP (defaulting to 10 which hopefully is the
4980 	 * IW).
4981 	 */
4982 	if (rack->rc_gp_filled == 0) {
4983 		/*
4984 		 * The initial estimate is special. We
4985 		 * have blasted out an IW worth of packets
4986 		 * without a real valid ack ts results. We
4987 		 * then setup the app_limited_needs_set flag,
4988 		 * this should get the first ack in (probably 2
4989 		 * MSS worth) to be recorded as the timestamp.
4990 		 * We thus allow a smaller number of bytes i.e.
4991 		 * IW - 2MSS.
4992 		 */
4993 		reqbytes -= (2 * segsiz);
4994 		/* Also lets fill previous for our first measurement to be neutral */
4995 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4996 	}
4997 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4998 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4999 					   rack->r_ctl.rc_app_limited_cnt,
5000 					   0, 0, 10, __LINE__, NULL, quality);
5001 		goto skip_measurement;
5002 	}
5003 	/*
5004 	 * We now need to calculate the Timely like status so
5005 	 * we can update (possibly) the b/w multipliers.
5006 	 */
5007 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5008 	if (rack->rc_gp_filled == 0) {
5009 		/* No previous reading */
5010 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5011 	} else {
5012 		if (rack->measure_saw_probe_rtt == 0) {
5013 			/*
5014 			 * We don't want a probertt to be counted
5015 			 * since it will be negative incorrectly. We
5016 			 * expect to be reducing the RTT when we
5017 			 * pace at a slower rate.
5018 			 */
5019 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5020 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5021 		}
5022 	}
5023 	timely_says = rack_make_timely_judgement(rack,
5024 	    rack->r_ctl.rc_gp_srtt,
5025 	    rack->r_ctl.rc_rtt_diff,
5026 	    rack->r_ctl.rc_prev_gp_srtt
5027 	);
5028 	bytes_ps *= HPTS_USEC_IN_SEC;
5029 	bytes_ps /= utim;
5030 	if (bytes_ps > rack->r_ctl.last_max_bw) {
5031 		/*
5032 		 * Something is on path playing
5033 		 * since this b/w is not possible based
5034 		 * on our BDP (highest rwnd and lowest rtt
5035 		 * we saw in the measurement window).
5036 		 *
5037 		 * Another option here would be to
5038 		 * instead skip the measurement.
5039 		 */
5040 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5041 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
5042 					   11, __LINE__, NULL, quality);
5043 		bytes_ps = rack->r_ctl.last_max_bw;
5044 	}
5045 	/* We store gp for b/w in bytes per second */
5046 	if (rack->rc_gp_filled == 0) {
5047 		/* Initial measurement */
5048 		if (bytes_ps) {
5049 			rack->r_ctl.gp_bw = bytes_ps;
5050 			rack->rc_gp_filled = 1;
5051 			rack->r_ctl.num_measurements = 1;
5052 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5053 		} else {
5054 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5055 						   rack->r_ctl.rc_app_limited_cnt,
5056 						   0, 0, 10, __LINE__, NULL, quality);
5057 		}
5058 		if (tcp_in_hpts(rack->rc_tp) &&
5059 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5060 			/*
5061 			 * Ok we can't trust the pacer in this case
5062 			 * where we transition from un-paced to paced.
5063 			 * Or for that matter when the burst mitigation
5064 			 * was making a wild guess and got it wrong.
5065 			 * Stop the pacer and clear up all the aggregate
5066 			 * delays etc.
5067 			 */
5068 			tcp_hpts_remove(rack->rc_tp);
5069 			rack->r_ctl.rc_hpts_flags = 0;
5070 			rack->r_ctl.rc_last_output_to = 0;
5071 		}
5072 		did_add = 2;
5073 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5074 		/* Still a small number run an average */
5075 		rack->r_ctl.gp_bw += bytes_ps;
5076 		addpart = rack->r_ctl.num_measurements;
5077 		rack->r_ctl.num_measurements++;
5078 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5079 			/* We have collected enough to move forward */
5080 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5081 		}
5082 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5083 		did_add = 3;
5084 	} else {
5085 		/*
5086 		 * We want to take 1/wma of the goodput and add in to 7/8th
5087 		 * of the old value weighted by the srtt. So if your measurement
5088 		 * period is say 2 SRTT's long you would get 1/4 as the
5089 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5090 		 *
5091 		 * But we must be careful not to take too much i.e. if the
5092 		 * srtt is say 20ms and the measurement is taken over
5093 		 * 400ms our weight would be 400/20 i.e. 20. On the
5094 		 * other hand if we get a measurement over 1ms with a
5095 		 * 10ms rtt we only want to take a much smaller portion.
5096 		 */
5097 		uint8_t meth;
5098 
5099 		if (rack->r_ctl.num_measurements < 0xff) {
5100 			rack->r_ctl.num_measurements++;
5101 		}
5102 		srtt = (uint64_t)tp->t_srtt;
5103 		if (srtt == 0) {
5104 			/*
5105 			 * Strange why did t_srtt go back to zero?
5106 			 */
5107 			if (rack->r_ctl.rc_rack_min_rtt)
5108 				srtt = rack->r_ctl.rc_rack_min_rtt;
5109 			else
5110 				srtt = HPTS_USEC_IN_MSEC;
5111 		}
5112 		/*
5113 		 * XXXrrs: Note for reviewers, in playing with
5114 		 * dynamic pacing I discovered this GP calculation
5115 		 * as done originally leads to some undesired results.
5116 		 * Basically you can get longer measurements contributing
5117 		 * too much to the WMA. Thus I changed it if you are doing
5118 		 * dynamic adjustments to only do the aportioned adjustment
5119 		 * if we have a very small (time wise) measurement. Longer
5120 		 * measurements just get there weight (defaulting to 1/8)
5121 		 * add to the WMA. We may want to think about changing
5122 		 * this to always do that for both sides i.e. dynamic
5123 		 * and non-dynamic... but considering lots of folks
5124 		 * were playing with this I did not want to change the
5125 		 * calculation per.se. without your thoughts.. Lawerence?
5126 		 * Peter??
5127 		 */
5128 		if (rack->rc_gp_dyn_mul == 0) {
5129 			subpart = rack->r_ctl.gp_bw * utim;
5130 			subpart /= (srtt * 8);
5131 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5132 				/*
5133 				 * The b/w update takes no more
5134 				 * away then 1/2 our running total
5135 				 * so factor it in.
5136 				 */
5137 				addpart = bytes_ps * utim;
5138 				addpart /= (srtt * 8);
5139 				meth = 1;
5140 			} else {
5141 				/*
5142 				 * Don't allow a single measurement
5143 				 * to account for more than 1/2 of the
5144 				 * WMA. This could happen on a retransmission
5145 				 * where utim becomes huge compared to
5146 				 * srtt (multiple retransmissions when using
5147 				 * the sending rate which factors in all the
5148 				 * transmissions from the first one).
5149 				 */
5150 				subpart = rack->r_ctl.gp_bw / 2;
5151 				addpart = bytes_ps / 2;
5152 				meth = 2;
5153 			}
5154 			rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5155 			resid_bw = rack->r_ctl.gp_bw - subpart;
5156 			rack->r_ctl.gp_bw = resid_bw + addpart;
5157 			did_add = 1;
5158 		} else {
5159 			if ((utim / srtt) <= 1) {
5160 				/*
5161 				 * The b/w update was over a small period
5162 				 * of time. The idea here is to prevent a small
5163 				 * measurement time period from counting
5164 				 * too much. So we scale it based on the
5165 				 * time so it attributes less than 1/rack_wma_divisor
5166 				 * of its measurement.
5167 				 */
5168 				subpart = rack->r_ctl.gp_bw * utim;
5169 				subpart /= (srtt * rack_wma_divisor);
5170 				addpart = bytes_ps * utim;
5171 				addpart /= (srtt * rack_wma_divisor);
5172 				meth = 3;
5173 			} else {
5174 				/*
5175 				 * The scaled measurement was long
5176 				 * enough so lets just add in the
5177 				 * portion of the measurement i.e. 1/rack_wma_divisor
5178 				 */
5179 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5180 				addpart = bytes_ps / rack_wma_divisor;
5181 				meth = 4;
5182 			}
5183 			if ((rack->measure_saw_probe_rtt == 0) ||
5184 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5185 				/*
5186 				 * For probe-rtt we only add it in
5187 				 * if its larger, all others we just
5188 				 * add in.
5189 				 */
5190 				did_add = 1;
5191 				rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5192 				resid_bw = rack->r_ctl.gp_bw - subpart;
5193 				rack->r_ctl.gp_bw = resid_bw + addpart;
5194 			}
5195 		}
5196 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5197 	}
5198 	/*
5199 	 * We only watch the growth of the GP during the initial startup
5200 	 * or first-slowstart that ensues. If we ever needed to watch
5201 	 * growth of gp outside of that period all we need to do is
5202 	 * remove the first clause of this if (rc_initial_ss_comp).
5203 	 */
5204 	if ((rack->rc_initial_ss_comp == 0) &&
5205 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5206 		uint64_t gp_est;
5207 
5208 		gp_est = bytes_ps;
5209 		if (tcp_bblogging_on(rack->rc_tp)) {
5210 			union tcp_log_stackspecific log;
5211 			struct timeval tv;
5212 
5213 			memset(&log, 0, sizeof(log));
5214 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5215 			log.u_bbr.flex1 = rack->r_ctl.current_round;
5216 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5217 			log.u_bbr.delRate = gp_est;
5218 			log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5219 			log.u_bbr.flex8 = 41;
5220 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5221 					    0, &log, false, NULL, __func__, __LINE__,&tv);
5222 		}
5223 		if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5224 		    (rack->r_ctl.last_gpest == 0)) {
5225 			/*
5226 			 * The round we get our measurement averaging going
5227 			 * is the base round so it always is the source point
5228 			 * for when we had our first increment. From there on
5229 			 * we only record the round that had a rise.
5230 			 */
5231 			rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5232 			rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5233 		} else if (gp_est >= rack->r_ctl.last_gpest) {
5234 			/*
5235 			 * Test to see if its gone up enough
5236 			 * to set the round count up to now. Note
5237 			 * that on the seeding of the 4th measurement we
5238 			 */
5239 			gp_est *= 1000;
5240 			gp_est /= rack->r_ctl.last_gpest;
5241 			if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5242 				/*
5243 				 * We went up enough to record the round.
5244 				 */
5245 				if (tcp_bblogging_on(rack->rc_tp)) {
5246 					union tcp_log_stackspecific log;
5247 					struct timeval tv;
5248 
5249 					memset(&log, 0, sizeof(log));
5250 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5251 					log.u_bbr.flex1 = rack->r_ctl.current_round;
5252 					log.u_bbr.flex2 = (uint32_t)gp_est;
5253 					log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5254 					log.u_bbr.delRate = gp_est;
5255 					log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5256 					log.u_bbr.flex8 = 42;
5257 					(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5258 							    0, &log, false, NULL, __func__, __LINE__,&tv);
5259 				}
5260 				rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5261 				if (rack->r_ctl.use_gp_not_last == 1)
5262 					rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5263 				else
5264 					rack->r_ctl.last_gpest = bytes_ps;
5265 			}
5266 		}
5267 	}
5268 	if ((rack->gp_ready == 0) &&
5269 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5270 		/* We have enough measurements now */
5271 		rack->gp_ready = 1;
5272 		if (rack->dgp_on ||
5273 		    rack->rack_hibeta)
5274 			rack_set_cc_pacing(rack);
5275 		if (rack->defer_options)
5276 			rack_apply_deferred_options(rack);
5277 	}
5278 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5279 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5280 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5281 
5282 	if ((rack->measure_saw_probe_rtt == 0) &&
5283 	    rack->rc_gp_rtt_set) {
5284 		if (rack->rc_skip_timely == 0) {
5285 			rack_update_multiplier(rack, timely_says, bytes_ps,
5286 					       rack->r_ctl.rc_gp_srtt,
5287 					       rack->r_ctl.rc_rtt_diff);
5288 		}
5289 	}
5290 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5291 				   rack_get_bw(rack), 3, line, NULL, quality);
5292 	rack_log_pacing_delay_calc(rack,
5293 				   bytes, /* flex2 */
5294 				   tim, /* flex1 */
5295 				   bytes_ps, /* bw_inuse */
5296 				   rack->r_ctl.gp_bw, /* delRate */
5297 				   rack_get_lt_bw(rack), /* rttProp */
5298 				   20, line, NULL, 0);
5299 	/* reset the gp srtt and setup the new prev */
5300 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5301 	/* Record the lost count for the next measurement */
5302 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5303 skip_measurement:
5304 	/*
5305 	 * We restart our diffs based on the gpsrtt in the
5306 	 * measurement window.
5307 	 */
5308 	rack->rc_gp_rtt_set = 0;
5309 	rack->rc_gp_saw_rec = 0;
5310 	rack->rc_gp_saw_ca = 0;
5311 	rack->rc_gp_saw_ss = 0;
5312 	rack->rc_dragged_bottom = 0;
5313 	if (quality == RACK_QUALITY_HIGH) {
5314 		/*
5315 		 * Gput in the stats world is in kbps where bytes_ps is
5316 		 * bytes per second so we do ((x * 8)/ 1000).
5317 		 */
5318 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5319 #ifdef STATS
5320 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5321 					 gput);
5322 		/*
5323 		 * XXXLAS: This is a temporary hack, and should be
5324 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5325 		 * API to deal with chained VOIs.
5326 		 */
5327 		if (tp->t_stats_gput_prev > 0)
5328 			stats_voi_update_abs_s32(tp->t_stats,
5329 						 VOI_TCP_GPUT_ND,
5330 						 ((gput - tp->t_stats_gput_prev) * 100) /
5331 						 tp->t_stats_gput_prev);
5332 #endif
5333 		tp->t_stats_gput_prev = gput;
5334 	}
5335 	tp->t_flags &= ~TF_GPUTINPROG;
5336 	/*
5337 	 * Now are we app limited now and there is space from where we
5338 	 * were to where we want to go?
5339 	 *
5340 	 * We don't do the other case i.e. non-applimited here since
5341 	 * the next send will trigger us picking up the missing data.
5342 	 */
5343 	if (rack->r_ctl.rc_first_appl &&
5344 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5345 	    rack->r_ctl.rc_app_limited_cnt &&
5346 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5347 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5348 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5349 		/*
5350 		 * Yep there is enough outstanding to make a measurement here.
5351 		 */
5352 		struct rack_sendmap *rsm;
5353 
5354 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5355 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5356 		tp->gput_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
5357 		rack->app_limited_needs_set = 0;
5358 		tp->gput_seq = th_ack;
5359 		if (rack->in_probe_rtt)
5360 			rack->measure_saw_probe_rtt = 1;
5361 		else if ((rack->measure_saw_probe_rtt) &&
5362 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5363 			rack->measure_saw_probe_rtt = 0;
5364 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5365 			/* There is a full window to gain info from */
5366 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5367 		} else {
5368 			/* We can only measure up to the applimited point */
5369 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5370 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5371 				/*
5372 				 * We don't have enough to make a measurement.
5373 				 */
5374 				tp->t_flags &= ~TF_GPUTINPROG;
5375 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5376 							   0, 0, 0, 6, __LINE__, NULL, quality);
5377 				return;
5378 			}
5379 		}
5380 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5381 			/*
5382 			 * We will get no more data into the SB
5383 			 * this means we need to have the data available
5384 			 * before we start a measurement.
5385 			 */
5386 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5387 				/* Nope not enough data. */
5388 				return;
5389 			}
5390 		}
5391 		tp->t_flags |= TF_GPUTINPROG;
5392 		/*
5393 		 * Now we need to find the timestamp of the send at tp->gput_seq
5394 		 * for the send based measurement.
5395 		 */
5396 		rack->r_ctl.rc_gp_cumack_ts = 0;
5397 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5398 		if (rsm) {
5399 			/* Ok send-based limit is set */
5400 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5401 				/*
5402 				 * Move back to include the earlier part
5403 				 * so our ack time lines up right (this may
5404 				 * make an overlapping measurement but thats
5405 				 * ok).
5406 				 */
5407 				tp->gput_seq = rsm->r_start;
5408 			}
5409 			if (rsm->r_flags & RACK_ACKED) {
5410 				struct rack_sendmap *nrsm;
5411 
5412 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5413 				tp->gput_seq = rsm->r_end;
5414 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5415 				if (nrsm)
5416 					rsm = nrsm;
5417 				else {
5418 					rack->app_limited_needs_set = 1;
5419 				}
5420 			} else
5421 				rack->app_limited_needs_set = 1;
5422 			/* We always go from the first send */
5423 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5424 		} else {
5425 			/*
5426 			 * If we don't find the rsm due to some
5427 			 * send-limit set the current time, which
5428 			 * basically disables the send-limit.
5429 			 */
5430 			struct timeval tv;
5431 
5432 			microuptime(&tv);
5433 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5434 		}
5435 		rack_tend_gp_marks(tp, rack);
5436 		rack_log_pacing_delay_calc(rack,
5437 					   tp->gput_seq,
5438 					   tp->gput_ack,
5439 					   (uintptr_t)rsm,
5440 					   tp->gput_ts,
5441 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5442 					   9,
5443 					   __LINE__, rsm, quality);
5444 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5445 	} else {
5446 		/*
5447 		 * To make sure proper timestamp merging occurs, we need to clear
5448 		 * all GP marks if we don't start a measurement.
5449 		 */
5450 		rack_clear_gp_marks(tp, rack);
5451 	}
5452 }
5453 
5454 /*
5455  * CC wrapper hook functions
5456  */
5457 static void
5458 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5459     uint16_t type, int32_t post_recovery)
5460 {
5461 	uint32_t prior_cwnd, acked;
5462 	struct tcp_log_buffer *lgb = NULL;
5463 	uint8_t labc_to_use, quality;
5464 
5465 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5466 	tp->t_ccv.nsegs = nsegs;
5467 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5468 	if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5469 		uint32_t max;
5470 
5471 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5472 		if (tp->t_ccv.bytes_this_ack > max) {
5473 			tp->t_ccv.bytes_this_ack = max;
5474 		}
5475 	}
5476 #ifdef STATS
5477 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5478 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5479 #endif
5480 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5481 		/*
5482 		 * We will ack all the data, time to end any
5483 		 * lt_bw_up we have running until something
5484 		 * new is sent. Note we need to use the actual
5485 		 * ack_rcv_time which with pacing may be different.
5486 		 */
5487 		uint64_t tmark;
5488 
5489 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5490 		rack->r_ctl.lt_seq = tp->snd_max;
5491 		tmark = tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time);
5492 		if (tmark >= rack->r_ctl.lt_timemark) {
5493 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5494 		}
5495 		rack->r_ctl.lt_timemark = tmark;
5496 		rack->lt_bw_up = 0;
5497 	}
5498 	quality = RACK_QUALITY_NONE;
5499 	if ((tp->t_flags & TF_GPUTINPROG) &&
5500 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5501 		/* Measure the Goodput */
5502 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5503 	}
5504 	/* Which way our we limited, if not cwnd limited no advance in CA */
5505 	if (tp->snd_cwnd <= tp->snd_wnd)
5506 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5507 	else
5508 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5509 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5510 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5511 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5512 		/* For the setting of a window past use the actual scwnd we are using */
5513 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5514 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5515 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5516 		}
5517 	} else {
5518 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5519 		tp->t_bytes_acked = 0;
5520 	}
5521 	prior_cwnd = tp->snd_cwnd;
5522 	if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5523 	    (rack_client_low_buf && rack->client_bufferlvl &&
5524 	    (rack->client_bufferlvl < rack_client_low_buf)))
5525 		labc_to_use = rack->rc_labc;
5526 	else
5527 		labc_to_use = rack_max_abc_post_recovery;
5528 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5529 		union tcp_log_stackspecific log;
5530 		struct timeval tv;
5531 
5532 		memset(&log, 0, sizeof(log));
5533 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5534 		log.u_bbr.flex1 = th_ack;
5535 		log.u_bbr.flex2 = tp->t_ccv.flags;
5536 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5537 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5538 		log.u_bbr.flex5 = labc_to_use;
5539 		log.u_bbr.flex6 = prior_cwnd;
5540 		log.u_bbr.flex7 = V_tcp_do_newsack;
5541 		log.u_bbr.flex8 = 1;
5542 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5543 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5544 	}
5545 	if (CC_ALGO(tp)->ack_received != NULL) {
5546 		/* XXXLAS: Find a way to live without this */
5547 		tp->t_ccv.curack = th_ack;
5548 		tp->t_ccv.labc = labc_to_use;
5549 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5550 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5551 	}
5552 	if (lgb) {
5553 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5554 	}
5555 	if (rack->r_must_retran) {
5556 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5557 			/*
5558 			 * We now are beyond the rxt point so lets disable
5559 			 * the flag.
5560 			 */
5561 			rack->r_ctl.rc_out_at_rto = 0;
5562 			rack->r_must_retran = 0;
5563 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5564 			/*
5565 			 * Only decrement the rc_out_at_rto if the cwnd advances
5566 			 * at least a whole segment. Otherwise next time the peer
5567 			 * acks, we won't be able to send this generaly happens
5568 			 * when we are in Congestion Avoidance.
5569 			 */
5570 			if (acked <= rack->r_ctl.rc_out_at_rto){
5571 				rack->r_ctl.rc_out_at_rto -= acked;
5572 			} else {
5573 				rack->r_ctl.rc_out_at_rto = 0;
5574 			}
5575 		}
5576 	}
5577 #ifdef STATS
5578 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5579 #endif
5580 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5581 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5582 	}
5583 	if ((rack->rc_initial_ss_comp == 0) &&
5584 	    (tp->snd_cwnd >= tp->snd_ssthresh)) {
5585 		/*
5586 		 * The cwnd has grown beyond ssthresh we have
5587 		 * entered ca and completed our first Slowstart.
5588 		 */
5589 		rack->rc_initial_ss_comp = 1;
5590 	}
5591 }
5592 
5593 static void
5594 tcp_rack_partialack(struct tcpcb *tp)
5595 {
5596 	struct tcp_rack *rack;
5597 
5598 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5599 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5600 	/*
5601 	 * If we are doing PRR and have enough
5602 	 * room to send <or> we are pacing and prr
5603 	 * is disabled we will want to see if we
5604 	 * can send data (by setting r_wanted_output to
5605 	 * true).
5606 	 */
5607 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5608 	    rack->rack_no_prr)
5609 		rack->r_wanted_output = 1;
5610 }
5611 
5612 static void
5613 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
5614 {
5615 	/*
5616 	 * Now exit recovery.
5617 	 */
5618 	EXIT_RECOVERY(tp->t_flags);
5619 }
5620 
5621 static void
5622 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5623 {
5624 	struct tcp_rack *rack;
5625 	uint32_t orig_cwnd;
5626 
5627 	orig_cwnd = tp->snd_cwnd;
5628 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5629 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5630 	/* only alert CC if we alerted when we entered */
5631 	if (CC_ALGO(tp)->post_recovery != NULL) {
5632 		tp->t_ccv.curack = th_ack;
5633 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5634 		if (tp->snd_cwnd < tp->snd_ssthresh) {
5635 			/*
5636 			 * Rack has burst control and pacing
5637 			 * so lets not set this any lower than
5638 			 * snd_ssthresh per RFC-6582 (option 2).
5639 			 */
5640 			tp->snd_cwnd = tp->snd_ssthresh;
5641 		}
5642 	}
5643 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5644 		union tcp_log_stackspecific log;
5645 		struct timeval tv;
5646 
5647 		memset(&log, 0, sizeof(log));
5648 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5649 		log.u_bbr.flex1 = th_ack;
5650 		log.u_bbr.flex2 = tp->t_ccv.flags;
5651 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5652 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5653 		log.u_bbr.flex5 = V_tcp_abc_l_var;
5654 		log.u_bbr.flex6 = orig_cwnd;
5655 		log.u_bbr.flex7 = V_tcp_do_newsack;
5656 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5657 		log.u_bbr.flex8 = 2;
5658 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5659 			       0, &log, false, NULL, __func__, __LINE__, &tv);
5660 	}
5661 	if ((rack->rack_no_prr == 0) &&
5662 	    (rack->no_prr_addback == 0) &&
5663 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
5664 		/*
5665 		 * Suck the next prr cnt back into cwnd, but
5666 		 * only do that if we are not application limited.
5667 		 */
5668 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5669 			/*
5670 			 * We are allowed to add back to the cwnd the amount we did
5671 			 * not get out if:
5672 			 * a) no_prr_addback is off.
5673 			 * b) we are not app limited
5674 			 * c) we are doing prr
5675 			 * <and>
5676 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5677 			 */
5678 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5679 					    rack->r_ctl.rc_prr_sndcnt);
5680 		}
5681 		rack->r_ctl.rc_prr_sndcnt = 0;
5682 		rack_log_to_prr(rack, 1, 0, __LINE__);
5683 	}
5684 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5685 	tp->snd_recover = tp->snd_una;
5686 	if (rack->r_ctl.dsack_persist) {
5687 		rack->r_ctl.dsack_persist--;
5688 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5689 			rack->r_ctl.num_dsack = 0;
5690 		}
5691 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5692 	}
5693 	if (rack->rto_from_rec == 1) {
5694 		rack->rto_from_rec = 0;
5695 		if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
5696 			tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
5697 	}
5698 	rack_exit_recovery(tp, rack, 1);
5699 }
5700 
5701 static void
5702 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5703 {
5704 	struct tcp_rack *rack;
5705 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5706 
5707 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5708 #ifdef STATS
5709 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5710 #endif
5711 	if (IN_RECOVERY(tp->t_flags) == 0) {
5712 		in_rec_at_entry = 0;
5713 		ssthresh_enter = tp->snd_ssthresh;
5714 		cwnd_enter = tp->snd_cwnd;
5715 	} else
5716 		in_rec_at_entry = 1;
5717 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5718 	switch (type) {
5719 	case CC_NDUPACK:
5720 		tp->t_flags &= ~TF_WASFRECOVERY;
5721 		tp->t_flags &= ~TF_WASCRECOVERY;
5722 		if (!IN_FASTRECOVERY(tp->t_flags)) {
5723 			/* Check if this is the end of the initial Start-up i.e. initial slow-start */
5724 			if (rack->rc_initial_ss_comp == 0) {
5725 				/* Yep it is the end of the initial slowstart */
5726 				rack->rc_initial_ss_comp = 1;
5727 			}
5728 			rack->r_ctl.rc_prr_delivered = 0;
5729 			rack->r_ctl.rc_prr_out = 0;
5730 			rack->r_fast_output = 0;
5731 			if (rack->rack_no_prr == 0) {
5732 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5733 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5734 			}
5735 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5736 			tp->snd_recover = tp->snd_max;
5737 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5738 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5739 		}
5740 		break;
5741 	case CC_ECN:
5742 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5743 		    /*
5744 		     * Allow ECN reaction on ACK to CWR, if
5745 		     * that data segment was also CE marked.
5746 		     */
5747 		    SEQ_GEQ(ack, tp->snd_recover)) {
5748 			EXIT_CONGRECOVERY(tp->t_flags);
5749 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5750 			rack->r_fast_output = 0;
5751 			tp->snd_recover = tp->snd_max + 1;
5752 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5753 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5754 		}
5755 		break;
5756 	case CC_RTO:
5757 		tp->t_dupacks = 0;
5758 		tp->t_bytes_acked = 0;
5759 		rack->r_fast_output = 0;
5760 		if (IN_RECOVERY(tp->t_flags))
5761 			rack_exit_recovery(tp, rack, 2);
5762 		orig_cwnd = tp->snd_cwnd;
5763 		rack_log_to_prr(rack, 16, orig_cwnd, line);
5764 		if (CC_ALGO(tp)->cong_signal == NULL) {
5765 			/* TSNH */
5766 			tp->snd_ssthresh = max(2,
5767 			    min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5768 			    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5769 			tp->snd_cwnd = ctf_fixed_maxseg(tp);
5770 		}
5771 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5772 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5773 		break;
5774 	case CC_RTO_ERR:
5775 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5776 		/* RTO was unnecessary, so reset everything. */
5777 		tp->snd_cwnd = tp->snd_cwnd_prev;
5778 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5779 		tp->snd_recover = tp->snd_recover_prev;
5780 		if (tp->t_flags & TF_WASFRECOVERY) {
5781 			ENTER_FASTRECOVERY(tp->t_flags);
5782 			tp->t_flags &= ~TF_WASFRECOVERY;
5783 		}
5784 		if (tp->t_flags & TF_WASCRECOVERY) {
5785 			ENTER_CONGRECOVERY(tp->t_flags);
5786 			tp->t_flags &= ~TF_WASCRECOVERY;
5787 		}
5788 		tp->snd_nxt = tp->snd_max;
5789 		tp->t_badrxtwin = 0;
5790 		break;
5791 	}
5792 	if ((CC_ALGO(tp)->cong_signal != NULL) &&
5793 	    (type != CC_RTO)){
5794 		tp->t_ccv.curack = ack;
5795 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5796 	}
5797 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5798 		rack_log_to_prr(rack, 15, cwnd_enter, line);
5799 		rack->r_ctl.dsack_byte_cnt = 0;
5800 		rack->r_ctl.retran_during_recovery = 0;
5801 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5802 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5803 		rack->r_ent_rec_ns = 1;
5804 	}
5805 }
5806 
5807 static inline void
5808 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5809 {
5810 	uint32_t i_cwnd;
5811 
5812 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5813 
5814 	if (CC_ALGO(tp)->after_idle != NULL)
5815 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
5816 
5817 	if (tp->snd_cwnd == 1)
5818 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
5819 	else
5820 		i_cwnd = rc_init_window(rack);
5821 
5822 	/*
5823 	 * Being idle is no different than the initial window. If the cc
5824 	 * clamps it down below the initial window raise it to the initial
5825 	 * window.
5826 	 */
5827 	if (tp->snd_cwnd < i_cwnd) {
5828 		tp->snd_cwnd = i_cwnd;
5829 	}
5830 }
5831 
5832 /*
5833  * Indicate whether this ack should be delayed.  We can delay the ack if
5834  * following conditions are met:
5835  *	- There is no delayed ack timer in progress.
5836  *	- Our last ack wasn't a 0-sized window. We never want to delay
5837  *	  the ack that opens up a 0-sized window.
5838  *	- LRO wasn't used for this segment. We make sure by checking that the
5839  *	  segment size is not larger than the MSS.
5840  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
5841  *	  connection.
5842  */
5843 #define DELAY_ACK(tp, tlen)			 \
5844 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5845 	((tp->t_flags & TF_DELACK) == 0) &&	 \
5846 	(tlen <= tp->t_maxseg) &&		 \
5847 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5848 
5849 static struct rack_sendmap *
5850 rack_find_lowest_rsm(struct tcp_rack *rack)
5851 {
5852 	struct rack_sendmap *rsm;
5853 
5854 	/*
5855 	 * Walk the time-order transmitted list looking for an rsm that is
5856 	 * not acked. This will be the one that was sent the longest time
5857 	 * ago that is still outstanding.
5858 	 */
5859 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5860 		if (rsm->r_flags & RACK_ACKED) {
5861 			continue;
5862 		}
5863 		goto finish;
5864 	}
5865 finish:
5866 	return (rsm);
5867 }
5868 
5869 static struct rack_sendmap *
5870 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5871 {
5872 	struct rack_sendmap *prsm;
5873 
5874 	/*
5875 	 * Walk the sequence order list backward until we hit and arrive at
5876 	 * the highest seq not acked. In theory when this is called it
5877 	 * should be the last segment (which it was not).
5878 	 */
5879 	prsm = rsm;
5880 
5881 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
5882 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5883 			continue;
5884 		}
5885 		return (prsm);
5886 	}
5887 	return (NULL);
5888 }
5889 
5890 static uint32_t
5891 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
5892 {
5893 	int32_t lro;
5894 	uint32_t thresh;
5895 
5896 	/*
5897 	 * lro is the flag we use to determine if we have seen reordering.
5898 	 * If it gets set we have seen reordering. The reorder logic either
5899 	 * works in one of two ways:
5900 	 *
5901 	 * If reorder-fade is configured, then we track the last time we saw
5902 	 * re-ordering occur. If we reach the point where enough time as
5903 	 * passed we no longer consider reordering as occurring.
5904 	 *
5905 	 * Or if reorder-face is 0, then once we see reordering we consider
5906 	 * the connection to alway be subject to reordering and just set lro
5907 	 * to 1.
5908 	 *
5909 	 * In the end if lro is non-zero we add the extra time for
5910 	 * reordering in.
5911 	 */
5912 	if (srtt == 0)
5913 		srtt = 1;
5914 	if (rack->r_ctl.rc_reorder_ts) {
5915 		if (rack->r_ctl.rc_reorder_fade) {
5916 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5917 				lro = cts - rack->r_ctl.rc_reorder_ts;
5918 				if (lro == 0) {
5919 					/*
5920 					 * No time as passed since the last
5921 					 * reorder, mark it as reordering.
5922 					 */
5923 					lro = 1;
5924 				}
5925 			} else {
5926 				/* Negative time? */
5927 				lro = 0;
5928 			}
5929 			if (lro > rack->r_ctl.rc_reorder_fade) {
5930 				/* Turn off reordering seen too */
5931 				rack->r_ctl.rc_reorder_ts = 0;
5932 				lro = 0;
5933 			}
5934 		} else {
5935 			/* Reodering does not fade */
5936 			lro = 1;
5937 		}
5938 	} else {
5939 		lro = 0;
5940 	}
5941 	if (rack->rc_rack_tmr_std_based == 0) {
5942 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
5943 	} else {
5944 		/* Standards based pkt-delay is 1/4 srtt */
5945 		thresh = srtt +  (srtt >> 2);
5946 	}
5947 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5948 		/* It must be set, if not you get 1/4 rtt */
5949 		if (rack->r_ctl.rc_reorder_shift)
5950 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5951 		else
5952 			thresh += (srtt >> 2);
5953 	}
5954 	if (rack->rc_rack_use_dsack &&
5955 	    lro &&
5956 	    (rack->r_ctl.num_dsack > 0)) {
5957 		/*
5958 		 * We only increase the reordering window if we
5959 		 * have seen reordering <and> we have a DSACK count.
5960 		 */
5961 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5962 		if (log_allowed)
5963 			rack_log_dsack_event(rack, 4, line, srtt, thresh);
5964 	}
5965 	/* SRTT * 2 is the ceiling */
5966 	if (thresh > (srtt * 2)) {
5967 		thresh = srtt * 2;
5968 	}
5969 	/* And we don't want it above the RTO max either */
5970 	if (thresh > rack_rto_max) {
5971 		thresh = rack_rto_max;
5972 	}
5973 	if (log_allowed)
5974 		rack_log_dsack_event(rack, 6, line,  srtt, thresh);
5975 	return (thresh);
5976 }
5977 
5978 static uint32_t
5979 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5980 		     struct rack_sendmap *rsm, uint32_t srtt)
5981 {
5982 	struct rack_sendmap *prsm;
5983 	uint32_t thresh, len;
5984 	int segsiz;
5985 
5986 	if (srtt == 0)
5987 		srtt = 1;
5988 	if (rack->r_ctl.rc_tlp_threshold)
5989 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5990 	else
5991 		thresh = (srtt * 2);
5992 
5993 	/* Get the previous sent packet, if any */
5994 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5995 	len = rsm->r_end - rsm->r_start;
5996 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5997 		/* Exactly like the ID */
5998 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5999 			uint32_t alt_thresh;
6000 			/*
6001 			 * Compensate for delayed-ack with the d-ack time.
6002 			 */
6003 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6004 			if (alt_thresh > thresh)
6005 				thresh = alt_thresh;
6006 		}
6007 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6008 		/* 2.1 behavior */
6009 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6010 		if (prsm && (len <= segsiz)) {
6011 			/*
6012 			 * Two packets outstanding, thresh should be (2*srtt) +
6013 			 * possible inter-packet delay (if any).
6014 			 */
6015 			uint32_t inter_gap = 0;
6016 			int idx, nidx;
6017 
6018 			idx = rsm->r_rtr_cnt - 1;
6019 			nidx = prsm->r_rtr_cnt - 1;
6020 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6021 				/* Yes it was sent later (or at the same time) */
6022 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6023 			}
6024 			thresh += inter_gap;
6025 		} else if (len <= segsiz) {
6026 			/*
6027 			 * Possibly compensate for delayed-ack.
6028 			 */
6029 			uint32_t alt_thresh;
6030 
6031 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6032 			if (alt_thresh > thresh)
6033 				thresh = alt_thresh;
6034 		}
6035 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6036 		/* 2.2 behavior */
6037 		if (len <= segsiz) {
6038 			uint32_t alt_thresh;
6039 			/*
6040 			 * Compensate for delayed-ack with the d-ack time.
6041 			 */
6042 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6043 			if (alt_thresh > thresh)
6044 				thresh = alt_thresh;
6045 		}
6046 	}
6047 	/* Not above an RTO */
6048 	if (thresh > tp->t_rxtcur) {
6049 		thresh = tp->t_rxtcur;
6050 	}
6051 	/* Not above a RTO max */
6052 	if (thresh > rack_rto_max) {
6053 		thresh = rack_rto_max;
6054 	}
6055 	/* Apply user supplied min TLP */
6056 	if (thresh < rack_tlp_min) {
6057 		thresh = rack_tlp_min;
6058 	}
6059 	return (thresh);
6060 }
6061 
6062 static uint32_t
6063 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6064 {
6065 	/*
6066 	 * We want the rack_rtt which is the
6067 	 * last rtt we measured. However if that
6068 	 * does not exist we fallback to the srtt (which
6069 	 * we probably will never do) and then as a last
6070 	 * resort we use RACK_INITIAL_RTO if no srtt is
6071 	 * yet set.
6072 	 */
6073 	if (rack->rc_rack_rtt)
6074 		return (rack->rc_rack_rtt);
6075 	else if (tp->t_srtt == 0)
6076 		return (RACK_INITIAL_RTO);
6077 	return (tp->t_srtt);
6078 }
6079 
6080 static struct rack_sendmap *
6081 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6082 {
6083 	/*
6084 	 * Check to see that we don't need to fall into recovery. We will
6085 	 * need to do so if our oldest transmit is past the time we should
6086 	 * have had an ack.
6087 	 */
6088 	struct tcp_rack *rack;
6089 	struct rack_sendmap *rsm;
6090 	int32_t idx;
6091 	uint32_t srtt, thresh;
6092 
6093 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6094 	if (tqhash_empty(rack->r_ctl.tqh)) {
6095 		return (NULL);
6096 	}
6097 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6098 	if (rsm == NULL)
6099 		return (NULL);
6100 
6101 
6102 	if (rsm->r_flags & RACK_ACKED) {
6103 		rsm = rack_find_lowest_rsm(rack);
6104 		if (rsm == NULL)
6105 			return (NULL);
6106 	}
6107 	idx = rsm->r_rtr_cnt - 1;
6108 	srtt = rack_grab_rtt(tp, rack);
6109 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6110 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6111 		return (NULL);
6112 	}
6113 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6114 		return (NULL);
6115 	}
6116 	/* Ok if we reach here we are over-due and this guy can be sent */
6117 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6118 	return (rsm);
6119 }
6120 
6121 static uint32_t
6122 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6123 {
6124 	int32_t t;
6125 	int32_t tt;
6126 	uint32_t ret_val;
6127 
6128 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6129 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6130  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6131 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6132 	ret_val = (uint32_t)tt;
6133 	return (ret_val);
6134 }
6135 
6136 static uint32_t
6137 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6138 {
6139 	/*
6140 	 * Start the FR timer, we do this based on getting the first one in
6141 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6142 	 * events we need to stop the running timer (if its running) before
6143 	 * starting the new one.
6144 	 */
6145 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6146 	uint32_t srtt_cur;
6147 	int32_t idx;
6148 	int32_t is_tlp_timer = 0;
6149 	struct rack_sendmap *rsm;
6150 
6151 	if (rack->t_timers_stopped) {
6152 		/* All timers have been stopped none are to run */
6153 		return (0);
6154 	}
6155 	if (rack->rc_in_persist) {
6156 		/* We can't start any timer in persists */
6157 		return (rack_get_persists_timer_val(tp, rack));
6158 	}
6159 	rack->rc_on_min_to = 0;
6160 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6161 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6162 		goto activate_rxt;
6163 	}
6164 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6165 	if ((rsm == NULL) || sup_rack) {
6166 		/* Nothing on the send map or no rack */
6167 activate_rxt:
6168 		time_since_sent = 0;
6169 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6170 		if (rsm) {
6171 			/*
6172 			 * Should we discount the RTX timer any?
6173 			 *
6174 			 * We want to discount it the smallest amount.
6175 			 * If a timer (Rack/TLP or RXT) has gone off more
6176 			 * recently thats the discount we want to use (now - timer time).
6177 			 * If the retransmit of the oldest packet was more recent then
6178 			 * we want to use that (now - oldest-packet-last_transmit_time).
6179 			 *
6180 			 */
6181 			idx = rsm->r_rtr_cnt - 1;
6182 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6183 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6184 			else
6185 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6186 			if (TSTMP_GT(cts, tstmp_touse))
6187 			    time_since_sent = cts - tstmp_touse;
6188 		}
6189 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6190 		    sbavail(&tptosocket(tp)->so_snd)) {
6191 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6192 			to = tp->t_rxtcur;
6193 			if (to > time_since_sent)
6194 				to -= time_since_sent;
6195 			else
6196 				to = rack->r_ctl.rc_min_to;
6197 			if (to == 0)
6198 				to = 1;
6199 			/* Special case for KEEPINIT */
6200 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6201 			    (TP_KEEPINIT(tp) != 0) &&
6202 			    rsm) {
6203 				/*
6204 				 * We have to put a ceiling on the rxt timer
6205 				 * of the keep-init timeout.
6206 				 */
6207 				uint32_t max_time, red;
6208 
6209 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6210 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6211 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6212 					if (red < max_time)
6213 						max_time -= red;
6214 					else
6215 						max_time = 1;
6216 				}
6217 				/* Reduce timeout to the keep value if needed */
6218 				if (max_time < to)
6219 					to = max_time;
6220 			}
6221 			return (to);
6222 		}
6223 		return (0);
6224 	}
6225 	if (rsm->r_flags & RACK_ACKED) {
6226 		rsm = rack_find_lowest_rsm(rack);
6227 		if (rsm == NULL) {
6228 			/* No lowest? */
6229 			goto activate_rxt;
6230 		}
6231 	}
6232 	/* Convert from ms to usecs */
6233 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6234 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6235 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6236 		if ((tp->t_flags & TF_SENTFIN) &&
6237 		    ((tp->snd_max - tp->snd_una) == 1) &&
6238 		    (rsm->r_flags & RACK_HAS_FIN)) {
6239 			/*
6240 			 * We don't start a rack timer if all we have is a
6241 			 * FIN outstanding.
6242 			 */
6243 			goto activate_rxt;
6244 		}
6245 		if ((rack->use_rack_rr == 0) &&
6246 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6247 		    (rack->rack_no_prr == 0) &&
6248 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6249 			/*
6250 			 * We are not cheating, in recovery  and
6251 			 * not enough ack's to yet get our next
6252 			 * retransmission out.
6253 			 *
6254 			 * Note that classified attackers do not
6255 			 * get to use the rack-cheat.
6256 			 */
6257 			goto activate_tlp;
6258 		}
6259 		srtt = rack_grab_rtt(tp, rack);
6260 		thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6261 		idx = rsm->r_rtr_cnt - 1;
6262 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6263 		if (SEQ_GEQ(exp, cts)) {
6264 			to = exp - cts;
6265 			if (to < rack->r_ctl.rc_min_to) {
6266 				to = rack->r_ctl.rc_min_to;
6267 				if (rack->r_rr_config == 3)
6268 					rack->rc_on_min_to = 1;
6269 			}
6270 		} else {
6271 			to = rack->r_ctl.rc_min_to;
6272 			if (rack->r_rr_config == 3)
6273 				rack->rc_on_min_to = 1;
6274 		}
6275 	} else {
6276 		/* Ok we need to do a TLP not RACK */
6277 activate_tlp:
6278 		if ((rack->rc_tlp_in_progress != 0) &&
6279 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6280 			/*
6281 			 * The previous send was a TLP and we have sent
6282 			 * N TLP's without sending new data.
6283 			 */
6284 			goto activate_rxt;
6285 		}
6286 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6287 		if (rsm == NULL) {
6288 			/* We found no rsm to TLP with. */
6289 			goto activate_rxt;
6290 		}
6291 		if (rsm->r_flags & RACK_HAS_FIN) {
6292 			/* If its a FIN we dont do TLP */
6293 			rsm = NULL;
6294 			goto activate_rxt;
6295 		}
6296 		idx = rsm->r_rtr_cnt - 1;
6297 		time_since_sent = 0;
6298 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6299 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6300 		else
6301 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6302 		if (TSTMP_GT(cts, tstmp_touse))
6303 		    time_since_sent = cts - tstmp_touse;
6304 		is_tlp_timer = 1;
6305 		if (tp->t_srtt) {
6306 			if ((rack->rc_srtt_measure_made == 0) &&
6307 			    (tp->t_srtt == 1)) {
6308 				/*
6309 				 * If another stack as run and set srtt to 1,
6310 				 * then the srtt was 0, so lets use the initial.
6311 				 */
6312 				srtt = RACK_INITIAL_RTO;
6313 			} else {
6314 				srtt_cur = tp->t_srtt;
6315 				srtt = srtt_cur;
6316 			}
6317 		} else
6318 			srtt = RACK_INITIAL_RTO;
6319 		/*
6320 		 * If the SRTT is not keeping up and the
6321 		 * rack RTT has spiked we want to use
6322 		 * the last RTT not the smoothed one.
6323 		 */
6324 		if (rack_tlp_use_greater &&
6325 		    tp->t_srtt &&
6326 		    (srtt < rack_grab_rtt(tp, rack))) {
6327 			srtt = rack_grab_rtt(tp, rack);
6328 		}
6329 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6330 		if (thresh > time_since_sent) {
6331 			to = thresh - time_since_sent;
6332 		} else {
6333 			to = rack->r_ctl.rc_min_to;
6334 			rack_log_alt_to_to_cancel(rack,
6335 						  thresh,		/* flex1 */
6336 						  time_since_sent,	/* flex2 */
6337 						  tstmp_touse,		/* flex3 */
6338 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6339 						  (uint32_t)rsm->r_tim_lastsent[idx],
6340 						  srtt,
6341 						  idx, 99);
6342 		}
6343 		if (to < rack_tlp_min) {
6344 			to = rack_tlp_min;
6345 		}
6346 		if (to > TICKS_2_USEC(tcp_rexmit_max)) {
6347 			/*
6348 			 * If the TLP time works out to larger than the max
6349 			 * RTO lets not do TLP.. just RTO.
6350 			 */
6351 			goto activate_rxt;
6352 		}
6353 	}
6354 	if (is_tlp_timer == 0) {
6355 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6356 	} else {
6357 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6358 	}
6359 	if (to == 0)
6360 		to = 1;
6361 	return (to);
6362 }
6363 
6364 static void
6365 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6366 {
6367 	if (rack->rc_in_persist == 0) {
6368 		if (tp->t_flags & TF_GPUTINPROG) {
6369 			/*
6370 			 * Stop the goodput now, the calling of the
6371 			 * measurement function clears the flag.
6372 			 */
6373 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6374 						    RACK_QUALITY_PERSIST);
6375 		}
6376 #ifdef NETFLIX_SHARED_CWND
6377 		if (rack->r_ctl.rc_scw) {
6378 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6379 			rack->rack_scwnd_is_idle = 1;
6380 		}
6381 #endif
6382 		rack->r_ctl.rc_went_idle_time = cts;
6383 		if (rack->r_ctl.rc_went_idle_time == 0)
6384 			rack->r_ctl.rc_went_idle_time = 1;
6385 		if (rack->lt_bw_up) {
6386 			/* Suspend our LT BW measurement */
6387 			uint64_t tmark;
6388 
6389 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6390 			rack->r_ctl.lt_seq = snd_una;
6391 			tmark = tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time);
6392 			if (tmark >= rack->r_ctl.lt_timemark) {
6393 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6394 			}
6395 			rack->r_ctl.lt_timemark = tmark;
6396 			rack->lt_bw_up = 0;
6397 			rack->r_persist_lt_bw_off = 1;
6398 		}
6399 		rack_timer_cancel(tp, rack, cts, __LINE__);
6400 		rack->r_ctl.persist_lost_ends = 0;
6401 		rack->probe_not_answered = 0;
6402 		rack->forced_ack = 0;
6403 		tp->t_rxtshift = 0;
6404 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6405 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6406 		rack->rc_in_persist = 1;
6407 	}
6408 }
6409 
6410 static void
6411 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6412 {
6413 	if (tcp_in_hpts(rack->rc_tp)) {
6414 		tcp_hpts_remove(rack->rc_tp);
6415 		rack->r_ctl.rc_hpts_flags = 0;
6416 	}
6417 #ifdef NETFLIX_SHARED_CWND
6418 	if (rack->r_ctl.rc_scw) {
6419 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6420 		rack->rack_scwnd_is_idle = 0;
6421 	}
6422 #endif
6423 	if (rack->rc_gp_dyn_mul &&
6424 	    (rack->use_fixed_rate == 0) &&
6425 	    (rack->rc_always_pace)) {
6426 		/*
6427 		 * Do we count this as if a probe-rtt just
6428 		 * finished?
6429 		 */
6430 		uint32_t time_idle, idle_min;
6431 
6432 		time_idle = cts - rack->r_ctl.rc_went_idle_time;
6433 		idle_min = rack_min_probertt_hold;
6434 		if (rack_probertt_gpsrtt_cnt_div) {
6435 			uint64_t extra;
6436 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6437 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
6438 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6439 			idle_min += (uint32_t)extra;
6440 		}
6441 		if (time_idle >= idle_min) {
6442 			/* Yes, we count it as a probe-rtt. */
6443 			uint32_t us_cts;
6444 
6445 			us_cts = tcp_get_usecs(NULL);
6446 			if (rack->in_probe_rtt == 0) {
6447 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6448 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6449 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6450 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6451 			} else {
6452 				rack_exit_probertt(rack, us_cts);
6453 			}
6454 		}
6455 	}
6456 	if (rack->r_persist_lt_bw_off) {
6457 		/* Continue where we left off */
6458 		rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
6459 		rack->lt_bw_up = 1;
6460 		rack->r_persist_lt_bw_off = 0;
6461 	}
6462 	rack->rc_in_persist = 0;
6463 	rack->r_ctl.rc_went_idle_time = 0;
6464 	tp->t_rxtshift = 0;
6465 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6466 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6467 	rack->r_ctl.rc_agg_delayed = 0;
6468 	rack->r_early = 0;
6469 	rack->r_late = 0;
6470 	rack->r_ctl.rc_agg_early = 0;
6471 }
6472 
6473 static void
6474 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6475 		   struct hpts_diag *diag, struct timeval *tv)
6476 {
6477 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6478 		union tcp_log_stackspecific log;
6479 
6480 		memset(&log, 0, sizeof(log));
6481 		log.u_bbr.flex1 = diag->p_nxt_slot;
6482 		log.u_bbr.flex2 = diag->p_cur_slot;
6483 		log.u_bbr.flex3 = diag->slot_req;
6484 		log.u_bbr.flex4 = diag->inp_hptsslot;
6485 		log.u_bbr.flex5 = diag->slot_remaining;
6486 		log.u_bbr.flex6 = diag->need_new_to;
6487 		log.u_bbr.flex7 = diag->p_hpts_active;
6488 		log.u_bbr.flex8 = diag->p_on_min_sleep;
6489 		/* Hijack other fields as needed */
6490 		log.u_bbr.epoch = diag->have_slept;
6491 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
6492 		log.u_bbr.pkts_out = diag->co_ret;
6493 		log.u_bbr.applimited = diag->hpts_sleep_time;
6494 		log.u_bbr.delivered = diag->p_prev_slot;
6495 		log.u_bbr.inflight = diag->p_runningslot;
6496 		log.u_bbr.bw_inuse = diag->wheel_slot;
6497 		log.u_bbr.rttProp = diag->wheel_cts;
6498 		log.u_bbr.timeStamp = cts;
6499 		log.u_bbr.delRate = diag->maxslots;
6500 		log.u_bbr.cur_del_rate = diag->p_curtick;
6501 		log.u_bbr.cur_del_rate <<= 32;
6502 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
6503 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6504 		    &rack->rc_inp->inp_socket->so_rcv,
6505 		    &rack->rc_inp->inp_socket->so_snd,
6506 		    BBR_LOG_HPTSDIAG, 0,
6507 		    0, &log, false, tv);
6508 	}
6509 
6510 }
6511 
6512 static void
6513 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6514 {
6515 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6516 		union tcp_log_stackspecific log;
6517 		struct timeval tv;
6518 
6519 		memset(&log, 0, sizeof(log));
6520 		log.u_bbr.flex1 = sb->sb_flags;
6521 		log.u_bbr.flex2 = len;
6522 		log.u_bbr.flex3 = sb->sb_state;
6523 		log.u_bbr.flex8 = type;
6524 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6525 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6526 		    &rack->rc_inp->inp_socket->so_rcv,
6527 		    &rack->rc_inp->inp_socket->so_snd,
6528 		    TCP_LOG_SB_WAKE, 0,
6529 		    len, &log, false, &tv);
6530 	}
6531 }
6532 
6533 static void
6534 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6535       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6536 {
6537 	struct hpts_diag diag;
6538 	struct inpcb *inp = tptoinpcb(tp);
6539 	struct timeval tv;
6540 	uint32_t delayed_ack = 0;
6541 	uint32_t hpts_timeout;
6542 	uint32_t entry_slot = slot;
6543 	uint8_t stopped;
6544 	uint32_t left = 0;
6545 	uint32_t us_cts;
6546 
6547 	if ((tp->t_state == TCPS_CLOSED) ||
6548 	    (tp->t_state == TCPS_LISTEN)) {
6549 		return;
6550 	}
6551 	if (tcp_in_hpts(tp)) {
6552 		/* Already on the pacer */
6553 		return;
6554 	}
6555 	stopped = rack->rc_tmr_stopped;
6556 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6557 		left = rack->r_ctl.rc_timer_exp - cts;
6558 	}
6559 	rack->r_ctl.rc_timer_exp = 0;
6560 	rack->r_ctl.rc_hpts_flags = 0;
6561 	us_cts = tcp_get_usecs(&tv);
6562 	/* Now early/late accounting */
6563 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6564 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6565 		/*
6566 		 * We have a early carry over set,
6567 		 * we can always add more time so we
6568 		 * can always make this compensation.
6569 		 *
6570 		 * Note if ack's are allowed to wake us do not
6571 		 * penalize the next timer for being awoke
6572 		 * by an ack aka the rc_agg_early (non-paced mode).
6573 		 */
6574 		slot += rack->r_ctl.rc_agg_early;
6575 		rack->r_early = 0;
6576 		rack->r_ctl.rc_agg_early = 0;
6577 	}
6578 	if ((rack->r_late) &&
6579 	    ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
6580 		/*
6581 		 * This is harder, we can
6582 		 * compensate some but it
6583 		 * really depends on what
6584 		 * the current pacing time is.
6585 		 */
6586 		if (rack->r_ctl.rc_agg_delayed >= slot) {
6587 			/*
6588 			 * We can't compensate for it all.
6589 			 * And we have to have some time
6590 			 * on the clock. We always have a min
6591 			 * 10 slots (10 x 10 i.e. 100 usecs).
6592 			 */
6593 			if (slot <= HPTS_USECS_PER_SLOT) {
6594 				/* We gain delay */
6595 				rack->r_ctl.rc_agg_delayed += (HPTS_USECS_PER_SLOT - slot);
6596 				slot = HPTS_USECS_PER_SLOT;
6597 			} else {
6598 				/* We take off some */
6599 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_USECS_PER_SLOT);
6600 				slot = HPTS_USECS_PER_SLOT;
6601 			}
6602 		} else {
6603 			slot -= rack->r_ctl.rc_agg_delayed;
6604 			rack->r_ctl.rc_agg_delayed = 0;
6605 			/* Make sure we have 100 useconds at minimum */
6606 			if (slot < HPTS_USECS_PER_SLOT) {
6607 				rack->r_ctl.rc_agg_delayed = HPTS_USECS_PER_SLOT - slot;
6608 				slot = HPTS_USECS_PER_SLOT;
6609 			}
6610 			if (rack->r_ctl.rc_agg_delayed == 0)
6611 				rack->r_late = 0;
6612 		}
6613 	} else if (rack->r_late) {
6614 		/* r_use_hpts_min is on and so is DGP */
6615 		uint32_t max_red;
6616 
6617 		max_red = (slot * rack->r_ctl.max_reduction) / 100;
6618 		if (max_red >= rack->r_ctl.rc_agg_delayed) {
6619 			slot -= rack->r_ctl.rc_agg_delayed;
6620 			rack->r_ctl.rc_agg_delayed = 0;
6621 		} else {
6622 			slot -= max_red;
6623 			rack->r_ctl.rc_agg_delayed -= max_red;
6624 		}
6625 	}
6626 	if ((rack->r_use_hpts_min == 1) &&
6627 	    (slot > 0) &&
6628 	    (rack->dgp_on == 1)) {
6629 		/*
6630 		 * We are enforcing a min pacing timer
6631 		 * based on our hpts min timeout.
6632 		 */
6633 		uint32_t min;
6634 
6635 		min = get_hpts_min_sleep_time();
6636 		if (min > slot) {
6637 			slot = min;
6638 		}
6639 	}
6640 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6641 	if (tp->t_flags & TF_DELACK) {
6642 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
6643 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6644 	}
6645 	if (delayed_ack && ((hpts_timeout == 0) ||
6646 			    (delayed_ack < hpts_timeout)))
6647 		hpts_timeout = delayed_ack;
6648 	else
6649 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6650 	/*
6651 	 * If no timers are going to run and we will fall off the hptsi
6652 	 * wheel, we resort to a keep-alive timer if its configured.
6653 	 */
6654 	if ((hpts_timeout == 0) &&
6655 	    (slot == 0)) {
6656 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6657 		    (tp->t_state <= TCPS_CLOSING)) {
6658 			/*
6659 			 * Ok we have no timer (persists, rack, tlp, rxt  or
6660 			 * del-ack), we don't have segments being paced. So
6661 			 * all that is left is the keepalive timer.
6662 			 */
6663 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6664 				/* Get the established keep-alive time */
6665 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6666 			} else {
6667 				/*
6668 				 * Get the initial setup keep-alive time,
6669 				 * note that this is probably not going to
6670 				 * happen, since rack will be running a rxt timer
6671 				 * if a SYN of some sort is outstanding. It is
6672 				 * actually handled in rack_timeout_rxt().
6673 				 */
6674 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6675 			}
6676 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6677 			if (rack->in_probe_rtt) {
6678 				/*
6679 				 * We want to instead not wake up a long time from
6680 				 * now but to wake up about the time we would
6681 				 * exit probe-rtt and initiate a keep-alive ack.
6682 				 * This will get us out of probe-rtt and update
6683 				 * our min-rtt.
6684 				 */
6685 				hpts_timeout = rack_min_probertt_hold;
6686 			}
6687 		}
6688 	}
6689 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6690 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6691 		/*
6692 		 * RACK, TLP, persists and RXT timers all are restartable
6693 		 * based on actions input .. i.e we received a packet (ack
6694 		 * or sack) and that changes things (rw, or snd_una etc).
6695 		 * Thus we can restart them with a new value. For
6696 		 * keep-alive, delayed_ack we keep track of what was left
6697 		 * and restart the timer with a smaller value.
6698 		 */
6699 		if (left < hpts_timeout)
6700 			hpts_timeout = left;
6701 	}
6702 	if (hpts_timeout) {
6703 		/*
6704 		 * Hack alert for now we can't time-out over 2,147,483
6705 		 * seconds (a bit more than 596 hours), which is probably ok
6706 		 * :).
6707 		 */
6708 		if (hpts_timeout > 0x7ffffffe)
6709 			hpts_timeout = 0x7ffffffe;
6710 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6711 	}
6712 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6713 	if ((rack->gp_ready == 0) &&
6714 	    (rack->use_fixed_rate == 0) &&
6715 	    (hpts_timeout < slot) &&
6716 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6717 		/*
6718 		 * We have no good estimate yet for the
6719 		 * old clunky burst mitigation or the
6720 		 * real pacing. And the tlp or rxt is smaller
6721 		 * than the pacing calculation. Lets not
6722 		 * pace that long since we know the calculation
6723 		 * so far is not accurate.
6724 		 */
6725 		slot = hpts_timeout;
6726 	}
6727 	/**
6728 	 * Turn off all the flags for queuing by default. The
6729 	 * flags have important meanings to what happens when
6730 	 * LRO interacts with the transport. Most likely (by default now)
6731 	 * mbuf_queueing and ack compression are on. So the transport
6732 	 * has a couple of flags that control what happens (if those
6733 	 * are not on then these flags won't have any effect since it
6734 	 * won't go through the queuing LRO path).
6735 	 *
6736 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6737 	 *                        pacing output, so don't disturb. But
6738 	 *                        it also means LRO can wake me if there
6739 	 *                        is a SACK arrival.
6740 	 *
6741 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6742 	 *                       with the above flag (QUEUE_READY) and
6743 	 *                       when present it says don't even wake me
6744 	 *                       if a SACK arrives.
6745 	 *
6746 	 * The idea behind these flags is that if we are pacing we
6747 	 * set the MBUF_QUEUE_READY and only get woken up if
6748 	 * a SACK arrives (which could change things) or if
6749 	 * our pacing timer expires. If, however, we have a rack
6750 	 * timer running, then we don't even want a sack to wake
6751 	 * us since the rack timer has to expire before we can send.
6752 	 *
6753 	 * Other cases should usually have none of the flags set
6754 	 * so LRO can call into us.
6755 	 */
6756 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6757 	if (slot) {
6758 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6759 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6760 		/*
6761 		 * A pacing timer (slot) is being set, in
6762 		 * such a case we cannot send (we are blocked by
6763 		 * the timer). So lets tell LRO that it should not
6764 		 * wake us unless there is a SACK. Note this only
6765 		 * will be effective if mbuf queueing is on or
6766 		 * compressed acks are being processed.
6767 		 */
6768 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6769 		/*
6770 		 * But wait if we have a Rack timer running
6771 		 * even a SACK should not disturb us (with
6772 		 * the exception of r_rr_config 3).
6773 		 */
6774 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6775 		    (IN_RECOVERY(tp->t_flags))) {
6776 			if (rack->r_rr_config != 3)
6777 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6778 			else if (rack->rc_pace_dnd) {
6779 				/*
6780 				 * When DND is on, we only let a sack
6781 				 * interrupt us if we are not in recovery.
6782 				 *
6783 				 * If DND is off, then we never hit here
6784 				 * and let all sacks wake us up.
6785 				 *
6786 				 */
6787 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6788 			}
6789 		}
6790 		if (rack->rc_ack_can_sendout_data) {
6791 			/*
6792 			 * Ahh but wait, this is that special case
6793 			 * where the pacing timer can be disturbed
6794 			 * backout the changes (used for non-paced
6795 			 * burst limiting).
6796 			 */
6797 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6798 			    TF2_MBUF_QUEUE_READY);
6799 		}
6800 		if ((rack->use_rack_rr) &&
6801 		    (rack->r_rr_config < 2) &&
6802 		    ((hpts_timeout) && (hpts_timeout < slot))) {
6803 			/*
6804 			 * Arrange for the hpts to kick back in after the
6805 			 * t-o if the t-o does not cause a send.
6806 			 */
6807 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6808 						   __LINE__, &diag);
6809 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6810 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6811 		} else {
6812 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
6813 						   __LINE__, &diag);
6814 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6815 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6816 		}
6817 	} else if (hpts_timeout) {
6818 		/*
6819 		 * With respect to t_flags2(?) here, lets let any new acks wake
6820 		 * us up here. Since we are not pacing (no pacing timer), output
6821 		 * can happen so we should let it. If its a Rack timer, then any inbound
6822 		 * packet probably won't change the sending (we will be blocked)
6823 		 * but it may change the prr stats so letting it in (the set defaults
6824 		 * at the start of this block) are good enough.
6825 		 */
6826 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6827 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6828 					   __LINE__, &diag);
6829 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6830 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6831 	} else {
6832 		/* No timer starting */
6833 #ifdef INVARIANTS
6834 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6835 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6836 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6837 		}
6838 #endif
6839 	}
6840 	rack->rc_tmr_stopped = 0;
6841 	if (slot)
6842 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6843 }
6844 
6845 static void
6846 rack_mark_lost(struct tcpcb *tp,
6847     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
6848 {
6849 	struct rack_sendmap *nrsm;
6850 	uint32_t thresh,  exp;
6851 
6852 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
6853 	nrsm = rsm;
6854 	TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
6855 		if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
6856 			/* Got up to all that were marked sack-passed */
6857 			break;
6858 		}
6859 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
6860 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
6861 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
6862 				/* We now consider it lost */
6863 				nrsm->r_flags |= RACK_WAS_LOST;
6864 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
6865 			} else {
6866 				/* Past here it won't be lost so stop */
6867 				break;
6868 			}
6869 		}
6870 	}
6871 }
6872 
6873 /*
6874  * RACK Timer, here we simply do logging and house keeping.
6875  * the normal rack_output() function will call the
6876  * appropriate thing to check if we need to do a RACK retransmit.
6877  * We return 1, saying don't proceed with rack_output only
6878  * when all timers have been stopped (destroyed PCB?).
6879  */
6880 static int
6881 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6882 {
6883 	/*
6884 	 * This timer simply provides an internal trigger to send out data.
6885 	 * The check_recovery_mode call will see if there are needed
6886 	 * retransmissions, if so we will enter fast-recovery. The output
6887 	 * call may or may not do the same thing depending on sysctl
6888 	 * settings.
6889 	 */
6890 	struct rack_sendmap *rsm;
6891 
6892 	counter_u64_add(rack_to_tot, 1);
6893 	if (rack->r_state && (rack->r_state != tp->t_state))
6894 		rack_set_state(tp, rack);
6895 	rack->rc_on_min_to = 0;
6896 	rsm = rack_check_recovery_mode(tp, cts);
6897 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6898 	if (rsm) {
6899 		/* We need to stroke any lost that are now declared as lost */
6900 		rack_mark_lost(tp, rack, rsm, cts);
6901 		rack->r_ctl.rc_resend = rsm;
6902 		rack->r_timer_override = 1;
6903 		if (rack->use_rack_rr) {
6904 			/*
6905 			 * Don't accumulate extra pacing delay
6906 			 * we are allowing the rack timer to
6907 			 * over-ride pacing i.e. rrr takes precedence
6908 			 * if the pacing interval is longer than the rrr
6909 			 * time (in other words we get the min pacing
6910 			 * time versus rrr pacing time).
6911 			 */
6912 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6913 		}
6914 	}
6915 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6916 	if (rsm == NULL) {
6917 		/* restart a timer and return 1 */
6918 		rack_start_hpts_timer(rack, tp, cts,
6919 				      0, 0, 0);
6920 		return (1);
6921 	}
6922 	return (0);
6923 }
6924 
6925 
6926 
6927 static void
6928 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6929 {
6930 
6931 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6932 		/*
6933 		 * The trailing space changed, mbufs can grow
6934 		 * at the tail but they can't shrink from
6935 		 * it, KASSERT that. Adjust the orig_m_len to
6936 		 * compensate for this change.
6937 		 */
6938 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
6939 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
6940 			 rsm->m,
6941 			 rsm,
6942 			 (intmax_t)M_TRAILINGROOM(rsm->m),
6943 			 rsm->orig_t_space,
6944 			 rsm->orig_m_len,
6945 			 rsm->m->m_len));
6946 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
6947 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
6948 	}
6949 	if (rsm->m->m_len < rsm->orig_m_len) {
6950 		/*
6951 		 * Mbuf shrank, trimmed off the top by an ack, our
6952 		 * offset changes.
6953 		 */
6954 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
6955 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
6956 			 rsm->m, rsm->m->m_len,
6957 			 rsm, rsm->orig_m_len,
6958 			 rsm->soff));
6959 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
6960 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6961 		else
6962 			rsm->soff = 0;
6963 		rsm->orig_m_len = rsm->m->m_len;
6964 #ifdef INVARIANTS
6965 	} else if (rsm->m->m_len > rsm->orig_m_len) {
6966 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
6967 		      rsm, rsm->m);
6968 #endif
6969 	}
6970 }
6971 
6972 static void
6973 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6974 {
6975 	struct mbuf *m;
6976 	uint32_t soff;
6977 
6978 	if (src_rsm->m &&
6979 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
6980 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
6981 		/* Fix up the orig_m_len and possibly the mbuf offset */
6982 		rack_adjust_orig_mlen(src_rsm);
6983 	}
6984 	m = src_rsm->m;
6985 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6986 	while (soff >= m->m_len) {
6987 		/* Move out past this mbuf */
6988 		soff -= m->m_len;
6989 		m = m->m_next;
6990 		KASSERT((m != NULL),
6991 			("rsm:%p nrsm:%p hit at soff:%u null m",
6992 			 src_rsm, rsm, soff));
6993 		if (m == NULL) {
6994 			/* This should *not* happen which is why there is a kassert */
6995 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
6996 					       (src_rsm->r_start - rack->rc_tp->snd_una),
6997 					       &src_rsm->soff);
6998 			src_rsm->orig_m_len = src_rsm->m->m_len;
6999 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7000 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7001 					   (rsm->r_start - rack->rc_tp->snd_una),
7002 					   &rsm->soff);
7003 			rsm->orig_m_len = rsm->m->m_len;
7004 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7005 			return;
7006 		}
7007 	}
7008 	rsm->m = m;
7009 	rsm->soff = soff;
7010 	rsm->orig_m_len = m->m_len;
7011 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7012 }
7013 
7014 static __inline void
7015 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7016 	       struct rack_sendmap *rsm, uint32_t start)
7017 {
7018 	int idx;
7019 
7020 	nrsm->r_start = start;
7021 	nrsm->r_end = rsm->r_end;
7022 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7023 	nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7024 	nrsm->r_flags = rsm->r_flags;
7025 	nrsm->r_dupack = rsm->r_dupack;
7026 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7027 	nrsm->r_rtr_bytes = 0;
7028 	nrsm->r_fas = rsm->r_fas;
7029 	nrsm->r_bas = rsm->r_bas;
7030 	tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7031 	nrsm->r_just_ret = rsm->r_just_ret;
7032 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7033 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7034 	}
7035 	/* Now if we have SYN flag we keep it on the left edge */
7036 	if (nrsm->r_flags & RACK_HAS_SYN)
7037 		nrsm->r_flags &= ~RACK_HAS_SYN;
7038 	/* Now if we have a FIN flag we keep it on the right edge */
7039 	if (rsm->r_flags & RACK_HAS_FIN)
7040 		rsm->r_flags &= ~RACK_HAS_FIN;
7041 	/* Push bit must go to the right edge as well */
7042 	if (rsm->r_flags & RACK_HAD_PUSH)
7043 		rsm->r_flags &= ~RACK_HAD_PUSH;
7044 	/* Update the count if app limited */
7045 	if (nrsm->r_flags & RACK_APP_LIMITED)
7046 		rack->r_ctl.rc_app_limited_cnt++;
7047 	/* Clone over the state of the hw_tls flag */
7048 	nrsm->r_hw_tls = rsm->r_hw_tls;
7049 	/*
7050 	 * Now we need to find nrsm's new location in the mbuf chain
7051 	 * we basically calculate a new offset, which is soff +
7052 	 * how much is left in original rsm. Then we walk out the mbuf
7053 	 * chain to find the righ position, it may be the same mbuf
7054 	 * or maybe not.
7055 	 */
7056 	KASSERT(((rsm->m != NULL) ||
7057 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7058 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7059 	if (rsm->m)
7060 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7061 }
7062 
7063 static struct rack_sendmap *
7064 rack_merge_rsm(struct tcp_rack *rack,
7065 	       struct rack_sendmap *l_rsm,
7066 	       struct rack_sendmap *r_rsm)
7067 {
7068 	/*
7069 	 * We are merging two ack'd RSM's,
7070 	 * the l_rsm is on the left (lower seq
7071 	 * values) and the r_rsm is on the right
7072 	 * (higher seq value). The simplest way
7073 	 * to merge these is to move the right
7074 	 * one into the left. I don't think there
7075 	 * is any reason we need to try to find
7076 	 * the oldest (or last oldest retransmitted).
7077 	 */
7078 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7079 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7080 	tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7081 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7082 		l_rsm->r_dupack = r_rsm->r_dupack;
7083 	if (r_rsm->r_rtr_bytes)
7084 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7085 	if (r_rsm->r_in_tmap) {
7086 		/* This really should not happen */
7087 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7088 		r_rsm->r_in_tmap = 0;
7089 	}
7090 
7091 	/* Now the flags */
7092 	if (r_rsm->r_flags & RACK_HAS_FIN)
7093 		l_rsm->r_flags |= RACK_HAS_FIN;
7094 	if (r_rsm->r_flags & RACK_TLP)
7095 		l_rsm->r_flags |= RACK_TLP;
7096 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7097 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7098 	if ((r_rsm->r_flags & RACK_APP_LIMITED) &&
7099 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7100 		/*
7101 		 * If both are app-limited then let the
7102 		 * free lower the count. If right is app
7103 		 * limited and left is not, transfer.
7104 		 */
7105 		l_rsm->r_flags |= RACK_APP_LIMITED;
7106 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7107 		if (r_rsm == rack->r_ctl.rc_first_appl)
7108 			rack->r_ctl.rc_first_appl = l_rsm;
7109 	}
7110 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7111 	/*
7112 	 * We keep the largest value, which is the newest
7113 	 * send. We do this in case a segment that is
7114 	 * joined together and not part of a GP estimate
7115 	 * later gets expanded into the GP estimate.
7116 	 *
7117 	 * We prohibit the merging of unlike kinds i.e.
7118 	 * all pieces that are in the GP estimate can be
7119 	 * merged and all pieces that are not in a GP estimate
7120 	 * can be merged, but not disimilar pieces. Combine
7121 	 * this with taking the highest here and we should
7122 	 * be ok unless of course the client reneges. Then
7123 	 * all bets are off.
7124 	 */
7125 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7126 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7127 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7128 	}
7129 	/*
7130 	 * When merging two RSM's we also need to consider the ack time and keep
7131 	 * newest. If the ack gets merged into a measurement then that is the
7132 	 * one we will want to be using.
7133 	 */
7134 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7135 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7136 
7137 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7138 		/* Transfer the split limit to the map we free */
7139 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7140 		l_rsm->r_limit_type = 0;
7141 	}
7142 	rack_free(rack, r_rsm);
7143 	l_rsm->r_flags |= RACK_MERGED;
7144 	return (l_rsm);
7145 }
7146 
7147 /*
7148  * TLP Timer, here we simply setup what segment we want to
7149  * have the TLP expire on, the normal rack_output() will then
7150  * send it out.
7151  *
7152  * We return 1, saying don't proceed with rack_output only
7153  * when all timers have been stopped (destroyed PCB?).
7154  */
7155 static int
7156 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7157 {
7158 	/*
7159 	 * Tail Loss Probe.
7160 	 */
7161 	struct rack_sendmap *rsm = NULL;
7162 	int insret __diagused;
7163 	struct socket *so = tptosocket(tp);
7164 	uint32_t amm;
7165 	uint32_t out, avail;
7166 	int collapsed_win = 0;
7167 
7168 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7169 		/* Its not time yet */
7170 		return (0);
7171 	}
7172 	if (ctf_progress_timeout_check(tp, true)) {
7173 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7174 		return (-ETIMEDOUT);	/* tcp_drop() */
7175 	}
7176 	/*
7177 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7178 	 * need to figure out how to force a full MSS segment out.
7179 	 */
7180 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7181 	rack->r_ctl.retran_during_recovery = 0;
7182 	rack->r_might_revert = 0;
7183 	rack->r_ctl.dsack_byte_cnt = 0;
7184 	counter_u64_add(rack_tlp_tot, 1);
7185 	if (rack->r_state && (rack->r_state != tp->t_state))
7186 		rack_set_state(tp, rack);
7187 	avail = sbavail(&so->so_snd);
7188 	out = tp->snd_max - tp->snd_una;
7189 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7190 		/* special case, we need a retransmission */
7191 		collapsed_win = 1;
7192 		goto need_retran;
7193 	}
7194 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7195 		rack->r_ctl.dsack_persist--;
7196 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7197 			rack->r_ctl.num_dsack = 0;
7198 		}
7199 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7200 	}
7201 	if ((tp->t_flags & TF_GPUTINPROG) &&
7202 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7203 		/*
7204 		 * If this is the second in a row
7205 		 * TLP and we are doing a measurement
7206 		 * its time to abandon the measurement.
7207 		 * Something is likely broken on
7208 		 * the clients network and measuring a
7209 		 * broken network does us no good.
7210 		 */
7211 		tp->t_flags &= ~TF_GPUTINPROG;
7212 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7213 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7214 					   tp->gput_seq,
7215 					   0, 0, 18, __LINE__, NULL, 0);
7216 	}
7217 	/*
7218 	 * Check our send oldest always settings, and if
7219 	 * there is an oldest to send jump to the need_retran.
7220 	 */
7221 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7222 		goto need_retran;
7223 
7224 	if (avail > out) {
7225 		/* New data is available */
7226 		amm = avail - out;
7227 		if (amm > ctf_fixed_maxseg(tp)) {
7228 			amm = ctf_fixed_maxseg(tp);
7229 			if ((amm + out) > tp->snd_wnd) {
7230 				/* We are rwnd limited */
7231 				goto need_retran;
7232 			}
7233 		} else if (amm < ctf_fixed_maxseg(tp)) {
7234 			/* not enough to fill a MTU */
7235 			goto need_retran;
7236 		}
7237 		if (IN_FASTRECOVERY(tp->t_flags)) {
7238 			/* Unlikely */
7239 			if (rack->rack_no_prr == 0) {
7240 				if (out + amm <= tp->snd_wnd) {
7241 					rack->r_ctl.rc_prr_sndcnt = amm;
7242 					rack->r_ctl.rc_tlp_new_data = amm;
7243 					rack_log_to_prr(rack, 4, 0, __LINE__);
7244 				}
7245 			} else
7246 				goto need_retran;
7247 		} else {
7248 			/* Set the send-new override */
7249 			if (out + amm <= tp->snd_wnd)
7250 				rack->r_ctl.rc_tlp_new_data = amm;
7251 			else
7252 				goto need_retran;
7253 		}
7254 		rack->r_ctl.rc_tlpsend = NULL;
7255 		counter_u64_add(rack_tlp_newdata, 1);
7256 		goto send;
7257 	}
7258 need_retran:
7259 	/*
7260 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7261 	 * optionally the first un-acked segment.
7262 	 */
7263 	if (collapsed_win == 0) {
7264 		if (rack_always_send_oldest)
7265 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7266 		else {
7267 			rsm = tqhash_max(rack->r_ctl.tqh);
7268 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7269 				rsm = rack_find_high_nonack(rack, rsm);
7270 			}
7271 		}
7272 		if (rsm == NULL) {
7273 #ifdef TCP_BLACKBOX
7274 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7275 #endif
7276 			goto out;
7277 		}
7278 	} else {
7279 		/*
7280 		 * We had a collapsed window, lets find
7281 		 * the point before the collapse.
7282 		 */
7283 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7284 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7285 		else {
7286 			rsm = tqhash_min(rack->r_ctl.tqh);
7287 		}
7288 		if (rsm == NULL) {
7289 			/* Huh */
7290 			goto out;
7291 		}
7292 	}
7293 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7294 		/*
7295 		 * We need to split this the last segment in two.
7296 		 */
7297 		struct rack_sendmap *nrsm;
7298 
7299 		nrsm = rack_alloc_full_limit(rack);
7300 		if (nrsm == NULL) {
7301 			/*
7302 			 * No memory to split, we will just exit and punt
7303 			 * off to the RXT timer.
7304 			 */
7305 			goto out;
7306 		}
7307 		rack_clone_rsm(rack, nrsm, rsm,
7308 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7309 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7310 #ifndef INVARIANTS
7311 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7312 #else
7313 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7314 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7315 			      nrsm, insret, rack, rsm);
7316 		}
7317 #endif
7318 		if (rsm->r_in_tmap) {
7319 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7320 			nrsm->r_in_tmap = 1;
7321 		}
7322 		rsm = nrsm;
7323 	}
7324 	rack->r_ctl.rc_tlpsend = rsm;
7325 send:
7326 	/* Make sure output path knows we are doing a TLP */
7327 	*doing_tlp = 1;
7328 	rack->r_timer_override = 1;
7329 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7330 	return (0);
7331 out:
7332 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7333 	return (0);
7334 }
7335 
7336 /*
7337  * Delayed ack Timer, here we simply need to setup the
7338  * ACK_NOW flag and remove the DELACK flag. From there
7339  * the output routine will send the ack out.
7340  *
7341  * We only return 1, saying don't proceed, if all timers
7342  * are stopped (destroyed PCB?).
7343  */
7344 static int
7345 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7346 {
7347 
7348 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7349 	tp->t_flags &= ~TF_DELACK;
7350 	tp->t_flags |= TF_ACKNOW;
7351 	KMOD_TCPSTAT_INC(tcps_delack);
7352 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7353 	return (0);
7354 }
7355 
7356 static inline int
7357 rack_send_ack_challange(struct tcp_rack *rack)
7358 {
7359 	struct tcptemp *t_template;
7360 
7361 	t_template = tcpip_maketemplate(rack->rc_inp);
7362 	if (t_template) {
7363 		if (rack->forced_ack == 0) {
7364 			rack->forced_ack = 1;
7365 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7366 		} else {
7367 			rack->probe_not_answered = 1;
7368 		}
7369 		tcp_respond(rack->rc_tp, t_template->tt_ipgen,
7370 			    &t_template->tt_t, (struct mbuf *)NULL,
7371 			    rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
7372 		free(t_template, M_TEMP);
7373 		/* This does send an ack so kill any D-ack timer */
7374 		if (rack->rc_tp->t_flags & TF_DELACK)
7375 			rack->rc_tp->t_flags &= ~TF_DELACK;
7376 		return(1);
7377 	} else
7378 		return (0);
7379 
7380 }
7381 
7382 /*
7383  * Persists timer, here we simply send the
7384  * same thing as a keepalive will.
7385  * the one byte send.
7386  *
7387  * We only return 1, saying don't proceed, if all timers
7388  * are stopped (destroyed PCB?).
7389  */
7390 static int
7391 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7392 {
7393 	int32_t retval = 1;
7394 
7395 	if (rack->rc_in_persist == 0)
7396 		return (0);
7397 	if (ctf_progress_timeout_check(tp, false)) {
7398 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7399 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7400 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7401 		return (-ETIMEDOUT);	/* tcp_drop() */
7402 	}
7403 	/*
7404 	 * Persistence timer into zero window. Force a byte to be output, if
7405 	 * possible.
7406 	 */
7407 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
7408 	/*
7409 	 * Hack: if the peer is dead/unreachable, we do not time out if the
7410 	 * window is closed.  After a full backoff, drop the connection if
7411 	 * the idle time (no responses to probes) reaches the maximum
7412 	 * backoff that we would use if retransmitting.
7413 	 */
7414 	if (tp->t_rxtshift >= V_tcp_retries &&
7415 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7416 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7417 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7418 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7419 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7420 		retval = -ETIMEDOUT;	/* tcp_drop() */
7421 		goto out;
7422 	}
7423 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7424 	    tp->snd_una == tp->snd_max)
7425 		rack_exit_persist(tp, rack, cts);
7426 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7427 	/*
7428 	 * If the user has closed the socket then drop a persisting
7429 	 * connection after a much reduced timeout.
7430 	 */
7431 	if (tp->t_state > TCPS_CLOSE_WAIT &&
7432 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7433 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7434 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7435 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7436 		retval = -ETIMEDOUT;	/* tcp_drop() */
7437 		goto out;
7438 	}
7439 	if (rack_send_ack_challange(rack)) {
7440 		/* only set it if we were answered */
7441 		if (rack->probe_not_answered) {
7442 			counter_u64_add(rack_persists_loss, 1);
7443 			rack->r_ctl.persist_lost_ends++;
7444 		}
7445 		counter_u64_add(rack_persists_sends, 1);
7446 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7447 	}
7448 	if (tp->t_rxtshift < V_tcp_retries)
7449 		tp->t_rxtshift++;
7450 out:
7451 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7452 	rack_start_hpts_timer(rack, tp, cts,
7453 			      0, 0, 0);
7454 	return (retval);
7455 }
7456 
7457 /*
7458  * If a keepalive goes off, we had no other timers
7459  * happening. We always return 1 here since this
7460  * routine either drops the connection or sends
7461  * out a segment with respond.
7462  */
7463 static int
7464 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7465 {
7466 	struct inpcb *inp = tptoinpcb(tp);
7467 
7468 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7469 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7470 	/*
7471 	 * Keep-alive timer went off; send something or drop connection if
7472 	 * idle for too long.
7473 	 */
7474 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
7475 	if (tp->t_state < TCPS_ESTABLISHED)
7476 		goto dropit;
7477 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7478 	    tp->t_state <= TCPS_CLOSING) {
7479 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7480 			goto dropit;
7481 		/*
7482 		 * Send a packet designed to force a response if the peer is
7483 		 * up and reachable: either an ACK if the connection is
7484 		 * still alive, or an RST if the peer has closed the
7485 		 * connection due to timeout or reboot. Using sequence
7486 		 * number tp->snd_una-1 causes the transmitted zero-length
7487 		 * segment to lie outside the receive window; by the
7488 		 * protocol spec, this requires the correspondent TCP to
7489 		 * respond.
7490 		 */
7491 		KMOD_TCPSTAT_INC(tcps_keepprobe);
7492 		rack_send_ack_challange(rack);
7493 	}
7494 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7495 	return (1);
7496 dropit:
7497 	KMOD_TCPSTAT_INC(tcps_keepdrops);
7498 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7499 	return (-ETIMEDOUT);	/* tcp_drop() */
7500 }
7501 
7502 /*
7503  * Retransmit helper function, clear up all the ack
7504  * flags and take care of important book keeping.
7505  */
7506 static void
7507 rack_remxt_tmr(struct tcpcb *tp)
7508 {
7509 	/*
7510 	 * The retransmit timer went off, all sack'd blocks must be
7511 	 * un-acked.
7512 	 */
7513 	struct rack_sendmap *rsm, *trsm = NULL;
7514 	struct tcp_rack *rack;
7515 
7516 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7517 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7518 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7519 	rack->r_timer_override = 1;
7520 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7521 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7522 	rack->r_late = 0;
7523 	rack->r_early = 0;
7524 	rack->r_ctl.rc_agg_delayed = 0;
7525 	rack->r_ctl.rc_agg_early = 0;
7526 	if (rack->r_state && (rack->r_state != tp->t_state))
7527 		rack_set_state(tp, rack);
7528 	if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
7529 		/*
7530 		 * We do not clear the scoreboard until we have had
7531 		 * more than rack_rxt_scoreboard_clear_thresh time-outs.
7532 		 */
7533 		rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7534 		if (rack->r_ctl.rc_resend != NULL)
7535 			rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7536 
7537 		return;
7538 	}
7539 	/*
7540 	 * Ideally we would like to be able to
7541 	 * mark SACK-PASS on anything not acked here.
7542 	 *
7543 	 * However, if we do that we would burst out
7544 	 * all that data 1ms apart. This would be unwise,
7545 	 * so for now we will just let the normal rxt timer
7546 	 * and tlp timer take care of it.
7547 	 *
7548 	 * Also we really need to stick them back in sequence
7549 	 * order. This way we send in the proper order and any
7550 	 * sacks that come floating in will "re-ack" the data.
7551 	 * To do this we zap the tmap with an INIT and then
7552 	 * walk through and place every rsm in the tail queue
7553 	 * hash table back in its seq ordered place.
7554 	 */
7555 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
7556 
7557 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7558 		rsm->r_dupack = 0;
7559 		if (rack_verbose_logging)
7560 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7561 		/* We must re-add it back to the tlist */
7562 		if (trsm == NULL) {
7563 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7564 		} else {
7565 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7566 		}
7567 		rsm->r_in_tmap = 1;
7568 		trsm = rsm;
7569 		if (rsm->r_flags & RACK_ACKED)
7570 			rsm->r_flags |= RACK_WAS_ACKED;
7571 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
7572 		rsm->r_flags |= RACK_MUST_RXT;
7573 	}
7574 	/* zero the lost since it's all gone */
7575 	rack->r_ctl.rc_considered_lost = 0;
7576 	/* Clear the count (we just un-acked them) */
7577 	rack->r_ctl.rc_sacked = 0;
7578 	rack->r_ctl.rc_sacklast = NULL;
7579 	/* Clear the tlp rtx mark */
7580 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7581 	if (rack->r_ctl.rc_resend != NULL)
7582 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7583 	rack->r_ctl.rc_prr_sndcnt = 0;
7584 	rack_log_to_prr(rack, 6, 0, __LINE__);
7585 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7586 	if (rack->r_ctl.rc_resend != NULL)
7587 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7588 	if (((tp->t_flags & TF_SACK_PERMIT) == 0) &&
7589 	    ((tp->t_flags & TF_SENTFIN) == 0)) {
7590 		/*
7591 		 * For non-sack customers new data
7592 		 * needs to go out as retransmits until
7593 		 * we retransmit up to snd_max.
7594 		 */
7595 		rack->r_must_retran = 1;
7596 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7597 							    rack->r_ctl.rc_sacked);
7598 	}
7599 }
7600 
7601 static void
7602 rack_convert_rtts(struct tcpcb *tp)
7603 {
7604 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7605 	tp->t_rxtcur = RACK_REXMTVAL(tp);
7606 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7607 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7608 	}
7609 	if (tp->t_rxtcur > rack_rto_max) {
7610 		tp->t_rxtcur = rack_rto_max;
7611 	}
7612 }
7613 
7614 static void
7615 rack_cc_conn_init(struct tcpcb *tp)
7616 {
7617 	struct tcp_rack *rack;
7618 	uint32_t srtt;
7619 
7620 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7621 	srtt = tp->t_srtt;
7622 	cc_conn_init(tp);
7623 	/*
7624 	 * Now convert to rack's internal format,
7625 	 * if required.
7626 	 */
7627 	if ((srtt == 0) && (tp->t_srtt != 0))
7628 		rack_convert_rtts(tp);
7629 	/*
7630 	 * We want a chance to stay in slowstart as
7631 	 * we create a connection. TCP spec says that
7632 	 * initially ssthresh is infinite. For our
7633 	 * purposes that is the snd_wnd.
7634 	 */
7635 	if (tp->snd_ssthresh < tp->snd_wnd) {
7636 		tp->snd_ssthresh = tp->snd_wnd;
7637 	}
7638 	/*
7639 	 * We also want to assure a IW worth of
7640 	 * data can get inflight.
7641 	 */
7642 	if (rc_init_window(rack) < tp->snd_cwnd)
7643 		tp->snd_cwnd = rc_init_window(rack);
7644 }
7645 
7646 /*
7647  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7648  * we will setup to retransmit the lowest seq number outstanding.
7649  */
7650 static int
7651 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7652 {
7653 	struct inpcb *inp = tptoinpcb(tp);
7654 	int32_t rexmt;
7655 	int32_t retval = 0;
7656 	bool isipv6;
7657 
7658 	if ((tp->t_flags & TF_GPUTINPROG) &&
7659 	    (tp->t_rxtshift)) {
7660 		/*
7661 		 * We have had a second timeout
7662 		 * measurements on successive rxt's are not profitable.
7663 		 * It is unlikely to be of any use (the network is
7664 		 * broken or the client went away).
7665 		 */
7666 		tp->t_flags &= ~TF_GPUTINPROG;
7667 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7668 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7669 					   tp->gput_seq,
7670 					   0, 0, 18, __LINE__, NULL, 0);
7671 	}
7672 	if (ctf_progress_timeout_check(tp, false)) {
7673 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7674 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7675 		return (-ETIMEDOUT);	/* tcp_drop() */
7676 	}
7677 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7678 	rack->r_ctl.retran_during_recovery = 0;
7679 	rack->rc_ack_required = 1;
7680 	rack->r_ctl.dsack_byte_cnt = 0;
7681 	if (IN_RECOVERY(tp->t_flags) &&
7682 	    (rack->rto_from_rec == 0)) {
7683 		/*
7684 		 * Mark that we had a rto while in recovery
7685 		 * and save the ssthresh so if we go back
7686 		 * into recovery we will have a chance
7687 		 * to slowstart back to the level.
7688 		 */
7689 		rack->rto_from_rec = 1;
7690 		rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
7691 	}
7692 	if (IN_FASTRECOVERY(tp->t_flags))
7693 		tp->t_flags |= TF_WASFRECOVERY;
7694 	else
7695 		tp->t_flags &= ~TF_WASFRECOVERY;
7696 	if (IN_CONGRECOVERY(tp->t_flags))
7697 		tp->t_flags |= TF_WASCRECOVERY;
7698 	else
7699 		tp->t_flags &= ~TF_WASCRECOVERY;
7700 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7701 	    (tp->snd_una == tp->snd_max)) {
7702 		/* Nothing outstanding .. nothing to do */
7703 		return (0);
7704 	}
7705 	if (rack->r_ctl.dsack_persist) {
7706 		rack->r_ctl.dsack_persist--;
7707 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7708 			rack->r_ctl.num_dsack = 0;
7709 		}
7710 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7711 	}
7712 	/*
7713 	 * Rack can only run one timer  at a time, so we cannot
7714 	 * run a KEEPINIT (gating SYN sending) and a retransmit
7715 	 * timer for the SYN. So if we are in a front state and
7716 	 * have a KEEPINIT timer we need to check the first transmit
7717 	 * against now to see if we have exceeded the KEEPINIT time
7718 	 * (if one is set).
7719 	 */
7720 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7721 	    (TP_KEEPINIT(tp) != 0)) {
7722 		struct rack_sendmap *rsm;
7723 
7724 		rsm = tqhash_min(rack->r_ctl.tqh);
7725 		if (rsm) {
7726 			/* Ok we have something outstanding to test keepinit with */
7727 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7728 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7729 				/* We have exceeded the KEEPINIT time */
7730 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7731 				goto drop_it;
7732 			}
7733 		}
7734 	}
7735 	/*
7736 	 * Retransmission timer went off.  Message has not been acked within
7737 	 * retransmit interval.  Back off to a longer retransmit interval
7738 	 * and retransmit one segment.
7739 	 */
7740 	if ((rack->r_ctl.rc_resend == NULL) ||
7741 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7742 		/*
7743 		 * If the rwnd collapsed on
7744 		 * the one we are retransmitting
7745 		 * it does not count against the
7746 		 * rxt count.
7747 		 */
7748 		tp->t_rxtshift++;
7749 	}
7750 	rack_remxt_tmr(tp);
7751 	if (tp->t_rxtshift > V_tcp_retries) {
7752 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7753 drop_it:
7754 		tp->t_rxtshift = V_tcp_retries;
7755 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7756 		/* XXXGL: previously t_softerror was casted to uint16_t */
7757 		MPASS(tp->t_softerror >= 0);
7758 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7759 		goto out;	/* tcp_drop() */
7760 	}
7761 	if (tp->t_state == TCPS_SYN_SENT) {
7762 		/*
7763 		 * If the SYN was retransmitted, indicate CWND to be limited
7764 		 * to 1 segment in cc_conn_init().
7765 		 */
7766 		tp->snd_cwnd = 1;
7767 	} else if (tp->t_rxtshift == 1) {
7768 		/*
7769 		 * first retransmit; record ssthresh and cwnd so they can be
7770 		 * recovered if this turns out to be a "bad" retransmit. A
7771 		 * retransmit is considered "bad" if an ACK for this segment
7772 		 * is received within RTT/2 interval; the assumption here is
7773 		 * that the ACK was already in flight.  See "On Estimating
7774 		 * End-to-End Network Path Properties" by Allman and Paxson
7775 		 * for more details.
7776 		 */
7777 		tp->snd_cwnd_prev = tp->snd_cwnd;
7778 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
7779 		tp->snd_recover_prev = tp->snd_recover;
7780 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7781 		tp->t_flags |= TF_PREVVALID;
7782 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7783 		tp->t_flags &= ~TF_PREVVALID;
7784 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7785 	if ((tp->t_state == TCPS_SYN_SENT) ||
7786 	    (tp->t_state == TCPS_SYN_RECEIVED))
7787 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7788 	else
7789 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7790 
7791 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7792 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7793 	/*
7794 	 * We enter the path for PLMTUD if connection is established or, if
7795 	 * connection is FIN_WAIT_1 status, reason for the last is that if
7796 	 * amount of data we send is very small, we could send it in couple
7797 	 * of packets and process straight to FIN. In that case we won't
7798 	 * catch ESTABLISHED state.
7799 	 */
7800 #ifdef INET6
7801 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7802 #else
7803 	isipv6 = false;
7804 #endif
7805 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
7806 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7807 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7808 	    ((tp->t_state == TCPS_ESTABLISHED) ||
7809 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
7810 		/*
7811 		 * Idea here is that at each stage of mtu probe (usually,
7812 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7813 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7814 		 * should take care of that.
7815 		 */
7816 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7817 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7818 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7819 		    tp->t_rxtshift % 2 == 0)) {
7820 			/*
7821 			 * Enter Path MTU Black-hole Detection mechanism: -
7822 			 * Disable Path MTU Discovery (IP "DF" bit). -
7823 			 * Reduce MTU to lower value than what we negotiated
7824 			 * with peer.
7825 			 */
7826 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7827 				/* Record that we may have found a black hole. */
7828 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7829 				/* Keep track of previous MSS. */
7830 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7831 			}
7832 
7833 			/*
7834 			 * Reduce the MSS to blackhole value or to the
7835 			 * default in an attempt to retransmit.
7836 			 */
7837 #ifdef INET6
7838 			if (isipv6 &&
7839 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7840 				/* Use the sysctl tuneable blackhole MSS. */
7841 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7842 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7843 			} else if (isipv6) {
7844 				/* Use the default MSS. */
7845 				tp->t_maxseg = V_tcp_v6mssdflt;
7846 				/*
7847 				 * Disable Path MTU Discovery when we switch
7848 				 * to minmss.
7849 				 */
7850 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7851 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7852 			}
7853 #endif
7854 #if defined(INET6) && defined(INET)
7855 			else
7856 #endif
7857 #ifdef INET
7858 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7859 				/* Use the sysctl tuneable blackhole MSS. */
7860 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7861 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7862 			} else {
7863 				/* Use the default MSS. */
7864 				tp->t_maxseg = V_tcp_mssdflt;
7865 				/*
7866 				 * Disable Path MTU Discovery when we switch
7867 				 * to minmss.
7868 				 */
7869 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7870 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7871 			}
7872 #endif
7873 		} else {
7874 			/*
7875 			 * If further retransmissions are still unsuccessful
7876 			 * with a lowered MTU, maybe this isn't a blackhole
7877 			 * and we restore the previous MSS and blackhole
7878 			 * detection flags. The limit '6' is determined by
7879 			 * giving each probe stage (1448, 1188, 524) 2
7880 			 * chances to recover.
7881 			 */
7882 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7883 			    (tp->t_rxtshift >= 6)) {
7884 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7885 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7886 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7887 				if (tp->t_maxseg < V_tcp_mssdflt) {
7888 					/*
7889 					 * The MSS is so small we should not
7890 					 * process incoming SACK's since we are
7891 					 * subject to attack in such a case.
7892 					 */
7893 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
7894 				} else {
7895 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
7896 				}
7897 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7898 			}
7899 		}
7900 	}
7901 	/*
7902 	 * Disable RFC1323 and SACK if we haven't got any response to
7903 	 * our third SYN to work-around some broken terminal servers
7904 	 * (most of which have hopefully been retired) that have bad VJ
7905 	 * header compression code which trashes TCP segments containing
7906 	 * unknown-to-them TCP options.
7907 	 */
7908 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7909 	    (tp->t_rxtshift == 3))
7910 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7911 	/*
7912 	 * If we backed off this far, our srtt estimate is probably bogus.
7913 	 * Clobber it so we'll take the next rtt measurement as our srtt;
7914 	 * move the current srtt into rttvar to keep the current retransmit
7915 	 * times until then.
7916 	 */
7917 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7918 #ifdef INET6
7919 		if ((inp->inp_vflag & INP_IPV6) != 0)
7920 			in6_losing(inp);
7921 		else
7922 #endif
7923 			in_losing(inp);
7924 		tp->t_rttvar += tp->t_srtt;
7925 		tp->t_srtt = 0;
7926 	}
7927 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7928 	tp->snd_recover = tp->snd_max;
7929 	tp->t_flags |= TF_ACKNOW;
7930 	tp->t_rtttime = 0;
7931 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7932 out:
7933 	return (retval);
7934 }
7935 
7936 static int
7937 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7938 {
7939 	int32_t ret = 0;
7940 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7941 
7942 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7943 	    (tp->t_flags & TF_GPUTINPROG)) {
7944 		/*
7945 		 * We have a goodput in progress
7946 		 * and we have entered a late state.
7947 		 * Do we have enough data in the sb
7948 		 * to handle the GPUT request?
7949 		 */
7950 		uint32_t bytes;
7951 
7952 		bytes = tp->gput_ack - tp->gput_seq;
7953 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
7954 			bytes += tp->gput_seq - tp->snd_una;
7955 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7956 			/*
7957 			 * There are not enough bytes in the socket
7958 			 * buffer that have been sent to cover this
7959 			 * measurement. Cancel it.
7960 			 */
7961 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7962 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
7963 						   tp->gput_seq,
7964 						   0, 0, 18, __LINE__, NULL, 0);
7965 			tp->t_flags &= ~TF_GPUTINPROG;
7966 		}
7967 	}
7968 	if (timers == 0) {
7969 		return (0);
7970 	}
7971 	if (tp->t_state == TCPS_LISTEN) {
7972 		/* no timers on listen sockets */
7973 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7974 			return (0);
7975 		return (1);
7976 	}
7977 	if ((timers & PACE_TMR_RACK) &&
7978 	    rack->rc_on_min_to) {
7979 		/*
7980 		 * For the rack timer when we
7981 		 * are on a min-timeout (which means rrr_conf = 3)
7982 		 * we don't want to check the timer. It may
7983 		 * be going off for a pace and thats ok we
7984 		 * want to send the retransmit (if its ready).
7985 		 *
7986 		 * If its on a normal rack timer (non-min) then
7987 		 * we will check if its expired.
7988 		 */
7989 		goto skip_time_check;
7990 	}
7991 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7992 		uint32_t left;
7993 
7994 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
7995 			ret = -1;
7996 			rack_log_to_processing(rack, cts, ret, 0);
7997 			return (0);
7998 		}
7999 		if (hpts_calling == 0) {
8000 			/*
8001 			 * A user send or queued mbuf (sack) has called us? We
8002 			 * return 0 and let the pacing guards
8003 			 * deal with it if they should or
8004 			 * should not cause a send.
8005 			 */
8006 			ret = -2;
8007 			rack_log_to_processing(rack, cts, ret, 0);
8008 			return (0);
8009 		}
8010 		/*
8011 		 * Ok our timer went off early and we are not paced false
8012 		 * alarm, go back to sleep. We make sure we don't have
8013 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8014 		 * flag in place.
8015 		 */
8016 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8017 		ret = -3;
8018 		left = rack->r_ctl.rc_timer_exp - cts;
8019 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8020 		rack_log_to_processing(rack, cts, ret, left);
8021 		return (1);
8022 	}
8023 skip_time_check:
8024 	rack->rc_tmr_stopped = 0;
8025 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8026 	if (timers & PACE_TMR_DELACK) {
8027 		ret = rack_timeout_delack(tp, rack, cts);
8028 	} else if (timers & PACE_TMR_RACK) {
8029 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8030 		rack->r_fast_output = 0;
8031 		ret = rack_timeout_rack(tp, rack, cts);
8032 	} else if (timers & PACE_TMR_TLP) {
8033 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8034 		rack->r_fast_output = 0;
8035 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8036 	} else if (timers & PACE_TMR_RXT) {
8037 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8038 		rack->r_fast_output = 0;
8039 		ret = rack_timeout_rxt(tp, rack, cts);
8040 	} else if (timers & PACE_TMR_PERSIT) {
8041 		ret = rack_timeout_persist(tp, rack, cts);
8042 	} else if (timers & PACE_TMR_KEEP) {
8043 		ret = rack_timeout_keepalive(tp, rack, cts);
8044 	}
8045 	rack_log_to_processing(rack, cts, ret, timers);
8046 	return (ret);
8047 }
8048 
8049 static void
8050 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8051 {
8052 	struct timeval tv;
8053 	uint32_t us_cts, flags_on_entry;
8054 	uint8_t hpts_removed = 0;
8055 
8056 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8057 	us_cts = tcp_get_usecs(&tv);
8058 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8059 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8060 	     ((tp->snd_max - tp->snd_una) == 0))) {
8061 		tcp_hpts_remove(rack->rc_tp);
8062 		hpts_removed = 1;
8063 		/* If we were not delayed cancel out the flag. */
8064 		if ((tp->snd_max - tp->snd_una) == 0)
8065 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8066 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8067 	}
8068 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8069 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8070 		if (tcp_in_hpts(rack->rc_tp) &&
8071 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8072 			/*
8073 			 * Canceling timer's when we have no output being
8074 			 * paced. We also must remove ourselves from the
8075 			 * hpts.
8076 			 */
8077 			tcp_hpts_remove(rack->rc_tp);
8078 			hpts_removed = 1;
8079 		}
8080 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8081 	}
8082 	if (hpts_removed == 0)
8083 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8084 }
8085 
8086 static int
8087 rack_stopall(struct tcpcb *tp)
8088 {
8089 	struct tcp_rack *rack;
8090 
8091 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8092 	rack->t_timers_stopped = 1;
8093 
8094 	tcp_hpts_remove(tp);
8095 
8096 	return (0);
8097 }
8098 
8099 static void
8100 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8101 {
8102 	/*
8103 	 * Assure no timers are running.
8104 	 */
8105 	if (tcp_timer_active(tp, TT_PERSIST)) {
8106 		/* We enter in persists, set the flag appropriately */
8107 		rack->rc_in_persist = 1;
8108 	}
8109 	if (tcp_in_hpts(rack->rc_tp)) {
8110 		tcp_hpts_remove(rack->rc_tp);
8111 	}
8112 }
8113 
8114 static void
8115 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8116     struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8117 {
8118 	int32_t idx;
8119 
8120 	rsm->r_rtr_cnt++;
8121 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8122 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8123 		rsm->r_flags |= RACK_OVERMAX;
8124 	}
8125 	rsm->r_act_rxt_cnt++;
8126 	/* Peg the count/index */
8127 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8128 	rsm->r_dupack = 0;
8129 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8130 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8131 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8132 	}
8133 	if (rsm->r_flags & RACK_WAS_LOST) {
8134 		/*
8135 		 * We retransmitted it putting it back in flight
8136 		 * remove the lost desgination and reduce the
8137 		 * bytes considered lost.
8138 		 */
8139 		rsm->r_flags &= ~RACK_WAS_LOST;
8140 		KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8141 			("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
8142 		if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8143 			rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8144 		else
8145 			rack->r_ctl.rc_considered_lost = 0;
8146 	}
8147 	idx = rsm->r_rtr_cnt - 1;
8148 	rsm->r_tim_lastsent[idx] = ts;
8149 	/*
8150 	 * Here we don't add in the len of send, since its already
8151 	 * in snduna <->snd_max.
8152 	 */
8153 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8154 				     rack->r_ctl.rc_sacked);
8155 	if (rsm->r_flags & RACK_ACKED) {
8156 		/* Problably MTU discovery messing with us */
8157 		rsm->r_flags &= ~RACK_ACKED;
8158 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8159 	}
8160 	if (rsm->r_in_tmap) {
8161 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8162 		rsm->r_in_tmap = 0;
8163 	}
8164 	/* Lets make sure it really is in or not the GP window */
8165 	rack_mark_in_gp_win(tp, rsm);
8166 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8167 	rsm->r_in_tmap = 1;
8168 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8169 	/* Take off the must retransmit flag, if its on */
8170 	if (rsm->r_flags & RACK_MUST_RXT) {
8171 		if (rack->r_must_retran)
8172 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8173 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8174 			/*
8175 			 * We have retransmitted all we need. Clear
8176 			 * any must retransmit flags.
8177 			 */
8178 			rack->r_must_retran = 0;
8179 			rack->r_ctl.rc_out_at_rto = 0;
8180 		}
8181 		rsm->r_flags &= ~RACK_MUST_RXT;
8182 	}
8183 	/* Remove any collapsed flag */
8184 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8185 	if (rsm->r_flags & RACK_SACK_PASSED) {
8186 		/* We have retransmitted due to the SACK pass */
8187 		rsm->r_flags &= ~RACK_SACK_PASSED;
8188 		rsm->r_flags |= RACK_WAS_SACKPASS;
8189 	}
8190 }
8191 
8192 static uint32_t
8193 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8194     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8195 {
8196 	/*
8197 	 * We (re-)transmitted starting at rsm->r_start for some length
8198 	 * (possibly less than r_end.
8199 	 */
8200 	struct rack_sendmap *nrsm;
8201 	int insret __diagused;
8202 	uint32_t c_end;
8203 	int32_t len;
8204 
8205 	len = *lenp;
8206 	c_end = rsm->r_start + len;
8207 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8208 		/*
8209 		 * We retransmitted the whole piece or more than the whole
8210 		 * slopping into the next rsm.
8211 		 */
8212 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8213 		if (c_end == rsm->r_end) {
8214 			*lenp = 0;
8215 			return (0);
8216 		} else {
8217 			int32_t act_len;
8218 
8219 			/* Hangs over the end return whats left */
8220 			act_len = rsm->r_end - rsm->r_start;
8221 			*lenp = (len - act_len);
8222 			return (rsm->r_end);
8223 		}
8224 		/* We don't get out of this block. */
8225 	}
8226 	/*
8227 	 * Here we retransmitted less than the whole thing which means we
8228 	 * have to split this into what was transmitted and what was not.
8229 	 */
8230 	nrsm = rack_alloc_full_limit(rack);
8231 	if (nrsm == NULL) {
8232 		/*
8233 		 * We can't get memory, so lets not proceed.
8234 		 */
8235 		*lenp = 0;
8236 		return (0);
8237 	}
8238 	/*
8239 	 * So here we are going to take the original rsm and make it what we
8240 	 * retransmitted. nrsm will be the tail portion we did not
8241 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8242 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8243 	 * 1, 6 and the new piece will be 6, 11.
8244 	 */
8245 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8246 	nrsm->r_dupack = 0;
8247 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8248 #ifndef INVARIANTS
8249 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8250 #else
8251 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8252 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8253 		      nrsm, insret, rack, rsm);
8254 	}
8255 #endif
8256 	if (rsm->r_in_tmap) {
8257 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8258 		nrsm->r_in_tmap = 1;
8259 	}
8260 	rsm->r_flags &= (~RACK_HAS_FIN);
8261 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8262 	/* Log a split of rsm into rsm and nrsm */
8263 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8264 	*lenp = 0;
8265 	return (0);
8266 }
8267 
8268 static void
8269 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8270 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8271 		struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
8272 		uint32_t s_moff, int hw_tls, int segsiz)
8273 {
8274 	struct tcp_rack *rack;
8275 	struct rack_sendmap *rsm, *nrsm;
8276 	int insret __diagused;
8277 
8278 	register uint32_t snd_max, snd_una;
8279 
8280 	/*
8281 	 * Add to the RACK log of packets in flight or retransmitted. If
8282 	 * there is a TS option we will use the TS echoed, if not we will
8283 	 * grab a TS.
8284 	 *
8285 	 * Retransmissions will increment the count and move the ts to its
8286 	 * proper place. Note that if options do not include TS's then we
8287 	 * won't be able to effectively use the ACK for an RTT on a retran.
8288 	 *
8289 	 * Notes about r_start and r_end. Lets consider a send starting at
8290 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8291 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8292 	 * This means that r_end is actually the first sequence for the next
8293 	 * slot (11).
8294 	 *
8295 	 */
8296 	/*
8297 	 * If err is set what do we do XXXrrs? should we not add the thing?
8298 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8299 	 * i.e. proceed with add ** do this for now.
8300 	 */
8301 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8302 	if (err)
8303 		/*
8304 		 * We don't log errors -- we could but snd_max does not
8305 		 * advance in this case either.
8306 		 */
8307 		return;
8308 
8309 	if (th_flags & TH_RST) {
8310 		/*
8311 		 * We don't log resets and we return immediately from
8312 		 * sending
8313 		 */
8314 		return;
8315 	}
8316 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8317 	snd_una = tp->snd_una;
8318 	snd_max = tp->snd_max;
8319 	if (th_flags & (TH_SYN | TH_FIN)) {
8320 		/*
8321 		 * The call to rack_log_output is made before bumping
8322 		 * snd_max. This means we can record one extra byte on a SYN
8323 		 * or FIN if seq_out is adding more on and a FIN is present
8324 		 * (and we are not resending).
8325 		 */
8326 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8327 			len++;
8328 		if (th_flags & TH_FIN)
8329 			len++;
8330 	}
8331 	if (SEQ_LEQ((seq_out + len), snd_una)) {
8332 		/* Are sending an old segment to induce an ack (keep-alive)? */
8333 		return;
8334 	}
8335 	if (SEQ_LT(seq_out, snd_una)) {
8336 		/* huh? should we panic? */
8337 		uint32_t end;
8338 
8339 		end = seq_out + len;
8340 		seq_out = snd_una;
8341 		if (SEQ_GEQ(end, seq_out))
8342 			len = end - seq_out;
8343 		else
8344 			len = 0;
8345 	}
8346 	if (len == 0) {
8347 		/* We don't log zero window probes */
8348 		return;
8349 	}
8350 	if (IN_FASTRECOVERY(tp->t_flags)) {
8351 		rack->r_ctl.rc_prr_out += len;
8352 	}
8353 	/* First question is it a retransmission or new? */
8354 	if (seq_out == snd_max) {
8355 		/* Its new */
8356 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8357 again:
8358 		rsm = rack_alloc(rack);
8359 		if (rsm == NULL) {
8360 			/*
8361 			 * Hmm out of memory and the tcb got destroyed while
8362 			 * we tried to wait.
8363 			 */
8364 			return;
8365 		}
8366 		if (th_flags & TH_FIN) {
8367 			rsm->r_flags = RACK_HAS_FIN|add_flag;
8368 		} else {
8369 			rsm->r_flags = add_flag;
8370 		}
8371 		if (hw_tls)
8372 			rsm->r_hw_tls = 1;
8373 		rsm->r_tim_lastsent[0] = cts;
8374 		rsm->r_rtr_cnt = 1;
8375  		rsm->r_act_rxt_cnt = 0;
8376 		rsm->r_rtr_bytes = 0;
8377 		if (th_flags & TH_SYN) {
8378 			/* The data space is one beyond snd_una */
8379 			rsm->r_flags |= RACK_HAS_SYN;
8380 		}
8381 		rsm->r_start = seq_out;
8382 		rsm->r_end = rsm->r_start + len;
8383 		rack_mark_in_gp_win(tp, rsm);
8384 		rsm->r_dupack = 0;
8385 		/*
8386 		 * save off the mbuf location that
8387 		 * sndmbuf_noadv returned (which is
8388 		 * where we started copying from)..
8389 		 */
8390 		rsm->m = s_mb;
8391 		rsm->soff = s_moff;
8392 		/*
8393 		 * Here we do add in the len of send, since its not yet
8394 		 * reflected in in snduna <->snd_max
8395 		 */
8396 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8397 					      rack->r_ctl.rc_sacked) +
8398 			      (rsm->r_end - rsm->r_start));
8399 		if ((rack->rc_initial_ss_comp == 0) &&
8400 		    (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
8401 			   rack->r_ctl.ss_hi_fs = rsm->r_fas;
8402 		}
8403 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8404 		if (rsm->m) {
8405 			if (rsm->m->m_len <= rsm->soff) {
8406 				/*
8407 				 * XXXrrs Question, will this happen?
8408 				 *
8409 				 * If sbsndptr is set at the correct place
8410 				 * then s_moff should always be somewhere
8411 				 * within rsm->m. But if the sbsndptr was
8412 				 * off then that won't be true. If it occurs
8413 				 * we need to walkout to the correct location.
8414 				 */
8415 				struct mbuf *lm;
8416 
8417 				lm = rsm->m;
8418 				while (lm->m_len <= rsm->soff) {
8419 					rsm->soff -= lm->m_len;
8420 					lm = lm->m_next;
8421 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8422 							     __func__, rack, s_moff, s_mb, rsm->soff));
8423 				}
8424 				rsm->m = lm;
8425 			}
8426 			rsm->orig_m_len = rsm->m->m_len;
8427 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8428 		} else {
8429 			rsm->orig_m_len = 0;
8430 			rsm->orig_t_space = 0;
8431 		}
8432 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8433 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8434 		/* Log a new rsm */
8435 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8436 #ifndef INVARIANTS
8437 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
8438 #else
8439 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8440 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8441 			      nrsm, insret, rack, rsm);
8442 		}
8443 #endif
8444 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8445 		rsm->r_in_tmap = 1;
8446 		if (rsm->r_flags & RACK_IS_PCM) {
8447 			rack->r_ctl.pcm_i.send_time = cts;
8448 			rack->r_ctl.pcm_i.eseq = rsm->r_end;
8449 			/* First time through we set the start too */
8450 			if (rack->pcm_in_progress == 0)
8451 				rack->r_ctl.pcm_i.sseq = rsm->r_start;
8452 		}
8453 		/*
8454 		 * Special case detection, is there just a single
8455 		 * packet outstanding when we are not in recovery?
8456 		 *
8457 		 * If this is true mark it so.
8458 		 */
8459 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8460 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8461 			struct rack_sendmap *prsm;
8462 
8463 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8464 			if (prsm)
8465 				prsm->r_one_out_nr = 1;
8466 		}
8467 		return;
8468 	}
8469 	/*
8470 	 * If we reach here its a retransmission and we need to find it.
8471 	 */
8472 more:
8473 	if (hintrsm && (hintrsm->r_start == seq_out)) {
8474 		rsm = hintrsm;
8475 		hintrsm = NULL;
8476 	} else {
8477 		/* No hints sorry */
8478 		rsm = NULL;
8479 	}
8480 	if ((rsm) && (rsm->r_start == seq_out)) {
8481 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8482 		if (len == 0) {
8483 			return;
8484 		} else {
8485 			goto more;
8486 		}
8487 	}
8488 	/* Ok it was not the last pointer go through it the hard way. */
8489 refind:
8490 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8491 	if (rsm) {
8492 		if (rsm->r_start == seq_out) {
8493 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8494 			if (len == 0) {
8495 				return;
8496 			} else {
8497 				goto refind;
8498 			}
8499 		}
8500 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8501 			/* Transmitted within this piece */
8502 			/*
8503 			 * Ok we must split off the front and then let the
8504 			 * update do the rest
8505 			 */
8506 			nrsm = rack_alloc_full_limit(rack);
8507 			if (nrsm == NULL) {
8508 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8509 				return;
8510 			}
8511 			/*
8512 			 * copy rsm to nrsm and then trim the front of rsm
8513 			 * to not include this part.
8514 			 */
8515 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
8516 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8517 #ifndef INVARIANTS
8518 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8519 #else
8520 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8521 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8522 				      nrsm, insret, rack, rsm);
8523 			}
8524 #endif
8525 			if (rsm->r_in_tmap) {
8526 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8527 				nrsm->r_in_tmap = 1;
8528 			}
8529 			rsm->r_flags &= (~RACK_HAS_FIN);
8530 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8531 			if (len == 0) {
8532 				return;
8533 			} else if (len > 0)
8534 				goto refind;
8535 		}
8536 	}
8537 	/*
8538 	 * Hmm not found in map did they retransmit both old and on into the
8539 	 * new?
8540 	 */
8541 	if (seq_out == tp->snd_max) {
8542 		goto again;
8543 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
8544 #ifdef INVARIANTS
8545 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8546 		       seq_out, len, tp->snd_una, tp->snd_max);
8547 		printf("Starting Dump of all rack entries\n");
8548 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8549 			printf("rsm:%p start:%u end:%u\n",
8550 			       rsm, rsm->r_start, rsm->r_end);
8551 		}
8552 		printf("Dump complete\n");
8553 		panic("seq_out not found rack:%p tp:%p",
8554 		      rack, tp);
8555 #endif
8556 	} else {
8557 #ifdef INVARIANTS
8558 		/*
8559 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8560 		 * flag)
8561 		 */
8562 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8563 		      seq_out, len, tp->snd_max, tp);
8564 #endif
8565 	}
8566 }
8567 
8568 /*
8569  * Record one of the RTT updates from an ack into
8570  * our sample structure.
8571  */
8572 
8573 static void
8574 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8575 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8576 {
8577 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8578 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8579 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8580 	}
8581 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8582 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8583 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8584 	}
8585 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8586 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8587 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
8588 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8589 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8590 	}
8591 	if ((confidence == 1) &&
8592 	    ((rsm == NULL) ||
8593 	     (rsm->r_just_ret) ||
8594 	     (rsm->r_one_out_nr &&
8595 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8596 		/*
8597 		 * If the rsm had a just return
8598 		 * hit it then we can't trust the
8599 		 * rtt measurement for buffer deterimination
8600 		 * Note that a confidence of 2, indicates
8601 		 * SACK'd which overrides the r_just_ret or
8602 		 * the r_one_out_nr. If it was a CUM-ACK and
8603 		 * we had only two outstanding, but get an
8604 		 * ack for only 1. Then that also lowers our
8605 		 * confidence.
8606 		 */
8607 		confidence = 0;
8608 	}
8609 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8610 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8611 		if (rack->r_ctl.rack_rs.confidence == 0) {
8612 			/*
8613 			 * We take anything with no current confidence
8614 			 * saved.
8615 			 */
8616 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8617 			rack->r_ctl.rack_rs.confidence = confidence;
8618 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8619 		} else if (confidence != 0) {
8620 			/*
8621 			 * Once we have a confident number,
8622 			 * we can update it with a smaller
8623 			 * value since this confident number
8624 			 * may include the DSACK time until
8625 			 * the next segment (the second one) arrived.
8626 			 */
8627 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8628 			rack->r_ctl.rack_rs.confidence = confidence;
8629 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8630 		}
8631 	}
8632 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8633 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8634 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8635 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
8636 }
8637 
8638 /*
8639  * Collect new round-trip time estimate
8640  * and update averages and current timeout.
8641  */
8642 static void
8643 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8644 {
8645 	int32_t delta;
8646 	int32_t rtt;
8647 
8648 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8649 		/* No valid sample */
8650 		return;
8651 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8652 		/* We are to use the lowest RTT seen in a single ack */
8653 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8654 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8655 		/* We are to use the highest RTT seen in a single ack */
8656 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8657 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8658 		/* We are to use the average RTT seen in a single ack */
8659 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8660 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8661 	} else {
8662 #ifdef INVARIANTS
8663 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8664 #endif
8665 		return;
8666 	}
8667 	if (rtt == 0)
8668 		rtt = 1;
8669 	if (rack->rc_gp_rtt_set == 0) {
8670 		/*
8671 		 * With no RTT we have to accept
8672 		 * even one we are not confident of.
8673 		 */
8674 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8675 		rack->rc_gp_rtt_set = 1;
8676 	} else if (rack->r_ctl.rack_rs.confidence) {
8677 		/* update the running gp srtt */
8678 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8679 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8680 	}
8681 	if (rack->r_ctl.rack_rs.confidence) {
8682 		/*
8683 		 * record the low and high for highly buffered path computation,
8684 		 * we only do this if we are confident (not a retransmission).
8685 		 */
8686 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8687 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8688 		}
8689 		if (rack->rc_highly_buffered == 0) {
8690 			/*
8691 			 * Currently once we declare a path has
8692 			 * highly buffered there is no going
8693 			 * back, which may be a problem...
8694 			 */
8695 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8696 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8697 						     rack->r_ctl.rc_highest_us_rtt,
8698 						     rack->r_ctl.rc_lowest_us_rtt,
8699 						     RACK_RTTS_SEEHBP);
8700 				rack->rc_highly_buffered = 1;
8701 			}
8702 		}
8703 	}
8704 	if ((rack->r_ctl.rack_rs.confidence) ||
8705 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8706 		/*
8707 		 * If we are highly confident of it <or> it was
8708 		 * never retransmitted we accept it as the last us_rtt.
8709 		 */
8710 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8711 		/* The lowest rtt can be set if its was not retransmited */
8712 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8713 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8714 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
8715 				rack->r_ctl.rc_lowest_us_rtt = 1;
8716 		}
8717 	}
8718 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8719 	if (tp->t_srtt != 0) {
8720 		/*
8721 		 * We keep a simple srtt in microseconds, like our rtt
8722 		 * measurement. We don't need to do any tricks with shifting
8723 		 * etc. Instead we just add in 1/8th of the new measurement
8724 		 * and subtract out 1/8 of the old srtt. We do the same with
8725 		 * the variance after finding the absolute value of the
8726 		 * difference between this sample and the current srtt.
8727 		 */
8728 		delta = tp->t_srtt - rtt;
8729 		/* Take off 1/8th of the current sRTT */
8730 		tp->t_srtt -= (tp->t_srtt >> 3);
8731 		/* Add in 1/8th of the new RTT just measured */
8732 		tp->t_srtt += (rtt >> 3);
8733 		if (tp->t_srtt <= 0)
8734 			tp->t_srtt = 1;
8735 		/* Now lets make the absolute value of the variance */
8736 		if (delta < 0)
8737 			delta = -delta;
8738 		/* Subtract out 1/8th */
8739 		tp->t_rttvar -= (tp->t_rttvar >> 3);
8740 		/* Add in 1/8th of the new variance we just saw */
8741 		tp->t_rttvar += (delta >> 3);
8742 		if (tp->t_rttvar <= 0)
8743 			tp->t_rttvar = 1;
8744 	} else {
8745 		/*
8746 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
8747 		 * variance to half the rtt (so our first retransmit happens
8748 		 * at 3*rtt).
8749 		 */
8750 		tp->t_srtt = rtt;
8751 		tp->t_rttvar = rtt >> 1;
8752 	}
8753 	rack->rc_srtt_measure_made = 1;
8754 	KMOD_TCPSTAT_INC(tcps_rttupdated);
8755 	if (tp->t_rttupdated < UCHAR_MAX)
8756 		tp->t_rttupdated++;
8757 #ifdef STATS
8758 	if (rack_stats_gets_ms_rtt == 0) {
8759 		/* Send in the microsecond rtt used for rxt timeout purposes */
8760 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8761 	} else if (rack_stats_gets_ms_rtt == 1) {
8762 		/* Send in the millisecond rtt used for rxt timeout purposes */
8763 		int32_t ms_rtt;
8764 
8765 		/* Round up */
8766 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8767 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8768 	} else if (rack_stats_gets_ms_rtt == 2) {
8769 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
8770 		int32_t ms_rtt;
8771 
8772 		/* Round up */
8773 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8774 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8775 	}  else {
8776 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
8777 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8778 	}
8779 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8780 #endif
8781 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_msec(&rack->r_ctl.act_rcv_time);
8782 	/*
8783 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8784 	 * way we do the smoothing, srtt and rttvar will each average +1/2
8785 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
8786 	 * tick of rounding and 1 extra tick because of +-1/2 tick
8787 	 * uncertainty in the firing of the timer.  The bias will give us
8788 	 * exactly the 1.5 tick we need.  But, because the bias is
8789 	 * statistical, we have to test that we don't drop below the minimum
8790 	 * feasible timer (which is 2 ticks).
8791 	 */
8792 	tp->t_rxtshift = 0;
8793 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8794 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8795 	rack_log_rtt_sample(rack, rtt);
8796 	tp->t_softerror = 0;
8797 }
8798 
8799 
8800 static void
8801 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8802 {
8803 	/*
8804 	 * Apply to filter the inbound us-rtt at us_cts.
8805 	 */
8806 	uint32_t old_rtt;
8807 
8808 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8809 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8810 			       us_rtt, us_cts);
8811 	if (old_rtt > us_rtt) {
8812 		/* We just hit a new lower rtt time */
8813 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8814 				     __LINE__, RACK_RTTS_NEWRTT);
8815 		/*
8816 		 * Only count it if its lower than what we saw within our
8817 		 * calculated range.
8818 		 */
8819 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8820 			if (rack_probertt_lower_within &&
8821 			    rack->rc_gp_dyn_mul &&
8822 			    (rack->use_fixed_rate == 0) &&
8823 			    (rack->rc_always_pace)) {
8824 				/*
8825 				 * We are seeing a new lower rtt very close
8826 				 * to the time that we would have entered probe-rtt.
8827 				 * This is probably due to the fact that a peer flow
8828 				 * has entered probe-rtt. Lets go in now too.
8829 				 */
8830 				uint32_t val;
8831 
8832 				val = rack_probertt_lower_within * rack_time_between_probertt;
8833 				val /= 100;
8834 				if ((rack->in_probe_rtt == 0) &&
8835 				    (rack->rc_skip_timely == 0) &&
8836 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
8837 					rack_enter_probertt(rack, us_cts);
8838 				}
8839 			}
8840 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8841 		}
8842 	}
8843 }
8844 
8845 static int
8846 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8847     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8848 {
8849 	uint32_t us_rtt;
8850 	int32_t i, all;
8851 	uint32_t t, len_acked;
8852 
8853 	if ((rsm->r_flags & RACK_ACKED) ||
8854 	    (rsm->r_flags & RACK_WAS_ACKED))
8855 		/* Already done */
8856 		return (0);
8857 	if (rsm->r_no_rtt_allowed) {
8858 		/* Not allowed */
8859 		return (0);
8860 	}
8861 	if (ack_type == CUM_ACKED) {
8862 		if (SEQ_GT(th_ack, rsm->r_end)) {
8863 			len_acked = rsm->r_end - rsm->r_start;
8864 			all = 1;
8865 		} else {
8866 			len_acked = th_ack - rsm->r_start;
8867 			all = 0;
8868 		}
8869 	} else {
8870 		len_acked = rsm->r_end - rsm->r_start;
8871 		all = 0;
8872 	}
8873 	if (rsm->r_rtr_cnt == 1) {
8874 
8875 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8876 		if ((int)t <= 0)
8877 			t = 1;
8878 		if (!tp->t_rttlow || tp->t_rttlow > t)
8879 			tp->t_rttlow = t;
8880 		if (!rack->r_ctl.rc_rack_min_rtt ||
8881 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8882 			rack->r_ctl.rc_rack_min_rtt = t;
8883 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8884 				rack->r_ctl.rc_rack_min_rtt = 1;
8885 			}
8886 		}
8887 		if (TSTMP_GT(tcp_tv_to_usec(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8888 			us_rtt = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8889 		else
8890 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8891 		if (us_rtt == 0)
8892 			us_rtt = 1;
8893 		if (CC_ALGO(tp)->rttsample != NULL) {
8894 			/* Kick the RTT to the CC */
8895 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8896 		}
8897 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usec(&rack->r_ctl.act_rcv_time));
8898 		if (ack_type == SACKED) {
8899 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8900 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8901 		} else {
8902 			/*
8903 			 * We need to setup what our confidence
8904 			 * is in this ack.
8905 			 *
8906 			 * If the rsm was app limited and it is
8907 			 * less than a mss in length (the end
8908 			 * of the send) then we have a gap. If we
8909 			 * were app limited but say we were sending
8910 			 * multiple MSS's then we are more confident
8911 			 * int it.
8912 			 *
8913 			 * When we are not app-limited then we see if
8914 			 * the rsm is being included in the current
8915 			 * measurement, we tell this by the app_limited_needs_set
8916 			 * flag.
8917 			 *
8918 			 * Note that being cwnd blocked is not applimited
8919 			 * as well as the pacing delay between packets which
8920 			 * are sending only 1 or 2 MSS's also will show up
8921 			 * in the RTT. We probably need to examine this algorithm
8922 			 * a bit more and enhance it to account for the delay
8923 			 * between rsm's. We could do that by saving off the
8924 			 * pacing delay of each rsm (in an rsm) and then
8925 			 * factoring that in somehow though for now I am
8926 			 * not sure how :)
8927 			 */
8928 			int calc_conf = 0;
8929 
8930 			if (rsm->r_flags & RACK_APP_LIMITED) {
8931 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8932 					calc_conf = 0;
8933 				else
8934 					calc_conf = 1;
8935 			} else if (rack->app_limited_needs_set == 0) {
8936 				calc_conf = 1;
8937 			} else {
8938 				calc_conf = 0;
8939 			}
8940 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8941 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8942 					    calc_conf, rsm, rsm->r_rtr_cnt);
8943 		}
8944 		if ((rsm->r_flags & RACK_TLP) &&
8945 		    (!IN_FASTRECOVERY(tp->t_flags))) {
8946 			/* Segment was a TLP and our retrans matched */
8947 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8948 				rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
8949 			}
8950 		}
8951 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8952 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8953 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8954 			/* New more recent rack_tmit_time */
8955 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8956 			if (rack->r_ctl.rc_rack_tmit_time == 0)
8957 				rack->r_ctl.rc_rack_tmit_time = 1;
8958 			rack->rc_rack_rtt = t;
8959 		}
8960 		return (1);
8961 	}
8962 	/*
8963 	 * We clear the soft/rxtshift since we got an ack.
8964 	 * There is no assurance we will call the commit() function
8965 	 * so we need to clear these to avoid incorrect handling.
8966 	 */
8967 	tp->t_rxtshift = 0;
8968 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8969 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8970 	tp->t_softerror = 0;
8971 	if (to && (to->to_flags & TOF_TS) &&
8972 	    (ack_type == CUM_ACKED) &&
8973 	    (to->to_tsecr) &&
8974 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8975 		/*
8976 		 * Now which timestamp does it match? In this block the ACK
8977 		 * must be coming from a previous transmission.
8978 		 */
8979 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
8980 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8981 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8982 				if ((int)t <= 0)
8983 					t = 1;
8984 				if (CC_ALGO(tp)->rttsample != NULL) {
8985 					/*
8986 					 * Kick the RTT to the CC, here
8987 					 * we lie a bit in that we know the
8988 					 * retransmission is correct even though
8989 					 * we retransmitted. This is because
8990 					 * we match the timestamps.
8991 					 */
8992 					if (TSTMP_GT(tcp_tv_to_usec(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8993 						us_rtt = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8994 					else
8995 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8996 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8997 				}
8998 				if ((i + 1) < rsm->r_rtr_cnt) {
8999 					/*
9000 					 * The peer ack'd from our previous
9001 					 * transmission. We have a spurious
9002 					 * retransmission and thus we dont
9003 					 * want to update our rack_rtt.
9004 					 *
9005 					 * Hmm should there be a CC revert here?
9006 					 *
9007 					 */
9008 					return (0);
9009 				}
9010 				if (!tp->t_rttlow || tp->t_rttlow > t)
9011 					tp->t_rttlow = t;
9012 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9013 					rack->r_ctl.rc_rack_min_rtt = t;
9014 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9015 						rack->r_ctl.rc_rack_min_rtt = 1;
9016 					}
9017 				}
9018 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9019 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9020 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9021 					/* New more recent rack_tmit_time */
9022 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9023 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9024 						rack->r_ctl.rc_rack_tmit_time = 1;
9025 					rack->rc_rack_rtt = t;
9026 				}
9027 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9028 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9029 						    rsm->r_rtr_cnt);
9030 				return (1);
9031 			}
9032 		}
9033 		/* If we are logging log out the sendmap */
9034 		if (tcp_bblogging_on(rack->rc_tp)) {
9035 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9036 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9037 			}
9038 		}
9039 		goto ts_not_found;
9040 	} else {
9041 		/*
9042 		 * Ok its a SACK block that we retransmitted. or a windows
9043 		 * machine without timestamps. We can tell nothing from the
9044 		 * time-stamp since its not there or the time the peer last
9045 		 * received a segment that moved forward its cum-ack point.
9046 		 */
9047 ts_not_found:
9048 		i = rsm->r_rtr_cnt - 1;
9049 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9050 		if ((int)t <= 0)
9051 			t = 1;
9052 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9053 			/*
9054 			 * We retransmitted and the ack came back in less
9055 			 * than the smallest rtt we have observed. We most
9056 			 * likely did an improper retransmit as outlined in
9057 			 * 6.2 Step 2 point 2 in the rack-draft so we
9058 			 * don't want to update our rack_rtt. We in
9059 			 * theory (in future) might want to think about reverting our
9060 			 * cwnd state but we won't for now.
9061 			 */
9062 			return (0);
9063 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9064 			/*
9065 			 * We retransmitted it and the retransmit did the
9066 			 * job.
9067 			 */
9068 			if (!rack->r_ctl.rc_rack_min_rtt ||
9069 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9070 				rack->r_ctl.rc_rack_min_rtt = t;
9071 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9072 					rack->r_ctl.rc_rack_min_rtt = 1;
9073 				}
9074 			}
9075 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9076 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9077 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9078 				/* New more recent rack_tmit_time */
9079 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9080 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9081 					rack->r_ctl.rc_rack_tmit_time = 1;
9082 				rack->rc_rack_rtt = t;
9083 			}
9084 			return (1);
9085 		}
9086 	}
9087 	return (0);
9088 }
9089 
9090 /*
9091  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9092  */
9093 static void
9094 rack_log_sack_passed(struct tcpcb *tp,
9095     struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9096 {
9097 	struct rack_sendmap *nrsm;
9098 	uint32_t thresh;
9099 
9100 	/* Get our rxt threshold for lost consideration */
9101 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9102 	/* Now start looking at rsm's */
9103 	nrsm = rsm;
9104 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9105 	    rack_head, r_tnext) {
9106 		if (nrsm == rsm) {
9107 			/* Skip original segment he is acked */
9108 			continue;
9109 		}
9110 		if (nrsm->r_flags & RACK_ACKED) {
9111 			/*
9112 			 * Skip ack'd segments, though we
9113 			 * should not see these, since tmap
9114 			 * should not have ack'd segments.
9115 			 */
9116 			continue;
9117 		}
9118 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9119 			/*
9120 			 * If the peer dropped the rwnd on
9121 			 * these then we don't worry about them.
9122 			 */
9123 			continue;
9124 		}
9125 		/* Check lost state */
9126 		if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9127 			uint32_t exp;
9128 
9129 			exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9130 			if (TSTMP_LT(exp, cts) || (exp == cts)) {
9131 				/* We consider it lost */
9132 				nrsm->r_flags |= RACK_WAS_LOST;
9133 				rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9134 			}
9135 		}
9136 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9137 			/*
9138 			 * We found one that is already marked
9139 			 * passed, we have been here before and
9140 			 * so all others below this are marked.
9141 			 */
9142 			break;
9143 		}
9144 		nrsm->r_flags |= RACK_SACK_PASSED;
9145 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9146 	}
9147 }
9148 
9149 static void
9150 rack_need_set_test(struct tcpcb *tp,
9151 		   struct tcp_rack *rack,
9152 		   struct rack_sendmap *rsm,
9153 		   tcp_seq th_ack,
9154 		   int line,
9155 		   int use_which)
9156 {
9157 	struct rack_sendmap *s_rsm;
9158 
9159 	if ((tp->t_flags & TF_GPUTINPROG) &&
9160 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9161 		/*
9162 		 * We were app limited, and this ack
9163 		 * butts up or goes beyond the point where we want
9164 		 * to start our next measurement. We need
9165 		 * to record the new gput_ts as here and
9166 		 * possibly update the start sequence.
9167 		 */
9168 		uint32_t seq, ts;
9169 
9170 		if (rsm->r_rtr_cnt > 1) {
9171 			/*
9172 			 * This is a retransmit, can we
9173 			 * really make any assessment at this
9174 			 * point?  We are not really sure of
9175 			 * the timestamp, is it this or the
9176 			 * previous transmission?
9177 			 *
9178 			 * Lets wait for something better that
9179 			 * is not retransmitted.
9180 			 */
9181 			return;
9182 		}
9183 		seq = tp->gput_seq;
9184 		ts = tp->gput_ts;
9185 		rack->app_limited_needs_set = 0;
9186 		tp->gput_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
9187 		/* Do we start at a new end? */
9188 		if ((use_which == RACK_USE_BEG) &&
9189 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9190 			/*
9191 			 * When we get an ACK that just eats
9192 			 * up some of the rsm, we set RACK_USE_BEG
9193 			 * since whats at r_start (i.e. th_ack)
9194 			 * is left unacked and thats where the
9195 			 * measurement now starts.
9196 			 */
9197 			tp->gput_seq = rsm->r_start;
9198 		}
9199 		if ((use_which == RACK_USE_END) &&
9200 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9201 			/*
9202 			 * We use the end when the cumack
9203 			 * is moving forward and completely
9204 			 * deleting the rsm passed so basically
9205 			 * r_end holds th_ack.
9206 			 *
9207 			 * For SACK's we also want to use the end
9208 			 * since this piece just got sacked and
9209 			 * we want to target anything after that
9210 			 * in our measurement.
9211 			 */
9212 			tp->gput_seq = rsm->r_end;
9213 		}
9214 		if (use_which == RACK_USE_END_OR_THACK) {
9215 			/*
9216 			 * special case for ack moving forward,
9217 			 * not a sack, we need to move all the
9218 			 * way up to where this ack cum-ack moves
9219 			 * to.
9220 			 */
9221 			if (SEQ_GT(th_ack, rsm->r_end))
9222 				tp->gput_seq = th_ack;
9223 			else
9224 				tp->gput_seq = rsm->r_end;
9225 		}
9226 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9227 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9228 		else
9229 			s_rsm = NULL;
9230 		/*
9231 		 * Pick up the correct send time if we can the rsm passed in
9232 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9233 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9234 		 * find a different seq i.e. the next send up.
9235 		 *
9236 		 * If that has not been sent, s_rsm will be NULL and we must
9237 		 * arrange it so this function will get called again by setting
9238 		 * app_limited_needs_set.
9239 		 */
9240 		if (s_rsm)
9241 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9242 		else {
9243 			/* If we hit here we have to have *not* sent tp->gput_seq */
9244 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9245 			/* Set it up so we will go through here again */
9246 			rack->app_limited_needs_set = 1;
9247 		}
9248 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9249 			/*
9250 			 * We moved beyond this guy's range, re-calculate
9251 			 * the new end point.
9252 			 */
9253 			if (rack->rc_gp_filled == 0) {
9254 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9255 			} else {
9256 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9257 			}
9258 		}
9259 		/*
9260 		 * We are moving the goal post, we may be able to clear the
9261 		 * measure_saw_probe_rtt flag.
9262 		 */
9263 		if ((rack->in_probe_rtt == 0) &&
9264 		    (rack->measure_saw_probe_rtt) &&
9265 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9266 			rack->measure_saw_probe_rtt = 0;
9267 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9268 					   seq, tp->gput_seq,
9269 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9270 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9271 					   5, line, NULL, 0);
9272 		if (rack->rc_gp_filled &&
9273 		    ((tp->gput_ack - tp->gput_seq) <
9274 		     max(rc_init_window(rack), (MIN_GP_WIN *
9275 						ctf_fixed_maxseg(tp))))) {
9276 			uint32_t ideal_amount;
9277 
9278 			ideal_amount = rack_get_measure_window(tp, rack);
9279 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9280 				/*
9281 				 * There is no sense of continuing this measurement
9282 				 * because its too small to gain us anything we
9283 				 * trust. Skip it and that way we can start a new
9284 				 * measurement quicker.
9285 				 */
9286 				tp->t_flags &= ~TF_GPUTINPROG;
9287 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9288 							   0, 0,
9289 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9290 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9291 							   6, __LINE__, NULL, 0);
9292 			} else {
9293 				/*
9294 				 * Reset the window further out.
9295 				 */
9296 				tp->gput_ack = tp->gput_seq + ideal_amount;
9297 			}
9298 		}
9299 		rack_tend_gp_marks(tp, rack);
9300 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9301 	}
9302 }
9303 
9304 static inline int
9305 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9306 {
9307 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9308 		/* Behind our TLP definition or right at */
9309 		return (0);
9310 	}
9311 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9312 		/* The start is beyond or right at our end of TLP definition */
9313 		return (0);
9314 	}
9315 	/* It has to be a sub-part of the original TLP recorded */
9316 	return (1);
9317 }
9318 
9319 static uint32_t
9320 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9321 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9322 		   uint32_t segsiz)
9323 {
9324 	uint32_t start, end, changed = 0;
9325 	struct rack_sendmap stack_map;
9326 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
9327 	int insret __diagused;
9328 	int32_t used_ref = 1;
9329 	int can_use_hookery = 0;
9330 
9331 	start = sack->start;
9332 	end = sack->end;
9333 	rsm = *prsm;
9334 
9335 do_rest_ofb:
9336 	if ((rsm == NULL) ||
9337 	    (SEQ_LT(end, rsm->r_start)) ||
9338 	    (SEQ_GEQ(start, rsm->r_end)) ||
9339 	    (SEQ_LT(start, rsm->r_start))) {
9340 		/*
9341 		 * We are not in the right spot,
9342 		 * find the correct spot in the tree.
9343 		 */
9344 		used_ref = 0;
9345 		rsm = tqhash_find(rack->r_ctl.tqh, start);
9346 	}
9347 	if (rsm == NULL) {
9348 		/* TSNH */
9349 		goto out;
9350 	}
9351 	/* Ok we have an ACK for some piece of this rsm */
9352 	if (rsm->r_start != start) {
9353 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9354 			/*
9355 			 * Before any splitting or hookery is
9356 			 * done is it a TLP of interest i.e. rxt?
9357 			 */
9358 			if ((rsm->r_flags & RACK_TLP) &&
9359 			    (rsm->r_rtr_cnt > 1)) {
9360 				/*
9361 				 * We are splitting a rxt TLP, check
9362 				 * if we need to save off the start/end
9363 				 */
9364 				if (rack->rc_last_tlp_acked_set &&
9365 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9366 					/*
9367 					 * We already turned this on since we are inside
9368 					 * the previous one was a partially sack now we
9369 					 * are getting another one (maybe all of it).
9370 					 *
9371 					 */
9372 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9373 					/*
9374 					 * Lets make sure we have all of it though.
9375 					 */
9376 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9377 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9378 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9379 								     rack->r_ctl.last_tlp_acked_end);
9380 					}
9381 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9382 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9383 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9384 								     rack->r_ctl.last_tlp_acked_end);
9385 					}
9386 				} else {
9387 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9388 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9389 					rack->rc_last_tlp_past_cumack = 0;
9390 					rack->rc_last_tlp_acked_set = 1;
9391 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9392 				}
9393 			}
9394 			/**
9395 			 * Need to split this in two pieces the before and after,
9396 			 * the before remains in the map, the after must be
9397 			 * added. In other words we have:
9398 			 * rsm        |--------------|
9399 			 * sackblk        |------->
9400 			 * rsm will become
9401 			 *     rsm    |---|
9402 			 * and nrsm will be  the sacked piece
9403 			 *     nrsm       |----------|
9404 			 *
9405 			 * But before we start down that path lets
9406 			 * see if the sack spans over on top of
9407 			 * the next guy and it is already sacked.
9408 			 *
9409 			 */
9410 			/*
9411 			 * Hookery can only be used if the two entries
9412 			 * are in the same bucket and neither one of
9413 			 * them staddle the bucket line.
9414 			 */
9415 			next = tqhash_next(rack->r_ctl.tqh, rsm);
9416 			if (next &&
9417 			    (rsm->bindex == next->bindex) &&
9418 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9419 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
9420 			    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9421 			    ((next->r_flags & RACK_IS_PCM) == 0) &&
9422 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
9423 			    (next->r_flags & RACK_IN_GP_WIN))
9424 				can_use_hookery = 1;
9425 			else
9426 				can_use_hookery = 0;
9427 			if (next && can_use_hookery &&
9428 			    (next->r_flags & RACK_ACKED) &&
9429 			    SEQ_GEQ(end, next->r_start)) {
9430 				/**
9431 				 * So the next one is already acked, and
9432 				 * we can thus by hookery use our stack_map
9433 				 * to reflect the piece being sacked and
9434 				 * then adjust the two tree entries moving
9435 				 * the start and ends around. So we start like:
9436 				 *  rsm     |------------|             (not-acked)
9437 				 *  next                 |-----------| (acked)
9438 				 *  sackblk        |-------->
9439 				 *  We want to end like so:
9440 				 *  rsm     |------|                   (not-acked)
9441 				 *  next           |-----------------| (acked)
9442 				 *  nrsm           |-----|
9443 				 * Where nrsm is a temporary stack piece we
9444 				 * use to update all the gizmos.
9445 				 */
9446 				/* Copy up our fudge block */
9447 				nrsm = &stack_map;
9448 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9449 				/* Now adjust our tree blocks */
9450 				tqhash_update_end(rack->r_ctl.tqh, rsm, start);
9451 				next->r_start = start;
9452  				rsm->r_flags |= RACK_SHUFFLED;
9453 				next->r_flags |= RACK_SHUFFLED;
9454 				/* Now we must adjust back where next->m is */
9455 				rack_setup_offset_for_rsm(rack, rsm, next);
9456 				/*
9457 				 * Which timestamp do we keep? It is rather
9458 				 * important in GP measurements to have the
9459 				 * accurate end of the send window.
9460 				 *
9461 				 * We keep the largest value, which is the newest
9462 				 * send. We do this in case a segment that is
9463 				 * joined together and not part of a GP estimate
9464 				 * later gets expanded into the GP estimate.
9465 				 *
9466 				 * We prohibit the merging of unlike kinds i.e.
9467 				 * all pieces that are in the GP estimate can be
9468 				 * merged and all pieces that are not in a GP estimate
9469 				 * can be merged, but not disimilar pieces. Combine
9470 				 * this with taking the highest here and we should
9471 				 * be ok unless of course the client reneges. Then
9472 				 * all bets are off.
9473 				 */
9474 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9475 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9476 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9477 				/*
9478 				 * And we must keep the newest ack arrival time.
9479 				 */
9480 				if (next->r_ack_arrival <
9481 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9482 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9483 
9484 
9485 				/* We don't need to adjust rsm, it did not change */
9486 				/* Clear out the dup ack count of the remainder */
9487 				rsm->r_dupack = 0;
9488 				rsm->r_just_ret = 0;
9489 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9490 				/* Now lets make sure our fudge block is right */
9491 				nrsm->r_start = start;
9492 				/* Now lets update all the stats and such */
9493 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9494 				if (rack->app_limited_needs_set)
9495 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9496 				changed += (nrsm->r_end - nrsm->r_start);
9497 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9498 				if (rsm->r_flags & RACK_WAS_LOST) {
9499 					int my_chg;
9500 
9501 					my_chg = (nrsm->r_end - nrsm->r_start);
9502 					KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9503 						("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9504 					if (my_chg <= rack->r_ctl.rc_considered_lost)
9505 						rack->r_ctl.rc_considered_lost -= my_chg;
9506 					else
9507 						rack->r_ctl.rc_considered_lost = 0;
9508 				}
9509 				if (nrsm->r_flags & RACK_SACK_PASSED) {
9510 					rack->r_ctl.rc_reorder_ts = cts;
9511 					if (rack->r_ctl.rc_reorder_ts == 0)
9512 						rack->r_ctl.rc_reorder_ts = 1;
9513 				}
9514 				/*
9515 				 * Now we want to go up from rsm (the
9516 				 * one left un-acked) to the next one
9517 				 * in the tmap. We do this so when
9518 				 * we walk backwards we include marking
9519 				 * sack-passed on rsm (The one passed in
9520 				 * is skipped since it is generally called
9521 				 * on something sacked before removing it
9522 				 * from the tmap).
9523 				 */
9524 				if (rsm->r_in_tmap) {
9525 					nrsm = TAILQ_NEXT(rsm, r_tnext);
9526 					/*
9527 					 * Now that we have the next
9528 					 * one walk backwards from there.
9529 					 */
9530 					if (nrsm && nrsm->r_in_tmap)
9531 						rack_log_sack_passed(tp, rack, nrsm, cts);
9532 				}
9533 				/* Now are we done? */
9534 				if (SEQ_LT(end, next->r_end) ||
9535 				    (end == next->r_end)) {
9536 					/* Done with block */
9537 					goto out;
9538 				}
9539 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9540 				counter_u64_add(rack_sack_used_next_merge, 1);
9541 				/* Postion for the next block */
9542 				start = next->r_end;
9543 				rsm = tqhash_next(rack->r_ctl.tqh, next);
9544 				if (rsm == NULL)
9545 					goto out;
9546 			} else {
9547 				/**
9548 				 * We can't use any hookery here, so we
9549 				 * need to split the map. We enter like
9550 				 * so:
9551 				 *  rsm      |--------|
9552 				 *  sackblk       |----->
9553 				 * We will add the new block nrsm and
9554 				 * that will be the new portion, and then
9555 				 * fall through after reseting rsm. So we
9556 				 * split and look like this:
9557 				 *  rsm      |----|
9558 				 *  sackblk       |----->
9559 				 *  nrsm          |---|
9560 				 * We then fall through reseting
9561 				 * rsm to nrsm, so the next block
9562 				 * picks it up.
9563 				 */
9564 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9565 				if (nrsm == NULL) {
9566 					/*
9567 					 * failed XXXrrs what can we do but loose the sack
9568 					 * info?
9569 					 */
9570 					goto out;
9571 				}
9572 				counter_u64_add(rack_sack_splits, 1);
9573 				rack_clone_rsm(rack, nrsm, rsm, start);
9574 				rsm->r_just_ret = 0;
9575 #ifndef INVARIANTS
9576 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9577 #else
9578 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9579 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9580 					      nrsm, insret, rack, rsm);
9581 				}
9582 #endif
9583 				if (rsm->r_in_tmap) {
9584 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9585 					nrsm->r_in_tmap = 1;
9586 				}
9587 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9588 				rsm->r_flags &= (~RACK_HAS_FIN);
9589 				/* Position us to point to the new nrsm that starts the sack blk */
9590 				rsm = nrsm;
9591 			}
9592 		} else {
9593 			/* Already sacked this piece */
9594 			counter_u64_add(rack_sack_skipped_acked, 1);
9595 			if (end == rsm->r_end) {
9596 				/* Done with block */
9597 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9598 				goto out;
9599 			} else if (SEQ_LT(end, rsm->r_end)) {
9600 				/* A partial sack to a already sacked block */
9601 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9602 				goto out;
9603 			} else {
9604 				/*
9605 				 * The end goes beyond this guy
9606 				 * reposition the start to the
9607 				 * next block.
9608 				 */
9609 				start = rsm->r_end;
9610 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9611 				if (rsm == NULL)
9612 					goto out;
9613 			}
9614 		}
9615 	}
9616 	if (SEQ_GEQ(end, rsm->r_end)) {
9617 		/**
9618 		 * The end of this block is either beyond this guy or right
9619 		 * at this guy. I.e.:
9620 		 *  rsm ---                 |-----|
9621 		 *  end                     |-----|
9622 		 *  <or>
9623 		 *  end                     |---------|
9624 		 */
9625 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9626 			/*
9627 			 * Is it a TLP of interest?
9628 			 */
9629 			if ((rsm->r_flags & RACK_TLP) &&
9630 			    (rsm->r_rtr_cnt > 1)) {
9631 				/*
9632 				 * We are splitting a rxt TLP, check
9633 				 * if we need to save off the start/end
9634 				 */
9635 				if (rack->rc_last_tlp_acked_set &&
9636 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9637 					/*
9638 					 * We already turned this on since we are inside
9639 					 * the previous one was a partially sack now we
9640 					 * are getting another one (maybe all of it).
9641 					 */
9642 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9643 					/*
9644 					 * Lets make sure we have all of it though.
9645 					 */
9646 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9647 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9648 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9649 								     rack->r_ctl.last_tlp_acked_end);
9650 					}
9651 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9652 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9653 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9654 								     rack->r_ctl.last_tlp_acked_end);
9655 					}
9656 				} else {
9657 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9658 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9659 					rack->rc_last_tlp_past_cumack = 0;
9660 					rack->rc_last_tlp_acked_set = 1;
9661 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9662 				}
9663 			}
9664 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9665 			changed += (rsm->r_end - rsm->r_start);
9666 			/* You get a count for acking a whole segment or more */
9667 			if (rsm->r_flags & RACK_WAS_LOST) {
9668 				int my_chg;
9669 
9670 				my_chg = (rsm->r_end - rsm->r_start);
9671 				rsm->r_flags &= ~RACK_WAS_LOST;
9672 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9673 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9674 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9675 					rack->r_ctl.rc_considered_lost -= my_chg;
9676 				else
9677 					rack->r_ctl.rc_considered_lost = 0;
9678 			}
9679 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9680 			if (rsm->r_in_tmap) /* should be true */
9681 				rack_log_sack_passed(tp, rack, rsm, cts);
9682 			/* Is Reordering occuring? */
9683 			if (rsm->r_flags & RACK_SACK_PASSED) {
9684 				rsm->r_flags &= ~RACK_SACK_PASSED;
9685 				rack->r_ctl.rc_reorder_ts = cts;
9686 				if (rack->r_ctl.rc_reorder_ts == 0)
9687 					rack->r_ctl.rc_reorder_ts = 1;
9688 			}
9689 			if (rack->app_limited_needs_set)
9690 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9691 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9692 			rsm->r_flags |= RACK_ACKED;
9693 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9694 			if (rsm->r_in_tmap) {
9695 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9696 				rsm->r_in_tmap = 0;
9697 			}
9698 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9699 		} else {
9700 			counter_u64_add(rack_sack_skipped_acked, 1);
9701 		}
9702 		if (end == rsm->r_end) {
9703 			/* This block only - done, setup for next */
9704 			goto out;
9705 		}
9706 		/*
9707 		 * There is more not coverend by this rsm move on
9708 		 * to the next block in the tail queue hash table.
9709 		 */
9710 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9711 		start = rsm->r_end;
9712 		rsm = nrsm;
9713 		if (rsm == NULL)
9714 			goto out;
9715 		goto do_rest_ofb;
9716 	}
9717 	/**
9718 	 * The end of this sack block is smaller than
9719 	 * our rsm i.e.:
9720 	 *  rsm ---                 |-----|
9721 	 *  end                     |--|
9722 	 */
9723 	if ((rsm->r_flags & RACK_ACKED) == 0) {
9724 		/*
9725 		 * Is it a TLP of interest?
9726 		 */
9727 		if ((rsm->r_flags & RACK_TLP) &&
9728 		    (rsm->r_rtr_cnt > 1)) {
9729 			/*
9730 			 * We are splitting a rxt TLP, check
9731 			 * if we need to save off the start/end
9732 			 */
9733 			if (rack->rc_last_tlp_acked_set &&
9734 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9735 				/*
9736 				 * We already turned this on since we are inside
9737 				 * the previous one was a partially sack now we
9738 				 * are getting another one (maybe all of it).
9739 				 */
9740 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9741 				/*
9742 				 * Lets make sure we have all of it though.
9743 				 */
9744 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9745 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9746 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9747 							     rack->r_ctl.last_tlp_acked_end);
9748 				}
9749 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9750 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9751 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9752 							     rack->r_ctl.last_tlp_acked_end);
9753 				}
9754 			} else {
9755 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9756 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9757 				rack->rc_last_tlp_past_cumack = 0;
9758 				rack->rc_last_tlp_acked_set = 1;
9759 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9760 			}
9761 		}
9762 		/*
9763 		 * Hookery can only be used if the two entries
9764 		 * are in the same bucket and neither one of
9765 		 * them staddle the bucket line.
9766 		 */
9767 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9768 		if (prev &&
9769 		    (rsm->bindex == prev->bindex) &&
9770 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9771 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
9772 		    ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9773 		    ((prev->r_flags & RACK_IS_PCM) == 0) &&
9774 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
9775 		    (prev->r_flags & RACK_IN_GP_WIN))
9776 			can_use_hookery = 1;
9777 		else
9778 			can_use_hookery = 0;
9779 		if (prev && can_use_hookery &&
9780 		    (prev->r_flags & RACK_ACKED)) {
9781 			/**
9782 			 * Goal, we want the right remainder of rsm to shrink
9783 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
9784 			 * We want to expand prev to go all the way
9785 			 * to prev->r_end <- end.
9786 			 * so in the tree we have before:
9787 			 *   prev     |--------|         (acked)
9788 			 *   rsm               |-------| (non-acked)
9789 			 *   sackblk           |-|
9790 			 * We churn it so we end up with
9791 			 *   prev     |----------|       (acked)
9792 			 *   rsm                 |-----| (non-acked)
9793 			 *   nrsm              |-| (temporary)
9794 			 *
9795 			 * Note if either prev/rsm is a TLP we don't
9796 			 * do this.
9797 			 */
9798 			nrsm = &stack_map;
9799 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9800 			tqhash_update_end(rack->r_ctl.tqh, prev, end);
9801 			rsm->r_start = end;
9802 			rsm->r_flags |= RACK_SHUFFLED;
9803 			prev->r_flags |= RACK_SHUFFLED;
9804 			/* Now adjust nrsm (stack copy) to be
9805 			 * the one that is the small
9806 			 * piece that was "sacked".
9807 			 */
9808 			nrsm->r_end = end;
9809 			rsm->r_dupack = 0;
9810 			/*
9811 			 * Which timestamp do we keep? It is rather
9812 			 * important in GP measurements to have the
9813 			 * accurate end of the send window.
9814 			 *
9815 			 * We keep the largest value, which is the newest
9816 			 * send. We do this in case a segment that is
9817 			 * joined together and not part of a GP estimate
9818 			 * later gets expanded into the GP estimate.
9819 			 *
9820 			 * We prohibit the merging of unlike kinds i.e.
9821 			 * all pieces that are in the GP estimate can be
9822 			 * merged and all pieces that are not in a GP estimate
9823 			 * can be merged, but not disimilar pieces. Combine
9824 			 * this with taking the highest here and we should
9825 			 * be ok unless of course the client reneges. Then
9826 			 * all bets are off.
9827 			 */
9828 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9829 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9830 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9831 			}
9832 			/*
9833 			 * And we must keep the newest ack arrival time.
9834 			 */
9835 
9836 			if(prev->r_ack_arrival <
9837 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9838 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9839 
9840 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9841 			/*
9842 			 * Now that the rsm has had its start moved forward
9843 			 * lets go ahead and get its new place in the world.
9844 			 */
9845 			rack_setup_offset_for_rsm(rack, prev, rsm);
9846 			/*
9847 			 * Now nrsm is our new little piece
9848 			 * that is acked (which was merged
9849 			 * to prev). Update the rtt and changed
9850 			 * based on that. Also check for reordering.
9851 			 */
9852 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9853 			if (rack->app_limited_needs_set)
9854 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9855 			changed += (nrsm->r_end - nrsm->r_start);
9856 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9857 			if (rsm->r_flags & RACK_WAS_LOST) {
9858 				int my_chg;
9859 
9860 				my_chg = (nrsm->r_end - nrsm->r_start);
9861 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9862 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9863 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9864 					rack->r_ctl.rc_considered_lost -= my_chg;
9865 				else
9866 					rack->r_ctl.rc_considered_lost = 0;
9867 			}
9868 			if (nrsm->r_flags & RACK_SACK_PASSED) {
9869 				rack->r_ctl.rc_reorder_ts = cts;
9870 				if (rack->r_ctl.rc_reorder_ts == 0)
9871 					rack->r_ctl.rc_reorder_ts = 1;
9872 			}
9873 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9874 			rsm = prev;
9875 			counter_u64_add(rack_sack_used_prev_merge, 1);
9876 		} else {
9877 			/**
9878 			 * This is the case where our previous
9879 			 * block is not acked either, so we must
9880 			 * split the block in two.
9881 			 */
9882 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9883 			if (nrsm == NULL) {
9884 				/* failed rrs what can we do but loose the sack info? */
9885 				goto out;
9886 			}
9887 			if ((rsm->r_flags & RACK_TLP) &&
9888 			    (rsm->r_rtr_cnt > 1)) {
9889 				/*
9890 				 * We are splitting a rxt TLP, check
9891 				 * if we need to save off the start/end
9892 				 */
9893 				if (rack->rc_last_tlp_acked_set &&
9894 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9895 					/*
9896 					 * We already turned this on since this block is inside
9897 					 * the previous one was a partially sack now we
9898 					 * are getting another one (maybe all of it).
9899 					 */
9900 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9901 					/*
9902 					 * Lets make sure we have all of it though.
9903 					 */
9904 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9905 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9906 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9907 								     rack->r_ctl.last_tlp_acked_end);
9908 					}
9909 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9910 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9911 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9912 								     rack->r_ctl.last_tlp_acked_end);
9913 					}
9914 				} else {
9915 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9916 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9917 					rack->rc_last_tlp_acked_set = 1;
9918 					rack->rc_last_tlp_past_cumack = 0;
9919 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9920 				}
9921 			}
9922 			/**
9923 			 * In this case nrsm becomes
9924 			 * nrsm->r_start = end;
9925 			 * nrsm->r_end = rsm->r_end;
9926 			 * which is un-acked.
9927 			 * <and>
9928 			 * rsm->r_end = nrsm->r_start;
9929 			 * i.e. the remaining un-acked
9930 			 * piece is left on the left
9931 			 * hand side.
9932 			 *
9933 			 * So we start like this
9934 			 * rsm      |----------| (not acked)
9935 			 * sackblk  |---|
9936 			 * build it so we have
9937 			 * rsm      |---|         (acked)
9938 			 * nrsm         |------|  (not acked)
9939 			 */
9940 			counter_u64_add(rack_sack_splits, 1);
9941 			rack_clone_rsm(rack, nrsm, rsm, end);
9942 			rsm->r_flags &= (~RACK_HAS_FIN);
9943 			rsm->r_just_ret = 0;
9944 #ifndef INVARIANTS
9945 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9946 #else
9947 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9948 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
9949 				      nrsm, insret, rack, rsm);
9950 			}
9951 #endif
9952 			if (rsm->r_in_tmap) {
9953 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9954 				nrsm->r_in_tmap = 1;
9955 			}
9956 			nrsm->r_dupack = 0;
9957 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
9958 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9959 			changed += (rsm->r_end - rsm->r_start);
9960 			if (rsm->r_flags & RACK_WAS_LOST) {
9961 				int my_chg;
9962 
9963 				my_chg = (rsm->r_end - rsm->r_start);
9964 				rsm->r_flags &= ~RACK_WAS_LOST;
9965 				KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9966 					("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
9967 				if (my_chg <= rack->r_ctl.rc_considered_lost)
9968 					rack->r_ctl.rc_considered_lost -= my_chg;
9969 				else
9970 					rack->r_ctl.rc_considered_lost = 0;
9971 			}
9972 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9973 
9974 			if (rsm->r_in_tmap) /* should be true */
9975 				rack_log_sack_passed(tp, rack, rsm, cts);
9976 			/* Is Reordering occuring? */
9977 			if (rsm->r_flags & RACK_SACK_PASSED) {
9978 				rsm->r_flags &= ~RACK_SACK_PASSED;
9979 				rack->r_ctl.rc_reorder_ts = cts;
9980 				if (rack->r_ctl.rc_reorder_ts == 0)
9981 					rack->r_ctl.rc_reorder_ts = 1;
9982 			}
9983 			if (rack->app_limited_needs_set)
9984 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9985 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9986 			rsm->r_flags |= RACK_ACKED;
9987 			rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9988 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
9989 			if (rsm->r_in_tmap) {
9990 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9991 				rsm->r_in_tmap = 0;
9992 			}
9993 		}
9994 	} else if (start != end){
9995 		/*
9996 		 * The block was already acked.
9997 		 */
9998 		counter_u64_add(rack_sack_skipped_acked, 1);
9999 	}
10000 out:
10001 	if (rsm &&
10002 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10003 	    (rsm->r_flags & RACK_ACKED)) {
10004 		/*
10005 		 * Now can we merge where we worked
10006 		 * with either the previous or
10007 		 * next block?
10008 		 */
10009 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10010 		while (next) {
10011 			if (next->r_flags & RACK_TLP)
10012 				break;
10013 			/* Only allow merges between ones in or out of GP window */
10014 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10015 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10016 				break;
10017 			}
10018 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10019 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10020 				break;
10021 			}
10022 			if (rsm->bindex != next->bindex)
10023 				break;
10024 			if (rsm->r_flags & RACK_STRADDLE)
10025 				break;
10026 			if (rsm->r_flags & RACK_IS_PCM)
10027 				break;
10028 			if (next->r_flags & RACK_STRADDLE)
10029 				break;
10030 			if (next->r_flags & RACK_IS_PCM)
10031 				break;
10032 			if (next->r_flags & RACK_ACKED) {
10033 				/* yep this and next can be merged */
10034 				rsm = rack_merge_rsm(rack, rsm, next);
10035 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10036 			} else
10037 				break;
10038 		}
10039 		/* Now what about the previous? */
10040 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10041 		while (prev) {
10042 			if (prev->r_flags & RACK_TLP)
10043 				break;
10044 			/* Only allow merges between ones in or out of GP window */
10045 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10046 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10047 				break;
10048 			}
10049 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10050 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10051 				break;
10052 			}
10053 			if (rsm->bindex != prev->bindex)
10054 				break;
10055 			if (rsm->r_flags & RACK_STRADDLE)
10056 				break;
10057 			if (rsm->r_flags & RACK_IS_PCM)
10058 				break;
10059 			if (prev->r_flags & RACK_STRADDLE)
10060 				break;
10061 			if (prev->r_flags & RACK_IS_PCM)
10062 				break;
10063 			if (prev->r_flags & RACK_ACKED) {
10064 				/* yep the previous and this can be merged */
10065 				rsm = rack_merge_rsm(rack, prev, rsm);
10066 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10067 			} else
10068 				break;
10069 		}
10070 	}
10071 	if (used_ref == 0) {
10072 		counter_u64_add(rack_sack_proc_all, 1);
10073 	} else {
10074 		counter_u64_add(rack_sack_proc_short, 1);
10075 	}
10076 	/* Save off the next one for quick reference. */
10077 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10078 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10079 	return (changed);
10080 }
10081 
10082 static void inline
10083 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10084 {
10085 	struct rack_sendmap *tmap;
10086 
10087 	tmap = NULL;
10088 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10089 		/* Its no longer sacked, mark it so */
10090 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10091 #ifdef INVARIANTS
10092 		if (rsm->r_in_tmap) {
10093 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10094 			      rack, rsm, rsm->r_flags);
10095 		}
10096 #endif
10097 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10098 		/* Rebuild it into our tmap */
10099 		if (tmap == NULL) {
10100 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10101 			tmap = rsm;
10102 		} else {
10103 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10104 			tmap = rsm;
10105 		}
10106 		tmap->r_in_tmap = 1;
10107 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10108 	}
10109 	/*
10110 	 * Now lets possibly clear the sack filter so we start
10111 	 * recognizing sacks that cover this area.
10112 	 */
10113 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10114 
10115 }
10116 
10117 
10118 static void inline
10119 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10120 {
10121 	/*
10122 	 * We look at advancing the end send time for our GP
10123 	 * measurement tracking only as the cumulative acknowledgment
10124 	 * moves forward. You might wonder about this, why not
10125 	 * at every transmission or retransmission within the
10126 	 * GP window update the rc_gp_cumack_ts? Well its rather
10127 	 * nuanced but basically the GP window *may* expand (as
10128 	 * it does below) or worse and harder to track it may shrink.
10129 	 *
10130 	 * This last makes it impossible to track at the time of
10131 	 * the send, since you may set forward your rc_gp_cumack_ts
10132 	 * when you send, because that send *is* in your currently
10133 	 * "guessed" window, but then it shrinks. Now which was
10134 	 * the send time of the last bytes in the window, by the
10135 	 * time you ask that question that part of the sendmap
10136 	 * is freed. So you don't know and you will have too
10137 	 * long of send window. Instead by updating the time
10138 	 * marker only when the cumack advances this assures us
10139 	 * that we will have only the sends in the window of our
10140 	 * GP measurement.
10141 	 *
10142 	 * Another complication from this is the
10143 	 * merging of sendmap entries. During SACK processing this
10144 	 * can happen to conserve the sendmap size. That breaks
10145 	 * everything down in tracking the send window of the GP
10146 	 * estimate. So to prevent that and keep it working with
10147 	 * a tiny bit more limited merging, we only allow like
10148 	 * types to be merged. I.e. if two sends are in the GP window
10149 	 * then its ok to merge them together. If two sends are not
10150 	 * in the GP window its ok to merge them together too. Though
10151 	 * one send in and one send out cannot be merged. We combine
10152 	 * this with never allowing the shrinking of the GP window when
10153 	 * we are in recovery so that we can properly calculate the
10154 	 * sending times.
10155 	 *
10156 	 * This all of course seems complicated, because it is.. :)
10157 	 *
10158 	 * The cum-ack is being advanced upon the sendmap.
10159 	 * If we are not doing a GP estimate don't
10160 	 * proceed.
10161 	 */
10162 	uint64_t ts;
10163 
10164 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10165 		return;
10166 	/*
10167 	 * If this sendmap entry is going
10168 	 * beyond the measurement window we had picked,
10169 	 * expand the measurement window by that much.
10170 	 */
10171 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10172 		tp->gput_ack = rsm->r_end;
10173 	}
10174 	/*
10175 	 * If we have not setup a ack, then we
10176 	 * have no idea if the newly acked pieces
10177 	 * will be "in our seq measurement range". If
10178 	 * it is when we clear the app_limited_needs_set
10179 	 * flag the timestamp will be updated.
10180 	 */
10181 	if (rack->app_limited_needs_set)
10182 		return;
10183 	/*
10184 	 * Finally, we grab out the latest timestamp
10185 	 * that this packet was sent and then see
10186 	 * if:
10187 	 *  a) The packet touches are newly defined GP range.
10188 	 *  b) The time is greater than (newer) than the
10189 	 *     one we currently have. If so we update
10190 	 *     our sending end time window.
10191 	 *
10192 	 * Note we *do not* do this at send time. The reason
10193 	 * is that if you do you *may* pick up a newer timestamp
10194 	 * for a range you are not going to measure. We project
10195 	 * out how far and then sometimes modify that to be
10196 	 * smaller. If that occurs then you will have a send
10197 	 * that does not belong to the range included.
10198 	 */
10199 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10200 	    rack->r_ctl.rc_gp_cumack_ts)
10201 		return;
10202 	if (rack_in_gp_window(tp, rsm)) {
10203 		rack->r_ctl.rc_gp_cumack_ts = ts;
10204 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10205 			       __LINE__, from, rsm);
10206 	}
10207 }
10208 
10209 static void
10210 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)
10211 {
10212 	struct rack_sendmap *rsm;
10213 	/*
10214 	 * The ACK point is advancing to th_ack, we must drop off
10215 	 * the packets in the rack log and calculate any eligble
10216 	 * RTT's.
10217 	 */
10218 
10219 	if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
10220 		/*
10221 		 * If we have some sack blocks in the filter
10222 		 * lets prune them out by calling sfb with no blocks.
10223 		 */
10224 		sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
10225 	}
10226 	if (SEQ_GT(th_ack, tp->snd_una)) {
10227 		/* Clear any app ack remembered settings */
10228 		rack->r_ctl.cleared_app_ack = 0;
10229 	}
10230 	rack->r_wanted_output = 1;
10231 	if (SEQ_GT(th_ack, tp->snd_una))
10232 		rack->r_ctl.last_cumack_advance = acktime;
10233 
10234 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10235 	if ((rack->rc_last_tlp_acked_set == 1)&&
10236 	    (rack->rc_last_tlp_past_cumack == 1) &&
10237 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10238 		/*
10239 		 * We have reached the point where our last rack
10240 		 * tlp retransmit sequence is ahead of the cum-ack.
10241 		 * This can only happen when the cum-ack moves all
10242 		 * the way around (its been a full 2^^31+1 bytes
10243 		 * or more since we sent a retransmitted TLP). Lets
10244 		 * turn off the valid flag since its not really valid.
10245 		 *
10246 		 * Note since sack's also turn on this event we have
10247 		 * a complication, we have to wait to age it out until
10248 		 * the cum-ack is by the TLP before checking which is
10249 		 * what the next else clause does.
10250 		 */
10251 		rack_log_dsack_event(rack, 9, __LINE__,
10252 				     rack->r_ctl.last_tlp_acked_start,
10253 				     rack->r_ctl.last_tlp_acked_end);
10254 		rack->rc_last_tlp_acked_set = 0;
10255 		rack->rc_last_tlp_past_cumack = 0;
10256 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10257 		   (rack->rc_last_tlp_past_cumack == 0) &&
10258 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10259 		/*
10260 		 * It is safe to start aging TLP's out.
10261 		 */
10262 		rack->rc_last_tlp_past_cumack = 1;
10263 	}
10264 	/* We do the same for the tlp send seq as well */
10265 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10266 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10267 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10268 		rack_log_dsack_event(rack, 9, __LINE__,
10269 				     rack->r_ctl.last_sent_tlp_seq,
10270 				     (rack->r_ctl.last_sent_tlp_seq +
10271 				      rack->r_ctl.last_sent_tlp_len));
10272 		rack->rc_last_sent_tlp_seq_valid = 0;
10273 		rack->rc_last_sent_tlp_past_cumack = 0;
10274 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10275 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10276 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10277 		/*
10278 		 * It is safe to start aging TLP's send.
10279 		 */
10280 		rack->rc_last_sent_tlp_past_cumack = 1;
10281 	}
10282 more:
10283 	rsm = tqhash_min(rack->r_ctl.tqh);
10284 	if (rsm == NULL) {
10285 		if ((th_ack - 1) == tp->iss) {
10286 			/*
10287 			 * For the SYN incoming case we will not
10288 			 * have called tcp_output for the sending of
10289 			 * the SYN, so there will be no map. All
10290 			 * other cases should probably be a panic.
10291 			 */
10292 			return;
10293 		}
10294 		if (tp->t_flags & TF_SENTFIN) {
10295 			/* if we sent a FIN we often will not have map */
10296 			return;
10297 		}
10298 #ifdef INVARIANTS
10299 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
10300 		      tp,
10301 		      tp->t_state, th_ack, rack,
10302 		      tp->snd_una, tp->snd_max);
10303 #endif
10304 		return;
10305 	}
10306 	if (SEQ_LT(th_ack, rsm->r_start)) {
10307 		/* Huh map is missing this */
10308 #ifdef INVARIANTS
10309 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10310 		       rsm->r_start,
10311 		       th_ack, tp->t_state, rack->r_state);
10312 #endif
10313 		return;
10314 	}
10315 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10316 
10317 	/* Now was it a retransmitted TLP? */
10318 	if ((rsm->r_flags & RACK_TLP) &&
10319 	    (rsm->r_rtr_cnt > 1)) {
10320 		/*
10321 		 * Yes, this rsm was a TLP and retransmitted, remember that
10322 		 * since if a DSACK comes back on this we don't want
10323 		 * to think of it as a reordered segment. This may
10324 		 * get updated again with possibly even other TLPs
10325 		 * in flight, but thats ok. Only when we don't send
10326 		 * a retransmitted TLP for 1/2 the sequences space
10327 		 * will it get turned off (above).
10328 		 */
10329 		if (rack->rc_last_tlp_acked_set &&
10330 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10331 			/*
10332 			 * We already turned this on since the end matches,
10333 			 * the previous one was a partially ack now we
10334 			 * are getting another one (maybe all of it).
10335 			 */
10336 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10337 			/*
10338 			 * Lets make sure we have all of it though.
10339 			 */
10340 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10341 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10342 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10343 						     rack->r_ctl.last_tlp_acked_end);
10344 			}
10345 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10346 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10347 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10348 						     rack->r_ctl.last_tlp_acked_end);
10349 			}
10350 		} else {
10351 			rack->rc_last_tlp_past_cumack = 1;
10352 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10353 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10354 			rack->rc_last_tlp_acked_set = 1;
10355 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10356 		}
10357 	}
10358 	/* Now do we consume the whole thing? */
10359 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10360 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
10361 		/* Its all consumed. */
10362 		uint32_t left;
10363 		uint8_t newly_acked;
10364 
10365 		if (rsm->r_flags & RACK_WAS_LOST) {
10366 			/*
10367 			 * This can happen when we marked it as lost
10368 			 * and yet before retransmitting we get an ack
10369 			 * which can happen due to reordering.
10370 			 */
10371 			rsm->r_flags &= ~RACK_WAS_LOST;
10372 			KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
10373 				("rsm:%p rack:%p rc_considered_lost goes negative", rsm,  rack));
10374 			if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
10375 				rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
10376 			else
10377 				rack->r_ctl.rc_considered_lost = 0;
10378 		}
10379 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10380 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10381 		rsm->r_rtr_bytes = 0;
10382 		/*
10383 		 * Record the time of highest cumack sent if its in our measurement
10384 		 * window and possibly bump out the end.
10385 		 */
10386 		rack_rsm_sender_update(rack, tp, rsm, 4);
10387 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10388 		if (rsm->r_in_tmap) {
10389 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10390 			rsm->r_in_tmap = 0;
10391 		}
10392 		newly_acked = 1;
10393 		if (rsm->r_flags & RACK_ACKED) {
10394 			/*
10395 			 * It was acked on the scoreboard -- remove
10396 			 * it from total
10397 			 */
10398 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10399 			newly_acked = 0;
10400 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
10401 			/*
10402 			 * There are segments ACKED on the
10403 			 * scoreboard further up. We are seeing
10404 			 * reordering.
10405 			 */
10406 			rsm->r_flags &= ~RACK_SACK_PASSED;
10407 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10408 			rsm->r_flags |= RACK_ACKED;
10409 			rack->r_ctl.rc_reorder_ts = cts;
10410 			if (rack->r_ctl.rc_reorder_ts == 0)
10411 				rack->r_ctl.rc_reorder_ts = 1;
10412 			if (rack->r_ent_rec_ns) {
10413 				/*
10414 				 * We have sent no more, and we saw an sack
10415 				 * then ack arrive.
10416 				 */
10417 				rack->r_might_revert = 1;
10418 			}
10419 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10420 		} else {
10421 			rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10422 		}
10423 		if ((rsm->r_flags & RACK_TO_REXT) &&
10424 		    (tp->t_flags & TF_RCVD_TSTMP) &&
10425 		    (to->to_flags & TOF_TS) &&
10426 		    (to->to_tsecr != 0) &&
10427 		    (tp->t_flags & TF_PREVVALID)) {
10428 			/*
10429 			 * We can use the timestamp to see
10430 			 * if this retransmission was from the
10431 			 * first transmit. If so we made a mistake.
10432 			 */
10433 			tp->t_flags &= ~TF_PREVVALID;
10434 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10435 				/* The first transmit is what this ack is for */
10436 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10437 			}
10438 		}
10439 		left = th_ack - rsm->r_end;
10440 		if (rack->app_limited_needs_set && newly_acked)
10441 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10442 		/* Free back to zone */
10443 		rack_free(rack, rsm);
10444 		if (left) {
10445 			goto more;
10446 		}
10447 		/* Check for reneging */
10448 		rsm = tqhash_min(rack->r_ctl.tqh);
10449 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10450 			/*
10451 			 * The peer has moved snd_una up to
10452 			 * the edge of this send, i.e. one
10453 			 * that it had previously acked. The only
10454 			 * way that can be true if the peer threw
10455 			 * away data (space issues) that it had
10456 			 * previously sacked (else it would have
10457 			 * given us snd_una up to (rsm->r_end).
10458 			 * We need to undo the acked markings here.
10459 			 *
10460 			 * Note we have to look to make sure th_ack is
10461 			 * our rsm->r_start in case we get an old ack
10462 			 * where th_ack is behind snd_una.
10463 			 */
10464 			rack_peer_reneges(rack, rsm, th_ack);
10465 		}
10466 		return;
10467 	}
10468 	if (rsm->r_flags & RACK_ACKED) {
10469 		/*
10470 		 * It was acked on the scoreboard -- remove it from
10471 		 * total for the part being cum-acked.
10472 		 */
10473 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10474 	} else {
10475 		rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
10476 	}
10477 	/* And what about the lost flag? */
10478 	if (rsm->r_flags & RACK_WAS_LOST) {
10479 		/*
10480 		 * This can happen when we marked it as lost
10481 		 * and yet before retransmitting we get an ack
10482 		 * which can happen due to reordering. In this
10483 		 * case its only a partial ack of the send.
10484 		 */
10485 		KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
10486 			("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm,  rack, th_ack));
10487 		if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
10488 			rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
10489 		else
10490 			rack->r_ctl.rc_considered_lost = 0;
10491 	}
10492 	/*
10493 	 * Clear the dup ack count for
10494 	 * the piece that remains.
10495 	 */
10496 	rsm->r_dupack = 0;
10497 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10498 	if (rsm->r_rtr_bytes) {
10499 		/*
10500 		 * It was retransmitted adjust the
10501 		 * sack holes for what was acked.
10502 		 */
10503 		int ack_am;
10504 
10505 		ack_am = (th_ack - rsm->r_start);
10506 		if (ack_am >= rsm->r_rtr_bytes) {
10507 			rack->r_ctl.rc_holes_rxt -= ack_am;
10508 			rsm->r_rtr_bytes -= ack_am;
10509 		}
10510 	}
10511 	/*
10512 	 * Update where the piece starts and record
10513 	 * the time of send of highest cumack sent if
10514 	 * its in our GP range.
10515 	 */
10516 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10517 	/* Now we need to move our offset forward too */
10518 	if (rsm->m &&
10519 	    ((rsm->orig_m_len != rsm->m->m_len) ||
10520 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10521 		/* Fix up the orig_m_len and possibly the mbuf offset */
10522 		rack_adjust_orig_mlen(rsm);
10523 	}
10524 	rsm->soff += (th_ack - rsm->r_start);
10525 	rack_rsm_sender_update(rack, tp, rsm, 5);
10526 	/* The trim will move th_ack into r_start for us */
10527 	tqhash_trim(rack->r_ctl.tqh, th_ack);
10528 	/* Now do we need to move the mbuf fwd too? */
10529 	{
10530 		struct mbuf *m;
10531 		uint32_t soff;
10532 
10533 		m = rsm->m;
10534 		soff = rsm->soff;
10535 		if (m) {
10536 			while (soff >= m->m_len) {
10537 				soff -= m->m_len;
10538 				KASSERT((m->m_next != NULL),
10539 					(" rsm:%p  off:%u soff:%u m:%p",
10540 					 rsm, rsm->soff, soff, m));
10541 				m = m->m_next;
10542 				if (m == NULL) {
10543 					/*
10544 					 * This is a fall-back that prevents a panic. In reality
10545 					 * we should be able to walk the mbuf's and find our place.
10546 					 * At this point snd_una has not been updated with the sbcut() yet
10547 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
10548 					 * should work fine. This is undesirable since we will take cache
10549 					 * hits to access the socket buffer. And even more puzzling is that
10550 					 * it happens occasionally. It should not :(
10551 					 */
10552 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10553 						      (rsm->r_start - tp->snd_una),
10554 						      &soff);
10555 					break;
10556 				}
10557 			}
10558 			/*
10559 			 * Now save in our updated values.
10560 			 */
10561 			rsm->m = m;
10562 			rsm->soff = soff;
10563 			rsm->orig_m_len = rsm->m->m_len;
10564 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10565 		}
10566 	}
10567 	if (rack->app_limited_needs_set &&
10568 	    SEQ_GEQ(th_ack, tp->gput_seq))
10569 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10570 }
10571 
10572 static void
10573 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10574 {
10575 	struct rack_sendmap *rsm;
10576 	int sack_pass_fnd = 0;
10577 
10578 	if (rack->r_might_revert) {
10579 		/*
10580 		 * Ok we have reordering, have not sent anything, we
10581 		 * might want to revert the congestion state if nothing
10582 		 * further has SACK_PASSED on it. Lets check.
10583 		 *
10584 		 * We also get here when we have DSACKs come in for
10585 		 * all the data that we FR'd. Note that a rxt or tlp
10586 		 * timer clears this from happening.
10587 		 */
10588 
10589 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10590 			if (rsm->r_flags & RACK_SACK_PASSED) {
10591 				sack_pass_fnd = 1;
10592 				break;
10593 			}
10594 		}
10595 		if (sack_pass_fnd == 0) {
10596 			/*
10597 			 * We went into recovery
10598 			 * incorrectly due to reordering!
10599 			 */
10600 			int orig_cwnd;
10601 
10602 			rack->r_ent_rec_ns = 0;
10603 			orig_cwnd = tp->snd_cwnd;
10604 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10605 			tp->snd_recover = tp->snd_una;
10606 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10607 			if (IN_RECOVERY(tp->t_flags)) {
10608 				rack_exit_recovery(tp, rack, 3);
10609 				if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
10610 					/*
10611 					 * We were in recovery, had an RTO
10612 					 * and then re-entered recovery (more sack's arrived)
10613 					 * and we have properly recorded the old ssthresh from
10614 					 * the first recovery. We want to be able to slow-start
10615 					 * back to this level. The ssthresh from the timeout
10616 					 * and then back into recovery will end up most likely
10617 					 * to be min(cwnd=1mss, 2mss). Which makes it basically
10618 					 * so we get no slow-start after our RTO.
10619 					 */
10620 					rack->rto_from_rec = 0;
10621 					if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
10622 						tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
10623 				}
10624 			}
10625 		}
10626 		rack->r_might_revert = 0;
10627 	}
10628 }
10629 
10630 
10631 static int
10632 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10633 {
10634 
10635 	uint32_t am, l_end;
10636 	int was_tlp = 0;
10637 
10638 	if (SEQ_GT(end, start))
10639 		am = end - start;
10640 	else
10641 		am = 0;
10642 	if ((rack->rc_last_tlp_acked_set ) &&
10643 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10644 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10645 		/*
10646 		 * The DSACK is because of a TLP which we don't
10647 		 * do anything with the reordering window over since
10648 		 * it was not reordering that caused the DSACK but
10649 		 * our previous retransmit TLP.
10650 		 */
10651 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
10652 		was_tlp = 1;
10653 		goto skip_dsack_round;
10654 	}
10655 	if (rack->rc_last_sent_tlp_seq_valid) {
10656 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10657 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10658 		    (SEQ_LEQ(end, l_end))) {
10659 			/*
10660 			 * This dsack is from the last sent TLP, ignore it
10661 			 * for reordering purposes.
10662 			 */
10663 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
10664 			was_tlp = 1;
10665 			goto skip_dsack_round;
10666 		}
10667 	}
10668 	if (rack->rc_dsack_round_seen == 0) {
10669 		rack->rc_dsack_round_seen = 1;
10670 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10671 		rack->r_ctl.num_dsack++;
10672 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
10673 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10674 	}
10675 skip_dsack_round:
10676 	/*
10677 	 * We keep track of how many DSACK blocks we get
10678 	 * after a recovery incident.
10679 	 */
10680 	rack->r_ctl.dsack_byte_cnt += am;
10681 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10682 	    rack->r_ctl.retran_during_recovery &&
10683 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10684 		/*
10685 		 * False recovery most likely culprit is reordering. If
10686 		 * nothing else is missing we need to revert.
10687 		 */
10688 		rack->r_might_revert = 1;
10689 		rack_handle_might_revert(rack->rc_tp, rack);
10690 		rack->r_might_revert = 0;
10691 		rack->r_ctl.retran_during_recovery = 0;
10692 		rack->r_ctl.dsack_byte_cnt = 0;
10693 	}
10694 	return (was_tlp);
10695 }
10696 
10697 static uint32_t
10698 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10699 {
10700 	return (((tp->snd_max - snd_una) -
10701 		 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
10702 }
10703 
10704 static int32_t
10705 rack_compute_pipe(struct tcpcb *tp)
10706 {
10707 	return ((int32_t)do_rack_compute_pipe(tp,
10708 					      (struct tcp_rack *)tp->t_fb_ptr,
10709 					      tp->snd_una));
10710 }
10711 
10712 static void
10713 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10714 {
10715 	/* Deal with changed and PRR here (in recovery only) */
10716 	uint32_t pipe, snd_una;
10717 
10718 	rack->r_ctl.rc_prr_delivered += changed;
10719 
10720 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10721 		/*
10722 		 * It is all outstanding, we are application limited
10723 		 * and thus we don't need more room to send anything.
10724 		 * Note we use tp->snd_una here and not th_ack because
10725 		 * the data as yet not been cut from the sb.
10726 		 */
10727 		rack->r_ctl.rc_prr_sndcnt = 0;
10728 		return;
10729 	}
10730 	/* Compute prr_sndcnt */
10731 	if (SEQ_GT(tp->snd_una, th_ack)) {
10732 		snd_una = tp->snd_una;
10733 	} else {
10734 		snd_una = th_ack;
10735 	}
10736 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
10737 	if (pipe > tp->snd_ssthresh) {
10738 		long sndcnt;
10739 
10740 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10741 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
10742 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10743 		else {
10744 			rack->r_ctl.rc_prr_sndcnt = 0;
10745 			rack_log_to_prr(rack, 9, 0, __LINE__);
10746 			sndcnt = 0;
10747 		}
10748 		sndcnt++;
10749 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10750 			sndcnt -= rack->r_ctl.rc_prr_out;
10751 		else
10752 			sndcnt = 0;
10753 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
10754 		rack_log_to_prr(rack, 10, 0, __LINE__);
10755 	} else {
10756 		uint32_t limit;
10757 
10758 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10759 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10760 		else
10761 			limit = 0;
10762 		if (changed > limit)
10763 			limit = changed;
10764 		limit += ctf_fixed_maxseg(tp);
10765 		if (tp->snd_ssthresh > pipe) {
10766 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10767 			rack_log_to_prr(rack, 11, 0, __LINE__);
10768 		} else {
10769 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10770 			rack_log_to_prr(rack, 12, 0, __LINE__);
10771 		}
10772 	}
10773 }
10774 
10775 static void
10776 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
10777 	     int *dsack_seen, int *sacks_seen)
10778 {
10779 	uint32_t changed;
10780 	struct tcp_rack *rack;
10781 	struct rack_sendmap *rsm;
10782 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
10783 	register uint32_t th_ack;
10784 	int32_t i, j, k, num_sack_blks = 0;
10785 	uint32_t cts, acked, ack_point;
10786 	int loop_start = 0;
10787 	uint32_t tsused;
10788 	uint32_t segsiz;
10789 
10790 
10791 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10792 	if (tcp_get_flags(th) & TH_RST) {
10793 		/* We don't log resets */
10794 		return;
10795 	}
10796 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10797 	cts = tcp_get_usecs(NULL);
10798 	rsm = tqhash_min(rack->r_ctl.tqh);
10799 	changed = 0;
10800 	th_ack = th->th_ack;
10801 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
10802 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
10803 		/*
10804 		 * You only get credit for
10805 		 * MSS and greater (and you get extra
10806 		 * credit for larger cum-ack moves).
10807 		 */
10808 		int ac;
10809 
10810 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
10811 		counter_u64_add(rack_ack_total, ac);
10812 	}
10813 	if (SEQ_GT(th_ack, tp->snd_una)) {
10814 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
10815 		tp->t_acktime = ticks;
10816 	}
10817 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
10818 		changed = th_ack - rsm->r_start;
10819 	if (changed) {
10820 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
10821 				       tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time));
10822 	}
10823 	if ((to->to_flags & TOF_SACK) == 0) {
10824 		/* We are done nothing left and no sack. */
10825 		rack_handle_might_revert(tp, rack);
10826 		/*
10827 		 * For cases where we struck a dup-ack
10828 		 * with no SACK, add to the changes so
10829 		 * PRR will work right.
10830 		 */
10831 		if (dup_ack_struck && (changed == 0)) {
10832 			changed += ctf_fixed_maxseg(rack->rc_tp);
10833 		}
10834 		goto out;
10835 	}
10836 	/* Sack block processing */
10837 	if (SEQ_GT(th_ack, tp->snd_una))
10838 		ack_point = th_ack;
10839 	else
10840 		ack_point = tp->snd_una;
10841 	for (i = 0; i < to->to_nsacks; i++) {
10842 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
10843 		      &sack, sizeof(sack));
10844 		sack.start = ntohl(sack.start);
10845 		sack.end = ntohl(sack.end);
10846 		if (SEQ_GT(sack.end, sack.start) &&
10847 		    SEQ_GT(sack.start, ack_point) &&
10848 		    SEQ_LT(sack.start, tp->snd_max) &&
10849 		    SEQ_GT(sack.end, ack_point) &&
10850 		    SEQ_LEQ(sack.end, tp->snd_max)) {
10851 			sack_blocks[num_sack_blks] = sack;
10852 			num_sack_blks++;
10853 		} else if (SEQ_LEQ(sack.start, th_ack) &&
10854 			   SEQ_LEQ(sack.end, th_ack)) {
10855 			int was_tlp;
10856 
10857 			if (dsack_seen != NULL)
10858 				*dsack_seen = 1;
10859 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
10860 			/*
10861 			 * Its a D-SACK block.
10862 			 */
10863 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
10864 		}
10865 	}
10866 	if (rack->rc_dsack_round_seen) {
10867 		/* Is the dsack roound over? */
10868 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
10869 			/* Yes it is */
10870 			rack->rc_dsack_round_seen = 0;
10871 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
10872 		}
10873 	}
10874 	/*
10875 	 * Sort the SACK blocks so we can update the rack scoreboard with
10876 	 * just one pass.
10877 	 */
10878 	num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
10879 					 num_sack_blks, th->th_ack);
10880 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
10881 	if (sacks_seen != NULL)
10882 		*sacks_seen = num_sack_blks;
10883 	if (num_sack_blks == 0) {
10884 		/* Nothing to sack, but we need to update counts */
10885 		goto out_with_totals;
10886 	}
10887 	/* Its a sack of some sort */
10888 	if (num_sack_blks < 2) {
10889 		/* Only one, we don't need to sort */
10890 		goto do_sack_work;
10891 	}
10892 	/* Sort the sacks */
10893 	for (i = 0; i < num_sack_blks; i++) {
10894 		for (j = i + 1; j < num_sack_blks; j++) {
10895 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
10896 				sack = sack_blocks[i];
10897 				sack_blocks[i] = sack_blocks[j];
10898 				sack_blocks[j] = sack;
10899 			}
10900 		}
10901 	}
10902 	/*
10903 	 * Now are any of the sack block ends the same (yes some
10904 	 * implementations send these)?
10905 	 */
10906 again:
10907 	if (num_sack_blks == 0)
10908 		goto out_with_totals;
10909 	if (num_sack_blks > 1) {
10910 		for (i = 0; i < num_sack_blks; i++) {
10911 			for (j = i + 1; j < num_sack_blks; j++) {
10912 				if (sack_blocks[i].end == sack_blocks[j].end) {
10913 					/*
10914 					 * Ok these two have the same end we
10915 					 * want the smallest end and then
10916 					 * throw away the larger and start
10917 					 * again.
10918 					 */
10919 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
10920 						/*
10921 						 * The second block covers
10922 						 * more area use that
10923 						 */
10924 						sack_blocks[i].start = sack_blocks[j].start;
10925 					}
10926 					/*
10927 					 * Now collapse out the dup-sack and
10928 					 * lower the count
10929 					 */
10930 					for (k = (j + 1); k < num_sack_blks; k++) {
10931 						sack_blocks[j].start = sack_blocks[k].start;
10932 						sack_blocks[j].end = sack_blocks[k].end;
10933 						j++;
10934 					}
10935 					num_sack_blks--;
10936 					goto again;
10937 				}
10938 			}
10939 		}
10940 	}
10941 do_sack_work:
10942 	/*
10943 	 * First lets look to see if
10944 	 * we have retransmitted and
10945 	 * can use the transmit next?
10946 	 */
10947 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10948 	if (rsm &&
10949 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
10950 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
10951 		/*
10952 		 * We probably did the FR and the next
10953 		 * SACK in continues as we would expect.
10954 		 */
10955 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, segsiz);
10956 		if (acked) {
10957 			rack->r_wanted_output = 1;
10958 			changed += acked;
10959 		}
10960 		if (num_sack_blks == 1) {
10961 			/*
10962 			 * This is what we would expect from
10963 			 * a normal implementation to happen
10964 			 * after we have retransmitted the FR,
10965 			 * i.e the sack-filter pushes down
10966 			 * to 1 block and the next to be retransmitted
10967 			 * is the sequence in the sack block (has more
10968 			 * are acked). Count this as ACK'd data to boost
10969 			 * up the chances of recovering any false positives.
10970 			 */
10971 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
10972 			counter_u64_add(rack_express_sack, 1);
10973 			goto out_with_totals;
10974 		} else {
10975 			/*
10976 			 * Start the loop through the
10977 			 * rest of blocks, past the first block.
10978 			 */
10979 			loop_start = 1;
10980 		}
10981 	}
10982 	counter_u64_add(rack_sack_total, 1);
10983 	rsm = rack->r_ctl.rc_sacklast;
10984 	for (i = loop_start; i < num_sack_blks; i++) {
10985 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts,  segsiz);
10986 		if (acked) {
10987 			rack->r_wanted_output = 1;
10988 			changed += acked;
10989 		}
10990 	}
10991 out_with_totals:
10992 	if (num_sack_blks > 1) {
10993 		/*
10994 		 * You get an extra stroke if
10995 		 * you have more than one sack-blk, this
10996 		 * could be where we are skipping forward
10997 		 * and the sack-filter is still working, or
10998 		 * it could be an attacker constantly
10999 		 * moving us.
11000 		 */
11001 		counter_u64_add(rack_move_some, 1);
11002 	}
11003 out:
11004 	if (changed) {
11005 		/* Something changed cancel the rack timer */
11006 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11007 	}
11008 	tsused = tcp_get_usecs(NULL);
11009 	rsm = tcp_rack_output(tp, rack, tsused);
11010 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11011 	    rsm &&
11012 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11013 		/* Enter recovery */
11014 		entered_recovery = 1;
11015 		rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
11016 		/*
11017 		 * When we enter recovery we need to assure we send
11018 		 * one packet.
11019 		 */
11020 		if (rack->rack_no_prr == 0) {
11021 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11022 			rack_log_to_prr(rack, 8, 0, __LINE__);
11023 		}
11024 		rack->r_timer_override = 1;
11025 		rack->r_early = 0;
11026 		rack->r_ctl.rc_agg_early = 0;
11027 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11028 		   rsm &&
11029 		   (rack->r_rr_config == 3)) {
11030 		/*
11031 		 * Assure we can output and we get no
11032 		 * remembered pace time except the retransmit.
11033 		 */
11034 		rack->r_timer_override = 1;
11035 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11036 		rack->r_ctl.rc_resend = rsm;
11037 	}
11038 	if (IN_FASTRECOVERY(tp->t_flags) &&
11039 	    (rack->rack_no_prr == 0) &&
11040 	    (entered_recovery == 0)) {
11041 		rack_update_prr(tp, rack, changed, th_ack);
11042 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11043 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11044 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11045 			/*
11046 			 * If you are pacing output you don't want
11047 			 * to override.
11048 			 */
11049 			rack->r_early = 0;
11050 			rack->r_ctl.rc_agg_early = 0;
11051 			rack->r_timer_override = 1;
11052 		}
11053 	}
11054 }
11055 
11056 static void
11057 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
11058 {
11059 	struct rack_sendmap *rsm;
11060 
11061 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11062 	while (rsm) {
11063 		/*
11064 		 * We need to skip anything already set
11065 		 * to be retransmitted.
11066 		 */
11067 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
11068 		    (rsm->r_flags & RACK_MUST_RXT)) {
11069 			rsm = TAILQ_NEXT(rsm, r_tnext);
11070 			continue;
11071 		}
11072 		break;
11073 	}
11074 	if (rsm && (rsm->r_dupack < 0xff)) {
11075 		rsm->r_dupack++;
11076 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11077 			struct timeval tv;
11078 			uint32_t cts;
11079 			/*
11080 			 * Here we see if we need to retransmit. For
11081 			 * a SACK type connection if enough time has passed
11082 			 * we will get a return of the rsm. For a non-sack
11083 			 * connection we will get the rsm returned if the
11084 			 * dupack value is 3 or more.
11085 			 */
11086 			cts = tcp_get_usecs(&tv);
11087 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11088 			if (rack->r_ctl.rc_resend != NULL) {
11089 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11090 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11091 							 th_ack,  __LINE__);
11092 				}
11093 				rack->r_wanted_output = 1;
11094 				rack->r_timer_override = 1;
11095 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11096 			}
11097 		} else {
11098 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11099 		}
11100 	}
11101 }
11102 
11103 static void
11104 rack_check_bottom_drag(struct tcpcb *tp,
11105 		       struct tcp_rack *rack,
11106 		       struct socket *so)
11107 {
11108 	/*
11109 	 * So what is dragging bottom?
11110 	 *
11111 	 * Dragging bottom means you were under pacing and had a
11112 	 * delay in processing inbound acks waiting on our pacing
11113 	 * timer to expire. While you were waiting all of the acknowledgments
11114 	 * for the packets you sent have arrived. This means we are pacing
11115 	 * way underneath the bottleneck to the point where our Goodput
11116 	 * measurements stop working, since they require more than one
11117 	 * ack (usually at least 8 packets worth with multiple acks so we can
11118 	 * gauge the inter-ack times). If that occurs we have a real problem
11119 	 * since we are stuck in a hole that we can't get out of without
11120 	 * something speeding us up.
11121 	 *
11122 	 * We also check to see if we are widdling down to just one segment
11123 	 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
11124 	 * then we are adding the delayed ack interval into our measurments and
11125 	 * we need to speed up slightly.
11126 	 */
11127 	uint32_t segsiz, minseg;
11128 
11129 	segsiz = ctf_fixed_maxseg(tp);
11130 	minseg = segsiz;
11131 	if (tp->snd_max == tp->snd_una) {
11132 		/*
11133 		 * We are doing dynamic pacing and we are way
11134 		 * under. Basically everything got acked while
11135 		 * we were still waiting on the pacer to expire.
11136 		 *
11137 		 * This means we need to boost the b/w in
11138 		 * addition to any earlier boosting of
11139 		 * the multiplier.
11140 		 */
11141 		uint64_t lt_bw;
11142 
11143 		tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
11144 		lt_bw = rack_get_lt_bw(rack);
11145 		rack->rc_dragged_bottom = 1;
11146 		rack_validate_multipliers_at_or_above100(rack);
11147 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11148 		    (rack->dis_lt_bw == 0) &&
11149 		    (rack->use_lesser_lt_bw == 0) &&
11150 		    (lt_bw > 0)) {
11151 			/*
11152 			 * Lets use the long-term b/w we have
11153 			 * been getting as a base.
11154 			 */
11155 			if (rack->rc_gp_filled == 0) {
11156 				if (lt_bw > ONE_POINT_TWO_MEG) {
11157 					/*
11158 					 * If we have no measurement
11159 					 * don't let us set in more than
11160 					 * 1.2Mbps. If we are still too
11161 					 * low after pacing with this we
11162 					 * will hopefully have a max b/w
11163 					 * available to sanity check things.
11164 					 */
11165 					lt_bw = ONE_POINT_TWO_MEG;
11166 				}
11167 				rack->r_ctl.rc_rtt_diff = 0;
11168 				rack->r_ctl.gp_bw = lt_bw;
11169 				rack->rc_gp_filled = 1;
11170 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11171 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11172 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11173 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11174 				rack->r_ctl.rc_rtt_diff = 0;
11175 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11176 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11177 				rack->r_ctl.gp_bw = lt_bw;
11178 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11179 			} else
11180 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11181 			if ((rack->gp_ready == 0) &&
11182 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11183 				/* We have enough measurements now */
11184 				rack->gp_ready = 1;
11185 				if (rack->dgp_on ||
11186 				    rack->rack_hibeta)
11187 					rack_set_cc_pacing(rack);
11188 				if (rack->defer_options)
11189 					rack_apply_deferred_options(rack);
11190 			}
11191 		} else {
11192 			/*
11193 			 * zero rtt possibly?, settle for just an old increase.
11194 			 */
11195 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11196 		}
11197 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11198 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11199 					       minseg)) &&
11200 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11201 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11202 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11203 		    (segsiz * rack_req_segs))) {
11204 		/*
11205 		 * We are doing dynamic GP pacing and
11206 		 * we have everything except 1MSS or less
11207 		 * bytes left out. We are still pacing away.
11208 		 * And there is data that could be sent, This
11209 		 * means we are inserting delayed ack time in
11210 		 * our measurements because we are pacing too slow.
11211 		 */
11212 		rack_validate_multipliers_at_or_above100(rack);
11213 		rack->rc_dragged_bottom = 1;
11214 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11215 	}
11216 }
11217 
11218 #ifdef TCP_REQUEST_TRK
11219 static void
11220 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11221 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11222 {
11223 	int do_log;
11224 
11225 	do_log = tcp_bblogging_on(rack->rc_tp);
11226 	if (do_log == 0) {
11227 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11228 			return;
11229 		/* We only allow the three below with point logging on */
11230 		if ((mod != HYBRID_LOG_RULES_APP) &&
11231 		    (mod != HYBRID_LOG_RULES_SET) &&
11232 		    (mod != HYBRID_LOG_REQ_COMP))
11233 			return;
11234 
11235 	}
11236 	if (do_log) {
11237 		union tcp_log_stackspecific log;
11238 		struct timeval tv;
11239 
11240 		/* Convert our ms to a microsecond */
11241 		memset(&log, 0, sizeof(log));
11242 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11243 		log.u_bbr.flex1 = seq;
11244 		log.u_bbr.cwnd_gain = line;
11245 		if (cur != NULL) {
11246 			uint64_t off;
11247 
11248 			log.u_bbr.flex2 = cur->start_seq;
11249 			log.u_bbr.flex3 = cur->end_seq;
11250 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11251 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11252 			log.u_bbr.flex6 = cur->flags;
11253 			log.u_bbr.pkts_out = cur->hybrid_flags;
11254 			log.u_bbr.rttProp = cur->timestamp;
11255 			log.u_bbr.cur_del_rate = cur->cspr;
11256 			log.u_bbr.bw_inuse = cur->start;
11257 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11258 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11259 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11260 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11261 			log.u_bbr.inhpts = 1;
11262 #ifdef TCP_REQUEST_TRK
11263 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11264 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11265 #endif
11266 		} else {
11267 			log.u_bbr.flex2 = err;
11268 		}
11269 		/*
11270 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11271 		 */
11272 		log.u_bbr.flex7 = rack->rc_catch_up;
11273 		log.u_bbr.flex7 <<= 1;
11274 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11275 		log.u_bbr.flex7 <<= 1;
11276 		log.u_bbr.flex7 |= rack->dgp_on;
11277 		/*
11278 		 * Compose bbr_state to be a bit wise 0000ADHF
11279 		 * where A is the always_pace flag
11280 		 * where D is the dgp_on flag
11281 		 * where H is the hybrid_mode on flag
11282 		 * where F is the use_fixed_rate flag.
11283 		 */
11284 		log.u_bbr.bbr_state = rack->rc_always_pace;
11285 		log.u_bbr.bbr_state <<= 1;
11286 		log.u_bbr.bbr_state |= rack->dgp_on;
11287 		log.u_bbr.bbr_state <<= 1;
11288 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
11289 		log.u_bbr.bbr_state <<= 1;
11290 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
11291 		log.u_bbr.flex8 = mod;
11292 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11293 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11294 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11295 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11296 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11297 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11298 		tcp_log_event(rack->rc_tp, NULL,
11299 		    &rack->rc_inp->inp_socket->so_rcv,
11300 		    &rack->rc_inp->inp_socket->so_snd,
11301 		    TCP_HYBRID_PACING_LOG, 0,
11302 	            0, &log, false, NULL, __func__, __LINE__, &tv);
11303 	}
11304 }
11305 #endif
11306 
11307 #ifdef TCP_REQUEST_TRK
11308 static void
11309 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11310 {
11311 	struct tcp_sendfile_track *rc_cur, *orig_ent;
11312 	struct tcpcb *tp;
11313 	int err = 0;
11314 
11315 	orig_ent = rack->r_ctl.rc_last_sft;
11316 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11317 	if (rc_cur == NULL) {
11318 		/* If not in the beginning what about the end piece */
11319 		if (rack->rc_hybrid_mode)
11320 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11321 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11322 	} else {
11323 		err = 12345;
11324 	}
11325 	/* If we find no parameters we are in straight DGP mode */
11326 	if(rc_cur == NULL) {
11327 		/* None found for this seq, just DGP for now */
11328 		if (rack->rc_hybrid_mode) {
11329 			rack->r_ctl.client_suggested_maxseg = 0;
11330 			rack->rc_catch_up = 0;
11331 			if (rack->cspr_is_fcc == 0)
11332 				rack->r_ctl.bw_rate_cap = 0;
11333 			else
11334 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11335 		}
11336 		if (rack->rc_hybrid_mode) {
11337 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11338 		}
11339 		if (rack->r_ctl.rc_last_sft) {
11340 			rack->r_ctl.rc_last_sft = NULL;
11341 		}
11342 		return;
11343 	}
11344 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11345 		/* This entry was never setup for hybrid pacing on/off etc */
11346 		if (rack->rc_hybrid_mode) {
11347 			rack->r_ctl.client_suggested_maxseg = 0;
11348 			rack->rc_catch_up = 0;
11349 			rack->r_ctl.bw_rate_cap = 0;
11350 		}
11351 		if (rack->r_ctl.rc_last_sft) {
11352 			rack->r_ctl.rc_last_sft = NULL;
11353 		}
11354 		if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11355 			rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
11356 			rc_cur->first_send = cts;
11357 			rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
11358 			rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11359 		}
11360 		return;
11361 	}
11362 	/*
11363 	 * Ok if we have a new entry *or* have never
11364 	 * set up an entry we need to proceed. If
11365 	 * we have already set it up this entry we
11366 	 * just continue along with what we already
11367 	 * setup.
11368 	 */
11369 	tp = rack->rc_tp;
11370 	if ((rack->r_ctl.rc_last_sft != NULL) &&
11371 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
11372 		/* Its already in place */
11373 		if (rack->rc_hybrid_mode)
11374 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11375 		return;
11376 	}
11377 	if (rack->rc_hybrid_mode == 0) {
11378 		rack->r_ctl.rc_last_sft = rc_cur;
11379 		if (orig_ent) {
11380 			orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11381 			orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11382 			orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11383 		}
11384 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11385 		return;
11386 	}
11387 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11388 		/* Compensate for all the header overhead's */
11389 		if (rack->cspr_is_fcc == 0)
11390 			rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
11391 		else
11392 			rack->r_ctl.fillcw_cap =  rack_compensate_for_linerate(rack, rc_cur->cspr);
11393 	} else {
11394 		if (rack->rc_hybrid_mode) {
11395 			if (rack->cspr_is_fcc == 0)
11396 				rack->r_ctl.bw_rate_cap = 0;
11397 			else
11398 				rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11399 		}
11400 	}
11401 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11402 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11403 	else
11404 		rack->r_ctl.client_suggested_maxseg = 0;
11405 	if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
11406 		/*
11407 		 * It is the same timestamp as the previous one
11408 		 * add the hybrid flag that will indicate we use
11409 		 * sendtime not arrival time for catch-up mode.
11410 		 */
11411 		rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
11412 	}
11413 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11414 	    (rc_cur->cspr > 0)) {
11415 		uint64_t len;
11416 
11417 		rack->rc_catch_up = 1;
11418 		/*
11419 		 * Calculate the deadline time, first set the
11420 		 * time to when the request arrived.
11421 		 */
11422 		if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
11423 			/*
11424 			 * For cases where its a duplicate tm (we received more
11425 			 * than one request for a tm) we want to use now, the point
11426 			 * where we are just sending the first bit of the request.
11427 			 */
11428 			rc_cur->deadline = cts;
11429 		} else {
11430 			/*
11431 			 * Here we have a different tm from the last request
11432 			 * so we want to use arrival time as our base.
11433 			 */
11434 			rc_cur->deadline = rc_cur->localtime;
11435 		}
11436 		/*
11437 		 * Next calculate the length and compensate for
11438 		 * TLS if need be.
11439 		 */
11440 		len = rc_cur->end - rc_cur->start;
11441 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11442 			/*
11443 			 * This session is doing TLS. Take a swag guess
11444 			 * at the overhead.
11445 			 */
11446 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11447 		}
11448 		/*
11449 		 * Now considering the size, and the cspr, what is the time that
11450 		 * would be required at the cspr rate. Here we use the raw
11451 		 * cspr value since the client only looks at the raw data. We
11452 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
11453 		 * That will get made up for in the CU pacing rate set.
11454 		 */
11455 		len *= HPTS_USEC_IN_SEC;
11456 		len /= rc_cur->cspr;
11457 		rc_cur->deadline += len;
11458 	} else {
11459 		rack->rc_catch_up = 0;
11460 		rc_cur->deadline = 0;
11461 	}
11462 	if (rack->r_ctl.client_suggested_maxseg != 0) {
11463 		/*
11464 		 * We need to reset the max pace segs if we have a
11465 		 * client_suggested_maxseg.
11466 		 */
11467 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
11468 	}
11469 	if (orig_ent) {
11470 		orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11471 		orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11472 		orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11473 	}
11474 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11475 	/* Remember it for next time and for CU mode */
11476 	rack->r_ctl.rc_last_sft = rc_cur;
11477 	rack->r_ctl.last_tm_mark = rc_cur->timestamp;
11478 }
11479 #endif
11480 
11481 static void
11482 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11483 {
11484 #ifdef TCP_REQUEST_TRK
11485 	struct tcp_sendfile_track *ent;
11486 
11487 	ent = rack->r_ctl.rc_last_sft;
11488 	if ((ent == NULL) ||
11489 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11490 	    (SEQ_GEQ(seq, ent->end_seq))) {
11491 		/* Time to update the track. */
11492 		rack_set_dgp_hybrid_mode(rack, seq, len, cts);
11493 		ent = rack->r_ctl.rc_last_sft;
11494 	}
11495 	/* Out of all */
11496 	if (ent == NULL) {
11497 		return;
11498 	}
11499 	if (SEQ_LT(ent->end_seq, (seq + len))) {
11500 		/*
11501 		 * This is the case where our end_seq guess
11502 		 * was wrong. This is usually due to TLS having
11503 		 * more bytes then our guess. It could also be the
11504 		 * case that the client sent in two requests closely
11505 		 * and the SB is full of both so we are sending part
11506 		 * of each (end|beg). In such a case lets move this
11507 		 * guys end to match the end of this send. That
11508 		 * way it will complete when all of it is acked.
11509 		 */
11510 		ent->end_seq = (seq + len);
11511 		if (rack->rc_hybrid_mode)
11512 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11513 	}
11514 	/* Now validate we have set the send time of this one */
11515 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11516 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11517 		ent->first_send = cts;
11518 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11519 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11520 	}
11521 #endif
11522 }
11523 
11524 static void
11525 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11526 {
11527 	/*
11528 	 * The fast output path is enabled and we
11529 	 * have moved the cumack forward. Lets see if
11530 	 * we can expand forward the fast path length by
11531 	 * that amount. What we would ideally like to
11532 	 * do is increase the number of bytes in the
11533 	 * fast path block (left_to_send) by the
11534 	 * acked amount. However we have to gate that
11535 	 * by two factors:
11536 	 * 1) The amount outstanding and the rwnd of the peer
11537 	 *    (i.e. we don't want to exceed the rwnd of the peer).
11538 	 *    <and>
11539 	 * 2) The amount of data left in the socket buffer (i.e.
11540 	 *    we can't send beyond what is in the buffer).
11541 	 *
11542 	 * Note that this does not take into account any increase
11543 	 * in the cwnd. We will only extend the fast path by
11544 	 * what was acked.
11545 	 */
11546 	uint32_t new_total, gating_val;
11547 
11548 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11549 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11550 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11551 	if (new_total <= gating_val) {
11552 		/* We can increase left_to_send by the acked amount */
11553 		counter_u64_add(rack_extended_rfo, 1);
11554 		rack->r_ctl.fsb.left_to_send = new_total;
11555 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11556 			("rack:%p left_to_send:%u sbavail:%u out:%u",
11557 			 rack, rack->r_ctl.fsb.left_to_send,
11558 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
11559 			 (tp->snd_max - tp->snd_una)));
11560 
11561 	}
11562 }
11563 
11564 static void
11565 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11566 {
11567 	/*
11568 	 * Here any sendmap entry that points to the
11569 	 * beginning mbuf must be adjusted to the correct
11570 	 * offset. This must be called with:
11571 	 * 1) The socket buffer locked
11572 	 * 2) snd_una adjusted to its new position.
11573 	 *
11574 	 * Note that (2) implies rack_ack_received has also
11575 	 * been called and all the sbcut's have been done.
11576 	 *
11577 	 * We grab the first mbuf in the socket buffer and
11578 	 * then go through the front of the sendmap, recalculating
11579 	 * the stored offset for any sendmap entry that has
11580 	 * that mbuf. We must use the sb functions to do this
11581 	 * since its possible an add was done has well as
11582 	 * the subtraction we may have just completed. This should
11583 	 * not be a penalty though, since we just referenced the sb
11584 	 * to go in and trim off the mbufs that we freed (of course
11585 	 * there will be a penalty for the sendmap references though).
11586 	 *
11587 	 * Note also with INVARIANT on, we validate with a KASSERT
11588 	 * that the first sendmap entry has a soff of 0.
11589 	 *
11590 	 */
11591 	struct mbuf *m;
11592 	struct rack_sendmap *rsm;
11593 	tcp_seq snd_una;
11594 #ifdef INVARIANTS
11595 	int first_processed = 0;
11596 #endif
11597 
11598 	snd_una = rack->rc_tp->snd_una;
11599 	SOCKBUF_LOCK_ASSERT(sb);
11600 	m = sb->sb_mb;
11601 	rsm = tqhash_min(rack->r_ctl.tqh);
11602 	if ((rsm == NULL) || (m == NULL)) {
11603 		/* Nothing outstanding */
11604 		return;
11605 	}
11606 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
11607 	KASSERT((rsm->m == m),
11608 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11609 		 rack, sb, rsm));
11610 	while (rsm->m && (rsm->m == m)) {
11611 		/* one to adjust */
11612 #ifdef INVARIANTS
11613 		struct mbuf *tm;
11614 		uint32_t soff;
11615 
11616 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11617 		if ((rsm->orig_m_len != m->m_len) ||
11618 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
11619 			rack_adjust_orig_mlen(rsm);
11620 		}
11621 		if (first_processed == 0) {
11622 			KASSERT((rsm->soff == 0),
11623 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
11624 				 rack, rsm));
11625 			first_processed = 1;
11626 		}
11627 		if ((rsm->soff != soff) || (rsm->m != tm)) {
11628 			/*
11629 			 * This is not a fatal error, we anticipate it
11630 			 * might happen (the else code), so we count it here
11631 			 * so that under invariant we can see that it really
11632 			 * does happen.
11633 			 */
11634 			counter_u64_add(rack_adjust_map_bw, 1);
11635 		}
11636 		rsm->m = tm;
11637 		rsm->soff = soff;
11638 		if (tm) {
11639 			rsm->orig_m_len = rsm->m->m_len;
11640 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11641 		} else {
11642 			rsm->orig_m_len = 0;
11643 			rsm->orig_t_space = 0;
11644 		}
11645 #else
11646 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11647 		if (rsm->m) {
11648 			rsm->orig_m_len = rsm->m->m_len;
11649 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11650 		} else {
11651 			rsm->orig_m_len = 0;
11652 			rsm->orig_t_space = 0;
11653 		}
11654 #endif
11655 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11656 		if (rsm == NULL)
11657 			break;
11658 	}
11659 }
11660 
11661 #ifdef TCP_REQUEST_TRK
11662 static inline void
11663 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11664 {
11665 	struct tcp_sendfile_track *ent;
11666 	int i;
11667 
11668 	if ((rack->rc_hybrid_mode == 0) &&
11669 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11670 		/*
11671 		 * Just do normal completions hybrid pacing is not on
11672 		 * and CLDL is off as well.
11673 		 */
11674 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
11675 		return;
11676 	}
11677 	/*
11678 	 * Originally I was just going to find the th_ack associated
11679 	 * with an entry. But then I realized a large strech ack could
11680 	 * in theory ack two or more requests at once. So instead we
11681 	 * need to find all entries that are completed by th_ack not
11682 	 * just a single entry and do our logging.
11683 	 */
11684 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11685 	while (ent != NULL) {
11686 		/*
11687 		 * We may be doing hybrid pacing or CLDL and need more details possibly
11688 		 * so we do it manually instead of calling
11689 		 * tcp_req_check_for_comp()
11690 		 */
11691 		uint64_t laa, tim, data, cbw, ftim;
11692 
11693 		/* Ok this ack frees it */
11694 		rack_log_hybrid(rack, th_ack,
11695 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11696 		rack_log_hybrid_sends(rack, ent, __LINE__);
11697 		/* calculate the time based on the ack arrival */
11698 		data = ent->end - ent->start;
11699 		laa = tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time);
11700 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
11701 			if (ent->first_send > ent->localtime)
11702 				ftim = ent->first_send;
11703 			else
11704 				ftim = ent->localtime;
11705 		} else {
11706 			/* TSNH */
11707 			ftim = ent->localtime;
11708 		}
11709 		if (laa > ent->localtime)
11710 			tim = laa - ftim;
11711 		else
11712 			tim = 0;
11713 		cbw = data * HPTS_USEC_IN_SEC;
11714 		if (tim > 0)
11715 			cbw /= tim;
11716 		else
11717 			cbw = 0;
11718 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
11719 		/*
11720 		 * Check to see if we are freeing what we are pointing to send wise
11721 		 * if so be sure to NULL the pointer so we know we are no longer
11722 		 * set to anything.
11723 		 */
11724 		if (ent == rack->r_ctl.rc_last_sft) {
11725 			rack->r_ctl.rc_last_sft = NULL;
11726 			if (rack->rc_hybrid_mode) {
11727 				rack->rc_catch_up = 0;
11728 				if (rack->cspr_is_fcc == 0)
11729 					rack->r_ctl.bw_rate_cap = 0;
11730 				else
11731 					rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11732 				rack->r_ctl.client_suggested_maxseg = 0;
11733 			}
11734 		}
11735 		/* Generate the log that the tcp_netflix call would have */
11736 		tcp_req_log_req_info(rack->rc_tp, ent,
11737 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
11738 		/* Free it and see if there is another one */
11739 		tcp_req_free_a_slot(rack->rc_tp, ent);
11740 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11741 	}
11742 }
11743 #endif
11744 
11745 
11746 /*
11747  * Return value of 1, we do not need to call rack_process_data().
11748  * return value of 0, rack_process_data can be called.
11749  * For ret_val if its 0 the TCP is locked, if its non-zero
11750  * its unlocked and probably unsafe to touch the TCB.
11751  */
11752 static int
11753 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11754     struct tcpcb *tp, struct tcpopt *to,
11755     uint32_t tiwin, int32_t tlen,
11756     int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
11757 {
11758 	int32_t ourfinisacked = 0;
11759 	int32_t nsegs, acked_amount;
11760 	int32_t acked;
11761 	struct mbuf *mfree;
11762 	struct tcp_rack *rack;
11763 	int32_t under_pacing = 0;
11764 	int32_t post_recovery = 0;
11765 	uint32_t p_cwnd;
11766 
11767 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11768 
11769 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11770 	if (SEQ_GEQ(tp->snd_una, tp->iss + (65535 << tp->snd_scale))) {
11771 		/* Checking SEG.ACK against ISS is definitely redundant. */
11772 		tp->t_flags2 |= TF2_NO_ISS_CHECK;
11773 	}
11774 	if (!V_tcp_insecure_ack) {
11775 		tcp_seq seq_min;
11776 		bool ghost_ack_check;
11777 
11778 		if (tp->t_flags2 & TF2_NO_ISS_CHECK) {
11779 			/* Check for too old ACKs (RFC 5961, Section 5.2). */
11780 			seq_min = tp->snd_una - tp->max_sndwnd;
11781 			ghost_ack_check = false;
11782 		} else {
11783 			if (SEQ_GT(tp->iss + 1, tp->snd_una - tp->max_sndwnd)) {
11784 				/* Checking for ghost ACKs is stricter. */
11785 				seq_min = tp->iss + 1;
11786 				ghost_ack_check = true;
11787 			} else {
11788 				/*
11789 				 * Checking for too old ACKs (RFC 5961,
11790 				 * Section 5.2) is stricter.
11791 				 */
11792 				seq_min = tp->snd_una - tp->max_sndwnd;
11793 				ghost_ack_check = false;
11794 			}
11795 		}
11796 		if (SEQ_LT(th->th_ack, seq_min)) {
11797 			if (ghost_ack_check)
11798 				TCPSTAT_INC(tcps_rcvghostack);
11799 			else
11800 				TCPSTAT_INC(tcps_rcvacktooold);
11801 			/* Send challenge ACK. */
11802 			ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11803 			rack->r_wanted_output = 1;
11804 			return (1);
11805 		}
11806 	}
11807 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
11808 		ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11809 		rack->r_wanted_output = 1;
11810 		return (1);
11811 	}
11812 	if (rack->gp_ready &&
11813 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11814 		under_pacing = 1;
11815 	}
11816 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
11817 		int in_rec, dup_ack_struck = 0;
11818 		int dsack_seen = 0, sacks_seen = 0;
11819 
11820 		in_rec = IN_FASTRECOVERY(tp->t_flags);
11821 		if (rack->rc_in_persist) {
11822 			tp->t_rxtshift = 0;
11823 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11824 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11825 		}
11826 
11827 		if ((th->th_ack == tp->snd_una) &&
11828 		    (tiwin == tp->snd_wnd) &&
11829 		    (orig_tlen == 0) &&
11830 		    ((to->to_flags & TOF_SACK) == 0)) {
11831 			rack_strike_dupack(rack, th->th_ack);
11832 			dup_ack_struck = 1;
11833 		}
11834 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
11835 			     dup_ack_struck, &dsack_seen, &sacks_seen);
11836 
11837 	}
11838 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11839 		/*
11840 		 * Old ack, behind (or duplicate to) the last one rcv'd
11841 		 * Note: We mark reordering is occuring if its
11842 		 * less than and we have not closed our window.
11843 		 */
11844 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
11845 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
11846 			if (rack->r_ctl.rc_reorder_ts == 0)
11847 				rack->r_ctl.rc_reorder_ts = 1;
11848 		}
11849 		return (0);
11850 	}
11851 	/*
11852 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
11853 	 * something we sent.
11854 	 */
11855 	if (tp->t_flags & TF_NEEDSYN) {
11856 		/*
11857 		 * T/TCP: Connection was half-synchronized, and our SYN has
11858 		 * been ACK'd (so connection is now fully synchronized).  Go
11859 		 * to non-starred state, increment snd_una for ACK of SYN,
11860 		 * and check if we can do window scaling.
11861 		 */
11862 		tp->t_flags &= ~TF_NEEDSYN;
11863 		tp->snd_una++;
11864 		/* Do window scaling? */
11865 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11866 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11867 			tp->rcv_scale = tp->request_r_scale;
11868 			/* Send window already scaled. */
11869 		}
11870 	}
11871 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11872 
11873 	acked = BYTES_THIS_ACK(tp, th);
11874 	if (acked) {
11875 		/*
11876 		 * Any time we move the cum-ack forward clear
11877 		 * keep-alive tied probe-not-answered. The
11878 		 * persists clears its own on entry.
11879 		 */
11880 		rack->probe_not_answered = 0;
11881 	}
11882 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11883 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11884 	/*
11885 	 * If we just performed our first retransmit, and the ACK arrives
11886 	 * within our recovery window, then it was a mistake to do the
11887 	 * retransmit in the first place.  Recover our original cwnd and
11888 	 * ssthresh, and proceed to transmit where we left off.
11889 	 */
11890 	if ((tp->t_flags & TF_PREVVALID) &&
11891 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11892 		tp->t_flags &= ~TF_PREVVALID;
11893 		if (tp->t_rxtshift == 1 &&
11894 		    (int)(ticks - tp->t_badrxtwin) < 0)
11895 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11896 	}
11897 	if (acked) {
11898 		/* assure we are not backed off */
11899 		tp->t_rxtshift = 0;
11900 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11901 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11902 		rack->rc_tlp_in_progress = 0;
11903 		rack->r_ctl.rc_tlp_cnt_out = 0;
11904 		/*
11905 		 * If it is the RXT timer we want to
11906 		 * stop it, so we can restart a TLP.
11907 		 */
11908 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11909 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11910 #ifdef TCP_REQUEST_TRK
11911 		rack_req_check_for_comp(rack, th->th_ack);
11912 #endif
11913 	}
11914 	/*
11915 	 * If we have a timestamp reply, update smoothed round trip time. If
11916 	 * no timestamp is present but transmit timer is running and timed
11917 	 * sequence number was acked, update smoothed round trip time. Since
11918 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
11919 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
11920 	 * timer.
11921 	 *
11922 	 * Some boxes send broken timestamp replies during the SYN+ACK
11923 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11924 	 * and blow up the retransmit timer.
11925 	 */
11926 	/*
11927 	 * If all outstanding data is acked, stop retransmit timer and
11928 	 * remember to restart (more output or persist). If there is more
11929 	 * data to be acked, restart retransmit timer, using current
11930 	 * (possibly backed-off) value.
11931 	 */
11932 	if (acked == 0) {
11933 		if (ofia)
11934 			*ofia = ourfinisacked;
11935 		return (0);
11936 	}
11937 	if (IN_RECOVERY(tp->t_flags)) {
11938 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
11939 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
11940 			tcp_rack_partialack(tp);
11941 		} else {
11942 			rack_post_recovery(tp, th->th_ack);
11943 			post_recovery = 1;
11944 			/*
11945 			 * Grab the segsiz, multiply by 2 and add the snd_cwnd
11946 			 * that is the max the CC should add if we are exiting
11947 			 * recovery and doing a late add.
11948 			 */
11949 			p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
11950 			p_cwnd <<= 1;
11951 			p_cwnd += tp->snd_cwnd;
11952 		}
11953 	} else if ((rack->rto_from_rec == 1) &&
11954 		   SEQ_GEQ(th->th_ack, tp->snd_recover)) {
11955 		/*
11956 		 * We were in recovery, hit a rxt timeout
11957 		 * and never re-entered recovery. The timeout(s)
11958 		 * made up all the lost data. In such a case
11959 		 * we need to clear the rto_from_rec flag.
11960 		 */
11961 		rack->rto_from_rec = 0;
11962 	}
11963 	/*
11964 	 * Let the congestion control algorithm update congestion control
11965 	 * related information. This typically means increasing the
11966 	 * congestion window.
11967 	 */
11968 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
11969 	if (post_recovery &&
11970 	    (tp->snd_cwnd > p_cwnd)) {
11971 		/* Must be non-newreno (cubic) getting too ahead of itself */
11972 		tp->snd_cwnd = p_cwnd;
11973 	}
11974 	SOCK_SENDBUF_LOCK(so);
11975 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
11976 	tp->snd_wnd -= acked_amount;
11977 	mfree = sbcut_locked(&so->so_snd, acked_amount);
11978 	if ((sbused(&so->so_snd) == 0) &&
11979 	    (acked > acked_amount) &&
11980 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
11981 	    (tp->t_flags & TF_SENTFIN)) {
11982 		/*
11983 		 * We must be sure our fin
11984 		 * was sent and acked (we can be
11985 		 * in FIN_WAIT_1 without having
11986 		 * sent the fin).
11987 		 */
11988 		ourfinisacked = 1;
11989 	}
11990 	tp->snd_una = th->th_ack;
11991 	/* wakeups? */
11992 	if (acked_amount && sbavail(&so->so_snd))
11993 		rack_adjust_sendmap_head(rack, &so->so_snd);
11994 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11995 	/* NB: sowwakeup_locked() does an implicit unlock. */
11996 	sowwakeup_locked(so);
11997 	m_freem(mfree);
11998 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
11999 		tp->snd_recover = tp->snd_una;
12000 
12001 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12002 		tp->snd_nxt = tp->snd_max;
12003 	}
12004 	if (under_pacing &&
12005 	    (rack->use_fixed_rate == 0) &&
12006 	    (rack->in_probe_rtt == 0) &&
12007 	    rack->rc_gp_dyn_mul &&
12008 	    rack->rc_always_pace) {
12009 		/* Check if we are dragging bottom */
12010 		rack_check_bottom_drag(tp, rack, so);
12011 	}
12012 	if (tp->snd_una == tp->snd_max) {
12013 		/* Nothing left outstanding */
12014 		tp->t_flags &= ~TF_PREVVALID;
12015 		if (rack->r_ctl.rc_went_idle_time == 0)
12016 			rack->r_ctl.rc_went_idle_time = 1;
12017 		rack->r_ctl.retran_during_recovery = 0;
12018 		rack->r_ctl.dsack_byte_cnt = 0;
12019 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12020 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12021 			tp->t_acktime = 0;
12022 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12023 		rack->rc_suspicious = 0;
12024 		/* Set need output so persist might get set */
12025 		rack->r_wanted_output = 1;
12026 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12027 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12028 		    (sbavail(&so->so_snd) == 0) &&
12029 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12030 			/*
12031 			 * The socket was gone and the
12032 			 * peer sent data (now or in the past), time to
12033 			 * reset him.
12034 			 */
12035 			*ret_val = 1;
12036 			/* tcp_close will kill the inp pre-log the Reset */
12037 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12038 			tp = tcp_close(tp);
12039 			ctf_do_dropwithreset(m, tp, th, tlen);
12040 			return (1);
12041 		}
12042 	}
12043 	if (ofia)
12044 		*ofia = ourfinisacked;
12045 	return (0);
12046 }
12047 
12048 
12049 static void
12050 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12051 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12052 {
12053 	if (tcp_bblogging_on(rack->rc_tp)) {
12054 		union tcp_log_stackspecific log;
12055 		struct timeval tv;
12056 
12057 		memset(&log, 0, sizeof(log));
12058 		log.u_bbr.flex1 = cnt;
12059 		log.u_bbr.flex2 = split;
12060 		log.u_bbr.flex3 = out;
12061 		log.u_bbr.flex4 = line;
12062 		log.u_bbr.flex5 = rack->r_must_retran;
12063 		log.u_bbr.flex6 = flags;
12064 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12065 		log.u_bbr.flex8 = dir;	/*
12066 					 * 1 is collapsed, 0 is uncollapsed,
12067 					 * 2 is log of a rsm being marked, 3 is a split.
12068 					 */
12069 		if (rsm == NULL)
12070 			log.u_bbr.rttProp = 0;
12071 		else
12072 			log.u_bbr.rttProp = (uintptr_t)rsm;
12073 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12074 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12075 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12076 		    &rack->rc_inp->inp_socket->so_rcv,
12077 		    &rack->rc_inp->inp_socket->so_snd,
12078 		    TCP_RACK_LOG_COLLAPSE, 0,
12079 		    0, &log, false, &tv);
12080 	}
12081 }
12082 
12083 static void
12084 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12085 {
12086 	/*
12087 	 * Here all we do is mark the collapsed point and set the flag.
12088 	 * This may happen again and again, but there is no
12089 	 * sense splitting our map until we know where the
12090 	 * peer finally lands in the collapse.
12091 	 */
12092 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12093 	if ((rack->rc_has_collapsed == 0) ||
12094 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12095 		counter_u64_add(rack_collapsed_win_seen, 1);
12096 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12097 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12098 	rack->rc_has_collapsed = 1;
12099 	rack->r_collapse_point_valid = 1;
12100 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12101 }
12102 
12103 static void
12104 rack_un_collapse_window(struct tcp_rack *rack, int line)
12105 {
12106 	struct rack_sendmap *nrsm, *rsm;
12107 	int cnt = 0, split = 0;
12108 	int insret __diagused;
12109 
12110 
12111 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12112 	rack->rc_has_collapsed = 0;
12113 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12114 	if (rsm == NULL) {
12115 		/* Nothing to do maybe the peer ack'ed it all */
12116 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12117 		return;
12118 	}
12119 	/* Now do we need to split this one? */
12120 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12121 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12122 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12123 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12124 		if (nrsm == NULL) {
12125 			/* We can't get a rsm, mark all? */
12126 			nrsm = rsm;
12127 			goto no_split;
12128 		}
12129 		/* Clone it */
12130 		split = 1;
12131 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12132 #ifndef INVARIANTS
12133 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12134 #else
12135 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12136 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12137 			      nrsm, insret, rack, rsm);
12138 		}
12139 #endif
12140 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12141 				 rack->r_ctl.last_collapse_point, __LINE__);
12142 		if (rsm->r_in_tmap) {
12143 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12144 			nrsm->r_in_tmap = 1;
12145 		}
12146 		/*
12147 		 * Set in the new RSM as the
12148 		 * collapsed starting point
12149 		 */
12150 		rsm = nrsm;
12151 	}
12152 
12153 no_split:
12154 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12155 		cnt++;
12156 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12157 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12158 		cnt++;
12159 	}
12160 	if (cnt) {
12161 		counter_u64_add(rack_collapsed_win, 1);
12162 	}
12163 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12164 }
12165 
12166 static void
12167 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12168 			int32_t tlen, int32_t tfo_syn)
12169 {
12170 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12171 		rack_timer_cancel(tp, rack,
12172 				  rack->r_ctl.rc_rcvtime, __LINE__);
12173 		tp->t_flags |= TF_DELACK;
12174 	} else {
12175 		rack->r_wanted_output = 1;
12176 		tp->t_flags |= TF_ACKNOW;
12177 	}
12178 }
12179 
12180 static void
12181 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12182 {
12183 	/*
12184 	 * If fast output is in progress, lets validate that
12185 	 * the new window did not shrink on us and make it
12186 	 * so fast output should end.
12187 	 */
12188 	if (rack->r_fast_output) {
12189 		uint32_t out;
12190 
12191 		/*
12192 		 * Calculate what we will send if left as is
12193 		 * and compare that to our send window.
12194 		 */
12195 		out = ctf_outstanding(tp);
12196 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12197 			/* ok we have an issue */
12198 			if (out >= tp->snd_wnd) {
12199 				/* Turn off fast output the window is met or collapsed */
12200 				rack->r_fast_output = 0;
12201 			} else {
12202 				/* we have some room left */
12203 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12204 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12205 					/* If not at least 1 full segment never mind */
12206 					rack->r_fast_output = 0;
12207 				}
12208 			}
12209 		}
12210 	}
12211 }
12212 
12213 /*
12214  * Return value of 1, the TCB is unlocked and most
12215  * likely gone, return value of 0, the TCP is still
12216  * locked.
12217  */
12218 static int
12219 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12220     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12221     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12222 {
12223 	/*
12224 	 * Update window information. Don't look at window if no ACK: TAC's
12225 	 * send garbage on first SYN.
12226 	 */
12227 	int32_t nsegs;
12228 	int32_t tfo_syn;
12229 	struct tcp_rack *rack;
12230 
12231 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12232 
12233 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12234 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12235 	if ((thflags & TH_ACK) &&
12236 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12237 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12238 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12239 		/* keep track of pure window updates */
12240 		if (tlen == 0 &&
12241 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12242 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12243 		tp->snd_wnd = tiwin;
12244 		rack_validate_fo_sendwin_up(tp, rack);
12245 		tp->snd_wl1 = th->th_seq;
12246 		tp->snd_wl2 = th->th_ack;
12247 		if (tp->snd_wnd > tp->max_sndwnd)
12248 			tp->max_sndwnd = tp->snd_wnd;
12249 		rack->r_wanted_output = 1;
12250 	} else if (thflags & TH_ACK) {
12251 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12252 			tp->snd_wnd = tiwin;
12253 			rack_validate_fo_sendwin_up(tp, rack);
12254 			tp->snd_wl1 = th->th_seq;
12255 			tp->snd_wl2 = th->th_ack;
12256 		}
12257 	}
12258 	if (tp->snd_wnd < ctf_outstanding(tp))
12259 		/* The peer collapsed the window */
12260 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12261 	else if (rack->rc_has_collapsed)
12262 		rack_un_collapse_window(rack, __LINE__);
12263 	if ((rack->r_collapse_point_valid) &&
12264 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12265 		rack->r_collapse_point_valid = 0;
12266 	/* Was persist timer active and now we have window space? */
12267 	if ((rack->rc_in_persist != 0) &&
12268 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12269 				rack->r_ctl.rc_pace_min_segs))) {
12270 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12271 		tp->snd_nxt = tp->snd_max;
12272 		/* Make sure we output to start the timer */
12273 		rack->r_wanted_output = 1;
12274 	}
12275 	/* Do we enter persists? */
12276 	if ((rack->rc_in_persist == 0) &&
12277 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12278 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12279 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12280 	    sbavail(&tptosocket(tp)->so_snd) &&
12281 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12282 		/*
12283 		 * Here the rwnd is less than
12284 		 * the pacing size, we are established,
12285 		 * nothing is outstanding, and there is
12286 		 * data to send. Enter persists.
12287 		 */
12288 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12289 	}
12290 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12291 		m_freem(m);
12292 		return (0);
12293 	}
12294 	/*
12295 	 * don't process the URG bit, ignore them drag
12296 	 * along the up.
12297 	 */
12298 	tp->rcv_up = tp->rcv_nxt;
12299 
12300 	/*
12301 	 * Process the segment text, merging it into the TCP sequencing
12302 	 * queue, and arranging for acknowledgment of receipt if necessary.
12303 	 * This process logically involves adjusting tp->rcv_wnd as data is
12304 	 * presented to the user (this happens in tcp_usrreq.c, case
12305 	 * PRU_RCVD).  If a FIN has already been received on this connection
12306 	 * then we just ignore the text.
12307 	 */
12308 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12309 	    (tp->t_flags & TF_FASTOPEN));
12310 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12311 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12312 		tcp_seq save_start = th->th_seq;
12313 		tcp_seq save_rnxt  = tp->rcv_nxt;
12314 		int     save_tlen  = tlen;
12315 
12316 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12317 		/*
12318 		 * Insert segment which includes th into TCP reassembly
12319 		 * queue with control block tp.  Set thflags to whether
12320 		 * reassembly now includes a segment with FIN.  This handles
12321 		 * the common case inline (segment is the next to be
12322 		 * received on an established connection, and the queue is
12323 		 * empty), avoiding linkage into and removal from the queue
12324 		 * and repetition of various conversions. Set DELACK for
12325 		 * segments received in order, but ack immediately when
12326 		 * segments are out of order (so fast retransmit can work).
12327 		 */
12328 		if (th->th_seq == tp->rcv_nxt &&
12329 		    SEGQ_EMPTY(tp) &&
12330 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
12331 		    tfo_syn)) {
12332 #ifdef NETFLIX_SB_LIMITS
12333 			u_int mcnt, appended;
12334 
12335 			if (so->so_rcv.sb_shlim) {
12336 				mcnt = m_memcnt(m);
12337 				appended = 0;
12338 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12339 				    CFO_NOSLEEP, NULL) == false) {
12340 					counter_u64_add(tcp_sb_shlim_fails, 1);
12341 					m_freem(m);
12342 					return (0);
12343 				}
12344 			}
12345 #endif
12346 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12347 			tp->rcv_nxt += tlen;
12348 			if (tlen &&
12349 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12350 			    (tp->t_fbyte_in == 0)) {
12351 				tp->t_fbyte_in = ticks;
12352 				if (tp->t_fbyte_in == 0)
12353 					tp->t_fbyte_in = 1;
12354 				if (tp->t_fbyte_out && tp->t_fbyte_in)
12355 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12356 			}
12357 			thflags = tcp_get_flags(th) & TH_FIN;
12358 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12359 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12360 			SOCK_RECVBUF_LOCK(so);
12361 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12362 				m_freem(m);
12363 			} else {
12364 				int32_t newsize;
12365 
12366 				if (tlen > 0) {
12367 					newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12368 					if (newsize)
12369 						if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12370 							so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12371 				}
12372 #ifdef NETFLIX_SB_LIMITS
12373 				appended =
12374 #endif
12375 					sbappendstream_locked(&so->so_rcv, m, 0);
12376 			}
12377 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12378 			/* NB: sorwakeup_locked() does an implicit unlock. */
12379 			sorwakeup_locked(so);
12380 #ifdef NETFLIX_SB_LIMITS
12381 			if (so->so_rcv.sb_shlim && appended != mcnt)
12382 				counter_fo_release(so->so_rcv.sb_shlim,
12383 				    mcnt - appended);
12384 #endif
12385 		} else {
12386 			/*
12387 			 * XXX: Due to the header drop above "th" is
12388 			 * theoretically invalid by now.  Fortunately
12389 			 * m_adj() doesn't actually frees any mbufs when
12390 			 * trimming from the head.
12391 			 */
12392 			tcp_seq temp = save_start;
12393 
12394 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
12395 			tp->t_flags |= TF_ACKNOW;
12396 			if (tp->t_flags & TF_WAKESOR) {
12397 				tp->t_flags &= ~TF_WAKESOR;
12398 				/* NB: sorwakeup_locked() does an implicit unlock. */
12399 				sorwakeup_locked(so);
12400 			}
12401 		}
12402 		if ((tp->t_flags & TF_SACK_PERMIT) &&
12403 		    (save_tlen > 0) &&
12404 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
12405 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12406 				/*
12407 				 * DSACK actually handled in the fastpath
12408 				 * above.
12409 				 */
12410 				tcp_update_sack_list(tp, save_start,
12411 				    save_start + save_tlen);
12412 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12413 				if ((tp->rcv_numsacks >= 1) &&
12414 				    (tp->sackblks[0].end == save_start)) {
12415 					/*
12416 					 * Partial overlap, recorded at todrop
12417 					 * above.
12418 					 */
12419 					tcp_update_sack_list(tp,
12420 					    tp->sackblks[0].start,
12421 					    tp->sackblks[0].end);
12422 				} else {
12423 					tcp_update_dsack_list(tp, save_start,
12424 					    save_start + save_tlen);
12425 				}
12426 			} else if (tlen >= save_tlen) {
12427 				/* Update of sackblks. */
12428 				tcp_update_dsack_list(tp, save_start,
12429 				    save_start + save_tlen);
12430 			} else if (tlen > 0) {
12431 				tcp_update_dsack_list(tp, save_start,
12432 				    save_start + tlen);
12433 			}
12434 		}
12435 	} else {
12436 		m_freem(m);
12437 		thflags &= ~TH_FIN;
12438 	}
12439 
12440 	/*
12441 	 * If FIN is received ACK the FIN and let the user know that the
12442 	 * connection is closing.
12443 	 */
12444 	if (thflags & TH_FIN) {
12445 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12446 			/* The socket upcall is handled by socantrcvmore. */
12447 			socantrcvmore(so);
12448 			/*
12449 			 * If connection is half-synchronized (ie NEEDSYN
12450 			 * flag on) then delay ACK, so it may be piggybacked
12451 			 * when SYN is sent. Otherwise, since we received a
12452 			 * FIN then no more input can be expected, send ACK
12453 			 * now.
12454 			 */
12455 			if (tp->t_flags & TF_NEEDSYN) {
12456 				rack_timer_cancel(tp, rack,
12457 				    rack->r_ctl.rc_rcvtime, __LINE__);
12458 				tp->t_flags |= TF_DELACK;
12459 			} else {
12460 				tp->t_flags |= TF_ACKNOW;
12461 			}
12462 			tp->rcv_nxt++;
12463 		}
12464 		switch (tp->t_state) {
12465 			/*
12466 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12467 			 * CLOSE_WAIT state.
12468 			 */
12469 		case TCPS_SYN_RECEIVED:
12470 			tp->t_starttime = ticks;
12471 			/* FALLTHROUGH */
12472 		case TCPS_ESTABLISHED:
12473 			rack_timer_cancel(tp, rack,
12474 			    rack->r_ctl.rc_rcvtime, __LINE__);
12475 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
12476 			break;
12477 
12478 			/*
12479 			 * If still in FIN_WAIT_1 STATE FIN has not been
12480 			 * acked so enter the CLOSING state.
12481 			 */
12482 		case TCPS_FIN_WAIT_1:
12483 			rack_timer_cancel(tp, rack,
12484 			    rack->r_ctl.rc_rcvtime, __LINE__);
12485 			tcp_state_change(tp, TCPS_CLOSING);
12486 			break;
12487 
12488 			/*
12489 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12490 			 * starting the time-wait timer, turning off the
12491 			 * other standard timers.
12492 			 */
12493 		case TCPS_FIN_WAIT_2:
12494 			rack_timer_cancel(tp, rack,
12495 			    rack->r_ctl.rc_rcvtime, __LINE__);
12496 			tcp_twstart(tp);
12497 			return (1);
12498 		}
12499 	}
12500 	/*
12501 	 * Return any desired output.
12502 	 */
12503 	if ((tp->t_flags & TF_ACKNOW) ||
12504 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12505 		rack->r_wanted_output = 1;
12506 	}
12507 	return (0);
12508 }
12509 
12510 /*
12511  * Here nothing is really faster, its just that we
12512  * have broken out the fast-data path also just like
12513  * the fast-ack.
12514  */
12515 static int
12516 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12517     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12518     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12519 {
12520 	int32_t nsegs;
12521 	int32_t newsize = 0;	/* automatic sockbuf scaling */
12522 	struct tcp_rack *rack;
12523 #ifdef NETFLIX_SB_LIMITS
12524 	u_int mcnt, appended;
12525 #endif
12526 
12527 	/*
12528 	 * If last ACK falls within this segment's sequence numbers, record
12529 	 * the timestamp. NOTE that the test is modified according to the
12530 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12531 	 */
12532 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12533 		return (0);
12534 	}
12535 	if (tiwin && tiwin != tp->snd_wnd) {
12536 		return (0);
12537 	}
12538 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12539 		return (0);
12540 	}
12541 	if (__predict_false((to->to_flags & TOF_TS) &&
12542 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12543 		return (0);
12544 	}
12545 	if (__predict_false((th->th_ack != tp->snd_una))) {
12546 		return (0);
12547 	}
12548 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12549 		return (0);
12550 	}
12551 	if ((to->to_flags & TOF_TS) != 0 &&
12552 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12553 		tp->ts_recent_age = tcp_ts_getticks();
12554 		tp->ts_recent = to->to_tsval;
12555 	}
12556 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12557 	/*
12558 	 * This is a pure, in-sequence data packet with nothing on the
12559 	 * reassembly queue and we have enough buffer space to take it.
12560 	 */
12561 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12562 
12563 #ifdef NETFLIX_SB_LIMITS
12564 	if (so->so_rcv.sb_shlim) {
12565 		mcnt = m_memcnt(m);
12566 		appended = 0;
12567 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12568 		    CFO_NOSLEEP, NULL) == false) {
12569 			counter_u64_add(tcp_sb_shlim_fails, 1);
12570 			m_freem(m);
12571 			return (1);
12572 		}
12573 	}
12574 #endif
12575 	/* Clean receiver SACK report if present */
12576 	if (tp->rcv_numsacks)
12577 		tcp_clean_sackreport(tp);
12578 	KMOD_TCPSTAT_INC(tcps_preddat);
12579 	tp->rcv_nxt += tlen;
12580 	if (tlen &&
12581 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12582 	    (tp->t_fbyte_in == 0)) {
12583 		tp->t_fbyte_in = ticks;
12584 		if (tp->t_fbyte_in == 0)
12585 			tp->t_fbyte_in = 1;
12586 		if (tp->t_fbyte_out && tp->t_fbyte_in)
12587 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12588 	}
12589 	/*
12590 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12591 	 */
12592 	tp->snd_wl1 = th->th_seq;
12593 	/*
12594 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12595 	 */
12596 	tp->rcv_up = tp->rcv_nxt;
12597 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12598 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12599 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12600 
12601 	/* Add data to socket buffer. */
12602 	SOCK_RECVBUF_LOCK(so);
12603 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12604 		m_freem(m);
12605 	} else {
12606 		/*
12607 		 * Set new socket buffer size. Give up when limit is
12608 		 * reached.
12609 		 */
12610 		if (newsize)
12611 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12612 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12613 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12614 #ifdef NETFLIX_SB_LIMITS
12615 		appended =
12616 #endif
12617 			sbappendstream_locked(&so->so_rcv, m, 0);
12618 		ctf_calc_rwin(so, tp);
12619 	}
12620 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12621 	/* NB: sorwakeup_locked() does an implicit unlock. */
12622 	sorwakeup_locked(so);
12623 #ifdef NETFLIX_SB_LIMITS
12624 	if (so->so_rcv.sb_shlim && mcnt != appended)
12625 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12626 #endif
12627 	rack_handle_delayed_ack(tp, rack, tlen, 0);
12628 	if (tp->snd_una == tp->snd_max)
12629 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12630 	return (1);
12631 }
12632 
12633 /*
12634  * This subfunction is used to try to highly optimize the
12635  * fast path. We again allow window updates that are
12636  * in sequence to remain in the fast-path. We also add
12637  * in the __predict's to attempt to help the compiler.
12638  * Note that if we return a 0, then we can *not* process
12639  * it and the caller should push the packet into the
12640  * slow-path.
12641  */
12642 static int
12643 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12644     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12645     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12646 {
12647 	int32_t acked;
12648 	int32_t nsegs;
12649 	int32_t under_pacing = 0;
12650 	struct tcp_rack *rack;
12651 
12652 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12653 		/* Old ack, behind (or duplicate to) the last one rcv'd */
12654 		return (0);
12655 	}
12656 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12657 		/* Above what we have sent? */
12658 		return (0);
12659 	}
12660 	if (__predict_false(tiwin == 0)) {
12661 		/* zero window */
12662 		return (0);
12663 	}
12664 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12665 		/* We need a SYN or a FIN, unlikely.. */
12666 		return (0);
12667 	}
12668 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12669 		/* Timestamp is behind .. old ack with seq wrap? */
12670 		return (0);
12671 	}
12672 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12673 		/* Still recovering */
12674 		return (0);
12675 	}
12676 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12677 	if (rack->r_ctl.rc_sacked) {
12678 		/* We have sack holes on our scoreboard */
12679 		return (0);
12680 	}
12681 	/* Ok if we reach here, we can process a fast-ack */
12682 	if (rack->gp_ready &&
12683 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12684 		under_pacing = 1;
12685 	}
12686 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12687 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12688 	/* Did the window get updated? */
12689 	if (tiwin != tp->snd_wnd) {
12690 		tp->snd_wnd = tiwin;
12691 		rack_validate_fo_sendwin_up(tp, rack);
12692 		tp->snd_wl1 = th->th_seq;
12693 		if (tp->snd_wnd > tp->max_sndwnd)
12694 			tp->max_sndwnd = tp->snd_wnd;
12695 	}
12696 	/* Do we exit persists? */
12697 	if ((rack->rc_in_persist != 0) &&
12698 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12699 			       rack->r_ctl.rc_pace_min_segs))) {
12700 		rack_exit_persist(tp, rack, cts);
12701 	}
12702 	/* Do we enter persists? */
12703 	if ((rack->rc_in_persist == 0) &&
12704 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12705 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12706 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12707 	    sbavail(&tptosocket(tp)->so_snd) &&
12708 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12709 		/*
12710 		 * Here the rwnd is less than
12711 		 * the pacing size, we are established,
12712 		 * nothing is outstanding, and there is
12713 		 * data to send. Enter persists.
12714 		 */
12715 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12716 	}
12717 	/*
12718 	 * If last ACK falls within this segment's sequence numbers, record
12719 	 * the timestamp. NOTE that the test is modified according to the
12720 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12721 	 */
12722 	if ((to->to_flags & TOF_TS) != 0 &&
12723 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12724 		tp->ts_recent_age = tcp_ts_getticks();
12725 		tp->ts_recent = to->to_tsval;
12726 	}
12727 	/*
12728 	 * This is a pure ack for outstanding data.
12729 	 */
12730 	KMOD_TCPSTAT_INC(tcps_predack);
12731 
12732 	/*
12733 	 * "bad retransmit" recovery.
12734 	 */
12735 	if ((tp->t_flags & TF_PREVVALID) &&
12736 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12737 		tp->t_flags &= ~TF_PREVVALID;
12738 		if (tp->t_rxtshift == 1 &&
12739 		    (int)(ticks - tp->t_badrxtwin) < 0)
12740 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12741 	}
12742 	/*
12743 	 * Recalculate the transmit timer / rtt.
12744 	 *
12745 	 * Some boxes send broken timestamp replies during the SYN+ACK
12746 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12747 	 * and blow up the retransmit timer.
12748 	 */
12749 	acked = BYTES_THIS_ACK(tp, th);
12750 
12751 #ifdef TCP_HHOOK
12752 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12753 	hhook_run_tcp_est_in(tp, th, to);
12754 #endif
12755 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12756 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12757 	if (acked) {
12758 		struct mbuf *mfree;
12759 
12760 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12761 		SOCK_SENDBUF_LOCK(so);
12762 		mfree = sbcut_locked(&so->so_snd, acked);
12763 		tp->snd_una = th->th_ack;
12764 		/* Note we want to hold the sb lock through the sendmap adjust */
12765 		rack_adjust_sendmap_head(rack, &so->so_snd);
12766 		/* Wake up the socket if we have room to write more */
12767 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12768 		sowwakeup_locked(so);
12769 		m_freem(mfree);
12770 		tp->t_rxtshift = 0;
12771 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12772 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12773 		rack->rc_tlp_in_progress = 0;
12774 		rack->r_ctl.rc_tlp_cnt_out = 0;
12775 		/*
12776 		 * If it is the RXT timer we want to
12777 		 * stop it, so we can restart a TLP.
12778 		 */
12779 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12780 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12781 
12782 #ifdef TCP_REQUEST_TRK
12783 		rack_req_check_for_comp(rack, th->th_ack);
12784 #endif
12785 	}
12786 	/*
12787 	 * Let the congestion control algorithm update congestion control
12788 	 * related information. This typically means increasing the
12789 	 * congestion window.
12790 	 */
12791 	if (tp->snd_wnd < ctf_outstanding(tp)) {
12792 		/* The peer collapsed the window */
12793 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12794 	} else if (rack->rc_has_collapsed)
12795 		rack_un_collapse_window(rack, __LINE__);
12796 	if ((rack->r_collapse_point_valid) &&
12797 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
12798 		rack->r_collapse_point_valid = 0;
12799 	/*
12800 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
12801 	 */
12802 	tp->snd_wl2 = th->th_ack;
12803 	tp->t_dupacks = 0;
12804 	m_freem(m);
12805 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
12806 
12807 	/*
12808 	 * If all outstanding data are acked, stop retransmit timer,
12809 	 * otherwise restart timer using current (possibly backed-off)
12810 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
12811 	 * If data are ready to send, let tcp_output decide between more
12812 	 * output or persist.
12813 	 */
12814 	if (under_pacing &&
12815 	    (rack->use_fixed_rate == 0) &&
12816 	    (rack->in_probe_rtt == 0) &&
12817 	    rack->rc_gp_dyn_mul &&
12818 	    rack->rc_always_pace) {
12819 		/* Check if we are dragging bottom */
12820 		rack_check_bottom_drag(tp, rack, so);
12821 	}
12822 	if (tp->snd_una == tp->snd_max) {
12823 		tp->t_flags &= ~TF_PREVVALID;
12824 		rack->r_ctl.retran_during_recovery = 0;
12825 		rack->rc_suspicious = 0;
12826 		rack->r_ctl.dsack_byte_cnt = 0;
12827 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12828 		if (rack->r_ctl.rc_went_idle_time == 0)
12829 			rack->r_ctl.rc_went_idle_time = 1;
12830 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12831 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12832 			tp->t_acktime = 0;
12833 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12834 	}
12835 	if (acked && rack->r_fast_output)
12836 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
12837 	if (sbavail(&so->so_snd)) {
12838 		rack->r_wanted_output = 1;
12839 	}
12840 	return (1);
12841 }
12842 
12843 /*
12844  * Return value of 1, the TCB is unlocked and most
12845  * likely gone, return value of 0, the TCP is still
12846  * locked.
12847  */
12848 static int
12849 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
12850     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12851     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12852 {
12853 	int32_t ret_val = 0;
12854 	int32_t orig_tlen = tlen;
12855 	int32_t todrop;
12856 	int32_t ourfinisacked = 0;
12857 	struct tcp_rack *rack;
12858 
12859 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12860 
12861 	ctf_calc_rwin(so, tp);
12862 	/*
12863 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
12864 	 * SYN, drop the input. if seg contains a RST, then drop the
12865 	 * connection. if seg does not contain SYN, then drop it. Otherwise
12866 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
12867 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
12868 	 * contains an ECE and ECN support is enabled, the stream is ECN
12869 	 * capable. if SYN has been acked change to ESTABLISHED else
12870 	 * SYN_RCVD state arrange for segment to be acked (eventually)
12871 	 * continue processing rest of data/controls.
12872 	 */
12873 	if ((thflags & TH_ACK) &&
12874 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
12875 	    SEQ_GT(th->th_ack, tp->snd_max))) {
12876 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
12877 		ctf_do_dropwithreset(m, tp, th, tlen);
12878 		return (1);
12879 	}
12880 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
12881 		TCP_PROBE5(connect__refused, NULL, tp,
12882 		    mtod(m, const char *), tp, th);
12883 		tp = tcp_drop(tp, ECONNREFUSED);
12884 		ctf_do_drop(m, tp);
12885 		return (1);
12886 	}
12887 	if (thflags & TH_RST) {
12888 		ctf_do_drop(m, tp);
12889 		return (1);
12890 	}
12891 	if (!(thflags & TH_SYN)) {
12892 		ctf_do_drop(m, tp);
12893 		return (1);
12894 	}
12895 	tp->irs = th->th_seq;
12896 	tcp_rcvseqinit(tp);
12897 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12898 	if (thflags & TH_ACK) {
12899 		int tfo_partial = 0;
12900 
12901 		KMOD_TCPSTAT_INC(tcps_connects);
12902 		soisconnected(so);
12903 #ifdef MAC
12904 		mac_socketpeer_set_from_mbuf(m, so);
12905 #endif
12906 		/* Do window scaling on this connection? */
12907 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12908 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12909 			tp->rcv_scale = tp->request_r_scale;
12910 		}
12911 		tp->rcv_adv += min(tp->rcv_wnd,
12912 		    TCP_MAXWIN << tp->rcv_scale);
12913 		/*
12914 		 * If not all the data that was sent in the TFO SYN
12915 		 * has been acked, resend the remainder right away.
12916 		 */
12917 		if ((tp->t_flags & TF_FASTOPEN) &&
12918 		    (tp->snd_una != tp->snd_max)) {
12919 			/* Was it a partial ack? */
12920 			if (SEQ_LT(th->th_ack, tp->snd_max))
12921 				tfo_partial = 1;
12922 		}
12923 		/*
12924 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
12925 		 * will be turned on later.
12926 		 */
12927 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
12928 			rack_timer_cancel(tp, rack,
12929 					  rack->r_ctl.rc_rcvtime, __LINE__);
12930 			tp->t_flags |= TF_DELACK;
12931 		} else {
12932 			rack->r_wanted_output = 1;
12933 			tp->t_flags |= TF_ACKNOW;
12934 		}
12935 
12936 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
12937 
12938 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
12939 			/*
12940 			 * We advance snd_una for the
12941 			 * fast open case. If th_ack is
12942 			 * acknowledging data beyond
12943 			 * snd_una we can't just call
12944 			 * ack-processing since the
12945 			 * data stream in our send-map
12946 			 * will start at snd_una + 1 (one
12947 			 * beyond the SYN). If its just
12948 			 * equal we don't need to do that
12949 			 * and there is no send_map.
12950 			 */
12951 			tp->snd_una++;
12952 			if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
12953 				/*
12954 				 * We sent a SYN with data, and thus have a
12955 				 * sendmap entry with a SYN set. Lets find it
12956 				 * and take off the send bit and the byte and
12957 				 * set it up to be what we send (send it next).
12958 				 */
12959 				struct rack_sendmap *rsm;
12960 
12961 				rsm = tqhash_min(rack->r_ctl.tqh);
12962 				if (rsm) {
12963 					if (rsm->r_flags & RACK_HAS_SYN) {
12964 						rsm->r_flags &= ~RACK_HAS_SYN;
12965 						rsm->r_start++;
12966 					}
12967 					rack->r_ctl.rc_resend = rsm;
12968 				}
12969 			}
12970 		}
12971 		/*
12972 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
12973 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
12974 		 */
12975 		tp->t_starttime = ticks;
12976 		if (tp->t_flags & TF_NEEDFIN) {
12977 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
12978 			tp->t_flags &= ~TF_NEEDFIN;
12979 			thflags &= ~TH_SYN;
12980 		} else {
12981 			tcp_state_change(tp, TCPS_ESTABLISHED);
12982 			TCP_PROBE5(connect__established, NULL, tp,
12983 			    mtod(m, const char *), tp, th);
12984 			rack_cc_conn_init(tp);
12985 		}
12986 	} else {
12987 		/*
12988 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
12989 		 * open.  If segment contains CC option and there is a
12990 		 * cached CC, apply TAO test. If it succeeds, connection is *
12991 		 * half-synchronized. Otherwise, do 3-way handshake:
12992 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
12993 		 * there was no CC option, clear cached CC value.
12994 		 */
12995 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
12996 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
12997 	}
12998 	/*
12999 	 * Advance th->th_seq to correspond to first data byte. If data,
13000 	 * trim to stay within window, dropping FIN if necessary.
13001 	 */
13002 	th->th_seq++;
13003 	if (tlen > tp->rcv_wnd) {
13004 		todrop = tlen - tp->rcv_wnd;
13005 		m_adj(m, -todrop);
13006 		tlen = tp->rcv_wnd;
13007 		thflags &= ~TH_FIN;
13008 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13009 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13010 	}
13011 	tp->snd_wl1 = th->th_seq - 1;
13012 	tp->rcv_up = th->th_seq;
13013 	/*
13014 	 * Client side of transaction: already sent SYN and data. If the
13015 	 * remote host used T/TCP to validate the SYN, our data will be
13016 	 * ACK'd; if so, enter normal data segment processing in the middle
13017 	 * of step 5, ack processing. Otherwise, goto step 6.
13018 	 */
13019 	if (thflags & TH_ACK) {
13020 		/* For syn-sent we need to possibly update the rtt */
13021 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13022 			uint32_t t, mcts;
13023 
13024 			mcts = tcp_ts_getticks();
13025 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13026 			if (!tp->t_rttlow || tp->t_rttlow > t)
13027 				tp->t_rttlow = t;
13028 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13029 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13030 			tcp_rack_xmit_timer_commit(rack, tp);
13031 		}
13032 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
13033 			return (ret_val);
13034 		/* We may have changed to FIN_WAIT_1 above */
13035 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13036 			/*
13037 			 * In FIN_WAIT_1 STATE in addition to the processing
13038 			 * for the ESTABLISHED state if our FIN is now
13039 			 * acknowledged then enter FIN_WAIT_2.
13040 			 */
13041 			if (ourfinisacked) {
13042 				/*
13043 				 * If we can't receive any more data, then
13044 				 * closing user can proceed. Starting the
13045 				 * timer is contrary to the specification,
13046 				 * but if we don't get a FIN we'll hang
13047 				 * forever.
13048 				 *
13049 				 * XXXjl: we should release the tp also, and
13050 				 * use a compressed state.
13051 				 */
13052 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13053 					soisdisconnected(so);
13054 					tcp_timer_activate(tp, TT_2MSL,
13055 					    (tcp_fast_finwait2_recycle ?
13056 					    tcp_finwait2_timeout :
13057 					    TP_MAXIDLE(tp)));
13058 				}
13059 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13060 			}
13061 		}
13062 	}
13063 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13064 	   tiwin, thflags, nxt_pkt));
13065 }
13066 
13067 /*
13068  * Return value of 1, the TCB is unlocked and most
13069  * likely gone, return value of 0, the TCP is still
13070  * locked.
13071  */
13072 static int
13073 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13074     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13075     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13076 {
13077 	struct tcp_rack *rack;
13078 	int32_t orig_tlen = tlen;
13079 	int32_t ret_val = 0;
13080 	int32_t ourfinisacked = 0;
13081 
13082 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13083 	ctf_calc_rwin(so, tp);
13084 	if ((thflags & TH_RST) ||
13085 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13086 		return (ctf_process_rst(m, th, so, tp));
13087 	if ((thflags & TH_ACK) &&
13088 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13089 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13090 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13091 		ctf_do_dropwithreset(m, tp, th, tlen);
13092 		return (1);
13093 	}
13094 	if (tp->t_flags & TF_FASTOPEN) {
13095 		/*
13096 		 * When a TFO connection is in SYN_RECEIVED, the
13097 		 * only valid packets are the initial SYN, a
13098 		 * retransmit/copy of the initial SYN (possibly with
13099 		 * a subset of the original data), a valid ACK, a
13100 		 * FIN, or a RST.
13101 		 */
13102 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13103 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13104 			ctf_do_dropwithreset(m, tp, th, tlen);
13105 			return (1);
13106 		} else if (thflags & TH_SYN) {
13107 			/* non-initial SYN is ignored */
13108 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13109 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13110 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13111 				ctf_do_drop(m, NULL);
13112 				return (0);
13113 			}
13114 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13115 			ctf_do_drop(m, NULL);
13116 			return (0);
13117 		}
13118 	}
13119 
13120 	/*
13121 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13122 	 * it's less than ts_recent, drop it.
13123 	 */
13124 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13125 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13126 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13127 			return (ret_val);
13128 	}
13129 	/*
13130 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13131 	 * this connection before trimming the data to fit the receive
13132 	 * window.  Check the sequence number versus IRS since we know the
13133 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13134 	 * "LAND" DoS attack.
13135 	 */
13136 	if (SEQ_LT(th->th_seq, tp->irs)) {
13137 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13138 		ctf_do_dropwithreset(m, tp, th, tlen);
13139 		return (1);
13140 	}
13141 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13142 		return (ret_val);
13143 	}
13144 	/*
13145 	 * If last ACK falls within this segment's sequence numbers, record
13146 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13147 	 * from the latest proposal of the tcplw@cray.com list (Braden
13148 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13149 	 * with our earlier PAWS tests, so this check should be solely
13150 	 * predicated on the sequence space of this segment. 3) That we
13151 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13152 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13153 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13154 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13155 	 * p.869. In such cases, we can still calculate the RTT correctly
13156 	 * when RCV.NXT == Last.ACK.Sent.
13157 	 */
13158 	if ((to->to_flags & TOF_TS) != 0 &&
13159 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13160 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13161 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13162 		tp->ts_recent_age = tcp_ts_getticks();
13163 		tp->ts_recent = to->to_tsval;
13164 	}
13165 	tp->snd_wnd = tiwin;
13166 	rack_validate_fo_sendwin_up(tp, rack);
13167 	/*
13168 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13169 	 * is on (half-synchronized state), then queue data for later
13170 	 * processing; else drop segment and return.
13171 	 */
13172 	if ((thflags & TH_ACK) == 0) {
13173 		if (tp->t_flags & TF_FASTOPEN) {
13174 			rack_cc_conn_init(tp);
13175 		}
13176 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13177 		    tiwin, thflags, nxt_pkt));
13178 	}
13179 	KMOD_TCPSTAT_INC(tcps_connects);
13180 	if (tp->t_flags & TF_SONOTCONN) {
13181 		tp->t_flags &= ~TF_SONOTCONN;
13182 		soisconnected(so);
13183 	}
13184 	/* Do window scaling? */
13185 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13186 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13187 		tp->rcv_scale = tp->request_r_scale;
13188 	}
13189 	/*
13190 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13191 	 * FIN-WAIT-1
13192 	 */
13193 	tp->t_starttime = ticks;
13194 	if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
13195 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13196 		tp->t_tfo_pending = NULL;
13197 	}
13198 	if (tp->t_flags & TF_NEEDFIN) {
13199 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13200 		tp->t_flags &= ~TF_NEEDFIN;
13201 	} else {
13202 		tcp_state_change(tp, TCPS_ESTABLISHED);
13203 		TCP_PROBE5(accept__established, NULL, tp,
13204 		    mtod(m, const char *), tp, th);
13205 		/*
13206 		 * TFO connections call cc_conn_init() during SYN
13207 		 * processing.  Calling it again here for such connections
13208 		 * is not harmless as it would undo the snd_cwnd reduction
13209 		 * that occurs when a TFO SYN|ACK is retransmitted.
13210 		 */
13211 		if (!(tp->t_flags & TF_FASTOPEN))
13212 			rack_cc_conn_init(tp);
13213 	}
13214 	/*
13215 	 * Account for the ACK of our SYN prior to
13216 	 * regular ACK processing below, except for
13217 	 * simultaneous SYN, which is handled later.
13218 	 */
13219 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13220 		tp->snd_una++;
13221 	/*
13222 	 * If segment contains data or ACK, will call tcp_reass() later; if
13223 	 * not, do so now to pass queued data to user.
13224 	 */
13225 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13226 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13227 		    (struct mbuf *)0);
13228 		if (tp->t_flags & TF_WAKESOR) {
13229 			tp->t_flags &= ~TF_WAKESOR;
13230 			/* NB: sorwakeup_locked() does an implicit unlock. */
13231 			sorwakeup_locked(so);
13232 		}
13233 	}
13234 	tp->snd_wl1 = th->th_seq - 1;
13235 	/* For syn-recv we need to possibly update the rtt */
13236 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13237 		uint32_t t, mcts;
13238 
13239 		mcts = tcp_ts_getticks();
13240 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13241 		if (!tp->t_rttlow || tp->t_rttlow > t)
13242 			tp->t_rttlow = t;
13243 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13244 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13245 		tcp_rack_xmit_timer_commit(rack, tp);
13246 	}
13247 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13248 		return (ret_val);
13249 	}
13250 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13251 		/* We could have went to FIN_WAIT_1 (or EST) above */
13252 		/*
13253 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13254 		 * ESTABLISHED state if our FIN is now acknowledged then
13255 		 * enter FIN_WAIT_2.
13256 		 */
13257 		if (ourfinisacked) {
13258 			/*
13259 			 * If we can't receive any more data, then closing
13260 			 * user can proceed. Starting the timer is contrary
13261 			 * to the specification, but if we don't get a FIN
13262 			 * we'll hang forever.
13263 			 *
13264 			 * XXXjl: we should release the tp also, and use a
13265 			 * compressed state.
13266 			 */
13267 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13268 				soisdisconnected(so);
13269 				tcp_timer_activate(tp, TT_2MSL,
13270 				    (tcp_fast_finwait2_recycle ?
13271 				    tcp_finwait2_timeout :
13272 				    TP_MAXIDLE(tp)));
13273 			}
13274 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13275 		}
13276 	}
13277 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13278 	    tiwin, thflags, nxt_pkt));
13279 }
13280 
13281 /*
13282  * Return value of 1, the TCB is unlocked and most
13283  * likely gone, return value of 0, the TCP is still
13284  * locked.
13285  */
13286 static int
13287 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13288     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13289     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13290 {
13291 	int32_t ret_val = 0;
13292 	int32_t orig_tlen = tlen;
13293 	struct tcp_rack *rack;
13294 
13295 	/*
13296 	 * Header prediction: check for the two common cases of a
13297 	 * uni-directional data xfer.  If the packet has no control flags,
13298 	 * is in-sequence, the window didn't change and we're not
13299 	 * retransmitting, it's a candidate.  If the length is zero and the
13300 	 * ack moved forward, we're the sender side of the xfer.  Just free
13301 	 * the data acked & wake any higher level process that was blocked
13302 	 * waiting for space.  If the length is non-zero and the ack didn't
13303 	 * move, we're the receiver side.  If we're getting packets in-order
13304 	 * (the reassembly queue is empty), add the data toc The socket
13305 	 * buffer and note that we need a delayed ack. Make sure that the
13306 	 * hidden state-flags are also off. Since we check for
13307 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13308 	 */
13309 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13310 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13311 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13312 	    __predict_true(SEGQ_EMPTY(tp)) &&
13313 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13314 		if (tlen == 0) {
13315 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13316 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13317 				return (0);
13318 			}
13319 		} else {
13320 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13321 			    tiwin, nxt_pkt, iptos)) {
13322 				return (0);
13323 			}
13324 		}
13325 	}
13326 	ctf_calc_rwin(so, tp);
13327 
13328 	if ((thflags & TH_RST) ||
13329 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13330 		return (ctf_process_rst(m, th, so, tp));
13331 
13332 	/*
13333 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13334 	 * synchronized state.
13335 	 */
13336 	if (thflags & TH_SYN) {
13337 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13338 		return (ret_val);
13339 	}
13340 	/*
13341 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13342 	 * it's less than ts_recent, drop it.
13343 	 */
13344 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13345 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13346 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13347 			return (ret_val);
13348 	}
13349 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13350 		return (ret_val);
13351 	}
13352 	/*
13353 	 * If last ACK falls within this segment's sequence numbers, record
13354 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13355 	 * from the latest proposal of the tcplw@cray.com list (Braden
13356 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13357 	 * with our earlier PAWS tests, so this check should be solely
13358 	 * predicated on the sequence space of this segment. 3) That we
13359 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13360 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13361 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13362 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13363 	 * p.869. In such cases, we can still calculate the RTT correctly
13364 	 * when RCV.NXT == Last.ACK.Sent.
13365 	 */
13366 	if ((to->to_flags & TOF_TS) != 0 &&
13367 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13368 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13369 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13370 		tp->ts_recent_age = tcp_ts_getticks();
13371 		tp->ts_recent = to->to_tsval;
13372 	}
13373 	/*
13374 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13375 	 * is on (half-synchronized state), then queue data for later
13376 	 * processing; else drop segment and return.
13377 	 */
13378 	if ((thflags & TH_ACK) == 0) {
13379 		if (tp->t_flags & TF_NEEDSYN) {
13380 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13381 			    tiwin, thflags, nxt_pkt));
13382 
13383 		} else if (tp->t_flags & TF_ACKNOW) {
13384 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13385 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13386 			return (ret_val);
13387 		} else {
13388 			ctf_do_drop(m, NULL);
13389 			return (0);
13390 		}
13391 	}
13392 	/*
13393 	 * Ack processing.
13394 	 */
13395 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13396 		return (ret_val);
13397 	}
13398 	if (sbavail(&so->so_snd)) {
13399 		if (ctf_progress_timeout_check(tp, true)) {
13400 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13401 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
13402 			return (1);
13403 		}
13404 	}
13405 	/* State changes only happen in rack_process_data() */
13406 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13407 	    tiwin, thflags, nxt_pkt));
13408 }
13409 
13410 /*
13411  * Return value of 1, the TCB is unlocked and most
13412  * likely gone, return value of 0, the TCP is still
13413  * locked.
13414  */
13415 static int
13416 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13417     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13418     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13419 {
13420 	int32_t ret_val = 0;
13421 	int32_t orig_tlen = tlen;
13422 
13423 	ctf_calc_rwin(so, tp);
13424 	if ((thflags & TH_RST) ||
13425 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13426 		return (ctf_process_rst(m, th, so, tp));
13427 	/*
13428 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13429 	 * synchronized state.
13430 	 */
13431 	if (thflags & TH_SYN) {
13432 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13433 		return (ret_val);
13434 	}
13435 	/*
13436 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13437 	 * it's less than ts_recent, drop it.
13438 	 */
13439 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13440 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13441 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13442 			return (ret_val);
13443 	}
13444 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13445 		return (ret_val);
13446 	}
13447 	/*
13448 	 * If last ACK falls within this segment's sequence numbers, record
13449 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13450 	 * from the latest proposal of the tcplw@cray.com list (Braden
13451 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13452 	 * with our earlier PAWS tests, so this check should be solely
13453 	 * predicated on the sequence space of this segment. 3) That we
13454 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13455 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13456 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13457 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13458 	 * p.869. In such cases, we can still calculate the RTT correctly
13459 	 * when RCV.NXT == Last.ACK.Sent.
13460 	 */
13461 	if ((to->to_flags & TOF_TS) != 0 &&
13462 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13463 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13464 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13465 		tp->ts_recent_age = tcp_ts_getticks();
13466 		tp->ts_recent = to->to_tsval;
13467 	}
13468 	/*
13469 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13470 	 * is on (half-synchronized state), then queue data for later
13471 	 * processing; else drop segment and return.
13472 	 */
13473 	if ((thflags & TH_ACK) == 0) {
13474 		if (tp->t_flags & TF_NEEDSYN) {
13475 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13476 			    tiwin, thflags, nxt_pkt));
13477 
13478 		} else if (tp->t_flags & TF_ACKNOW) {
13479 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13480 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13481 			return (ret_val);
13482 		} else {
13483 			ctf_do_drop(m, NULL);
13484 			return (0);
13485 		}
13486 	}
13487 	/*
13488 	 * Ack processing.
13489 	 */
13490 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13491 		return (ret_val);
13492 	}
13493 	if (sbavail(&so->so_snd)) {
13494 		if (ctf_progress_timeout_check(tp, true)) {
13495 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13496 						tp, tick, PROGRESS_DROP, __LINE__);
13497 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
13498 			return (1);
13499 		}
13500 	}
13501 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13502 	    tiwin, thflags, nxt_pkt));
13503 }
13504 
13505 static int
13506 rack_check_data_after_close(struct mbuf *m,
13507     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13508 {
13509 	struct tcp_rack *rack;
13510 
13511 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13512 	if (rack->rc_allow_data_af_clo == 0) {
13513 	close_now:
13514 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13515 		/* tcp_close will kill the inp pre-log the Reset */
13516 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13517 		tp = tcp_close(tp);
13518 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13519 		ctf_do_dropwithreset(m, tp, th, *tlen);
13520 		return (1);
13521 	}
13522 	if (sbavail(&so->so_snd) == 0)
13523 		goto close_now;
13524 	/* Ok we allow data that is ignored and a followup reset */
13525 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13526 	tp->rcv_nxt = th->th_seq + *tlen;
13527 	tp->t_flags2 |= TF2_DROP_AF_DATA;
13528 	rack->r_wanted_output = 1;
13529 	*tlen = 0;
13530 	return (0);
13531 }
13532 
13533 /*
13534  * Return value of 1, the TCB is unlocked and most
13535  * likely gone, return value of 0, the TCP is still
13536  * locked.
13537  */
13538 static int
13539 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13540     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13541     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13542 {
13543 	int32_t ret_val = 0;
13544 	int32_t orig_tlen = tlen;
13545 	int32_t ourfinisacked = 0;
13546 
13547 	ctf_calc_rwin(so, tp);
13548 
13549 	if ((thflags & TH_RST) ||
13550 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13551 		return (ctf_process_rst(m, th, so, tp));
13552 	/*
13553 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13554 	 * synchronized state.
13555 	 */
13556 	if (thflags & TH_SYN) {
13557 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13558 		return (ret_val);
13559 	}
13560 	/*
13561 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13562 	 * it's less than ts_recent, drop it.
13563 	 */
13564 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13565 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13566 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13567 			return (ret_val);
13568 	}
13569 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13570 		return (ret_val);
13571 	}
13572 	/*
13573 	 * If new data are received on a connection after the user processes
13574 	 * are gone, then RST the other end.
13575 	 */
13576 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13577 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13578 		return (1);
13579 	/*
13580 	 * If last ACK falls within this segment's sequence numbers, record
13581 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13582 	 * from the latest proposal of the tcplw@cray.com list (Braden
13583 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13584 	 * with our earlier PAWS tests, so this check should be solely
13585 	 * predicated on the sequence space of this segment. 3) That we
13586 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13587 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13588 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13589 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13590 	 * p.869. In such cases, we can still calculate the RTT correctly
13591 	 * when RCV.NXT == Last.ACK.Sent.
13592 	 */
13593 	if ((to->to_flags & TOF_TS) != 0 &&
13594 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13595 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13596 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13597 		tp->ts_recent_age = tcp_ts_getticks();
13598 		tp->ts_recent = to->to_tsval;
13599 	}
13600 	/*
13601 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13602 	 * is on (half-synchronized state), then queue data for later
13603 	 * processing; else drop segment and return.
13604 	 */
13605 	if ((thflags & TH_ACK) == 0) {
13606 		if (tp->t_flags & TF_NEEDSYN) {
13607 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13608 			    tiwin, thflags, nxt_pkt));
13609 		} else if (tp->t_flags & TF_ACKNOW) {
13610 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13611 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13612 			return (ret_val);
13613 		} else {
13614 			ctf_do_drop(m, NULL);
13615 			return (0);
13616 		}
13617 	}
13618 	/*
13619 	 * Ack processing.
13620 	 */
13621 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13622 		return (ret_val);
13623 	}
13624 	if (ourfinisacked) {
13625 		/*
13626 		 * If we can't receive any more data, then closing user can
13627 		 * proceed. Starting the timer is contrary to the
13628 		 * specification, but if we don't get a FIN we'll hang
13629 		 * forever.
13630 		 *
13631 		 * XXXjl: we should release the tp also, and use a
13632 		 * compressed state.
13633 		 */
13634 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13635 			soisdisconnected(so);
13636 			tcp_timer_activate(tp, TT_2MSL,
13637 			    (tcp_fast_finwait2_recycle ?
13638 			    tcp_finwait2_timeout :
13639 			    TP_MAXIDLE(tp)));
13640 		}
13641 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
13642 	}
13643 	if (sbavail(&so->so_snd)) {
13644 		if (ctf_progress_timeout_check(tp, true)) {
13645 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13646 						tp, tick, PROGRESS_DROP, __LINE__);
13647 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
13648 			return (1);
13649 		}
13650 	}
13651 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13652 	    tiwin, thflags, nxt_pkt));
13653 }
13654 
13655 /*
13656  * Return value of 1, the TCB is unlocked and most
13657  * likely gone, return value of 0, the TCP is still
13658  * locked.
13659  */
13660 static int
13661 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13662     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13663     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13664 {
13665 	int32_t ret_val = 0;
13666 	int32_t orig_tlen = tlen;
13667 	int32_t ourfinisacked = 0;
13668 
13669 	ctf_calc_rwin(so, tp);
13670 
13671 	if ((thflags & TH_RST) ||
13672 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13673 		return (ctf_process_rst(m, th, so, tp));
13674 	/*
13675 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13676 	 * synchronized state.
13677 	 */
13678 	if (thflags & TH_SYN) {
13679 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13680 		return (ret_val);
13681 	}
13682 	/*
13683 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13684 	 * it's less than ts_recent, drop it.
13685 	 */
13686 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13687 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13688 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13689 			return (ret_val);
13690 	}
13691 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13692 		return (ret_val);
13693 	}
13694 	/*
13695 	 * If last ACK falls within this segment's sequence numbers, record
13696 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13697 	 * from the latest proposal of the tcplw@cray.com list (Braden
13698 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13699 	 * with our earlier PAWS tests, so this check should be solely
13700 	 * predicated on the sequence space of this segment. 3) That we
13701 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13702 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13703 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13704 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13705 	 * p.869. In such cases, we can still calculate the RTT correctly
13706 	 * when RCV.NXT == Last.ACK.Sent.
13707 	 */
13708 	if ((to->to_flags & TOF_TS) != 0 &&
13709 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13710 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13711 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13712 		tp->ts_recent_age = tcp_ts_getticks();
13713 		tp->ts_recent = to->to_tsval;
13714 	}
13715 	/*
13716 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13717 	 * is on (half-synchronized state), then queue data for later
13718 	 * processing; else drop segment and return.
13719 	 */
13720 	if ((thflags & TH_ACK) == 0) {
13721 		if (tp->t_flags & TF_NEEDSYN) {
13722 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13723 			    tiwin, thflags, nxt_pkt));
13724 		} else if (tp->t_flags & TF_ACKNOW) {
13725 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13726 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13727 			return (ret_val);
13728 		} else {
13729 			ctf_do_drop(m, NULL);
13730 			return (0);
13731 		}
13732 	}
13733 	/*
13734 	 * Ack processing.
13735 	 */
13736 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13737 		return (ret_val);
13738 	}
13739 	if (ourfinisacked) {
13740 		tcp_twstart(tp);
13741 		m_freem(m);
13742 		return (1);
13743 	}
13744 	if (sbavail(&so->so_snd)) {
13745 		if (ctf_progress_timeout_check(tp, true)) {
13746 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13747 						tp, tick, PROGRESS_DROP, __LINE__);
13748 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
13749 			return (1);
13750 		}
13751 	}
13752 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13753 	    tiwin, thflags, nxt_pkt));
13754 }
13755 
13756 /*
13757  * Return value of 1, the TCB is unlocked and most
13758  * likely gone, return value of 0, the TCP is still
13759  * locked.
13760  */
13761 static int
13762 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13763     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13764     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13765 {
13766 	int32_t ret_val = 0;
13767 	int32_t orig_tlen;
13768 	int32_t ourfinisacked = 0;
13769 
13770 	ctf_calc_rwin(so, tp);
13771 
13772 	if ((thflags & TH_RST) ||
13773 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13774 		return (ctf_process_rst(m, th, so, tp));
13775 	/*
13776 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13777 	 * synchronized state.
13778 	 */
13779 	if (thflags & TH_SYN) {
13780 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13781 		return (ret_val);
13782 	}
13783 	/*
13784 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13785 	 * it's less than ts_recent, drop it.
13786 	 */
13787 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13788 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13789 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13790 			return (ret_val);
13791 	}
13792 	orig_tlen = tlen;
13793 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13794 		return (ret_val);
13795 	}
13796 	/*
13797 	 * If last ACK falls within this segment's sequence numbers, record
13798 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13799 	 * from the latest proposal of the tcplw@cray.com list (Braden
13800 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13801 	 * with our earlier PAWS tests, so this check should be solely
13802 	 * predicated on the sequence space of this segment. 3) That we
13803 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13804 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13805 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13806 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13807 	 * p.869. In such cases, we can still calculate the RTT correctly
13808 	 * when RCV.NXT == Last.ACK.Sent.
13809 	 */
13810 	if ((to->to_flags & TOF_TS) != 0 &&
13811 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13812 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13813 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13814 		tp->ts_recent_age = tcp_ts_getticks();
13815 		tp->ts_recent = to->to_tsval;
13816 	}
13817 	/*
13818 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13819 	 * is on (half-synchronized state), then queue data for later
13820 	 * processing; else drop segment and return.
13821 	 */
13822 	if ((thflags & TH_ACK) == 0) {
13823 		if (tp->t_flags & TF_NEEDSYN) {
13824 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13825 			    tiwin, thflags, nxt_pkt));
13826 		} else if (tp->t_flags & TF_ACKNOW) {
13827 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13828 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13829 			return (ret_val);
13830 		} else {
13831 			ctf_do_drop(m, NULL);
13832 			return (0);
13833 		}
13834 	}
13835 	/*
13836 	 * case TCPS_LAST_ACK: Ack processing.
13837 	 */
13838 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13839 		return (ret_val);
13840 	}
13841 	if (ourfinisacked) {
13842 		tp = tcp_close(tp);
13843 		ctf_do_drop(m, tp);
13844 		return (1);
13845 	}
13846 	if (sbavail(&so->so_snd)) {
13847 		if (ctf_progress_timeout_check(tp, true)) {
13848 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13849 						tp, tick, PROGRESS_DROP, __LINE__);
13850 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
13851 			return (1);
13852 		}
13853 	}
13854 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13855 	    tiwin, thflags, nxt_pkt));
13856 }
13857 
13858 /*
13859  * Return value of 1, the TCB is unlocked and most
13860  * likely gone, return value of 0, the TCP is still
13861  * locked.
13862  */
13863 static int
13864 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
13865     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13866     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13867 {
13868 	int32_t ret_val = 0;
13869 	int32_t orig_tlen = tlen;
13870 	int32_t ourfinisacked = 0;
13871 
13872 	ctf_calc_rwin(so, tp);
13873 
13874 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
13875 	if ((thflags & TH_RST) ||
13876 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13877 		return (ctf_process_rst(m, th, so, tp));
13878 	/*
13879 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13880 	 * synchronized state.
13881 	 */
13882 	if (thflags & TH_SYN) {
13883 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13884 		return (ret_val);
13885 	}
13886 	/*
13887 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13888 	 * it's less than ts_recent, drop it.
13889 	 */
13890 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13891 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13892 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13893 			return (ret_val);
13894 	}
13895 	if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13896 		return (ret_val);
13897 	}
13898 	/*
13899 	 * If new data are received on a connection after the user processes
13900 	 * are gone, then RST the other end.
13901 	 */
13902 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13903 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13904 		return (1);
13905 	/*
13906 	 * If last ACK falls within this segment's sequence numbers, record
13907 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13908 	 * from the latest proposal of the tcplw@cray.com list (Braden
13909 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13910 	 * with our earlier PAWS tests, so this check should be solely
13911 	 * predicated on the sequence space of this segment. 3) That we
13912 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13913 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13914 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13915 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13916 	 * p.869. In such cases, we can still calculate the RTT correctly
13917 	 * when RCV.NXT == Last.ACK.Sent.
13918 	 */
13919 	if ((to->to_flags & TOF_TS) != 0 &&
13920 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13921 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13922 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13923 		tp->ts_recent_age = tcp_ts_getticks();
13924 		tp->ts_recent = to->to_tsval;
13925 	}
13926 	/*
13927 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13928 	 * is on (half-synchronized state), then queue data for later
13929 	 * processing; else drop segment and return.
13930 	 */
13931 	if ((thflags & TH_ACK) == 0) {
13932 		if (tp->t_flags & TF_NEEDSYN) {
13933 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13934 			    tiwin, thflags, nxt_pkt));
13935 		} else if (tp->t_flags & TF_ACKNOW) {
13936 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13937 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13938 			return (ret_val);
13939 		} else {
13940 			ctf_do_drop(m, NULL);
13941 			return (0);
13942 		}
13943 	}
13944 	/*
13945 	 * Ack processing.
13946 	 */
13947 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13948 		return (ret_val);
13949 	}
13950 	if (sbavail(&so->so_snd)) {
13951 		if (ctf_progress_timeout_check(tp, true)) {
13952 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13953 						tp, tick, PROGRESS_DROP, __LINE__);
13954 			ctf_do_dropwithreset_conn(m, tp, th, tlen);
13955 			return (1);
13956 		}
13957 	}
13958 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13959 	    tiwin, thflags, nxt_pkt));
13960 }
13961 
13962 static void inline
13963 rack_clear_rate_sample(struct tcp_rack *rack)
13964 {
13965 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
13966 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
13967 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
13968 }
13969 
13970 static void
13971 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
13972 {
13973 	uint64_t bw_est, rate_wanted;
13974 	int chged = 0;
13975 	uint32_t user_max, orig_min, orig_max;
13976 
13977 #ifdef TCP_REQUEST_TRK
13978 	if (rack->rc_hybrid_mode &&
13979 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
13980 	    (rack_hybrid_allow_set_maxseg == 1) &&
13981 	    (rack->r_ctl.rc_last_sft != NULL)) {
13982 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
13983 		return;
13984 	}
13985 #endif
13986 	orig_min = rack->r_ctl.rc_pace_min_segs;
13987 	orig_max = rack->r_ctl.rc_pace_max_segs;
13988 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
13989 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
13990 		chged = 1;
13991 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
13992 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
13993 		if (user_max != rack->r_ctl.rc_pace_max_segs)
13994 			chged = 1;
13995 	}
13996 	if (rack->rc_force_max_seg) {
13997 		rack->r_ctl.rc_pace_max_segs = user_max;
13998 	} else if (rack->use_fixed_rate) {
13999 		bw_est = rack_get_bw(rack);
14000 		if ((rack->r_ctl.crte == NULL) ||
14001 		    (bw_est != rack->r_ctl.crte->rate)) {
14002 			rack->r_ctl.rc_pace_max_segs = user_max;
14003 		} else {
14004 			/* We are pacing right at the hardware rate */
14005 			uint32_t segsiz, pace_one;
14006 
14007 			if (rack_pace_one_seg ||
14008 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14009 				pace_one = 1;
14010 			else
14011 				pace_one = 0;
14012 			segsiz = min(ctf_fixed_maxseg(tp),
14013 				     rack->r_ctl.rc_pace_min_segs);
14014 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14015 				tp, bw_est, segsiz, pace_one,
14016 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14017 		}
14018 	} else if (rack->rc_always_pace) {
14019 		if (rack->r_ctl.gp_bw ||
14020 		    rack->r_ctl.init_rate) {
14021 			/* We have a rate of some sort set */
14022 			uint32_t  orig;
14023 
14024 			bw_est = rack_get_bw(rack);
14025 			orig = rack->r_ctl.rc_pace_max_segs;
14026 			if (fill_override)
14027 				rate_wanted = *fill_override;
14028 			else
14029 				rate_wanted = rack_get_gp_est(rack);
14030 			if (rate_wanted) {
14031 				/* We have something */
14032 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14033 										   rate_wanted,
14034 										   ctf_fixed_maxseg(rack->rc_tp));
14035 			} else
14036 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14037 			if (orig != rack->r_ctl.rc_pace_max_segs)
14038 				chged = 1;
14039 		} else if ((rack->r_ctl.gp_bw == 0) &&
14040 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14041 			/*
14042 			 * If we have nothing limit us to bursting
14043 			 * out IW sized pieces.
14044 			 */
14045 			chged = 1;
14046 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14047 		}
14048 	}
14049 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14050 		chged = 1;
14051 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14052 	}
14053 	if (chged)
14054 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14055 }
14056 
14057 
14058 static void
14059 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14060 {
14061 #ifdef INET6
14062 	struct ip6_hdr *ip6 = NULL;
14063 #endif
14064 #ifdef INET
14065 	struct ip *ip = NULL;
14066 #endif
14067 	struct udphdr *udp = NULL;
14068 
14069 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14070 #ifdef INET6
14071 	if (rack->r_is_v6) {
14072 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14073 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14074 		if (tp->t_port) {
14075 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14076 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14077 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14078 			udp->uh_dport = tp->t_port;
14079 			rack->r_ctl.fsb.udp = udp;
14080 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14081 		} else
14082 		{
14083 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14084 			rack->r_ctl.fsb.udp = NULL;
14085 		}
14086 		tcpip_fillheaders(rack->rc_inp,
14087 				  tp->t_port,
14088 				  ip6, rack->r_ctl.fsb.th);
14089 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14090 	} else
14091 #endif				/* INET6 */
14092 #ifdef INET
14093 	{
14094 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14095 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14096 		if (tp->t_port) {
14097 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14098 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14099 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14100 			udp->uh_dport = tp->t_port;
14101 			rack->r_ctl.fsb.udp = udp;
14102 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14103 		} else
14104 		{
14105 			rack->r_ctl.fsb.udp = NULL;
14106 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14107 		}
14108 		tcpip_fillheaders(rack->rc_inp,
14109 				  tp->t_port,
14110 				  ip, rack->r_ctl.fsb.th);
14111 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14112 	}
14113 #endif
14114 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14115 	    (long)TCP_MAXWIN << tp->rcv_scale);
14116 	rack->r_fsb_inited = 1;
14117 }
14118 
14119 static int
14120 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14121 {
14122 	/*
14123 	 * Allocate the larger of spaces V6 if available else just
14124 	 * V4 and include udphdr (overbook)
14125 	 */
14126 #ifdef INET6
14127 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14128 #else
14129 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14130 #endif
14131 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14132 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14133 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14134 		return (ENOMEM);
14135 	}
14136 	rack->r_fsb_inited = 0;
14137 	return (0);
14138 }
14139 
14140 static void
14141 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14142 {
14143 	/*
14144 	 * Types of logs (mod value)
14145 	 * 20 - Initial round setup
14146 	 * 21 - Rack declares a new round.
14147 	 */
14148 	struct tcpcb *tp;
14149 
14150 	tp = rack->rc_tp;
14151 	if (tcp_bblogging_on(tp)) {
14152 		union tcp_log_stackspecific log;
14153 		struct timeval tv;
14154 
14155 		memset(&log, 0, sizeof(log));
14156 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14157 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14158 		log.u_bbr.flex3 = high_seq;
14159 		log.u_bbr.flex4 = tp->snd_max;
14160 		log.u_bbr.flex8 = mod;
14161 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14162 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14163 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14164 		TCP_LOG_EVENTP(tp, NULL,
14165 		    &tptosocket(tp)->so_rcv,
14166 		    &tptosocket(tp)->so_snd,
14167 		    TCP_HYSTART, 0,
14168 		    0, &log, false, &tv);
14169 	}
14170 }
14171 
14172 static void
14173 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14174 {
14175 	rack->rack_deferred_inited = 1;
14176 	rack->r_ctl.roundends = tp->snd_max;
14177 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14178 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14179 }
14180 
14181 static void
14182 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14183 {
14184 	/* Retransmit bit controls.
14185 	 *
14186 	 * The setting of these values control one of
14187 	 * three settings you can have and dictate
14188 	 * how rack does retransmissions. Note this
14189 	 * is in *any* mode i.e. pacing on or off DGP
14190 	 * fixed rate pacing, or just bursting rack.
14191 	 *
14192 	 * 1 - Use full sized retransmits i.e. limit
14193 	 *     the size to whatever the pace_max_segments
14194 	 *     size is.
14195 	 *
14196 	 * 2 - Use pacer min granularity as a guide to
14197 	 *     the size combined with the current calculated
14198 	 *     goodput b/w measurement. So for example if
14199 	 *     the goodput is measured at 20Mbps we would
14200 	 *     calculate 8125 (pacer minimum 250usec in
14201 	 *     that b/w) and then round it up to the next
14202 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14203 	 *
14204 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14205 	 *     fall here too if we are setting via rack_init()).
14206 	 *
14207 	 */
14208 	if (ctl == 1) {
14209 		rack->full_size_rxt = 1;
14210 		rack->shape_rxt_to_pacing_min  = 0;
14211 	} else if (ctl == 2) {
14212 		rack->full_size_rxt = 0;
14213 		rack->shape_rxt_to_pacing_min  = 1;
14214 	} else {
14215 		rack->full_size_rxt = 0;
14216 		rack->shape_rxt_to_pacing_min  = 0;
14217 	}
14218 }
14219 
14220 static void
14221 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14222 		  uint32_t flex1,
14223 		  uint32_t flex2,
14224 		  uint32_t flex3)
14225 {
14226 	if (tcp_bblogging_on(rack->rc_tp)) {
14227 		union tcp_log_stackspecific log;
14228 		struct timeval tv;
14229 
14230 		memset(&log, 0, sizeof(log));
14231 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14232 		log.u_bbr.flex8 = mod;
14233 		log.u_bbr.flex1 = flex1;
14234 		log.u_bbr.flex2 = flex2;
14235 		log.u_bbr.flex3 = flex3;
14236 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14237 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14238 	}
14239 }
14240 
14241 static int
14242 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14243 {
14244 	struct tcp_rack *rack;
14245 	struct rack_sendmap *rsm;
14246 	int i;
14247 
14248 
14249 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14250 	switch (reqr->req) {
14251 	case TCP_QUERY_SENDMAP:
14252 		if ((reqr->req_param == tp->snd_max) ||
14253 		    (tp->snd_max == tp->snd_una)){
14254 			/* Unlikely */
14255 			return (0);
14256 		}
14257 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14258 		if (rsm == NULL) {
14259 			/* Can't find that seq -- unlikely */
14260 			return (0);
14261 		}
14262 		reqr->sendmap_start = rsm->r_start;
14263 		reqr->sendmap_end = rsm->r_end;
14264 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14265 		reqr->sendmap_fas = rsm->r_fas;
14266 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14267 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14268 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14269 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14270 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14271 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14272 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14273 		reqr->sendmap_dupacks = rsm->r_dupack;
14274 		rack_log_chg_info(tp, rack, 1,
14275 				  rsm->r_start,
14276 				  rsm->r_end,
14277 				  rsm->r_flags);
14278 		return(1);
14279 		break;
14280 	case TCP_QUERY_TIMERS_UP:
14281 		if (rack->r_ctl.rc_hpts_flags == 0) {
14282 			/* no timers up */
14283 			return (0);
14284 		}
14285 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14286 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14287 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14288 		}
14289 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14290 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14291 		}
14292 		rack_log_chg_info(tp, rack, 2,
14293 				  rack->r_ctl.rc_hpts_flags,
14294 				  rack->r_ctl.rc_last_output_to,
14295 				  rack->r_ctl.rc_timer_exp);
14296 		return (1);
14297 		break;
14298 	case TCP_QUERY_RACK_TIMES:
14299 		/* Reordering items */
14300 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14301 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14302 		/* Timerstamps and timers */
14303 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14304 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14305 		reqr->rack_rtt = rack->rc_rack_rtt;
14306 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14307 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14308 		/* PRR data */
14309 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
14310 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14311 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14312 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14313 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14314 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14315 		/* TLP and persists info */
14316 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14317 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14318 		if (rack->rc_in_persist) {
14319 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14320 			reqr->rack_in_persist = 1;
14321 		} else {
14322 			reqr->rack_time_went_idle = 0;
14323 			reqr->rack_in_persist = 0;
14324 		}
14325 		if (rack->r_wanted_output)
14326 			reqr->rack_wanted_output = 1;
14327 		else
14328 			reqr->rack_wanted_output = 0;
14329 		return (1);
14330 		break;
14331 	default:
14332 		return (-EINVAL);
14333 	}
14334 }
14335 
14336 static void
14337 rack_switch_failed(struct tcpcb *tp)
14338 {
14339 	/*
14340 	 * This method gets called if a stack switch was
14341 	 * attempted and it failed. We are left
14342 	 * but our hpts timers were stopped and we
14343 	 * need to validate time units and t_flags2.
14344 	 */
14345 	struct tcp_rack *rack;
14346 	struct timeval tv;
14347 	uint32_t cts;
14348 	uint32_t toval;
14349 	struct hpts_diag diag;
14350 
14351 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14352 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14353 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14354 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14355 	else
14356 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14357 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14358 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14359 	if (tp->t_in_hpts > IHPTS_NONE) {
14360 		/* Strange */
14361 		return;
14362 	}
14363 	cts = tcp_get_usecs(&tv);
14364 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14365 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14366 			toval = rack->r_ctl.rc_last_output_to - cts;
14367 		} else {
14368 			/* one slot please */
14369 			toval = HPTS_USECS_PER_SLOT;
14370 		}
14371 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14372 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14373 			toval = rack->r_ctl.rc_timer_exp - cts;
14374 		} else {
14375 			/* one slot please */
14376 			toval = HPTS_USECS_PER_SLOT;
14377 		}
14378 	} else
14379 		toval = HPTS_USECS_PER_SLOT;
14380 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14381 				   __LINE__, &diag);
14382 	rack_log_hpts_diag(rack, cts, &diag, &tv);
14383 }
14384 
14385 static int
14386 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14387 {
14388 	struct rack_sendmap *rsm, *ersm;
14389 	int insret __diagused;
14390 	/*
14391 	 * When initing outstanding, we must be quite careful
14392 	 * to not refer to tp->t_fb_ptr. This has the old rack
14393 	 * pointer in it, not the "new" one (when we are doing
14394 	 * a stack switch).
14395 	 */
14396 
14397 
14398 	if (tp->t_fb->tfb_chg_query == NULL) {
14399 		/* Create a send map for the current outstanding data */
14400 
14401 		rsm = rack_alloc(rack);
14402 		if (rsm == NULL) {
14403 			uma_zfree(rack_pcb_zone, ptr);
14404 			return (ENOMEM);
14405 		}
14406 		rsm->r_no_rtt_allowed = 1;
14407 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14408 		rsm->r_rtr_cnt = 1;
14409 		rsm->r_rtr_bytes = 0;
14410 		if (tp->t_flags & TF_SENTFIN)
14411 			rsm->r_flags |= RACK_HAS_FIN;
14412 		rsm->r_end = tp->snd_max;
14413 		if (tp->snd_una == tp->iss) {
14414 			/* The data space is one beyond snd_una */
14415 			rsm->r_flags |= RACK_HAS_SYN;
14416 			rsm->r_start = tp->iss;
14417 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14418 		} else
14419 			rsm->r_start = tp->snd_una;
14420 		rsm->r_dupack = 0;
14421 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14422 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14423 			if (rsm->m) {
14424 				rsm->orig_m_len = rsm->m->m_len;
14425 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14426 			} else {
14427 				rsm->orig_m_len = 0;
14428 				rsm->orig_t_space = 0;
14429 			}
14430 		} else {
14431 			/*
14432 			 * This can happen if we have a stand-alone FIN or
14433 			 *  SYN.
14434 			 */
14435 			rsm->m = NULL;
14436 			rsm->orig_m_len = 0;
14437 			rsm->orig_t_space = 0;
14438 			rsm->soff = 0;
14439 		}
14440 #ifdef INVARIANTS
14441 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14442 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14443 			      insret, rack, rsm);
14444 		}
14445 #else
14446 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14447 #endif
14448 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14449 		rsm->r_in_tmap = 1;
14450 	} else {
14451 		/* We have a query mechanism, lets use it */
14452 		struct tcp_query_resp qr;
14453 		int i;
14454 		tcp_seq at;
14455 
14456 		at = tp->snd_una;
14457 		while (at != tp->snd_max) {
14458 			memset(&qr, 0, sizeof(qr));
14459 			qr.req = TCP_QUERY_SENDMAP;
14460 			qr.req_param = at;
14461 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14462 				break;
14463 			/* Move forward */
14464 			at = qr.sendmap_end;
14465 			/* Now lets build the entry for this one */
14466 			rsm = rack_alloc(rack);
14467 			if (rsm == NULL) {
14468 				uma_zfree(rack_pcb_zone, ptr);
14469 				return (ENOMEM);
14470 			}
14471 			memset(rsm, 0, sizeof(struct rack_sendmap));
14472 			/* Now configure the rsm and insert it */
14473 			rsm->r_dupack = qr.sendmap_dupacks;
14474 			rsm->r_start = qr.sendmap_start;
14475 			rsm->r_end = qr.sendmap_end;
14476 			if (qr.sendmap_fas)
14477 				rsm->r_fas = qr.sendmap_end;
14478 			else
14479 				rsm->r_fas = rsm->r_start - tp->snd_una;
14480 			/*
14481 			 * We have carefully aligned the bits
14482 			 * so that all we have to do is copy over
14483 			 * the bits with the mask.
14484 			 */
14485 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14486 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14487 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14488 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14489 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
14490 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
14491 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14492 					   (rsm->r_start - tp->snd_una), &rsm->soff);
14493 			if (rsm->m) {
14494 				rsm->orig_m_len = rsm->m->m_len;
14495 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14496 			} else {
14497 				rsm->orig_m_len = 0;
14498 				rsm->orig_t_space = 0;
14499 			}
14500 #ifdef INVARIANTS
14501 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14502 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14503 				      insret, rack, rsm);
14504 			}
14505 #else
14506 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14507 #endif
14508 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
14509 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14510 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14511 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14512 						/*
14513 						 * If the existing ersm was sent at
14514 						 * a later time than the new one, then
14515 						 * the new one should appear ahead of this
14516 						 * ersm.
14517 						 */
14518 						rsm->r_in_tmap = 1;
14519 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14520 						break;
14521 					}
14522 				}
14523 				if (rsm->r_in_tmap == 0) {
14524 					/*
14525 					 * Not found so shove it on the tail.
14526 					 */
14527 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14528 					rsm->r_in_tmap = 1;
14529 				}
14530  			} else {
14531 				if ((rack->r_ctl.rc_sacklast == NULL) ||
14532 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14533 					rack->r_ctl.rc_sacklast = rsm;
14534 				}
14535 			}
14536 			rack_log_chg_info(tp, rack, 3,
14537 					  rsm->r_start,
14538 					  rsm->r_end,
14539 					  rsm->r_flags);
14540 		}
14541 	}
14542 	return (0);
14543 }
14544 
14545 
14546 static int32_t
14547 rack_init(struct tcpcb *tp, void **ptr)
14548 {
14549 	struct inpcb *inp = tptoinpcb(tp);
14550 	struct tcp_rack *rack = NULL;
14551 	uint32_t iwin, snt, us_cts;
14552 	size_t sz;
14553 	int err, no_query;
14554 
14555 	tcp_hpts_init(tp);
14556 
14557 	/*
14558 	 * First are we the initial or are we a switched stack?
14559 	 * If we are initing via tcp_newtcppcb the ptr passed
14560 	 * will be tp->t_fb_ptr. If its a stack switch that
14561 	 * has a previous stack we can query it will be a local
14562 	 * var that will in the end be set into t_fb_ptr.
14563 	 */
14564 	if (ptr == &tp->t_fb_ptr)
14565 		no_query = 1;
14566 	else
14567 		no_query = 0;
14568 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14569 	if (*ptr == NULL) {
14570 		/*
14571 		 * We need to allocate memory but cant. The INP and INP_INFO
14572 		 * locks and they are recursive (happens during setup. So a
14573 		 * scheme to drop the locks fails :(
14574 		 *
14575 		 */
14576 		return(ENOMEM);
14577 	}
14578 	memset(*ptr, 0, sizeof(struct tcp_rack));
14579 	rack = (struct tcp_rack *)*ptr;
14580 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14581 	if (rack->r_ctl.tqh == NULL) {
14582 		uma_zfree(rack_pcb_zone, rack);
14583 		return(ENOMEM);
14584 	}
14585 	tqhash_init(rack->r_ctl.tqh);
14586 	TAILQ_INIT(&rack->r_ctl.rc_free);
14587 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
14588 	rack->rc_tp = tp;
14589 	rack->rc_inp = inp;
14590 	/* Set the flag */
14591 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14592 	/* Probably not needed but lets be sure */
14593 	rack_clear_rate_sample(rack);
14594 	/*
14595 	 * Save off the default values, socket options will poke
14596 	 * at these if pacing is not on or we have not yet
14597 	 * reached where pacing is on (gp_ready/fixed enabled).
14598 	 * When they get set into the CC module (when gp_ready
14599 	 * is enabled or we enable fixed) then we will set these
14600 	 * values into the CC and place in here the old values
14601 	 * so we have a restoral. Then we will set the flag
14602 	 * rc_pacing_cc_set. That way whenever we turn off pacing
14603 	 * or switch off this stack, we will know to go restore
14604 	 * the saved values.
14605 	 *
14606 	 * We specifically put into the beta the ecn value for pacing.
14607 	 */
14608 	rack->rc_new_rnd_needed = 1;
14609 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14610 	/* We want abe like behavior as well */
14611 
14612 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14613 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14614 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14615 	if (rack_fill_cw_state)
14616 		rack->rc_pace_to_cwnd = 1;
14617 	if (rack_pacing_min_seg)
14618 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14619 	if (use_rack_rr)
14620 		rack->use_rack_rr = 1;
14621 	if (rack_dnd_default) {
14622 		rack->rc_pace_dnd = 1;
14623 	}
14624 	if (V_tcp_delack_enabled)
14625 		tp->t_delayed_ack = 1;
14626 	else
14627 		tp->t_delayed_ack = 0;
14628 #ifdef TCP_ACCOUNTING
14629 	if (rack_tcp_accounting) {
14630 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14631 	}
14632 #endif
14633 	rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
14634 	sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
14635 	rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
14636 	if (rack->r_ctl.pcm_s == NULL) {
14637 		rack->r_ctl.pcm_i.cnt_alloc = 0;
14638 	}
14639 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14640 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14641 	if (rack_enable_shared_cwnd)
14642 		rack->rack_enable_scwnd = 1;
14643 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14644 	rack->rc_user_set_max_segs = rack_hptsi_segments;
14645 	rack->r_ctl.max_reduction = rack_max_reduce;
14646 	rack->rc_force_max_seg = 0;
14647 	TAILQ_INIT(&rack->r_ctl.opt_list);
14648 	rack->r_ctl.rc_saved_beta = V_newreno_beta_ecn;
14649 	rack->r_ctl.rc_saved_beta_ecn = V_newreno_beta_ecn;
14650 	if (rack_hibeta_setting) {
14651 		rack->rack_hibeta = 1;
14652 		if ((rack_hibeta_setting >= 50) &&
14653 		    (rack_hibeta_setting <= 100)) {
14654 			rack->r_ctl.rc_saved_beta = rack_hibeta_setting;
14655 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
14656 		}
14657 	} else {
14658 		rack->r_ctl.saved_hibeta = 50;
14659 	}
14660 	/*
14661 	 * We initialize to all ones so we never match 0
14662 	 * just in case the client sends in 0, it hopefully
14663 	 * will never have all 1's in ms :-)
14664 	 */
14665 	rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
14666 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14667 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14668 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14669 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14670 	rack->r_ctl.rc_highest_us_rtt = 0;
14671 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14672 	rack->pcm_enabled = rack_pcm_is_enabled;
14673 	if (rack_fillcw_bw_cap)
14674 		rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
14675 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14676 	if (rack_use_cmp_acks)
14677 		rack->r_use_cmp_ack = 1;
14678 	if (rack_disable_prr)
14679 		rack->rack_no_prr = 1;
14680 	if (rack_gp_no_rec_chg)
14681 		rack->rc_gp_no_rec_chg = 1;
14682 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14683 		rack->r_ctl.pacing_method |= RACK_REG_PACING;
14684 		rack->rc_always_pace = 1;
14685 		if (rack->rack_hibeta)
14686 			rack_set_cc_pacing(rack);
14687 	} else
14688 		rack->rc_always_pace = 0;
14689 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14690 		rack->r_mbuf_queue = 1;
14691 	else
14692 		rack->r_mbuf_queue = 0;
14693 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
14694 	if (rack_limits_scwnd)
14695 		rack->r_limit_scw = 1;
14696 	else
14697 		rack->r_limit_scw = 0;
14698 	rack_init_retransmit_value(rack, rack_rxt_controls);
14699 	rack->rc_labc = V_tcp_abc_l_var;
14700 	if (rack_honors_hpts_min_to)
14701 		rack->r_use_hpts_min = 1;
14702 	if (tp->snd_una != 0) {
14703 		rack->rc_sendvars_notset = 0;
14704 		/*
14705 		 * Make sure any TCP timers are not running.
14706 		 */
14707 		tcp_timer_stop(tp);
14708 	} else {
14709 		/*
14710 		 * Server side, we are called from the
14711 		 * syn-cache. This means none of the
14712 		 * snd_una/max are set yet so we have
14713 		 * to defer this until the first send.
14714 		 */
14715 		rack->rc_sendvars_notset = 1;
14716 	}
14717 
14718 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14719 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14720 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14721 	rack->r_ctl.rc_min_to = rack_min_to;
14722 	microuptime(&rack->r_ctl.act_rcv_time);
14723 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14724 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14725 	if (rack_hw_up_only)
14726 		rack->r_up_only = 1;
14727 	if (rack_do_dyn_mul) {
14728 		/* When dynamic adjustment is on CA needs to start at 100% */
14729 		rack->rc_gp_dyn_mul = 1;
14730 		if (rack_do_dyn_mul >= 100)
14731 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14732 	} else
14733 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14734 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14735 	if (rack_timely_off) {
14736 		rack->rc_skip_timely = 1;
14737 	}
14738 	if (rack->rc_skip_timely) {
14739 		rack->r_ctl.rack_per_of_gp_rec = 90;
14740 		rack->r_ctl.rack_per_of_gp_ca = 100;
14741 		rack->r_ctl.rack_per_of_gp_ss = 250;
14742 	}
14743 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14744 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_msec(&rack->r_ctl.act_rcv_time);
14745 	rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_msec(&rack->r_ctl.act_rcv_time);
14746 
14747 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14748 				rack_probertt_filter_life);
14749 	us_cts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
14750 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14751 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
14752 	rack->r_ctl.rc_went_idle_time = us_cts;
14753 	rack->r_ctl.rc_time_probertt_starts = 0;
14754 
14755 	rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
14756 	if (rack_rnd_cnt_req  & 0x10000)
14757 		rack->r_ctl.gate_to_fs = 1;
14758 	rack->r_ctl.gp_gain_req = rack_gp_gain_req;
14759 	if ((rack_rnd_cnt_req & 0x100) > 0) {
14760 
14761 	}
14762 	if (rack_dsack_std_based & 0x1) {
14763 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14764 		rack->rc_rack_tmr_std_based = 1;
14765 	}
14766 	if (rack_dsack_std_based & 0x2) {
14767 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
14768 		rack->rc_rack_use_dsack = 1;
14769 	}
14770 	/* We require at least one measurement, even if the sysctl is 0 */
14771 	if (rack_req_measurements)
14772 		rack->r_ctl.req_measurements = rack_req_measurements;
14773 	else
14774 		rack->r_ctl.req_measurements = 1;
14775 	if (rack_enable_hw_pacing)
14776 		rack->rack_hdw_pace_ena = 1;
14777 	if (rack_hw_rate_caps)
14778 		rack->r_rack_hw_rate_caps = 1;
14779 	if (rack_non_rxt_use_cr)
14780 		rack->rack_rec_nonrxt_use_cr = 1;
14781 	/* Lets setup the fsb block */
14782 	err = rack_init_fsb(tp, rack);
14783 	if (err) {
14784 		uma_zfree(rack_pcb_zone, *ptr);
14785 		*ptr = NULL;
14786 		return (err);
14787 	}
14788 	if (rack_do_hystart) {
14789 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
14790 		if (rack_do_hystart > 1)
14791 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
14792 		if (rack_do_hystart > 2)
14793 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
14794 	}
14795 	/* Log what we will do with queries */
14796 	rack_log_chg_info(tp, rack, 7,
14797 			  no_query, 0, 0);
14798 	if (rack_def_profile)
14799 		rack_set_profile(rack, rack_def_profile);
14800 	/* Cancel the GP measurement in progress */
14801 	tp->t_flags &= ~TF_GPUTINPROG;
14802 	if ((tp->t_state != TCPS_CLOSED) &&
14803 	    (tp->t_state != TCPS_TIME_WAIT)) {
14804 		/*
14805 		 * We are already open, we may
14806 		 * need to adjust a few things.
14807 		 */
14808 		if (SEQ_GT(tp->snd_max, tp->iss))
14809 			snt = tp->snd_max - tp->iss;
14810 		else
14811 			snt = 0;
14812 		iwin = rc_init_window(rack);
14813 		if ((snt < iwin) &&
14814 		    (no_query == 1)) {
14815 			/* We are not past the initial window
14816 			 * on the first init (i.e. a stack switch
14817 			 * has not yet occured) so we need to make
14818 			 * sure cwnd and ssthresh is correct.
14819 			 */
14820 			if (tp->snd_cwnd < iwin)
14821 				tp->snd_cwnd = iwin;
14822 			/*
14823 			 * If we are within the initial window
14824 			 * we want ssthresh to be unlimited. Setting
14825 			 * it to the rwnd (which the default stack does
14826 			 * and older racks) is not really a good idea
14827 			 * since we want to be in SS and grow both the
14828 			 * cwnd and the rwnd (via dynamic rwnd growth). If
14829 			 * we set it to the rwnd then as the peer grows its
14830 			 * rwnd we will be stuck in CA and never hit SS.
14831 			 *
14832 			 * Its far better to raise it up high (this takes the
14833 			 * risk that there as been a loss already, probably
14834 			 * we should have an indicator in all stacks of loss
14835 			 * but we don't), but considering the normal use this
14836 			 * is a risk worth taking. The consequences of not
14837 			 * hitting SS are far worse than going one more time
14838 			 * into it early on (before we have sent even a IW).
14839 			 * It is highly unlikely that we will have had a loss
14840 			 * before getting the IW out.
14841 			 */
14842 			tp->snd_ssthresh = 0xffffffff;
14843 		}
14844 		/*
14845 		 * Any init based on sequence numbers
14846 		 * should be done in the deferred init path
14847 		 * since we can be CLOSED and not have them
14848 		 * inited when rack_init() is called. We
14849 		 * are not closed so lets call it.
14850 		 */
14851 		rack_deferred_init(tp, rack);
14852 	}
14853 	if ((tp->t_state != TCPS_CLOSED) &&
14854 	    (tp->t_state != TCPS_TIME_WAIT) &&
14855 	    (no_query == 0) &&
14856 	    (tp->snd_una != tp->snd_max))  {
14857 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
14858 		if (err) {
14859 			*ptr = NULL;
14860 			return(err);
14861 		}
14862 	}
14863 	rack_stop_all_timers(tp, rack);
14864 	/* Setup all the t_flags2 */
14865 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14866 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14867 	else
14868 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14869 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14870 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14871 	/*
14872 	 * Timers in Rack are kept in microseconds so lets
14873 	 * convert any initial incoming variables
14874 	 * from ticks into usecs. Note that we
14875 	 * also change the values of t_srtt and t_rttvar, if
14876 	 * they are non-zero. They are kept with a 5
14877 	 * bit decimal so we have to carefully convert
14878 	 * these to get the full precision.
14879 	 */
14880 	rack_convert_rtts(tp);
14881 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
14882 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
14883 		/* We do not start any timers on DROPPED connections */
14884 		if (tp->t_fb->tfb_chg_query == NULL) {
14885 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14886 		} else {
14887 			struct tcp_query_resp qr;
14888 			int ret;
14889 
14890 			memset(&qr, 0, sizeof(qr));
14891 
14892 			/* Get the misc time stamps and such for rack */
14893 			qr.req = TCP_QUERY_RACK_TIMES;
14894 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14895 			if (ret == 1) {
14896 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
14897 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
14898 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
14899 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
14900 				rack->rc_rack_rtt = qr.rack_rtt;
14901 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
14902 				rack->r_ctl.rc_sacked = qr.rack_sacked;
14903 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
14904 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
14905 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
14906 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
14907 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
14908 				if (qr.rack_tlp_out) {
14909 					rack->rc_tlp_in_progress = 1;
14910 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
14911 				} else {
14912 					rack->rc_tlp_in_progress = 0;
14913 					rack->r_ctl.rc_tlp_cnt_out = 0;
14914 				}
14915 				if (qr.rack_srtt_measured)
14916 					rack->rc_srtt_measure_made = 1;
14917 				if (qr.rack_in_persist == 1) {
14918 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
14919 #ifdef NETFLIX_SHARED_CWND
14920 					if (rack->r_ctl.rc_scw) {
14921 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
14922 						rack->rack_scwnd_is_idle = 1;
14923 					}
14924 #endif
14925 					rack->r_ctl.persist_lost_ends = 0;
14926 					rack->probe_not_answered = 0;
14927 					rack->forced_ack = 0;
14928 					tp->t_rxtshift = 0;
14929 					rack->rc_in_persist = 1;
14930 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
14931 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
14932 				}
14933 				if (qr.rack_wanted_output)
14934 					rack->r_wanted_output = 1;
14935 				rack_log_chg_info(tp, rack, 6,
14936 						  qr.rack_min_rtt,
14937 						  qr.rack_rtt,
14938 						  qr.rack_reorder_ts);
14939 			}
14940 			/* Get the old stack timers */
14941 			qr.req_param = 0;
14942 			qr.req = TCP_QUERY_TIMERS_UP;
14943 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14944 			if (ret) {
14945 				/*
14946 				 * non-zero return means we have a timer('s)
14947 				 * to start. Zero means no timer (no keepalive
14948 				 * I suppose).
14949 				 */
14950 				uint32_t tov = 0;
14951 
14952 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
14953 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
14954 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
14955 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
14956 						tov = qr.timer_pacing_to - us_cts;
14957 					else
14958 						tov = HPTS_USECS_PER_SLOT;
14959 				}
14960 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
14961 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
14962 					if (tov == 0) {
14963 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
14964 							tov = qr.timer_timer_exp - us_cts;
14965 						else
14966 							tov = HPTS_USECS_PER_SLOT;
14967 					}
14968 				}
14969 				rack_log_chg_info(tp, rack, 4,
14970 						  rack->r_ctl.rc_hpts_flags,
14971 						  rack->r_ctl.rc_last_output_to,
14972 						  rack->r_ctl.rc_timer_exp);
14973 				if (tov) {
14974 					struct hpts_diag diag;
14975 
14976 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
14977 								   __LINE__, &diag);
14978 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
14979 				}
14980 			}
14981 		}
14982 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
14983 				     __LINE__, RACK_RTTS_INIT);
14984 	}
14985 	return (0);
14986 }
14987 
14988 static int
14989 rack_handoff_ok(struct tcpcb *tp)
14990 {
14991 	if ((tp->t_state == TCPS_CLOSED) ||
14992 	    (tp->t_state == TCPS_LISTEN)) {
14993 		/* Sure no problem though it may not stick */
14994 		return (0);
14995 	}
14996 	if ((tp->t_state == TCPS_SYN_SENT) ||
14997 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
14998 		/*
14999 		 * We really don't know if you support sack,
15000 		 * you have to get to ESTAB or beyond to tell.
15001 		 */
15002 		return (EAGAIN);
15003 	}
15004 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15005 		/*
15006 		 * Rack will only send a FIN after all data is acknowledged.
15007 		 * So in this case we have more data outstanding. We can't
15008 		 * switch stacks until either all data and only the FIN
15009 		 * is left (in which case rack_init() now knows how
15010 		 * to deal with that) <or> all is acknowledged and we
15011 		 * are only left with incoming data, though why you
15012 		 * would want to switch to rack after all data is acknowledged
15013 		 * I have no idea (rrs)!
15014 		 */
15015 		return (EAGAIN);
15016 	}
15017 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15018 		return (0);
15019 	}
15020 	/*
15021 	 * If we reach here we don't do SACK on this connection so we can
15022 	 * never do rack.
15023 	 */
15024 	return (EINVAL);
15025 }
15026 
15027 static void
15028 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15029 {
15030 
15031 	if (tp->t_fb_ptr) {
15032 		uint32_t cnt_free = 0;
15033 		struct tcp_rack *rack;
15034 		struct rack_sendmap *rsm;
15035 
15036 		tcp_handle_orphaned_packets(tp);
15037 		tp->t_flags &= ~TF_FORCEDATA;
15038 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15039 		rack_log_pacing_delay_calc(rack,
15040 					   0,
15041 					   0,
15042 					   0,
15043 					   rack_get_gp_est(rack), /* delRate */
15044 					   rack_get_lt_bw(rack), /* rttProp */
15045 					   20, __LINE__, NULL, 0);
15046 #ifdef NETFLIX_SHARED_CWND
15047 		if (rack->r_ctl.rc_scw) {
15048 			uint32_t limit;
15049 
15050 			if (rack->r_limit_scw)
15051 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15052 			else
15053 				limit = 0;
15054 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15055 						  rack->r_ctl.rc_scw_index,
15056 						  limit);
15057 			rack->r_ctl.rc_scw = NULL;
15058 		}
15059 #endif
15060 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15061 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15062 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15063 			rack->r_ctl.fsb.th = NULL;
15064 		}
15065 		if (rack->rc_always_pace == 1) {
15066 			rack_remove_pacing(rack);
15067 		}
15068 		/* Clean up any options if they were not applied */
15069 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15070 			struct deferred_opt_list *dol;
15071 
15072 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15073 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15074 			free(dol, M_TCPDO);
15075 		}
15076 		/* rack does not use force data but other stacks may clear it */
15077 		if (rack->r_ctl.crte != NULL) {
15078 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15079 			rack->rack_hdrw_pacing = 0;
15080 			rack->r_ctl.crte = NULL;
15081 		}
15082 #ifdef TCP_BLACKBOX
15083 		tcp_log_flowend(tp);
15084 #endif
15085 		/*
15086 		 * Lets take a different approach to purging just
15087 		 * get each one and free it like a cum-ack would and
15088 		 * not use a foreach loop.
15089 		 */
15090 		rsm = tqhash_min(rack->r_ctl.tqh);
15091 		while (rsm) {
15092 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15093 			rack->r_ctl.rc_num_maps_alloced--;
15094 			uma_zfree(rack_zone, rsm);
15095 			rsm = tqhash_min(rack->r_ctl.tqh);
15096 		}
15097 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15098 		while (rsm) {
15099 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15100 			rack->r_ctl.rc_num_maps_alloced--;
15101 			rack->rc_free_cnt--;
15102 			cnt_free++;
15103 			uma_zfree(rack_zone, rsm);
15104 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15105 		}
15106 		if (rack->r_ctl.pcm_s != NULL) {
15107 			free(rack->r_ctl.pcm_s, M_TCPPCM);
15108 			rack->r_ctl.pcm_s = NULL;
15109 			rack->r_ctl.pcm_i.cnt_alloc = 0;
15110 			rack->r_ctl.pcm_i.cnt = 0;
15111 		}
15112 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15113 		    (tcp_bblogging_on(tp))) {
15114 			union tcp_log_stackspecific log;
15115 			struct timeval tv;
15116 
15117 			memset(&log, 0, sizeof(log));
15118 			log.u_bbr.flex8 = 10;
15119 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15120 			log.u_bbr.flex2 = rack->rc_free_cnt;
15121 			log.u_bbr.flex3 = cnt_free;
15122 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15123 			rsm = tqhash_min(rack->r_ctl.tqh);
15124 			log.u_bbr.delRate = (uintptr_t)rsm;
15125 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15126 			log.u_bbr.cur_del_rate = (uintptr_t)rsm;
15127 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15128 			log.u_bbr.pkt_epoch = __LINE__;
15129 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15130 					     0, &log, false, NULL, NULL, 0, &tv);
15131 		}
15132 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15133 			("rack:%p num_aloc:%u after freeing all?",
15134 			 rack,
15135 			 rack->r_ctl.rc_num_maps_alloced));
15136 		rack->rc_free_cnt = 0;
15137 		free(rack->r_ctl.tqh, M_TCPFSB);
15138 		rack->r_ctl.tqh = NULL;
15139 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15140 		tp->t_fb_ptr = NULL;
15141 	}
15142 	/* Make sure snd_nxt is correctly set */
15143 	tp->snd_nxt = tp->snd_max;
15144 }
15145 
15146 static void
15147 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15148 {
15149 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15150 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15151 	}
15152 	switch (tp->t_state) {
15153 	case TCPS_SYN_SENT:
15154 		rack->r_state = TCPS_SYN_SENT;
15155 		rack->r_substate = rack_do_syn_sent;
15156 		break;
15157 	case TCPS_SYN_RECEIVED:
15158 		rack->r_state = TCPS_SYN_RECEIVED;
15159 		rack->r_substate = rack_do_syn_recv;
15160 		break;
15161 	case TCPS_ESTABLISHED:
15162 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15163 		rack->r_state = TCPS_ESTABLISHED;
15164 		rack->r_substate = rack_do_established;
15165 		break;
15166 	case TCPS_CLOSE_WAIT:
15167 		rack->r_state = TCPS_CLOSE_WAIT;
15168 		rack->r_substate = rack_do_close_wait;
15169 		break;
15170 	case TCPS_FIN_WAIT_1:
15171 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15172 		rack->r_state = TCPS_FIN_WAIT_1;
15173 		rack->r_substate = rack_do_fin_wait_1;
15174 		break;
15175 	case TCPS_CLOSING:
15176 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15177 		rack->r_state = TCPS_CLOSING;
15178 		rack->r_substate = rack_do_closing;
15179 		break;
15180 	case TCPS_LAST_ACK:
15181 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15182 		rack->r_state = TCPS_LAST_ACK;
15183 		rack->r_substate = rack_do_lastack;
15184 		break;
15185 	case TCPS_FIN_WAIT_2:
15186 		rack->r_state = TCPS_FIN_WAIT_2;
15187 		rack->r_substate = rack_do_fin_wait_2;
15188 		break;
15189 	case TCPS_LISTEN:
15190 	case TCPS_CLOSED:
15191 	case TCPS_TIME_WAIT:
15192 	default:
15193 		break;
15194 	};
15195 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15196 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15197 
15198 }
15199 
15200 static void
15201 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15202 {
15203 	/*
15204 	 * We received an ack, and then did not
15205 	 * call send or were bounced out due to the
15206 	 * hpts was running. Now a timer is up as well, is
15207 	 * it the right timer?
15208 	 */
15209 	struct rack_sendmap *rsm;
15210 	int tmr_up;
15211 
15212 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15213 	if (tcp_in_hpts(rack->rc_tp) == 0) {
15214 		/*
15215 		 * Ok we probably need some timer up, but no
15216 		 * matter what the mask we are not in hpts. We
15217 		 * may have received an old ack and thus did nothing.
15218 		 */
15219 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15220 		rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15221 		return;
15222 	}
15223 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15224 		return;
15225 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15226 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15227 	    (tmr_up == PACE_TMR_RXT)) {
15228 		/* Should be an RXT */
15229 		return;
15230 	}
15231 	if (rsm == NULL) {
15232 		/* Nothing outstanding? */
15233 		if (tp->t_flags & TF_DELACK) {
15234 			if (tmr_up == PACE_TMR_DELACK)
15235 				/* We are supposed to have delayed ack up and we do */
15236 				return;
15237 		} else if (((V_tcp_always_keepalive ||
15238 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15239 			    (tp->t_state <= TCPS_CLOSING)) &&
15240 			   (tmr_up == PACE_TMR_KEEP) &&
15241 			   (tp->snd_max == tp->snd_una)) {
15242 			/* We should have keep alive up and we do */
15243 			return;
15244 		}
15245 	}
15246 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15247 		   ((tmr_up == PACE_TMR_TLP) ||
15248 		    (tmr_up == PACE_TMR_RACK) ||
15249 		    (tmr_up == PACE_TMR_RXT))) {
15250 		/*
15251 		 * Either a Rack, TLP or RXT is fine if  we
15252 		 * have outstanding data.
15253 		 */
15254 		return;
15255 	} else if (tmr_up == PACE_TMR_DELACK) {
15256 		/*
15257 		 * If the delayed ack was going to go off
15258 		 * before the rtx/tlp/rack timer were going to
15259 		 * expire, then that would be the timer in control.
15260 		 * Note we don't check the time here trusting the
15261 		 * code is correct.
15262 		 */
15263 		return;
15264 	}
15265 	/*
15266 	 * Ok the timer originally started is not what we want now.
15267 	 * We will force the hpts to be stopped if any, and restart
15268 	 * with the slot set to what was in the saved slot.
15269 	 */
15270 	if (tcp_in_hpts(rack->rc_tp)) {
15271 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15272 			uint32_t us_cts;
15273 
15274 			us_cts = tcp_get_usecs(NULL);
15275 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15276 				rack->r_early = 1;
15277 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15278 			}
15279 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15280 		}
15281 		tcp_hpts_remove(rack->rc_tp);
15282 	}
15283 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15284 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15285 }
15286 
15287 
15288 static void
15289 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15290 {
15291 	if ((SEQ_LT(tp->snd_wl1, seq) ||
15292 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15293 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15294 		/* keep track of pure window updates */
15295 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15296 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15297 		tp->snd_wnd = tiwin;
15298 		rack_validate_fo_sendwin_up(tp, rack);
15299 		tp->snd_wl1 = seq;
15300 		tp->snd_wl2 = ack;
15301 		if (tp->snd_wnd > tp->max_sndwnd)
15302 			tp->max_sndwnd = tp->snd_wnd;
15303 	    rack->r_wanted_output = 1;
15304 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15305 		tp->snd_wnd = tiwin;
15306 		rack_validate_fo_sendwin_up(tp, rack);
15307 		tp->snd_wl1 = seq;
15308 		tp->snd_wl2 = ack;
15309 	} else {
15310 		/* Not a valid win update */
15311 		return;
15312 	}
15313 	if (tp->snd_wnd > tp->max_sndwnd)
15314 		tp->max_sndwnd = tp->snd_wnd;
15315 	/* Do we exit persists? */
15316 	if ((rack->rc_in_persist != 0) &&
15317 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15318 				rack->r_ctl.rc_pace_min_segs))) {
15319 		rack_exit_persist(tp, rack, cts);
15320 	}
15321 	/* Do we enter persists? */
15322 	if ((rack->rc_in_persist == 0) &&
15323 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15324 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
15325 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15326 	    sbavail(&tptosocket(tp)->so_snd) &&
15327 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15328 		/*
15329 		 * Here the rwnd is less than
15330 		 * the pacing size, we are established,
15331 		 * nothing is outstanding, and there is
15332 		 * data to send. Enter persists.
15333 		 */
15334 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15335 	}
15336 }
15337 
15338 static void
15339 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15340 {
15341 
15342 	if (tcp_bblogging_on(rack->rc_tp)) {
15343 		struct inpcb *inp = tptoinpcb(tp);
15344 		union tcp_log_stackspecific log;
15345 		struct timeval ltv;
15346 		char tcp_hdr_buf[60];
15347 		struct tcphdr *th;
15348 		struct timespec ts;
15349 		uint32_t orig_snd_una;
15350 		uint8_t xx = 0;
15351 
15352 #ifdef TCP_REQUEST_TRK
15353 		struct tcp_sendfile_track *tcp_req;
15354 
15355 		if (SEQ_GT(ae->ack, tp->snd_una)) {
15356 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15357 		} else {
15358 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15359 		}
15360 #endif
15361 		memset(&log, 0, sizeof(log));
15362 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15363 		if (rack->rack_no_prr == 0)
15364 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15365 		else
15366 			log.u_bbr.flex1 = 0;
15367 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15368 		log.u_bbr.use_lt_bw <<= 1;
15369 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
15370 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15371 		log.u_bbr.bbr_state = rack->rc_free_cnt;
15372 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15373 		log.u_bbr.pkts_out = tp->t_maxseg;
15374 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15375 		log.u_bbr.flex7 = 1;
15376 		log.u_bbr.lost = ae->flags;
15377 		log.u_bbr.cwnd_gain = ackval;
15378 		log.u_bbr.pacing_gain = 0x2;
15379 		if (ae->flags & TSTMP_HDWR) {
15380 			/* Record the hardware timestamp if present */
15381 			log.u_bbr.flex3 = M_TSTMP;
15382 			ts.tv_sec = ae->timestamp / 1000000000;
15383 			ts.tv_nsec = ae->timestamp % 1000000000;
15384 			ltv.tv_sec = ts.tv_sec;
15385 			ltv.tv_usec = ts.tv_nsec / 1000;
15386 			log.u_bbr.lt_epoch = tcp_tv_to_usec(&ltv);
15387 		} else if (ae->flags & TSTMP_LRO) {
15388 			/* Record the LRO the arrival timestamp */
15389 			log.u_bbr.flex3 = M_TSTMP_LRO;
15390 			ts.tv_sec = ae->timestamp / 1000000000;
15391 			ts.tv_nsec = ae->timestamp % 1000000000;
15392 			ltv.tv_sec = ts.tv_sec;
15393 			ltv.tv_usec = ts.tv_nsec / 1000;
15394 			log.u_bbr.flex5 = tcp_tv_to_usec(&ltv);
15395 		}
15396 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15397 		/* Log the rcv time */
15398 		log.u_bbr.delRate = ae->timestamp;
15399 #ifdef TCP_REQUEST_TRK
15400 		log.u_bbr.applimited = tp->t_tcpreq_closed;
15401 		log.u_bbr.applimited <<= 8;
15402 		log.u_bbr.applimited |= tp->t_tcpreq_open;
15403 		log.u_bbr.applimited <<= 8;
15404 		log.u_bbr.applimited |= tp->t_tcpreq_req;
15405 		if (tcp_req) {
15406 			/* Copy out any client req info */
15407 			/* seconds */
15408 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15409 			/* useconds */
15410 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15411 			log.u_bbr.rttProp = tcp_req->timestamp;
15412 			log.u_bbr.cur_del_rate = tcp_req->start;
15413 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15414 				log.u_bbr.flex8 |= 1;
15415 			} else {
15416 				log.u_bbr.flex8 |= 2;
15417 				log.u_bbr.bw_inuse = tcp_req->end;
15418 			}
15419 			log.u_bbr.flex6 = tcp_req->start_seq;
15420 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15421 				log.u_bbr.flex8 |= 4;
15422 				log.u_bbr.epoch = tcp_req->end_seq;
15423 			}
15424 		}
15425 #endif
15426 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15427 		th = (struct tcphdr *)tcp_hdr_buf;
15428 		th->th_seq = ae->seq;
15429 		th->th_ack = ae->ack;
15430 		th->th_win = ae->win;
15431 		/* Now fill in the ports */
15432 		th->th_sport = inp->inp_fport;
15433 		th->th_dport = inp->inp_lport;
15434 		tcp_set_flags(th, ae->flags);
15435 		/* Now do we have a timestamp option? */
15436 		if (ae->flags & HAS_TSTMP) {
15437 			u_char *cp;
15438 			uint32_t val;
15439 
15440 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15441 			cp = (u_char *)(th + 1);
15442 			*cp = TCPOPT_NOP;
15443 			cp++;
15444 			*cp = TCPOPT_NOP;
15445 			cp++;
15446 			*cp = TCPOPT_TIMESTAMP;
15447 			cp++;
15448 			*cp = TCPOLEN_TIMESTAMP;
15449 			cp++;
15450 			val = htonl(ae->ts_value);
15451 			bcopy((char *)&val,
15452 			      (char *)cp, sizeof(uint32_t));
15453 			val = htonl(ae->ts_echo);
15454 			bcopy((char *)&val,
15455 			      (char *)(cp + 4), sizeof(uint32_t));
15456 		} else
15457 			th->th_off = (sizeof(struct tcphdr) >> 2);
15458 
15459 		/*
15460 		 * For sane logging we need to play a little trick.
15461 		 * If the ack were fully processed we would have moved
15462 		 * snd_una to high_seq, but since compressed acks are
15463 		 * processed in two phases, at this point (logging) snd_una
15464 		 * won't be advanced. So we would see multiple acks showing
15465 		 * the advancement. We can prevent that by "pretending" that
15466 		 * snd_una was advanced and then un-advancing it so that the
15467 		 * logging code has the right value for tlb_snd_una.
15468 		 */
15469 		if (tp->snd_una != high_seq) {
15470 			orig_snd_una = tp->snd_una;
15471 			tp->snd_una = high_seq;
15472 			xx = 1;
15473 		} else
15474 			xx = 0;
15475 		TCP_LOG_EVENTP(tp, th,
15476 			       &tptosocket(tp)->so_rcv,
15477 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15478 			       0, &log, true, &ltv);
15479 		if (xx) {
15480 			tp->snd_una = orig_snd_una;
15481 		}
15482 	}
15483 
15484 }
15485 
15486 static void
15487 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15488 {
15489 	uint32_t us_rtt;
15490 	/*
15491 	 * A persist or keep-alive was forced out, update our
15492 	 * min rtt time. Note now worry about lost responses.
15493 	 * When a subsequent keep-alive or persist times out
15494 	 * and forced_ack is still on, then the last probe
15495 	 * was not responded to. In such cases we have a
15496 	 * sysctl that controls the behavior. Either we apply
15497 	 * the rtt but with reduced confidence (0). Or we just
15498 	 * plain don't apply the rtt estimate. Having data flow
15499 	 * will clear the probe_not_answered flag i.e. cum-ack
15500 	 * move forward <or> exiting and reentering persists.
15501 	 */
15502 
15503 	rack->forced_ack = 0;
15504 	rack->rc_tp->t_rxtshift = 0;
15505 	if ((rack->rc_in_persist &&
15506 	     (tiwin == rack->rc_tp->snd_wnd)) ||
15507 	    (rack->rc_in_persist == 0)) {
15508 		/*
15509 		 * In persists only apply the RTT update if this is
15510 		 * a response to our window probe. And that
15511 		 * means the rwnd sent must match the current
15512 		 * snd_wnd. If it does not, then we got a
15513 		 * window update ack instead. For keepalive
15514 		 * we allow the answer no matter what the window.
15515 		 *
15516 		 * Note that if the probe_not_answered is set then
15517 		 * the forced_ack_ts is the oldest one i.e. the first
15518 		 * probe sent that might have been lost. This assures
15519 		 * us that if we do calculate an RTT it is longer not
15520 		 * some short thing.
15521 		 */
15522 		if (rack->rc_in_persist)
15523 			counter_u64_add(rack_persists_acks, 1);
15524 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15525 		if (us_rtt == 0)
15526 			us_rtt = 1;
15527 		if (rack->probe_not_answered == 0) {
15528 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15529 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15530 		} else {
15531 			/* We have a retransmitted probe here too */
15532 			if (rack_apply_rtt_with_reduced_conf) {
15533 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15534 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15535 			}
15536 		}
15537 	}
15538 }
15539 
15540 static void
15541 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15542 {
15543 	/*
15544 	 * The next send has occurred mark the end of the round
15545 	 * as when that data gets acknowledged. We can
15546 	 * also do common things we might need to do when
15547 	 * a round begins.
15548 	 */
15549 	rack->r_ctl.roundends = tp->snd_max;
15550 	rack->rc_new_rnd_needed = 0;
15551 	rack_log_hystart_event(rack, tp->snd_max, 4);
15552 }
15553 
15554 
15555 static void
15556 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
15557 	     uint32_t flex3)
15558 {
15559 	if (tcp_bblogging_on(rack->rc_tp)) {
15560 		union tcp_log_stackspecific log;
15561 		struct timeval tv;
15562 
15563 		(void)tcp_get_usecs(&tv);
15564 		memset(&log, 0, sizeof(log));
15565 		log.u_bbr.timeStamp = tcp_tv_to_usec(&tv);
15566 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15567 		log.u_bbr.flex8 = mod;
15568 		log.u_bbr.flex1 = flex1;
15569 		log.u_bbr.flex2 = flex2;
15570 		log.u_bbr.flex3 = flex3;
15571 		log.u_bbr.flex4 = rack_pcm_every_n_rounds;
15572 		log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
15573 		log.u_bbr.bbr_substate = rack->pcm_needed;
15574 		log.u_bbr.bbr_substate <<= 1;
15575 		log.u_bbr.bbr_substate |= rack->pcm_in_progress;
15576 		log.u_bbr.bbr_substate <<= 1;
15577 		log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
15578 		(void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
15579 				    0, &log, false, NULL, NULL, 0, &tv);
15580 	}
15581 }
15582 
15583 static void
15584 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15585 {
15586 	/*
15587 	 * The round (current_round) has ended. We now
15588 	 * setup for the next round by incrementing the
15589 	 * round numnber and doing any round specific
15590 	 * things.
15591 	 */
15592 	rack_log_hystart_event(rack, high_seq, 21);
15593 	rack->r_ctl.current_round++;
15594 	/* New round (current_round) begins at next send */
15595 	rack->rc_new_rnd_needed = 1;
15596 	if ((rack->pcm_enabled == 1) &&
15597 	    (rack->pcm_needed == 0) &&
15598 	    (rack->pcm_in_progress == 0)) {
15599 		/*
15600 		 * If we have enabled PCM, then we need to
15601 		 * check if the round has adanced to the state
15602 		 * where one is required.
15603 		 */
15604 		int rnds;
15605 
15606 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
15607 		if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
15608 			rack->pcm_needed = 1;
15609 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15610 		} else if (rack_verbose_logging) {
15611 			rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15612 		}
15613 	}
15614 	if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
15615 		/* We have hystart enabled send the round info in */
15616 		if (CC_ALGO(tp)->newround != NULL) {
15617 			CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
15618 		}
15619 	}
15620 	/*
15621 	 * For DGP an initial startup check. We want to validate
15622 	 * that we are not just pushing on slow-start and just
15623 	 * not gaining.. i.e. filling buffers without getting any
15624 	 * boost in b/w during the inital slow-start.
15625 	 */
15626 	if (rack->dgp_on &&
15627 	    (rack->rc_initial_ss_comp == 0) &&
15628 	    (tp->snd_cwnd < tp->snd_ssthresh) &&
15629 	    (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
15630 	    (rack->r_ctl.gp_rnd_thresh > 0) &&
15631 	    ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
15632 
15633 		/*
15634 		 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
15635 		 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
15636 		 * exit SS.
15637 		 *
15638 		 * Pick up the flight size now as we enter slowstart (not the
15639 		 * cwnd which may be inflated).
15640 		 */
15641 		rack->rc_initial_ss_comp = 1;
15642 
15643 		if (tcp_bblogging_on(rack->rc_tp)) {
15644 			union tcp_log_stackspecific log;
15645 			struct timeval tv;
15646 
15647 			memset(&log, 0, sizeof(log));
15648 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15649 			log.u_bbr.flex1 = rack->r_ctl.current_round;
15650 			log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
15651 			log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
15652 			log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
15653 			log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
15654 			log.u_bbr.flex8 = 40;
15655 			(void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
15656 					    0, &log, false, NULL, __func__, __LINE__,&tv);
15657 		}
15658 		if ((rack->r_ctl.gate_to_fs == 1) &&
15659 		     (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
15660 			tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
15661 		}
15662 		tp->snd_ssthresh = tp->snd_cwnd - 1;
15663 		/* Turn off any fast output running */
15664 		rack->r_fast_output = 0;
15665 	}
15666 }
15667 
15668 static int
15669 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15670 {
15671 	/*
15672 	 * Handle a "special" compressed ack mbuf. Each incoming
15673 	 * ack has only four possible dispositions:
15674 	 *
15675 	 * A) It moves the cum-ack forward
15676 	 * B) It is behind the cum-ack.
15677 	 * C) It is a window-update ack.
15678 	 * D) It is a dup-ack.
15679 	 *
15680 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15681 	 * in the incoming mbuf. We also need to still pay attention
15682 	 * to nxt_pkt since there may be another packet after this
15683 	 * one.
15684 	 */
15685 #ifdef TCP_ACCOUNTING
15686 	uint64_t ts_val;
15687 	uint64_t rdstc;
15688 #endif
15689 	int segsiz;
15690 	struct timespec ts;
15691 	struct tcp_rack *rack;
15692 	struct tcp_ackent *ae;
15693 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15694 	int cnt, i, did_out, ourfinisacked = 0;
15695 	struct tcpopt to_holder, *to = NULL;
15696 #ifdef TCP_ACCOUNTING
15697 	int win_up_req = 0;
15698 #endif
15699 	int nsegs = 0;
15700 	int under_pacing = 0;
15701 	int post_recovery = 0;
15702 #ifdef TCP_ACCOUNTING
15703 	sched_pin();
15704 #endif
15705 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15706 	if (rack->gp_ready &&
15707 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15708 		under_pacing = 1;
15709 
15710 	if (rack->r_state != tp->t_state)
15711 		rack_set_state(tp, rack);
15712 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15713 	    (tp->t_flags & TF_GPUTINPROG)) {
15714 		/*
15715 		 * We have a goodput in progress
15716 		 * and we have entered a late state.
15717 		 * Do we have enough data in the sb
15718 		 * to handle the GPUT request?
15719 		 */
15720 		uint32_t bytes;
15721 
15722 		bytes = tp->gput_ack - tp->gput_seq;
15723 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
15724 			bytes += tp->gput_seq - tp->snd_una;
15725 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15726 			/*
15727 			 * There are not enough bytes in the socket
15728 			 * buffer that have been sent to cover this
15729 			 * measurement. Cancel it.
15730 			 */
15731 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15732 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
15733 						   tp->gput_seq,
15734 						   0, 0, 18, __LINE__, NULL, 0);
15735 			tp->t_flags &= ~TF_GPUTINPROG;
15736 		}
15737 	}
15738 	to = &to_holder;
15739 	to->to_flags = 0;
15740 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15741 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15742 	cnt = m->m_len / sizeof(struct tcp_ackent);
15743 	counter_u64_add(rack_multi_single_eq, cnt);
15744 	high_seq = tp->snd_una;
15745 	the_win = tp->snd_wnd;
15746 	win_seq = tp->snd_wl1;
15747 	win_upd_ack = tp->snd_wl2;
15748 	cts = tcp_tv_to_usec(tv);
15749 	ms_cts = tcp_tv_to_msec(tv);
15750 	rack->r_ctl.rc_rcvtime = cts;
15751 	segsiz = ctf_fixed_maxseg(tp);
15752 	if ((rack->rc_gp_dyn_mul) &&
15753 	    (rack->use_fixed_rate == 0) &&
15754 	    (rack->rc_always_pace)) {
15755 		/* Check in on probertt */
15756 		rack_check_probe_rtt(rack, cts);
15757 	}
15758 	for (i = 0; i < cnt; i++) {
15759 #ifdef TCP_ACCOUNTING
15760 		ts_val = get_cyclecount();
15761 #endif
15762 		rack_clear_rate_sample(rack);
15763 		ae = ((mtod(m, struct tcp_ackent *)) + i);
15764 		if (ae->flags & TH_FIN)
15765 			rack_log_pacing_delay_calc(rack,
15766 						   0,
15767 						   0,
15768 						   0,
15769 						   rack_get_gp_est(rack), /* delRate */
15770 						   rack_get_lt_bw(rack), /* rttProp */
15771 						   20, __LINE__, NULL, 0);
15772 		/* Setup the window */
15773 		tiwin = ae->win << tp->snd_scale;
15774 		if (tiwin > rack->r_ctl.rc_high_rwnd)
15775 			rack->r_ctl.rc_high_rwnd = tiwin;
15776 		/* figure out the type of ack */
15777 		if (SEQ_LT(ae->ack, high_seq)) {
15778 			/* Case B*/
15779 			ae->ack_val_set = ACK_BEHIND;
15780 		} else if (SEQ_GT(ae->ack, high_seq)) {
15781 			/* Case A */
15782 			ae->ack_val_set = ACK_CUMACK;
15783 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15784 			/* Case D */
15785 			ae->ack_val_set = ACK_DUPACK;
15786 		} else {
15787 			/* Case C */
15788 			ae->ack_val_set = ACK_RWND;
15789 		}
15790 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15791 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15792 		/* Validate timestamp */
15793 		if (ae->flags & HAS_TSTMP) {
15794 			/* Setup for a timestamp */
15795 			to->to_flags = TOF_TS;
15796 			ae->ts_echo -= tp->ts_offset;
15797 			to->to_tsecr = ae->ts_echo;
15798 			to->to_tsval = ae->ts_value;
15799 			/*
15800 			 * If echoed timestamp is later than the current time, fall back to
15801 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
15802 			 * were used when this connection was established.
15803 			 */
15804 			if (TSTMP_GT(ae->ts_echo, ms_cts))
15805 				to->to_tsecr = 0;
15806 			if (tp->ts_recent &&
15807 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15808 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15809 #ifdef TCP_ACCOUNTING
15810 					rdstc = get_cyclecount();
15811 					if (rdstc > ts_val) {
15812 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15813 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15814 						}
15815 					}
15816 #endif
15817 					continue;
15818 				}
15819 			}
15820 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15821 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15822 				tp->ts_recent_age = tcp_ts_getticks();
15823 				tp->ts_recent = ae->ts_value;
15824 			}
15825 		} else {
15826 			/* Setup for a no options */
15827 			to->to_flags = 0;
15828 		}
15829 		/* Update the rcv time and perform idle reduction possibly */
15830 		if  (tp->t_idle_reduce &&
15831 		     (tp->snd_max == tp->snd_una) &&
15832 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15833 			counter_u64_add(rack_input_idle_reduces, 1);
15834 			rack_cc_after_idle(rack, tp);
15835 		}
15836 		tp->t_rcvtime = ticks;
15837 		/* Now what about ECN of a chain of pure ACKs? */
15838 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
15839 			tcp_packets_this_ack(tp, ae->ack),
15840 			ae->codepoint))
15841 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15842 #ifdef TCP_ACCOUNTING
15843 		/* Count for the specific type of ack in */
15844 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15845 			tp->tcp_cnt_counters[ae->ack_val_set]++;
15846 		}
15847 #endif
15848 		/*
15849 		 * Note how we could move up these in the determination
15850 		 * above, but we don't so that way the timestamp checks (and ECN)
15851 		 * is done first before we do any processing on the ACK.
15852 		 * The non-compressed path through the code has this
15853 		 * weakness (noted by @jtl) that it actually does some
15854 		 * processing before verifying the timestamp information.
15855 		 * We don't take that path here which is why we set
15856 		 * the ack_val_set first, do the timestamp and ecn
15857 		 * processing, and then look at what we have setup.
15858 		 */
15859 		if (ae->ack_val_set == ACK_BEHIND) {
15860 			/*
15861 			 * Case B flag reordering, if window is not closed
15862 			 * or it could be a keep-alive or persists
15863 			 */
15864 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15865 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
15866 				if (rack->r_ctl.rc_reorder_ts == 0)
15867 					rack->r_ctl.rc_reorder_ts = 1;
15868 			}
15869 		} else if (ae->ack_val_set == ACK_DUPACK) {
15870 			/* Case D */
15871 			rack_strike_dupack(rack, ae->ack);
15872 		} else if (ae->ack_val_set == ACK_RWND) {
15873 			/* Case C */
15874 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15875 				ts.tv_sec = ae->timestamp / 1000000000;
15876 				ts.tv_nsec = ae->timestamp % 1000000000;
15877 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15878 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15879 			} else {
15880 				rack->r_ctl.act_rcv_time = *tv;
15881 			}
15882 			if (rack->forced_ack) {
15883 				rack_handle_probe_response(rack, tiwin,
15884 							   tcp_tv_to_usec(&rack->r_ctl.act_rcv_time));
15885 			}
15886 #ifdef TCP_ACCOUNTING
15887 			win_up_req = 1;
15888 #endif
15889 			win_upd_ack = ae->ack;
15890 			win_seq = ae->seq;
15891 			the_win = tiwin;
15892 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15893 		} else {
15894 			/* Case A */
15895 			if (SEQ_GT(ae->ack, tp->snd_max)) {
15896 				/*
15897 				 * We just send an ack since the incoming
15898 				 * ack is beyond the largest seq we sent.
15899 				 */
15900 				if ((tp->t_flags & TF_ACKNOW) == 0) {
15901 					ctf_ack_war_checks(tp);
15902 					if (tp->t_flags && TF_ACKNOW)
15903 						rack->r_wanted_output = 1;
15904 				}
15905 			} else {
15906 				nsegs++;
15907 				/* If the window changed setup to update */
15908 				if (tiwin != tp->snd_wnd) {
15909 					win_upd_ack = ae->ack;
15910 					win_seq = ae->seq;
15911 					the_win = tiwin;
15912 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15913 				}
15914 #ifdef TCP_ACCOUNTING
15915 				/* Account for the acks */
15916 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15917 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
15918 				}
15919 #endif
15920 				high_seq = ae->ack;
15921 				/* Setup our act_rcv_time */
15922 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15923 					ts.tv_sec = ae->timestamp / 1000000000;
15924 					ts.tv_nsec = ae->timestamp % 1000000000;
15925 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15926 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15927 				} else {
15928 					rack->r_ctl.act_rcv_time = *tv;
15929 				}
15930 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
15931 						       tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time));
15932 #ifdef TCP_REQUEST_TRK
15933 				rack_req_check_for_comp(rack, high_seq);
15934 #endif
15935 				if (rack->rc_dsack_round_seen) {
15936 					/* Is the dsack round over? */
15937 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
15938 						/* Yes it is */
15939 						rack->rc_dsack_round_seen = 0;
15940 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
15941 					}
15942 				}
15943 			}
15944 		}
15945 		/* And lets be sure to commit the rtt measurements for this ack */
15946 		tcp_rack_xmit_timer_commit(rack, tp);
15947 #ifdef TCP_ACCOUNTING
15948 		rdstc = get_cyclecount();
15949 		if (rdstc > ts_val) {
15950 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15951 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15952 				if (ae->ack_val_set == ACK_CUMACK)
15953 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
15954 			}
15955 		}
15956 #endif
15957 	}
15958 #ifdef TCP_ACCOUNTING
15959 	ts_val = get_cyclecount();
15960 #endif
15961 	/* Tend to any collapsed window */
15962 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
15963 		/* The peer collapsed the window */
15964 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
15965 	} else if (rack->rc_has_collapsed)
15966 		rack_un_collapse_window(rack, __LINE__);
15967 	if ((rack->r_collapse_point_valid) &&
15968 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
15969 		rack->r_collapse_point_valid = 0;
15970 	acked_amount = acked = (high_seq - tp->snd_una);
15971 	if (acked) {
15972 		/*
15973 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
15974 		 * causes issues when we are just going app limited. Lets
15975 		 * instead use SEQ_GT <or> where its equal but more data
15976 		 * is outstanding.
15977 		 *
15978 		 * Also make sure we are on the last ack of a series. We
15979 		 * have to have all the ack's processed in queue to know
15980 		 * if there is something left outstanding.
15981 		 *
15982 		 */
15983 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
15984 		    (rack->rc_new_rnd_needed == 0) &&
15985 		    (nxt_pkt == 0)) {
15986 			/*
15987 			 * We have crossed into a new round with
15988 			 * this th_ack value.
15989 			 */
15990 			rack_new_round_setup(tp, rack, high_seq);
15991 		}
15992 		/*
15993 		 * Clear the probe not answered flag
15994 		 * since cum-ack moved forward.
15995 		 */
15996 		rack->probe_not_answered = 0;
15997 		if (tp->t_flags & TF_NEEDSYN) {
15998 			/*
15999 			 * T/TCP: Connection was half-synchronized, and our SYN has
16000 			 * been ACK'd (so connection is now fully synchronized).  Go
16001 			 * to non-starred state, increment snd_una for ACK of SYN,
16002 			 * and check if we can do window scaling.
16003 			 */
16004 			tp->t_flags &= ~TF_NEEDSYN;
16005 			tp->snd_una++;
16006 			acked_amount = acked = (high_seq - tp->snd_una);
16007 		}
16008 		if (acked > sbavail(&so->so_snd))
16009 			acked_amount = sbavail(&so->so_snd);
16010 		if (IN_FASTRECOVERY(tp->t_flags) &&
16011 		    (rack->rack_no_prr == 0))
16012 			rack_update_prr(tp, rack, acked_amount, high_seq);
16013 		if (IN_RECOVERY(tp->t_flags)) {
16014 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16015 			    (SEQ_LT(high_seq, tp->snd_max))) {
16016 				tcp_rack_partialack(tp);
16017 			} else {
16018 				rack_post_recovery(tp, high_seq);
16019 				post_recovery = 1;
16020 			}
16021 		}  else if ((rack->rto_from_rec == 1) &&
16022 			    SEQ_GEQ(high_seq, tp->snd_recover)) {
16023 			/*
16024 			 * We were in recovery, hit a rxt timeout
16025 			 * and never re-entered recovery. The timeout(s)
16026 			 * made up all the lost data. In such a case
16027 			 * we need to clear the rto_from_rec flag.
16028 			 */
16029 			rack->rto_from_rec = 0;
16030 		}
16031 		/* Handle the rack-log-ack part (sendmap) */
16032 		if ((sbused(&so->so_snd) == 0) &&
16033 		    (acked > acked_amount) &&
16034 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16035 		    (tp->t_flags & TF_SENTFIN)) {
16036 			/*
16037 			 * We must be sure our fin
16038 			 * was sent and acked (we can be
16039 			 * in FIN_WAIT_1 without having
16040 			 * sent the fin).
16041 			 */
16042 			ourfinisacked = 1;
16043 			/*
16044 			 * Lets make sure snd_una is updated
16045 			 * since most likely acked_amount = 0 (it
16046 			 * should be).
16047 			 */
16048 			tp->snd_una = high_seq;
16049 		}
16050 		/* Did we make a RTO error? */
16051 		if ((tp->t_flags & TF_PREVVALID) &&
16052 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16053 			tp->t_flags &= ~TF_PREVVALID;
16054 			if (tp->t_rxtshift == 1 &&
16055 			    (int)(ticks - tp->t_badrxtwin) < 0)
16056 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16057 		}
16058 		/* Handle the data in the socket buffer */
16059 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16060 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16061 		if (acked_amount > 0) {
16062 			uint32_t p_cwnd;
16063 			struct mbuf *mfree;
16064 
16065 			if (post_recovery) {
16066 				/*
16067 				 * Grab the segsiz, multiply by 2 and add the snd_cwnd
16068 				 * that is the max the CC should add if we are exiting
16069 				 * recovery and doing a late add.
16070 				 */
16071 				p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16072 				p_cwnd <<= 1;
16073 				p_cwnd += tp->snd_cwnd;
16074 			}
16075 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
16076 			if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
16077 				/* Must be non-newreno (cubic) getting too ahead of itself */
16078 				tp->snd_cwnd = p_cwnd;
16079 			}
16080 			SOCK_SENDBUF_LOCK(so);
16081 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16082 			tp->snd_una = high_seq;
16083 			/* Note we want to hold the sb lock through the sendmap adjust */
16084 			rack_adjust_sendmap_head(rack, &so->so_snd);
16085 			/* Wake up the socket if we have room to write more */
16086 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16087 			sowwakeup_locked(so);
16088 			m_freem(mfree);
16089 		}
16090 		/* update progress */
16091 		tp->t_acktime = ticks;
16092 		rack_log_progress_event(rack, tp, tp->t_acktime,
16093 					PROGRESS_UPDATE, __LINE__);
16094 		/* Clear out shifts and such */
16095 		tp->t_rxtshift = 0;
16096 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16097 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16098 		rack->rc_tlp_in_progress = 0;
16099 		rack->r_ctl.rc_tlp_cnt_out = 0;
16100 		/* Send recover and snd_nxt must be dragged along */
16101 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16102 			tp->snd_recover = tp->snd_una;
16103 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16104 			tp->snd_nxt = tp->snd_max;
16105 		/*
16106 		 * If the RXT timer is running we want to
16107 		 * stop it, so we can restart a TLP (or new RXT).
16108 		 */
16109 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16110 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16111 		tp->snd_wl2 = high_seq;
16112 		tp->t_dupacks = 0;
16113 		if (under_pacing &&
16114 		    (rack->use_fixed_rate == 0) &&
16115 		    (rack->in_probe_rtt == 0) &&
16116 		    rack->rc_gp_dyn_mul &&
16117 		    rack->rc_always_pace) {
16118 			/* Check if we are dragging bottom */
16119 			rack_check_bottom_drag(tp, rack, so);
16120 		}
16121 		if (tp->snd_una == tp->snd_max) {
16122 			tp->t_flags &= ~TF_PREVVALID;
16123 			rack->r_ctl.retran_during_recovery = 0;
16124 			rack->rc_suspicious = 0;
16125 			rack->r_ctl.dsack_byte_cnt = 0;
16126 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16127 			if (rack->r_ctl.rc_went_idle_time == 0)
16128 				rack->r_ctl.rc_went_idle_time = 1;
16129 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16130 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16131 				tp->t_acktime = 0;
16132 			/* Set so we might enter persists... */
16133 			rack->r_wanted_output = 1;
16134 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16135 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16136 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16137 			    (sbavail(&so->so_snd) == 0) &&
16138 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16139 				/*
16140 				 * The socket was gone and the
16141 				 * peer sent data (not now in the past), time to
16142 				 * reset him.
16143 				 */
16144 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16145 				/* tcp_close will kill the inp pre-log the Reset */
16146 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16147 #ifdef TCP_ACCOUNTING
16148 				rdstc = get_cyclecount();
16149 				if (rdstc > ts_val) {
16150 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16151 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16152 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16153 					}
16154 				}
16155 #endif
16156 				m_freem(m);
16157 				tp = tcp_close(tp);
16158 				if (tp == NULL) {
16159 #ifdef TCP_ACCOUNTING
16160 					sched_unpin();
16161 #endif
16162 					return (1);
16163 				}
16164 				/*
16165 				 * We would normally do drop-with-reset which would
16166 				 * send back a reset. We can't since we don't have
16167 				 * all the needed bits. Instead lets arrange for
16168 				 * a call to tcp_output(). That way since we
16169 				 * are in the closed state we will generate a reset.
16170 				 *
16171 				 * Note if tcp_accounting is on we don't unpin since
16172 				 * we do that after the goto label.
16173 				 */
16174 				goto send_out_a_rst;
16175 			}
16176 			if ((sbused(&so->so_snd) == 0) &&
16177 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16178 			    (tp->t_flags & TF_SENTFIN)) {
16179 				/*
16180 				 * If we can't receive any more data, then closing user can
16181 				 * proceed. Starting the timer is contrary to the
16182 				 * specification, but if we don't get a FIN we'll hang
16183 				 * forever.
16184 				 *
16185 				 */
16186 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16187 					soisdisconnected(so);
16188 					tcp_timer_activate(tp, TT_2MSL,
16189 							   (tcp_fast_finwait2_recycle ?
16190 							    tcp_finwait2_timeout :
16191 							    TP_MAXIDLE(tp)));
16192 				}
16193 				if (ourfinisacked == 0) {
16194 					/*
16195 					 * We don't change to fin-wait-2 if we have our fin acked
16196 					 * which means we are probably in TCPS_CLOSING.
16197 					 */
16198 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16199 				}
16200 			}
16201 		}
16202 		/* Wake up the socket if we have room to write more */
16203 		if (sbavail(&so->so_snd)) {
16204 			rack->r_wanted_output = 1;
16205 			if (ctf_progress_timeout_check(tp, true)) {
16206 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16207 							tp, tick, PROGRESS_DROP, __LINE__);
16208 				/*
16209 				 * We cheat here and don't send a RST, we should send one
16210 				 * when the pacer drops the connection.
16211 				 */
16212 #ifdef TCP_ACCOUNTING
16213 				rdstc = get_cyclecount();
16214 				if (rdstc > ts_val) {
16215 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16216 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16217 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16218 					}
16219 				}
16220 				sched_unpin();
16221 #endif
16222 				(void)tcp_drop(tp, ETIMEDOUT);
16223 				m_freem(m);
16224 				return (1);
16225 			}
16226 		}
16227 		if (ourfinisacked) {
16228 			switch(tp->t_state) {
16229 			case TCPS_CLOSING:
16230 #ifdef TCP_ACCOUNTING
16231 				rdstc = get_cyclecount();
16232 				if (rdstc > ts_val) {
16233 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16234 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16235 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16236 					}
16237 				}
16238 				sched_unpin();
16239 #endif
16240 				tcp_twstart(tp);
16241 				m_freem(m);
16242 				return (1);
16243 				break;
16244 			case TCPS_LAST_ACK:
16245 #ifdef TCP_ACCOUNTING
16246 				rdstc = get_cyclecount();
16247 				if (rdstc > ts_val) {
16248 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16249 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16250 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16251 					}
16252 				}
16253 				sched_unpin();
16254 #endif
16255 				tp = tcp_close(tp);
16256 				ctf_do_drop(m, tp);
16257 				return (1);
16258 				break;
16259 			case TCPS_FIN_WAIT_1:
16260 #ifdef TCP_ACCOUNTING
16261 				rdstc = get_cyclecount();
16262 				if (rdstc > ts_val) {
16263 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16264 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16265 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16266 					}
16267 				}
16268 #endif
16269 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16270 					soisdisconnected(so);
16271 					tcp_timer_activate(tp, TT_2MSL,
16272 							   (tcp_fast_finwait2_recycle ?
16273 							    tcp_finwait2_timeout :
16274 							    TP_MAXIDLE(tp)));
16275 				}
16276 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
16277 				break;
16278 			default:
16279 				break;
16280 			}
16281 		}
16282 		if (rack->r_fast_output) {
16283 			/*
16284 			 * We re doing fast output.. can we expand that?
16285 			 */
16286 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16287 		}
16288 #ifdef TCP_ACCOUNTING
16289 		rdstc = get_cyclecount();
16290 		if (rdstc > ts_val) {
16291 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16292 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16293 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16294 			}
16295 		}
16296 
16297 	} else if (win_up_req) {
16298 		rdstc = get_cyclecount();
16299 		if (rdstc > ts_val) {
16300 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16301 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16302 			}
16303 		}
16304 #endif
16305 	}
16306 	/* Now is there a next packet, if so we are done */
16307 	m_freem(m);
16308 	did_out = 0;
16309 	if (nxt_pkt) {
16310 #ifdef TCP_ACCOUNTING
16311 		sched_unpin();
16312 #endif
16313 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16314 		return (0);
16315 	}
16316 	rack_handle_might_revert(tp, rack);
16317 	ctf_calc_rwin(so, tp);
16318 	if ((rack->r_wanted_output != 0) ||
16319 	    (rack->r_fast_output != 0) ||
16320 	    (tp->t_flags & TF_ACKNOW )) {
16321 	send_out_a_rst:
16322 		if (tcp_output(tp) < 0) {
16323 #ifdef TCP_ACCOUNTING
16324 			sched_unpin();
16325 #endif
16326 			return (1);
16327 		}
16328 		did_out = 1;
16329 	}
16330 	if (tp->t_flags2 & TF2_HPTS_CALLS)
16331 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
16332 	rack_free_trim(rack);
16333 #ifdef TCP_ACCOUNTING
16334 	sched_unpin();
16335 #endif
16336 	rack_timer_audit(tp, rack, &so->so_snd);
16337 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16338 	return (0);
16339 }
16340 
16341 #define	TCP_LRO_TS_OPTION \
16342     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16343 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16344 
16345 static int
16346 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16347     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16348     struct timeval *tv)
16349 {
16350 	struct inpcb *inp = tptoinpcb(tp);
16351 	struct socket *so = tptosocket(tp);
16352 #ifdef TCP_ACCOUNTING
16353 	uint64_t ts_val;
16354 #endif
16355 	int32_t thflags, retval, did_out = 0;
16356 	int32_t way_out = 0;
16357 	/*
16358 	 * cts - is the current time from tv (caller gets ts) in microseconds.
16359 	 * ms_cts - is the current time from tv in milliseconds.
16360 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16361 	 */
16362 	uint32_t cts, us_cts, ms_cts;
16363 	uint32_t tiwin;
16364 	struct timespec ts;
16365 	struct tcpopt to;
16366 	struct tcp_rack *rack;
16367 	struct rack_sendmap *rsm;
16368 	int32_t prev_state = 0;
16369 	int no_output = 0;
16370 	int slot_remaining = 0;
16371 #ifdef TCP_ACCOUNTING
16372 	int ack_val_set = 0xf;
16373 #endif
16374 	int nsegs;
16375 
16376 	NET_EPOCH_ASSERT();
16377 	INP_WLOCK_ASSERT(inp);
16378 
16379 	/*
16380 	 * tv passed from common code is from either M_TSTMP_LRO or
16381 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16382 	 */
16383 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16384 	if (rack->rack_deferred_inited == 0) {
16385 		/*
16386 		 * If we are the connecting socket we will
16387 		 * hit rack_init() when no sequence numbers
16388 		 * are setup. This makes it so we must defer
16389 		 * some initialization. Call that now.
16390 		 */
16391 		rack_deferred_init(tp, rack);
16392 	}
16393 	/*
16394 	 * Check to see if we need to skip any output plans. This
16395 	 * can happen in the non-LRO path where we are pacing and
16396 	 * must process the ack coming in but need to defer sending
16397 	 * anything becase a pacing timer is running.
16398 	 */
16399 	us_cts = tcp_tv_to_usec(tv);
16400 	if (m->m_flags & M_ACKCMP) {
16401 		/*
16402 		 * All compressed ack's are ack's by definition so
16403 		 * remove any ack required flag and then do the processing.
16404 		 */
16405 		rack->rc_ack_required = 0;
16406 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16407 	}
16408 	thflags = tcp_get_flags(th);
16409 	if ((rack->rc_always_pace == 1) &&
16410 	    (rack->rc_ack_can_sendout_data == 0) &&
16411 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16412 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16413 		/*
16414 		 * Ok conditions are right for queuing the packets
16415 		 * but we do have to check the flags in the inp, it
16416 		 * could be, if a sack is present, we want to be awoken and
16417 		 * so should process the packets.
16418 		 */
16419 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16420 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16421 			no_output = 1;
16422 		} else {
16423 			/*
16424 			 * If there is no options, or just a
16425 			 * timestamp option, we will want to queue
16426 			 * the packets. This is the same that LRO does
16427 			 * and will need to change with accurate ECN.
16428 			 */
16429 			uint32_t *ts_ptr;
16430 			int optlen;
16431 
16432 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16433 			ts_ptr = (uint32_t *)(th + 1);
16434 			if ((optlen == 0) ||
16435 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
16436 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
16437 				no_output = 1;
16438 		}
16439 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16440 			/*
16441 			 * It is unrealistic to think we can pace in less than
16442 			 * the minimum granularity of the pacer (def:250usec). So
16443 			 * if we have less than that time remaining we should go
16444 			 * ahead and allow output to be "early". We will attempt to
16445 			 * make up for it in any pacing time we try to apply on
16446 			 * the outbound packet.
16447 			 */
16448 			no_output = 0;
16449 		}
16450 	}
16451 	/*
16452 	 * If there is a RST or FIN lets dump out the bw
16453 	 * with a FIN the connection may go on but we
16454 	 * may not.
16455 	 */
16456 	if ((thflags & TH_FIN) || (thflags & TH_RST))
16457 		rack_log_pacing_delay_calc(rack,
16458 					   rack->r_ctl.gp_bw,
16459 					   0,
16460 					   0,
16461 					   rack_get_gp_est(rack), /* delRate */
16462 					   rack_get_lt_bw(rack), /* rttProp */
16463 					   20, __LINE__, NULL, 0);
16464 	if (m->m_flags & M_ACKCMP) {
16465 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16466 	}
16467 	cts = tcp_tv_to_usec(tv);
16468 	ms_cts =  tcp_tv_to_msec(tv);
16469 	nsegs = m->m_pkthdr.lro_nsegs;
16470 	counter_u64_add(rack_proc_non_comp_ack, 1);
16471 #ifdef TCP_ACCOUNTING
16472 	sched_pin();
16473 	if (thflags & TH_ACK)
16474 		ts_val = get_cyclecount();
16475 #endif
16476 	if ((m->m_flags & M_TSTMP) ||
16477 	    (m->m_flags & M_TSTMP_LRO)) {
16478 		mbuf_tstmp2timespec(m, &ts);
16479 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16480 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16481 	} else
16482 		rack->r_ctl.act_rcv_time = *tv;
16483 	kern_prefetch(rack, &prev_state);
16484 	prev_state = 0;
16485 	/*
16486 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
16487 	 * the scale is zero.
16488 	 */
16489 	tiwin = th->th_win << tp->snd_scale;
16490 #ifdef TCP_ACCOUNTING
16491 	if (thflags & TH_ACK) {
16492 		/*
16493 		 * We have a tradeoff here. We can either do what we are
16494 		 * doing i.e. pinning to this CPU and then doing the accounting
16495 		 * <or> we could do a critical enter, setup the rdtsc and cpu
16496 		 * as in below, and then validate we are on the same CPU on
16497 		 * exit. I have choosen to not do the critical enter since
16498 		 * that often will gain you a context switch, and instead lock
16499 		 * us (line above this if) to the same CPU with sched_pin(). This
16500 		 * means we may be context switched out for a higher priority
16501 		 * interupt but we won't be moved to another CPU.
16502 		 *
16503 		 * If this occurs (which it won't very often since we most likely
16504 		 * are running this code in interupt context and only a higher
16505 		 * priority will bump us ... clock?) we will falsely add in
16506 		 * to the time the interupt processing time plus the ack processing
16507 		 * time. This is ok since its a rare event.
16508 		 */
16509 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16510 						    ctf_fixed_maxseg(tp));
16511 	}
16512 #endif
16513 	/*
16514 	 * Parse options on any incoming segment.
16515 	 */
16516 	memset(&to, 0, sizeof(to));
16517 	tcp_dooptions(&to, (u_char *)(th + 1),
16518 	    (th->th_off << 2) - sizeof(struct tcphdr),
16519 	    (thflags & TH_SYN) ? TO_SYN : 0);
16520 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16521 	    __func__));
16522 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16523 	    __func__));
16524 	if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
16525 		/*
16526 		 * We don't look at sack's from the
16527 		 * peer because the MSS is too small which
16528 		 * can subject us to an attack.
16529 		 */
16530 		to.to_flags &= ~TOF_SACK;
16531 	}
16532 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16533 	    (tp->t_flags & TF_GPUTINPROG)) {
16534 		/*
16535 		 * We have a goodput in progress
16536 		 * and we have entered a late state.
16537 		 * Do we have enough data in the sb
16538 		 * to handle the GPUT request?
16539 		 */
16540 		uint32_t bytes;
16541 
16542 		bytes = tp->gput_ack - tp->gput_seq;
16543 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16544 			bytes += tp->gput_seq - tp->snd_una;
16545 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16546 			/*
16547 			 * There are not enough bytes in the socket
16548 			 * buffer that have been sent to cover this
16549 			 * measurement. Cancel it.
16550 			 */
16551 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16552 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16553 						   tp->gput_seq,
16554 						   0, 0, 18, __LINE__, NULL, 0);
16555 			tp->t_flags &= ~TF_GPUTINPROG;
16556 		}
16557 	}
16558 	if (tcp_bblogging_on(rack->rc_tp)) {
16559 		union tcp_log_stackspecific log;
16560 		struct timeval ltv;
16561 #ifdef TCP_REQUEST_TRK
16562 		struct tcp_sendfile_track *tcp_req;
16563 
16564 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
16565 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16566 		} else {
16567 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16568 		}
16569 #endif
16570 		memset(&log, 0, sizeof(log));
16571 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16572 		if (rack->rack_no_prr == 0)
16573 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16574 		else
16575 			log.u_bbr.flex1 = 0;
16576 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16577 		log.u_bbr.use_lt_bw <<= 1;
16578 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16579 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16580 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16581 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16582 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16583 		log.u_bbr.flex3 = m->m_flags;
16584 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16585 		log.u_bbr.lost = thflags;
16586 		log.u_bbr.pacing_gain = 0x1;
16587 #ifdef TCP_ACCOUNTING
16588 		log.u_bbr.cwnd_gain = ack_val_set;
16589 #endif
16590 		log.u_bbr.flex7 = 2;
16591 		if (m->m_flags & M_TSTMP) {
16592 			/* Record the hardware timestamp if present */
16593 			mbuf_tstmp2timespec(m, &ts);
16594 			ltv.tv_sec = ts.tv_sec;
16595 			ltv.tv_usec = ts.tv_nsec / 1000;
16596 			log.u_bbr.lt_epoch = tcp_tv_to_usec(&ltv);
16597 		} else if (m->m_flags & M_TSTMP_LRO) {
16598 			/* Record the LRO the arrival timestamp */
16599 			mbuf_tstmp2timespec(m, &ts);
16600 			ltv.tv_sec = ts.tv_sec;
16601 			ltv.tv_usec = ts.tv_nsec / 1000;
16602 			log.u_bbr.flex5 = tcp_tv_to_usec(&ltv);
16603 		}
16604 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16605 		/* Log the rcv time */
16606 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16607 #ifdef TCP_REQUEST_TRK
16608 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16609 		log.u_bbr.applimited <<= 8;
16610 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16611 		log.u_bbr.applimited <<= 8;
16612 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16613 		if (tcp_req) {
16614 			/* Copy out any client req info */
16615 			/* seconds */
16616 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16617 			/* useconds */
16618 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16619 			log.u_bbr.rttProp = tcp_req->timestamp;
16620 			log.u_bbr.cur_del_rate = tcp_req->start;
16621 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16622 				log.u_bbr.flex8 |= 1;
16623 			} else {
16624 				log.u_bbr.flex8 |= 2;
16625 				log.u_bbr.bw_inuse = tcp_req->end;
16626 			}
16627 			log.u_bbr.flex6 = tcp_req->start_seq;
16628 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16629 				log.u_bbr.flex8 |= 4;
16630 				log.u_bbr.epoch = tcp_req->end_seq;
16631 			}
16632 		}
16633 #endif
16634 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16635 		    tlen, &log, true, &ltv);
16636 	}
16637 	/* Remove ack required flag if set, we have one  */
16638 	if (thflags & TH_ACK)
16639 		rack->rc_ack_required = 0;
16640 	rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16641 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16642 		way_out = 4;
16643 		retval = 0;
16644 		m_freem(m);
16645 		goto done_with_input;
16646 	}
16647 	/*
16648 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16649 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16650 	 */
16651 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16652 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16653 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16654 		ctf_do_dropwithreset(m, tp, th, tlen);
16655 #ifdef TCP_ACCOUNTING
16656 		sched_unpin();
16657 #endif
16658 		return (1);
16659 	}
16660 	/*
16661 	 * If timestamps were negotiated during SYN/ACK and a
16662 	 * segment without a timestamp is received, silently drop
16663 	 * the segment, unless it is a RST segment or missing timestamps are
16664 	 * tolerated.
16665 	 * See section 3.2 of RFC 7323.
16666 	 */
16667 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16668 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16669 		way_out = 5;
16670 		retval = 0;
16671 		m_freem(m);
16672 		goto done_with_input;
16673 	}
16674 	/*
16675 	 * Segment received on connection. Reset idle time and keep-alive
16676 	 * timer. XXX: This should be done after segment validation to
16677 	 * ignore broken/spoofed segs.
16678 	 */
16679 	if  (tp->t_idle_reduce &&
16680 	     (tp->snd_max == tp->snd_una) &&
16681 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16682 		counter_u64_add(rack_input_idle_reduces, 1);
16683 		rack_cc_after_idle(rack, tp);
16684 	}
16685 	tp->t_rcvtime = ticks;
16686 #ifdef STATS
16687 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16688 #endif
16689 	if (tiwin > rack->r_ctl.rc_high_rwnd)
16690 		rack->r_ctl.rc_high_rwnd = tiwin;
16691 	/*
16692 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16693 	 * this to occur after we've validated the segment.
16694 	 */
16695 	if (tcp_ecn_input_segment(tp, thflags, tlen,
16696 	    tcp_packets_this_ack(tp, th->th_ack),
16697 	    iptos))
16698 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16699 
16700 	/*
16701 	 * If echoed timestamp is later than the current time, fall back to
16702 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16703 	 * were used when this connection was established.
16704 	 */
16705 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16706 		to.to_tsecr -= tp->ts_offset;
16707 		if (TSTMP_GT(to.to_tsecr, ms_cts))
16708 			to.to_tsecr = 0;
16709 	}
16710 	if ((rack->r_rcvpath_rtt_up == 1) &&
16711 	    (to.to_flags & TOF_TS) &&
16712 	    (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
16713 		uint32_t rtt = 0;
16714 
16715 		/*
16716 		 * We are receiving only and thus not sending
16717 		 * data to do an RTT. We set a flag when we first
16718 		 * sent this TS to the peer. We now have it back
16719 		 * and have an RTT to share. We log it as a conf
16720 		 * 4, we are not so sure about it.. since we
16721 		 * may have lost an ack.
16722 		 */
16723 		if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
16724 		    rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
16725 		rack->r_rcvpath_rtt_up = 0;
16726 		/* Submit and commit the timer */
16727 		if (rtt > 0) {
16728 			tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
16729 			tcp_rack_xmit_timer_commit(rack, tp);
16730 		}
16731 	}
16732 	/*
16733 	 * If its the first time in we need to take care of options and
16734 	 * verify we can do SACK for rack!
16735 	 */
16736 	if (rack->r_state == 0) {
16737 		/* Should be init'd by rack_init() */
16738 		KASSERT(rack->rc_inp != NULL,
16739 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
16740 		if (rack->rc_inp == NULL) {
16741 			rack->rc_inp = inp;
16742 		}
16743 
16744 		/*
16745 		 * Process options only when we get SYN/ACK back. The SYN
16746 		 * case for incoming connections is handled in tcp_syncache.
16747 		 * According to RFC1323 the window field in a SYN (i.e., a
16748 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16749 		 * this is traditional behavior, may need to be cleaned up.
16750 		 */
16751 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16752 			/* Handle parallel SYN for ECN */
16753 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16754 			if ((to.to_flags & TOF_SCALE) &&
16755 			    (tp->t_flags & TF_REQ_SCALE)) {
16756 				tp->t_flags |= TF_RCVD_SCALE;
16757 				tp->snd_scale = to.to_wscale;
16758 			} else
16759 				tp->t_flags &= ~TF_REQ_SCALE;
16760 			/*
16761 			 * Initial send window.  It will be updated with the
16762 			 * next incoming segment to the scaled value.
16763 			 */
16764 			tp->snd_wnd = th->th_win;
16765 			rack_validate_fo_sendwin_up(tp, rack);
16766 			if ((to.to_flags & TOF_TS) &&
16767 			    (tp->t_flags & TF_REQ_TSTMP)) {
16768 				tp->t_flags |= TF_RCVD_TSTMP;
16769 				tp->ts_recent = to.to_tsval;
16770 				tp->ts_recent_age = cts;
16771 			} else
16772 				tp->t_flags &= ~TF_REQ_TSTMP;
16773 			if (to.to_flags & TOF_MSS) {
16774 				tcp_mss(tp, to.to_mss);
16775 			}
16776 			if ((tp->t_flags & TF_SACK_PERMIT) &&
16777 			    (to.to_flags & TOF_SACKPERM) == 0)
16778 				tp->t_flags &= ~TF_SACK_PERMIT;
16779 			if (tp->t_flags & TF_FASTOPEN) {
16780 				if (to.to_flags & TOF_FASTOPEN) {
16781 					uint16_t mss;
16782 
16783 					if (to.to_flags & TOF_MSS)
16784 						mss = to.to_mss;
16785 					else
16786 						if ((inp->inp_vflag & INP_IPV6) != 0)
16787 							mss = TCP6_MSS;
16788 						else
16789 							mss = TCP_MSS;
16790 					tcp_fastopen_update_cache(tp, mss,
16791 					    to.to_tfo_len, to.to_tfo_cookie);
16792 				} else
16793 					tcp_fastopen_disable_path(tp);
16794 			}
16795 		}
16796 		/*
16797 		 * At this point we are at the initial call. Here we decide
16798 		 * if we are doing RACK or not. We do this by seeing if
16799 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
16800 		 * The code now does do dup-ack counting so if you don't
16801 		 * switch back you won't get rack & TLP, but you will still
16802 		 * get this stack.
16803 		 */
16804 
16805 		if ((rack_sack_not_required == 0) &&
16806 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16807 			tcp_switch_back_to_default(tp);
16808 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16809 			    tlen, iptos);
16810 #ifdef TCP_ACCOUNTING
16811 			sched_unpin();
16812 #endif
16813 			return (1);
16814 		}
16815 		tcp_set_hpts(tp);
16816 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16817 	}
16818 	if (thflags & TH_FIN)
16819 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16820 	us_cts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
16821 	if ((rack->rc_gp_dyn_mul) &&
16822 	    (rack->use_fixed_rate == 0) &&
16823 	    (rack->rc_always_pace)) {
16824 		/* Check in on probertt */
16825 		rack_check_probe_rtt(rack, cts);
16826 	}
16827 	rack_clear_rate_sample(rack);
16828 	if ((rack->forced_ack) &&
16829 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
16830 		rack_handle_probe_response(rack, tiwin, us_cts);
16831 	}
16832 	/*
16833 	 * This is the one exception case where we set the rack state
16834 	 * always. All other times (timers etc) we must have a rack-state
16835 	 * set (so we assure we have done the checks above for SACK).
16836 	 */
16837 	rack->r_ctl.rc_rcvtime = cts;
16838 	if (rack->r_state != tp->t_state)
16839 		rack_set_state(tp, rack);
16840 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
16841 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16842 		kern_prefetch(rsm, &prev_state);
16843 	prev_state = rack->r_state;
16844 	if ((thflags & TH_RST) &&
16845 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16846 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16847 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16848 		/* The connection will be killed by a reset check the tracepoint */
16849 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16850 	}
16851 	retval = (*rack->r_substate) (m, th, so,
16852 	    tp, &to, drop_hdrlen,
16853 	    tlen, tiwin, thflags, nxt_pkt, iptos);
16854 	if (retval == 0) {
16855 		/*
16856 		 * If retval is 1 the tcb is unlocked and most likely the tp
16857 		 * is gone.
16858 		 */
16859 		INP_WLOCK_ASSERT(inp);
16860 		if ((rack->rc_gp_dyn_mul) &&
16861 		    (rack->rc_always_pace) &&
16862 		    (rack->use_fixed_rate == 0) &&
16863 		    rack->in_probe_rtt &&
16864 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
16865 			/*
16866 			 * If we are going for target, lets recheck before
16867 			 * we output.
16868 			 */
16869 			rack_check_probe_rtt(rack, cts);
16870 		}
16871 		if (rack->set_pacing_done_a_iw == 0) {
16872 			/* How much has been acked? */
16873 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16874 				/* We have enough to set in the pacing segment size */
16875 				rack->set_pacing_done_a_iw = 1;
16876 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
16877 			}
16878 		}
16879 		tcp_rack_xmit_timer_commit(rack, tp);
16880 #ifdef TCP_ACCOUNTING
16881 		/*
16882 		 * If we set the ack_val_se to what ack processing we are doing
16883 		 * we also want to track how many cycles we burned. Note
16884 		 * the bits after tcp_output we let be "free". This is because
16885 		 * we are also tracking the tcp_output times as well. Note the
16886 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
16887 		 * 0xf cannot be returned and is what we initialize it too to
16888 		 * indicate we are not doing the tabulations.
16889 		 */
16890 		if (ack_val_set != 0xf) {
16891 			uint64_t crtsc;
16892 
16893 			crtsc = get_cyclecount();
16894 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16895 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16896 			}
16897 		}
16898 #endif
16899 		if ((nxt_pkt == 0) && (no_output == 0)) {
16900 			if ((rack->r_wanted_output != 0) ||
16901 			    (tp->t_flags & TF_ACKNOW) ||
16902 			    (rack->r_fast_output != 0)) {
16903 
16904 do_output_now:
16905 				if (tcp_output(tp) < 0) {
16906 #ifdef TCP_ACCOUNTING
16907 					sched_unpin();
16908 #endif
16909 					return (1);
16910 				}
16911 				did_out = 1;
16912 			}
16913 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16914 			rack_free_trim(rack);
16915 		} else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
16916 			goto do_output_now;
16917 		} else if ((no_output == 1) &&
16918 			   (nxt_pkt == 0) &&
16919 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
16920 			/*
16921 			 * We are not in hpts and we had a pacing timer up. Use
16922 			 * the remaining time (slot_remaining) to restart the timer.
16923 			 */
16924 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
16925 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
16926 			rack_free_trim(rack);
16927 		}
16928 		/* Clear the flag, it may have been cleared by output but we may not have  */
16929 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
16930 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
16931 		/*
16932 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16933 		 * causes issues when we are just going app limited. Lets
16934 		 * instead use SEQ_GT <or> where its equal but more data
16935 		 * is outstanding.
16936 		 *
16937 		 * Also make sure we are on the last ack of a series. We
16938 		 * have to have all the ack's processed in queue to know
16939 		 * if there is something left outstanding.
16940 		 */
16941 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
16942 		    (rack->rc_new_rnd_needed == 0) &&
16943 		    (nxt_pkt == 0)) {
16944 			/*
16945 			 * We have crossed into a new round with
16946 			 * the new snd_unae.
16947 			 */
16948 			rack_new_round_setup(tp, rack, tp->snd_una);
16949 		}
16950 		if ((nxt_pkt == 0) &&
16951 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
16952 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
16953 		     (tp->t_flags & TF_DELACK) ||
16954 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16955 		      (tp->t_state <= TCPS_CLOSING)))) {
16956 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
16957 			if ((tp->snd_max == tp->snd_una) &&
16958 			    ((tp->t_flags & TF_DELACK) == 0) &&
16959 			    (tcp_in_hpts(rack->rc_tp)) &&
16960 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
16961 				/* keep alive not needed if we are hptsi output yet */
16962 				;
16963 			} else {
16964 				int late = 0;
16965 				if (tcp_in_hpts(tp)) {
16966 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16967 						us_cts = tcp_get_usecs(NULL);
16968 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
16969 							rack->r_early = 1;
16970 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
16971 						} else
16972 							late = 1;
16973 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16974 					}
16975 					tcp_hpts_remove(tp);
16976 				}
16977 				if (late && (did_out == 0)) {
16978 					/*
16979 					 * We are late in the sending
16980 					 * and we did not call the output
16981 					 * (this probably should not happen).
16982 					 */
16983 					goto do_output_now;
16984 				}
16985 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16986 			}
16987 			way_out = 1;
16988 		} else if (nxt_pkt == 0) {
16989 			/* Do we have the correct timer running? */
16990 			rack_timer_audit(tp, rack, &so->so_snd);
16991 			way_out = 2;
16992 		}
16993 	done_with_input:
16994 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
16995 		if (did_out)
16996 			rack->r_wanted_output = 0;
16997 	}
16998 
16999 #ifdef TCP_ACCOUNTING
17000 	sched_unpin();
17001 #endif
17002 	return (retval);
17003 }
17004 
17005 static void
17006 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17007     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17008 {
17009 	struct timeval tv;
17010 
17011 	/* First lets see if we have old packets */
17012 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17013 		if (ctf_do_queued_segments(tp, 1)) {
17014 			m_freem(m);
17015 			return;
17016 		}
17017 	}
17018 	if (m->m_flags & M_TSTMP_LRO) {
17019 		mbuf_tstmp2timeval(m, &tv);
17020 	} else {
17021 		/* Should not be should we kassert instead? */
17022 		tcp_get_usecs(&tv);
17023 	}
17024 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17025 	    &tv) == 0) {
17026 		INP_WUNLOCK(tptoinpcb(tp));
17027 	}
17028 }
17029 
17030 struct rack_sendmap *
17031 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17032 {
17033 	struct rack_sendmap *rsm = NULL;
17034 	int32_t idx;
17035 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17036 
17037 	/* Return the next guy to be re-transmitted */
17038 	if (tqhash_empty(rack->r_ctl.tqh)) {
17039 		return (NULL);
17040 	}
17041 	if (tp->t_flags & TF_SENTFIN) {
17042 		/* retran the end FIN? */
17043 		return (NULL);
17044 	}
17045 	/* ok lets look at this one */
17046 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17047 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17048 		return (rsm);
17049 	}
17050 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17051 		goto check_it;
17052 	}
17053 	rsm = rack_find_lowest_rsm(rack);
17054 	if (rsm == NULL) {
17055 		return (NULL);
17056 	}
17057 check_it:
17058 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
17059 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17060 		/*
17061 		 * No sack so we automatically do the 3 strikes and
17062 		 * retransmit (no rack timer would be started).
17063 		 */
17064 		return (rsm);
17065 	}
17066 	if (rsm->r_flags & RACK_ACKED) {
17067 		return (NULL);
17068 	}
17069 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17070 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17071 		/* Its not yet ready */
17072 		return (NULL);
17073 	}
17074 	srtt = rack_grab_rtt(tp, rack);
17075 	idx = rsm->r_rtr_cnt - 1;
17076 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17077 	thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
17078 	if ((tsused == ts_low) ||
17079 	    (TSTMP_LT(tsused, ts_low))) {
17080 		/* No time since sending */
17081 		return (NULL);
17082 	}
17083 	if ((tsused - ts_low) < thresh) {
17084 		/* It has not been long enough yet */
17085 		return (NULL);
17086 	}
17087 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17088 	    ((rsm->r_flags & RACK_SACK_PASSED))) {
17089 		/*
17090 		 * We have passed the dup-ack threshold <or>
17091 		 * a SACK has indicated this is missing.
17092 		 * Note that if you are a declared attacker
17093 		 * it is only the dup-ack threshold that
17094 		 * will cause retransmits.
17095 		 */
17096 		/* log retransmit reason */
17097 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17098 		rack->r_fast_output = 0;
17099 		return (rsm);
17100 	}
17101 	return (NULL);
17102 }
17103 
17104 static void
17105 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
17106 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17107 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17108 {
17109 	if (tcp_bblogging_on(rack->rc_tp)) {
17110 		union tcp_log_stackspecific log;
17111 		struct timeval tv;
17112 
17113 		if (rack_verbose_logging == 0) {
17114 			/*
17115 			 * We are not verbose screen out all but
17116 			 * ones we always want.
17117 			 */
17118 			if ((method != 2) &&
17119 			    (method != 3) &&
17120 			    (method != 7) &&
17121 			    (method != 89) &&
17122 			    (method != 14) &&
17123 			    (method != 20)) {
17124 				return;
17125 			}
17126 		}
17127 		memset(&log, 0, sizeof(log));
17128 		log.u_bbr.flex1 = slot;
17129 		log.u_bbr.flex2 = len;
17130 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17131 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17132 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17133 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17134 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17135 		log.u_bbr.use_lt_bw <<= 1;
17136 		log.u_bbr.use_lt_bw |= rack->r_late;
17137 		log.u_bbr.use_lt_bw <<= 1;
17138 		log.u_bbr.use_lt_bw |= rack->r_early;
17139 		log.u_bbr.use_lt_bw <<= 1;
17140 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17141 		log.u_bbr.use_lt_bw <<= 1;
17142 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17143 		log.u_bbr.use_lt_bw <<= 1;
17144 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17145 		log.u_bbr.use_lt_bw <<= 1;
17146 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17147 		log.u_bbr.use_lt_bw <<= 1;
17148 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17149 		log.u_bbr.pkt_epoch = line;
17150 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17151 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17152 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17153 		log.u_bbr.bw_inuse = bw_est;
17154 		log.u_bbr.delRate = bw;
17155 		if (rack->r_ctl.gp_bw == 0)
17156 			log.u_bbr.cur_del_rate = 0;
17157 		else
17158 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17159 		log.u_bbr.rttProp = len_time;
17160 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17161 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17162 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17163 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17164 			/* We are in slow start */
17165 			log.u_bbr.flex7 = 1;
17166 		} else {
17167 			/* we are on congestion avoidance */
17168 			log.u_bbr.flex7 = 0;
17169 		}
17170 		log.u_bbr.flex8 = method;
17171 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17172 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17173 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17174 		log.u_bbr.cwnd_gain <<= 1;
17175 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17176 		log.u_bbr.cwnd_gain <<= 1;
17177 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17178 		log.u_bbr.cwnd_gain <<= 1;
17179 		log.u_bbr.cwnd_gain |= rack->use_fixed_rate;
17180 		log.u_bbr.cwnd_gain <<= 1;
17181 		log.u_bbr.cwnd_gain |= rack->rc_always_pace;
17182 		log.u_bbr.cwnd_gain <<= 1;
17183 		log.u_bbr.cwnd_gain |= rack->gp_ready;
17184 		log.u_bbr.bbr_substate = quality;
17185 		log.u_bbr.bbr_state = rack->dgp_on;
17186 		log.u_bbr.bbr_state <<= 1;
17187 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17188 		log.u_bbr.bbr_state <<= 2;
17189 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17190 		    &rack->rc_inp->inp_socket->so_rcv,
17191 		    &rack->rc_inp->inp_socket->so_snd,
17192 		    BBR_LOG_HPTSI_CALC, 0,
17193 		    0, &log, false, &tv);
17194 	}
17195 }
17196 
17197 static uint32_t
17198 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17199 {
17200 	uint32_t new_tso, user_max, pace_one;
17201 
17202 	user_max = rack->rc_user_set_max_segs * mss;
17203 	if (rack->rc_force_max_seg) {
17204 		return (user_max);
17205 	}
17206 	if (rack->use_fixed_rate &&
17207 	    ((rack->r_ctl.crte == NULL) ||
17208 	     (bw != rack->r_ctl.crte->rate))) {
17209 		/* Use the user mss since we are not exactly matched */
17210 		return (user_max);
17211 	}
17212 	if (rack_pace_one_seg ||
17213 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17214 		pace_one = 1;
17215 	else
17216 		pace_one = 0;
17217 
17218 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17219 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17220 	if (new_tso > user_max)
17221 		new_tso = user_max;
17222 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17223 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17224 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17225 	}
17226 	if (rack->r_ctl.rc_user_set_min_segs &&
17227 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17228 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17229 	return (new_tso);
17230 }
17231 
17232 static uint64_t
17233 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17234 {
17235 	uint64_t reduced_win;
17236 	uint32_t gain;
17237 
17238 	if (window_input < rc_init_window(rack)) {
17239 		/*
17240 		 * The cwnd is collapsed to
17241 		 * nearly zero, maybe because of a time-out?
17242 		 * Lets drop back to the lt-bw.
17243 		 */
17244 		reduced_win = rack_get_lt_bw(rack);
17245 		/* Set the flag so the caller knows its a rate and not a reduced window */
17246 		*rate_set = 1;
17247 		gain = 100;
17248 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
17249 		/*
17250 		 * If we are in recover our cwnd needs to be less for
17251 		 * our pacing consideration.
17252 		 */
17253 		if (rack->rack_hibeta == 0) {
17254 			reduced_win = window_input / 2;
17255 			gain = 50;
17256 		} else {
17257 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
17258 			reduced_win /= 100;
17259 			gain = rack->r_ctl.saved_hibeta;
17260 		}
17261 	} else {
17262 		/*
17263 		 * Apply Timely factor to increase/decrease the
17264 		 * amount we are pacing at.
17265 		 */
17266 		gain = rack_get_output_gain(rack, NULL);
17267 		if (gain > rack_gain_p5_ub) {
17268 			gain = rack_gain_p5_ub;
17269 		}
17270 		reduced_win = window_input * gain;
17271 		reduced_win /= 100;
17272 	}
17273 	if (gain_b != NULL)
17274 		*gain_b = gain;
17275 	/*
17276 	 * What is being returned here is a trimmed down
17277 	 * window values in all cases where rate_set is left
17278 	 * at 0. In one case we actually return the rate (lt_bw).
17279 	 * the "reduced_win" is returned as a slimmed down cwnd that
17280 	 * is then calculated by the caller into a rate when rate_set
17281 	 * is 0.
17282 	 */
17283 	return (reduced_win);
17284 }
17285 
17286 static int32_t
17287 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)
17288 {
17289 	uint64_t lentim, fill_bw;
17290 
17291 	rack->r_via_fill_cw = 0;
17292 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17293 		return (slot);
17294 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17295 		return (slot);
17296 	if (rack->r_ctl.rc_last_us_rtt == 0)
17297 		return (slot);
17298 	if (rack->rc_pace_fill_if_rttin_range &&
17299 	    (rack->r_ctl.rc_last_us_rtt >=
17300 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17301 		/* The rtt is huge, N * smallest, lets not fill */
17302 		return (slot);
17303 	}
17304 	if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
17305 		return (slot);
17306 	/*
17307 	 * first lets calculate the b/w based on the last us-rtt
17308 	 * and the the smallest send window.
17309 	 */
17310 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17311 	if (rack->rc_fillcw_apply_discount) {
17312 		uint32_t rate_set = 0;
17313 
17314 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17315 		if (rate_set) {
17316 			goto at_lt_bw;
17317 		}
17318 	}
17319 	/* Take the rwnd if its smaller */
17320 	if (fill_bw > rack->rc_tp->snd_wnd)
17321 		fill_bw = rack->rc_tp->snd_wnd;
17322 	/* Now lets make it into a b/w */
17323 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17324 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17325 	/* Adjust to any cap */
17326 	if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
17327 		fill_bw = rack->r_ctl.fillcw_cap;
17328 
17329 at_lt_bw:
17330 	if (rack_bw_multipler > 0) {
17331 		/*
17332 		 * We want to limit fill-cw to the some multiplier
17333 		 * of the max(lt_bw, gp_est). The normal default
17334 		 * is 0 for off, so a sysctl has enabled it.
17335 		 */
17336 		uint64_t lt_bw, gp, rate;
17337 
17338 		gp = rack_get_gp_est(rack);
17339 		lt_bw = rack_get_lt_bw(rack);
17340 		if (lt_bw > gp)
17341 			rate = lt_bw;
17342 		else
17343 			rate = gp;
17344 		rate *= rack_bw_multipler;
17345 		rate /= 100;
17346 		if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
17347 			union tcp_log_stackspecific log;
17348 			struct timeval tv;
17349 
17350 			memset(&log, 0, sizeof(log));
17351 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17352 			log.u_bbr.flex1 = rack_bw_multipler;
17353 			log.u_bbr.flex2 = len;
17354 			log.u_bbr.cur_del_rate = gp;
17355 			log.u_bbr.delRate = lt_bw;
17356 			log.u_bbr.bw_inuse = rate;
17357 			log.u_bbr.rttProp = fill_bw;
17358 			log.u_bbr.flex8 = 44;
17359 			tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
17360 				      BBR_LOG_CWND, 0,
17361 				      0, &log, false, NULL,
17362 				      __func__, __LINE__, &tv);
17363 		}
17364 		if (fill_bw > rate)
17365 			fill_bw = rate;
17366 	}
17367 	/* We are below the min b/w */
17368 	if (non_paced)
17369 		*rate_wanted = fill_bw;
17370 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17371 		return (slot);
17372 	rack->r_via_fill_cw = 1;
17373 	if (rack->r_rack_hw_rate_caps &&
17374 	    (rack->r_ctl.crte != NULL)) {
17375 		uint64_t high_rate;
17376 
17377 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17378 		if (fill_bw > high_rate) {
17379 			/* We are capping bw at the highest rate table entry */
17380 			if (*rate_wanted > high_rate) {
17381 				/* The original rate was also capped */
17382 				rack->r_via_fill_cw = 0;
17383 			}
17384 			rack_log_hdwr_pacing(rack,
17385 					     fill_bw, high_rate, __LINE__,
17386 					     0, 3);
17387 			fill_bw = high_rate;
17388 			if (capped)
17389 				*capped = 1;
17390 		}
17391 	} else if ((rack->r_ctl.crte == NULL) &&
17392 		   (rack->rack_hdrw_pacing == 0) &&
17393 		   (rack->rack_hdw_pace_ena) &&
17394 		   rack->r_rack_hw_rate_caps &&
17395 		   (rack->rack_attempt_hdwr_pace == 0) &&
17396 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
17397 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17398 		/*
17399 		 * Ok we may have a first attempt that is greater than our top rate
17400 		 * lets check.
17401 		 */
17402 		uint64_t high_rate;
17403 
17404 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17405 		if (high_rate) {
17406 			if (fill_bw > high_rate) {
17407 				fill_bw = high_rate;
17408 				if (capped)
17409 					*capped = 1;
17410 			}
17411 		}
17412 	}
17413 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17414 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17415 				   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17416 		fill_bw = rack->r_ctl.bw_rate_cap;
17417 	}
17418 	/*
17419 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
17420 	 * in an rtt (unless it was capped), what does that
17421 	 * time wise equate too?
17422 	 */
17423 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17424 	lentim /= fill_bw;
17425 	*rate_wanted = fill_bw;
17426 	if (non_paced || (lentim < slot)) {
17427 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17428 					   0, lentim, 12, __LINE__, NULL, 0);
17429 		return ((int32_t)lentim);
17430 	} else
17431 		return (slot);
17432 }
17433 
17434 static int32_t
17435 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
17436 {
17437 	uint64_t srtt;
17438 	int32_t slot = 0;
17439 	int can_start_hw_pacing = 1;
17440 	int err;
17441 	int pace_one;
17442 
17443 	if (rack_pace_one_seg ||
17444 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17445 		pace_one = 1;
17446 	else
17447 		pace_one = 0;
17448 	if (rack->rc_always_pace == 0) {
17449 		/*
17450 		 * We use the most optimistic possible cwnd/srtt for
17451 		 * sending calculations. This will make our
17452 		 * calculation anticipate getting more through
17453 		 * quicker then possible. But thats ok we don't want
17454 		 * the peer to have a gap in data sending.
17455 		 */
17456 		uint64_t cwnd, tr_perms = 0;
17457 		int32_t reduce;
17458 
17459 	old_method:
17460 		/*
17461 		 * We keep no precise pacing with the old method
17462 		 * instead we use the pacer to mitigate bursts.
17463 		 */
17464 		if (rack->r_ctl.rc_rack_min_rtt)
17465 			srtt = rack->r_ctl.rc_rack_min_rtt;
17466 		else
17467 			srtt = max(tp->t_srtt, 1);
17468 		if (rack->r_ctl.rc_rack_largest_cwnd)
17469 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17470 		else
17471 			cwnd = rack->r_ctl.cwnd_to_use;
17472 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
17473 		tr_perms = (cwnd * 1000) / srtt;
17474 		if (tr_perms == 0) {
17475 			tr_perms = ctf_fixed_maxseg(tp);
17476 		}
17477 		/*
17478 		 * Calculate how long this will take to drain, if
17479 		 * the calculation comes out to zero, thats ok we
17480 		 * will use send_a_lot to possibly spin around for
17481 		 * more increasing tot_len_this_send to the point
17482 		 * that its going to require a pace, or we hit the
17483 		 * cwnd. Which in that case we are just waiting for
17484 		 * a ACK.
17485 		 */
17486 		slot = len / tr_perms;
17487 		/* Now do we reduce the time so we don't run dry? */
17488 		if (slot && rack_slot_reduction) {
17489 			reduce = (slot / rack_slot_reduction);
17490 			if (reduce < slot) {
17491 				slot -= reduce;
17492 			} else
17493 				slot = 0;
17494 		} else
17495 			reduce = 0;
17496 		slot *= HPTS_USEC_IN_MSEC;
17497 		if (rack->rc_pace_to_cwnd) {
17498 			uint64_t rate_wanted = 0;
17499 
17500 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17501 			rack->rc_ack_can_sendout_data = 1;
17502 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17503 		} else
17504 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17505 		/*******************************************************/
17506 		/* RRS: We insert non-paced call to stats here for len */
17507 		/*******************************************************/
17508 	} else {
17509 		uint64_t bw_est, res, lentim, rate_wanted;
17510 		uint32_t segs, oh;
17511 		int capped = 0;
17512 		int prev_fill;
17513 
17514 		if ((rack->r_rr_config == 1) && rsm) {
17515 			return (rack->r_ctl.rc_min_to);
17516 		}
17517 		if (rack->use_fixed_rate) {
17518 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17519 		} else if ((rack->r_ctl.init_rate == 0) &&
17520 			   (rack->r_ctl.gp_bw == 0)) {
17521 			/* no way to yet do an estimate */
17522 			bw_est = rate_wanted = 0;
17523 		} else if (rack->dgp_on) {
17524 			bw_est = rack_get_bw(rack);
17525 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17526 		} else {
17527 			uint32_t gain, rate_set = 0;
17528 
17529 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17530 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17531 			if (rate_set == 0) {
17532 				if (rate_wanted > rack->rc_tp->snd_wnd)
17533 					rate_wanted = rack->rc_tp->snd_wnd;
17534 				/* Now lets make it into a b/w */
17535 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17536 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17537 			}
17538 			bw_est = rate_wanted;
17539 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17540 						   rack->r_ctl.cwnd_to_use,
17541 						   rate_wanted, bw_est,
17542 						   rack->r_ctl.rc_last_us_rtt,
17543 						   88, __LINE__, NULL, gain);
17544 		}
17545 		if (((bw_est == 0) || (rate_wanted == 0) || (rack->gp_ready == 0)) &&
17546 		    (rack->use_fixed_rate == 0)) {
17547 			/*
17548 			 * No way yet to make a b/w estimate or
17549 			 * our raise is set incorrectly.
17550 			 */
17551 			goto old_method;
17552 		}
17553 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
17554 		/* We need to account for all the overheads */
17555 		segs = (len + segsiz - 1) / segsiz;
17556 		/*
17557 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17558 		 * and how much data we put in each packet. Yes this
17559 		 * means we may be off if we are larger than 1500 bytes
17560 		 * or smaller. But this just makes us more conservative.
17561 		 */
17562 
17563 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17564 		if (rack->r_is_v6) {
17565 #ifdef INET6
17566 			oh += sizeof(struct ip6_hdr);
17567 #endif
17568 		} else {
17569 #ifdef INET
17570 			oh += sizeof(struct ip);
17571 #endif
17572 		}
17573 		/* We add a fixed 14 for the ethernet header */
17574 		oh += 14;
17575 		segs *= oh;
17576 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17577 		res = lentim / rate_wanted;
17578 		slot = (uint32_t)res;
17579 		if (rack_hw_rate_min &&
17580 		    (rate_wanted < rack_hw_rate_min)) {
17581 			can_start_hw_pacing = 0;
17582 			if (rack->r_ctl.crte) {
17583 				/*
17584 				 * Ok we need to release it, we
17585 				 * have fallen too low.
17586 				 */
17587 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17588 				rack->r_ctl.crte = NULL;
17589 				rack->rack_attempt_hdwr_pace = 0;
17590 				rack->rack_hdrw_pacing = 0;
17591 			}
17592 		}
17593 		if (rack->r_ctl.crte &&
17594 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17595 			/*
17596 			 * We want more than the hardware can give us,
17597 			 * don't start any hw pacing.
17598 			 */
17599 			can_start_hw_pacing = 0;
17600 			if (rack->r_rack_hw_rate_caps == 0) {
17601 				/*
17602 				 * Ok we need to release it, we
17603 				 * want more than the card can give us and
17604 				 * no rate cap is in place. Set it up so
17605 				 * when we want less we can retry.
17606 				 */
17607 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17608 				rack->r_ctl.crte = NULL;
17609 				rack->rack_attempt_hdwr_pace = 0;
17610 				rack->rack_hdrw_pacing = 0;
17611 			}
17612 		}
17613 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17614 			/*
17615 			 * We lost our rate somehow, this can happen
17616 			 * if the interface changed underneath us.
17617 			 */
17618 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17619 			rack->r_ctl.crte = NULL;
17620 			/* Lets re-allow attempting to setup pacing */
17621 			rack->rack_hdrw_pacing = 0;
17622 			rack->rack_attempt_hdwr_pace = 0;
17623 			rack_log_hdwr_pacing(rack,
17624 					     rate_wanted, bw_est, __LINE__,
17625 					     0, 6);
17626 		}
17627 		prev_fill = rack->r_via_fill_cw;
17628 		if ((rack->rc_pace_to_cwnd) &&
17629 		    (capped == 0) &&
17630 		    (rack->dgp_on == 1) &&
17631 		    (rack->use_fixed_rate == 0) &&
17632 		    (rack->in_probe_rtt == 0) &&
17633 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17634 			/*
17635 			 * We want to pace at our rate *or* faster to
17636 			 * fill the cwnd to the max if its not full.
17637 			 */
17638 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17639 			/* Re-check to make sure we are not exceeding our max b/w */
17640 			if ((rack->r_ctl.crte != NULL) &&
17641 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17642 				/*
17643 				 * We want more than the hardware can give us,
17644 				 * don't start any hw pacing.
17645 				 */
17646 				can_start_hw_pacing = 0;
17647 				if (rack->r_rack_hw_rate_caps == 0) {
17648 					/*
17649 					 * Ok we need to release it, we
17650 					 * want more than the card can give us and
17651 					 * no rate cap is in place. Set it up so
17652 					 * when we want less we can retry.
17653 					 */
17654 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17655 					rack->r_ctl.crte = NULL;
17656 					rack->rack_attempt_hdwr_pace = 0;
17657 					rack->rack_hdrw_pacing = 0;
17658 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17659 				}
17660 			}
17661 		}
17662 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17663 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17664 			if ((rack->rack_hdw_pace_ena) &&
17665 			    (can_start_hw_pacing > 0) &&
17666 			    (rack->rack_hdrw_pacing == 0) &&
17667 			    (rack->rack_attempt_hdwr_pace == 0)) {
17668 				/*
17669 				 * Lets attempt to turn on hardware pacing
17670 				 * if we can.
17671 				 */
17672 				rack->rack_attempt_hdwr_pace = 1;
17673 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17674 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
17675 								       rate_wanted,
17676 								       RS_PACING_GEQ,
17677 								       &err, &rack->r_ctl.crte_prev_rate);
17678 				if (rack->r_ctl.crte) {
17679 					rack->rack_hdrw_pacing = 1;
17680 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17681 													   pace_one, rack->r_ctl.crte,
17682 													   NULL, rack->r_ctl.pace_len_divisor);
17683 					rack_log_hdwr_pacing(rack,
17684 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17685 							     err, 0);
17686 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17687 				} else {
17688 					counter_u64_add(rack_hw_pace_init_fail, 1);
17689 				}
17690 			} else if (rack->rack_hdrw_pacing &&
17691 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17692 				/* Do we need to adjust our rate? */
17693 				const struct tcp_hwrate_limit_table *nrte;
17694 
17695 				if (rack->r_up_only &&
17696 				    (rate_wanted < rack->r_ctl.crte->rate)) {
17697 					/**
17698 					 * We have four possible states here
17699 					 * having to do with the previous time
17700 					 * and this time.
17701 					 *   previous  |  this-time
17702 					 * A)     0      |     0   -- fill_cw not in the picture
17703 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
17704 					 * C)     1      |     1   -- all rates from fill_cw
17705 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
17706 					 *
17707 					 * For case A, C and D we don't allow a drop. But for
17708 					 * case B where we now our on our steady rate we do
17709 					 * allow a drop.
17710 					 *
17711 					 */
17712 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17713 						goto done_w_hdwr;
17714 				}
17715 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
17716 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17717 					if (rack_hw_rate_to_low &&
17718 					    (bw_est < rack_hw_rate_to_low)) {
17719 						/*
17720 						 * The pacing rate is too low for hardware, but
17721 						 * do allow hardware pacing to be restarted.
17722 						 */
17723 						rack_log_hdwr_pacing(rack,
17724 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
17725 								     0, 5);
17726 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17727 						rack->r_ctl.crte = NULL;
17728 						rack->rack_attempt_hdwr_pace = 0;
17729 						rack->rack_hdrw_pacing = 0;
17730 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17731 						goto done_w_hdwr;
17732 					}
17733 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17734 								   rack->rc_tp,
17735 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
17736 								   rate_wanted,
17737 								   RS_PACING_GEQ,
17738 								   &err, &rack->r_ctl.crte_prev_rate);
17739 					if (nrte == NULL) {
17740 						/*
17741 						 * Lost the rate, lets drop hardware pacing
17742 						 * period.
17743 						 */
17744 						rack->rack_hdrw_pacing = 0;
17745 						rack->r_ctl.crte = NULL;
17746 						rack_log_hdwr_pacing(rack,
17747 								     rate_wanted, 0, __LINE__,
17748 								     err, 1);
17749 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17750 						counter_u64_add(rack_hw_pace_lost, 1);
17751 					} else if (nrte != rack->r_ctl.crte) {
17752 						rack->r_ctl.crte = nrte;
17753 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17754 														   segsiz, pace_one, rack->r_ctl.crte,
17755 														   NULL, rack->r_ctl.pace_len_divisor);
17756 						rack_log_hdwr_pacing(rack,
17757 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17758 								     err, 2);
17759 						rack->r_ctl.last_hw_bw_req = rate_wanted;
17760 					}
17761 				} else {
17762 					/* We just need to adjust the segment size */
17763 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17764 					rack_log_hdwr_pacing(rack,
17765 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17766 							     0, 4);
17767 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17768 				}
17769 			}
17770 		}
17771 	done_w_hdwr:
17772 		if (rack_limit_time_with_srtt &&
17773 		    (rack->use_fixed_rate == 0) &&
17774 		    (rack->rack_hdrw_pacing == 0)) {
17775 			/*
17776 			 * Sanity check, we do not allow the pacing delay
17777 			 * to be longer than the SRTT of the path. If it is
17778 			 * a slow path, then adding a packet should increase
17779 			 * the RTT and compensate for this i.e. the srtt will
17780 			 * be greater so the allowed pacing time will be greater.
17781 			 *
17782 			 * Note this restriction is not for where a peak rate
17783 			 * is set, we are doing fixed pacing or hardware pacing.
17784 			 */
17785 			if (rack->rc_tp->t_srtt)
17786 				srtt = rack->rc_tp->t_srtt;
17787 			else
17788 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
17789 			if (srtt < (uint64_t)slot) {
17790 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17791 				slot = srtt;
17792 			}
17793 		}
17794 		/*******************************************************************/
17795 		/* RRS: We insert paced call to stats here for len and rate_wanted */
17796 		/*******************************************************************/
17797 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17798 	}
17799 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17800 		/*
17801 		 * If this rate is seeing enobufs when it
17802 		 * goes to send then either the nic is out
17803 		 * of gas or we are mis-estimating the time
17804 		 * somehow and not letting the queue empty
17805 		 * completely. Lets add to the pacing time.
17806 		 */
17807 		int hw_boost_delay;
17808 
17809 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17810 		if (hw_boost_delay > rack_enobuf_hw_max)
17811 			hw_boost_delay = rack_enobuf_hw_max;
17812 		else if (hw_boost_delay < rack_enobuf_hw_min)
17813 			hw_boost_delay = rack_enobuf_hw_min;
17814 		slot += hw_boost_delay;
17815 	}
17816 	return (slot);
17817 }
17818 
17819 static void
17820 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17821     tcp_seq startseq, uint32_t sb_offset)
17822 {
17823 	struct rack_sendmap *my_rsm = NULL;
17824 
17825 	if (tp->t_state < TCPS_ESTABLISHED) {
17826 		/*
17827 		 * We don't start any measurements if we are
17828 		 * not at least established.
17829 		 */
17830 		return;
17831 	}
17832 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
17833 		/*
17834 		 * We will get no more data into the SB
17835 		 * this means we need to have the data available
17836 		 * before we start a measurement.
17837 		 */
17838 
17839 		if (sbavail(&tptosocket(tp)->so_snd) <
17840 		    max(rc_init_window(rack),
17841 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17842 			/* Nope not enough data */
17843 			return;
17844 		}
17845 	}
17846 	tp->t_flags |= TF_GPUTINPROG;
17847 	rack->r_ctl.rc_gp_cumack_ts = 0;
17848 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17849 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17850 	tp->gput_seq = startseq;
17851 	rack->app_limited_needs_set = 0;
17852 	if (rack->in_probe_rtt)
17853 		rack->measure_saw_probe_rtt = 1;
17854 	else if ((rack->measure_saw_probe_rtt) &&
17855 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17856 		rack->measure_saw_probe_rtt = 0;
17857 	if (rack->rc_gp_filled)
17858 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
17859 	else {
17860 		/* Special case initial measurement */
17861 		struct timeval tv;
17862 
17863 		tp->gput_ts = tcp_get_usecs(&tv);
17864 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17865 	}
17866 	/*
17867 	 * We take a guess out into the future,
17868 	 * if we have no measurement and no
17869 	 * initial rate, we measure the first
17870 	 * initial-windows worth of data to
17871 	 * speed up getting some GP measurement and
17872 	 * thus start pacing.
17873 	 */
17874 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17875 		rack->app_limited_needs_set = 1;
17876 		tp->gput_ack = startseq + max(rc_init_window(rack),
17877 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17878 		rack_log_pacing_delay_calc(rack,
17879 					   tp->gput_seq,
17880 					   tp->gput_ack,
17881 					   0,
17882 					   tp->gput_ts,
17883 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17884 					   9,
17885 					   __LINE__, NULL, 0);
17886 		rack_tend_gp_marks(tp, rack);
17887 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17888 		return;
17889 	}
17890 	if (sb_offset) {
17891 		/*
17892 		 * We are out somewhere in the sb
17893 		 * can we use the already outstanding data?
17894 		 */
17895 
17896 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
17897 			/*
17898 			 * Yes first one is good and in this case
17899 			 * the tp->gput_ts is correctly set based on
17900 			 * the last ack that arrived (no need to
17901 			 * set things up when an ack comes in).
17902 			 */
17903 			my_rsm = tqhash_min(rack->r_ctl.tqh);
17904 			if ((my_rsm == NULL) ||
17905 			    (my_rsm->r_rtr_cnt != 1)) {
17906 				/* retransmission? */
17907 				goto use_latest;
17908 			}
17909 		} else {
17910 			if (rack->r_ctl.rc_first_appl == NULL) {
17911 				/*
17912 				 * If rc_first_appl is NULL
17913 				 * then the cnt should be 0.
17914 				 * This is probably an error, maybe
17915 				 * a KASSERT would be approprate.
17916 				 */
17917 				goto use_latest;
17918 			}
17919 			/*
17920 			 * If we have a marker pointer to the last one that is
17921 			 * app limited we can use that, but we need to set
17922 			 * things up so that when it gets ack'ed we record
17923 			 * the ack time (if its not already acked).
17924 			 */
17925 			rack->app_limited_needs_set = 1;
17926 			/*
17927 			 * We want to get to the rsm that is either
17928 			 * next with space i.e. over 1 MSS or the one
17929 			 * after that (after the app-limited).
17930 			 */
17931 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17932 			if (my_rsm) {
17933 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17934 					/* Have to use the next one */
17935 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17936 				else {
17937 					/* Use after the first MSS of it is acked */
17938 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17939 					goto start_set;
17940 				}
17941 			}
17942 			if ((my_rsm == NULL) ||
17943 			    (my_rsm->r_rtr_cnt != 1)) {
17944 				/*
17945 				 * Either its a retransmit or
17946 				 * the last is the app-limited one.
17947 				 */
17948 				goto use_latest;
17949 			}
17950 		}
17951 		tp->gput_seq = my_rsm->r_start;
17952 start_set:
17953 		if (my_rsm->r_flags & RACK_ACKED) {
17954 			/*
17955 			 * This one has been acked use the arrival ack time
17956 			 */
17957 			struct rack_sendmap *nrsm;
17958 
17959 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17960 			rack->app_limited_needs_set = 0;
17961 			/*
17962 			 * Ok in this path we need to use the r_end now
17963 			 * since this guy is the starting ack.
17964 			 */
17965 			tp->gput_seq = my_rsm->r_end;
17966 			/*
17967 			 * We also need to adjust up the sendtime
17968 			 * to the send of the next data after my_rsm.
17969 			 */
17970 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17971 			if (nrsm != NULL)
17972 				my_rsm = nrsm;
17973 			else {
17974 				/*
17975 				 * The next as not been sent, thats the
17976 				 * case for using the latest.
17977 				 */
17978 				goto use_latest;
17979 			}
17980 		}
17981 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17982 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
17983 		rack->r_ctl.rc_gp_cumack_ts = 0;
17984 		if ((rack->r_ctl.cleared_app_ack == 1) &&
17985 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.cleared_app_ack_seq))) {
17986 			/*
17987 			 * We just cleared an application limited period
17988 			 * so the next seq out needs to skip the first
17989 			 * ack.
17990 			 */
17991 			rack->app_limited_needs_set = 1;
17992 			rack->r_ctl.cleared_app_ack = 0;
17993 		}
17994 		rack_log_pacing_delay_calc(rack,
17995 					   tp->gput_seq,
17996 					   tp->gput_ack,
17997 					   (uintptr_t)my_rsm,
17998 					   tp->gput_ts,
17999 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18000 					   9,
18001 					   __LINE__, my_rsm, 0);
18002 		/* Now lets make sure all are marked as they should be */
18003 		rack_tend_gp_marks(tp, rack);
18004 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18005 		return;
18006 	}
18007 
18008 use_latest:
18009 	/*
18010 	 * We don't know how long we may have been
18011 	 * idle or if this is the first-send. Lets
18012 	 * setup the flag so we will trim off
18013 	 * the first ack'd data so we get a true
18014 	 * measurement.
18015 	 */
18016 	rack->app_limited_needs_set = 1;
18017 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18018 	rack->r_ctl.rc_gp_cumack_ts = 0;
18019 	/* Find this guy so we can pull the send time */
18020 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18021 	if (my_rsm) {
18022 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18023 		if (my_rsm->r_flags & RACK_ACKED) {
18024 			/*
18025 			 * Unlikely since its probably what was
18026 			 * just transmitted (but I am paranoid).
18027 			 */
18028 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18029 			rack->app_limited_needs_set = 0;
18030 		}
18031 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18032 			/* This also is unlikely */
18033 			tp->gput_seq = my_rsm->r_start;
18034 		}
18035 	} else {
18036 		/*
18037 		 * TSNH unless we have some send-map limit,
18038 		 * and even at that it should not be hitting
18039 		 * that limit (we should have stopped sending).
18040 		 */
18041 		struct timeval tv;
18042 
18043 		microuptime(&tv);
18044 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18045 	}
18046 	rack_tend_gp_marks(tp, rack);
18047 	rack_log_pacing_delay_calc(rack,
18048 				   tp->gput_seq,
18049 				   tp->gput_ack,
18050 				   (uintptr_t)my_rsm,
18051 				   tp->gput_ts,
18052 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18053 				   9, __LINE__, NULL, 0);
18054 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18055 }
18056 
18057 static inline uint32_t
18058 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18059     uint32_t avail, int32_t sb_offset)
18060 {
18061 	uint32_t len;
18062 	uint32_t sendwin;
18063 
18064 	if (tp->snd_wnd > cwnd_to_use)
18065 		sendwin = cwnd_to_use;
18066 	else
18067 		sendwin = tp->snd_wnd;
18068 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18069 		/* We never want to go over our peers rcv-window */
18070 		len = 0;
18071 	} else {
18072 		uint32_t flight;
18073 
18074 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18075 		if (flight >= sendwin) {
18076 			/*
18077 			 * We have in flight what we are allowed by cwnd (if
18078 			 * it was rwnd blocking it would have hit above out
18079 			 * >= tp->snd_wnd).
18080 			 */
18081 			return (0);
18082 		}
18083 		len = sendwin - flight;
18084 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18085 			/* We would send too much (beyond the rwnd) */
18086 			len = tp->snd_wnd - ctf_outstanding(tp);
18087 		}
18088 		if ((len + sb_offset) > avail) {
18089 			/*
18090 			 * We don't have that much in the SB, how much is
18091 			 * there?
18092 			 */
18093 			len = avail - sb_offset;
18094 		}
18095 	}
18096 	return (len);
18097 }
18098 
18099 static void
18100 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18101 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18102 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18103 {
18104 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18105 		union tcp_log_stackspecific log;
18106 		struct timeval tv;
18107 
18108 		memset(&log, 0, sizeof(log));
18109 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18110 		log.u_bbr.flex1 = error;
18111 		log.u_bbr.flex2 = flags;
18112 		log.u_bbr.flex3 = rsm_is_null;
18113 		log.u_bbr.flex4 = ipoptlen;
18114 		log.u_bbr.flex5 = tp->rcv_numsacks;
18115 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18116 		log.u_bbr.flex7 = optlen;
18117 		log.u_bbr.flex8 = rack->r_fsb_inited;
18118 		log.u_bbr.applimited = rack->r_fast_output;
18119 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18120 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18121 		log.u_bbr.cwnd_gain = mode;
18122 		log.u_bbr.pkts_out = orig_len;
18123 		log.u_bbr.lt_epoch = len;
18124 		log.u_bbr.delivered = line;
18125 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18126 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18127 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18128 			       len, &log, false, NULL, __func__, __LINE__, &tv);
18129 	}
18130 }
18131 
18132 
18133 static struct mbuf *
18134 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18135 		   struct rack_fast_send_blk *fsb,
18136 		   int32_t seglimit, int32_t segsize, int hw_tls)
18137 {
18138 #ifdef KERN_TLS
18139 	struct ktls_session *tls, *ntls;
18140 #ifdef INVARIANTS
18141 	struct mbuf *start;
18142 #endif
18143 #endif
18144 	struct mbuf *m, *n, **np, *smb;
18145 	struct mbuf *top;
18146 	int32_t off, soff;
18147 	int32_t len = *plen;
18148 	int32_t fragsize;
18149 	int32_t len_cp = 0;
18150 	uint32_t mlen, frags;
18151 
18152 	soff = off = the_off;
18153 	smb = m = the_m;
18154 	np = &top;
18155 	top = NULL;
18156 #ifdef KERN_TLS
18157 	if (hw_tls && (m->m_flags & M_EXTPG))
18158 		tls = m->m_epg_tls;
18159 	else
18160 		tls = NULL;
18161 #ifdef INVARIANTS
18162 	start = m;
18163 #endif
18164 #endif
18165 	while (len > 0) {
18166 		if (m == NULL) {
18167 			*plen = len_cp;
18168 			break;
18169 		}
18170 #ifdef KERN_TLS
18171 		if (hw_tls) {
18172 			if (m->m_flags & M_EXTPG)
18173 				ntls = m->m_epg_tls;
18174 			else
18175 				ntls = NULL;
18176 
18177 			/*
18178 			 * Avoid mixing TLS records with handshake
18179 			 * data or TLS records from different
18180 			 * sessions.
18181 			 */
18182 			if (tls != ntls) {
18183 				MPASS(m != start);
18184 				*plen = len_cp;
18185 				break;
18186 			}
18187 		}
18188 #endif
18189 		mlen = min(len, m->m_len - off);
18190 		if (seglimit) {
18191 			/*
18192 			 * For M_EXTPG mbufs, add 3 segments
18193 			 * + 1 in case we are crossing page boundaries
18194 			 * + 2 in case the TLS hdr/trailer are used
18195 			 * It is cheaper to just add the segments
18196 			 * than it is to take the cache miss to look
18197 			 * at the mbuf ext_pgs state in detail.
18198 			 */
18199 			if (m->m_flags & M_EXTPG) {
18200 				fragsize = min(segsize, PAGE_SIZE);
18201 				frags = 3;
18202 			} else {
18203 				fragsize = segsize;
18204 				frags = 0;
18205 			}
18206 
18207 			/* Break if we really can't fit anymore. */
18208 			if ((frags + 1) >= seglimit) {
18209 				*plen =	len_cp;
18210 				break;
18211 			}
18212 
18213 			/*
18214 			 * Reduce size if you can't copy the whole
18215 			 * mbuf. If we can't copy the whole mbuf, also
18216 			 * adjust len so the loop will end after this
18217 			 * mbuf.
18218 			 */
18219 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18220 				mlen = (seglimit - frags - 1) * fragsize;
18221 				len = mlen;
18222 				*plen = len_cp + len;
18223 			}
18224 			frags += howmany(mlen, fragsize);
18225 			if (frags == 0)
18226 				frags++;
18227 			seglimit -= frags;
18228 			KASSERT(seglimit > 0,
18229 			    ("%s: seglimit went too low", __func__));
18230 		}
18231 		n = m_get(M_NOWAIT, m->m_type);
18232 		*np = n;
18233 		if (n == NULL)
18234 			goto nospace;
18235 		n->m_len = mlen;
18236 		soff += mlen;
18237 		len_cp += n->m_len;
18238 		if (m->m_flags & (M_EXT | M_EXTPG)) {
18239 			n->m_data = m->m_data + off;
18240 			mb_dupcl(n, m);
18241 		} else {
18242 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18243 			    (u_int)n->m_len);
18244 		}
18245 		len -= n->m_len;
18246 		off = 0;
18247 		m = m->m_next;
18248 		np = &n->m_next;
18249 		if (len || (soff == smb->m_len)) {
18250 			/*
18251 			 * We have more so we move forward  or
18252 			 * we have consumed the entire mbuf and
18253 			 * len has fell to 0.
18254 			 */
18255 			soff = 0;
18256 			smb = m;
18257 		}
18258 
18259 	}
18260 	if (fsb != NULL) {
18261 		fsb->m = smb;
18262 		fsb->off = soff;
18263 		if (smb) {
18264 			/*
18265 			 * Save off the size of the mbuf. We do
18266 			 * this so that we can recognize when it
18267 			 * has been trimmed by sbcut() as acks
18268 			 * come in.
18269 			 */
18270 			fsb->o_m_len = smb->m_len;
18271 			fsb->o_t_len = M_TRAILINGROOM(smb);
18272 		} else {
18273 			/*
18274 			 * This is the case where the next mbuf went to NULL. This
18275 			 * means with this copy we have sent everything in the sb.
18276 			 * In theory we could clear the fast_output flag, but lets
18277 			 * not since its possible that we could get more added
18278 			 * and acks that call the extend function which would let
18279 			 * us send more.
18280 			 */
18281 			fsb->o_m_len = 0;
18282 			fsb->o_t_len = 0;
18283 		}
18284 	}
18285 	return (top);
18286 nospace:
18287 	if (top)
18288 		m_freem(top);
18289 	return (NULL);
18290 
18291 }
18292 
18293 /*
18294  * This is a copy of m_copym(), taking the TSO segment size/limit
18295  * constraints into account, and advancing the sndptr as it goes.
18296  */
18297 static struct mbuf *
18298 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18299 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18300 {
18301 	struct mbuf *m, *n;
18302 	int32_t soff;
18303 
18304 	m = rack->r_ctl.fsb.m;
18305 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18306 		/*
18307 		 * The trailing space changed, mbufs can grow
18308 		 * at the tail but they can't shrink from
18309 		 * it, KASSERT that. Adjust the orig_m_len to
18310 		 * compensate for this change.
18311 		 */
18312 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18313 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18314 			 m,
18315 			 rack,
18316 			 (intmax_t)M_TRAILINGROOM(m),
18317 			 rack->r_ctl.fsb.o_t_len,
18318 			 rack->r_ctl.fsb.o_m_len,
18319 			 m->m_len));
18320 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18321 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18322 	}
18323 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18324 		/*
18325 		 * Mbuf shrank, trimmed off the top by an ack, our
18326 		 * offset changes.
18327 		 */
18328 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18329 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18330 			 m, m->m_len,
18331 			 rack, rack->r_ctl.fsb.o_m_len,
18332 			 rack->r_ctl.fsb.off));
18333 
18334 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18335 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18336 		else
18337 			rack->r_ctl.fsb.off = 0;
18338 		rack->r_ctl.fsb.o_m_len = m->m_len;
18339 #ifdef INVARIANTS
18340 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18341 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
18342 		      rack, m);
18343 #endif
18344 	}
18345 	soff = rack->r_ctl.fsb.off;
18346 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18347 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18348 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18349 				 __FUNCTION__,
18350 				 rack, *plen, m, m->m_len));
18351 	/* Save off the right location before we copy and advance */
18352 	*s_soff = soff;
18353 	*s_mb = rack->r_ctl.fsb.m;
18354 	n = rack_fo_base_copym(m, soff, plen,
18355 			       &rack->r_ctl.fsb,
18356 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18357 	return (n);
18358 }
18359 
18360 /* Log the buffer level */
18361 static void
18362 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18363 		     int len, struct timeval *tv,
18364 		     uint32_t cts)
18365 {
18366 	uint32_t p_rate = 0, p_queue = 0, err = 0;
18367 	union tcp_log_stackspecific log;
18368 
18369 #ifdef RATELIMIT
18370 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18371 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
18372 #endif
18373 	memset(&log, 0, sizeof(log));
18374 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18375 	log.u_bbr.flex1 = p_rate;
18376 	log.u_bbr.flex2 = p_queue;
18377 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18378 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18379 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18380 	log.u_bbr.flex7 = 99;
18381 	log.u_bbr.flex8 = 0;
18382 	log.u_bbr.pkts_out = err;
18383 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
18384 	log.u_bbr.timeStamp = cts;
18385 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18386 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18387 		       len, &log, false, NULL, __func__, __LINE__, tv);
18388 
18389 }
18390 
18391 static uint32_t
18392 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18393 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18394 {
18395 	uint64_t lentime = 0;
18396 #ifdef RATELIMIT
18397 	uint32_t p_rate = 0, p_queue = 0, err;
18398 	union tcp_log_stackspecific log;
18399 	uint64_t bw;
18400 
18401 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18402 	/* Failed or queue is zero */
18403 	if (err || (p_queue == 0)) {
18404 		lentime = 0;
18405 		goto out;
18406 	}
18407 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18408 	if (err) {
18409 		lentime = 0;
18410 		goto out;
18411 	}
18412 	/*
18413 	 * If we reach here we have some bytes in
18414 	 * the queue. The number returned is a value
18415 	 * between 0 and 0xffff where ffff is full
18416 	 * and 0 is empty. So how best to make this into
18417 	 * something usable?
18418 	 *
18419 	 * The "safer" way is lets take the b/w gotten
18420 	 * from the query (which should be our b/w rate)
18421 	 * and pretend that a full send (our rc_pace_max_segs)
18422 	 * is outstanding. We factor it so its as if a full
18423 	 * number of our MSS segment is terms of full
18424 	 * ethernet segments are outstanding.
18425 	 */
18426 	bw = p_rate / 8;
18427 	if (bw) {
18428 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18429 		lentime *= ETHERNET_SEGMENT_SIZE;
18430 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18431 		lentime /= bw;
18432 	} else {
18433 		/* TSNH -- KASSERT? */
18434 		lentime = 0;
18435 	}
18436 out:
18437 	if (tcp_bblogging_on(tp)) {
18438 		memset(&log, 0, sizeof(log));
18439 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18440 		log.u_bbr.flex1 = p_rate;
18441 		log.u_bbr.flex2 = p_queue;
18442 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18443 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18444 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18445 		log.u_bbr.flex7 = 99;
18446 		log.u_bbr.flex8 = 0;
18447 		log.u_bbr.pkts_out = err;
18448 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
18449 		log.u_bbr.cur_del_rate = lentime;
18450 		log.u_bbr.timeStamp = cts;
18451 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18452 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18453 			       len, &log, false, NULL, __func__, __LINE__,tv);
18454 	}
18455 #endif
18456 	return ((uint32_t)lentime);
18457 }
18458 
18459 static int
18460 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18461 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18462 {
18463 	/*
18464 	 * Enter the fast retransmit path. We are given that a sched_pin is
18465 	 * in place (if accounting is compliled in) and the cycle count taken
18466 	 * at the entry is in the ts_val. The concept her is that the rsm
18467 	 * now holds the mbuf offsets and such so we can directly transmit
18468 	 * without a lot of overhead, the len field is already set for
18469 	 * us to prohibit us from sending too much (usually its 1MSS).
18470 	 */
18471 	struct ip *ip = NULL;
18472 	struct udphdr *udp = NULL;
18473 	struct tcphdr *th = NULL;
18474 	struct mbuf *m = NULL;
18475 	struct inpcb *inp;
18476 	uint8_t *cpto;
18477 	struct tcp_log_buffer *lgb;
18478 #ifdef TCP_ACCOUNTING
18479 	uint64_t crtsc;
18480 	int cnt_thru = 1;
18481 #endif
18482 	struct tcpopt to;
18483 	u_char opt[TCP_MAXOLEN];
18484 	uint32_t hdrlen, optlen;
18485 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18486 	uint16_t flags;
18487 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
18488 	uint32_t if_hw_tsomaxsegsize;
18489 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18490 
18491 #ifdef INET6
18492 	struct ip6_hdr *ip6 = NULL;
18493 
18494 	if (rack->r_is_v6) {
18495 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18496 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18497 	} else
18498 #endif				/* INET6 */
18499 	{
18500 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18501 		hdrlen = sizeof(struct tcpiphdr);
18502 	}
18503 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18504 		goto failed;
18505 	}
18506 	if (doing_tlp) {
18507 		/* Its a TLP add the flag, it may already be there but be sure */
18508 		rsm->r_flags |= RACK_TLP;
18509 	} else {
18510 		/* If it was a TLP it is not not on this retransmit */
18511 		rsm->r_flags &= ~RACK_TLP;
18512 	}
18513 	startseq = rsm->r_start;
18514 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18515 	inp = rack->rc_inp;
18516 	to.to_flags = 0;
18517 	flags = tcp_outflags[tp->t_state];
18518 	if (flags & (TH_SYN|TH_RST)) {
18519 		goto failed;
18520 	}
18521 	if (rsm->r_flags & RACK_HAS_FIN) {
18522 		/* We can't send a FIN here */
18523 		goto failed;
18524 	}
18525 	if (flags & TH_FIN) {
18526 		/* We never send a FIN */
18527 		flags &= ~TH_FIN;
18528 	}
18529 	if (tp->t_flags & TF_RCVD_TSTMP) {
18530 		to.to_tsval = ms_cts + tp->ts_offset;
18531 		to.to_tsecr = tp->ts_recent;
18532 		to.to_flags = TOF_TS;
18533 	}
18534 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18535 	/* TCP-MD5 (RFC2385). */
18536 	if (tp->t_flags & TF_SIGNATURE)
18537 		to.to_flags |= TOF_SIGNATURE;
18538 #endif
18539 	optlen = tcp_addoptions(&to, opt);
18540 	hdrlen += optlen;
18541 	udp = rack->r_ctl.fsb.udp;
18542 	if (udp)
18543 		hdrlen += sizeof(struct udphdr);
18544 	if (rack->r_ctl.rc_pace_max_segs)
18545 		max_val = rack->r_ctl.rc_pace_max_segs;
18546 	else if (rack->rc_user_set_max_segs)
18547 		max_val = rack->rc_user_set_max_segs * segsiz;
18548 	else
18549 		max_val = len;
18550 	if ((tp->t_flags & TF_TSO) &&
18551 	    V_tcp_do_tso &&
18552 	    (len > segsiz) &&
18553 	    (tp->t_port == 0))
18554 		tso = 1;
18555 #ifdef INET6
18556 	if (MHLEN < hdrlen + max_linkhdr)
18557 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18558 	else
18559 #endif
18560 		m = m_gethdr(M_NOWAIT, MT_DATA);
18561 	if (m == NULL)
18562 		goto failed;
18563 	m->m_data += max_linkhdr;
18564 	m->m_len = hdrlen;
18565 	th = rack->r_ctl.fsb.th;
18566 	/* Establish the len to send */
18567 	if (len > max_val)
18568 		len = max_val;
18569 	if ((tso) && (len + optlen > segsiz)) {
18570 		uint32_t if_hw_tsomax;
18571 		int32_t max_len;
18572 
18573 		/* extract TSO information */
18574 		if_hw_tsomax = tp->t_tsomax;
18575 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18576 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18577 		/*
18578 		 * Check if we should limit by maximum payload
18579 		 * length:
18580 		 */
18581 		if (if_hw_tsomax != 0) {
18582 			/* compute maximum TSO length */
18583 			max_len = (if_hw_tsomax - hdrlen -
18584 				   max_linkhdr);
18585 			if (max_len <= 0) {
18586 				goto failed;
18587 			} else if (len > max_len) {
18588 				len = max_len;
18589 			}
18590 		}
18591 		if (len <= segsiz) {
18592 			/*
18593 			 * In case there are too many small fragments don't
18594 			 * use TSO:
18595 			 */
18596 			tso = 0;
18597 		}
18598 	} else {
18599 		tso = 0;
18600 	}
18601 	if ((tso == 0) && (len > segsiz))
18602 		len = segsiz;
18603 	(void)tcp_get_usecs(tv);
18604 	if ((len == 0) ||
18605 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
18606 		goto failed;
18607 	}
18608 	th->th_seq = htonl(rsm->r_start);
18609 	th->th_ack = htonl(tp->rcv_nxt);
18610 	/*
18611 	 * The PUSH bit should only be applied
18612 	 * if the full retransmission is made. If
18613 	 * we are sending less than this is the
18614 	 * left hand edge and should not have
18615 	 * the PUSH bit.
18616 	 */
18617 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
18618 	    (len == (rsm->r_end - rsm->r_start)))
18619 		flags |= TH_PUSH;
18620 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18621 	if (th->th_win == 0) {
18622 		tp->t_sndzerowin++;
18623 		tp->t_flags |= TF_RXWIN0SENT;
18624 	} else
18625 		tp->t_flags &= ~TF_RXWIN0SENT;
18626 	if (rsm->r_flags & RACK_TLP) {
18627 		/*
18628 		 * TLP should not count in retran count, but
18629 		 * in its own bin
18630 		 */
18631 		counter_u64_add(rack_tlp_retran, 1);
18632 		counter_u64_add(rack_tlp_retran_bytes, len);
18633 	} else {
18634 		tp->t_sndrexmitpack++;
18635 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18636 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18637 	}
18638 #ifdef STATS
18639 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18640 				 len);
18641 #endif
18642 	if (rsm->m == NULL)
18643 		goto failed;
18644 	if (rsm->m &&
18645 	    ((rsm->orig_m_len != rsm->m->m_len) ||
18646 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18647 		/* Fix up the orig_m_len and possibly the mbuf offset */
18648 		rack_adjust_orig_mlen(rsm);
18649 	}
18650 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18651 	if (len <= segsiz) {
18652 		/*
18653 		 * Must have ran out of mbufs for the copy
18654 		 * shorten it to no longer need tso. Lets
18655 		 * not put on sendalot since we are low on
18656 		 * mbufs.
18657 		 */
18658 		tso = 0;
18659 	}
18660 	if ((m->m_next == NULL) || (len <= 0)){
18661 		goto failed;
18662 	}
18663 	if (udp) {
18664 		if (rack->r_is_v6)
18665 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
18666 		else
18667 			ulen = hdrlen + len - sizeof(struct ip);
18668 		udp->uh_ulen = htons(ulen);
18669 	}
18670 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18671 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18672 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18673 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
18674 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18675 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18676 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18677 #ifdef INET6
18678 		if (rack->r_is_v6) {
18679 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18680 		    ip6->ip6_flow |= htonl(ect << 20);
18681 		}
18682 		else
18683 #endif
18684 		{
18685 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
18686 		    ip->ip_tos |= ect;
18687 		}
18688 	}
18689 	if (rack->r_ctl.crte != NULL) {
18690 		/* See if we can send via the hw queue */
18691 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18692 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
18693 		if (slot == 0)
18694 			ip_sendflag = 0;
18695 	}
18696 	tcp_set_flags(th, flags);
18697 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18698 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18699 	if (to.to_flags & TOF_SIGNATURE) {
18700 		/*
18701 		 * Calculate MD5 signature and put it into the place
18702 		 * determined before.
18703 		 * NOTE: since TCP options buffer doesn't point into
18704 		 * mbuf's data, calculate offset and use it.
18705 		 */
18706 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18707 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18708 			/*
18709 			 * Do not send segment if the calculation of MD5
18710 			 * digest has failed.
18711 			 */
18712 			goto failed;
18713 		}
18714 	}
18715 #endif
18716 #ifdef INET6
18717 	if (rack->r_is_v6) {
18718 		if (tp->t_port) {
18719 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18720 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18721 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18722 			th->th_sum = htons(0);
18723 			UDPSTAT_INC(udps_opackets);
18724 		} else {
18725 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18726 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18727 			th->th_sum = in6_cksum_pseudo(ip6,
18728 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18729 						      0);
18730 		}
18731 	}
18732 #endif
18733 #if defined(INET6) && defined(INET)
18734 	else
18735 #endif
18736 #ifdef INET
18737 	{
18738 		if (tp->t_port) {
18739 			m->m_pkthdr.csum_flags = CSUM_UDP;
18740 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18741 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18742 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18743 			th->th_sum = htons(0);
18744 			UDPSTAT_INC(udps_opackets);
18745 		} else {
18746 			m->m_pkthdr.csum_flags = CSUM_TCP;
18747 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18748 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18749 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18750 									IPPROTO_TCP + len + optlen));
18751 		}
18752 		/* IP version must be set here for ipv4/ipv6 checking later */
18753 		KASSERT(ip->ip_v == IPVERSION,
18754 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18755 	}
18756 #endif
18757 	if (tso) {
18758 		/*
18759 		 * Here we use segsiz since we have no added options besides
18760 		 * any standard timestamp options (no DSACKs or SACKS are sent
18761 		 * via either fast-path).
18762 		 */
18763 		KASSERT(len > segsiz,
18764 			("%s: len <= tso_segsz tp:%p", __func__, tp));
18765 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18766 		m->m_pkthdr.tso_segsz = segsiz;
18767 	}
18768 #ifdef INET6
18769 	if (rack->r_is_v6) {
18770 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18771 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18772 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18773 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18774 		else
18775 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18776 	}
18777 #endif
18778 #if defined(INET) && defined(INET6)
18779 	else
18780 #endif
18781 #ifdef INET
18782 	{
18783 		ip->ip_len = htons(m->m_pkthdr.len);
18784 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18785 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18786 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18787 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18788 				ip->ip_off |= htons(IP_DF);
18789 			}
18790 		} else {
18791 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18792 		}
18793 	}
18794 #endif
18795 	if (doing_tlp == 0) {
18796 		/* Set we retransmitted */
18797 		rack->rc_gp_saw_rec = 1;
18798 	} else {
18799 		/* Its a TLP set ca or ss */
18800 		if (tp->snd_cwnd > tp->snd_ssthresh) {
18801 			/* Set we sent in CA */
18802 			rack->rc_gp_saw_ca = 1;
18803 		} else {
18804 			/* Set we sent in SS */
18805 			rack->rc_gp_saw_ss = 1;
18806 		}
18807 	}
18808 	/* Time to copy in our header */
18809 	cpto = mtod(m, uint8_t *);
18810 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18811 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18812 	if (optlen) {
18813 		bcopy(opt, th + 1, optlen);
18814 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18815 	} else {
18816 		th->th_off = sizeof(struct tcphdr) >> 2;
18817 	}
18818 	if (tcp_bblogging_on(rack->rc_tp)) {
18819 		union tcp_log_stackspecific log;
18820 
18821 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18822 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18823 			counter_u64_add(rack_collapsed_win_rxt, 1);
18824 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18825 		}
18826 		memset(&log, 0, sizeof(log));
18827 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18828 		if (rack->rack_no_prr)
18829 			log.u_bbr.flex1 = 0;
18830 		else
18831 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18832 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18833 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18834 		log.u_bbr.flex4 = max_val;
18835 		/* Save off the early/late values */
18836 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18837 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18838 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18839 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18840 		if (doing_tlp == 0)
18841 			log.u_bbr.flex8 = 1;
18842 		else
18843 			log.u_bbr.flex8 = 2;
18844 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18845 		log.u_bbr.flex7 = 55;
18846 		log.u_bbr.pkts_out = tp->t_maxseg;
18847 		log.u_bbr.timeStamp = cts;
18848 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18849 		if (rsm->r_rtr_cnt > 0) {
18850 			/*
18851 			 * When we have a retransmit we want to log the
18852 			 * burst at send and flight at send from before.
18853 			 */
18854 			log.u_bbr.flex5 = rsm->r_fas;
18855 			log.u_bbr.bbr_substate = rsm->r_bas;
18856 		} else {
18857 			/*
18858 			 * This is currently unlikely until we do the
18859 			 * packet pair probes but I will add it for completeness.
18860 			 */
18861 			log.u_bbr.flex5 = log.u_bbr.inflight;
18862 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18863 		}
18864 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18865 		log.u_bbr.delivered = 0;
18866 		log.u_bbr.rttProp = (uintptr_t)rsm;
18867 		log.u_bbr.delRate = rsm->r_flags;
18868 		log.u_bbr.delRate <<= 31;
18869 		log.u_bbr.delRate |= rack->r_must_retran;
18870 		log.u_bbr.delRate <<= 1;
18871 		log.u_bbr.delRate |= 1;
18872 		log.u_bbr.pkt_epoch = __LINE__;
18873 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18874 				     len, &log, false, NULL, __func__, __LINE__, tv);
18875 	} else
18876 		lgb = NULL;
18877 	if ((rack->r_ctl.crte != NULL) &&
18878 	    tcp_bblogging_on(tp)) {
18879 		rack_log_queue_level(tp, rack, len, tv, cts);
18880 	}
18881 #ifdef INET6
18882 	if (rack->r_is_v6) {
18883 		error = ip6_output(m, inp->in6p_outputopts,
18884 				   &inp->inp_route6,
18885 				   ip_sendflag, NULL, NULL, inp);
18886 	}
18887 	else
18888 #endif
18889 #ifdef INET
18890 	{
18891 		error = ip_output(m, NULL,
18892 				  &inp->inp_route,
18893 				  ip_sendflag, 0, inp);
18894 	}
18895 #endif
18896 	m = NULL;
18897 	if (lgb) {
18898 		lgb->tlb_errno = error;
18899 		lgb = NULL;
18900 	}
18901 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
18902 	tp->snd_nxt = tp->snd_max;
18903 	if (error) {
18904 		goto failed;
18905 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18906 		rack->rc_hw_nobuf = 0;
18907 		rack->r_ctl.rc_agg_delayed = 0;
18908 		rack->r_early = 0;
18909 		rack->r_late = 0;
18910 		rack->r_ctl.rc_agg_early = 0;
18911 	}
18912 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18913 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18914 	if (doing_tlp) {
18915 		rack->rc_tlp_in_progress = 1;
18916 		rack->r_ctl.rc_tlp_cnt_out++;
18917 	}
18918 	if (error == 0) {
18919 		counter_u64_add(rack_total_bytes, len);
18920 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18921 		if (doing_tlp) {
18922 			rack->rc_last_sent_tlp_past_cumack = 0;
18923 			rack->rc_last_sent_tlp_seq_valid = 1;
18924 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18925 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18926 		}
18927 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18928 			rack->r_ctl.rc_prr_sndcnt -= len;
18929 		else
18930 			rack->r_ctl.rc_prr_sndcnt = 0;
18931 	}
18932 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18933 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
18934 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18935 		rack->r_ctl.retran_during_recovery += len;
18936 	{
18937 		int idx;
18938 
18939 		idx = (len / segsiz) + 3;
18940 		if (idx >= TCP_MSS_ACCT_ATIMER)
18941 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18942 		else
18943 			counter_u64_add(rack_out_size[idx], 1);
18944 	}
18945 	if (tp->t_rtttime == 0) {
18946 		tp->t_rtttime = ticks;
18947 		tp->t_rtseq = startseq;
18948 		KMOD_TCPSTAT_INC(tcps_segstimed);
18949 	}
18950 	counter_u64_add(rack_fto_rsm_send, 1);
18951 	if (error && (error == ENOBUFS)) {
18952 		if (rack->r_ctl.crte != NULL) {
18953 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18954 			if (tcp_bblogging_on(rack->rc_tp))
18955 				rack_log_queue_level(tp, rack, len, tv, cts);
18956 		} else
18957 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18958 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18959 		if (rack->rc_enobuf < 0x7f)
18960 			rack->rc_enobuf++;
18961 		if (slot < (10 * HPTS_USEC_IN_MSEC))
18962 			slot = 10 * HPTS_USEC_IN_MSEC;
18963 		if (rack->r_ctl.crte != NULL) {
18964 			counter_u64_add(rack_saw_enobuf_hw, 1);
18965 			tcp_rl_log_enobuf(rack->r_ctl.crte);
18966 		}
18967 		counter_u64_add(rack_saw_enobuf, 1);
18968 	} else {
18969 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
18970 	}
18971 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
18972 #ifdef TCP_ACCOUNTING
18973 	crtsc = get_cyclecount();
18974 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18975 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
18976 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18977 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
18978 	}
18979 	sched_unpin();
18980 #endif
18981 	return (0);
18982 failed:
18983 	if (m)
18984 		m_free(m);
18985 	return (-1);
18986 }
18987 
18988 static void
18989 rack_sndbuf_autoscale(struct tcp_rack *rack)
18990 {
18991 	/*
18992 	 * Automatic sizing of send socket buffer.  Often the send buffer
18993 	 * size is not optimally adjusted to the actual network conditions
18994 	 * at hand (delay bandwidth product).  Setting the buffer size too
18995 	 * small limits throughput on links with high bandwidth and high
18996 	 * delay (eg. trans-continental/oceanic links).  Setting the
18997 	 * buffer size too big consumes too much real kernel memory,
18998 	 * especially with many connections on busy servers.
18999 	 *
19000 	 * The criteria to step up the send buffer one notch are:
19001 	 *  1. receive window of remote host is larger than send buffer
19002 	 *     (with a fudge factor of 5/4th);
19003 	 *  2. send buffer is filled to 7/8th with data (so we actually
19004 	 *     have data to make use of it);
19005 	 *  3. send buffer fill has not hit maximal automatic size;
19006 	 *  4. our send window (slow start and cogestion controlled) is
19007 	 *     larger than sent but unacknowledged data in send buffer.
19008 	 *
19009 	 * Note that the rack version moves things much faster since
19010 	 * we want to avoid hitting cache lines in the rack_fast_output()
19011 	 * path so this is called much less often and thus moves
19012 	 * the SB forward by a percentage.
19013 	 */
19014 	struct socket *so;
19015 	struct tcpcb *tp;
19016 	uint32_t sendwin, scaleup;
19017 
19018 	tp = rack->rc_tp;
19019 	so = rack->rc_inp->inp_socket;
19020 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19021 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19022 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19023 		    sbused(&so->so_snd) >=
19024 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19025 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19026 		    sendwin >= (sbused(&so->so_snd) -
19027 		    (tp->snd_max - tp->snd_una))) {
19028 			if (rack_autosndbuf_inc)
19029 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19030 			else
19031 				scaleup = V_tcp_autosndbuf_inc;
19032 			if (scaleup < V_tcp_autosndbuf_inc)
19033 				scaleup = V_tcp_autosndbuf_inc;
19034 			scaleup += so->so_snd.sb_hiwat;
19035 			if (scaleup > V_tcp_autosndbuf_max)
19036 				scaleup = V_tcp_autosndbuf_max;
19037 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19038 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19039 		}
19040 	}
19041 }
19042 
19043 static int
19044 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19045 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long *tot_len, int *send_err, int line)
19046 {
19047 	/*
19048 	 * Enter to do fast output. We are given that the sched_pin is
19049 	 * in place (if accounting is compiled in) and the cycle count taken
19050 	 * at entry is in place in ts_val. The idea here is that
19051 	 * we know how many more bytes needs to be sent (presumably either
19052 	 * during pacing or to fill the cwnd and that was greater than
19053 	 * the max-burst). We have how much to send and all the info we
19054 	 * need to just send.
19055 	 */
19056 #ifdef INET
19057 	struct ip *ip = NULL;
19058 #endif
19059 	struct udphdr *udp = NULL;
19060 	struct tcphdr *th = NULL;
19061 	struct mbuf *m, *s_mb;
19062 	struct inpcb *inp;
19063 	uint8_t *cpto;
19064 	struct tcp_log_buffer *lgb;
19065 #ifdef TCP_ACCOUNTING
19066 	uint64_t crtsc;
19067 #endif
19068 	struct tcpopt to;
19069 	u_char opt[TCP_MAXOLEN];
19070 	uint32_t hdrlen, optlen;
19071 #ifdef TCP_ACCOUNTING
19072 	int cnt_thru = 1;
19073 #endif
19074 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19075 	uint16_t flags;
19076 	uint32_t s_soff;
19077 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19078 	uint32_t if_hw_tsomaxsegsize;
19079 	uint32_t add_flag = RACK_SENT_FP;
19080 #ifdef INET6
19081 	struct ip6_hdr *ip6 = NULL;
19082 
19083 	if (rack->r_is_v6) {
19084 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19085 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19086 	} else
19087 #endif				/* INET6 */
19088 	{
19089 #ifdef INET
19090 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19091 		hdrlen = sizeof(struct tcpiphdr);
19092 #endif
19093 	}
19094 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19095 		m = NULL;
19096 		goto failed;
19097 	}
19098 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19099 	startseq = tp->snd_max;
19100 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19101 	inp = rack->rc_inp;
19102 	len = rack->r_ctl.fsb.left_to_send;
19103 	to.to_flags = 0;
19104 	flags = rack->r_ctl.fsb.tcp_flags;
19105 	if (tp->t_flags & TF_RCVD_TSTMP) {
19106 		to.to_tsval = ms_cts + tp->ts_offset;
19107 		to.to_tsecr = tp->ts_recent;
19108 		to.to_flags = TOF_TS;
19109 	}
19110 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19111 	/* TCP-MD5 (RFC2385). */
19112 	if (tp->t_flags & TF_SIGNATURE)
19113 		to.to_flags |= TOF_SIGNATURE;
19114 #endif
19115 	optlen = tcp_addoptions(&to, opt);
19116 	hdrlen += optlen;
19117 	udp = rack->r_ctl.fsb.udp;
19118 	if (udp)
19119 		hdrlen += sizeof(struct udphdr);
19120 	if (rack->r_ctl.rc_pace_max_segs)
19121 		max_val = rack->r_ctl.rc_pace_max_segs;
19122 	else if (rack->rc_user_set_max_segs)
19123 		max_val = rack->rc_user_set_max_segs * segsiz;
19124 	else
19125 		max_val = len;
19126 	if ((tp->t_flags & TF_TSO) &&
19127 	    V_tcp_do_tso &&
19128 	    (len > segsiz) &&
19129 	    (tp->t_port == 0))
19130 		tso = 1;
19131 again:
19132 #ifdef INET6
19133 	if (MHLEN < hdrlen + max_linkhdr)
19134 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19135 	else
19136 #endif
19137 		m = m_gethdr(M_NOWAIT, MT_DATA);
19138 	if (m == NULL)
19139 		goto failed;
19140 	m->m_data += max_linkhdr;
19141 	m->m_len = hdrlen;
19142 	th = rack->r_ctl.fsb.th;
19143 	/* Establish the len to send */
19144 	if (len > max_val)
19145 		len = max_val;
19146 	if ((tso) && (len + optlen > segsiz)) {
19147 		uint32_t if_hw_tsomax;
19148 		int32_t max_len;
19149 
19150 		/* extract TSO information */
19151 		if_hw_tsomax = tp->t_tsomax;
19152 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19153 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19154 		/*
19155 		 * Check if we should limit by maximum payload
19156 		 * length:
19157 		 */
19158 		if (if_hw_tsomax != 0) {
19159 			/* compute maximum TSO length */
19160 			max_len = (if_hw_tsomax - hdrlen -
19161 				   max_linkhdr);
19162 			if (max_len <= 0) {
19163 				goto failed;
19164 			} else if (len > max_len) {
19165 				len = max_len;
19166 			}
19167 		}
19168 		if (len <= segsiz) {
19169 			/*
19170 			 * In case there are too many small fragments don't
19171 			 * use TSO:
19172 			 */
19173 			tso = 0;
19174 		}
19175 	} else {
19176 		tso = 0;
19177 	}
19178 	if ((tso == 0) && (len > segsiz))
19179 		len = segsiz;
19180 	(void)tcp_get_usecs(tv);
19181 	if ((len == 0) ||
19182 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19183 		goto failed;
19184 	}
19185 	sb_offset = tp->snd_max - tp->snd_una;
19186 	th->th_seq = htonl(tp->snd_max);
19187 	th->th_ack = htonl(tp->rcv_nxt);
19188 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19189 	if (th->th_win == 0) {
19190 		tp->t_sndzerowin++;
19191 		tp->t_flags |= TF_RXWIN0SENT;
19192 	} else
19193 		tp->t_flags &= ~TF_RXWIN0SENT;
19194 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
19195 	KMOD_TCPSTAT_INC(tcps_sndpack);
19196 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19197 #ifdef STATS
19198 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19199 				 len);
19200 #endif
19201 	if (rack->r_ctl.fsb.m == NULL)
19202 		goto failed;
19203 
19204 	/* s_mb and s_soff are saved for rack_log_output */
19205 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19206 				    &s_mb, &s_soff);
19207 	if (len <= segsiz) {
19208 		/*
19209 		 * Must have ran out of mbufs for the copy
19210 		 * shorten it to no longer need tso. Lets
19211 		 * not put on sendalot since we are low on
19212 		 * mbufs.
19213 		 */
19214 		tso = 0;
19215 	}
19216 	if (rack->r_ctl.fsb.rfo_apply_push &&
19217 	    (len == rack->r_ctl.fsb.left_to_send)) {
19218 		flags |= TH_PUSH;
19219 		add_flag |= RACK_HAD_PUSH;
19220 	}
19221 	if ((m->m_next == NULL) || (len <= 0)){
19222 		goto failed;
19223 	}
19224 	if (udp) {
19225 		if (rack->r_is_v6)
19226 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19227 		else
19228 			ulen = hdrlen + len - sizeof(struct ip);
19229 		udp->uh_ulen = htons(ulen);
19230 	}
19231 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19232 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19233 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19234 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
19235 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19236 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19237 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19238 #ifdef INET6
19239 		if (rack->r_is_v6) {
19240 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19241 			ip6->ip6_flow |= htonl(ect << 20);
19242 		}
19243 		else
19244 #endif
19245 		{
19246 #ifdef INET
19247 			ip->ip_tos &= ~IPTOS_ECN_MASK;
19248 			ip->ip_tos |= ect;
19249 #endif
19250 		}
19251 	}
19252 	tcp_set_flags(th, flags);
19253 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19254 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19255 	if (to.to_flags & TOF_SIGNATURE) {
19256 		/*
19257 		 * Calculate MD5 signature and put it into the place
19258 		 * determined before.
19259 		 * NOTE: since TCP options buffer doesn't point into
19260 		 * mbuf's data, calculate offset and use it.
19261 		 */
19262 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19263 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19264 			/*
19265 			 * Do not send segment if the calculation of MD5
19266 			 * digest has failed.
19267 			 */
19268 			goto failed;
19269 		}
19270 	}
19271 #endif
19272 #ifdef INET6
19273 	if (rack->r_is_v6) {
19274 		if (tp->t_port) {
19275 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19276 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19277 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19278 			th->th_sum = htons(0);
19279 			UDPSTAT_INC(udps_opackets);
19280 		} else {
19281 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19282 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19283 			th->th_sum = in6_cksum_pseudo(ip6,
19284 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19285 						      0);
19286 		}
19287 	}
19288 #endif
19289 #if defined(INET6) && defined(INET)
19290 	else
19291 #endif
19292 #ifdef INET
19293 	{
19294 		if (tp->t_port) {
19295 			m->m_pkthdr.csum_flags = CSUM_UDP;
19296 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19297 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19298 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19299 			th->th_sum = htons(0);
19300 			UDPSTAT_INC(udps_opackets);
19301 		} else {
19302 			m->m_pkthdr.csum_flags = CSUM_TCP;
19303 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19304 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19305 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19306 									IPPROTO_TCP + len + optlen));
19307 		}
19308 		/* IP version must be set here for ipv4/ipv6 checking later */
19309 		KASSERT(ip->ip_v == IPVERSION,
19310 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19311 	}
19312 #endif
19313 	if (tso) {
19314 		/*
19315 		 * Here we use segsiz since we have no added options besides
19316 		 * any standard timestamp options (no DSACKs or SACKS are sent
19317 		 * via either fast-path).
19318 		 */
19319 		KASSERT(len > segsiz,
19320 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19321 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19322 		m->m_pkthdr.tso_segsz = segsiz;
19323 	}
19324 #ifdef INET6
19325 	if (rack->r_is_v6) {
19326 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19327 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19328 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19329 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19330 		else
19331 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19332 	}
19333 #endif
19334 #if defined(INET) && defined(INET6)
19335 	else
19336 #endif
19337 #ifdef INET
19338 	{
19339 		ip->ip_len = htons(m->m_pkthdr.len);
19340 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19341 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19342 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19343 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19344 				ip->ip_off |= htons(IP_DF);
19345 			}
19346 		} else {
19347 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19348 		}
19349 	}
19350 #endif
19351 	if (tp->snd_cwnd > tp->snd_ssthresh) {
19352 		/* Set we sent in CA */
19353 		rack->rc_gp_saw_ca = 1;
19354 	} else {
19355 		/* Set we sent in SS */
19356 		rack->rc_gp_saw_ss = 1;
19357 	}
19358 	/* Time to copy in our header */
19359 	cpto = mtod(m, uint8_t *);
19360 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19361 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19362 	if (optlen) {
19363 		bcopy(opt, th + 1, optlen);
19364 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19365 	} else {
19366 		th->th_off = sizeof(struct tcphdr) >> 2;
19367 	}
19368 	if ((rack->r_ctl.crte != NULL) &&
19369 	    tcp_bblogging_on(tp)) {
19370 		rack_log_queue_level(tp, rack, len, tv, cts);
19371 	}
19372 	if (tcp_bblogging_on(rack->rc_tp)) {
19373 		union tcp_log_stackspecific log;
19374 
19375 		memset(&log, 0, sizeof(log));
19376 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19377 		if (rack->rack_no_prr)
19378 			log.u_bbr.flex1 = 0;
19379 		else
19380 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19381 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19382 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19383 		log.u_bbr.flex4 = max_val;
19384 		/* Save off the early/late values */
19385 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19386 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19387 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19388 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19389 		log.u_bbr.flex8 = 0;
19390 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19391 		log.u_bbr.flex7 = 44;
19392 		log.u_bbr.pkts_out = tp->t_maxseg;
19393 		log.u_bbr.timeStamp = cts;
19394 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19395 		log.u_bbr.flex5 = log.u_bbr.inflight;
19396 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19397 		log.u_bbr.delivered = rack->r_ctl.fsb.left_to_send;
19398 		log.u_bbr.rttProp = 0;
19399 		log.u_bbr.delRate = rack->r_must_retran;
19400 		log.u_bbr.delRate <<= 1;
19401 		log.u_bbr.pkt_epoch = line;
19402 		/* For fast output no retrans so just inflight and how many mss we send */
19403 		log.u_bbr.flex5 = log.u_bbr.inflight;
19404 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19405 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19406 				     len, &log, false, NULL, __func__, __LINE__, tv);
19407 	} else
19408 		lgb = NULL;
19409 #ifdef INET6
19410 	if (rack->r_is_v6) {
19411 		error = ip6_output(m, inp->in6p_outputopts,
19412 				   &inp->inp_route6,
19413 				   0, NULL, NULL, inp);
19414 	}
19415 #endif
19416 #if defined(INET) && defined(INET6)
19417 	else
19418 #endif
19419 #ifdef INET
19420 	{
19421 		error = ip_output(m, NULL,
19422 				  &inp->inp_route,
19423 				  0, 0, inp);
19424 	}
19425 #endif
19426 	if (lgb) {
19427 		lgb->tlb_errno = error;
19428 		lgb = NULL;
19429 	}
19430 	if (error) {
19431 		*send_err = error;
19432 		m = NULL;
19433 		goto failed;
19434 	} else if (rack->rc_hw_nobuf) {
19435 		rack->rc_hw_nobuf = 0;
19436 		rack->r_ctl.rc_agg_delayed = 0;
19437 		rack->r_early = 0;
19438 		rack->r_late = 0;
19439 		rack->r_ctl.rc_agg_early = 0;
19440 	}
19441 	if ((error == 0) && (rack->lt_bw_up == 0)) {
19442 		/* Unlikely */
19443 		rack->r_ctl.lt_timemark = tcp_tv_to_lusec(tv);
19444 		rack->r_ctl.lt_seq = tp->snd_una;
19445 		rack->lt_bw_up = 1;
19446 	} else if ((error == 0) &&
19447 		   (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
19448 		/*
19449 		 * Need to record what we have since we are
19450 		 * approaching seq wrap.
19451 		 */
19452 		struct timeval tv;
19453 		uint64_t tmark;
19454 
19455 		rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
19456 		rack->r_ctl.lt_seq = tp->snd_una;
19457 		tmark = tcp_get_u64_usecs(&tv);
19458 		if (tmark > rack->r_ctl.lt_timemark) {
19459 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
19460 			rack->r_ctl.lt_timemark = tmark;
19461 		}
19462 	}
19463 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19464 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19465 	if (tp->snd_una == tp->snd_max) {
19466 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19467 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19468 		tp->t_acktime = ticks;
19469 	}
19470 	counter_u64_add(rack_total_bytes, len);
19471 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19472 
19473 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19474 	*tot_len += len;
19475 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
19476 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19477 	tp->snd_max += len;
19478 	tp->snd_nxt = tp->snd_max;
19479 	if (rack->rc_new_rnd_needed) {
19480 		rack_new_round_starts(tp, rack, tp->snd_max);
19481 	}
19482 	{
19483 		int idx;
19484 
19485 		idx = (len / segsiz) + 3;
19486 		if (idx >= TCP_MSS_ACCT_ATIMER)
19487 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19488 		else
19489 			counter_u64_add(rack_out_size[idx], 1);
19490 	}
19491 	if (len <= rack->r_ctl.fsb.left_to_send)
19492 		rack->r_ctl.fsb.left_to_send -= len;
19493 	else
19494 		rack->r_ctl.fsb.left_to_send = 0;
19495 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
19496 		rack->r_fast_output = 0;
19497 		rack->r_ctl.fsb.left_to_send = 0;
19498 		/* At the end of fast_output scale up the sb */
19499 		SOCK_SENDBUF_LOCK(rack->rc_inp->inp_socket);
19500 		rack_sndbuf_autoscale(rack);
19501 		SOCK_SENDBUF_UNLOCK(rack->rc_inp->inp_socket);
19502 	}
19503 	if (tp->t_rtttime == 0) {
19504 		tp->t_rtttime = ticks;
19505 		tp->t_rtseq = startseq;
19506 		KMOD_TCPSTAT_INC(tcps_segstimed);
19507 	}
19508 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19509 	    (max_val > len) &&
19510 	    (*tot_len < rack->r_ctl.rc_pace_max_segs) &&
19511 	    (tso == 0)) {
19512 		max_val -= len;
19513 		len = segsiz;
19514 		th = rack->r_ctl.fsb.th;
19515 #ifdef TCP_ACCOUNTING
19516 		cnt_thru++;
19517 #endif
19518 		goto again;
19519 	}
19520 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19521 	counter_u64_add(rack_fto_send, 1);
19522 	slot = rack_get_pacing_delay(rack, tp, *tot_len, NULL, segsiz, __LINE__);
19523 	rack_start_hpts_timer(rack, tp, cts, slot, *tot_len, 0);
19524 #ifdef TCP_ACCOUNTING
19525 	crtsc = get_cyclecount();
19526 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19527 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19528 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19529 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((*tot_len + segsiz - 1) / segsiz);
19530 	}
19531 	sched_unpin();
19532 #endif
19533 	return (0);
19534 failed:
19535 	if (m)
19536 		m_free(m);
19537 	rack->r_fast_output = 0;
19538 	return (-1);
19539 }
19540 
19541 static inline void
19542 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19543 		       struct sockbuf *sb,
19544 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19545 		       bool hw_tls,
19546 		       uint16_t flags)
19547 {
19548 	rack->r_fast_output = 1;
19549 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19550 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19551 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19552 	rack->r_ctl.fsb.tcp_flags = flags;
19553 	rack->r_ctl.fsb.left_to_send = orig_len - len;
19554 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19555 		/* Less than a full sized pace, lets not  */
19556 		rack->r_fast_output = 0;
19557 		return;
19558 	} else {
19559 		/* Round down to the nearest pace_max_seg */
19560 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19561 	}
19562 	if (hw_tls)
19563 		rack->r_ctl.fsb.hw_tls = 1;
19564 	else
19565 		rack->r_ctl.fsb.hw_tls = 0;
19566 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19567 		("rack:%p left_to_send:%u sbavail:%u out:%u",
19568 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19569 		 (tp->snd_max - tp->snd_una)));
19570 	if (rack->r_ctl.fsb.left_to_send < segsiz)
19571 		rack->r_fast_output = 0;
19572 	else {
19573 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19574 			rack->r_ctl.fsb.rfo_apply_push = 1;
19575 		else
19576 			rack->r_ctl.fsb.rfo_apply_push = 0;
19577 	}
19578 }
19579 
19580 static uint32_t
19581 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19582 {
19583 	uint64_t min_time;
19584 	uint32_t maxlen;
19585 
19586 	min_time = (uint64_t)get_hpts_min_sleep_time();
19587 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19588 	maxlen = roundup(maxlen, segsiz);
19589 	return (maxlen);
19590 }
19591 
19592 static struct rack_sendmap *
19593 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19594 {
19595 	struct rack_sendmap *rsm = NULL;
19596 	int thresh;
19597 
19598 restart:
19599 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19600 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19601 		/* Nothing, strange turn off validity  */
19602 		rack->r_collapse_point_valid = 0;
19603 		return (NULL);
19604 	}
19605 	/* Can we send it yet? */
19606 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19607 		/*
19608 		 * Receiver window has not grown enough for
19609 		 * the segment to be put on the wire.
19610 		 */
19611 		return (NULL);
19612 	}
19613 	if (rsm->r_flags & RACK_ACKED) {
19614 		/*
19615 		 * It has been sacked, lets move to the
19616 		 * next one if possible.
19617 		 */
19618 		rack->r_ctl.last_collapse_point = rsm->r_end;
19619 		/* Are we done? */
19620 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19621 			    rack->r_ctl.high_collapse_point)) {
19622 			rack->r_collapse_point_valid = 0;
19623 			return (NULL);
19624 		}
19625 		goto restart;
19626 	}
19627 	/* Now has it been long enough ? */
19628 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
19629 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19630 		rack_log_collapse(rack, rsm->r_start,
19631 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19632 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
19633 		return (rsm);
19634 	}
19635 	/* Not enough time */
19636 	rack_log_collapse(rack, rsm->r_start,
19637 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19638 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
19639 	return (NULL);
19640 }
19641 
19642 static inline void
19643 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19644 {
19645 	if ((rack->full_size_rxt == 0) &&
19646 	    (rack->shape_rxt_to_pacing_min == 0) &&
19647 	    (*len >= segsiz)) {
19648 		*len = segsiz;
19649 	} else if (rack->shape_rxt_to_pacing_min &&
19650 		 rack->gp_ready) {
19651 		/* We use pacing min as shaping len req */
19652 		uint32_t maxlen;
19653 
19654 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19655 		if (*len > maxlen)
19656 			*len = maxlen;
19657 	} else {
19658 		/*
19659 		 * The else is full_size_rxt is on so send it all
19660 		 * note we do need to check this for exceeding
19661 		 * our max segment size due to the fact that
19662 		 * we do sometimes merge chunks together i.e.
19663 		 * we cannot just assume that we will never have
19664 		 * a chunk greater than pace_max_seg
19665 		 */
19666 		if (*len > pace_max_seg)
19667 			*len = pace_max_seg;
19668 	}
19669 }
19670 
19671 static int
19672 rack_output(struct tcpcb *tp)
19673 {
19674 	struct socket *so;
19675 	uint32_t recwin;
19676 	uint32_t sb_offset, s_moff = 0;
19677 	int32_t len, error = 0;
19678 	uint16_t flags;
19679 	struct mbuf *m, *s_mb = NULL;
19680 	struct mbuf *mb;
19681 	uint32_t if_hw_tsomaxsegcount = 0;
19682 	uint32_t if_hw_tsomaxsegsize;
19683 	int32_t segsiz, minseg;
19684 	long tot_len_this_send = 0;
19685 #ifdef INET
19686 	struct ip *ip = NULL;
19687 #endif
19688 	struct udphdr *udp = NULL;
19689 	struct tcp_rack *rack;
19690 	struct tcphdr *th;
19691 	uint8_t pass = 0;
19692 	uint8_t mark = 0;
19693 	uint8_t check_done = 0;
19694 	uint8_t wanted_cookie = 0;
19695 	u_char opt[TCP_MAXOLEN];
19696 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
19697 	uint32_t rack_seq;
19698 
19699 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19700 	unsigned ipsec_optlen = 0;
19701 
19702 #endif
19703 	int32_t idle, sendalot;
19704 	uint32_t tot_idle;
19705 	int32_t sub_from_prr = 0;
19706 	volatile int32_t sack_rxmit;
19707 	struct rack_sendmap *rsm = NULL;
19708 	int32_t tso, mtu;
19709 	struct tcpopt to;
19710 	int32_t slot = 0;
19711 	int32_t sup_rack = 0;
19712 	uint32_t cts, ms_cts, delayed, early;
19713 	uint32_t add_flag = RACK_SENT_SP;
19714 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19715 	uint8_t doing_tlp = 0;
19716 	uint32_t cwnd_to_use, pace_max_seg;
19717 	int32_t do_a_prefetch = 0;
19718 	int32_t prefetch_rsm = 0;
19719 	int32_t orig_len = 0;
19720 	struct timeval tv;
19721 	int32_t prefetch_so_done = 0;
19722 	struct tcp_log_buffer *lgb;
19723 	struct inpcb *inp = tptoinpcb(tp);
19724 	struct sockbuf *sb;
19725 	uint64_t ts_val = 0;
19726 #ifdef TCP_ACCOUNTING
19727 	uint64_t crtsc;
19728 #endif
19729 #ifdef INET6
19730 	struct ip6_hdr *ip6 = NULL;
19731 	int32_t isipv6;
19732 #endif
19733 	bool hpts_calling, hw_tls = false;
19734 
19735 	NET_EPOCH_ASSERT();
19736 	INP_WLOCK_ASSERT(inp);
19737 
19738 	/* setup and take the cache hits here */
19739 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19740 #ifdef TCP_ACCOUNTING
19741 	sched_pin();
19742 	ts_val = get_cyclecount();
19743 #endif
19744 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19745 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
19746 #ifdef TCP_OFFLOAD
19747 	if (tp->t_flags & TF_TOE) {
19748 #ifdef TCP_ACCOUNTING
19749 		sched_unpin();
19750 #endif
19751 		return (tcp_offload_output(tp));
19752 	}
19753 #endif
19754 	if (rack->rack_deferred_inited == 0) {
19755 		/*
19756 		 * If we are the connecting socket we will
19757 		 * hit rack_init() when no sequence numbers
19758 		 * are setup. This makes it so we must defer
19759 		 * some initialization. Call that now.
19760 		 */
19761 		rack_deferred_init(tp, rack);
19762 	}
19763 	/*
19764 	 * For TFO connections in SYN_RECEIVED, only allow the initial
19765 	 * SYN|ACK and those sent by the retransmit timer.
19766 	 */
19767 	if ((tp->t_flags & TF_FASTOPEN) &&
19768 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
19769 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
19770 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
19771 #ifdef TCP_ACCOUNTING
19772 		sched_unpin();
19773 #endif
19774 		return (0);
19775 	}
19776 #ifdef INET6
19777 	if (rack->r_state) {
19778 		/* Use the cache line loaded if possible */
19779 		isipv6 = rack->r_is_v6;
19780 	} else {
19781 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19782 	}
19783 #endif
19784 	early = 0;
19785 	cts = tcp_get_usecs(&tv);
19786 	ms_cts = tcp_tv_to_msec(&tv);
19787 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19788 	    tcp_in_hpts(rack->rc_tp)) {
19789 		/*
19790 		 * We are on the hpts for some timer but not hptsi output.
19791 		 * Remove from the hpts unconditionally.
19792 		 */
19793 		rack_timer_cancel(tp, rack, cts, __LINE__);
19794 	}
19795 	/* Are we pacing and late? */
19796 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19797 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19798 		/* We are delayed */
19799 		delayed = cts - rack->r_ctl.rc_last_output_to;
19800 	} else {
19801 		delayed = 0;
19802 	}
19803 	/* Do the timers, which may override the pacer */
19804 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19805 		int retval;
19806 
19807 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
19808 					     &doing_tlp);
19809 		if (retval != 0) {
19810 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19811 #ifdef TCP_ACCOUNTING
19812 			sched_unpin();
19813 #endif
19814 			/*
19815 			 * If timers want tcp_drop(), then pass error out,
19816 			 * otherwise suppress it.
19817 			 */
19818 			return (retval < 0 ? retval : 0);
19819 		}
19820 	}
19821 	if (rack->rc_in_persist) {
19822 		if (tcp_in_hpts(rack->rc_tp) == 0) {
19823 			/* Timer is not running */
19824 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19825 		}
19826 #ifdef TCP_ACCOUNTING
19827 		sched_unpin();
19828 #endif
19829 		return (0);
19830 	}
19831 	if ((rack->rc_ack_required == 1) &&
19832 	    (rack->r_timer_override == 0)){
19833 		/* A timeout occurred and no ack has arrived */
19834 		if (tcp_in_hpts(rack->rc_tp) == 0) {
19835 			/* Timer is not running */
19836 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19837 		}
19838 #ifdef TCP_ACCOUNTING
19839 		sched_unpin();
19840 #endif
19841 		return (0);
19842 	}
19843 	if ((rack->r_timer_override) ||
19844 	    (rack->rc_ack_can_sendout_data) ||
19845 	    (delayed) ||
19846 	    (tp->t_state < TCPS_ESTABLISHED)) {
19847 		rack->rc_ack_can_sendout_data = 0;
19848 		if (tcp_in_hpts(rack->rc_tp))
19849 			tcp_hpts_remove(rack->rc_tp);
19850 	} else if (tcp_in_hpts(rack->rc_tp)) {
19851 		/*
19852 		 * On the hpts you can't pass even if ACKNOW is on, we will
19853 		 * when the hpts fires.
19854 		 */
19855 #ifdef TCP_ACCOUNTING
19856 		crtsc = get_cyclecount();
19857 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19858 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19859 			tp->tcp_cnt_counters[SND_BLOCKED]++;
19860 		}
19861 		sched_unpin();
19862 #endif
19863 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19864 		return (0);
19865 	}
19866 	/* Finish out both pacing early and late accounting */
19867 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19868 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19869 		early = rack->r_ctl.rc_last_output_to - cts;
19870 	} else
19871 		early = 0;
19872 	if (delayed && (rack->rc_always_pace == 1)) {
19873 		rack->r_ctl.rc_agg_delayed += delayed;
19874 		rack->r_late = 1;
19875 	} else if (early && (rack->rc_always_pace == 1)) {
19876 		rack->r_ctl.rc_agg_early += early;
19877 		rack->r_early = 1;
19878 	} else if (rack->rc_always_pace == 0) {
19879 		/* Non-paced we are not late */
19880 		rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
19881 		rack->r_early = rack->r_late = 0;
19882 	}
19883 	/* Now that early/late accounting is done turn off the flag */
19884 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19885 	rack->r_wanted_output = 0;
19886 	rack->r_timer_override = 0;
19887 	if ((tp->t_state != rack->r_state) &&
19888 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
19889 		rack_set_state(tp, rack);
19890 	}
19891 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19892 	minseg = segsiz;
19893 	if (rack->r_ctl.rc_pace_max_segs == 0)
19894 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
19895 	else
19896 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
19897 	if ((rack->r_fast_output) &&
19898 	    (doing_tlp == 0) &&
19899 	    (tp->rcv_numsacks == 0)) {
19900 		int ret;
19901 
19902 		error = 0;
19903 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, &tot_len_this_send, &error, __LINE__);
19904 		if (ret > 0)
19905 			return(ret);
19906 		else if (error) {
19907 			inp = rack->rc_inp;
19908 			so = inp->inp_socket;
19909 			sb = &so->so_snd;
19910 			goto nomore;
19911 		} else {
19912 			/* Return == 0, if there is more we can send tot_len wise fall through and send */
19913 			if (tot_len_this_send >= pace_max_seg)
19914 				return (ret);
19915 #ifdef TCP_ACCOUNTING
19916 			/* We need to re-pin since fast_output un-pined */
19917 			sched_pin();
19918 			ts_val = get_cyclecount();
19919 #endif
19920 			/* Fall back out so we can send any more that may bring us to pace_max_seg */
19921 		}
19922 	}
19923 	inp = rack->rc_inp;
19924 	/*
19925 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
19926 	 * only allow the initial SYN or SYN|ACK and those sent
19927 	 * by the retransmit timer.
19928 	 */
19929 	if ((tp->t_flags & TF_FASTOPEN) &&
19930 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
19931 	     (tp->t_state == TCPS_SYN_SENT)) &&
19932 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19933 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
19934 		rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19935 #ifdef TCP_ACCOUNTING
19936 		sched_unpin();
19937 #endif
19938 		return (0);
19939 	}
19940 	/*
19941 	 * Determine length of data that should be transmitted, and flags
19942 	 * that will be used. If there is some data or critical controls
19943 	 * (SYN, RST) to send, then transmit; otherwise, investigate
19944 	 * further.
19945 	 */
19946 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19947 	if (tp->t_idle_reduce) {
19948 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19949 			rack_cc_after_idle(rack, tp);
19950 	}
19951 	tp->t_flags &= ~TF_LASTIDLE;
19952 	if (idle) {
19953 		if (tp->t_flags & TF_MORETOCOME) {
19954 			tp->t_flags |= TF_LASTIDLE;
19955 			idle = 0;
19956 		}
19957 	}
19958 	if ((tp->snd_una == tp->snd_max) &&
19959 	    rack->r_ctl.rc_went_idle_time &&
19960 	    (cts > rack->r_ctl.rc_went_idle_time)) {
19961 		tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
19962 		if (tot_idle > rack_min_probertt_hold) {
19963 			/* Count as a probe rtt */
19964 			if (rack->in_probe_rtt == 0) {
19965 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
19966 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19967 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19968 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19969 			} else {
19970 				rack_exit_probertt(rack, cts);
19971 			}
19972 		}
19973 	} else
19974 		tot_idle = 0;
19975 	if (rack_use_fsb &&
19976 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
19977 	    (rack->r_fsb_inited == 0) &&
19978 	    (rack->r_state != TCPS_CLOSED))
19979 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
19980 	if (rack->rc_sendvars_notset == 1) {
19981 		rack->rc_sendvars_notset = 0;
19982 		/*
19983 		 * Make sure any TCP timers (keep-alive) is not running.
19984 		 */
19985 		tcp_timer_stop(tp);
19986 	}
19987 	if ((rack->rack_no_prr == 1) &&
19988 	    (rack->rc_always_pace == 0)) {
19989 		/*
19990 		 * Sanity check before sending, if we have
19991 		 * no-pacing enabled and prr is turned off that
19992 		 * is a logistics error. Correct this by turnning
19993 		 * prr back on. A user *must* set some form of
19994 		 * pacing in order to turn PRR off. We do this
19995 		 * in the output path so that we can avoid socket
19996 		 * option ordering issues that would occur if we
19997 		 * tried to do it while setting rack_no_prr on.
19998 		 */
19999 		rack->rack_no_prr = 0;
20000 	}
20001 	if ((rack->pcm_enabled == 1) &&
20002 	    (rack->pcm_needed == 0) &&
20003 	    (tot_idle > 0)) {
20004 		/*
20005 		 * We have been idle some micro seconds. We need
20006 		 * to factor this in to see if a PCM is needed.
20007 		 */
20008 		uint32_t rtts_idle, rnds;
20009 
20010 		if (tp->t_srtt)
20011 			rtts_idle = tot_idle / tp->t_srtt;
20012 		else
20013 			rtts_idle = 0;
20014 		rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
20015 		rack->r_ctl.pcm_idle_rounds += rtts_idle;
20016 		if ((rnds + rack->r_ctl.pcm_idle_rounds)  >= rack_pcm_every_n_rounds) {
20017 			rack->pcm_needed = 1;
20018 			rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
20019 		}
20020 	}
20021 again:
20022 	sendalot = 0;
20023 	cts = tcp_get_usecs(&tv);
20024 	ms_cts = tcp_tv_to_msec(&tv);
20025 	tso = 0;
20026 	mtu = 0;
20027 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
20028 	    (rack->r_ctl.pcm_max_seg == 0)) {
20029 		/*
20030 		 * We set in our first send so we know that the ctf_fixed_maxseg
20031 		 * has been fully set. If we do it in rack_init() we most likely
20032 		 * see 512 bytes so we end up at 5120, not desirable.
20033 		 */
20034 		rack->r_ctl.pcm_max_seg = rc_init_window(rack);
20035 		if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
20036 			/*
20037 			 * Assure our initial PCM probe is at least 10 MSS.
20038 			 */
20039 			rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
20040 		}
20041 	}
20042 	if ((rack->r_ctl.pcm_max_seg != 0) && (rack->pcm_needed == 1)) {
20043 		uint32_t rw_avail, cwa;
20044 
20045 		if (tp->snd_wnd > ctf_outstanding(tp))
20046 			rw_avail = tp->snd_wnd - ctf_outstanding(tp);
20047 		else
20048 			rw_avail = 0;
20049 		if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
20050 			cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20051 		else
20052 			cwa = 0;
20053 		if ((cwa >= rack->r_ctl.pcm_max_seg) &&
20054 		    (rw_avail > rack->r_ctl.pcm_max_seg)) {
20055 			/* Raise up the max seg for this trip through */
20056 			pace_max_seg = rack->r_ctl.pcm_max_seg;
20057 			/* Disable any fast output */
20058 			rack->r_fast_output = 0;
20059 		}
20060 		if (rack_verbose_logging) {
20061 			rack_log_pcm(rack, 4,
20062 				     cwa, rack->r_ctl.pcm_max_seg, rw_avail);
20063 		}
20064 	}
20065 	sb_offset = tp->snd_max - tp->snd_una;
20066 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20067 	flags = tcp_outflags[tp->t_state];
20068 	while (rack->rc_free_cnt < rack_free_cache) {
20069 		rsm = rack_alloc(rack);
20070 		if (rsm == NULL) {
20071 			if (hpts_calling)
20072 				/* Retry in a ms */
20073 				slot = (1 * HPTS_USEC_IN_MSEC);
20074 			so = inp->inp_socket;
20075 			sb = &so->so_snd;
20076 			goto just_return_nolock;
20077 		}
20078 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20079 		rack->rc_free_cnt++;
20080 		rsm = NULL;
20081 	}
20082 	sack_rxmit = 0;
20083 	len = 0;
20084 	rsm = NULL;
20085 	if (flags & TH_RST) {
20086 		SOCK_SENDBUF_LOCK(inp->inp_socket);
20087 		so = inp->inp_socket;
20088 		sb = &so->so_snd;
20089 		goto send;
20090 	}
20091 	if (rack->r_ctl.rc_resend) {
20092 		/* Retransmit timer */
20093 		rsm = rack->r_ctl.rc_resend;
20094 		rack->r_ctl.rc_resend = NULL;
20095 		len = rsm->r_end - rsm->r_start;
20096 		sack_rxmit = 1;
20097 		sendalot = 0;
20098 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20099 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20100 			 __func__, __LINE__,
20101 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20102 		sb_offset = rsm->r_start - tp->snd_una;
20103 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20104 	} else if (rack->r_collapse_point_valid &&
20105 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20106 		/*
20107 		 * If an RSM is returned then enough time has passed
20108 		 * for us to retransmit it. Move up the collapse point,
20109 		 * since this rsm has its chance to retransmit now.
20110 		 */
20111 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20112 		rack->r_ctl.last_collapse_point = rsm->r_end;
20113 		/* Are we done? */
20114 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20115 			    rack->r_ctl.high_collapse_point))
20116 			rack->r_collapse_point_valid = 0;
20117 		sack_rxmit = 1;
20118 		/* We are not doing a TLP */
20119 		doing_tlp = 0;
20120 		len = rsm->r_end - rsm->r_start;
20121 		sb_offset = rsm->r_start - tp->snd_una;
20122 		sendalot = 0;
20123 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20124 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20125 		/* We have a retransmit that takes precedence */
20126 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20127 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20128 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20129 			/* Enter recovery if not induced by a time-out */
20130 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20131 		}
20132 #ifdef INVARIANTS
20133 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20134 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20135 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
20136 		}
20137 #endif
20138 		len = rsm->r_end - rsm->r_start;
20139 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20140 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20141 			 __func__, __LINE__,
20142 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20143 		sb_offset = rsm->r_start - tp->snd_una;
20144 		sendalot = 0;
20145 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20146 		if (len > 0) {
20147 			sack_rxmit = 1;
20148 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20149 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20150 					 min(len, segsiz));
20151 		}
20152 	} else if (rack->r_ctl.rc_tlpsend) {
20153 		/* Tail loss probe */
20154 		long cwin;
20155 		long tlen;
20156 
20157 		/*
20158 		 * Check if we can do a TLP with a RACK'd packet
20159 		 * this can happen if we are not doing the rack
20160 		 * cheat and we skipped to a TLP and it
20161 		 * went off.
20162 		 */
20163 		rsm = rack->r_ctl.rc_tlpsend;
20164 		/* We are doing a TLP make sure the flag is preent */
20165 		rsm->r_flags |= RACK_TLP;
20166 		rack->r_ctl.rc_tlpsend = NULL;
20167 		sack_rxmit = 1;
20168 		tlen = rsm->r_end - rsm->r_start;
20169 		if (tlen > segsiz)
20170 			tlen = segsiz;
20171 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20172 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20173 			 __func__, __LINE__,
20174 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20175 		sb_offset = rsm->r_start - tp->snd_una;
20176 		cwin = min(tp->snd_wnd, tlen);
20177 		len = cwin;
20178 	}
20179 	if (rack->r_must_retran &&
20180 	    (doing_tlp == 0) &&
20181 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20182 	    (rsm == NULL)) {
20183 		/*
20184 		 * There are two different ways that we
20185 		 * can get into this block:
20186 		 * a) This is a non-sack connection, we had a time-out
20187 		 *    and thus r_must_retran was set and everything
20188 		 *    left outstanding as been marked for retransmit.
20189 		 * b) The MTU of the path shrank, so that everything
20190 		 *    was marked to be retransmitted with the smaller
20191 		 *    mtu and r_must_retran was set.
20192 		 *
20193 		 * This means that we expect the sendmap (outstanding)
20194 		 * to all be marked must. We can use the tmap to
20195 		 * look at them.
20196 		 *
20197 		 */
20198 		int sendwin, flight;
20199 
20200 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20201 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20202 		if (flight >= sendwin) {
20203 			/*
20204 			 * We can't send yet.
20205 			 */
20206 			so = inp->inp_socket;
20207 			sb = &so->so_snd;
20208 			goto just_return_nolock;
20209 		}
20210 		/*
20211 		 * This is the case a/b mentioned above. All
20212 		 * outstanding/not-acked should be marked.
20213 		 * We can use the tmap to find them.
20214 		 */
20215 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20216 		if (rsm == NULL) {
20217 			/* TSNH */
20218 			rack->r_must_retran = 0;
20219 			rack->r_ctl.rc_out_at_rto = 0;
20220 			so = inp->inp_socket;
20221 			sb = &so->so_snd;
20222 			goto just_return_nolock;
20223 		}
20224 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20225 			/*
20226 			 * The first one does not have the flag, did we collapse
20227 			 * further up in our list?
20228 			 */
20229 			rack->r_must_retran = 0;
20230 			rack->r_ctl.rc_out_at_rto = 0;
20231 			rsm = NULL;
20232 			sack_rxmit = 0;
20233 		} else {
20234 			sack_rxmit = 1;
20235 			len = rsm->r_end - rsm->r_start;
20236 			sb_offset = rsm->r_start - tp->snd_una;
20237 			sendalot = 0;
20238 			if ((rack->full_size_rxt == 0) &&
20239 			    (rack->shape_rxt_to_pacing_min == 0) &&
20240 			    (len >= segsiz))
20241 				len = segsiz;
20242 			else if (rack->shape_rxt_to_pacing_min &&
20243 				 rack->gp_ready) {
20244 				/* We use pacing min as shaping len req */
20245 				uint32_t maxlen;
20246 
20247 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20248 				if (len > maxlen)
20249 					len = maxlen;
20250 			}
20251 			/*
20252 			 * Delay removing the flag RACK_MUST_RXT so
20253 			 * that the fastpath for retransmit will
20254 			 * work with this rsm.
20255 			 */
20256 		}
20257 	}
20258 	/*
20259 	 * Enforce a connection sendmap count limit if set
20260 	 * as long as we are not retransmiting.
20261 	 */
20262 	if ((rsm == NULL) &&
20263 	    (V_tcp_map_entries_limit > 0) &&
20264 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20265 		counter_u64_add(rack_to_alloc_limited, 1);
20266 		if (!rack->alloc_limit_reported) {
20267 			rack->alloc_limit_reported = 1;
20268 			counter_u64_add(rack_alloc_limited_conns, 1);
20269 		}
20270 		so = inp->inp_socket;
20271 		sb = &so->so_snd;
20272 		goto just_return_nolock;
20273 	}
20274 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20275 		/* we are retransmitting the fin */
20276 		len--;
20277 		if (len) {
20278 			/*
20279 			 * When retransmitting data do *not* include the
20280 			 * FIN. This could happen from a TLP probe.
20281 			 */
20282 			flags &= ~TH_FIN;
20283 		}
20284 	}
20285 	if (rsm && rack->r_fsb_inited &&
20286 	    rack_use_rsm_rfo &&
20287 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20288 		int ret;
20289 
20290 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20291 		if (ret == 0)
20292 			return (0);
20293 	}
20294 	so = inp->inp_socket;
20295 	sb = &so->so_snd;
20296 	if (do_a_prefetch == 0) {
20297 		kern_prefetch(sb, &do_a_prefetch);
20298 		do_a_prefetch = 1;
20299 	}
20300 #ifdef NETFLIX_SHARED_CWND
20301 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20302 	    rack->rack_enable_scwnd) {
20303 		/* We are doing cwnd sharing */
20304 		if (rack->gp_ready &&
20305 		    (rack->rack_attempted_scwnd == 0) &&
20306 		    (rack->r_ctl.rc_scw == NULL) &&
20307 		    tp->t_lib) {
20308 			/* The pcbid is in, lets make an attempt */
20309 			counter_u64_add(rack_try_scwnd, 1);
20310 			rack->rack_attempted_scwnd = 1;
20311 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20312 								   &rack->r_ctl.rc_scw_index,
20313 								   segsiz);
20314 		}
20315 		if (rack->r_ctl.rc_scw &&
20316 		    (rack->rack_scwnd_is_idle == 1) &&
20317 		    sbavail(&so->so_snd)) {
20318 			/* we are no longer out of data */
20319 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20320 			rack->rack_scwnd_is_idle = 0;
20321 		}
20322 		if (rack->r_ctl.rc_scw) {
20323 			/* First lets update and get the cwnd */
20324 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20325 										       rack->r_ctl.rc_scw_index,
20326 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
20327 		}
20328 	}
20329 #endif
20330 	/*
20331 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20332 	 * state flags.
20333 	 */
20334 	if (tp->t_flags & TF_NEEDFIN)
20335 		flags |= TH_FIN;
20336 	if (tp->t_flags & TF_NEEDSYN)
20337 		flags |= TH_SYN;
20338 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20339 		void *end_rsm;
20340 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20341 		if (end_rsm)
20342 			kern_prefetch(end_rsm, &prefetch_rsm);
20343 		prefetch_rsm = 1;
20344 	}
20345 	SOCK_SENDBUF_LOCK(so);
20346 	if ((sack_rxmit == 0) &&
20347 	    (TCPS_HAVEESTABLISHED(tp->t_state) ||
20348 	    (tp->t_flags & TF_FASTOPEN))) {
20349 		/*
20350 		 * We are not retransmitting (sack_rxmit is 0) so we
20351 		 * are sending new data. This is always based on snd_max.
20352 		 * Now in theory snd_max may be equal to snd_una, if so
20353 		 * then nothing is outstanding and the offset would be 0.
20354 		 */
20355 		uint32_t avail;
20356 
20357 		avail = sbavail(sb);
20358 		if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
20359 			sb_offset = tp->snd_max - tp->snd_una;
20360 		else
20361 			sb_offset = 0;
20362 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20363 			if (rack->r_ctl.rc_tlp_new_data) {
20364 				/* TLP is forcing out new data */
20365 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20366 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20367 				}
20368 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20369 					if (tp->snd_wnd > sb_offset)
20370 						len = tp->snd_wnd - sb_offset;
20371 					else
20372 						len = 0;
20373 				} else {
20374 					len = rack->r_ctl.rc_tlp_new_data;
20375 				}
20376 				rack->r_ctl.rc_tlp_new_data = 0;
20377 			}  else {
20378 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20379 			}
20380 			if ((rack->r_ctl.crte == NULL) &&
20381 			    IN_FASTRECOVERY(tp->t_flags) &&
20382 			    (rack->full_size_rxt == 0) &&
20383 			    (rack->shape_rxt_to_pacing_min == 0) &&
20384 			    (len > segsiz)) {
20385 				/*
20386 				 * For prr=off, we need to send only 1 MSS
20387 				 * at a time. We do this because another sack could
20388 				 * be arriving that causes us to send retransmits and
20389 				 * we don't want to be on a long pace due to a larger send
20390 				 * that keeps us from sending out the retransmit.
20391 				 */
20392 				len = segsiz;
20393 			} else if (rack->shape_rxt_to_pacing_min &&
20394 				   rack->gp_ready) {
20395 				/* We use pacing min as shaping len req */
20396 				uint32_t maxlen;
20397 
20398 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20399 				if (len > maxlen)
20400 					len = maxlen;
20401 			}/* The else is full_size_rxt is on so send it all */
20402 		} else {
20403 			uint32_t outstanding;
20404 			/*
20405 			 * We are inside of a Fast recovery episode, this
20406 			 * is caused by a SACK or 3 dup acks. At this point
20407 			 * we have sent all the retransmissions and we rely
20408 			 * on PRR to dictate what we will send in the form of
20409 			 * new data.
20410 			 */
20411 
20412 			outstanding = tp->snd_max - tp->snd_una;
20413 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20414 				if (tp->snd_wnd > outstanding) {
20415 					len = tp->snd_wnd - outstanding;
20416 					/* Check to see if we have the data */
20417 					if ((sb_offset + len) > avail) {
20418 						/* It does not all fit */
20419 						if (avail > sb_offset)
20420 							len = avail - sb_offset;
20421 						else
20422 							len = 0;
20423 					}
20424 				} else {
20425 					len = 0;
20426 				}
20427 			} else if (avail > sb_offset) {
20428 				len = avail - sb_offset;
20429 			} else {
20430 				len = 0;
20431 			}
20432 			if (len > 0) {
20433 				if (len > rack->r_ctl.rc_prr_sndcnt) {
20434 					len = rack->r_ctl.rc_prr_sndcnt;
20435 				}
20436 				if (len > 0) {
20437 					sub_from_prr = 1;
20438 				}
20439 			}
20440 			if (len > segsiz) {
20441 				/*
20442 				 * We should never send more than a MSS when
20443 				 * retransmitting or sending new data in prr
20444 				 * mode unless the override flag is on. Most
20445 				 * likely the PRR algorithm is not going to
20446 				 * let us send a lot as well :-)
20447 				 */
20448 				if (rack->r_ctl.rc_prr_sendalot == 0) {
20449 					len = segsiz;
20450 				}
20451 			} else if (len < segsiz) {
20452 				/*
20453 				 * Do we send any? The idea here is if the
20454 				 * send empty's the socket buffer we want to
20455 				 * do it. However if not then lets just wait
20456 				 * for our prr_sndcnt to get bigger.
20457 				 */
20458 				long leftinsb;
20459 
20460 				leftinsb = sbavail(sb) - sb_offset;
20461 				if (leftinsb > len) {
20462 					/* This send does not empty the sb */
20463 					len = 0;
20464 				}
20465 			}
20466 		}
20467 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20468 		/*
20469 		 * If you have not established
20470 		 * and are not doing FAST OPEN
20471 		 * no data please.
20472 		 */
20473 		if ((sack_rxmit == 0) &&
20474 		    !(tp->t_flags & TF_FASTOPEN)) {
20475 			len = 0;
20476 			sb_offset = 0;
20477 		}
20478 	}
20479 	if (prefetch_so_done == 0) {
20480 		kern_prefetch(so, &prefetch_so_done);
20481 		prefetch_so_done = 1;
20482 	}
20483 	orig_len = len;
20484 	/*
20485 	 * Lop off SYN bit if it has already been sent.  However, if this is
20486 	 * SYN-SENT state and if segment contains data and if we don't know
20487 	 * that foreign host supports TAO, suppress sending segment.
20488 	 */
20489 	if ((flags & TH_SYN) &&
20490 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
20491 	    ((sack_rxmit == 0) &&
20492 	     (tp->t_rxtshift == 0))) {
20493 		/*
20494 		 * When sending additional segments following a TFO SYN|ACK,
20495 		 * do not include the SYN bit.
20496 		 */
20497 		if ((tp->t_flags & TF_FASTOPEN) &&
20498 		    (tp->t_state == TCPS_SYN_RECEIVED))
20499 			flags &= ~TH_SYN;
20500 	}
20501 	/*
20502 	 * Be careful not to send data and/or FIN on SYN segments. This
20503 	 * measure is needed to prevent interoperability problems with not
20504 	 * fully conformant TCP implementations.
20505 	 */
20506 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20507 		len = 0;
20508 		flags &= ~TH_FIN;
20509 	}
20510 	/*
20511 	 * On TFO sockets, ensure no data is sent in the following cases:
20512 	 *
20513 	 *  - When retransmitting SYN|ACK on a passively-created socket
20514 	 *
20515 	 *  - When retransmitting SYN on an actively created socket
20516 	 *
20517 	 *  - When sending a zero-length cookie (cookie request) on an
20518 	 *    actively created socket
20519 	 *
20520 	 *  - When the socket is in the CLOSED state (RST is being sent)
20521 	 */
20522 	if ((tp->t_flags & TF_FASTOPEN) &&
20523 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20524 	     ((tp->t_state == TCPS_SYN_SENT) &&
20525 	      (tp->t_tfo_client_cookie_len == 0)) ||
20526 	     (flags & TH_RST))) {
20527 		sack_rxmit = 0;
20528 		len = 0;
20529 	}
20530 	/* Without fast-open there should never be data sent on a SYN */
20531 	if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
20532 		len = 0;
20533 	}
20534 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20535 		/* We only send 1 MSS if we have a DSACK block */
20536 		add_flag |= RACK_SENT_W_DSACK;
20537 		len = segsiz;
20538 	}
20539 	if (len <= 0) {
20540 		/*
20541 		 * We have nothing to send, or the window shrank, or
20542 		 * is closed, do we need to go into persists?
20543 		 */
20544 		len = 0;
20545 		if ((tp->snd_wnd == 0) &&
20546 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20547 		    (tp->snd_una == tp->snd_max) &&
20548 		    (sb_offset < (int)sbavail(sb))) {
20549 			rack_enter_persist(tp, rack, cts, tp->snd_una);
20550 		}
20551 	} else if ((rsm == NULL) &&
20552 		   (doing_tlp == 0) &&
20553 		   (len < pace_max_seg)) {
20554 		/*
20555 		 * We are not sending a maximum sized segment for
20556 		 * some reason. Should we not send anything (think
20557 		 * sws or persists)?
20558 		 */
20559 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20560 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20561 		    (len < minseg) &&
20562 		    (len < (int)(sbavail(sb) - sb_offset))) {
20563 			/*
20564 			 * Here the rwnd is less than
20565 			 * the minimum pacing size, this is not a retransmit,
20566 			 * we are established and
20567 			 * the send is not the last in the socket buffer
20568 			 * we send nothing, and we may enter persists
20569 			 * if nothing is outstanding.
20570 			 */
20571 			len = 0;
20572 			if (tp->snd_max == tp->snd_una) {
20573 				/*
20574 				 * Nothing out we can
20575 				 * go into persists.
20576 				 */
20577 				rack_enter_persist(tp, rack, cts, tp->snd_una);
20578 			}
20579 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20580 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20581 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20582 			   (len < minseg)) {
20583 			/*
20584 			 * Here we are not retransmitting, and
20585 			 * the cwnd is not so small that we could
20586 			 * not send at least a min size (rxt timer
20587 			 * not having gone off), We have 2 segments or
20588 			 * more already in flight, its not the tail end
20589 			 * of the socket buffer  and the cwnd is blocking
20590 			 * us from sending out a minimum pacing segment size.
20591 			 * Lets not send anything.
20592 			 */
20593 			len = 0;
20594 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20595 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20596 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20597 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20598 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
20599 			/*
20600 			 * Here we have a send window but we have
20601 			 * filled it up and we can't send another pacing segment.
20602 			 * We also have in flight more than 2 segments
20603 			 * and we are not completing the sb i.e. we allow
20604 			 * the last bytes of the sb to go out even if
20605 			 * its not a full pacing segment.
20606 			 */
20607 			len = 0;
20608 		} else if ((rack->r_ctl.crte != NULL) &&
20609 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20610 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20611 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20612 			   (len < (int)(sbavail(sb) - sb_offset))) {
20613 			/*
20614 			 * Here we are doing hardware pacing, this is not a TLP,
20615 			 * we are not sending a pace max segment size, there is rwnd
20616 			 * room to send at least N pace_max_seg, the cwnd is greater
20617 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20618 			 * more segments in flight and its not the tail of the socket buffer.
20619 			 *
20620 			 * We don't want to send instead we need to get more ack's in to
20621 			 * allow us to send a full pacing segment. Normally, if we are pacing
20622 			 * about the right speed, we should have finished our pacing
20623 			 * send as most of the acks have come back if we are at the
20624 			 * right rate. This is a bit fuzzy since return path delay
20625 			 * can delay the acks, which is why we want to make sure we
20626 			 * have cwnd space to have a bit more than a max pace segments in flight.
20627 			 *
20628 			 * If we have not gotten our acks back we are pacing at too high a
20629 			 * rate delaying will not hurt and will bring our GP estimate down by
20630 			 * injecting the delay. If we don't do this we will send
20631 			 * 2 MSS out in response to the acks being clocked in which
20632 			 * defeats the point of hw-pacing (i.e. to help us get
20633 			 * larger TSO's out).
20634 			 */
20635 			len = 0;
20636 		}
20637 
20638 	}
20639 	/* len will be >= 0 after this point. */
20640 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20641 	rack_sndbuf_autoscale(rack);
20642 	/*
20643 	 * Decide if we can use TCP Segmentation Offloading (if supported by
20644 	 * hardware).
20645 	 *
20646 	 * TSO may only be used if we are in a pure bulk sending state.  The
20647 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20648 	 * options prevent using TSO.  With TSO the TCP header is the same
20649 	 * (except for the sequence number) for all generated packets.  This
20650 	 * makes it impossible to transmit any options which vary per
20651 	 * generated segment or packet.
20652 	 *
20653 	 * IPv4 handling has a clear separation of ip options and ip header
20654 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20655 	 * the right thing below to provide length of just ip options and thus
20656 	 * checking for ipoptlen is enough to decide if ip options are present.
20657 	 */
20658 	ipoptlen = 0;
20659 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20660 	/*
20661 	 * Pre-calculate here as we save another lookup into the darknesses
20662 	 * of IPsec that way and can actually decide if TSO is ok.
20663 	 */
20664 #ifdef INET6
20665 	if (isipv6 && IPSEC_ENABLED(ipv6))
20666 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20667 #ifdef INET
20668 	else
20669 #endif
20670 #endif				/* INET6 */
20671 #ifdef INET
20672 		if (IPSEC_ENABLED(ipv4))
20673 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20674 #endif				/* INET */
20675 #endif
20676 
20677 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20678 	ipoptlen += ipsec_optlen;
20679 #endif
20680 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20681 	    (tp->t_port == 0) &&
20682 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
20683 	    sack_rxmit == 0 &&
20684 	    ipoptlen == 0)
20685 		tso = 1;
20686 	{
20687 		uint32_t outstanding __unused;
20688 
20689 		outstanding = tp->snd_max - tp->snd_una;
20690 		if (tp->t_flags & TF_SENTFIN) {
20691 			/*
20692 			 * If we sent a fin, snd_max is 1 higher than
20693 			 * snd_una
20694 			 */
20695 			outstanding--;
20696 		}
20697 		if (sack_rxmit) {
20698 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20699 				flags &= ~TH_FIN;
20700 		}
20701 	}
20702 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20703 		      (long)TCP_MAXWIN << tp->rcv_scale);
20704 
20705 	/*
20706 	 * Sender silly window avoidance.   We transmit under the following
20707 	 * conditions when len is non-zero:
20708 	 *
20709 	 * - We have a full segment (or more with TSO) - This is the last
20710 	 * buffer in a write()/send() and we are either idle or running
20711 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
20712 	 * then 1/2 the maximum send window's worth of data (receiver may be
20713 	 * limited the window size) - we need to retransmit
20714 	 */
20715 	if (len) {
20716 		if (len >= segsiz) {
20717 			goto send;
20718 		}
20719 		/*
20720 		 * NOTE! on localhost connections an 'ack' from the remote
20721 		 * end may occur synchronously with the output and cause us
20722 		 * to flush a buffer queued with moretocome.  XXX
20723 		 *
20724 		 */
20725 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
20726 		    (idle || (tp->t_flags & TF_NODELAY)) &&
20727 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20728 		    (tp->t_flags & TF_NOPUSH) == 0) {
20729 			pass = 2;
20730 			goto send;
20731 		}
20732 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
20733 			pass = 22;
20734 			goto send;
20735 		}
20736 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20737 			pass = 4;
20738 			goto send;
20739 		}
20740 		if (sack_rxmit) {
20741 			pass = 6;
20742 			goto send;
20743 		}
20744 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20745 		    (ctf_outstanding(tp) < (segsiz * 2))) {
20746 			/*
20747 			 * We have less than two MSS outstanding (delayed ack)
20748 			 * and our rwnd will not let us send a full sized
20749 			 * MSS. Lets go ahead and let this small segment
20750 			 * out because we want to try to have at least two
20751 			 * packets inflight to not be caught by delayed ack.
20752 			 */
20753 			pass = 12;
20754 			goto send;
20755 		}
20756 	}
20757 	/*
20758 	 * Sending of standalone window updates.
20759 	 *
20760 	 * Window updates are important when we close our window due to a
20761 	 * full socket buffer and are opening it again after the application
20762 	 * reads data from it.  Once the window has opened again and the
20763 	 * remote end starts to send again the ACK clock takes over and
20764 	 * provides the most current window information.
20765 	 *
20766 	 * We must avoid the silly window syndrome whereas every read from
20767 	 * the receive buffer, no matter how small, causes a window update
20768 	 * to be sent.  We also should avoid sending a flurry of window
20769 	 * updates when the socket buffer had queued a lot of data and the
20770 	 * application is doing small reads.
20771 	 *
20772 	 * Prevent a flurry of pointless window updates by only sending an
20773 	 * update when we can increase the advertized window by more than
20774 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
20775 	 * full or is very small be more aggressive and send an update
20776 	 * whenever we can increase by two mss sized segments. In all other
20777 	 * situations the ACK's to new incoming data will carry further
20778 	 * window increases.
20779 	 *
20780 	 * Don't send an independent window update if a delayed ACK is
20781 	 * pending (it will get piggy-backed on it) or the remote side
20782 	 * already has done a half-close and won't send more data.  Skip
20783 	 * this if the connection is in T/TCP half-open state.
20784 	 */
20785 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20786 	    !(tp->t_flags & TF_DELACK) &&
20787 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
20788 		/*
20789 		 * "adv" is the amount we could increase the window, taking
20790 		 * into account that we are limited by TCP_MAXWIN <<
20791 		 * tp->rcv_scale.
20792 		 */
20793 		int32_t adv;
20794 		int oldwin;
20795 
20796 		adv = recwin;
20797 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20798 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
20799 			if (adv > oldwin)
20800 				adv -= oldwin;
20801 			else {
20802 				/* We can't increase the window */
20803 				adv = 0;
20804 			}
20805 		} else
20806 			oldwin = 0;
20807 
20808 		/*
20809 		 * If the new window size ends up being the same as or less
20810 		 * than the old size when it is scaled, then don't force
20811 		 * a window update.
20812 		 */
20813 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20814 			goto dontupdate;
20815 
20816 		if (adv >= (int32_t)(2 * segsiz) &&
20817 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20818 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20819 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20820 			pass = 7;
20821 			goto send;
20822 		}
20823 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20824 			pass = 23;
20825 			goto send;
20826 		}
20827 	}
20828 dontupdate:
20829 
20830 	/*
20831 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
20832 	 * is also a catch-all for the retransmit timer timeout case.
20833 	 */
20834 	if (tp->t_flags & TF_ACKNOW) {
20835 		pass = 8;
20836 		goto send;
20837 	}
20838 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20839 		pass = 9;
20840 		goto send;
20841 	}
20842 	/*
20843 	 * If our state indicates that FIN should be sent and we have not
20844 	 * yet done so, then we need to send.
20845 	 */
20846 	if ((flags & TH_FIN) &&
20847 	    (tp->snd_max == tp->snd_una)) {
20848 		pass = 11;
20849 		goto send;
20850 	}
20851 	/*
20852 	 * No reason to send a segment, just return.
20853 	 */
20854 just_return:
20855 	SOCK_SENDBUF_UNLOCK(so);
20856 just_return_nolock:
20857 	{
20858 		int app_limited = CTF_JR_SENT_DATA;
20859 
20860 		if ((tp->t_flags & TF_FASTOPEN) == 0 &&
20861 		    (flags & TH_FIN) &&
20862 		    (len == 0) &&
20863 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
20864 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
20865 			/*
20866 			 * Ok less than or right at a MSS is
20867 			 * outstanding. The original FreeBSD stack would
20868 			 * have sent a FIN, which can speed things up for
20869 			 * a transactional application doing a MSG_WAITALL.
20870 			 * To speed things up since we do *not* send a FIN
20871 			 * if data is outstanding, we send a "challenge ack".
20872 			 * The idea behind that is instead of having to have
20873 			 * the peer wait for the delayed-ack timer to run off
20874 			 * we send an ack that makes the peer send us an ack.
20875 			 */
20876 			rack_send_ack_challange(rack);
20877 		}
20878 		if (tot_len_this_send > 0) {
20879 			rack->r_ctl.fsb.recwin = recwin;
20880 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
20881 			if ((error == 0) &&
20882 			    rack_use_rfo &&
20883 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
20884 			    (ipoptlen == 0) &&
20885 			    rack->r_fsb_inited &&
20886 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
20887 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
20888 			    (doing_tlp == 0) &&
20889 			    (rack->r_must_retran == 0) &&
20890 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
20891 			    (len > 0) && (orig_len > 0) &&
20892 			    (orig_len > len) &&
20893 			    ((orig_len - len) >= segsiz) &&
20894 			    ((optlen == 0) ||
20895 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20896 				/* We can send at least one more MSS using our fsb */
20897 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
20898 						       segsiz, pace_max_seg, hw_tls, flags);
20899 			} else
20900 				rack->r_fast_output = 0;
20901 			rack_log_fsb(rack, tp, so, flags,
20902 				     ipoptlen, orig_len, len, 0,
20903 				     1, optlen, __LINE__, 1);
20904 			/* Assure when we leave that snd_nxt will point to top */
20905 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20906 				tp->snd_nxt = tp->snd_max;
20907 		} else {
20908 			int end_window = 0;
20909 			uint32_t seq = tp->gput_ack;
20910 
20911 			rsm = tqhash_max(rack->r_ctl.tqh);
20912 			if (rsm) {
20913 				/*
20914 				 * Mark the last sent that we just-returned (hinting
20915 				 * that delayed ack may play a role in any rtt measurement).
20916 				 */
20917 				rsm->r_just_ret = 1;
20918 			}
20919 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20920 			rack->r_ctl.rc_agg_delayed = 0;
20921 			rack->r_early = 0;
20922 			rack->r_late = 0;
20923 			rack->r_ctl.rc_agg_early = 0;
20924 			if ((ctf_outstanding(tp) +
20925 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20926 				 minseg)) >= tp->snd_wnd) {
20927 				/* We are limited by the rwnd */
20928 				app_limited = CTF_JR_RWND_LIMITED;
20929 				if (IN_FASTRECOVERY(tp->t_flags))
20930 					rack->r_ctl.rc_prr_sndcnt = 0;
20931 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
20932 				/* We are limited by whats available -- app limited */
20933 				app_limited = CTF_JR_APP_LIMITED;
20934 				if (IN_FASTRECOVERY(tp->t_flags))
20935 					rack->r_ctl.rc_prr_sndcnt = 0;
20936 			} else if ((idle == 0) &&
20937 				   ((tp->t_flags & TF_NODELAY) == 0) &&
20938 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20939 				   (len < segsiz)) {
20940 				/*
20941 				 * No delay is not on and the
20942 				 * user is sending less than 1MSS. This
20943 				 * brings out SWS avoidance so we
20944 				 * don't send. Another app-limited case.
20945 				 */
20946 				app_limited = CTF_JR_APP_LIMITED;
20947 			} else if (tp->t_flags & TF_NOPUSH) {
20948 				/*
20949 				 * The user has requested no push of
20950 				 * the last segment and we are
20951 				 * at the last segment. Another app
20952 				 * limited case.
20953 				 */
20954 				app_limited = CTF_JR_APP_LIMITED;
20955 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20956 				/* Its the cwnd */
20957 				app_limited = CTF_JR_CWND_LIMITED;
20958 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
20959 				   (rack->rack_no_prr == 0) &&
20960 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20961 				app_limited = CTF_JR_PRR;
20962 			} else {
20963 				/* Now why here are we not sending? */
20964 #ifdef NOW
20965 #ifdef INVARIANTS
20966 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20967 #endif
20968 #endif
20969 				app_limited = CTF_JR_ASSESSING;
20970 			}
20971 			/*
20972 			 * App limited in some fashion, for our pacing GP
20973 			 * measurements we don't want any gap (even cwnd).
20974 			 * Close  down the measurement window.
20975 			 */
20976 			if (rack_cwnd_block_ends_measure &&
20977 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
20978 			     (app_limited == CTF_JR_PRR))) {
20979 				/*
20980 				 * The reason we are not sending is
20981 				 * the cwnd (or prr). We have been configured
20982 				 * to end the measurement window in
20983 				 * this case.
20984 				 */
20985 				end_window = 1;
20986 			} else if (rack_rwnd_block_ends_measure &&
20987 				   (app_limited == CTF_JR_RWND_LIMITED)) {
20988 				/*
20989 				 * We are rwnd limited and have been
20990 				 * configured to end the measurement
20991 				 * window in this case.
20992 				 */
20993 				end_window = 1;
20994 			} else if (app_limited == CTF_JR_APP_LIMITED) {
20995 				/*
20996 				 * A true application limited period, we have
20997 				 * ran out of data.
20998 				 */
20999 				end_window = 1;
21000 			} else if (app_limited == CTF_JR_ASSESSING) {
21001 				/*
21002 				 * In the assessing case we hit the end of
21003 				 * the if/else and had no known reason
21004 				 * This will panic us under invariants..
21005 				 *
21006 				 * If we get this out in logs we need to
21007 				 * investagate which reason we missed.
21008 				 */
21009 				end_window = 1;
21010 			}
21011 			if (end_window) {
21012 				uint8_t log = 0;
21013 
21014 				/* Adjust the Gput measurement */
21015 				if ((tp->t_flags & TF_GPUTINPROG) &&
21016 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21017 					tp->gput_ack = tp->snd_max;
21018 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21019 						/*
21020 						 * There is not enough to measure.
21021 						 */
21022 						tp->t_flags &= ~TF_GPUTINPROG;
21023 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21024 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
21025 									   tp->gput_seq,
21026 									   0, 0, 18, __LINE__, NULL, 0);
21027 					} else
21028 						log = 1;
21029 				}
21030 				/* Mark the last packet as app limited */
21031 				rsm = tqhash_max(rack->r_ctl.tqh);
21032 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21033 					if (rack->r_ctl.rc_app_limited_cnt == 0)
21034 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21035 					else {
21036 						/*
21037 						 * Go out to the end app limited and mark
21038 						 * this new one as next and move the end_appl up
21039 						 * to this guy.
21040 						 */
21041 						if (rack->r_ctl.rc_end_appl)
21042 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21043 						rack->r_ctl.rc_end_appl = rsm;
21044 					}
21045 					rsm->r_flags |= RACK_APP_LIMITED;
21046 					rack->r_ctl.rc_app_limited_cnt++;
21047 				}
21048 				if (log)
21049 					rack_log_pacing_delay_calc(rack,
21050 								   rack->r_ctl.rc_app_limited_cnt, seq,
21051 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21052 			}
21053 		}
21054 		/* Check if we need to go into persists or not */
21055 		if ((tp->snd_max == tp->snd_una) &&
21056 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
21057 		    sbavail(sb) &&
21058 		    (sbavail(sb) > tp->snd_wnd) &&
21059 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21060 			/* Yes lets make sure to move to persist before timer-start */
21061 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21062 		}
21063 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21064 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21065 	}
21066 #ifdef NETFLIX_SHARED_CWND
21067 	if ((sbavail(sb) == 0) &&
21068 	    rack->r_ctl.rc_scw) {
21069 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21070 		rack->rack_scwnd_is_idle = 1;
21071 	}
21072 #endif
21073 #ifdef TCP_ACCOUNTING
21074 	if (tot_len_this_send > 0) {
21075 		crtsc = get_cyclecount();
21076 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21077 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
21078 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21079 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21080 		}
21081 	} else {
21082 		crtsc = get_cyclecount();
21083 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21084 			tp->tcp_cnt_counters[SND_LIMITED]++;
21085 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21086 		}
21087 	}
21088 	sched_unpin();
21089 #endif
21090 	return (0);
21091 
21092 send:
21093 	if ((rack->r_ctl.crte != NULL) &&
21094 	    (rsm == NULL) &&
21095 	    ((rack->rc_hw_nobuf == 1) ||
21096 	     (rack_hw_check_queue && (check_done == 0)))) {
21097 		/*
21098 		 * We only want to do this once with the hw_check_queue,
21099 		 * for the enobuf case we would only do it once if
21100 		 * we come around to again, the flag will be clear.
21101 		 */
21102 		check_done = 1;
21103 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21104 		if (slot) {
21105 			rack->r_ctl.rc_agg_delayed = 0;
21106 			rack->r_ctl.rc_agg_early = 0;
21107 			rack->r_early = 0;
21108 			rack->r_late = 0;
21109 			SOCK_SENDBUF_UNLOCK(so);
21110 			goto skip_all_send;
21111 		}
21112 	}
21113 	if (rsm || sack_rxmit)
21114 		counter_u64_add(rack_nfto_resend, 1);
21115 	else
21116 		counter_u64_add(rack_non_fto_send, 1);
21117 	if ((flags & TH_FIN) &&
21118 	    sbavail(sb)) {
21119 		/*
21120 		 * We do not transmit a FIN
21121 		 * with data outstanding. We
21122 		 * need to make it so all data
21123 		 * is acked first.
21124 		 */
21125 		flags &= ~TH_FIN;
21126 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21127 		    (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21128 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21129 			/*
21130 			 * Ok less than or right at a MSS is
21131 			 * outstanding. The original FreeBSD stack would
21132 			 * have sent a FIN, which can speed things up for
21133 			 * a transactional application doing a MSG_WAITALL.
21134 			 * To speed things up since we do *not* send a FIN
21135 			 * if data is outstanding, we send a "challenge ack".
21136 			 * The idea behind that is instead of having to have
21137 			 * the peer wait for the delayed-ack timer to run off
21138 			 * we send an ack that makes the peer send us an ack.
21139 			 */
21140 			rack_send_ack_challange(rack);
21141 		}
21142 	}
21143 	/* Enforce stack imposed max seg size if we have one */
21144 	if (pace_max_seg &&
21145 	    (len > pace_max_seg)) {
21146 		mark = 1;
21147 		len = pace_max_seg;
21148 	}
21149 	if ((rsm == NULL) &&
21150 	    (rack->pcm_in_progress == 0) &&
21151 	    (rack->r_ctl.pcm_max_seg > 0) &&
21152 	    (len >= rack->r_ctl.pcm_max_seg)) {
21153 		/* It is large enough for a measurement */
21154 		add_flag |= RACK_IS_PCM;
21155 		rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg,  add_flag);
21156 	} else if (rack_verbose_logging) {
21157 		rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg,  add_flag);
21158 	}
21159 
21160 	SOCKBUF_LOCK_ASSERT(sb);
21161 	if (len > 0) {
21162 		if (len >= segsiz)
21163 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21164 		else
21165 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21166 	}
21167 	/*
21168 	 * Before ESTABLISHED, force sending of initial options unless TCP
21169 	 * set not to do any options. NOTE: we assume that the IP/TCP header
21170 	 * plus TCP options always fit in a single mbuf, leaving room for a
21171 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21172 	 * + optlen <= MCLBYTES
21173 	 */
21174 	optlen = 0;
21175 #ifdef INET6
21176 	if (isipv6)
21177 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21178 	else
21179 #endif
21180 		hdrlen = sizeof(struct tcpiphdr);
21181 
21182 	/*
21183 	 * Ok what seq are we sending from. If we have
21184 	 * no rsm to use, then we look at various bits,
21185 	 * if we are putting out a SYN it will be ISS.
21186 	 * If we are retransmitting a FIN it will
21187 	 * be snd_max-1 else its snd_max.
21188 	 */
21189 	if (rsm == NULL) {
21190 		if (flags & TH_SYN)
21191 			rack_seq = tp->iss;
21192 		else if ((flags & TH_FIN) &&
21193 			 (tp->t_flags & TF_SENTFIN))
21194 			rack_seq = tp->snd_max - 1;
21195 		else
21196 			rack_seq = tp->snd_max;
21197 	} else {
21198 		rack_seq = rsm->r_start;
21199 	}
21200 	/*
21201 	 * Compute options for segment. We only have to care about SYN and
21202 	 * established connection segments.  Options for SYN-ACK segments
21203 	 * are handled in TCP syncache.
21204 	 */
21205 	to.to_flags = 0;
21206 	if ((tp->t_flags & TF_NOOPT) == 0) {
21207 		/* Maximum segment size. */
21208 		if (flags & TH_SYN) {
21209 			to.to_mss = tcp_mssopt(&inp->inp_inc);
21210 			if (tp->t_port)
21211 				to.to_mss -= V_tcp_udp_tunneling_overhead;
21212 			to.to_flags |= TOF_MSS;
21213 
21214 			/*
21215 			 * On SYN or SYN|ACK transmits on TFO connections,
21216 			 * only include the TFO option if it is not a
21217 			 * retransmit, as the presence of the TFO option may
21218 			 * have caused the original SYN or SYN|ACK to have
21219 			 * been dropped by a middlebox.
21220 			 */
21221 			if ((tp->t_flags & TF_FASTOPEN) &&
21222 			    (tp->t_rxtshift == 0)) {
21223 				if (tp->t_state == TCPS_SYN_RECEIVED) {
21224 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21225 					to.to_tfo_cookie =
21226 						(u_int8_t *)&tp->t_tfo_cookie.server;
21227 					to.to_flags |= TOF_FASTOPEN;
21228 					wanted_cookie = 1;
21229 				} else if (tp->t_state == TCPS_SYN_SENT) {
21230 					to.to_tfo_len =
21231 						tp->t_tfo_client_cookie_len;
21232 					to.to_tfo_cookie =
21233 						tp->t_tfo_cookie.client;
21234 					to.to_flags |= TOF_FASTOPEN;
21235 					wanted_cookie = 1;
21236 					/*
21237 					 * If we wind up having more data to
21238 					 * send with the SYN than can fit in
21239 					 * one segment, don't send any more
21240 					 * until the SYN|ACK comes back from
21241 					 * the other end.
21242 					 */
21243 					sendalot = 0;
21244 				}
21245 			}
21246 		}
21247 		/* Window scaling. */
21248 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21249 			to.to_wscale = tp->request_r_scale;
21250 			to.to_flags |= TOF_SCALE;
21251 		}
21252 		/* Timestamps. */
21253 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
21254 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21255 			uint32_t ts_to_use;
21256 
21257 			if ((rack->r_rcvpath_rtt_up == 1) &&
21258 			    (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
21259 				/*
21260 				 * When we are doing a rcv_rtt probe all
21261 				 * other timestamps use the next msec. This
21262 				 * is safe since our previous ack is in the
21263 				 * air and we will just have a few more
21264 				 * on the next ms. This assures that only
21265 				 * the one ack has the ms_cts that was on
21266 				 * our ack-probe.
21267 				 */
21268 				ts_to_use = ms_cts + 1;
21269 			} else {
21270 				ts_to_use = ms_cts;
21271 			}
21272 			to.to_tsval = ts_to_use + tp->ts_offset;
21273 			to.to_tsecr = tp->ts_recent;
21274 			to.to_flags |= TOF_TS;
21275 			if ((len == 0) &&
21276 			    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21277 			    ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
21278 			    (tp->snd_una == tp->snd_max) &&
21279 			    (flags & TH_ACK) &&
21280 			    (sbavail(sb) == 0) &&
21281 			    (rack->r_ctl.current_round != 0) &&
21282 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21283 			    (rack->r_rcvpath_rtt_up == 0)) {
21284 				rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
21285 				rack->r_ctl.last_time_of_arm_rcv = cts;
21286 				rack->r_rcvpath_rtt_up = 1;
21287 				/* Subtract 1 from seq to force a response */
21288 				rack_seq--;
21289 			}
21290 		}
21291 		/* Set receive buffer autosizing timestamp. */
21292 		if (tp->rfbuf_ts == 0 &&
21293 		    (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
21294 			tp->rfbuf_ts = ms_cts;
21295 		}
21296 		/* Selective ACK's. */
21297 		if (tp->t_flags & TF_SACK_PERMIT) {
21298 			if (flags & TH_SYN)
21299 				to.to_flags |= TOF_SACKPERM;
21300 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21301 				 tp->rcv_numsacks > 0) {
21302 				to.to_flags |= TOF_SACK;
21303 				to.to_nsacks = tp->rcv_numsacks;
21304 				to.to_sacks = (u_char *)tp->sackblks;
21305 			}
21306 		}
21307 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21308 		/* TCP-MD5 (RFC2385). */
21309 		if (tp->t_flags & TF_SIGNATURE)
21310 			to.to_flags |= TOF_SIGNATURE;
21311 #endif
21312 
21313 		/* Processing the options. */
21314 		hdrlen += optlen = tcp_addoptions(&to, opt);
21315 		/*
21316 		 * If we wanted a TFO option to be added, but it was unable
21317 		 * to fit, ensure no data is sent.
21318 		 */
21319 		if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
21320 		    !(to.to_flags & TOF_FASTOPEN))
21321 			len = 0;
21322 	}
21323 	if (tp->t_port) {
21324 		if (V_tcp_udp_tunneling_port == 0) {
21325 			/* The port was removed?? */
21326 			SOCK_SENDBUF_UNLOCK(so);
21327 #ifdef TCP_ACCOUNTING
21328 			crtsc = get_cyclecount();
21329 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21330 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21331 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21332 			}
21333 			sched_unpin();
21334 #endif
21335 			return (EHOSTUNREACH);
21336 		}
21337 		hdrlen += sizeof(struct udphdr);
21338 	}
21339 #ifdef INET6
21340 	if (isipv6)
21341 		ipoptlen = ip6_optlen(inp);
21342 	else
21343 #endif
21344 		if (inp->inp_options)
21345 			ipoptlen = inp->inp_options->m_len -
21346 				offsetof(struct ipoption, ipopt_list);
21347 		else
21348 			ipoptlen = 0;
21349 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21350 	ipoptlen += ipsec_optlen;
21351 #endif
21352 
21353 	/*
21354 	 * Adjust data length if insertion of options will bump the packet
21355 	 * length beyond the t_maxseg length. Clear the FIN bit because we
21356 	 * cut off the tail of the segment.
21357 	 */
21358 	if (len + optlen + ipoptlen > tp->t_maxseg) {
21359 		if (tso) {
21360 			uint32_t if_hw_tsomax;
21361 			uint32_t moff;
21362 			int32_t max_len;
21363 
21364 			/* extract TSO information */
21365 			if_hw_tsomax = tp->t_tsomax;
21366 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21367 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21368 			KASSERT(ipoptlen == 0,
21369 				("%s: TSO can't do IP options", __func__));
21370 
21371 			/*
21372 			 * Check if we should limit by maximum payload
21373 			 * length:
21374 			 */
21375 			if (if_hw_tsomax != 0) {
21376 				/* compute maximum TSO length */
21377 				max_len = (if_hw_tsomax - hdrlen -
21378 					   max_linkhdr);
21379 				if (max_len <= 0) {
21380 					len = 0;
21381 				} else if (len > max_len) {
21382 					if (doing_tlp == 0)
21383 						sendalot = 1;
21384 					len = max_len;
21385 					mark = 2;
21386 				}
21387 			}
21388 			/*
21389 			 * Prevent the last segment from being fractional
21390 			 * unless the send sockbuf can be emptied:
21391 			 */
21392 			max_len = (tp->t_maxseg - optlen);
21393 			if ((sb_offset + len) < sbavail(sb)) {
21394 				moff = len % (u_int)max_len;
21395 				if (moff != 0) {
21396 					mark = 3;
21397 					len -= moff;
21398 				}
21399 			}
21400 			/*
21401 			 * In case there are too many small fragments don't
21402 			 * use TSO:
21403 			 */
21404 			if (len <= max_len) {
21405 				mark = 4;
21406 				tso = 0;
21407 			}
21408 			/*
21409 			 * Send the FIN in a separate segment after the bulk
21410 			 * sending is done. We don't trust the TSO
21411 			 * implementations to clear the FIN flag on all but
21412 			 * the last segment.
21413 			 */
21414 			if (tp->t_flags & TF_NEEDFIN) {
21415 				sendalot = 4;
21416 			}
21417 		} else {
21418 			mark = 5;
21419 			if (optlen + ipoptlen >= tp->t_maxseg) {
21420 				/*
21421 				 * Since we don't have enough space to put
21422 				 * the IP header chain and the TCP header in
21423 				 * one packet as required by RFC 7112, don't
21424 				 * send it. Also ensure that at least one
21425 				 * byte of the payload can be put into the
21426 				 * TCP segment.
21427 				 */
21428 				SOCK_SENDBUF_UNLOCK(so);
21429 				error = EMSGSIZE;
21430 				sack_rxmit = 0;
21431 				goto out;
21432 			}
21433 			len = tp->t_maxseg - optlen - ipoptlen;
21434 			sendalot = 5;
21435 		}
21436 	} else {
21437 		tso = 0;
21438 		mark = 6;
21439 	}
21440 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21441 		("%s: len > IP_MAXPACKET", __func__));
21442 #ifdef DIAGNOSTIC
21443 #ifdef INET6
21444 	if (max_linkhdr + hdrlen > MCLBYTES)
21445 #else
21446 		if (max_linkhdr + hdrlen > MHLEN)
21447 #endif
21448 			panic("tcphdr too big");
21449 #endif
21450 
21451 	/*
21452 	 * This KASSERT is here to catch edge cases at a well defined place.
21453 	 * Before, those had triggered (random) panic conditions further
21454 	 * down.
21455 	 */
21456 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21457 	if ((len == 0) &&
21458 	    (flags & TH_FIN) &&
21459 	    (sbused(sb))) {
21460 		/*
21461 		 * We have outstanding data, don't send a fin by itself!.
21462 		 *
21463 		 * Check to see if we need to send a challenge ack.
21464 		 */
21465 		if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21466 		    ((tp->snd_max - tp->snd_una) <= segsiz)) {
21467 			/*
21468 			 * Ok less than or right at a MSS is
21469 			 * outstanding. The original FreeBSD stack would
21470 			 * have sent a FIN, which can speed things up for
21471 			 * a transactional application doing a MSG_WAITALL.
21472 			 * To speed things up since we do *not* send a FIN
21473 			 * if data is outstanding, we send a "challenge ack".
21474 			 * The idea behind that is instead of having to have
21475 			 * the peer wait for the delayed-ack timer to run off
21476 			 * we send an ack that makes the peer send us an ack.
21477 			 */
21478 			rack_send_ack_challange(rack);
21479 		}
21480 		goto just_return;
21481 	}
21482 	/*
21483 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
21484 	 * and initialize the header from the template for sends on this
21485 	 * connection.
21486 	 */
21487 	hw_tls = tp->t_nic_ktls_xmit != 0;
21488 	if (len) {
21489 		uint32_t max_val;
21490 		uint32_t moff;
21491 
21492 		if (pace_max_seg)
21493 			max_val = pace_max_seg;
21494 		else
21495 			max_val = len;
21496 		/*
21497 		 * We allow a limit on sending with hptsi.
21498 		 */
21499 		if (len > max_val) {
21500 			mark = 7;
21501 			len = max_val;
21502 		}
21503 #ifdef INET6
21504 		if (MHLEN < hdrlen + max_linkhdr)
21505 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21506 		else
21507 #endif
21508 			m = m_gethdr(M_NOWAIT, MT_DATA);
21509 
21510 		if (m == NULL) {
21511 			SOCK_SENDBUF_UNLOCK(so);
21512 			error = ENOBUFS;
21513 			sack_rxmit = 0;
21514 			goto out;
21515 		}
21516 		m->m_data += max_linkhdr;
21517 		m->m_len = hdrlen;
21518 
21519 		/*
21520 		 * Start the m_copy functions from the closest mbuf to the
21521 		 * sb_offset in the socket buffer chain.
21522 		 */
21523 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
21524 		s_mb = mb;
21525 		s_moff = moff;
21526 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21527 			m_copydata(mb, moff, (int)len,
21528 				   mtod(m, caddr_t)+hdrlen);
21529 			/*
21530 			 * If we are not retransmitting advance the
21531 			 * sndptr to help remember the next place in
21532 			 * the sb.
21533 			 */
21534 			if (rsm == NULL)
21535 				sbsndptr_adv(sb, mb, len);
21536 			m->m_len += len;
21537 		} else {
21538 			struct sockbuf *msb;
21539 
21540 			/*
21541 			 * If we are not retransmitting pass in msb so
21542 			 * the socket buffer can be advanced. Otherwise
21543 			 * set it to NULL if its a retransmission since
21544 			 * we don't want to change the sb remembered
21545 			 * location.
21546 			 */
21547 			if (rsm == NULL)
21548 				msb = sb;
21549 			else
21550 				msb = NULL;
21551 			m->m_next = tcp_m_copym(
21552 				mb, moff, &len,
21553 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21554 				((rsm == NULL) ? hw_tls : 0));
21555 			if (len <= (tp->t_maxseg - optlen)) {
21556 				/*
21557 				 * Must have ran out of mbufs for the copy
21558 				 * shorten it to no longer need tso. Lets
21559 				 * not put on sendalot since we are low on
21560 				 * mbufs.
21561 				 */
21562 				tso = 0;
21563 			}
21564 			if (m->m_next == NULL) {
21565 				SOCK_SENDBUF_UNLOCK(so);
21566 				(void)m_free(m);
21567 				error = ENOBUFS;
21568 				sack_rxmit = 0;
21569 				goto out;
21570 			}
21571 		}
21572 		if (sack_rxmit) {
21573 			if (rsm && (rsm->r_flags & RACK_TLP)) {
21574 				/*
21575 				 * TLP should not count in retran count, but
21576 				 * in its own bin
21577 				 */
21578 				counter_u64_add(rack_tlp_retran, 1);
21579 				counter_u64_add(rack_tlp_retran_bytes, len);
21580 			} else {
21581 				tp->t_sndrexmitpack++;
21582 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21583 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21584 			}
21585 #ifdef STATS
21586 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21587 						 len);
21588 #endif
21589 		} else {
21590 			KMOD_TCPSTAT_INC(tcps_sndpack);
21591 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21592 #ifdef STATS
21593 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21594 						 len);
21595 #endif
21596 		}
21597 		/*
21598 		 * If we're sending everything we've got, set PUSH. (This
21599 		 * will keep happy those implementations which only give
21600 		 * data to the user when a buffer fills or a PUSH comes in.)
21601 		 */
21602 		if (sb_offset + len == sbused(sb) &&
21603 		    sbused(sb) &&
21604 		    !(flags & TH_SYN)) {
21605 			flags |= TH_PUSH;
21606 			add_flag |= RACK_HAD_PUSH;
21607 		}
21608 		SOCK_SENDBUF_UNLOCK(so);
21609 	} else {
21610 		SOCK_SENDBUF_UNLOCK(so);
21611 		if (tp->t_flags & TF_ACKNOW)
21612 			KMOD_TCPSTAT_INC(tcps_sndacks);
21613 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
21614 			KMOD_TCPSTAT_INC(tcps_sndctrl);
21615 		else
21616 			KMOD_TCPSTAT_INC(tcps_sndwinup);
21617 
21618 		m = m_gethdr(M_NOWAIT, MT_DATA);
21619 		if (m == NULL) {
21620 			error = ENOBUFS;
21621 			sack_rxmit = 0;
21622 			goto out;
21623 		}
21624 #ifdef INET6
21625 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21626 		    MHLEN >= hdrlen) {
21627 			M_ALIGN(m, hdrlen);
21628 		} else
21629 #endif
21630 			m->m_data += max_linkhdr;
21631 		m->m_len = hdrlen;
21632 	}
21633 	SOCK_SENDBUF_UNLOCK_ASSERT(so);
21634 	m->m_pkthdr.rcvif = (struct ifnet *)0;
21635 #ifdef MAC
21636 	mac_inpcb_create_mbuf(inp, m);
21637 #endif
21638 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21639 #ifdef INET6
21640 		if (isipv6)
21641 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21642 		else
21643 #endif				/* INET6 */
21644 #ifdef INET
21645 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21646 #endif
21647 		th = rack->r_ctl.fsb.th;
21648 		udp = rack->r_ctl.fsb.udp;
21649 		if (udp) {
21650 #ifdef INET6
21651 			if (isipv6)
21652 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21653 			else
21654 #endif				/* INET6 */
21655 				ulen = hdrlen + len - sizeof(struct ip);
21656 			udp->uh_ulen = htons(ulen);
21657 		}
21658 	} else {
21659 #ifdef INET6
21660 		if (isipv6) {
21661 			ip6 = mtod(m, struct ip6_hdr *);
21662 			if (tp->t_port) {
21663 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21664 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21665 				udp->uh_dport = tp->t_port;
21666 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21667 				udp->uh_ulen = htons(ulen);
21668 				th = (struct tcphdr *)(udp + 1);
21669 			} else
21670 				th = (struct tcphdr *)(ip6 + 1);
21671 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
21672 		} else
21673 #endif				/* INET6 */
21674 		{
21675 #ifdef INET
21676 			ip = mtod(m, struct ip *);
21677 			if (tp->t_port) {
21678 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21679 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21680 				udp->uh_dport = tp->t_port;
21681 				ulen = hdrlen + len - sizeof(struct ip);
21682 				udp->uh_ulen = htons(ulen);
21683 				th = (struct tcphdr *)(udp + 1);
21684 			} else
21685 				th = (struct tcphdr *)(ip + 1);
21686 			tcpip_fillheaders(inp, tp->t_port, ip, th);
21687 #endif
21688 		}
21689 	}
21690 	/*
21691 	 * If we are starting a connection, send ECN setup SYN packet. If we
21692 	 * are on a retransmit, we may resend those bits a number of times
21693 	 * as per RFC 3168.
21694 	 */
21695 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21696 		flags |= tcp_ecn_output_syn_sent(tp);
21697 	}
21698 	/* Also handle parallel SYN for ECN */
21699 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
21700 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21701 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21702 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21703 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
21704 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21705 #ifdef INET6
21706 		if (isipv6) {
21707 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21708 			ip6->ip6_flow |= htonl(ect << 20);
21709 		}
21710 		else
21711 #endif
21712 		{
21713 #ifdef INET
21714 			ip->ip_tos &= ~IPTOS_ECN_MASK;
21715 			ip->ip_tos |= ect;
21716 #endif
21717 		}
21718 	}
21719 	th->th_seq = htonl(rack_seq);
21720 	th->th_ack = htonl(tp->rcv_nxt);
21721 	tcp_set_flags(th, flags);
21722 	/*
21723 	 * Calculate receive window.  Don't shrink window, but avoid silly
21724 	 * window syndrome.
21725 	 * If a RST segment is sent, advertise a window of zero.
21726 	 */
21727 	if (flags & TH_RST) {
21728 		recwin = 0;
21729 	} else {
21730 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21731 		    recwin < (long)segsiz) {
21732 			recwin = 0;
21733 		}
21734 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21735 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21736 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21737 	}
21738 
21739 	/*
21740 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21741 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21742 	 * handled in syncache.
21743 	 */
21744 	if (flags & TH_SYN)
21745 		th->th_win = htons((u_short)
21746 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21747 	else {
21748 		/* Avoid shrinking window with window scaling. */
21749 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
21750 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21751 	}
21752 	/*
21753 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21754 	 * window.  This may cause the remote transmitter to stall.  This
21755 	 * flag tells soreceive() to disable delayed acknowledgements when
21756 	 * draining the buffer.  This can occur if the receiver is
21757 	 * attempting to read more data than can be buffered prior to
21758 	 * transmitting on the connection.
21759 	 */
21760 	if (th->th_win == 0) {
21761 		tp->t_sndzerowin++;
21762 		tp->t_flags |= TF_RXWIN0SENT;
21763 	} else
21764 		tp->t_flags &= ~TF_RXWIN0SENT;
21765 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
21766 	/* Now are we using fsb?, if so copy the template data to the mbuf */
21767 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21768 		uint8_t *cpto;
21769 
21770 		cpto = mtod(m, uint8_t *);
21771 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21772 		/*
21773 		 * We have just copied in:
21774 		 * IP/IP6
21775 		 * <optional udphdr>
21776 		 * tcphdr (no options)
21777 		 *
21778 		 * We need to grab the correct pointers into the mbuf
21779 		 * for both the tcp header, and possibly the udp header (if tunneling).
21780 		 * We do this by using the offset in the copy buffer and adding it
21781 		 * to the mbuf base pointer (cpto).
21782 		 */
21783 #ifdef INET6
21784 		if (isipv6)
21785 			ip6 = mtod(m, struct ip6_hdr *);
21786 		else
21787 #endif				/* INET6 */
21788 #ifdef INET
21789 			ip = mtod(m, struct ip *);
21790 #endif
21791 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21792 		/* If we have a udp header lets set it into the mbuf as well */
21793 		if (udp)
21794 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21795 	}
21796 	if (optlen) {
21797 		bcopy(opt, th + 1, optlen);
21798 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21799 	}
21800 	/*
21801 	 * Put TCP length in extended header, and then checksum extended
21802 	 * header and data.
21803 	 */
21804 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
21805 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21806 	if (to.to_flags & TOF_SIGNATURE) {
21807 		/*
21808 		 * Calculate MD5 signature and put it into the place
21809 		 * determined before.
21810 		 * NOTE: since TCP options buffer doesn't point into
21811 		 * mbuf's data, calculate offset and use it.
21812 		 */
21813 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21814 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21815 			/*
21816 			 * Do not send segment if the calculation of MD5
21817 			 * digest has failed.
21818 			 */
21819 			goto out;
21820 		}
21821 	}
21822 #endif
21823 #ifdef INET6
21824 	if (isipv6) {
21825 		/*
21826 		 * ip6_plen is not need to be filled now, and will be filled
21827 		 * in ip6_output.
21828 		 */
21829 		if (tp->t_port) {
21830 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21831 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21832 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21833 			th->th_sum = htons(0);
21834 			UDPSTAT_INC(udps_opackets);
21835 		} else {
21836 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21837 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21838 			th->th_sum = in6_cksum_pseudo(ip6,
21839 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21840 						      0);
21841 		}
21842 	}
21843 #endif
21844 #if defined(INET6) && defined(INET)
21845 	else
21846 #endif
21847 #ifdef INET
21848 	{
21849 		if (tp->t_port) {
21850 			m->m_pkthdr.csum_flags = CSUM_UDP;
21851 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21852 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21853 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21854 			th->th_sum = htons(0);
21855 			UDPSTAT_INC(udps_opackets);
21856 		} else {
21857 			m->m_pkthdr.csum_flags = CSUM_TCP;
21858 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21859 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
21860 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21861 									IPPROTO_TCP + len + optlen));
21862 		}
21863 		/* IP version must be set here for ipv4/ipv6 checking later */
21864 		KASSERT(ip->ip_v == IPVERSION,
21865 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
21866 	}
21867 #endif
21868 	/*
21869 	 * Enable TSO and specify the size of the segments. The TCP pseudo
21870 	 * header checksum is always provided. XXX: Fixme: This is currently
21871 	 * not the case for IPv6.
21872 	 */
21873 	if (tso) {
21874 		/*
21875 		 * Here we must use t_maxseg and the optlen since
21876 		 * the optlen may include SACK's (or DSACK).
21877 		 */
21878 		KASSERT(len > tp->t_maxseg - optlen,
21879 			("%s: len <= tso_segsz", __func__));
21880 		m->m_pkthdr.csum_flags |= CSUM_TSO;
21881 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21882 	}
21883 	KASSERT(len + hdrlen == m_length(m, NULL),
21884 		("%s: mbuf chain different than expected: %d + %u != %u",
21885 		 __func__, len, hdrlen, m_length(m, NULL)));
21886 
21887 #ifdef TCP_HHOOK
21888 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21889 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
21890 #endif
21891 	if ((rack->r_ctl.crte != NULL) &&
21892 	    (rack->rc_hw_nobuf == 0) &&
21893 	    tcp_bblogging_on(tp)) {
21894 		rack_log_queue_level(tp, rack, len, &tv, cts);
21895 	}
21896 	/* We're getting ready to send; log now. */
21897 	if (tcp_bblogging_on(rack->rc_tp)) {
21898 		union tcp_log_stackspecific log;
21899 
21900 		memset(&log, 0, sizeof(log));
21901 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21902 		if (rack->rack_no_prr)
21903 			log.u_bbr.flex1 = 0;
21904 		else
21905 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21906 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21907 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21908 		log.u_bbr.flex4 = orig_len;
21909 		/* Save off the early/late values */
21910 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21911 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21912 		log.u_bbr.bw_inuse = rack_get_bw(rack);
21913 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21914 		log.u_bbr.flex8 = 0;
21915 		if (rsm) {
21916 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21917 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21918 				counter_u64_add(rack_collapsed_win_rxt, 1);
21919 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21920 			}
21921 			if (doing_tlp)
21922 				log.u_bbr.flex8 = 2;
21923 			else
21924 				log.u_bbr.flex8 = 1;
21925 		} else {
21926 			if (doing_tlp)
21927 				log.u_bbr.flex8 = 3;
21928 		}
21929 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21930 		log.u_bbr.flex7 = mark;
21931 		log.u_bbr.flex7 <<= 8;
21932 		log.u_bbr.flex7 |= pass;
21933 		log.u_bbr.pkts_out = tp->t_maxseg;
21934 		log.u_bbr.timeStamp = cts;
21935 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21936 		if (rsm && (rsm->r_rtr_cnt > 0)) {
21937 			/*
21938 			 * When we have a retransmit we want to log the
21939 			 * burst at send and flight at send from before.
21940 			 */
21941 			log.u_bbr.flex5 = rsm->r_fas;
21942 			log.u_bbr.bbr_substate = rsm->r_bas;
21943 		} else {
21944 			/*
21945 			 * New transmits we log in flex5 the inflight again as
21946 			 * well as the number of segments in our send in the
21947 			 * substate field.
21948 			 */
21949 			log.u_bbr.flex5 = log.u_bbr.inflight;
21950 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21951 		}
21952 		log.u_bbr.lt_epoch = cwnd_to_use;
21953 		log.u_bbr.delivered = sendalot;
21954 		log.u_bbr.rttProp = (uintptr_t)rsm;
21955 		log.u_bbr.pkt_epoch = __LINE__;
21956 		if (rsm) {
21957 			log.u_bbr.delRate = rsm->r_flags;
21958 			log.u_bbr.delRate <<= 31;
21959 			log.u_bbr.delRate |= rack->r_must_retran;
21960 			log.u_bbr.delRate <<= 1;
21961 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21962 		} else {
21963 			log.u_bbr.delRate = rack->r_must_retran;
21964 			log.u_bbr.delRate <<= 1;
21965 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21966 		}
21967 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
21968 				    len, &log, false, NULL, __func__, __LINE__, &tv);
21969 	} else
21970 		lgb = NULL;
21971 
21972 	/*
21973 	 * Fill in IP length and desired time to live and send to IP level.
21974 	 * There should be a better way to handle ttl and tos; we could keep
21975 	 * them in the template, but need a way to checksum without them.
21976 	 */
21977 	/*
21978 	 * m->m_pkthdr.len should have been set before cksum calcuration,
21979 	 * because in6_cksum() need it.
21980 	 */
21981 #ifdef INET6
21982 	if (isipv6) {
21983 		/*
21984 		 * we separately set hoplimit for every segment, since the
21985 		 * user might want to change the value via setsockopt. Also,
21986 		 * desired default hop limit might be changed via Neighbor
21987 		 * Discovery.
21988 		 */
21989 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
21990 
21991 		/*
21992 		 * Set the packet size here for the benefit of DTrace
21993 		 * probes. ip6_output() will set it properly; it's supposed
21994 		 * to include the option header lengths as well.
21995 		 */
21996 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
21997 
21998 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
21999 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22000 		else
22001 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22002 
22003 		if (tp->t_state == TCPS_SYN_SENT)
22004 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22005 
22006 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22007 		/* TODO: IPv6 IP6TOS_ECT bit on */
22008 		error = ip6_output(m,
22009 				   inp->in6p_outputopts,
22010 				   &inp->inp_route6,
22011 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22012 				   NULL, NULL, inp);
22013 
22014 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22015 			mtu = inp->inp_route6.ro_nh->nh_mtu;
22016 	}
22017 #endif				/* INET6 */
22018 #if defined(INET) && defined(INET6)
22019 	else
22020 #endif
22021 #ifdef INET
22022 	{
22023 		ip->ip_len = htons(m->m_pkthdr.len);
22024 #ifdef INET6
22025 		if (inp->inp_vflag & INP_IPV6PROTO)
22026 			ip->ip_ttl = in6_selecthlim(inp, NULL);
22027 #endif				/* INET6 */
22028 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22029 		/*
22030 		 * If we do path MTU discovery, then we set DF on every
22031 		 * packet. This might not be the best thing to do according
22032 		 * to RFC3390 Section 2. However the tcp hostcache migitates
22033 		 * the problem so it affects only the first tcp connection
22034 		 * with a host.
22035 		 *
22036 		 * NB: Don't set DF on small MTU/MSS to have a safe
22037 		 * fallback.
22038 		 */
22039 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22040 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22041 			if (tp->t_port == 0 || len < V_tcp_minmss) {
22042 				ip->ip_off |= htons(IP_DF);
22043 			}
22044 		} else {
22045 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22046 		}
22047 
22048 		if (tp->t_state == TCPS_SYN_SENT)
22049 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22050 
22051 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
22052 
22053 		error = ip_output(m,
22054 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22055 				  inp->inp_options,
22056 #else
22057 				  NULL,
22058 #endif
22059 				  &inp->inp_route,
22060 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22061 				  inp);
22062 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22063 			mtu = inp->inp_route.ro_nh->nh_mtu;
22064 	}
22065 #endif				/* INET */
22066 	if (lgb) {
22067 		lgb->tlb_errno = error;
22068 		lgb = NULL;
22069 	}
22070 
22071 out:
22072 	/*
22073 	 * In transmit state, time the transmission and arrange for the
22074 	 * retransmit.  In persist state, just set snd_max.
22075 	 */
22076 	if ((rsm == NULL) &&  doing_tlp)
22077 		add_flag |= RACK_TLP;
22078 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22079 			rack_to_usec_ts(&tv),
22080 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22081 	if (error == 0) {
22082 		if (add_flag & RACK_IS_PCM) {
22083 			/* We just launched a PCM */
22084 			/* rrs here log */
22085 			rack->pcm_in_progress = 1;
22086 			rack->pcm_needed = 0;
22087 			rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg,  add_flag);
22088 		}
22089 		if (rsm == NULL) {
22090 			if (rack->lt_bw_up == 0) {
22091 				rack->r_ctl.lt_timemark = tcp_tv_to_lusec(&tv);
22092 				rack->r_ctl.lt_seq = tp->snd_una;
22093 				rack->lt_bw_up = 1;
22094 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22095 				/*
22096 				 * Need to record what we have since we are
22097 				 * approaching seq wrap.
22098 				 */
22099 				uint64_t tmark;
22100 
22101 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22102 				rack->r_ctl.lt_seq = tp->snd_una;
22103 				tmark = tcp_get_u64_usecs(&tv);
22104 				if (tmark > rack->r_ctl.lt_timemark) {
22105 					rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22106 					rack->r_ctl.lt_timemark = tmark;
22107 				}
22108 			}
22109 		}
22110 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
22111 		counter_u64_add(rack_total_bytes, len);
22112 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22113 		if (rsm && doing_tlp) {
22114 			rack->rc_last_sent_tlp_past_cumack = 0;
22115 			rack->rc_last_sent_tlp_seq_valid = 1;
22116 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22117 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22118 		}
22119 		if (rack->rc_hw_nobuf) {
22120 			rack->rc_hw_nobuf = 0;
22121 			rack->r_ctl.rc_agg_delayed = 0;
22122 			rack->r_early = 0;
22123 			rack->r_late = 0;
22124 			rack->r_ctl.rc_agg_early = 0;
22125 		}
22126 		if (rsm && (doing_tlp == 0)) {
22127 			/* Set we retransmitted */
22128 			rack->rc_gp_saw_rec = 1;
22129 		} else {
22130 			if (cwnd_to_use > tp->snd_ssthresh) {
22131 				/* Set we sent in CA */
22132 				rack->rc_gp_saw_ca = 1;
22133 			} else {
22134 				/* Set we sent in SS */
22135 				rack->rc_gp_saw_ss = 1;
22136 			}
22137 		}
22138 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22139 		    (tp->t_flags & TF_SACK_PERMIT) &&
22140 		    tp->rcv_numsacks > 0)
22141 			tcp_clean_dsack_blocks(tp);
22142 		tot_len_this_send += len;
22143 		if (len == 0) {
22144 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22145 		} else {
22146 			int idx;
22147 
22148 			idx = (len / segsiz) + 3;
22149 			if (idx >= TCP_MSS_ACCT_ATIMER)
22150 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22151 			else
22152 				counter_u64_add(rack_out_size[idx], 1);
22153 		}
22154 	}
22155 	if ((rack->rack_no_prr == 0) &&
22156 	    sub_from_prr &&
22157 	    (error == 0)) {
22158 		if (rack->r_ctl.rc_prr_sndcnt >= len)
22159 			rack->r_ctl.rc_prr_sndcnt -= len;
22160 		else
22161 			rack->r_ctl.rc_prr_sndcnt = 0;
22162 	}
22163 	sub_from_prr = 0;
22164 	if (rsm != NULL) {
22165 		if (doing_tlp)
22166 			/* Make sure the TLP is added */
22167 			rsm->r_flags |= RACK_TLP;
22168 		else
22169 			/* If its a resend without TLP then it must not have the flag */
22170 			rsm->r_flags &= ~RACK_TLP;
22171  	}
22172 	if ((error == 0) &&
22173 	    (len > 0) &&
22174 	    (tp->snd_una == tp->snd_max))
22175 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
22176 
22177 	{
22178 		/*
22179 		 * This block is not associated with the above error == 0 test.
22180 		 * It is used to advance snd_max if we have a new transmit.
22181 		 */
22182 		tcp_seq startseq = tp->snd_max;
22183 
22184 
22185 		if (rsm && (doing_tlp == 0))
22186 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22187 		if (error)
22188 			/* We don't log or do anything with errors */
22189 			goto nomore;
22190 		if (doing_tlp == 0) {
22191 			if (rsm == NULL) {
22192 				/*
22193 				 * Not a retransmission of some
22194 				 * sort, new data is going out so
22195 				 * clear our TLP count and flag.
22196 				 */
22197 				rack->rc_tlp_in_progress = 0;
22198 				rack->r_ctl.rc_tlp_cnt_out = 0;
22199 			}
22200 		} else {
22201 			/*
22202 			 * We have just sent a TLP, mark that it is true
22203 			 * and make sure our in progress is set so we
22204 			 * continue to check the count.
22205 			 */
22206 			rack->rc_tlp_in_progress = 1;
22207 			rack->r_ctl.rc_tlp_cnt_out++;
22208 		}
22209 		/*
22210 		 * If we are retransmitting we are done, snd_max
22211 		 * does not get updated.
22212 		 */
22213 		if (sack_rxmit)
22214 			goto nomore;
22215 		if ((tp->snd_una == tp->snd_max) && (len > 0)) {
22216 			/*
22217 			 * Update the time we just added data since
22218 			 * nothing was outstanding.
22219 			 */
22220 			rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22221 			tp->t_acktime = ticks;
22222 		}
22223 		/*
22224 		 * Now for special SYN/FIN handling.
22225 		 */
22226 		if (flags & (TH_SYN | TH_FIN)) {
22227 			if ((flags & TH_SYN) &&
22228 			    ((tp->t_flags & TF_SENTSYN) == 0)) {
22229 				tp->snd_max++;
22230 				tp->t_flags |= TF_SENTSYN;
22231 			}
22232 			if ((flags & TH_FIN) &&
22233 			    ((tp->t_flags & TF_SENTFIN) == 0)) {
22234 				tp->snd_max++;
22235 				tp->t_flags |= TF_SENTFIN;
22236 			}
22237 		}
22238 		tp->snd_max += len;
22239 		if (rack->rc_new_rnd_needed) {
22240 			rack_new_round_starts(tp, rack, tp->snd_max);
22241 		}
22242 		/*
22243 		 * Time this transmission if not a retransmission and
22244 		 * not currently timing anything.
22245 		 * This is only relevant in case of switching back to
22246 		 * the base stack.
22247 		 */
22248 		if (tp->t_rtttime == 0) {
22249 			tp->t_rtttime = ticks;
22250 			tp->t_rtseq = startseq;
22251 			KMOD_TCPSTAT_INC(tcps_segstimed);
22252 		}
22253 		if (len &&
22254 		    ((tp->t_flags & TF_GPUTINPROG) == 0))
22255 			rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22256 		/*
22257 		 * If we are doing FO we need to update the mbuf position and subtract
22258 		 * this happens when the peer sends us duplicate information and
22259 		 * we thus want to send a DSACK.
22260 		 *
22261 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22262 		 * turned off? If not then we are going to echo multiple DSACK blocks
22263 		 * out (with the TSO), which we should not be doing.
22264 		 */
22265 		if (rack->r_fast_output && len) {
22266 			if (rack->r_ctl.fsb.left_to_send > len)
22267 				rack->r_ctl.fsb.left_to_send -= len;
22268 			else
22269 				rack->r_ctl.fsb.left_to_send = 0;
22270 			if (rack->r_ctl.fsb.left_to_send < segsiz)
22271 				rack->r_fast_output = 0;
22272 			if (rack->r_fast_output) {
22273 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22274 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22275 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22276 			}
22277 		}
22278 		if (rack_pcm_blast == 0) {
22279 			if ((orig_len > len) &&
22280 			    (add_flag & RACK_IS_PCM) &&
22281 			    (len < pace_max_seg) &&
22282 			    ((pace_max_seg - len) > segsiz)) {
22283 				/*
22284 				 * We are doing a PCM measurement and we did
22285 				 * not get enough data in the TSO to meet the
22286 				 * burst requirement.
22287 				 */
22288 				uint32_t n_len;
22289 
22290 				n_len = (orig_len - len);
22291 				orig_len -= len;
22292 				pace_max_seg -= len;
22293 				len = n_len;
22294 				sb_offset = tp->snd_max - tp->snd_una;
22295 				/* Re-lock for the next spin */
22296 				SOCK_SENDBUF_LOCK(so);
22297 				goto send;
22298 			}
22299 		} else {
22300 			if ((orig_len > len) &&
22301 			    (add_flag & RACK_IS_PCM) &&
22302 			    ((orig_len - len) > segsiz)) {
22303 				/*
22304 				 * We are doing a PCM measurement and we did
22305 				 * not get enough data in the TSO to meet the
22306 				 * burst requirement.
22307 				 */
22308 				uint32_t n_len;
22309 
22310 				n_len = (orig_len - len);
22311 				orig_len -= len;
22312 				len = n_len;
22313 				sb_offset = tp->snd_max - tp->snd_una;
22314 				/* Re-lock for the next spin */
22315 				SOCK_SENDBUF_LOCK(so);
22316 				goto send;
22317 			}
22318 		}
22319 	}
22320 nomore:
22321 	if (error) {
22322 		rack->r_ctl.rc_agg_delayed = 0;
22323 		rack->r_early = 0;
22324 		rack->r_late = 0;
22325 		rack->r_ctl.rc_agg_early = 0;
22326 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
22327 		/*
22328 		 * Failures do not advance the seq counter above. For the
22329 		 * case of ENOBUFS we will fall out and retry in 1ms with
22330 		 * the hpts. Everything else will just have to retransmit
22331 		 * with the timer.
22332 		 *
22333 		 * In any case, we do not want to loop around for another
22334 		 * send without a good reason.
22335 		 */
22336 		sendalot = 0;
22337 		switch (error) {
22338 		case EPERM:
22339 		case EACCES:
22340 			tp->t_softerror = error;
22341 #ifdef TCP_ACCOUNTING
22342 			crtsc = get_cyclecount();
22343 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22344 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22345 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22346 			}
22347 			sched_unpin();
22348 #endif
22349 			return (error);
22350 		case ENOBUFS:
22351 			/*
22352 			 * Pace us right away to retry in a some
22353 			 * time
22354 			 */
22355 			if (rack->r_ctl.crte != NULL) {
22356 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22357 				if (tcp_bblogging_on(rack->rc_tp))
22358 					rack_log_queue_level(tp, rack, len, &tv, cts);
22359 			} else
22360 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22361 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22362 			if (rack->rc_enobuf < 0x7f)
22363 				rack->rc_enobuf++;
22364 			if (slot < (10 * HPTS_USEC_IN_MSEC))
22365 				slot = 10 * HPTS_USEC_IN_MSEC;
22366 			if (rack->r_ctl.crte != NULL) {
22367 				counter_u64_add(rack_saw_enobuf_hw, 1);
22368 				tcp_rl_log_enobuf(rack->r_ctl.crte);
22369 			}
22370 			counter_u64_add(rack_saw_enobuf, 1);
22371 			goto enobufs;
22372 		case EMSGSIZE:
22373 			/*
22374 			 * For some reason the interface we used initially
22375 			 * to send segments changed to another or lowered
22376 			 * its MTU. If TSO was active we either got an
22377 			 * interface without TSO capabilits or TSO was
22378 			 * turned off. If we obtained mtu from ip_output()
22379 			 * then update it and try again.
22380 			 */
22381 			if (tso)
22382 				tp->t_flags &= ~TF_TSO;
22383 			if (mtu != 0) {
22384 				int saved_mtu;
22385 
22386 				saved_mtu = tp->t_maxseg;
22387 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
22388 				if (saved_mtu > tp->t_maxseg) {
22389 					goto again;
22390 				}
22391 			}
22392 			slot = 10 * HPTS_USEC_IN_MSEC;
22393 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22394 #ifdef TCP_ACCOUNTING
22395 			crtsc = get_cyclecount();
22396 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22397 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22398 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22399 			}
22400 			sched_unpin();
22401 #endif
22402 			return (error);
22403 		case ENETUNREACH:
22404 			counter_u64_add(rack_saw_enetunreach, 1);
22405 			/* FALLTHROUGH */
22406 		case EHOSTDOWN:
22407 		case EHOSTUNREACH:
22408 		case ENETDOWN:
22409 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
22410 				tp->t_softerror = error;
22411 				error = 0;
22412 			}
22413 			/* FALLTHROUGH */
22414 		default:
22415 			slot = 10 * HPTS_USEC_IN_MSEC;
22416 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22417 #ifdef TCP_ACCOUNTING
22418 			crtsc = get_cyclecount();
22419 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22420 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22421 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22422 			}
22423 			sched_unpin();
22424 #endif
22425 			return (error);
22426 		}
22427 	} else {
22428 		rack->rc_enobuf = 0;
22429 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22430 			rack->r_ctl.retran_during_recovery += len;
22431 	}
22432 	KMOD_TCPSTAT_INC(tcps_sndtotal);
22433 
22434 	/*
22435 	 * Data sent (as far as we can tell). If this advertises a larger
22436 	 * window than any other segment, then remember the size of the
22437 	 * advertised window. Any pending ACK has now been sent.
22438 	 */
22439 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22440 		tp->rcv_adv = tp->rcv_nxt + recwin;
22441 
22442 	tp->last_ack_sent = tp->rcv_nxt;
22443 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22444 enobufs:
22445 	if (sendalot) {
22446 		/* Do we need to turn off sendalot? */
22447 		if (pace_max_seg &&
22448 		    (tot_len_this_send >= pace_max_seg)) {
22449 			/* We hit our max. */
22450 			sendalot = 0;
22451 		}
22452 	}
22453 	if ((error == 0) && (flags & TH_FIN))
22454 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22455 	if (flags & TH_RST) {
22456 		/*
22457 		 * We don't send again after sending a RST.
22458 		 */
22459 		slot = 0;
22460 		sendalot = 0;
22461 		if (error == 0)
22462 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22463 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22464 		/*
22465 		 * Get our pacing rate, if an error
22466 		 * occurred in sending (ENOBUF) we would
22467 		 * hit the else if with slot preset. Other
22468 		 * errors return.
22469 		 */
22470 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
22471 	}
22472 	/* We have sent clear the flag */
22473 	rack->r_ent_rec_ns = 0;
22474 	if (rack->r_must_retran) {
22475 		if (rsm) {
22476 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22477 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22478 				/*
22479 				 * We have retransmitted all.
22480 				 */
22481 				rack->r_must_retran = 0;
22482 				rack->r_ctl.rc_out_at_rto = 0;
22483 			}
22484 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22485 			/*
22486 			 * Sending new data will also kill
22487 			 * the loop.
22488 			 */
22489 			rack->r_must_retran = 0;
22490 			rack->r_ctl.rc_out_at_rto = 0;
22491 		}
22492 	}
22493 	rack->r_ctl.fsb.recwin = recwin;
22494 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22495 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22496 		/*
22497 		 * We hit an RTO and now have past snd_max at the RTO
22498 		 * clear all the WAS flags.
22499 		 */
22500 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22501 	}
22502 	if (slot) {
22503 		/* set the rack tcb into the slot N */
22504 		if ((error == 0) &&
22505 		    rack_use_rfo &&
22506 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22507 		    (rsm == NULL) &&
22508 		    (ipoptlen == 0) &&
22509 		    (doing_tlp == 0) &&
22510 		    rack->r_fsb_inited &&
22511 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22512 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22513 		    (rack->r_must_retran == 0) &&
22514 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22515 		    (len > 0) && (orig_len > 0) &&
22516 		    (orig_len > len) &&
22517 		    ((orig_len - len) >= segsiz) &&
22518 		    ((optlen == 0) ||
22519 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22520 			/* We can send at least one more MSS using our fsb */
22521 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22522 					       segsiz, pace_max_seg, hw_tls, flags);
22523 		} else
22524 			rack->r_fast_output = 0;
22525 		rack_log_fsb(rack, tp, so, flags,
22526 			     ipoptlen, orig_len, len, error,
22527 			     (rsm == NULL), optlen, __LINE__, 2);
22528 	} else if (sendalot) {
22529 		int ret;
22530 
22531 		sack_rxmit = 0;
22532 		if ((error == 0) &&
22533 		    rack_use_rfo &&
22534 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22535 		    (rsm == NULL) &&
22536 		    (doing_tlp == 0) &&
22537 		    (ipoptlen == 0) &&
22538 		    (rack->r_must_retran == 0) &&
22539 		    rack->r_fsb_inited &&
22540 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22541 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22542 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22543 		    (len > 0) && (orig_len > 0) &&
22544 		    (orig_len > len) &&
22545 		    ((orig_len - len) >= segsiz) &&
22546 		    ((optlen == 0) ||
22547 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22548 			/* we can use fast_output for more */
22549 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22550 					       segsiz, pace_max_seg, hw_tls, flags);
22551 			if (rack->r_fast_output) {
22552 				error = 0;
22553 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, &tot_len_this_send, &error, __LINE__);
22554 				if (ret >= 0)
22555 					return (ret);
22556 			        else if (error)
22557 					goto nomore;
22558 
22559 			}
22560 		}
22561 		goto again;
22562 	}
22563 skip_all_send:
22564 	/* Assure when we leave that snd_nxt will point to top */
22565 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22566 		tp->snd_nxt = tp->snd_max;
22567 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22568 #ifdef TCP_ACCOUNTING
22569 	crtsc = get_cyclecount() - ts_val;
22570 	if (tot_len_this_send) {
22571 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22572 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22573 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22574 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22575 		}
22576 	} else {
22577 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22578 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
22579 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22580 		}
22581 	}
22582 	sched_unpin();
22583 #endif
22584 	if (error == ENOBUFS)
22585 		error = 0;
22586 	return (error);
22587 }
22588 
22589 static void
22590 rack_update_seg(struct tcp_rack *rack)
22591 {
22592 	uint32_t orig_val;
22593 
22594 	orig_val = rack->r_ctl.rc_pace_max_segs;
22595 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22596 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
22597 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22598 }
22599 
22600 static void
22601 rack_mtu_change(struct tcpcb *tp)
22602 {
22603 	/*
22604 	 * The MSS may have changed
22605 	 */
22606 	struct tcp_rack *rack;
22607 	struct rack_sendmap *rsm;
22608 
22609 	rack = (struct tcp_rack *)tp->t_fb_ptr;
22610 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22611 		/*
22612 		 * The MTU has changed we need to resend everything
22613 		 * since all we have sent is lost. We first fix
22614 		 * up the mtu though.
22615 		 */
22616 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
22617 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
22618 		rack_remxt_tmr(tp);
22619 		rack->r_fast_output = 0;
22620 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22621 						rack->r_ctl.rc_sacked);
22622 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22623 		rack->r_must_retran = 1;
22624 		/* Mark all inflight to needing to be rxt'd */
22625 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22626 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22627 		}
22628 	}
22629 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22630 	/* We don't use snd_nxt to retransmit */
22631 	tp->snd_nxt = tp->snd_max;
22632 }
22633 
22634 static int
22635 rack_set_dgp(struct tcp_rack *rack)
22636 {
22637 	if (rack->dgp_on == 1)
22638 		return(0);
22639 	if ((rack->use_fixed_rate == 1) &&
22640 	    (rack->rc_always_pace == 1)) {
22641 		/*
22642 		 * We are already pacing another
22643 		 * way.
22644 		 */
22645 		return (EBUSY);
22646 	}
22647 	if (rack->rc_always_pace == 1) {
22648 		rack_remove_pacing(rack);
22649 	}
22650 	if (tcp_incr_dgp_pacing_cnt() == 0)
22651 		return (ENOSPC);
22652 	rack->r_ctl.pacing_method |= RACK_DGP_PACING;
22653 	rack->rc_fillcw_apply_discount = 0;
22654 	rack->dgp_on = 1;
22655 	rack->rc_always_pace = 1;
22656 	rack->rc_pace_dnd = 1;
22657 	rack->use_fixed_rate = 0;
22658 	if (rack->gp_ready)
22659 		rack_set_cc_pacing(rack);
22660 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22661 	rack->rack_attempt_hdwr_pace = 0;
22662 	/* rxt settings */
22663 	rack->full_size_rxt = 1;
22664 	rack->shape_rxt_to_pacing_min  = 0;
22665 	/* cmpack=1 */
22666 	rack->r_use_cmp_ack = 1;
22667 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22668 	    rack->r_use_cmp_ack)
22669 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22670 	/* scwnd=1 */
22671 	rack->rack_enable_scwnd = 1;
22672 	/* dynamic=100 */
22673 	rack->rc_gp_dyn_mul = 1;
22674 	/* gp_inc_ca */
22675 	rack->r_ctl.rack_per_of_gp_ca = 100;
22676 	/* rrr_conf=3 */
22677 	rack->r_rr_config = 3;
22678 	/* npush=2 */
22679 	rack->r_ctl.rc_no_push_at_mrtt = 2;
22680 	/* fillcw=1 */
22681 	rack->rc_pace_to_cwnd = 1;
22682 	rack->rc_pace_fill_if_rttin_range = 0;
22683 	rack->rtt_limit_mul = 0;
22684 	/* noprr=1 */
22685 	rack->rack_no_prr = 1;
22686 	/* lscwnd=1 */
22687 	rack->r_limit_scw = 1;
22688 	/* gp_inc_rec */
22689 	rack->r_ctl.rack_per_of_gp_rec = 90;
22690 	return (0);
22691 }
22692 
22693 static int
22694 rack_set_profile(struct tcp_rack *rack, int prof)
22695 {
22696 	int err = EINVAL;
22697 	if (prof == 1) {
22698 		/*
22699 		 * Profile 1 is "standard" DGP. It ignores
22700 		 * client buffer level.
22701 		 */
22702 		err = rack_set_dgp(rack);
22703 		if (err)
22704 			return (err);
22705 	} else if (prof == 6) {
22706 		err = rack_set_dgp(rack);
22707 		if (err)
22708 			return (err);
22709 		/*
22710 		 * Profile 6 tweaks DGP so that it will apply to
22711 		 * fill-cw the same settings that profile5 does
22712 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22713 		 */
22714 		rack->rc_fillcw_apply_discount = 1;
22715 	} else if (prof == 0) {
22716 		/* This changes things back to the default settings */
22717 		if (rack->rc_always_pace == 1) {
22718 			rack_remove_pacing(rack);
22719 		} else {
22720 			/* Make sure any stray flags are off */
22721 			rack->dgp_on = 0;
22722 			rack->rc_hybrid_mode = 0;
22723 			rack->use_fixed_rate = 0;
22724 		}
22725 		err = 0;
22726 		if (rack_fill_cw_state)
22727 			rack->rc_pace_to_cwnd = 1;
22728 		else
22729 			rack->rc_pace_to_cwnd = 0;
22730 
22731 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22732 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
22733 			rack->rc_always_pace = 1;
22734 			if (rack->rack_hibeta)
22735 				rack_set_cc_pacing(rack);
22736 		} else
22737 			rack->rc_always_pace = 0;
22738 		if (rack_dsack_std_based & 0x1) {
22739 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22740 			rack->rc_rack_tmr_std_based = 1;
22741 		}
22742 		if (rack_dsack_std_based & 0x2) {
22743 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22744 			rack->rc_rack_use_dsack = 1;
22745 		}
22746 		if (rack_use_cmp_acks)
22747 			rack->r_use_cmp_ack = 1;
22748 		else
22749 			rack->r_use_cmp_ack = 0;
22750 		if (rack_disable_prr)
22751 			rack->rack_no_prr = 1;
22752 		else
22753 			rack->rack_no_prr = 0;
22754 		if (rack_gp_no_rec_chg)
22755 			rack->rc_gp_no_rec_chg = 1;
22756 		else
22757 			rack->rc_gp_no_rec_chg = 0;
22758 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22759 			rack->r_mbuf_queue = 1;
22760 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22761 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22762 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22763 		} else {
22764 			rack->r_mbuf_queue = 0;
22765 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22766 		}
22767 		if (rack_enable_shared_cwnd)
22768 			rack->rack_enable_scwnd = 1;
22769 		else
22770 			rack->rack_enable_scwnd = 0;
22771 		if (rack_do_dyn_mul) {
22772 			/* When dynamic adjustment is on CA needs to start at 100% */
22773 			rack->rc_gp_dyn_mul = 1;
22774 			if (rack_do_dyn_mul >= 100)
22775 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22776 		} else {
22777 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22778 			rack->rc_gp_dyn_mul = 0;
22779 		}
22780 		rack->r_rr_config = 0;
22781 		rack->r_ctl.rc_no_push_at_mrtt = 0;
22782 		rack->rc_pace_fill_if_rttin_range = 0;
22783 		rack->rtt_limit_mul = 0;
22784 
22785 		if (rack_enable_hw_pacing)
22786 			rack->rack_hdw_pace_ena = 1;
22787 		else
22788 			rack->rack_hdw_pace_ena = 0;
22789 		if (rack_disable_prr)
22790 			rack->rack_no_prr = 1;
22791 		else
22792 			rack->rack_no_prr = 0;
22793 		if (rack_limits_scwnd)
22794 			rack->r_limit_scw  = 1;
22795 		else
22796 			rack->r_limit_scw  = 0;
22797 		rack_init_retransmit_value(rack, rack_rxt_controls);
22798 		err = 0;
22799 	}
22800 	return (err);
22801 }
22802 
22803 static int
22804 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22805 {
22806 	struct deferred_opt_list *dol;
22807 
22808 	dol = malloc(sizeof(struct deferred_opt_list),
22809 		     M_TCPDO, M_NOWAIT|M_ZERO);
22810 	if (dol == NULL) {
22811 		/*
22812 		 * No space yikes -- fail out..
22813 		 */
22814 		return (0);
22815 	}
22816 	dol->optname = sopt_name;
22817 	dol->optval = loptval;
22818 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22819 	return (1);
22820 }
22821 
22822 static int
22823 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22824 {
22825 #ifdef TCP_REQUEST_TRK
22826 	struct tcp_sendfile_track *sft;
22827 	struct timeval tv;
22828 	tcp_seq seq;
22829 	int err;
22830 
22831 	microuptime(&tv);
22832 
22833 	/* Make sure no fixed rate is on */
22834 	rack->use_fixed_rate = 0;
22835 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22836 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22837 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22838 	/* Now allocate or find our entry that will have these settings */
22839 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusec(&tv), 0);
22840 	if (sft == NULL) {
22841 		rack->rc_tp->tcp_hybrid_error++;
22842 		/* no space, where would it have gone? */
22843 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22844 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22845 		return (ENOSPC);
22846 	}
22847 	/* mask our internal flags */
22848 	hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
22849 	/* The seq will be snd_una + everything in the buffer */
22850 	seq = sft->start_seq;
22851 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22852 		/* Disabling hybrid pacing */
22853 		if (rack->rc_hybrid_mode) {
22854 			rack_set_profile(rack, 0);
22855 			rack->rc_tp->tcp_hybrid_stop++;
22856 		}
22857 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22858 		return (0);
22859 	}
22860 	if (rack->dgp_on == 0) {
22861 		/*
22862 		 * If we have not yet turned DGP on, do so
22863 		 * now setting pure DGP mode, no buffer level
22864 		 * response.
22865 		 */
22866 		if ((err = rack_set_profile(rack, 1)) != 0){
22867 			/* Failed to turn pacing on */
22868 			rack->rc_tp->tcp_hybrid_error++;
22869 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22870 			return (err);
22871 		}
22872 	}
22873 	/*
22874 	 * Now we must switch to hybrid mode as well which also
22875 	 * means moving to regular pacing.
22876 	 */
22877 	if (rack->rc_hybrid_mode == 0) {
22878 		/* First time */
22879 		if (tcp_can_enable_pacing()) {
22880 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
22881 			rack->rc_hybrid_mode = 1;
22882 		} else {
22883 			return (ENOSPC);
22884 		}
22885 		if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
22886 			/*
22887 			 * This should be true.
22888 			 */
22889 			tcp_dec_dgp_pacing_cnt();
22890 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
22891 		}
22892 	}
22893 	/* Now set in our flags */
22894 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22895 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22896 		sft->cspr = hybrid->cspr;
22897 	else
22898 		sft->cspr = 0;
22899 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22900 		sft->hint_maxseg = hybrid->hint_maxseg;
22901 	else
22902 		sft->hint_maxseg = 0;
22903 	rack->rc_tp->tcp_hybrid_start++;
22904 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22905 	return (0);
22906 #else
22907 	return (ENOTSUP);
22908 #endif
22909 }
22910 
22911 static int
22912 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
22913 {
22914 	/* We pulled a SSI info log out what was there */
22915 	si->bytes_transmitted = tp->t_sndbytes;
22916 	si->bytes_retransmitted = tp->t_snd_rxt_bytes;
22917 	return (0);
22918 }
22919 
22920 static int
22921 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22922 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22923 
22924 {
22925 	struct epoch_tracker et;
22926 	struct sockopt sopt;
22927 	struct cc_newreno_opts opt;
22928 	uint64_t val;
22929 	int error = 0;
22930 	uint16_t ca, ss;
22931 
22932 	switch (sopt_name) {
22933 	case TCP_RACK_SET_RXT_OPTIONS:
22934 		if (optval <= 2) {
22935 			rack_init_retransmit_value(rack, optval);
22936 		} else {
22937 			/*
22938 			 * You must send in 0, 1 or 2 all else is
22939 			 * invalid.
22940 			 */
22941 			error = EINVAL;
22942 		}
22943 		break;
22944 	case TCP_RACK_DSACK_OPT:
22945 		RACK_OPTS_INC(tcp_rack_dsack_opt);
22946 		if (optval & 0x1) {
22947 			rack->rc_rack_tmr_std_based = 1;
22948 		} else {
22949 			rack->rc_rack_tmr_std_based = 0;
22950 		}
22951 		if (optval & 0x2) {
22952 			rack->rc_rack_use_dsack = 1;
22953 		} else {
22954 			rack->rc_rack_use_dsack = 0;
22955 		}
22956 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
22957 		break;
22958 	case TCP_RACK_PACING_DIVISOR:
22959 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
22960 		if (optval == 0) {
22961 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
22962 		} else {
22963 			if (optval < RL_MIN_DIVISOR)
22964 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
22965 			else
22966 				rack->r_ctl.pace_len_divisor = optval;
22967 		}
22968 		break;
22969 	case TCP_RACK_HI_BETA:
22970 		RACK_OPTS_INC(tcp_rack_hi_beta);
22971 		if (optval > 0) {
22972 			rack->rack_hibeta = 1;
22973 			if ((optval >= 50) &&
22974 			    (optval <= 100)) {
22975 				/*
22976 				 * User wants to set a custom beta.
22977 				 */
22978 				rack->r_ctl.saved_hibeta = optval;
22979 				if (rack->rc_pacing_cc_set)
22980 					rack_undo_cc_pacing(rack);
22981 				rack->r_ctl.rc_saved_beta = optval;
22982 			}
22983 			if (rack->rc_pacing_cc_set == 0)
22984 				rack_set_cc_pacing(rack);
22985 		} else {
22986 			rack->rack_hibeta = 0;
22987 			if (rack->rc_pacing_cc_set)
22988 				rack_undo_cc_pacing(rack);
22989 		}
22990 		break;
22991 	case TCP_RACK_PACING_BETA:
22992 		error = EINVAL;
22993 		break;
22994 	case TCP_RACK_TIMER_SLOP:
22995 		RACK_OPTS_INC(tcp_rack_timer_slop);
22996 		rack->r_ctl.timer_slop = optval;
22997 		if (rack->rc_tp->t_srtt) {
22998 			/*
22999 			 * If we have an SRTT lets update t_rxtcur
23000 			 * to have the new slop.
23001 			 */
23002 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23003 					   rack_rto_min, rack_rto_max,
23004 					   rack->r_ctl.timer_slop);
23005 		}
23006 		break;
23007 	case TCP_RACK_PACING_BETA_ECN:
23008 		RACK_OPTS_INC(tcp_rack_beta_ecn);
23009 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23010 			/* This only works for newreno. */
23011 			error = EINVAL;
23012 			break;
23013 		}
23014 		if (rack->rc_pacing_cc_set) {
23015 			/*
23016 			 * Set them into the real CC module
23017 			 * whats in the rack pcb is the old values
23018 			 * to be used on restoral/
23019 			 */
23020 			sopt.sopt_dir = SOPT_SET;
23021 			opt.name = CC_NEWRENO_BETA_ECN;
23022 			opt.val = optval;
23023 			if (CC_ALGO(tp)->ctl_output != NULL)
23024 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23025 			else
23026 				error = ENOENT;
23027 		} else {
23028 			/*
23029 			 * Not pacing yet so set it into our local
23030 			 * rack pcb storage.
23031 			 */
23032 			rack->r_ctl.rc_saved_beta_ecn = optval;
23033 		}
23034 		break;
23035 	case TCP_DEFER_OPTIONS:
23036 		RACK_OPTS_INC(tcp_defer_opt);
23037 		if (optval) {
23038 			if (rack->gp_ready) {
23039 				/* Too late */
23040 				error = EINVAL;
23041 				break;
23042 			}
23043 			rack->defer_options = 1;
23044 		} else
23045 			rack->defer_options = 0;
23046 		break;
23047 	case TCP_RACK_MEASURE_CNT:
23048 		RACK_OPTS_INC(tcp_rack_measure_cnt);
23049 		if (optval && (optval <= 0xff)) {
23050 			rack->r_ctl.req_measurements = optval;
23051 		} else
23052 			error = EINVAL;
23053 		break;
23054 	case TCP_REC_ABC_VAL:
23055 		RACK_OPTS_INC(tcp_rec_abc_val);
23056 		if (optval > 0)
23057 			rack->r_use_labc_for_rec = 1;
23058 		else
23059 			rack->r_use_labc_for_rec = 0;
23060 		break;
23061 	case TCP_RACK_ABC_VAL:
23062 		RACK_OPTS_INC(tcp_rack_abc_val);
23063 		if ((optval > 0) && (optval < 255))
23064 			rack->rc_labc = optval;
23065 		else
23066 			error = EINVAL;
23067 		break;
23068 	case TCP_HDWR_UP_ONLY:
23069 		RACK_OPTS_INC(tcp_pacing_up_only);
23070 		if (optval)
23071 			rack->r_up_only = 1;
23072 		else
23073 			rack->r_up_only = 0;
23074 		break;
23075 	case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
23076 		RACK_OPTS_INC(tcp_fillcw_rate_cap);
23077 		rack->r_ctl.fillcw_cap = loptval;
23078 		break;
23079 	case TCP_PACING_RATE_CAP:
23080 		RACK_OPTS_INC(tcp_pacing_rate_cap);
23081 		if ((rack->dgp_on == 1) &&
23082 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23083 			/*
23084 			 * If we are doing DGP we need to switch
23085 			 * to using the pacing limit.
23086 			 */
23087 			if (tcp_can_enable_pacing() == 0) {
23088 				error = ENOSPC;
23089 				break;
23090 			}
23091 			/*
23092 			 * Now change up the flags and counts to be correct.
23093 			 */
23094 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23095 			tcp_dec_dgp_pacing_cnt();
23096 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23097 		}
23098 		rack->r_ctl.bw_rate_cap = loptval;
23099 		break;
23100 	case TCP_HYBRID_PACING:
23101 		if (hybrid == NULL) {
23102 			error = EINVAL;
23103 			break;
23104 		}
23105 		if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
23106 			error = EPERM;
23107 			break;
23108 		}
23109 		error = process_hybrid_pacing(rack, hybrid);
23110 		break;
23111 	case TCP_SIDECHAN_DIS:			/*  URL:scodm */
23112 		if (optval)
23113 			rack->r_ctl.side_chan_dis_mask = optval;
23114 		else
23115 			rack->r_ctl.side_chan_dis_mask = 0;
23116 		break;
23117 	case TCP_RACK_PROFILE:
23118 		RACK_OPTS_INC(tcp_profile);
23119 		error = rack_set_profile(rack, optval);
23120 		break;
23121 	case TCP_USE_CMP_ACKS:
23122 		RACK_OPTS_INC(tcp_use_cmp_acks);
23123 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23124 			/* You can't turn it off once its on! */
23125 			error = EINVAL;
23126 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23127 			rack->r_use_cmp_ack = 1;
23128 			rack->r_mbuf_queue = 1;
23129 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23130 		}
23131 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23132 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
23133 		break;
23134 	case TCP_SHARED_CWND_TIME_LIMIT:
23135 		RACK_OPTS_INC(tcp_lscwnd);
23136 		if (optval)
23137 			rack->r_limit_scw = 1;
23138 		else
23139 			rack->r_limit_scw = 0;
23140 		break;
23141 	case TCP_RACK_DGP_IN_REC:
23142 		error = EINVAL;
23143 		break;
23144  	case TCP_RACK_PACE_TO_FILL:
23145 		RACK_OPTS_INC(tcp_fillcw);
23146 		if (optval == 0)
23147 			rack->rc_pace_to_cwnd = 0;
23148 		else {
23149 			rack->rc_pace_to_cwnd = 1;
23150 		}
23151 		if ((optval >= rack_gp_rtt_maxmul) &&
23152 		    rack_gp_rtt_maxmul &&
23153 		    (optval < 0xf)) {
23154 			rack->rc_pace_fill_if_rttin_range = 1;
23155 			rack->rtt_limit_mul = optval;
23156 		} else {
23157 			rack->rc_pace_fill_if_rttin_range = 0;
23158 			rack->rtt_limit_mul = 0;
23159 		}
23160 		break;
23161 	case TCP_RACK_NO_PUSH_AT_MAX:
23162 		RACK_OPTS_INC(tcp_npush);
23163 		if (optval == 0)
23164 			rack->r_ctl.rc_no_push_at_mrtt = 0;
23165 		else if (optval < 0xff)
23166 			rack->r_ctl.rc_no_push_at_mrtt = optval;
23167 		else
23168 			error = EINVAL;
23169 		break;
23170 	case TCP_SHARED_CWND_ENABLE:
23171 		RACK_OPTS_INC(tcp_rack_scwnd);
23172 		if (optval == 0)
23173 			rack->rack_enable_scwnd = 0;
23174 		else
23175 			rack->rack_enable_scwnd = 1;
23176 		break;
23177 	case TCP_RACK_MBUF_QUEUE:
23178 		/* Now do we use the LRO mbuf-queue feature */
23179 		RACK_OPTS_INC(tcp_rack_mbufq);
23180 		if (optval || rack->r_use_cmp_ack)
23181 			rack->r_mbuf_queue = 1;
23182 		else
23183 			rack->r_mbuf_queue = 0;
23184 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23185 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23186 		else
23187 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23188 		break;
23189 	case TCP_RACK_NONRXT_CFG_RATE:
23190 		RACK_OPTS_INC(tcp_rack_cfg_rate);
23191 		if (optval == 0)
23192 			rack->rack_rec_nonrxt_use_cr = 0;
23193 		else
23194 			rack->rack_rec_nonrxt_use_cr = 1;
23195 		break;
23196 	case TCP_NO_PRR:
23197 		RACK_OPTS_INC(tcp_rack_noprr);
23198 		if (optval == 0)
23199 			rack->rack_no_prr = 0;
23200 		else if (optval == 1)
23201 			rack->rack_no_prr = 1;
23202 		else if (optval == 2)
23203 			rack->no_prr_addback = 1;
23204 		else
23205 			error = EINVAL;
23206 		break;
23207 	case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
23208 		if (optval > 0)
23209 			rack->cspr_is_fcc = 1;
23210 		else
23211 			rack->cspr_is_fcc = 0;
23212 		break;
23213 	case TCP_TIMELY_DYN_ADJ:
23214 		RACK_OPTS_INC(tcp_timely_dyn);
23215 		if (optval == 0)
23216 			rack->rc_gp_dyn_mul = 0;
23217 		else {
23218 			rack->rc_gp_dyn_mul = 1;
23219 			if (optval >= 100) {
23220 				/*
23221 				 * If the user sets something 100 or more
23222 				 * its the gp_ca value.
23223 				 */
23224 				rack->r_ctl.rack_per_of_gp_ca  = optval;
23225 			}
23226 		}
23227 		break;
23228 	case TCP_RACK_DO_DETECTION:
23229 		error = EINVAL;
23230 		break;
23231 	case TCP_RACK_TLP_USE:
23232 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23233 			error = EINVAL;
23234 			break;
23235 		}
23236 		RACK_OPTS_INC(tcp_tlp_use);
23237 		rack->rack_tlp_threshold_use = optval;
23238 		break;
23239 	case TCP_RACK_TLP_REDUCE:
23240 		/* RACK TLP cwnd reduction (bool) */
23241 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
23242 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23243 		break;
23244 		/*  Pacing related ones */
23245 	case TCP_RACK_PACE_ALWAYS:
23246 		/*
23247 		 * zero is old rack method, 1 is new
23248 		 * method using a pacing rate.
23249 		 */
23250 		RACK_OPTS_INC(tcp_rack_pace_always);
23251 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23252 			error = EPERM;
23253 			break;
23254 		}
23255 		if (optval > 0) {
23256 			if (rack->rc_always_pace) {
23257 				error = EALREADY;
23258 				break;
23259 			} else if (tcp_can_enable_pacing()) {
23260 				rack->r_ctl.pacing_method |= RACK_REG_PACING;
23261 				rack->rc_always_pace = 1;
23262 				if (rack->rack_hibeta)
23263 					rack_set_cc_pacing(rack);
23264 			}
23265 			else {
23266 				error = ENOSPC;
23267 				break;
23268 			}
23269 		} else {
23270 			if (rack->rc_always_pace == 1) {
23271 				rack_remove_pacing(rack);
23272 			}
23273 		}
23274 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23275 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23276 		else
23277 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23278 		/* A rate may be set irate or other, if so set seg size */
23279 		rack_update_seg(rack);
23280 		break;
23281 	case TCP_BBR_RACK_INIT_RATE:
23282 		RACK_OPTS_INC(tcp_initial_rate);
23283 		val = optval;
23284 		/* Change from kbits per second to bytes per second */
23285 		val *= 1000;
23286 		val /= 8;
23287 		rack->r_ctl.init_rate = val;
23288 		if (rack->rc_always_pace)
23289 			rack_update_seg(rack);
23290 		break;
23291 	case TCP_BBR_IWINTSO:
23292 		error = EINVAL;
23293 		break;
23294 	case TCP_RACK_FORCE_MSEG:
23295 		RACK_OPTS_INC(tcp_rack_force_max_seg);
23296 		if (optval)
23297 			rack->rc_force_max_seg = 1;
23298 		else
23299 			rack->rc_force_max_seg = 0;
23300 		break;
23301 	case TCP_RACK_PACE_MIN_SEG:
23302 		RACK_OPTS_INC(tcp_rack_min_seg);
23303 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23304 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23305 		break;
23306 	case TCP_RACK_PACE_MAX_SEG:
23307 		/* Max segments size in a pace in bytes */
23308 		RACK_OPTS_INC(tcp_rack_max_seg);
23309 		if ((rack->dgp_on == 1) &&
23310 		    (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23311 			/*
23312 			 * If we set a max-seg and are doing DGP then
23313 			 * we now fall under the pacing limits not the
23314 			 * DGP ones.
23315 			 */
23316 			if (tcp_can_enable_pacing() == 0) {
23317 				error = ENOSPC;
23318 				break;
23319 			}
23320 			/*
23321 			 * Now change up the flags and counts to be correct.
23322 			 */
23323 			rack->r_ctl.pacing_method |= RACK_REG_PACING;
23324 			tcp_dec_dgp_pacing_cnt();
23325 			rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23326 		}
23327 		if (optval <= MAX_USER_SET_SEG)
23328 			rack->rc_user_set_max_segs = optval;
23329 		else
23330 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23331 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23332 		break;
23333 	case TCP_RACK_PACE_RATE_REC:
23334 		/* Set the fixed pacing rate in Bytes per second ca */
23335 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23336 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23337 			error = EPERM;
23338 			break;
23339 		}
23340 		if (rack->dgp_on) {
23341 			/*
23342 			 * We are already pacing another
23343 			 * way.
23344 			 */
23345 			error = EBUSY;
23346 			break;
23347 		}
23348 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23349 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23350 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23351 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23352 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23353 		rack->use_fixed_rate = 1;
23354 		if (rack->rack_hibeta)
23355 			rack_set_cc_pacing(rack);
23356 		rack_log_pacing_delay_calc(rack,
23357 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23358 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23359 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23360 					   __LINE__, NULL,0);
23361 		break;
23362 
23363 	case TCP_RACK_PACE_RATE_SS:
23364 		/* Set the fixed pacing rate in Bytes per second ca */
23365 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23366 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23367 			error = EPERM;
23368 			break;
23369 		}
23370 		if (rack->dgp_on) {
23371 			/*
23372 			 * We are already pacing another
23373 			 * way.
23374 			 */
23375 			error = EBUSY;
23376 			break;
23377 		}
23378 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23379 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23380 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23381 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23382 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23383 		rack->use_fixed_rate = 1;
23384 		if (rack->rack_hibeta)
23385 			rack_set_cc_pacing(rack);
23386 		rack_log_pacing_delay_calc(rack,
23387 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23388 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23389 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23390 					   __LINE__, NULL, 0);
23391 		break;
23392 
23393 	case TCP_RACK_PACE_RATE_CA:
23394 		/* Set the fixed pacing rate in Bytes per second ca */
23395 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23396 		if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23397 			error = EPERM;
23398 			break;
23399 		}
23400 		if (rack->dgp_on) {
23401 			/*
23402 			 * We are already pacing another
23403 			 * way.
23404 			 */
23405 			error = EBUSY;
23406 			break;
23407 		}
23408 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23409 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23410 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23411 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23412 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23413 		rack->use_fixed_rate = 1;
23414 		if (rack->rack_hibeta)
23415 			rack_set_cc_pacing(rack);
23416 		rack_log_pacing_delay_calc(rack,
23417 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23418 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23419 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23420 					   __LINE__, NULL, 0);
23421 		break;
23422 	case TCP_RACK_GP_INCREASE_REC:
23423 		RACK_OPTS_INC(tcp_gp_inc_rec);
23424 		rack->r_ctl.rack_per_of_gp_rec = optval;
23425 		rack_log_pacing_delay_calc(rack,
23426 					   rack->r_ctl.rack_per_of_gp_ss,
23427 					   rack->r_ctl.rack_per_of_gp_ca,
23428 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23429 					   __LINE__, NULL, 0);
23430 		break;
23431 	case TCP_RACK_GP_INCREASE_CA:
23432 		RACK_OPTS_INC(tcp_gp_inc_ca);
23433 		ca = optval;
23434 		if (ca < 100) {
23435 			/*
23436 			 * We don't allow any reduction
23437 			 * over the GP b/w.
23438 			 */
23439 			error = EINVAL;
23440 			break;
23441 		}
23442 		rack->r_ctl.rack_per_of_gp_ca = ca;
23443 		rack_log_pacing_delay_calc(rack,
23444 					   rack->r_ctl.rack_per_of_gp_ss,
23445 					   rack->r_ctl.rack_per_of_gp_ca,
23446 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23447 					   __LINE__, NULL, 0);
23448 		break;
23449 	case TCP_RACK_GP_INCREASE_SS:
23450 		RACK_OPTS_INC(tcp_gp_inc_ss);
23451 		ss = optval;
23452 		if (ss < 100) {
23453 			/*
23454 			 * We don't allow any reduction
23455 			 * over the GP b/w.
23456 			 */
23457 			error = EINVAL;
23458 			break;
23459 		}
23460 		rack->r_ctl.rack_per_of_gp_ss = ss;
23461 		rack_log_pacing_delay_calc(rack,
23462 					   rack->r_ctl.rack_per_of_gp_ss,
23463 					   rack->r_ctl.rack_per_of_gp_ca,
23464 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23465 					   __LINE__, NULL, 0);
23466 		break;
23467 	case TCP_RACK_RR_CONF:
23468 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23469 		if (optval && optval <= 3)
23470 			rack->r_rr_config = optval;
23471 		else
23472 			rack->r_rr_config = 0;
23473 		break;
23474 	case TCP_PACING_DND:			/*  URL:dnd */
23475 		if (optval > 0)
23476 			rack->rc_pace_dnd = 1;
23477 		else
23478 			rack->rc_pace_dnd = 0;
23479 		break;
23480 	case TCP_HDWR_RATE_CAP:
23481 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
23482 		if (optval) {
23483 			if (rack->r_rack_hw_rate_caps == 0)
23484 				rack->r_rack_hw_rate_caps = 1;
23485 			else
23486 				error = EALREADY;
23487 		} else {
23488 			rack->r_rack_hw_rate_caps = 0;
23489 		}
23490 		break;
23491 	case TCP_DGP_UPPER_BOUNDS:
23492 	{
23493 		uint8_t val;
23494 		val = optval & 0x0000ff;
23495 		rack->r_ctl.rack_per_upper_bound_ca = val;
23496 		val = (optval >> 16) & 0x0000ff;
23497 		rack->r_ctl.rack_per_upper_bound_ss = val;
23498 		break;
23499 	}
23500 	case TCP_SS_EEXIT:			/*  URL:eexit */
23501 		if (optval > 0) {
23502 			rack->r_ctl.gp_rnd_thresh =  optval & 0x0ff;
23503 			if (optval & 0x10000) {
23504 				rack->r_ctl.gate_to_fs = 1;
23505 			} else {
23506 				rack->r_ctl.gate_to_fs = 0;
23507 			}
23508 			if (optval & 0x20000) {
23509 				rack->r_ctl.use_gp_not_last = 1;
23510 			} else {
23511 				rack->r_ctl.use_gp_not_last = 0;
23512 			}
23513 			if (optval & 0xfffc0000) {
23514 				uint32_t v;
23515 
23516 				v = (optval >> 18) & 0x00003fff;
23517 				if (v >= 1000)
23518 					rack->r_ctl.gp_gain_req = v;
23519 			}
23520 		} else {
23521 			/* We do not do ss early exit at all */
23522 			rack->rc_initial_ss_comp = 1;
23523 			rack->r_ctl.gp_rnd_thresh = 0;
23524 		}
23525 		break;
23526 	case TCP_RACK_SPLIT_LIMIT:
23527 		RACK_OPTS_INC(tcp_split_limit);
23528 		rack->r_ctl.rc_split_limit = optval;
23529 		break;
23530 	case TCP_BBR_HDWR_PACE:
23531 		RACK_OPTS_INC(tcp_hdwr_pacing);
23532 		if (optval){
23533 			if (rack->rack_hdrw_pacing == 0) {
23534 				rack->rack_hdw_pace_ena = 1;
23535 				rack->rack_attempt_hdwr_pace = 0;
23536 			} else
23537 				error = EALREADY;
23538 		} else {
23539 			rack->rack_hdw_pace_ena = 0;
23540 #ifdef RATELIMIT
23541 			if (rack->r_ctl.crte != NULL) {
23542 				rack->rack_hdrw_pacing = 0;
23543 				rack->rack_attempt_hdwr_pace = 0;
23544 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23545 				rack->r_ctl.crte = NULL;
23546 			}
23547 #endif
23548 		}
23549 		break;
23550 		/*  End Pacing related ones */
23551 	case TCP_RACK_PRR_SENDALOT:
23552 		/* Allow PRR to send more than one seg */
23553 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
23554 		rack->r_ctl.rc_prr_sendalot = optval;
23555 		break;
23556 	case TCP_RACK_MIN_TO:
23557 		/* Minimum time between rack t-o's in ms */
23558 		RACK_OPTS_INC(tcp_rack_min_to);
23559 		rack->r_ctl.rc_min_to = optval;
23560 		break;
23561 	case TCP_RACK_EARLY_SEG:
23562 		/* If early recovery max segments */
23563 		RACK_OPTS_INC(tcp_rack_early_seg);
23564 		rack->r_ctl.rc_early_recovery_segs = optval;
23565 		break;
23566 	case TCP_RACK_ENABLE_HYSTART:
23567 	{
23568 		if (optval) {
23569 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23570 			if (rack_do_hystart > RACK_HYSTART_ON)
23571 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23572 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23573 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23574 		} else {
23575 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23576 		}
23577 	}
23578 	break;
23579 	case TCP_RACK_REORD_THRESH:
23580 		/* RACK reorder threshold (shift amount) */
23581 		RACK_OPTS_INC(tcp_rack_reord_thresh);
23582 		if ((optval > 0) && (optval < 31))
23583 			rack->r_ctl.rc_reorder_shift = optval;
23584 		else
23585 			error = EINVAL;
23586 		break;
23587 	case TCP_RACK_REORD_FADE:
23588 		/* Does reordering fade after ms time */
23589 		RACK_OPTS_INC(tcp_rack_reord_fade);
23590 		rack->r_ctl.rc_reorder_fade = optval;
23591 		break;
23592 	case TCP_RACK_TLP_THRESH:
23593 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
23594 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
23595 		if (optval)
23596 			rack->r_ctl.rc_tlp_threshold = optval;
23597 		else
23598 			error = EINVAL;
23599 		break;
23600 	case TCP_BBR_USE_RACK_RR:
23601 		RACK_OPTS_INC(tcp_rack_rr);
23602 		if (optval)
23603 			rack->use_rack_rr = 1;
23604 		else
23605 			rack->use_rack_rr = 0;
23606 		break;
23607 	case TCP_RACK_PKT_DELAY:
23608 		/* RACK added ms i.e. rack-rtt + reord + N */
23609 		RACK_OPTS_INC(tcp_rack_pkt_delay);
23610 		rack->r_ctl.rc_pkt_delay = optval;
23611 		break;
23612 	case TCP_DELACK:
23613 		RACK_OPTS_INC(tcp_rack_delayed_ack);
23614 		if (optval == 0)
23615 			tp->t_delayed_ack = 0;
23616 		else
23617 			tp->t_delayed_ack = 1;
23618 		if (tp->t_flags & TF_DELACK) {
23619 			tp->t_flags &= ~TF_DELACK;
23620 			tp->t_flags |= TF_ACKNOW;
23621 			NET_EPOCH_ENTER(et);
23622 			rack_output(tp);
23623 			NET_EPOCH_EXIT(et);
23624 		}
23625 		break;
23626 
23627 	case TCP_BBR_RACK_RTT_USE:
23628 		RACK_OPTS_INC(tcp_rack_rtt_use);
23629 		if ((optval != USE_RTT_HIGH) &&
23630 		    (optval != USE_RTT_LOW) &&
23631 		    (optval != USE_RTT_AVG))
23632 			error = EINVAL;
23633 		else
23634 			rack->r_ctl.rc_rate_sample_method = optval;
23635 		break;
23636 	case TCP_HONOR_HPTS_MIN:
23637 		RACK_OPTS_INC(tcp_honor_hpts);
23638 		if (optval) {
23639 			rack->r_use_hpts_min = 1;
23640 			/*
23641 			 * Must be between 2 - 80% to be a reduction else
23642 			 * we keep the default (10%).
23643 			 */
23644 			if ((optval > 1) && (optval <= 80)) {
23645 				rack->r_ctl.max_reduction = optval;
23646 			}
23647 		} else
23648 			rack->r_use_hpts_min = 0;
23649 		break;
23650 	case TCP_REC_IS_DYN:			/*  URL:dynrec */
23651 		RACK_OPTS_INC(tcp_dyn_rec);
23652 		if (optval)
23653 			rack->rc_gp_no_rec_chg = 1;
23654 		else
23655 			rack->rc_gp_no_rec_chg = 0;
23656 		break;
23657 	case TCP_NO_TIMELY:
23658 		RACK_OPTS_INC(tcp_notimely);
23659 		if (optval) {
23660 			rack->rc_skip_timely = 1;
23661 			rack->r_ctl.rack_per_of_gp_rec = 90;
23662 			rack->r_ctl.rack_per_of_gp_ca = 100;
23663 			rack->r_ctl.rack_per_of_gp_ss = 250;
23664 		} else {
23665 			rack->rc_skip_timely = 0;
23666 		}
23667 		break;
23668 	case TCP_GP_USE_LTBW:
23669 		if (optval == 0) {
23670 			rack->use_lesser_lt_bw = 0;
23671 			rack->dis_lt_bw = 1;
23672 		} else if (optval == 1) {
23673 			rack->use_lesser_lt_bw = 1;
23674 			rack->dis_lt_bw = 0;
23675 		} else if (optval == 2) {
23676 			rack->use_lesser_lt_bw = 0;
23677 			rack->dis_lt_bw = 0;
23678 		}
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 static void
23695 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
23696 {
23697 	/*
23698 	 * A new connection has been created (tp) and
23699 	 * the parent is the inpcb given. We want to
23700 	 * apply a read-lock to the parent (we are already
23701 	 * holding a write lock on the tp) and copy anything
23702 	 * out of the rack specific data as long as its tfb is
23703 	 * the same as ours i.e. we are the same stack. Otherwise
23704 	 * we just return.
23705 	 */
23706 	struct tcpcb *par;
23707 	struct tcp_rack *dest, *src;
23708 	int cnt = 0;
23709 
23710 	par = intotcpcb(parent);
23711 	if (par->t_fb != tp->t_fb) {
23712 		/* Not the same stack */
23713 		tcp_log_socket_option(tp, 0, 0, 1);
23714 		return;
23715 	}
23716 	/* Ok if we reach here lets setup the two rack pointers */
23717 	dest = (struct tcp_rack *)tp->t_fb_ptr;
23718 	src = (struct tcp_rack *)par->t_fb_ptr;
23719 	if ((src == NULL) || (dest == NULL)) {
23720 		/* Huh? */
23721 		tcp_log_socket_option(tp, 0, 0, 2);
23722 		return;
23723 	}
23724 	/* Now copy out anything we wish to inherit i.e. things in socket-options */
23725 	/* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
23726 	if ((src->dgp_on) && (dest->dgp_on == 0)) {
23727 		/* Profile 1 had to be set via sock opt */
23728 		rack_set_dgp(dest);
23729 		cnt++;
23730 	}
23731 	/* TCP_RACK_SET_RXT_OPTIONS */
23732 	if (dest->full_size_rxt != src->full_size_rxt) {
23733 		dest->full_size_rxt = src->full_size_rxt;
23734 		cnt++;
23735 	}
23736 	if (dest->shape_rxt_to_pacing_min  != src->shape_rxt_to_pacing_min) {
23737 		dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
23738 		cnt++;
23739 	}
23740 	/* TCP_RACK_DSACK_OPT */
23741 	if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
23742 		dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
23743 		cnt++;
23744 	}
23745 	if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
23746 		dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
23747 		cnt++;
23748 	}
23749 	/* TCP_RACK_PACING_DIVISOR */
23750 	if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
23751 		dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
23752 		cnt++;
23753 	}
23754 	/* TCP_RACK_HI_BETA */
23755 	if (src->rack_hibeta != dest->rack_hibeta) {
23756 		cnt++;
23757 		if (src->rack_hibeta) {
23758 			dest->r_ctl.rc_saved_beta = src->r_ctl.rc_saved_beta;
23759 			dest->rack_hibeta = 1;
23760 		} else {
23761 			dest->rack_hibeta = 0;
23762 		}
23763 	}
23764 	/* TCP_RACK_TIMER_SLOP */
23765 	if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
23766 		dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
23767 		cnt++;
23768 	}
23769 	/* TCP_RACK_PACING_BETA_ECN */
23770 	if (dest->r_ctl.rc_saved_beta_ecn != src->r_ctl.rc_saved_beta_ecn) {
23771 		dest->r_ctl.rc_saved_beta_ecn = src->r_ctl.rc_saved_beta_ecn;
23772 		cnt++;
23773 	}
23774 	/* We do not do TCP_DEFER_OPTIONS */
23775 	/* TCP_RACK_MEASURE_CNT */
23776 	if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
23777 		dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
23778 		cnt++;
23779 	}
23780 	/* TCP_HDWR_UP_ONLY */
23781 	if (dest->r_up_only != src->r_up_only) {
23782 		dest->r_up_only = src->r_up_only;
23783 		cnt++;
23784 	}
23785 	/* TCP_FILLCW_RATE_CAP */
23786 	if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
23787 		dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
23788 		cnt++;
23789 	}
23790 	/* TCP_PACING_RATE_CAP */
23791 	if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
23792 		dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
23793 		cnt++;
23794 	}
23795 	/* A listener can't set TCP_HYBRID_PACING */
23796 	/* TCP_SIDECHAN_DIS */
23797 	if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
23798 		dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
23799 		cnt++;
23800 	}
23801 	/* TCP_SHARED_CWND_TIME_LIMIT */
23802 	if (dest->r_limit_scw != src->r_limit_scw) {
23803 		dest->r_limit_scw = src->r_limit_scw;
23804 		cnt++;
23805 	}
23806 	/* TCP_RACK_PACE_TO_FILL */
23807 	if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
23808 		dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
23809 		cnt++;
23810 	}
23811 	if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
23812 		dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
23813 		cnt++;
23814 	}
23815 	if (dest->rtt_limit_mul != src->rtt_limit_mul) {
23816 		dest->rtt_limit_mul = src->rtt_limit_mul;
23817 		cnt++;
23818 	}
23819 	/* TCP_RACK_NO_PUSH_AT_MAX */
23820 	if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
23821 		dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
23822 		cnt++;
23823 	}
23824 	/* TCP_SHARED_CWND_ENABLE */
23825 	if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
23826 		dest->rack_enable_scwnd = src->rack_enable_scwnd;
23827 		cnt++;
23828 	}
23829 	/* TCP_USE_CMP_ACKS */
23830 	if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
23831 		dest->r_use_cmp_ack = src->r_use_cmp_ack;
23832 		cnt++;
23833 	}
23834 
23835 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23836 		dest->r_mbuf_queue = src->r_mbuf_queue;
23837 		cnt++;
23838 	}
23839 	/* TCP_RACK_MBUF_QUEUE */
23840 	if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23841 		dest->r_mbuf_queue = src->r_mbuf_queue;
23842 		cnt++;
23843 	}
23844 	if  (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
23845 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23846 	} else {
23847 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23848 	}
23849 	if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
23850 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
23851 	}
23852 	/* TCP_RACK_NONRXT_CFG_RATE */
23853 	if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
23854 		dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
23855 		cnt++;
23856 	}
23857 	/* TCP_NO_PRR */
23858 	if (dest->rack_no_prr != src->rack_no_prr) {
23859 		dest->rack_no_prr = src->rack_no_prr;
23860 		cnt++;
23861 	}
23862 	if (dest->no_prr_addback != src->no_prr_addback) {
23863 		dest->no_prr_addback = src->no_prr_addback;
23864 		cnt++;
23865 	}
23866 	/* RACK_CSPR_IS_FCC */
23867 	if (dest->cspr_is_fcc != src->cspr_is_fcc) {
23868 		dest->cspr_is_fcc = src->cspr_is_fcc;
23869 		cnt++;
23870 	}
23871 	/* TCP_TIMELY_DYN_ADJ */
23872 	if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
23873 		dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
23874 		cnt++;
23875 	}
23876 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23877 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23878 		cnt++;
23879 	}
23880 	/* TCP_RACK_TLP_USE */
23881 	if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
23882 		dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
23883 		cnt++;
23884 	}
23885 	/* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
23886 	/* TCP_BBR_RACK_INIT_RATE */
23887 	if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
23888 		dest->r_ctl.init_rate = src->r_ctl.init_rate;
23889 		cnt++;
23890 	}
23891 	/* TCP_RACK_FORCE_MSEG */
23892 	if (dest->rc_force_max_seg != src->rc_force_max_seg) {
23893 		dest->rc_force_max_seg = src->rc_force_max_seg;
23894 		cnt++;
23895 	}
23896 	/* TCP_RACK_PACE_MIN_SEG */
23897 	if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
23898 		dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
23899 		cnt++;
23900 	}
23901 	/* we don't allow TCP_RACK_PACE_MAX_SEG */
23902 	/* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS,  TCP_RACK_PACE_RATE_CA */
23903 	if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
23904 		dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
23905 		cnt++;
23906 	}
23907 	if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
23908 		dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
23909 		cnt++;
23910 	}
23911 	if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
23912 		dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
23913 		cnt++;
23914 	}
23915 	/* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
23916 	if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
23917 		dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
23918 		cnt++;
23919 	}
23920 	if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23921 		dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23922 		cnt++;
23923 	}
23924 
23925 	if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
23926 		dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
23927 		cnt++;
23928 	}
23929 	/* TCP_RACK_RR_CONF */
23930 	if (dest->r_rr_config != src->r_rr_config) {
23931 		dest->r_rr_config = src->r_rr_config;
23932 		cnt++;
23933 	}
23934 	/* TCP_PACING_DND */
23935 	if (dest->rc_pace_dnd != src->rc_pace_dnd) {
23936 		dest->rc_pace_dnd = src->rc_pace_dnd;
23937 		cnt++;
23938 	}
23939 	/* TCP_HDWR_RATE_CAP */
23940 	if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
23941 		dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
23942 		cnt++;
23943 	}
23944 	/* TCP_DGP_UPPER_BOUNDS */
23945 	if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
23946 		dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
23947 		cnt++;
23948 	}
23949 	if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
23950 		dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
23951 		cnt++;
23952 	}
23953 	/* TCP_SS_EEXIT */
23954 	if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
23955 		dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
23956 		cnt++;
23957 	}
23958 	if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
23959 		dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
23960 		cnt++;
23961 	}
23962 	if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
23963 		dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
23964 		cnt++;
23965 	}
23966 	if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
23967 		dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
23968 		cnt++;
23969 	}
23970 	/* TCP_BBR_HDWR_PACE */
23971 	if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
23972 		dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
23973 		cnt++;
23974 	}
23975 	if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
23976 		dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
23977 		cnt++;
23978 	}
23979 	/* TCP_RACK_PRR_SENDALOT */
23980 	if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
23981 		dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
23982 		cnt++;
23983 	}
23984 	/* TCP_RACK_MIN_TO */
23985 	if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
23986 		dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
23987 		cnt++;
23988 	}
23989 	/* TCP_RACK_EARLY_SEG */
23990 	if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
23991 		dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
23992 		cnt++;
23993 	}
23994 	/* TCP_RACK_ENABLE_HYSTART */
23995 	if (par->t_ccv.flags != tp->t_ccv.flags) {
23996 		cnt++;
23997 		if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
23998 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23999 			if (rack_do_hystart > RACK_HYSTART_ON)
24000 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
24001 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
24002 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
24003 		} else {
24004 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
24005 		}
24006 	}
24007 	/* TCP_RACK_REORD_THRESH */
24008 	if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
24009 		dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
24010 		cnt++;
24011 	}
24012 	/* TCP_RACK_REORD_FADE */
24013 	if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
24014 		dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
24015 		cnt++;
24016 	}
24017 	/* TCP_RACK_TLP_THRESH */
24018 	if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
24019 		dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
24020 		cnt++;
24021 	}
24022 	/* TCP_BBR_USE_RACK_RR */
24023 	if (dest->use_rack_rr != src->use_rack_rr) {
24024 		dest->use_rack_rr = src->use_rack_rr;
24025 		cnt++;
24026 	}
24027 	/* TCP_RACK_PKT_DELAY */
24028 	if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
24029 		dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
24030 		cnt++;
24031 	}
24032 	/* TCP_DELACK will get copied via the main code if applicable */
24033 	/* TCP_BBR_RACK_RTT_USE */
24034 	if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
24035 		dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
24036 		cnt++;
24037 	}
24038 	/* TCP_HONOR_HPTS_MIN */
24039 	if (dest->r_use_hpts_min != src->r_use_hpts_min) {
24040 		dest->r_use_hpts_min = src->r_use_hpts_min;
24041 		cnt++;
24042 	}
24043 	if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
24044 		dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
24045 		cnt++;
24046 	}
24047 	/* TCP_REC_IS_DYN */
24048 	if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
24049 		dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
24050 		cnt++;
24051 	}
24052 	if (dest->rc_skip_timely != src->rc_skip_timely) {
24053 		dest->rc_skip_timely = src->rc_skip_timely;
24054 		cnt++;
24055 	}
24056 	/* TCP_DATA_AFTER_CLOSE */
24057 	if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
24058 		dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
24059 		cnt++;
24060 	}
24061 	/* TCP_GP_USE_LTBW */
24062 	if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
24063 		dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
24064 		cnt++;
24065 	}
24066 	if (dest->dis_lt_bw != src->dis_lt_bw) {
24067 		dest->dis_lt_bw = src->dis_lt_bw;
24068 		cnt++;
24069 	}
24070 	tcp_log_socket_option(tp, 0, cnt, 0);
24071 }
24072 
24073 
24074 static void
24075 rack_apply_deferred_options(struct tcp_rack *rack)
24076 {
24077 	struct deferred_opt_list *dol, *sdol;
24078 	uint32_t s_optval;
24079 
24080 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
24081 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
24082 		/* Disadvantage of deferal is you loose the error return */
24083 		s_optval = (uint32_t)dol->optval;
24084 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
24085 		free(dol, M_TCPDO);
24086 	}
24087 }
24088 
24089 static void
24090 rack_hw_tls_change(struct tcpcb *tp, int chg)
24091 {
24092 	/* Update HW tls state */
24093 	struct tcp_rack *rack;
24094 
24095 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24096 	if (chg)
24097 		rack->r_ctl.fsb.hw_tls = 1;
24098 	else
24099 		rack->r_ctl.fsb.hw_tls = 0;
24100 }
24101 
24102 static int
24103 rack_pru_options(struct tcpcb *tp, int flags)
24104 {
24105 	if (flags & PRUS_OOB)
24106 		return (EOPNOTSUPP);
24107 	return (0);
24108 }
24109 
24110 static bool
24111 rack_wake_check(struct tcpcb *tp)
24112 {
24113 	struct tcp_rack *rack;
24114 	struct timeval tv;
24115 	uint32_t cts;
24116 
24117 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24118 	if (rack->r_ctl.rc_hpts_flags) {
24119 		cts = tcp_get_usecs(&tv);
24120 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
24121 			/*
24122 			 * Pacing timer is up, check if we are ready.
24123 			 */
24124 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
24125 				return (true);
24126 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
24127 			/*
24128 			 * A timer is up, check if we are ready.
24129 			 */
24130 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
24131 				return (true);
24132 		}
24133 	}
24134 	return (false);
24135 }
24136 
24137 static struct tcp_function_block __tcp_rack = {
24138 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
24139 	.tfb_tcp_output = rack_output,
24140 	.tfb_do_queued_segments = ctf_do_queued_segments,
24141 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
24142 	.tfb_tcp_do_segment = rack_do_segment,
24143 	.tfb_tcp_ctloutput = rack_ctloutput,
24144 	.tfb_tcp_fb_init = rack_init,
24145 	.tfb_tcp_fb_fini = rack_fini,
24146 	.tfb_tcp_timer_stop_all = rack_stopall,
24147 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
24148 	.tfb_tcp_handoff_ok = rack_handoff_ok,
24149 	.tfb_tcp_mtu_chg = rack_mtu_change,
24150 	.tfb_pru_options = rack_pru_options,
24151 	.tfb_hwtls_change = rack_hw_tls_change,
24152 	.tfb_chg_query = rack_chg_query,
24153 	.tfb_switch_failed = rack_switch_failed,
24154 	.tfb_early_wake_check = rack_wake_check,
24155 	.tfb_compute_pipe = rack_compute_pipe,
24156 	.tfb_stack_info = rack_stack_information,
24157 	.tfb_inherit = rack_inherit,
24158 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP | TCP_FUNC_DEFAULT_OK,
24159 
24160 };
24161 
24162 /*
24163  * rack_ctloutput() must drop the inpcb lock before performing copyin on
24164  * socket option arguments.  When it re-acquires the lock after the copy, it
24165  * has to revalidate that the connection is still valid for the socket
24166  * option.
24167  */
24168 static int
24169 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24170 {
24171 	struct inpcb *inp = tptoinpcb(tp);
24172 #ifdef INET
24173 	struct ip *ip;
24174 #endif
24175 	struct tcp_rack *rack;
24176 	struct tcp_hybrid_req hybrid;
24177 	uint64_t loptval;
24178 	int32_t error = 0, optval;
24179 
24180 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24181 	if (rack == NULL) {
24182 		INP_WUNLOCK(inp);
24183 		return (EINVAL);
24184 	}
24185 #ifdef INET
24186 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
24187 #endif
24188 
24189 	switch (sopt->sopt_level) {
24190 #ifdef INET6
24191 	case IPPROTO_IPV6:
24192 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
24193 		switch (sopt->sopt_name) {
24194 		case IPV6_USE_MIN_MTU:
24195 			tcp6_use_min_mtu(tp);
24196 			break;
24197 		}
24198 		INP_WUNLOCK(inp);
24199 		return (0);
24200 #endif
24201 #ifdef INET
24202 	case IPPROTO_IP:
24203 		switch (sopt->sopt_name) {
24204 		case IP_TOS:
24205 			/*
24206 			 * The DSCP codepoint has changed, update the fsb.
24207 			 */
24208 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
24209 			break;
24210 		case IP_TTL:
24211 			/*
24212 			 * The TTL has changed, update the fsb.
24213 			 */
24214 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
24215 			break;
24216 		}
24217 		INP_WUNLOCK(inp);
24218 		return (0);
24219 #endif
24220 #ifdef SO_PEERPRIO
24221 	case SOL_SOCKET:
24222 		switch (sopt->sopt_name) {
24223 		case SO_PEERPRIO:			/*  SC-URL:bs */
24224 			/* Already read in and sanity checked in sosetopt(). */
24225 			if (inp->inp_socket) {
24226 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
24227 			}
24228 			break;
24229 		}
24230 		INP_WUNLOCK(inp);
24231 		return (0);
24232 #endif
24233 	case IPPROTO_TCP:
24234 		switch (sopt->sopt_name) {
24235 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
24236 		/*  Pacing related ones */
24237 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
24238 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
24239 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
24240 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
24241 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
24242 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
24243 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
24244 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
24245 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
24246 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
24247 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
24248 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
24249 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
24250 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
24251 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
24252 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
24253 		case TCP_FILLCW_RATE_CAP:		/*  URL:fillcw_cap */
24254 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
24255 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
24256 			/* End pacing related */
24257 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
24258 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
24259 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
24260 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
24261 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
24262 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
24263 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
24264 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
24265 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
24266 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
24267 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
24268 		case TCP_NO_PRR:			/*  URL:noprr */
24269 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
24270 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
24271 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
24272 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
24273 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
24274 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
24275 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
24276 		case TCP_RACK_PROFILE:			/*  URL:profile */
24277 		case TCP_SIDECHAN_DIS:			/*  URL:scodm */
24278 		case TCP_HYBRID_PACING:			/*  URL:pacing=hybrid */
24279 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
24280 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
24281 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
24282 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
24283 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
24284 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
24285 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
24286 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
24287 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
24288 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
24289 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
24290 		case TCP_SS_EEXIT:			/*  URL:eexit */
24291 		case TCP_DGP_UPPER_BOUNDS:		/*  URL:upper */
24292 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
24293 		case TCP_PACING_DND:			/*  URL:dnd */
24294 		case TCP_NO_TIMELY:			/*  URL:notimely */
24295 		case RACK_CSPR_IS_FCC:			/*  URL:csprisfcc */
24296 		case TCP_HONOR_HPTS_MIN:		/*  URL:hptsmin */
24297 		case TCP_REC_IS_DYN:			/*  URL:dynrec */
24298 		case TCP_GP_USE_LTBW:			/*  URL:useltbw */
24299 			goto process_opt;
24300 			break;
24301 		default:
24302 			/* Filter off all unknown options to the base stack */
24303 			return (tcp_default_ctloutput(tp, sopt));
24304 			break;
24305 		}
24306 	default:
24307 		INP_WUNLOCK(inp);
24308 		return (0);
24309 	}
24310 process_opt:
24311 	INP_WUNLOCK(inp);
24312 	if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24313 	    (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
24314 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
24315 		/*
24316 		 * We truncate it down to 32 bits for the socket-option trace this
24317 		 * means rates > 34Gbps won't show right, but thats probably ok.
24318 		 */
24319 		optval = (uint32_t)loptval;
24320 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
24321 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
24322 	} else {
24323 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
24324 		/* Save it in 64 bit form too */
24325 		loptval = optval;
24326 	}
24327 	if (error)
24328 		return (error);
24329 	INP_WLOCK(inp);
24330 	if (tp->t_fb != &__tcp_rack) {
24331 		INP_WUNLOCK(inp);
24332 		return (ENOPROTOOPT);
24333 	}
24334 	if (rack->defer_options && (rack->gp_ready == 0) &&
24335 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
24336 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
24337 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
24338 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
24339 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
24340 		/* Options are being deferred */
24341 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
24342 			INP_WUNLOCK(inp);
24343 			return (0);
24344 		} else {
24345 			/* No memory to defer, fail */
24346 			INP_WUNLOCK(inp);
24347 			return (ENOMEM);
24348 		}
24349 	}
24350 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
24351 	INP_WUNLOCK(inp);
24352 	return (error);
24353 }
24354 
24355 static void
24356 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
24357 {
24358 
24359 	INP_WLOCK_ASSERT(tptoinpcb(tp));
24360 	bzero(ti, sizeof(*ti));
24361 
24362 	ti->tcpi_state = tp->t_state;
24363 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
24364 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
24365 	if (tp->t_flags & TF_SACK_PERMIT)
24366 		ti->tcpi_options |= TCPI_OPT_SACK;
24367 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
24368 		ti->tcpi_options |= TCPI_OPT_WSCALE;
24369 		ti->tcpi_snd_wscale = tp->snd_scale;
24370 		ti->tcpi_rcv_wscale = tp->rcv_scale;
24371 	}
24372 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
24373 		ti->tcpi_options |= TCPI_OPT_ECN;
24374 	if (tp->t_flags & TF_FASTOPEN)
24375 		ti->tcpi_options |= TCPI_OPT_TFO;
24376 	/* still kept in ticks is t_rcvtime */
24377 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
24378 	/* Since we hold everything in precise useconds this is easy */
24379 	ti->tcpi_rtt = tp->t_srtt;
24380 	ti->tcpi_rttvar = tp->t_rttvar;
24381 	ti->tcpi_rto = tp->t_rxtcur;
24382 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
24383 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
24384 	/*
24385 	 * FreeBSD-specific extension fields for tcp_info.
24386 	 */
24387 	ti->tcpi_rcv_space = tp->rcv_wnd;
24388 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
24389 	ti->tcpi_snd_wnd = tp->snd_wnd;
24390 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
24391 	ti->tcpi_snd_nxt = tp->snd_nxt;
24392 	ti->tcpi_snd_mss = tp->t_maxseg;
24393 	ti->tcpi_rcv_mss = tp->t_maxseg;
24394 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24395 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24396 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24397 	ti->tcpi_total_tlp = tp->t_sndtlppack;
24398 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24399 	ti->tcpi_rttmin = tp->t_rttlow;
24400 #ifdef NETFLIX_STATS
24401 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24402 #endif
24403 #ifdef TCP_OFFLOAD
24404 	if (tp->t_flags & TF_TOE) {
24405 		ti->tcpi_options |= TCPI_OPT_TOE;
24406 		tcp_offload_tcp_info(tp, ti);
24407 	}
24408 #endif
24409 }
24410 
24411 static int
24412 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24413 {
24414 	struct inpcb *inp = tptoinpcb(tp);
24415 	struct tcp_rack *rack;
24416 	int32_t error, optval;
24417 	uint64_t val, loptval;
24418 	struct	tcp_info ti;
24419 	/*
24420 	 * Because all our options are either boolean or an int, we can just
24421 	 * pull everything into optval and then unlock and copy. If we ever
24422 	 * add a option that is not a int, then this will have quite an
24423 	 * impact to this routine.
24424 	 */
24425 	error = 0;
24426 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24427 	if (rack == NULL) {
24428 		INP_WUNLOCK(inp);
24429 		return (EINVAL);
24430 	}
24431 	switch (sopt->sopt_name) {
24432 	case TCP_INFO:
24433 		/* First get the info filled */
24434 		rack_fill_info(tp, &ti);
24435 		/* Fix up the rtt related fields if needed */
24436 		INP_WUNLOCK(inp);
24437 		error = sooptcopyout(sopt, &ti, sizeof ti);
24438 		return (error);
24439 	/*
24440 	 * Beta is the congestion control value for NewReno that influences how
24441 	 * much of a backoff happens when loss is detected. It is normally set
24442 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24443 	 * when you exit recovery.
24444 	 */
24445 	case TCP_RACK_PACING_BETA:
24446 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24447 			error = EINVAL;
24448 		else if (rack->rc_pacing_cc_set == 0)
24449 			optval = rack->r_ctl.rc_saved_beta;
24450 		else {
24451 			/*
24452 			 * Reach out into the CC data and report back what
24453 			 * I have previously set. Yeah it looks hackish but
24454 			 * we don't want to report the saved values.
24455 			 */
24456 			if (tp->t_ccv.cc_data)
24457 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24458 			else
24459 				error = EINVAL;
24460 		}
24461 		break;
24462 	/*
24463 	 * Beta_ecn is the congestion control value for NewReno that influences how
24464 	 * much of a backoff happens when a ECN mark is detected. It is normally set
24465 	 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24466 	 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24467 	 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24468 	 */
24469 	case TCP_RACK_PACING_BETA_ECN:
24470 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24471 			error = EINVAL;
24472 		else if (rack->rc_pacing_cc_set == 0)
24473 			optval = rack->r_ctl.rc_saved_beta_ecn;
24474 		else {
24475 			/*
24476 			 * Reach out into the CC data and report back what
24477 			 * I have previously set. Yeah it looks hackish but
24478 			 * we don't want to report the saved values.
24479 			 */
24480 			if (tp->t_ccv.cc_data)
24481 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24482 			else
24483 				error = EINVAL;
24484 		}
24485 		break;
24486 	case TCP_RACK_DSACK_OPT:
24487 		optval = 0;
24488 		if (rack->rc_rack_tmr_std_based) {
24489 			optval |= 1;
24490 		}
24491 		if (rack->rc_rack_use_dsack) {
24492 			optval |= 2;
24493 		}
24494 		break;
24495 	case TCP_RACK_ENABLE_HYSTART:
24496 	{
24497 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24498 			optval = RACK_HYSTART_ON;
24499 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24500 				optval = RACK_HYSTART_ON_W_SC;
24501 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24502 				optval = RACK_HYSTART_ON_W_SC_C;
24503 		} else {
24504 			optval = RACK_HYSTART_OFF;
24505 		}
24506 	}
24507 	break;
24508 	case TCP_RACK_DGP_IN_REC:
24509 		error = EINVAL;
24510 		break;
24511 	case TCP_RACK_HI_BETA:
24512 		optval = rack->rack_hibeta;
24513 		break;
24514 	case TCP_DEFER_OPTIONS:
24515 		optval = rack->defer_options;
24516 		break;
24517 	case TCP_RACK_MEASURE_CNT:
24518 		optval = rack->r_ctl.req_measurements;
24519 		break;
24520 	case TCP_REC_ABC_VAL:
24521 		optval = rack->r_use_labc_for_rec;
24522 		break;
24523 	case TCP_RACK_ABC_VAL:
24524 		optval = rack->rc_labc;
24525 		break;
24526 	case TCP_HDWR_UP_ONLY:
24527 		optval= rack->r_up_only;
24528 		break;
24529 	case TCP_FILLCW_RATE_CAP:
24530 		loptval = rack->r_ctl.fillcw_cap;
24531 		break;
24532 	case TCP_PACING_RATE_CAP:
24533 		loptval = rack->r_ctl.bw_rate_cap;
24534 		break;
24535 	case TCP_RACK_PROFILE:
24536 		/* You cannot retrieve a profile, its write only */
24537 		error = EINVAL;
24538 		break;
24539 	case TCP_SIDECHAN_DIS:
24540 		optval = rack->r_ctl.side_chan_dis_mask;
24541 		break;
24542 	case TCP_HYBRID_PACING:
24543 		/* You cannot retrieve hybrid pacing information, its write only */
24544 		error = EINVAL;
24545 		break;
24546 	case TCP_USE_CMP_ACKS:
24547 		optval = rack->r_use_cmp_ack;
24548 		break;
24549 	case TCP_RACK_PACE_TO_FILL:
24550 		optval = rack->rc_pace_to_cwnd;
24551 		break;
24552 	case TCP_RACK_NO_PUSH_AT_MAX:
24553 		optval = rack->r_ctl.rc_no_push_at_mrtt;
24554 		break;
24555 	case TCP_SHARED_CWND_ENABLE:
24556 		optval = rack->rack_enable_scwnd;
24557 		break;
24558 	case TCP_RACK_NONRXT_CFG_RATE:
24559 		optval = rack->rack_rec_nonrxt_use_cr;
24560 		break;
24561 	case TCP_NO_PRR:
24562 		if (rack->rack_no_prr  == 1)
24563 			optval = 1;
24564 		else if (rack->no_prr_addback == 1)
24565 			optval = 2;
24566 		else
24567 			optval = 0;
24568 		break;
24569 	case TCP_GP_USE_LTBW:
24570 		if (rack->dis_lt_bw) {
24571 			/* It is not used */
24572 			optval = 0;
24573 		} else if (rack->use_lesser_lt_bw) {
24574 			/* we use min() */
24575 			optval = 1;
24576 		} else {
24577 			/* we use max() */
24578 			optval = 2;
24579 		}
24580 		break;
24581 	case TCP_RACK_DO_DETECTION:
24582 		error = EINVAL;
24583 		break;
24584 	case TCP_RACK_MBUF_QUEUE:
24585 		/* Now do we use the LRO mbuf-queue feature */
24586 		optval = rack->r_mbuf_queue;
24587 		break;
24588 	case RACK_CSPR_IS_FCC:
24589 		optval = rack->cspr_is_fcc;
24590 		break;
24591 	case TCP_TIMELY_DYN_ADJ:
24592 		optval = rack->rc_gp_dyn_mul;
24593 		break;
24594 	case TCP_BBR_IWINTSO:
24595 		error = EINVAL;
24596 		break;
24597 	case TCP_RACK_TLP_REDUCE:
24598 		/* RACK TLP cwnd reduction (bool) */
24599 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24600 		break;
24601 	case TCP_BBR_RACK_INIT_RATE:
24602 		val = rack->r_ctl.init_rate;
24603 		/* convert to kbits per sec */
24604 		val *= 8;
24605 		val /= 1000;
24606 		optval = (uint32_t)val;
24607 		break;
24608 	case TCP_RACK_FORCE_MSEG:
24609 		optval = rack->rc_force_max_seg;
24610 		break;
24611 	case TCP_RACK_PACE_MIN_SEG:
24612 		optval = rack->r_ctl.rc_user_set_min_segs;
24613 		break;
24614 	case TCP_RACK_PACE_MAX_SEG:
24615 		/* Max segments in a pace */
24616 		optval = rack->rc_user_set_max_segs;
24617 		break;
24618 	case TCP_RACK_PACE_ALWAYS:
24619 		/* Use the always pace method */
24620 		optval = rack->rc_always_pace;
24621 		break;
24622 	case TCP_RACK_PRR_SENDALOT:
24623 		/* Allow PRR to send more than one seg */
24624 		optval = rack->r_ctl.rc_prr_sendalot;
24625 		break;
24626 	case TCP_RACK_MIN_TO:
24627 		/* Minimum time between rack t-o's in ms */
24628 		optval = rack->r_ctl.rc_min_to;
24629 		break;
24630 	case TCP_RACK_SPLIT_LIMIT:
24631 		optval = rack->r_ctl.rc_split_limit;
24632 		break;
24633 	case TCP_RACK_EARLY_SEG:
24634 		/* If early recovery max segments */
24635 		optval = rack->r_ctl.rc_early_recovery_segs;
24636 		break;
24637 	case TCP_RACK_REORD_THRESH:
24638 		/* RACK reorder threshold (shift amount) */
24639 		optval = rack->r_ctl.rc_reorder_shift;
24640 		break;
24641 	case TCP_SS_EEXIT:
24642 		if (rack->r_ctl.gp_rnd_thresh) {
24643 			uint32_t v;
24644 
24645 			v = rack->r_ctl.gp_gain_req;
24646 			v <<= 17;
24647 			optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
24648 			if (rack->r_ctl.gate_to_fs == 1)
24649 				optval |= 0x10000;
24650 		} else
24651 			optval = 0;
24652 		break;
24653 	case TCP_RACK_REORD_FADE:
24654 		/* Does reordering fade after ms time */
24655 		optval = rack->r_ctl.rc_reorder_fade;
24656 		break;
24657 	case TCP_BBR_USE_RACK_RR:
24658 		/* Do we use the rack cheat for rxt */
24659 		optval = rack->use_rack_rr;
24660 		break;
24661 	case TCP_RACK_RR_CONF:
24662 		optval = rack->r_rr_config;
24663 		break;
24664 	case TCP_HDWR_RATE_CAP:
24665 		optval = rack->r_rack_hw_rate_caps;
24666 		break;
24667 	case TCP_BBR_HDWR_PACE:
24668 		optval = rack->rack_hdw_pace_ena;
24669 		break;
24670 	case TCP_RACK_TLP_THRESH:
24671 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24672 		optval = rack->r_ctl.rc_tlp_threshold;
24673 		break;
24674 	case TCP_RACK_PKT_DELAY:
24675 		/* RACK added ms i.e. rack-rtt + reord + N */
24676 		optval = rack->r_ctl.rc_pkt_delay;
24677 		break;
24678 	case TCP_RACK_TLP_USE:
24679 		optval = rack->rack_tlp_threshold_use;
24680 		break;
24681 	case TCP_PACING_DND:
24682 		optval = rack->rc_pace_dnd;
24683 		break;
24684 	case TCP_RACK_PACE_RATE_CA:
24685 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24686 		break;
24687 	case TCP_RACK_PACE_RATE_SS:
24688 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24689 		break;
24690 	case TCP_RACK_PACE_RATE_REC:
24691 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24692 		break;
24693 	case TCP_DGP_UPPER_BOUNDS:
24694 		optval = rack->r_ctl.rack_per_upper_bound_ss;
24695 		optval <<= 16;
24696 		optval |= rack->r_ctl.rack_per_upper_bound_ca;
24697 		break;
24698 	case TCP_RACK_GP_INCREASE_SS:
24699 		optval = rack->r_ctl.rack_per_of_gp_ca;
24700 		break;
24701 	case TCP_RACK_GP_INCREASE_CA:
24702 		optval = rack->r_ctl.rack_per_of_gp_ss;
24703 		break;
24704 	case TCP_RACK_PACING_DIVISOR:
24705 		optval = rack->r_ctl.pace_len_divisor;
24706 		break;
24707 	case TCP_BBR_RACK_RTT_USE:
24708 		optval = rack->r_ctl.rc_rate_sample_method;
24709 		break;
24710 	case TCP_DELACK:
24711 		optval = tp->t_delayed_ack;
24712 		break;
24713 	case TCP_DATA_AFTER_CLOSE:
24714 		optval = rack->rc_allow_data_af_clo;
24715 		break;
24716 	case TCP_SHARED_CWND_TIME_LIMIT:
24717 		optval = rack->r_limit_scw;
24718 		break;
24719 	case TCP_HONOR_HPTS_MIN:
24720 		if (rack->r_use_hpts_min)
24721 			optval = rack->r_ctl.max_reduction;
24722 		else
24723 			optval = 0;
24724 		break;
24725 	case TCP_REC_IS_DYN:
24726 		optval = rack->rc_gp_no_rec_chg;
24727 		break;
24728 	case TCP_NO_TIMELY:
24729 		optval = rack->rc_skip_timely;
24730 		break;
24731 	case TCP_RACK_TIMER_SLOP:
24732 		optval = rack->r_ctl.timer_slop;
24733 		break;
24734 	default:
24735 		return (tcp_default_ctloutput(tp, sopt));
24736 		break;
24737 	}
24738 	INP_WUNLOCK(inp);
24739 	if (error == 0) {
24740 		if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24741 		    (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
24742 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
24743 		else
24744 			error = sooptcopyout(sopt, &optval, sizeof optval);
24745 	}
24746 	return (error);
24747 }
24748 
24749 static int
24750 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24751 {
24752 	if (sopt->sopt_dir == SOPT_SET) {
24753 		return (rack_set_sockopt(tp, sopt));
24754 	} else if (sopt->sopt_dir == SOPT_GET) {
24755 		return (rack_get_sockopt(tp, sopt));
24756 	} else {
24757 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24758 	}
24759 }
24760 
24761 static const char *rack_stack_names[] = {
24762 	__XSTRING(STACKNAME),
24763 #ifdef STACKALIAS
24764 	__XSTRING(STACKALIAS),
24765 #endif
24766 };
24767 
24768 static int
24769 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24770 {
24771 	memset(mem, 0, size);
24772 	return (0);
24773 }
24774 
24775 static void
24776 rack_dtor(void *mem, int32_t size, void *arg)
24777 {
24778 
24779 }
24780 
24781 static bool rack_mod_inited = false;
24782 
24783 static int
24784 tcp_addrack(module_t mod, int32_t type, void *data)
24785 {
24786 	int32_t err = 0;
24787 	int num_stacks;
24788 
24789 	switch (type) {
24790 	case MOD_LOAD:
24791 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24792 		    sizeof(struct rack_sendmap),
24793 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24794 
24795 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24796 		    sizeof(struct tcp_rack),
24797 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24798 
24799 		sysctl_ctx_init(&rack_sysctl_ctx);
24800 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24801 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24802 		    OID_AUTO,
24803 #ifdef STACKALIAS
24804 		    __XSTRING(STACKALIAS),
24805 #else
24806 		    __XSTRING(STACKNAME),
24807 #endif
24808 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24809 		    "");
24810 		if (rack_sysctl_root == NULL) {
24811 			printf("Failed to add sysctl node\n");
24812 			err = EFAULT;
24813 			goto free_uma;
24814 		}
24815 		rack_init_sysctls();
24816 		num_stacks = nitems(rack_stack_names);
24817 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24818 		    rack_stack_names, &num_stacks);
24819 		if (err) {
24820 			printf("Failed to register %s stack name for "
24821 			    "%s module\n", rack_stack_names[num_stacks],
24822 			    __XSTRING(MODNAME));
24823 			sysctl_ctx_free(&rack_sysctl_ctx);
24824 free_uma:
24825 			uma_zdestroy(rack_zone);
24826 			uma_zdestroy(rack_pcb_zone);
24827 			rack_counter_destroy();
24828 			printf("Failed to register rack module -- err:%d\n", err);
24829 			return (err);
24830 		}
24831 		tcp_lro_reg_mbufq();
24832 		rack_mod_inited = true;
24833 		break;
24834 	case MOD_QUIESCE:
24835 		err = deregister_tcp_functions(&__tcp_rack, true, false);
24836 		break;
24837 	case MOD_UNLOAD:
24838 		err = deregister_tcp_functions(&__tcp_rack, false, true);
24839 		if (err == EBUSY)
24840 			break;
24841 		if (rack_mod_inited) {
24842 			uma_zdestroy(rack_zone);
24843 			uma_zdestroy(rack_pcb_zone);
24844 			sysctl_ctx_free(&rack_sysctl_ctx);
24845 			rack_counter_destroy();
24846 			rack_mod_inited = false;
24847 		}
24848 		tcp_lro_dereg_mbufq();
24849 		err = 0;
24850 		break;
24851 	default:
24852 		return (EOPNOTSUPP);
24853 	}
24854 	return (err);
24855 }
24856 
24857 static moduledata_t tcp_rack = {
24858 	.name = __XSTRING(MODNAME),
24859 	.evhand = tcp_addrack,
24860 	.priv = 0
24861 };
24862 
24863 MODULE_VERSION(MODNAME, 1);
24864 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24865 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24866 
24867 #endif /* #if !defined(INET) && !defined(INET6) */
24868