2 * Copyright (c) 2016-2020 Netflix, Inc.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
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.
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
27 * Author: Randall Stewart <rrs@netflix.com>
28 * This work is based on the ACM Queue paper
29 * BBR - Congestion Based Congestion Control
30 * and also numerous discussions with Neal, Yuchung and Van.
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
37 #include "opt_inet6.h"
38 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
40 #include "opt_ratelimit.h"
41 #include <sys/param.h>
43 #include <sys/module.h>
44 #include <sys/kernel.h>
45 #include <sys/libkern.h>
47 #include <sys/hhook.h>
49 #include <sys/malloc.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/sysctl.h>
55 #include <sys/systm.h>
57 #include <sys/qmath.h>
59 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
61 #include <sys/refcount.h>
62 #include <sys/queue.h>
63 #include <sys/eventhandler.h>
65 #include <sys/kthread.h>
67 #include <sys/mutex.h>
68 #include <sys/tim_filter.h>
70 #include <sys/protosw.h>
72 #include <sys/kern_prefetch.h>
74 #include <net/route.h>
75 #include <net/route/nhop.h>
78 #define TCPSTATES /* for logging */
80 #include <netinet/in.h>
81 #include <netinet/in_kdtrace.h>
82 #include <netinet/in_pcb.h>
83 #include <netinet/ip.h>
84 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
85 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
86 #include <netinet/ip_var.h>
87 #include <netinet/ip6.h>
88 #include <netinet6/in6_pcb.h>
89 #include <netinet6/ip6_var.h>
91 #include <netinet/tcp.h>
92 #include <netinet/tcp_fsm.h>
93 #include <netinet/tcp_seq.h>
94 #include <netinet/tcp_timer.h>
95 #include <netinet/tcp_var.h>
96 #include <netinet/tcpip.h>
97 #include <netinet/tcp_hpts.h>
98 #include <netinet/cc/cc.h>
99 #include <netinet/tcp_log_buf.h>
100 #include <netinet/tcp_ratelimit.h>
101 #include <netinet/tcp_lro.h>
103 #include <netinet/tcp_debug.h>
104 #endif /* TCPDEBUG */
106 #include <netinet/tcp_offload.h>
109 #include <netinet6/tcp6_var.h>
111 #include <netinet/tcp_fastopen.h>
113 #include <netipsec/ipsec_support.h>
115 #include <net/if_var.h>
116 #include <net/ethernet.h>
118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
119 #include <netipsec/ipsec.h>
120 #include <netipsec/ipsec6.h>
123 #include <netinet/udp.h>
124 #include <netinet/udp_var.h>
125 #include <machine/in_cksum.h>
128 #include <security/mac/mac_framework.h>
131 #include "sack_filter.h"
133 #include "rack_bbr_common.h"
135 uma_zone_t bbr_pcb_zone;
137 struct sysctl_ctx_list bbr_sysctl_ctx;
138 struct sysctl_oid *bbr_sysctl_root;
140 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \
142 if ((u_long)(tv) < (u_long)(tvmin)) \
144 if ((u_long)(tv) > (u_long)(tvmax)) \
148 /*#define BBR_INVARIANT 1*/
153 static uint32_t bbr_def_init_win = 10;
154 static int32_t bbr_persist_min = 250000; /* 250ms */
155 static int32_t bbr_persist_max = 1000000; /* 1 Second */
156 static int32_t bbr_cwnd_may_shrink = 0;
157 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP;
158 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT;
159 static int32_t bbr_hardware_pacing_limit = 8000;
160 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */
161 static int32_t bbr_no_retran = 0;
163 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */
164 static int32_t bbr_max_net_error_cnt = 10;
165 /* Should the following be dynamic too -- loss wise */
166 static int32_t bbr_rtt_gain_thresh = 0;
167 /* Measurement controls */
168 static int32_t bbr_use_google_algo = 1;
169 static int32_t bbr_ts_limiting = 1;
170 static int32_t bbr_ts_can_raise = 0;
171 static int32_t bbr_do_red = 600;
172 static int32_t bbr_red_scale = 20000;
173 static int32_t bbr_red_mul = 1;
174 static int32_t bbr_red_div = 2;
175 static int32_t bbr_red_growth_restrict = 1;
176 static int32_t bbr_target_is_bbunit = 0;
177 static int32_t bbr_drop_limit = 0;
179 * How much gain do we need to see to
182 static int32_t bbr_marks_rxt_sack_passed = 0;
183 static int32_t bbr_start_exit = 25;
184 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */
185 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */
186 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */
187 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this
188 * if we go back ever to where the pacer
189 * has priority over timers.
191 static int32_t bbr_policer_call_from_rack_to = 0;
192 static int32_t bbr_policer_detection_enabled = 1;
193 static int32_t bbr_min_measurements_req = 1; /* We need at least 2
194 * measurments before we are
195 * "good" note that 2 == 1.
196 * This is because we use a >
197 * comparison. This means if
198 * min_measure was 0, it takes
199 * num-measures > min(0) and
200 * you get 1 measurement and
201 * you are good. Set to 1, you
203 * measurements (this is done
204 * to prevent it from being ok
205 * to have no measurements). */
206 static int32_t bbr_no_pacing_until = 4;
208 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */
209 static int32_t bbr_min_peer_delta = 20; /* 20 units */
210 static int32_t bbr_delta_percent = 150; /* 15.0 % */
212 static int32_t bbr_target_cwnd_mult_limit = 8;
214 * bbr_cwnd_min_val is the number of
215 * segments we hold to in the RTT probe
218 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS;
220 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS;
222 static int32_t bbr_gain_to_target = 1;
223 static int32_t bbr_gain_gets_extra_too = 1;
225 * bbr_high_gain is the 2/ln(2) value we need
226 * to double the sending rate in startup. This
227 * is used for both cwnd and hptsi gain's.
229 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1;
230 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1;
231 static int32_t bbr_use_lower_gain_in_startup = 1;
233 /* thresholds for reduction on drain in sub-states/drain */
234 static int32_t bbr_drain_rtt = BBR_SRTT;
235 static int32_t bbr_drain_floor = 88;
236 static int32_t google_allow_early_out = 1;
237 static int32_t google_consider_lost = 1;
238 static int32_t bbr_drain_drop_mul = 4;
239 static int32_t bbr_drain_drop_div = 5;
240 static int32_t bbr_rand_ot = 50;
241 static int32_t bbr_can_force_probertt = 0;
242 static int32_t bbr_can_adjust_probertt = 1;
243 static int32_t bbr_probertt_sets_rtt = 0;
244 static int32_t bbr_can_use_ts_for_rtt = 1;
245 static int32_t bbr_is_ratio = 0;
246 static int32_t bbr_sub_drain_app_limit = 1;
247 static int32_t bbr_prtt_slam_cwnd = 1;
248 static int32_t bbr_sub_drain_slam_cwnd = 1;
249 static int32_t bbr_slam_cwnd_in_main_drain = 1;
250 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter
252 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4);
254 * bbr_drain_gain is the reverse of the high_gain
255 * designed to drain back out the standing queue
256 * that is formed in startup by causing a larger
257 * hptsi gain and thus drainging the packets
260 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885;
261 static int32_t bbr_rttprobe_gain = 192;
264 * The cwnd_gain is the default cwnd gain applied when
265 * calculating a target cwnd. Note that the cwnd is
266 * a secondary factor in the way BBR works (see the
267 * paper and think about it, it will take some time).
268 * Basically the hptsi_gain spreads the packets out
269 * so you never get more than BDP to the peer even
270 * if the cwnd is high. In our implemenation that
271 * means in non-recovery/retransmission scenarios
272 * cwnd will never be reached by the flight-size.
274 static int32_t bbr_cwnd_gain = BBR_UNIT * 2;
275 static int32_t bbr_tlp_type_to_use = BBR_SRTT;
276 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */
277 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */
278 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */
279 static int32_t bbr_ignore_data_after_close = 1;
280 static int16_t bbr_hptsi_gain[] = {
290 int32_t bbr_use_rack_resend_cheat = 1;
291 int32_t bbr_sends_full_iwnd = 1;
293 #define BBR_HPTSI_GAIN_MAX 8
295 * The BBR module incorporates a number of
296 * TCP ideas that have been put out into the IETF
297 * over the last few years:
298 * - Yuchung Cheng's RACK TCP (for which its named) that
299 * will stop us using the number of dup acks and instead
300 * use time as the gage of when we retransmit.
301 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
302 * of Dukkipati et.al.
303 * - Van Jacobson's et.al BBR.
305 * RACK depends on SACK, so if an endpoint arrives that
306 * cannot do SACK the state machine below will shuttle the
307 * connection back to using the "default" TCP stack that is
310 * To implement BBR and RACK the original TCP stack was first decomposed
311 * into a functional state machine with individual states
312 * for each of the possible TCP connection states. The do_segement
313 * functions role in life is to mandate the connection supports SACK
314 * initially and then assure that the RACK state matches the conenction
315 * state before calling the states do_segment function. Data processing
316 * of inbound segments also now happens in the hpts_do_segment in general
317 * with only one exception. This is so we can keep the connection on
320 * Each state is simplified due to the fact that the original do_segment
321 * has been decomposed and we *know* what state we are in (no
322 * switches on the state) and all tests for SACK are gone. This
323 * greatly simplifies what each state does.
325 * TCP output is also over-written with a new version since it
326 * must maintain the new rack scoreboard and has had hptsi
327 * integrated as a requirment. Still todo is to eliminate the
328 * use of the callout_() system and use the hpts for all
331 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */
332 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */
333 static const int32_t bbr_min_req_free = 2; /* The min we must have on the
335 static int32_t bbr_tlp_thresh = 1;
336 static int32_t bbr_reorder_thresh = 2;
337 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def
338 * 60,000,000 - 60 seconds */
339 static int32_t bbr_pkt_delay = 1000;
340 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */
341 static int32_t bbr_incr_timers = 1;
343 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */
344 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */
345 static int32_t bbr_exit_startup_at_loss = 1;
348 * bbr_lt_bw_ratio is 1/8th
349 * bbr_lt_bw_diff is < 4 Kbit/sec
351 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */
352 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */
353 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use
355 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure
357 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */
358 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */
359 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */
361 static int32_t bbr_verbose_logging = 0;
363 * Currently regular tcp has a rto_min of 30ms
364 * the backoff goes 12 times so that ends up
365 * being a total of 122.850 seconds before a
366 * connection is killed.
368 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */
369 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */
371 /****************************************************/
372 /* DEFAULT TSO SIZING (cpu performance impacting) */
373 /****************************************************/
374 /* What amount is our formula using to get TSO size */
375 static int32_t bbr_hptsi_per_second = 1000;
378 * For hptsi under bbr_cross_over connections what is delay
379 * target 7ms (in usec) combined with a seg_max of 2
380 * gets us close to identical google behavior in
381 * TSO size selection (possibly more 1MSS sends).
383 static int32_t bbr_hptsi_segments_delay_tar = 7000;
385 /* Does pacing delay include overhead's in its time calculations? */
386 static int32_t bbr_include_enet_oh = 0;
387 static int32_t bbr_include_ip_oh = 1;
388 static int32_t bbr_include_tcp_oh = 1;
389 static int32_t bbr_google_discount = 10;
391 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */
392 static int32_t bbr_state_is_pkt_epoch = 0;
393 static int32_t bbr_state_drain_2_tar = 1;
394 /* What is the max the 0 - bbr_cross_over MBPS TSO target
395 * can reach using our delay target. Note that this
396 * value becomes the floor for the cross over
399 static int32_t bbr_hptsi_segments_max = 2;
400 static int32_t bbr_hptsi_segments_floor = 1;
401 static int32_t bbr_hptsi_utter_max = 0;
403 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */
404 static int32_t bbr_hptsi_bytes_min = 1460;
405 static int32_t bbr_all_get_min = 0;
407 /* Cross over point from algo-a to algo-b */
408 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS;
410 /* Do we deal with our restart state? */
411 static int32_t bbr_uses_idle_restart = 0;
412 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */
414 /* Do we allow hardware pacing? */
415 static int32_t bbr_allow_hdwr_pacing = 0;
416 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */
417 static int32_t bbr_hdwr_pace_floor = 1;
418 static int32_t bbr_hdwr_pacing_delay_cnt = 10;
420 /****************************************************/
421 static int32_t bbr_resends_use_tso = 0;
422 static int32_t bbr_tlp_max_resend = 2;
423 static int32_t bbr_sack_block_limit = 128;
425 #define BBR_MAX_STAT 19
426 counter_u64_t bbr_state_time[BBR_MAX_STAT];
427 counter_u64_t bbr_state_lost[BBR_MAX_STAT];
428 counter_u64_t bbr_state_resend[BBR_MAX_STAT];
429 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE];
430 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE];
431 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE];
432 counter_u64_t bbr_flows_whdwr_pacing;
433 counter_u64_t bbr_flows_nohdwr_pacing;
435 counter_u64_t bbr_nohdwr_pacing_enobuf;
436 counter_u64_t bbr_hdwr_pacing_enobuf;
438 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr);
441 * Static defintions we need for forward declarations.
444 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain,
445 uint32_t useconds_time, uint64_t bw);
447 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain);
449 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win);
451 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses);
453 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line,
456 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain);
458 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch,
459 int32_t pkt_epoch, uint32_t losses);
461 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm);
462 static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp);
464 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
465 struct bbr_sendmap *rsm, uint32_t srtt,
468 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts,
471 bbr_set_state_target(struct tcp_bbr *bbr, int line);
473 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line);
476 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line);
479 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts);
482 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts);
485 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt,
486 uint32_t line, uint8_t is_start, uint16_t set);
488 static struct bbr_sendmap *
489 bbr_find_lowest_rsm(struct tcp_bbr *bbr);
490 static __inline uint32_t
491 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type);
493 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which);
496 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
497 uint32_t thresh, uint32_t to);
499 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag);
502 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot,
503 uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay);
506 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr,
507 uint32_t cts, int32_t line);
509 bbr_stop_all_timers(struct tcpcb *tp);
511 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts);
513 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts);
515 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts);
518 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
519 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod);
522 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp,
525 static inline uint8_t
526 bbr_state_val(struct tcp_bbr *bbr)
528 return(bbr->rc_bbr_substate);
531 static inline uint32_t
532 get_min_cwnd(struct tcp_bbr *bbr)
536 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
537 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
538 return (bbr_cwnd_min_val_hs * mss);
540 return (bbr_cwnd_min_val * mss);
544 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
549 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
550 if (tp->t_srtt == 0) {
551 srtt = (uint64_t)BBR_INITIAL_RTO;
554 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
555 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
557 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
558 bbr_persist_min, bbr_persist_max);
559 return ((uint32_t)ret_val);
563 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
566 * Start the FR timer, we do this based on getting the first one in
567 * the rc_tmap. Note that if its NULL we must stop the timer. in all
568 * events we need to stop the running timer (if its running) before
569 * starting the new one.
571 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
573 int32_t is_tlp_timer = 0;
574 struct bbr_sendmap *rsm;
576 if (bbr->rc_all_timers_stopped) {
577 /* All timers have been stopped none are to run */
580 if (bbr->rc_in_persist) {
581 /* We can't start any timer in persists */
582 return (bbr_get_persists_timer_val(tp, bbr));
584 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
586 ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
587 (tp->t_state < TCPS_ESTABLISHED)) {
588 /* Nothing on the send map */
590 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
594 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
596 idx = rsm->r_rtr_cnt - 1;
597 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
598 tstmp_touse = rsm->r_tim_lastsent[idx];
600 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
601 if (TSTMP_GT(tstmp_touse, cts))
602 time_since_sent = cts - tstmp_touse;
604 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
606 tov = BBR_INITIAL_RTO;
608 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
609 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
611 tov *= tcp_backoff[tp->t_rxtshift];
612 if (tov > time_since_sent)
613 tov -= time_since_sent;
615 tov = bbr->r_ctl.rc_min_to;
616 TCPT_RANGESET_NOSLOP(to, tov,
617 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
618 (bbr->rc_max_rto_sec * USECS_IN_SECOND));
619 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
624 if (rsm->r_flags & BBR_ACKED) {
625 rsm = bbr_find_lowest_rsm(bbr);
631 /* Convert from ms to usecs */
632 if (rsm->r_flags & BBR_SACK_PASSED) {
633 if ((tp->t_flags & TF_SENTFIN) &&
634 ((tp->snd_max - tp->snd_una) == 1) &&
635 (rsm->r_flags & BBR_HAS_FIN)) {
637 * We don't start a bbr rack timer if all we have is
642 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
643 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
644 idx = rsm->r_rtr_cnt - 1;
645 exp = rsm->r_tim_lastsent[idx] + thresh;
646 if (SEQ_GEQ(exp, cts)) {
648 if (to < bbr->r_ctl.rc_min_to) {
649 to = bbr->r_ctl.rc_min_to;
652 to = bbr->r_ctl.rc_min_to;
655 /* Ok we need to do a TLP not RACK */
656 if (bbr->rc_tlp_in_progress != 0) {
658 * The previous send was a TLP.
662 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
664 /* We found no rsm to TLP with. */
667 if (rsm->r_flags & BBR_HAS_FIN) {
668 /* If its a FIN we don't do TLP */
673 idx = rsm->r_rtr_cnt - 1;
674 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
675 tstmp_touse = rsm->r_tim_lastsent[idx];
677 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
678 if (TSTMP_GT(tstmp_touse, cts))
679 time_since_sent = cts - tstmp_touse;
681 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
682 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
683 if (thresh > time_since_sent)
684 to = thresh - time_since_sent;
686 to = bbr->r_ctl.rc_min_to;
687 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
689 * If the TLP time works out to larger than the max
690 * RTO lets not do TLP.. just RTO.
694 if ((bbr->rc_tlp_rtx_out == 1) &&
695 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
697 * Second retransmit of the same TLP
700 bbr->rc_tlp_rtx_out = 0;
703 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
705 * The tail is no longer the last one I did a probe
708 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
709 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
712 if (is_tlp_timer == 0) {
713 BBR_STAT_INC(bbr_to_arm_rack);
714 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
716 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
717 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
719 * We have exceeded how many times we can retran the
720 * current TLP timer, switch to the RTO timer.
724 BBR_STAT_INC(bbr_to_arm_tlp);
725 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
731 static inline int32_t
732 bbr_minseg(struct tcp_bbr *bbr)
734 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
738 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
741 struct hpts_diag diag;
742 uint32_t delayed_ack = 0;
744 uint32_t hpts_timeout;
746 int32_t delay_calc = 0;
747 uint32_t prev_delay = 0;
750 if (inp->inp_in_hpts) {
751 /* A previous call is already set up */
754 if ((tp->t_state == TCPS_CLOSED) ||
755 (tp->t_state == TCPS_LISTEN)) {
758 stopped = bbr->rc_tmr_stopped;
759 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
760 left = bbr->r_ctl.rc_timer_exp - cts;
762 bbr->r_ctl.rc_hpts_flags = 0;
763 bbr->r_ctl.rc_timer_exp = 0;
764 prev_delay = bbr->r_ctl.rc_last_delay_val;
765 if (bbr->r_ctl.rc_last_delay_val &&
768 * If a previous pacer delay was in place we
769 * are not coming from the output side (where
770 * we calculate a delay, more likely a timer).
772 slot = bbr->r_ctl.rc_last_delay_val;
773 if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
774 /* Compensate for time passed */
775 delay_calc = cts - bbr->rc_pacer_started;
776 if (delay_calc <= slot)
780 /* Do we have early to make up for by pushing out the pacing time? */
781 if (bbr->r_agg_early_set) {
782 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
783 slot += bbr->r_ctl.rc_agg_early;
784 bbr->r_ctl.rc_agg_early = 0;
785 bbr->r_agg_early_set = 0;
787 /* Are we running a total debt that needs to be compensated for? */
788 if (bbr->r_ctl.rc_hptsi_agg_delay) {
789 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
790 /* We nuke the delay */
791 slot -= bbr->r_ctl.rc_hptsi_agg_delay;
792 bbr->r_ctl.rc_hptsi_agg_delay = 0;
794 /* We nuke some of the delay, put in a minimal 100usecs */
795 bbr->r_ctl.rc_hptsi_agg_delay -= slot;
796 bbr->r_ctl.rc_last_delay_val = slot = 100;
799 bbr->r_ctl.rc_last_delay_val = slot;
800 hpts_timeout = bbr_timer_start(tp, bbr, cts);
801 if (tp->t_flags & TF_DELACK) {
802 if (bbr->rc_in_persist == 0) {
803 delayed_ack = bbr_delack_time;
806 * We are in persists and have
807 * gotten a new data element.
809 if (hpts_timeout > bbr_delack_time) {
811 * Lets make the persists timer (which acks)
812 * be the smaller of hpts_timeout and bbr_delack_time.
814 hpts_timeout = bbr_delack_time;
819 ((hpts_timeout == 0) ||
820 (delayed_ack < hpts_timeout))) {
821 /* We need a Delayed ack timer */
822 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
823 hpts_timeout = delayed_ack;
826 /* Mark that we have a pacing timer up */
827 BBR_STAT_INC(bbr_paced_segments);
828 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
831 * If no timers are going to run and we will fall off thfe hptsi
832 * wheel, we resort to a keep-alive timer if its configured.
834 if ((hpts_timeout == 0) &&
836 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
837 (tp->t_state <= TCPS_CLOSING)) {
839 * Ok we have no timer (persists, rack, tlp, rxt or
840 * del-ack), we don't have segments being paced. So
841 * all that is left is the keepalive timer.
843 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
844 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
846 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
848 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
851 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
852 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
854 * RACK, TLP, persists and RXT timers all are restartable
855 * based on actions input .. i.e we received a packet (ack
856 * or sack) and that changes things (rw, or snd_una etc).
857 * Thus we can restart them with a new value. For
858 * keep-alive, delayed_ack we keep track of what was left
859 * and restart the timer with a smaller value.
861 if (left < hpts_timeout)
864 if (bbr->r_ctl.rc_incr_tmrs && slot &&
865 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
867 * If configured to do so, and the timer is either
868 * the TLP or RXT timer, we need to increase the timeout
869 * by the pacing time. Consider the bottleneck at my
870 * machine as an example, we are sending something
871 * to start a TLP on. The last packet won't be emitted
872 * fully until the pacing time (the bottleneck will hold
873 * the data in place). Once the packet is emitted that
874 * is when we want to start waiting for the TLP. This
875 * is most evident with hardware pacing (where the nic
876 * is holding the packet(s) before emitting). But it
877 * can also show up in the network so we do it for all
878 * cases. Technically we would take off one packet from
879 * this extra delay but this is easier and being more
880 * conservative is probably better.
882 hpts_timeout += slot;
886 * Hack alert for now we can't time-out over 2147 seconds (a
887 * bit more than 35min)
889 if (hpts_timeout > 0x7ffffffe)
890 hpts_timeout = 0x7ffffffe;
891 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
893 bbr->r_ctl.rc_timer_exp = 0;
895 (bbr->rc_use_google ||
896 bbr->output_error_seen ||
897 (slot <= hpts_timeout)) ) {
899 * Tell LRO that it can queue packets while
902 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
903 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
904 (bbr->rc_cwnd_limited == 0)) {
906 * If we are not cwnd limited and we
907 * are running a rack timer we put on
908 * the do not disturbe even for sack.
910 inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
912 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
913 bbr->rc_pacer_started = cts;
915 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
917 bbr->rc_timer_first = 0;
918 bbr->bbr_timer_src = frm;
919 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
920 bbr_log_hpts_diag(bbr, cts, &diag);
921 } else if (hpts_timeout) {
922 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
925 * We add the flag here as well if the slot is set,
926 * since hpts will call in to clear the queue first before
927 * calling the output routine (which does our timers).
928 * We don't want to set the flag if its just a timer
929 * else the arrival of data might (that causes us
930 * to send more) might get delayed. Imagine being
931 * on a keep-alive timer and a request comes in for
935 bbr->rc_pacer_started = cts;
936 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
937 (bbr->rc_cwnd_limited == 0)) {
939 * For a rack timer, don't wake us even
940 * if a sack arrives as long as we are
943 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
944 inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
946 /* All other timers wake us up */
947 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
948 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
950 bbr->bbr_timer_src = frm;
951 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
952 bbr_log_hpts_diag(bbr, cts, &diag);
953 bbr->rc_timer_first = 1;
955 bbr->rc_tmr_stopped = 0;
956 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
960 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
963 * We received an ack, and then did not call send or were bounced
964 * out due to the hpts was running. Now a timer is up as well, is it
968 struct bbr_sendmap *rsm;
969 uint32_t hpts_timeout;
972 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
973 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
975 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
976 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
977 (tmr_up == PACE_TMR_RXT)) {
978 /* Should be an RXT */
983 /* Nothing outstanding? */
984 if (tp->t_flags & TF_DELACK) {
985 if (tmr_up == PACE_TMR_DELACK)
987 * We are supposed to have delayed ack up
991 } else if (sbavail(&inp->inp_socket->so_snd) &&
992 (tmr_up == PACE_TMR_RXT)) {
994 * if we hit enobufs then we would expect the
995 * possiblity of nothing outstanding and the RXT up
996 * (and the hptsi timer).
999 } else if (((V_tcp_always_keepalive ||
1000 inp->inp_socket->so_options & SO_KEEPALIVE) &&
1001 (tp->t_state <= TCPS_CLOSING)) &&
1002 (tmr_up == PACE_TMR_KEEP) &&
1003 (tp->snd_max == tp->snd_una)) {
1004 /* We should have keep alive up and we do */
1008 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1009 if ((tp->t_flags & TF_SENTFIN) &&
1010 ((tp->snd_max - tp->snd_una) == 1) &&
1011 (rsm->r_flags & BBR_HAS_FIN)) {
1012 /* needs to be a RXT */
1013 if (tmr_up == PACE_TMR_RXT)
1017 } else if (tmr_up == PACE_TMR_RACK)
1021 } else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1022 /* Rack timer has priority if we have data out */
1024 } else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1025 ((tmr_up == PACE_TMR_TLP) ||
1026 (tmr_up == PACE_TMR_RXT))) {
1028 * Either a TLP or RXT is fine if no sack-passed is in place
1029 * and data is outstanding.
1032 } else if (tmr_up == PACE_TMR_DELACK) {
1034 * If the delayed ack was going to go off before the
1035 * rtx/tlp/rack timer were going to expire, then that would
1036 * be the timer in control. Note we don't check the time
1037 * here trusting the code is correct.
1041 if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1042 ((tmr_up == PACE_TMR_RXT) ||
1043 (tmr_up == PACE_TMR_TLP) ||
1044 (tmr_up == PACE_TMR_RACK))) {
1046 * We have outstanding data and
1047 * we *do* have a RACK, TLP or RXT
1048 * timer running. We won't restart
1049 * anything here since thats probably ok we
1050 * will get called with some timer here shortly.
1055 * Ok the timer originally started is not what we want now. We will
1056 * force the hpts to be stopped if any, and restart with the slot
1057 * set to what was in the saved slot.
1060 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1061 if (inp->inp_in_hpts)
1062 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
1063 bbr_timer_cancel(bbr, __LINE__, cts);
1064 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1068 * Output is hptsi so we just need to switch the type of
1069 * timer. We don't bother with keep-alive, since when we
1070 * jump through the output, it will start the keep-alive if
1073 * We only need a delayed-ack added and or the hpts_timeout.
1075 hpts_timeout = bbr_timer_start(tp, bbr, cts);
1076 if (tp->t_flags & TF_DELACK) {
1077 if (hpts_timeout == 0) {
1078 hpts_timeout = bbr_delack_time;
1079 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1081 else if (hpts_timeout > bbr_delack_time) {
1082 hpts_timeout = bbr_delack_time;
1083 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1087 if (hpts_timeout > 0x7ffffffe)
1088 hpts_timeout = 0x7ffffffe;
1089 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1094 int32_t bbr_clear_lost = 0;
1097 * Considers the two time values now (cts) and earlier.
1098 * If cts is smaller than earlier, we could have
1099 * had a sequence wrap (our counter wraps every
1100 * 70 min or so) or it could be just clock skew
1101 * getting us two differnt time values. Clock skew
1102 * will show up within 10ms or so. So in such
1103 * a case (where cts is behind earlier time by
1104 * less than 10ms) we return 0. Otherwise we
1105 * return the true difference between them.
1107 static inline uint32_t
1108 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1110 * Given two timestamps, the current time stamp cts, and some other
1111 * time-stamp taken in theory earlier return the difference. The
1112 * trick is here sometimes locking will get the other timestamp
1113 * after the cts. If this occurs we need to return 0.
1115 if (TSTMP_GEQ(cts, earlier_time))
1116 return (cts - earlier_time);
1118 * cts is behind earlier_time if its less than 10ms consider it 0.
1119 * If its more than 10ms difference then we had a time wrap. Else
1120 * its just the normal locking foo. I wonder if we should not go to
1121 * 64bit TS and get rid of this issue.
1123 if (TSTMP_GEQ((cts + 10000), earlier_time))
1126 * Ok the time must have wrapped. So we need to answer a large
1127 * amount of time, which the normal subtraction should do.
1129 return (cts - earlier_time);
1133 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1138 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1139 if (error || req->newptr == NULL)
1142 error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1146 #ifdef BBR_INVARIANTS
1147 printf("Clearing BBR lost counters\n");
1149 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1150 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1151 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1152 } else if (stat == 2) {
1153 #ifdef BBR_INVARIANTS
1154 printf("Clearing BBR option counters\n");
1156 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1157 } else if (stat == 3) {
1158 #ifdef BBR_INVARIANTS
1159 printf("Clearing BBR stats counters\n");
1161 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1162 } else if (stat == 4) {
1163 #ifdef BBR_INVARIANTS
1164 printf("Clearing BBR out-size counters\n");
1166 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1173 bbr_init_sysctls(void)
1175 struct sysctl_oid *bbr_probertt;
1176 struct sysctl_oid *bbr_hptsi;
1177 struct sysctl_oid *bbr_measure;
1178 struct sysctl_oid *bbr_cwnd;
1179 struct sysctl_oid *bbr_timeout;
1180 struct sysctl_oid *bbr_states;
1181 struct sysctl_oid *bbr_startup;
1182 struct sysctl_oid *bbr_policer;
1184 /* Probe rtt controls */
1185 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1186 SYSCTL_CHILDREN(bbr_sysctl_root),
1189 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1191 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1192 SYSCTL_CHILDREN(bbr_probertt),
1193 OID_AUTO, "gain", CTLFLAG_RW,
1194 &bbr_rttprobe_gain, 192,
1195 "What is the filter gain drop in probe_rtt (0=disable)?");
1196 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1197 SYSCTL_CHILDREN(bbr_probertt),
1198 OID_AUTO, "cwnd", CTLFLAG_RW,
1199 &bbr_rtt_probe_cwndtarg, 4,
1200 "How many mss's are outstanding during probe-rtt");
1201 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1202 SYSCTL_CHILDREN(bbr_probertt),
1203 OID_AUTO, "int", CTLFLAG_RW,
1204 &bbr_rtt_probe_limit, 4000000,
1205 "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1206 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1207 SYSCTL_CHILDREN(bbr_probertt),
1208 OID_AUTO, "mintime", CTLFLAG_RW,
1209 &bbr_rtt_probe_time, 200000,
1210 "How many microseconds in probe-rtt");
1211 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1212 SYSCTL_CHILDREN(bbr_probertt),
1213 OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1214 &bbr_filter_len_sec, 6,
1215 "How long in seconds does the rttProp filter run?");
1216 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1217 SYSCTL_CHILDREN(bbr_probertt),
1218 OID_AUTO, "drain_rtt", CTLFLAG_RW,
1219 &bbr_drain_rtt, BBR_SRTT,
1220 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1221 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1222 SYSCTL_CHILDREN(bbr_probertt),
1223 OID_AUTO, "can_force", CTLFLAG_RW,
1224 &bbr_can_force_probertt, 0,
1225 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1226 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1227 SYSCTL_CHILDREN(bbr_probertt),
1228 OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1229 &bbr_probertt_sets_rtt, 0,
1230 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1231 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1232 SYSCTL_CHILDREN(bbr_probertt),
1233 OID_AUTO, "can_adjust", CTLFLAG_RW,
1234 &bbr_can_adjust_probertt, 1,
1235 "Can we dynamically adjust the probe-rtt limits and times?");
1236 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1237 SYSCTL_CHILDREN(bbr_probertt),
1238 OID_AUTO, "is_ratio", CTLFLAG_RW,
1240 "is the limit to filter a ratio?");
1241 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1242 SYSCTL_CHILDREN(bbr_probertt),
1243 OID_AUTO, "use_cwnd", CTLFLAG_RW,
1244 &bbr_prtt_slam_cwnd, 0,
1245 "Should we set/recover cwnd?");
1246 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1247 SYSCTL_CHILDREN(bbr_probertt),
1248 OID_AUTO, "can_use_ts", CTLFLAG_RW,
1249 &bbr_can_use_ts_for_rtt, 1,
1250 "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1252 /* Pacing controls */
1253 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1254 SYSCTL_CHILDREN(bbr_sysctl_root),
1257 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1259 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1260 SYSCTL_CHILDREN(bbr_hptsi),
1261 OID_AUTO, "hw_pacing", CTLFLAG_RW,
1262 &bbr_allow_hdwr_pacing, 1,
1263 "Do we allow hardware pacing?");
1264 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1265 SYSCTL_CHILDREN(bbr_hptsi),
1266 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1267 &bbr_hardware_pacing_limit, 4000,
1268 "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1269 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1270 SYSCTL_CHILDREN(bbr_hptsi),
1271 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1272 &bbr_hdwr_pace_adjust, 2,
1273 "Multiplier to calculated tso size?");
1274 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1275 SYSCTL_CHILDREN(bbr_hptsi),
1276 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1277 &bbr_hdwr_pace_floor, 1,
1278 "Do we invoke the hardware pacing floor?");
1279 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1280 SYSCTL_CHILDREN(bbr_hptsi),
1281 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1282 &bbr_hdwr_pacing_delay_cnt, 10,
1283 "How many packets must be sent after hdwr pacing is enabled");
1284 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1285 SYSCTL_CHILDREN(bbr_hptsi),
1286 OID_AUTO, "bw_cross", CTLFLAG_RW,
1287 &bbr_cross_over, 3000000,
1288 "What is the point where we cross over to linux like TSO size set");
1289 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1290 SYSCTL_CHILDREN(bbr_hptsi),
1291 OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1292 &bbr_hptsi_segments_delay_tar, 7000,
1293 "What is the worse case delay target for hptsi < 48Mbp connections");
1294 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1295 SYSCTL_CHILDREN(bbr_hptsi),
1296 OID_AUTO, "enet_oh", CTLFLAG_RW,
1297 &bbr_include_enet_oh, 0,
1298 "Do we include the ethernet overhead in calculating pacing delay?");
1299 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1300 SYSCTL_CHILDREN(bbr_hptsi),
1301 OID_AUTO, "ip_oh", CTLFLAG_RW,
1302 &bbr_include_ip_oh, 1,
1303 "Do we include the IP overhead in calculating pacing delay?");
1304 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1305 SYSCTL_CHILDREN(bbr_hptsi),
1306 OID_AUTO, "tcp_oh", CTLFLAG_RW,
1307 &bbr_include_tcp_oh, 0,
1308 "Do we include the TCP overhead in calculating pacing delay?");
1309 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1310 SYSCTL_CHILDREN(bbr_hptsi),
1311 OID_AUTO, "google_discount", CTLFLAG_RW,
1312 &bbr_google_discount, 10,
1313 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1314 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1315 SYSCTL_CHILDREN(bbr_hptsi),
1316 OID_AUTO, "all_get_min", CTLFLAG_RW,
1317 &bbr_all_get_min, 0,
1318 "If you are less than a MSS do you just get the min?");
1319 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1320 SYSCTL_CHILDREN(bbr_hptsi),
1321 OID_AUTO, "tso_min", CTLFLAG_RW,
1322 &bbr_hptsi_bytes_min, 1460,
1323 "For 0 -> 24Mbps what is floor number of segments for TSO");
1324 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1325 SYSCTL_CHILDREN(bbr_hptsi),
1326 OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1327 &bbr_hptsi_segments_max, 6,
1328 "For 0 -> 24Mbps what is top number of segments for TSO");
1329 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1330 SYSCTL_CHILDREN(bbr_hptsi),
1331 OID_AUTO, "seg_floor", CTLFLAG_RW,
1332 &bbr_hptsi_segments_floor, 1,
1333 "Minimum TSO size we will fall too in segments");
1334 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1335 SYSCTL_CHILDREN(bbr_hptsi),
1336 OID_AUTO, "utter_max", CTLFLAG_RW,
1337 &bbr_hptsi_utter_max, 0,
1338 "The absolute maximum that any pacing (outside of hardware) can be");
1339 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1340 SYSCTL_CHILDREN(bbr_hptsi),
1341 OID_AUTO, "seg_divisor", CTLFLAG_RW,
1342 &bbr_hptsi_per_second, 100,
1343 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1344 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1345 SYSCTL_CHILDREN(bbr_hptsi),
1346 OID_AUTO, "srtt_mul", CTLFLAG_RW,
1347 &bbr_hptsi_max_mul, 1,
1348 "The multiplier for pace len max");
1349 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1350 SYSCTL_CHILDREN(bbr_hptsi),
1351 OID_AUTO, "srtt_div", CTLFLAG_RW,
1352 &bbr_hptsi_max_div, 2,
1353 "The divisor for pace len max");
1354 /* Measurement controls */
1355 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1356 SYSCTL_CHILDREN(bbr_sysctl_root),
1359 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1360 "Measurement controls");
1361 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1362 SYSCTL_CHILDREN(bbr_measure),
1363 OID_AUTO, "min_i_bw", CTLFLAG_RW,
1364 &bbr_initial_bw_bps, 62500,
1365 "Minimum initial b/w in bytes per second");
1366 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1367 SYSCTL_CHILDREN(bbr_measure),
1368 OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1369 &bbr_sack_not_required, 0,
1370 "Do we allow bbr to run on connections not supporting SACK?");
1371 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1372 SYSCTL_CHILDREN(bbr_measure),
1373 OID_AUTO, "use_google", CTLFLAG_RW,
1374 &bbr_use_google_algo, 0,
1375 "Use has close to google V1.0 has possible?");
1376 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1377 SYSCTL_CHILDREN(bbr_measure),
1378 OID_AUTO, "ts_limiting", CTLFLAG_RW,
1379 &bbr_ts_limiting, 1,
1380 "Do we attempt to use the peers timestamp to limit b/w caculations?");
1381 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1382 SYSCTL_CHILDREN(bbr_measure),
1383 OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1384 &bbr_ts_can_raise, 0,
1385 "Can we raise the b/w via timestamp b/w calculation?");
1386 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1387 SYSCTL_CHILDREN(bbr_measure),
1388 OID_AUTO, "ts_delta", CTLFLAG_RW,
1389 &bbr_min_usec_delta, 20000,
1390 "How long in usec between ts of our sends in ts validation code?");
1391 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1392 SYSCTL_CHILDREN(bbr_measure),
1393 OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1394 &bbr_min_peer_delta, 20,
1395 "What min numerical value should be between the peer deltas?");
1396 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1397 SYSCTL_CHILDREN(bbr_measure),
1398 OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1399 &bbr_delta_percent, 150,
1400 "What percentage (150 = 15.0) do we allow variance for?");
1401 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1402 SYSCTL_CHILDREN(bbr_measure),
1403 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1404 &bbr_min_measurements_req, 1,
1405 "What is the minimum measurment count we need before we switch to our b/w estimate");
1406 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1407 SYSCTL_CHILDREN(bbr_measure),
1408 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1409 &bbr_no_pacing_until, 4,
1410 "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1411 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1412 SYSCTL_CHILDREN(bbr_measure),
1413 OID_AUTO, "quanta", CTLFLAG_RW,
1415 "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1416 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1417 SYSCTL_CHILDREN(bbr_measure),
1418 OID_AUTO, "noretran", CTLFLAG_RW,
1420 "Should google mode not use retransmission measurements for the b/w estimation?");
1421 /* State controls */
1422 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1423 SYSCTL_CHILDREN(bbr_sysctl_root),
1426 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1428 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1429 SYSCTL_CHILDREN(bbr_states),
1430 OID_AUTO, "idle_restart", CTLFLAG_RW,
1431 &bbr_uses_idle_restart, 0,
1432 "Do we use a new special idle_restart state to ramp back up quickly?");
1433 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1434 SYSCTL_CHILDREN(bbr_states),
1435 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1436 &bbr_idle_restart_threshold, 100000,
1437 "How long must we be idle before we restart??");
1438 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1439 SYSCTL_CHILDREN(bbr_states),
1440 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1441 &bbr_state_is_pkt_epoch, 0,
1442 "Do we use a pkt-epoch for substate if 0 rttProp?");
1443 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1444 SYSCTL_CHILDREN(bbr_states),
1445 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1446 &bbr_rtt_gain_thresh, 0,
1447 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1448 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1449 SYSCTL_CHILDREN(bbr_states),
1450 OID_AUTO, "drain_floor", CTLFLAG_RW,
1451 &bbr_drain_floor, 88,
1452 "What is the lowest we can drain (pg) too?");
1453 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1454 SYSCTL_CHILDREN(bbr_states),
1455 OID_AUTO, "drain_2_target", CTLFLAG_RW,
1456 &bbr_state_drain_2_tar, 1,
1457 "Do we drain to target in drain substate?");
1458 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1459 SYSCTL_CHILDREN(bbr_states),
1460 OID_AUTO, "gain_2_target", CTLFLAG_RW,
1461 &bbr_gain_to_target, 1,
1462 "Does probe bw gain to target??");
1463 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1464 SYSCTL_CHILDREN(bbr_states),
1465 OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1466 &bbr_gain_gets_extra_too, 1,
1467 "Does probe bw gain get the extra time too?");
1468 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1469 SYSCTL_CHILDREN(bbr_states),
1470 OID_AUTO, "ld_div", CTLFLAG_RW,
1471 &bbr_drain_drop_div, 5,
1472 "Long drain drop divider?");
1473 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1474 SYSCTL_CHILDREN(bbr_states),
1475 OID_AUTO, "ld_mul", CTLFLAG_RW,
1476 &bbr_drain_drop_mul, 4,
1477 "Long drain drop multiplier?");
1478 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1479 SYSCTL_CHILDREN(bbr_states),
1480 OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1482 "Random discount of the ot?");
1483 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1484 SYSCTL_CHILDREN(bbr_states),
1485 OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1486 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1487 "How many packet-epochs does the b/w delivery rate last?");
1488 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1489 SYSCTL_CHILDREN(bbr_states),
1490 OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1491 &bbr_sub_drain_app_limit, 0,
1492 "Does our sub-state drain invoke app limited if its long?");
1493 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1494 SYSCTL_CHILDREN(bbr_states),
1495 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1496 &bbr_sub_drain_slam_cwnd, 0,
1497 "Should we set/recover cwnd for sub-state drain?");
1498 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1499 SYSCTL_CHILDREN(bbr_states),
1500 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1501 &bbr_slam_cwnd_in_main_drain, 0,
1502 "Should we set/recover cwnd for main-state drain?");
1503 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1504 SYSCTL_CHILDREN(bbr_states),
1505 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1506 &google_allow_early_out, 1,
1507 "Should we allow google probe-bw/drain to exit early at flight target?");
1508 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1509 SYSCTL_CHILDREN(bbr_states),
1510 OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1511 &google_consider_lost, 1,
1512 "Should we have losses exit gain of probebw in google mode??");
1513 /* Startup controls */
1514 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1515 SYSCTL_CHILDREN(bbr_sysctl_root),
1518 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1519 "Startup controls");
1520 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1521 SYSCTL_CHILDREN(bbr_startup),
1522 OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1523 &bbr_sends_full_iwnd, 1,
1524 "Do we not pace but burst out initial windows has our TSO size?");
1525 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1526 SYSCTL_CHILDREN(bbr_startup),
1527 OID_AUTO, "loss_threshold", CTLFLAG_RW,
1528 &bbr_startup_loss_thresh, 2000,
1529 "In startup what is the loss threshold in a pe that will exit us from startup?");
1530 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1531 SYSCTL_CHILDREN(bbr_startup),
1532 OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1533 &bbr_use_lower_gain_in_startup, 1,
1534 "Should we use a lower hptsi gain if we see loss in startup?");
1535 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1536 SYSCTL_CHILDREN(bbr_startup),
1537 OID_AUTO, "gain", CTLFLAG_RW,
1538 &bbr_start_exit, 25,
1539 "What gain percent do we need to see to stay in startup??");
1540 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1541 SYSCTL_CHILDREN(bbr_startup),
1542 OID_AUTO, "low_gain", CTLFLAG_RW,
1543 &bbr_low_start_exit, 15,
1544 "What gain percent do we need to see to stay in the lower gain startup??");
1545 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1546 SYSCTL_CHILDREN(bbr_startup),
1547 OID_AUTO, "loss_exit", CTLFLAG_RW,
1548 &bbr_exit_startup_at_loss, 1,
1549 "Should we exit startup at loss in an epoch if we are not gaining?");
1551 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1552 SYSCTL_CHILDREN(bbr_sysctl_root),
1555 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1557 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1558 SYSCTL_CHILDREN(bbr_cwnd),
1559 OID_AUTO, "tar_rtt", CTLFLAG_RW,
1560 &bbr_cwndtarget_rtt_touse, 0,
1561 "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1562 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1563 SYSCTL_CHILDREN(bbr_cwnd),
1564 OID_AUTO, "may_shrink", CTLFLAG_RW,
1565 &bbr_cwnd_may_shrink, 0,
1566 "Can the cwnd shrink if it would grow to more than the target?");
1567 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1568 SYSCTL_CHILDREN(bbr_cwnd),
1569 OID_AUTO, "max_target_limit", CTLFLAG_RW,
1570 &bbr_target_cwnd_mult_limit, 8,
1571 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1572 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1573 SYSCTL_CHILDREN(bbr_cwnd),
1574 OID_AUTO, "highspeed_min", CTLFLAG_RW,
1575 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1576 "What is the high-speed min cwnd (rttProp under 1ms)");
1577 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1578 SYSCTL_CHILDREN(bbr_cwnd),
1579 OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1580 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1581 "What is the min cwnd (rttProp > 1ms)");
1582 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1583 SYSCTL_CHILDREN(bbr_cwnd),
1584 OID_AUTO, "initwin", CTLFLAG_RW,
1585 &bbr_def_init_win, 10,
1586 "What is the BBR initial window, if 0 use tcp version");
1587 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1588 SYSCTL_CHILDREN(bbr_cwnd),
1589 OID_AUTO, "do_loss_red", CTLFLAG_RW,
1591 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1592 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1593 SYSCTL_CHILDREN(bbr_cwnd),
1594 OID_AUTO, "red_scale", CTLFLAG_RW,
1595 &bbr_red_scale, 20000,
1596 "What RTT do we scale with?");
1597 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1598 SYSCTL_CHILDREN(bbr_cwnd),
1599 OID_AUTO, "red_growslow", CTLFLAG_RW,
1600 &bbr_red_growth_restrict, 1,
1601 "Do we restrict cwnd growth for whats in flight?");
1602 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1603 SYSCTL_CHILDREN(bbr_cwnd),
1604 OID_AUTO, "red_div", CTLFLAG_RW,
1606 "If we reduce whats the divisor?");
1607 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1608 SYSCTL_CHILDREN(bbr_cwnd),
1609 OID_AUTO, "red_mul", CTLFLAG_RW,
1611 "If we reduce whats the mulitiplier?");
1612 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1613 SYSCTL_CHILDREN(bbr_cwnd),
1614 OID_AUTO, "target_is_unit", CTLFLAG_RW,
1615 &bbr_target_is_bbunit, 0,
1616 "Is the state target the pacing_gain or BBR_UNIT?");
1617 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1618 SYSCTL_CHILDREN(bbr_cwnd),
1619 OID_AUTO, "drop_limit", CTLFLAG_RW,
1621 "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1623 /* Timeout controls */
1624 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1625 SYSCTL_CHILDREN(bbr_sysctl_root),
1628 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1629 "Time out controls");
1630 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1631 SYSCTL_CHILDREN(bbr_timeout),
1632 OID_AUTO, "delack", CTLFLAG_RW,
1633 &bbr_delack_time, 100000,
1634 "BBR's delayed ack time");
1635 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1636 SYSCTL_CHILDREN(bbr_timeout),
1637 OID_AUTO, "tlp_uses", CTLFLAG_RW,
1638 &bbr_tlp_type_to_use, 3,
1639 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1640 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1641 SYSCTL_CHILDREN(bbr_timeout),
1642 OID_AUTO, "persmin", CTLFLAG_RW,
1643 &bbr_persist_min, 250000,
1644 "What is the minimum time in microseconds between persists");
1645 SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1646 SYSCTL_CHILDREN(bbr_timeout),
1647 OID_AUTO, "persmax", CTLFLAG_RW,
1648 &bbr_persist_max, 1000000,
1649 "What is the largest delay in microseconds between persists");
1650 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1651 SYSCTL_CHILDREN(bbr_timeout),
1652 OID_AUTO, "tlp_minto", CTLFLAG_RW,
1653 &bbr_tlp_min, 10000,
1654 "TLP Min timeout in usecs");
1655 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1656 SYSCTL_CHILDREN(bbr_timeout),
1657 OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1658 &bbr_delayed_ack_time, 200000,
1659 "TLP delayed ack compensation value");
1660 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1661 SYSCTL_CHILDREN(bbr_sysctl_root),
1662 OID_AUTO, "minrto", CTLFLAG_RW,
1663 &bbr_rto_min_ms, 30,
1664 "Minimum RTO in ms");
1665 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1666 SYSCTL_CHILDREN(bbr_timeout),
1667 OID_AUTO, "maxrto", CTLFLAG_RW,
1668 &bbr_rto_max_sec, 4,
1669 "Maxiumum RTO in seconds -- should be at least as large as min_rto");
1670 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1671 SYSCTL_CHILDREN(bbr_timeout),
1672 OID_AUTO, "tlp_retry", CTLFLAG_RW,
1673 &bbr_tlp_max_resend, 2,
1674 "How many times does TLP retry a single segment or multiple with no ACK");
1675 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1676 SYSCTL_CHILDREN(bbr_timeout),
1677 OID_AUTO, "minto", CTLFLAG_RW,
1679 "Minimum rack timeout in useconds");
1680 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1681 SYSCTL_CHILDREN(bbr_timeout),
1682 OID_AUTO, "pktdelay", CTLFLAG_RW,
1683 &bbr_pkt_delay, 1000,
1684 "Extra RACK time (in useconds) besides reordering thresh");
1685 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1686 SYSCTL_CHILDREN(bbr_timeout),
1687 OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1688 &bbr_incr_timers, 1,
1689 "Increase the RXT/TLP timer by the pacing time used?");
1690 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1691 SYSCTL_CHILDREN(bbr_timeout),
1692 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1693 &bbr_marks_rxt_sack_passed, 0,
1694 "Mark sack passed on all those not ack'd when a RXT hits?");
1695 /* Policer controls */
1696 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1697 SYSCTL_CHILDREN(bbr_sysctl_root),
1700 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1701 "Policer controls");
1702 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1703 SYSCTL_CHILDREN(bbr_policer),
1704 OID_AUTO, "detect_enable", CTLFLAG_RW,
1705 &bbr_policer_detection_enabled, 1,
1706 "Is policer detection enabled??");
1707 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1708 SYSCTL_CHILDREN(bbr_policer),
1709 OID_AUTO, "min_pes", CTLFLAG_RW,
1710 &bbr_lt_intvl_min_rtts, 4,
1711 "Minimum number of PE's?");
1712 SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1713 SYSCTL_CHILDREN(bbr_policer),
1714 OID_AUTO, "bwdiff", CTLFLAG_RW,
1715 &bbr_lt_bw_diff, (4000/8),
1716 "Minimal bw diff?");
1717 SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1718 SYSCTL_CHILDREN(bbr_policer),
1719 OID_AUTO, "bwratio", CTLFLAG_RW,
1720 &bbr_lt_bw_ratio, 8,
1721 "Minimal bw diff?");
1722 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1723 SYSCTL_CHILDREN(bbr_policer),
1724 OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1725 &bbr_policer_call_from_rack_to, 0,
1726 "Do we call the policer detection code from a rack-timeout?");
1727 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1728 SYSCTL_CHILDREN(bbr_policer),
1729 OID_AUTO, "false_postive", CTLFLAG_RW,
1730 &bbr_lt_intvl_fp, 0,
1731 "What packet epoch do we do false-postive detection at (0=no)?");
1732 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1733 SYSCTL_CHILDREN(bbr_policer),
1734 OID_AUTO, "loss_thresh", CTLFLAG_RW,
1735 &bbr_lt_loss_thresh, 196,
1736 "Loss threshold 196 = 19.6%?");
1737 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1738 SYSCTL_CHILDREN(bbr_policer),
1739 OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1740 &bbr_lt_fd_thresh, 100,
1741 "What percentage is the false detection threshold (150=15.0)?");
1743 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1744 SYSCTL_CHILDREN(bbr_sysctl_root),
1745 OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1746 &bbr_use_rack_resend_cheat, 0,
1747 "Do we burst 1ms between sends on retransmissions (like rack)?");
1748 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1749 SYSCTL_CHILDREN(bbr_sysctl_root),
1750 OID_AUTO, "error_paceout", CTLFLAG_RW,
1751 &bbr_error_base_paceout, 10000,
1752 "When we hit an error what is the min to pace out in usec's?");
1753 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1754 SYSCTL_CHILDREN(bbr_sysctl_root),
1755 OID_AUTO, "kill_paceout", CTLFLAG_RW,
1756 &bbr_max_net_error_cnt, 10,
1757 "When we hit this many errors in a row, kill the session?");
1758 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1759 SYSCTL_CHILDREN(bbr_sysctl_root),
1760 OID_AUTO, "data_after_close", CTLFLAG_RW,
1761 &bbr_ignore_data_after_close, 1,
1762 "Do we hold off sending a RST until all pending data is ack'd");
1763 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1764 SYSCTL_CHILDREN(bbr_sysctl_root),
1765 OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1766 &bbr_resends_use_tso, 0,
1767 "Can resends use TSO?");
1768 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1769 SYSCTL_CHILDREN(bbr_sysctl_root),
1770 OID_AUTO, "sblklimit", CTLFLAG_RW,
1771 &bbr_sack_block_limit, 128,
1772 "When do we start ignoring small sack blocks");
1773 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1774 SYSCTL_CHILDREN(bbr_sysctl_root),
1775 OID_AUTO, "bb_verbose", CTLFLAG_RW,
1776 &bbr_verbose_logging, 0,
1777 "Should BBR black box logging be verbose");
1778 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1779 SYSCTL_CHILDREN(bbr_sysctl_root),
1780 OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1781 &bbr_reorder_thresh, 2,
1782 "What factor for rack will be added when seeing reordering (shift right)");
1783 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1784 SYSCTL_CHILDREN(bbr_sysctl_root),
1785 OID_AUTO, "reorder_fade", CTLFLAG_RW,
1786 &bbr_reorder_fade, 0,
1787 "Does reorder detection fade, if so how many ms (0 means never)");
1788 SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1789 SYSCTL_CHILDREN(bbr_sysctl_root),
1790 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1792 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1793 /* Stats and counters */
1794 /* The pacing counters for hdwr/software can't be in the array */
1795 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1796 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1797 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1798 SYSCTL_CHILDREN(bbr_sysctl_root),
1799 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1800 &bbr_hdwr_pacing_enobuf,
1801 "Total number of enobufs for hardware paced flows");
1802 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1803 SYSCTL_CHILDREN(bbr_sysctl_root),
1804 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1805 &bbr_nohdwr_pacing_enobuf,
1806 "Total number of enobufs for non-hardware paced flows");
1808 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1809 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1810 SYSCTL_CHILDREN(bbr_sysctl_root),
1811 OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1812 &bbr_flows_whdwr_pacing,
1813 "Total number of hardware paced flows");
1814 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1815 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1816 SYSCTL_CHILDREN(bbr_sysctl_root),
1817 OID_AUTO, "software_pacing", CTLFLAG_RD,
1818 &bbr_flows_nohdwr_pacing,
1819 "Total number of software paced flows");
1820 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1821 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1822 OID_AUTO, "stats", CTLFLAG_RD,
1823 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1824 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1825 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1826 OID_AUTO, "opts", CTLFLAG_RD,
1827 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1828 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1829 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1830 OID_AUTO, "lost", CTLFLAG_RD,
1831 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1832 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1833 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1834 OID_AUTO, "stateresend", CTLFLAG_RD,
1835 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1836 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1837 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1838 OID_AUTO, "statetime", CTLFLAG_RD,
1839 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1840 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1841 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1842 OID_AUTO, "outsize", CTLFLAG_RD,
1843 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1844 SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1845 SYSCTL_CHILDREN(bbr_sysctl_root),
1846 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1847 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1851 bbr_counter_destroy(void)
1853 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1854 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1855 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1856 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1857 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1858 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1859 counter_u64_free(bbr_nohdwr_pacing_enobuf);
1860 counter_u64_free(bbr_hdwr_pacing_enobuf);
1861 counter_u64_free(bbr_flows_whdwr_pacing);
1862 counter_u64_free(bbr_flows_nohdwr_pacing);
1866 static __inline void
1867 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1869 memset(l, 0, sizeof(union tcp_log_stackspecific));
1870 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1871 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1872 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1873 l->bw_inuse = bbr_get_bw(bbr);
1874 l->inflight = ctf_flight_size(bbr->rc_tp,
1875 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1876 l->applimited = bbr->r_ctl.r_app_limited_until;
1877 l->delivered = bbr->r_ctl.rc_delivered;
1879 l->lost = bbr->r_ctl.rc_lost;
1880 l->bbr_state = bbr->rc_bbr_state;
1881 l->bbr_substate = bbr_state_val(bbr);
1882 l->epoch = bbr->r_ctl.rc_rtt_epoch;
1883 l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1884 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1885 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1886 l->inhpts = bbr->rc_inp->inp_in_hpts;
1887 l->ininput = bbr->rc_inp->inp_in_input;
1888 l->use_lt_bw = bbr->rc_lt_use_bw;
1889 l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1890 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1894 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1896 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1897 union tcp_log_stackspecific log;
1899 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1900 log.u_bbr.flex1 = 0;
1901 log.u_bbr.flex2 = 0;
1902 log.u_bbr.flex5 = 0;
1903 log.u_bbr.flex3 = 0;
1904 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1905 log.u_bbr.flex7 = reason;
1906 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1907 log.u_bbr.flex8 = 0;
1908 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1909 &bbr->rc_inp->inp_socket->so_rcv,
1910 &bbr->rc_inp->inp_socket->so_snd,
1911 BBR_LOG_BW_RED_EV, 0,
1912 0, &log, false, &bbr->rc_tv);
1917 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1919 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1920 union tcp_log_stackspecific log;
1922 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1923 log.u_bbr.flex1 = seq;
1924 log.u_bbr.flex2 = count;
1925 log.u_bbr.flex8 = mode;
1926 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1927 &bbr->rc_inp->inp_socket->so_rcv,
1928 &bbr->rc_inp->inp_socket->so_snd,
1930 0, &log, false, &bbr->rc_tv);
1935 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1936 uint8_t reason, uint32_t p_maxseg, int len)
1938 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1939 union tcp_log_stackspecific log;
1941 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1942 log.u_bbr.flex1 = p_maxseg;
1943 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1944 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1945 log.u_bbr.flex4 = reason;
1946 log.u_bbr.flex5 = bbr->rc_in_persist;
1947 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1948 log.u_bbr.flex7 = p_maxseg;
1949 log.u_bbr.flex8 = bbr->rc_in_persist;
1950 log.u_bbr.pkts_out = 0;
1951 log.u_bbr.applimited = len;
1952 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1953 &bbr->rc_inp->inp_socket->so_rcv,
1954 &bbr->rc_inp->inp_socket->so_snd,
1956 tlen, &log, false, &bbr->rc_tv);
1961 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1963 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1964 union tcp_log_stackspecific log;
1966 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1967 log.u_bbr.flex1 = seq;
1968 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1969 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1970 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1971 &bbr->rc_inp->inp_socket->so_rcv,
1972 &bbr->rc_inp->inp_socket->so_snd,
1974 0, &log, false, &bbr->rc_tv);
1979 bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts)
1981 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1982 union tcp_log_stackspecific log;
1984 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1985 log.u_bbr.flex1 = tso;
1986 log.u_bbr.flex2 = maxseg;
1987 log.u_bbr.flex3 = mtu;
1988 log.u_bbr.flex4 = csum_flags;
1989 TCP_LOG_EVENTP(tp, NULL,
1990 &bbr->rc_inp->inp_socket->so_rcv,
1991 &bbr->rc_inp->inp_socket->so_snd,
1993 0, &log, false, &bbr->rc_tv);
1998 bbr_log_flowend(struct tcp_bbr *bbr)
2000 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2001 union tcp_log_stackspecific log;
2002 struct sockbuf *r, *s;
2005 if (bbr->rc_inp->inp_socket) {
2006 r = &bbr->rc_inp->inp_socket->so_rcv;
2007 s = &bbr->rc_inp->inp_socket->so_snd;
2011 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2012 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2015 0, &log, false, &tv);
2020 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2021 uint32_t lost, uint32_t del)
2023 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2024 union tcp_log_stackspecific log;
2026 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2027 log.u_bbr.flex1 = lost;
2028 log.u_bbr.flex2 = del;
2029 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2030 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2031 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2032 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2033 log.u_bbr.flex7 = line;
2034 log.u_bbr.flex8 = 0;
2035 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2036 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2037 &bbr->rc_inp->inp_socket->so_rcv,
2038 &bbr->rc_inp->inp_socket->so_snd,
2039 BBR_LOG_PKT_EPOCH, 0,
2040 0, &log, false, &bbr->rc_tv);
2045 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2047 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2048 union tcp_log_stackspecific log;
2050 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2051 log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2052 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2053 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2054 log.u_bbr.flex7 = line;
2055 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2056 &bbr->rc_inp->inp_socket->so_rcv,
2057 &bbr->rc_inp->inp_socket->so_snd,
2058 BBR_LOG_TIME_EPOCH, 0,
2059 0, &log, false, &bbr->rc_tv);
2064 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2066 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2067 union tcp_log_stackspecific log;
2069 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2070 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2071 log.u_bbr.flex2 = new_tar;
2072 log.u_bbr.flex3 = line;
2073 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2074 log.u_bbr.flex5 = bbr_quanta;
2075 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2076 log.u_bbr.flex7 = bbr->rc_last_options;
2077 log.u_bbr.flex8 = meth;
2078 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2079 &bbr->rc_inp->inp_socket->so_rcv,
2080 &bbr->rc_inp->inp_socket->so_snd,
2081 BBR_LOG_STATE_TARGET, 0,
2082 0, &log, false, &bbr->rc_tv);
2088 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2090 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2091 union tcp_log_stackspecific log;
2093 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2094 log.u_bbr.flex1 = line;
2095 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2096 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2097 if (bbr_state_is_pkt_epoch)
2098 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2100 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2101 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2102 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2103 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2104 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2105 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2106 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2107 &bbr->rc_inp->inp_socket->so_rcv,
2108 &bbr->rc_inp->inp_socket->so_snd,
2110 0, &log, false, &bbr->rc_tv);
2115 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2116 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2118 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2119 union tcp_log_stackspecific log;
2121 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2122 log.u_bbr.flex1 = line;
2123 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2124 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2125 log.u_bbr.flex4 = applied;
2126 log.u_bbr.flex5 = rtt;
2127 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2128 log.u_bbr.flex7 = cond;
2129 log.u_bbr.flex8 = reas;
2130 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2131 &bbr->rc_inp->inp_socket->so_rcv,
2132 &bbr->rc_inp->inp_socket->so_snd,
2133 BBR_LOG_RTT_SHRINKS, 0,
2134 0, &log, false, &bbr->rc_tv);
2139 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2141 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2142 union tcp_log_stackspecific log;
2144 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2145 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2146 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2147 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2148 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2149 &bbr->rc_inp->inp_socket->so_rcv,
2150 &bbr->rc_inp->inp_socket->so_snd,
2152 0, &log, false, &bbr->rc_tv);
2157 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2158 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2160 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2161 union tcp_log_stackspecific log;
2163 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2164 log.u_bbr.flex1 = line;
2165 log.u_bbr.flex2 = prev_acked;
2166 log.u_bbr.flex3 = bytes_this_ack;
2167 log.u_bbr.flex4 = chg;
2168 log.u_bbr.flex5 = th_ack;
2169 log.u_bbr.flex6 = target;
2170 log.u_bbr.flex8 = meth;
2171 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2172 &bbr->rc_inp->inp_socket->so_rcv,
2173 &bbr->rc_inp->inp_socket->so_snd,
2175 0, &log, false, &bbr->rc_tv);
2180 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2183 * Log the rtt sample we are applying to the srtt algorithm in
2186 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2187 union tcp_log_stackspecific log;
2189 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2190 log.u_bbr.flex1 = rtt;
2191 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2192 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2193 log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2194 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2195 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2196 log.u_bbr.flex6 = tsin;
2197 log.u_bbr.flex7 = 0;
2198 log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2199 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2200 &bbr->rc_inp->inp_socket->so_rcv,
2201 &bbr->rc_inp->inp_socket->so_snd,
2203 0, &log, false, &bbr->rc_tv);
2208 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2210 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2211 union tcp_log_stackspecific log;
2213 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2214 log.u_bbr.flex1 = time_in;
2215 log.u_bbr.flex2 = line;
2216 log.u_bbr.flex8 = enter_exit;
2217 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2218 &bbr->rc_inp->inp_socket->so_rcv,
2219 &bbr->rc_inp->inp_socket->so_snd,
2221 0, &log, false, &bbr->rc_tv);
2225 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2227 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2228 union tcp_log_stackspecific log;
2230 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2231 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2232 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2233 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2234 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2235 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2236 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2237 &bbr->rc_inp->inp_socket->so_rcv,
2238 &bbr->rc_inp->inp_socket->so_snd,
2239 BBR_LOG_ACKCLEAR, 0,
2240 0, &log, false, &bbr->rc_tv);
2245 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2246 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2248 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2249 union tcp_log_stackspecific log;
2252 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2253 log.u_bbr.flex1 = nsegs;
2254 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2258 log.u_bbr.flex3 = m->m_flags;
2259 if (m->m_flags & M_TSTMP) {
2260 mbuf_tstmp2timespec(m, &ts);
2261 tv.tv_sec = ts.tv_sec;
2262 tv.tv_usec = ts.tv_nsec / 1000;
2263 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2265 log.u_bbr.lt_epoch = 0;
2267 if (m->m_flags & M_TSTMP_LRO) {
2268 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000;
2269 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000;
2270 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2272 /* No arrival timestamp */
2273 log.u_bbr.flex5 = 0;
2276 log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2278 log.u_bbr.flex3 = 0;
2279 log.u_bbr.flex5 = 0;
2280 log.u_bbr.flex6 = 0;
2281 log.u_bbr.pkts_out = 0;
2283 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2284 log.u_bbr.flex7 = bbr->r_wanted_output;
2285 log.u_bbr.flex8 = bbr->rc_in_persist;
2286 TCP_LOG_EVENTP(bbr->rc_tp, th,
2287 &bbr->rc_inp->inp_socket->so_rcv,
2288 &bbr->rc_inp->inp_socket->so_snd,
2290 tlen, &log, true, &bbr->rc_tv);
2295 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2297 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2298 union tcp_log_stackspecific log;
2300 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2301 log.u_bbr.flex1 = did_out;
2302 log.u_bbr.flex2 = nxt_pkt;
2303 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2304 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2305 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2306 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2307 log.u_bbr.flex7 = bbr->r_wanted_output;
2308 log.u_bbr.flex8 = bbr->rc_in_persist;
2309 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2310 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2311 &bbr->rc_inp->inp_socket->so_rcv,
2312 &bbr->rc_inp->inp_socket->so_snd,
2313 BBR_LOG_DOSEG_DONE, 0,
2314 0, &log, true, &bbr->rc_tv);
2319 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2320 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2322 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2323 union tcp_log_stackspecific log;
2325 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2326 log.u_bbr.flex1 = line;
2327 log.u_bbr.flex2 = o_len;
2328 log.u_bbr.flex3 = segcnt;
2329 log.u_bbr.flex4 = segsiz;
2330 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2331 &bbr->rc_inp->inp_socket->so_rcv,
2332 &bbr->rc_inp->inp_socket->so_snd,
2333 BBR_LOG_ENOBUF_JMP, ENOBUFS,
2334 len, &log, true, &bbr->rc_tv);
2339 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2341 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2342 union tcp_log_stackspecific log;
2344 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2345 log.u_bbr.flex1 = timers;
2346 log.u_bbr.flex2 = ret;
2347 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2348 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2349 log.u_bbr.flex5 = cts;
2350 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2351 log.u_bbr.flex8 = hpts_calling;
2352 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2353 &bbr->rc_inp->inp_socket->so_rcv,
2354 &bbr->rc_inp->inp_socket->so_snd,
2355 BBR_LOG_TO_PROCESS, 0,
2356 0, &log, false, &bbr->rc_tv);
2361 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2363 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2364 union tcp_log_stackspecific log;
2367 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2368 log.u_bbr.flex1 = bbr->bbr_timer_src;
2369 log.u_bbr.flex2 = 0;
2370 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2371 ar = (uint64_t)(bbr->r_ctl.rc_resend);
2373 ar &= 0x00000000ffffffff;
2374 log.u_bbr.flex4 = (uint32_t)ar;
2375 ar = (uint64_t)bbr->r_ctl.rc_resend;
2376 ar &= 0x00000000ffffffff;
2377 log.u_bbr.flex5 = (uint32_t)ar;
2378 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2379 log.u_bbr.flex8 = to_num;
2380 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2381 &bbr->rc_inp->inp_socket->so_rcv,
2382 &bbr->rc_inp->inp_socket->so_snd,
2384 0, &log, false, &bbr->rc_tv);
2389 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2391 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2392 union tcp_log_stackspecific log;
2394 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2395 log.u_bbr.flex1 = flex1;
2396 log.u_bbr.flex2 = flex2;
2397 log.u_bbr.flex3 = flex3;
2398 log.u_bbr.flex4 = 0;
2399 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2400 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2401 log.u_bbr.flex8 = reason;
2402 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2403 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2404 &bbr->rc_inp->inp_socket->so_rcv,
2405 &bbr->rc_inp->inp_socket->so_snd,
2407 0, &log, false, &bbr->rc_tv);
2412 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2414 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2415 union tcp_log_stackspecific log;
2417 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2418 log.u_bbr.flex1 = diag->p_nxt_slot;
2419 log.u_bbr.flex2 = diag->p_cur_slot;
2420 log.u_bbr.flex3 = diag->slot_req;
2421 log.u_bbr.flex4 = diag->inp_hptsslot;
2422 log.u_bbr.flex5 = diag->slot_remaining;
2423 log.u_bbr.flex6 = diag->need_new_to;
2424 log.u_bbr.flex7 = diag->p_hpts_active;
2425 log.u_bbr.flex8 = diag->p_on_min_sleep;
2426 /* Hijack other fields as needed */
2427 log.u_bbr.epoch = diag->have_slept;
2428 log.u_bbr.lt_epoch = diag->yet_to_sleep;
2429 log.u_bbr.pkts_out = diag->co_ret;
2430 log.u_bbr.applimited = diag->hpts_sleep_time;
2431 log.u_bbr.delivered = diag->p_prev_slot;
2432 log.u_bbr.inflight = diag->p_runningtick;
2433 log.u_bbr.bw_inuse = diag->wheel_tick;
2434 log.u_bbr.rttProp = diag->wheel_cts;
2435 log.u_bbr.delRate = diag->maxticks;
2436 log.u_bbr.cur_del_rate = diag->p_curtick;
2437 log.u_bbr.cur_del_rate <<= 32;
2438 log.u_bbr.cur_del_rate |= diag->p_lasttick;
2439 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2440 &bbr->rc_inp->inp_socket->so_rcv,
2441 &bbr->rc_inp->inp_socket->so_snd,
2442 BBR_LOG_HPTSDIAG, 0,
2443 0, &log, false, &bbr->rc_tv);
2448 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2449 uint32_t thresh, uint32_t to)
2451 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2452 union tcp_log_stackspecific log;
2454 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2455 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2456 log.u_bbr.flex2 = time_since_sent;
2457 log.u_bbr.flex3 = srtt;
2458 log.u_bbr.flex4 = thresh;
2459 log.u_bbr.flex5 = to;
2460 log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2461 log.u_bbr.flex8 = mode;
2462 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2463 &bbr->rc_inp->inp_socket->so_rcv,
2464 &bbr->rc_inp->inp_socket->so_snd,
2465 BBR_LOG_TIMERPREP, 0,
2466 0, &log, false, &bbr->rc_tv);
2471 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2472 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2474 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2475 union tcp_log_stackspecific log;
2477 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2478 log.u_bbr.flex1 = usecs;
2479 log.u_bbr.flex2 = len;
2480 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2481 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2483 log.u_bbr.flex5 = (1 << 2);
2485 log.u_bbr.flex5 = 0;
2486 log.u_bbr.flex6 = override;
2487 log.u_bbr.flex7 = gain;
2488 log.u_bbr.flex8 = mod;
2489 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2490 &bbr->rc_inp->inp_socket->so_rcv,
2491 &bbr->rc_inp->inp_socket->so_snd,
2492 BBR_LOG_HPTSI_CALC, 0,
2493 len, &log, false, &bbr->rc_tv);
2498 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2500 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2501 union tcp_log_stackspecific log;
2503 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2505 log.u_bbr.flex1 = bbr->bbr_timer_src;
2506 log.u_bbr.flex2 = to;
2507 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2508 log.u_bbr.flex4 = slot;
2509 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2510 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2511 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2512 log.u_bbr.flex8 = which;
2513 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2514 &bbr->rc_inp->inp_socket->so_rcv,
2515 &bbr->rc_inp->inp_socket->so_snd,
2516 BBR_LOG_TIMERSTAR, 0,
2517 0, &log, false, &bbr->rc_tv);
2522 bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm)
2524 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2525 union tcp_log_stackspecific log;
2527 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2528 log.u_bbr.flex1 = thresh;
2529 log.u_bbr.flex2 = lro;
2530 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2531 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2532 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2533 log.u_bbr.flex6 = srtt;
2534 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2535 log.u_bbr.flex8 = frm;
2536 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2537 &bbr->rc_inp->inp_socket->so_rcv,
2538 &bbr->rc_inp->inp_socket->so_snd,
2539 BBR_LOG_THRESH_CALC, 0,
2540 0, &log, false, &bbr->rc_tv);
2545 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2547 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2548 union tcp_log_stackspecific log;
2550 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2551 log.u_bbr.flex1 = line;
2552 log.u_bbr.flex2 = bbr->bbr_timer_src;
2553 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2554 log.u_bbr.flex4 = bbr->rc_in_persist;
2555 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2556 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2557 log.u_bbr.flex8 = hpts_removed;
2558 log.u_bbr.pkts_out = bbr->rc_pacer_started;
2559 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2560 &bbr->rc_inp->inp_socket->so_rcv,
2561 &bbr->rc_inp->inp_socket->so_snd,
2562 BBR_LOG_TIMERCANC, 0,
2563 0, &log, false, &bbr->rc_tv);
2568 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2570 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2571 union tcp_log_stackspecific log;
2573 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2574 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2575 log.u_bbr.flex2 = (peer_delta >> 32);
2576 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2577 log.u_bbr.flex4 = (delta >> 32);
2578 log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2579 log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2580 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2581 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2582 &bbr->rc_inp->inp_socket->so_rcv,
2583 &bbr->rc_inp->inp_socket->so_snd,
2584 BBR_LOG_TSTMP_VAL, 0,
2585 0, &log, false, &bbr->rc_tv);
2590 bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr)
2592 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2593 union tcp_log_stackspecific log;
2595 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2596 log.u_bbr.flex1 = tsosz;
2597 log.u_bbr.flex2 = tls;
2598 log.u_bbr.flex3 = tcp_min_hptsi_time;
2599 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2600 log.u_bbr.flex5 = old_val;
2601 log.u_bbr.flex6 = maxseg;
2602 log.u_bbr.flex7 = bbr->rc_no_pacing;
2603 log.u_bbr.flex7 <<= 1;
2604 log.u_bbr.flex7 |= bbr->rc_past_init_win;
2606 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2608 log.u_bbr.flex8 = bbr->rc_use_google;
2609 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2610 &bbr->rc_inp->inp_socket->so_rcv,
2611 &bbr->rc_inp->inp_socket->so_snd,
2613 0, &log, false, &bbr->rc_tv);
2618 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2619 uint32_t flags, uint32_t line)
2621 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2622 union tcp_log_stackspecific log;
2624 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2625 log.u_bbr.flex1 = line;
2626 log.u_bbr.flex2 = rsm->r_start;
2627 log.u_bbr.flex3 = rsm->r_end;
2628 log.u_bbr.flex4 = rsm->r_delivered;
2629 log.u_bbr.flex5 = rsm->r_rtr_cnt;
2630 log.u_bbr.flex6 = rsm->r_dupack;
2631 log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2632 log.u_bbr.flex8 = rsm->r_flags;
2633 /* Hijack the pkts_out fids */
2634 log.u_bbr.applimited = flags;
2635 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2636 &bbr->rc_inp->inp_socket->so_rcv,
2637 &bbr->rc_inp->inp_socket->so_snd,
2639 0, &log, false, &bbr->rc_tv);
2644 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2645 uint32_t flex3, uint32_t flex2, uint32_t flex5,
2646 uint32_t flex6, uint32_t pkts_out, int flex7,
2647 uint32_t flex4, uint32_t flex1)
2650 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2651 union tcp_log_stackspecific log;
2653 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2654 log.u_bbr.flex1 = flex1;
2655 log.u_bbr.flex2 = flex2;
2656 log.u_bbr.flex3 = flex3;
2657 log.u_bbr.flex4 = flex4;
2658 log.u_bbr.flex5 = flex5;
2659 log.u_bbr.flex6 = flex6;
2660 log.u_bbr.flex7 = flex7;
2661 /* Hijack the pkts_out fids */
2662 log.u_bbr.pkts_out = pkts_out;
2663 log.u_bbr.flex8 = flex8;
2664 if (bbr->rc_ack_was_delayed)
2665 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2667 log.u_bbr.epoch = 0;
2668 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2669 &bbr->rc_inp->inp_socket->so_rcv,
2670 &bbr->rc_inp->inp_socket->so_snd,
2672 flex2, &log, false, &bbr->rc_tv);
2677 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2678 uint32_t newbw, uint32_t obw, uint32_t diff,
2681 if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2682 union tcp_log_stackspecific log;
2684 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2685 log.u_bbr.flex1 = reason;
2686 log.u_bbr.flex2 = newbw;
2687 log.u_bbr.flex3 = obw;
2688 log.u_bbr.flex4 = diff;
2689 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2690 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2691 log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2692 log.u_bbr.pkts_out = tim;
2693 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2694 if (bbr->rc_lt_use_bw == 0)
2695 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2697 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2698 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2699 &bbr->rc_inp->inp_socket->so_rcv,
2700 &bbr->rc_inp->inp_socket->so_snd,
2702 0, &log, false, &bbr->rc_tv);
2707 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2709 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2710 union tcp_log_stackspecific log;
2712 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2713 log.u_bbr.flex1 = line;
2714 log.u_bbr.flex2 = tick;
2715 log.u_bbr.flex3 = tp->t_maxunacktime;
2716 log.u_bbr.flex4 = tp->t_acktime;
2717 log.u_bbr.flex8 = event;
2718 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2719 &bbr->rc_inp->inp_socket->so_rcv,
2720 &bbr->rc_inp->inp_socket->so_snd,
2721 BBR_LOG_PROGRESS, 0,
2722 0, &log, false, &bbr->rc_tv);
2727 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2728 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2731 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2732 union tcp_log_stackspecific log;
2734 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2735 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2736 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2737 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff);
2738 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2739 log.u_bbr.bw_inuse = rate;
2740 log.u_bbr.flex5 = line;
2741 log.u_bbr.flex6 = error;
2742 log.u_bbr.flex8 = bbr->skip_gain;
2743 log.u_bbr.flex8 <<= 1;
2744 log.u_bbr.flex8 |= bbr->gain_is_limited;
2745 log.u_bbr.flex8 <<= 1;
2746 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2747 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2748 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2749 &bbr->rc_inp->inp_socket->so_rcv,
2750 &bbr->rc_inp->inp_socket->so_snd,
2751 BBR_LOG_HDWR_PACE, 0,
2752 0, &log, false, &bbr->rc_tv);
2757 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay)
2759 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2760 union tcp_log_stackspecific log;
2762 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2763 log.u_bbr.flex1 = slot;
2764 log.u_bbr.flex2 = del_by;
2765 log.u_bbr.flex3 = prev_delay;
2766 log.u_bbr.flex4 = line;
2767 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2768 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2769 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2770 log.u_bbr.flex8 = bbr->rc_in_persist;
2771 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2772 &bbr->rc_inp->inp_socket->so_rcv,
2773 &bbr->rc_inp->inp_socket->so_snd,
2775 len, &log, false, &bbr->rc_tv);
2780 bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags)
2782 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2783 union tcp_log_stackspecific log;
2785 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2786 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2787 log.u_bbr.flex2 = 0;
2788 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2789 log.u_bbr.flex4 = end;
2790 log.u_bbr.flex5 = seq;
2791 log.u_bbr.flex6 = t;
2792 log.u_bbr.flex7 = match;
2793 log.u_bbr.flex8 = flags;
2794 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2795 &bbr->rc_inp->inp_socket->so_rcv,
2796 &bbr->rc_inp->inp_socket->so_snd,
2798 0, &log, false, &bbr->rc_tv);
2803 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2805 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2806 union tcp_log_stackspecific log;
2808 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2809 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2810 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2811 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2812 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2813 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2814 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2815 log.u_bbr.flex7 = 0;
2816 log.u_bbr.flex8 = entry_method;
2817 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2818 &bbr->rc_inp->inp_socket->so_rcv,
2819 &bbr->rc_inp->inp_socket->so_snd,
2820 BBR_LOG_EXIT_GAIN, 0,
2821 0, &log, false, &bbr->rc_tv);
2826 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2828 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2829 union tcp_log_stackspecific log;
2831 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2833 log.u_bbr.flex1 = 0;
2834 log.u_bbr.flex2 = 0;
2835 log.u_bbr.flex3 = 0;
2836 log.u_bbr.flex4 = 0;
2837 log.u_bbr.flex7 = 0;
2838 log.u_bbr.flex8 = settings_desired;
2840 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2841 &bbr->rc_inp->inp_socket->so_rcv,
2842 &bbr->rc_inp->inp_socket->so_snd,
2843 BBR_LOG_SETTINGS_CHG, 0,
2844 0, &log, false, &bbr->rc_tv);
2849 * Returns the bw from the our filter.
2851 static inline uint64_t
2852 bbr_get_full_bw(struct tcp_bbr *bbr)
2856 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2862 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2867 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2868 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2871 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2876 calclr *= (uint64_t)1000;
2877 calclr /= (uint64_t)del;
2879 /* Nothing delivered? 100.0% loss */
2882 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr;
2883 if (IN_RECOVERY(bbr->rc_tp->t_flags))
2884 bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2885 bbr->r_ctl.rc_pkt_epoch++;
2886 if (bbr->rc_no_pacing &&
2887 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2888 bbr->rc_no_pacing = 0;
2889 tcp_bbr_tso_size_check(bbr, cts);
2891 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2892 bbr->r_ctl.rc_pkt_epoch_time = cts;
2893 /* What was our loss rate */
2894 bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2895 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2896 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2900 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2902 uint32_t epoch_time;
2904 /* Tick the RTT clock */
2905 bbr->r_ctl.rc_rtt_epoch++;
2906 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2907 bbr_log_time_epoch(bbr, cts, line, epoch_time);
2908 bbr->r_ctl.rc_rcv_epoch_start = cts;
2912 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2914 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2915 bbr->rc_is_pkt_epoch_now = 1;
2920 * Returns the bw from either the b/w filter
2921 * or from the lt_bw (if the connection is being
2924 static inline uint64_t
2925 __bbr_get_bw(struct tcp_bbr *bbr)
2927 uint64_t bw, min_bw;
2929 int gm_measure_cnt = 1;
2932 * For startup we make, like google, a
2933 * minimum b/w. This is generated from the
2934 * IW and the rttProp. We do fall back to srtt
2935 * if for some reason (initial handshake) we don't
2936 * have a rttProp. We, in the worst case, fall back
2937 * to the configured min_bw (rc_initial_hptsi_bw).
2939 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2940 /* Attempt first to use rttProp */
2941 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2942 if (rtt && (rtt < 0xffffffff)) {
2944 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2945 ((uint64_t)1000000);
2947 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2948 min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2951 } else if (bbr->rc_tp->t_srtt != 0) {
2952 /* No rttProp, use srtt? */
2953 rtt = bbr_get_rtt(bbr, BBR_SRTT);
2956 min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2961 if ((bbr->rc_past_init_win == 0) &&
2962 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2963 bbr->rc_past_init_win = 1;
2964 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1))
2966 if (gm_measure_cnt &&
2967 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2968 (bbr->rc_past_init_win == 0))) {
2969 /* For google we use our guess rate until we get 1 measurement */
2972 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2973 if (rtt && (rtt < 0xffffffff)) {
2975 * We have an RTT measurment. Use that in
2976 * combination with our initial window to calculate
2979 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2980 ((uint64_t)1000000);
2982 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2983 bw = bbr->r_ctl.rc_initial_hptsi_bw;
2986 /* Drop back to the 40 and punt to a default */
2987 bw = bbr->r_ctl.rc_initial_hptsi_bw;
2990 /* Probably should panic */
2997 if (bbr->rc_lt_use_bw)
2998 bw = bbr->r_ctl.rc_lt_bw;
2999 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
3000 bw = bbr->r_ctl.red_bw;
3002 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3003 if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
3005 * Enforce user set rate limit, keep in mind that
3006 * t_peakrate_thr is in B/s already
3008 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3011 /* We should not be at 0, go to the initial window then */
3012 goto use_initial_window;
3015 /* Probably should panic */
3022 static inline uint64_t
3023 bbr_get_bw(struct tcp_bbr *bbr)
3027 bw = __bbr_get_bw(bbr);
3032 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3034 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3035 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3036 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3037 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3041 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3043 bbr->rc_lt_is_sampling = 0;
3044 bbr->rc_lt_use_bw = 0;
3045 bbr->r_ctl.rc_lt_bw = 0;
3046 bbr_reset_lt_bw_interval(bbr, cts);
3050 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3054 /* Do we have a previous sample? */
3055 if (bbr->r_ctl.rc_lt_bw) {
3056 /* Get the diff in bytes per second */
3057 if (bbr->r_ctl.rc_lt_bw > bw)
3058 diff = bbr->r_ctl.rc_lt_bw - bw;
3060 diff = bw - bbr->r_ctl.rc_lt_bw;
3061 if ((diff <= bbr_lt_bw_diff) ||
3062 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3063 /* Consider us policed */
3066 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3067 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */
3068 bbr->rc_lt_use_bw = 1;
3069 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3071 * Use pkt based epoch for measuring length of
3074 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3076 * reason 4 is we need to start consider being
3079 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3083 bbr->r_ctl.rc_lt_bw = bw;
3084 bbr_reset_lt_bw_interval(bbr, cts);
3085 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3089 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3091 uint32_t ran, deduct;
3093 ran = arc4random_uniform(bbr_rand_ot);
3095 deduct = bbr->r_ctl.rc_level_state_extra / ran;
3096 bbr->r_ctl.rc_level_state_extra -= deduct;
3100 * Return randomly the starting state
3101 * to use in probebw.
3104 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3109 /* Initialize the offset to 0 */
3110 bbr->r_ctl.rc_exta_time_gd = 0;
3111 bbr->rc_hit_state_1 = 0;
3112 bbr->r_ctl.rc_level_state_extra = 0;
3113 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3115 * The math works funny here :) the return value is used to set the
3116 * substate and then the state change is called which increments by
3117 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3118 * we fully enter the state. Note that the (8 - 1 - ran) assures that
3119 * we return 1 - 7, so we dont return 0 and end up starting in
3122 ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3124 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3125 bbr_set_epoch(bbr, cts, __LINE__);
3127 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3132 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3134 uint32_t diff, d_time;
3135 uint64_t del_time, bw, lost, delivered;
3137 if (bbr->r_use_policer == 0)
3139 if (bbr->rc_lt_use_bw) {
3140 /* We are using lt bw do we stop yet? */
3141 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3142 if (diff > bbr_lt_bw_max_rtts) {
3145 bbr_reset_lt_bw_sampling(bbr, cts);
3146 if (bbr->rc_filled_pipe) {
3147 bbr_set_epoch(bbr, cts, __LINE__);
3148 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3149 bbr_substate_change(bbr, cts, __LINE__, 0);
3150 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3151 bbr_log_type_statechange(bbr, cts, __LINE__);
3154 * This should not happen really
3155 * unless we remove the startup/drain
3156 * restrictions above.
3158 bbr->rc_bbr_state = BBR_STATE_STARTUP;
3159 bbr_set_epoch(bbr, cts, __LINE__);
3160 bbr->r_ctl.rc_bbr_state_time = cts;
3161 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3162 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3163 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3164 bbr_set_state_target(bbr, __LINE__);
3165 bbr_log_type_statechange(bbr, cts, __LINE__);
3167 /* reason 0 is to stop using lt-bw */
3168 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3171 if (bbr_lt_intvl_fp == 0) {
3172 /* Not doing false-postive detection */
3175 /* False positive detection */
3176 if (diff == bbr_lt_intvl_fp) {
3177 /* At bbr_lt_intvl_fp we record the lost */
3178 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3179 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3180 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3181 /* Now is our loss rate still high? */
3182 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3183 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3184 if ((delivered == 0) ||
3185 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3186 /* No still below our threshold */
3187 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3189 /* Yikes its still high, it must be a false positive */
3190 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3197 * Wait for the first loss before sampling, to let the policer
3198 * exhaust its tokens and estimate the steady-state rate allowed by
3199 * the policer. Starting samples earlier includes bursts that
3200 * over-estimate the bw.
3202 if (bbr->rc_lt_is_sampling == 0) {
3203 /* reason 1 is to begin doing the sampling */
3204 if (loss_detected == 0)
3206 bbr_reset_lt_bw_interval(bbr, cts);
3207 bbr->rc_lt_is_sampling = 1;
3208 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3211 /* Now how long were we delivering long term last> */
3212 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3213 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3217 /* To avoid underestimates, reset sampling if we run out of data. */
3218 if (bbr->r_ctl.r_app_limited_until) {
3219 /* Can not measure in app-limited state */
3220 bbr_reset_lt_bw_sampling(bbr, cts);
3221 /* reason 2 is to reset sampling due to app limits */
3222 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3225 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3226 if (diff < bbr_lt_intvl_min_rtts) {
3228 * need more samples (we don't
3229 * start on a round like linux so
3232 /* 6 is not_enough time or no-loss */
3233 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3236 if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3238 * For now if we wait too long, reset all sampling. We need
3239 * to do some research here, its possible that we should
3240 * base this on how much loss as occurred.. something like
3241 * if its under 10% (or some thresh) reset all otherwise
3242 * don't. Thats for phase II I guess.
3244 bbr_reset_lt_bw_sampling(bbr, cts);
3245 /* reason 3 is to reset sampling due too long of sampling */
3246 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3250 * End sampling interval when a packet is lost, so we estimate the
3251 * policer tokens were exhausted. Stopping the sampling before the
3252 * tokens are exhausted under-estimates the policed rate.
3254 if (loss_detected == 0) {
3255 /* 6 is not_enough time or no-loss */
3256 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3259 /* Calculate packets lost and delivered in sampling interval. */
3260 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3261 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3262 if ((delivered == 0) ||
3263 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3264 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3267 if (d_time < 1000) {
3268 /* Not enough time. wait */
3269 /* 6 is not_enough time or no-loss */
3270 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3273 if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3275 bbr_reset_lt_bw_sampling(bbr, cts);
3276 /* reason 3 is to reset sampling due too long of sampling */
3277 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3282 bw *= (uint64_t)USECS_IN_SECOND;
3284 bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3288 * Allocate a sendmap from our zone.
3290 static struct bbr_sendmap *
3291 bbr_alloc(struct tcp_bbr *bbr)
3293 struct bbr_sendmap *rsm;
3295 BBR_STAT_INC(bbr_to_alloc);
3296 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3298 bbr->r_ctl.rc_num_maps_alloced++;
3301 if (bbr->r_ctl.rc_free_cnt) {
3302 BBR_STAT_INC(bbr_to_alloc_emerg);
3303 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3304 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3305 bbr->r_ctl.rc_free_cnt--;
3308 BBR_STAT_INC(bbr_to_alloc_failed);
3312 static struct bbr_sendmap *
3313 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3315 if ((V_tcp_map_entries_limit > 0) &&
3316 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3317 BBR_STAT_INC(bbr_alloc_limited);
3318 if (!bbr->alloc_limit_reported) {
3319 bbr->alloc_limit_reported = 1;
3320 BBR_STAT_INC(bbr_alloc_limited_conns);
3324 return (bbr_alloc(bbr));
3327 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3328 static struct bbr_sendmap *
3329 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3331 struct bbr_sendmap *rsm;
3334 /* currently there is only one limit type */
3335 if (V_tcp_map_split_limit > 0 &&
3336 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3337 BBR_STAT_INC(bbr_split_limited);
3338 if (!bbr->alloc_limit_reported) {
3339 bbr->alloc_limit_reported = 1;
3340 BBR_STAT_INC(bbr_alloc_limited_conns);
3346 /* allocate and mark in the limit type, if set */
3347 rsm = bbr_alloc(bbr);
3348 if (rsm != NULL && limit_type) {
3349 rsm->r_limit_type = limit_type;
3350 bbr->r_ctl.rc_num_split_allocs++;
3356 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3358 if (rsm->r_limit_type) {
3359 /* currently there is only one limit type */
3360 bbr->r_ctl.rc_num_split_allocs--;
3362 if (rsm->r_is_smallmap)
3363 bbr->r_ctl.rc_num_small_maps_alloced--;
3364 if (bbr->r_ctl.rc_tlp_send == rsm)
3365 bbr->r_ctl.rc_tlp_send = NULL;
3366 if (bbr->r_ctl.rc_resend == rsm) {
3367 bbr->r_ctl.rc_resend = NULL;
3369 if (bbr->r_ctl.rc_next == rsm)
3370 bbr->r_ctl.rc_next = NULL;
3371 if (bbr->r_ctl.rc_sacklast == rsm)
3372 bbr->r_ctl.rc_sacklast = NULL;
3373 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3374 memset(rsm, 0, sizeof(struct bbr_sendmap));
3375 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3376 rsm->r_limit_type = 0;
3377 bbr->r_ctl.rc_free_cnt++;
3380 bbr->r_ctl.rc_num_maps_alloced--;
3381 uma_zfree(bbr_zone, rsm);
3388 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3390 * Calculate the bytes in flight needed given the bw (in bytes per
3391 * second) and the specifyed rtt in useconds. We need to put out the
3392 * returned value per RTT to match that rate. Gain will normally
3393 * raise it up from there.
3395 * This should not overflow as long as the bandwidth is below 1
3396 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3397 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3399 uint64_t usec_per_sec;
3401 usec_per_sec = USECS_IN_SECOND;
3402 return ((rtt * bw) / usec_per_sec);
3406 * Return the initial cwnd.
3409 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3413 if (bbr->rc_init_win) {
3414 i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3415 } else if (V_tcp_initcwnd_segments)
3416 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3417 max(2 * tp->t_maxseg, 14600));
3418 else if (V_tcp_do_rfc3390)
3419 i_cwnd = min(4 * tp->t_maxseg,
3420 max(2 * tp->t_maxseg, 4380));
3422 /* Per RFC5681 Section 3.1 */
3423 if (tp->t_maxseg > 2190)
3424 i_cwnd = 2 * tp->t_maxseg;
3425 else if (tp->t_maxseg > 1095)
3426 i_cwnd = 3 * tp->t_maxseg;
3428 i_cwnd = 4 * tp->t_maxseg;
3434 * Given a specified gain, return the target
3435 * cwnd based on that gain.
3438 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3443 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3444 (bbr_get_full_bw(bbr) == 0)) {
3445 /* No measurements yet */
3446 return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3449 * Get bytes per RTT needed (rttProp is normally in
3450 * bbr_cwndtarget_rtt_touse)
3452 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3453 /* Get the bdp from the two values */
3454 bdp = bbr_get_bw_delay_prod(rtt, bw);
3455 /* Now apply the gain */
3456 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3462 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3466 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3467 /* Get the base cwnd with gain rounded to a mss */
3468 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3470 * Add in N (2 default since we do not have a
3471 * fq layer to trap packets in) quanta's per the I-D
3472 * section 4.2.3.2 quanta adjust.
3474 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3475 if (bbr->rc_use_google) {
3476 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3477 (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3479 * The linux implementation adds
3480 * an extra 2 x mss in gain cycle which
3481 * is documented no-where except in the code.
3482 * so we add more for Neal undocumented feature
3486 if ((cwnd / mss) & 0x1) {
3487 /* Round up for odd num mss */
3491 /* Are we below the min cwnd? */
3492 if (cwnd < get_min_cwnd(bbr))
3493 return (get_min_cwnd(bbr));
3498 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3506 bbr_get_header_oh(struct tcp_bbr *bbr)
3511 if (bbr->r_ctl.rc_inc_tcp_oh) {
3512 /* Do we include TCP overhead? */
3513 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3515 if (bbr->r_ctl.rc_inc_ip_oh) {
3516 /* Do we include IP overhead? */
3519 seg_oh += sizeof(struct ip6_hdr);
3523 seg_oh += sizeof(struct ip);
3526 if (bbr->r_ctl.rc_inc_enet_oh) {
3527 /* Do we include the ethernet overhead? */
3528 seg_oh += sizeof(struct ether_header);
3534 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3536 uint64_t divor, res, tim;
3538 if (useconds_time == 0)
3540 gain = bbr_gain_adjust(bbr, gain);
3541 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3542 tim = useconds_time;
3543 res = (tim * bw * gain) / divor;
3546 return ((uint32_t)res);
3550 * Given a gain and a length return the delay in useconds that
3551 * should be used to evenly space out packets
3552 * on the connection (based on the gain factor).
3555 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3557 uint64_t bw, lentim, res;
3558 uint32_t usecs, srtt, over = 0;
3559 uint32_t seg_oh, num_segs, maxseg;
3564 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3565 num_segs = (len + maxseg - 1) / maxseg;
3566 if (bbr->rc_use_google == 0) {
3567 seg_oh = bbr_get_header_oh(bbr);
3568 len += (num_segs * seg_oh);
3570 gain = bbr_gain_adjust(bbr, gain);
3571 bw = bbr_get_bw(bbr);
3572 if (bbr->rc_use_google) {
3576 * Reduce the b/w by the google discount
3579 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3580 cbw /= (uint64_t)1000;
3581 /* We don't apply a discount if it results in 0 */
3585 lentim = ((uint64_t)len *
3586 (uint64_t)USECS_IN_SECOND *
3587 (uint64_t)BBR_UNIT);
3588 res = lentim / ((uint64_t)gain * bw);
3591 usecs = (uint32_t)res;
3592 srtt = bbr_get_rtt(bbr, BBR_SRTT);
3593 if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3594 (bbr->rc_use_google == 0) &&
3595 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3597 * We cannot let the delay be more than 1/2 the srtt time.
3598 * Otherwise we cannot pace out or send properly.
3600 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3601 BBR_STAT_INC(bbr_hpts_min_time);
3604 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3609 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3610 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3612 INP_WLOCK_ASSERT(tp->t_inpcb);
3614 uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3618 if ((tp->t_flags & TF_GPUTINPROG) &&
3619 SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3621 * Strech acks and compressed acks will cause this to
3622 * oscillate but we are doing it the same way as the main
3623 * stack so it will be compariable (though possibly not
3627 int64_t gput, time_stamp;
3629 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3630 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3631 cgput = gput / time_stamp;
3632 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3634 if (tp->t_stats_gput_prev > 0)
3635 stats_voi_update_abs_s32(tp->t_stats,
3637 ((gput - tp->t_stats_gput_prev) * 100) /
3638 tp->t_stats_gput_prev);
3639 tp->t_flags &= ~TF_GPUTINPROG;
3640 tp->t_stats_gput_prev = cgput;
3643 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3644 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3645 /* We don't change anything in probe-rtt */
3648 maxseg = tp->t_maxseg - bbr->rc_last_options;
3649 saved_bytes = bytes_this_ack;
3650 bytes_this_ack += sack_changed;
3651 if (bytes_this_ack > prev_acked) {
3652 bytes_this_ack -= prev_acked;
3654 * A byte ack'd gives us a full mss
3655 * to be like linux i.e. they count packets.
3657 if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3658 bytes_this_ack = maxseg;
3663 cwnd = tp->snd_cwnd;
3664 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3666 target_cwnd = bbr_get_target_cwnd(bbr,
3668 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3670 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3671 if (IN_RECOVERY(tp->t_flags) &&
3672 (bbr->bbr_prev_in_rec == 0)) {
3674 * We are entering recovery and
3675 * thus packet conservation.
3677 bbr->pkt_conservation = 1;
3678 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3679 cwnd = ctf_flight_size(tp,
3680 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3683 if (IN_RECOVERY(tp->t_flags)) {
3686 bbr->bbr_prev_in_rec = 1;
3687 if (cwnd > losses) {
3693 flight = ctf_flight_size(tp,
3694 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3695 bbr_log_type_cwndupd(bbr, flight, 0,
3696 losses, 10, 0, 0, line);
3697 if (bbr->pkt_conservation) {
3700 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3701 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3705 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3706 /* Clear packet conservation after an rttProp */
3707 bbr->pkt_conservation = 0;
3709 if ((flight + bytes_this_ack) > cwnd)
3710 cwnd = flight + bytes_this_ack;
3711 if (cwnd < get_min_cwnd(bbr))
3712 cwnd = get_min_cwnd(bbr);
3713 tp->snd_cwnd = cwnd;
3714 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3715 prev_acked, 1, target_cwnd, th->th_ack, line);
3720 bbr->bbr_prev_in_rec = 0;
3721 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3722 bbr->r_ctl.restrict_growth--;
3723 if (bytes_this_ack > maxseg)
3724 bytes_this_ack = maxseg;
3726 if (bbr->rc_filled_pipe) {
3728 * Here we have exited startup and filled the pipe. We will
3729 * thus allow the cwnd to shrink to the target. We hit here
3735 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3738 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3742 * Here we are still in startup, we increase cwnd by what
3745 if ((cwnd < target_cwnd) ||
3746 (bbr->rc_past_init_win == 0)) {
3748 cwnd += bytes_this_ack;
3751 * Method 4 means we are at target so no gain in
3752 * startup and past the initial window.
3757 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3758 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3762 tcp_bbr_partialack(struct tcpcb *tp)
3764 struct tcp_bbr *bbr;
3766 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3767 INP_WLOCK_ASSERT(tp->t_inpcb);
3768 if (ctf_flight_size(tp,
3769 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
3771 bbr->r_wanted_output = 1;
3776 bbr_post_recovery(struct tcpcb *tp)
3778 struct tcp_bbr *bbr;
3781 INP_WLOCK_ASSERT(tp->t_inpcb);
3782 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3784 * Here we just exit recovery.
3786 EXIT_RECOVERY(tp->t_flags);
3787 /* Lock in our b/w reduction for the specified number of pkt-epochs */
3788 bbr->r_recovery_bw = 0;
3789 tp->snd_recover = tp->snd_una;
3790 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3791 bbr->pkt_conservation = 0;
3792 if (bbr->rc_use_google == 0) {
3794 * For non-google mode lets
3795 * go ahead and make sure we clear
3796 * the recovery state so if we
3797 * bounce back in to recovery we
3800 bbr->bbr_prev_in_rec = 0;
3802 bbr_log_type_exit_rec(bbr);
3803 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3804 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3805 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3807 /* For probe-rtt case lets fix up its saved_cwnd */
3808 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3809 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3810 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3813 flight = ctf_flight_size(tp,
3814 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3815 if ((bbr->rc_use_google == 0) &&
3817 uint64_t val, lr2use;
3818 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3821 if (bbr_get_rtt(bbr, BBR_SRTT)) {
3822 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3823 val /= bbr_get_rtt(bbr, BBR_SRTT);
3824 ratio = (uint32_t)val;
3828 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3829 bbr->r_ctl.recovery_lr, 21,
3831 bbr->r_ctl.rc_red_cwnd_pe,
3833 if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3835 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3836 bbr_prtt_slam_cwnd) ||
3837 (bbr_sub_drain_slam_cwnd &&
3838 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3839 bbr->rc_hit_state_1 &&
3840 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3841 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3842 bbr_slam_cwnd_in_main_drain)) {
3844 * Here we must poke at the saved cwnd
3845 * as well as the cwnd.
3847 cwnd = bbr->r_ctl.rc_saved_cwnd;
3848 cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3850 cwnd = tp->snd_cwnd;
3851 cwnd_p = &tp->snd_cwnd;
3853 maxseg = tp->t_maxseg - bbr->rc_last_options;
3854 /* Add the overall lr with the recovery lr */
3855 if (bbr->r_ctl.rc_lost == 0)
3857 else if (bbr->r_ctl.rc_delivered == 0)
3860 lr2use = bbr->r_ctl.rc_lost * 1000;
3861 lr2use /= bbr->r_ctl.rc_delivered;
3863 lr2use += bbr->r_ctl.recovery_lr;
3864 acks_inflight = (flight / (maxseg * 2));
3865 if (bbr_red_scale) {
3866 lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3867 lr2use /= bbr_red_scale;
3868 if ((bbr_red_growth_restrict) &&
3869 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3870 bbr->r_ctl.restrict_growth += acks_inflight;
3873 val = (uint64_t)cwnd * lr2use;
3876 newcwnd = roundup((cwnd - val), maxseg);
3880 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3881 val /= (uint64_t)bbr_red_div;
3882 newcwnd = roundup((uint32_t)val, maxseg);
3884 /* with standard delayed acks how many acks can I expect? */
3885 if (bbr_drop_limit == 0) {
3887 * Anticpate how much we will
3888 * raise the cwnd based on the acks.
3890 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3891 /* We do enforce the min (with the acks) */
3892 newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3896 * A strict drop limit of N is is inplace
3898 if (newcwnd < (bbr_drop_limit * maxseg)) {
3899 newcwnd = bbr_drop_limit * maxseg;
3902 /* For the next N acks do we restrict the growth */
3904 if (tp->snd_cwnd > newcwnd)
3905 tp->snd_cwnd = newcwnd;
3906 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3908 bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3909 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3912 bbr->r_ctl.recovery_lr = 0;
3913 if (flight <= tp->snd_cwnd) {
3914 bbr->r_wanted_output = 1;
3916 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3920 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3922 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3923 /* Limit the drop in b/w to 1/2 our current filter. */
3924 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3925 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3926 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3927 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3928 tcp_bbr_tso_size_check(bbr, cts);
3932 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3934 struct tcp_bbr *bbr;
3936 INP_WLOCK_ASSERT(tp->t_inpcb);
3938 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
3940 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3943 if (!IN_RECOVERY(tp->t_flags)) {
3944 tp->snd_recover = tp->snd_max;
3945 /* Start a new epoch */
3946 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3947 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3949 * Move forward the lt epoch
3950 * so it won't count the truncated
3953 bbr->r_ctl.rc_lt_epoch++;
3955 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
3957 * Just like the policer detection code
3958 * if we are in startup we must push
3959 * forward the last startup epoch
3960 * to hide the truncated PE.
3962 bbr->r_ctl.rc_bbr_last_startup_epoch++;
3964 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
3965 ENTER_RECOVERY(tp->t_flags);
3966 bbr->rc_tlp_rtx_out = 0;
3967 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
3968 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3969 if (bbr->rc_inp->inp_in_hpts &&
3970 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
3972 * When we enter recovery, we need to restart
3973 * any timers. This may mean we gain an agg
3974 * early, which will be made up for at the last
3977 bbr->rc_timer_first = 1;
3978 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
3981 * Calculate a new cwnd based on to the current
3982 * delivery rate with no gain. We get the bdp
3983 * without gaining it up like we normally would and
3984 * we use the last cur_del_rate.
3986 if ((bbr->rc_use_google == 0) &&
3987 (bbr->r_ctl.bbr_rttprobe_gain_val ||
3988 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
3989 tp->snd_cwnd = ctf_flight_size(tp,
3990 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3991 (tp->t_maxseg - bbr->rc_last_options);
3992 if (tp->snd_cwnd < get_min_cwnd(bbr)) {
3993 /* We always gate to min cwnd */
3994 tp->snd_cwnd = get_min_cwnd(bbr);
3996 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
3998 bbr_log_type_enter_rec(bbr, rsm->r_start);
4002 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4003 /* RTO was unnecessary, so reset everything. */
4004 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
4005 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
4006 tp->snd_cwnd = tp->snd_cwnd_prev;
4007 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4008 tp->snd_recover = tp->snd_recover_prev;
4009 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4010 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4012 tp->t_badrxtwin = 0;
4018 * Indicate whether this ack should be delayed. We can delay the ack if
4019 * following conditions are met:
4020 * - There is no delayed ack timer in progress.
4021 * - Our last ack wasn't a 0-sized window. We never want to delay
4022 * the ack that opens up a 0-sized window.
4023 * - LRO wasn't used for this segment. We make sure by checking that the
4024 * segment size is not larger than the MSS.
4025 * - Delayed acks are enabled or this is a half-synchronized T/TCP
4027 * - The data being acked is less than a full segment (a stretch ack
4028 * of more than a segment we should ack.
4029 * - nsegs is 1 (if its more than that we received more than 1 ack).
4031 #define DELAY_ACK(tp, bbr, nsegs) \
4032 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4033 ((tp->t_flags & TF_DELACK) == 0) && \
4034 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \
4035 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4038 * Return the lowest RSM in the map of
4039 * packets still in flight that is not acked.
4040 * This should normally find on the first one
4041 * since we remove packets from the send
4042 * map after they are marked ACKED.
4044 static struct bbr_sendmap *
4045 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4047 struct bbr_sendmap *rsm;
4050 * Walk the time-order transmitted list looking for an rsm that is
4051 * not acked. This will be the one that was sent the longest time
4052 * ago that is still outstanding.
4054 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4055 if (rsm->r_flags & BBR_ACKED) {
4064 static struct bbr_sendmap *
4065 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4067 struct bbr_sendmap *prsm;
4070 * Walk the sequence order list backward until we hit and arrive at
4071 * the highest seq not acked. In theory when this is called it
4072 * should be the last segment (which it was not).
4075 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4076 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4085 * Returns to the caller the number of microseconds that
4086 * the packet can be outstanding before we think we
4087 * should have had an ack returned.
4090 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4093 * lro is the flag we use to determine if we have seen reordering.
4094 * If it gets set we have seen reordering. The reorder logic either
4095 * works in one of two ways:
4097 * If reorder-fade is configured, then we track the last time we saw
4098 * re-ordering occur. If we reach the point where enough time as
4099 * passed we no longer consider reordering has occuring.
4101 * Or if reorder-face is 0, then once we see reordering we consider
4102 * the connection to alway be subject to reordering and just set lro
4105 * In the end if lro is non-zero we add the extra time for
4109 uint32_t thresh, t_rxtcur;
4113 if (bbr->r_ctl.rc_reorder_ts) {
4114 if (bbr->r_ctl.rc_reorder_fade) {
4115 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4116 lro = cts - bbr->r_ctl.rc_reorder_ts;
4119 * No time as passed since the last
4120 * reorder, mark it as reordering.
4125 /* Negative time? */
4128 if (lro > bbr->r_ctl.rc_reorder_fade) {
4129 /* Turn off reordering seen too */
4130 bbr->r_ctl.rc_reorder_ts = 0;
4134 /* Reodering does not fade */
4140 thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4142 /* It must be set, if not you get 1/4 rtt */
4143 if (bbr->r_ctl.rc_reorder_shift)
4144 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4146 thresh += (srtt >> 2);
4150 /* We don't let the rack timeout be above a RTO */
4151 if ((bbr->rc_tp)->t_srtt == 0)
4152 t_rxtcur = BBR_INITIAL_RTO;
4154 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4155 if (thresh > t_rxtcur) {
4158 /* And we don't want it above the RTO max either */
4159 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4160 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4162 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4167 * Return to the caller the amount of time in mico-seconds
4168 * that should be used for the TLP timer from the last
4169 * send time of this packet.
4172 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4173 struct bbr_sendmap *rsm, uint32_t srtt,
4176 uint32_t thresh, len, maxseg, t_rxtcur;
4177 struct bbr_sendmap *prsm;
4181 if (bbr->rc_tlp_threshold)
4182 thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4184 thresh = (srtt * 2);
4185 maxseg = tp->t_maxseg - bbr->rc_last_options;
4186 /* Get the previous sent packet, if any */
4187 len = rsm->r_end - rsm->r_start;
4190 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4191 if (prsm && (len <= maxseg)) {
4193 * Two packets outstanding, thresh should be (2*srtt) +
4194 * possible inter-packet delay (if any).
4196 uint32_t inter_gap = 0;
4199 idx = rsm->r_rtr_cnt - 1;
4200 nidx = prsm->r_rtr_cnt - 1;
4201 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4202 /* Yes it was sent later (or at the same time) */
4203 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4205 thresh += inter_gap;
4206 } else if (len <= maxseg) {
4208 * Possibly compensate for delayed-ack.
4210 uint32_t alt_thresh;
4212 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4213 if (alt_thresh > thresh)
4214 thresh = alt_thresh;
4216 /* Not above the current RTO */
4217 if (tp->t_srtt == 0)
4218 t_rxtcur = BBR_INITIAL_RTO;
4220 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4222 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4223 /* Not above an RTO */
4224 if (thresh > t_rxtcur) {
4227 /* Not above a RTO max */
4228 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4229 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4231 /* And now apply the user TLP min */
4232 if (thresh < bbr_tlp_min) {
4233 thresh = bbr_tlp_min;
4239 * Return one of three RTTs to use (in microseconds).
4241 static __inline uint32_t
4242 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4247 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4248 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4249 /* We have no rtt at all */
4250 if (bbr->rc_tp->t_srtt == 0)
4251 f_rtt = BBR_INITIAL_RTO;
4253 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4255 * Since we don't know how good the rtt is apply a
4258 if (f_rtt < bbr_delayed_ack_time) {
4259 f_rtt = bbr_delayed_ack_time;
4262 /* Take the filter version or last measured pkt-rtt */
4263 if (rtt_type == BBR_RTT_PROP) {
4265 } else if (rtt_type == BBR_RTT_PKTRTT) {
4266 if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4267 srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4269 /* No pkt rtt yet */
4272 } else if (rtt_type == BBR_RTT_RACK) {
4273 srtt = bbr->r_ctl.rc_last_rtt;
4274 /* We need to add in any internal delay for our timer */
4275 if (bbr->rc_ack_was_delayed)
4276 srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4277 } else if (rtt_type == BBR_SRTT) {
4278 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4282 #ifdef BBR_INVARIANTS
4283 panic("Unknown rtt request type %d", rtt_type);
4290 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4294 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4296 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4297 /* It is lost (past time) */
4304 * Return a sendmap if we need to retransmit something.
4306 static struct bbr_sendmap *
4307 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4310 * Check to see that we don't need to fall into recovery. We will
4311 * need to do so if our oldest transmit is past the time we should
4315 struct bbr_sendmap *rsm;
4318 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4319 /* Nothing outstanding that we know of */
4322 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4324 /* Nothing in the transmit map */
4327 if (tp->t_flags & TF_SENTFIN) {
4328 /* Fin restricted, don't find anything once a fin is sent */
4331 if (rsm->r_flags & BBR_ACKED) {
4333 * Ok the first one is acked (this really should not happen
4334 * since we remove the from the tmap once they are acked)
4336 rsm = bbr_find_lowest_rsm(bbr);
4340 idx = rsm->r_rtr_cnt - 1;
4341 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4342 /* Send timestamp is the same or less? can't be ready */
4345 /* Get our RTT time */
4346 if (bbr_is_lost(bbr, rsm, cts) &&
4347 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4348 (rsm->r_flags & BBR_SACK_PASSED))) {
4349 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4350 rsm->r_flags |= BBR_MARKED_LOST;
4351 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4352 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4354 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4355 #ifdef BBR_INVARIANTS
4356 if ((rsm->r_end - rsm->r_start) == 0)
4357 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4365 * RACK Timer, here we simply do logging and house keeping.
4366 * the normal bbr_output_wtime() function will call the
4367 * appropriate thing to check if we need to do a RACK retransmit.
4368 * We return 1, saying don't proceed with bbr_output_wtime only
4369 * when all timers have been stopped (destroyed PCB?).
4372 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4375 * This timer simply provides an internal trigger to send out data.
4376 * The check_recovery_mode call will see if there are needed
4377 * retransmissions, if so we will enter fast-recovery. The output
4378 * call may or may not do the same thing depending on sysctl
4383 if (bbr->rc_all_timers_stopped) {
4386 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4387 /* Its not time yet */
4390 BBR_STAT_INC(bbr_to_tot);
4391 lost = bbr->r_ctl.rc_lost;
4392 if (bbr->r_state && (bbr->r_state != tp->t_state))
4393 bbr_set_state(tp, bbr, 0);
4394 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4395 if (bbr->r_ctl.rc_resend == NULL) {
4396 /* Lets do the check here */
4397 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4399 if (bbr_policer_call_from_rack_to)
4400 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4401 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4405 static __inline void
4406 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4410 nrsm->r_start = start;
4411 nrsm->r_end = rsm->r_end;
4412 nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4413 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed;
4414 nrsm->r_flags = rsm->r_flags;
4415 /* We don't transfer forward the SYN flag */
4416 nrsm->r_flags &= ~BBR_HAS_SYN;
4417 /* We move forward the FIN flag, not that this should happen */
4418 rsm->r_flags &= ~BBR_HAS_FIN;
4419 nrsm->r_dupack = rsm->r_dupack;
4420 nrsm->r_rtr_bytes = 0;
4421 nrsm->r_is_gain = rsm->r_is_gain;
4422 nrsm->r_is_drain = rsm->r_is_drain;
4423 nrsm->r_delivered = rsm->r_delivered;
4424 nrsm->r_ts_valid = rsm->r_ts_valid;
4425 nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4426 nrsm->r_del_time = rsm->r_del_time;
4427 nrsm->r_app_limited = rsm->r_app_limited;
4428 nrsm->r_first_sent_time = rsm->r_first_sent_time;
4429 nrsm->r_flight_at_send = rsm->r_flight_at_send;
4430 /* We split a piece the lower section looses any just_ret flag. */
4431 nrsm->r_bbr_state = rsm->r_bbr_state;
4432 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4433 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4435 rsm->r_end = nrsm->r_start;
4436 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4438 /* Check if we got too small */
4439 if ((rsm->r_is_smallmap == 0) &&
4440 ((rsm->r_end - rsm->r_start) <= idx)) {
4441 bbr->r_ctl.rc_num_small_maps_alloced++;
4442 rsm->r_is_smallmap = 1;
4444 /* Check the new one as well */
4445 if ((nrsm->r_end - nrsm->r_start) <= idx) {
4446 bbr->r_ctl.rc_num_small_maps_alloced++;
4447 nrsm->r_is_smallmap = 1;
4452 bbr_sack_mergable(struct bbr_sendmap *at,
4453 uint32_t start, uint32_t end)
4456 * Given a sack block defined by
4457 * start and end, and a current postion
4458 * at. Return 1 if either side of at
4459 * would show that the block is mergable
4460 * to that side. A block to be mergable
4461 * must have overlap with the start/end
4462 * and be in the SACK'd state.
4464 struct bbr_sendmap *l_rsm;
4465 struct bbr_sendmap *r_rsm;
4467 /* first get the either side blocks */
4468 l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4469 r_rsm = TAILQ_NEXT(at, r_next);
4470 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4471 /* Potentially mergeable */
4472 if ((l_rsm->r_end == start) ||
4473 (SEQ_LT(start, l_rsm->r_end) &&
4474 SEQ_GT(end, l_rsm->r_end))) {
4485 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4486 /* Potentially mergeable */
4487 if ((r_rsm->r_start == end) ||
4488 (SEQ_LT(start, r_rsm->r_start) &&
4489 SEQ_GT(end, r_rsm->r_start))) {
4491 * map blk |---------|
4494 * map blk |---------|
4495 * sack blk |-------|
4503 static struct bbr_sendmap *
4504 bbr_merge_rsm(struct tcp_bbr *bbr,
4505 struct bbr_sendmap *l_rsm,
4506 struct bbr_sendmap *r_rsm)
4509 * We are merging two ack'd RSM's,
4510 * the l_rsm is on the left (lower seq
4511 * values) and the r_rsm is on the right
4512 * (higher seq value). The simplest way
4513 * to merge these is to move the right
4514 * one into the left. I don't think there
4515 * is any reason we need to try to find
4516 * the oldest (or last oldest retransmitted).
4518 l_rsm->r_end = r_rsm->r_end;
4519 if (l_rsm->r_dupack < r_rsm->r_dupack)
4520 l_rsm->r_dupack = r_rsm->r_dupack;
4521 if (r_rsm->r_rtr_bytes)
4522 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4523 if (r_rsm->r_in_tmap) {
4524 /* This really should not happen */
4525 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4527 if (r_rsm->r_app_limited)
4528 l_rsm->r_app_limited = r_rsm->r_app_limited;
4530 if (r_rsm->r_flags & BBR_HAS_FIN)
4531 l_rsm->r_flags |= BBR_HAS_FIN;
4532 if (r_rsm->r_flags & BBR_TLP)
4533 l_rsm->r_flags |= BBR_TLP;
4534 if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4535 l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4536 if (r_rsm->r_flags & BBR_MARKED_LOST) {
4537 /* This really should not happen */
4538 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4540 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4541 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4542 /* Transfer the split limit to the map we free */
4543 r_rsm->r_limit_type = l_rsm->r_limit_type;
4544 l_rsm->r_limit_type = 0;
4546 bbr_free(bbr, r_rsm);
4551 * TLP Timer, here we simply setup what segment we want to
4552 * have the TLP expire on, the normal bbr_output_wtime() will then
4555 * We return 1, saying don't proceed with bbr_output_wtime only
4556 * when all timers have been stopped (destroyed PCB?).
4559 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4564 struct bbr_sendmap *rsm = NULL;
4567 uint32_t out, avail;
4569 int collapsed_win = 0;
4571 if (bbr->rc_all_timers_stopped) {
4574 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4575 /* Its not time yet */
4578 if (ctf_progress_timeout_check(tp, true)) {
4579 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4580 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4583 /* Did we somehow get into persists? */
4584 if (bbr->rc_in_persist) {
4587 if (bbr->r_state && (bbr->r_state != tp->t_state))
4588 bbr_set_state(tp, bbr, 0);
4589 BBR_STAT_INC(bbr_tlp_tot);
4590 maxseg = tp->t_maxseg - bbr->rc_last_options;
4592 * A TLP timer has expired. We have been idle for 2 rtts. So we now
4593 * need to figure out how to force a full MSS segment out.
4595 so = tp->t_inpcb->inp_socket;
4596 avail = sbavail(&so->so_snd);
4597 out = ctf_outstanding(tp);
4598 if (out > tp->snd_wnd) {
4599 /* special case, we need a retransmission */
4604 /* New data is available */
4608 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4609 /* not enough to fill a MTU and no-delay is off */
4612 /* Set the send-new override */
4613 if ((out + amm) <= tp->snd_wnd) {
4614 bbr->rc_tlp_new_data = 1;
4618 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4619 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4620 bbr->r_ctl.rc_tlp_send = NULL;
4622 BBR_STAT_INC(bbr_tlp_newdata);
4627 * Ok we need to arrange the last un-acked segment to be re-sent, or
4628 * optionally the first un-acked segment.
4630 if (collapsed_win == 0) {
4631 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4632 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4633 rsm = bbr_find_high_nonack(bbr, rsm);
4640 * We must find the last segment
4641 * that was acceptable by the client.
4643 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4644 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4650 /* None? if so send the first */
4651 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4656 if ((rsm->r_end - rsm->r_start) > maxseg) {
4658 * We need to split this the last segment in two.
4660 struct bbr_sendmap *nrsm;
4662 nrsm = bbr_alloc_full_limit(bbr);
4665 * We can't get memory to split, we can either just
4666 * not split it. Or retransmit the whole piece, lets
4667 * do the large send (BTLP :-) ).
4671 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4672 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4673 if (rsm->r_in_tmap) {
4674 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4675 nrsm->r_in_tmap = 1;
4677 rsm->r_flags &= (~BBR_HAS_FIN);
4681 bbr->r_ctl.rc_tlp_send = rsm;
4682 bbr->rc_tlp_rtx_out = 1;
4683 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4684 bbr->r_ctl.rc_tlp_seg_send_cnt++;
4687 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4688 bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4691 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4693 * Can't [re]/transmit a segment we have retranmitted the
4694 * max times. We need the retransmit timer to take over.
4697 bbr->rc_tlp_new_data = 0;
4698 bbr->r_ctl.rc_tlp_send = NULL;
4700 rsm->r_flags &= ~BBR_TLP;
4701 BBR_STAT_INC(bbr_tlp_retran_fail);
4704 rsm->r_flags |= BBR_TLP;
4706 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4707 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4709 * We have retransmitted to many times for TLP. Switch to
4710 * the regular RTO timer
4714 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4715 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4720 * Delayed ack Timer, here we simply need to setup the
4721 * ACK_NOW flag and remove the DELACK flag. From there
4722 * the output routine will send the ack out.
4724 * We only return 1, saying don't proceed, if all timers
4725 * are stopped (destroyed PCB?).
4728 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4730 if (bbr->rc_all_timers_stopped) {
4733 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4734 tp->t_flags &= ~TF_DELACK;
4735 tp->t_flags |= TF_ACKNOW;
4736 KMOD_TCPSTAT_INC(tcps_delack);
4737 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4742 * Here we send a KEEP-ALIVE like probe to the
4743 * peer, we do not send data.
4745 * We only return 1, saying don't proceed, if all timers
4746 * are stopped (destroyed PCB?).
4749 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4751 struct tcptemp *t_template;
4754 if (bbr->rc_all_timers_stopped) {
4757 if (bbr->rc_in_persist == 0)
4759 KASSERT(tp->t_inpcb != NULL,
4760 ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
4762 * Persistence timer into zero window. Force a byte to be output, if
4765 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4766 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4767 KMOD_TCPSTAT_INC(tcps_persisttimeo);
4769 * Have we exceeded the user specified progress time?
4771 if (ctf_progress_timeout_check(tp, true)) {
4772 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4773 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4777 * Hack: if the peer is dead/unreachable, we do not time out if the
4778 * window is closed. After a full backoff, drop the connection if
4779 * the idle time (no responses to probes) reaches the maximum
4780 * backoff that we would use if retransmitting.
4782 if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4783 (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4784 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4785 KMOD_TCPSTAT_INC(tcps_persistdrop);
4786 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4787 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4790 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4791 tp->snd_una == tp->snd_max) {
4792 bbr_exit_persist(tp, bbr, cts, __LINE__);
4797 * If the user has closed the socket then drop a persisting
4798 * connection after a much reduced timeout.
4800 if (tp->t_state > TCPS_CLOSE_WAIT &&
4801 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4802 KMOD_TCPSTAT_INC(tcps_persistdrop);
4803 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4804 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4807 t_template = tcpip_maketemplate(bbr->rc_inp);
4809 tcp_respond(tp, t_template->tt_ipgen,
4810 &t_template->tt_t, (struct mbuf *)NULL,
4811 tp->rcv_nxt, tp->snd_una - 1, 0);
4812 /* This sends an ack */
4813 if (tp->t_flags & TF_DELACK)
4814 tp->t_flags &= ~TF_DELACK;
4815 free(t_template, M_TEMP);
4817 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4819 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4825 * If a keepalive goes off, we had no other timers
4826 * happening. We always return 1 here since this
4827 * routine either drops the connection or sends
4828 * out a segment with respond.
4831 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4833 struct tcptemp *t_template;
4836 if (bbr->rc_all_timers_stopped) {
4839 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4841 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4843 * Keep-alive timer went off; send something or drop connection if
4844 * idle for too long.
4846 KMOD_TCPSTAT_INC(tcps_keeptimeo);
4847 if (tp->t_state < TCPS_ESTABLISHED)
4849 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4850 tp->t_state <= TCPS_CLOSING) {
4851 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4854 * Send a packet designed to force a response if the peer is
4855 * up and reachable: either an ACK if the connection is
4856 * still alive, or an RST if the peer has closed the
4857 * connection due to timeout or reboot. Using sequence
4858 * number tp->snd_una-1 causes the transmitted zero-length
4859 * segment to lie outside the receive window; by the
4860 * protocol spec, this requires the correspondent TCP to
4863 KMOD_TCPSTAT_INC(tcps_keepprobe);
4864 t_template = tcpip_maketemplate(inp);
4866 tcp_respond(tp, t_template->tt_ipgen,
4867 &t_template->tt_t, (struct mbuf *)NULL,
4868 tp->rcv_nxt, tp->snd_una - 1, 0);
4869 free(t_template, M_TEMP);
4872 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4875 KMOD_TCPSTAT_INC(tcps_keepdrops);
4876 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4877 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4882 * Retransmit helper function, clear up all the ack
4883 * flags and take care of important book keeping.
4886 bbr_remxt_tmr(struct tcpcb *tp)
4889 * The retransmit timer went off, all sack'd blocks must be
4892 struct bbr_sendmap *rsm, *trsm = NULL;
4893 struct tcp_bbr *bbr;
4896 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4897 cts = tcp_get_usecs(&bbr->rc_tv);
4898 lost = bbr->r_ctl.rc_lost;
4899 if (bbr->r_state && (bbr->r_state != tp->t_state))
4900 bbr_set_state(tp, bbr, 0);
4902 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4903 if (rsm->r_flags & BBR_ACKED) {
4907 if (rsm->r_in_tmap == 0) {
4908 /* We must re-add it back to the tlist */
4910 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4912 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4916 old_flags = rsm->r_flags;
4917 rsm->r_flags |= BBR_RXT_CLEARED;
4918 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4919 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4921 if ((tp->t_state < TCPS_ESTABLISHED) &&
4922 (rsm->r_start == tp->snd_una)) {
4924 * Special case for TCP FO. Where
4925 * we sent more data beyond the snd_max.
4926 * We don't mark that as lost and stop here.
4930 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4931 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4932 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4934 if (bbr_marks_rxt_sack_passed) {
4936 * With this option, we will rack out
4937 * in 1ms increments the rest of the packets.
4939 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4940 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4943 * With this option we only mark them lost
4944 * and remove all sack'd markings. We will run
4945 * another RXT or a TLP. This will cause
4946 * us to eventually send more based on what
4949 rsm->r_flags |= BBR_MARKED_LOST;
4950 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4951 rsm->r_flags &= ~BBR_SACK_PASSED;
4956 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4957 /* Clear the count (we just un-acked them) */
4958 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
4959 bbr->rc_tlp_new_data = 0;
4960 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4961 /* zap the behindness on a rxt */
4962 bbr->r_ctl.rc_hptsi_agg_delay = 0;
4963 bbr->r_agg_early_set = 0;
4964 bbr->r_ctl.rc_agg_early = 0;
4965 bbr->rc_tlp_rtx_out = 0;
4966 bbr->r_ctl.rc_sacked = 0;
4967 bbr->r_ctl.rc_sacklast = NULL;
4968 bbr->r_timer_override = 1;
4969 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4973 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
4974 * we will setup to retransmit the lowest seq number outstanding.
4977 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4983 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
4984 if (bbr->rc_all_timers_stopped) {
4987 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
4988 (tp->snd_una == tp->snd_max)) {
4989 /* Nothing outstanding .. nothing to do */
4993 * Retransmission timer went off. Message has not been acked within
4994 * retransmit interval. Back off to a longer retransmit interval
4995 * and retransmit one segment.
4997 if (ctf_progress_timeout_check(tp, true)) {
4999 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
5000 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
5004 if ((bbr->r_ctl.rc_resend == NULL) ||
5005 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
5007 * If the rwnd collapsed on
5008 * the one we are retransmitting
5009 * it does not count against the
5014 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5015 tp->t_rxtshift = TCP_MAXRXTSHIFT;
5016 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5018 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5019 tcp_set_inp_to_drop(bbr->rc_inp,
5020 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5023 if (tp->t_state == TCPS_SYN_SENT) {
5025 * If the SYN was retransmitted, indicate CWND to be limited
5026 * to 1 segment in cc_conn_init().
5029 } else if (tp->t_rxtshift == 1) {
5031 * first retransmit; record ssthresh and cwnd so they can be
5032 * recovered if this turns out to be a "bad" retransmit. A
5033 * retransmit is considered "bad" if an ACK for this segment
5034 * is received within RTT/2 interval; the assumption here is
5035 * that the ACK was already in flight. See "On Estimating
5036 * End-to-End Network Path Properties" by Allman and Paxson
5039 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5040 if (!IN_RECOVERY(tp->t_flags)) {
5041 tp->snd_cwnd_prev = tp->snd_cwnd;
5042 tp->snd_ssthresh_prev = tp->snd_ssthresh;
5043 tp->snd_recover_prev = tp->snd_recover;
5044 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5045 tp->t_flags |= TF_PREVVALID;
5047 tp->t_flags &= ~TF_PREVVALID;
5049 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5051 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5052 tp->t_flags &= ~TF_PREVVALID;
5054 KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5055 if ((tp->t_state == TCPS_SYN_SENT) ||
5056 (tp->t_state == TCPS_SYN_RECEIVED))
5057 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5059 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5060 TCPT_RANGESET(tp->t_rxtcur, rexmt,
5061 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5062 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5064 * We enter the path for PLMTUD if connection is established or, if
5065 * connection is FIN_WAIT_1 status, reason for the last is that if
5066 * amount of data we send is very small, we could send it in couple
5067 * of packets and process straight to FIN. In that case we won't
5068 * catch ESTABLISHED state.
5071 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5075 if (((V_tcp_pmtud_blackhole_detect == 1) ||
5076 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5077 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5078 ((tp->t_state == TCPS_ESTABLISHED) ||
5079 (tp->t_state == TCPS_FIN_WAIT_1))) {
5081 * Idea here is that at each stage of mtu probe (usually,
5082 * 1448 -> 1188 -> 524) should be given 2 chances to recover
5083 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5084 * should take care of that.
5086 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5087 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5088 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5089 tp->t_rxtshift % 2 == 0)) {
5091 * Enter Path MTU Black-hole Detection mechanism: -
5092 * Disable Path MTU Discovery (IP "DF" bit). -
5093 * Reduce MTU to lower value than what we negotiated
5096 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5098 * Record that we may have found a black
5101 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5102 /* Keep track of previous MSS. */
5103 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5106 * Reduce the MSS to blackhole value or to the
5107 * default in an attempt to retransmit.
5110 isipv6 = bbr->r_is_v6;
5112 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5113 /* Use the sysctl tuneable blackhole MSS. */
5114 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5115 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5116 } else if (isipv6) {
5117 /* Use the default MSS. */
5118 tp->t_maxseg = V_tcp_v6mssdflt;
5120 * Disable Path MTU Discovery when we switch
5123 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5124 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5127 #if defined(INET6) && defined(INET)
5131 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5132 /* Use the sysctl tuneable blackhole MSS. */
5133 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5134 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5136 /* Use the default MSS. */
5137 tp->t_maxseg = V_tcp_mssdflt;
5139 * Disable Path MTU Discovery when we switch
5142 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5143 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5148 * If further retransmissions are still unsuccessful
5149 * with a lowered MTU, maybe this isn't a blackhole
5150 * and we restore the previous MSS and blackhole
5151 * detection flags. The limit '6' is determined by
5152 * giving each probe stage (1448, 1188, 524) 2
5153 * chances to recover.
5155 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5156 (tp->t_rxtshift >= 6)) {
5157 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5158 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5159 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5160 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5165 * Disable RFC1323 and SACK if we haven't got any response to our
5166 * third SYN to work-around some broken terminal servers (most of
5167 * which have hopefully been retired) that have bad VJ header
5168 * compression code which trashes TCP segments containing
5169 * unknown-to-them TCP options.
5171 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5172 (tp->t_rxtshift == 3))
5173 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5175 * If we backed off this far, our srtt estimate is probably bogus.
5176 * Clobber it so we'll take the next rtt measurement as our srtt;
5177 * move the current srtt into rttvar to keep the current retransmit
5180 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5183 in6_losing(tp->t_inpcb);
5186 in_losing(tp->t_inpcb);
5187 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5190 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5191 tp->snd_recover = tp->snd_max;
5192 tp->t_flags |= TF_ACKNOW;
5199 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5202 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5207 if (tp->t_state == TCPS_LISTEN) {
5208 /* no timers on listen sockets */
5209 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5213 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5216 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5218 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5221 if (hpts_calling == 0) {
5223 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5227 * Ok our timer went off early and we are not paced false
5228 * alarm, go back to sleep.
5230 left = bbr->r_ctl.rc_timer_exp - cts;
5232 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5233 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left));
5236 bbr->rc_tmr_stopped = 0;
5237 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5238 if (timers & PACE_TMR_DELACK) {
5239 ret = bbr_timeout_delack(tp, bbr, cts);
5240 } else if (timers & PACE_TMR_PERSIT) {
5241 ret = bbr_timeout_persist(tp, bbr, cts);
5242 } else if (timers & PACE_TMR_RACK) {
5243 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5244 ret = bbr_timeout_rack(tp, bbr, cts);
5245 } else if (timers & PACE_TMR_TLP) {
5246 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5247 ret = bbr_timeout_tlp(tp, bbr, cts);
5248 } else if (timers & PACE_TMR_RXT) {
5249 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5250 ret = bbr_timeout_rxt(tp, bbr, cts);
5251 } else if (timers & PACE_TMR_KEEP) {
5252 ret = bbr_timeout_keepalive(tp, bbr, cts);
5254 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5259 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5261 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5262 uint8_t hpts_removed = 0;
5264 if (bbr->rc_inp->inp_in_hpts &&
5265 (bbr->rc_timer_first == 1)) {
5267 * If we are canceling timer's when we have the
5268 * timer ahead of the output being paced. We also
5269 * must remove ourselves from the hpts.
5272 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
5273 if (bbr->r_ctl.rc_last_delay_val) {
5274 /* Update the last hptsi delay too */
5275 uint32_t time_since_send;
5277 if (TSTMP_GT(cts, bbr->rc_pacer_started))
5278 time_since_send = cts - bbr->rc_pacer_started;
5280 time_since_send = 0;
5281 if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5282 /* Cut down our slot time */
5283 bbr->r_ctl.rc_last_delay_val -= time_since_send;
5285 bbr->r_ctl.rc_last_delay_val = 0;
5287 bbr->rc_pacer_started = cts;
5290 bbr->rc_timer_first = 0;
5291 bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5292 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5293 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5298 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5300 struct tcp_bbr *bbr;
5302 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5303 bbr->rc_all_timers_stopped = 1;
5308 * stop all timers always returning 0.
5311 bbr_stopall(struct tcpcb *tp)
5317 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5323 * return true if a bbr timer (rack or tlp) is active.
5326 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type)
5332 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5334 struct bbr_sendmap *rsm;
5336 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5337 if ((rsm == NULL) || (u_rsm == rsm))
5339 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5343 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5344 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5350 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5351 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5352 rsm->r_flags |= BBR_OVERMAX;
5354 if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5355 /* Take off the collapsed flag at rxt */
5356 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5358 if (rsm->r_flags & BBR_MARKED_LOST) {
5359 /* We have retransmitted, its no longer lost */
5360 rsm->r_flags &= ~BBR_MARKED_LOST;
5361 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5363 if (rsm->r_flags & BBR_RXT_CLEARED) {
5365 * We hit a RXT timer on it and
5366 * we cleared the "acked" flag.
5367 * We now have it going back into
5368 * flight, we can remove the cleared
5369 * flag and possibly do accounting on
5372 rsm->r_flags &= ~BBR_RXT_CLEARED;
5374 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5375 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5376 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5378 idx = rsm->r_rtr_cnt - 1;
5379 rsm->r_tim_lastsent[idx] = cts;
5380 rsm->r_pacing_delay = pacing_time;
5381 rsm->r_delivered = bbr->r_ctl.rc_delivered;
5382 rsm->r_ts_valid = bbr->rc_ts_valid;
5383 if (bbr->rc_ts_valid)
5384 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5385 if (bbr->r_ctl.r_app_limited_until)
5386 rsm->r_app_limited = 1;
5388 rsm->r_app_limited = 0;
5389 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5390 rsm->r_bbr_state = bbr_state_val(bbr);
5392 rsm->r_bbr_state = 8;
5393 if (rsm->r_flags & BBR_ACKED) {
5394 /* Problably MTU discovery messing with us */
5397 old_flags = rsm->r_flags;
5398 rsm->r_flags &= ~BBR_ACKED;
5399 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5400 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5401 if (bbr->r_ctl.rc_sacked == 0)
5402 bbr->r_ctl.rc_sacklast = NULL;
5404 if (rsm->r_in_tmap) {
5405 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5407 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5409 if (rsm->r_flags & BBR_SACK_PASSED) {
5410 /* We have retransmitted due to the SACK pass */
5411 rsm->r_flags &= ~BBR_SACK_PASSED;
5412 rsm->r_flags |= BBR_WAS_SACKPASS;
5414 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5415 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5416 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5417 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5418 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5420 rsm->r_is_drain = 0;
5421 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5422 rsm->r_is_drain = 1;
5425 rsm->r_is_drain = 0;
5428 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5432 * Returns 0, or the sequence where we stopped
5433 * updating. We also update the lenp to be the amount
5438 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5439 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5442 * We (re-)transmitted starting at rsm->r_start for some length
5443 * (possibly less than r_end.
5445 struct bbr_sendmap *nrsm;
5450 c_end = rsm->r_start + len;
5451 if (SEQ_GEQ(c_end, rsm->r_end)) {
5453 * We retransmitted the whole piece or more than the whole
5454 * slopping into the next rsm.
5456 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5457 if (c_end == rsm->r_end) {
5463 /* Hangs over the end return whats left */
5464 act_len = rsm->r_end - rsm->r_start;
5465 *lenp = (len - act_len);
5466 return (rsm->r_end);
5468 /* We don't get out of this block. */
5471 * Here we retransmitted less than the whole thing which means we
5472 * have to split this into what was transmitted and what was not.
5474 nrsm = bbr_alloc_full_limit(bbr);
5480 * So here we are going to take the original rsm and make it what we
5481 * retransmitted. nrsm will be the tail portion we did not
5482 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5483 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5484 * 1, 6 and the new piece will be 6, 11.
5486 bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5487 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5489 if (rsm->r_in_tmap) {
5490 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5491 nrsm->r_in_tmap = 1;
5493 rsm->r_flags &= (~BBR_HAS_FIN);
5494 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5500 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5504 bw = bbr_get_bw(bbr);
5505 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5506 bw /= (uint64_t)BBR_UNIT;
5511 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5512 uint64_t act_rate, uint64_t rate_wanted)
5515 * We could not get a full gains worth
5518 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5519 /* we can't even get the real rate */
5523 bbr->gain_is_limited = 0;
5524 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5526 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5528 /* We can use a lower gain */
5530 bbr->gain_is_limited = 1;
5535 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5537 const struct tcp_hwrate_limit_table *nrte;
5538 int error, rate = -1;
5540 if (bbr->r_ctl.crte == NULL)
5542 if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5543 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5544 /* Lost our routes? */
5545 /* Clear the way for a re-attempt */
5546 bbr->bbr_attempt_hdwr_pace = 0;
5548 bbr->gain_is_limited = 0;
5550 bbr->bbr_hdrw_pacing = 0;
5551 counter_u64_add(bbr_flows_whdwr_pacing, -1);
5552 counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5553 tcp_bbr_tso_size_check(bbr, cts);
5556 rate = bbr_get_hardware_rate(bbr);
5557 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5559 bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5561 (RS_PACING_GEQ|RS_PACING_SUB_OK),
5566 if (nrte != bbr->r_ctl.crte) {
5567 bbr->r_ctl.crte = nrte;
5569 BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5570 if (bbr->r_ctl.crte->rate < rate) {
5571 /* We have a problem */
5572 bbr_setup_less_of_rate(bbr, cts,
5573 bbr->r_ctl.crte->rate, rate);
5576 bbr->gain_is_limited = 0;
5580 /* A failure should release the tag */
5581 BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5582 bbr->gain_is_limited = 0;
5584 bbr->bbr_hdrw_pacing = 0;
5586 bbr_type_log_hdwr_pacing(bbr,
5587 bbr->r_ctl.crte->ptbl->rs_ifp,
5589 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5597 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5600 * If we have hardware pacing support
5601 * we need to factor that in for our
5604 const struct tcp_hwrate_limit_table *rlp;
5605 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5607 if ((bbr->bbr_hdrw_pacing == 0) ||
5608 (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5609 (bbr->r_ctl.crte == NULL))
5611 if (bbr->hw_pacing_set == 0) {
5612 /* Not yet by the hdwr pacing count delay */
5615 if (bbr_hdwr_pace_adjust == 0) {
5619 rlp = bbr->r_ctl.crte;
5620 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5621 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5623 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5625 * So lets first get the
5626 * time we will take between
5627 * TSO sized sends currently without
5630 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5631 bbr->r_ctl.rc_pace_max_segs, cts, 1);
5632 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5633 hdwr_delay *= rlp->time_between;
5634 if (cur_delay > hdwr_delay)
5635 delta = cur_delay - hdwr_delay;
5638 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5639 (bbr->r_ctl.rc_pace_max_segs / maxseg),
5642 (delta < (max(rlp->time_between,
5643 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5645 * Now lets divide by the pacing
5646 * time between each segment the
5647 * hardware sends rounding up and
5648 * derive a bytes from that. We multiply
5649 * that by bbr_hdwr_pace_adjust to get
5650 * more bang for our buck.
5652 * The goal is to have the software pacer
5653 * waiting no more than an additional
5654 * pacing delay if we can (without the
5655 * compensation i.e. x bbr_hdwr_pace_adjust).
5657 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5658 (bbr->r_ctl.rc_pace_max_segs/maxseg));
5659 seg_sz *= bbr_hdwr_pace_adjust;
5660 if (bbr_hdwr_pace_floor &&
5661 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5662 /* Currently hardware paces
5663 * out rs_min_seg segments at a time.
5664 * We need to make sure we always send at least
5665 * a full burst of bbr_hdwr_pace_floor down.
5667 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5670 } else if (delta == 0) {
5672 * The highest pacing rate is
5673 * above our b/w gained. This means
5674 * we probably are going quite fast at
5675 * the hardware highest rate. Lets just multiply
5676 * the calculated TSO size by the
5677 * multiplier factor (its probably
5678 * 4 segments in the default config for
5681 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5682 if (bbr_hdwr_pace_floor &&
5683 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5684 /* Currently hardware paces
5685 * out rs_min_seg segments at a time.
5686 * We need to make sure we always send at least
5687 * a full burst of bbr_hdwr_pace_floor down.
5689 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5693 * The pacing time difference is so
5694 * big that the hardware will
5695 * pace out more rapidly then we
5696 * really want and then we
5697 * will have a long delay. Lets just keep
5698 * the same TSO size so its as if
5699 * we were not using hdwr pacing (we
5700 * just gain a bit of spacing from the
5701 * hardware if seg_sz > 1).
5703 seg_sz = bbr->r_ctl.rc_pace_max_segs;
5705 if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5708 new_tso = bbr->r_ctl.rc_pace_max_segs;
5709 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5710 new_tso = PACE_MAX_IP_BYTES - maxseg;
5712 if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5713 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5714 bbr->r_ctl.rc_pace_max_segs = new_tso;
5719 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5722 uint32_t old_tso = 0, new_tso;
5723 uint32_t maxseg, bytes;
5726 * Google/linux uses the following algorithm to determine
5727 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5729 * bytes = bw_in_bytes_per_second / 1000
5730 * bytes = min(bytes, 64k)
5731 * tso_segs = bytes / MSS
5736 * tso_segs = max(tso_segs, min_tso_segs)
5738 * * Note apply a device specific limit (we apply this in the
5740 * Note that before the initial measurement is made google bursts out
5741 * a full iwnd just like new-reno/cubic.
5743 * We do not use this algorithm. Instead we
5744 * use a two phased approach:
5746 * if ( bw <= per-tcb-cross-over)
5747 * goal_tso = calculate how much with this bw we
5748 * can send in goal-time seconds.
5749 * if (goal_tso > mss)
5750 * seg = goal_tso / mss
5754 * if (tso > per-tcb-max)
5756 * else if ( bw > 512Mbps)
5757 * tso = max-tso (64k/mss)
5759 * goal_tso = bw / per-tcb-divsor
5760 * seg = (goal_tso + mss-1)/mss
5763 * if (tso < per-tcb-floor)
5764 * tso = per-tcb-floor
5765 * if (tso > per-tcb-utter_max)
5766 * tso = per-tcb-utter_max
5768 * Note the default per-tcb-divisor is 1000 (same as google).
5769 * the goal cross over is 30Mbps however. To recreate googles
5770 * algorithm you need to set:
5772 * cross-over = 23,168,000 bps
5775 * per-tcb-divisor = 1000
5778 * This will get you "google bbr" behavior with respect to tso size.
5780 * Note we do set anything TSO size until we are past the initial
5781 * window. Before that we gnerally use either a single MSS
5782 * or we use the full IW size (so we burst a IW at a time)
5785 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5786 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5788 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5790 old_tso = bbr->r_ctl.rc_pace_max_segs;
5791 if (bbr->rc_past_init_win == 0) {
5793 * Not enough data has been acknowledged to make a
5794 * judgement. Set up the initial TSO based on if we
5795 * are sending a full IW at once or not.
5797 if (bbr->rc_use_google)
5798 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5799 else if (bbr->bbr_init_win_cheat)
5800 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5802 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5803 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5804 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5805 if (bbr->r_ctl.rc_pace_max_segs == 0) {
5806 bbr->r_ctl.rc_pace_max_segs = maxseg;
5808 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5809 bbr_adjust_for_hw_pacing(bbr, cts);
5813 * Now lets set the TSO goal based on our delivery rate in
5814 * bytes per second. Note we only do this if
5815 * we have acked at least the initial cwnd worth of data.
5817 bw = bbr_get_bw(bbr);
5818 if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5819 (bbr->rc_use_google == 0)) {
5820 /* We clamp to one MSS in recovery */
5822 } else if (bbr->rc_use_google) {
5825 /* Google considers the gain too */
5826 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5827 bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5831 if (bytes > (64 * 1024))
5833 new_tso = bytes / maxseg;
5834 if (bw < ONE_POINT_TWO_MEG)
5838 if (new_tso < min_tso_segs)
5839 new_tso = min_tso_segs;
5841 } else if (bbr->rc_no_pacing) {
5842 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5843 } else if (bw <= bbr->r_ctl.bbr_cross_over) {
5845 * Calculate the worse case b/w TSO if we are inserting no
5846 * more than a delay_target number of TSO's.
5848 uint32_t tso_len, min_tso;
5850 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5851 if (tso_len > maxseg) {
5852 new_tso = tso_len / maxseg;
5853 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5854 new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5858 * less than a full sized frame yikes.. long rtt or
5861 min_tso = bbr_minseg(bbr);
5862 if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5863 new_tso = rounddown(tso_len, min_tso);
5867 } else if (bw > FIVETWELVE_MBPS) {
5869 * This guy is so fast b/w wise that we can TSO as large as
5870 * possible of segments that the NIC will allow.
5872 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5875 * This formula is based on attempting to send a segment or
5876 * more every bbr_hptsi_per_second. The default is 1000
5877 * which means you are targeting what you can send every 1ms
5878 * based on the peers bw.
5880 * If the number drops to say 500, then you are looking more
5881 * at 2ms and you will raise how much we send in a single
5882 * TSO thus saving CPU (less bbr_output_wtime() calls). The
5883 * trade off of course is you will send more at once and
5884 * thus tend to clump up the sends into larger "bursts"
5887 bw /= bbr->r_ctl.bbr_hptsi_per_second;
5888 new_tso = roundup(bw, (uint64_t)maxseg);
5890 * Gate the floor to match what our lower than 48Mbps
5891 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5892 * becomes the floor for this calculation.
5894 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5895 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5897 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5898 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5899 if (new_tso > PACE_MAX_IP_BYTES)
5900 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5901 /* Enforce an utter maximum. */
5902 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5903 new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5905 if (old_tso != new_tso) {
5906 /* Only log changes */
5907 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5908 bbr->r_ctl.rc_pace_max_segs = new_tso;
5910 /* We have hardware pacing! */
5911 bbr_adjust_for_hw_pacing(bbr, cts);
5915 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
5916 uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts,
5917 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
5921 struct bbr_sendmap *rsm, *nrsm;
5922 register uint32_t snd_max, snd_una;
5923 uint32_t pacing_time;
5925 * Add to the RACK log of packets in flight or retransmitted. If
5926 * there is a TS option we will use the TS echoed, if not we will
5929 * Retransmissions will increment the count and move the ts to its
5930 * proper place. Note that if options do not include TS's then we
5931 * won't be able to effectively use the ACK for an RTT on a retran.
5933 * Notes about r_start and r_end. Lets consider a send starting at
5934 * sequence 1 for 10 bytes. In such an example the r_start would be
5935 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
5936 * This means that r_end is actually the first sequence for the next
5940 INP_WLOCK_ASSERT(tp->t_inpcb);
5943 * We don't log errors -- we could but snd_max does not
5944 * advance in this case either.
5948 if (th_flags & TH_RST) {
5950 * We don't log resets and we return immediately from
5956 snd_una = tp->snd_una;
5957 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
5959 * The call to bbr_log_output is made before bumping
5960 * snd_max. This means we can record one extra byte on a SYN
5961 * or FIN if seq_out is adding more on and a FIN is present
5962 * (and we are not resending).
5964 if ((th_flags & TH_SYN) && (tp->iss == seq_out))
5966 if (th_flags & TH_FIN)
5969 if (SEQ_LEQ((seq_out + len), snd_una)) {
5970 /* Are sending an old segment to induce an ack (keep-alive)? */
5973 if (SEQ_LT(seq_out, snd_una)) {
5974 /* huh? should we panic? */
5977 end = seq_out + len;
5979 len = end - seq_out;
5981 snd_max = tp->snd_max;
5983 /* We don't log zero window probes */
5986 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
5987 /* First question is it a retransmission? */
5988 if (seq_out == snd_max) {
5990 rsm = bbr_alloc(bbr);
5995 if (th_flags & TH_SYN)
5996 rsm->r_flags |= BBR_HAS_SYN;
5997 if (th_flags & TH_FIN)
5998 rsm->r_flags |= BBR_HAS_FIN;
5999 rsm->r_tim_lastsent[0] = cts;
6001 rsm->r_rtr_bytes = 0;
6002 rsm->r_start = seq_out;
6003 rsm->r_end = rsm->r_start + len;
6005 rsm->r_delivered = bbr->r_ctl.rc_delivered;
6006 rsm->r_pacing_delay = pacing_time;
6007 rsm->r_ts_valid = bbr->rc_ts_valid;
6008 if (bbr->rc_ts_valid)
6009 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
6010 rsm->r_del_time = bbr->r_ctl.rc_del_time;
6011 if (bbr->r_ctl.r_app_limited_until)
6012 rsm->r_app_limited = 1;
6014 rsm->r_app_limited = 0;
6015 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
6016 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
6017 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
6019 * Here we must also add in this rsm since snd_max
6020 * is updated after we return from a new send.
6022 rsm->r_flight_at_send += len;
6023 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
6024 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
6026 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
6027 rsm->r_bbr_state = bbr_state_val(bbr);
6029 rsm->r_bbr_state = 8;
6030 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
6032 rsm->r_is_drain = 0;
6033 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
6034 rsm->r_is_drain = 1;
6037 rsm->r_is_drain = 0;
6043 * If we reach here its a retransmission and we need to find it.
6046 if (hintrsm && (hintrsm->r_start == seq_out)) {
6049 } else if (bbr->r_ctl.rc_next) {
6050 /* We have a hint from a previous run */
6051 rsm = bbr->r_ctl.rc_next;
6053 /* No hints sorry */
6056 if ((rsm) && (rsm->r_start == seq_out)) {
6058 * We used rc_next or hintrsm to retransmit, hopefully the
6061 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6068 /* Ok it was not the last pointer go through it the hard way. */
6069 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6070 if (rsm->r_start == seq_out) {
6071 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6072 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6079 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6080 /* Transmitted within this piece */
6082 * Ok we must split off the front and then let the
6083 * update do the rest
6085 nrsm = bbr_alloc_full_limit(bbr);
6087 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6091 * copy rsm to nrsm and then trim the front of rsm
6092 * to not include this part.
6094 bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6095 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6096 if (rsm->r_in_tmap) {
6097 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6098 nrsm->r_in_tmap = 1;
6100 rsm->r_flags &= (~BBR_HAS_FIN);
6101 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6108 * Hmm not found in map did they retransmit both old and on into the
6111 if (seq_out == tp->snd_max) {
6113 } else if (SEQ_LT(seq_out, tp->snd_max)) {
6114 #ifdef BBR_INVARIANTS
6115 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6116 seq_out, len, tp->snd_una, tp->snd_max);
6117 printf("Starting Dump of all rack entries\n");
6118 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6119 printf("rsm:%p start:%u end:%u\n",
6120 rsm, rsm->r_start, rsm->r_end);
6122 printf("Dump complete\n");
6123 panic("seq_out not found rack:%p tp:%p",
6127 #ifdef BBR_INVARIANTS
6129 * Hmm beyond sndmax? (only if we are using the new rtt-pack
6132 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6133 seq_out, len, tp->snd_max, tp);
6139 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6142 * Collapse timeout back the cum-ack moved.
6145 tp->t_softerror = 0;
6149 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6152 bbr->r_ctl.cur_rtt = rtt_usecs;
6153 bbr->r_ctl.ts_in = tsin;
6155 bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6159 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6162 * We have in our bbr control:
6163 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6164 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6165 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6166 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6168 * Now we can calculate the time between the sends by doing:
6170 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6172 * And the peer's time between receiving them by doing:
6174 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6176 * We want to figure out if the timestamp values are in msec, 10msec or usec.
6177 * We also may find that we can't use the timestamps if say we see
6178 * that the peer_delta indicates that though we may have taken 10ms to
6179 * pace out the data, it only saw 1ms between the two packets. This would
6180 * indicate that somewhere on the path is a batching entity that is giving
6181 * out time-slices of the actual b/w. This would mean we could not use
6182 * reliably the peers timestamps.
6184 * We expect delta > peer_delta initially. Until we figure out the
6185 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6186 * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6187 * then its 10ms vs our usec. If the peer is running a usec clock we would
6188 * put a 1 there. If the value is faster then ours, we will disable the
6189 * use of timestamps (though we could revist this later if we find it to be not
6190 * just an isolated one or two flows)).
6192 * To detect the batching middle boxes we will come up with our compensation and
6193 * if with it in place, we find the peer is drastically off (by some margin) in
6194 * the smaller direction, then we will assume the worst case and disable use of timestamps.
6197 uint64_t delta, peer_delta, delta_up;
6199 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6200 if (delta < bbr_min_usec_delta) {
6202 * Have not seen a min amount of time
6203 * between our send times so we can
6204 * make a determination of the timestamp
6209 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6210 if (peer_delta < bbr_min_peer_delta) {
6212 * We may have enough in the form of
6213 * our delta but the peers number
6214 * has not changed that much. It could
6215 * be its clock ratio is such that
6216 * we need more data (10ms tick) or
6217 * there may be other compression scenarios
6218 * going on. In any event we need the
6219 * spread to be larger.
6223 /* Ok lets first see which way our delta is going */
6224 if (peer_delta > delta) {
6225 /* Very unlikely, the peer without
6226 * compensation shows that it saw
6227 * the two sends arrive further apart
6228 * then we saw then in micro-seconds.
6230 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6231 /* well it looks like the peer is a micro-second clock. */
6232 bbr->rc_ts_clock_set = 1;
6233 bbr->r_ctl.bbr_peer_tsratio = 1;
6235 bbr->rc_ts_cant_be_used = 1;
6236 bbr->rc_ts_clock_set = 1;
6240 /* Ok we know that the peer_delta is smaller than our send distance */
6241 bbr->rc_ts_clock_set = 1;
6242 /* First question is it within the percentage that they are using usec time? */
6243 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6244 if ((peer_delta + delta_up) >= delta) {
6245 /* Its a usec clock */
6246 bbr->r_ctl.bbr_peer_tsratio = 1;
6247 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6250 /* Ok if not usec, what about 10usec (though unlikely)? */
6251 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6252 if (((peer_delta * 10) + delta_up) >= delta) {
6253 bbr->r_ctl.bbr_peer_tsratio = 10;
6254 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6257 /* And what about 100usec (though again unlikely)? */
6258 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6259 if (((peer_delta * 100) + delta_up) >= delta) {
6260 bbr->r_ctl.bbr_peer_tsratio = 100;
6261 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6264 /* And how about 1 msec (the most likely one)? */
6265 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6266 if (((peer_delta * 1000) + delta_up) >= delta) {
6267 bbr->r_ctl.bbr_peer_tsratio = 1000;
6268 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6271 /* Ok if not msec could it be 10 msec? */
6272 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6273 if (((peer_delta * 10000) + delta_up) >= delta) {
6274 bbr->r_ctl.bbr_peer_tsratio = 10000;
6277 /* If we fall down here the clock tick so slowly we can't use it */
6278 bbr->rc_ts_cant_be_used = 1;
6279 bbr->r_ctl.bbr_peer_tsratio = 0;
6280 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6284 * Collect new round-trip time estimate
6285 * and update averages and current timeout.
6288 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6294 if (bbr->rtt_valid == 0)
6295 /* No valid sample */
6298 rtt = bbr->r_ctl.cur_rtt;
6299 tsin = bbr->r_ctl.ts_in;
6300 if (bbr->rc_prtt_set_ts) {
6302 * We are to force feed the rttProp filter due
6303 * to an entry into PROBE_RTT. This assures
6304 * that the times are sync'd between when we
6305 * go into PROBE_RTT and the filter expiration.
6307 * Google does not use a true filter, so they do
6308 * this implicitly since they only keep one value
6309 * and when they enter probe-rtt they update the
6310 * value to the newest rtt.
6314 bbr->rc_prtt_set_ts = 0;
6315 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6317 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6319 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6321 if (bbr->rc_ack_was_delayed)
6322 rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6324 if (rtt < bbr->r_ctl.rc_lowest_rtt)
6325 bbr->r_ctl.rc_lowest_rtt = rtt;
6326 bbr_log_rtt_sample(bbr, rtt, tsin);
6327 if (bbr->r_init_rtt) {
6329 * The initial rtt is not-trusted, nuke it and lets get
6330 * our first valid measurement in.
6332 bbr->r_init_rtt = 0;
6335 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6337 * So we have not yet figured out
6338 * what the peers TSTMP value is
6339 * in (most likely ms). We need a
6340 * series of cum-ack's to determine
6343 if (bbr->rc_ack_is_cumack) {
6344 if (bbr->rc_ts_data_set) {
6345 /* Lets attempt to determine the timestamp granularity. */
6346 bbr_make_timestamp_determination(bbr);
6348 bbr->rc_ts_data_set = 1;
6349 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6350 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6354 * We have to have consecutive acks
6355 * reset any "filled" state to none.
6357 bbr->rc_ts_data_set = 0;
6361 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6364 if (tp->t_srtt != 0) {
6366 * srtt is stored as fixed point with 5 bits after the
6367 * binary point (i.e., scaled by 8). The following magic is
6368 * equivalent to the smoothing algorithm in rfc793 with an
6369 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6370 * Adjust rtt to origin 0.
6373 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6374 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6376 tp->t_srtt += delta;
6377 if (tp->t_srtt <= 0)
6381 * We accumulate a smoothed rtt variance (actually, a
6382 * smoothed mean difference), then set the retransmit timer
6383 * to smoothed rtt + 4 times the smoothed variance. rttvar
6384 * is stored as fixed point with 4 bits after the binary
6385 * point (scaled by 16). The following is equivalent to
6386 * rfc793 smoothing with an alpha of .75 (rttvar =
6387 * rttvar*3/4 + |delta| / 4). This replaces rfc793's
6392 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6393 tp->t_rttvar += delta;
6394 if (tp->t_rttvar <= 0)
6396 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6397 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6400 * No rtt measurement yet - use the unsmoothed rtt. Set the
6401 * variance to half the rtt (so our first retransmit happens
6404 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6405 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6406 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6408 KMOD_TCPSTAT_INC(tcps_rttupdated);
6411 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6414 * the retransmit should happen at rtt + 4 * rttvar. Because of the
6415 * way we do the smoothing, srtt and rttvar will each average +1/2
6416 * tick of bias. When we compute the retransmit timer, we want 1/2
6417 * tick of rounding and 1 extra tick because of +-1/2 tick
6418 * uncertainty in the firing of the timer. The bias will give us
6419 * exactly the 1.5 tick we need. But, because the bias is
6420 * statistical, we have to test that we don't drop below the minimum
6421 * feasible timer (which is 2 ticks).
6423 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6424 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6425 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6428 * We received an ack for a packet that wasn't retransmitted; it is
6429 * probably safe to discard any error indications we've received
6430 * recently. This isn't quite right, but close enough for now (a
6431 * route might have failed after we sent a segment, and the return
6432 * path might not be symmetrical).
6434 tp->t_softerror = 0;
6435 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6436 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6437 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6441 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6443 bbr->r_ctl.rc_rtt_shrinks = cts;
6444 if (bbr_can_force_probertt &&
6445 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6446 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6448 * We should enter probe-rtt its been too long
6449 * since we have been there.
6451 bbr_enter_probe_rtt(bbr, cts, __LINE__);
6453 bbr_check_probe_rtt_limits(bbr, cts);
6457 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6461 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6462 /* We never apply a zero measurment */
6463 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6467 if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6468 bbr->r_ctl.r_measurement_count++;
6469 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6470 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6471 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6472 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6475 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6476 if (bbr->bbr_hdrw_pacing) {
6478 * Apply a new rate to the hardware
6481 bbr_update_hardware_pacing_rate(bbr, cts);
6483 bbr_set_state_target(bbr, __LINE__);
6484 tcp_bbr_tso_size_check(bbr, cts);
6485 if (bbr->r_recovery_bw) {
6486 bbr_setup_red_bw(bbr, cts);
6487 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6489 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6490 tcp_bbr_tso_size_check(bbr, cts);
6494 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6496 if (bbr->rc_in_persist == 0) {
6497 /* We log only when not in persist */
6498 /* Translate to a Bytes Per Second */
6499 uint64_t tim, bw, ts_diff, ts_bw;
6500 uint32_t upper, lower, delivered;
6502 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6503 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6507 * Now that we have processed the tim (skipping the sample
6508 * or possibly updating the time, go ahead and
6509 * calculate the cdr.
6511 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6512 bw = (uint64_t)delivered;
6513 bw *= (uint64_t)USECS_IN_SECOND;
6516 /* We must have a calculatable amount */
6519 upper = (bw >> 32) & 0x00000000ffffffff;
6520 lower = bw & 0x00000000ffffffff;
6522 * If we are using this b/w shove it in now so we
6523 * can see in the trace viewer if it gets over-ridden.
6525 if (rsm->r_ts_valid &&
6527 bbr->rc_ts_clock_set &&
6528 (bbr->rc_ts_cant_be_used == 0) &&
6529 bbr->rc_use_ts_limit) {
6530 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6531 ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6532 if ((delivered == 0) ||
6534 /* Can't use the ts */
6535 bbr_log_type_bbrupd(bbr, 61, cts,
6537 bbr->r_ctl.last_inbound_ts,
6538 rsm->r_del_ack_ts, 0,
6539 0, 0, 0, delivered);
6541 ts_bw = (uint64_t)delivered;
6542 ts_bw *= (uint64_t)USECS_IN_SECOND;
6544 bbr_log_type_bbrupd(bbr, 62, cts,
6546 (ts_bw & 0xffffffff), 0, 0,
6547 0, 0, ts_diff, delivered);
6548 if ((bbr->ts_can_raise) &&
6550 bbr_log_type_bbrupd(bbr, 8, cts,
6554 (bw & 0x00000000ffffffff),
6557 } else if (ts_bw && (ts_bw < bw)) {
6558 bbr_log_type_bbrupd(bbr, 7, cts,
6562 (bw & 0x00000000ffffffff),
6568 if (rsm->r_first_sent_time &&
6569 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6572 * We use what was in flight at the time of our
6573 * send and the size of this send to figure
6574 * out what we have been sending at (amount).
6575 * For the time we take from the time of
6576 * the send of the first send outstanding
6577 * until this send plus this sends pacing
6578 * time. This gives us a good calculation
6579 * as to the rate we have been sending at.
6582 sbw = (uint64_t)(rsm->r_flight_at_send);
6583 sbw *= (uint64_t)USECS_IN_SECOND;
6584 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6585 sti += rsm->r_pacing_delay;
6588 bbr_log_type_bbrupd(bbr, 6, cts,
6593 rsm->r_first_sent_time, 0, (sbw >> 32),
6598 /* Use the google algorithm for b/w measurements */
6599 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6600 if ((rsm->r_app_limited == 0) ||
6601 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6602 tcp_bbr_commit_bw(bbr, cts);
6603 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6604 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time);
6610 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6612 if (bbr->rc_in_persist == 0) {
6613 /* We log only when not in persist */
6614 /* Translate to a Bytes Per Second */
6616 uint32_t upper, lower, delivered;
6619 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6620 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6624 * Now that we have processed the tim (skipping the sample
6625 * or possibly updating the time, go ahead and
6626 * calculate the cdr.
6628 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6629 bw = (uint64_t)delivered;
6630 bw *= (uint64_t)USECS_IN_SECOND;
6632 if (tim < bbr->r_ctl.rc_lowest_rtt) {
6633 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6634 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6638 upper = (bw >> 32) & 0x00000000ffffffff;
6639 lower = bw & 0x00000000ffffffff;
6641 * If we are using this b/w shove it in now so we
6642 * can see in the trace viewer if it gets over-ridden.
6644 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6645 /* Gate by the sending rate */
6646 if (rsm->r_first_sent_time &&
6647 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6650 * We use what was in flight at the time of our
6651 * send and the size of this send to figure
6652 * out what we have been sending at (amount).
6653 * For the time we take from the time of
6654 * the send of the first send outstanding
6655 * until this send plus this sends pacing
6656 * time. This gives us a good calculation
6657 * as to the rate we have been sending at.
6660 sbw = (uint64_t)(rsm->r_flight_at_send);
6661 sbw *= (uint64_t)USECS_IN_SECOND;
6662 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6663 sti += rsm->r_pacing_delay;
6666 bbr_log_type_bbrupd(bbr, 6, cts,
6671 rsm->r_first_sent_time, 0, (sbw >> 32),
6676 (sti < bbr->r_ctl.rc_lowest_rtt)) {
6677 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6678 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6683 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6684 if ((no_apply == 0) &&
6685 ((rsm->r_app_limited == 0) ||
6686 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6687 tcp_bbr_commit_bw(bbr, cts);
6688 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6689 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time);
6695 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6696 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6698 uint64_t old_rttprop;
6700 /* Update our delivery time and amount */
6701 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6702 bbr->r_ctl.rc_del_time = cts;
6705 * 0 means its a retransmit, for now we don't use these for
6710 if ((bbr->rc_use_google == 0) &&
6711 (match != BBR_RTT_BY_EXACTMATCH) &&
6712 (match != BBR_RTT_BY_TIMESTAMP)){
6714 * We get a lot of rtt updates, lets not pay attention to
6715 * any that are not an exact match. That way we don't have
6716 * to worry about timestamps and the whole nonsense of
6717 * unsure if its a retransmission etc (if we ever had the
6718 * timestamp fixed to always have the last thing sent this
6719 * would not be a issue).
6723 if ((bbr_no_retran && bbr->rc_use_google) &&
6724 (match != BBR_RTT_BY_EXACTMATCH) &&
6725 (match != BBR_RTT_BY_TIMESTAMP)){
6727 * We only do measurements in google mode
6728 * with bbr_no_retran on for sure things.
6732 /* Only update srtt if we know by exact match */
6733 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6734 if (ack_type == BBR_CUM_ACKED)
6735 bbr->rc_ack_is_cumack = 1;
6737 bbr->rc_ack_is_cumack = 0;
6738 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6740 * Note the following code differs to the original
6741 * BBR spec. It calls for <= not <. However after a
6742 * long discussion in email with Neal, he acknowledged
6743 * that it should be < than so that we will have flows
6744 * going into probe-rtt (we were seeing cases where that
6745 * did not happen and caused ugly things to occur). We
6746 * have added this agreed upon fix to our code base.
6748 if (rtt < old_rttprop) {
6749 /* Update when we last saw a rtt drop */
6750 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6751 bbr_set_reduced_rtt(bbr, cts, __LINE__);
6753 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6754 match, rsm->r_start, rsm->r_flags);
6755 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6756 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6758 * The RTT-prop moved, reset the target (may be a
6759 * nop for some states).
6761 bbr_set_state_target(bbr, __LINE__);
6762 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6763 bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6764 __LINE__, BBR_RTTS_NEW_TARGET, 0);
6765 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6767 bbr_check_probe_rtt_limits(bbr, cts);
6769 if ((bbr->rc_use_google == 0) &&
6770 (match == BBR_RTT_BY_TIMESTAMP)) {
6772 * We don't do b/w update with
6773 * these since they are not really
6778 if (bbr->r_ctl.r_app_limited_until &&
6779 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6780 /* We are no longer app-limited */
6781 bbr->r_ctl.r_app_limited_until = 0;
6783 if (bbr->rc_use_google) {
6784 bbr_google_measurement(bbr, rsm, rtt, cts);
6786 bbr_nf_measurement(bbr, rsm, rtt, cts);
6791 * Convert a timestamp that the main stack
6792 * uses (milliseconds) into one that bbr uses
6793 * (microseconds). Return that converted timestamp.
6796 bbr_ts_convert(uint32_t cts) {
6799 sec = cts / MS_IN_USEC;
6800 msec = cts - (MS_IN_USEC * sec);
6801 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6805 * Return 0 if we did not update the RTT time, return
6809 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6810 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6813 uint32_t t, uts = 0;
6815 if ((rsm->r_flags & BBR_ACKED) ||
6816 (rsm->r_flags & BBR_WAS_RENEGED) ||
6817 (rsm->r_flags & BBR_RXT_CLEARED)) {
6821 if (rsm->r_rtt_not_allowed) {
6825 if (rsm->r_rtr_cnt == 1) {
6827 * Only one transmit. Hopefully the normal case.
6829 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6830 t = cts - rsm->r_tim_lastsent[0];
6835 bbr->r_ctl.rc_last_rtt = t;
6836 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6837 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6840 /* Convert to usecs */
6841 if ((bbr_can_use_ts_for_rtt == 1) &&
6842 (bbr->rc_use_google == 1) &&
6843 (ack_type == BBR_CUM_ACKED) &&
6844 (to->to_flags & TOF_TS) &&
6845 (to->to_tsecr != 0)) {
6846 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
6850 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6851 BBR_RTT_BY_TIMESTAMP,
6852 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
6856 uts = bbr_ts_convert(to->to_tsecr);
6857 if ((to->to_flags & TOF_TS) &&
6858 (to->to_tsecr != 0) &&
6859 (ack_type == BBR_CUM_ACKED) &&
6860 ((rsm->r_flags & BBR_OVERMAX) == 0)) {
6862 * Now which timestamp does it match? In this block the ACK
6863 * may be coming from a previous transmission.
6867 fudge = BBR_TIMER_FUDGE;
6868 for (i = 0; i < rsm->r_rtr_cnt; i++) {
6869 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
6870 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
6871 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6872 t = cts - rsm->r_tim_lastsent[i];
6877 bbr->r_ctl.rc_last_rtt = t;
6878 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
6879 rsm->r_tim_lastsent[i], ack_type, to);
6880 if ((i + 1) < rsm->r_rtr_cnt) {
6883 } else if (rsm->r_flags & BBR_TLP) {
6884 bbr->rc_tlp_rtx_out = 0;
6889 /* Fall through if we can't find a matching timestamp */
6892 * Ok its a SACK block that we retransmitted. or a windows
6893 * machine without timestamps. We can tell nothing from the
6894 * time-stamp since its not there or the time the peer last
6895 * recieved a segment that moved forward its cum-ack point.
6897 * Lets look at the last retransmit and see what we can tell
6898 * (with BBR for space we only keep 2 note we have to keep
6899 * at least 2 so the map can not be condensed more).
6901 i = rsm->r_rtr_cnt - 1;
6902 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6903 t = cts - rsm->r_tim_lastsent[i];
6906 if (t < bbr->r_ctl.rc_lowest_rtt) {
6908 * We retransmitted and the ack came back in less
6909 * than the smallest rtt we have observed in the
6910 * windowed rtt. We most likey did an improper
6911 * retransmit as outlined in 4.2 Step 3 point 2 in
6914 * Use the prior transmission to update all the
6915 * information as long as there is only one prior
6918 if ((rsm->r_flags & BBR_OVERMAX) == 0) {
6919 #ifdef BBR_INVARIANTS
6920 if (rsm->r_rtr_cnt == 1)
6921 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
6923 i = rsm->r_rtr_cnt - 2;
6924 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6925 t = cts - rsm->r_tim_lastsent[i];
6928 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
6929 rsm->r_tim_lastsent[i], ack_type, to);
6933 * Too many prior transmissions, just
6934 * updated BBR delivered
6937 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6938 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6942 * We retransmitted it and the retransmit did the
6945 if (rsm->r_flags & BBR_TLP)
6946 bbr->rc_tlp_rtx_out = 0;
6947 if ((rsm->r_flags & BBR_OVERMAX) == 0)
6948 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
6949 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
6951 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6952 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6959 * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
6962 bbr_log_sack_passed(struct tcpcb *tp,
6963 struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
6965 struct bbr_sendmap *nrsm;
6968 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
6969 bbr_head, r_tnext) {
6971 /* Skip orginal segment he is acked */
6974 if (nrsm->r_flags & BBR_ACKED) {
6975 /* Skip ack'd segments */
6978 if (nrsm->r_flags & BBR_SACK_PASSED) {
6980 * We found one that is already marked
6981 * passed, we have been here before and
6982 * so all others below this are marked.
6986 BBR_STAT_INC(bbr_sack_passed);
6987 nrsm->r_flags |= BBR_SACK_PASSED;
6988 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
6989 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
6990 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
6991 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
6992 nrsm->r_flags |= BBR_MARKED_LOST;
6994 nrsm->r_flags &= ~BBR_WAS_SACKPASS;
6999 * Returns the number of bytes that were
7000 * newly ack'd by sack blocks.
7003 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
7004 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
7007 uint32_t start, end, maxseg, changed = 0;
7008 struct bbr_sendmap *rsm, *nrsm;
7009 int32_t used_ref = 1;
7010 uint8_t went_back = 0, went_fwd = 0;
7012 maxseg = tp->t_maxseg - bbr->rc_last_options;
7013 start = sack->start;
7019 /* Do we locate the block behind where we last were? */
7020 if (rsm && SEQ_LT(start, rsm->r_start)) {
7022 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
7023 if (SEQ_GEQ(start, rsm->r_start) &&
7024 SEQ_LT(start, rsm->r_end)) {
7032 * Ok lets locate the block where this guy is fwd from rsm (if its
7035 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
7036 if (SEQ_GEQ(start, rsm->r_start) &&
7037 SEQ_LT(start, rsm->r_end)) {
7044 * This happens when we get duplicate sack blocks with the
7045 * same end. For example SACK 4: 100 SACK 3: 100 The sort
7046 * will not change there location so we would just start at
7047 * the end of the first one and get lost.
7049 if (tp->t_flags & TF_SENTFIN) {
7051 * Check to see if we have not logged the FIN that
7054 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7055 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7057 * Ok we did not get the FIN logged.
7065 #ifdef BBR_INVARIANTS
7066 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7067 tp, bbr, sack, to, prsm);
7073 BBR_STAT_INC(bbr_sack_proc_restart);
7075 goto start_at_beginning;
7077 /* Ok we have an ACK for some piece of rsm */
7078 if (rsm->r_start != start) {
7080 * Need to split this in two pieces the before and after.
7082 if (bbr_sack_mergable(rsm, start, end))
7083 nrsm = bbr_alloc_full_limit(bbr);
7085 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7087 /* We could not allocate ignore the sack */
7092 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7095 bbr_clone_rsm(bbr, nrsm, rsm, start);
7096 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7097 if (rsm->r_in_tmap) {
7098 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7099 nrsm->r_in_tmap = 1;
7101 rsm->r_flags &= (~BBR_HAS_FIN);
7104 if (SEQ_GEQ(end, rsm->r_end)) {
7106 * The end of this block is either beyond this guy or right
7109 if ((rsm->r_flags & BBR_ACKED) == 0) {
7110 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7111 changed += (rsm->r_end - rsm->r_start);
7112 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7113 bbr_log_sack_passed(tp, bbr, rsm);
7114 if (rsm->r_flags & BBR_MARKED_LOST) {
7115 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7117 /* Is Reordering occuring? */
7118 if (rsm->r_flags & BBR_SACK_PASSED) {
7119 BBR_STAT_INC(bbr_reorder_seen);
7120 bbr->r_ctl.rc_reorder_ts = cts;
7121 if (rsm->r_flags & BBR_MARKED_LOST) {
7122 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7123 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7124 /* LT sampling also needs adjustment */
7125 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7128 rsm->r_flags |= BBR_ACKED;
7129 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7130 if (rsm->r_in_tmap) {
7131 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7135 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7136 if (end == rsm->r_end) {
7137 /* This block only - done */
7140 /* There is more not coverend by this rsm move on */
7142 nrsm = TAILQ_NEXT(rsm, r_next);
7147 if (rsm->r_flags & BBR_ACKED) {
7148 /* Been here done that */
7151 /* Ok we need to split off this one at the tail */
7152 if (bbr_sack_mergable(rsm, start, end))
7153 nrsm = bbr_alloc_full_limit(bbr);
7155 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7157 /* failed XXXrrs what can we do but loose the sack info? */
7162 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7166 bbr_clone_rsm(bbr, nrsm, rsm, end);
7167 /* The sack block does not cover this guy fully */
7168 rsm->r_flags &= (~BBR_HAS_FIN);
7169 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7170 if (rsm->r_in_tmap) {
7171 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7172 nrsm->r_in_tmap = 1;
7175 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7176 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7177 changed += (rsm->r_end - rsm->r_start);
7178 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7179 bbr_log_sack_passed(tp, bbr, rsm);
7180 /* Is Reordering occuring? */
7181 if (rsm->r_flags & BBR_MARKED_LOST) {
7182 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7184 if (rsm->r_flags & BBR_SACK_PASSED) {
7185 BBR_STAT_INC(bbr_reorder_seen);
7186 bbr->r_ctl.rc_reorder_ts = cts;
7187 if (rsm->r_flags & BBR_MARKED_LOST) {
7188 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7189 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7190 /* LT sampling also needs adjustment */
7191 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7194 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7195 rsm->r_flags |= BBR_ACKED;
7196 if (rsm->r_in_tmap) {
7197 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7201 if (rsm && (rsm->r_flags & BBR_ACKED)) {
7203 * Now can we merge this newly acked
7204 * block with either the previous or
7207 nrsm = TAILQ_NEXT(rsm, r_next);
7209 (nrsm->r_flags & BBR_ACKED)) {
7210 /* yep this and next can be merged */
7211 rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7213 /* Now what about the previous? */
7214 nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7216 (nrsm->r_flags & BBR_ACKED)) {
7217 /* yep the previous and this can be merged */
7218 rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7221 if (used_ref == 0) {
7222 BBR_STAT_INC(bbr_sack_proc_all);
7224 BBR_STAT_INC(bbr_sack_proc_short);
7226 if (went_fwd && went_back) {
7227 BBR_STAT_INC(bbr_sack_search_both);
7228 } else if (went_fwd) {
7229 BBR_STAT_INC(bbr_sack_search_fwd);
7230 } else if (went_back) {
7231 BBR_STAT_INC(bbr_sack_search_back);
7233 /* Save off where the next seq is */
7235 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7237 bbr->r_ctl.rc_sacklast = NULL;
7243 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7245 struct bbr_sendmap *tmap;
7247 BBR_STAT_INC(bbr_reneges_seen);
7249 while (rsm && (rsm->r_flags & BBR_ACKED)) {
7250 /* Its no longer sacked, mark it so */
7252 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7253 #ifdef BBR_INVARIANTS
7254 if (rsm->r_in_tmap) {
7255 panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7256 bbr, rsm, rsm->r_flags);
7259 oflags = rsm->r_flags;
7260 if (rsm->r_flags & BBR_MARKED_LOST) {
7261 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7262 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7263 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7264 /* LT sampling also needs adjustment */
7265 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7267 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7268 rsm->r_flags |= BBR_WAS_RENEGED;
7269 rsm->r_flags |= BBR_RXT_CLEARED;
7270 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7271 /* Rebuild it into our tmap */
7273 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7276 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7279 tmap->r_in_tmap = 1;
7281 * XXXrrs Delivered? Should we do anything here?
7283 * Of course we don't on a rxt timeout so maybe its ok that
7288 rsm = TAILQ_NEXT(rsm, r_next);
7291 * Now lets possibly clear the sack filter so we start recognizing
7292 * sacks that cover this area.
7294 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7298 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7300 struct tcp_bbr *bbr;
7301 struct bbr_sendmap *rsm;
7304 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7305 cts = bbr->r_ctl.rc_rcvtime;
7306 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7307 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7308 if ((rsm->r_end - rsm->r_start) <= 1) {
7309 /* Log out the SYN completely */
7310 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7311 rsm->r_rtr_bytes = 0;
7312 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7313 if (rsm->r_in_tmap) {
7314 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7317 if (bbr->r_ctl.rc_next == rsm) {
7318 /* scoot along the marker */
7319 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7322 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7325 /* There is more (Fast open)? strip out SYN. */
7326 rsm->r_flags &= ~BBR_HAS_SYN;
7333 * Returns the number of bytes that were
7334 * acknowledged by SACK blocks.
7338 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7339 uint32_t *prev_acked)
7341 uint32_t changed, last_seq, entered_recovery = 0;
7342 struct tcp_bbr *bbr;
7343 struct bbr_sendmap *rsm;
7344 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7345 register uint32_t th_ack;
7346 int32_t i, j, k, new_sb, num_sack_blks = 0;
7347 uint32_t cts, acked, ack_point, sack_changed = 0;
7348 uint32_t p_maxseg, maxseg, p_acked = 0;
7350 INP_WLOCK_ASSERT(tp->t_inpcb);
7351 if (th->th_flags & TH_RST) {
7352 /* We don't log resets */
7355 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7356 cts = bbr->r_ctl.rc_rcvtime;
7358 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7360 maxseg = tp->t_maxseg - bbr->rc_last_options;
7361 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7362 th_ack = th->th_ack;
7363 if (SEQ_GT(th_ack, tp->snd_una)) {
7364 acked = th_ack - tp->snd_una;
7365 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7366 bbr->rc_tp->t_acktime = ticks;
7369 if (SEQ_LEQ(th_ack, tp->snd_una)) {
7370 /* Only sent here for sack processing */
7373 if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7374 changed = th_ack - rsm->r_start;
7375 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7377 * For the SYN incoming case we will not have called
7378 * tcp_output for the sending of the SYN, so there will be
7379 * no map. All other cases should probably be a panic.
7381 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7383 * We have a timestamp that can be used to generate
7386 uint32_t ts, now, rtt;
7388 ts = bbr_ts_convert(to->to_tsecr);
7389 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7393 bbr_log_type_bbrrttprop(bbr, rtt,
7395 BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7396 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7398 bbr->r_wanted_output = 1;
7402 } else if (rsm == NULL) {
7407 * The ACK point is advancing to th_ack, we must drop off
7408 * the packets in the rack log and calculate any eligble
7411 bbr->r_wanted_output = 1;
7414 if (tp->t_flags & TF_SENTFIN) {
7415 /* if we send a FIN we will not hav a map */
7418 #ifdef BBR_INVARIANTS
7419 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7421 th, tp->t_state, bbr,
7422 tp->snd_una, tp->snd_max, changed);
7427 if (SEQ_LT(th_ack, rsm->r_start)) {
7428 /* Huh map is missing this */
7429 #ifdef BBR_INVARIANTS
7430 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7432 th_ack, tp->t_state,
7434 panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7437 } else if (th_ack == rsm->r_start) {
7438 /* None here to ack */
7442 * Clear the dup ack counter, it will
7443 * either be freed or if there is some
7444 * remaining we need to start it at zero.
7447 /* Now do we consume the whole thing? */
7448 if (SEQ_GEQ(th_ack, rsm->r_end)) {
7449 /* Its all consumed. */
7452 if (rsm->r_flags & BBR_ACKED) {
7454 * It was acked on the scoreboard -- remove it from
7457 p_acked += (rsm->r_end - rsm->r_start);
7458 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7459 if (bbr->r_ctl.rc_sacked == 0)
7460 bbr->r_ctl.rc_sacklast = NULL;
7462 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7463 if (rsm->r_flags & BBR_MARKED_LOST) {
7464 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7466 if (rsm->r_flags & BBR_SACK_PASSED) {
7468 * There are acked segments ACKED on the
7469 * scoreboard further up. We are seeing
7472 BBR_STAT_INC(bbr_reorder_seen);
7473 bbr->r_ctl.rc_reorder_ts = cts;
7474 if (rsm->r_flags & BBR_MARKED_LOST) {
7475 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7476 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7477 /* LT sampling also needs adjustment */
7478 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7481 rsm->r_flags &= ~BBR_MARKED_LOST;
7483 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7484 rsm->r_rtr_bytes = 0;
7485 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7486 if (rsm->r_in_tmap) {
7487 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7490 if (bbr->r_ctl.rc_next == rsm) {
7491 /* scoot along the marker */
7492 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7494 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7495 /* Adjust the packet counts */
7496 left = th_ack - rsm->r_end;
7497 /* Free back to zone */
7500 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7505 if (rsm->r_flags & BBR_ACKED) {
7507 * It was acked on the scoreboard -- remove it from total
7508 * for the part being cum-acked.
7510 p_acked += (rsm->r_end - rsm->r_start);
7511 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7512 if (bbr->r_ctl.rc_sacked == 0)
7513 bbr->r_ctl.rc_sacklast = NULL;
7516 * It was acked up to th_ack point for the first time
7518 struct bbr_sendmap lrsm;
7520 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7521 lrsm.r_end = th_ack;
7522 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7524 if ((rsm->r_flags & BBR_MARKED_LOST) &&
7525 ((rsm->r_flags & BBR_ACKED) == 0)) {
7527 * It was marked lost and partly ack'd now
7528 * for the first time. We lower the rc_lost_bytes
7529 * and still leave it MARKED.
7531 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7533 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7534 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7535 rsm->r_rtr_bytes = 0;
7536 /* adjust packet count */
7537 rsm->r_start = th_ack;
7539 /* Check for reneging */
7540 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7541 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7543 * The peer has moved snd_una up to the edge of this send,
7544 * i.e. one that it had previously acked. The only way that
7545 * can be true if the peer threw away data (space issues)
7546 * that it had previously sacked (else it would have given
7547 * us snd_una up to (rsm->r_end). We need to undo the acked
7550 * Note we have to look to make sure th_ack is our
7551 * rsm->r_start in case we get an old ack where th_ack is
7554 bbr_peer_reneges(bbr, rsm, th->th_ack);
7556 if ((to->to_flags & TOF_SACK) == 0) {
7557 /* We are done nothing left to log */
7560 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7562 last_seq = rsm->r_end;
7564 last_seq = tp->snd_max;
7566 /* Sack block processing */
7567 if (SEQ_GT(th_ack, tp->snd_una))
7570 ack_point = tp->snd_una;
7571 for (i = 0; i < to->to_nsacks; i++) {
7572 bcopy((to->to_sacks + i * TCPOLEN_SACK),
7573 &sack, sizeof(sack));
7574 sack.start = ntohl(sack.start);
7575 sack.end = ntohl(sack.end);
7576 if (SEQ_GT(sack.end, sack.start) &&
7577 SEQ_GT(sack.start, ack_point) &&
7578 SEQ_LT(sack.start, tp->snd_max) &&
7579 SEQ_GT(sack.end, ack_point) &&
7580 SEQ_LEQ(sack.end, tp->snd_max)) {
7581 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7582 (SEQ_LT(sack.end, last_seq)) &&
7583 ((sack.end - sack.start) < (p_maxseg / 8))) {
7585 * Not the last piece and its smaller than
7586 * 1/8th of a p_maxseg. We ignore this.
7588 BBR_STAT_INC(bbr_runt_sacks);
7591 sack_blocks[num_sack_blks] = sack;
7593 #ifdef NETFLIX_STATS
7594 } else if (SEQ_LEQ(sack.start, th_ack) &&
7595 SEQ_LEQ(sack.end, th_ack)) {
7597 * Its a D-SACK block.
7599 tcp_record_dsack(sack.start, sack.end);
7603 if (num_sack_blks == 0)
7606 * Sort the SACK blocks so we can update the rack scoreboard with
7609 new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7610 num_sack_blks, th->th_ack);
7611 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7612 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7613 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7614 num_sack_blks = new_sb;
7615 if (num_sack_blks < 2) {
7618 /* Sort the sacks */
7619 for (i = 0; i < num_sack_blks; i++) {
7620 for (j = i + 1; j < num_sack_blks; j++) {
7621 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7622 sack = sack_blocks[i];
7623 sack_blocks[i] = sack_blocks[j];
7624 sack_blocks[j] = sack;
7629 * Now are any of the sack block ends the same (yes some
7630 * implememtations send these)?
7633 if (num_sack_blks > 1) {
7634 for (i = 0; i < num_sack_blks; i++) {
7635 for (j = i + 1; j < num_sack_blks; j++) {
7636 if (sack_blocks[i].end == sack_blocks[j].end) {
7638 * Ok these two have the same end we
7639 * want the smallest end and then
7640 * throw away the larger and start
7643 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7645 * The second block covers
7646 * more area use that
7648 sack_blocks[i].start = sack_blocks[j].start;
7651 * Now collapse out the dup-sack and
7654 for (k = (j + 1); k < num_sack_blks; k++) {
7655 sack_blocks[j].start = sack_blocks[k].start;
7656 sack_blocks[j].end = sack_blocks[k].end;
7666 rsm = bbr->r_ctl.rc_sacklast;
7667 for (i = 0; i < num_sack_blks; i++) {
7668 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7670 bbr->r_wanted_output = 1;
7672 sack_changed += acked;
7676 *prev_acked = p_acked;
7677 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7679 * Ok we have a high probability that we need to go in to
7680 * recovery since we have data sack'd
7682 struct bbr_sendmap *rsm;
7684 rsm = bbr_check_recovery_mode(tp, bbr, cts);
7686 /* Enter recovery */
7687 entered_recovery = 1;
7688 bbr->r_wanted_output = 1;
7690 * When we enter recovery we need to assure we send
7693 if (bbr->r_ctl.rc_resend == NULL) {
7694 bbr->r_ctl.rc_resend = rsm;
7698 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7700 * See if we need to rack-retransmit anything if so set it
7701 * up as the thing to resend assuming something else is not
7702 * already in that position.
7704 if (bbr->r_ctl.rc_resend == NULL) {
7705 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7709 * We return the amount that changed via sack, this is used by the
7710 * ack-received code to augment what was changed between th_ack <->
7713 return (sack_changed);
7717 bbr_strike_dupack(struct tcp_bbr *bbr)
7719 struct bbr_sendmap *rsm;
7721 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7722 if (rsm && (rsm->r_dupack < 0xff)) {
7724 if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7725 bbr->r_wanted_output = 1;
7730 * Return value of 1, we do not need to call bbr_process_data().
7731 * return value of 0, bbr_process_data can be called.
7732 * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7733 * its unlocked and probably unsafe to touch the TCB.
7736 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7737 struct tcpcb *tp, struct tcpopt *to,
7738 uint32_t tiwin, int32_t tlen,
7739 int32_t * ofia, int32_t thflags, int32_t * ret_val)
7741 int32_t ourfinisacked = 0;
7742 int32_t acked_amount;
7745 uint32_t lost, sack_changed = 0;
7747 struct tcp_bbr *bbr;
7748 uint32_t prev_acked = 0;
7750 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7751 lost = bbr->r_ctl.rc_lost;
7752 nsegs = max(1, m->m_pkthdr.lro_nsegs);
7753 if (SEQ_GT(th->th_ack, tp->snd_max)) {
7754 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7755 bbr->r_wanted_output = 1;
7758 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7759 /* Process the ack */
7760 if (bbr->rc_in_persist)
7762 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7763 bbr_strike_dupack(bbr);
7764 sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7766 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7767 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7769 * Old ack, behind the last one rcv'd or a duplicate ack
7772 if (th->th_ack == tp->snd_una) {
7773 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7774 if (bbr->r_state == TCPS_SYN_SENT) {
7776 * Special case on where we sent SYN. When
7777 * the SYN-ACK is processed in syn_sent
7778 * state it bumps the snd_una. This causes
7779 * us to hit here even though we did ack 1
7782 * Go through the nothing left case so we
7791 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7792 * something we sent.
7794 if (tp->t_flags & TF_NEEDSYN) {
7796 * T/TCP: Connection was half-synchronized, and our SYN has
7797 * been ACK'd (so connection is now fully synchronized). Go
7798 * to non-starred state, increment snd_una for ACK of SYN,
7799 * and check if we can do window scaling.
7801 tp->t_flags &= ~TF_NEEDSYN;
7803 /* Do window scaling? */
7804 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7805 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7806 tp->rcv_scale = tp->request_r_scale;
7807 /* Send window already scaled. */
7810 INP_WLOCK_ASSERT(tp->t_inpcb);
7812 acked = BYTES_THIS_ACK(tp, th);
7813 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7814 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7817 * If we just performed our first retransmit, and the ACK arrives
7818 * within our recovery window, then it was a mistake to do the
7819 * retransmit in the first place. Recover our original cwnd and
7820 * ssthresh, and proceed to transmit where we left off.
7822 if (tp->t_flags & TF_PREVVALID) {
7823 tp->t_flags &= ~TF_PREVVALID;
7824 if (tp->t_rxtshift == 1 &&
7825 (int)(ticks - tp->t_badrxtwin) < 0)
7826 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7828 SOCKBUF_LOCK(&so->so_snd);
7829 acked_amount = min(acked, (int)sbavail(&so->so_snd));
7830 tp->snd_wnd -= acked_amount;
7831 mfree = sbcut_locked(&so->so_snd, acked_amount);
7832 /* NB: sowwakeup_locked() does an implicit unlock. */
7833 sowwakeup_locked(so);
7835 if (SEQ_GT(th->th_ack, tp->snd_una)) {
7836 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7838 tp->snd_una = th->th_ack;
7839 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7840 if (IN_RECOVERY(tp->t_flags)) {
7841 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7842 (SEQ_LT(th->th_ack, tp->snd_max))) {
7843 tcp_bbr_partialack(tp);
7845 bbr_post_recovery(tp);
7848 if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
7849 tp->snd_recover = tp->snd_una;
7851 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
7852 tp->snd_nxt = tp->snd_max;
7854 if (tp->snd_una == tp->snd_max) {
7855 /* Nothing left outstanding */
7857 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
7858 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
7859 bbr->rc_tp->t_acktime = 0;
7860 if ((sbused(&so->so_snd) == 0) &&
7861 (tp->t_flags & TF_SENTFIN)) {
7864 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
7865 if (bbr->rc_in_persist == 0) {
7866 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
7868 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
7869 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
7871 * We invalidate the last ack here since we
7872 * don't want to transfer forward the time
7873 * for our sum's calculations.
7875 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7876 (sbavail(&so->so_snd) == 0) &&
7877 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
7879 * The socket was gone and the peer sent data, time
7883 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
7884 /* tcp_close will kill the inp pre-log the Reset */
7885 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
7887 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
7888 BBR_STAT_INC(bbr_dropped_af_data);
7891 /* Set need output so persist might get set */
7892 bbr->r_wanted_output = 1;
7895 *ofia = ourfinisacked;
7900 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7902 if (bbr->rc_in_persist == 0) {
7903 bbr_timer_cancel(bbr, __LINE__, cts);
7904 bbr->r_ctl.rc_last_delay_val = 0;
7906 bbr->rc_in_persist = 1;
7907 bbr->r_ctl.rc_went_idle_time = cts;
7908 /* We should be capped when rw went to 0 but just in case */
7909 bbr_log_type_pesist(bbr, cts, 0, line, 1);
7910 /* Time freezes for the state, so do the accounting now */
7911 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
7914 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
7915 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7918 idx = bbr_state_val(bbr);
7919 counter_u64_add(bbr_state_time[(idx + 5)], time_in);
7921 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
7924 bbr->r_ctl.rc_bbr_state_time = cts;
7929 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
7932 * Note that if idle time does not exceed our
7933 * threshold, we do nothing continuing the state
7934 * transitions we were last walking through.
7936 if (idle_time >= bbr_idle_restart_threshold) {
7937 if (bbr->rc_use_idle_restart) {
7938 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
7940 * Set our target using BBR_UNIT, so
7941 * we increase at a dramatic rate but
7942 * we stop when we get the pipe
7943 * full again for our current b/w estimate.
7945 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
7946 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
7947 bbr_set_state_target(bbr, __LINE__);
7948 /* Now setup our gains to ramp up */
7949 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
7950 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
7951 bbr_log_type_statechange(bbr, cts, __LINE__);
7952 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7953 bbr_substate_change(bbr, cts, __LINE__, 1);
7959 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7963 if (bbr->rc_in_persist == 0)
7965 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
7966 bbr->rc_in_persist = 0;
7967 bbr->rc_hit_state_1 = 0;
7968 bbr->r_ctl.rc_del_time = cts;
7970 * We invalidate the last ack here since we
7971 * don't want to transfer forward the time
7972 * for our sum's calculations.
7974 if (bbr->rc_inp->inp_in_hpts) {
7975 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
7976 bbr->rc_timer_first = 0;
7977 bbr->r_ctl.rc_hpts_flags = 0;
7978 bbr->r_ctl.rc_last_delay_val = 0;
7979 bbr->r_ctl.rc_hptsi_agg_delay = 0;
7980 bbr->r_agg_early_set = 0;
7981 bbr->r_ctl.rc_agg_early = 0;
7983 bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
7984 if (idle_time >= bbr_rtt_probe_time) {
7986 * This qualifies as a RTT_PROBE session since we drop the
7987 * data outstanding to nothing and waited more than
7988 * bbr_rtt_probe_time.
7990 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
7991 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
7995 * If in probeBW and we have persisted more than an RTT lets do
7998 /* Force a time based epoch */
7999 bbr_set_epoch(bbr, cts, __LINE__);
8001 * Setup the lost so we don't count anything against the guy
8002 * we have been stuck with during persists.
8004 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
8005 /* Time un-freezes for the state */
8006 bbr->r_ctl.rc_bbr_state_time = cts;
8007 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
8008 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
8010 * If we are going back to probe-bw
8011 * or probe_rtt, we may need to possibly
8012 * do a fast restart.
8014 bbr_restart_after_idle(bbr, cts, idle_time);
8019 bbr_collapsed_window(struct tcp_bbr *bbr)
8022 * Now we must walk the
8023 * send map and divide the
8024 * ones left stranded. These
8025 * guys can't cause us to abort
8026 * the connection and are really
8027 * "unsent". However if a buggy
8028 * client actually did keep some
8029 * of the data i.e. collapsed the win
8030 * and refused to ack and then opened
8031 * the win and acked that data. We would
8032 * get into an ack war, the simplier
8033 * method then of just pretending we
8034 * did not send those segments something
8037 struct bbr_sendmap *rsm, *nrsm;
8043 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8044 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8045 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8046 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8047 /* Find the first seq past or at maxseq */
8048 if (rsm->r_flags & BBR_RWND_COLLAPSED)
8049 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8050 if (SEQ_GEQ(max_seq, rsm->r_start) &&
8051 SEQ_GEQ(rsm->r_end, max_seq)) {
8056 bbr->rc_has_collapsed = 0;
8058 /* Nothing to do strange */
8064 * We don't want to split if splitting
8065 * would generate too many small segments
8066 * less we let an attacker fragment our
8067 * send_map and leave us out of memory.
8069 if ((max_seq != rsm->r_start) &&
8070 (max_seq != rsm->r_end)){
8074 res1 = max_seq - rsm->r_start;
8075 res2 = rsm->r_end - max_seq;
8076 if ((res1 >= (maxseg/8)) &&
8077 (res2 >= (maxseg/8))) {
8078 /* No small pieces here */
8080 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8081 /* We are under the limit */
8085 /* Ok do we need to split this rsm? */
8086 if (max_seq == rsm->r_start) {
8087 /* It's this guy no split required */
8089 } else if (max_seq == rsm->r_end) {
8090 /* It's the next one no split required. */
8091 nrsm = TAILQ_NEXT(rsm, r_next);
8096 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8097 /* yep we need to split it */
8098 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8100 /* failed XXXrrs what can we do mark the whole? */
8105 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8106 bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8107 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8108 if (rsm->r_in_tmap) {
8109 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8110 nrsm->r_in_tmap = 1;
8114 * Split not allowed just start here just
8120 BBR_STAT_INC(bbr_collapsed_win);
8121 /* reuse fnd as a count */
8123 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8124 nrsm->r_flags |= BBR_RWND_COLLAPSED;
8126 bbr->rc_has_collapsed = 1;
8128 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8132 bbr_un_collapse_window(struct tcp_bbr *bbr)
8134 struct bbr_sendmap *rsm;
8137 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8138 if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8139 /* Clear the flag */
8140 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8145 bbr_log_type_rwnd_collapse(bbr,
8146 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8147 bbr->rc_has_collapsed = 0;
8151 * Return value of 1, the TCB is unlocked and most
8152 * likely gone, return value of 0, the TCB is still
8156 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8157 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8158 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8161 * Update window information. Don't look at window if no ACK: TAC's
8162 * send garbage on first SYN.
8166 struct tcp_bbr *bbr;
8168 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8169 INP_WLOCK_ASSERT(tp->t_inpcb);
8170 nsegs = max(1, m->m_pkthdr.lro_nsegs);
8171 if ((thflags & TH_ACK) &&
8172 (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8173 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8174 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8175 /* keep track of pure window updates */
8177 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8178 KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8179 tp->snd_wnd = tiwin;
8180 tp->snd_wl1 = th->th_seq;
8181 tp->snd_wl2 = th->th_ack;
8182 if (tp->snd_wnd > tp->max_sndwnd)
8183 tp->max_sndwnd = tp->snd_wnd;
8184 bbr->r_wanted_output = 1;
8185 } else if (thflags & TH_ACK) {
8186 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8187 tp->snd_wnd = tiwin;
8188 tp->snd_wl1 = th->th_seq;
8189 tp->snd_wl2 = th->th_ack;
8192 if (tp->snd_wnd < ctf_outstanding(tp))
8193 /* The peer collapsed its window on us */
8194 bbr_collapsed_window(bbr);
8195 else if (bbr->rc_has_collapsed)
8196 bbr_un_collapse_window(bbr);
8197 /* Was persist timer active and now we have window space? */
8198 if ((bbr->rc_in_persist != 0) &&
8199 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8200 bbr_minseg(bbr)))) {
8202 * Make the rate persist at end of persist mode if idle long
8205 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8207 /* Make sure we output to start the timer */
8208 bbr->r_wanted_output = 1;
8210 /* Do we need to enter persist? */
8211 if ((bbr->rc_in_persist == 0) &&
8212 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8213 TCPS_HAVEESTABLISHED(tp->t_state) &&
8214 (tp->snd_max == tp->snd_una) &&
8215 sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8216 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8217 /* No send window.. we must enter persist */
8218 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8220 if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8225 * We don't support urgent data but
8226 * drag along the up just to make sure
8227 * if there is a stack switch no one
8230 tp->rcv_up = tp->rcv_nxt;
8231 INP_WLOCK_ASSERT(tp->t_inpcb);
8234 * Process the segment text, merging it into the TCP sequencing
8235 * queue, and arranging for acknowledgment of receipt if necessary.
8236 * This process logically involves adjusting tp->rcv_wnd as data is
8237 * presented to the user (this happens in tcp_usrreq.c, case
8238 * PRU_RCVD). If a FIN has already been received on this connection
8239 * then we just ignore the text.
8241 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8242 IS_FASTOPEN(tp->t_flags));
8243 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8244 TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8245 tcp_seq save_start = th->th_seq;
8246 tcp_seq save_rnxt = tp->rcv_nxt;
8247 int save_tlen = tlen;
8249 m_adj(m, drop_hdrlen); /* delayed header drop */
8251 * Insert segment which includes th into TCP reassembly
8252 * queue with control block tp. Set thflags to whether
8253 * reassembly now includes a segment with FIN. This handles
8254 * the common case inline (segment is the next to be
8255 * received on an established connection, and the queue is
8256 * empty), avoiding linkage into and removal from the queue
8257 * and repetition of various conversions. Set DELACK for
8258 * segments received in order, but ack immediately when
8259 * segments are out of order (so fast retransmit can work).
8261 if (th->th_seq == tp->rcv_nxt &&
8263 (TCPS_HAVEESTABLISHED(tp->t_state) ||
8265 #ifdef NETFLIX_SB_LIMITS
8266 u_int mcnt, appended;
8268 if (so->so_rcv.sb_shlim) {
8271 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8272 CFO_NOSLEEP, NULL) == false) {
8273 counter_u64_add(tcp_sb_shlim_fails, 1);
8280 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8281 bbr->bbr_segs_rcvd += max(1, nsegs);
8282 tp->t_flags |= TF_DELACK;
8283 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8285 bbr->r_wanted_output = 1;
8286 tp->t_flags |= TF_ACKNOW;
8288 tp->rcv_nxt += tlen;
8290 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8291 (tp->t_fbyte_in == 0)) {
8292 tp->t_fbyte_in = ticks;
8293 if (tp->t_fbyte_in == 0)
8295 if (tp->t_fbyte_out && tp->t_fbyte_in)
8296 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8298 thflags = th->th_flags & TH_FIN;
8299 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8300 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8301 SOCKBUF_LOCK(&so->so_rcv);
8302 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8305 #ifdef NETFLIX_SB_LIMITS
8308 sbappendstream_locked(&so->so_rcv, m, 0);
8309 /* NB: sorwakeup_locked() does an implicit unlock. */
8310 sorwakeup_locked(so);
8311 #ifdef NETFLIX_SB_LIMITS
8312 if (so->so_rcv.sb_shlim && appended != mcnt)
8313 counter_fo_release(so->so_rcv.sb_shlim,
8319 * XXX: Due to the header drop above "th" is
8320 * theoretically invalid by now. Fortunately
8321 * m_adj() doesn't actually frees any mbufs when
8322 * trimming from the head.
8324 tcp_seq temp = save_start;
8326 thflags = tcp_reass(tp, th, &temp, &tlen, m);
8327 tp->t_flags |= TF_ACKNOW;
8328 if (tp->t_flags & TF_WAKESOR) {
8329 tp->t_flags &= ~TF_WAKESOR;
8330 /* NB: sorwakeup_locked() does an implicit unlock. */
8331 sorwakeup_locked(so);
8334 if ((tp->t_flags & TF_SACK_PERMIT) &&
8336 TCPS_HAVEESTABLISHED(tp->t_state)) {
8337 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8339 * DSACK actually handled in the fastpath
8342 tcp_update_sack_list(tp, save_start,
8343 save_start + save_tlen);
8344 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8345 if ((tp->rcv_numsacks >= 1) &&
8346 (tp->sackblks[0].end == save_start)) {
8348 * Partial overlap, recorded at todrop
8351 tcp_update_sack_list(tp,
8352 tp->sackblks[0].start,
8353 tp->sackblks[0].end);
8355 tcp_update_dsack_list(tp, save_start,
8356 save_start + save_tlen);
8358 } else if (tlen >= save_tlen) {
8359 /* Update of sackblks. */
8360 tcp_update_dsack_list(tp, save_start,
8361 save_start + save_tlen);
8362 } else if (tlen > 0) {
8363 tcp_update_dsack_list(tp, save_start,
8373 * If FIN is received ACK the FIN and let the user know that the
8374 * connection is closing.
8376 if (thflags & TH_FIN) {
8377 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8378 /* The socket upcall is handled by socantrcvmore. */
8381 * If connection is half-synchronized (ie NEEDSYN
8382 * flag on) then delay ACK, so it may be piggybacked
8383 * when SYN is sent. Otherwise, since we received a
8384 * FIN then no more input can be expected, send ACK
8387 if (tp->t_flags & TF_NEEDSYN) {
8388 tp->t_flags |= TF_DELACK;
8389 bbr_timer_cancel(bbr,
8390 __LINE__, bbr->r_ctl.rc_rcvtime);
8392 tp->t_flags |= TF_ACKNOW;
8396 switch (tp->t_state) {
8398 * In SYN_RECEIVED and ESTABLISHED STATES enter the
8401 case TCPS_SYN_RECEIVED:
8402 tp->t_starttime = ticks;
8404 case TCPS_ESTABLISHED:
8405 tcp_state_change(tp, TCPS_CLOSE_WAIT);
8409 * If still in FIN_WAIT_1 STATE FIN has not been
8410 * acked so enter the CLOSING state.
8412 case TCPS_FIN_WAIT_1:
8413 tcp_state_change(tp, TCPS_CLOSING);
8417 * In FIN_WAIT_2 state enter the TIME_WAIT state,
8418 * starting the time-wait timer, turning off the
8419 * other standard timers.
8421 case TCPS_FIN_WAIT_2:
8422 bbr->rc_timer_first = 1;
8423 bbr_timer_cancel(bbr,
8424 __LINE__, bbr->r_ctl.rc_rcvtime);
8425 INP_WLOCK_ASSERT(tp->t_inpcb);
8431 * Return any desired output.
8433 if ((tp->t_flags & TF_ACKNOW) ||
8434 (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8435 bbr->r_wanted_output = 1;
8437 INP_WLOCK_ASSERT(tp->t_inpcb);
8442 * Here nothing is really faster, its just that we
8443 * have broken out the fast-data path also just like
8444 * the fast-ack. Return 1 if we processed the packet
8445 * return 0 if you need to take the "slow-path".
8448 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8449 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8450 uint32_t tiwin, int32_t nxt_pkt)
8453 int32_t newsize = 0; /* automatic sockbuf scaling */
8454 struct tcp_bbr *bbr;
8455 #ifdef NETFLIX_SB_LIMITS
8456 u_int mcnt, appended;
8460 * The size of tcp_saveipgen must be the size of the max ip header,
8463 u_char tcp_saveipgen[IP6_HDR_LEN];
8464 struct tcphdr tcp_savetcp;
8468 /* On the hpts and we would have called output */
8469 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8472 * If last ACK falls within this segment's sequence numbers, record
8473 * the timestamp. NOTE that the test is modified according to the
8474 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8476 if (bbr->r_ctl.rc_resend != NULL) {
8479 if (tiwin && tiwin != tp->snd_wnd) {
8482 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8485 if (__predict_false((to->to_flags & TOF_TS) &&
8486 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8489 if (__predict_false((th->th_ack != tp->snd_una))) {
8492 if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8495 if ((to->to_flags & TOF_TS) != 0 &&
8496 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8497 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8498 tp->ts_recent = to->to_tsval;
8501 * This is a pure, in-sequence data packet with nothing on the
8502 * reassembly queue and we have enough buffer space to take it.
8504 nsegs = max(1, m->m_pkthdr.lro_nsegs);
8506 #ifdef NETFLIX_SB_LIMITS
8507 if (so->so_rcv.sb_shlim) {
8510 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8511 CFO_NOSLEEP, NULL) == false) {
8512 counter_u64_add(tcp_sb_shlim_fails, 1);
8518 /* Clean receiver SACK report if present */
8519 if (tp->rcv_numsacks)
8520 tcp_clean_sackreport(tp);
8521 KMOD_TCPSTAT_INC(tcps_preddat);
8522 tp->rcv_nxt += tlen;
8524 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8525 (tp->t_fbyte_in == 0)) {
8526 tp->t_fbyte_in = ticks;
8527 if (tp->t_fbyte_in == 0)
8529 if (tp->t_fbyte_out && tp->t_fbyte_in)
8530 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8533 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8535 tp->snd_wl1 = th->th_seq;
8537 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8539 tp->rcv_up = tp->rcv_nxt;
8540 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8541 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8543 if (so->so_options & SO_DEBUG)
8544 tcp_trace(TA_INPUT, ostate, tp,
8545 (void *)tcp_saveipgen, &tcp_savetcp, 0);
8547 newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8549 /* Add data to socket buffer. */
8550 SOCKBUF_LOCK(&so->so_rcv);
8551 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8555 * Set new socket buffer size. Give up when limit is
8559 if (!sbreserve_locked(&so->so_rcv,
8561 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8562 m_adj(m, drop_hdrlen); /* delayed header drop */
8564 #ifdef NETFLIX_SB_LIMITS
8567 sbappendstream_locked(&so->so_rcv, m, 0);
8568 ctf_calc_rwin(so, tp);
8570 /* NB: sorwakeup_locked() does an implicit unlock. */
8571 sorwakeup_locked(so);
8572 #ifdef NETFLIX_SB_LIMITS
8573 if (so->so_rcv.sb_shlim && mcnt != appended)
8574 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8576 if (DELAY_ACK(tp, bbr, nsegs)) {
8577 bbr->bbr_segs_rcvd += max(1, nsegs);
8578 tp->t_flags |= TF_DELACK;
8579 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8581 bbr->r_wanted_output = 1;
8582 tp->t_flags |= TF_ACKNOW;
8588 * This subfunction is used to try to highly optimize the
8589 * fast path. We again allow window updates that are
8590 * in sequence to remain in the fast-path. We also add
8591 * in the __predict's to attempt to help the compiler.
8592 * Note that if we return a 0, then we can *not* process
8593 * it and the caller should push the packet into the
8594 * slow-path. If we return 1, then all is well and
8595 * the packet is fully processed.
8598 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8599 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8600 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8604 uint32_t sack_changed;
8607 * The size of tcp_saveipgen must be the size of the max ip header,
8610 u_char tcp_saveipgen[IP6_HDR_LEN];
8611 struct tcphdr tcp_savetcp;
8615 uint32_t prev_acked = 0;
8616 struct tcp_bbr *bbr;
8618 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8619 /* Old ack, behind (or duplicate to) the last one rcv'd */
8622 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8623 /* Above what we have sent? */
8626 if (__predict_false(tiwin == 0)) {
8630 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8631 /* We need a SYN or a FIN, unlikely.. */
8634 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8635 /* Timestamp is behind .. old ack with seq wrap? */
8638 if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8639 /* Still recovering */
8642 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8643 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8644 /* We are retransmitting */
8647 if (__predict_false(bbr->rc_in_persist != 0)) {
8648 /* In persist mode */
8651 if (bbr->r_ctl.rc_sacked) {
8652 /* We have sack holes on our scoreboard */
8655 /* Ok if we reach here, we can process a fast-ack */
8656 nsegs = max(1, m->m_pkthdr.lro_nsegs);
8657 sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8659 * We never detect loss in fast ack [we can't
8660 * have a sack and can't be in recovery so
8661 * we always pass 0 (nothing detected)].
8663 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8664 /* Did the window get updated? */
8665 if (tiwin != tp->snd_wnd) {
8666 tp->snd_wnd = tiwin;
8667 tp->snd_wl1 = th->th_seq;
8668 if (tp->snd_wnd > tp->max_sndwnd)
8669 tp->max_sndwnd = tp->snd_wnd;
8671 /* Do we need to exit persists? */
8672 if ((bbr->rc_in_persist != 0) &&
8673 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8674 bbr_minseg(bbr)))) {
8675 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8676 bbr->r_wanted_output = 1;
8678 /* Do we need to enter persists? */
8679 if ((bbr->rc_in_persist == 0) &&
8680 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8681 TCPS_HAVEESTABLISHED(tp->t_state) &&
8682 (tp->snd_max == tp->snd_una) &&
8683 sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8684 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8685 /* No send window.. we must enter persist */
8686 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8689 * If last ACK falls within this segment's sequence numbers, record
8690 * the timestamp. NOTE that the test is modified according to the
8691 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8693 if ((to->to_flags & TOF_TS) != 0 &&
8694 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8695 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8696 tp->ts_recent = to->to_tsval;
8699 * This is a pure ack for outstanding data.
8701 KMOD_TCPSTAT_INC(tcps_predack);
8704 * "bad retransmit" recovery.
8706 if (tp->t_flags & TF_PREVVALID) {
8707 tp->t_flags &= ~TF_PREVVALID;
8708 if (tp->t_rxtshift == 1 &&
8709 (int)(ticks - tp->t_badrxtwin) < 0)
8710 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8713 * Recalculate the transmit timer / rtt.
8715 * Some boxes send broken timestamp replies during the SYN+ACK
8716 * phase, ignore timestamps of 0 or we could calculate a huge RTT
8717 * and blow up the retransmit timer.
8719 acked = BYTES_THIS_ACK(tp, th);
8722 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8723 hhook_run_tcp_est_in(tp, th, to);
8726 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8727 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8728 sbdrop(&so->so_snd, acked);
8730 if (SEQ_GT(th->th_ack, tp->snd_una))
8731 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8732 tp->snd_una = th->th_ack;
8733 if (tp->snd_wnd < ctf_outstanding(tp))
8734 /* The peer collapsed its window on us */
8735 bbr_collapsed_window(bbr);
8736 else if (bbr->rc_has_collapsed)
8737 bbr_un_collapse_window(bbr);
8739 if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8740 tp->snd_recover = tp->snd_una;
8742 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8744 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8746 tp->snd_wl2 = th->th_ack;
8749 * If all outstanding data are acked, stop retransmit timer,
8750 * otherwise restart timer using current (possibly backed-off)
8751 * value. If process is waiting for space, wakeup/selwakeup/signal.
8752 * If data are ready to send, let tcp_output decide between more
8753 * output or persist.
8756 if (so->so_options & SO_DEBUG)
8757 tcp_trace(TA_INPUT, ostate, tp,
8758 (void *)tcp_saveipgen,
8761 /* Wake up the socket if we have room to write more */
8763 if (tp->snd_una == tp->snd_max) {
8764 /* Nothing left outstanding */
8765 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8766 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8767 bbr->rc_tp->t_acktime = 0;
8768 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8769 if (bbr->rc_in_persist == 0) {
8770 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8772 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8773 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8775 * We invalidate the last ack here since we
8776 * don't want to transfer forward the time
8777 * for our sum's calculations.
8779 bbr->r_wanted_output = 1;
8781 if (sbavail(&so->so_snd)) {
8782 bbr->r_wanted_output = 1;
8788 * Return value of 1, the TCB is unlocked and most
8789 * likely gone, return value of 0, the TCB is still
8793 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8794 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8795 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8798 int32_t ourfinisacked = 0;
8799 struct tcp_bbr *bbr;
8800 int32_t ret_val = 0;
8802 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8803 ctf_calc_rwin(so, tp);
8805 * If the state is SYN_SENT: if seg contains an ACK, but not for our
8806 * SYN, drop the input. if seg contains a RST, then drop the
8807 * connection. if seg does not contain SYN, then drop it. Otherwise
8808 * this is an acceptable SYN segment initialize tp->rcv_nxt and
8809 * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8810 * not support ECN so we will not say we are capable. if SYN has
8811 * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8812 * segment to be acked (eventually) continue processing rest of
8813 * data/controls, beginning with URG
8815 if ((thflags & TH_ACK) &&
8816 (SEQ_LEQ(th->th_ack, tp->iss) ||
8817 SEQ_GT(th->th_ack, tp->snd_max))) {
8818 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8819 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8822 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8823 TCP_PROBE5(connect__refused, NULL, tp,
8824 mtod(m, const char *), tp, th);
8825 tp = tcp_drop(tp, ECONNREFUSED);
8829 if (thflags & TH_RST) {
8833 if (!(thflags & TH_SYN)) {
8837 tp->irs = th->th_seq;
8839 if (thflags & TH_ACK) {
8840 int tfo_partial = 0;
8842 KMOD_TCPSTAT_INC(tcps_connects);
8845 mac_socketpeer_set_from_mbuf(m, so);
8847 /* Do window scaling on this connection? */
8848 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8849 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8850 tp->rcv_scale = tp->request_r_scale;
8852 tp->rcv_adv += min(tp->rcv_wnd,
8853 TCP_MAXWIN << tp->rcv_scale);
8855 * If not all the data that was sent in the TFO SYN
8856 * has been acked, resend the remainder right away.
8858 if (IS_FASTOPEN(tp->t_flags) &&
8859 (tp->snd_una != tp->snd_max)) {
8860 tp->snd_nxt = th->th_ack;
8864 * If there's data, delay ACK; if there's also a FIN ACKNOW
8865 * will be turned on later.
8867 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8868 bbr->bbr_segs_rcvd += 1;
8869 tp->t_flags |= TF_DELACK;
8870 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8872 bbr->r_wanted_output = 1;
8873 tp->t_flags |= TF_ACKNOW;
8875 if (SEQ_GT(th->th_ack, tp->iss)) {
8878 * handle it specially.
8880 bbr_log_syn(tp, to);
8882 if (SEQ_GT(th->th_ack, tp->snd_una)) {
8884 * We advance snd_una for the
8885 * fast open case. If th_ack is
8886 * acknowledging data beyond
8887 * snd_una we can't just call
8888 * ack-processing since the
8889 * data stream in our send-map
8890 * will start at snd_una + 1 (one
8891 * beyond the SYN). If its just
8892 * equal we don't need to do that
8893 * and there is no send_map.
8898 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
8899 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
8901 tp->t_starttime = ticks;
8902 if (tp->t_flags & TF_NEEDFIN) {
8903 tcp_state_change(tp, TCPS_FIN_WAIT_1);
8904 tp->t_flags &= ~TF_NEEDFIN;
8907 tcp_state_change(tp, TCPS_ESTABLISHED);
8908 TCP_PROBE5(connect__established, NULL, tp,
8909 mtod(m, const char *), tp, th);
8914 * Received initial SYN in SYN-SENT[*] state => simultaneous
8915 * open. If segment contains CC option and there is a
8916 * cached CC, apply TAO test. If it succeeds, connection is *
8917 * half-synchronized. Otherwise, do 3-way handshake:
8918 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
8919 * there was no CC option, clear cached CC value.
8921 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
8922 tcp_state_change(tp, TCPS_SYN_RECEIVED);
8924 INP_WLOCK_ASSERT(tp->t_inpcb);
8926 * Advance th->th_seq to correspond to first data byte. If data,
8927 * trim to stay within window, dropping FIN if necessary.
8930 if (tlen > tp->rcv_wnd) {
8931 todrop = tlen - tp->rcv_wnd;
8935 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
8936 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
8938 tp->snd_wl1 = th->th_seq - 1;
8939 tp->rcv_up = th->th_seq;
8941 * Client side of transaction: already sent SYN and data. If the
8942 * remote host used T/TCP to validate the SYN, our data will be
8943 * ACK'd; if so, enter normal data segment processing in the middle
8944 * of step 5, ack processing. Otherwise, goto step 6.
8946 if (thflags & TH_ACK) {
8947 if ((to->to_flags & TOF_TS) != 0) {
8950 t = tcp_tv_to_mssectick(&bbr->rc_tv);
8951 if (TSTMP_GEQ(t, to->to_tsecr)) {
8952 rtt = t - to->to_tsecr;
8957 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
8958 apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
8959 rtt, bbr->r_ctl.rc_rcvtime);
8962 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
8964 /* We may have changed to FIN_WAIT_1 above */
8965 if (tp->t_state == TCPS_FIN_WAIT_1) {
8967 * In FIN_WAIT_1 STATE in addition to the processing
8968 * for the ESTABLISHED state if our FIN is now
8969 * acknowledged then enter FIN_WAIT_2.
8971 if (ourfinisacked) {
8973 * If we can't receive any more data, then
8974 * closing user can proceed. Starting the
8975 * timer is contrary to the specification,
8976 * but if we don't get a FIN we'll hang
8979 * XXXjl: we should release the tp also, and
8980 * use a compressed state.
8982 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8983 soisdisconnected(so);
8984 tcp_timer_activate(tp, TT_2MSL,
8985 (tcp_fast_finwait2_recycle ?
8986 tcp_finwait2_timeout :
8989 tcp_state_change(tp, TCPS_FIN_WAIT_2);
8993 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
8994 tiwin, thflags, nxt_pkt));
8998 * Return value of 1, the TCB is unlocked and most
8999 * likely gone, return value of 0, the TCB is still
9003 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9004 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9005 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9007 int32_t ourfinisacked = 0;
9009 struct tcp_bbr *bbr;
9011 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9012 ctf_calc_rwin(so, tp);
9013 if ((thflags & TH_ACK) &&
9014 (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9015 SEQ_GT(th->th_ack, tp->snd_max))) {
9016 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9017 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9020 if (IS_FASTOPEN(tp->t_flags)) {
9022 * When a TFO connection is in SYN_RECEIVED, the only valid
9023 * packets are the initial SYN, a retransmit/copy of the
9024 * initial SYN (possibly with a subset of the original
9025 * data), a valid ACK, a FIN, or a RST.
9027 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
9028 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9029 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9031 } else if (thflags & TH_SYN) {
9032 /* non-initial SYN is ignored */
9033 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9034 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9035 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9036 ctf_do_drop(m, NULL);
9039 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9040 ctf_do_drop(m, NULL);
9044 if ((thflags & TH_RST) ||
9045 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9046 return (ctf_process_rst(m, th, so, tp));
9048 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9049 * it's less than ts_recent, drop it.
9051 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9052 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9053 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9057 * In the SYN-RECEIVED state, validate that the packet belongs to
9058 * this connection before trimming the data to fit the receive
9059 * window. Check the sequence number versus IRS since we know the
9060 * sequence numbers haven't wrapped. This is a partial fix for the
9061 * "LAND" DoS attack.
9063 if (SEQ_LT(th->th_seq, tp->irs)) {
9064 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9065 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9068 INP_WLOCK_ASSERT(tp->t_inpcb);
9069 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9073 * If last ACK falls within this segment's sequence numbers, record
9074 * its timestamp. NOTE: 1) That the test incorporates suggestions
9075 * from the latest proposal of the tcplw@cray.com list (Braden
9076 * 1993/04/26). 2) That updating only on newer timestamps interferes
9077 * with our earlier PAWS tests, so this check should be solely
9078 * predicated on the sequence space of this segment. 3) That we
9079 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9080 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9081 * SEG.Len, This modified check allows us to overcome RFC1323's
9082 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9083 * p.869. In such cases, we can still calculate the RTT correctly
9084 * when RCV.NXT == Last.ACK.Sent.
9086 if ((to->to_flags & TOF_TS) != 0 &&
9087 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9088 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9089 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9090 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9091 tp->ts_recent = to->to_tsval;
9093 tp->snd_wnd = tiwin;
9095 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9096 * is on (half-synchronized state), then queue data for later
9097 * processing; else drop segment and return.
9099 if ((thflags & TH_ACK) == 0) {
9100 if (IS_FASTOPEN(tp->t_flags)) {
9103 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9104 tiwin, thflags, nxt_pkt));
9106 KMOD_TCPSTAT_INC(tcps_connects);
9108 /* Do window scaling? */
9109 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9110 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9111 tp->rcv_scale = tp->request_r_scale;
9114 * ok for the first time in lets see if we can use the ts to figure
9115 * out what the initial RTT was.
9117 if ((to->to_flags & TOF_TS) != 0) {
9120 t = tcp_tv_to_mssectick(&bbr->rc_tv);
9121 if (TSTMP_GEQ(t, to->to_tsecr)) {
9122 rtt = t - to->to_tsecr;
9127 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9128 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9131 /* Drop off any SYN in the send map (probably not there) */
9132 if (thflags & TH_ACK)
9133 bbr_log_syn(tp, to);
9134 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9135 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9136 tp->t_tfo_pending = NULL;
9139 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* ->
9142 tp->t_starttime = ticks;
9143 if (tp->t_flags & TF_NEEDFIN) {
9144 tcp_state_change(tp, TCPS_FIN_WAIT_1);
9145 tp->t_flags &= ~TF_NEEDFIN;
9147 tcp_state_change(tp, TCPS_ESTABLISHED);
9148 TCP_PROBE5(accept__established, NULL, tp,
9149 mtod(m, const char *), tp, th);
9151 * TFO connections call cc_conn_init() during SYN
9152 * processing. Calling it again here for such connections
9153 * is not harmless as it would undo the snd_cwnd reduction
9154 * that occurs when a TFO SYN|ACK is retransmitted.
9156 if (!IS_FASTOPEN(tp->t_flags))
9160 * Account for the ACK of our SYN prior to
9161 * regular ACK processing below, except for
9162 * simultaneous SYN, which is handled later.
9164 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9167 * If segment contains data or ACK, will call tcp_reass() later; if
9168 * not, do so now to pass queued data to user.
9170 if (tlen == 0 && (thflags & TH_FIN) == 0) {
9171 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9173 if (tp->t_flags & TF_WAKESOR) {
9174 tp->t_flags &= ~TF_WAKESOR;
9175 /* NB: sorwakeup_locked() does an implicit unlock. */
9176 sorwakeup_locked(so);
9179 tp->snd_wl1 = th->th_seq - 1;
9180 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9183 if (tp->t_state == TCPS_FIN_WAIT_1) {
9184 /* We could have went to FIN_WAIT_1 (or EST) above */
9186 * In FIN_WAIT_1 STATE in addition to the processing for the
9187 * ESTABLISHED state if our FIN is now acknowledged then
9190 if (ourfinisacked) {
9192 * If we can't receive any more data, then closing
9193 * user can proceed. Starting the timer is contrary
9194 * to the specification, but if we don't get a FIN
9195 * we'll hang forever.
9197 * XXXjl: we should release the tp also, and use a
9200 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9201 soisdisconnected(so);
9202 tcp_timer_activate(tp, TT_2MSL,
9203 (tcp_fast_finwait2_recycle ?
9204 tcp_finwait2_timeout :
9207 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9210 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9211 tiwin, thflags, nxt_pkt));
9215 * Return value of 1, the TCB is unlocked and most
9216 * likely gone, return value of 0, the TCB is still
9220 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9221 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9222 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9224 struct tcp_bbr *bbr;
9228 * Header prediction: check for the two common cases of a
9229 * uni-directional data xfer. If the packet has no control flags,
9230 * is in-sequence, the window didn't change and we're not
9231 * retransmitting, it's a candidate. If the length is zero and the
9232 * ack moved forward, we're the sender side of the xfer. Just free
9233 * the data acked & wake any higher level process that was blocked
9234 * waiting for space. If the length is non-zero and the ack didn't
9235 * move, we're the receiver side. If we're getting packets in-order
9236 * (the reassembly queue is empty), add the data toc The socket
9237 * buffer and note that we need a delayed ack. Make sure that the
9238 * hidden state-flags are also off. Since we check for
9239 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9241 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9242 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9244 * If we have delived under 4 segments increase the initial
9245 * window if raised by the peer. We use this to determine
9246 * dynamic and static rwnd's at the end of a connection.
9248 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9250 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9251 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9252 __predict_true(SEGQ_EMPTY(tp)) &&
9253 __predict_true(th->th_seq == tp->rcv_nxt)) {
9255 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9256 tiwin, nxt_pkt, iptos)) {
9260 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9266 ctf_calc_rwin(so, tp);
9268 if ((thflags & TH_RST) ||
9269 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9270 return (ctf_process_rst(m, th, so, tp));
9272 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9273 * synchronized state.
9275 if (thflags & TH_SYN) {
9276 ctf_challenge_ack(m, th, tp, &ret_val);
9280 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9281 * it's less than ts_recent, drop it.
9283 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9284 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9285 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9288 INP_WLOCK_ASSERT(tp->t_inpcb);
9289 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9293 * If last ACK falls within this segment's sequence numbers, record
9294 * its timestamp. NOTE: 1) That the test incorporates suggestions
9295 * from the latest proposal of the tcplw@cray.com list (Braden
9296 * 1993/04/26). 2) That updating only on newer timestamps interferes
9297 * with our earlier PAWS tests, so this check should be solely
9298 * predicated on the sequence space of this segment. 3) That we
9299 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9300 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9301 * SEG.Len, This modified check allows us to overcome RFC1323's
9302 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9303 * p.869. In such cases, we can still calculate the RTT correctly
9304 * when RCV.NXT == Last.ACK.Sent.
9306 if ((to->to_flags & TOF_TS) != 0 &&
9307 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9308 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9309 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9310 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9311 tp->ts_recent = to->to_tsval;
9314 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9315 * is on (half-synchronized state), then queue data for later
9316 * processing; else drop segment and return.
9318 if ((thflags & TH_ACK) == 0) {
9319 if (tp->t_flags & TF_NEEDSYN) {
9320 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9321 tiwin, thflags, nxt_pkt));
9322 } else if (tp->t_flags & TF_ACKNOW) {
9323 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9324 bbr->r_wanted_output = 1;
9327 ctf_do_drop(m, NULL);
9334 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9337 if (sbavail(&so->so_snd)) {
9338 if (ctf_progress_timeout_check(tp, true)) {
9339 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9340 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9344 /* State changes only happen in bbr_process_data() */
9345 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9346 tiwin, thflags, nxt_pkt));
9350 * Return value of 1, the TCB is unlocked and most
9351 * likely gone, return value of 0, the TCB is still
9355 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9356 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9357 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9359 struct tcp_bbr *bbr;
9362 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9363 ctf_calc_rwin(so, tp);
9364 if ((thflags & TH_RST) ||
9365 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9366 return (ctf_process_rst(m, th, so, tp));
9368 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9369 * synchronized state.
9371 if (thflags & TH_SYN) {
9372 ctf_challenge_ack(m, th, tp, &ret_val);
9376 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9377 * it's less than ts_recent, drop it.
9379 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9380 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9381 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9384 INP_WLOCK_ASSERT(tp->t_inpcb);
9385 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9389 * If last ACK falls within this segment's sequence numbers, record
9390 * its timestamp. NOTE: 1) That the test incorporates suggestions
9391 * from the latest proposal of the tcplw@cray.com list (Braden
9392 * 1993/04/26). 2) That updating only on newer timestamps interferes
9393 * with our earlier PAWS tests, so this check should be solely
9394 * predicated on the sequence space of this segment. 3) That we
9395 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9396 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9397 * SEG.Len, This modified check allows us to overcome RFC1323's
9398 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9399 * p.869. In such cases, we can still calculate the RTT correctly
9400 * when RCV.NXT == Last.ACK.Sent.
9402 if ((to->to_flags & TOF_TS) != 0 &&
9403 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9404 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9405 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9406 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9407 tp->ts_recent = to->to_tsval;
9410 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9411 * is on (half-synchronized state), then queue data for later
9412 * processing; else drop segment and return.
9414 if ((thflags & TH_ACK) == 0) {
9415 if (tp->t_flags & TF_NEEDSYN) {
9416 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9417 tiwin, thflags, nxt_pkt));
9418 } else if (tp->t_flags & TF_ACKNOW) {
9419 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9420 bbr->r_wanted_output = 1;
9423 ctf_do_drop(m, NULL);
9430 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9433 if (sbavail(&so->so_snd)) {
9434 if (ctf_progress_timeout_check(tp, true)) {
9435 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9436 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9440 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9441 tiwin, thflags, nxt_pkt));
9445 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9446 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9449 if (bbr->rc_allow_data_af_clo == 0) {
9451 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9452 /* tcp_close will kill the inp pre-log the Reset */
9453 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9455 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9456 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9459 if (sbavail(&so->so_snd) == 0)
9461 /* Ok we allow data that is ignored and a followup reset */
9462 tp->rcv_nxt = th->th_seq + *tlen;
9463 tp->t_flags2 |= TF2_DROP_AF_DATA;
9464 bbr->r_wanted_output = 1;
9470 * Return value of 1, the TCB is unlocked and most
9471 * likely gone, return value of 0, the TCB is still
9475 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9476 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9477 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9479 int32_t ourfinisacked = 0;
9481 struct tcp_bbr *bbr;
9483 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9484 ctf_calc_rwin(so, tp);
9485 if ((thflags & TH_RST) ||
9486 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9487 return (ctf_process_rst(m, th, so, tp));
9489 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9490 * synchronized state.
9492 if (thflags & TH_SYN) {
9493 ctf_challenge_ack(m, th, tp, &ret_val);
9497 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9498 * it's less than ts_recent, drop it.
9500 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9501 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9502 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9505 INP_WLOCK_ASSERT(tp->t_inpcb);
9506 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9510 * If new data are received on a connection after the user processes
9511 * are gone, then RST the other end.
9513 if ((so->so_state & SS_NOFDREF) && tlen) {
9515 * We call a new function now so we might continue and setup
9516 * to reset at all data being ack'd.
9518 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9522 * If last ACK falls within this segment's sequence numbers, record
9523 * its timestamp. NOTE: 1) That the test incorporates suggestions
9524 * from the latest proposal of the tcplw@cray.com list (Braden
9525 * 1993/04/26). 2) That updating only on newer timestamps interferes
9526 * with our earlier PAWS tests, so this check should be solely
9527 * predicated on the sequence space of this segment. 3) That we
9528 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9529 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9530 * SEG.Len, This modified check allows us to overcome RFC1323's
9531 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9532 * p.869. In such cases, we can still calculate the RTT correctly
9533 * when RCV.NXT == Last.ACK.Sent.
9535 if ((to->to_flags & TOF_TS) != 0 &&
9536 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9537 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9538 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9539 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9540 tp->ts_recent = to->to_tsval;
9543 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9544 * is on (half-synchronized state), then queue data for later
9545 * processing; else drop segment and return.
9547 if ((thflags & TH_ACK) == 0) {
9548 if (tp->t_flags & TF_NEEDSYN) {
9549 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9550 tiwin, thflags, nxt_pkt));
9551 } else if (tp->t_flags & TF_ACKNOW) {
9552 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9553 bbr->r_wanted_output = 1;
9556 ctf_do_drop(m, NULL);
9563 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9566 if (ourfinisacked) {
9568 * If we can't receive any more data, then closing user can
9569 * proceed. Starting the timer is contrary to the
9570 * specification, but if we don't get a FIN we'll hang
9573 * XXXjl: we should release the tp also, and use a
9576 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9577 soisdisconnected(so);
9578 tcp_timer_activate(tp, TT_2MSL,
9579 (tcp_fast_finwait2_recycle ?
9580 tcp_finwait2_timeout :
9583 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9585 if (sbavail(&so->so_snd)) {
9586 if (ctf_progress_timeout_check(tp, true)) {
9587 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9588 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9592 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9593 tiwin, thflags, nxt_pkt));
9597 * Return value of 1, the TCB is unlocked and most
9598 * likely gone, return value of 0, the TCB is still
9602 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9603 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9604 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9606 int32_t ourfinisacked = 0;
9608 struct tcp_bbr *bbr;
9610 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9611 ctf_calc_rwin(so, tp);
9612 if ((thflags & TH_RST) ||
9613 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9614 return (ctf_process_rst(m, th, so, tp));
9616 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9617 * synchronized state.
9619 if (thflags & TH_SYN) {
9620 ctf_challenge_ack(m, th, tp, &ret_val);
9624 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9625 * it's less than ts_recent, drop it.
9627 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9628 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9629 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9632 INP_WLOCK_ASSERT(tp->t_inpcb);
9633 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9637 * If new data are received on a connection after the user processes
9638 * are gone, then RST the other end.
9640 if ((so->so_state & SS_NOFDREF) && tlen) {
9642 * We call a new function now so we might continue and setup
9643 * to reset at all data being ack'd.
9645 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9649 * If last ACK falls within this segment's sequence numbers, record
9650 * its timestamp. NOTE: 1) That the test incorporates suggestions
9651 * from the latest proposal of the tcplw@cray.com list (Braden
9652 * 1993/04/26). 2) That updating only on newer timestamps interferes
9653 * with our earlier PAWS tests, so this check should be solely
9654 * predicated on the sequence space of this segment. 3) That we
9655 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9656 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9657 * SEG.Len, This modified check allows us to overcome RFC1323's
9658 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9659 * p.869. In such cases, we can still calculate the RTT correctly
9660 * when RCV.NXT == Last.ACK.Sent.
9662 if ((to->to_flags & TOF_TS) != 0 &&
9663 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9664 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9665 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9666 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9667 tp->ts_recent = to->to_tsval;
9670 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9671 * is on (half-synchronized state), then queue data for later
9672 * processing; else drop segment and return.
9674 if ((thflags & TH_ACK) == 0) {
9675 if (tp->t_flags & TF_NEEDSYN) {
9676 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9677 tiwin, thflags, nxt_pkt));
9678 } else if (tp->t_flags & TF_ACKNOW) {
9679 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9680 bbr->r_wanted_output = 1;
9683 ctf_do_drop(m, NULL);
9690 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9693 if (ourfinisacked) {
9698 if (sbavail(&so->so_snd)) {
9699 if (ctf_progress_timeout_check(tp, true)) {
9700 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9701 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9705 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9706 tiwin, thflags, nxt_pkt));
9710 * Return value of 1, the TCB is unlocked and most
9711 * likely gone, return value of 0, the TCB is still
9715 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9716 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9717 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9719 int32_t ourfinisacked = 0;
9721 struct tcp_bbr *bbr;
9723 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9724 ctf_calc_rwin(so, tp);
9725 if ((thflags & TH_RST) ||
9726 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9727 return (ctf_process_rst(m, th, so, tp));
9729 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9730 * synchronized state.
9732 if (thflags & TH_SYN) {
9733 ctf_challenge_ack(m, th, tp, &ret_val);
9737 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9738 * it's less than ts_recent, drop it.
9740 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9741 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9742 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9745 INP_WLOCK_ASSERT(tp->t_inpcb);
9746 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9750 * If new data are received on a connection after the user processes
9751 * are gone, then RST the other end.
9753 if ((so->so_state & SS_NOFDREF) && tlen) {
9755 * We call a new function now so we might continue and setup
9756 * to reset at all data being ack'd.
9758 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9762 * If last ACK falls within this segment's sequence numbers, record
9763 * its timestamp. NOTE: 1) That the test incorporates suggestions
9764 * from the latest proposal of the tcplw@cray.com list (Braden
9765 * 1993/04/26). 2) That updating only on newer timestamps interferes
9766 * with our earlier PAWS tests, so this check should be solely
9767 * predicated on the sequence space of this segment. 3) That we
9768 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9769 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9770 * SEG.Len, This modified check allows us to overcome RFC1323's
9771 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9772 * p.869. In such cases, we can still calculate the RTT correctly
9773 * when RCV.NXT == Last.ACK.Sent.
9775 if ((to->to_flags & TOF_TS) != 0 &&
9776 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9777 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9778 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9779 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9780 tp->ts_recent = to->to_tsval;
9783 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9784 * is on (half-synchronized state), then queue data for later
9785 * processing; else drop segment and return.
9787 if ((thflags & TH_ACK) == 0) {
9788 if (tp->t_flags & TF_NEEDSYN) {
9789 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9790 tiwin, thflags, nxt_pkt));
9791 } else if (tp->t_flags & TF_ACKNOW) {
9792 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9793 bbr->r_wanted_output = 1;
9796 ctf_do_drop(m, NULL);
9801 * case TCPS_LAST_ACK: Ack processing.
9803 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9806 if (ourfinisacked) {
9811 if (sbavail(&so->so_snd)) {
9812 if (ctf_progress_timeout_check(tp, true)) {
9813 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9814 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9818 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9819 tiwin, thflags, nxt_pkt));
9823 * Return value of 1, the TCB is unlocked and most
9824 * likely gone, return value of 0, the TCB is still
9828 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9829 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9830 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9832 int32_t ourfinisacked = 0;
9834 struct tcp_bbr *bbr;
9836 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9837 ctf_calc_rwin(so, tp);
9838 /* Reset receive buffer auto scaling when not in bulk receive mode. */
9839 if ((thflags & TH_RST) ||
9840 (tp->t_fin_is_rst && (thflags & TH_FIN)))
9841 return (ctf_process_rst(m, th, so, tp));
9844 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9845 * synchronized state.
9847 if (thflags & TH_SYN) {
9848 ctf_challenge_ack(m, th, tp, &ret_val);
9851 INP_WLOCK_ASSERT(tp->t_inpcb);
9853 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9854 * it's less than ts_recent, drop it.
9856 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9857 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9858 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9861 INP_WLOCK_ASSERT(tp->t_inpcb);
9862 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9866 * If new data are received on a connection after the user processes
9867 * are gone, then we may RST the other end depending on the outcome
9868 * of bbr_check_data_after_close.
9870 if ((so->so_state & SS_NOFDREF) &&
9873 * We call a new function now so we might continue and setup
9874 * to reset at all data being ack'd.
9876 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9879 INP_WLOCK_ASSERT(tp->t_inpcb);
9881 * If last ACK falls within this segment's sequence numbers, record
9882 * its timestamp. NOTE: 1) That the test incorporates suggestions
9883 * from the latest proposal of the tcplw@cray.com list (Braden
9884 * 1993/04/26). 2) That updating only on newer timestamps interferes
9885 * with our earlier PAWS tests, so this check should be solely
9886 * predicated on the sequence space of this segment. 3) That we
9887 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9888 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9889 * SEG.Len, This modified check allows us to overcome RFC1323's
9890 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9891 * p.869. In such cases, we can still calculate the RTT correctly
9892 * when RCV.NXT == Last.ACK.Sent.
9894 INP_WLOCK_ASSERT(tp->t_inpcb);
9895 if ((to->to_flags & TOF_TS) != 0 &&
9896 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9897 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9898 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9899 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9900 tp->ts_recent = to->to_tsval;
9903 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
9904 * is on (half-synchronized state), then queue data for later
9905 * processing; else drop segment and return.
9907 if ((thflags & TH_ACK) == 0) {
9908 if (tp->t_flags & TF_NEEDSYN) {
9909 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9910 tiwin, thflags, nxt_pkt));
9911 } else if (tp->t_flags & TF_ACKNOW) {
9912 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9913 bbr->r_wanted_output = 1;
9916 ctf_do_drop(m, NULL);
9923 INP_WLOCK_ASSERT(tp->t_inpcb);
9924 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9927 if (sbavail(&so->so_snd)) {
9928 if (ctf_progress_timeout_check(tp, true)) {
9929 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9930 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9934 INP_WLOCK_ASSERT(tp->t_inpcb);
9935 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9936 tiwin, thflags, nxt_pkt));
9940 bbr_stop_all_timers(struct tcpcb *tp)
9942 struct tcp_bbr *bbr;
9945 * Assure no timers are running.
9947 if (tcp_timer_active(tp, TT_PERSIST)) {
9948 /* We enter in persists, set the flag appropriately */
9949 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9950 bbr->rc_in_persist = 1;
9952 tcp_timer_suspend(tp, TT_PERSIST);
9953 tcp_timer_suspend(tp, TT_REXMT);
9954 tcp_timer_suspend(tp, TT_KEEP);
9955 tcp_timer_suspend(tp, TT_DELACK);
9959 bbr_google_mode_on(struct tcp_bbr *bbr)
9961 bbr->rc_use_google = 1;
9962 bbr->rc_no_pacing = 0;
9963 bbr->r_ctl.bbr_google_discount = bbr_google_discount;
9964 bbr->r_use_policer = bbr_policer_detection_enabled;
9965 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
9966 bbr->bbr_use_rack_cheat = 0;
9967 bbr->r_ctl.rc_incr_tmrs = 0;
9968 bbr->r_ctl.rc_inc_tcp_oh = 0;
9969 bbr->r_ctl.rc_inc_ip_oh = 0;
9970 bbr->r_ctl.rc_inc_enet_oh = 0;
9971 reset_time(&bbr->r_ctl.rc_delrate,
9972 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
9973 reset_time_small(&bbr->r_ctl.rc_rttprop,
9974 (11 * USECS_IN_SECOND));
9975 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
9979 bbr_google_mode_off(struct tcp_bbr *bbr)
9981 bbr->rc_use_google = 0;
9982 bbr->r_ctl.bbr_google_discount = 0;
9983 bbr->no_pacing_until = bbr_no_pacing_until;
9984 bbr->r_use_policer = 0;
9985 if (bbr->no_pacing_until)
9986 bbr->rc_no_pacing = 1;
9988 bbr->rc_no_pacing = 0;
9989 if (bbr_use_rack_resend_cheat)
9990 bbr->bbr_use_rack_cheat = 1;
9992 bbr->bbr_use_rack_cheat = 0;
9993 if (bbr_incr_timers)
9994 bbr->r_ctl.rc_incr_tmrs = 1;
9996 bbr->r_ctl.rc_incr_tmrs = 0;
9997 if (bbr_include_tcp_oh)
9998 bbr->r_ctl.rc_inc_tcp_oh = 1;
10000 bbr->r_ctl.rc_inc_tcp_oh = 0;
10001 if (bbr_include_ip_oh)
10002 bbr->r_ctl.rc_inc_ip_oh = 1;
10004 bbr->r_ctl.rc_inc_ip_oh = 0;
10005 if (bbr_include_enet_oh)
10006 bbr->r_ctl.rc_inc_enet_oh = 1;
10008 bbr->r_ctl.rc_inc_enet_oh = 0;
10009 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10010 reset_time(&bbr->r_ctl.rc_delrate,
10011 bbr_num_pktepo_for_del_limit);
10012 reset_time_small(&bbr->r_ctl.rc_rttprop,
10013 (bbr_filter_len_sec * USECS_IN_SECOND));
10014 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10017 * Return 0 on success, non-zero on failure
10018 * which indicates the error (usually no memory).
10021 bbr_init(struct tcpcb *tp)
10023 struct tcp_bbr *bbr = NULL;
10027 tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
10028 if (tp->t_fb_ptr == NULL) {
10030 * We need to allocate memory but cant. The INP and INP_INFO
10031 * locks and they are recusive (happens during setup. So a
10032 * scheme to drop the locks fails :(
10037 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10038 bbr->rtt_valid = 0;
10040 inp->inp_flags2 |= INP_CANNOT_DO_ECN;
10041 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10042 TAILQ_INIT(&bbr->r_ctl.rc_map);
10043 TAILQ_INIT(&bbr->r_ctl.rc_free);
10044 TAILQ_INIT(&bbr->r_ctl.rc_tmap);
10047 bbr->rc_inp = tp->t_inpcb;
10049 cts = tcp_get_usecs(&bbr->rc_tv);
10051 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
10052 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
10053 bbr->rc_tlp_threshold = bbr_tlp_thresh;
10054 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
10055 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
10056 bbr->r_ctl.rc_min_to = bbr_min_to;
10057 bbr->rc_bbr_state = BBR_STATE_STARTUP;
10058 bbr->r_ctl.bbr_lost_at_state = 0;
10059 bbr->r_ctl.rc_lost_at_startup = 0;
10060 bbr->rc_all_timers_stopped = 0;
10061 bbr->r_ctl.rc_bbr_lastbtlbw = 0;
10062 bbr->r_ctl.rc_pkt_epoch_del = 0;
10063 bbr->r_ctl.rc_pkt_epoch = 0;
10064 bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
10065 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
10066 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
10067 bbr->r_ctl.rc_went_idle_time = cts;
10068 bbr->rc_pacer_started = cts;
10069 bbr->r_ctl.rc_pkt_epoch_time = cts;
10070 bbr->r_ctl.rc_rcvtime = cts;
10071 bbr->r_ctl.rc_bbr_state_time = cts;
10072 bbr->r_ctl.rc_del_time = cts;
10073 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
10074 bbr->r_ctl.last_in_probertt = cts;
10075 bbr->skip_gain = 0;
10076 bbr->gain_is_limited = 0;
10077 bbr->no_pacing_until = bbr_no_pacing_until;
10078 if (bbr->no_pacing_until)
10079 bbr->rc_no_pacing = 1;
10080 if (bbr_use_google_algo) {
10081 bbr->rc_no_pacing = 0;
10082 bbr->rc_use_google = 1;
10083 bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10084 bbr->r_use_policer = bbr_policer_detection_enabled;
10086 bbr->rc_use_google = 0;
10087 bbr->r_ctl.bbr_google_discount = 0;
10088 bbr->r_use_policer = 0;
10090 if (bbr_ts_limiting)
10091 bbr->rc_use_ts_limit = 1;
10093 bbr->rc_use_ts_limit = 0;
10094 if (bbr_ts_can_raise)
10095 bbr->ts_can_raise = 1;
10097 bbr->ts_can_raise = 0;
10098 if (V_tcp_delack_enabled == 1)
10099 tp->t_delayed_ack = 2;
10100 else if (V_tcp_delack_enabled == 0)
10101 tp->t_delayed_ack = 0;
10102 else if (V_tcp_delack_enabled < 100)
10103 tp->t_delayed_ack = V_tcp_delack_enabled;
10105 tp->t_delayed_ack = 2;
10106 if (bbr->rc_use_google == 0)
10107 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10109 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10110 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10111 bbr->rc_max_rto_sec = bbr_rto_max_sec;
10112 bbr->rc_init_win = bbr_def_init_win;
10113 if (tp->t_flags & TF_REQ_TSTMP)
10114 bbr->rc_last_options = TCP_TS_OVERHEAD;
10115 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10116 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10117 bbr->r_init_rtt = 1;
10119 counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10120 if (bbr_allow_hdwr_pacing)
10121 bbr->bbr_hdw_pace_ena = 1;
10123 bbr->bbr_hdw_pace_ena = 0;
10124 if (bbr_sends_full_iwnd)
10125 bbr->bbr_init_win_cheat = 1;
10127 bbr->bbr_init_win_cheat = 0;
10128 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10129 bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10130 bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10131 bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10132 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10133 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10134 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10135 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10136 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10137 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10138 bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10139 bbr->r_ctl.rc_rtt_shrinks = cts;
10140 if (bbr->rc_use_google) {
10141 setup_time_filter(&bbr->r_ctl.rc_delrate,
10143 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10144 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10145 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10147 setup_time_filter(&bbr->r_ctl.rc_delrate,
10149 bbr_num_pktepo_for_del_limit);
10150 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10151 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10153 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10154 if (bbr_uses_idle_restart)
10155 bbr->rc_use_idle_restart = 1;
10157 bbr->rc_use_idle_restart = 0;
10158 bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10159 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10160 if (bbr_resends_use_tso)
10161 bbr->rc_resends_use_tso = 1;
10162 #ifdef NETFLIX_PEAKRATE
10163 tp->t_peakrate_thr = tp->t_maxpeakrate;
10165 if (tp->snd_una != tp->snd_max) {
10166 /* Create a send map for the current outstanding data */
10167 struct bbr_sendmap *rsm;
10169 rsm = bbr_alloc(bbr);
10171 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10172 tp->t_fb_ptr = NULL;
10175 rsm->r_rtt_not_allowed = 1;
10176 rsm->r_tim_lastsent[0] = cts;
10177 rsm->r_rtr_cnt = 1;
10178 rsm->r_rtr_bytes = 0;
10179 rsm->r_start = tp->snd_una;
10180 rsm->r_end = tp->snd_max;
10182 rsm->r_delivered = bbr->r_ctl.rc_delivered;
10183 rsm->r_ts_valid = 0;
10184 rsm->r_del_ack_ts = tp->ts_recent;
10185 rsm->r_del_time = cts;
10186 if (bbr->r_ctl.r_app_limited_until)
10187 rsm->r_app_limited = 1;
10189 rsm->r_app_limited = 0;
10190 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10191 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10192 rsm->r_in_tmap = 1;
10193 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10194 rsm->r_bbr_state = bbr_state_val(bbr);
10196 rsm->r_bbr_state = 8;
10198 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10199 bbr->bbr_use_rack_cheat = 1;
10200 if (bbr_incr_timers && (bbr->rc_use_google == 0))
10201 bbr->r_ctl.rc_incr_tmrs = 1;
10202 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10203 bbr->r_ctl.rc_inc_tcp_oh = 1;
10204 if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10205 bbr->r_ctl.rc_inc_ip_oh = 1;
10206 if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10207 bbr->r_ctl.rc_inc_enet_oh = 1;
10209 bbr_log_type_statechange(bbr, cts, __LINE__);
10210 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10214 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10215 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10217 /* announce the settings and state */
10218 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10219 tcp_bbr_tso_size_check(bbr, cts);
10221 * Now call the generic function to start a timer. This will place
10222 * the TCB on the hptsi wheel if a timer is needed with appropriate
10225 bbr_stop_all_timers(tp);
10226 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10231 * Return 0 if we can accept the connection. Return
10232 * non-zero if we can't handle the connection. A EAGAIN
10233 * means you need to wait until the connection is up.
10234 * a EADDRNOTAVAIL means we can never handle the connection
10238 bbr_handoff_ok(struct tcpcb *tp)
10240 if ((tp->t_state == TCPS_CLOSED) ||
10241 (tp->t_state == TCPS_LISTEN)) {
10242 /* Sure no problem though it may not stick */
10245 if ((tp->t_state == TCPS_SYN_SENT) ||
10246 (tp->t_state == TCPS_SYN_RECEIVED)) {
10248 * We really don't know you have to get to ESTAB or beyond
10253 if (tp->t_flags & TF_SENTFIN)
10255 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10259 * If we reach here we don't do SACK on this connection so we can
10266 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10268 if (tp->t_fb_ptr) {
10270 struct tcp_bbr *bbr;
10271 struct bbr_sendmap *rsm;
10273 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10274 if (bbr->r_ctl.crte)
10275 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10276 bbr_log_flowend(bbr);
10279 /* Backout any flags2 we applied */
10280 tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN;
10281 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10282 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10284 if (bbr->bbr_hdrw_pacing)
10285 counter_u64_add(bbr_flows_whdwr_pacing, -1);
10287 counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10288 if (bbr->r_ctl.crte != NULL) {
10289 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
10290 bbr->r_ctl.crte = NULL;
10292 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10294 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10295 uma_zfree(bbr_zone, rsm);
10296 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10298 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10300 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10301 uma_zfree(bbr_zone, rsm);
10302 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10304 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10305 if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10306 BBR_STAT_INC(bbr_dynamic_rwnd);
10308 BBR_STAT_INC(bbr_static_rwnd);
10309 bbr->r_ctl.rc_free_cnt = 0;
10310 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10311 tp->t_fb_ptr = NULL;
10313 /* Make sure snd_nxt is correctly set */
10314 tp->snd_nxt = tp->snd_max;
10318 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10320 switch (tp->t_state) {
10321 case TCPS_SYN_SENT:
10322 bbr->r_state = TCPS_SYN_SENT;
10323 bbr->r_substate = bbr_do_syn_sent;
10325 case TCPS_SYN_RECEIVED:
10326 bbr->r_state = TCPS_SYN_RECEIVED;
10327 bbr->r_substate = bbr_do_syn_recv;
10329 case TCPS_ESTABLISHED:
10330 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10331 bbr->r_state = TCPS_ESTABLISHED;
10332 bbr->r_substate = bbr_do_established;
10334 case TCPS_CLOSE_WAIT:
10335 bbr->r_state = TCPS_CLOSE_WAIT;
10336 bbr->r_substate = bbr_do_close_wait;
10338 case TCPS_FIN_WAIT_1:
10339 bbr->r_state = TCPS_FIN_WAIT_1;
10340 bbr->r_substate = bbr_do_fin_wait_1;
10343 bbr->r_state = TCPS_CLOSING;
10344 bbr->r_substate = bbr_do_closing;
10346 case TCPS_LAST_ACK:
10347 bbr->r_state = TCPS_LAST_ACK;
10348 bbr->r_substate = bbr_do_lastack;
10350 case TCPS_FIN_WAIT_2:
10351 bbr->r_state = TCPS_FIN_WAIT_2;
10352 bbr->r_substate = bbr_do_fin_wait_2;
10356 case TCPS_TIME_WAIT:
10363 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10366 * Now what state are we going into now? Is there adjustments
10369 int32_t old_state, old_gain;
10371 old_state = bbr_state_val(bbr);
10372 old_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
10373 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10374 /* Save the lowest srtt we saw in our end of the sub-state */
10375 bbr->rc_hit_state_1 = 0;
10376 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10377 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10379 bbr->rc_bbr_substate++;
10380 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10381 /* Cycle back to first state-> gain */
10382 bbr->rc_bbr_substate = 0;
10384 if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10386 * We enter the gain(5/4) cycle (possibly less if
10387 * shallow buffer detection is enabled)
10389 if (bbr->skip_gain) {
10391 * Hardware pacing has set our rate to
10392 * the max and limited our b/w just
10393 * do level i.e. no gain.
10395 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10396 } else if (bbr->gain_is_limited &&
10397 bbr->bbr_hdrw_pacing &&
10400 * We can't gain above the hardware pacing
10401 * rate which is less than our rate + the gain
10402 * calculate the gain needed to reach the hardware
10405 uint64_t bw, rate, gain_calc;
10407 bw = bbr_get_bw(bbr);
10408 rate = bbr->r_ctl.crte->rate;
10410 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10411 gain_calc = (rate * BBR_UNIT) / bw;
10412 if (gain_calc < BBR_UNIT)
10413 gain_calc = BBR_UNIT;
10414 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10416 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10419 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10420 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10421 bbr->r_ctl.rc_bbr_state_atflight = cts;
10423 bbr->r_ctl.rc_bbr_state_atflight = 0;
10424 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10425 bbr->rc_hit_state_1 = 1;
10426 bbr->r_ctl.rc_exta_time_gd = 0;
10427 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10428 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10429 if (bbr_state_drain_2_tar) {
10430 bbr->r_ctl.rc_bbr_state_atflight = 0;
10432 bbr->r_ctl.rc_bbr_state_atflight = cts;
10433 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10435 /* All other cycles hit here 2-7 */
10436 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10437 if (bbr_sub_drain_slam_cwnd &&
10438 (bbr->rc_use_google == 0) &&
10439 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10440 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10441 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10443 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10444 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10445 bbr_get_rtt(bbr, BBR_RTT_PROP));
10447 bbr->r_ctl.rc_exta_time_gd = 0;
10448 if (bbr->r_ctl.rc_exta_time_gd) {
10449 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10450 /* Now chop up the time for each state (div by 7) */
10451 bbr->r_ctl.rc_level_state_extra /= 7;
10452 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10453 /* Add a randomization */
10454 bbr_randomize_extra_state_time(bbr);
10458 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10459 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10461 if (bbr->rc_use_google) {
10462 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10464 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10465 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10467 bbr_log_type_statechange(bbr, cts, line);
10469 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10472 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10473 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10474 counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10476 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10479 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10480 bbr_set_state_target(bbr, __LINE__);
10481 if (bbr_sub_drain_slam_cwnd &&
10482 (bbr->rc_use_google == 0) &&
10483 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10484 /* Slam down the cwnd */
10485 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10486 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10487 if (bbr_sub_drain_app_limit) {
10488 /* Go app limited if we are on a long drain */
10489 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10490 ctf_flight_size(bbr->rc_tp,
10491 (bbr->r_ctl.rc_sacked +
10492 bbr->r_ctl.rc_lost_bytes)));
10494 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10496 if (bbr->rc_lt_use_bw) {
10497 /* In policed mode we clamp pacing_gain to BBR_UNIT */
10498 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10500 /* Google changes TSO size every cycle */
10501 if (bbr->rc_use_google)
10502 tcp_bbr_tso_size_check(bbr, cts);
10503 bbr->r_ctl.gain_epoch = cts;
10504 bbr->r_ctl.rc_bbr_state_time = cts;
10505 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10509 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10511 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10512 (google_allow_early_out == 1) &&
10513 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10514 /* We have reached out target flight size possibly early */
10517 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10520 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10522 * Must be a rttProp movement forward before
10523 * we can change states.
10527 if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10529 * The needed time has passed but for
10530 * the gain cycle extra rules apply:
10531 * 1) If we have seen loss, we exit
10532 * 2) If we have not reached the target
10533 * we stay in GAIN (gain-to-target).
10535 if (google_consider_lost && losses)
10537 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10542 /* For gain we must reach our target, all others last 1 rttProp */
10543 bbr_substate_change(bbr, cts, __LINE__, 1);
10547 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10549 uint32_t flight, bbr_cur_cycle_time;
10551 if (bbr->rc_use_google) {
10552 bbr_set_probebw_google_gains(bbr, cts, losses);
10557 * Never alow cts to be 0 we
10558 * do this so we can judge if
10559 * we have set a timestamp.
10563 if (bbr_state_is_pkt_epoch)
10564 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10566 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10568 if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10569 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10570 flight = ctf_flight_size(bbr->rc_tp,
10571 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10572 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10573 /* Keep it slam down */
10574 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10575 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10576 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10578 if (bbr_sub_drain_app_limit) {
10579 /* Go app limited if we are on a long drain */
10580 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10583 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10584 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10585 (flight >= bbr->r_ctl.flightsize_at_drain))) {
10587 * Still here after the same time as
10588 * the gain. We need to drain harder
10589 * for the next srtt. Reduce by a set amount
10590 * the gain drop is capped at DRAIN states
10593 bbr->r_ctl.flightsize_at_drain = flight;
10594 if (bbr_drain_drop_mul &&
10595 bbr_drain_drop_div &&
10596 (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10597 /* Use your specific drop value (def 4/5 = 20%) */
10598 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10599 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10601 /* You get drop of 20% */
10602 bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10603 bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10605 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10606 /* Reduce our gain again to the bottom */
10607 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10609 bbr_log_exit_gain(bbr, cts, 4);
10611 * Extend out so we wait another
10612 * epoch before dropping again.
10614 bbr->r_ctl.gain_epoch = cts;
10616 if (flight <= bbr->r_ctl.rc_target_at_state) {
10617 if (bbr_sub_drain_slam_cwnd &&
10618 (bbr->rc_use_google == 0) &&
10619 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10620 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10621 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10623 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10624 bbr_log_exit_gain(bbr, cts, 3);
10628 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10629 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10632 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10633 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >=
10634 bbr->rc_tp->snd_wnd)) {
10635 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10636 bbr_log_exit_gain(bbr, cts, 2);
10640 * We fall through and return always one of two things has
10642 * 1) We are still not at target
10644 * 2) We reached the target and set rc_bbr_state_atflight
10645 * which means we no longer hit this block
10646 * next time we are called.
10651 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10653 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10654 /* Less than a full time-period has passed */
10657 if (bbr->r_ctl.rc_level_state_extra &&
10658 (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10659 ((cts - bbr->r_ctl.rc_bbr_state_time) <
10660 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10661 /* Less than a full time-period + extra has passed */
10664 if (bbr_gain_gets_extra_too &&
10665 bbr->r_ctl.rc_level_state_extra &&
10666 (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10667 ((cts - bbr->r_ctl.rc_bbr_state_time) <
10668 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10669 /* Less than a full time-period + extra has passed */
10672 bbr_substate_change(bbr, cts, __LINE__, 1);
10676 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10680 if (bbr->rc_use_google) {
10681 /* Google just uses the cwnd target */
10682 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10684 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10685 bbr->r_ctl.rc_pace_max_segs);
10686 /* Get the base cwnd with gain rounded to a mss */
10687 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10689 /* Make sure it is within our min */
10690 if (tar < get_min_cwnd(bbr))
10691 return (get_min_cwnd(bbr));
10697 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10699 uint32_t tar, meth;
10701 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10702 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10703 /* Special case using old probe-rtt method */
10704 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10707 /* Non-probe-rtt case and reduced probe-rtt */
10708 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10709 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10710 /* For gain cycle we use the hptsi gain */
10711 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10713 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10715 * If configured, or for google all other states
10718 tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10722 * Or we set a target based on the pacing gain
10723 * for non-google mode and default (non-configured).
10724 * Note we don't set a target goal below drain (192).
10726 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) {
10727 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10730 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10735 bbr_log_set_of_state_target(bbr, tar, line, meth);
10736 bbr->r_ctl.rc_target_at_state = tar;
10740 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10742 /* Change to probe_rtt */
10745 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10746 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10747 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10748 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10749 + bbr->r_ctl.rc_delivered);
10750 /* Setup so we force feed the filter */
10751 if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10752 bbr->rc_prtt_set_ts = 1;
10753 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10754 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10755 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10757 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10758 bbr->r_ctl.rc_rtt_shrinks = cts;
10759 bbr->r_ctl.last_in_probertt = cts;
10760 bbr->r_ctl.rc_probertt_srttchktim = cts;
10761 bbr->r_ctl.rc_bbr_state_time = cts;
10762 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10763 /* We need to force the filter to update */
10765 if ((bbr_sub_drain_slam_cwnd) &&
10766 bbr->rc_hit_state_1 &&
10767 (bbr->rc_use_google == 0) &&
10768 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10769 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10770 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10772 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10773 /* Update the lost */
10774 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10775 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10776 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */
10777 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10778 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10779 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10780 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10781 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10782 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10785 * We bring it down slowly by using a hptsi gain that is
10786 * probably 75%. This will slowly float down our outstanding
10787 * without tampering with the cwnd.
10789 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10790 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10791 bbr_set_state_target(bbr, __LINE__);
10792 if (bbr_prtt_slam_cwnd &&
10793 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10794 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10795 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10798 if (ctf_flight_size(bbr->rc_tp,
10799 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10800 bbr->r_ctl.rc_target_at_state) {
10801 /* We are at target */
10802 bbr->r_ctl.rc_bbr_enters_probertt = cts;
10804 /* We need to come down to reach target before our time begins */
10805 bbr->r_ctl.rc_bbr_enters_probertt = 0;
10807 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10808 BBR_STAT_INC(bbr_enter_probertt);
10809 bbr_log_exit_gain(bbr, cts, 0);
10810 bbr_log_type_statechange(bbr, cts, line);
10814 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10817 * Sanity check on probe-rtt intervals.
10818 * In crazy situations where we are competing
10819 * against new-reno flows with huge buffers
10820 * our rtt-prop interval could come to dominate
10821 * things if we can't get through a full set
10822 * of cycles, we need to adjust it.
10824 if (bbr_can_adjust_probertt &&
10825 (bbr->rc_use_google == 0)) {
10827 uint32_t cur_rttp, fval, newval, baseval;
10829 /* Are we to small and go into probe-rtt to often? */
10830 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10831 cur_rttp = roundup(baseval, USECS_IN_SECOND);
10832 fval = bbr_filter_len_sec * USECS_IN_SECOND;
10833 if (bbr_is_ratio == 0) {
10834 if (fval > bbr_rtt_probe_limit)
10835 newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10841 mul = fval / bbr_rtt_probe_limit;
10842 newval = cur_rttp * mul;
10844 if (cur_rttp > bbr->r_ctl.rc_probertt_int) {
10845 bbr->r_ctl.rc_probertt_int = cur_rttp;
10846 reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10850 * No adjustments were made
10851 * do we need to shrink it?
10853 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
10854 if (cur_rttp <= bbr_rtt_probe_limit) {
10856 * Things have calmed down lets
10857 * shrink all the way to default
10859 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10860 reset_time_small(&bbr->r_ctl.rc_rttprop,
10861 (bbr_filter_len_sec * USECS_IN_SECOND));
10862 cur_rttp = bbr_rtt_probe_limit;
10863 newval = (bbr_filter_len_sec * USECS_IN_SECOND);
10867 * Well does some adjustment make sense?
10869 if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
10870 /* We can reduce interval time some */
10871 bbr->r_ctl.rc_probertt_int = cur_rttp;
10872 reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10879 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
10884 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
10886 /* Exit probe-rtt */
10888 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
10889 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10890 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10892 bbr_log_exit_gain(bbr, cts, 1);
10893 bbr->rc_hit_state_1 = 0;
10894 bbr->r_ctl.rc_rtt_shrinks = cts;
10895 bbr->r_ctl.last_in_probertt = cts;
10896 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
10897 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10898 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
10899 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
10900 bbr->r_ctl.rc_delivered);
10901 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10904 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10905 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10907 if (bbr->rc_filled_pipe) {
10908 /* Switch to probe_bw */
10909 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
10910 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
10911 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10912 bbr_substate_change(bbr, cts, __LINE__, 0);
10913 bbr_log_type_statechange(bbr, cts, __LINE__);
10915 /* Back to startup */
10916 bbr->rc_bbr_state = BBR_STATE_STARTUP;
10917 bbr->r_ctl.rc_bbr_state_time = cts;
10919 * We don't want to give a complete free 3
10920 * measurements until we exit, so we use
10921 * the number of pe's we were in probe-rtt
10922 * to add to the startup_epoch. That way
10923 * we will still retain the old state.
10925 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
10926 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10927 /* Make sure to use the lower pg when shifting back in */
10928 if (bbr->r_ctl.rc_lost &&
10929 bbr_use_lower_gain_in_startup &&
10930 (bbr->rc_use_google == 0))
10931 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
10933 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
10934 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
10935 /* Probably not needed but set it anyway */
10936 bbr_set_state_target(bbr, __LINE__);
10937 bbr_log_type_statechange(bbr, cts, __LINE__);
10938 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10939 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
10941 bbr_check_probe_rtt_limits(bbr, cts);
10944 static int32_t inline
10945 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
10947 if ((bbr->rc_past_init_win == 1) &&
10948 (bbr->rc_in_persist == 0) &&
10949 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
10952 if (bbr_can_force_probertt &&
10953 (bbr->rc_in_persist == 0) &&
10954 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
10955 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
10962 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch)
10964 uint64_t btlbw, gain;
10965 if (pkt_epoch == 0) {
10967 * Need to be on a pkt-epoch to continue.
10971 btlbw = bbr_get_full_bw(bbr);
10972 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
10973 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
10974 if (btlbw >= gain) {
10975 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
10976 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10977 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
10978 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
10980 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
10982 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10983 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
10987 static int32_t inline
10988 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
10990 /* Have we gained 25% in the last 3 packet based epoch's? */
10991 uint64_t btlbw, gain;
10993 int delta, rtt_gain;
10995 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
10996 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
10998 * This qualifies as a RTT_PROBE session since we drop the
10999 * data outstanding to nothing and waited more than
11000 * bbr_rtt_probe_time.
11002 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11003 bbr_set_reduced_rtt(bbr, cts, __LINE__);
11005 if (bbr_should_enter_probe_rtt(bbr, cts)) {
11006 bbr_enter_probe_rtt(bbr, cts, __LINE__);
11009 if (bbr->rc_use_google)
11010 return (bbr_google_startup(bbr, cts, pkt_epoch));
11012 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11013 (bbr_use_lower_gain_in_startup)) {
11014 /* Drop to a lower gain 1.5 x since we saw loss */
11015 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11017 if (pkt_epoch == 0) {
11019 * Need to be on a pkt-epoch to continue.
11023 if (bbr_rtt_gain_thresh) {
11025 * Do we allow a flow to stay
11026 * in startup with no loss and no
11027 * gain in rtt over a set threshold?
11029 if (bbr->r_ctl.rc_pkt_epoch_rtt &&
11030 bbr->r_ctl.startup_last_srtt &&
11031 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
11032 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
11033 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
11036 if ((bbr->r_ctl.startup_last_srtt == 0) ||
11037 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
11038 /* First time or new lower value */
11039 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
11041 if ((bbr->r_ctl.rc_lost == 0) &&
11042 (rtt_gain < bbr_rtt_gain_thresh)) {
11044 * No loss, and we are under
11045 * our gain threhold for
11048 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11049 bbr->r_ctl.rc_bbr_last_startup_epoch++;
11050 bbr_log_startup_event(bbr, cts, rtt_gain,
11051 delta, bbr->r_ctl.startup_last_srtt, 10);
11055 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
11056 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
11057 (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
11059 * We only assess if we have a new measurment when
11060 * we have no loss and are not in recovery.
11061 * Drag up by one our last_startup epoch so we will hold
11062 * the number of non-gain we have already accumulated.
11064 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11065 bbr->r_ctl.rc_bbr_last_startup_epoch++;
11066 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11067 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
11070 /* Case where we reduced the lost (bad retransmit) */
11071 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
11072 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11073 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
11074 btlbw = bbr_get_full_bw(bbr);
11075 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
11076 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11077 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11079 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11080 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11082 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
11083 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11084 if (btlbw >= gain) {
11085 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11086 /* Update the lost so we won't exit in next set of tests */
11087 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11088 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11089 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11091 if ((bbr->rc_loss_exit &&
11092 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11093 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11094 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11096 * If we had no gain, we had loss and that loss was above
11097 * our threshould, the rwnd is not constrained, and we have
11098 * had at least 3 packet epochs exit. Note that this is
11099 * switched off by sysctl. Google does not do this by the
11102 if ((ctf_flight_size(bbr->rc_tp,
11103 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11104 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11106 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11107 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11109 /* Just record an updated loss value */
11110 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11111 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11112 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11115 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11116 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11118 /* Return 1 to exit the startup state. */
11121 /* Stay in startup */
11122 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11123 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11128 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11131 * A tick occurred in the rtt epoch do we need to do anything?
11133 #ifdef BBR_INVARIANTS
11134 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11135 (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11136 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11137 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11138 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11140 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11143 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11144 /* Do we exit the startup state? */
11145 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11148 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11149 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11150 bbr->rc_filled_pipe = 1;
11151 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11152 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11153 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11154 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11157 if (bbr->rc_no_pacing)
11158 bbr->rc_no_pacing = 0;
11159 bbr->r_ctl.rc_bbr_state_time = cts;
11160 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11161 bbr->rc_bbr_state = BBR_STATE_DRAIN;
11162 bbr_set_state_target(bbr, __LINE__);
11163 if ((bbr->rc_use_google == 0) &&
11164 bbr_slam_cwnd_in_main_drain) {
11165 /* Here we don't have to worry about probe-rtt */
11166 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11167 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11168 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11170 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11171 bbr_log_type_statechange(bbr, cts, __LINE__);
11172 if (ctf_flight_size(bbr->rc_tp,
11173 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11174 bbr->r_ctl.rc_target_at_state) {
11176 * Switch to probe_bw if we are already
11179 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11180 bbr_substate_change(bbr, cts, __LINE__, 0);
11181 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11182 bbr_log_type_statechange(bbr, cts, __LINE__);
11185 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11190 inflight = ctf_flight_size(tp,
11191 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11192 if (inflight >= bbr->r_ctl.rc_target_at_state) {
11193 /* We have reached a flight of the cwnd target */
11194 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11195 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11196 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11197 bbr_set_state_target(bbr, __LINE__);
11199 * Rig it so we don't do anything crazy and
11200 * start fresh with a new randomization.
11202 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11203 bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11204 bbr_substate_change(bbr, cts, __LINE__, 1);
11206 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11207 /* Has in-flight reached the bdp (or less)? */
11212 inflight = ctf_flight_size(tp,
11213 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11214 if ((bbr->rc_use_google == 0) &&
11215 bbr_slam_cwnd_in_main_drain &&
11216 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11218 * Here we don't have to worry about probe-rtt
11219 * re-slam it, but keep it slammed down.
11221 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11222 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11224 if (inflight <= bbr->r_ctl.rc_target_at_state) {
11225 /* We have drained */
11226 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11227 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11228 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11231 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11232 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11234 if ((bbr->rc_use_google == 0) &&
11235 bbr_slam_cwnd_in_main_drain &&
11236 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11237 /* Restore the cwnd */
11238 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11239 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11241 /* Setup probe-rtt has being done now RRS-HERE */
11242 bbr->r_ctl.rc_rtt_shrinks = cts;
11243 bbr->r_ctl.last_in_probertt = cts;
11244 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11245 /* Randomly pick a sub-state */
11246 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11247 bbr_substate_change(bbr, cts, __LINE__, 0);
11248 bbr_log_type_statechange(bbr, cts, __LINE__);
11250 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11253 flight = ctf_flight_size(bbr->rc_tp,
11254 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11255 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11256 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11257 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11259 * We must keep cwnd at the desired MSS.
11261 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11262 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11263 } else if ((bbr_prtt_slam_cwnd) &&
11264 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11266 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11267 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11269 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11270 /* Has outstanding reached our target? */
11271 if (flight <= bbr->r_ctl.rc_target_at_state) {
11272 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11273 bbr->r_ctl.rc_bbr_enters_probertt = cts;
11274 /* If time is exactly 0, be 1usec off */
11275 if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11276 bbr->r_ctl.rc_bbr_enters_probertt = 1;
11277 if (bbr->rc_use_google == 0) {
11279 * Restore any lowering that as occurred to
11282 if (bbr->r_ctl.bbr_rttprobe_gain_val)
11283 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11285 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11288 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11289 (bbr->rc_use_google == 0) &&
11290 bbr->r_ctl.bbr_rttprobe_gain_val &&
11291 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11292 (flight >= bbr->r_ctl.flightsize_at_drain))) {
11294 * We have doddled with our current hptsi
11295 * gain an srtt and have still not made it
11296 * to target, or we have increased our flight.
11297 * Lets reduce the gain by xx%
11298 * flooring the reduce at DRAIN (based on
11303 bbr->r_ctl.flightsize_at_drain = flight;
11304 bbr->r_ctl.rc_probertt_srttchktim = cts;
11305 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11306 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11307 /* Reduce our gain again */
11308 bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11309 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11310 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11311 /* one more chance before we give up */
11312 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11313 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11315 /* At the very bottom */
11316 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11320 if (bbr->r_ctl.rc_bbr_enters_probertt &&
11321 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11322 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11323 /* Time to exit probe RTT normally */
11324 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11326 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11327 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11328 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11330 * This qualifies as a RTT_PROBE session since we
11331 * drop the data outstanding to nothing and waited
11332 * more than bbr_rtt_probe_time.
11334 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11335 bbr_set_reduced_rtt(bbr, cts, __LINE__);
11337 if (bbr_should_enter_probe_rtt(bbr, cts)) {
11338 bbr_enter_probe_rtt(bbr, cts, __LINE__);
11340 bbr_set_probebw_gains(bbr, cts, losses);
11346 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11350 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11351 bbr_set_epoch(bbr, cts, line);
11352 /* At each epoch doe lt bw sampling */
11355 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11359 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11360 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11361 int32_t nxt_pkt, struct timeval *tv)
11363 int32_t thflags, retval;
11364 uint32_t cts, lcts;
11367 struct tcp_bbr *bbr;
11368 struct bbr_sendmap *rsm;
11369 struct timeval ltv;
11370 int32_t did_out = 0;
11371 int32_t in_recovery;
11373 int32_t prev_state;
11376 nsegs = max(1, m->m_pkthdr.lro_nsegs);
11377 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11378 /* add in our stats */
11379 kern_prefetch(bbr, &prev_state);
11381 thflags = th->th_flags;
11383 * If this is either a state-changing packet or current state isn't
11384 * established, we require a write lock on tcbinfo. Otherwise, we
11385 * allow the tcbinfo to be in either alocked or unlocked, as the
11386 * caller may have unnecessarily acquired a write lock due to a
11389 INP_WLOCK_ASSERT(tp->t_inpcb);
11390 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11392 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11395 tp->t_rcvtime = ticks;
11397 * Unscale the window into a 32-bit value. For the SYN_SENT state
11398 * the scale is zero.
11400 tiwin = th->th_win << tp->snd_scale;
11402 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11405 if (m->m_flags & M_TSTMP) {
11406 /* Prefer the hardware timestamp if present */
11407 struct timespec ts;
11409 mbuf_tstmp2timespec(m, &ts);
11410 bbr->rc_tv.tv_sec = ts.tv_sec;
11411 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11412 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11413 } else if (m->m_flags & M_TSTMP_LRO) {
11414 /* Next the arrival timestamp */
11415 struct timespec ts;
11417 mbuf_tstmp2timespec(m, &ts);
11418 bbr->rc_tv.tv_sec = ts.tv_sec;
11419 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11420 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11423 * Ok just get the current time.
11425 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11428 * Parse options on any incoming segment.
11430 tcp_dooptions(&to, (u_char *)(th + 1),
11431 (th->th_off << 2) - sizeof(struct tcphdr),
11432 (thflags & TH_SYN) ? TO_SYN : 0);
11435 * If timestamps were negotiated during SYN/ACK and a
11436 * segment without a timestamp is received, silently drop
11437 * the segment, unless it is a RST segment or missing timestamps are
11439 * See section 3.2 of RFC 7323.
11441 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
11442 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
11445 goto done_with_input;
11448 * If echoed timestamp is later than the current time, fall back to
11449 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
11450 * were used when this connection was established.
11452 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11453 to.to_tsecr -= tp->ts_offset;
11454 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11458 * If its the first time in we need to take care of options and
11459 * verify we can do SACK for rack!
11461 if (bbr->r_state == 0) {
11463 * Process options only when we get SYN/ACK back. The SYN
11464 * case for incoming connections is handled in tcp_syncache.
11465 * According to RFC1323 the window field in a SYN (i.e., a
11466 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11467 * this is traditional behavior, may need to be cleaned up.
11469 if (bbr->rc_inp == NULL) {
11470 bbr->rc_inp = tp->t_inpcb;
11473 * We need to init rc_inp here since its not init'd when
11474 * bbr_init is called
11476 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11477 if ((to.to_flags & TOF_SCALE) &&
11478 (tp->t_flags & TF_REQ_SCALE)) {
11479 tp->t_flags |= TF_RCVD_SCALE;
11480 tp->snd_scale = to.to_wscale;
11482 tp->t_flags &= ~TF_REQ_SCALE;
11484 * Initial send window. It will be updated with the
11485 * next incoming segment to the scaled value.
11487 tp->snd_wnd = th->th_win;
11488 if ((to.to_flags & TOF_TS) &&
11489 (tp->t_flags & TF_REQ_TSTMP)) {
11490 tp->t_flags |= TF_RCVD_TSTMP;
11491 tp->ts_recent = to.to_tsval;
11492 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11494 tp->t_flags &= ~TF_REQ_TSTMP;
11495 if (to.to_flags & TOF_MSS)
11496 tcp_mss(tp, to.to_mss);
11497 if ((tp->t_flags & TF_SACK_PERMIT) &&
11498 (to.to_flags & TOF_SACKPERM) == 0)
11499 tp->t_flags &= ~TF_SACK_PERMIT;
11500 if (IS_FASTOPEN(tp->t_flags)) {
11501 if (to.to_flags & TOF_FASTOPEN) {
11504 if (to.to_flags & TOF_MSS)
11507 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11511 tcp_fastopen_update_cache(tp, mss,
11512 to.to_tfo_len, to.to_tfo_cookie);
11514 tcp_fastopen_disable_path(tp);
11518 * At this point we are at the initial call. Here we decide
11519 * if we are doing RACK or not. We do this by seeing if
11520 * TF_SACK_PERMIT is set, if not rack is *not* possible and
11521 * we switch to the default code.
11523 if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11525 tcp_switch_back_to_default(tp);
11526 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11531 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11532 tcp_set_hpts(tp->t_inpcb);
11533 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11535 if (thflags & TH_ACK) {
11536 /* Track ack types */
11537 if (to.to_flags & TOF_SACK)
11538 BBR_STAT_INC(bbr_acks_with_sacks);
11540 BBR_STAT_INC(bbr_plain_acks);
11543 * This is the one exception case where we set the rack state
11544 * always. All other times (timers etc) we must have a rack-state
11545 * set (so we assure we have done the checks above for SACK).
11547 if (thflags & TH_FIN)
11548 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11549 if (bbr->r_state != tp->t_state)
11550 bbr_set_state(tp, bbr, tiwin);
11552 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11553 kern_prefetch(rsm, &prev_state);
11554 prev_state = bbr->r_state;
11555 bbr->rc_ack_was_delayed = 0;
11556 lost = bbr->r_ctl.rc_lost;
11557 bbr->rc_is_pkt_epoch_now = 0;
11558 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11559 /* Get the real time into lcts and figure the real delay */
11560 lcts = tcp_get_usecs(<v);
11561 if (TSTMP_GT(lcts, cts)) {
11562 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11563 bbr->rc_ack_was_delayed = 1;
11564 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11565 bbr->r_ctl.highest_hdwr_delay))
11566 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11568 bbr->r_ctl.rc_ack_hdwr_delay = 0;
11569 bbr->rc_ack_was_delayed = 0;
11572 bbr->r_ctl.rc_ack_hdwr_delay = 0;
11573 bbr->rc_ack_was_delayed = 0;
11575 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11576 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11579 goto done_with_input;
11582 * If a segment with the ACK-bit set arrives in the SYN-SENT state
11583 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11585 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11586 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11587 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11588 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11591 in_recovery = IN_RECOVERY(tp->t_flags);
11592 if (tiwin > bbr->r_ctl.rc_high_rwnd)
11593 bbr->r_ctl.rc_high_rwnd = tiwin;
11594 #ifdef BBR_INVARIANTS
11595 if ((tp->t_inpcb->inp_flags & INP_DROPPED) ||
11596 (tp->t_inpcb->inp_flags2 & INP_FREED)) {
11597 panic("tp:%p bbr:%p given a dropped inp:%p",
11598 tp, bbr, tp->t_inpcb);
11601 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11602 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11603 bbr->rtt_valid = 0;
11604 if (to.to_flags & TOF_TS) {
11605 bbr->rc_ts_valid = 1;
11606 bbr->r_ctl.last_inbound_ts = to.to_tsval;
11608 bbr->rc_ts_valid = 0;
11609 bbr->r_ctl.last_inbound_ts = 0;
11611 retval = (*bbr->r_substate) (m, th, so,
11612 tp, &to, drop_hdrlen,
11613 tlen, tiwin, thflags, nxt_pkt, iptos);
11614 #ifdef BBR_INVARIANTS
11615 if ((retval == 0) &&
11616 (tp->t_inpcb == NULL)) {
11617 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11618 retval, tp, prev_state);
11622 BBR_STAT_INC(bbr_rlock_left_ret0);
11624 BBR_STAT_INC(bbr_rlock_left_ret1);
11627 * If retval is 1 the tcb is unlocked and most likely the tp
11630 INP_WLOCK_ASSERT(tp->t_inpcb);
11631 tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11632 if (bbr->rc_is_pkt_epoch_now)
11633 bbr_set_pktepoch(bbr, cts, __LINE__);
11634 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11635 if (nxt_pkt == 0) {
11636 if (bbr->r_wanted_output != 0) {
11637 bbr->rc_output_starts_timer = 0;
11639 (void)tp->t_fb->tfb_tcp_output(tp);
11641 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11643 if ((nxt_pkt == 0) &&
11644 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11645 (SEQ_GT(tp->snd_max, tp->snd_una) ||
11646 (tp->t_flags & TF_DELACK) ||
11647 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11648 (tp->t_state <= TCPS_CLOSING)))) {
11650 * We could not send (probably in the hpts but
11651 * stopped the timer)?
11653 if ((tp->snd_max == tp->snd_una) &&
11654 ((tp->t_flags & TF_DELACK) == 0) &&
11655 (bbr->rc_inp->inp_in_hpts) &&
11656 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11658 * keep alive not needed if we are hptsi
11663 if (bbr->rc_inp->inp_in_hpts) {
11664 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
11665 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11666 (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11669 del = lcts - bbr->rc_pacer_started;
11670 if (bbr->r_ctl.rc_last_delay_val > del) {
11671 BBR_STAT_INC(bbr_force_timer_start);
11672 bbr->r_ctl.rc_last_delay_val -= del;
11673 bbr->rc_pacer_started = lcts;
11676 bbr->r_ctl.rc_last_delay_val = 0;
11677 BBR_STAT_INC(bbr_force_output);
11678 (void)tp->t_fb->tfb_tcp_output(tp);
11682 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11685 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11686 /* Do we have the correct timer running? */
11687 bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11689 /* Do we have a new state */
11690 if (bbr->r_state != tp->t_state)
11691 bbr_set_state(tp, bbr, tiwin);
11693 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11695 bbr->r_wanted_output = 0;
11696 #ifdef BBR_INVARIANTS
11697 if (tp->t_inpcb == NULL) {
11698 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11700 retval, tp, prev_state);
11708 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11709 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11714 /* First lets see if we have old packets */
11715 if (tp->t_in_pkt) {
11716 if (ctf_do_queued_segments(so, tp, 1)) {
11721 if (m->m_flags & M_TSTMP_LRO) {
11722 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11723 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11725 /* Should not be should we kassert instead? */
11726 tcp_get_usecs(&tv);
11728 retval = bbr_do_segment_nounlock(m, th, so, tp,
11729 drop_hdrlen, tlen, iptos, 0, &tv);
11731 INP_WUNLOCK(tp->t_inpcb);
11736 * Return how much data can be sent without violating the
11740 static inline uint32_t
11741 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11742 uint32_t avail, int32_t sb_offset, uint32_t cts)
11746 if (ctf_outstanding(tp) >= tp->snd_wnd) {
11747 /* We never want to go over our peers rcv-window */
11752 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11753 if (flight >= sendwin) {
11755 * We have in flight what we are allowed by cwnd (if
11756 * it was rwnd blocking it would have hit above out
11761 len = sendwin - flight;
11762 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11763 /* We would send too much (beyond the rwnd) */
11764 len = tp->snd_wnd - ctf_outstanding(tp);
11766 if ((len + sb_offset) > avail) {
11768 * We don't have that much in the SB, how much is
11771 len = avail - sb_offset;
11778 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11780 #ifdef NETFLIX_STATS
11781 KMOD_TCPSTAT_INC(tcps_sndpack_error);
11782 KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len);
11787 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11790 bbr_do_error_accounting(tp, bbr, rsm, len, error);
11794 if (rsm->r_flags & BBR_TLP) {
11796 * TLP should not count in retran count, but in its
11799 #ifdef NETFLIX_STATS
11800 KMOD_TCPSTAT_INC(tcps_tlpresends);
11801 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11805 tp->t_sndrexmitpack++;
11806 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11807 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11809 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11814 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11817 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11818 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11819 /* Non probe_bw log in 1, 2, or 4. */
11820 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11823 * Log our probe state 3, and log also 5-13 to show
11824 * us the recovery sub-state for the send. This
11825 * means that 3 == (5+6+7+8+9+10+11+12+13)
11827 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11828 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11830 /* Place in both 16's the totals of retransmitted */
11831 counter_u64_add(bbr_state_lost[16], len);
11832 counter_u64_add(bbr_state_resend[16], len);
11833 /* Place in 17's the total sent */
11834 counter_u64_add(bbr_state_resend[17], len);
11835 counter_u64_add(bbr_state_lost[17], len);
11839 KMOD_TCPSTAT_INC(tcps_sndpack);
11840 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11841 /* Place in 17's the total sent */
11842 counter_u64_add(bbr_state_resend[17], len);
11843 counter_u64_add(bbr_state_lost[17], len);
11845 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11852 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11854 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11856 * Limit the cwnd to not be above N x the target plus whats
11857 * is outstanding. The target is based on the current b/w
11862 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11863 target += ctf_outstanding(tp);
11864 target *= bbr_target_cwnd_mult_limit;
11865 if (tp->snd_cwnd > target)
11866 tp->snd_cwnd = target;
11867 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11872 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
11875 * "adv" is the amount we could increase the window, taking into
11876 * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
11882 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
11883 oldwin = (tp->rcv_adv - tp->rcv_nxt);
11887 /* We can't increase the window */
11894 * If the new window size ends up being the same as or less
11895 * than the old size when it is scaled, then don't force
11898 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
11901 if (adv >= (2 * maxseg) &&
11902 (adv >= (so->so_rcv.sb_hiwat / 4) ||
11903 recwin <= (so->so_rcv.sb_hiwat / 8) ||
11904 so->so_rcv.sb_hiwat <= 8 * maxseg)) {
11907 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
11913 * Return 0 on success and a errno on failure to send.
11914 * Note that a 0 return may not mean we sent anything
11915 * if the TCB was on the hpts. A non-zero return
11916 * does indicate the error we got from ip[6]_output.
11919 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
11924 uint32_t recwin, sendwin;
11926 int32_t flags, abandon, error = 0;
11927 struct tcp_log_buffer *lgb = NULL;
11930 uint32_t if_hw_tsomaxsegcount = 0;
11931 uint32_t if_hw_tsomaxsegsize = 0;
11932 uint32_t if_hw_tsomax = 0;
11933 struct ip *ip = NULL;
11935 struct ipovly *ipov = NULL;
11937 struct tcp_bbr *bbr;
11939 struct udphdr *udp = NULL;
11940 u_char opt[TCP_MAXOLEN];
11941 unsigned ipoptlen, optlen, hdrlen;
11944 uint32_t delay_calc=0;
11945 uint8_t doing_tlp = 0;
11946 uint8_t local_options;
11947 #ifdef BBR_INVARIANTS
11948 uint8_t doing_retran_from = 0;
11949 uint8_t picked_up_retran = 0;
11951 uint8_t wanted_cookie = 0;
11952 uint8_t more_to_rxt=0;
11953 int32_t prefetch_so_done = 0;
11954 int32_t prefetch_rsm = 0;
11955 uint32_t what_we_can = 0;
11956 uint32_t tot_len = 0;
11957 uint32_t rtr_cnt = 0;
11958 uint32_t maxseg, pace_max_segs, p_maxseg;
11959 int32_t csum_flags;
11961 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11962 unsigned ipsec_optlen = 0;
11965 volatile int32_t sack_rxmit;
11966 struct bbr_sendmap *rsm = NULL;
11971 struct sockbuf *sb;
11972 uint32_t hpts_calling;
11974 struct ip6_hdr *ip6 = NULL;
11977 uint8_t app_limited = BBR_JR_SENT_DATA;
11978 uint8_t filled_all = 0;
11979 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11980 /* We take a cache hit here */
11981 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
11982 cts = tcp_tv_to_usectick(&bbr->rc_tv);
11984 so = inp->inp_socket;
11986 if (sb->sb_flags & SB_TLS_IFNET)
11990 kern_prefetch(sb, &maxseg);
11991 maxseg = tp->t_maxseg - bbr->rc_last_options;
11992 if (bbr_minseg(bbr) < maxseg) {
11993 tcp_bbr_tso_size_check(bbr, cts);
11995 /* Remove any flags that indicate we are pacing on the inp */
11996 pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
11997 p_maxseg = min(maxseg, pace_max_segs);
11998 INP_WLOCK_ASSERT(inp);
12000 if (tp->t_flags & TF_TOE)
12001 return (tcp_offload_output(tp));
12005 if (bbr->r_state) {
12006 /* Use the cache line loaded if possible */
12007 isipv6 = bbr->r_is_v6;
12009 isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
12012 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
12013 inp->inp_in_hpts) {
12015 * We are on the hpts for some timer but not hptsi output.
12016 * Possibly remove from the hpts so we can send/recv etc.
12018 if ((tp->t_flags & TF_ACKNOW) == 0) {
12020 * No immediate demand right now to send an ack, but
12021 * the user may have read, making room for new data
12022 * (a window update). If so we may want to cancel
12023 * whatever timer is running (KEEP/DEL-ACK?) and
12024 * continue to send out a window update. Or we may
12025 * have gotten more data into the socket buffer to
12028 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12029 (long)TCP_MAXWIN << tp->rcv_scale);
12030 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
12031 ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
12032 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
12033 (tp->snd_max - tp->snd_una))) {
12035 * Nothing new to send and no window update
12036 * is needed to send. Lets just return and
12037 * let the timer-run off.
12042 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12043 bbr_timer_cancel(bbr, __LINE__, cts);
12045 if (bbr->r_ctl.rc_last_delay_val) {
12046 /* Calculate a rough delay for early escape to sending */
12047 if (SEQ_GT(cts, bbr->rc_pacer_started))
12048 delay_calc = cts - bbr->rc_pacer_started;
12049 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12050 delay_calc -= bbr->r_ctl.rc_last_delay_val;
12054 /* Mark that we have called bbr_output(). */
12055 if ((bbr->r_timer_override) ||
12056 (tp->t_state < TCPS_ESTABLISHED)) {
12057 /* Timeouts or early states are exempt */
12058 if (inp->inp_in_hpts)
12059 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12060 } else if (inp->inp_in_hpts) {
12061 if ((bbr->r_ctl.rc_last_delay_val) &&
12062 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12065 * We were being paced for output and the delay has
12066 * already exceeded when we were supposed to be
12067 * called, lets go ahead and pull out of the hpts
12070 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
12071 bbr->r_ctl.rc_last_delay_val = 0;
12072 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12073 } else if (tp->t_state == TCPS_CLOSED) {
12074 bbr->r_ctl.rc_last_delay_val = 0;
12075 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12078 * On the hpts, you shall not pass! even if ACKNOW
12079 * is on, we will when the hpts fires, unless of
12080 * course we are overdue.
12082 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
12086 bbr->rc_cwnd_limited = 0;
12087 if (bbr->r_ctl.rc_last_delay_val) {
12088 /* recalculate the real delay and deal with over/under */
12089 if (SEQ_GT(cts, bbr->rc_pacer_started))
12090 delay_calc = cts - bbr->rc_pacer_started;
12093 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12094 /* Setup the delay which will be added in */
12095 delay_calc -= bbr->r_ctl.rc_last_delay_val;
12098 * We are early setup to adjust
12101 uint64_t merged_val;
12103 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
12104 bbr->r_agg_early_set = 1;
12105 if (bbr->r_ctl.rc_hptsi_agg_delay) {
12106 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
12107 /* Nope our previous late cancels out the early */
12108 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12109 bbr->r_agg_early_set = 0;
12110 bbr->r_ctl.rc_agg_early = 0;
12112 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12113 bbr->r_ctl.rc_hptsi_agg_delay = 0;
12116 merged_val = bbr->rc_pacer_started;
12118 merged_val |= bbr->r_ctl.rc_last_delay_val;
12119 bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12120 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12121 bbr->r_agg_early_set, 3);
12122 bbr->r_ctl.rc_last_delay_val = 0;
12123 BBR_STAT_INC(bbr_early);
12127 /* We were not delayed due to hptsi */
12128 if (bbr->r_agg_early_set)
12129 bbr->r_ctl.rc_agg_early = 0;
12130 bbr->r_agg_early_set = 0;
12135 * We had a hptsi delay which means we are falling behind on
12136 * sending at the expected rate. Calculate an extra amount
12137 * of data we can send, if any, to put us back on track.
12139 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12140 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12142 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12144 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12145 if ((tp->snd_una == tp->snd_max) &&
12146 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12149 * Ok we have been idle with nothing outstanding
12150 * we possibly need to start fresh with either a new
12151 * suite of states or a fast-ramp up.
12153 bbr_restart_after_idle(bbr,
12154 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12157 * Now was there a hptsi delay where we are behind? We only count
12158 * being behind if: a) We are not in recovery. b) There was a delay.
12159 * <and> c) We had room to send something.
12162 hpts_calling = inp->inp_hpts_calls;
12163 inp->inp_hpts_calls = 0;
12164 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12165 if (bbr_process_timers(tp, bbr, cts, hpts_calling)) {
12166 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12170 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12171 if (hpts_calling &&
12172 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12173 bbr->r_ctl.rc_last_delay_val = 0;
12175 bbr->r_timer_override = 0;
12176 bbr->r_wanted_output = 0;
12178 * For TFO connections in SYN_RECEIVED, only allow the initial
12179 * SYN|ACK and those sent by the retransmit timer.
12181 if (IS_FASTOPEN(tp->t_flags) &&
12182 ((tp->t_state == TCPS_SYN_RECEIVED) ||
12183 (tp->t_state == TCPS_SYN_SENT)) &&
12184 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
12185 (tp->t_rxtshift == 0)) { /* not a retransmit */
12187 goto just_return_nolock;
12190 * Before sending anything check for a state update. For hpts
12191 * calling without input this is important. If its input calling
12192 * then this was already done.
12194 if (bbr->rc_use_google == 0)
12195 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12198 * If we've recently taken a timeout, snd_max will be greater than
12199 * snd_max. BBR in general does not pay much attention to snd_nxt
12200 * for historic reasons the persist timer still uses it. This means
12201 * we have to look at it. All retransmissions that are not persits
12202 * use the rsm that needs to be sent so snd_nxt is ignored. At the
12203 * end of this routine we pull snd_nxt always up to snd_max.
12206 #ifdef BBR_INVARIANTS
12207 doing_retran_from = picked_up_retran = 0;
12213 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12214 sb_offset = tp->snd_max - tp->snd_una;
12215 flags = tcp_outflags[tp->t_state];
12219 if (flags & TH_RST) {
12224 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12225 /* We need to always have one in reserve */
12226 rsm = bbr_alloc(bbr);
12229 /* Lie to get on the hpts */
12230 tot_len = tp->t_maxseg;
12232 /* Retry in a ms */
12234 goto just_return_nolock;
12236 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12237 bbr->r_ctl.rc_free_cnt++;
12240 /* What do we send, a resend? */
12241 if (bbr->r_ctl.rc_resend == NULL) {
12242 /* Check for rack timeout */
12243 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12244 if (bbr->r_ctl.rc_resend) {
12245 #ifdef BBR_INVARIANTS
12246 picked_up_retran = 1;
12248 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12251 if (bbr->r_ctl.rc_resend) {
12252 rsm = bbr->r_ctl.rc_resend;
12253 #ifdef BBR_INVARIANTS
12254 doing_retran_from = 1;
12256 /* Remove any TLP flags its a RACK or T-O */
12257 rsm->r_flags &= ~BBR_TLP;
12258 bbr->r_ctl.rc_resend = NULL;
12259 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12260 #ifdef BBR_INVARIANTS
12261 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12262 tp, bbr, rsm, rsm->r_start, tp->snd_una);
12263 goto recheck_resend;
12267 goto recheck_resend;
12271 if (rsm->r_flags & BBR_HAS_SYN) {
12272 /* Only retransmit a SYN by itself */
12274 if ((flags & TH_SYN) == 0) {
12275 /* Huh something is wrong */
12277 if (rsm->r_start == rsm->r_end) {
12278 /* Clean it up, somehow we missed the ack? */
12279 bbr_log_syn(tp, NULL);
12281 /* TFO with data? */
12282 rsm->r_flags &= ~BBR_HAS_SYN;
12283 len = rsm->r_end - rsm->r_start;
12286 /* Retransmitting SYN */
12292 len = rsm->r_end - rsm->r_start;
12293 if ((bbr->rc_resends_use_tso == 0) &&
12298 sb_offset = rsm->r_start - tp->snd_una;
12301 KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12302 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12305 /* I dont think this can happen */
12307 goto recheck_resend;
12309 BBR_STAT_INC(bbr_resends_set);
12310 } else if (bbr->r_ctl.rc_tlp_send) {
12315 rsm = bbr->r_ctl.rc_tlp_send;
12316 bbr->r_ctl.rc_tlp_send = NULL;
12318 len = rsm->r_end - rsm->r_start;
12320 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12323 if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12324 #ifdef BBR_INVARIANTS
12325 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12326 tp, bbr, tp->snd_una, rsm, rsm->r_start);
12330 goto recheck_resend;
12333 sb_offset = rsm->r_start - tp->snd_una;
12334 BBR_STAT_INC(bbr_tlp_set);
12337 * Enforce a connection sendmap count limit if set
12338 * as long as we are not retransmiting.
12340 if ((rsm == NULL) &&
12341 (V_tcp_map_entries_limit > 0) &&
12342 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12343 BBR_STAT_INC(bbr_alloc_limited);
12344 if (!bbr->alloc_limit_reported) {
12345 bbr->alloc_limit_reported = 1;
12346 BBR_STAT_INC(bbr_alloc_limited_conns);
12348 goto just_return_nolock;
12350 #ifdef BBR_INVARIANTS
12351 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12352 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12353 tp, bbr, rsm, sb_offset, len);
12357 * Get standard flags, and add SYN or FIN if requested by 'hidden'
12360 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12362 if (tp->t_flags & TF_NEEDSYN)
12365 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12366 /* we are retransmitting the fin */
12370 * When retransmitting data do *not* include the
12371 * FIN. This could happen from a TLP probe if we
12372 * allowed data with a FIN.
12377 if (flags & TH_FIN)
12380 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12383 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12385 kern_prefetch(end_rsm, &prefetch_rsm);
12390 * If snd_nxt == snd_max and we have transmitted a FIN, the
12391 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12392 * negative length. This can also occur when TCP opens up its
12393 * congestion window while receiving additional duplicate acks after
12394 * fast-retransmit because TCP will reset snd_nxt to snd_max after
12395 * the fast-retransmit.
12397 * In the normal retransmit-FIN-only case, however, snd_nxt will be
12398 * set to snd_una, the sb_offset will be 0, and the length may wind
12401 * If sack_rxmit is true we are retransmitting from the scoreboard
12402 * in which case len is already set.
12404 if (sack_rxmit == 0) {
12407 avail = sbavail(sb);
12408 if (SEQ_GT(tp->snd_max, tp->snd_una))
12409 sb_offset = tp->snd_max - tp->snd_una;
12412 if (bbr->rc_tlp_new_data) {
12413 /* TLP is forcing out new data */
12419 if (tlplen > (uint32_t)(avail - sb_offset)) {
12420 tlplen = (uint32_t)(avail - sb_offset);
12422 if (tlplen > tp->snd_wnd) {
12427 bbr->rc_tlp_new_data = 0;
12429 what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12430 if ((len < p_maxseg) &&
12431 (bbr->rc_in_persist == 0) &&
12432 (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12433 ((avail - sb_offset) >= p_maxseg)) {
12435 * We are not completing whats in the socket
12436 * buffer (i.e. there is at least a segment
12437 * waiting to send) and we have 2 or more
12438 * segments outstanding. There is no sense
12439 * of sending a little piece. Lets defer and
12440 * and wait until we can send a whole
12445 if (bbr->rc_in_persist) {
12447 * We are in persists, figure out if
12448 * a retransmit is available (maybe the previous
12449 * persists we sent) or if we have to send new
12452 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12454 len = rsm->r_end - rsm->r_start;
12455 if (rsm->r_flags & BBR_HAS_FIN)
12457 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12460 BBR_STAT_INC(bbr_persist_reneg);
12462 * XXXrrs we could force the len to
12463 * 1 byte here to cause the chunk to
12464 * split apart.. but that would then
12465 * mean we always retransmit it as
12466 * one byte even after the window
12470 sb_offset = rsm->r_start - tp->snd_una;
12473 * First time through in persists or peer
12474 * acked our one byte. Though we do have
12475 * to have something in the sb.
12485 if (prefetch_so_done == 0) {
12486 kern_prefetch(so, &prefetch_so_done);
12487 prefetch_so_done = 1;
12490 * Lop off SYN bit if it has already been sent. However, if this is
12491 * SYN-SENT state and if segment contains data and if we don't know
12492 * that foreign host supports TAO, suppress sending segment.
12494 if ((flags & TH_SYN) && (rsm == NULL) &&
12495 SEQ_GT(tp->snd_max, tp->snd_una)) {
12496 if (tp->t_state != TCPS_SYN_RECEIVED)
12499 * When sending additional segments following a TFO SYN|ACK,
12500 * do not include the SYN bit.
12502 if (IS_FASTOPEN(tp->t_flags) &&
12503 (tp->t_state == TCPS_SYN_RECEIVED))
12505 sb_offset--, len++;
12506 if (sbavail(sb) == 0)
12508 } else if ((flags & TH_SYN) && rsm) {
12510 * Subtract one from the len for the SYN being
12516 * Be careful not to send data and/or FIN on SYN segments. This
12517 * measure is needed to prevent interoperability problems with not
12518 * fully conformant TCP implementations.
12520 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12525 * On TFO sockets, ensure no data is sent in the following cases:
12527 * - When retransmitting SYN|ACK on a passively-created socket
12528 * - When retransmitting SYN on an actively created socket
12529 * - When sending a zero-length cookie (cookie request) on an
12530 * actively created socket
12531 * - When the socket is in the CLOSED state (RST is being sent)
12533 if (IS_FASTOPEN(tp->t_flags) &&
12534 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12535 ((tp->t_state == TCPS_SYN_SENT) &&
12536 (tp->t_tfo_client_cookie_len == 0)) ||
12537 (flags & TH_RST))) {
12542 /* Without fast-open there should never be data sent on a SYN */
12543 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12547 * If FIN has been sent but not acked, but we haven't been
12548 * called to retransmit, len will be < 0. Otherwise, window
12549 * shrank after we sent into it. If window shrank to 0,
12550 * cancel pending retransmit, pull snd_nxt back to (closed)
12551 * window, and set the persist timer if it isn't already
12552 * going. If the window didn't close completely, just wait
12555 * We also do a general check here to ensure that we will
12556 * set the persist timer when we have data to send, but a
12557 * 0-byte window. This makes sure the persist timer is set
12558 * even if the packet hits one of the "goto send" lines
12562 if ((tp->snd_wnd == 0) &&
12563 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12564 (tp->snd_una == tp->snd_max) &&
12565 (sb_offset < (int)sbavail(sb))) {
12567 * Not enough room in the rwnd to send
12568 * a paced segment out.
12570 bbr_enter_persist(tp, bbr, cts, __LINE__);
12572 } else if ((rsm == NULL) &&
12573 (doing_tlp == 0) &&
12574 (len < bbr->r_ctl.rc_pace_max_segs)) {
12576 * We are not sending a full segment for
12577 * some reason. Should we not send anything (think
12578 * sws or persists)?
12580 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12581 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12582 (len < (int)(sbavail(sb) - sb_offset))) {
12584 * Here the rwnd is less than
12585 * the pacing size, this is not a retransmit,
12586 * we are established and
12587 * the send is not the last in the socket buffer
12588 * lets not send, and possibly enter persists.
12591 if (tp->snd_max == tp->snd_una)
12592 bbr_enter_persist(tp, bbr, cts, __LINE__);
12593 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12594 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12595 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12596 (len < (int)(sbavail(sb) - sb_offset)) &&
12597 (len < bbr_minseg(bbr))) {
12599 * Here we are not retransmitting, and
12600 * the cwnd is not so small that we could
12601 * not send at least a min size (rxt timer
12602 * not having gone off), We have 2 segments or
12603 * more already in flight, its not the tail end
12604 * of the socket buffer and the cwnd is blocking
12605 * us from sending out minimum pacing segment size.
12606 * Lets not send anything.
12608 bbr->rc_cwnd_limited = 1;
12610 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12611 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12612 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12613 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12614 (len < (int)(sbavail(sb) - sb_offset)) &&
12615 (TCPS_HAVEESTABLISHED(tp->t_state))) {
12617 * Here we have a send window but we have
12618 * filled it up and we can't send another pacing segment.
12619 * We also have in flight more than 2 segments
12620 * and we are not completing the sb i.e. we allow
12621 * the last bytes of the sb to go out even if
12622 * its not a full pacing segment.
12627 /* len will be >= 0 after this point. */
12628 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12629 tcp_sndbuf_autoscale(tp, so, sendwin);
12633 if (bbr->rc_in_persist &&
12636 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12638 * We are in persist, not doing a retransmit and don't have enough space
12639 * yet to send a full TSO. So is it at the end of the sb
12640 * if so we need to send else nuke to 0 and don't send.
12643 if (sbavail(sb) > sb_offset)
12644 sbleft = sbavail(sb) - sb_offset;
12647 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12648 /* not at end of sb lets not send */
12653 * Decide if we can use TCP Segmentation Offloading (if supported by
12656 * TSO may only be used if we are in a pure bulk sending state. The
12657 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12658 * options prevent using TSO. With TSO the TCP header is the same
12659 * (except for the sequence number) for all generated packets. This
12660 * makes it impossible to transmit any options which vary per
12661 * generated segment or packet.
12663 * IPv4 handling has a clear separation of ip options and ip header
12664 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12665 * does the right thing below to provide length of just ip options
12666 * and thus checking for ipoptlen is enough to decide if ip options
12671 ipoptlen = ip6_optlen(inp);
12674 if (inp->inp_options)
12675 ipoptlen = inp->inp_options->m_len -
12676 offsetof(struct ipoption, ipopt_list);
12679 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12681 * Pre-calculate here as we save another lookup into the darknesses
12682 * of IPsec that way and can actually decide if TSO is ok.
12685 if (isipv6 && IPSEC_ENABLED(ipv6))
12686 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12692 if (IPSEC_ENABLED(ipv4))
12693 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12696 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12697 ipoptlen += ipsec_optlen;
12699 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12701 (tp->t_port == 0) &&
12702 ((tp->t_flags & TF_SIGNATURE) == 0) &&
12703 tp->rcv_numsacks == 0 &&
12707 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12708 (long)TCP_MAXWIN << tp->rcv_scale);
12710 * Sender silly window avoidance. We transmit under the following
12711 * conditions when len is non-zero:
12713 * - We have a full segment (or more with TSO) - This is the last
12714 * buffer in a write()/send() and we are either idle or running
12715 * NODELAY - we've timed out (e.g. persist timer) - we have more
12716 * then 1/2 the maximum send window's worth of data (receiver may be
12717 * limited the window size) - we need to retransmit
12724 if (len >= p_maxseg)
12727 * NOTE! on localhost connections an 'ack' from the remote
12728 * end may occur synchronously with the output and cause us
12729 * to flush a buffer queued with moretocome. XXX
12732 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */
12733 ((tp->t_flags & TF_NODELAY) ||
12734 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12735 (tp->t_flags & TF_NOPUSH) == 0) {
12738 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */
12741 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12746 * Sending of standalone window updates.
12748 * Window updates are important when we close our window due to a
12749 * full socket buffer and are opening it again after the application
12750 * reads data from it. Once the window has opened again and the
12751 * remote end starts to send again the ACK clock takes over and
12752 * provides the most current window information.
12754 * We must avoid the silly window syndrome whereas every read from
12755 * the receive buffer, no matter how small, causes a window update
12756 * to be sent. We also should avoid sending a flurry of window
12757 * updates when the socket buffer had queued a lot of data and the
12758 * application is doing small reads.
12760 * Prevent a flurry of pointless window updates by only sending an
12761 * update when we can increase the advertized window by more than
12762 * 1/4th of the socket buffer capacity. When the buffer is getting
12763 * full or is very small be more aggressive and send an update
12764 * whenever we can increase by two mss sized segments. In all other
12765 * situations the ACK's to new incoming data will carry further
12766 * window increases.
12768 * Don't send an independent window update if a delayed ACK is
12769 * pending (it will get piggy-backed on it) or the remote side
12770 * already has done a half-close and won't send more data. Skip
12771 * this if the connection is in T/TCP half-open state.
12773 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12774 !(tp->t_flags & TF_DELACK) &&
12775 !TCPS_HAVERCVDFIN(tp->t_state)) {
12776 /* Check to see if we should do a window update */
12777 if (bbr_window_update_needed(tp, so, recwin, maxseg))
12781 * Send if we owe the peer an ACK, RST, SYN. ACKNOW
12782 * is also a catch-all for the retransmit timer timeout case.
12784 if (tp->t_flags & TF_ACKNOW) {
12787 if (flags & TH_RST) {
12788 /* Always send a RST if one is due */
12791 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) {
12795 * If our state indicates that FIN should be sent and we have not
12796 * yet done so, then we need to send.
12798 if (flags & TH_FIN &&
12799 ((tp->t_flags & TF_SENTFIN) == 0)) {
12803 * No reason to send a segment, just return.
12806 SOCKBUF_UNLOCK(sb);
12807 just_return_nolock:
12809 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
12810 if (bbr->rc_no_pacing)
12812 if (tot_len == 0) {
12813 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
12815 BBR_STAT_INC(bbr_rwnd_limited);
12816 app_limited = BBR_JR_RWND_LIMITED;
12817 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12818 if ((bbr->rc_in_persist == 0) &&
12819 TCPS_HAVEESTABLISHED(tp->t_state) &&
12820 (tp->snd_max == tp->snd_una) &&
12821 sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
12822 /* No send window.. we must enter persist */
12823 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
12825 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
12826 BBR_STAT_INC(bbr_app_limited);
12827 app_limited = BBR_JR_APP_LIMITED;
12828 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12829 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12830 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
12831 BBR_STAT_INC(bbr_cwnd_limited);
12832 app_limited = BBR_JR_CWND_LIMITED;
12833 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12834 bbr->r_ctl.rc_lost_bytes)));
12835 bbr->rc_cwnd_limited = 1;
12837 BBR_STAT_INC(bbr_app_limited);
12838 app_limited = BBR_JR_APP_LIMITED;
12839 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12841 bbr->r_ctl.rc_hptsi_agg_delay = 0;
12842 bbr->r_agg_early_set = 0;
12843 bbr->r_ctl.rc_agg_early = 0;
12844 bbr->r_ctl.rc_last_delay_val = 0;
12845 } else if (bbr->rc_use_google == 0)
12846 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12847 /* Are we app limited? */
12848 if ((app_limited == BBR_JR_APP_LIMITED) ||
12849 (app_limited == BBR_JR_RWND_LIMITED)) {
12851 * We are application limited.
12853 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12854 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
12857 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12858 /* Dont update the time if we did not send */
12859 bbr->r_ctl.rc_last_delay_val = 0;
12860 bbr->rc_output_starts_timer = 1;
12861 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
12862 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
12863 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12864 /* Make sure snd_nxt is drug up */
12865 tp->snd_nxt = tp->snd_max;
12870 if (doing_tlp == 0) {
12872 * Data not a TLP, and its not the rxt firing. If it is the
12873 * rxt firing, we want to leave the tlp_in_progress flag on
12874 * so we don't send another TLP. It has to be a rack timer
12875 * or normal send (response to acked data) to clear the tlp
12876 * in progress flag.
12878 bbr->rc_tlp_in_progress = 0;
12879 bbr->rc_tlp_rtx_out = 0;
12884 bbr->rc_tlp_in_progress = 1;
12886 bbr_timer_cancel(bbr, __LINE__, cts);
12888 if (sbused(sb) > 0) {
12890 * This is sub-optimal. We only send a stand alone
12891 * FIN on its own segment.
12893 if (flags & TH_FIN) {
12895 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
12896 /* Lets not send this */
12904 * We do *not* send a FIN on a retransmit if it has data.
12905 * The if clause here where len > 1 should never come true.
12908 (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
12909 (flags & TH_FIN))) {
12914 SOCKBUF_LOCK_ASSERT(sb);
12916 if ((tp->snd_una == tp->snd_max) &&
12917 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
12919 * This qualifies as a RTT_PROBE session since we
12920 * drop the data outstanding to nothing and waited
12921 * more than bbr_rtt_probe_time.
12923 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
12924 bbr_set_reduced_rtt(bbr, cts, __LINE__);
12927 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
12929 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
12932 * Before ESTABLISHED, force sending of initial options unless TCP
12933 * set not to do any options. NOTE: we assume that the IP/TCP header
12934 * plus TCP options always fit in a single mbuf, leaving room for a
12935 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
12936 * + optlen <= MCLBYTES
12941 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12944 hdrlen = sizeof(struct tcpiphdr);
12947 * Compute options for segment. We only have to care about SYN and
12948 * established connection segments. Options for SYN-ACK segments
12949 * are handled in TCP syncache.
12953 if ((tp->t_flags & TF_NOOPT) == 0) {
12954 /* Maximum segment size. */
12955 if (flags & TH_SYN) {
12956 to.to_mss = tcp_mssopt(&inp->inp_inc);
12958 to.to_mss -= V_tcp_udp_tunneling_overhead;
12959 to.to_flags |= TOF_MSS;
12961 * On SYN or SYN|ACK transmits on TFO connections,
12962 * only include the TFO option if it is not a
12963 * retransmit, as the presence of the TFO option may
12964 * have caused the original SYN or SYN|ACK to have
12965 * been dropped by a middlebox.
12967 if (IS_FASTOPEN(tp->t_flags) &&
12968 (tp->t_rxtshift == 0)) {
12969 if (tp->t_state == TCPS_SYN_RECEIVED) {
12970 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
12972 (u_int8_t *)&tp->t_tfo_cookie.server;
12973 to.to_flags |= TOF_FASTOPEN;
12975 } else if (tp->t_state == TCPS_SYN_SENT) {
12977 tp->t_tfo_client_cookie_len;
12979 tp->t_tfo_cookie.client;
12980 to.to_flags |= TOF_FASTOPEN;
12985 /* Window scaling. */
12986 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
12987 to.to_wscale = tp->request_r_scale;
12988 to.to_flags |= TOF_SCALE;
12991 if ((tp->t_flags & TF_RCVD_TSTMP) ||
12992 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
12993 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
12994 to.to_tsecr = tp->ts_recent;
12995 to.to_flags |= TOF_TS;
12996 local_options += TCPOLEN_TIMESTAMP + 2;
12998 /* Set receive buffer autosizing timestamp. */
12999 if (tp->rfbuf_ts == 0 &&
13000 (so->so_rcv.sb_flags & SB_AUTOSIZE))
13001 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv);
13002 /* Selective ACK's. */
13003 if (flags & TH_SYN)
13004 to.to_flags |= TOF_SACKPERM;
13005 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13006 tp->rcv_numsacks > 0) {
13007 to.to_flags |= TOF_SACK;
13008 to.to_nsacks = tp->rcv_numsacks;
13009 to.to_sacks = (u_char *)tp->sackblks;
13011 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13012 /* TCP-MD5 (RFC2385). */
13013 if (tp->t_flags & TF_SIGNATURE)
13014 to.to_flags |= TOF_SIGNATURE;
13015 #endif /* TCP_SIGNATURE */
13017 /* Processing the options. */
13018 hdrlen += (optlen = tcp_addoptions(&to, opt));
13020 * If we wanted a TFO option to be added, but it was unable
13021 * to fit, ensure no data is sent.
13023 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13024 !(to.to_flags & TOF_FASTOPEN))
13028 if (V_tcp_udp_tunneling_port == 0) {
13029 /* The port was removed?? */
13030 SOCKBUF_UNLOCK(&so->so_snd);
13031 return (EHOSTUNREACH);
13033 hdrlen += sizeof(struct udphdr);
13037 ipoptlen = ip6_optlen(tp->t_inpcb);
13040 if (tp->t_inpcb->inp_options)
13041 ipoptlen = tp->t_inpcb->inp_options->m_len -
13042 offsetof(struct ipoption, ipopt_list);
13046 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13047 ipoptlen += ipsec_optlen;
13049 if (bbr->rc_last_options != local_options) {
13051 * Cache the options length this generally does not change
13052 * on a connection. We use this to calculate TSO.
13054 bbr->rc_last_options = local_options;
13056 maxseg = tp->t_maxseg - (ipoptlen + optlen);
13057 p_maxseg = min(maxseg, pace_max_segs);
13059 * Adjust data length if insertion of options will bump the packet
13060 * length beyond the t_maxseg length. Clear the FIN bit because we
13061 * cut off the tail of the segment.
13063 if (len > maxseg) {
13064 if (len != 0 && (flags & TH_FIN)) {
13071 /* extract TSO information */
13072 if_hw_tsomax = tp->t_tsomax;
13073 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13074 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13075 KASSERT(ipoptlen == 0,
13076 ("%s: TSO can't do IP options", __func__));
13079 * Check if we should limit by maximum payload
13082 if (if_hw_tsomax != 0) {
13083 /* compute maximum TSO length */
13084 max_len = (if_hw_tsomax - hdrlen -
13086 if (max_len <= 0) {
13088 } else if (len > max_len) {
13093 * Prevent the last segment from being fractional
13094 * unless the send sockbuf can be emptied:
13096 if ((sb_offset + len) < sbavail(sb)) {
13097 moff = len % (uint32_t)maxseg;
13103 * In case there are too many small fragments don't
13106 if (len <= maxseg) {
13111 /* Not doing TSO */
13112 if (optlen + ipoptlen >= tp->t_maxseg) {
13114 * Since we don't have enough space to put
13115 * the IP header chain and the TCP header in
13116 * one packet as required by RFC 7112, don't
13117 * send it. Also ensure that at least one
13118 * byte of the payload can be put into the
13121 SOCKBUF_UNLOCK(&so->so_snd);
13129 /* Not doing TSO */
13130 if_hw_tsomaxsegcount = 0;
13133 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13134 ("%s: len > IP_MAXPACKET", __func__));
13137 if (max_linkhdr + hdrlen > MCLBYTES)
13139 if (max_linkhdr + hdrlen > MHLEN)
13141 panic("tcphdr too big");
13144 * This KASSERT is here to catch edge cases at a well defined place.
13145 * Before, those had triggered (random) panic conditions further
13148 #ifdef BBR_INVARIANTS
13150 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13151 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13152 rsm, tp, bbr, rsm->r_start, tp->snd_una);
13156 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13158 (flags & TH_FIN) &&
13161 * We have outstanding data, don't send a fin by itself!.
13167 * Grab a header mbuf, attaching a copy of data to be transmitted,
13168 * and initialize the header from the template for sends on this
13176 * We place a limit on sending with hptsi.
13178 if ((rsm == NULL) && len > pace_max_segs)
13179 len = pace_max_segs;
13183 if (MHLEN < hdrlen + max_linkhdr)
13184 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13187 m = m_gethdr(M_NOWAIT, MT_DATA);
13190 BBR_STAT_INC(bbr_failed_mbuf_aloc);
13191 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13192 SOCKBUF_UNLOCK(sb);
13197 m->m_data += max_linkhdr;
13200 * Start the m_copy functions from the closest mbuf to the
13201 * sb_offset in the socket buffer chain.
13203 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13204 #ifdef BBR_INVARIANTS
13205 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13206 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13207 tp, bbr, len, sb_offset, sbavail(sb), rsm,
13214 * In this messed up situation we have two choices,
13215 * a) pretend the send worked, and just start timers
13216 * and what not (not good since that may lead us
13217 * back here a lot). <or> b) Send the lowest segment
13218 * in the map. <or> c) Drop the connection. Lets do
13219 * <b> which if it continues to happen will lead to
13220 * <c> via timeouts.
13222 BBR_STAT_INC(bbr_offset_recovery);
13223 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13230 if (rsm->r_start != tp->snd_una) {
13232 * Things are really messed up, <c>
13233 * is the only thing to do.
13235 BBR_STAT_INC(bbr_offset_drop);
13236 tcp_set_inp_to_drop(inp, EFAULT);
13237 SOCKBUF_UNLOCK(sb);
13241 len = rsm->r_end - rsm->r_start;
13243 if (len > sbavail(sb))
13248 mb = sbsndptr_noadv(sb, sb_offset, &moff);
13249 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13250 m_copydata(mb, moff, (int)len,
13251 mtod(m, caddr_t)+hdrlen);
13253 sbsndptr_adv(sb, mb, len);
13256 struct sockbuf *msb;
13262 #ifdef BBR_INVARIANTS
13263 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13265 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ",
13266 tp, bbr, len, moff,
13268 tp->snd_una, rsm->r_flags, rsm->r_start,
13271 doing_tlp, sack_rxmit);
13273 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13274 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13279 m->m_next = tcp_m_copym(
13281 if_hw_tsomaxsegcount,
13282 if_hw_tsomaxsegsize, msb,
13283 ((rsm == NULL) ? hw_tls : 0)
13284 #ifdef NETFLIX_COPY_ARGS
13288 if (len <= maxseg) {
13290 * Must have ran out of mbufs for the copy
13291 * shorten it to no longer need tso. Lets
13292 * not put on sendalot since we are low on
13297 if (m->m_next == NULL) {
13298 SOCKBUF_UNLOCK(sb);
13305 #ifdef BBR_INVARIANTS
13306 if (tso && len < maxseg) {
13307 panic("tp:%p tso on, but len:%d < maxseg:%d",
13310 if (tso && if_hw_tsomaxsegcount) {
13311 int32_t seg_cnt = 0;
13319 if (seg_cnt > if_hw_tsomaxsegcount) {
13320 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13325 * If we're sending everything we've got, set PUSH. (This
13326 * will keep happy those implementations which only give
13327 * data to the user when a buffer fills or a PUSH comes in.)
13329 if (sb_offset + len == sbused(sb) &&
13331 !(flags & TH_SYN)) {
13334 SOCKBUF_UNLOCK(sb);
13336 SOCKBUF_UNLOCK(sb);
13337 if (tp->t_flags & TF_ACKNOW)
13338 KMOD_TCPSTAT_INC(tcps_sndacks);
13339 else if (flags & (TH_SYN | TH_FIN | TH_RST))
13340 KMOD_TCPSTAT_INC(tcps_sndctrl);
13342 KMOD_TCPSTAT_INC(tcps_sndwinup);
13344 m = m_gethdr(M_NOWAIT, MT_DATA);
13346 BBR_STAT_INC(bbr_failed_mbuf_aloc);
13347 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13349 /* Fudge the send time since we could not send */
13354 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13356 M_ALIGN(m, hdrlen);
13359 m->m_data += max_linkhdr;
13362 SOCKBUF_UNLOCK_ASSERT(sb);
13363 m->m_pkthdr.rcvif = (struct ifnet *)0;
13365 mac_inpcb_create_mbuf(inp, m);
13369 ip6 = mtod(m, struct ip6_hdr *);
13371 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
13372 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13373 udp->uh_dport = tp->t_port;
13374 ulen = hdrlen + len - sizeof(struct ip6_hdr);
13375 udp->uh_ulen = htons(ulen);
13376 th = (struct tcphdr *)(udp + 1);
13378 th = (struct tcphdr *)(ip6 + 1);
13380 tcpip_fillheaders(inp, tp->t_port, ip6, th);
13384 ip = mtod(m, struct ip *);
13386 ipov = (struct ipovly *)ip;
13389 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
13390 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13391 udp->uh_dport = tp->t_port;
13392 ulen = hdrlen + len - sizeof(struct ip);
13393 udp->uh_ulen = htons(ulen);
13394 th = (struct tcphdr *)(udp + 1);
13396 th = (struct tcphdr *)(ip + 1);
13398 tcpip_fillheaders(inp, tp->t_port, ip, th);
13401 * If we are doing retransmissions, then snd_nxt will not reflect
13402 * the first unsent octet. For ACK only packets, we do not want the
13403 * sequence number of the retransmitted packet, we want the sequence
13404 * number of the next unsent octet. So, if there is no data (and no
13405 * SYN or FIN), use snd_max instead of snd_nxt when filling in
13406 * ti_seq. But if we are in persist state, snd_max might reflect
13407 * one byte beyond the right edge of the window, so use snd_nxt in
13408 * that case, since we know we aren't doing a retransmission.
13409 * (retransmit and persist are mutually exclusive...)
13411 if (sack_rxmit == 0) {
13412 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13413 /* New data (including new persists) */
13414 th->th_seq = htonl(tp->snd_max);
13415 bbr_seq = tp->snd_max;
13416 } else if (flags & TH_SYN) {
13417 /* Syn's always send from iss */
13418 th->th_seq = htonl(tp->iss);
13420 } else if (flags & TH_FIN) {
13421 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13423 * If we sent the fin already its 1 minus
13426 th->th_seq = (htonl(tp->snd_max - 1));
13427 bbr_seq = (tp->snd_max - 1);
13429 /* First time FIN use snd_max */
13430 th->th_seq = htonl(tp->snd_max);
13431 bbr_seq = tp->snd_max;
13435 * len == 0 and not persist we use snd_max, sending
13436 * an ack unless we have sent the fin then its 1
13440 * XXXRRS Question if we are in persists and we have
13441 * nothing outstanding to send and we have not sent
13442 * a FIN, we will send an ACK. In such a case it
13443 * might be better to send (tp->snd_una - 1) which
13444 * would force the peer to ack.
13446 if (tp->t_flags & TF_SENTFIN) {
13447 th->th_seq = htonl(tp->snd_max - 1);
13448 bbr_seq = (tp->snd_max - 1);
13450 th->th_seq = htonl(tp->snd_max);
13451 bbr_seq = tp->snd_max;
13455 /* All retransmits use the rsm to guide the send */
13456 th->th_seq = htonl(rsm->r_start);
13457 bbr_seq = rsm->r_start;
13459 th->th_ack = htonl(tp->rcv_nxt);
13461 bcopy(opt, th + 1, optlen);
13462 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13464 th->th_flags = flags;
13466 * Calculate receive window. Don't shrink window, but avoid silly
13469 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13472 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13473 recwin < (tp->rcv_adv - tp->rcv_nxt))
13474 recwin = (tp->rcv_adv - tp->rcv_nxt);
13475 if (recwin > TCP_MAXWIN << tp->rcv_scale)
13476 recwin = TCP_MAXWIN << tp->rcv_scale;
13479 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13480 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is
13481 * handled in syncache.
13483 if (flags & TH_SYN)
13484 th->th_win = htons((u_short)
13485 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13487 /* Avoid shrinking window with window scaling. */
13488 recwin = roundup2(recwin, 1 << tp->rcv_scale);
13489 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13492 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13493 * window. This may cause the remote transmitter to stall. This
13494 * flag tells soreceive() to disable delayed acknowledgements when
13495 * draining the buffer. This can occur if the receiver is
13496 * attempting to read more data than can be buffered prior to
13497 * transmitting on the connection.
13499 if (th->th_win == 0) {
13500 tp->t_sndzerowin++;
13501 tp->t_flags |= TF_RXWIN0SENT;
13503 tp->t_flags &= ~TF_RXWIN0SENT;
13505 * We don't support urgent data, but drag along
13506 * the pointer in case of a stack switch.
13508 tp->snd_up = tp->snd_una;
13510 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13511 if (to.to_flags & TOF_SIGNATURE) {
13513 * Calculate MD5 signature and put it into the place
13514 * determined before. NOTE: since TCP options buffer doesn't
13515 * point into mbuf's data, calculate offset and use it.
13517 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13518 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13520 * Do not send segment if the calculation of MD5
13521 * digest has failed.
13529 * Put TCP length in extended header, and then checksum extended
13532 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
13536 * ip6_plen is not need to be filled now, and will be filled
13540 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13541 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13542 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13543 th->th_sum = htons(0);
13544 UDPSTAT_INC(udps_opackets);
13546 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13547 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13548 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13549 optlen + len, IPPROTO_TCP, 0);
13553 #if defined(INET6) && defined(INET)
13559 m->m_pkthdr.csum_flags = CSUM_UDP;
13560 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13561 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13562 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13563 th->th_sum = htons(0);
13564 UDPSTAT_INC(udps_opackets);
13566 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13567 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13568 th->th_sum = in_pseudo(ip->ip_src.s_addr,
13569 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13570 IPPROTO_TCP + len + optlen));
13572 /* IP version must be set here for ipv4/ipv6 checking later */
13573 KASSERT(ip->ip_v == IPVERSION,
13574 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13579 * Enable TSO and specify the size of the segments. The TCP pseudo
13580 * header checksum is always provided. XXX: Fixme: This is currently
13581 * not the case for IPv6.
13584 KASSERT(len > maxseg,
13585 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13586 m->m_pkthdr.csum_flags |= CSUM_TSO;
13587 csum_flags |= CSUM_TSO;
13588 m->m_pkthdr.tso_segsz = maxseg;
13590 KASSERT(len + hdrlen == m_length(m, NULL),
13591 ("%s: mbuf chain different than expected: %d + %u != %u",
13592 __func__, len, hdrlen, m_length(m, NULL)));
13595 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13596 hhook_run_tcp_est_out(tp, th, &to, len, tso);
13602 if (so->so_options & SO_DEBUG) {
13609 save = ipov->ih_len;
13610 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen +
13611 * (th->th_off << 2) */ );
13613 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13617 ipov->ih_len = save;
13619 #endif /* TCPDEBUG */
13621 /* Log to the black box */
13622 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13623 union tcp_log_stackspecific log;
13625 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13626 /* Record info on type of transmission */
13627 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13628 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13629 log.u_bbr.flex3 = maxseg;
13630 log.u_bbr.flex4 = delay_calc;
13631 /* Encode filled_all into the upper flex5 bit */
13632 log.u_bbr.flex5 = bbr->rc_past_init_win;
13633 log.u_bbr.flex5 <<= 1;
13634 log.u_bbr.flex5 |= bbr->rc_no_pacing;
13635 log.u_bbr.flex5 <<= 29;
13637 log.u_bbr.flex5 |= 0x80000000;
13638 log.u_bbr.flex5 |= tp->t_maxseg;
13639 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13640 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13641 /* lets poke in the low and the high here for debugging */
13642 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13643 if (rsm || sack_rxmit) {
13645 log.u_bbr.flex8 = 2;
13647 log.u_bbr.flex8 = 1;
13649 log.u_bbr.flex8 = 0;
13651 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13652 len, &log, false, NULL, NULL, 0, tv);
13657 * Fill in IP length and desired time to live and send to IP level.
13658 * There should be a better way to handle ttl and tos; we could keep
13659 * them in the template, but need a way to checksum without them.
13662 * m->m_pkthdr.len should have been set before cksum calcuration,
13663 * because in6_cksum() need it.
13668 * we separately set hoplimit for every segment, since the
13669 * user might want to change the value via setsockopt. Also,
13670 * desired default hop limit might be changed via Neighbor
13673 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13676 * Set the packet size here for the benefit of DTrace
13677 * probes. ip6_output() will set it properly; it's supposed
13678 * to include the option header lengths as well.
13680 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13682 if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13683 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13685 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13687 if (tp->t_state == TCPS_SYN_SENT)
13688 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13690 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13691 /* TODO: IPv6 IP6TOS_ECT bit on */
13692 error = ip6_output(m, inp->in6p_outputopts,
13694 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13697 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13698 mtu = inp->inp_route6.ro_nh->nh_mtu;
13701 #if defined(INET) && defined(INET6)
13706 ip->ip_len = htons(m->m_pkthdr.len);
13709 ip->ip_ttl = in6_selecthlim(inp, NULL);
13712 * If we do path MTU discovery, then we set DF on every
13713 * packet. This might not be the best thing to do according
13714 * to RFC3390 Section 2. However the tcp hostcache migitates
13715 * the problem so it affects only the first tcp connection
13718 * NB: Don't set DF on small MTU/MSS to have a safe
13721 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13722 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13723 if (tp->t_port == 0 || len < V_tcp_minmss) {
13724 ip->ip_off |= htons(IP_DF);
13727 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13730 if (tp->t_state == TCPS_SYN_SENT)
13731 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13733 TCP_PROBE5(send, NULL, tp, ip, tp, th);
13735 error = ip_output(m, inp->inp_options, &inp->inp_route,
13736 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13738 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13739 mtu = inp->inp_route.ro_nh->nh_mtu;
13745 lgb->tlb_errno = error;
13749 * In transmit state, time the transmission and arrange for the
13750 * retransmit. In persist state, just set snd_max.
13753 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp);
13754 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13755 (tp->t_flags & TF_SACK_PERMIT) &&
13756 tp->rcv_numsacks > 0)
13757 tcp_clean_dsack_blocks(tp);
13758 /* We sent an ack clear the bbr_segs_rcvd count */
13759 bbr->output_error_seen = 0;
13760 bbr->oerror_cnt = 0;
13761 bbr->bbr_segs_rcvd = 0;
13763 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13764 /* Do accounting for new sends */
13765 if ((len > 0) && (rsm == NULL)) {
13767 if (tp->snd_una == tp->snd_max) {
13769 * Special case to match google, when
13770 * nothing is in flight the delivered
13771 * time does get updated to the current
13772 * time (see tcp_rate_bsd.c).
13774 bbr->r_ctl.rc_del_time = cts;
13776 if (len >= maxseg) {
13777 idx = (len / maxseg) + 3;
13778 if (idx >= TCP_MSS_ACCT_ATIMER)
13779 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13781 counter_u64_add(bbr_out_size[idx], 1);
13783 /* smaller than a MSS */
13784 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13785 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13786 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13787 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13793 * We must do the send accounting before we log the output,
13794 * otherwise the state of the rsm could change and we account to the
13798 bbr_do_send_accounting(tp, bbr, rsm, len, error);
13800 if (tp->snd_una == tp->snd_max)
13801 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
13804 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
13805 cts, mb, &abandon, rsm, 0, sb);
13808 * If bbr_log_output destroys the TCB or sees a TH_RST being
13809 * sent we should hit this condition.
13813 if (bbr->rc_in_persist == 0) {
13815 * Advance snd_nxt over sequence space of this segment.
13818 /* We don't log or do anything with errors */
13821 if (tp->snd_una == tp->snd_max &&
13822 (len || (flags & (TH_SYN | TH_FIN)))) {
13824 * Update the time we just added data since none was
13827 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13828 bbr->rc_tp->t_acktime = ticks;
13830 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
13831 if (flags & TH_SYN) {
13833 * Smack the snd_max to iss + 1
13834 * if its a FO we will add len below.
13836 tp->snd_max = tp->iss + 1;
13838 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13840 tp->t_flags |= TF_SENTFIN;
13843 if (sack_rxmit == 0)
13844 tp->snd_max += len;
13846 if ((error == 0) && len)
13849 /* Persists case */
13850 int32_t xlen = len;
13855 if (flags & TH_SYN)
13857 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13859 tp->t_flags |= TF_SENTFIN;
13861 if (xlen && (tp->snd_una == tp->snd_max)) {
13863 * Update the time we just added data since none was
13866 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13867 bbr->rc_tp->t_acktime = ticks;
13869 if (sack_rxmit == 0)
13870 tp->snd_max += xlen;
13871 tot_len += (len + optlen + ipoptlen);
13876 * Failures do not advance the seq counter above. For the
13877 * case of ENOBUFS we will fall out and become ack-clocked.
13878 * capping the cwnd at the current flight.
13879 * Everything else will just have to retransmit with the timer
13882 SOCKBUF_UNLOCK_ASSERT(sb);
13883 BBR_STAT_INC(bbr_saw_oerr);
13884 /* Clear all delay/early tracks */
13885 bbr->r_ctl.rc_hptsi_agg_delay = 0;
13886 bbr->r_ctl.rc_agg_early = 0;
13887 bbr->r_agg_early_set = 0;
13888 bbr->output_error_seen = 1;
13889 if (bbr->oerror_cnt < 0xf)
13891 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
13892 /* drop the session */
13893 tcp_set_inp_to_drop(inp, ENETDOWN);
13898 * Make this guy have to get ack's to send
13899 * more but lets make sure we don't
13900 * slam him below a T-O (1MSS).
13902 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
13903 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13904 bbr->r_ctl.rc_lost_bytes)) - maxseg;
13905 if (tp->snd_cwnd < maxseg)
13906 tp->snd_cwnd = maxseg;
13908 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
13909 BBR_STAT_INC(bbr_saw_enobuf);
13910 if (bbr->bbr_hdrw_pacing)
13911 counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
13913 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
13915 * Here even in the enobuf's case we want to do our
13916 * state update. The reason being we may have been
13917 * called by the input function. If so we have had
13924 * For some reason the interface we used initially
13925 * to send segments changed to another or lowered
13926 * its MTU. If TSO was active we either got an
13927 * interface without TSO capabilits or TSO was
13928 * turned off. If we obtained mtu from ip_output()
13929 * then update it and try again.
13931 /* Turn on tracing (or try to) */
13935 old_maxseg = tp->t_maxseg;
13936 BBR_STAT_INC(bbr_saw_emsgsiz);
13937 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
13939 tcp_mss_update(tp, -1, mtu, NULL, NULL);
13940 if (old_maxseg <= tp->t_maxseg) {
13941 /* Huh it did not shrink? */
13942 tp->t_maxseg = old_maxseg - 40;
13943 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
13946 * Nuke all other things that can interfere
13949 if ((tot_len + len) && (len >= tp->t_maxseg)) {
13950 slot = bbr_get_pacing_delay(bbr,
13951 bbr->r_ctl.rc_bbr_hptsi_gain,
13952 (tot_len + len), cts, 0);
13953 if (slot < bbr_error_base_paceout)
13954 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13956 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
13957 bbr->rc_output_starts_timer = 1;
13958 bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
13963 tp->t_softerror = error;
13969 if (TCPS_HAVERCVDSYN(tp->t_state)) {
13970 tp->t_softerror = error;
13974 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
13975 bbr->rc_output_starts_timer = 1;
13976 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
13980 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
13983 (bbr->rc_in_persist == 0)) {
13984 tp->gput_seq = bbr_seq;
13985 tp->gput_ack = bbr_seq +
13986 min(sbavail(&so->so_snd) - sb_offset, sendwin);
13988 tp->t_flags |= TF_GPUTINPROG;
13991 KMOD_TCPSTAT_INC(tcps_sndtotal);
13992 if ((bbr->bbr_hdw_pace_ena) &&
13993 (bbr->bbr_attempt_hdwr_pace == 0) &&
13994 (bbr->rc_past_init_win) &&
13995 (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
13996 (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
13997 (inp->inp_route.ro_nh &&
13998 inp->inp_route.ro_nh->nh_ifp)) {
14000 * We are past the initial window and
14001 * have at least one measurement so we
14002 * could use hardware pacing if its available.
14003 * We have an interface and we have not attempted
14004 * to setup hardware pacing, lets try to now.
14006 uint64_t rate_wanted;
14009 rate_wanted = bbr_get_hardware_rate(bbr);
14010 bbr->bbr_attempt_hdwr_pace = 1;
14011 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
14012 inp->inp_route.ro_nh->nh_ifp,
14014 (RS_PACING_GEQ|RS_PACING_SUB_OK),
14016 if (bbr->r_ctl.crte) {
14017 bbr_type_log_hdwr_pacing(bbr,
14018 bbr->r_ctl.crte->ptbl->rs_ifp,
14020 bbr->r_ctl.crte->rate,
14021 __LINE__, cts, err);
14022 BBR_STAT_INC(bbr_hdwr_rl_add_ok);
14023 counter_u64_add(bbr_flows_nohdwr_pacing, -1);
14024 counter_u64_add(bbr_flows_whdwr_pacing, 1);
14025 bbr->bbr_hdrw_pacing = 1;
14026 /* Now what is our gain status? */
14027 if (bbr->r_ctl.crte->rate < rate_wanted) {
14028 /* We have a problem */
14029 bbr_setup_less_of_rate(bbr, cts,
14030 bbr->r_ctl.crte->rate, rate_wanted);
14033 bbr->gain_is_limited = 0;
14034 bbr->skip_gain = 0;
14036 tcp_bbr_tso_size_check(bbr, cts);
14038 bbr_type_log_hdwr_pacing(bbr,
14039 inp->inp_route.ro_nh->nh_ifp,
14042 __LINE__, cts, err);
14043 BBR_STAT_INC(bbr_hdwr_rl_add_fail);
14046 if (bbr->bbr_hdrw_pacing) {
14048 * Worry about cases where the route
14049 * changes or something happened that we
14050 * lost our hardware pacing possibly during
14051 * the last ip_output call.
14053 if (inp->inp_snd_tag == NULL) {
14054 /* A change during ip output disabled hw pacing? */
14055 bbr->bbr_hdrw_pacing = 0;
14056 } else if ((inp->inp_route.ro_nh == NULL) ||
14057 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
14059 * We had an interface or route change,
14060 * detach from the current hdwr pacing
14061 * and setup to re-attempt next go
14064 bbr->bbr_hdrw_pacing = 0;
14065 bbr->bbr_attempt_hdwr_pace = 0;
14066 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
14067 tcp_bbr_tso_size_check(bbr, cts);
14071 * Data sent (as far as we can tell). If this advertises a larger
14072 * window than any other segment, then remember the size of the
14073 * advertised window. Any pending ACK has now been sent.
14075 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14076 tp->rcv_adv = tp->rcv_nxt + recwin;
14078 tp->last_ack_sent = tp->rcv_nxt;
14079 if ((error == 0) &&
14080 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
14081 (doing_tlp == 0) &&
14084 ((flags & TH_RST) == 0) &&
14085 ((flags & TH_SYN) == 0) &&
14086 (IN_RECOVERY(tp->t_flags) == 0) &&
14087 (bbr->rc_in_persist == 0) &&
14088 (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
14090 * For non-tso we need to goto again until we have sent out
14091 * enough data to match what we are hptsi out every hptsi
14094 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14095 /* Make sure snd_nxt is drug up */
14096 tp->snd_nxt = tp->snd_max;
14104 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14108 if ((error == 0) && (flags & TH_FIN))
14109 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14110 if ((error == 0) && (flags & TH_RST))
14111 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14112 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
14114 * Calculate/Re-Calculate the hptsi slot in usecs based on
14115 * what we have sent so far
14117 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
14118 if (bbr->rc_no_pacing)
14121 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14123 if (bbr->rc_use_google == 0)
14124 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
14125 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14126 bbr->r_ctl.rc_lost_bytes)));
14127 bbr->rc_output_starts_timer = 1;
14128 if (bbr->bbr_use_rack_cheat &&
14130 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14131 /* Rack cheats and shotguns out all rxt's 1ms apart */
14135 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14137 * We don't change the tso size until some number of sends
14138 * to give the hardware commands time to get down
14139 * to the interface.
14141 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14142 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14143 bbr->hw_pacing_set = 1;
14144 tcp_bbr_tso_size_check(bbr, cts);
14147 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14148 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14149 /* Make sure snd_nxt is drug up */
14150 tp->snd_nxt = tp->snd_max;
14157 * See bbr_output_wtime() for return values.
14160 bbr_output(struct tcpcb *tp)
14164 struct tcp_bbr *bbr;
14166 NET_EPOCH_ASSERT();
14168 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14169 INP_WLOCK_ASSERT(tp->t_inpcb);
14170 (void)tcp_get_usecs(&tv);
14171 ret = bbr_output_wtime(tp, &tv);
14176 bbr_mtu_chg(struct tcpcb *tp)
14178 struct tcp_bbr *bbr;
14179 struct bbr_sendmap *rsm, *frsm = NULL;
14183 * The MTU has changed. a) Clear the sack filter. b) Mark everything
14184 * over the current size as SACK_PASS so a retransmit will occur.
14187 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14188 maxseg = tp->t_maxseg - bbr->rc_last_options;
14189 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14190 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14191 /* Don't mess with ones acked (by sack?) */
14192 if (rsm->r_flags & BBR_ACKED)
14194 if ((rsm->r_end - rsm->r_start) > maxseg) {
14196 * We mark sack-passed on all the previous large
14197 * sends we did. This will force them to retransmit.
14199 rsm->r_flags |= BBR_SACK_PASSED;
14200 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14201 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14202 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14203 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14204 rsm->r_flags |= BBR_MARKED_LOST;
14211 bbr->r_ctl.rc_resend = frsm;
14216 bbr_pru_options(struct tcpcb *tp, int flags)
14218 if (flags & PRUS_OOB)
14219 return (EOPNOTSUPP);
14223 struct tcp_function_block __tcp_bbr = {
14224 .tfb_tcp_block_name = __XSTRING(STACKNAME),
14225 .tfb_tcp_output = bbr_output,
14226 .tfb_do_queued_segments = ctf_do_queued_segments,
14227 .tfb_do_segment_nounlock = bbr_do_segment_nounlock,
14228 .tfb_tcp_do_segment = bbr_do_segment,
14229 .tfb_tcp_ctloutput = bbr_ctloutput,
14230 .tfb_tcp_fb_init = bbr_init,
14231 .tfb_tcp_fb_fini = bbr_fini,
14232 .tfb_tcp_timer_stop_all = bbr_stopall,
14233 .tfb_tcp_timer_activate = bbr_timer_activate,
14234 .tfb_tcp_timer_active = bbr_timer_active,
14235 .tfb_tcp_timer_stop = bbr_timer_stop,
14236 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
14237 .tfb_tcp_handoff_ok = bbr_handoff_ok,
14238 .tfb_tcp_mtu_chg = bbr_mtu_chg,
14239 .tfb_pru_options = bbr_pru_options,
14243 * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14244 * socket option arguments. When it re-acquires the lock after the copy, it
14245 * has to revalidate that the connection is still valid for the socket
14249 bbr_set_sockopt(struct socket *so, struct sockopt *sopt,
14250 struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14252 struct epoch_tracker et;
14253 int32_t error = 0, optval;
14255 switch (sopt->sopt_name) {
14256 case TCP_RACK_PACE_MAX_SEG:
14257 case TCP_RACK_MIN_TO:
14258 case TCP_RACK_REORD_THRESH:
14259 case TCP_RACK_REORD_FADE:
14260 case TCP_RACK_TLP_THRESH:
14261 case TCP_RACK_PKT_DELAY:
14262 case TCP_BBR_ALGORITHM:
14263 case TCP_BBR_TSLIMITS:
14264 case TCP_BBR_IWINTSO:
14265 case TCP_BBR_RECFORCE:
14266 case TCP_BBR_STARTUP_PG:
14267 case TCP_BBR_DRAIN_PG:
14268 case TCP_BBR_RWND_IS_APP:
14269 case TCP_BBR_PROBE_RTT_INT:
14270 case TCP_BBR_PROBE_RTT_GAIN:
14271 case TCP_BBR_PROBE_RTT_LEN:
14272 case TCP_BBR_STARTUP_LOSS_EXIT:
14273 case TCP_BBR_USEDEL_RATE:
14274 case TCP_BBR_MIN_RTO:
14275 case TCP_BBR_MAX_RTO:
14276 case TCP_BBR_PACE_PER_SEC:
14278 case TCP_BBR_PACE_DEL_TAR:
14279 case TCP_BBR_SEND_IWND_IN_TSO:
14280 case TCP_BBR_EXTRA_STATE:
14281 case TCP_BBR_UTTER_MAX_TSO:
14282 case TCP_BBR_MIN_TOPACEOUT:
14283 case TCP_BBR_FLOOR_MIN_TSO:
14284 case TCP_BBR_TSTMP_RAISES:
14285 case TCP_BBR_POLICER_DETECT:
14286 case TCP_BBR_USE_RACK_CHEAT:
14287 case TCP_DATA_AFTER_CLOSE:
14288 case TCP_BBR_HDWR_PACE:
14289 case TCP_BBR_PACE_SEG_MAX:
14290 case TCP_BBR_PACE_SEG_MIN:
14291 case TCP_BBR_PACE_CROSS:
14292 case TCP_BBR_PACE_OH:
14293 #ifdef NETFLIX_PEAKRATE
14294 case TCP_MAXPEAKRATE:
14296 case TCP_BBR_TMR_PACE_OH:
14297 case TCP_BBR_RACK_RTT_USE:
14298 case TCP_BBR_RETRAN_WTSO:
14301 return (tcp_default_ctloutput(so, sopt, inp, tp));
14305 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14309 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14311 return (ECONNRESET);
14313 tp = intotcpcb(inp);
14314 if (tp->t_fb != &__tcp_bbr) {
14316 return (ENOPROTOOPT);
14318 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14319 switch (sopt->sopt_name) {
14320 case TCP_BBR_PACE_PER_SEC:
14321 BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14322 bbr->r_ctl.bbr_hptsi_per_second = optval;
14324 case TCP_BBR_PACE_DEL_TAR:
14325 BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14326 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14328 case TCP_BBR_PACE_SEG_MAX:
14329 BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14330 bbr->r_ctl.bbr_hptsi_segments_max = optval;
14332 case TCP_BBR_PACE_SEG_MIN:
14333 BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14334 bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14336 case TCP_BBR_PACE_CROSS:
14337 BBR_OPTS_INC(tcp_bbr_pace_cross);
14338 bbr->r_ctl.bbr_cross_over = optval;
14340 case TCP_BBR_ALGORITHM:
14341 BBR_OPTS_INC(tcp_bbr_algorithm);
14342 if (optval && (bbr->rc_use_google == 0)) {
14343 /* Turn on the google mode */
14344 bbr_google_mode_on(bbr);
14345 if ((optval > 3) && (optval < 500)) {
14347 * Must be at least greater than .3%
14348 * and must be less than 50.0%.
14350 bbr->r_ctl.bbr_google_discount = optval;
14352 } else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14353 /* Turn off the google mode */
14354 bbr_google_mode_off(bbr);
14357 case TCP_BBR_TSLIMITS:
14358 BBR_OPTS_INC(tcp_bbr_tslimits);
14360 bbr->rc_use_ts_limit = 1;
14361 else if (optval == 0)
14362 bbr->rc_use_ts_limit = 0;
14367 case TCP_BBR_IWINTSO:
14368 BBR_OPTS_INC(tcp_bbr_iwintso);
14369 if ((optval >= 0) && (optval < 128)) {
14372 bbr->rc_init_win = optval;
14373 twin = bbr_initial_cwnd(bbr, tp);
14374 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14375 tp->snd_cwnd = twin;
14381 case TCP_BBR_STARTUP_PG:
14382 BBR_OPTS_INC(tcp_bbr_startup_pg);
14383 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14384 bbr->r_ctl.rc_startup_pg = optval;
14385 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14386 bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14391 case TCP_BBR_DRAIN_PG:
14392 BBR_OPTS_INC(tcp_bbr_drain_pg);
14393 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14394 bbr->r_ctl.rc_drain_pg = optval;
14398 case TCP_BBR_PROBE_RTT_LEN:
14399 BBR_OPTS_INC(tcp_bbr_probertt_len);
14401 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14405 case TCP_BBR_PROBE_RTT_GAIN:
14406 BBR_OPTS_INC(tcp_bbr_probertt_gain);
14407 if (optval <= BBR_UNIT)
14408 bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14412 case TCP_BBR_PROBE_RTT_INT:
14413 BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14415 bbr->r_ctl.rc_probertt_int = optval;
14419 case TCP_BBR_MIN_TOPACEOUT:
14420 BBR_OPTS_INC(tcp_bbr_topaceout);
14422 bbr->no_pacing_until = 0;
14423 bbr->rc_no_pacing = 0;
14424 } else if (optval <= 0x00ff) {
14425 bbr->no_pacing_until = optval;
14426 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14427 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14428 /* Turn on no pacing */
14429 bbr->rc_no_pacing = 1;
14434 case TCP_BBR_STARTUP_LOSS_EXIT:
14435 BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14436 bbr->rc_loss_exit = optval;
14438 case TCP_BBR_USEDEL_RATE:
14441 case TCP_BBR_MIN_RTO:
14442 BBR_OPTS_INC(tcp_bbr_min_rto);
14443 bbr->r_ctl.rc_min_rto_ms = optval;
14445 case TCP_BBR_MAX_RTO:
14446 BBR_OPTS_INC(tcp_bbr_max_rto);
14447 bbr->rc_max_rto_sec = optval;
14449 case TCP_RACK_MIN_TO:
14450 /* Minimum time between rack t-o's in ms */
14451 BBR_OPTS_INC(tcp_rack_min_to);
14452 bbr->r_ctl.rc_min_to = optval;
14454 case TCP_RACK_REORD_THRESH:
14455 /* RACK reorder threshold (shift amount) */
14456 BBR_OPTS_INC(tcp_rack_reord_thresh);
14457 if ((optval > 0) && (optval < 31))
14458 bbr->r_ctl.rc_reorder_shift = optval;
14462 case TCP_RACK_REORD_FADE:
14463 /* Does reordering fade after ms time */
14464 BBR_OPTS_INC(tcp_rack_reord_fade);
14465 bbr->r_ctl.rc_reorder_fade = optval;
14467 case TCP_RACK_TLP_THRESH:
14468 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14469 BBR_OPTS_INC(tcp_rack_tlp_thresh);
14471 bbr->rc_tlp_threshold = optval;
14475 case TCP_BBR_USE_RACK_CHEAT:
14476 BBR_OPTS_INC(tcp_use_rackcheat);
14477 if (bbr->rc_use_google) {
14481 BBR_OPTS_INC(tcp_rack_cheat);
14483 bbr->bbr_use_rack_cheat = 1;
14485 bbr->bbr_use_rack_cheat = 0;
14487 case TCP_BBR_FLOOR_MIN_TSO:
14488 BBR_OPTS_INC(tcp_utter_max_tso);
14489 if ((optval >= 0) && (optval < 40))
14490 bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14494 case TCP_BBR_UTTER_MAX_TSO:
14495 BBR_OPTS_INC(tcp_utter_max_tso);
14496 if ((optval >= 0) && (optval < 0xffff))
14497 bbr->r_ctl.bbr_utter_max = optval;
14502 case TCP_BBR_EXTRA_STATE:
14503 BBR_OPTS_INC(tcp_extra_state);
14505 bbr->rc_use_idle_restart = 1;
14507 bbr->rc_use_idle_restart = 0;
14509 case TCP_BBR_SEND_IWND_IN_TSO:
14510 BBR_OPTS_INC(tcp_iwnd_tso);
14512 bbr->bbr_init_win_cheat = 1;
14513 if (bbr->rc_past_init_win == 0) {
14515 cts = tcp_get_usecs(&bbr->rc_tv);
14516 tcp_bbr_tso_size_check(bbr, cts);
14519 bbr->bbr_init_win_cheat = 0;
14521 case TCP_BBR_HDWR_PACE:
14522 BBR_OPTS_INC(tcp_hdwr_pacing);
14524 bbr->bbr_hdw_pace_ena = 1;
14525 bbr->bbr_attempt_hdwr_pace = 0;
14527 bbr->bbr_hdw_pace_ena = 0;
14529 if (bbr->r_ctl.crte != NULL) {
14530 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp);
14531 bbr->r_ctl.crte = NULL;
14538 BBR_OPTS_INC(tcp_delack);
14539 if (optval < 100) {
14540 if (optval == 0) /* off */
14541 tp->t_delayed_ack = 0;
14542 else if (optval == 1) /* on which is 2 */
14543 tp->t_delayed_ack = 2;
14544 else /* higher than 2 and less than 100 */
14545 tp->t_delayed_ack = optval;
14546 if (tp->t_flags & TF_DELACK) {
14547 tp->t_flags &= ~TF_DELACK;
14548 tp->t_flags |= TF_ACKNOW;
14549 NET_EPOCH_ENTER(et);
14551 NET_EPOCH_EXIT(et);
14556 case TCP_RACK_PKT_DELAY:
14557 /* RACK added ms i.e. rack-rtt + reord + N */
14558 BBR_OPTS_INC(tcp_rack_pkt_delay);
14559 bbr->r_ctl.rc_pkt_delay = optval;
14561 #ifdef NETFLIX_PEAKRATE
14562 case TCP_MAXPEAKRATE:
14563 BBR_OPTS_INC(tcp_maxpeak);
14564 error = tcp_set_maxpeakrate(tp, optval);
14566 tp->t_peakrate_thr = tp->t_maxpeakrate;
14569 case TCP_BBR_RETRAN_WTSO:
14570 BBR_OPTS_INC(tcp_retran_wtso);
14572 bbr->rc_resends_use_tso = 1;
14574 bbr->rc_resends_use_tso = 0;
14576 case TCP_DATA_AFTER_CLOSE:
14577 BBR_OPTS_INC(tcp_data_ac);
14579 bbr->rc_allow_data_af_clo = 1;
14581 bbr->rc_allow_data_af_clo = 0;
14583 case TCP_BBR_POLICER_DETECT:
14584 BBR_OPTS_INC(tcp_policer_det);
14585 if (bbr->rc_use_google == 0)
14588 bbr->r_use_policer = 1;
14590 bbr->r_use_policer = 0;
14593 case TCP_BBR_TSTMP_RAISES:
14594 BBR_OPTS_INC(tcp_ts_raises);
14596 bbr->ts_can_raise = 1;
14598 bbr->ts_can_raise = 0;
14600 case TCP_BBR_TMR_PACE_OH:
14601 BBR_OPTS_INC(tcp_pacing_oh_tmr);
14602 if (bbr->rc_use_google) {
14606 bbr->r_ctl.rc_incr_tmrs = 1;
14608 bbr->r_ctl.rc_incr_tmrs = 0;
14611 case TCP_BBR_PACE_OH:
14612 BBR_OPTS_INC(tcp_pacing_oh);
14613 if (bbr->rc_use_google) {
14616 if (optval > (BBR_INCL_TCP_OH|
14618 BBR_INCL_ENET_OH)) {
14622 if (optval & BBR_INCL_TCP_OH)
14623 bbr->r_ctl.rc_inc_tcp_oh = 1;
14625 bbr->r_ctl.rc_inc_tcp_oh = 0;
14626 if (optval & BBR_INCL_IP_OH)
14627 bbr->r_ctl.rc_inc_ip_oh = 1;
14629 bbr->r_ctl.rc_inc_ip_oh = 0;
14630 if (optval & BBR_INCL_ENET_OH)
14631 bbr->r_ctl.rc_inc_enet_oh = 1;
14633 bbr->r_ctl.rc_inc_enet_oh = 0;
14637 return (tcp_default_ctloutput(so, sopt, inp, tp));
14640 #ifdef NETFLIX_STATS
14641 tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14648 * return 0 on success, error-num on failure
14651 bbr_get_sockopt(struct socket *so, struct sockopt *sopt,
14652 struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14654 int32_t error, optval;
14657 * Because all our options are either boolean or an int, we can just
14658 * pull everything into optval and then unlock and copy. If we ever
14659 * add a option that is not a int, then this will have quite an
14660 * impact to this routine.
14662 switch (sopt->sopt_name) {
14663 case TCP_BBR_PACE_PER_SEC:
14664 optval = bbr->r_ctl.bbr_hptsi_per_second;
14666 case TCP_BBR_PACE_DEL_TAR:
14667 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14669 case TCP_BBR_PACE_SEG_MAX:
14670 optval = bbr->r_ctl.bbr_hptsi_segments_max;
14672 case TCP_BBR_MIN_TOPACEOUT:
14673 optval = bbr->no_pacing_until;
14675 case TCP_BBR_PACE_SEG_MIN:
14676 optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14678 case TCP_BBR_PACE_CROSS:
14679 optval = bbr->r_ctl.bbr_cross_over;
14681 case TCP_BBR_ALGORITHM:
14682 optval = bbr->rc_use_google;
14684 case TCP_BBR_TSLIMITS:
14685 optval = bbr->rc_use_ts_limit;
14687 case TCP_BBR_IWINTSO:
14688 optval = bbr->rc_init_win;
14690 case TCP_BBR_STARTUP_PG:
14691 optval = bbr->r_ctl.rc_startup_pg;
14693 case TCP_BBR_DRAIN_PG:
14694 optval = bbr->r_ctl.rc_drain_pg;
14696 case TCP_BBR_PROBE_RTT_INT:
14697 optval = bbr->r_ctl.rc_probertt_int;
14699 case TCP_BBR_PROBE_RTT_LEN:
14700 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14702 case TCP_BBR_PROBE_RTT_GAIN:
14703 optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14705 case TCP_BBR_STARTUP_LOSS_EXIT:
14706 optval = bbr->rc_loss_exit;
14708 case TCP_BBR_USEDEL_RATE:
14711 case TCP_BBR_MIN_RTO:
14712 optval = bbr->r_ctl.rc_min_rto_ms;
14714 case TCP_BBR_MAX_RTO:
14715 optval = bbr->rc_max_rto_sec;
14717 case TCP_RACK_PACE_MAX_SEG:
14718 /* Max segments in a pace */
14719 optval = bbr->r_ctl.rc_pace_max_segs;
14721 case TCP_RACK_MIN_TO:
14722 /* Minimum time between rack t-o's in ms */
14723 optval = bbr->r_ctl.rc_min_to;
14725 case TCP_RACK_REORD_THRESH:
14726 /* RACK reorder threshold (shift amount) */
14727 optval = bbr->r_ctl.rc_reorder_shift;
14729 case TCP_RACK_REORD_FADE:
14730 /* Does reordering fade after ms time */
14731 optval = bbr->r_ctl.rc_reorder_fade;
14733 case TCP_BBR_USE_RACK_CHEAT:
14734 /* Do we use the rack cheat for rxt */
14735 optval = bbr->bbr_use_rack_cheat;
14737 case TCP_BBR_FLOOR_MIN_TSO:
14738 optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14740 case TCP_BBR_UTTER_MAX_TSO:
14741 optval = bbr->r_ctl.bbr_utter_max;
14743 case TCP_BBR_SEND_IWND_IN_TSO:
14744 /* Do we send TSO size segments initially */
14745 optval = bbr->bbr_init_win_cheat;
14747 case TCP_BBR_EXTRA_STATE:
14748 optval = bbr->rc_use_idle_restart;
14750 case TCP_RACK_TLP_THRESH:
14751 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14752 optval = bbr->rc_tlp_threshold;
14754 case TCP_RACK_PKT_DELAY:
14755 /* RACK added ms i.e. rack-rtt + reord + N */
14756 optval = bbr->r_ctl.rc_pkt_delay;
14758 case TCP_BBR_RETRAN_WTSO:
14759 optval = bbr->rc_resends_use_tso;
14761 case TCP_DATA_AFTER_CLOSE:
14762 optval = bbr->rc_allow_data_af_clo;
14765 optval = tp->t_delayed_ack;
14767 case TCP_BBR_HDWR_PACE:
14768 optval = bbr->bbr_hdw_pace_ena;
14770 case TCP_BBR_POLICER_DETECT:
14771 optval = bbr->r_use_policer;
14773 case TCP_BBR_TSTMP_RAISES:
14774 optval = bbr->ts_can_raise;
14776 case TCP_BBR_TMR_PACE_OH:
14777 optval = bbr->r_ctl.rc_incr_tmrs;
14779 case TCP_BBR_PACE_OH:
14781 if (bbr->r_ctl.rc_inc_tcp_oh)
14782 optval |= BBR_INCL_TCP_OH;
14783 if (bbr->r_ctl.rc_inc_ip_oh)
14784 optval |= BBR_INCL_IP_OH;
14785 if (bbr->r_ctl.rc_inc_enet_oh)
14786 optval |= BBR_INCL_ENET_OH;
14789 return (tcp_default_ctloutput(so, sopt, inp, tp));
14793 error = sooptcopyout(sopt, &optval, sizeof optval);
14798 * return 0 on success, error-num on failure
14801 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14803 int32_t error = EINVAL;
14804 struct tcp_bbr *bbr;
14806 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14811 if (sopt->sopt_dir == SOPT_SET) {
14812 return (bbr_set_sockopt(so, sopt, inp, tp, bbr));
14813 } else if (sopt->sopt_dir == SOPT_GET) {
14814 return (bbr_get_sockopt(so, sopt, inp, tp, bbr));
14821 static const char *bbr_stack_names[] = {
14822 __XSTRING(STACKNAME),
14824 __XSTRING(STACKALIAS),
14828 static bool bbr_mod_inited = false;
14831 tcp_addbbr(module_t mod, int32_t type, void *data)
14838 printf("Attempting to load " __XSTRING(MODNAME) "\n");
14839 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14840 sizeof(struct bbr_sendmap),
14841 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
14842 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14843 sizeof(struct tcp_bbr),
14844 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14845 sysctl_ctx_init(&bbr_sysctl_ctx);
14846 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
14847 SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14850 __XSTRING(STACKALIAS),
14852 __XSTRING(STACKNAME),
14854 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14856 if (bbr_sysctl_root == NULL) {
14857 printf("Failed to add sysctl node\n");
14861 bbr_init_sysctls();
14862 num_stacks = nitems(bbr_stack_names);
14863 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
14864 bbr_stack_names, &num_stacks);
14866 printf("Failed to register %s stack name for "
14867 "%s module\n", bbr_stack_names[num_stacks],
14868 __XSTRING(MODNAME));
14869 sysctl_ctx_free(&bbr_sysctl_ctx);
14871 uma_zdestroy(bbr_zone);
14872 uma_zdestroy(bbr_pcb_zone);
14873 bbr_counter_destroy();
14874 printf("Failed to register " __XSTRING(MODNAME)
14875 " module err:%d\n", err);
14878 tcp_lro_reg_mbufq();
14879 bbr_mod_inited = true;
14880 printf(__XSTRING(MODNAME) " is now available\n");
14883 err = deregister_tcp_functions(&__tcp_bbr, true, false);
14886 err = deregister_tcp_functions(&__tcp_bbr, false, true);
14889 if (bbr_mod_inited) {
14890 uma_zdestroy(bbr_zone);
14891 uma_zdestroy(bbr_pcb_zone);
14892 sysctl_ctx_free(&bbr_sysctl_ctx);
14893 bbr_counter_destroy();
14894 printf(__XSTRING(MODNAME)
14895 " is now no longer available\n");
14896 bbr_mod_inited = false;
14898 tcp_lro_dereg_mbufq();
14902 return (EOPNOTSUPP);
14907 static moduledata_t tcp_bbr = {
14908 .name = __XSTRING(MODNAME),
14909 .evhand = tcp_addbbr,
14913 MODULE_VERSION(MODNAME, 1);
14914 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14915 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);