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 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_tcpdebug.h"
34 #include "opt_ratelimit.h"
35 #include <sys/param.h>
37 #include <sys/module.h>
38 #include <sys/kernel.h>
40 #include <sys/hhook.h>
43 #include <sys/malloc.h>
45 #include <sys/mutex.h>
47 #include <sys/proc.h> /* for proc0 declaration */
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
53 #include <sys/qmath.h>
55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
59 #include <sys/refcount.h>
60 #include <sys/queue.h>
61 #include <sys/tim_filter.h>
63 #include <sys/kthread.h>
64 #include <sys/kern_prefetch.h>
65 #include <sys/protosw.h>
67 #include <sys/sched.h>
68 #include <machine/cpu.h>
72 #include <net/route.h>
73 #include <net/route/nhop.h>
76 #define TCPSTATES /* for logging */
78 #include <netinet/in.h>
79 #include <netinet/in_kdtrace.h>
80 #include <netinet/in_pcb.h>
81 #include <netinet/ip.h>
82 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
83 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
84 #include <netinet/ip_var.h>
85 #include <netinet/ip6.h>
86 #include <netinet6/in6_pcb.h>
87 #include <netinet6/ip6_var.h>
88 #include <netinet/tcp.h>
90 #include <netinet/tcp_fsm.h>
91 #include <netinet/tcp_log_buf.h>
92 #include <netinet/tcp_seq.h>
93 #include <netinet/tcp_timer.h>
94 #include <netinet/tcp_var.h>
95 #include <netinet/tcp_hpts.h>
96 #include <netinet/tcp_ratelimit.h>
97 #include <netinet/tcp_accounting.h>
98 #include <netinet/tcpip.h>
99 #include <netinet/cc/cc.h>
100 #include <netinet/cc/cc_newreno.h>
101 #include <netinet/tcp_fastopen.h>
102 #include <netinet/tcp_lro.h>
103 #ifdef NETFLIX_SHARED_CWND
104 #include <netinet/tcp_shared_cwnd.h>
107 #include <netinet/tcp_debug.h>
108 #endif /* TCPDEBUG */
110 #include <netinet/tcp_offload.h>
113 #include <netinet6/tcp6_var.h>
116 #include <netipsec/ipsec_support.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>
130 #include "sack_filter.h"
131 #include "tcp_rack.h"
132 #include "rack_bbr_common.h"
134 uma_zone_t rack_zone;
135 uma_zone_t rack_pcb_zone;
138 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t)))
141 VNET_DECLARE(uint32_t, newreno_beta);
142 VNET_DECLARE(uint32_t, newreno_beta_ecn);
143 #define V_newreno_beta VNET(newreno_beta)
144 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
147 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
148 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
150 struct sysctl_ctx_list rack_sysctl_ctx;
151 struct sysctl_oid *rack_sysctl_root;
157 * The RACK module incorporates a number of
158 * TCP ideas that have been put out into the IETF
159 * over the last few years:
160 * - Matt Mathis's Rate Halving which slowly drops
161 * the congestion window so that the ack clock can
162 * be maintained during a recovery.
163 * - Yuchung Cheng's RACK TCP (for which its named) that
164 * will stop us using the number of dup acks and instead
165 * use time as the gage of when we retransmit.
166 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
167 * of Dukkipati et.al.
168 * RACK depends on SACK, so if an endpoint arrives that
169 * cannot do SACK the state machine below will shuttle the
170 * connection back to using the "default" TCP stack that is
173 * To implement RACK the original TCP stack was first decomposed
174 * into a functional state machine with individual states
175 * for each of the possible TCP connection states. The do_segement
176 * functions role in life is to mandate the connection supports SACK
177 * initially and then assure that the RACK state matches the conenction
178 * state before calling the states do_segment function. Each
179 * state is simplified due to the fact that the original do_segment
180 * has been decomposed and we *know* what state we are in (no
181 * switches on the state) and all tests for SACK are gone. This
182 * greatly simplifies what each state does.
184 * TCP output is also over-written with a new version since it
185 * must maintain the new rack scoreboard.
188 static int32_t rack_tlp_thresh = 1;
189 static int32_t rack_tlp_limit = 2; /* No more than 2 TLPs w-out new data */
190 static int32_t rack_tlp_use_greater = 1;
191 static int32_t rack_reorder_thresh = 2;
192 static int32_t rack_reorder_fade = 60000000; /* 0 - never fade, def 60,000,000
194 static uint8_t rack_req_measurements = 1;
195 /* Attack threshold detections */
196 static uint32_t rack_highest_sack_thresh_seen = 0;
197 static uint32_t rack_highest_move_thresh_seen = 0;
198 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
199 static int32_t rack_hw_pace_extra_slots = 2; /* 2 extra MSS time betweens */
200 static int32_t rack_hw_rate_caps = 1; /* 1; */
201 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
202 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
203 static int32_t rack_hw_up_only = 1;
204 static int32_t rack_stats_gets_ms_rtt = 1;
205 static int32_t rack_prr_addbackmax = 2;
207 static int32_t rack_pkt_delay = 1000;
208 static int32_t rack_send_a_lot_in_prr = 1;
209 static int32_t rack_min_to = 1000; /* Number of microsecond min timeout */
210 static int32_t rack_verbose_logging = 0;
211 static int32_t rack_ignore_data_after_close = 1;
212 static int32_t rack_enable_shared_cwnd = 1;
213 static int32_t rack_use_cmp_acks = 1;
214 static int32_t rack_use_fsb = 1;
215 static int32_t rack_use_rfo = 1;
216 static int32_t rack_use_rsm_rfo = 1;
217 static int32_t rack_max_abc_post_recovery = 2;
218 static int32_t rack_client_low_buf = 0;
219 #ifdef TCP_ACCOUNTING
220 static int32_t rack_tcp_accounting = 0;
222 static int32_t rack_limits_scwnd = 1;
223 static int32_t rack_enable_mqueue_for_nonpaced = 0;
224 static int32_t rack_disable_prr = 0;
225 static int32_t use_rack_rr = 1;
226 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
227 static int32_t rack_persist_min = 250000; /* 250usec */
228 static int32_t rack_persist_max = 2000000; /* 2 Second in usec's */
229 static int32_t rack_sack_not_required = 1; /* set to one to allow non-sack to use rack */
230 static int32_t rack_default_init_window = 0; /* Use system default */
231 static int32_t rack_limit_time_with_srtt = 0;
232 static int32_t rack_autosndbuf_inc = 20; /* In percentage form */
233 static int32_t rack_enobuf_hw_boost_mult = 2; /* How many times the hw rate we boost slot using time_between */
234 static int32_t rack_enobuf_hw_max = 12000; /* 12 ms in usecs */
235 static int32_t rack_enobuf_hw_min = 10000; /* 10 ms in usecs */
236 static int32_t rack_hw_rwnd_factor = 2; /* How many max_segs the rwnd must be before we hold off sending */
238 * Currently regular tcp has a rto_min of 30ms
239 * the backoff goes 12 times so that ends up
240 * being a total of 122.850 seconds before a
241 * connection is killed.
243 static uint32_t rack_def_data_window = 20;
244 static uint32_t rack_goal_bdp = 2;
245 static uint32_t rack_min_srtts = 1;
246 static uint32_t rack_min_measure_usec = 0;
247 static int32_t rack_tlp_min = 10000; /* 10ms */
248 static int32_t rack_rto_min = 30000; /* 30,000 usec same as main freebsd */
249 static int32_t rack_rto_max = 4000000; /* 4 seconds in usec's */
250 static const int32_t rack_free_cache = 2;
251 static int32_t rack_hptsi_segments = 40;
252 static int32_t rack_rate_sample_method = USE_RTT_LOW;
253 static int32_t rack_pace_every_seg = 0;
254 static int32_t rack_delayed_ack_time = 40000; /* 40ms in usecs */
255 static int32_t rack_slot_reduction = 4;
256 static int32_t rack_wma_divisor = 8; /* For WMA calculation */
257 static int32_t rack_cwnd_block_ends_measure = 0;
258 static int32_t rack_rwnd_block_ends_measure = 0;
259 static int32_t rack_def_profile = 0;
261 static int32_t rack_lower_cwnd_at_tlp = 0;
262 static int32_t rack_limited_retran = 0;
263 static int32_t rack_always_send_oldest = 0;
264 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
266 static uint16_t rack_per_of_gp_ss = 250; /* 250 % slow-start */
267 static uint16_t rack_per_of_gp_ca = 200; /* 200 % congestion-avoidance */
268 static uint16_t rack_per_of_gp_rec = 200; /* 200 % of bw */
271 static uint16_t rack_per_of_gp_probertt = 60; /* 60% of bw */
272 static uint16_t rack_per_of_gp_lowthresh = 40; /* 40% is bottom */
273 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
274 static uint16_t rack_atexit_prtt_hbp = 130; /* Clamp to 130% on exit prtt if highly buffered path */
275 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
277 static uint32_t rack_max_drain_wait = 2; /* How man gp srtt's before we give up draining */
278 static uint32_t rack_must_drain = 1; /* How many GP srtt's we *must* wait */
279 static uint32_t rack_probertt_use_min_rtt_entry = 1; /* Use the min to calculate the goal else gp_srtt */
280 static uint32_t rack_probertt_use_min_rtt_exit = 0;
281 static uint32_t rack_probe_rtt_sets_cwnd = 0;
282 static uint32_t rack_probe_rtt_safety_val = 2000000; /* No more than 2 sec in probe-rtt */
283 static uint32_t rack_time_between_probertt = 9600000; /* 9.6 sec in usecs */
284 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0; /* How many srtt periods does probe-rtt last top fraction */
285 static uint32_t rack_probertt_gpsrtt_cnt_div = 0; /* How many srtt periods does probe-rtt last bottom fraction */
286 static uint32_t rack_min_probertt_hold = 40000; /* Equal to delayed ack time */
287 static uint32_t rack_probertt_filter_life = 10000000;
288 static uint32_t rack_probertt_lower_within = 10;
289 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds) to count as a lowering */
290 static int32_t rack_pace_one_seg = 0; /* Shall we pace for less than 1.4Meg 1MSS at a time */
291 static int32_t rack_probertt_clear_is = 1;
292 static int32_t rack_max_drain_hbp = 1; /* Extra drain times gpsrtt for highly buffered paths */
293 static int32_t rack_hbp_thresh = 3; /* what is the divisor max_rtt/min_rtt to decided a hbp */
296 static int32_t rack_max_per_above = 30; /* When we go to increment stop if above 100+this% */
298 /* Timely information */
299 /* Combine these two gives the range of 'no change' to bw */
300 /* ie the up/down provide the upper and lower bound */
301 static int32_t rack_gp_per_bw_mul_up = 2; /* 2% */
302 static int32_t rack_gp_per_bw_mul_down = 4; /* 4% */
303 static int32_t rack_gp_rtt_maxmul = 3; /* 3 x maxmin */
304 static int32_t rack_gp_rtt_minmul = 1; /* minrtt + (minrtt/mindiv) is lower rtt */
305 static int32_t rack_gp_rtt_mindiv = 4; /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
306 static int32_t rack_gp_decrease_per = 20; /* 20% decrease in multipler */
307 static int32_t rack_gp_increase_per = 2; /* 2% increase in multipler */
308 static int32_t rack_per_lower_bound = 50; /* Don't allow to drop below this multiplier */
309 static int32_t rack_per_upper_bound_ss = 0; /* Don't allow SS to grow above this */
310 static int32_t rack_per_upper_bound_ca = 0; /* Don't allow CA to grow above this */
311 static int32_t rack_do_dyn_mul = 0; /* Are the rack gp multipliers dynamic */
312 static int32_t rack_gp_no_rec_chg = 1; /* Prohibit recovery from reducing it's multiplier */
313 static int32_t rack_timely_dec_clear = 6; /* Do we clear decrement count at a value (6)? */
314 static int32_t rack_timely_max_push_rise = 3; /* One round of pushing */
315 static int32_t rack_timely_max_push_drop = 3; /* Three round of pushing */
316 static int32_t rack_timely_min_segs = 4; /* 4 segment minimum */
317 static int32_t rack_use_max_for_nobackoff = 0;
318 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
319 static int32_t rack_timely_no_stopping = 0;
320 static int32_t rack_down_raise_thresh = 100;
321 static int32_t rack_req_segs = 1;
322 static uint64_t rack_bw_rate_cap = 0;
324 /* Weird delayed ack mode */
325 static int32_t rack_use_imac_dack = 0;
326 /* Rack specific counters */
327 counter_u64_t rack_badfr;
328 counter_u64_t rack_badfr_bytes;
329 counter_u64_t rack_rtm_prr_retran;
330 counter_u64_t rack_rtm_prr_newdata;
331 counter_u64_t rack_timestamp_mismatch;
332 counter_u64_t rack_reorder_seen;
333 counter_u64_t rack_paced_segments;
334 counter_u64_t rack_unpaced_segments;
335 counter_u64_t rack_calc_zero;
336 counter_u64_t rack_calc_nonzero;
337 counter_u64_t rack_saw_enobuf;
338 counter_u64_t rack_saw_enobuf_hw;
339 counter_u64_t rack_saw_enetunreach;
340 counter_u64_t rack_per_timer_hole;
341 counter_u64_t rack_large_ackcmp;
342 counter_u64_t rack_small_ackcmp;
344 counter_u64_t rack_adjust_map_bw;
346 /* Tail loss probe counters */
347 counter_u64_t rack_tlp_tot;
348 counter_u64_t rack_tlp_newdata;
349 counter_u64_t rack_tlp_retran;
350 counter_u64_t rack_tlp_retran_bytes;
351 counter_u64_t rack_tlp_retran_fail;
352 counter_u64_t rack_to_tot;
353 counter_u64_t rack_to_arm_rack;
354 counter_u64_t rack_to_arm_tlp;
355 counter_u64_t rack_hot_alloc;
356 counter_u64_t rack_to_alloc;
357 counter_u64_t rack_to_alloc_hard;
358 counter_u64_t rack_to_alloc_emerg;
359 counter_u64_t rack_to_alloc_limited;
360 counter_u64_t rack_alloc_limited_conns;
361 counter_u64_t rack_split_limited;
363 #define MAX_NUM_OF_CNTS 13
364 counter_u64_t rack_proc_comp_ack[MAX_NUM_OF_CNTS];
365 counter_u64_t rack_multi_single_eq;
366 counter_u64_t rack_proc_non_comp_ack;
368 counter_u64_t rack_fto_send;
369 counter_u64_t rack_fto_rsm_send;
370 counter_u64_t rack_nfto_resend;
371 counter_u64_t rack_non_fto_send;
372 counter_u64_t rack_extended_rfo;
374 counter_u64_t rack_sack_proc_all;
375 counter_u64_t rack_sack_proc_short;
376 counter_u64_t rack_sack_proc_restart;
377 counter_u64_t rack_sack_attacks_detected;
378 counter_u64_t rack_sack_attacks_reversed;
379 counter_u64_t rack_sack_used_next_merge;
380 counter_u64_t rack_sack_splits;
381 counter_u64_t rack_sack_used_prev_merge;
382 counter_u64_t rack_sack_skipped_acked;
383 counter_u64_t rack_ack_total;
384 counter_u64_t rack_express_sack;
385 counter_u64_t rack_sack_total;
386 counter_u64_t rack_move_none;
387 counter_u64_t rack_move_some;
389 counter_u64_t rack_used_tlpmethod;
390 counter_u64_t rack_used_tlpmethod2;
391 counter_u64_t rack_enter_tlp_calc;
392 counter_u64_t rack_input_idle_reduces;
393 counter_u64_t rack_collapsed_win;
394 counter_u64_t rack_tlp_does_nada;
395 counter_u64_t rack_try_scwnd;
396 counter_u64_t rack_hw_pace_init_fail;
397 counter_u64_t rack_hw_pace_lost;
398 counter_u64_t rack_sbsndptr_right;
399 counter_u64_t rack_sbsndptr_wrong;
401 /* Temp CPU counters */
402 counter_u64_t rack_find_high;
404 counter_u64_t rack_progress_drops;
405 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
406 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
409 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
411 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax) do { \
412 (tv) = (value) + TICKS_2_USEC(tcp_rexmit_slop); \
413 if ((u_long)(tv) < (u_long)(tvmin)) \
415 if ((u_long)(tv) > (u_long)(tvmax)) \
420 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line);
423 rack_process_ack(struct mbuf *m, struct tcphdr *th,
424 struct socket *so, struct tcpcb *tp, struct tcpopt *to,
425 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
427 rack_process_data(struct mbuf *m, struct tcphdr *th,
428 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
429 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
431 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
432 uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
433 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
434 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
436 static struct rack_sendmap *
437 rack_check_recovery_mode(struct tcpcb *tp,
440 rack_cong_signal(struct tcpcb *tp,
441 uint32_t type, uint32_t ack);
442 static void rack_counter_destroy(void);
444 rack_ctloutput(struct socket *so, struct sockopt *sopt,
445 struct inpcb *inp, struct tcpcb *tp);
446 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
448 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
450 rack_do_segment(struct mbuf *m, struct tcphdr *th,
451 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
453 static void rack_dtor(void *mem, int32_t size, void *arg);
455 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
456 uint32_t flex1, uint32_t flex2,
457 uint32_t flex3, uint32_t flex4,
458 uint32_t flex5, uint32_t flex6,
459 uint16_t flex7, uint8_t mod);
461 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
462 uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm);
463 static struct rack_sendmap *
464 rack_find_high_nonack(struct tcp_rack *rack,
465 struct rack_sendmap *rsm);
466 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
467 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
468 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
470 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
471 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
473 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
474 tcp_seq th_ack, int line);
476 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
477 static int32_t rack_handoff_ok(struct tcpcb *tp);
478 static int32_t rack_init(struct tcpcb *tp);
479 static void rack_init_sysctls(void);
481 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
482 struct tcphdr *th, int entered_rec, int dup_ack_struck);
484 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
485 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
486 struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff);
489 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
490 struct rack_sendmap *rsm);
491 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
492 static int32_t rack_output(struct tcpcb *tp);
495 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
496 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
497 uint32_t cts, int *moved_two);
498 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
499 static void rack_remxt_tmr(struct tcpcb *tp);
501 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
502 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
503 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
504 static int32_t rack_stopall(struct tcpcb *tp);
506 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
508 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
509 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
510 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
512 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
513 struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
515 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
516 struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
518 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
519 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
520 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
522 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
523 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
524 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
526 rack_do_closing(struct mbuf *m, struct tcphdr *th,
527 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
528 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
530 rack_do_established(struct mbuf *m, struct tcphdr *th,
531 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
532 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
534 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
535 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
536 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
538 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
539 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
540 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
542 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
543 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
544 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
546 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
547 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
548 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
550 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
551 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
552 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
554 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
555 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
556 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
557 struct rack_sendmap *
558 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
560 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
561 uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
563 tcp_rack_partialack(struct tcpcb *tp);
565 rack_set_profile(struct tcp_rack *rack, int prof);
567 rack_apply_deferred_options(struct tcp_rack *rack);
569 int32_t rack_clear_counter=0;
572 rack_set_cc_pacing(struct tcp_rack *rack)
575 struct cc_newreno_opts opt;
576 struct newreno old, *ptr;
580 if (rack->rc_pacing_cc_set)
584 if (tp->cc_algo == NULL) {
586 printf("No cc algorithm?\n");
589 rack->rc_pacing_cc_set = 1;
590 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
591 /* Not new-reno we can't play games with beta! */
592 printf("cc_algo:%s is not NEWRENO:%s\n",
593 tp->cc_algo->name, CCALGONAME_NEWRENO);
596 ptr = ((struct newreno *)tp->ccv->cc_data);
597 if (CC_ALGO(tp)->ctl_output == NULL) {
598 /* Huh, why does new_reno no longer have a set function? */
599 printf("no ctl_output for algo:%s\n", tp->cc_algo->name);
603 /* Just the default values */
604 old.beta = V_newreno_beta_ecn;
605 old.beta_ecn = V_newreno_beta_ecn;
606 old.newreno_flags = 0;
608 old.beta = ptr->beta;
609 old.beta_ecn = ptr->beta_ecn;
610 old.newreno_flags = ptr->newreno_flags;
612 sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
613 sopt.sopt_dir = SOPT_SET;
614 opt.name = CC_NEWRENO_BETA;
615 opt.val = rack->r_ctl.rc_saved_beta.beta;
616 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
618 printf("Error returned by ctl_output %d\n", error);
622 * Hack alert we need to set in our newreno_flags
623 * so that Abe behavior is also applied.
625 ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN;
626 opt.name = CC_NEWRENO_BETA_ECN;
627 opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
628 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
630 printf("Error returned by ctl_output %d\n", error);
633 /* Save off the original values for restoral */
634 memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
636 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
637 union tcp_log_stackspecific log;
640 ptr = ((struct newreno *)tp->ccv->cc_data);
641 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
642 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
644 log.u_bbr.flex1 = ptr->beta;
645 log.u_bbr.flex2 = ptr->beta_ecn;
646 log.u_bbr.flex3 = ptr->newreno_flags;
648 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
649 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
650 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
651 log.u_bbr.flex7 = rack->gp_ready;
652 log.u_bbr.flex7 <<= 1;
653 log.u_bbr.flex7 |= rack->use_fixed_rate;
654 log.u_bbr.flex7 <<= 1;
655 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
656 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
658 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
659 0, &log, false, NULL, NULL, 0, &tv);
664 rack_undo_cc_pacing(struct tcp_rack *rack)
666 struct newreno old, *ptr;
669 if (rack->rc_pacing_cc_set == 0)
672 rack->rc_pacing_cc_set = 0;
673 if (tp->cc_algo == NULL)
676 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
677 /* Not new-reno nothing to do! */
680 ptr = ((struct newreno *)tp->ccv->cc_data);
683 * This happens at rack_fini() if the
684 * cc module gets freed on us. In that
685 * case we loose our "new" settings but
686 * thats ok, since the tcb is going away anyway.
690 /* Grab out our set values */
691 memcpy(&old, ptr, sizeof(struct newreno));
692 /* Copy back in the original values */
693 memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
694 /* Now save back the values we had set in (for when pacing is restored) */
695 memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
696 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
697 union tcp_log_stackspecific log;
700 ptr = ((struct newreno *)tp->ccv->cc_data);
701 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
702 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
703 log.u_bbr.flex1 = ptr->beta;
704 log.u_bbr.flex2 = ptr->beta_ecn;
705 log.u_bbr.flex3 = ptr->newreno_flags;
706 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
707 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
708 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
709 log.u_bbr.flex7 = rack->gp_ready;
710 log.u_bbr.flex7 <<= 1;
711 log.u_bbr.flex7 |= rack->use_fixed_rate;
712 log.u_bbr.flex7 <<= 1;
713 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
714 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
716 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
717 0, &log, false, NULL, NULL, 0, &tv);
721 #ifdef NETFLIX_PEAKRATE
723 rack_update_peakrate_thr(struct tcpcb *tp)
725 /* Keep in mind that t_maxpeakrate is in B/s. */
727 peak = uqmax((tp->t_maxseg * 2),
728 (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
729 tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
734 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
740 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
741 if (error || req->newptr == NULL)
744 error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
749 printf("Clearing RACK counters\n");
751 counter_u64_zero(rack_badfr);
752 counter_u64_zero(rack_badfr_bytes);
753 counter_u64_zero(rack_rtm_prr_retran);
754 counter_u64_zero(rack_rtm_prr_newdata);
755 counter_u64_zero(rack_timestamp_mismatch);
756 counter_u64_zero(rack_reorder_seen);
757 counter_u64_zero(rack_tlp_tot);
758 counter_u64_zero(rack_tlp_newdata);
759 counter_u64_zero(rack_tlp_retran);
760 counter_u64_zero(rack_tlp_retran_bytes);
761 counter_u64_zero(rack_tlp_retran_fail);
762 counter_u64_zero(rack_to_tot);
763 counter_u64_zero(rack_to_arm_rack);
764 counter_u64_zero(rack_to_arm_tlp);
765 counter_u64_zero(rack_paced_segments);
766 counter_u64_zero(rack_calc_zero);
767 counter_u64_zero(rack_calc_nonzero);
768 counter_u64_zero(rack_unpaced_segments);
769 counter_u64_zero(rack_saw_enobuf);
770 counter_u64_zero(rack_saw_enobuf_hw);
771 counter_u64_zero(rack_saw_enetunreach);
772 counter_u64_zero(rack_per_timer_hole);
773 counter_u64_zero(rack_large_ackcmp);
774 counter_u64_zero(rack_small_ackcmp);
776 counter_u64_zero(rack_adjust_map_bw);
778 counter_u64_zero(rack_to_alloc_hard);
779 counter_u64_zero(rack_to_alloc_emerg);
780 counter_u64_zero(rack_sack_proc_all);
781 counter_u64_zero(rack_fto_send);
782 counter_u64_zero(rack_fto_rsm_send);
783 counter_u64_zero(rack_extended_rfo);
784 counter_u64_zero(rack_hw_pace_init_fail);
785 counter_u64_zero(rack_hw_pace_lost);
786 counter_u64_zero(rack_sbsndptr_wrong);
787 counter_u64_zero(rack_sbsndptr_right);
788 counter_u64_zero(rack_non_fto_send);
789 counter_u64_zero(rack_nfto_resend);
790 counter_u64_zero(rack_sack_proc_short);
791 counter_u64_zero(rack_sack_proc_restart);
792 counter_u64_zero(rack_to_alloc);
793 counter_u64_zero(rack_to_alloc_limited);
794 counter_u64_zero(rack_alloc_limited_conns);
795 counter_u64_zero(rack_split_limited);
796 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
797 counter_u64_zero(rack_proc_comp_ack[i]);
799 counter_u64_zero(rack_multi_single_eq);
800 counter_u64_zero(rack_proc_non_comp_ack);
801 counter_u64_zero(rack_find_high);
802 counter_u64_zero(rack_sack_attacks_detected);
803 counter_u64_zero(rack_sack_attacks_reversed);
804 counter_u64_zero(rack_sack_used_next_merge);
805 counter_u64_zero(rack_sack_used_prev_merge);
806 counter_u64_zero(rack_sack_splits);
807 counter_u64_zero(rack_sack_skipped_acked);
808 counter_u64_zero(rack_ack_total);
809 counter_u64_zero(rack_express_sack);
810 counter_u64_zero(rack_sack_total);
811 counter_u64_zero(rack_move_none);
812 counter_u64_zero(rack_move_some);
813 counter_u64_zero(rack_used_tlpmethod);
814 counter_u64_zero(rack_used_tlpmethod2);
815 counter_u64_zero(rack_enter_tlp_calc);
816 counter_u64_zero(rack_progress_drops);
817 counter_u64_zero(rack_tlp_does_nada);
818 counter_u64_zero(rack_try_scwnd);
819 counter_u64_zero(rack_collapsed_win);
821 rack_clear_counter = 0;
826 rack_init_sysctls(void)
829 struct sysctl_oid *rack_counters;
830 struct sysctl_oid *rack_attack;
831 struct sysctl_oid *rack_pacing;
832 struct sysctl_oid *rack_timely;
833 struct sysctl_oid *rack_timers;
834 struct sysctl_oid *rack_tlp;
835 struct sysctl_oid *rack_misc;
836 struct sysctl_oid *rack_measure;
837 struct sysctl_oid *rack_probertt;
838 struct sysctl_oid *rack_hw_pacing;
840 rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
841 SYSCTL_CHILDREN(rack_sysctl_root),
844 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
845 "Rack Sack Attack Counters and Controls");
846 rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
847 SYSCTL_CHILDREN(rack_sysctl_root),
850 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
852 SYSCTL_ADD_S32(&rack_sysctl_ctx,
853 SYSCTL_CHILDREN(rack_sysctl_root),
854 OID_AUTO, "rate_sample_method", CTLFLAG_RW,
855 &rack_rate_sample_method , USE_RTT_LOW,
856 "What method should we use for rate sampling 0=high, 1=low ");
857 /* Probe rtt related controls */
858 rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
859 SYSCTL_CHILDREN(rack_sysctl_root),
862 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
863 "ProbeRTT related Controls");
864 SYSCTL_ADD_U16(&rack_sysctl_ctx,
865 SYSCTL_CHILDREN(rack_probertt),
866 OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
867 &rack_atexit_prtt_hbp, 130,
868 "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
869 SYSCTL_ADD_U16(&rack_sysctl_ctx,
870 SYSCTL_CHILDREN(rack_probertt),
871 OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
872 &rack_atexit_prtt, 130,
873 "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
874 SYSCTL_ADD_U16(&rack_sysctl_ctx,
875 SYSCTL_CHILDREN(rack_probertt),
876 OID_AUTO, "gp_per_mul", CTLFLAG_RW,
877 &rack_per_of_gp_probertt, 60,
878 "What percentage of goodput do we pace at in probertt");
879 SYSCTL_ADD_U16(&rack_sysctl_ctx,
880 SYSCTL_CHILDREN(rack_probertt),
881 OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
882 &rack_per_of_gp_probertt_reduce, 10,
883 "What percentage of goodput do we reduce every gp_srtt");
884 SYSCTL_ADD_U16(&rack_sysctl_ctx,
885 SYSCTL_CHILDREN(rack_probertt),
886 OID_AUTO, "gp_per_low", CTLFLAG_RW,
887 &rack_per_of_gp_lowthresh, 40,
888 "What percentage of goodput do we allow the multiplier to fall to");
889 SYSCTL_ADD_U32(&rack_sysctl_ctx,
890 SYSCTL_CHILDREN(rack_probertt),
891 OID_AUTO, "time_between", CTLFLAG_RW,
892 & rack_time_between_probertt, 96000000,
893 "How many useconds between the lowest rtt falling must past before we enter probertt");
894 SYSCTL_ADD_U32(&rack_sysctl_ctx,
895 SYSCTL_CHILDREN(rack_probertt),
896 OID_AUTO, "safety", CTLFLAG_RW,
897 &rack_probe_rtt_safety_val, 2000000,
898 "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
899 SYSCTL_ADD_U32(&rack_sysctl_ctx,
900 SYSCTL_CHILDREN(rack_probertt),
901 OID_AUTO, "sets_cwnd", CTLFLAG_RW,
902 &rack_probe_rtt_sets_cwnd, 0,
903 "Do we set the cwnd too (if always_lower is on)");
904 SYSCTL_ADD_U32(&rack_sysctl_ctx,
905 SYSCTL_CHILDREN(rack_probertt),
906 OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
907 &rack_max_drain_wait, 2,
908 "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
909 SYSCTL_ADD_U32(&rack_sysctl_ctx,
910 SYSCTL_CHILDREN(rack_probertt),
911 OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
913 "We must drain this many gp_srtt's waiting for flight to reach goal");
914 SYSCTL_ADD_U32(&rack_sysctl_ctx,
915 SYSCTL_CHILDREN(rack_probertt),
916 OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
917 &rack_probertt_use_min_rtt_entry, 1,
918 "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
919 SYSCTL_ADD_U32(&rack_sysctl_ctx,
920 SYSCTL_CHILDREN(rack_probertt),
921 OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
922 &rack_probertt_use_min_rtt_exit, 0,
923 "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
924 SYSCTL_ADD_U32(&rack_sysctl_ctx,
925 SYSCTL_CHILDREN(rack_probertt),
926 OID_AUTO, "length_div", CTLFLAG_RW,
927 &rack_probertt_gpsrtt_cnt_div, 0,
928 "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
929 SYSCTL_ADD_U32(&rack_sysctl_ctx,
930 SYSCTL_CHILDREN(rack_probertt),
931 OID_AUTO, "length_mul", CTLFLAG_RW,
932 &rack_probertt_gpsrtt_cnt_mul, 0,
933 "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
934 SYSCTL_ADD_U32(&rack_sysctl_ctx,
935 SYSCTL_CHILDREN(rack_probertt),
936 OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
937 &rack_min_probertt_hold, 200000,
938 "What is the minimum time we hold probertt at target");
939 SYSCTL_ADD_U32(&rack_sysctl_ctx,
940 SYSCTL_CHILDREN(rack_probertt),
941 OID_AUTO, "filter_life", CTLFLAG_RW,
942 &rack_probertt_filter_life, 10000000,
943 "What is the time for the filters life in useconds");
944 SYSCTL_ADD_U32(&rack_sysctl_ctx,
945 SYSCTL_CHILDREN(rack_probertt),
946 OID_AUTO, "lower_within", CTLFLAG_RW,
947 &rack_probertt_lower_within, 10,
948 "If the rtt goes lower within this percentage of the time, go into probe-rtt");
949 SYSCTL_ADD_U32(&rack_sysctl_ctx,
950 SYSCTL_CHILDREN(rack_probertt),
951 OID_AUTO, "must_move", CTLFLAG_RW,
952 &rack_min_rtt_movement, 250,
953 "How much is the minimum movement in rtt to count as a drop for probertt purposes");
954 SYSCTL_ADD_U32(&rack_sysctl_ctx,
955 SYSCTL_CHILDREN(rack_probertt),
956 OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
957 &rack_probertt_clear_is, 1,
958 "Do we clear I/S counts on exiting probe-rtt");
959 SYSCTL_ADD_S32(&rack_sysctl_ctx,
960 SYSCTL_CHILDREN(rack_probertt),
961 OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
962 &rack_max_drain_hbp, 1,
963 "How many extra drain gpsrtt's do we get in highly buffered paths");
964 SYSCTL_ADD_S32(&rack_sysctl_ctx,
965 SYSCTL_CHILDREN(rack_probertt),
966 OID_AUTO, "hbp_threshold", CTLFLAG_RW,
968 "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
969 /* Pacing related sysctls */
970 rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
971 SYSCTL_CHILDREN(rack_sysctl_root),
974 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
975 "Pacing related Controls");
976 SYSCTL_ADD_S32(&rack_sysctl_ctx,
977 SYSCTL_CHILDREN(rack_pacing),
978 OID_AUTO, "max_pace_over", CTLFLAG_RW,
979 &rack_max_per_above, 30,
980 "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
981 SYSCTL_ADD_S32(&rack_sysctl_ctx,
982 SYSCTL_CHILDREN(rack_pacing),
983 OID_AUTO, "pace_to_one", CTLFLAG_RW,
984 &rack_pace_one_seg, 0,
985 "Do we allow low b/w pacing of 1MSS instead of two");
986 SYSCTL_ADD_S32(&rack_sysctl_ctx,
987 SYSCTL_CHILDREN(rack_pacing),
988 OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
989 &rack_limit_time_with_srtt, 0,
990 "Do we limit pacing time based on srtt");
991 SYSCTL_ADD_S32(&rack_sysctl_ctx,
992 SYSCTL_CHILDREN(rack_pacing),
993 OID_AUTO, "init_win", CTLFLAG_RW,
994 &rack_default_init_window, 0,
995 "Do we have a rack initial window 0 = system default");
996 SYSCTL_ADD_U16(&rack_sysctl_ctx,
997 SYSCTL_CHILDREN(rack_pacing),
998 OID_AUTO, "gp_per_ss", CTLFLAG_RW,
999 &rack_per_of_gp_ss, 250,
1000 "If non zero, what percentage of goodput to pace at in slow start");
1001 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1002 SYSCTL_CHILDREN(rack_pacing),
1003 OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1004 &rack_per_of_gp_ca, 150,
1005 "If non zero, what percentage of goodput to pace at in congestion avoidance");
1006 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1007 SYSCTL_CHILDREN(rack_pacing),
1008 OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1009 &rack_per_of_gp_rec, 200,
1010 "If non zero, what percentage of goodput to pace at in recovery");
1011 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1012 SYSCTL_CHILDREN(rack_pacing),
1013 OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1014 &rack_hptsi_segments, 40,
1015 "What size is the max for TSO segments in pacing and burst mitigation");
1016 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1017 SYSCTL_CHILDREN(rack_pacing),
1018 OID_AUTO, "burst_reduces", CTLFLAG_RW,
1019 &rack_slot_reduction, 4,
1020 "When doing only burst mitigation what is the reduce divisor");
1021 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1022 SYSCTL_CHILDREN(rack_sysctl_root),
1023 OID_AUTO, "use_pacing", CTLFLAG_RW,
1024 &rack_pace_every_seg, 0,
1025 "If set we use pacing, if clear we use only the original burst mitigation");
1026 SYSCTL_ADD_U64(&rack_sysctl_ctx,
1027 SYSCTL_CHILDREN(rack_pacing),
1028 OID_AUTO, "rate_cap", CTLFLAG_RW,
1029 &rack_bw_rate_cap, 0,
1030 "If set we apply this value to the absolute rate cap used by pacing");
1031 SYSCTL_ADD_U8(&rack_sysctl_ctx,
1032 SYSCTL_CHILDREN(rack_sysctl_root),
1033 OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1034 &rack_req_measurements, 1,
1035 "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1036 /* Hardware pacing */
1037 rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1038 SYSCTL_CHILDREN(rack_sysctl_root),
1041 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1042 "Pacing related Controls");
1043 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1044 SYSCTL_CHILDREN(rack_hw_pacing),
1045 OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1046 &rack_hw_rwnd_factor, 2,
1047 "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1048 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1049 SYSCTL_CHILDREN(rack_hw_pacing),
1050 OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1051 &rack_enobuf_hw_boost_mult, 2,
1052 "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1053 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1054 SYSCTL_CHILDREN(rack_hw_pacing),
1055 OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1056 &rack_enobuf_hw_max, 2,
1057 "What is the max boost the pacing time if we see a ENOBUFS?");
1058 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1059 SYSCTL_CHILDREN(rack_hw_pacing),
1060 OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1061 &rack_enobuf_hw_min, 2,
1062 "What is the min boost the pacing time if we see a ENOBUFS?");
1063 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1064 SYSCTL_CHILDREN(rack_hw_pacing),
1065 OID_AUTO, "enable", CTLFLAG_RW,
1066 &rack_enable_hw_pacing, 0,
1067 "Should RACK attempt to use hw pacing?");
1068 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1069 SYSCTL_CHILDREN(rack_hw_pacing),
1070 OID_AUTO, "rate_cap", CTLFLAG_RW,
1071 &rack_hw_rate_caps, 1,
1072 "Does the highest hardware pacing rate cap the rate we will send at??");
1073 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1074 SYSCTL_CHILDREN(rack_hw_pacing),
1075 OID_AUTO, "rate_min", CTLFLAG_RW,
1076 &rack_hw_rate_min, 0,
1077 "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1078 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1079 SYSCTL_CHILDREN(rack_hw_pacing),
1080 OID_AUTO, "rate_to_low", CTLFLAG_RW,
1081 &rack_hw_rate_to_low, 0,
1082 "If we fall below this rate, dis-engage hw pacing?");
1083 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1084 SYSCTL_CHILDREN(rack_hw_pacing),
1085 OID_AUTO, "up_only", CTLFLAG_RW,
1086 &rack_hw_up_only, 1,
1087 "Do we allow hw pacing to lower the rate selected?");
1088 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1089 SYSCTL_CHILDREN(rack_hw_pacing),
1090 OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1091 &rack_hw_pace_extra_slots, 2,
1092 "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1093 rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1094 SYSCTL_CHILDREN(rack_sysctl_root),
1097 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1098 "Rack Timely RTT Controls");
1099 /* Timely based GP dynmics */
1100 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101 SYSCTL_CHILDREN(rack_timely),
1102 OID_AUTO, "upper", CTLFLAG_RW,
1103 &rack_gp_per_bw_mul_up, 2,
1104 "Rack timely upper range for equal b/w (in percentage)");
1105 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106 SYSCTL_CHILDREN(rack_timely),
1107 OID_AUTO, "lower", CTLFLAG_RW,
1108 &rack_gp_per_bw_mul_down, 4,
1109 "Rack timely lower range for equal b/w (in percentage)");
1110 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111 SYSCTL_CHILDREN(rack_timely),
1112 OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1113 &rack_gp_rtt_maxmul, 3,
1114 "Rack timely multipler of lowest rtt for rtt_max");
1115 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116 SYSCTL_CHILDREN(rack_timely),
1117 OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1118 &rack_gp_rtt_mindiv, 4,
1119 "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1120 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121 SYSCTL_CHILDREN(rack_timely),
1122 OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1123 &rack_gp_rtt_minmul, 1,
1124 "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1125 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126 SYSCTL_CHILDREN(rack_timely),
1127 OID_AUTO, "decrease", CTLFLAG_RW,
1128 &rack_gp_decrease_per, 20,
1129 "Rack timely decrease percentage of our GP multiplication factor");
1130 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131 SYSCTL_CHILDREN(rack_timely),
1132 OID_AUTO, "increase", CTLFLAG_RW,
1133 &rack_gp_increase_per, 2,
1134 "Rack timely increase perentage of our GP multiplication factor");
1135 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136 SYSCTL_CHILDREN(rack_timely),
1137 OID_AUTO, "lowerbound", CTLFLAG_RW,
1138 &rack_per_lower_bound, 50,
1139 "Rack timely lowest percentage we allow GP multiplier to fall to");
1140 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141 SYSCTL_CHILDREN(rack_timely),
1142 OID_AUTO, "upperboundss", CTLFLAG_RW,
1143 &rack_per_upper_bound_ss, 0,
1144 "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1145 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146 SYSCTL_CHILDREN(rack_timely),
1147 OID_AUTO, "upperboundca", CTLFLAG_RW,
1148 &rack_per_upper_bound_ca, 0,
1149 "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1150 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151 SYSCTL_CHILDREN(rack_timely),
1152 OID_AUTO, "dynamicgp", CTLFLAG_RW,
1153 &rack_do_dyn_mul, 0,
1154 "Rack timely do we enable dynmaic timely goodput by default");
1155 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156 SYSCTL_CHILDREN(rack_timely),
1157 OID_AUTO, "no_rec_red", CTLFLAG_RW,
1158 &rack_gp_no_rec_chg, 1,
1159 "Rack timely do we prohibit the recovery multiplier from being lowered");
1160 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161 SYSCTL_CHILDREN(rack_timely),
1162 OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1163 &rack_timely_dec_clear, 6,
1164 "Rack timely what threshold do we count to before another boost during b/w decent");
1165 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166 SYSCTL_CHILDREN(rack_timely),
1167 OID_AUTO, "max_push_rise", CTLFLAG_RW,
1168 &rack_timely_max_push_rise, 3,
1169 "Rack timely how many times do we push up with b/w increase");
1170 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1171 SYSCTL_CHILDREN(rack_timely),
1172 OID_AUTO, "max_push_drop", CTLFLAG_RW,
1173 &rack_timely_max_push_drop, 3,
1174 "Rack timely how many times do we push back on b/w decent");
1175 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1176 SYSCTL_CHILDREN(rack_timely),
1177 OID_AUTO, "min_segs", CTLFLAG_RW,
1178 &rack_timely_min_segs, 4,
1179 "Rack timely when setting the cwnd what is the min num segments");
1180 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1181 SYSCTL_CHILDREN(rack_timely),
1182 OID_AUTO, "noback_max", CTLFLAG_RW,
1183 &rack_use_max_for_nobackoff, 0,
1184 "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1185 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1186 SYSCTL_CHILDREN(rack_timely),
1187 OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1188 &rack_timely_int_timely_only, 0,
1189 "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1190 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1191 SYSCTL_CHILDREN(rack_timely),
1192 OID_AUTO, "nonstop", CTLFLAG_RW,
1193 &rack_timely_no_stopping, 0,
1194 "Rack timely don't stop increase");
1195 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1196 SYSCTL_CHILDREN(rack_timely),
1197 OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1198 &rack_down_raise_thresh, 100,
1199 "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1200 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1201 SYSCTL_CHILDREN(rack_timely),
1202 OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1204 "Bottom dragging if not these many segments outstanding and room");
1206 /* TLP and Rack related parameters */
1207 rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1208 SYSCTL_CHILDREN(rack_sysctl_root),
1211 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1212 "TLP and Rack related Controls");
1213 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214 SYSCTL_CHILDREN(rack_tlp),
1215 OID_AUTO, "use_rrr", CTLFLAG_RW,
1217 "Do we use Rack Rapid Recovery");
1218 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219 SYSCTL_CHILDREN(rack_tlp),
1220 OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1221 &rack_max_abc_post_recovery, 2,
1222 "Since we do early recovery, do we override the l_abc to a value, if so what?");
1223 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224 SYSCTL_CHILDREN(rack_tlp),
1225 OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1226 &rack_non_rxt_use_cr, 0,
1227 "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1228 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 SYSCTL_CHILDREN(rack_tlp),
1230 OID_AUTO, "tlpmethod", CTLFLAG_RW,
1231 &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1232 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1233 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234 SYSCTL_CHILDREN(rack_tlp),
1235 OID_AUTO, "limit", CTLFLAG_RW,
1237 "How many TLP's can be sent without sending new data");
1238 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239 SYSCTL_CHILDREN(rack_tlp),
1240 OID_AUTO, "use_greater", CTLFLAG_RW,
1241 &rack_tlp_use_greater, 1,
1242 "Should we use the rack_rtt time if its greater than srtt");
1243 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244 SYSCTL_CHILDREN(rack_tlp),
1245 OID_AUTO, "tlpminto", CTLFLAG_RW,
1246 &rack_tlp_min, 10000,
1247 "TLP minimum timeout per the specification (in microseconds)");
1248 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249 SYSCTL_CHILDREN(rack_tlp),
1250 OID_AUTO, "send_oldest", CTLFLAG_RW,
1251 &rack_always_send_oldest, 0,
1252 "Should we always send the oldest TLP and RACK-TLP");
1253 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254 SYSCTL_CHILDREN(rack_tlp),
1255 OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1256 &rack_limited_retran, 0,
1257 "How many times can a rack timeout drive out sends");
1258 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259 SYSCTL_CHILDREN(rack_tlp),
1260 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1261 &rack_lower_cwnd_at_tlp, 0,
1262 "When a TLP completes a retran should we enter recovery");
1263 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264 SYSCTL_CHILDREN(rack_tlp),
1265 OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1266 &rack_reorder_thresh, 2,
1267 "What factor for rack will be added when seeing reordering (shift right)");
1268 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269 SYSCTL_CHILDREN(rack_tlp),
1270 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1271 &rack_tlp_thresh, 1,
1272 "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1273 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1274 SYSCTL_CHILDREN(rack_tlp),
1275 OID_AUTO, "reorder_fade", CTLFLAG_RW,
1276 &rack_reorder_fade, 60000000,
1277 "Does reorder detection fade, if so how many microseconds (0 means never)");
1278 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1279 SYSCTL_CHILDREN(rack_tlp),
1280 OID_AUTO, "pktdelay", CTLFLAG_RW,
1281 &rack_pkt_delay, 1000,
1282 "Extra RACK time (in microseconds) besides reordering thresh");
1284 /* Timer related controls */
1285 rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1286 SYSCTL_CHILDREN(rack_sysctl_root),
1289 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1290 "Timer related controls");
1291 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1292 SYSCTL_CHILDREN(rack_timers),
1293 OID_AUTO, "persmin", CTLFLAG_RW,
1294 &rack_persist_min, 250000,
1295 "What is the minimum time in microseconds between persists");
1296 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1297 SYSCTL_CHILDREN(rack_timers),
1298 OID_AUTO, "persmax", CTLFLAG_RW,
1299 &rack_persist_max, 2000000,
1300 "What is the largest delay in microseconds between persists");
1301 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1302 SYSCTL_CHILDREN(rack_timers),
1303 OID_AUTO, "delayed_ack", CTLFLAG_RW,
1304 &rack_delayed_ack_time, 40000,
1305 "Delayed ack time (40ms in microseconds)");
1306 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1307 SYSCTL_CHILDREN(rack_timers),
1308 OID_AUTO, "minrto", CTLFLAG_RW,
1309 &rack_rto_min, 30000,
1310 "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1311 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1312 SYSCTL_CHILDREN(rack_timers),
1313 OID_AUTO, "maxrto", CTLFLAG_RW,
1314 &rack_rto_max, 4000000,
1315 "Maxiumum RTO in microseconds -- should be at least as large as min_rto");
1316 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1317 SYSCTL_CHILDREN(rack_timers),
1318 OID_AUTO, "minto", CTLFLAG_RW,
1320 "Minimum rack timeout in microseconds");
1321 /* Measure controls */
1322 rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1323 SYSCTL_CHILDREN(rack_sysctl_root),
1326 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1327 "Measure related controls");
1328 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1329 SYSCTL_CHILDREN(rack_measure),
1330 OID_AUTO, "wma_divisor", CTLFLAG_RW,
1331 &rack_wma_divisor, 8,
1332 "When doing b/w calculation what is the divisor for the WMA");
1333 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1334 SYSCTL_CHILDREN(rack_measure),
1335 OID_AUTO, "end_cwnd", CTLFLAG_RW,
1336 &rack_cwnd_block_ends_measure, 0,
1337 "Does a cwnd just-return end the measurement window (app limited)");
1338 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1339 SYSCTL_CHILDREN(rack_measure),
1340 OID_AUTO, "end_rwnd", CTLFLAG_RW,
1341 &rack_rwnd_block_ends_measure, 0,
1342 "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1343 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1344 SYSCTL_CHILDREN(rack_measure),
1345 OID_AUTO, "min_target", CTLFLAG_RW,
1346 &rack_def_data_window, 20,
1347 "What is the minimum target window (in mss) for a GP measurements");
1348 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1349 SYSCTL_CHILDREN(rack_measure),
1350 OID_AUTO, "goal_bdp", CTLFLAG_RW,
1352 "What is the goal BDP to measure");
1353 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1354 SYSCTL_CHILDREN(rack_measure),
1355 OID_AUTO, "min_srtts", CTLFLAG_RW,
1357 "What is the goal BDP to measure");
1358 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1359 SYSCTL_CHILDREN(rack_measure),
1360 OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1361 &rack_min_measure_usec, 0,
1362 "What is the Minimum time time for a measurement if 0, this is off");
1363 /* Misc rack controls */
1364 rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1365 SYSCTL_CHILDREN(rack_sysctl_root),
1368 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1369 "Misc related controls");
1370 #ifdef TCP_ACCOUNTING
1371 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1372 SYSCTL_CHILDREN(rack_misc),
1373 OID_AUTO, "tcp_acct", CTLFLAG_RW,
1374 &rack_tcp_accounting, 0,
1375 "Should we turn on TCP accounting for all rack sessions?");
1377 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378 SYSCTL_CHILDREN(rack_misc),
1379 OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1380 &rack_prr_addbackmax, 2,
1381 "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1382 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383 SYSCTL_CHILDREN(rack_misc),
1384 OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1385 &rack_stats_gets_ms_rtt, 1,
1386 "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1387 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388 SYSCTL_CHILDREN(rack_misc),
1389 OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1390 &rack_client_low_buf, 0,
1391 "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1392 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393 SYSCTL_CHILDREN(rack_misc),
1394 OID_AUTO, "defprofile", CTLFLAG_RW,
1395 &rack_def_profile, 0,
1396 "Should RACK use a default profile (0=no, num == profile num)?");
1397 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398 SYSCTL_CHILDREN(rack_misc),
1399 OID_AUTO, "cmpack", CTLFLAG_RW,
1400 &rack_use_cmp_acks, 1,
1401 "Should RACK have LRO send compressed acks");
1402 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1403 SYSCTL_CHILDREN(rack_misc),
1404 OID_AUTO, "fsb", CTLFLAG_RW,
1406 "Should RACK use the fast send block?");
1407 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1408 SYSCTL_CHILDREN(rack_misc),
1409 OID_AUTO, "rfo", CTLFLAG_RW,
1411 "Should RACK use rack_fast_output()?");
1412 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1413 SYSCTL_CHILDREN(rack_misc),
1414 OID_AUTO, "rsmrfo", CTLFLAG_RW,
1415 &rack_use_rsm_rfo, 1,
1416 "Should RACK use rack_fast_rsm_output()?");
1417 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1418 SYSCTL_CHILDREN(rack_misc),
1419 OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1420 &rack_enable_shared_cwnd, 1,
1421 "Should RACK try to use the shared cwnd on connections where allowed");
1422 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1423 SYSCTL_CHILDREN(rack_misc),
1424 OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1425 &rack_limits_scwnd, 1,
1426 "Should RACK place low end time limits on the shared cwnd feature");
1427 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1428 SYSCTL_CHILDREN(rack_misc),
1429 OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1430 &rack_enable_mqueue_for_nonpaced, 0,
1431 "Should RACK use mbuf queuing for non-paced connections");
1432 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1433 SYSCTL_CHILDREN(rack_misc),
1434 OID_AUTO, "iMac_dack", CTLFLAG_RW,
1435 &rack_use_imac_dack, 0,
1436 "Should RACK try to emulate iMac delayed ack");
1437 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1438 SYSCTL_CHILDREN(rack_misc),
1439 OID_AUTO, "no_prr", CTLFLAG_RW,
1440 &rack_disable_prr, 0,
1441 "Should RACK not use prr and only pace (must have pacing on)");
1442 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443 SYSCTL_CHILDREN(rack_misc),
1444 OID_AUTO, "bb_verbose", CTLFLAG_RW,
1445 &rack_verbose_logging, 0,
1446 "Should RACK black box logging be verbose");
1447 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448 SYSCTL_CHILDREN(rack_misc),
1449 OID_AUTO, "data_after_close", CTLFLAG_RW,
1450 &rack_ignore_data_after_close, 1,
1451 "Do we hold off sending a RST until all pending data is ack'd");
1452 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453 SYSCTL_CHILDREN(rack_misc),
1454 OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1455 &rack_sack_not_required, 1,
1456 "Do we allow rack to run on connections not supporting SACK");
1457 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1458 SYSCTL_CHILDREN(rack_misc),
1459 OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1460 &rack_send_a_lot_in_prr, 1,
1461 "Send a lot in prr");
1462 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1463 SYSCTL_CHILDREN(rack_misc),
1464 OID_AUTO, "autoscale", CTLFLAG_RW,
1465 &rack_autosndbuf_inc, 20,
1466 "What percentage should rack scale up its snd buffer by?");
1467 /* Sack Attacker detection stuff */
1468 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1469 SYSCTL_CHILDREN(rack_attack),
1470 OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1471 &rack_highest_sack_thresh_seen, 0,
1472 "Highest sack to ack ratio seen");
1473 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1474 SYSCTL_CHILDREN(rack_attack),
1475 OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1476 &rack_highest_move_thresh_seen, 0,
1477 "Highest move to non-move ratio seen");
1478 rack_ack_total = counter_u64_alloc(M_WAITOK);
1479 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1480 SYSCTL_CHILDREN(rack_attack),
1481 OID_AUTO, "acktotal", CTLFLAG_RD,
1483 "Total number of Ack's");
1484 rack_express_sack = counter_u64_alloc(M_WAITOK);
1485 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1486 SYSCTL_CHILDREN(rack_attack),
1487 OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1489 "Total expresss number of Sack's");
1490 rack_sack_total = counter_u64_alloc(M_WAITOK);
1491 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1492 SYSCTL_CHILDREN(rack_attack),
1493 OID_AUTO, "sacktotal", CTLFLAG_RD,
1495 "Total number of SACKs");
1496 rack_move_none = counter_u64_alloc(M_WAITOK);
1497 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1498 SYSCTL_CHILDREN(rack_attack),
1499 OID_AUTO, "move_none", CTLFLAG_RD,
1501 "Total number of SACK index reuse of postions under threshold");
1502 rack_move_some = counter_u64_alloc(M_WAITOK);
1503 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1504 SYSCTL_CHILDREN(rack_attack),
1505 OID_AUTO, "move_some", CTLFLAG_RD,
1507 "Total number of SACK index reuse of postions over threshold");
1508 rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1509 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1510 SYSCTL_CHILDREN(rack_attack),
1511 OID_AUTO, "attacks", CTLFLAG_RD,
1512 &rack_sack_attacks_detected,
1513 "Total number of SACK attackers that had sack disabled");
1514 rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1515 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516 SYSCTL_CHILDREN(rack_attack),
1517 OID_AUTO, "reversed", CTLFLAG_RD,
1518 &rack_sack_attacks_reversed,
1519 "Total number of SACK attackers that were later determined false positive");
1520 rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1521 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522 SYSCTL_CHILDREN(rack_attack),
1523 OID_AUTO, "nextmerge", CTLFLAG_RD,
1524 &rack_sack_used_next_merge,
1525 "Total number of times we used the next merge");
1526 rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1527 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1528 SYSCTL_CHILDREN(rack_attack),
1529 OID_AUTO, "prevmerge", CTLFLAG_RD,
1530 &rack_sack_used_prev_merge,
1531 "Total number of times we used the prev merge");
1533 rack_fto_send = counter_u64_alloc(M_WAITOK);
1534 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1535 SYSCTL_CHILDREN(rack_counters),
1536 OID_AUTO, "fto_send", CTLFLAG_RD,
1537 &rack_fto_send, "Total number of rack_fast_output sends");
1538 rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1539 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1540 SYSCTL_CHILDREN(rack_counters),
1541 OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1542 &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1543 rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1544 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1545 SYSCTL_CHILDREN(rack_counters),
1546 OID_AUTO, "nfto_resend", CTLFLAG_RD,
1547 &rack_nfto_resend, "Total number of rack_output retransmissions");
1548 rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1549 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1550 SYSCTL_CHILDREN(rack_counters),
1551 OID_AUTO, "nfto_send", CTLFLAG_RD,
1552 &rack_non_fto_send, "Total number of rack_output first sends");
1553 rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1554 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1555 SYSCTL_CHILDREN(rack_counters),
1556 OID_AUTO, "rfo_extended", CTLFLAG_RD,
1557 &rack_extended_rfo, "Total number of times we extended rfo");
1559 rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1560 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1561 SYSCTL_CHILDREN(rack_counters),
1562 OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1563 &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1564 rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1566 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1567 SYSCTL_CHILDREN(rack_counters),
1568 OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1569 &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1573 rack_badfr = counter_u64_alloc(M_WAITOK);
1574 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1575 SYSCTL_CHILDREN(rack_counters),
1576 OID_AUTO, "badfr", CTLFLAG_RD,
1577 &rack_badfr, "Total number of bad FRs");
1578 rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1579 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1580 SYSCTL_CHILDREN(rack_counters),
1581 OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1582 &rack_badfr_bytes, "Total number of bad FRs");
1583 rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1584 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585 SYSCTL_CHILDREN(rack_counters),
1586 OID_AUTO, "prrsndret", CTLFLAG_RD,
1587 &rack_rtm_prr_retran,
1588 "Total number of prr based retransmits");
1589 rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1590 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1591 SYSCTL_CHILDREN(rack_counters),
1592 OID_AUTO, "prrsndnew", CTLFLAG_RD,
1593 &rack_rtm_prr_newdata,
1594 "Total number of prr based new transmits");
1595 rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1596 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1597 SYSCTL_CHILDREN(rack_counters),
1598 OID_AUTO, "tsnf", CTLFLAG_RD,
1599 &rack_timestamp_mismatch,
1600 "Total number of timestamps that we could not find the reported ts");
1601 rack_find_high = counter_u64_alloc(M_WAITOK);
1602 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1603 SYSCTL_CHILDREN(rack_counters),
1604 OID_AUTO, "findhigh", CTLFLAG_RD,
1606 "Total number of FIN causing find-high");
1607 rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1608 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1609 SYSCTL_CHILDREN(rack_counters),
1610 OID_AUTO, "reordering", CTLFLAG_RD,
1612 "Total number of times we added delay due to reordering");
1613 rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1614 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1615 SYSCTL_CHILDREN(rack_counters),
1616 OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1618 "Total number of tail loss probe expirations");
1619 rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1620 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1621 SYSCTL_CHILDREN(rack_counters),
1622 OID_AUTO, "tlp_new", CTLFLAG_RD,
1624 "Total number of tail loss probe sending new data");
1625 rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1626 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1627 SYSCTL_CHILDREN(rack_counters),
1628 OID_AUTO, "tlp_retran", CTLFLAG_RD,
1630 "Total number of tail loss probe sending retransmitted data");
1631 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1632 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1633 SYSCTL_CHILDREN(rack_counters),
1634 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1635 &rack_tlp_retran_bytes,
1636 "Total bytes of tail loss probe sending retransmitted data");
1637 rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1638 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1639 SYSCTL_CHILDREN(rack_counters),
1640 OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1641 &rack_tlp_retran_fail,
1642 "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1643 rack_to_tot = counter_u64_alloc(M_WAITOK);
1644 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1645 SYSCTL_CHILDREN(rack_counters),
1646 OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1648 "Total number of times the rack to expired");
1649 rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1650 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1651 SYSCTL_CHILDREN(rack_counters),
1652 OID_AUTO, "arm_rack", CTLFLAG_RD,
1654 "Total number of times the rack timer armed");
1655 rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1656 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1657 SYSCTL_CHILDREN(rack_counters),
1658 OID_AUTO, "arm_tlp", CTLFLAG_RD,
1660 "Total number of times the tlp timer armed");
1661 rack_calc_zero = counter_u64_alloc(M_WAITOK);
1662 rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1663 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1664 SYSCTL_CHILDREN(rack_counters),
1665 OID_AUTO, "calc_zero", CTLFLAG_RD,
1667 "Total number of times pacing time worked out to zero");
1668 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1669 SYSCTL_CHILDREN(rack_counters),
1670 OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1672 "Total number of times pacing time worked out to non-zero");
1673 rack_paced_segments = counter_u64_alloc(M_WAITOK);
1674 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1675 SYSCTL_CHILDREN(rack_counters),
1676 OID_AUTO, "paced", CTLFLAG_RD,
1677 &rack_paced_segments,
1678 "Total number of times a segment send caused hptsi");
1679 rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1680 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681 SYSCTL_CHILDREN(rack_counters),
1682 OID_AUTO, "unpaced", CTLFLAG_RD,
1683 &rack_unpaced_segments,
1684 "Total number of times a segment did not cause hptsi");
1685 rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1686 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687 SYSCTL_CHILDREN(rack_counters),
1688 OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1690 "Total number of times a sends returned enobuf for non-hdwr paced connections");
1691 rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1692 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693 SYSCTL_CHILDREN(rack_counters),
1694 OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1695 &rack_saw_enobuf_hw,
1696 "Total number of times a send returned enobuf for hdwr paced connections");
1697 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1698 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1699 SYSCTL_CHILDREN(rack_counters),
1700 OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1701 &rack_saw_enetunreach,
1702 "Total number of times a send received a enetunreachable");
1703 rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1704 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1705 SYSCTL_CHILDREN(rack_counters),
1706 OID_AUTO, "alloc_hot", CTLFLAG_RD,
1708 "Total allocations from the top of our list");
1709 rack_to_alloc = counter_u64_alloc(M_WAITOK);
1710 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1711 SYSCTL_CHILDREN(rack_counters),
1712 OID_AUTO, "allocs", CTLFLAG_RD,
1714 "Total allocations of tracking structures");
1715 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1716 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1717 SYSCTL_CHILDREN(rack_counters),
1718 OID_AUTO, "allochard", CTLFLAG_RD,
1719 &rack_to_alloc_hard,
1720 "Total allocations done with sleeping the hard way");
1721 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1722 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1723 SYSCTL_CHILDREN(rack_counters),
1724 OID_AUTO, "allocemerg", CTLFLAG_RD,
1725 &rack_to_alloc_emerg,
1726 "Total allocations done from emergency cache");
1727 rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1728 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1729 SYSCTL_CHILDREN(rack_counters),
1730 OID_AUTO, "alloc_limited", CTLFLAG_RD,
1731 &rack_to_alloc_limited,
1732 "Total allocations dropped due to limit");
1733 rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1734 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1735 SYSCTL_CHILDREN(rack_counters),
1736 OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1737 &rack_alloc_limited_conns,
1738 "Connections with allocations dropped due to limit");
1739 rack_split_limited = counter_u64_alloc(M_WAITOK);
1740 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1741 SYSCTL_CHILDREN(rack_counters),
1742 OID_AUTO, "split_limited", CTLFLAG_RD,
1743 &rack_split_limited,
1744 "Split allocations dropped due to limit");
1746 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1748 sprintf(name, "cmp_ack_cnt_%d", i);
1749 rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1750 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1751 SYSCTL_CHILDREN(rack_counters),
1752 OID_AUTO, name, CTLFLAG_RD,
1753 &rack_proc_comp_ack[i],
1754 "Number of compressed acks we processed");
1756 rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1757 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1758 SYSCTL_CHILDREN(rack_counters),
1759 OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1761 "Number of TCP connections with large mbuf's for compressed acks");
1762 rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1763 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1764 SYSCTL_CHILDREN(rack_counters),
1765 OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1767 "Number of TCP connections with small mbuf's for compressed acks");
1769 rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1770 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1771 SYSCTL_CHILDREN(rack_counters),
1772 OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1773 &rack_adjust_map_bw,
1774 "Number of times we hit the case where the sb went up and down on a sendmap entry");
1776 rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1777 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1778 SYSCTL_CHILDREN(rack_counters),
1779 OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1780 &rack_multi_single_eq,
1781 "Number of compressed acks total represented");
1782 rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1783 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1784 SYSCTL_CHILDREN(rack_counters),
1785 OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1786 &rack_proc_non_comp_ack,
1787 "Number of non compresseds acks that we processed");
1790 rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1791 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1792 SYSCTL_CHILDREN(rack_counters),
1793 OID_AUTO, "sack_long", CTLFLAG_RD,
1794 &rack_sack_proc_all,
1795 "Total times we had to walk whole list for sack processing");
1796 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1797 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1798 SYSCTL_CHILDREN(rack_counters),
1799 OID_AUTO, "sack_restart", CTLFLAG_RD,
1800 &rack_sack_proc_restart,
1801 "Total times we had to walk whole list due to a restart");
1802 rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1803 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804 SYSCTL_CHILDREN(rack_counters),
1805 OID_AUTO, "sack_short", CTLFLAG_RD,
1806 &rack_sack_proc_short,
1807 "Total times we took shortcut for sack processing");
1808 rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1809 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1810 SYSCTL_CHILDREN(rack_counters),
1811 OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1812 &rack_enter_tlp_calc,
1813 "Total times we called calc-tlp");
1814 rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1815 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1816 SYSCTL_CHILDREN(rack_counters),
1817 OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1818 &rack_used_tlpmethod,
1819 "Total number of runt sacks");
1820 rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1821 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1822 SYSCTL_CHILDREN(rack_counters),
1823 OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1824 &rack_used_tlpmethod2,
1825 "Total number of times we hit TLP method 2");
1826 rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1827 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1828 SYSCTL_CHILDREN(rack_attack),
1829 OID_AUTO, "skipacked", CTLFLAG_RD,
1830 &rack_sack_skipped_acked,
1831 "Total number of times we skipped previously sacked");
1832 rack_sack_splits = counter_u64_alloc(M_WAITOK);
1833 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1834 SYSCTL_CHILDREN(rack_attack),
1835 OID_AUTO, "ofsplit", CTLFLAG_RD,
1837 "Total number of times we did the old fashion tree split");
1838 rack_progress_drops = counter_u64_alloc(M_WAITOK);
1839 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1840 SYSCTL_CHILDREN(rack_counters),
1841 OID_AUTO, "prog_drops", CTLFLAG_RD,
1842 &rack_progress_drops,
1843 "Total number of progress drops");
1844 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1845 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1846 SYSCTL_CHILDREN(rack_counters),
1847 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1848 &rack_input_idle_reduces,
1849 "Total number of idle reductions on input");
1850 rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1851 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1852 SYSCTL_CHILDREN(rack_counters),
1853 OID_AUTO, "collapsed_win", CTLFLAG_RD,
1854 &rack_collapsed_win,
1855 "Total number of collapsed windows");
1856 rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1857 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1858 SYSCTL_CHILDREN(rack_counters),
1859 OID_AUTO, "tlp_nada", CTLFLAG_RD,
1860 &rack_tlp_does_nada,
1861 "Total number of nada tlp calls");
1862 rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1863 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1864 SYSCTL_CHILDREN(rack_counters),
1865 OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1867 "Total number of scwnd attempts");
1869 rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1870 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1871 SYSCTL_CHILDREN(rack_counters),
1872 OID_AUTO, "timer_hole", CTLFLAG_RD,
1873 &rack_per_timer_hole,
1874 "Total persists start in timer hole");
1876 rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1877 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1878 SYSCTL_CHILDREN(rack_counters),
1879 OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1880 &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1881 rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1882 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1883 SYSCTL_CHILDREN(rack_counters),
1884 OID_AUTO, "sndptr_right", CTLFLAG_RD,
1885 &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1887 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1888 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1889 OID_AUTO, "outsize", CTLFLAG_RD,
1890 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1891 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1892 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1893 OID_AUTO, "opts", CTLFLAG_RD,
1894 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1895 SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1896 SYSCTL_CHILDREN(rack_sysctl_root),
1897 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1898 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1902 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1904 if (SEQ_GEQ(b->r_start, a->r_start) &&
1905 SEQ_LT(b->r_start, a->r_end)) {
1907 * The entry b is within the
1909 * a -- |-------------|
1914 * b -- |-----------|
1917 } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1919 * b falls as either the next
1920 * sequence block after a so a
1921 * is said to be smaller than b.
1931 * Whats left is where a is
1932 * larger than b. i.e:
1936 * b -- |--------------|
1941 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1942 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1945 rc_init_window(struct tcp_rack *rack)
1949 if (rack->rc_init_win == 0) {
1951 * Nothing set by the user, use the system stack
1954 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1956 win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1961 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1963 if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1964 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1965 else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1966 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1968 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1972 rack_get_bw(struct tcp_rack *rack)
1974 if (rack->use_fixed_rate) {
1975 /* Return the fixed pacing rate */
1976 return (rack_get_fixed_pacing_bw(rack));
1978 if (rack->r_ctl.gp_bw == 0) {
1980 * We have yet no b/w measurement,
1981 * if we have a user set initial bw
1982 * return it. If we don't have that and
1983 * we have an srtt, use the tcp IW (10) to
1984 * calculate a fictional b/w over the SRTT
1985 * which is more or less a guess. Note
1986 * we don't use our IW from rack on purpose
1987 * so if we have like IW=30, we are not
1988 * calculating a "huge" b/w.
1991 if (rack->r_ctl.init_rate)
1992 return (rack->r_ctl.init_rate);
1994 /* Has the user set a max peak rate? */
1995 #ifdef NETFLIX_PEAKRATE
1996 if (rack->rc_tp->t_maxpeakrate)
1997 return (rack->rc_tp->t_maxpeakrate);
1999 /* Ok lets come up with the IW guess, if we have a srtt */
2000 if (rack->rc_tp->t_srtt == 0) {
2002 * Go with old pacing method
2003 * i.e. burst mitigation only.
2007 /* Ok lets get the initial TCP win (not racks) */
2008 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2009 srtt = (uint64_t)rack->rc_tp->t_srtt;
2010 bw *= (uint64_t)USECS_IN_SECOND;
2012 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2013 bw = rack->r_ctl.bw_rate_cap;
2018 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2019 /* Averaging is done, we can return the value */
2020 bw = rack->r_ctl.gp_bw;
2022 /* Still doing initial average must calculate */
2023 bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2025 #ifdef NETFLIX_PEAKRATE
2026 if ((rack->rc_tp->t_maxpeakrate) &&
2027 (bw > rack->rc_tp->t_maxpeakrate)) {
2028 /* The user has set a peak rate to pace at
2029 * don't allow us to pace faster than that.
2031 return (rack->rc_tp->t_maxpeakrate);
2034 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2035 bw = rack->r_ctl.bw_rate_cap;
2041 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2043 if (rack->use_fixed_rate) {
2045 } else if (rack->in_probe_rtt && (rsm == NULL))
2046 return (rack->r_ctl.rack_per_of_gp_probertt);
2047 else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2048 rack->r_ctl.rack_per_of_gp_rec)) {
2050 /* a retransmission always use the recovery rate */
2051 return (rack->r_ctl.rack_per_of_gp_rec);
2052 } else if (rack->rack_rec_nonrxt_use_cr) {
2053 /* Directed to use the configured rate */
2054 goto configured_rate;
2055 } else if (rack->rack_no_prr &&
2056 (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2057 /* No PRR, lets just use the b/w estimate only */
2061 * Here we may have a non-retransmit but we
2062 * have no overrides, so just use the recovery
2063 * rate (prr is in effect).
2065 return (rack->r_ctl.rack_per_of_gp_rec);
2069 /* For the configured rate we look at our cwnd vs the ssthresh */
2070 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2071 return (rack->r_ctl.rack_per_of_gp_ss);
2073 return (rack->r_ctl.rack_per_of_gp_ca);
2077 rack_log_hdwr_pacing(struct tcp_rack *rack,
2078 uint64_t rate, uint64_t hw_rate, int line,
2079 int error, uint16_t mod)
2081 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2082 union tcp_log_stackspecific log;
2084 const struct ifnet *ifp;
2086 memset(&log, 0, sizeof(log));
2087 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2088 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2089 if (rack->r_ctl.crte) {
2090 ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2091 } else if (rack->rc_inp->inp_route.ro_nh &&
2092 rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2093 ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2097 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff);
2098 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2100 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2101 log.u_bbr.bw_inuse = rate;
2102 log.u_bbr.flex5 = line;
2103 log.u_bbr.flex6 = error;
2104 log.u_bbr.flex7 = mod;
2105 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2106 log.u_bbr.flex8 = rack->use_fixed_rate;
2107 log.u_bbr.flex8 <<= 1;
2108 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2109 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2110 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2111 if (rack->r_ctl.crte)
2112 log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2114 log.u_bbr.cur_del_rate = 0;
2115 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2116 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2117 &rack->rc_inp->inp_socket->so_rcv,
2118 &rack->rc_inp->inp_socket->so_snd,
2119 BBR_LOG_HDWR_PACE, 0,
2120 0, &log, false, &tv);
2125 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2128 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2130 uint64_t bw_est, high_rate;
2133 gain = (uint64_t)rack_get_output_gain(rack, rsm);
2135 bw_est /= (uint64_t)100;
2136 /* Never fall below the minimum (def 64kbps) */
2137 if (bw_est < RACK_MIN_BW)
2138 bw_est = RACK_MIN_BW;
2139 if (rack->r_rack_hw_rate_caps) {
2140 /* Rate caps are in place */
2141 if (rack->r_ctl.crte != NULL) {
2142 /* We have a hdwr rate already */
2143 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2144 if (bw_est >= high_rate) {
2145 /* We are capping bw at the highest rate table entry */
2146 rack_log_hdwr_pacing(rack,
2147 bw_est, high_rate, __LINE__,
2153 } else if ((rack->rack_hdrw_pacing == 0) &&
2154 (rack->rack_hdw_pace_ena) &&
2155 (rack->rack_attempt_hdwr_pace == 0) &&
2156 (rack->rc_inp->inp_route.ro_nh != NULL) &&
2157 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2159 * Special case, we have not yet attempted hardware
2160 * pacing, and yet we may, when we do, find out if we are
2161 * above the highest rate. We need to know the maxbw for the interface
2162 * in question (if it supports ratelimiting). We get back
2163 * a 0, if the interface is not found in the RL lists.
2165 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2167 /* Yep, we have a rate is it above this rate? */
2168 if (bw_est > high_rate) {
2180 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2182 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2183 union tcp_log_stackspecific log;
2186 if ((mod != 1) && (rack_verbose_logging == 0)) {
2188 * We get 3 values currently for mod
2189 * 1 - We are retransmitting and this tells the reason.
2190 * 2 - We are clearing a dup-ack count.
2191 * 3 - We are incrementing a dup-ack count.
2193 * The clear/increment are only logged
2194 * if you have BBverbose on.
2198 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2199 log.u_bbr.flex1 = tsused;
2200 log.u_bbr.flex2 = thresh;
2201 log.u_bbr.flex3 = rsm->r_flags;
2202 log.u_bbr.flex4 = rsm->r_dupack;
2203 log.u_bbr.flex5 = rsm->r_start;
2204 log.u_bbr.flex6 = rsm->r_end;
2205 log.u_bbr.flex8 = mod;
2206 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2207 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2208 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2209 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2210 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2211 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2212 log.u_bbr.pacing_gain = rack->r_must_retran;
2213 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2214 &rack->rc_inp->inp_socket->so_rcv,
2215 &rack->rc_inp->inp_socket->so_snd,
2216 BBR_LOG_SETTINGS_CHG, 0,
2217 0, &log, false, &tv);
2222 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2224 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2225 union tcp_log_stackspecific log;
2228 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2229 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2230 log.u_bbr.flex2 = to;
2231 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2232 log.u_bbr.flex4 = slot;
2233 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2234 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2235 log.u_bbr.flex7 = rack->rc_in_persist;
2236 log.u_bbr.flex8 = which;
2237 if (rack->rack_no_prr)
2238 log.u_bbr.pkts_out = 0;
2240 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2241 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2242 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2243 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2244 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2245 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2246 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2247 log.u_bbr.pacing_gain = rack->r_must_retran;
2248 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2249 log.u_bbr.lost = rack_rto_min;
2250 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2251 &rack->rc_inp->inp_socket->so_rcv,
2252 &rack->rc_inp->inp_socket->so_snd,
2253 BBR_LOG_TIMERSTAR, 0,
2254 0, &log, false, &tv);
2259 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2261 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2262 union tcp_log_stackspecific log;
2265 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2266 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2267 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2268 log.u_bbr.flex8 = to_num;
2269 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2270 log.u_bbr.flex2 = rack->rc_rack_rtt;
2272 log.u_bbr.flex3 = 0;
2274 log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2275 if (rack->rack_no_prr)
2276 log.u_bbr.flex5 = 0;
2278 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2279 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2280 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2281 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2282 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2283 log.u_bbr.pacing_gain = rack->r_must_retran;
2284 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2285 &rack->rc_inp->inp_socket->so_rcv,
2286 &rack->rc_inp->inp_socket->so_snd,
2288 0, &log, false, &tv);
2293 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2294 struct rack_sendmap *prev,
2295 struct rack_sendmap *rsm,
2296 struct rack_sendmap *next,
2297 int flag, uint32_t th_ack, int line)
2299 if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2300 union tcp_log_stackspecific log;
2303 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2304 log.u_bbr.flex8 = flag;
2305 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2306 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2307 log.u_bbr.cur_del_rate = (uint64_t)prev;
2308 log.u_bbr.delRate = (uint64_t)rsm;
2309 log.u_bbr.rttProp = (uint64_t)next;
2310 log.u_bbr.flex7 = 0;
2312 log.u_bbr.flex1 = prev->r_start;
2313 log.u_bbr.flex2 = prev->r_end;
2314 log.u_bbr.flex7 |= 0x4;
2317 log.u_bbr.flex3 = rsm->r_start;
2318 log.u_bbr.flex4 = rsm->r_end;
2319 log.u_bbr.flex7 |= 0x2;
2322 log.u_bbr.flex5 = next->r_start;
2323 log.u_bbr.flex6 = next->r_end;
2324 log.u_bbr.flex7 |= 0x1;
2326 log.u_bbr.applimited = line;
2327 log.u_bbr.pkts_out = th_ack;
2328 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2329 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2330 if (rack->rack_no_prr)
2333 log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2334 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2335 &rack->rc_inp->inp_socket->so_rcv,
2336 &rack->rc_inp->inp_socket->so_snd,
2338 0, &log, false, &tv);
2343 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2344 struct rack_sendmap *rsm, int conf)
2346 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2347 union tcp_log_stackspecific log;
2349 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2350 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2351 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2352 log.u_bbr.flex1 = t;
2353 log.u_bbr.flex2 = len;
2354 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2355 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2356 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2357 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2358 log.u_bbr.flex7 = conf;
2359 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2360 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2361 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2362 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2363 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2364 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2366 log.u_bbr.pkt_epoch = rsm->r_start;
2367 log.u_bbr.lost = rsm->r_end;
2368 log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2369 log.u_bbr.pacing_gain = rsm->r_flags;
2372 log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2374 log.u_bbr.cwnd_gain = 0;
2375 log.u_bbr.pacing_gain = 0;
2377 /* Write out general bits of interest rrs here */
2378 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2379 log.u_bbr.use_lt_bw <<= 1;
2380 log.u_bbr.use_lt_bw |= rack->forced_ack;
2381 log.u_bbr.use_lt_bw <<= 1;
2382 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2383 log.u_bbr.use_lt_bw <<= 1;
2384 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2385 log.u_bbr.use_lt_bw <<= 1;
2386 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2387 log.u_bbr.use_lt_bw <<= 1;
2388 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2389 log.u_bbr.use_lt_bw <<= 1;
2390 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2391 log.u_bbr.use_lt_bw <<= 1;
2392 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2393 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2394 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2395 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2396 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2397 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2398 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2399 log.u_bbr.bw_inuse <<= 32;
2401 log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2402 TCP_LOG_EVENTP(tp, NULL,
2403 &rack->rc_inp->inp_socket->so_rcv,
2404 &rack->rc_inp->inp_socket->so_snd,
2406 0, &log, false, &tv);
2413 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2416 * Log the rtt sample we are
2417 * applying to the srtt algorithm in
2420 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2421 union tcp_log_stackspecific log;
2424 /* Convert our ms to a microsecond */
2425 memset(&log, 0, sizeof(log));
2426 log.u_bbr.flex1 = rtt;
2427 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2428 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2429 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2430 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2431 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2432 log.u_bbr.flex7 = 1;
2433 log.u_bbr.flex8 = rack->sack_attack_disable;
2434 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2435 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2436 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2437 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2438 log.u_bbr.pacing_gain = rack->r_must_retran;
2440 * We capture in delRate the upper 32 bits as
2441 * the confidence level we had declared, and the
2442 * lower 32 bits as the actual RTT using the arrival
2445 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2446 log.u_bbr.delRate <<= 32;
2447 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2448 /* Lets capture all the things that make up t_rtxcur */
2449 log.u_bbr.applimited = rack_rto_min;
2450 log.u_bbr.epoch = rack_rto_max;
2451 log.u_bbr.lt_epoch = rtt;
2452 log.u_bbr.lost = rack_rto_min;
2453 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2454 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2455 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2456 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2457 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2458 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2459 &rack->rc_inp->inp_socket->so_rcv,
2460 &rack->rc_inp->inp_socket->so_snd,
2462 0, &log, false, &tv);
2467 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2469 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2470 union tcp_log_stackspecific log;
2473 /* Convert our ms to a microsecond */
2474 memset(&log, 0, sizeof(log));
2475 log.u_bbr.flex1 = rtt;
2476 log.u_bbr.flex2 = send_time;
2477 log.u_bbr.flex3 = ack_time;
2478 log.u_bbr.flex4 = where;
2479 log.u_bbr.flex7 = 2;
2480 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2481 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2482 &rack->rc_inp->inp_socket->so_rcv,
2483 &rack->rc_inp->inp_socket->so_snd,
2485 0, &log, false, &tv);
2492 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line)
2494 if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2495 union tcp_log_stackspecific log;
2498 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2499 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2500 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2501 log.u_bbr.flex1 = line;
2502 log.u_bbr.flex2 = tick;
2503 log.u_bbr.flex3 = tp->t_maxunacktime;
2504 log.u_bbr.flex4 = tp->t_acktime;
2505 log.u_bbr.flex8 = event;
2506 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2507 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2508 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2509 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2510 log.u_bbr.pacing_gain = rack->r_must_retran;
2511 TCP_LOG_EVENTP(tp, NULL,
2512 &rack->rc_inp->inp_socket->so_rcv,
2513 &rack->rc_inp->inp_socket->so_snd,
2514 BBR_LOG_PROGRESS, 0,
2515 0, &log, false, &tv);
2520 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2522 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2523 union tcp_log_stackspecific log;
2525 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2526 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2527 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2528 log.u_bbr.flex1 = slot;
2529 if (rack->rack_no_prr)
2530 log.u_bbr.flex2 = 0;
2532 log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2533 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2534 log.u_bbr.flex8 = rack->rc_in_persist;
2535 log.u_bbr.timeStamp = cts;
2536 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2537 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2538 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2539 log.u_bbr.pacing_gain = rack->r_must_retran;
2540 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2541 &rack->rc_inp->inp_socket->so_rcv,
2542 &rack->rc_inp->inp_socket->so_snd,
2544 0, &log, false, tv);
2549 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2551 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2552 union tcp_log_stackspecific log;
2555 memset(&log, 0, sizeof(log));
2556 log.u_bbr.flex1 = did_out;
2557 log.u_bbr.flex2 = nxt_pkt;
2558 log.u_bbr.flex3 = way_out;
2559 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2560 if (rack->rack_no_prr)
2561 log.u_bbr.flex5 = 0;
2563 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2564 log.u_bbr.flex6 = nsegs;
2565 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2566 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data; /* Do we have ack-can-send set */
2567 log.u_bbr.flex7 <<= 1;
2568 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2569 log.u_bbr.flex7 <<= 1;
2570 log.u_bbr.flex7 |= rack->r_wanted_output; /* Do we want output */
2571 log.u_bbr.flex8 = rack->rc_in_persist;
2572 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2573 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2574 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2575 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2576 log.u_bbr.use_lt_bw <<= 1;
2577 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2578 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2579 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2580 log.u_bbr.pacing_gain = rack->r_must_retran;
2581 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2582 &rack->rc_inp->inp_socket->so_rcv,
2583 &rack->rc_inp->inp_socket->so_snd,
2584 BBR_LOG_DOSEG_DONE, 0,
2585 0, &log, false, &tv);
2590 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2592 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2593 union tcp_log_stackspecific log;
2597 memset(&log, 0, sizeof(log));
2598 cts = tcp_get_usecs(&tv);
2599 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2600 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2601 log.u_bbr.flex4 = arg1;
2602 log.u_bbr.flex5 = arg2;
2603 log.u_bbr.flex6 = arg3;
2604 log.u_bbr.flex8 = frm;
2605 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2606 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2607 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2608 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2609 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2610 log.u_bbr.pacing_gain = rack->r_must_retran;
2611 TCP_LOG_EVENTP(tp, NULL,
2612 &tp->t_inpcb->inp_socket->so_rcv,
2613 &tp->t_inpcb->inp_socket->so_snd,
2614 TCP_HDWR_PACE_SIZE, 0,
2615 0, &log, false, &tv);
2620 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2621 uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2623 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2624 union tcp_log_stackspecific log;
2627 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2628 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2629 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2630 log.u_bbr.flex1 = slot;
2631 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2632 log.u_bbr.flex4 = reason;
2633 if (rack->rack_no_prr)
2634 log.u_bbr.flex5 = 0;
2636 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2637 log.u_bbr.flex7 = hpts_calling;
2638 log.u_bbr.flex8 = rack->rc_in_persist;
2639 log.u_bbr.lt_epoch = cwnd_to_use;
2640 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2641 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2642 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2643 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2644 log.u_bbr.pacing_gain = rack->r_must_retran;
2645 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2646 &rack->rc_inp->inp_socket->so_rcv,
2647 &rack->rc_inp->inp_socket->so_snd,
2649 tlen, &log, false, &tv);
2654 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2655 struct timeval *tv, uint32_t flags_on_entry)
2657 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2658 union tcp_log_stackspecific log;
2660 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2661 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2662 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2663 log.u_bbr.flex1 = line;
2664 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2665 log.u_bbr.flex3 = flags_on_entry;
2666 log.u_bbr.flex4 = us_cts;
2667 if (rack->rack_no_prr)
2668 log.u_bbr.flex5 = 0;
2670 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2671 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2672 log.u_bbr.flex7 = hpts_removed;
2673 log.u_bbr.flex8 = 1;
2674 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2675 log.u_bbr.timeStamp = us_cts;
2676 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2677 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2678 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2679 log.u_bbr.pacing_gain = rack->r_must_retran;
2680 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2681 &rack->rc_inp->inp_socket->so_rcv,
2682 &rack->rc_inp->inp_socket->so_snd,
2683 BBR_LOG_TIMERCANC, 0,
2684 0, &log, false, tv);
2689 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2690 uint32_t flex1, uint32_t flex2,
2691 uint32_t flex3, uint32_t flex4,
2692 uint32_t flex5, uint32_t flex6,
2693 uint16_t flex7, uint8_t mod)
2695 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2696 union tcp_log_stackspecific log;
2700 /* No you can't use 1, its for the real to cancel */
2703 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2704 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2705 log.u_bbr.flex1 = flex1;
2706 log.u_bbr.flex2 = flex2;
2707 log.u_bbr.flex3 = flex3;
2708 log.u_bbr.flex4 = flex4;
2709 log.u_bbr.flex5 = flex5;
2710 log.u_bbr.flex6 = flex6;
2711 log.u_bbr.flex7 = flex7;
2712 log.u_bbr.flex8 = mod;
2713 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2714 &rack->rc_inp->inp_socket->so_rcv,
2715 &rack->rc_inp->inp_socket->so_snd,
2716 BBR_LOG_TIMERCANC, 0,
2717 0, &log, false, &tv);
2722 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2724 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2725 union tcp_log_stackspecific log;
2728 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2729 log.u_bbr.flex1 = timers;
2730 log.u_bbr.flex2 = ret;
2731 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2732 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2733 log.u_bbr.flex5 = cts;
2734 if (rack->rack_no_prr)
2735 log.u_bbr.flex6 = 0;
2737 log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2738 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2739 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2740 log.u_bbr.pacing_gain = rack->r_must_retran;
2741 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2742 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2743 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2744 &rack->rc_inp->inp_socket->so_rcv,
2745 &rack->rc_inp->inp_socket->so_snd,
2746 BBR_LOG_TO_PROCESS, 0,
2747 0, &log, false, &tv);
2752 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2754 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2755 union tcp_log_stackspecific log;
2758 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2759 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2760 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2761 if (rack->rack_no_prr)
2762 log.u_bbr.flex3 = 0;
2764 log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2765 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2766 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2767 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2768 log.u_bbr.flex8 = frm;
2769 log.u_bbr.pkts_out = orig_cwnd;
2770 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2771 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2772 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2773 log.u_bbr.use_lt_bw <<= 1;
2774 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2775 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2776 &rack->rc_inp->inp_socket->so_rcv,
2777 &rack->rc_inp->inp_socket->so_snd,
2779 0, &log, false, &tv);
2783 #ifdef NETFLIX_EXP_DETECTION
2785 rack_log_sad(struct tcp_rack *rack, int event)
2787 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2788 union tcp_log_stackspecific log;
2791 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2792 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2793 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2794 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2795 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2796 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2797 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2798 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2799 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2800 log.u_bbr.lt_epoch |= rack->do_detection;
2801 log.u_bbr.applimited = tcp_map_minimum;
2802 log.u_bbr.flex7 = rack->sack_attack_disable;
2803 log.u_bbr.flex8 = event;
2804 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2805 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2806 log.u_bbr.delivered = tcp_sad_decay_val;
2807 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2808 &rack->rc_inp->inp_socket->so_rcv,
2809 &rack->rc_inp->inp_socket->so_snd,
2810 TCP_SAD_DETECTION, 0,
2811 0, &log, false, &tv);
2817 rack_counter_destroy(void)
2821 counter_u64_free(rack_fto_send);
2822 counter_u64_free(rack_fto_rsm_send);
2823 counter_u64_free(rack_nfto_resend);
2824 counter_u64_free(rack_hw_pace_init_fail);
2825 counter_u64_free(rack_hw_pace_lost);
2826 counter_u64_free(rack_non_fto_send);
2827 counter_u64_free(rack_extended_rfo);
2828 counter_u64_free(rack_ack_total);
2829 counter_u64_free(rack_express_sack);
2830 counter_u64_free(rack_sack_total);
2831 counter_u64_free(rack_move_none);
2832 counter_u64_free(rack_move_some);
2833 counter_u64_free(rack_sack_attacks_detected);
2834 counter_u64_free(rack_sack_attacks_reversed);
2835 counter_u64_free(rack_sack_used_next_merge);
2836 counter_u64_free(rack_sack_used_prev_merge);
2837 counter_u64_free(rack_badfr);
2838 counter_u64_free(rack_badfr_bytes);
2839 counter_u64_free(rack_rtm_prr_retran);
2840 counter_u64_free(rack_rtm_prr_newdata);
2841 counter_u64_free(rack_timestamp_mismatch);
2842 counter_u64_free(rack_find_high);
2843 counter_u64_free(rack_reorder_seen);
2844 counter_u64_free(rack_tlp_tot);
2845 counter_u64_free(rack_tlp_newdata);
2846 counter_u64_free(rack_tlp_retran);
2847 counter_u64_free(rack_tlp_retran_bytes);
2848 counter_u64_free(rack_tlp_retran_fail);
2849 counter_u64_free(rack_to_tot);
2850 counter_u64_free(rack_to_arm_rack);
2851 counter_u64_free(rack_to_arm_tlp);
2852 counter_u64_free(rack_calc_zero);
2853 counter_u64_free(rack_calc_nonzero);
2854 counter_u64_free(rack_paced_segments);
2855 counter_u64_free(rack_unpaced_segments);
2856 counter_u64_free(rack_saw_enobuf);
2857 counter_u64_free(rack_saw_enobuf_hw);
2858 counter_u64_free(rack_saw_enetunreach);
2859 counter_u64_free(rack_hot_alloc);
2860 counter_u64_free(rack_to_alloc);
2861 counter_u64_free(rack_to_alloc_hard);
2862 counter_u64_free(rack_to_alloc_emerg);
2863 counter_u64_free(rack_to_alloc_limited);
2864 counter_u64_free(rack_alloc_limited_conns);
2865 counter_u64_free(rack_split_limited);
2866 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2867 counter_u64_free(rack_proc_comp_ack[i]);
2869 counter_u64_free(rack_multi_single_eq);
2870 counter_u64_free(rack_proc_non_comp_ack);
2871 counter_u64_free(rack_sack_proc_all);
2872 counter_u64_free(rack_sack_proc_restart);
2873 counter_u64_free(rack_sack_proc_short);
2874 counter_u64_free(rack_enter_tlp_calc);
2875 counter_u64_free(rack_used_tlpmethod);
2876 counter_u64_free(rack_used_tlpmethod2);
2877 counter_u64_free(rack_sack_skipped_acked);
2878 counter_u64_free(rack_sack_splits);
2879 counter_u64_free(rack_progress_drops);
2880 counter_u64_free(rack_input_idle_reduces);
2881 counter_u64_free(rack_collapsed_win);
2882 counter_u64_free(rack_tlp_does_nada);
2883 counter_u64_free(rack_try_scwnd);
2884 counter_u64_free(rack_per_timer_hole);
2885 counter_u64_free(rack_large_ackcmp);
2886 counter_u64_free(rack_small_ackcmp);
2888 counter_u64_free(rack_adjust_map_bw);
2890 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2891 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2894 static struct rack_sendmap *
2895 rack_alloc(struct tcp_rack *rack)
2897 struct rack_sendmap *rsm;
2900 * First get the top of the list it in
2901 * theory is the "hottest" rsm we have,
2902 * possibly just freed by ack processing.
2904 if (rack->rc_free_cnt > rack_free_cache) {
2905 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2906 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2907 counter_u64_add(rack_hot_alloc, 1);
2908 rack->rc_free_cnt--;
2912 * Once we get under our free cache we probably
2913 * no longer have a "hot" one available. Lets
2916 rsm = uma_zalloc(rack_zone, M_NOWAIT);
2918 rack->r_ctl.rc_num_maps_alloced++;
2919 counter_u64_add(rack_to_alloc, 1);
2923 * Dig in to our aux rsm's (the last two) since
2924 * UMA failed to get us one.
2926 if (rack->rc_free_cnt) {
2927 counter_u64_add(rack_to_alloc_emerg, 1);
2928 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2929 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2930 rack->rc_free_cnt--;
2936 static struct rack_sendmap *
2937 rack_alloc_full_limit(struct tcp_rack *rack)
2939 if ((V_tcp_map_entries_limit > 0) &&
2940 (rack->do_detection == 0) &&
2941 (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2942 counter_u64_add(rack_to_alloc_limited, 1);
2943 if (!rack->alloc_limit_reported) {
2944 rack->alloc_limit_reported = 1;
2945 counter_u64_add(rack_alloc_limited_conns, 1);
2949 return (rack_alloc(rack));
2952 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2953 static struct rack_sendmap *
2954 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2956 struct rack_sendmap *rsm;
2959 /* currently there is only one limit type */
2960 if (V_tcp_map_split_limit > 0 &&
2961 (rack->do_detection == 0) &&
2962 rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2963 counter_u64_add(rack_split_limited, 1);
2964 if (!rack->alloc_limit_reported) {
2965 rack->alloc_limit_reported = 1;
2966 counter_u64_add(rack_alloc_limited_conns, 1);
2972 /* allocate and mark in the limit type, if set */
2973 rsm = rack_alloc(rack);
2974 if (rsm != NULL && limit_type) {
2975 rsm->r_limit_type = limit_type;
2976 rack->r_ctl.rc_num_split_allocs++;
2982 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2984 if (rsm->r_flags & RACK_APP_LIMITED) {
2985 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2986 rack->r_ctl.rc_app_limited_cnt--;
2989 if (rsm->r_limit_type) {
2990 /* currently there is only one limit type */
2991 rack->r_ctl.rc_num_split_allocs--;
2993 if (rsm == rack->r_ctl.rc_first_appl) {
2994 if (rack->r_ctl.rc_app_limited_cnt == 0)
2995 rack->r_ctl.rc_first_appl = NULL;
2997 /* Follow the next one out */
2998 struct rack_sendmap fe;
3000 fe.r_start = rsm->r_nseq_appl;
3001 rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3004 if (rsm == rack->r_ctl.rc_resend)
3005 rack->r_ctl.rc_resend = NULL;
3006 if (rsm == rack->r_ctl.rc_rsm_at_retran)
3007 rack->r_ctl.rc_rsm_at_retran = NULL;
3008 if (rsm == rack->r_ctl.rc_end_appl)
3009 rack->r_ctl.rc_end_appl = NULL;
3010 if (rack->r_ctl.rc_tlpsend == rsm)
3011 rack->r_ctl.rc_tlpsend = NULL;
3012 if (rack->r_ctl.rc_sacklast == rsm)
3013 rack->r_ctl.rc_sacklast = NULL;
3014 memset(rsm, 0, sizeof(struct rack_sendmap));
3015 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3016 rack->rc_free_cnt++;
3020 rack_free_trim(struct tcp_rack *rack)
3022 struct rack_sendmap *rsm;
3025 * Free up all the tail entries until
3026 * we get our list down to the limit.
3028 while (rack->rc_free_cnt > rack_free_cache) {
3029 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3030 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3031 rack->rc_free_cnt--;
3032 uma_zfree(rack_zone, rsm);
3038 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3040 uint64_t srtt, bw, len, tim;
3041 uint32_t segsiz, def_len, minl;
3043 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3044 def_len = rack_def_data_window * segsiz;
3045 if (rack->rc_gp_filled == 0) {
3047 * We have no measurement (IW is in flight?) so
3048 * we can only guess using our data_window sysctl
3049 * value (usually 100MSS).
3054 * Now we have a number of factors to consider.
3056 * 1) We have a desired BDP which is usually
3058 * 2) We have a minimum number of rtt's usually 1 SRTT
3059 * but we allow it too to be more.
3060 * 3) We want to make sure a measurement last N useconds (if
3061 * we have set rack_min_measure_usec.
3063 * We handle the first concern here by trying to create a data
3064 * window of max(rack_def_data_window, DesiredBDP). The
3065 * second concern we handle in not letting the measurement
3066 * window end normally until at least the required SRTT's
3067 * have gone by which is done further below in
3068 * rack_enough_for_measurement(). Finally the third concern
3069 * we also handle here by calculating how long that time
3070 * would take at the current BW and then return the
3071 * max of our first calculation and that length. Note
3072 * that if rack_min_measure_usec is 0, we don't deal
3073 * with concern 3. Also for both Concern 1 and 3 an
3074 * application limited period could end the measurement
3077 * So lets calculate the BDP with the "known" b/w using
3078 * the SRTT has our rtt and then multiply it by the
3081 bw = rack_get_bw(rack);
3082 srtt = (uint64_t)tp->t_srtt;
3084 len /= (uint64_t)HPTS_USEC_IN_SEC;
3085 len *= max(1, rack_goal_bdp);
3086 /* Now we need to round up to the nearest MSS */
3087 len = roundup(len, segsiz);
3088 if (rack_min_measure_usec) {
3089 /* Now calculate our min length for this b/w */
3090 tim = rack_min_measure_usec;
3091 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3094 minl = roundup(minl, segsiz);
3099 * Now if we have a very small window we want
3100 * to attempt to get the window that is
3101 * as small as possible. This happens on
3102 * low b/w connections and we don't want to
3103 * span huge numbers of rtt's between measurements.
3105 * We basically include 2 over our "MIN window" so
3106 * that the measurement can be shortened (possibly) by
3110 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3112 return (max((uint32_t)len, def_len));
3117 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
3119 uint32_t tim, srtts, segsiz;
3122 * Has enough time passed for the GP measurement to be valid?
3124 if ((tp->snd_max == tp->snd_una) ||
3125 (th_ack == tp->snd_max)){
3129 if (SEQ_LT(th_ack, tp->gput_seq)) {
3130 /* Not enough bytes yet */
3133 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3134 if (SEQ_LT(th_ack, tp->gput_ack) &&
3135 ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3136 /* Not enough bytes yet */
3139 if (rack->r_ctl.rc_first_appl &&
3140 (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
3142 * We are up to the app limited point
3143 * we have to measure irrespective of the time..
3147 /* Now what about time? */
3148 srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3149 tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3153 /* Nope not even a full SRTT has passed */
3158 rack_log_timely(struct tcp_rack *rack,
3159 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3160 uint64_t up_bnd, int line, uint8_t method)
3162 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3163 union tcp_log_stackspecific log;
3166 memset(&log, 0, sizeof(log));
3167 log.u_bbr.flex1 = logged;
3168 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3169 log.u_bbr.flex2 <<= 4;
3170 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3171 log.u_bbr.flex2 <<= 4;
3172 log.u_bbr.flex2 |= rack->rc_gp_incr;
3173 log.u_bbr.flex2 <<= 4;
3174 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3175 log.u_bbr.flex3 = rack->rc_gp_incr;
3176 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3177 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3178 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3179 log.u_bbr.flex7 = rack->rc_gp_bwred;
3180 log.u_bbr.flex8 = method;
3181 log.u_bbr.cur_del_rate = cur_bw;
3182 log.u_bbr.delRate = low_bnd;
3183 log.u_bbr.bw_inuse = up_bnd;
3184 log.u_bbr.rttProp = rack_get_bw(rack);
3185 log.u_bbr.pkt_epoch = line;
3186 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3187 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3188 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3189 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3190 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3191 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3192 log.u_bbr.cwnd_gain <<= 1;
3193 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3194 log.u_bbr.cwnd_gain <<= 1;
3195 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3196 log.u_bbr.cwnd_gain <<= 1;
3197 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3198 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3199 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3200 &rack->rc_inp->inp_socket->so_rcv,
3201 &rack->rc_inp->inp_socket->so_snd,
3203 0, &log, false, &tv);
3208 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3211 * Before we increase we need to know if
3212 * the estimate just made was less than
3213 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3215 * If we already are pacing at a fast enough
3216 * rate to push us faster there is no sense of
3219 * We first caculate our actual pacing rate (ss or ca multipler
3220 * times our cur_bw).
3222 * Then we take the last measured rate and multipy by our
3223 * maximum pacing overage to give us a max allowable rate.
3225 * If our act_rate is smaller than our max_allowable rate
3226 * then we should increase. Else we should hold steady.
3229 uint64_t act_rate, max_allow_rate;
3231 if (rack_timely_no_stopping)
3234 if ((cur_bw == 0) || (last_bw_est == 0)) {
3236 * Initial startup case or
3237 * everything is acked case.
3239 rack_log_timely(rack, mult, cur_bw, 0, 0,
3245 * We can always pace at or slightly above our rate.
3247 rack_log_timely(rack, mult, cur_bw, 0, 0,
3251 act_rate = cur_bw * (uint64_t)mult;
3253 max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3254 max_allow_rate /= 100;
3255 if (act_rate < max_allow_rate) {
3257 * Here the rate we are actually pacing at
3258 * is smaller than 10% above our last measurement.
3259 * This means we are pacing below what we would
3260 * like to try to achieve (plus some wiggle room).
3262 rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate,
3267 * Here we are already pacing at least rack_max_per_above(10%)
3268 * what we are getting back. This indicates most likely
3269 * that we are being limited (cwnd/rwnd/app) and can't
3270 * get any more b/w. There is no sense of trying to
3271 * raise up the pacing rate its not speeding us up
3272 * and we already are pacing faster than we are getting.
3274 rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate,
3281 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3284 * When we drag bottom, we want to assure
3285 * that no multiplier is below 1.0, if so
3286 * we want to restore it to at least that.
3288 if (rack->r_ctl.rack_per_of_gp_rec < 100) {
3289 /* This is unlikely we usually do not touch recovery */
3290 rack->r_ctl.rack_per_of_gp_rec = 100;
3292 if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3293 rack->r_ctl.rack_per_of_gp_ca = 100;
3295 if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3296 rack->r_ctl.rack_per_of_gp_ss = 100;
3301 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3303 if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3304 rack->r_ctl.rack_per_of_gp_ca = 100;
3306 if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3307 rack->r_ctl.rack_per_of_gp_ss = 100;
3312 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3314 int32_t calc, logged, plus;
3320 * override is passed when we are
3321 * loosing b/w and making one last
3322 * gasp at trying to not loose out
3323 * to a new-reno flow.
3327 /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3328 if (rack->rc_gp_incr &&
3329 ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3331 * Reset and get 5 strokes more before the boost. Note
3332 * that the count is 0 based so we have to add one.
3335 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3336 rack->rc_gp_timely_inc_cnt = 0;
3338 plus = (uint32_t)rack_gp_increase_per;
3339 /* Must be at least 1% increase for true timely increases */
3341 ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3343 if (rack->rc_gp_saw_rec &&
3344 (rack->rc_gp_no_rec_chg == 0) &&
3345 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3346 rack->r_ctl.rack_per_of_gp_rec)) {
3347 /* We have been in recovery ding it too */
3348 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3352 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3353 if (rack_per_upper_bound_ss &&
3354 (rack->rc_dragged_bottom == 0) &&
3355 (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3356 rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3358 if (rack->rc_gp_saw_ca &&
3359 (rack->rc_gp_saw_ss == 0) &&
3360 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3361 rack->r_ctl.rack_per_of_gp_ca)) {
3363 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3367 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3368 if (rack_per_upper_bound_ca &&
3369 (rack->rc_dragged_bottom == 0) &&
3370 (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3371 rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3373 if (rack->rc_gp_saw_ss &&
3374 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3375 rack->r_ctl.rack_per_of_gp_ss)) {
3377 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3380 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3381 if (rack_per_upper_bound_ss &&
3382 (rack->rc_dragged_bottom == 0) &&
3383 (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3384 rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3388 (rack->rc_gp_incr == 0)){
3389 /* Go into increment mode */
3390 rack->rc_gp_incr = 1;
3391 rack->rc_gp_timely_inc_cnt = 0;
3393 if (rack->rc_gp_incr &&
3395 (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3396 rack->rc_gp_timely_inc_cnt++;
3398 rack_log_timely(rack, logged, plus, 0, 0,
3403 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3406 * norm_grad = rtt_diff / minrtt;
3407 * new_per = curper * (1 - B * norm_grad)
3409 * B = rack_gp_decrease_per (default 10%)
3410 * rtt_dif = input var current rtt-diff
3411 * curper = input var current percentage
3412 * minrtt = from rack filter
3417 perf = (((uint64_t)curper * ((uint64_t)1000000 -
3418 ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3419 (((uint64_t)rtt_diff * (uint64_t)1000000)/
3420 (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3421 (uint64_t)1000000)) /
3423 if (perf > curper) {
3427 return ((uint32_t)perf);
3431 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3435 * result = curper * (1 - (B * ( 1 - ------ ))
3438 * B = rack_gp_decrease_per (default 10%)
3439 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3442 uint32_t highrttthresh;
3444 highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3446 perf = (((uint64_t)curper * ((uint64_t)1000000 -
3447 ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3448 ((uint64_t)highrttthresh * (uint64_t)1000000) /
3449 (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3454 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3456 uint64_t logvar, logvar2, logvar3;
3457 uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3459 if (rack->rc_gp_incr) {
3460 /* Turn off increment counting */
3461 rack->rc_gp_incr = 0;
3462 rack->rc_gp_timely_inc_cnt = 0;
3464 ss_red = ca_red = rec_red = 0;
3466 /* Calculate the reduction value */
3470 /* Must be at least 1% reduction */
3471 if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3472 /* We have been in recovery ding it too */
3473 if (timely_says == 2) {
3474 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3475 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3481 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3482 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3483 rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3484 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3486 rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3489 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3490 rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3493 if (rack->rc_gp_saw_ss) {
3495 if (timely_says == 2) {
3496 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3497 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3503 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3504 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3505 ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3506 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3509 rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3513 logvar2 = (uint32_t)rtt;
3515 logvar2 |= (uint32_t)rtt_diff;
3516 logvar3 = rack_gp_rtt_maxmul;
3518 logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3519 rack_log_timely(rack, timely_says,
3521 logvar, __LINE__, 10);
3523 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3524 rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3526 } else if (rack->rc_gp_saw_ca) {
3528 if (timely_says == 2) {
3529 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3530 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3536 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3537 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3538 ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3539 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3541 rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3546 logvar2 = (uint32_t)rtt;
3548 logvar2 |= (uint32_t)rtt_diff;
3549 logvar3 = rack_gp_rtt_maxmul;
3551 logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3552 rack_log_timely(rack, timely_says,
3554 logvar, __LINE__, 10);
3556 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3557 rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3560 if (rack->rc_gp_timely_dec_cnt < 0x7) {
3561 rack->rc_gp_timely_dec_cnt++;
3562 if (rack_timely_dec_clear &&
3563 (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3564 rack->rc_gp_timely_dec_cnt = 0;
3569 rack_log_timely(rack, logged, rec_red, rack_per_lower_bound, logvar,
3574 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3575 uint32_t rtt, uint32_t line, uint8_t reas)
3577 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3578 union tcp_log_stackspecific log;
3581 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3582 log.u_bbr.flex1 = line;
3583 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3584 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3585 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3586 log.u_bbr.flex5 = rtt;
3587 log.u_bbr.flex6 = rack->rc_highly_buffered;
3588 log.u_bbr.flex6 <<= 1;
3589 log.u_bbr.flex6 |= rack->forced_ack;
3590 log.u_bbr.flex6 <<= 1;
3591 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3592 log.u_bbr.flex6 <<= 1;
3593 log.u_bbr.flex6 |= rack->in_probe_rtt;
3594 log.u_bbr.flex6 <<= 1;
3595 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3596 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3597 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3598 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3599 log.u_bbr.flex8 = reas;
3600 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3601 log.u_bbr.delRate = rack_get_bw(rack);
3602 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3603 log.u_bbr.cur_del_rate <<= 32;
3604 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3605 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3606 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3607 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3608 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3609 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3610 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3611 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3612 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3613 log.u_bbr.rttProp = us_cts;
3614 log.u_bbr.rttProp <<= 32;
3615 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3616 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3617 &rack->rc_inp->inp_socket->so_rcv,
3618 &rack->rc_inp->inp_socket->so_snd,
3619 BBR_LOG_RTT_SHRINKS, 0,
3620 0, &log, false, &rack->r_ctl.act_rcv_time);
3625 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3629 bwdp = rack_get_bw(rack);
3630 bwdp *= (uint64_t)rtt;
3631 bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3632 rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3633 if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3635 * A window protocol must be able to have 4 packets
3636 * outstanding as the floor in order to function
3637 * (especially considering delayed ack :D).
3639 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3644 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3647 * ProbeRTT is a bit different in rack_pacing than in
3648 * BBR. It is like BBR in that it uses the lowering of
3649 * the RTT as a signal that we saw something new and
3650 * counts from there for how long between. But it is
3651 * different in that its quite simple. It does not
3652 * play with the cwnd and wait until we get down
3653 * to N segments outstanding and hold that for
3654 * 200ms. Instead it just sets the pacing reduction
3655 * rate to a set percentage (70 by default) and hold
3656 * that for a number of recent GP Srtt's.
3660 if (rack->rc_gp_dyn_mul == 0)
3663 if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3667 if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3668 SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3670 * Stop the goodput now, the idea here is
3671 * that future measurements with in_probe_rtt
3672 * won't register if they are not greater so
3673 * we want to get what info (if any) is available
3676 rack_do_goodput_measurement(rack->rc_tp, rack,
3677 rack->rc_tp->snd_una, __LINE__);
3679 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3680 rack->r_ctl.rc_time_probertt_entered = us_cts;
3681 segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3682 rack->r_ctl.rc_pace_min_segs);
3683 rack->in_probe_rtt = 1;
3684 rack->measure_saw_probe_rtt = 1;
3685 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3686 rack->r_ctl.rc_time_probertt_starts = 0;
3687 rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3688 if (rack_probertt_use_min_rtt_entry)
3689 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3691 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3692 rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3693 __LINE__, RACK_RTTS_ENTERPROBE);
3697 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3699 struct rack_sendmap *rsm;
3702 segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3703 rack->r_ctl.rc_pace_min_segs);
3704 rack->in_probe_rtt = 0;
3705 if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3706 SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3708 * Stop the goodput now, the idea here is
3709 * that future measurements with in_probe_rtt
3710 * won't register if they are not greater so
3711 * we want to get what info (if any) is available
3714 rack_do_goodput_measurement(rack->rc_tp, rack,
3715 rack->rc_tp->snd_una, __LINE__);
3716 } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3718 * We don't have enough data to make a measurement.
3719 * So lets just stop and start here after exiting
3720 * probe-rtt. We probably are not interested in
3721 * the results anyway.
3723 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3726 * Measurements through the current snd_max are going
3727 * to be limited by the slower pacing rate.
3729 * We need to mark these as app-limited so we
3730 * don't collapse the b/w.
3732 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3733 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3734 if (rack->r_ctl.rc_app_limited_cnt == 0)
3735 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3738 * Go out to the end app limited and mark
3739 * this new one as next and move the end_appl up
3742 if (rack->r_ctl.rc_end_appl)
3743 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3744 rack->r_ctl.rc_end_appl = rsm;
3746 rsm->r_flags |= RACK_APP_LIMITED;
3747 rack->r_ctl.rc_app_limited_cnt++;
3750 * Now, we need to examine our pacing rate multipliers.
3751 * If its under 100%, we need to kick it back up to
3752 * 100%. We also don't let it be over our "max" above
3753 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3754 * Note setting clamp_atexit_prtt to 0 has the effect
3755 * of setting CA/SS to 100% always at exit (which is
3756 * the default behavior).
3758 if (rack_probertt_clear_is) {
3759 rack->rc_gp_incr = 0;
3760 rack->rc_gp_bwred = 0;
3761 rack->rc_gp_timely_inc_cnt = 0;
3762 rack->rc_gp_timely_dec_cnt = 0;
3764 /* Do we do any clamping at exit? */
3765 if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3766 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3767 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3769 if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3770 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3771 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3774 * Lets set rtt_diff to 0, so that we will get a "boost"
3777 rack->r_ctl.rc_rtt_diff = 0;
3779 /* Clear all flags so we start fresh */
3780 rack->rc_tp->t_bytes_acked = 0;
3781 rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3783 * If configured to, set the cwnd and ssthresh to
3786 if (rack_probe_rtt_sets_cwnd) {
3790 /* Set ssthresh so we get into CA once we hit our target */
3791 if (rack_probertt_use_min_rtt_exit == 1) {
3792 /* Set to min rtt */
3793 rack_set_prtt_target(rack, segsiz,
3794 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3795 } else if (rack_probertt_use_min_rtt_exit == 2) {
3796 /* Set to current gp rtt */
3797 rack_set_prtt_target(rack, segsiz,
3798 rack->r_ctl.rc_gp_srtt);
3799 } else if (rack_probertt_use_min_rtt_exit == 3) {
3800 /* Set to entry gp rtt */
3801 rack_set_prtt_target(rack, segsiz,
3802 rack->r_ctl.rc_entry_gp_rtt);
3807 sum = rack->r_ctl.rc_entry_gp_rtt;
3809 sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3812 * A highly buffered path needs
3813 * cwnd space for timely to work.
3814 * Lets set things up as if
3815 * we are heading back here again.
3817 setval = rack->r_ctl.rc_entry_gp_rtt;
3818 } else if (sum >= 15) {
3820 * Lets take the smaller of the
3821 * two since we are just somewhat
3824 setval = rack->r_ctl.rc_gp_srtt;
3825 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3826 setval = rack->r_ctl.rc_entry_gp_rtt;
3829 * Here we are not highly buffered
3830 * and should pick the min we can to
3831 * keep from causing loss.
3833 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3835 rack_set_prtt_target(rack, segsiz,
3838 if (rack_probe_rtt_sets_cwnd > 1) {
3839 /* There is a percentage here to boost */
3840 ebdp = rack->r_ctl.rc_target_probertt_flight;
3841 ebdp *= rack_probe_rtt_sets_cwnd;
3843 setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3845 setto = rack->r_ctl.rc_target_probertt_flight;
3846 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3847 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3849 rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3851 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3852 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3854 rack_log_rtt_shrinks(rack, us_cts,
3855 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3856 __LINE__, RACK_RTTS_EXITPROBE);
3857 /* Clear times last so log has all the info */
3858 rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3859 rack->r_ctl.rc_time_probertt_entered = us_cts;
3860 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3861 rack->r_ctl.rc_time_of_last_probertt = us_cts;
3865 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3867 /* Check in on probe-rtt */
3868 if (rack->rc_gp_filled == 0) {
3869 /* We do not do p-rtt unless we have gp measurements */
3872 if (rack->in_probe_rtt) {
3873 uint64_t no_overflow;
3874 uint32_t endtime, must_stay;
3876 if (rack->r_ctl.rc_went_idle_time &&
3877 ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3879 * We went idle during prtt, just exit now.
3881 rack_exit_probertt(rack, us_cts);
3882 } else if (rack_probe_rtt_safety_val &&
3883 TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3884 ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3886 * Probe RTT safety value triggered!
3888 rack_log_rtt_shrinks(rack, us_cts,
3889 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3890 __LINE__, RACK_RTTS_SAFETY);
3891 rack_exit_probertt(rack, us_cts);
3893 /* Calculate the max we will wait */
3894 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3895 if (rack->rc_highly_buffered)
3896 endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3897 /* Calculate the min we must wait */
3898 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3899 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3900 TSTMP_LT(us_cts, endtime)) {
3902 /* Do we lower more? */
3904 if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3905 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3908 calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3911 calc *= rack_per_of_gp_probertt_reduce;
3912 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3914 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3915 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3917 /* We must reach target or the time set */
3920 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3921 if ((TSTMP_LT(us_cts, must_stay) &&
3922 rack->rc_highly_buffered) ||
3923 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3924 rack->r_ctl.rc_target_probertt_flight)) {
3925 /* We are not past the must_stay time */
3928 rack_log_rtt_shrinks(rack, us_cts,
3929 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3930 __LINE__, RACK_RTTS_REACHTARGET);
3931 rack->r_ctl.rc_time_probertt_starts = us_cts;
3932 if (rack->r_ctl.rc_time_probertt_starts == 0)
3933 rack->r_ctl.rc_time_probertt_starts = 1;
3934 /* Restore back to our rate we want to pace at in prtt */
3935 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3938 * Setup our end time, some number of gp_srtts plus 200ms.
3940 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3941 (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3942 if (rack_probertt_gpsrtt_cnt_div)
3943 endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3946 endtime += rack_min_probertt_hold;
3947 endtime += rack->r_ctl.rc_time_probertt_starts;
3948 if (TSTMP_GEQ(us_cts, endtime)) {
3949 /* yes, exit probertt */
3950 rack_exit_probertt(rack, us_cts);
3953 } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3954 /* Go into probertt, its been too long since we went lower */
3955 rack_enter_probertt(rack, us_cts);
3960 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3961 uint32_t rtt, int32_t rtt_diff)
3963 uint64_t cur_bw, up_bnd, low_bnd, subfr;
3966 if ((rack->rc_gp_dyn_mul == 0) ||
3967 (rack->use_fixed_rate) ||
3968 (rack->in_probe_rtt) ||
3969 (rack->rc_always_pace == 0)) {
3970 /* No dynamic GP multipler in play */
3973 losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3974 cur_bw = rack_get_bw(rack);
3975 /* Calculate our up and down range */
3976 up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3978 up_bnd += rack->r_ctl.last_gp_comp_bw;
3980 subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3982 low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3983 if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3985 * This is the case where our RTT is above
3986 * the max target and we have been configured
3987 * to just do timely no bonus up stuff in that case.
3989 * There are two configurations, set to 1, and we
3990 * just do timely if we are over our max. If its
3991 * set above 1 then we slam the multipliers down
3992 * to 100 and then decrement per timely.
3994 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
3996 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3997 rack_validate_multipliers_at_or_below_100(rack);
3998 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3999 } else if ((last_bw_est < low_bnd) && !losses) {
4001 * We are decreasing this is a bit complicated this
4002 * means we are loosing ground. This could be
4003 * because another flow entered and we are competing
4004 * for b/w with it. This will push the RTT up which
4005 * makes timely unusable unless we want to get shoved
4006 * into a corner and just be backed off (the age
4007 * old problem with delay based CC).
4009 * On the other hand if it was a route change we
4010 * would like to stay somewhat contained and not
4011 * blow out the buffers.
4013 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4015 rack->r_ctl.last_gp_comp_bw = cur_bw;
4016 if (rack->rc_gp_bwred == 0) {
4017 /* Go into reduction counting */
4018 rack->rc_gp_bwred = 1;
4019 rack->rc_gp_timely_dec_cnt = 0;
4021 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4022 (timely_says == 0)) {
4024 * Push another time with a faster pacing
4025 * to try to gain back (we include override to
4026 * get a full raise factor).
4028 if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4029 (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4030 (timely_says == 0) ||
4031 (rack_down_raise_thresh == 0)) {
4033 * Do an override up in b/w if we were
4034 * below the threshold or if the threshold
4035 * is zero we always do the raise.
4037 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4039 /* Log it stays the same */
4040 rack_log_timely(rack, 0, last_bw_est, low_bnd, 0,
4043 rack->rc_gp_timely_dec_cnt++;
4044 /* We are not incrementing really no-count */
4045 rack->rc_gp_incr = 0;
4046 rack->rc_gp_timely_inc_cnt = 0;
4049 * Lets just use the RTT
4050 * information and give up
4055 } else if ((timely_says != 2) &&
4057 (last_bw_est > up_bnd)) {
4059 * We are increasing b/w lets keep going, updating
4060 * our b/w and ignoring any timely input, unless
4061 * of course we are at our max raise (if there is one).
4064 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4066 rack->r_ctl.last_gp_comp_bw = cur_bw;
4067 if (rack->rc_gp_saw_ss &&
4068 rack_per_upper_bound_ss &&
4069 (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4071 * In cases where we can't go higher
4072 * we should just use timely.
4076 if (rack->rc_gp_saw_ca &&
4077 rack_per_upper_bound_ca &&
4078 (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4080 * In cases where we can't go higher
4081 * we should just use timely.
4085 rack->rc_gp_bwred = 0;
4086 rack->rc_gp_timely_dec_cnt = 0;
4087 /* You get a set number of pushes if timely is trying to reduce */
4088 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4089 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4091 /* Log it stays the same */
4092 rack_log_timely(rack, 0, last_bw_est, up_bnd, 0,
4098 * We are staying between the lower and upper range bounds
4099 * so use timely to decide.
4101 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4105 rack->rc_gp_incr = 0;
4106 rack->rc_gp_timely_inc_cnt = 0;
4107 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4109 (last_bw_est < low_bnd)) {
4110 /* We are loosing ground */
4111 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4112 rack->rc_gp_timely_dec_cnt++;
4113 /* We are not incrementing really no-count */
4114 rack->rc_gp_incr = 0;
4115 rack->rc_gp_timely_inc_cnt = 0;
4117 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4119 rack->rc_gp_bwred = 0;
4120 rack->rc_gp_timely_dec_cnt = 0;
4121 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4127 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4129 int32_t timely_says;
4130 uint64_t log_mult, log_rtt_a_diff;
4132 log_rtt_a_diff = rtt;
4133 log_rtt_a_diff <<= 32;
4134 log_rtt_a_diff |= (uint32_t)rtt_diff;
4135 if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4136 rack_gp_rtt_maxmul)) {
4137 /* Reduce the b/w multipler */
4139 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4141 log_mult |= prev_rtt;
4142 rack_log_timely(rack, timely_says, log_mult,
4143 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4144 log_rtt_a_diff, __LINE__, 4);
4145 } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4146 ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4147 max(rack_gp_rtt_mindiv , 1)))) {
4148 /* Increase the b/w multipler */
4149 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4150 ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4151 max(rack_gp_rtt_mindiv , 1));
4153 log_mult |= prev_rtt;
4155 rack_log_timely(rack, timely_says, log_mult ,
4156 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4157 log_rtt_a_diff, __LINE__, 5);
4160 * Use a gradient to find it the timely gradient
4162 * grad = rc_rtt_diff / min_rtt;
4164 * anything below or equal to 0 will be
4165 * a increase indication. Anything above
4166 * zero is a decrease. Note we take care
4167 * of the actual gradient calculation
4168 * in the reduction (its not needed for
4171 log_mult = prev_rtt;
4172 if (rtt_diff <= 0) {
4174 * Rttdiff is less than zero, increase the
4175 * b/w multipler (its 0 or negative)
4178 rack_log_timely(rack, timely_says, log_mult,
4179 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4181 /* Reduce the b/w multipler */
4183 rack_log_timely(rack, timely_says, log_mult,
4184 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4187 return (timely_says);
4191 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4192 tcp_seq th_ack, int line)
4194 uint64_t tim, bytes_ps, ltim, stim, utim;
4195 uint32_t segsiz, bytes, reqbytes, us_cts;
4196 int32_t gput, new_rtt_diff, timely_says;
4197 uint64_t resid_bw, subpart = 0, addpart = 0, srtt;
4200 us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4201 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4202 if (TSTMP_GEQ(us_cts, tp->gput_ts))
4203 tim = us_cts - tp->gput_ts;
4207 if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4208 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4212 * Use the larger of the send time or ack time. This prevents us
4213 * from being influenced by ack artifacts to come up with too
4214 * high of measurement. Note that since we are spanning over many more
4215 * bytes in most of our measurements hopefully that is less likely to
4221 utim = max(stim, 1);
4222 /* Lets get a msec time ltim too for the old stuff */
4223 ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4224 gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4225 reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4226 if ((tim == 0) && (stim == 0)) {
4228 * Invalid measurement time, maybe
4229 * all on one ack/one send?
4233 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4234 0, 0, 0, 10, __LINE__, NULL);
4235 goto skip_measurement;
4237 if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4238 /* We never made a us_rtt measurement? */
4241 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4242 0, 0, 0, 10, __LINE__, NULL);
4243 goto skip_measurement;
4246 * Calculate the maximum possible b/w this connection
4247 * could have. We base our calculation on the lowest
4248 * rtt we have seen during the measurement and the
4249 * largest rwnd the client has given us in that time. This
4250 * forms a BDP that is the maximum that we could ever
4251 * get to the client. Anything larger is not valid.
4253 * I originally had code here that rejected measurements
4254 * where the time was less than 1/2 the latest us_rtt.
4255 * But after thinking on that I realized its wrong since
4256 * say you had a 150Mbps or even 1Gbps link, and you
4257 * were a long way away.. example I am in Europe (100ms rtt)
4258 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4259 * bytes my time would be 1.2ms, and yet my rtt would say
4260 * the measurement was invalid the time was < 50ms. The
4261 * same thing is true for 150Mb (8ms of time).
4263 * A better way I realized is to look at what the maximum
4264 * the connection could possibly do. This is gated on
4265 * the lowest RTT we have seen and the highest rwnd.
4266 * We should in theory never exceed that, if we are
4267 * then something on the path is storing up packets
4268 * and then feeding them all at once to our endpoint
4269 * messing up our measurement.
4271 rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4272 rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4273 rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4274 if (SEQ_LT(th_ack, tp->gput_seq)) {
4275 /* No measurement can be made */
4278 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4279 0, 0, 0, 10, __LINE__, NULL);
4280 goto skip_measurement;
4282 bytes = (th_ack - tp->gput_seq);
4283 bytes_ps = (uint64_t)bytes;
4285 * Don't measure a b/w for pacing unless we have gotten at least
4286 * an initial windows worth of data in this measurement interval.
4288 * Small numbers of bytes get badly influenced by delayed ack and
4289 * other artifacts. Note we take the initial window or our
4290 * defined minimum GP (defaulting to 10 which hopefully is the
4293 if (rack->rc_gp_filled == 0) {
4295 * The initial estimate is special. We
4296 * have blasted out an IW worth of packets
4297 * without a real valid ack ts results. We
4298 * then setup the app_limited_needs_set flag,
4299 * this should get the first ack in (probably 2
4300 * MSS worth) to be recorded as the timestamp.
4301 * We thus allow a smaller number of bytes i.e.
4304 reqbytes -= (2 * segsiz);
4305 /* Also lets fill previous for our first measurement to be neutral */
4306 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4308 if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4309 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4310 rack->r_ctl.rc_app_limited_cnt,
4311 0, 0, 10, __LINE__, NULL);
4312 goto skip_measurement;
4315 * We now need to calculate the Timely like status so
4316 * we can update (possibly) the b/w multipliers.
4318 new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4319 if (rack->rc_gp_filled == 0) {
4320 /* No previous reading */
4321 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4323 if (rack->measure_saw_probe_rtt == 0) {
4325 * We don't want a probertt to be counted
4326 * since it will be negative incorrectly. We
4327 * expect to be reducing the RTT when we
4328 * pace at a slower rate.
4330 rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4331 rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4334 timely_says = rack_make_timely_judgement(rack,
4335 rack->r_ctl.rc_gp_srtt,
4336 rack->r_ctl.rc_rtt_diff,
4337 rack->r_ctl.rc_prev_gp_srtt
4339 bytes_ps *= HPTS_USEC_IN_SEC;
4341 if (bytes_ps > rack->r_ctl.last_max_bw) {
4343 * Something is on path playing
4344 * since this b/w is not possible based
4345 * on our BDP (highest rwnd and lowest rtt
4346 * we saw in the measurement window).
4348 * Another option here would be to
4349 * instead skip the measurement.
4351 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4352 bytes_ps, rack->r_ctl.last_max_bw, 0,
4353 11, __LINE__, NULL);
4354 bytes_ps = rack->r_ctl.last_max_bw;
4356 /* We store gp for b/w in bytes per second */
4357 if (rack->rc_gp_filled == 0) {
4358 /* Initial measurment */
4360 rack->r_ctl.gp_bw = bytes_ps;
4361 rack->rc_gp_filled = 1;
4362 rack->r_ctl.num_measurements = 1;
4363 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4365 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4366 rack->r_ctl.rc_app_limited_cnt,
4367 0, 0, 10, __LINE__, NULL);
4369 if (rack->rc_inp->inp_in_hpts &&
4370 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4372 * Ok we can't trust the pacer in this case
4373 * where we transition from un-paced to paced.
4374 * Or for that matter when the burst mitigation
4375 * was making a wild guess and got it wrong.
4376 * Stop the pacer and clear up all the aggregate
4379 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4380 rack->r_ctl.rc_hpts_flags = 0;
4381 rack->r_ctl.rc_last_output_to = 0;
4384 } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4385 /* Still a small number run an average */
4386 rack->r_ctl.gp_bw += bytes_ps;
4387 addpart = rack->r_ctl.num_measurements;
4388 rack->r_ctl.num_measurements++;
4389 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4390 /* We have collected enought to move forward */
4391 rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4396 * We want to take 1/wma of the goodput and add in to 7/8th
4397 * of the old value weighted by the srtt. So if your measurement
4398 * period is say 2 SRTT's long you would get 1/4 as the
4399 * value, if it was like 1/2 SRTT then you would get 1/16th.
4401 * But we must be careful not to take too much i.e. if the
4402 * srtt is say 20ms and the measurement is taken over
4403 * 400ms our weight would be 400/20 i.e. 20. On the
4404 * other hand if we get a measurement over 1ms with a
4405 * 10ms rtt we only want to take a much smaller portion.
4407 if (rack->r_ctl.num_measurements < 0xff) {
4408 rack->r_ctl.num_measurements++;
4410 srtt = (uint64_t)tp->t_srtt;
4413 * Strange why did t_srtt go back to zero?
4415 if (rack->r_ctl.rc_rack_min_rtt)
4416 srtt = rack->r_ctl.rc_rack_min_rtt;
4418 srtt = HPTS_USEC_IN_MSEC;
4421 * XXXrrs: Note for reviewers, in playing with
4422 * dynamic pacing I discovered this GP calculation
4423 * as done originally leads to some undesired results.
4424 * Basically you can get longer measurements contributing
4425 * too much to the WMA. Thus I changed it if you are doing
4426 * dynamic adjustments to only do the aportioned adjustment
4427 * if we have a very small (time wise) measurement. Longer
4428 * measurements just get there weight (defaulting to 1/8)
4429 * add to the WMA. We may want to think about changing
4430 * this to always do that for both sides i.e. dynamic
4431 * and non-dynamic... but considering lots of folks
4432 * were playing with this I did not want to change the
4433 * calculation per.se. without your thoughts.. Lawerence?
4436 if (rack->rc_gp_dyn_mul == 0) {
4437 subpart = rack->r_ctl.gp_bw * utim;
4438 subpart /= (srtt * 8);
4439 if (subpart < (rack->r_ctl.gp_bw / 2)) {
4441 * The b/w update takes no more
4442 * away then 1/2 our running total
4445 addpart = bytes_ps * utim;
4446 addpart /= (srtt * 8);
4449 * Don't allow a single measurement
4450 * to account for more than 1/2 of the
4451 * WMA. This could happen on a retransmission
4452 * where utim becomes huge compared to
4453 * srtt (multiple retransmissions when using
4454 * the sending rate which factors in all the
4455 * transmissions from the first one).
4457 subpart = rack->r_ctl.gp_bw / 2;
4458 addpart = bytes_ps / 2;
4460 resid_bw = rack->r_ctl.gp_bw - subpart;
4461 rack->r_ctl.gp_bw = resid_bw + addpart;
4464 if ((utim / srtt) <= 1) {
4466 * The b/w update was over a small period
4467 * of time. The idea here is to prevent a small
4468 * measurement time period from counting
4469 * too much. So we scale it based on the
4470 * time so it attributes less than 1/rack_wma_divisor
4471 * of its measurement.
4473 subpart = rack->r_ctl.gp_bw * utim;
4474 subpart /= (srtt * rack_wma_divisor);
4475 addpart = bytes_ps * utim;
4476 addpart /= (srtt * rack_wma_divisor);
4479 * The scaled measurement was long
4480 * enough so lets just add in the
4481 * portion of the measurment i.e. 1/rack_wma_divisor
4483 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4484 addpart = bytes_ps / rack_wma_divisor;
4486 if ((rack->measure_saw_probe_rtt == 0) ||
4487 (bytes_ps > rack->r_ctl.gp_bw)) {
4489 * For probe-rtt we only add it in
4490 * if its larger, all others we just
4494 resid_bw = rack->r_ctl.gp_bw - subpart;
4495 rack->r_ctl.gp_bw = resid_bw + addpart;
4499 if ((rack->gp_ready == 0) &&
4500 (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4501 /* We have enough measurements now */
4503 rack_set_cc_pacing(rack);
4504 if (rack->defer_options)
4505 rack_apply_deferred_options(rack);
4507 rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4508 rack_get_bw(rack), 22, did_add, NULL);
4509 /* We do not update any multipliers if we are in or have seen a probe-rtt */
4510 if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4511 rack_update_multiplier(rack, timely_says, bytes_ps,
4512 rack->r_ctl.rc_gp_srtt,
4513 rack->r_ctl.rc_rtt_diff);
4514 rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4515 rack_get_bw(rack), 3, line, NULL);
4516 /* reset the gp srtt and setup the new prev */
4517 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4518 /* Record the lost count for the next measurement */
4519 rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4521 * We restart our diffs based on the gpsrtt in the
4522 * measurement window.
4524 rack->rc_gp_rtt_set = 0;
4525 rack->rc_gp_saw_rec = 0;
4526 rack->rc_gp_saw_ca = 0;
4527 rack->rc_gp_saw_ss = 0;
4528 rack->rc_dragged_bottom = 0;
4532 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4535 * XXXLAS: This is a temporary hack, and should be
4536 * chained off VOI_TCP_GPUT when stats(9) grows an
4537 * API to deal with chained VOIs.
4539 if (tp->t_stats_gput_prev > 0)
4540 stats_voi_update_abs_s32(tp->t_stats,
4542 ((gput - tp->t_stats_gput_prev) * 100) /
4543 tp->t_stats_gput_prev);
4545 tp->t_flags &= ~TF_GPUTINPROG;
4546 tp->t_stats_gput_prev = gput;
4548 * Now are we app limited now and there is space from where we
4549 * were to where we want to go?
4551 * We don't do the other case i.e. non-applimited here since
4552 * the next send will trigger us picking up the missing data.
4554 if (rack->r_ctl.rc_first_appl &&
4555 TCPS_HAVEESTABLISHED(tp->t_state) &&
4556 rack->r_ctl.rc_app_limited_cnt &&
4557 (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4558 ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
4559 max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4561 * Yep there is enough outstanding to make a measurement here.
4563 struct rack_sendmap *rsm, fe;
4565 tp->t_flags |= TF_GPUTINPROG;
4566 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4567 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4568 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4569 rack->app_limited_needs_set = 0;
4570 tp->gput_seq = th_ack;
4571 if (rack->in_probe_rtt)
4572 rack->measure_saw_probe_rtt = 1;
4573 else if ((rack->measure_saw_probe_rtt) &&
4574 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4575 rack->measure_saw_probe_rtt = 0;
4576 if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
4577 /* There is a full window to gain info from */
4578 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4580 /* We can only measure up to the applimited point */
4581 tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
4584 * Now we need to find the timestamp of the send at tp->gput_seq
4585 * for the send based measurement.
4587 fe.r_start = tp->gput_seq;
4588 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4590 /* Ok send-based limit is set */
4591 if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4593 * Move back to include the earlier part
4594 * so our ack time lines up right (this may
4595 * make an overlapping measurement but thats
4598 tp->gput_seq = rsm->r_start;
4600 if (rsm->r_flags & RACK_ACKED)
4601 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4603 rack->app_limited_needs_set = 1;
4604 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4607 * If we don't find the rsm due to some
4608 * send-limit set the current time, which
4609 * basically disables the send-limit.
4614 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4616 rack_log_pacing_delay_calc(rack,
4621 rack->r_ctl.rc_app_limited_cnt,
4628 * CC wrapper hook functions
4631 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4632 uint16_t type, int32_t recovery)
4634 uint32_t prior_cwnd, acked;
4635 struct tcp_log_buffer *lgb = NULL;
4636 uint8_t labc_to_use;
4638 INP_WLOCK_ASSERT(tp->t_inpcb);
4639 tp->ccv->nsegs = nsegs;
4640 acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4641 if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4644 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4645 if (tp->ccv->bytes_this_ack > max) {
4646 tp->ccv->bytes_this_ack = max;
4650 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4651 ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4653 if ((tp->t_flags & TF_GPUTINPROG) &&
4654 rack_enough_for_measurement(tp, rack, th_ack)) {
4655 /* Measure the Goodput */
4656 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__);
4657 #ifdef NETFLIX_PEAKRATE
4658 if ((type == CC_ACK) &&
4659 (tp->t_maxpeakrate)) {
4661 * We update t_peakrate_thr. This gives us roughly
4662 * one update per round trip time. Note
4663 * it will only be used if pace_always is off i.e
4664 * we don't do this for paced flows.
4666 rack_update_peakrate_thr(tp);
4670 /* Which way our we limited, if not cwnd limited no advance in CA */
4671 if (tp->snd_cwnd <= tp->snd_wnd)
4672 tp->ccv->flags |= CCF_CWND_LIMITED;
4674 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4675 if (tp->snd_cwnd > tp->snd_ssthresh) {
4676 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4677 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4678 /* For the setting of a window past use the actual scwnd we are using */
4679 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4680 tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4681 tp->ccv->flags |= CCF_ABC_SENTAWND;
4684 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4685 tp->t_bytes_acked = 0;
4687 prior_cwnd = tp->snd_cwnd;
4688 if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4689 (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4690 labc_to_use = rack->rc_labc;
4692 labc_to_use = rack_max_abc_post_recovery;
4693 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4694 union tcp_log_stackspecific log;
4697 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4698 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4699 log.u_bbr.flex1 = th_ack;
4700 log.u_bbr.flex2 = tp->ccv->flags;
4701 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4702 log.u_bbr.flex4 = tp->ccv->nsegs;
4703 log.u_bbr.flex5 = labc_to_use;
4704 log.u_bbr.flex6 = prior_cwnd;
4705 log.u_bbr.flex7 = V_tcp_do_newsack;
4706 log.u_bbr.flex8 = 1;
4707 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4708 0, &log, false, NULL, NULL, 0, &tv);
4710 if (CC_ALGO(tp)->ack_received != NULL) {
4711 /* XXXLAS: Find a way to live without this */
4712 tp->ccv->curack = th_ack;
4713 tp->ccv->labc = labc_to_use;
4714 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4715 CC_ALGO(tp)->ack_received(tp->ccv, type);
4718 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4720 if (rack->r_must_retran) {
4721 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4723 * We now are beyond the rxt point so lets disable
4726 rack->r_ctl.rc_out_at_rto = 0;
4727 rack->r_must_retran = 0;
4728 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4730 * Only decrement the rc_out_at_rto if the cwnd advances
4731 * at least a whole segment. Otherwise next time the peer
4732 * acks, we won't be able to send this generaly happens
4733 * when we are in Congestion Avoidance.
4735 if (acked <= rack->r_ctl.rc_out_at_rto){
4736 rack->r_ctl.rc_out_at_rto -= acked;
4738 rack->r_ctl.rc_out_at_rto = 0;
4743 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4745 if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4746 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4748 #ifdef NETFLIX_PEAKRATE
4749 /* we enforce max peak rate if it is set and we are not pacing */
4750 if ((rack->rc_always_pace == 0) &&
4751 tp->t_peakrate_thr &&
4752 (tp->snd_cwnd > tp->t_peakrate_thr)) {
4753 tp->snd_cwnd = tp->t_peakrate_thr;
4759 tcp_rack_partialack(struct tcpcb *tp)
4761 struct tcp_rack *rack;
4763 rack = (struct tcp_rack *)tp->t_fb_ptr;
4764 INP_WLOCK_ASSERT(tp->t_inpcb);
4766 * If we are doing PRR and have enough
4767 * room to send <or> we are pacing and prr
4768 * is disabled we will want to see if we
4769 * can send data (by setting r_wanted_output to
4772 if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4774 rack->r_wanted_output = 1;
4778 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4780 struct tcp_rack *rack;
4783 orig_cwnd = tp->snd_cwnd;
4784 INP_WLOCK_ASSERT(tp->t_inpcb);
4785 rack = (struct tcp_rack *)tp->t_fb_ptr;
4786 /* only alert CC if we alerted when we entered */
4787 if (CC_ALGO(tp)->post_recovery != NULL) {
4788 tp->ccv->curack = th_ack;
4789 CC_ALGO(tp)->post_recovery(tp->ccv);
4790 if (tp->snd_cwnd < tp->snd_ssthresh) {
4792 * Rack has burst control and pacing
4793 * so lets not set this any lower than
4794 * snd_ssthresh per RFC-6582 (option 2).
4796 tp->snd_cwnd = tp->snd_ssthresh;
4799 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4800 union tcp_log_stackspecific log;
4803 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4804 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4805 log.u_bbr.flex1 = th_ack;
4806 log.u_bbr.flex2 = tp->ccv->flags;
4807 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4808 log.u_bbr.flex4 = tp->ccv->nsegs;
4809 log.u_bbr.flex5 = V_tcp_abc_l_var;
4810 log.u_bbr.flex6 = orig_cwnd;
4811 log.u_bbr.flex7 = V_tcp_do_newsack;
4812 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4813 log.u_bbr.flex8 = 2;
4814 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4815 0, &log, false, NULL, NULL, 0, &tv);
4817 if ((rack->rack_no_prr == 0) &&
4818 (rack->no_prr_addback == 0) &&
4819 (rack->r_ctl.rc_prr_sndcnt > 0)) {
4821 * Suck the next prr cnt back into cwnd, but
4822 * only do that if we are not application limited.
4824 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4826 * We are allowed to add back to the cwnd the amount we did
4828 * a) no_prr_addback is off.
4829 * b) we are not app limited
4830 * c) we are doing prr
4832 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4834 tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4835 rack->r_ctl.rc_prr_sndcnt);
4837 rack->r_ctl.rc_prr_sndcnt = 0;
4838 rack_log_to_prr(rack, 1, 0);
4840 rack_log_to_prr(rack, 14, orig_cwnd);
4841 tp->snd_recover = tp->snd_una;
4842 EXIT_RECOVERY(tp->t_flags);
4846 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4848 struct tcp_rack *rack;
4849 uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4851 INP_WLOCK_ASSERT(tp->t_inpcb);
4853 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4855 if (IN_RECOVERY(tp->t_flags) == 0) {
4856 in_rec_at_entry = 0;
4857 ssthresh_enter = tp->snd_ssthresh;
4858 cwnd_enter = tp->snd_cwnd;
4860 in_rec_at_entry = 1;
4861 rack = (struct tcp_rack *)tp->t_fb_ptr;
4864 tp->t_flags &= ~TF_WASFRECOVERY;
4865 tp->t_flags &= ~TF_WASCRECOVERY;
4866 if (!IN_FASTRECOVERY(tp->t_flags)) {
4867 rack->r_ctl.rc_prr_delivered = 0;
4868 rack->r_ctl.rc_prr_out = 0;
4869 if (rack->rack_no_prr == 0) {
4870 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4871 rack_log_to_prr(rack, 2, in_rec_at_entry);
4873 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4874 tp->snd_recover = tp->snd_max;
4875 if (tp->t_flags2 & TF2_ECN_PERMIT)
4876 tp->t_flags2 |= TF2_ECN_SND_CWR;
4880 if (!IN_CONGRECOVERY(tp->t_flags) ||
4882 * Allow ECN reaction on ACK to CWR, if
4883 * that data segment was also CE marked.
4885 SEQ_GEQ(ack, tp->snd_recover)) {
4886 EXIT_CONGRECOVERY(tp->t_flags);
4887 KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4888 tp->snd_recover = tp->snd_max + 1;
4889 if (tp->t_flags2 & TF2_ECN_PERMIT)
4890 tp->t_flags2 |= TF2_ECN_SND_CWR;
4895 tp->t_bytes_acked = 0;
4896 EXIT_RECOVERY(tp->t_flags);
4897 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4898 ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4899 orig_cwnd = tp->snd_cwnd;
4900 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4901 rack_log_to_prr(rack, 16, orig_cwnd);
4902 if (tp->t_flags2 & TF2_ECN_PERMIT)
4903 tp->t_flags2 |= TF2_ECN_SND_CWR;
4906 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4907 /* RTO was unnecessary, so reset everything. */
4908 tp->snd_cwnd = tp->snd_cwnd_prev;
4909 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4910 tp->snd_recover = tp->snd_recover_prev;
4911 if (tp->t_flags & TF_WASFRECOVERY) {
4912 ENTER_FASTRECOVERY(tp->t_flags);
4913 tp->t_flags &= ~TF_WASFRECOVERY;
4915 if (tp->t_flags & TF_WASCRECOVERY) {
4916 ENTER_CONGRECOVERY(tp->t_flags);
4917 tp->t_flags &= ~TF_WASCRECOVERY;
4919 tp->snd_nxt = tp->snd_max;
4920 tp->t_badrxtwin = 0;
4923 if ((CC_ALGO(tp)->cong_signal != NULL) &&
4925 tp->ccv->curack = ack;
4926 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4928 if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4929 rack_log_to_prr(rack, 15, cwnd_enter);
4930 rack->r_ctl.dsack_byte_cnt = 0;
4931 rack->r_ctl.retran_during_recovery = 0;
4932 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4933 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4934 rack->r_ent_rec_ns = 1;
4939 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4943 INP_WLOCK_ASSERT(tp->t_inpcb);
4945 #ifdef NETFLIX_STATS
4946 KMOD_TCPSTAT_INC(tcps_idle_restarts);
4947 if (tp->t_state == TCPS_ESTABLISHED)
4948 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4950 if (CC_ALGO(tp)->after_idle != NULL)
4951 CC_ALGO(tp)->after_idle(tp->ccv);
4953 if (tp->snd_cwnd == 1)
4954 i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */
4956 i_cwnd = rc_init_window(rack);
4959 * Being idle is no differnt than the initial window. If the cc
4960 * clamps it down below the initial window raise it to the initial
4963 if (tp->snd_cwnd < i_cwnd) {
4964 tp->snd_cwnd = i_cwnd;
4969 * Indicate whether this ack should be delayed. We can delay the ack if
4970 * following conditions are met:
4971 * - There is no delayed ack timer in progress.
4972 * - Our last ack wasn't a 0-sized window. We never want to delay
4973 * the ack that opens up a 0-sized window.
4974 * - LRO wasn't used for this segment. We make sure by checking that the
4975 * segment size is not larger than the MSS.
4976 * - Delayed acks are enabled or this is a half-synchronized T/TCP
4979 #define DELAY_ACK(tp, tlen) \
4980 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4981 ((tp->t_flags & TF_DELACK) == 0) && \
4982 (tlen <= tp->t_maxseg) && \
4983 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4985 static struct rack_sendmap *
4986 rack_find_lowest_rsm(struct tcp_rack *rack)
4988 struct rack_sendmap *rsm;
4991 * Walk the time-order transmitted list looking for an rsm that is
4992 * not acked. This will be the one that was sent the longest time
4993 * ago that is still outstanding.
4995 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4996 if (rsm->r_flags & RACK_ACKED) {
5005 static struct rack_sendmap *
5006 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5008 struct rack_sendmap *prsm;
5011 * Walk the sequence order list backward until we hit and arrive at
5012 * the highest seq not acked. In theory when this is called it
5013 * should be the last segment (which it was not).
5015 counter_u64_add(rack_find_high, 1);
5017 RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5018 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5027 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5033 * lro is the flag we use to determine if we have seen reordering.
5034 * If it gets set we have seen reordering. The reorder logic either
5035 * works in one of two ways:
5037 * If reorder-fade is configured, then we track the last time we saw
5038 * re-ordering occur. If we reach the point where enough time as
5039 * passed we no longer consider reordering has occuring.
5041 * Or if reorder-face is 0, then once we see reordering we consider
5042 * the connection to alway be subject to reordering and just set lro
5045 * In the end if lro is non-zero we add the extra time for
5050 if (rack->r_ctl.rc_reorder_ts) {
5051 if (rack->r_ctl.rc_reorder_fade) {
5052 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5053 lro = cts - rack->r_ctl.rc_reorder_ts;
5056 * No time as passed since the last
5057 * reorder, mark it as reordering.
5062 /* Negative time? */
5065 if (lro > rack->r_ctl.rc_reorder_fade) {
5066 /* Turn off reordering seen too */
5067 rack->r_ctl.rc_reorder_ts = 0;
5071 /* Reodering does not fade */
5077 thresh = srtt + rack->r_ctl.rc_pkt_delay;
5079 /* It must be set, if not you get 1/4 rtt */
5080 if (rack->r_ctl.rc_reorder_shift)
5081 thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5083 thresh += (srtt >> 2);
5087 /* We don't let the rack timeout be above a RTO */
5088 if (thresh > rack->rc_tp->t_rxtcur) {
5089 thresh = rack->rc_tp->t_rxtcur;
5091 /* And we don't want it above the RTO max either */
5092 if (thresh > rack_rto_max) {
5093 thresh = rack_rto_max;
5099 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5100 struct rack_sendmap *rsm, uint32_t srtt)
5102 struct rack_sendmap *prsm;
5103 uint32_t thresh, len;
5108 if (rack->r_ctl.rc_tlp_threshold)
5109 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5111 thresh = (srtt * 2);
5113 /* Get the previous sent packet, if any */
5114 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5115 counter_u64_add(rack_enter_tlp_calc, 1);
5116 len = rsm->r_end - rsm->r_start;
5117 if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5118 /* Exactly like the ID */
5119 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5120 uint32_t alt_thresh;
5122 * Compensate for delayed-ack with the d-ack time.
5124 counter_u64_add(rack_used_tlpmethod, 1);
5125 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5126 if (alt_thresh > thresh)
5127 thresh = alt_thresh;
5129 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5131 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5132 if (prsm && (len <= segsiz)) {
5134 * Two packets outstanding, thresh should be (2*srtt) +
5135 * possible inter-packet delay (if any).
5137 uint32_t inter_gap = 0;
5140 counter_u64_add(rack_used_tlpmethod, 1);
5141 idx = rsm->r_rtr_cnt - 1;
5142 nidx = prsm->r_rtr_cnt - 1;
5143 if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5144 /* Yes it was sent later (or at the same time) */
5145 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5147 thresh += inter_gap;
5148 } else if (len <= segsiz) {
5150 * Possibly compensate for delayed-ack.
5152 uint32_t alt_thresh;
5154 counter_u64_add(rack_used_tlpmethod2, 1);
5155 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5156 if (alt_thresh > thresh)
5157 thresh = alt_thresh;
5159 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5161 if (len <= segsiz) {
5162 uint32_t alt_thresh;
5164 * Compensate for delayed-ack with the d-ack time.
5166 counter_u64_add(rack_used_tlpmethod, 1);
5167 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5168 if (alt_thresh > thresh)
5169 thresh = alt_thresh;
5172 /* Not above an RTO */
5173 if (thresh > tp->t_rxtcur) {
5174 thresh = tp->t_rxtcur;
5176 /* Not above a RTO max */
5177 if (thresh > rack_rto_max) {
5178 thresh = rack_rto_max;
5180 /* Apply user supplied min TLP */
5181 if (thresh < rack_tlp_min) {
5182 thresh = rack_tlp_min;
5188 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5191 * We want the rack_rtt which is the
5192 * last rtt we measured. However if that
5193 * does not exist we fallback to the srtt (which
5194 * we probably will never do) and then as a last
5195 * resort we use RACK_INITIAL_RTO if no srtt is
5198 if (rack->rc_rack_rtt)
5199 return (rack->rc_rack_rtt);
5200 else if (tp->t_srtt == 0)
5201 return (RACK_INITIAL_RTO);
5202 return (tp->t_srtt);
5205 static struct rack_sendmap *
5206 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5209 * Check to see that we don't need to fall into recovery. We will
5210 * need to do so if our oldest transmit is past the time we should
5213 struct tcp_rack *rack;
5214 struct rack_sendmap *rsm;
5216 uint32_t srtt, thresh;
5218 rack = (struct tcp_rack *)tp->t_fb_ptr;
5219 if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5222 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5226 if (rsm->r_flags & RACK_ACKED) {
5227 rsm = rack_find_lowest_rsm(rack);
5231 idx = rsm->r_rtr_cnt - 1;
5232 srtt = rack_grab_rtt(tp, rack);
5233 thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5234 if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5237 if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5240 /* Ok if we reach here we are over-due and this guy can be sent */
5241 if (IN_RECOVERY(tp->t_flags) == 0) {
5243 * For the one that enters us into recovery record undo
5246 rack->r_ctl.rc_rsm_start = rsm->r_start;
5247 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5248 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5250 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5255 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5261 t = (tp->t_srtt + (tp->t_rttvar << 2));
5262 RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5263 rack_persist_min, rack_persist_max);
5264 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5266 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5267 ret_val = (uint32_t)tt;
5272 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5275 * Start the FR timer, we do this based on getting the first one in
5276 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5277 * events we need to stop the running timer (if its running) before
5278 * starting the new one.
5280 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5283 int32_t is_tlp_timer = 0;
5284 struct rack_sendmap *rsm;
5286 if (rack->t_timers_stopped) {
5287 /* All timers have been stopped none are to run */
5290 if (rack->rc_in_persist) {
5291 /* We can't start any timer in persists */
5292 return (rack_get_persists_timer_val(tp, rack));
5294 rack->rc_on_min_to = 0;
5295 if ((tp->t_state < TCPS_ESTABLISHED) ||
5296 ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5299 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5300 if ((rsm == NULL) || sup_rack) {
5301 /* Nothing on the send map or no rack */
5303 time_since_sent = 0;
5304 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5307 * Should we discount the RTX timer any?
5309 * We want to discount it the smallest amount.
5310 * If a timer (Rack/TLP or RXT) has gone off more
5311 * recently thats the discount we want to use (now - timer time).
5312 * If the retransmit of the oldest packet was more recent then
5313 * we want to use that (now - oldest-packet-last_transmit_time).
5316 idx = rsm->r_rtr_cnt - 1;
5317 if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5318 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5320 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5321 if (TSTMP_GT(cts, tstmp_touse))
5322 time_since_sent = cts - tstmp_touse;
5324 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5325 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5327 if (to > time_since_sent)
5328 to -= time_since_sent;
5330 to = rack->r_ctl.rc_min_to;
5333 /* Special case for KEEPINIT */
5334 if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5335 (TP_KEEPINIT(tp) != 0) &&
5338 * We have to put a ceiling on the rxt timer
5339 * of the keep-init timeout.
5341 uint32_t max_time, red;
5343 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5344 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5345 red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5351 /* Reduce timeout to the keep value if needed */
5359 if (rsm->r_flags & RACK_ACKED) {
5360 rsm = rack_find_lowest_rsm(rack);
5366 if (rack->sack_attack_disable) {
5368 * We don't want to do
5369 * any TLP's if you are an attacker.
5370 * Though if you are doing what
5371 * is expected you may still have
5372 * SACK-PASSED marks.
5376 /* Convert from ms to usecs */
5377 if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5378 if ((tp->t_flags & TF_SENTFIN) &&
5379 ((tp->snd_max - tp->snd_una) == 1) &&
5380 (rsm->r_flags & RACK_HAS_FIN)) {
5382 * We don't start a rack timer if all we have is a
5387 if ((rack->use_rack_rr == 0) &&
5388 (IN_FASTRECOVERY(tp->t_flags)) &&
5389 (rack->rack_no_prr == 0) &&
5390 (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) {
5392 * We are not cheating, in recovery and
5393 * not enough ack's to yet get our next
5394 * retransmission out.
5396 * Note that classified attackers do not
5397 * get to use the rack-cheat.
5401 srtt = rack_grab_rtt(tp, rack);
5402 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5403 idx = rsm->r_rtr_cnt - 1;
5404 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5405 if (SEQ_GEQ(exp, cts)) {
5407 if (to < rack->r_ctl.rc_min_to) {
5408 to = rack->r_ctl.rc_min_to;
5409 if (rack->r_rr_config == 3)
5410 rack->rc_on_min_to = 1;
5413 to = rack->r_ctl.rc_min_to;
5414 if (rack->r_rr_config == 3)
5415 rack->rc_on_min_to = 1;
5418 /* Ok we need to do a TLP not RACK */
5420 if ((rack->rc_tlp_in_progress != 0) &&
5421 (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5423 * The previous send was a TLP and we have sent
5424 * N TLP's without sending new data.
5428 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5430 /* We found no rsm to TLP with. */
5433 if (rsm->r_flags & RACK_HAS_FIN) {
5434 /* If its a FIN we dont do TLP */
5438 idx = rsm->r_rtr_cnt - 1;
5439 time_since_sent = 0;
5440 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5441 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5443 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5444 if (TSTMP_GT(cts, tstmp_touse))
5445 time_since_sent = cts - tstmp_touse;
5448 if ((rack->rc_srtt_measure_made == 0) &&
5449 (tp->t_srtt == 1)) {
5451 * If another stack as run and set srtt to 1,
5452 * then the srtt was 0, so lets use the initial.
5454 srtt = RACK_INITIAL_RTO;
5456 srtt_cur = tp->t_srtt;
5460 srtt = RACK_INITIAL_RTO;
5462 * If the SRTT is not keeping up and the
5463 * rack RTT has spiked we want to use
5464 * the last RTT not the smoothed one.
5466 if (rack_tlp_use_greater &&
5468 (srtt < rack_grab_rtt(tp, rack))) {
5469 srtt = rack_grab_rtt(tp, rack);
5471 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5472 if (thresh > time_since_sent) {
5473 to = thresh - time_since_sent;
5475 to = rack->r_ctl.rc_min_to;
5476 rack_log_alt_to_to_cancel(rack,
5478 time_since_sent, /* flex2 */
5479 tstmp_touse, /* flex3 */
5480 rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5481 (uint32_t)rsm->r_tim_lastsent[idx],
5485 if (to < rack_tlp_min) {
5488 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5490 * If the TLP time works out to larger than the max
5491 * RTO lets not do TLP.. just RTO.
5496 if (is_tlp_timer == 0) {
5497 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5499 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5507 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5509 if (rack->rc_in_persist == 0) {
5510 if (tp->t_flags & TF_GPUTINPROG) {
5512 * Stop the goodput now, the calling of the
5513 * measurement function clears the flag.
5515 rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
5517 #ifdef NETFLIX_SHARED_CWND
5518 if (rack->r_ctl.rc_scw) {
5519 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5520 rack->rack_scwnd_is_idle = 1;
5523 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5524 if (rack->r_ctl.rc_went_idle_time == 0)
5525 rack->r_ctl.rc_went_idle_time = 1;
5526 rack_timer_cancel(tp, rack, cts, __LINE__);
5528 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5529 rack_rto_min, rack_rto_max);
5530 rack->rc_in_persist = 1;
5535 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5537 if (rack->rc_inp->inp_in_hpts) {
5538 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5539 rack->r_ctl.rc_hpts_flags = 0;
5541 #ifdef NETFLIX_SHARED_CWND
5542 if (rack->r_ctl.rc_scw) {
5543 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5544 rack->rack_scwnd_is_idle = 0;
5547 if (rack->rc_gp_dyn_mul &&
5548 (rack->use_fixed_rate == 0) &&
5549 (rack->rc_always_pace)) {
5551 * Do we count this as if a probe-rtt just
5554 uint32_t time_idle, idle_min;
5556 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5557 idle_min = rack_min_probertt_hold;
5558 if (rack_probertt_gpsrtt_cnt_div) {
5560 extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5561 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5562 extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5563 idle_min += (uint32_t)extra;
5565 if (time_idle >= idle_min) {
5566 /* Yes, we count it as a probe-rtt. */
5569 us_cts = tcp_get_usecs(NULL);
5570 if (rack->in_probe_rtt == 0) {
5571 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5572 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5573 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5574 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5576 rack_exit_probertt(rack, us_cts);
5580 rack->rc_in_persist = 0;
5581 rack->r_ctl.rc_went_idle_time = 0;
5583 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5584 rack_rto_min, rack_rto_max);
5585 rack->r_ctl.rc_agg_delayed = 0;
5588 rack->r_ctl.rc_agg_early = 0;
5592 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5593 struct hpts_diag *diag, struct timeval *tv)
5595 if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5596 union tcp_log_stackspecific log;
5598 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5599 log.u_bbr.flex1 = diag->p_nxt_slot;
5600 log.u_bbr.flex2 = diag->p_cur_slot;
5601 log.u_bbr.flex3 = diag->slot_req;
5602 log.u_bbr.flex4 = diag->inp_hptsslot;
5603 log.u_bbr.flex5 = diag->slot_remaining;
5604 log.u_bbr.flex6 = diag->need_new_to;
5605 log.u_bbr.flex7 = diag->p_hpts_active;
5606 log.u_bbr.flex8 = diag->p_on_min_sleep;
5607 /* Hijack other fields as needed */
5608 log.u_bbr.epoch = diag->have_slept;
5609 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5610 log.u_bbr.pkts_out = diag->co_ret;
5611 log.u_bbr.applimited = diag->hpts_sleep_time;
5612 log.u_bbr.delivered = diag->p_prev_slot;
5613 log.u_bbr.inflight = diag->p_runningtick;
5614 log.u_bbr.bw_inuse = diag->wheel_tick;
5615 log.u_bbr.rttProp = diag->wheel_cts;
5616 log.u_bbr.timeStamp = cts;
5617 log.u_bbr.delRate = diag->maxticks;
5618 log.u_bbr.cur_del_rate = diag->p_curtick;
5619 log.u_bbr.cur_del_rate <<= 32;
5620 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5621 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5622 &rack->rc_inp->inp_socket->so_rcv,
5623 &rack->rc_inp->inp_socket->so_snd,
5624 BBR_LOG_HPTSDIAG, 0,
5625 0, &log, false, tv);
5631 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5633 if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5634 union tcp_log_stackspecific log;
5637 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5638 log.u_bbr.flex1 = sb->sb_flags;
5639 log.u_bbr.flex2 = len;
5640 log.u_bbr.flex3 = sb->sb_state;
5641 log.u_bbr.flex8 = type;
5642 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5643 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5644 &rack->rc_inp->inp_socket->so_rcv,
5645 &rack->rc_inp->inp_socket->so_snd,
5647 len, &log, false, &tv);
5652 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5653 int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5655 struct hpts_diag diag;
5658 uint32_t delayed_ack = 0;
5659 uint32_t hpts_timeout;
5660 uint32_t entry_slot = slot;
5666 if ((tp->t_state == TCPS_CLOSED) ||
5667 (tp->t_state == TCPS_LISTEN)) {
5670 if (inp->inp_in_hpts) {
5671 /* Already on the pacer */
5674 stopped = rack->rc_tmr_stopped;
5675 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5676 left = rack->r_ctl.rc_timer_exp - cts;
5678 rack->r_ctl.rc_timer_exp = 0;
5679 rack->r_ctl.rc_hpts_flags = 0;
5680 us_cts = tcp_get_usecs(&tv);
5681 /* Now early/late accounting */
5682 rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL);
5683 if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5685 * We have a early carry over set,
5686 * we can always add more time so we
5687 * can always make this compensation.
5689 * Note if ack's are allowed to wake us do not
5690 * penalize the next timer for being awoke
5691 * by an ack aka the rc_agg_early (non-paced mode).
5693 slot += rack->r_ctl.rc_agg_early;
5695 rack->r_ctl.rc_agg_early = 0;
5699 * This is harder, we can
5700 * compensate some but it
5701 * really depends on what
5702 * the current pacing time is.
5704 if (rack->r_ctl.rc_agg_delayed >= slot) {
5706 * We can't compensate for it all.
5707 * And we have to have some time
5708 * on the clock. We always have a min
5709 * 10 slots (10 x 10 i.e. 100 usecs).
5711 if (slot <= HPTS_TICKS_PER_USEC) {
5713 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
5714 slot = HPTS_TICKS_PER_USEC;
5716 /* We take off some */
5717 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
5718 slot = HPTS_TICKS_PER_USEC;
5721 slot -= rack->r_ctl.rc_agg_delayed;
5722 rack->r_ctl.rc_agg_delayed = 0;
5723 /* Make sure we have 100 useconds at minimum */
5724 if (slot < HPTS_TICKS_PER_USEC) {
5725 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
5726 slot = HPTS_TICKS_PER_USEC;
5728 if (rack->r_ctl.rc_agg_delayed == 0)
5733 /* We are pacing too */
5734 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5736 hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5737 #ifdef NETFLIX_EXP_DETECTION
5738 if (rack->sack_attack_disable &&
5739 (slot < tcp_sad_pacing_interval)) {
5741 * We have a potential attacker on
5742 * the line. We have possibly some
5743 * (or now) pacing time set. We want to
5744 * slow down the processing of sacks by some
5745 * amount (if it is an attacker). Set the default
5746 * slot for attackers in place (unless the orginal
5747 * interval is longer). Its stored in
5748 * micro-seconds, so lets convert to msecs.
5750 slot = tcp_sad_pacing_interval;
5753 if (tp->t_flags & TF_DELACK) {
5754 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5755 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5757 if (delayed_ack && ((hpts_timeout == 0) ||
5758 (delayed_ack < hpts_timeout)))
5759 hpts_timeout = delayed_ack;
5761 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5763 * If no timers are going to run and we will fall off the hptsi
5764 * wheel, we resort to a keep-alive timer if its configured.
5766 if ((hpts_timeout == 0) &&
5768 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5769 (tp->t_state <= TCPS_CLOSING)) {
5771 * Ok we have no timer (persists, rack, tlp, rxt or
5772 * del-ack), we don't have segments being paced. So
5773 * all that is left is the keepalive timer.
5775 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5776 /* Get the established keep-alive time */
5777 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5780 * Get the initial setup keep-alive time,
5781 * note that this is probably not going to
5782 * happen, since rack will be running a rxt timer
5783 * if a SYN of some sort is outstanding. It is
5784 * actually handled in rack_timeout_rxt().
5786 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5788 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5789 if (rack->in_probe_rtt) {
5791 * We want to instead not wake up a long time from
5792 * now but to wake up about the time we would
5793 * exit probe-rtt and initiate a keep-alive ack.
5794 * This will get us out of probe-rtt and update
5797 hpts_timeout = rack_min_probertt_hold;
5801 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5802 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5804 * RACK, TLP, persists and RXT timers all are restartable
5805 * based on actions input .. i.e we received a packet (ack
5806 * or sack) and that changes things (rw, or snd_una etc).
5807 * Thus we can restart them with a new value. For
5808 * keep-alive, delayed_ack we keep track of what was left
5809 * and restart the timer with a smaller value.
5811 if (left < hpts_timeout)
5812 hpts_timeout = left;
5816 * Hack alert for now we can't time-out over 2,147,483
5817 * seconds (a bit more than 596 hours), which is probably ok
5820 if (hpts_timeout > 0x7ffffffe)
5821 hpts_timeout = 0x7ffffffe;
5822 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5824 rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL);
5825 if ((rack->gp_ready == 0) &&
5826 (rack->use_fixed_rate == 0) &&
5827 (hpts_timeout < slot) &&
5828 (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5830 * We have no good estimate yet for the
5831 * old clunky burst mitigation or the
5832 * real pacing. And the tlp or rxt is smaller
5833 * than the pacing calculation. Lets not
5834 * pace that long since we know the calculation
5835 * so far is not accurate.
5837 slot = hpts_timeout;
5839 rack->r_ctl.last_pacing_time = slot;
5841 * Turn off all the flags for queuing by default. The
5842 * flags have important meanings to what happens when
5843 * LRO interacts with the transport. Most likely (by default now)
5844 * mbuf_queueing and ack compression are on. So the transport
5845 * has a couple of flags that control what happens (if those
5846 * are not on then these flags won't have any effect since it
5847 * won't go through the queuing LRO path).
5849 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5850 * pacing output, so don't disturb. But
5851 * it also means LRO can wake me if there
5852 * is a SACK arrival.
5854 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5855 * with the above flag (QUEUE_READY) and
5856 * when present it says don't even wake me
5857 * if a SACK arrives.
5859 * The idea behind these flags is that if we are pacing we
5860 * set the MBUF_QUEUE_READY and only get woken up if
5861 * a SACK arrives (which could change things) or if
5862 * our pacing timer expires. If, however, we have a rack
5863 * timer running, then we don't even want a sack to wake
5864 * us since the rack timer has to expire before we can send.
5866 * Other cases should usually have none of the flags set
5867 * so LRO can call into us.
5869 inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5871 rack->r_ctl.rc_last_output_to = us_cts + slot;
5873 * A pacing timer (slot) is being set, in
5874 * such a case we cannot send (we are blocked by
5875 * the timer). So lets tell LRO that it should not
5876 * wake us unless there is a SACK. Note this only
5877 * will be effective if mbuf queueing is on or
5878 * compressed acks are being processed.
5880 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5882 * But wait if we have a Rack timer running
5883 * even a SACK should not disturb us (with
5884 * the exception of r_rr_config 3).
5886 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5887 (rack->r_rr_config != 3))
5888 inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5889 if (rack->rc_ack_can_sendout_data) {
5891 * Ahh but wait, this is that special case
5892 * where the pacing timer can be disturbed
5893 * backout the changes (used for non-paced
5896 inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5898 if ((rack->use_rack_rr) &&
5899 (rack->r_rr_config < 2) &&
5900 ((hpts_timeout) && (hpts_timeout < slot))) {
5902 * Arrange for the hpts to kick back in after the
5903 * t-o if the t-o does not cause a send.
5905 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5907 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5908 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5910 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5912 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5913 rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5915 } else if (hpts_timeout) {
5917 * With respect to inp_flags2 here, lets let any new acks wake
5918 * us up here. Since we are not pacing (no pacing timer), output
5919 * can happen so we should let it. If its a Rack timer, then any inbound
5920 * packet probably won't change the sending (we will be blocked)
5921 * but it may change the prr stats so letting it in (the set defaults
5922 * at the start of this block) are good enough.
5924 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5926 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5927 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5929 /* No timer starting */
5931 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5932 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5933 tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5937 rack->rc_tmr_stopped = 0;
5939 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5943 * RACK Timer, here we simply do logging and house keeping.
5944 * the normal rack_output() function will call the
5945 * appropriate thing to check if we need to do a RACK retransmit.
5946 * We return 1, saying don't proceed with rack_output only
5947 * when all timers have been stopped (destroyed PCB?).
5950 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5953 * This timer simply provides an internal trigger to send out data.
5954 * The check_recovery_mode call will see if there are needed
5955 * retransmissions, if so we will enter fast-recovery. The output
5956 * call may or may not do the same thing depending on sysctl
5959 struct rack_sendmap *rsm;
5961 if (tp->t_timers->tt_flags & TT_STOPPED) {
5964 counter_u64_add(rack_to_tot, 1);
5965 if (rack->r_state && (rack->r_state != tp->t_state))
5966 rack_set_state(tp, rack);
5967 rack->rc_on_min_to = 0;
5968 rsm = rack_check_recovery_mode(tp, cts);
5969 rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5971 rack->r_ctl.rc_resend = rsm;
5972 rack->r_timer_override = 1;
5973 if (rack->use_rack_rr) {
5975 * Don't accumulate extra pacing delay
5976 * we are allowing the rack timer to
5977 * over-ride pacing i.e. rrr takes precedence
5978 * if the pacing interval is longer than the rrr
5979 * time (in other words we get the min pacing
5980 * time versus rrr pacing time).
5982 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5985 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5987 /* restart a timer and return 1 */
5988 rack_start_hpts_timer(rack, tp, cts,
5996 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5998 if (rsm->m->m_len > rsm->orig_m_len) {
6000 * Mbuf grew, caused by sbcompress, our offset does
6003 rsm->orig_m_len = rsm->m->m_len;
6004 } else if (rsm->m->m_len < rsm->orig_m_len) {
6006 * Mbuf shrank, trimmed off the top by an ack, our
6009 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6010 rsm->orig_m_len = rsm->m->m_len;
6015 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6020 if (src_rsm->orig_m_len != src_rsm->m->m_len) {
6021 /* Fix up the orig_m_len and possibly the mbuf offset */
6022 rack_adjust_orig_mlen(src_rsm);
6025 soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6026 while (soff >= m->m_len) {
6027 /* Move out past this mbuf */
6030 KASSERT((m != NULL),
6031 ("rsm:%p nrsm:%p hit at soff:%u null m",
6032 src_rsm, rsm, soff));
6036 rsm->orig_m_len = m->m_len;
6039 static __inline void
6040 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6041 struct rack_sendmap *rsm, uint32_t start)
6045 nrsm->r_start = start;
6046 nrsm->r_end = rsm->r_end;
6047 nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6048 nrsm->r_flags = rsm->r_flags;
6049 nrsm->r_dupack = rsm->r_dupack;
6050 nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6051 nrsm->r_rtr_bytes = 0;
6052 rsm->r_end = nrsm->r_start;
6053 nrsm->r_just_ret = rsm->r_just_ret;
6054 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6055 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6057 /* Now if we have SYN flag we keep it on the left edge */
6058 if (nrsm->r_flags & RACK_HAS_SYN)
6059 nrsm->r_flags &= ~RACK_HAS_SYN;
6060 /* Now if we have a FIN flag we keep it on the right edge */
6061 if (nrsm->r_flags & RACK_HAS_FIN)
6062 nrsm->r_flags &= ~RACK_HAS_FIN;
6064 * Now we need to find nrsm's new location in the mbuf chain
6065 * we basically calculate a new offset, which is soff +
6066 * how much is left in original rsm. Then we walk out the mbuf
6067 * chain to find the righ postion, it may be the same mbuf
6070 KASSERT(((rsm->m != NULL) ||
6071 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6072 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6074 rack_setup_offset_for_rsm(rsm, nrsm);
6077 static struct rack_sendmap *
6078 rack_merge_rsm(struct tcp_rack *rack,
6079 struct rack_sendmap *l_rsm,
6080 struct rack_sendmap *r_rsm)
6083 * We are merging two ack'd RSM's,
6084 * the l_rsm is on the left (lower seq
6085 * values) and the r_rsm is on the right
6086 * (higher seq value). The simplest way
6087 * to merge these is to move the right
6088 * one into the left. I don't think there
6089 * is any reason we need to try to find
6090 * the oldest (or last oldest retransmitted).
6092 struct rack_sendmap *rm;
6094 rack_log_map_chg(rack->rc_tp, rack, NULL,
6095 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6096 l_rsm->r_end = r_rsm->r_end;
6097 if (l_rsm->r_dupack < r_rsm->r_dupack)
6098 l_rsm->r_dupack = r_rsm->r_dupack;
6099 if (r_rsm->r_rtr_bytes)
6100 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6101 if (r_rsm->r_in_tmap) {
6102 /* This really should not happen */
6103 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6104 r_rsm->r_in_tmap = 0;
6108 if (r_rsm->r_flags & RACK_HAS_FIN)
6109 l_rsm->r_flags |= RACK_HAS_FIN;
6110 if (r_rsm->r_flags & RACK_TLP)
6111 l_rsm->r_flags |= RACK_TLP;
6112 if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6113 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6114 if ((r_rsm->r_flags & RACK_APP_LIMITED) &&
6115 ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6117 * If both are app-limited then let the
6118 * free lower the count. If right is app
6119 * limited and left is not, transfer.
6121 l_rsm->r_flags |= RACK_APP_LIMITED;
6122 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6123 if (r_rsm == rack->r_ctl.rc_first_appl)
6124 rack->r_ctl.rc_first_appl = l_rsm;
6126 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6129 panic("removing head in rack:%p rsm:%p rm:%p",
6133 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6134 /* Transfer the split limit to the map we free */
6135 r_rsm->r_limit_type = l_rsm->r_limit_type;
6136 l_rsm->r_limit_type = 0;
6138 rack_free(rack, r_rsm);
6143 * TLP Timer, here we simply setup what segment we want to
6144 * have the TLP expire on, the normal rack_output() will then
6147 * We return 1, saying don't proceed with rack_output only
6148 * when all timers have been stopped (destroyed PCB?).
6151 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6156 struct rack_sendmap *rsm = NULL;
6157 struct rack_sendmap *insret;
6160 uint32_t out, avail;
6161 int collapsed_win = 0;
6163 if (tp->t_timers->tt_flags & TT_STOPPED) {
6166 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6167 /* Its not time yet */
6170 if (ctf_progress_timeout_check(tp, true)) {
6171 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6172 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6176 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6177 * need to figure out how to force a full MSS segment out.
6179 rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6180 rack->r_ctl.retran_during_recovery = 0;
6181 rack->r_ctl.dsack_byte_cnt = 0;
6182 counter_u64_add(rack_tlp_tot, 1);
6183 if (rack->r_state && (rack->r_state != tp->t_state))
6184 rack_set_state(tp, rack);
6185 so = tp->t_inpcb->inp_socket;
6186 avail = sbavail(&so->so_snd);
6187 out = tp->snd_max - tp->snd_una;
6188 if (out > tp->snd_wnd) {
6189 /* special case, we need a retransmission */
6194 * Check our send oldest always settings, and if
6195 * there is an oldest to send jump to the need_retran.
6197 if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6201 /* New data is available */
6203 if (amm > ctf_fixed_maxseg(tp)) {
6204 amm = ctf_fixed_maxseg(tp);
6205 if ((amm + out) > tp->snd_wnd) {
6206 /* We are rwnd limited */
6209 } else if (amm < ctf_fixed_maxseg(tp)) {
6210 /* not enough to fill a MTU */
6213 if (IN_FASTRECOVERY(tp->t_flags)) {
6215 if (rack->rack_no_prr == 0) {
6216 if (out + amm <= tp->snd_wnd) {
6217 rack->r_ctl.rc_prr_sndcnt = amm;
6218 rack_log_to_prr(rack, 4, 0);
6223 /* Set the send-new override */
6224 if (out + amm <= tp->snd_wnd)
6225 rack->r_ctl.rc_tlp_new_data = amm;
6229 rack->r_ctl.rc_tlpsend = NULL;
6230 counter_u64_add(rack_tlp_newdata, 1);
6235 * Ok we need to arrange the last un-acked segment to be re-sent, or
6236 * optionally the first un-acked segment.
6238 if (collapsed_win == 0) {
6239 if (rack_always_send_oldest)
6240 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6242 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6243 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6244 rsm = rack_find_high_nonack(rack, rsm);
6248 counter_u64_add(rack_tlp_does_nada, 1);
6250 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6256 * We must find the last segment
6257 * that was acceptable by the client.
6259 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6260 if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6266 /* None? if so send the first */
6267 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6269 counter_u64_add(rack_tlp_does_nada, 1);
6271 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6277 if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6279 * We need to split this the last segment in two.
6281 struct rack_sendmap *nrsm;
6283 nrsm = rack_alloc_full_limit(rack);
6286 * No memory to split, we will just exit and punt
6287 * off to the RXT timer.
6289 counter_u64_add(rack_tlp_does_nada, 1);
6292 rack_clone_rsm(rack, nrsm, rsm,
6293 (rsm->r_end - ctf_fixed_maxseg(tp)));
6294 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6295 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6297 if (insret != NULL) {
6298 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6299 nrsm, insret, rack, rsm);
6302 if (rsm->r_in_tmap) {
6303 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6304 nrsm->r_in_tmap = 1;
6306 rsm->r_flags &= (~RACK_HAS_FIN);
6309 rack->r_ctl.rc_tlpsend = rsm;
6311 rack->r_timer_override = 1;
6312 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6315 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6320 * Delayed ack Timer, here we simply need to setup the
6321 * ACK_NOW flag and remove the DELACK flag. From there
6322 * the output routine will send the ack out.
6324 * We only return 1, saying don't proceed, if all timers
6325 * are stopped (destroyed PCB?).
6328 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6330 if (tp->t_timers->tt_flags & TT_STOPPED) {
6333 rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6334 tp->t_flags &= ~TF_DELACK;
6335 tp->t_flags |= TF_ACKNOW;
6336 KMOD_TCPSTAT_INC(tcps_delack);
6337 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6342 * Persists timer, here we simply send the
6343 * same thing as a keepalive will.
6344 * the one byte send.
6346 * We only return 1, saying don't proceed, if all timers
6347 * are stopped (destroyed PCB?).
6350 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6352 struct tcptemp *t_template;
6358 if (tp->t_timers->tt_flags & TT_STOPPED) {
6361 if (rack->rc_in_persist == 0)
6363 if (ctf_progress_timeout_check(tp, false)) {
6364 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6365 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6366 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6369 KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6371 * Persistence timer into zero window. Force a byte to be output, if
6374 KMOD_TCPSTAT_INC(tcps_persisttimeo);
6376 * Hack: if the peer is dead/unreachable, we do not time out if the
6377 * window is closed. After a full backoff, drop the connection if
6378 * the idle time (no responses to probes) reaches the maximum
6379 * backoff that we would use if retransmitting.
6381 if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6382 (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6383 TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6384 KMOD_TCPSTAT_INC(tcps_persistdrop);
6386 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6387 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6390 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6391 tp->snd_una == tp->snd_max)
6392 rack_exit_persist(tp, rack, cts);
6393 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6395 * If the user has closed the socket then drop a persisting
6396 * connection after a much reduced timeout.
6398 if (tp->t_state > TCPS_CLOSE_WAIT &&
6399 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6401 KMOD_TCPSTAT_INC(tcps_persistdrop);
6402 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6403 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6406 t_template = tcpip_maketemplate(rack->rc_inp);
6408 /* only set it if we were answered */
6409 if (rack->forced_ack == 0) {
6410 rack->forced_ack = 1;
6411 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6413 tcp_respond(tp, t_template->tt_ipgen,
6414 &t_template->tt_t, (struct mbuf *)NULL,
6415 tp->rcv_nxt, tp->snd_una - 1, 0);
6416 /* This sends an ack */
6417 if (tp->t_flags & TF_DELACK)
6418 tp->t_flags &= ~TF_DELACK;
6419 free(t_template, M_TEMP);
6421 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6424 rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6425 rack_start_hpts_timer(rack, tp, cts,
6431 * If a keepalive goes off, we had no other timers
6432 * happening. We always return 1 here since this
6433 * routine either drops the connection or sends
6434 * out a segment with respond.
6437 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6439 struct tcptemp *t_template;
6442 if (tp->t_timers->tt_flags & TT_STOPPED) {
6445 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6447 rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6449 * Keep-alive timer went off; send something or drop connection if
6450 * idle for too long.
6452 KMOD_TCPSTAT_INC(tcps_keeptimeo);
6453 if (tp->t_state < TCPS_ESTABLISHED)
6455 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6456 tp->t_state <= TCPS_CLOSING) {
6457 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6460 * Send a packet designed to force a response if the peer is
6461 * up and reachable: either an ACK if the connection is
6462 * still alive, or an RST if the peer has closed the
6463 * connection due to timeout or reboot. Using sequence
6464 * number tp->snd_una-1 causes the transmitted zero-length
6465 * segment to lie outside the receive window; by the
6466 * protocol spec, this requires the correspondent TCP to
6469 KMOD_TCPSTAT_INC(tcps_keepprobe);
6470 t_template = tcpip_maketemplate(inp);
6472 if (rack->forced_ack == 0) {
6473 rack->forced_ack = 1;
6474 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6476 tcp_respond(tp, t_template->tt_ipgen,
6477 &t_template->tt_t, (struct mbuf *)NULL,
6478 tp->rcv_nxt, tp->snd_una - 1, 0);
6479 free(t_template, M_TEMP);
6482 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6485 KMOD_TCPSTAT_INC(tcps_keepdrops);
6486 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6487 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6492 * Retransmit helper function, clear up all the ack
6493 * flags and take care of important book keeping.
6496 rack_remxt_tmr(struct tcpcb *tp)
6499 * The retransmit timer went off, all sack'd blocks must be
6502 struct rack_sendmap *rsm, *trsm = NULL;
6503 struct tcp_rack *rack;
6505 rack = (struct tcp_rack *)tp->t_fb_ptr;
6506 rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6507 rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6508 if (rack->r_state && (rack->r_state != tp->t_state))
6509 rack_set_state(tp, rack);
6511 * Ideally we would like to be able to
6512 * mark SACK-PASS on anything not acked here.
6514 * However, if we do that we would burst out
6515 * all that data 1ms apart. This would be unwise,
6516 * so for now we will just let the normal rxt timer
6517 * and tlp timer take care of it.
6519 * Also we really need to stick them back in sequence
6520 * order. This way we send in the proper order and any
6521 * sacks that come floating in will "re-ack" the data.
6522 * To do this we zap the tmap with an INIT and then
6523 * walk through and place every rsm in the RB tree
6524 * back in its seq ordered place.
6526 TAILQ_INIT(&rack->r_ctl.rc_tmap);
6527 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6529 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6530 /* We must re-add it back to the tlist */
6532 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6534 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6538 if (rsm->r_flags & RACK_ACKED)
6539 rsm->r_flags |= RACK_WAS_ACKED;
6540 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6542 /* Clear the count (we just un-acked them) */
6543 rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6544 rack->r_ctl.rc_sacked = 0;
6545 rack->r_ctl.rc_sacklast = NULL;
6546 rack->r_ctl.rc_agg_delayed = 0;
6548 rack->r_ctl.rc_agg_early = 0;
6550 /* Clear the tlp rtx mark */
6551 rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6552 if (rack->r_ctl.rc_resend != NULL)
6553 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6554 rack->r_ctl.rc_prr_sndcnt = 0;
6555 rack_log_to_prr(rack, 6, 0);
6556 rack->r_timer_override = 1;
6557 if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6558 #ifdef NETFLIX_EXP_DETECTION
6559 || (rack->sack_attack_disable != 0)
6561 ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6563 * For non-sack customers new data
6564 * needs to go out as retransmits until
6565 * we retransmit up to snd_max.
6567 rack->r_must_retran = 1;
6568 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6569 rack->r_ctl.rc_sacked);
6571 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6575 rack_convert_rtts(struct tcpcb *tp)
6577 if (tp->t_srtt > 1) {
6580 val = tp->t_srtt >> TCP_RTT_SHIFT;
6581 frac = tp->t_srtt & 0x1f;
6582 tp->t_srtt = TICKS_2_USEC(val);
6584 * frac is the fractional part of the srtt (if any)
6585 * but its in ticks and every bit represents
6590 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6592 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6600 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6601 frac = tp->t_rttvar & 0x1f;
6602 tp->t_rttvar = TICKS_2_USEC(val);
6604 * frac is the fractional part of the srtt (if any)
6605 * but its in ticks and every bit represents
6610 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6612 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6614 tp->t_rttvar += frac;
6617 tp->t_rxtcur = RACK_REXMTVAL(tp);
6618 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6619 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6621 if (tp->t_rxtcur > rack_rto_max) {
6622 tp->t_rxtcur = rack_rto_max;
6627 rack_cc_conn_init(struct tcpcb *tp)
6629 struct tcp_rack *rack;
6632 rack = (struct tcp_rack *)tp->t_fb_ptr;
6636 * Now convert to rack's internal format,
6639 if ((srtt == 0) && (tp->t_srtt != 0))
6640 rack_convert_rtts(tp);
6642 * We want a chance to stay in slowstart as
6643 * we create a connection. TCP spec says that
6644 * initially ssthresh is infinite. For our
6645 * purposes that is the snd_wnd.
6647 if (tp->snd_ssthresh < tp->snd_wnd) {
6648 tp->snd_ssthresh = tp->snd_wnd;
6651 * We also want to assure a IW worth of
6652 * data can get inflight.
6654 if (rc_init_window(rack) < tp->snd_cwnd)
6655 tp->snd_cwnd = rc_init_window(rack);
6659 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6660 * we will setup to retransmit the lowest seq number outstanding.
6663 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6671 if (tp->t_timers->tt_flags & TT_STOPPED) {
6674 if (ctf_progress_timeout_check(tp, false)) {
6675 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6676 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6677 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6680 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6681 rack->r_ctl.retran_during_recovery = 0;
6682 rack->r_ctl.dsack_byte_cnt = 0;
6683 if (IN_FASTRECOVERY(tp->t_flags))
6684 tp->t_flags |= TF_WASFRECOVERY;
6686 tp->t_flags &= ~TF_WASFRECOVERY;
6687 if (IN_CONGRECOVERY(tp->t_flags))
6688 tp->t_flags |= TF_WASCRECOVERY;
6690 tp->t_flags &= ~TF_WASCRECOVERY;
6691 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6692 (tp->snd_una == tp->snd_max)) {
6693 /* Nothing outstanding .. nothing to do */
6697 * Rack can only run one timer at a time, so we cannot
6698 * run a KEEPINIT (gating SYN sending) and a retransmit
6699 * timer for the SYN. So if we are in a front state and
6700 * have a KEEPINIT timer we need to check the first transmit
6701 * against now to see if we have exceeded the KEEPINIT time
6704 if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6705 (TP_KEEPINIT(tp) != 0)) {
6706 struct rack_sendmap *rsm;
6708 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6710 /* Ok we have something outstanding to test keepinit with */
6711 if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6712 ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6713 /* We have exceeded the KEEPINIT time */
6714 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6720 * Retransmission timer went off. Message has not been acked within
6721 * retransmit interval. Back off to a longer retransmit interval
6722 * and retransmit one segment.
6725 if ((rack->r_ctl.rc_resend == NULL) ||
6726 ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6728 * If the rwnd collapsed on
6729 * the one we are retransmitting
6730 * it does not count against the
6735 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6736 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6738 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6739 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6741 tcp_set_inp_to_drop(rack->rc_inp,
6742 (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6745 if (tp->t_state == TCPS_SYN_SENT) {
6747 * If the SYN was retransmitted, indicate CWND to be limited
6748 * to 1 segment in cc_conn_init().
6751 } else if (tp->t_rxtshift == 1) {
6753 * first retransmit; record ssthresh and cwnd so they can be
6754 * recovered if this turns out to be a "bad" retransmit. A
6755 * retransmit is considered "bad" if an ACK for this segment
6756 * is received within RTT/2 interval; the assumption here is
6757 * that the ACK was already in flight. See "On Estimating
6758 * End-to-End Network Path Properties" by Allman and Paxson
6761 tp->snd_cwnd_prev = tp->snd_cwnd;
6762 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6763 tp->snd_recover_prev = tp->snd_recover;
6764 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6765 tp->t_flags |= TF_PREVVALID;
6766 } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6767 tp->t_flags &= ~TF_PREVVALID;
6768 KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6769 if ((tp->t_state == TCPS_SYN_SENT) ||
6770 (tp->t_state == TCPS_SYN_RECEIVED))
6771 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6773 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6775 RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6776 max(rack_rto_min, rexmt), rack_rto_max);
6778 * We enter the path for PLMTUD if connection is established or, if
6779 * connection is FIN_WAIT_1 status, reason for the last is that if
6780 * amount of data we send is very small, we could send it in couple
6781 * of packets and process straight to FIN. In that case we won't
6782 * catch ESTABLISHED state.
6785 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6789 if (((V_tcp_pmtud_blackhole_detect == 1) ||
6790 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6791 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6792 ((tp->t_state == TCPS_ESTABLISHED) ||
6793 (tp->t_state == TCPS_FIN_WAIT_1))) {
6795 * Idea here is that at each stage of mtu probe (usually,
6796 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6797 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6798 * should take care of that.
6800 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6801 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6802 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6803 tp->t_rxtshift % 2 == 0)) {
6805 * Enter Path MTU Black-hole Detection mechanism: -
6806 * Disable Path MTU Discovery (IP "DF" bit). -
6807 * Reduce MTU to lower value than what we negotiated
6810 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6811 /* Record that we may have found a black hole. */
6812 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6813 /* Keep track of previous MSS. */
6814 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6818 * Reduce the MSS to blackhole value or to the
6819 * default in an attempt to retransmit.
6823 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6824 /* Use the sysctl tuneable blackhole MSS. */
6825 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6826 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6827 } else if (isipv6) {
6828 /* Use the default MSS. */
6829 tp->t_maxseg = V_tcp_v6mssdflt;
6831 * Disable Path MTU Discovery when we switch
6834 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6835 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6838 #if defined(INET6) && defined(INET)
6842 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6843 /* Use the sysctl tuneable blackhole MSS. */
6844 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6845 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6847 /* Use the default MSS. */
6848 tp->t_maxseg = V_tcp_mssdflt;
6850 * Disable Path MTU Discovery when we switch
6853 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6854 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6859 * If further retransmissions are still unsuccessful
6860 * with a lowered MTU, maybe this isn't a blackhole
6861 * and we restore the previous MSS and blackhole
6862 * detection flags. The limit '6' is determined by
6863 * giving each probe stage (1448, 1188, 524) 2
6864 * chances to recover.
6866 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6867 (tp->t_rxtshift >= 6)) {
6868 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6869 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6870 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6871 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6876 * Disable RFC1323 and SACK if we haven't got any response to
6877 * our third SYN to work-around some broken terminal servers
6878 * (most of which have hopefully been retired) that have bad VJ
6879 * header compression code which trashes TCP segments containing
6880 * unknown-to-them TCP options.
6882 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6883 (tp->t_rxtshift == 3))
6884 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6886 * If we backed off this far, our srtt estimate is probably bogus.
6887 * Clobber it so we'll take the next rtt measurement as our srtt;
6888 * move the current srtt into rttvar to keep the current retransmit
6891 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6893 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6894 in6_losing(tp->t_inpcb);
6897 in_losing(tp->t_inpcb);
6898 tp->t_rttvar += tp->t_srtt;
6901 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6902 tp->snd_recover = tp->snd_max;
6903 tp->t_flags |= TF_ACKNOW;
6905 rack_cong_signal(tp, CC_RTO, tp->snd_una);
6911 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
6914 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6919 if (tp->t_state == TCPS_LISTEN) {
6920 /* no timers on listen sockets */
6921 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6925 if ((timers & PACE_TMR_RACK) &&
6926 rack->rc_on_min_to) {
6928 * For the rack timer when we
6929 * are on a min-timeout (which means rrr_conf = 3)
6930 * we don't want to check the timer. It may
6931 * be going off for a pace and thats ok we
6932 * want to send the retransmit (if its ready).
6934 * If its on a normal rack timer (non-min) then
6935 * we will check if its expired.
6937 goto skip_time_check;
6939 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6942 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6944 rack_log_to_processing(rack, cts, ret, 0);
6947 if (hpts_calling == 0) {
6949 * A user send or queued mbuf (sack) has called us? We
6950 * return 0 and let the pacing guards
6951 * deal with it if they should or
6952 * should not cause a send.
6955 rack_log_to_processing(rack, cts, ret, 0);
6959 * Ok our timer went off early and we are not paced false
6960 * alarm, go back to sleep.
6963 left = rack->r_ctl.rc_timer_exp - cts;
6964 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6965 rack_log_to_processing(rack, cts, ret, left);
6969 rack->rc_tmr_stopped = 0;
6970 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6971 if (timers & PACE_TMR_DELACK) {
6972 ret = rack_timeout_delack(tp, rack, cts);
6973 } else if (timers & PACE_TMR_RACK) {
6974 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6975 rack->r_fast_output = 0;
6976 ret = rack_timeout_rack(tp, rack, cts);
6977 } else if (timers & PACE_TMR_TLP) {
6978 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6979 ret = rack_timeout_tlp(tp, rack, cts);
6980 } else if (timers & PACE_TMR_RXT) {
6981 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6982 rack->r_fast_output = 0;
6983 ret = rack_timeout_rxt(tp, rack, cts);
6984 } else if (timers & PACE_TMR_PERSIT) {
6985 ret = rack_timeout_persist(tp, rack, cts);
6986 } else if (timers & PACE_TMR_KEEP) {
6987 ret = rack_timeout_keepalive(tp, rack, cts);
6989 rack_log_to_processing(rack, cts, ret, timers);
6994 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6997 uint32_t us_cts, flags_on_entry;
6998 uint8_t hpts_removed = 0;
7000 flags_on_entry = rack->r_ctl.rc_hpts_flags;
7001 us_cts = tcp_get_usecs(&tv);
7002 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7003 ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7004 ((tp->snd_max - tp->snd_una) == 0))) {
7005 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7007 /* If we were not delayed cancel out the flag. */
7008 if ((tp->snd_max - tp->snd_una) == 0)
7009 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7010 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7012 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7013 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7014 if (rack->rc_inp->inp_in_hpts &&
7015 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7017 * Canceling timer's when we have no output being
7018 * paced. We also must remove ourselves from the
7021 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7024 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7026 if (hpts_removed == 0)
7027 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7031 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7037 rack_stopall(struct tcpcb *tp)
7039 struct tcp_rack *rack;
7040 rack = (struct tcp_rack *)tp->t_fb_ptr;
7041 rack->t_timers_stopped = 1;
7046 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7052 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7058 rack_stop_all_timers(struct tcpcb *tp)
7060 struct tcp_rack *rack;
7063 * Assure no timers are running.
7065 if (tcp_timer_active(tp, TT_PERSIST)) {
7066 /* We enter in persists, set the flag appropriately */
7067 rack = (struct tcp_rack *)tp->t_fb_ptr;
7068 rack->rc_in_persist = 1;
7070 tcp_timer_suspend(tp, TT_PERSIST);
7071 tcp_timer_suspend(tp, TT_REXMT);
7072 tcp_timer_suspend(tp, TT_KEEP);
7073 tcp_timer_suspend(tp, TT_DELACK);
7077 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7078 struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7081 uint16_t stripped_flags;
7084 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7086 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7087 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7088 rsm->r_flags |= RACK_OVERMAX;
7090 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7091 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7092 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7094 idx = rsm->r_rtr_cnt - 1;
7095 rsm->r_tim_lastsent[idx] = ts;
7096 stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7097 if (rsm->r_flags & RACK_ACKED) {
7098 /* Problably MTU discovery messing with us */
7099 rsm->r_flags &= ~RACK_ACKED;
7100 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7102 if (rsm->r_in_tmap) {
7103 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7106 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7108 if (rsm->r_flags & RACK_SACK_PASSED) {
7109 /* We have retransmitted due to the SACK pass */
7110 rsm->r_flags &= ~RACK_SACK_PASSED;
7111 rsm->r_flags |= RACK_WAS_SACKPASS;
7116 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7117 struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7120 * We (re-)transmitted starting at rsm->r_start for some length
7121 * (possibly less than r_end.
7123 struct rack_sendmap *nrsm, *insret;
7128 c_end = rsm->r_start + len;
7129 if (SEQ_GEQ(c_end, rsm->r_end)) {
7131 * We retransmitted the whole piece or more than the whole
7132 * slopping into the next rsm.
7134 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7135 if (c_end == rsm->r_end) {
7141 /* Hangs over the end return whats left */
7142 act_len = rsm->r_end - rsm->r_start;
7143 *lenp = (len - act_len);
7144 return (rsm->r_end);
7146 /* We don't get out of this block. */
7149 * Here we retransmitted less than the whole thing which means we
7150 * have to split this into what was transmitted and what was not.
7152 nrsm = rack_alloc_full_limit(rack);
7155 * We can't get memory, so lets not proceed.
7161 * So here we are going to take the original rsm and make it what we
7162 * retransmitted. nrsm will be the tail portion we did not
7163 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7164 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7165 * 1, 6 and the new piece will be 6, 11.
7167 rack_clone_rsm(rack, nrsm, rsm, c_end);
7169 rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7170 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7172 if (insret != NULL) {
7173 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7174 nrsm, insret, rack, rsm);
7177 if (rsm->r_in_tmap) {
7178 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7179 nrsm->r_in_tmap = 1;
7181 rsm->r_flags &= (~RACK_HAS_FIN);
7182 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7183 /* Log a split of rsm into rsm and nrsm */
7184 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7190 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7191 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7192 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff)
7194 struct tcp_rack *rack;
7195 struct rack_sendmap *rsm, *nrsm, *insret, fe;
7196 register uint32_t snd_max, snd_una;
7199 * Add to the RACK log of packets in flight or retransmitted. If
7200 * there is a TS option we will use the TS echoed, if not we will
7203 * Retransmissions will increment the count and move the ts to its
7204 * proper place. Note that if options do not include TS's then we
7205 * won't be able to effectively use the ACK for an RTT on a retran.
7207 * Notes about r_start and r_end. Lets consider a send starting at
7208 * sequence 1 for 10 bytes. In such an example the r_start would be
7209 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7210 * This means that r_end is actually the first sequence for the next
7215 * If err is set what do we do XXXrrs? should we not add the thing?
7216 * -- i.e. return if err != 0 or should we pretend we sent it? --
7217 * i.e. proceed with add ** do this for now.
7219 INP_WLOCK_ASSERT(tp->t_inpcb);
7222 * We don't log errors -- we could but snd_max does not
7223 * advance in this case either.
7227 if (th_flags & TH_RST) {
7229 * We don't log resets and we return immediately from
7234 rack = (struct tcp_rack *)tp->t_fb_ptr;
7235 snd_una = tp->snd_una;
7236 snd_max = tp->snd_max;
7237 if (th_flags & (TH_SYN | TH_FIN)) {
7239 * The call to rack_log_output is made before bumping
7240 * snd_max. This means we can record one extra byte on a SYN
7241 * or FIN if seq_out is adding more on and a FIN is present
7242 * (and we are not resending).
7244 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7246 if (th_flags & TH_FIN)
7248 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7250 * The add/update as not been done for the FIN/SYN
7253 snd_max = tp->snd_nxt;
7256 if (SEQ_LEQ((seq_out + len), snd_una)) {
7257 /* Are sending an old segment to induce an ack (keep-alive)? */
7260 if (SEQ_LT(seq_out, snd_una)) {
7261 /* huh? should we panic? */
7264 end = seq_out + len;
7266 if (SEQ_GEQ(end, seq_out))
7267 len = end - seq_out;
7272 /* We don't log zero window probes */
7275 rack->r_ctl.rc_time_last_sent = cts;
7276 if (IN_FASTRECOVERY(tp->t_flags)) {
7277 rack->r_ctl.rc_prr_out += len;
7279 /* First question is it a retransmission or new? */
7280 if (seq_out == snd_max) {
7283 rsm = rack_alloc(rack);
7286 * Hmm out of memory and the tcb got destroyed while
7291 if (th_flags & TH_FIN) {
7292 rsm->r_flags = RACK_HAS_FIN|add_flag;
7294 rsm->r_flags = add_flag;
7296 rsm->r_tim_lastsent[0] = cts;
7298 rsm->r_rtr_bytes = 0;
7299 if (th_flags & TH_SYN) {
7300 /* The data space is one beyond snd_una */
7301 rsm->r_flags |= RACK_HAS_SYN;
7303 rsm->r_start = seq_out;
7304 rsm->r_end = rsm->r_start + len;
7307 * save off the mbuf location that
7308 * sndmbuf_noadv returned (which is
7309 * where we started copying from)..
7313 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7315 if (rsm->m->m_len <= rsm->soff) {
7317 * XXXrrs Question, will this happen?
7319 * If sbsndptr is set at the correct place
7320 * then s_moff should always be somewhere
7321 * within rsm->m. But if the sbsndptr was
7322 * off then that won't be true. If it occurs
7323 * we need to walkout to the correct location.
7328 while (lm->m_len <= rsm->soff) {
7329 rsm->soff -= lm->m_len;
7331 KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7332 __func__, rack, s_moff, s_mb, rsm->soff));
7335 counter_u64_add(rack_sbsndptr_wrong, 1);
7337 counter_u64_add(rack_sbsndptr_right, 1);
7338 rsm->orig_m_len = rsm->m->m_len;
7340 rsm->orig_m_len = 0;
7341 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7343 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7344 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7346 if (insret != NULL) {
7347 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7348 nrsm, insret, rack, rsm);
7351 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7354 * Special case detection, is there just a single
7355 * packet outstanding when we are not in recovery?
7357 * If this is true mark it so.
7359 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7360 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7361 struct rack_sendmap *prsm;
7363 prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7365 prsm->r_one_out_nr = 1;
7370 * If we reach here its a retransmission and we need to find it.
7372 memset(&fe, 0, sizeof(fe));
7374 if (hintrsm && (hintrsm->r_start == seq_out)) {
7378 /* No hints sorry */
7381 if ((rsm) && (rsm->r_start == seq_out)) {
7382 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7389 /* Ok it was not the last pointer go through it the hard way. */
7391 fe.r_start = seq_out;
7392 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7394 if (rsm->r_start == seq_out) {
7395 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7402 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7403 /* Transmitted within this piece */
7405 * Ok we must split off the front and then let the
7406 * update do the rest
7408 nrsm = rack_alloc_full_limit(rack);
7410 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7414 * copy rsm to nrsm and then trim the front of rsm
7415 * to not include this part.
7417 rack_clone_rsm(rack, nrsm, rsm, seq_out);
7418 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7419 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7421 if (insret != NULL) {
7422 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7423 nrsm, insret, rack, rsm);
7426 if (rsm->r_in_tmap) {
7427 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7428 nrsm->r_in_tmap = 1;
7430 rsm->r_flags &= (~RACK_HAS_FIN);
7431 seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7439 * Hmm not found in map did they retransmit both old and on into the
7442 if (seq_out == tp->snd_max) {
7444 } else if (SEQ_LT(seq_out, tp->snd_max)) {
7446 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7447 seq_out, len, tp->snd_una, tp->snd_max);
7448 printf("Starting Dump of all rack entries\n");
7449 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7450 printf("rsm:%p start:%u end:%u\n",
7451 rsm, rsm->r_start, rsm->r_end);
7453 printf("Dump complete\n");
7454 panic("seq_out not found rack:%p tp:%p",
7460 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7463 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7464 seq_out, len, tp->snd_max, tp);
7470 * Record one of the RTT updates from an ack into
7471 * our sample structure.
7475 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7476 int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7478 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7479 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7480 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7482 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7483 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7484 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7486 if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7487 if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7488 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7489 if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7490 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7492 if ((confidence == 1) &&
7494 (rsm->r_just_ret) ||
7495 (rsm->r_one_out_nr &&
7496 len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7498 * If the rsm had a just return
7499 * hit it then we can't trust the
7500 * rtt measurement for buffer deterimination
7501 * Note that a confidence of 2, indicates
7502 * SACK'd which overrides the r_just_ret or
7503 * the r_one_out_nr. If it was a CUM-ACK and
7504 * we had only two outstanding, but get an
7505 * ack for only 1. Then that also lowers our
7510 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7511 (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7512 if (rack->r_ctl.rack_rs.confidence == 0) {
7514 * We take anything with no current confidence
7517 rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7518 rack->r_ctl.rack_rs.confidence = confidence;
7519 rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7520 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7522 * Once we have a confident number,
7523 * we can update it with a smaller
7524 * value since this confident number
7525 * may include the DSACK time until
7526 * the next segment (the second one) arrived.
7528 rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7529 rack->r_ctl.rack_rs.confidence = confidence;
7530 rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7533 rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7534 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7535 rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7536 rack->r_ctl.rack_rs.rs_rtt_cnt++;
7540 * Collect new round-trip time estimate
7541 * and update averages and current timeout.
7544 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7547 uint32_t o_srtt, o_var;
7548 int32_t hrtt_up = 0;
7551 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7552 /* No valid sample */
7554 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7555 /* We are to use the lowest RTT seen in a single ack */
7556 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7557 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7558 /* We are to use the highest RTT seen in a single ack */
7559 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7560 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7561 /* We are to use the average RTT seen in a single ack */
7562 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7563 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7566 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7572 if (rack->rc_gp_rtt_set == 0) {
7574 * With no RTT we have to accept
7575 * even one we are not confident of.
7577 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7578 rack->rc_gp_rtt_set = 1;
7579 } else if (rack->r_ctl.rack_rs.confidence) {
7580 /* update the running gp srtt */
7581 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7582 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7584 if (rack->r_ctl.rack_rs.confidence) {
7586 * record the low and high for highly buffered path computation,
7587 * we only do this if we are confident (not a retransmission).
7589 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7590 rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7593 if (rack->rc_highly_buffered == 0) {
7595 * Currently once we declare a path has
7596 * highly buffered there is no going
7597 * back, which may be a problem...
7599 if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7600 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7601 rack->r_ctl.rc_highest_us_rtt,
7602 rack->r_ctl.rc_lowest_us_rtt,
7604 rack->rc_highly_buffered = 1;
7608 if ((rack->r_ctl.rack_rs.confidence) ||
7609 (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7611 * If we are highly confident of it <or> it was
7612 * never retransmitted we accept it as the last us_rtt.
7614 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7615 /* The lowest rtt can be set if its was not retransmited */
7616 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7617 rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7618 if (rack->r_ctl.rc_lowest_us_rtt == 0)
7619 rack->r_ctl.rc_lowest_us_rtt = 1;
7622 o_srtt = tp->t_srtt;
7623 o_var = tp->t_rttvar;
7624 rack = (struct tcp_rack *)tp->t_fb_ptr;
7625 if (tp->t_srtt != 0) {
7627 * We keep a simple srtt in microseconds, like our rtt
7628 * measurement. We don't need to do any tricks with shifting
7629 * etc. Instead we just add in 1/8th of the new measurement
7630 * and subtract out 1/8 of the old srtt. We do the same with
7631 * the variance after finding the absolute value of the
7632 * difference between this sample and the current srtt.
7634 delta = tp->t_srtt - rtt;
7635 /* Take off 1/8th of the current sRTT */
7636 tp->t_srtt -= (tp->t_srtt >> 3);
7637 /* Add in 1/8th of the new RTT just measured */
7638 tp->t_srtt += (rtt >> 3);
7639 if (tp->t_srtt <= 0)
7641 /* Now lets make the absolute value of the variance */
7644 /* Subtract out 1/8th */
7645 tp->t_rttvar -= (tp->t_rttvar >> 3);
7646 /* Add in 1/8th of the new variance we just saw */
7647 tp->t_rttvar += (delta >> 3);
7648 if (tp->t_rttvar <= 0)
7650 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7651 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7654 * No rtt measurement yet - use the unsmoothed rtt. Set the
7655 * variance to half the rtt (so our first retransmit happens
7659 tp->t_rttvar = rtt >> 1;
7660 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7662 rack->rc_srtt_measure_made = 1;
7663 KMOD_TCPSTAT_INC(tcps_rttupdated);
7666 if (rack_stats_gets_ms_rtt == 0) {
7667 /* Send in the microsecond rtt used for rxt timeout purposes */
7668 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7669 } else if (rack_stats_gets_ms_rtt == 1) {
7670 /* Send in the millisecond rtt used for rxt timeout purposes */
7674 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7675 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7676 } else if (rack_stats_gets_ms_rtt == 2) {
7677 /* Send in the millisecond rtt has close to the path RTT as we can get */
7681 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7682 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7684 /* Send in the microsecond rtt has close to the path RTT as we can get */
7685 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7690 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7691 * way we do the smoothing, srtt and rttvar will each average +1/2
7692 * tick of bias. When we compute the retransmit timer, we want 1/2
7693 * tick of rounding and 1 extra tick because of +-1/2 tick
7694 * uncertainty in the firing of the timer. The bias will give us
7695 * exactly the 1.5 tick we need. But, because the bias is
7696 * statistical, we have to test that we don't drop below the minimum
7697 * feasible timer (which is 2 ticks).
7700 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7701 max(rack_rto_min, rtt + 2), rack_rto_max);
7702 rack_log_rtt_sample(rack, rtt);
7703 tp->t_softerror = 0;
7708 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7711 * Apply to filter the inbound us-rtt at us_cts.
7715 old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7716 apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7718 if (rack->r_ctl.last_pacing_time &&
7719 rack->rc_gp_dyn_mul &&
7720 (rack->r_ctl.last_pacing_time > us_rtt))
7721 rack->pacing_longer_than_rtt = 1;
7723 rack->pacing_longer_than_rtt = 0;
7724 if (old_rtt > us_rtt) {
7725 /* We just hit a new lower rtt time */
7726 rack_log_rtt_shrinks(rack, us_cts, old_rtt,
7727 __LINE__, RACK_RTTS_NEWRTT);
7729 * Only count it if its lower than what we saw within our
7732 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7733 if (rack_probertt_lower_within &&
7734 rack->rc_gp_dyn_mul &&
7735 (rack->use_fixed_rate == 0) &&
7736 (rack->rc_always_pace)) {
7738 * We are seeing a new lower rtt very close
7739 * to the time that we would have entered probe-rtt.
7740 * This is probably due to the fact that a peer flow
7741 * has entered probe-rtt. Lets go in now too.
7745 val = rack_probertt_lower_within * rack_time_between_probertt;
7747 if ((rack->in_probe_rtt == 0) &&
7748 ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7749 rack_enter_probertt(rack, us_cts);
7752 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7758 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7759 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7762 uint32_t t, len_acked;
7764 if ((rsm->r_flags & RACK_ACKED) ||
7765 (rsm->r_flags & RACK_WAS_ACKED))
7768 if (rsm->r_no_rtt_allowed) {
7772 if (ack_type == CUM_ACKED) {
7773 if (SEQ_GT(th_ack, rsm->r_end)) {
7774 len_acked = rsm->r_end - rsm->r_start;
7777 len_acked = th_ack - rsm->r_start;
7781 len_acked = rsm->r_end - rsm->r_start;
7784 if (rsm->r_rtr_cnt == 1) {
7787 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7790 if (!tp->t_rttlow || tp->t_rttlow > t)
7792 if (!rack->r_ctl.rc_rack_min_rtt ||
7793 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7794 rack->r_ctl.rc_rack_min_rtt = t;
7795 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7796 rack->r_ctl.rc_rack_min_rtt = 1;
7799 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7800 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7802 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7805 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7806 if (ack_type == SACKED) {
7807 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7808 tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7811 * We need to setup what our confidence
7814 * If the rsm was app limited and it is
7815 * less than a mss in length (the end
7816 * of the send) then we have a gap. If we
7817 * were app limited but say we were sending
7818 * multiple MSS's then we are more confident
7821 * When we are not app-limited then we see if
7822 * the rsm is being included in the current
7823 * measurement, we tell this by the app_limited_needs_set
7826 * Note that being cwnd blocked is not applimited
7827 * as well as the pacing delay between packets which
7828 * are sending only 1 or 2 MSS's also will show up
7829 * in the RTT. We probably need to examine this algorithm
7830 * a bit more and enhance it to account for the delay
7831 * between rsm's. We could do that by saving off the
7832 * pacing delay of each rsm (in an rsm) and then
7833 * factoring that in somehow though for now I am
7838 if (rsm->r_flags & RACK_APP_LIMITED) {
7839 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7843 } else if (rack->app_limited_needs_set == 0) {
7848 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7849 tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7850 calc_conf, rsm, rsm->r_rtr_cnt);
7852 if ((rsm->r_flags & RACK_TLP) &&
7853 (!IN_FASTRECOVERY(tp->t_flags))) {
7854 /* Segment was a TLP and our retrans matched */
7855 if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7856 rack->r_ctl.rc_rsm_start = tp->snd_max;
7857 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7858 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7859 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7862 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7863 /* New more recent rack_tmit_time */
7864 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7865 rack->rc_rack_rtt = t;
7870 * We clear the soft/rxtshift since we got an ack.
7871 * There is no assurance we will call the commit() function
7872 * so we need to clear these to avoid incorrect handling.
7875 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7876 rack_rto_min, rack_rto_max);
7877 tp->t_softerror = 0;
7878 if (to && (to->to_flags & TOF_TS) &&
7879 (ack_type == CUM_ACKED) &&
7881 ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7883 * Now which timestamp does it match? In this block the ACK
7884 * must be coming from a previous transmission.
7886 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7887 if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7888 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7891 if ((i + 1) < rsm->r_rtr_cnt) {
7893 * The peer ack'd from our previous
7894 * transmission. We have a spurious
7895 * retransmission and thus we dont
7896 * want to update our rack_rtt.
7900 if (!tp->t_rttlow || tp->t_rttlow > t)
7902 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7903 rack->r_ctl.rc_rack_min_rtt = t;
7904 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7905 rack->r_ctl.rc_rack_min_rtt = 1;
7908 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7909 (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7910 /* New more recent rack_tmit_time */
7911 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7912 rack->rc_rack_rtt = t;
7914 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7915 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7923 * Ok its a SACK block that we retransmitted. or a windows
7924 * machine without timestamps. We can tell nothing from the
7925 * time-stamp since its not there or the time the peer last
7926 * recieved a segment that moved forward its cum-ack point.
7929 i = rsm->r_rtr_cnt - 1;
7930 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7933 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7935 * We retransmitted and the ack came back in less
7936 * than the smallest rtt we have observed. We most
7937 * likely did an improper retransmit as outlined in
7938 * 6.2 Step 2 point 2 in the rack-draft so we
7939 * don't want to update our rack_rtt. We in
7940 * theory (in future) might want to think about reverting our
7941 * cwnd state but we won't for now.
7944 } else if (rack->r_ctl.rc_rack_min_rtt) {
7946 * We retransmitted it and the retransmit did the
7949 if (!rack->r_ctl.rc_rack_min_rtt ||
7950 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7951 rack->r_ctl.rc_rack_min_rtt = t;
7952 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7953 rack->r_ctl.rc_rack_min_rtt = 1;
7956 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7957 /* New more recent rack_tmit_time */
7958 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7959 rack->rc_rack_rtt = t;
7968 * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7971 rack_log_sack_passed(struct tcpcb *tp,
7972 struct tcp_rack *rack, struct rack_sendmap *rsm)
7974 struct rack_sendmap *nrsm;
7977 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7978 rack_head, r_tnext) {
7980 /* Skip orginal segment he is acked */
7983 if (nrsm->r_flags & RACK_ACKED) {
7985 * Skip ack'd segments, though we
7986 * should not see these, since tmap
7987 * should not have ack'd segments.
7991 if (nrsm->r_flags & RACK_SACK_PASSED) {
7993 * We found one that is already marked
7994 * passed, we have been here before and
7995 * so all others below this are marked.
7999 nrsm->r_flags |= RACK_SACK_PASSED;
8000 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8005 rack_need_set_test(struct tcpcb *tp,
8006 struct tcp_rack *rack,
8007 struct rack_sendmap *rsm,
8013 if ((tp->t_flags & TF_GPUTINPROG) &&
8014 SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8016 * We were app limited, and this ack
8017 * butts up or goes beyond the point where we want
8018 * to start our next measurement. We need
8019 * to record the new gput_ts as here and
8020 * possibly update the start sequence.
8024 if (rsm->r_rtr_cnt > 1) {
8026 * This is a retransmit, can we
8027 * really make any assessment at this
8028 * point? We are not really sure of
8029 * the timestamp, is it this or the
8030 * previous transmission?
8032 * Lets wait for something better that
8033 * is not retransmitted.
8039 rack->app_limited_needs_set = 0;
8040 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8041 /* Do we start at a new end? */
8042 if ((use_which == RACK_USE_BEG) &&
8043 SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8045 * When we get an ACK that just eats
8046 * up some of the rsm, we set RACK_USE_BEG
8047 * since whats at r_start (i.e. th_ack)
8048 * is left unacked and thats where the
8049 * measurement not starts.
8051 tp->gput_seq = rsm->r_start;
8052 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8054 if ((use_which == RACK_USE_END) &&
8055 SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8057 * We use the end when the cumack
8058 * is moving forward and completely
8059 * deleting the rsm passed so basically
8060 * r_end holds th_ack.
8062 * For SACK's we also want to use the end
8063 * since this piece just got sacked and
8064 * we want to target anything after that
8065 * in our measurement.
8067 tp->gput_seq = rsm->r_end;
8068 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8070 if (use_which == RACK_USE_END_OR_THACK) {
8072 * special case for ack moving forward,
8073 * not a sack, we need to move all the
8074 * way up to where this ack cum-ack moves
8077 if (SEQ_GT(th_ack, rsm->r_end))
8078 tp->gput_seq = th_ack;
8080 tp->gput_seq = rsm->r_end;
8081 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8083 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8085 * We moved beyond this guy's range, re-calculate
8086 * the new end point.
8088 if (rack->rc_gp_filled == 0) {
8089 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8091 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8095 * We are moving the goal post, we may be able to clear the
8096 * measure_saw_probe_rtt flag.
8098 if ((rack->in_probe_rtt == 0) &&
8099 (rack->measure_saw_probe_rtt) &&
8100 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8101 rack->measure_saw_probe_rtt = 0;
8102 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8103 seq, tp->gput_seq, 0, 5, line, NULL);
8104 if (rack->rc_gp_filled &&
8105 ((tp->gput_ack - tp->gput_seq) <
8106 max(rc_init_window(rack), (MIN_GP_WIN *
8107 ctf_fixed_maxseg(tp))))) {
8108 uint32_t ideal_amount;
8110 ideal_amount = rack_get_measure_window(tp, rack);
8111 if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8113 * There is no sense of continuing this measurement
8114 * because its too small to gain us anything we
8115 * trust. Skip it and that way we can start a new
8116 * measurement quicker.
8118 tp->t_flags &= ~TF_GPUTINPROG;
8119 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8120 0, 0, 0, 6, __LINE__, NULL);
8123 * Reset the window further out.
8125 tp->gput_ack = tp->gput_seq + ideal_amount;
8132 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8133 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8135 uint32_t start, end, changed = 0;
8136 struct rack_sendmap stack_map;
8137 struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8138 int32_t used_ref = 1;
8141 start = sack->start;
8144 memset(&fe, 0, sizeof(fe));
8146 if ((rsm == NULL) ||
8147 (SEQ_LT(end, rsm->r_start)) ||
8148 (SEQ_GEQ(start, rsm->r_end)) ||
8149 (SEQ_LT(start, rsm->r_start))) {
8151 * We are not in the right spot,
8152 * find the correct spot in the tree.
8156 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8163 /* Ok we have an ACK for some piece of this rsm */
8164 if (rsm->r_start != start) {
8165 if ((rsm->r_flags & RACK_ACKED) == 0) {
8167 * Need to split this in two pieces the before and after,
8168 * the before remains in the map, the after must be
8169 * added. In other words we have:
8170 * rsm |--------------|
8174 * and nrsm will be the sacked piece
8177 * But before we start down that path lets
8178 * see if the sack spans over on top of
8179 * the next guy and it is already sacked.
8181 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8182 if (next && (next->r_flags & RACK_ACKED) &&
8183 SEQ_GEQ(end, next->r_start)) {
8185 * So the next one is already acked, and
8186 * we can thus by hookery use our stack_map
8187 * to reflect the piece being sacked and
8188 * then adjust the two tree entries moving
8189 * the start and ends around. So we start like:
8190 * rsm |------------| (not-acked)
8191 * next |-----------| (acked)
8192 * sackblk |-------->
8193 * We want to end like so:
8194 * rsm |------| (not-acked)
8195 * next |-----------------| (acked)
8197 * Where nrsm is a temporary stack piece we
8198 * use to update all the gizmos.
8200 /* Copy up our fudge block */
8202 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8203 /* Now adjust our tree blocks */
8205 next->r_start = start;
8206 /* Now we must adjust back where next->m is */
8207 rack_setup_offset_for_rsm(rsm, next);
8209 /* We don't need to adjust rsm, it did not change */
8210 /* Clear out the dup ack count of the remainder */
8212 rsm->r_just_ret = 0;
8213 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8214 /* Now lets make sure our fudge block is right */
8215 nrsm->r_start = start;
8216 /* Now lets update all the stats and such */
8217 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8218 if (rack->app_limited_needs_set)
8219 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8220 changed += (nrsm->r_end - nrsm->r_start);
8221 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8222 if (nrsm->r_flags & RACK_SACK_PASSED) {
8223 counter_u64_add(rack_reorder_seen, 1);
8224 rack->r_ctl.rc_reorder_ts = cts;
8227 * Now we want to go up from rsm (the
8228 * one left un-acked) to the next one
8229 * in the tmap. We do this so when
8230 * we walk backwards we include marking
8231 * sack-passed on rsm (The one passed in
8232 * is skipped since it is generally called
8233 * on something sacked before removing it
8236 if (rsm->r_in_tmap) {
8237 nrsm = TAILQ_NEXT(rsm, r_tnext);
8239 * Now that we have the next
8240 * one walk backwards from there.
8242 if (nrsm && nrsm->r_in_tmap)
8243 rack_log_sack_passed(tp, rack, nrsm);
8245 /* Now are we done? */
8246 if (SEQ_LT(end, next->r_end) ||
8247 (end == next->r_end)) {
8248 /* Done with block */
8251 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8252 counter_u64_add(rack_sack_used_next_merge, 1);
8253 /* Postion for the next block */
8254 start = next->r_end;
8255 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8260 * We can't use any hookery here, so we
8261 * need to split the map. We enter like
8265 * We will add the new block nrsm and
8266 * that will be the new portion, and then
8267 * fall through after reseting rsm. So we
8268 * split and look like this:
8272 * We then fall through reseting
8273 * rsm to nrsm, so the next block
8276 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8279 * failed XXXrrs what can we do but loose the sack
8284 counter_u64_add(rack_sack_splits, 1);
8285 rack_clone_rsm(rack, nrsm, rsm, start);
8286 rsm->r_just_ret = 0;
8287 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8289 if (insret != NULL) {
8290 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8291 nrsm, insret, rack, rsm);
8294 if (rsm->r_in_tmap) {
8295 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8296 nrsm->r_in_tmap = 1;
8298 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8299 rsm->r_flags &= (~RACK_HAS_FIN);
8300 /* Position us to point to the new nrsm that starts the sack blk */
8304 /* Already sacked this piece */
8305 counter_u64_add(rack_sack_skipped_acked, 1);
8307 if (end == rsm->r_end) {
8308 /* Done with block */
8309 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8311 } else if (SEQ_LT(end, rsm->r_end)) {
8312 /* A partial sack to a already sacked block */
8314 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8318 * The end goes beyond this guy
8319 * repostion the start to the
8323 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8329 if (SEQ_GEQ(end, rsm->r_end)) {
8331 * The end of this block is either beyond this guy or right
8332 * at this guy. I.e.:
8338 if ((rsm->r_flags & RACK_ACKED) == 0) {
8339 rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8340 changed += (rsm->r_end - rsm->r_start);
8341 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8342 if (rsm->r_in_tmap) /* should be true */
8343 rack_log_sack_passed(tp, rack, rsm);
8344 /* Is Reordering occuring? */
8345 if (rsm->r_flags & RACK_SACK_PASSED) {
8346 rsm->r_flags &= ~RACK_SACK_PASSED;
8347 counter_u64_add(rack_reorder_seen, 1);
8348 rack->r_ctl.rc_reorder_ts = cts;
8350 if (rack->app_limited_needs_set)
8351 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8352 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8353 rsm->r_flags |= RACK_ACKED;
8354 rsm->r_flags &= ~RACK_TLP;
8355 if (rsm->r_in_tmap) {
8356 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8359 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8361 counter_u64_add(rack_sack_skipped_acked, 1);
8364 if (end == rsm->r_end) {
8365 /* This block only - done, setup for next */
8369 * There is more not coverend by this rsm move on
8370 * to the next block in the RB tree.
8372 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8380 * The end of this sack block is smaller than
8385 if ((rsm->r_flags & RACK_ACKED) == 0) {
8386 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8387 if (prev && (prev->r_flags & RACK_ACKED)) {
8389 * Goal, we want the right remainder of rsm to shrink
8390 * in place and span from (rsm->r_start = end) to rsm->r_end.
8391 * We want to expand prev to go all the way
8392 * to prev->r_end <- end.
8393 * so in the tree we have before:
8394 * prev |--------| (acked)
8395 * rsm |-------| (non-acked)
8397 * We churn it so we end up with
8398 * prev |----------| (acked)
8399 * rsm |-----| (non-acked)
8400 * nrsm |-| (temporary)
8403 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8406 /* Now adjust nrsm (stack copy) to be
8407 * the one that is the small
8408 * piece that was "sacked".
8412 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8414 * Now that the rsm has had its start moved forward
8415 * lets go ahead and get its new place in the world.
8417 rack_setup_offset_for_rsm(prev, rsm);
8419 * Now nrsm is our new little piece
8420 * that is acked (which was merged
8421 * to prev). Update the rtt and changed
8422 * based on that. Also check for reordering.
8424 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8425 if (rack->app_limited_needs_set)
8426 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8427 changed += (nrsm->r_end - nrsm->r_start);
8428 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8429 if (nrsm->r_flags & RACK_SACK_PASSED) {
8430 counter_u64_add(rack_reorder_seen, 1);
8431 rack->r_ctl.rc_reorder_ts = cts;
8433 rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8435 counter_u64_add(rack_sack_used_prev_merge, 1);
8438 * This is the case where our previous
8439 * block is not acked either, so we must
8440 * split the block in two.
8442 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8444 /* failed rrs what can we do but loose the sack info? */
8448 * In this case nrsm becomes
8449 * nrsm->r_start = end;
8450 * nrsm->r_end = rsm->r_end;
8451 * which is un-acked.
8453 * rsm->r_end = nrsm->r_start;
8454 * i.e. the remaining un-acked
8455 * piece is left on the left
8458 * So we start like this
8459 * rsm |----------| (not acked)
8461 * build it so we have
8463 * nrsm |------| (not acked)
8465 counter_u64_add(rack_sack_splits, 1);
8466 rack_clone_rsm(rack, nrsm, rsm, end);
8467 rsm->r_flags &= (~RACK_HAS_FIN);
8468 rsm->r_just_ret = 0;
8469 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8471 if (insret != NULL) {
8472 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8473 nrsm, insret, rack, rsm);
8476 if (rsm->r_in_tmap) {
8477 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8478 nrsm->r_in_tmap = 1;
8481 rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8482 rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8483 changed += (rsm->r_end - rsm->r_start);
8484 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8485 if (rsm->r_in_tmap) /* should be true */
8486 rack_log_sack_passed(tp, rack, rsm);
8487 /* Is Reordering occuring? */
8488 if (rsm->r_flags & RACK_SACK_PASSED) {
8489 rsm->r_flags &= ~RACK_SACK_PASSED;
8490 counter_u64_add(rack_reorder_seen, 1);
8491 rack->r_ctl.rc_reorder_ts = cts;
8493 if (rack->app_limited_needs_set)
8494 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8495 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8496 rsm->r_flags |= RACK_ACKED;
8497 rsm->r_flags &= ~RACK_TLP;
8498 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8499 if (rsm->r_in_tmap) {
8500 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8504 } else if (start != end){
8506 * The block was already acked.
8508 counter_u64_add(rack_sack_skipped_acked, 1);
8512 if (rsm && (rsm->r_flags & RACK_ACKED)) {
8514 * Now can we merge where we worked
8515 * with either the previous or
8518 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8520 if (next->r_flags & RACK_ACKED) {
8521 /* yep this and next can be merged */
8522 rsm = rack_merge_rsm(rack, rsm, next);
8523 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8527 /* Now what about the previous? */
8528 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8530 if (prev->r_flags & RACK_ACKED) {
8531 /* yep the previous and this can be merged */
8532 rsm = rack_merge_rsm(rack, prev, rsm);
8533 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8538 if (used_ref == 0) {
8539 counter_u64_add(rack_sack_proc_all, 1);
8541 counter_u64_add(rack_sack_proc_short, 1);
8543 /* Save off the next one for quick reference. */
8545 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8548 *prsm = rack->r_ctl.rc_sacklast = nrsm;
8549 /* Pass back the moved. */
8555 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8557 struct rack_sendmap *tmap;
8560 while (rsm && (rsm->r_flags & RACK_ACKED)) {
8561 /* Its no longer sacked, mark it so */
8562 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8564 if (rsm->r_in_tmap) {
8565 panic("rack:%p rsm:%p flags:0x%x in tmap?",
8566 rack, rsm, rsm->r_flags);
8569 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8570 /* Rebuild it into our tmap */
8572 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8575 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8578 tmap->r_in_tmap = 1;
8579 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8582 * Now lets possibly clear the sack filter so we start
8583 * recognizing sacks that cover this area.
8585 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8590 rack_do_decay(struct tcp_rack *rack)
8594 #define timersub(tvp, uvp, vvp) \
8596 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \
8597 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \
8598 if ((vvp)->tv_usec < 0) { \
8600 (vvp)->tv_usec += 1000000; \
8604 timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8607 rack->r_ctl.input_pkt++;
8608 if ((rack->rc_in_persist) ||
8609 (res.tv_sec >= 1) ||
8610 (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8612 * Check for decay of non-SAD,
8613 * we want all SAD detection metrics to
8614 * decay 1/4 per second (or more) passed.
8618 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8619 /* Update our saved tracking values */
8620 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8621 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8622 /* Now do we escape without decay? */
8623 #ifdef NETFLIX_EXP_DETECTION
8624 if (rack->rc_in_persist ||
8625 (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8626 (pkt_delta < tcp_sad_low_pps)){
8628 * We don't decay idle connections
8629 * or ones that have a low input pps.
8633 /* Decay the counters */
8634 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8636 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8638 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8640 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8647 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8649 struct rack_sendmap *rsm, *rm;
8652 * The ACK point is advancing to th_ack, we must drop off
8653 * the packets in the rack log and calculate any eligble
8656 rack->r_wanted_output = 1;
8658 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8660 if ((th_ack - 1) == tp->iss) {
8662 * For the SYN incoming case we will not
8663 * have called tcp_output for the sending of
8664 * the SYN, so there will be no map. All
8665 * other cases should probably be a panic.
8669 if (tp->t_flags & TF_SENTFIN) {
8670 /* if we sent a FIN we often will not have map */
8674 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8676 tp->t_state, th_ack, rack,
8677 tp->snd_una, tp->snd_max, tp->snd_nxt);
8681 if (SEQ_LT(th_ack, rsm->r_start)) {
8682 /* Huh map is missing this */
8684 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8686 th_ack, tp->t_state, rack->r_state);
8690 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8691 /* Now do we consume the whole thing? */
8692 if (SEQ_GEQ(th_ack, rsm->r_end)) {
8693 /* Its all consumed. */
8695 uint8_t newly_acked;
8697 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8698 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8699 rsm->r_rtr_bytes = 0;
8700 /* Record the time of highest cumack sent */
8701 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8702 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8705 panic("removing head in rack:%p rsm:%p rm:%p",
8709 if (rsm->r_in_tmap) {
8710 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8714 if (rsm->r_flags & RACK_ACKED) {
8716 * It was acked on the scoreboard -- remove
8719 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8721 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8723 * There are segments ACKED on the
8724 * scoreboard further up. We are seeing
8727 rsm->r_flags &= ~RACK_SACK_PASSED;
8728 counter_u64_add(rack_reorder_seen, 1);
8729 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8730 rsm->r_flags |= RACK_ACKED;
8731 rack->r_ctl.rc_reorder_ts = cts;
8732 if (rack->r_ent_rec_ns) {
8734 * We have sent no more, and we saw an sack
8737 rack->r_might_revert = 1;
8740 if ((rsm->r_flags & RACK_TO_REXT) &&
8741 (tp->t_flags & TF_RCVD_TSTMP) &&
8742 (to->to_flags & TOF_TS) &&
8743 (tp->t_flags & TF_PREVVALID)) {
8745 * We can use the timestamp to see
8746 * if this retransmission was from the
8747 * first transmit. If so we made a mistake.
8749 tp->t_flags &= ~TF_PREVVALID;
8750 if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8751 /* The first transmit is what this ack is for */
8752 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8755 left = th_ack - rsm->r_end;
8756 if (rack->app_limited_needs_set && newly_acked)
8757 rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8758 /* Free back to zone */
8759 rack_free(rack, rsm);
8763 /* Check for reneging */
8764 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8765 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8767 * The peer has moved snd_una up to
8768 * the edge of this send, i.e. one
8769 * that it had previously acked. The only
8770 * way that can be true if the peer threw
8771 * away data (space issues) that it had
8772 * previously sacked (else it would have
8773 * given us snd_una up to (rsm->r_end).
8774 * We need to undo the acked markings here.
8776 * Note we have to look to make sure th_ack is
8777 * our rsm->r_start in case we get an old ack
8778 * where th_ack is behind snd_una.
8780 rack_peer_reneges(rack, rsm, th_ack);
8784 if (rsm->r_flags & RACK_ACKED) {
8786 * It was acked on the scoreboard -- remove it from
8787 * total for the part being cum-acked.
8789 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8792 * Clear the dup ack count for
8793 * the piece that remains.
8796 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8797 if (rsm->r_rtr_bytes) {
8799 * It was retransmitted adjust the
8800 * sack holes for what was acked.
8804 ack_am = (th_ack - rsm->r_start);
8805 if (ack_am >= rsm->r_rtr_bytes) {
8806 rack->r_ctl.rc_holes_rxt -= ack_am;
8807 rsm->r_rtr_bytes -= ack_am;
8811 * Update where the piece starts and record
8812 * the time of send of highest cumack sent.
8814 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8815 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8816 /* Now we need to move our offset forward too */
8817 if (rsm->orig_m_len != rsm->m->m_len) {
8818 /* Fix up the orig_m_len and possibly the mbuf offset */
8819 rack_adjust_orig_mlen(rsm);
8821 rsm->soff += (th_ack - rsm->r_start);
8822 rsm->r_start = th_ack;
8823 /* Now do we need to move the mbuf fwd too? */
8824 while (rsm->soff >= rsm->m->m_len) {
8825 rsm->soff -= rsm->m->m_len;
8826 rsm->m = rsm->m->m_next;
8827 KASSERT((rsm->m != NULL),
8828 (" nrsm:%p hit at soff:%u null m",
8831 rsm->orig_m_len = rsm->m->m_len;
8832 if (rack->app_limited_needs_set)
8833 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8837 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8839 struct rack_sendmap *rsm;
8840 int sack_pass_fnd = 0;
8842 if (rack->r_might_revert) {
8844 * Ok we have reordering, have not sent anything, we
8845 * might want to revert the congestion state if nothing
8846 * further has SACK_PASSED on it. Lets check.
8848 * We also get here when we have DSACKs come in for
8849 * all the data that we FR'd. Note that a rxt or tlp
8850 * timer clears this from happening.
8853 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8854 if (rsm->r_flags & RACK_SACK_PASSED) {
8859 if (sack_pass_fnd == 0) {
8861 * We went into recovery
8862 * incorrectly due to reordering!
8866 rack->r_ent_rec_ns = 0;
8867 orig_cwnd = tp->snd_cwnd;
8868 tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8869 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8870 tp->snd_recover = tp->snd_una;
8871 rack_log_to_prr(rack, 14, orig_cwnd);
8872 EXIT_RECOVERY(tp->t_flags);
8874 rack->r_might_revert = 0;
8878 #ifdef NETFLIX_EXP_DETECTION
8880 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack, uint32_t bytes_this_ack, uint32_t segsiz)
8882 if ((rack->do_detection || tcp_force_detection) &&
8883 tcp_sack_to_ack_thresh &&
8884 tcp_sack_to_move_thresh &&
8885 ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8887 * We have thresholds set to find
8888 * possible attackers and disable sack.
8891 uint64_t ackratio, moveratio, movetotal;
8894 rack_log_sad(rack, 1);
8895 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8896 ackratio *= (uint64_t)(1000);
8897 if (rack->r_ctl.ack_count)
8898 ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8900 /* We really should not hit here */
8903 if ((rack->sack_attack_disable == 0) &&
8904 (ackratio > rack_highest_sack_thresh_seen))
8905 rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8906 movetotal = rack->r_ctl.sack_moved_extra;
8907 movetotal += rack->r_ctl.sack_noextra_move;
8908 moveratio = rack->r_ctl.sack_moved_extra;
8909 moveratio *= (uint64_t)1000;
8911 moveratio /= movetotal;
8913 /* No moves, thats pretty good */
8916 if ((rack->sack_attack_disable == 0) &&
8917 (moveratio > rack_highest_move_thresh_seen))
8918 rack_highest_move_thresh_seen = (uint32_t)moveratio;
8919 if (rack->sack_attack_disable == 0) {
8920 if ((ackratio > tcp_sack_to_ack_thresh) &&
8921 (moveratio > tcp_sack_to_move_thresh)) {
8922 /* Disable sack processing */
8923 rack->sack_attack_disable = 1;
8924 if (rack->r_rep_attack == 0) {
8925 rack->r_rep_attack = 1;
8926 counter_u64_add(rack_sack_attacks_detected, 1);
8928 if (tcp_attack_on_turns_on_logging) {
8930 * Turn on logging, used for debugging
8933 rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
8935 /* Clamp the cwnd at flight size */
8936 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
8937 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
8938 rack_log_sad(rack, 2);
8941 /* We are sack-disabled check for false positives */
8942 if ((ackratio <= tcp_restoral_thresh) ||
8943 (rack->r_ctl.rc_num_maps_alloced < tcp_map_minimum)) {
8944 rack->sack_attack_disable = 0;
8945 rack_log_sad(rack, 3);
8946 /* Restart counting */
8947 rack->r_ctl.sack_count = 0;
8948 rack->r_ctl.sack_moved_extra = 0;
8949 rack->r_ctl.sack_noextra_move = 1;
8950 rack->r_ctl.ack_count = max(1,
8951 (bytes_this_ack / segsiz));
8953 if (rack->r_rep_reverse == 0) {
8954 rack->r_rep_reverse = 1;
8955 counter_u64_add(rack_sack_attacks_reversed, 1);
8957 /* Restore the cwnd */
8958 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
8959 rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
8967 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
8972 if (SEQ_GT(end, start))
8977 * We keep track of how many DSACK blocks we get
8978 * after a recovery incident.
8980 rack->r_ctl.dsack_byte_cnt += am;
8981 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
8982 rack->r_ctl.retran_during_recovery &&
8983 (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
8985 * False recovery most likely culprit is reordering. If
8986 * nothing else is missing we need to revert.
8988 rack->r_might_revert = 1;
8989 rack_handle_might_revert(rack->rc_tp, rack);
8990 rack->r_might_revert = 0;
8991 rack->r_ctl.retran_during_recovery = 0;
8992 rack->r_ctl.dsack_byte_cnt = 0;
8997 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
8999 /* Deal with changed and PRR here (in recovery only) */
9000 uint32_t pipe, snd_una;
9002 rack->r_ctl.rc_prr_delivered += changed;
9004 if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9006 * It is all outstanding, we are application limited
9007 * and thus we don't need more room to send anything.
9008 * Note we use tp->snd_una here and not th_ack because
9009 * the data as yet not been cut from the sb.
9011 rack->r_ctl.rc_prr_sndcnt = 0;
9014 /* Compute prr_sndcnt */
9015 if (SEQ_GT(tp->snd_una, th_ack)) {
9016 snd_una = tp->snd_una;
9020 pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9021 if (pipe > tp->snd_ssthresh) {
9024 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9025 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9026 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9028 rack->r_ctl.rc_prr_sndcnt = 0;
9029 rack_log_to_prr(rack, 9, 0);
9033 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9034 sndcnt -= rack->r_ctl.rc_prr_out;
9037 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9038 rack_log_to_prr(rack, 10, 0);
9042 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9043 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9046 if (changed > limit)
9048 limit += ctf_fixed_maxseg(tp);
9049 if (tp->snd_ssthresh > pipe) {
9050 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9051 rack_log_to_prr(rack, 11, 0);
9053 rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9054 rack_log_to_prr(rack, 12, 0);
9060 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9063 struct tcp_rack *rack;
9064 struct rack_sendmap *rsm;
9065 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9066 register uint32_t th_ack;
9067 int32_t i, j, k, num_sack_blks = 0;
9068 uint32_t cts, acked, ack_point, sack_changed = 0;
9069 int loop_start = 0, moved_two = 0;
9073 INP_WLOCK_ASSERT(tp->t_inpcb);
9074 if (th->th_flags & TH_RST) {
9075 /* We don't log resets */
9078 rack = (struct tcp_rack *)tp->t_fb_ptr;
9079 cts = tcp_get_usecs(NULL);
9080 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9082 th_ack = th->th_ack;
9083 if (rack->sack_attack_disable == 0)
9084 rack_do_decay(rack);
9085 if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9087 * You only get credit for
9088 * MSS and greater (and you get extra
9089 * credit for larger cum-ack moves).
9093 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9094 rack->r_ctl.ack_count += ac;
9095 counter_u64_add(rack_ack_total, ac);
9097 if (rack->r_ctl.ack_count > 0xfff00000) {
9099 * reduce the number to keep us under
9102 rack->r_ctl.ack_count /= 2;
9103 rack->r_ctl.sack_count /= 2;
9105 if (SEQ_GT(th_ack, tp->snd_una)) {
9106 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9107 tp->t_acktime = ticks;
9109 if (rsm && SEQ_GT(th_ack, rsm->r_start))
9110 changed = th_ack - rsm->r_start;
9112 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9114 if ((to->to_flags & TOF_SACK) == 0) {
9115 /* We are done nothing left and no sack. */
9116 rack_handle_might_revert(tp, rack);
9118 * For cases where we struck a dup-ack
9119 * with no SACK, add to the changes so
9120 * PRR will work right.
9122 if (dup_ack_struck && (changed == 0)) {
9123 changed += ctf_fixed_maxseg(rack->rc_tp);
9127 /* Sack block processing */
9128 if (SEQ_GT(th_ack, tp->snd_una))
9131 ack_point = tp->snd_una;
9132 for (i = 0; i < to->to_nsacks; i++) {
9133 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9134 &sack, sizeof(sack));
9135 sack.start = ntohl(sack.start);
9136 sack.end = ntohl(sack.end);
9137 if (SEQ_GT(sack.end, sack.start) &&
9138 SEQ_GT(sack.start, ack_point) &&
9139 SEQ_LT(sack.start, tp->snd_max) &&
9140 SEQ_GT(sack.end, ack_point) &&
9141 SEQ_LEQ(sack.end, tp->snd_max)) {
9142 sack_blocks[num_sack_blks] = sack;
9144 #ifdef NETFLIX_STATS
9145 } else if (SEQ_LEQ(sack.start, th_ack) &&
9146 SEQ_LEQ(sack.end, th_ack)) {
9148 * Its a D-SACK block.
9150 tcp_record_dsack(sack.start, sack.end);
9152 rack_note_dsack(rack, sack.start, sack.end);
9156 * Sort the SACK blocks so we can update the rack scoreboard with
9159 num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9160 num_sack_blks, th->th_ack);
9161 ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9162 if (num_sack_blks == 0) {
9163 /* Nothing to sack (DSACKs?) */
9164 goto out_with_totals;
9166 if (num_sack_blks < 2) {
9167 /* Only one, we don't need to sort */
9170 /* Sort the sacks */
9171 for (i = 0; i < num_sack_blks; i++) {
9172 for (j = i + 1; j < num_sack_blks; j++) {
9173 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9174 sack = sack_blocks[i];
9175 sack_blocks[i] = sack_blocks[j];
9176 sack_blocks[j] = sack;
9181 * Now are any of the sack block ends the same (yes some
9182 * implementations send these)?
9185 if (num_sack_blks == 0)
9186 goto out_with_totals;
9187 if (num_sack_blks > 1) {
9188 for (i = 0; i < num_sack_blks; i++) {
9189 for (j = i + 1; j < num_sack_blks; j++) {
9190 if (sack_blocks[i].end == sack_blocks[j].end) {
9192 * Ok these two have the same end we
9193 * want the smallest end and then
9194 * throw away the larger and start
9197 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9199 * The second block covers
9200 * more area use that
9202 sack_blocks[i].start = sack_blocks[j].start;
9205 * Now collapse out the dup-sack and
9208 for (k = (j + 1); k < num_sack_blks; k++) {
9209 sack_blocks[j].start = sack_blocks[k].start;
9210 sack_blocks[j].end = sack_blocks[k].end;
9221 * First lets look to see if
9222 * we have retransmitted and
9223 * can use the transmit next?
9225 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9227 SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9228 SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9230 * We probably did the FR and the next
9231 * SACK in continues as we would expect.
9233 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9235 rack->r_wanted_output = 1;
9237 sack_changed += acked;
9239 if (num_sack_blks == 1) {
9241 * This is what we would expect from
9242 * a normal implementation to happen
9243 * after we have retransmitted the FR,
9244 * i.e the sack-filter pushes down
9245 * to 1 block and the next to be retransmitted
9246 * is the sequence in the sack block (has more
9247 * are acked). Count this as ACK'd data to boost
9248 * up the chances of recovering any false positives.
9250 rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9251 counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9252 counter_u64_add(rack_express_sack, 1);
9253 if (rack->r_ctl.ack_count > 0xfff00000) {
9255 * reduce the number to keep us under
9258 rack->r_ctl.ack_count /= 2;
9259 rack->r_ctl.sack_count /= 2;
9261 goto out_with_totals;
9264 * Start the loop through the
9265 * rest of blocks, past the first block.
9271 /* Its a sack of some sort */
9272 rack->r_ctl.sack_count++;
9273 if (rack->r_ctl.sack_count > 0xfff00000) {
9275 * reduce the number to keep us under
9278 rack->r_ctl.ack_count /= 2;
9279 rack->r_ctl.sack_count /= 2;
9281 counter_u64_add(rack_sack_total, 1);
9282 if (rack->sack_attack_disable) {
9283 /* An attacker disablement is in place */
9284 if (num_sack_blks > 1) {
9285 rack->r_ctl.sack_count += (num_sack_blks - 1);
9286 rack->r_ctl.sack_moved_extra++;
9287 counter_u64_add(rack_move_some, 1);
9288 if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9289 rack->r_ctl.sack_moved_extra /= 2;
9290 rack->r_ctl.sack_noextra_move /= 2;
9295 rsm = rack->r_ctl.rc_sacklast;
9296 for (i = loop_start; i < num_sack_blks; i++) {
9297 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9299 rack->r_wanted_output = 1;
9301 sack_changed += acked;
9305 * If we did not get a SACK for at least a MSS and
9306 * had to move at all, or if we moved more than our
9307 * threshold, it counts against the "extra" move.
9309 rack->r_ctl.sack_moved_extra += moved_two;
9310 counter_u64_add(rack_move_some, 1);
9313 * else we did not have to move
9314 * any more than we would expect.
9316 rack->r_ctl.sack_noextra_move++;
9317 counter_u64_add(rack_move_none, 1);
9319 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9321 * If the SACK was not a full MSS then
9322 * we add to sack_count the number of
9323 * MSS's (or possibly more than
9324 * a MSS if its a TSO send) we had to skip by.
9326 rack->r_ctl.sack_count += moved_two;
9327 counter_u64_add(rack_sack_total, moved_two);
9330 * Now we need to setup for the next
9331 * round. First we make sure we won't
9332 * exceed the size of our uint32_t on
9333 * the various counts, and then clear out
9336 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9337 (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9338 rack->r_ctl.sack_moved_extra /= 2;
9339 rack->r_ctl.sack_noextra_move /= 2;
9341 if (rack->r_ctl.sack_count > 0xfff00000) {
9342 rack->r_ctl.ack_count /= 2;
9343 rack->r_ctl.sack_count /= 2;
9348 if (num_sack_blks > 1) {
9350 * You get an extra stroke if
9351 * you have more than one sack-blk, this
9352 * could be where we are skipping forward
9353 * and the sack-filter is still working, or
9354 * it could be an attacker constantly
9357 rack->r_ctl.sack_moved_extra++;
9358 counter_u64_add(rack_move_some, 1);
9361 #ifdef NETFLIX_EXP_DETECTION
9362 rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9365 /* Something changed cancel the rack timer */
9366 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9368 tsused = tcp_get_usecs(NULL);
9369 rsm = tcp_rack_output(tp, rack, tsused);
9370 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9372 /* Enter recovery */
9373 rack->r_ctl.rc_rsm_start = rsm->r_start;
9374 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9375 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9376 entered_recovery = 1;
9377 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9379 * When we enter recovery we need to assure we send
9382 if (rack->rack_no_prr == 0) {
9383 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9384 rack_log_to_prr(rack, 8, 0);
9386 rack->r_timer_override = 1;
9388 rack->r_ctl.rc_agg_early = 0;
9389 } else if (IN_FASTRECOVERY(tp->t_flags) &&
9391 (rack->r_rr_config == 3)) {
9393 * Assure we can output and we get no
9394 * remembered pace time except the retransmit.
9396 rack->r_timer_override = 1;
9397 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9398 rack->r_ctl.rc_resend = rsm;
9400 if (IN_FASTRECOVERY(tp->t_flags) &&
9401 (rack->rack_no_prr == 0) &&
9402 (entered_recovery == 0)) {
9403 rack_update_prr(tp, rack, changed, th_ack);
9404 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9405 ((rack->rc_inp->inp_in_hpts == 0) &&
9406 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9408 * If you are pacing output you don't want
9412 rack->r_ctl.rc_agg_early = 0;
9413 rack->r_timer_override = 1;
9419 rack_strike_dupack(struct tcp_rack *rack)
9421 struct rack_sendmap *rsm;
9423 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9424 while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9425 rsm = TAILQ_NEXT(rsm, r_tnext);
9427 if (rsm && (rsm->r_dupack < 0xff)) {
9429 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9433 * Here we see if we need to retransmit. For
9434 * a SACK type connection if enough time has passed
9435 * we will get a return of the rsm. For a non-sack
9436 * connection we will get the rsm returned if the
9437 * dupack value is 3 or more.
9439 cts = tcp_get_usecs(&tv);
9440 rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9441 if (rack->r_ctl.rc_resend != NULL) {
9442 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9443 rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9444 rack->rc_tp->snd_una);
9446 rack->r_wanted_output = 1;
9447 rack->r_timer_override = 1;
9448 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9451 rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9457 rack_check_bottom_drag(struct tcpcb *tp,
9458 struct tcp_rack *rack,
9459 struct socket *so, int32_t acked)
9461 uint32_t segsiz, minseg;
9463 segsiz = ctf_fixed_maxseg(tp);
9466 if (tp->snd_max == tp->snd_una) {
9468 * We are doing dynamic pacing and we are way
9469 * under. Basically everything got acked while
9470 * we were still waiting on the pacer to expire.
9472 * This means we need to boost the b/w in
9473 * addition to any earlier boosting of
9476 rack->rc_dragged_bottom = 1;
9477 rack_validate_multipliers_at_or_above100(rack);
9479 * Lets use the segment bytes acked plus
9480 * the lowest RTT seen as the basis to
9481 * form a b/w estimate. This will be off
9482 * due to the fact that the true estimate
9483 * should be around 1/2 the time of the RTT
9484 * but we can settle for that.
9486 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9488 uint64_t bw, calc_bw, rtt;
9490 rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9492 /* no us sample is there a ms one? */
9493 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9494 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9496 goto no_measurement;
9500 calc_bw = bw * 1000000;
9502 if (rack->r_ctl.last_max_bw &&
9503 (rack->r_ctl.last_max_bw < calc_bw)) {
9505 * If we have a last calculated max bw
9508 calc_bw = rack->r_ctl.last_max_bw;
9510 /* now plop it in */
9511 if (rack->rc_gp_filled == 0) {
9512 if (calc_bw > ONE_POINT_TWO_MEG) {
9514 * If we have no measurement
9515 * don't let us set in more than
9516 * 1.2Mbps. If we are still too
9517 * low after pacing with this we
9518 * will hopefully have a max b/w
9519 * available to sanity check things.
9521 calc_bw = ONE_POINT_TWO_MEG;
9523 rack->r_ctl.rc_rtt_diff = 0;
9524 rack->r_ctl.gp_bw = calc_bw;
9525 rack->rc_gp_filled = 1;
9526 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9527 rack->r_ctl.num_measurements = RACK_REQ_AVG;
9528 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9529 } else if (calc_bw > rack->r_ctl.gp_bw) {
9530 rack->r_ctl.rc_rtt_diff = 0;
9531 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9532 rack->r_ctl.num_measurements = RACK_REQ_AVG;
9533 rack->r_ctl.gp_bw = calc_bw;
9534 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9536 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9537 if ((rack->gp_ready == 0) &&
9538 (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9539 /* We have enough measurements now */
9541 rack_set_cc_pacing(rack);
9542 if (rack->defer_options)
9543 rack_apply_deferred_options(rack);
9546 * For acks over 1mss we do a extra boost to simulate
9547 * where we would get 2 acks (we want 110 for the mul).
9550 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9553 * zero rtt possibly?, settle for just an old increase.
9556 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9558 } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9559 (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9561 (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9562 (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9563 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9564 (segsiz * rack_req_segs))) {
9566 * We are doing dynamic GP pacing and
9567 * we have everything except 1MSS or less
9568 * bytes left out. We are still pacing away.
9569 * And there is data that could be sent, This
9570 * means we are inserting delayed ack time in
9571 * our measurements because we are pacing too slow.
9573 rack_validate_multipliers_at_or_above100(rack);
9574 rack->rc_dragged_bottom = 1;
9575 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9582 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9585 * The fast output path is enabled and we
9586 * have moved the cumack forward. Lets see if
9587 * we can expand forward the fast path length by
9588 * that amount. What we would ideally like to
9589 * do is increase the number of bytes in the
9590 * fast path block (left_to_send) by the
9591 * acked amount. However we have to gate that
9593 * 1) The amount outstanding and the rwnd of the peer
9594 * (i.e. we don't want to exceed the rwnd of the peer).
9596 * 2) The amount of data left in the socket buffer (i.e.
9597 * we can't send beyond what is in the buffer).
9599 * Note that this does not take into account any increase
9600 * in the cwnd. We will only extend the fast path by
9603 uint32_t new_total, gating_val;
9605 new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9606 gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9607 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9608 if (new_total <= gating_val) {
9609 /* We can increase left_to_send by the acked amount */
9610 counter_u64_add(rack_extended_rfo, 1);
9611 rack->r_ctl.fsb.left_to_send = new_total;
9612 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9613 ("rack:%p left_to_send:%u sbavail:%u out:%u",
9614 rack, rack->r_ctl.fsb.left_to_send,
9615 sbavail(&rack->rc_inp->inp_socket->so_snd),
9616 (tp->snd_max - tp->snd_una)));
9622 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9625 * Here any sendmap entry that points to the
9626 * beginning mbuf must be adjusted to the correct
9627 * offset. This must be called with:
9628 * 1) The socket buffer locked
9629 * 2) snd_una adjusted to its new postion.
9631 * Note that (2) implies rack_ack_received has also
9634 * We grab the first mbuf in the socket buffer and
9635 * then go through the front of the sendmap, recalculating
9636 * the stored offset for any sendmap entry that has
9637 * that mbuf. We must use the sb functions to do this
9638 * since its possible an add was done has well as
9639 * the subtraction we may have just completed. This should
9640 * not be a penalty though, since we just referenced the sb
9641 * to go in and trim off the mbufs that we freed (of course
9642 * there will be a penalty for the sendmap references though).
9645 struct rack_sendmap *rsm;
9647 SOCKBUF_LOCK_ASSERT(sb);
9649 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9650 if ((rsm == NULL) || (m == NULL)) {
9651 /* Nothing outstanding */
9654 while (rsm->m == m) {
9660 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9661 if (rsm->orig_m_len != m->m_len) {
9662 rack_adjust_orig_mlen(rsm);
9664 if (rsm->soff != soff) {
9666 * This is not a fatal error, we anticipate it
9667 * might happen (the else code), so we count it here
9668 * so that under invariant we can see that it really
9671 counter_u64_add(rack_adjust_map_bw, 1);
9675 rsm->orig_m_len = rsm->m->m_len;
9677 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9678 rsm->orig_m_len = rsm->m->m_len;
9680 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9688 * Return value of 1, we do not need to call rack_process_data().
9689 * return value of 0, rack_process_data can be called.
9690 * For ret_val if its 0 the TCP is locked, if its non-zero
9691 * its unlocked and probably unsafe to touch the TCB.
9694 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9695 struct tcpcb *tp, struct tcpopt *to,
9696 uint32_t tiwin, int32_t tlen,
9697 int32_t * ofia, int32_t thflags, int32_t *ret_val)
9699 int32_t ourfinisacked = 0;
9700 int32_t nsegs, acked_amount;
9703 struct tcp_rack *rack;
9704 int32_t under_pacing = 0;
9705 int32_t recovery = 0;
9707 rack = (struct tcp_rack *)tp->t_fb_ptr;
9708 if (SEQ_GT(th->th_ack, tp->snd_max)) {
9709 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9710 &rack->r_ctl.challenge_ack_ts,
9711 &rack->r_ctl.challenge_ack_cnt);
9712 rack->r_wanted_output = 1;
9715 if (rack->gp_ready &&
9716 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9719 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9720 int in_rec, dup_ack_struck = 0;
9722 in_rec = IN_FASTRECOVERY(tp->t_flags);
9723 if (rack->rc_in_persist) {
9725 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9726 rack_rto_min, rack_rto_max);
9728 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9729 rack_strike_dupack(rack);
9732 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9734 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9736 * Old ack, behind (or duplicate to) the last one rcv'd
9737 * Note: We mark reordering is occuring if its
9738 * less than and we have not closed our window.
9740 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9741 counter_u64_add(rack_reorder_seen, 1);
9742 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9747 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9748 * something we sent.
9750 if (tp->t_flags & TF_NEEDSYN) {
9752 * T/TCP: Connection was half-synchronized, and our SYN has
9753 * been ACK'd (so connection is now fully synchronized). Go
9754 * to non-starred state, increment snd_una for ACK of SYN,
9755 * and check if we can do window scaling.
9757 tp->t_flags &= ~TF_NEEDSYN;
9759 /* Do window scaling? */
9760 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9761 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9762 tp->rcv_scale = tp->request_r_scale;
9763 /* Send window already scaled. */
9766 nsegs = max(1, m->m_pkthdr.lro_nsegs);
9767 INP_WLOCK_ASSERT(tp->t_inpcb);
9769 acked = BYTES_THIS_ACK(tp, th);
9770 KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9771 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9773 * If we just performed our first retransmit, and the ACK arrives
9774 * within our recovery window, then it was a mistake to do the
9775 * retransmit in the first place. Recover our original cwnd and
9776 * ssthresh, and proceed to transmit where we left off.
9778 if ((tp->t_flags & TF_PREVVALID) &&
9779 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9780 tp->t_flags &= ~TF_PREVVALID;
9781 if (tp->t_rxtshift == 1 &&
9782 (int)(ticks - tp->t_badrxtwin) < 0)
9783 rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9786 /* assure we are not backed off */
9788 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9789 rack_rto_min, rack_rto_max);
9790 rack->rc_tlp_in_progress = 0;
9791 rack->r_ctl.rc_tlp_cnt_out = 0;
9793 * If it is the RXT timer we want to
9794 * stop it, so we can restart a TLP.
9796 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9797 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9798 #ifdef NETFLIX_HTTP_LOGGING
9799 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9803 * If we have a timestamp reply, update smoothed round trip time. If
9804 * no timestamp is present but transmit timer is running and timed
9805 * sequence number was acked, update smoothed round trip time. Since
9806 * we now have an rtt measurement, cancel the timer backoff (cf.,
9807 * Phil Karn's retransmit alg.). Recompute the initial retransmit
9810 * Some boxes send broken timestamp replies during the SYN+ACK
9811 * phase, ignore timestamps of 0 or we could calculate a huge RTT
9812 * and blow up the retransmit timer.
9815 * If all outstanding data is acked, stop retransmit timer and
9816 * remember to restart (more output or persist). If there is more
9817 * data to be acked, restart retransmit timer, using current
9818 * (possibly backed-off) value.
9822 *ofia = ourfinisacked;
9825 if (IN_RECOVERY(tp->t_flags)) {
9826 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9827 (SEQ_LT(th->th_ack, tp->snd_max))) {
9828 tcp_rack_partialack(tp);
9830 rack_post_recovery(tp, th->th_ack);
9835 * Let the congestion control algorithm update congestion control
9836 * related information. This typically means increasing the
9837 * congestion window.
9839 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9840 SOCKBUF_LOCK(&so->so_snd);
9841 acked_amount = min(acked, (int)sbavail(&so->so_snd));
9842 tp->snd_wnd -= acked_amount;
9843 mfree = sbcut_locked(&so->so_snd, acked_amount);
9844 if ((sbused(&so->so_snd) == 0) &&
9845 (acked > acked_amount) &&
9846 (tp->t_state >= TCPS_FIN_WAIT_1) &&
9847 (tp->t_flags & TF_SENTFIN)) {
9849 * We must be sure our fin
9850 * was sent and acked (we can be
9851 * in FIN_WAIT_1 without having
9856 tp->snd_una = th->th_ack;
9857 if (acked_amount && sbavail(&so->so_snd))
9858 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9859 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9860 SOCKBUF_UNLOCK(&so->so_snd);
9861 tp->t_flags |= TF_WAKESOW;
9863 if (SEQ_GT(tp->snd_una, tp->snd_recover))
9864 tp->snd_recover = tp->snd_una;
9866 if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9867 tp->snd_nxt = tp->snd_una;
9870 (rack->use_fixed_rate == 0) &&
9871 (rack->in_probe_rtt == 0) &&
9872 rack->rc_gp_dyn_mul &&
9873 rack->rc_always_pace) {
9874 /* Check if we are dragging bottom */
9875 rack_check_bottom_drag(tp, rack, so, acked);
9877 if (tp->snd_una == tp->snd_max) {
9878 /* Nothing left outstanding */
9879 tp->t_flags &= ~TF_PREVVALID;
9880 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9881 rack->r_ctl.retran_during_recovery = 0;
9882 rack->r_ctl.dsack_byte_cnt = 0;
9883 if (rack->r_ctl.rc_went_idle_time == 0)
9884 rack->r_ctl.rc_went_idle_time = 1;
9885 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9886 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9888 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9889 /* Set need output so persist might get set */
9890 rack->r_wanted_output = 1;
9891 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9892 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9893 (sbavail(&so->so_snd) == 0) &&
9894 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9896 * The socket was gone and the
9897 * peer sent data (now or in the past), time to
9901 /* tcp_close will kill the inp pre-log the Reset */
9902 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9904 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
9909 *ofia = ourfinisacked;
9914 rack_collapsed_window(struct tcp_rack *rack)
9917 * Now we must walk the
9918 * send map and divide the
9919 * ones left stranded. These
9920 * guys can't cause us to abort
9921 * the connection and are really
9922 * "unsent". However if a buggy
9923 * client actually did keep some
9924 * of the data i.e. collapsed the win
9925 * and refused to ack and then opened
9926 * the win and acked that data. We would
9927 * get into an ack war, the simplier
9928 * method then of just pretending we
9929 * did not send those segments something
9932 struct rack_sendmap *rsm, *nrsm, fe, *insret;
9935 max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
9936 memset(&fe, 0, sizeof(fe));
9937 fe.r_start = max_seq;
9938 /* Find the first seq past or at maxseq */
9939 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
9941 /* Nothing to do strange */
9942 rack->rc_has_collapsed = 0;
9946 * Now do we need to split at
9947 * the collapse point?
9949 if (SEQ_GT(max_seq, rsm->r_start)) {
9950 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9952 /* We can't get a rsm, mark all? */
9957 rack_clone_rsm(rack, nrsm, rsm, max_seq);
9958 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
9960 if (insret != NULL) {
9961 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
9962 nrsm, insret, rack, rsm);
9965 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
9966 if (rsm->r_in_tmap) {
9967 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9968 nrsm->r_in_tmap = 1;
9971 * Set in the new RSM as the
9972 * collapsed starting point
9977 counter_u64_add(rack_collapsed_win, 1);
9978 RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
9979 nrsm->r_flags |= RACK_RWND_COLLAPSED;
9981 rack->rc_has_collapsed = 1;
9985 rack_un_collapse_window(struct tcp_rack *rack)
9987 struct rack_sendmap *rsm;
9989 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
9990 if (rsm->r_flags & RACK_RWND_COLLAPSED)
9991 rsm->r_flags &= ~RACK_RWND_COLLAPSED;
9995 rack->rc_has_collapsed = 0;
9999 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10000 int32_t tlen, int32_t tfo_syn)
10002 if (DELAY_ACK(tp, tlen) || tfo_syn) {
10003 if (rack->rc_dack_mode &&
10005 (rack->rc_dack_toggle == 1)) {
10006 goto no_delayed_ack;
10008 rack_timer_cancel(tp, rack,
10009 rack->r_ctl.rc_rcvtime, __LINE__);
10010 tp->t_flags |= TF_DELACK;
10013 rack->r_wanted_output = 1;
10014 tp->t_flags |= TF_ACKNOW;
10015 if (rack->rc_dack_mode) {
10016 if (tp->t_flags & TF_DELACK)
10017 rack->rc_dack_toggle = 1;
10019 rack->rc_dack_toggle = 0;
10025 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10028 * If fast output is in progress, lets validate that
10029 * the new window did not shrink on us and make it
10030 * so fast output should end.
10032 if (rack->r_fast_output) {
10036 * Calculate what we will send if left as is
10037 * and compare that to our send window.
10039 out = ctf_outstanding(tp);
10040 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10041 /* ok we have an issue */
10042 if (out >= tp->snd_wnd) {
10043 /* Turn off fast output the window is met or collapsed */
10044 rack->r_fast_output = 0;
10046 /* we have some room left */
10047 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10048 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10049 /* If not at least 1 full segment never mind */
10050 rack->r_fast_output = 0;
10058 * Return value of 1, the TCB is unlocked and most
10059 * likely gone, return value of 0, the TCP is still
10063 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10064 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10065 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10068 * Update window information. Don't look at window if no ACK: TAC's
10069 * send garbage on first SYN.
10073 struct tcp_rack *rack;
10075 rack = (struct tcp_rack *)tp->t_fb_ptr;
10076 INP_WLOCK_ASSERT(tp->t_inpcb);
10077 nsegs = max(1, m->m_pkthdr.lro_nsegs);
10078 if ((thflags & TH_ACK) &&
10079 (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10080 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10081 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10082 /* keep track of pure window updates */
10084 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10085 KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10086 tp->snd_wnd = tiwin;
10087 rack_validate_fo_sendwin_up(tp, rack);
10088 tp->snd_wl1 = th->th_seq;
10089 tp->snd_wl2 = th->th_ack;
10090 if (tp->snd_wnd > tp->max_sndwnd)
10091 tp->max_sndwnd = tp->snd_wnd;
10092 rack->r_wanted_output = 1;
10093 } else if (thflags & TH_ACK) {
10094 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10095 tp->snd_wnd = tiwin;
10096 rack_validate_fo_sendwin_up(tp, rack);
10097 tp->snd_wl1 = th->th_seq;
10098 tp->snd_wl2 = th->th_ack;
10101 if (tp->snd_wnd < ctf_outstanding(tp))
10102 /* The peer collapsed the window */
10103 rack_collapsed_window(rack);
10104 else if (rack->rc_has_collapsed)
10105 rack_un_collapse_window(rack);
10106 /* Was persist timer active and now we have window space? */
10107 if ((rack->rc_in_persist != 0) &&
10108 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10109 rack->r_ctl.rc_pace_min_segs))) {
10110 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10111 tp->snd_nxt = tp->snd_max;
10112 /* Make sure we output to start the timer */
10113 rack->r_wanted_output = 1;
10115 /* Do we enter persists? */
10116 if ((rack->rc_in_persist == 0) &&
10117 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10118 TCPS_HAVEESTABLISHED(tp->t_state) &&
10119 (tp->snd_max == tp->snd_una) &&
10120 sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10121 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10123 * Here the rwnd is less than
10124 * the pacing size, we are established,
10125 * nothing is outstanding, and there is
10126 * data to send. Enter persists.
10128 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10130 if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10135 * don't process the URG bit, ignore them drag
10138 tp->rcv_up = tp->rcv_nxt;
10139 INP_WLOCK_ASSERT(tp->t_inpcb);
10142 * Process the segment text, merging it into the TCP sequencing
10143 * queue, and arranging for acknowledgment of receipt if necessary.
10144 * This process logically involves adjusting tp->rcv_wnd as data is
10145 * presented to the user (this happens in tcp_usrreq.c, case
10146 * PRU_RCVD). If a FIN has already been received on this connection
10147 * then we just ignore the text.
10149 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10150 IS_FASTOPEN(tp->t_flags));
10151 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10152 TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10153 tcp_seq save_start = th->th_seq;
10154 tcp_seq save_rnxt = tp->rcv_nxt;
10155 int save_tlen = tlen;
10157 m_adj(m, drop_hdrlen); /* delayed header drop */
10159 * Insert segment which includes th into TCP reassembly
10160 * queue with control block tp. Set thflags to whether
10161 * reassembly now includes a segment with FIN. This handles
10162 * the common case inline (segment is the next to be
10163 * received on an established connection, and the queue is
10164 * empty), avoiding linkage into and removal from the queue
10165 * and repetition of various conversions. Set DELACK for
10166 * segments received in order, but ack immediately when
10167 * segments are out of order (so fast retransmit can work).
10169 if (th->th_seq == tp->rcv_nxt &&
10171 (TCPS_HAVEESTABLISHED(tp->t_state) ||
10173 #ifdef NETFLIX_SB_LIMITS
10174 u_int mcnt, appended;
10176 if (so->so_rcv.sb_shlim) {
10177 mcnt = m_memcnt(m);
10179 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10180 CFO_NOSLEEP, NULL) == false) {
10181 counter_u64_add(tcp_sb_shlim_fails, 1);
10187 rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10188 tp->rcv_nxt += tlen;
10190 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10191 (tp->t_fbyte_in == 0)) {
10192 tp->t_fbyte_in = ticks;
10193 if (tp->t_fbyte_in == 0)
10194 tp->t_fbyte_in = 1;
10195 if (tp->t_fbyte_out && tp->t_fbyte_in)
10196 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10198 thflags = th->th_flags & TH_FIN;
10199 KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10200 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10201 SOCKBUF_LOCK(&so->so_rcv);
10202 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10205 #ifdef NETFLIX_SB_LIMITS
10208 sbappendstream_locked(&so->so_rcv, m, 0);
10210 rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10211 SOCKBUF_UNLOCK(&so->so_rcv);
10212 tp->t_flags |= TF_WAKESOR;
10213 #ifdef NETFLIX_SB_LIMITS
10214 if (so->so_rcv.sb_shlim && appended != mcnt)
10215 counter_fo_release(so->so_rcv.sb_shlim,
10220 * XXX: Due to the header drop above "th" is
10221 * theoretically invalid by now. Fortunately
10222 * m_adj() doesn't actually frees any mbufs when
10223 * trimming from the head.
10225 tcp_seq temp = save_start;
10227 thflags = tcp_reass(tp, th, &temp, &tlen, m);
10228 tp->t_flags |= TF_ACKNOW;
10230 if ((tp->t_flags & TF_SACK_PERMIT) &&
10232 TCPS_HAVEESTABLISHED(tp->t_state)) {
10233 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10235 * DSACK actually handled in the fastpath
10238 RACK_OPTS_INC(tcp_sack_path_1);
10239 tcp_update_sack_list(tp, save_start,
10240 save_start + save_tlen);
10241 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10242 if ((tp->rcv_numsacks >= 1) &&
10243 (tp->sackblks[0].end == save_start)) {
10245 * Partial overlap, recorded at todrop
10248 RACK_OPTS_INC(tcp_sack_path_2a);
10249 tcp_update_sack_list(tp,
10250 tp->sackblks[0].start,
10251 tp->sackblks[0].end);
10253 RACK_OPTS_INC(tcp_sack_path_2b);
10254 tcp_update_dsack_list(tp, save_start,
10255 save_start + save_tlen);
10257 } else if (tlen >= save_tlen) {
10258 /* Update of sackblks. */
10259 RACK_OPTS_INC(tcp_sack_path_3);
10260 tcp_update_dsack_list(tp, save_start,
10261 save_start + save_tlen);
10262 } else if (tlen > 0) {
10263 RACK_OPTS_INC(tcp_sack_path_4);
10264 tcp_update_dsack_list(tp, save_start,
10265 save_start + tlen);
10270 thflags &= ~TH_FIN;
10274 * If FIN is received ACK the FIN and let the user know that the
10275 * connection is closing.
10277 if (thflags & TH_FIN) {
10278 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10280 /* The socket upcall is handled by socantrcvmore. */
10281 tp->t_flags &= ~TF_WAKESOR;
10283 * If connection is half-synchronized (ie NEEDSYN
10284 * flag on) then delay ACK, so it may be piggybacked
10285 * when SYN is sent. Otherwise, since we received a
10286 * FIN then no more input can be expected, send ACK
10289 if (tp->t_flags & TF_NEEDSYN) {
10290 rack_timer_cancel(tp, rack,
10291 rack->r_ctl.rc_rcvtime, __LINE__);
10292 tp->t_flags |= TF_DELACK;
10294 tp->t_flags |= TF_ACKNOW;
10298 switch (tp->t_state) {
10300 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10301 * CLOSE_WAIT state.
10303 case TCPS_SYN_RECEIVED:
10304 tp->t_starttime = ticks;
10306 case TCPS_ESTABLISHED:
10307 rack_timer_cancel(tp, rack,
10308 rack->r_ctl.rc_rcvtime, __LINE__);
10309 tcp_state_change(tp, TCPS_CLOSE_WAIT);
10313 * If still in FIN_WAIT_1 STATE FIN has not been
10314 * acked so enter the CLOSING state.
10316 case TCPS_FIN_WAIT_1:
10317 rack_timer_cancel(tp, rack,
10318 rack->r_ctl.rc_rcvtime, __LINE__);
10319 tcp_state_change(tp, TCPS_CLOSING);
10323 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10324 * starting the time-wait timer, turning off the
10325 * other standard timers.
10327 case TCPS_FIN_WAIT_2:
10328 rack_timer_cancel(tp, rack,
10329 rack->r_ctl.rc_rcvtime, __LINE__);
10335 * Return any desired output.
10337 if ((tp->t_flags & TF_ACKNOW) ||
10338 (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10339 rack->r_wanted_output = 1;
10341 INP_WLOCK_ASSERT(tp->t_inpcb);
10346 * Here nothing is really faster, its just that we
10347 * have broken out the fast-data path also just like
10351 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10352 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10353 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10356 int32_t newsize = 0; /* automatic sockbuf scaling */
10357 struct tcp_rack *rack;
10358 #ifdef NETFLIX_SB_LIMITS
10359 u_int mcnt, appended;
10363 * The size of tcp_saveipgen must be the size of the max ip header,
10366 u_char tcp_saveipgen[IP6_HDR_LEN];
10367 struct tcphdr tcp_savetcp;
10372 * If last ACK falls within this segment's sequence numbers, record
10373 * the timestamp. NOTE that the test is modified according to the
10374 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10376 if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10379 if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10382 if (tiwin && tiwin != tp->snd_wnd) {
10385 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10388 if (__predict_false((to->to_flags & TOF_TS) &&
10389 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10392 if (__predict_false((th->th_ack != tp->snd_una))) {
10395 if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10398 if ((to->to_flags & TOF_TS) != 0 &&
10399 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10400 tp->ts_recent_age = tcp_ts_getticks();
10401 tp->ts_recent = to->to_tsval;
10403 rack = (struct tcp_rack *)tp->t_fb_ptr;
10405 * This is a pure, in-sequence data packet with nothing on the
10406 * reassembly queue and we have enough buffer space to take it.
10408 nsegs = max(1, m->m_pkthdr.lro_nsegs);
10410 #ifdef NETFLIX_SB_LIMITS
10411 if (so->so_rcv.sb_shlim) {
10412 mcnt = m_memcnt(m);
10414 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10415 CFO_NOSLEEP, NULL) == false) {
10416 counter_u64_add(tcp_sb_shlim_fails, 1);
10422 /* Clean receiver SACK report if present */
10423 if (tp->rcv_numsacks)
10424 tcp_clean_sackreport(tp);
10425 KMOD_TCPSTAT_INC(tcps_preddat);
10426 tp->rcv_nxt += tlen;
10428 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10429 (tp->t_fbyte_in == 0)) {
10430 tp->t_fbyte_in = ticks;
10431 if (tp->t_fbyte_in == 0)
10432 tp->t_fbyte_in = 1;
10433 if (tp->t_fbyte_out && tp->t_fbyte_in)
10434 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10437 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10439 tp->snd_wl1 = th->th_seq;
10441 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10443 tp->rcv_up = tp->rcv_nxt;
10444 KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10445 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10447 if (so->so_options & SO_DEBUG)
10448 tcp_trace(TA_INPUT, ostate, tp,
10449 (void *)tcp_saveipgen, &tcp_savetcp, 0);
10451 newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10453 /* Add data to socket buffer. */
10454 SOCKBUF_LOCK(&so->so_rcv);
10455 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10459 * Set new socket buffer size. Give up when limit is
10463 if (!sbreserve_locked(&so->so_rcv,
10464 newsize, so, NULL))
10465 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10466 m_adj(m, drop_hdrlen); /* delayed header drop */
10467 #ifdef NETFLIX_SB_LIMITS
10470 sbappendstream_locked(&so->so_rcv, m, 0);
10471 ctf_calc_rwin(so, tp);
10473 rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10474 SOCKBUF_UNLOCK(&so->so_rcv);
10475 tp->t_flags |= TF_WAKESOR;
10476 #ifdef NETFLIX_SB_LIMITS
10477 if (so->so_rcv.sb_shlim && mcnt != appended)
10478 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10480 rack_handle_delayed_ack(tp, rack, tlen, 0);
10481 if (tp->snd_una == tp->snd_max)
10482 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10487 * This subfunction is used to try to highly optimize the
10488 * fast path. We again allow window updates that are
10489 * in sequence to remain in the fast-path. We also add
10490 * in the __predict's to attempt to help the compiler.
10491 * Note that if we return a 0, then we can *not* process
10492 * it and the caller should push the packet into the
10496 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10497 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10498 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10504 * The size of tcp_saveipgen must be the size of the max ip header,
10507 u_char tcp_saveipgen[IP6_HDR_LEN];
10508 struct tcphdr tcp_savetcp;
10511 int32_t under_pacing = 0;
10512 struct tcp_rack *rack;
10514 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10515 /* Old ack, behind (or duplicate to) the last one rcv'd */
10518 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10519 /* Above what we have sent? */
10522 if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10523 /* We are retransmitting */
10526 if (__predict_false(tiwin == 0)) {
10530 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10531 /* We need a SYN or a FIN, unlikely.. */
10534 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10535 /* Timestamp is behind .. old ack with seq wrap? */
10538 if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10539 /* Still recovering */
10542 rack = (struct tcp_rack *)tp->t_fb_ptr;
10543 if (rack->r_ctl.rc_sacked) {
10544 /* We have sack holes on our scoreboard */
10547 /* Ok if we reach here, we can process a fast-ack */
10548 if (rack->gp_ready &&
10549 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10552 nsegs = max(1, m->m_pkthdr.lro_nsegs);
10553 rack_log_ack(tp, to, th, 0, 0);
10554 /* Did the window get updated? */
10555 if (tiwin != tp->snd_wnd) {
10556 tp->snd_wnd = tiwin;
10557 rack_validate_fo_sendwin_up(tp, rack);
10558 tp->snd_wl1 = th->th_seq;
10559 if (tp->snd_wnd > tp->max_sndwnd)
10560 tp->max_sndwnd = tp->snd_wnd;
10562 /* Do we exit persists? */
10563 if ((rack->rc_in_persist != 0) &&
10564 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10565 rack->r_ctl.rc_pace_min_segs))) {
10566 rack_exit_persist(tp, rack, cts);
10568 /* Do we enter persists? */
10569 if ((rack->rc_in_persist == 0) &&
10570 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10571 TCPS_HAVEESTABLISHED(tp->t_state) &&
10572 (tp->snd_max == tp->snd_una) &&
10573 sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10574 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10576 * Here the rwnd is less than
10577 * the pacing size, we are established,
10578 * nothing is outstanding, and there is
10579 * data to send. Enter persists.
10581 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10584 * If last ACK falls within this segment's sequence numbers, record
10585 * the timestamp. NOTE that the test is modified according to the
10586 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10588 if ((to->to_flags & TOF_TS) != 0 &&
10589 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10590 tp->ts_recent_age = tcp_ts_getticks();
10591 tp->ts_recent = to->to_tsval;
10594 * This is a pure ack for outstanding data.
10596 KMOD_TCPSTAT_INC(tcps_predack);
10599 * "bad retransmit" recovery.
10601 if ((tp->t_flags & TF_PREVVALID) &&
10602 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10603 tp->t_flags &= ~TF_PREVVALID;
10604 if (tp->t_rxtshift == 1 &&
10605 (int)(ticks - tp->t_badrxtwin) < 0)
10606 rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10609 * Recalculate the transmit timer / rtt.
10611 * Some boxes send broken timestamp replies during the SYN+ACK
10612 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10613 * and blow up the retransmit timer.
10615 acked = BYTES_THIS_ACK(tp, th);
10618 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10619 hhook_run_tcp_est_in(tp, th, to);
10621 KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10622 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10624 struct mbuf *mfree;
10626 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10627 SOCKBUF_LOCK(&so->so_snd);
10628 mfree = sbcut_locked(&so->so_snd, acked);
10629 tp->snd_una = th->th_ack;
10630 /* Note we want to hold the sb lock through the sendmap adjust */
10631 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10632 /* Wake up the socket if we have room to write more */
10633 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10634 SOCKBUF_UNLOCK(&so->so_snd);
10635 tp->t_flags |= TF_WAKESOW;
10637 tp->t_rxtshift = 0;
10638 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10639 rack_rto_min, rack_rto_max);
10640 rack->rc_tlp_in_progress = 0;
10641 rack->r_ctl.rc_tlp_cnt_out = 0;
10643 * If it is the RXT timer we want to
10644 * stop it, so we can restart a TLP.
10646 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10647 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10648 #ifdef NETFLIX_HTTP_LOGGING
10649 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10653 * Let the congestion control algorithm update congestion control
10654 * related information. This typically means increasing the
10655 * congestion window.
10657 if (tp->snd_wnd < ctf_outstanding(tp)) {
10658 /* The peer collapsed the window */
10659 rack_collapsed_window(rack);
10660 } else if (rack->rc_has_collapsed)
10661 rack_un_collapse_window(rack);
10664 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10666 tp->snd_wl2 = th->th_ack;
10669 /* ND6_HINT(tp); *//* Some progress has been made. */
10672 * If all outstanding data are acked, stop retransmit timer,
10673 * otherwise restart timer using current (possibly backed-off)
10674 * value. If process is waiting for space, wakeup/selwakeup/signal.
10675 * If data are ready to send, let tcp_output decide between more
10676 * output or persist.
10679 if (so->so_options & SO_DEBUG)
10680 tcp_trace(TA_INPUT, ostate, tp,
10681 (void *)tcp_saveipgen,
10684 if (under_pacing &&
10685 (rack->use_fixed_rate == 0) &&
10686 (rack->in_probe_rtt == 0) &&
10687 rack->rc_gp_dyn_mul &&
10688 rack->rc_always_pace) {
10689 /* Check if we are dragging bottom */
10690 rack_check_bottom_drag(tp, rack, so, acked);
10692 if (tp->snd_una == tp->snd_max) {
10693 tp->t_flags &= ~TF_PREVVALID;
10694 rack->r_ctl.retran_during_recovery = 0;
10695 rack->r_ctl.dsack_byte_cnt = 0;
10696 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10697 if (rack->r_ctl.rc_went_idle_time == 0)
10698 rack->r_ctl.rc_went_idle_time = 1;
10699 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10700 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10702 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10704 if (acked && rack->r_fast_output)
10705 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10706 if (sbavail(&so->so_snd)) {
10707 rack->r_wanted_output = 1;
10713 * Return value of 1, the TCB is unlocked and most
10714 * likely gone, return value of 0, the TCP is still
10718 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10719 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10720 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10722 int32_t ret_val = 0;
10724 int32_t ourfinisacked = 0;
10725 struct tcp_rack *rack;
10727 ctf_calc_rwin(so, tp);
10729 * If the state is SYN_SENT: if seg contains an ACK, but not for our
10730 * SYN, drop the input. if seg contains a RST, then drop the
10731 * connection. if seg does not contain SYN, then drop it. Otherwise
10732 * this is an acceptable SYN segment initialize tp->rcv_nxt and
10733 * tp->irs if seg contains ack then advance tp->snd_una if seg
10734 * contains an ECE and ECN support is enabled, the stream is ECN
10735 * capable. if SYN has been acked change to ESTABLISHED else
10736 * SYN_RCVD state arrange for segment to be acked (eventually)
10737 * continue processing rest of data/controls.
10739 if ((thflags & TH_ACK) &&
10740 (SEQ_LEQ(th->th_ack, tp->iss) ||
10741 SEQ_GT(th->th_ack, tp->snd_max))) {
10742 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10743 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10746 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10747 TCP_PROBE5(connect__refused, NULL, tp,
10748 mtod(m, const char *), tp, th);
10749 tp = tcp_drop(tp, ECONNREFUSED);
10750 ctf_do_drop(m, tp);
10753 if (thflags & TH_RST) {
10754 ctf_do_drop(m, tp);
10757 if (!(thflags & TH_SYN)) {
10758 ctf_do_drop(m, tp);
10761 tp->irs = th->th_seq;
10762 tcp_rcvseqinit(tp);
10763 rack = (struct tcp_rack *)tp->t_fb_ptr;
10764 if (thflags & TH_ACK) {
10765 int tfo_partial = 0;
10767 KMOD_TCPSTAT_INC(tcps_connects);
10770 mac_socketpeer_set_from_mbuf(m, so);
10772 /* Do window scaling on this connection? */
10773 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10774 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10775 tp->rcv_scale = tp->request_r_scale;
10777 tp->rcv_adv += min(tp->rcv_wnd,
10778 TCP_MAXWIN << tp->rcv_scale);
10780 * If not all the data that was sent in the TFO SYN
10781 * has been acked, resend the remainder right away.
10783 if (IS_FASTOPEN(tp->t_flags) &&
10784 (tp->snd_una != tp->snd_max)) {
10785 tp->snd_nxt = th->th_ack;
10789 * If there's data, delay ACK; if there's also a FIN ACKNOW
10790 * will be turned on later.
10792 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10793 rack_timer_cancel(tp, rack,
10794 rack->r_ctl.rc_rcvtime, __LINE__);
10795 tp->t_flags |= TF_DELACK;
10797 rack->r_wanted_output = 1;
10798 tp->t_flags |= TF_ACKNOW;
10799 rack->rc_dack_toggle = 0;
10801 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10802 (V_tcp_do_ecn == 1)) {
10803 tp->t_flags2 |= TF2_ECN_PERMIT;
10804 KMOD_TCPSTAT_INC(tcps_ecn_shs);
10806 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10808 * We advance snd_una for the
10809 * fast open case. If th_ack is
10810 * acknowledging data beyond
10811 * snd_una we can't just call
10812 * ack-processing since the
10813 * data stream in our send-map
10814 * will start at snd_una + 1 (one
10815 * beyond the SYN). If its just
10816 * equal we don't need to do that
10817 * and there is no send_map.
10822 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10823 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10825 tp->t_starttime = ticks;
10826 if (tp->t_flags & TF_NEEDFIN) {
10827 tcp_state_change(tp, TCPS_FIN_WAIT_1);
10828 tp->t_flags &= ~TF_NEEDFIN;
10829 thflags &= ~TH_SYN;
10831 tcp_state_change(tp, TCPS_ESTABLISHED);
10832 TCP_PROBE5(connect__established, NULL, tp,
10833 mtod(m, const char *), tp, th);
10834 rack_cc_conn_init(tp);
10838 * Received initial SYN in SYN-SENT[*] state => simultaneous
10839 * open. If segment contains CC option and there is a
10840 * cached CC, apply TAO test. If it succeeds, connection is *
10841 * half-synchronized. Otherwise, do 3-way handshake:
10842 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10843 * there was no CC option, clear cached CC value.
10845 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10846 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10848 INP_WLOCK_ASSERT(tp->t_inpcb);
10850 * Advance th->th_seq to correspond to first data byte. If data,
10851 * trim to stay within window, dropping FIN if necessary.
10854 if (tlen > tp->rcv_wnd) {
10855 todrop = tlen - tp->rcv_wnd;
10857 tlen = tp->rcv_wnd;
10858 thflags &= ~TH_FIN;
10859 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10860 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10862 tp->snd_wl1 = th->th_seq - 1;
10863 tp->rcv_up = th->th_seq;
10865 * Client side of transaction: already sent SYN and data. If the
10866 * remote host used T/TCP to validate the SYN, our data will be
10867 * ACK'd; if so, enter normal data segment processing in the middle
10868 * of step 5, ack processing. Otherwise, goto step 6.
10870 if (thflags & TH_ACK) {
10871 /* For syn-sent we need to possibly update the rtt */
10872 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10875 mcts = tcp_ts_getticks();
10876 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10877 if (!tp->t_rttlow || tp->t_rttlow > t)
10879 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10880 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10881 tcp_rack_xmit_timer_commit(rack, tp);
10883 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10885 /* We may have changed to FIN_WAIT_1 above */
10886 if (tp->t_state == TCPS_FIN_WAIT_1) {
10888 * In FIN_WAIT_1 STATE in addition to the processing
10889 * for the ESTABLISHED state if our FIN is now
10890 * acknowledged then enter FIN_WAIT_2.
10892 if (ourfinisacked) {
10894 * If we can't receive any more data, then
10895 * closing user can proceed. Starting the
10896 * timer is contrary to the specification,
10897 * but if we don't get a FIN we'll hang
10900 * XXXjl: we should release the tp also, and
10901 * use a compressed state.
10903 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10904 soisdisconnected(so);
10905 tcp_timer_activate(tp, TT_2MSL,
10906 (tcp_fast_finwait2_recycle ?
10907 tcp_finwait2_timeout :
10910 tcp_state_change(tp, TCPS_FIN_WAIT_2);
10914 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10915 tiwin, thflags, nxt_pkt));
10919 * Return value of 1, the TCB is unlocked and most
10920 * likely gone, return value of 0, the TCP is still
10924 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
10925 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10926 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10928 struct tcp_rack *rack;
10929 int32_t ret_val = 0;
10930 int32_t ourfinisacked = 0;
10932 ctf_calc_rwin(so, tp);
10933 if ((thflags & TH_ACK) &&
10934 (SEQ_LEQ(th->th_ack, tp->snd_una) ||
10935 SEQ_GT(th->th_ack, tp->snd_max))) {
10936 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10937 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10940 rack = (struct tcp_rack *)tp->t_fb_ptr;
10941 if (IS_FASTOPEN(tp->t_flags)) {
10943 * When a TFO connection is in SYN_RECEIVED, the
10944 * only valid packets are the initial SYN, a
10945 * retransmit/copy of the initial SYN (possibly with
10946 * a subset of the original data), a valid ACK, a
10949 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
10950 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10951 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10953 } else if (thflags & TH_SYN) {
10954 /* non-initial SYN is ignored */
10955 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
10956 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
10957 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
10958 ctf_do_drop(m, NULL);
10961 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
10962 ctf_do_drop(m, NULL);
10966 if ((thflags & TH_RST) ||
10967 (tp->t_fin_is_rst && (thflags & TH_FIN)))
10968 return (ctf_process_rst(m, th, so, tp));
10970 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10971 * it's less than ts_recent, drop it.
10973 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10974 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10975 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10979 * In the SYN-RECEIVED state, validate that the packet belongs to
10980 * this connection before trimming the data to fit the receive
10981 * window. Check the sequence number versus IRS since we know the
10982 * sequence numbers haven't wrapped. This is a partial fix for the
10983 * "LAND" DoS attack.
10985 if (SEQ_LT(th->th_seq, tp->irs)) {
10986 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10987 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10990 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
10991 &rack->r_ctl.challenge_ack_ts,
10992 &rack->r_ctl.challenge_ack_cnt)) {
10996 * If last ACK falls within this segment's sequence numbers, record
10997 * its timestamp. NOTE: 1) That the test incorporates suggestions
10998 * from the latest proposal of the tcplw@cray.com list (Braden
10999 * 1993/04/26). 2) That updating only on newer timestamps interferes
11000 * with our earlier PAWS tests, so this check should be solely
11001 * predicated on the sequence space of this segment. 3) That we
11002 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11003 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11004 * SEG.Len, This modified check allows us to overcome RFC1323's
11005 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11006 * p.869. In such cases, we can still calculate the RTT correctly
11007 * when RCV.NXT == Last.ACK.Sent.
11009 if ((to->to_flags & TOF_TS) != 0 &&
11010 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11011 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11012 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11013 tp->ts_recent_age = tcp_ts_getticks();
11014 tp->ts_recent = to->to_tsval;
11016 tp->snd_wnd = tiwin;
11017 rack_validate_fo_sendwin_up(tp, rack);
11019 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11020 * is on (half-synchronized state), then queue data for later
11021 * processing; else drop segment and return.
11023 if ((thflags & TH_ACK) == 0) {
11024 if (IS_FASTOPEN(tp->t_flags)) {
11025 rack_cc_conn_init(tp);
11027 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11028 tiwin, thflags, nxt_pkt));
11030 KMOD_TCPSTAT_INC(tcps_connects);
11032 /* Do window scaling? */
11033 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11034 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11035 tp->rcv_scale = tp->request_r_scale;
11038 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* ->
11041 tp->t_starttime = ticks;
11042 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11043 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11044 tp->t_tfo_pending = NULL;
11046 if (tp->t_flags & TF_NEEDFIN) {
11047 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11048 tp->t_flags &= ~TF_NEEDFIN;
11050 tcp_state_change(tp, TCPS_ESTABLISHED);
11051 TCP_PROBE5(accept__established, NULL, tp,
11052 mtod(m, const char *), tp, th);
11054 * TFO connections call cc_conn_init() during SYN
11055 * processing. Calling it again here for such connections
11056 * is not harmless as it would undo the snd_cwnd reduction
11057 * that occurs when a TFO SYN|ACK is retransmitted.
11059 if (!IS_FASTOPEN(tp->t_flags))
11060 rack_cc_conn_init(tp);
11063 * Account for the ACK of our SYN prior to
11064 * regular ACK processing below, except for
11065 * simultaneous SYN, which is handled later.
11067 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11070 * If segment contains data or ACK, will call tcp_reass() later; if
11071 * not, do so now to pass queued data to user.
11073 if (tlen == 0 && (thflags & TH_FIN) == 0)
11074 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11076 tp->snd_wl1 = th->th_seq - 1;
11077 /* For syn-recv we need to possibly update the rtt */
11078 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11081 mcts = tcp_ts_getticks();
11082 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11083 if (!tp->t_rttlow || tp->t_rttlow > t)
11085 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11086 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11087 tcp_rack_xmit_timer_commit(rack, tp);
11089 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11092 if (tp->t_state == TCPS_FIN_WAIT_1) {
11093 /* We could have went to FIN_WAIT_1 (or EST) above */
11095 * In FIN_WAIT_1 STATE in addition to the processing for the
11096 * ESTABLISHED state if our FIN is now acknowledged then
11097 * enter FIN_WAIT_2.
11099 if (ourfinisacked) {
11101 * If we can't receive any more data, then closing
11102 * user can proceed. Starting the timer is contrary
11103 * to the specification, but if we don't get a FIN
11104 * we'll hang forever.
11106 * XXXjl: we should release the tp also, and use a
11107 * compressed state.
11109 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11110 soisdisconnected(so);
11111 tcp_timer_activate(tp, TT_2MSL,
11112 (tcp_fast_finwait2_recycle ?
11113 tcp_finwait2_timeout :
11116 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11119 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11120 tiwin, thflags, nxt_pkt));
11124 * Return value of 1, the TCB is unlocked and most
11125 * likely gone, return value of 0, the TCP is still
11129 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11130 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11131 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11133 int32_t ret_val = 0;
11134 struct tcp_rack *rack;
11137 * Header prediction: check for the two common cases of a
11138 * uni-directional data xfer. If the packet has no control flags,
11139 * is in-sequence, the window didn't change and we're not
11140 * retransmitting, it's a candidate. If the length is zero and the
11141 * ack moved forward, we're the sender side of the xfer. Just free
11142 * the data acked & wake any higher level process that was blocked
11143 * waiting for space. If the length is non-zero and the ack didn't
11144 * move, we're the receiver side. If we're getting packets in-order
11145 * (the reassembly queue is empty), add the data toc The socket
11146 * buffer and note that we need a delayed ack. Make sure that the
11147 * hidden state-flags are also off. Since we check for
11148 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11150 rack = (struct tcp_rack *)tp->t_fb_ptr;
11151 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11152 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11153 __predict_true(SEGQ_EMPTY(tp)) &&
11154 __predict_true(th->th_seq == tp->rcv_nxt)) {
11156 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11157 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11161 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11162 tiwin, nxt_pkt, iptos)) {
11167 ctf_calc_rwin(so, tp);
11169 if ((thflags & TH_RST) ||
11170 (tp->t_fin_is_rst && (thflags & TH_FIN)))
11171 return (ctf_process_rst(m, th, so, tp));
11174 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11175 * synchronized state.
11177 if (thflags & TH_SYN) {
11178 ctf_challenge_ack(m, th, tp, &ret_val);
11182 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11183 * it's less than ts_recent, drop it.
11185 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11186 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11187 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11190 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11191 &rack->r_ctl.challenge_ack_ts,
11192 &rack->r_ctl.challenge_ack_cnt)) {
11196 * If last ACK falls within this segment's sequence numbers, record
11197 * its timestamp. NOTE: 1) That the test incorporates suggestions
11198 * from the latest proposal of the tcplw@cray.com list (Braden
11199 * 1993/04/26). 2) That updating only on newer timestamps interferes
11200 * with our earlier PAWS tests, so this check should be solely
11201 * predicated on the sequence space of this segment. 3) That we
11202 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11203 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11204 * SEG.Len, This modified check allows us to overcome RFC1323's
11205 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11206 * p.869. In such cases, we can still calculate the RTT correctly
11207 * when RCV.NXT == Last.ACK.Sent.
11209 if ((to->to_flags & TOF_TS) != 0 &&
11210 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11211 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11212 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11213 tp->ts_recent_age = tcp_ts_getticks();
11214 tp->ts_recent = to->to_tsval;
11217 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11218 * is on (half-synchronized state), then queue data for later
11219 * processing; else drop segment and return.
11221 if ((thflags & TH_ACK) == 0) {
11222 if (tp->t_flags & TF_NEEDSYN) {
11223 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11224 tiwin, thflags, nxt_pkt));
11226 } else if (tp->t_flags & TF_ACKNOW) {
11227 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11228 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11231 ctf_do_drop(m, NULL);
11238 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11241 if (sbavail(&so->so_snd)) {
11242 if (ctf_progress_timeout_check(tp, true)) {
11243 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11244 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11245 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11249 /* State changes only happen in rack_process_data() */
11250 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11251 tiwin, thflags, nxt_pkt));
11255 * Return value of 1, the TCB is unlocked and most
11256 * likely gone, return value of 0, the TCP is still
11260 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11261 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11262 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11264 int32_t ret_val = 0;
11265 struct tcp_rack *rack;
11267 rack = (struct tcp_rack *)tp->t_fb_ptr;
11268 ctf_calc_rwin(so, tp);
11269 if ((thflags & TH_RST) ||
11270 (tp->t_fin_is_rst && (thflags & TH_FIN)))
11271 return (ctf_process_rst(m, th, so, tp));
11273 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11274 * synchronized state.
11276 if (thflags & TH_SYN) {
11277 ctf_challenge_ack(m, th, tp, &ret_val);
11281 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11282 * it's less than ts_recent, drop it.
11284 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11285 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11286 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11289 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11290 &rack->r_ctl.challenge_ack_ts,
11291 &rack->r_ctl.challenge_ack_cnt)) {
11295 * If last ACK falls within this segment's sequence numbers, record
11296 * its timestamp. NOTE: 1) That the test incorporates suggestions
11297 * from the latest proposal of the tcplw@cray.com list (Braden
11298 * 1993/04/26). 2) That updating only on newer timestamps interferes
11299 * with our earlier PAWS tests, so this check should be solely
11300 * predicated on the sequence space of this segment. 3) That we
11301 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11302 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11303 * SEG.Len, This modified check allows us to overcome RFC1323's
11304 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11305 * p.869. In such cases, we can still calculate the RTT correctly
11306 * when RCV.NXT == Last.ACK.Sent.
11308 if ((to->to_flags & TOF_TS) != 0 &&
11309 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11310 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11311 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11312 tp->ts_recent_age = tcp_ts_getticks();
11313 tp->ts_recent = to->to_tsval;
11316 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11317 * is on (half-synchronized state), then queue data for later
11318 * processing; else drop segment and return.
11320 if ((thflags & TH_ACK) == 0) {
11321 if (tp->t_flags & TF_NEEDSYN) {
11322 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11323 tiwin, thflags, nxt_pkt));
11325 } else if (tp->t_flags & TF_ACKNOW) {
11326 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11327 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11330 ctf_do_drop(m, NULL);
11337 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11340 if (sbavail(&so->so_snd)) {
11341 if (ctf_progress_timeout_check(tp, true)) {
11342 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11343 tp, tick, PROGRESS_DROP, __LINE__);
11344 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11345 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11349 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11350 tiwin, thflags, nxt_pkt));
11354 rack_check_data_after_close(struct mbuf *m,
11355 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11357 struct tcp_rack *rack;
11359 rack = (struct tcp_rack *)tp->t_fb_ptr;
11360 if (rack->rc_allow_data_af_clo == 0) {
11362 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11363 /* tcp_close will kill the inp pre-log the Reset */
11364 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11365 tp = tcp_close(tp);
11366 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11367 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11370 if (sbavail(&so->so_snd) == 0)
11372 /* Ok we allow data that is ignored and a followup reset */
11373 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11374 tp->rcv_nxt = th->th_seq + *tlen;
11375 tp->t_flags2 |= TF2_DROP_AF_DATA;
11376 rack->r_wanted_output = 1;
11382 * Return value of 1, the TCB is unlocked and most
11383 * likely gone, return value of 0, the TCP is still
11387 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11388 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11389 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11391 int32_t ret_val = 0;
11392 int32_t ourfinisacked = 0;
11393 struct tcp_rack *rack;
11395 rack = (struct tcp_rack *)tp->t_fb_ptr;
11396 ctf_calc_rwin(so, tp);
11398 if ((thflags & TH_RST) ||
11399 (tp->t_fin_is_rst && (thflags & TH_FIN)))
11400 return (ctf_process_rst(m, th, so, tp));
11402 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11403 * synchronized state.
11405 if (thflags & TH_SYN) {
11406 ctf_challenge_ack(m, th, tp, &ret_val);
11410 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11411 * it's less than ts_recent, drop it.
11413 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11414 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11415 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11418 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11419 &rack->r_ctl.challenge_ack_ts,
11420 &rack->r_ctl.challenge_ack_cnt)) {
11424 * If new data are received on a connection after the user processes
11425 * are gone, then RST the other end.
11427 if ((so->so_state & SS_NOFDREF) && tlen) {
11428 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11432 * If last ACK falls within this segment's sequence numbers, record
11433 * its timestamp. NOTE: 1) That the test incorporates suggestions
11434 * from the latest proposal of the tcplw@cray.com list (Braden
11435 * 1993/04/26). 2) That updating only on newer timestamps interferes
11436 * with our earlier PAWS tests, so this check should be solely
11437 * predicated on the sequence space of this segment. 3) That we
11438 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11439 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11440 * SEG.Len, This modified check allows us to overcome RFC1323's
11441 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11442 * p.869. In such cases, we can still calculate the RTT correctly
11443 * when RCV.NXT == Last.ACK.Sent.
11445 if ((to->to_flags & TOF_TS) != 0 &&
11446 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11447 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11448 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11449 tp->ts_recent_age = tcp_ts_getticks();
11450 tp->ts_recent = to->to_tsval;
11453 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11454 * is on (half-synchronized state), then queue data for later
11455 * processing; else drop segment and return.
11457 if ((thflags & TH_ACK) == 0) {
11458 if (tp->t_flags & TF_NEEDSYN) {
11459 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11460 tiwin, thflags, nxt_pkt));
11461 } else if (tp->t_flags & TF_ACKNOW) {
11462 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11463 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11466 ctf_do_drop(m, NULL);
11473 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11476 if (ourfinisacked) {
11478 * If we can't receive any more data, then closing user can
11479 * proceed. Starting the timer is contrary to the
11480 * specification, but if we don't get a FIN we'll hang
11483 * XXXjl: we should release the tp also, and use a
11484 * compressed state.
11486 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11487 soisdisconnected(so);
11488 tcp_timer_activate(tp, TT_2MSL,
11489 (tcp_fast_finwait2_recycle ?
11490 tcp_finwait2_timeout :
11493 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11495 if (sbavail(&so->so_snd)) {
11496 if (ctf_progress_timeout_check(tp, true)) {
11497 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11498 tp, tick, PROGRESS_DROP, __LINE__);
11499 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11500 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11504 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11505 tiwin, thflags, nxt_pkt));
11509 * Return value of 1, the TCB is unlocked and most
11510 * likely gone, return value of 0, the TCP is still
11514 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11515 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11516 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11518 int32_t ret_val = 0;
11519 int32_t ourfinisacked = 0;
11520 struct tcp_rack *rack;
11522 rack = (struct tcp_rack *)tp->t_fb_ptr;
11523 ctf_calc_rwin(so, tp);
11525 if ((thflags & TH_RST) ||
11526 (tp->t_fin_is_rst && (thflags & TH_FIN)))
11527 return (ctf_process_rst(m, th, so, tp));
11529 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11530 * synchronized state.
11532 if (thflags & TH_SYN) {
11533 ctf_challenge_ack(m, th, tp, &ret_val);
11537 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11538 * it's less than ts_recent, drop it.
11540 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11541 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11542 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11545 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11546 &rack->r_ctl.challenge_ack_ts,
11547 &rack->r_ctl.challenge_ack_cnt)) {
11551 * If new data are received on a connection after the user processes
11552 * are gone, then RST the other end.
11554 if ((so->so_state & SS_NOFDREF) && tlen) {
11555 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11559 * If last ACK falls within this segment's sequence numbers, record
11560 * its timestamp. NOTE: 1) That the test incorporates suggestions
11561 * from the latest proposal of the tcplw@cray.com list (Braden
11562 * 1993/04/26). 2) That updating only on newer timestamps interferes
11563 * with our earlier PAWS tests, so this check should be solely
11564 * predicated on the sequence space of this segment. 3) That we
11565 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11566 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11567 * SEG.Len, This modified check allows us to overcome RFC1323's
11568 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11569 * p.869. In such cases, we can still calculate the RTT correctly
11570 * when RCV.NXT == Last.ACK.Sent.
11572 if ((to->to_flags & TOF_TS) != 0 &&
11573 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11574 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11575 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11576 tp->ts_recent_age = tcp_ts_getticks();
11577 tp->ts_recent = to->to_tsval;
11580 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11581 * is on (half-synchronized state), then queue data for later
11582 * processing; else drop segment and return.
11584 if ((thflags & TH_ACK) == 0) {
11585 if (tp->t_flags & TF_NEEDSYN) {
11586 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11587 tiwin, thflags, nxt_pkt));
11588 } else if (tp->t_flags & TF_ACKNOW) {
11589 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11590 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11593 ctf_do_drop(m, NULL);
11600 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11603 if (ourfinisacked) {
11608 if (sbavail(&so->so_snd)) {
11609 if (ctf_progress_timeout_check(tp, true)) {
11610 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11611 tp, tick, PROGRESS_DROP, __LINE__);
11612 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11613 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11617 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11618 tiwin, thflags, nxt_pkt));
11622 * Return value of 1, the TCB is unlocked and most
11623 * likely gone, return value of 0, the TCP is still
11627 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11628 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11629 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11631 int32_t ret_val = 0;
11632 int32_t ourfinisacked = 0;
11633 struct tcp_rack *rack;
11635 rack = (struct tcp_rack *)tp->t_fb_ptr;
11636 ctf_calc_rwin(so, tp);
11638 if ((thflags & TH_RST) ||
11639 (tp->t_fin_is_rst && (thflags & TH_FIN)))
11640 return (ctf_process_rst(m, th, so, tp));
11642 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11643 * synchronized state.
11645 if (thflags & TH_SYN) {
11646 ctf_challenge_ack(m, th, tp, &ret_val);
11650 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11651 * it's less than ts_recent, drop it.
11653 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11654 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11655 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11658 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11659 &rack->r_ctl.challenge_ack_ts,
11660 &rack->r_ctl.challenge_ack_cnt)) {
11664 * If new data are received on a connection after the user processes
11665 * are gone, then RST the other end.
11667 if ((so->so_state & SS_NOFDREF) && tlen) {
11668 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11672 * If last ACK falls within this segment's sequence numbers, record
11673 * its timestamp. NOTE: 1) That the test incorporates suggestions
11674 * from the latest proposal of the tcplw@cray.com list (Braden
11675 * 1993/04/26). 2) That updating only on newer timestamps interferes
11676 * with our earlier PAWS tests, so this check should be solely
11677 * predicated on the sequence space of this segment. 3) That we
11678 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11679 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11680 * SEG.Len, This modified check allows us to overcome RFC1323's
11681 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11682 * p.869. In such cases, we can still calculate the RTT correctly
11683 * when RCV.NXT == Last.ACK.Sent.
11685 if ((to->to_flags & TOF_TS) != 0 &&
11686 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11687 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11688 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11689 tp->ts_recent_age = tcp_ts_getticks();
11690 tp->ts_recent = to->to_tsval;
11693 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11694 * is on (half-synchronized state), then queue data for later
11695 * processing; else drop segment and return.
11697 if ((thflags & TH_ACK) == 0) {
11698 if (tp->t_flags & TF_NEEDSYN) {
11699 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11700 tiwin, thflags, nxt_pkt));
11701 } else if (tp->t_flags & TF_ACKNOW) {
11702 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11703 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11706 ctf_do_drop(m, NULL);
11711 * case TCPS_LAST_ACK: Ack processing.
11713 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11716 if (ourfinisacked) {
11717 tp = tcp_close(tp);
11718 ctf_do_drop(m, tp);
11721 if (sbavail(&so->so_snd)) {
11722 if (ctf_progress_timeout_check(tp, true)) {
11723 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11724 tp, tick, PROGRESS_DROP, __LINE__);
11725 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11726 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11730 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11731 tiwin, thflags, nxt_pkt));
11735 * Return value of 1, the TCB is unlocked and most
11736 * likely gone, return value of 0, the TCP is still
11740 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11741 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11742 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11744 int32_t ret_val = 0;
11745 int32_t ourfinisacked = 0;
11746 struct tcp_rack *rack;
11748 rack = (struct tcp_rack *)tp->t_fb_ptr;
11749 ctf_calc_rwin(so, tp);
11751 /* Reset receive buffer auto scaling when not in bulk receive mode. */
11752 if ((thflags & TH_RST) ||
11753 (tp->t_fin_is_rst && (thflags & TH_FIN)))
11754 return (ctf_process_rst(m, th, so, tp));
11756 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11757 * synchronized state.
11759 if (thflags & TH_SYN) {
11760 ctf_challenge_ack(m, th, tp, &ret_val);
11764 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11765 * it's less than ts_recent, drop it.
11767 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11768 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11769 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11772 if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11773 &rack->r_ctl.challenge_ack_ts,
11774 &rack->r_ctl.challenge_ack_cnt)) {
11778 * If new data are received on a connection after the user processes
11779 * are gone, then RST the other end.
11781 if ((so->so_state & SS_NOFDREF) &&
11783 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11787 * If last ACK falls within this segment's sequence numbers, record
11788 * its timestamp. NOTE: 1) That the test incorporates suggestions
11789 * from the latest proposal of the tcplw@cray.com list (Braden
11790 * 1993/04/26). 2) That updating only on newer timestamps interferes
11791 * with our earlier PAWS tests, so this check should be solely
11792 * predicated on the sequence space of this segment. 3) That we
11793 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11794 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11795 * SEG.Len, This modified check allows us to overcome RFC1323's
11796 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11797 * p.869. In such cases, we can still calculate the RTT correctly
11798 * when RCV.NXT == Last.ACK.Sent.
11800 if ((to->to_flags & TOF_TS) != 0 &&
11801 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11802 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11803 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11804 tp->ts_recent_age = tcp_ts_getticks();
11805 tp->ts_recent = to->to_tsval;
11808 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
11809 * is on (half-synchronized state), then queue data for later
11810 * processing; else drop segment and return.
11812 if ((thflags & TH_ACK) == 0) {
11813 if (tp->t_flags & TF_NEEDSYN) {
11814 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11815 tiwin, thflags, nxt_pkt));
11816 } else if (tp->t_flags & TF_ACKNOW) {
11817 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11818 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11821 ctf_do_drop(m, NULL);
11828 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11831 if (sbavail(&so->so_snd)) {
11832 if (ctf_progress_timeout_check(tp, true)) {
11833 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11834 tp, tick, PROGRESS_DROP, __LINE__);
11835 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11836 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11840 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11841 tiwin, thflags, nxt_pkt));
11845 rack_clear_rate_sample(struct tcp_rack *rack)
11847 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11848 rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11849 rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11853 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11855 uint64_t bw_est, rate_wanted;
11857 uint32_t user_max, orig_min, orig_max;
11859 orig_min = rack->r_ctl.rc_pace_min_segs;
11860 orig_max = rack->r_ctl.rc_pace_max_segs;
11861 user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11862 if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11864 rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11865 if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11866 if (user_max != rack->r_ctl.rc_pace_max_segs)
11869 if (rack->rc_force_max_seg) {
11870 rack->r_ctl.rc_pace_max_segs = user_max;
11871 } else if (rack->use_fixed_rate) {
11872 bw_est = rack_get_bw(rack);
11873 if ((rack->r_ctl.crte == NULL) ||
11874 (bw_est != rack->r_ctl.crte->rate)) {
11875 rack->r_ctl.rc_pace_max_segs = user_max;
11877 /* We are pacing right at the hardware rate */
11880 segsiz = min(ctf_fixed_maxseg(tp),
11881 rack->r_ctl.rc_pace_min_segs);
11882 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11883 tp, bw_est, segsiz, 0,
11884 rack->r_ctl.crte, NULL);
11886 } else if (rack->rc_always_pace) {
11887 if (rack->r_ctl.gp_bw ||
11888 #ifdef NETFLIX_PEAKRATE
11889 rack->rc_tp->t_maxpeakrate ||
11891 rack->r_ctl.init_rate) {
11892 /* We have a rate of some sort set */
11895 bw_est = rack_get_bw(rack);
11896 orig = rack->r_ctl.rc_pace_max_segs;
11898 rate_wanted = *fill_override;
11900 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
11902 /* We have something */
11903 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
11905 ctf_fixed_maxseg(rack->rc_tp));
11907 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
11908 if (orig != rack->r_ctl.rc_pace_max_segs)
11910 } else if ((rack->r_ctl.gp_bw == 0) &&
11911 (rack->r_ctl.rc_pace_max_segs == 0)) {
11913 * If we have nothing limit us to bursting
11914 * out IW sized pieces.
11917 rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
11920 if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
11922 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
11925 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
11930 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
11933 struct ip6_hdr *ip6 = NULL;
11936 struct ip *ip = NULL;
11938 struct udphdr *udp = NULL;
11940 /* Ok lets fill in the fast block, it can only be used with no IP options! */
11942 if (rack->r_is_v6) {
11943 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
11944 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
11946 rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11947 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
11948 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11949 udp->uh_dport = tp->t_port;
11950 rack->r_ctl.fsb.udp = udp;
11951 rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11954 rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
11955 rack->r_ctl.fsb.udp = NULL;
11957 tcpip_fillheaders(rack->rc_inp,
11959 ip6, rack->r_ctl.fsb.th);
11963 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
11964 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
11966 rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11967 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
11968 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11969 udp->uh_dport = tp->t_port;
11970 rack->r_ctl.fsb.udp = udp;
11971 rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11974 rack->r_ctl.fsb.udp = NULL;
11975 rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
11977 tcpip_fillheaders(rack->rc_inp,
11979 ip, rack->r_ctl.fsb.th);
11981 rack->r_fsb_inited = 1;
11985 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
11988 * Allocate the larger of spaces V6 if available else just
11989 * V4 and include udphdr (overbook)
11992 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
11994 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
11996 rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
11997 M_TCPFSB, M_NOWAIT|M_ZERO);
11998 if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12001 rack->r_fsb_inited = 0;
12006 rack_init(struct tcpcb *tp)
12008 struct tcp_rack *rack = NULL;
12009 struct rack_sendmap *insret;
12010 uint32_t iwin, snt, us_cts;
12013 tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12014 if (tp->t_fb_ptr == NULL) {
12016 * We need to allocate memory but cant. The INP and INP_INFO
12017 * locks and they are recusive (happens during setup. So a
12018 * scheme to drop the locks fails :(
12023 memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12025 rack = (struct tcp_rack *)tp->t_fb_ptr;
12026 RB_INIT(&rack->r_ctl.rc_mtree);
12027 TAILQ_INIT(&rack->r_ctl.rc_free);
12028 TAILQ_INIT(&rack->r_ctl.rc_tmap);
12030 rack->rc_inp = tp->t_inpcb;
12032 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12033 /* Probably not needed but lets be sure */
12034 rack_clear_rate_sample(rack);
12036 * Save off the default values, socket options will poke
12037 * at these if pacing is not on or we have not yet
12038 * reached where pacing is on (gp_ready/fixed enabled).
12039 * When they get set into the CC module (when gp_ready
12040 * is enabled or we enable fixed) then we will set these
12041 * values into the CC and place in here the old values
12042 * so we have a restoral. Then we will set the flag
12043 * rc_pacing_cc_set. That way whenever we turn off pacing
12044 * or switch off this stack, we will know to go restore
12045 * the saved values.
12047 rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12048 rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12049 /* We want abe like behavior as well */
12050 rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12051 rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12052 rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12053 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12055 rack->use_rack_rr = 1;
12056 if (V_tcp_delack_enabled)
12057 tp->t_delayed_ack = 1;
12059 tp->t_delayed_ack = 0;
12060 #ifdef TCP_ACCOUNTING
12061 if (rack_tcp_accounting) {
12062 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12065 if (rack_enable_shared_cwnd)
12066 rack->rack_enable_scwnd = 1;
12067 rack->rc_user_set_max_segs = rack_hptsi_segments;
12068 rack->rc_force_max_seg = 0;
12069 if (rack_use_imac_dack)
12070 rack->rc_dack_mode = 1;
12071 TAILQ_INIT(&rack->r_ctl.opt_list);
12072 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12073 rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12074 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12075 rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12076 rack->r_ctl.rc_highest_us_rtt = 0;
12077 rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12078 if (rack_use_cmp_acks)
12079 rack->r_use_cmp_ack = 1;
12080 if (rack_disable_prr)
12081 rack->rack_no_prr = 1;
12082 if (rack_gp_no_rec_chg)
12083 rack->rc_gp_no_rec_chg = 1;
12084 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12085 rack->rc_always_pace = 1;
12086 if (rack->use_fixed_rate || rack->gp_ready)
12087 rack_set_cc_pacing(rack);
12089 rack->rc_always_pace = 0;
12090 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12091 rack->r_mbuf_queue = 1;
12093 rack->r_mbuf_queue = 0;
12094 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12095 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12097 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12098 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12099 if (rack_limits_scwnd)
12100 rack->r_limit_scw = 1;
12102 rack->r_limit_scw = 0;
12103 rack->rc_labc = V_tcp_abc_l_var;
12104 rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12105 rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12106 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12107 rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12108 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12109 rack->r_ctl.rc_min_to = rack_min_to;
12110 microuptime(&rack->r_ctl.act_rcv_time);
12111 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12112 rack->r_running_late = 0;
12113 rack->r_running_early = 0;
12114 rack->rc_init_win = rack_default_init_window;
12115 rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12116 if (rack_hw_up_only)
12117 rack->r_up_only = 1;
12118 if (rack_do_dyn_mul) {
12119 /* When dynamic adjustment is on CA needs to start at 100% */
12120 rack->rc_gp_dyn_mul = 1;
12121 if (rack_do_dyn_mul >= 100)
12122 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12124 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12125 rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12126 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12127 rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12128 setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12129 rack_probertt_filter_life);
12130 us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12131 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12132 rack->r_ctl.rc_time_of_last_probertt = us_cts;
12133 rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12134 rack->r_ctl.rc_time_probertt_starts = 0;
12135 /* We require at least one measurement, even if the sysctl is 0 */
12136 if (rack_req_measurements)
12137 rack->r_ctl.req_measurements = rack_req_measurements;
12139 rack->r_ctl.req_measurements = 1;
12140 if (rack_enable_hw_pacing)
12141 rack->rack_hdw_pace_ena = 1;
12142 if (rack_hw_rate_caps)
12143 rack->r_rack_hw_rate_caps = 1;
12144 /* Do we force on detection? */
12145 #ifdef NETFLIX_EXP_DETECTION
12146 if (tcp_force_detection)
12147 rack->do_detection = 1;
12150 rack->do_detection = 0;
12151 if (rack_non_rxt_use_cr)
12152 rack->rack_rec_nonrxt_use_cr = 1;
12153 err = rack_init_fsb(tp, rack);
12155 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12156 tp->t_fb_ptr = NULL;
12159 if (tp->snd_una != tp->snd_max) {
12160 /* Create a send map for the current outstanding data */
12161 struct rack_sendmap *rsm;
12163 rsm = rack_alloc(rack);
12165 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12166 tp->t_fb_ptr = NULL;
12169 rsm->r_no_rtt_allowed = 1;
12170 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12171 rsm->r_rtr_cnt = 1;
12172 rsm->r_rtr_bytes = 0;
12173 if (tp->t_flags & TF_SENTFIN) {
12174 rsm->r_end = tp->snd_max - 1;
12175 rsm->r_flags |= RACK_HAS_FIN;
12177 rsm->r_end = tp->snd_max;
12179 if (tp->snd_una == tp->iss) {
12180 /* The data space is one beyond snd_una */
12181 rsm->r_flags |= RACK_HAS_SYN;
12182 rsm->r_start = tp->iss;
12183 rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12185 rsm->r_start = tp->snd_una;
12187 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12188 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12189 rsm->orig_m_len = rsm->m->m_len;
12192 * This can happen if we have a stand-alone FIN or
12196 rsm->orig_m_len = 0;
12199 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12201 if (insret != NULL) {
12202 panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12203 insret, rack, rsm);
12206 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12207 rsm->r_in_tmap = 1;
12210 * Timers in Rack are kept in microseconds so lets
12211 * convert any initial incoming variables
12212 * from ticks into usecs. Note that we
12213 * also change the values of t_srtt and t_rttvar, if
12214 * they are non-zero. They are kept with a 5
12215 * bit decimal so we have to carefully convert
12216 * these to get the full precision.
12218 rack_convert_rtts(tp);
12219 tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12220 if (rack_def_profile)
12221 rack_set_profile(rack, rack_def_profile);
12222 /* Cancel the GP measurement in progress */
12223 tp->t_flags &= ~TF_GPUTINPROG;
12224 if (SEQ_GT(tp->snd_max, tp->iss))
12225 snt = tp->snd_max - tp->iss;
12228 iwin = rc_init_window(rack);
12230 /* We are not past the initial window
12231 * so we need to make sure cwnd is
12234 if (tp->snd_cwnd < iwin)
12235 tp->snd_cwnd = iwin;
12237 * If we are within the initial window
12238 * we want ssthresh to be unlimited. Setting
12239 * it to the rwnd (which the default stack does
12240 * and older racks) is not really a good idea
12241 * since we want to be in SS and grow both the
12242 * cwnd and the rwnd (via dynamic rwnd growth). If
12243 * we set it to the rwnd then as the peer grows its
12244 * rwnd we will be stuck in CA and never hit SS.
12246 * Its far better to raise it up high (this takes the
12247 * risk that there as been a loss already, probably
12248 * we should have an indicator in all stacks of loss
12249 * but we don't), but considering the normal use this
12250 * is a risk worth taking. The consequences of not
12251 * hitting SS are far worse than going one more time
12252 * into it early on (before we have sent even a IW).
12253 * It is highly unlikely that we will have had a loss
12254 * before getting the IW out.
12256 tp->snd_ssthresh = 0xffffffff;
12258 rack_stop_all_timers(tp);
12259 /* Lets setup the fsb block */
12260 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12261 rack_log_rtt_shrinks(rack, us_cts, tp->t_rxtcur,
12262 __LINE__, RACK_RTTS_INIT);
12267 rack_handoff_ok(struct tcpcb *tp)
12269 if ((tp->t_state == TCPS_CLOSED) ||
12270 (tp->t_state == TCPS_LISTEN)) {
12271 /* Sure no problem though it may not stick */
12274 if ((tp->t_state == TCPS_SYN_SENT) ||
12275 (tp->t_state == TCPS_SYN_RECEIVED)) {
12277 * We really don't know if you support sack,
12278 * you have to get to ESTAB or beyond to tell.
12282 if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12284 * Rack will only send a FIN after all data is acknowledged.
12285 * So in this case we have more data outstanding. We can't
12286 * switch stacks until either all data and only the FIN
12287 * is left (in which case rack_init() now knows how
12288 * to deal with that) <or> all is acknowledged and we
12289 * are only left with incoming data, though why you
12290 * would want to switch to rack after all data is acknowledged
12291 * I have no idea (rrs)!
12295 if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12299 * If we reach here we don't do SACK on this connection so we can
12307 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12311 if (tp->t_fb_ptr) {
12312 struct tcp_rack *rack;
12313 struct rack_sendmap *rsm, *nrsm, *rm;
12315 rack = (struct tcp_rack *)tp->t_fb_ptr;
12316 if (tp->t_in_pkt) {
12318 * Since we are switching we need to process any
12319 * inbound packets in case a compressed ack is
12320 * in queue or the new stack does not support
12321 * mbuf queuing. These packets in theory should
12322 * have been handled by the old stack anyway.
12324 if ((rack->rc_inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT)) ||
12325 (rack->rc_inp->inp_flags2 & INP_FREED)) {
12326 /* Kill all the packets */
12327 struct mbuf *save, *m;
12330 tp->t_in_pkt = NULL;
12331 tp->t_tail_pkt = NULL;
12333 save = m->m_nextpkt;
12334 m->m_nextpkt = NULL;
12339 /* Process all the packets */
12340 ctf_do_queued_segments(rack->rc_inp->inp_socket, rack->rc_tp, 0);
12342 if ((tp->t_inpcb) &&
12343 (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12346 /* Total if we used large or small (if ack-cmp was used). */
12347 if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12348 counter_u64_add(rack_large_ackcmp, 1);
12350 counter_u64_add(rack_small_ackcmp, 1);
12353 tp->t_flags &= ~TF_FORCEDATA;
12354 #ifdef NETFLIX_SHARED_CWND
12355 if (rack->r_ctl.rc_scw) {
12358 if (rack->r_limit_scw)
12359 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12362 tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12363 rack->r_ctl.rc_scw_index,
12365 rack->r_ctl.rc_scw = NULL;
12368 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12369 free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12370 rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12371 rack->r_ctl.fsb.th = NULL;
12373 /* Convert back to ticks, with */
12374 if (tp->t_srtt > 1) {
12375 uint32_t val, frac;
12377 val = USEC_2_TICKS(tp->t_srtt);
12378 frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12379 tp->t_srtt = val << TCP_RTT_SHIFT;
12381 * frac is the fractional part here is left
12382 * over from converting to hz and shifting.
12383 * We need to convert this to the 5 bit
12388 frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12390 frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12392 tp->t_srtt += frac;
12395 if (tp->t_rttvar) {
12396 uint32_t val, frac;
12398 val = USEC_2_TICKS(tp->t_rttvar);
12399 frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12400 tp->t_rttvar = val << TCP_RTTVAR_SHIFT;
12402 * frac is the fractional part here is left
12403 * over from converting to hz and shifting.
12404 * We need to convert this to the 5 bit
12409 frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12411 frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12413 tp->t_rttvar += frac;
12416 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12417 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12418 if (rack->rc_always_pace) {
12419 tcp_decrement_paced_conn();
12420 rack_undo_cc_pacing(rack);
12421 rack->rc_always_pace = 0;
12423 /* Clean up any options if they were not applied */
12424 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12425 struct deferred_opt_list *dol;
12427 dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12428 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12429 free(dol, M_TCPDO);
12431 /* rack does not use force data but other stacks may clear it */
12432 if (rack->r_ctl.crte != NULL) {
12433 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12434 rack->rack_hdrw_pacing = 0;
12435 rack->r_ctl.crte = NULL;
12437 #ifdef TCP_BLACKBOX
12438 tcp_log_flowend(tp);
12440 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12441 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12444 panic("At fini, rack:%p rsm:%p rm:%p",
12448 uma_zfree(rack_zone, rsm);
12450 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12452 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12453 uma_zfree(rack_zone, rsm);
12454 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12456 rack->rc_free_cnt = 0;
12457 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12458 tp->t_fb_ptr = NULL;
12461 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12462 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12463 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12464 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12465 /* Cancel the GP measurement in progress */
12466 tp->t_flags &= ~TF_GPUTINPROG;
12467 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12469 /* Make sure snd_nxt is correctly set */
12470 tp->snd_nxt = tp->snd_max;
12474 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12476 switch (tp->t_state) {
12477 case TCPS_SYN_SENT:
12478 rack->r_state = TCPS_SYN_SENT;
12479 rack->r_substate = rack_do_syn_sent;
12481 case TCPS_SYN_RECEIVED:
12482 rack->r_state = TCPS_SYN_RECEIVED;
12483 rack->r_substate = rack_do_syn_recv;
12485 case TCPS_ESTABLISHED:
12486 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12487 rack->r_state = TCPS_ESTABLISHED;
12488 rack->r_substate = rack_do_established;
12490 case TCPS_CLOSE_WAIT:
12491 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12492 rack->r_state = TCPS_CLOSE_WAIT;
12493 rack->r_substate = rack_do_close_wait;
12495 case TCPS_FIN_WAIT_1:
12496 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12497 rack->r_state = TCPS_FIN_WAIT_1;
12498 rack->r_substate = rack_do_fin_wait_1;
12501 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12502 rack->r_state = TCPS_CLOSING;
12503 rack->r_substate = rack_do_closing;
12505 case TCPS_LAST_ACK:
12506 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12507 rack->r_state = TCPS_LAST_ACK;
12508 rack->r_substate = rack_do_lastack;
12510 case TCPS_FIN_WAIT_2:
12511 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12512 rack->r_state = TCPS_FIN_WAIT_2;
12513 rack->r_substate = rack_do_fin_wait_2;
12517 case TCPS_TIME_WAIT:
12521 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12522 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12527 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12530 * We received an ack, and then did not
12531 * call send or were bounced out due to the
12532 * hpts was running. Now a timer is up as well, is
12533 * it the right timer?
12535 struct rack_sendmap *rsm;
12538 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12539 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12541 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12542 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12543 (tmr_up == PACE_TMR_RXT)) {
12544 /* Should be an RXT */
12548 /* Nothing outstanding? */
12549 if (tp->t_flags & TF_DELACK) {
12550 if (tmr_up == PACE_TMR_DELACK)
12551 /* We are supposed to have delayed ack up and we do */
12553 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12555 * if we hit enobufs then we would expect the possiblity
12556 * of nothing outstanding and the RXT up (and the hptsi timer).
12559 } else if (((V_tcp_always_keepalive ||
12560 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12561 (tp->t_state <= TCPS_CLOSING)) &&
12562 (tmr_up == PACE_TMR_KEEP) &&
12563 (tp->snd_max == tp->snd_una)) {
12564 /* We should have keep alive up and we do */
12568 if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12569 ((tmr_up == PACE_TMR_TLP) ||
12570 (tmr_up == PACE_TMR_RACK) ||
12571 (tmr_up == PACE_TMR_RXT))) {
12573 * Either a Rack, TLP or RXT is fine if we
12574 * have outstanding data.
12577 } else if (tmr_up == PACE_TMR_DELACK) {
12579 * If the delayed ack was going to go off
12580 * before the rtx/tlp/rack timer were going to
12581 * expire, then that would be the timer in control.
12582 * Note we don't check the time here trusting the
12588 * Ok the timer originally started is not what we want now.
12589 * We will force the hpts to be stopped if any, and restart
12590 * with the slot set to what was in the saved slot.
12592 if (rack->rc_inp->inp_in_hpts) {
12593 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12596 us_cts = tcp_get_usecs(NULL);
12597 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12599 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12601 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12603 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
12605 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12606 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12611 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq)
12613 tp->snd_wnd = tiwin;
12614 rack_validate_fo_sendwin_up(tp, rack);
12617 if (tp->snd_wnd > tp->max_sndwnd)
12618 tp->max_sndwnd = tp->snd_wnd;
12619 if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12620 /* The peer collapsed the window */
12621 rack_collapsed_window(rack);
12622 } else if (rack->rc_has_collapsed)
12623 rack_un_collapse_window(rack);
12624 /* Do we exit persists? */
12625 if ((rack->rc_in_persist != 0) &&
12626 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12627 rack->r_ctl.rc_pace_min_segs))) {
12628 rack_exit_persist(tp, rack, cts);
12630 /* Do we enter persists? */
12631 if ((rack->rc_in_persist == 0) &&
12632 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12633 TCPS_HAVEESTABLISHED(tp->t_state) &&
12634 (tp->snd_max == tp->snd_una) &&
12635 sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
12636 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
12638 * Here the rwnd is less than
12639 * the pacing size, we are established,
12640 * nothing is outstanding, and there is
12641 * data to send. Enter persists.
12643 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12648 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
12651 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
12652 union tcp_log_stackspecific log;
12653 struct timeval ltv;
12654 char tcp_hdr_buf[60];
12656 struct timespec ts;
12657 uint32_t orig_snd_una;
12660 #ifdef NETFLIX_HTTP_LOGGING
12661 struct http_sendfile_track *http_req;
12663 if (SEQ_GT(ae->ack, tp->snd_una)) {
12664 http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
12666 http_req = tcp_http_find_req_for_seq(tp, ae->ack);
12669 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
12670 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
12671 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
12672 if (rack->rack_no_prr == 0)
12673 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
12675 log.u_bbr.flex1 = 0;
12676 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
12677 log.u_bbr.use_lt_bw <<= 1;
12678 log.u_bbr.use_lt_bw |= rack->r_might_revert;
12679 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
12680 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
12681 log.u_bbr.pkts_out = tp->t_maxseg;
12682 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
12683 log.u_bbr.flex7 = 1;
12684 log.u_bbr.lost = ae->flags;
12685 log.u_bbr.cwnd_gain = ackval;
12686 log.u_bbr.pacing_gain = 0x2;
12687 if (ae->flags & TSTMP_HDWR) {
12688 /* Record the hardware timestamp if present */
12689 log.u_bbr.flex3 = M_TSTMP;
12690 ts.tv_sec = ae->timestamp / 1000000000;
12691 ts.tv_nsec = ae->timestamp % 1000000000;
12692 ltv.tv_sec = ts.tv_sec;
12693 ltv.tv_usec = ts.tv_nsec / 1000;
12694 log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v);
12695 } else if (ae->flags & TSTMP_LRO) {
12696 /* Record the LRO the arrival timestamp */
12697 log.u_bbr.flex3 = M_TSTMP_LRO;
12698 ts.tv_sec = ae->timestamp / 1000000000;
12699 ts.tv_nsec = ae->timestamp % 1000000000;
12700 ltv.tv_sec = ts.tv_sec;
12701 ltv.tv_usec = ts.tv_nsec / 1000;
12702 log.u_bbr.flex5 = tcp_tv_to_usectick(<v);
12704 log.u_bbr.timeStamp = tcp_get_usecs(<v);
12705 /* Log the rcv time */
12706 log.u_bbr.delRate = ae->timestamp;
12707 #ifdef NETFLIX_HTTP_LOGGING
12708 log.u_bbr.applimited = tp->t_http_closed;
12709 log.u_bbr.applimited <<= 8;
12710 log.u_bbr.applimited |= tp->t_http_open;
12711 log.u_bbr.applimited <<= 8;
12712 log.u_bbr.applimited |= tp->t_http_req;
12714 /* Copy out any client req info */
12716 log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
12718 log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
12719 log.u_bbr.rttProp = http_req->timestamp;
12720 log.u_bbr.cur_del_rate = http_req->start;
12721 if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
12722 log.u_bbr.flex8 |= 1;
12724 log.u_bbr.flex8 |= 2;
12725 log.u_bbr.bw_inuse = http_req->end;
12727 log.u_bbr.flex6 = http_req->start_seq;
12728 if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
12729 log.u_bbr.flex8 |= 4;
12730 log.u_bbr.epoch = http_req->end_seq;
12734 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
12735 th = (struct tcphdr *)tcp_hdr_buf;
12736 th->th_seq = ae->seq;
12737 th->th_ack = ae->ack;
12738 th->th_win = ae->win;
12739 /* Now fill in the ports */
12740 th->th_sport = tp->t_inpcb->inp_fport;
12741 th->th_dport = tp->t_inpcb->inp_lport;
12742 th->th_flags = ae->flags & 0xff;
12743 /* Now do we have a timestamp option? */
12744 if (ae->flags & HAS_TSTMP) {
12748 th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
12749 cp = (u_char *)(th + 1);
12754 *cp = TCPOPT_TIMESTAMP;
12756 *cp = TCPOLEN_TIMESTAMP;
12758 val = htonl(ae->ts_value);
12759 bcopy((char *)&val,
12760 (char *)cp, sizeof(uint32_t));
12761 val = htonl(ae->ts_echo);
12762 bcopy((char *)&val,
12763 (char *)(cp + 4), sizeof(uint32_t));
12765 th->th_off = (sizeof(struct tcphdr) >> 2);
12768 * For sane logging we need to play a little trick.
12769 * If the ack were fully processed we would have moved
12770 * snd_una to high_seq, but since compressed acks are
12771 * processed in two phases, at this point (logging) snd_una
12772 * won't be advanced. So we would see multiple acks showing
12773 * the advancement. We can prevent that by "pretending" that
12774 * snd_una was advanced and then un-advancing it so that the
12775 * logging code has the right value for tlb_snd_una.
12777 if (tp->snd_una != high_seq) {
12778 orig_snd_una = tp->snd_una;
12779 tp->snd_una = high_seq;
12783 TCP_LOG_EVENTP(tp, th,
12784 &tp->t_inpcb->inp_socket->so_rcv,
12785 &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
12786 0, &log, true, <v);
12788 tp->snd_una = orig_snd_una;
12795 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
12798 * Handle a "special" compressed ack mbuf. Each incoming
12799 * ack has only four possible dispositions:
12801 * A) It moves the cum-ack forward
12802 * B) It is behind the cum-ack.
12803 * C) It is a window-update ack.
12804 * D) It is a dup-ack.
12806 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
12807 * in the incoming mbuf. We also need to still pay attention
12808 * to nxt_pkt since there may be another packet after this
12811 #ifdef TCP_ACCOUNTING
12816 struct timespec ts;
12817 struct tcp_rack *rack;
12818 struct tcp_ackent *ae;
12819 uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
12820 int cnt, i, did_out, ourfinisacked = 0;
12821 int win_up_req = 0;
12822 struct tcpopt to_holder, *to = NULL;
12824 int under_pacing = 1;
12827 #ifdef TCP_ACCOUNTING
12830 rack = (struct tcp_rack *)tp->t_fb_ptr;
12831 if (rack->gp_ready &&
12832 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
12837 if (rack->r_state != tp->t_state)
12838 rack_set_state(tp, rack);
12841 KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
12842 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
12843 cnt = m->m_len / sizeof(struct tcp_ackent);
12845 if (idx >= MAX_NUM_OF_CNTS)
12846 idx = MAX_NUM_OF_CNTS - 1;
12847 counter_u64_add(rack_proc_comp_ack[idx], 1);
12848 counter_u64_add(rack_multi_single_eq, cnt);
12849 high_seq = tp->snd_una;
12850 the_win = tp->snd_wnd;
12851 win_seq = tp->snd_wl1;
12852 win_upd_ack = tp->snd_wl2;
12853 cts = us_cts = tcp_tv_to_usectick(tv);
12854 segsiz = ctf_fixed_maxseg(tp);
12855 if ((rack->rc_gp_dyn_mul) &&
12856 (rack->use_fixed_rate == 0) &&
12857 (rack->rc_always_pace)) {
12858 /* Check in on probertt */
12859 rack_check_probe_rtt(rack, us_cts);
12861 for (i = 0; i < cnt; i++) {
12862 #ifdef TCP_ACCOUNTING
12863 ts_val = get_cyclecount();
12865 rack_clear_rate_sample(rack);
12866 ae = ((mtod(m, struct tcp_ackent *)) + i);
12867 /* Setup the window */
12868 tiwin = ae->win << tp->snd_scale;
12869 /* figure out the type of ack */
12870 if (SEQ_LT(ae->ack, high_seq)) {
12872 ae->ack_val_set = ACK_BEHIND;
12873 } else if (SEQ_GT(ae->ack, high_seq)) {
12875 ae->ack_val_set = ACK_CUMACK;
12876 } else if (tiwin == the_win) {
12878 ae->ack_val_set = ACK_DUPACK;
12881 ae->ack_val_set = ACK_RWND;
12883 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
12884 /* Validate timestamp */
12885 if (ae->flags & HAS_TSTMP) {
12886 /* Setup for a timestamp */
12887 to->to_flags = TOF_TS;
12888 ae->ts_echo -= tp->ts_offset;
12889 to->to_tsecr = ae->ts_echo;
12890 to->to_tsval = ae->ts_value;
12892 * If echoed timestamp is later than the current time, fall back to
12893 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
12894 * were used when this connection was established.
12896 if (TSTMP_GT(ae->ts_echo, cts))
12898 if (tp->ts_recent &&
12899 TSTMP_LT(ae->ts_value, tp->ts_recent)) {
12900 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
12901 #ifdef TCP_ACCOUNTING
12902 rdstc = get_cyclecount();
12903 if (rdstc > ts_val) {
12904 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
12906 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12907 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
12914 if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
12915 SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
12916 tp->ts_recent_age = tcp_ts_getticks();
12917 tp->ts_recent = ae->ts_value;
12920 /* Setup for a no options */
12923 /* Update the rcv time and perform idle reduction possibly */
12924 if (tp->t_idle_reduce &&
12925 (tp->snd_max == tp->snd_una) &&
12926 ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
12927 counter_u64_add(rack_input_idle_reduces, 1);
12928 rack_cc_after_idle(rack, tp);
12930 tp->t_rcvtime = ticks;
12931 /* Now what about ECN? */
12932 if (tp->t_flags2 & TF2_ECN_PERMIT) {
12933 if (ae->flags & TH_CWR) {
12934 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
12935 tp->t_flags |= TF_ACKNOW;
12937 switch (ae->codepoint & IPTOS_ECN_MASK) {
12939 tp->t_flags2 |= TF2_ECN_SND_ECE;
12940 KMOD_TCPSTAT_INC(tcps_ecn_ce);
12942 case IPTOS_ECN_ECT0:
12943 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
12945 case IPTOS_ECN_ECT1:
12946 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
12950 /* Process a packet differently from RFC3168. */
12951 cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
12952 /* Congestion experienced. */
12953 if (ae->flags & TH_ECE) {
12954 rack_cong_signal(tp, CC_ECN, ae->ack);
12957 #ifdef TCP_ACCOUNTING
12958 /* Count for the specific type of ack in */
12959 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
12960 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12961 tp->tcp_cnt_counters[ae->ack_val_set]++;
12965 * Note how we could move up these in the determination
12966 * above, but we don't so that way the timestamp checks (and ECN)
12967 * is done first before we do any processing on the ACK.
12968 * The non-compressed path through the code has this
12969 * weakness (noted by @jtl) that it actually does some
12970 * processing before verifying the timestamp information.
12971 * We don't take that path here which is why we set
12972 * the ack_val_set first, do the timestamp and ecn
12973 * processing, and then look at what we have setup.
12975 if (ae->ack_val_set == ACK_BEHIND) {
12977 * Case B flag reordering, if window is not closed
12978 * or it could be a keep-alive or persists
12980 if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
12981 counter_u64_add(rack_reorder_seen, 1);
12982 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12984 } else if (ae->ack_val_set == ACK_DUPACK) {
12987 rack_strike_dupack(rack);
12988 } else if (ae->ack_val_set == ACK_RWND) {
12992 win_upd_ack = ae->ack;
12998 if (SEQ_GT(ae->ack, tp->snd_max)) {
13000 * We just send an ack since the incoming
13001 * ack is beyond the largest seq we sent.
13003 if ((tp->t_flags & TF_ACKNOW) == 0) {
13004 ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13005 if (tp->t_flags && TF_ACKNOW)
13006 rack->r_wanted_output = 1;
13010 /* If the window changed setup to update */
13011 if (tiwin != tp->snd_wnd) {
13013 win_upd_ack = ae->ack;
13017 #ifdef TCP_ACCOUNTING
13018 /* Account for the acks */
13019 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13020 tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13022 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13023 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13025 high_seq = ae->ack;
13026 /* Setup our act_rcv_time */
13027 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13028 ts.tv_sec = ae->timestamp / 1000000000;
13029 ts.tv_nsec = ae->timestamp % 1000000000;
13030 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13031 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13033 rack->r_ctl.act_rcv_time = *tv;
13035 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13038 /* And lets be sure to commit the rtt measurements for this ack */
13039 tcp_rack_xmit_timer_commit(rack, tp);
13040 #ifdef TCP_ACCOUNTING
13041 rdstc = get_cyclecount();
13042 if (rdstc > ts_val) {
13043 counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13044 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13045 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13046 if (ae->ack_val_set == ACK_CUMACK)
13047 tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13052 #ifdef TCP_ACCOUNTING
13053 ts_val = get_cyclecount();
13055 acked_amount = acked = (high_seq - tp->snd_una);
13057 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13060 if (rack->sack_attack_disable == 0)
13061 rack_do_decay(rack);
13062 if (acked >= segsiz) {
13064 * You only get credit for
13065 * MSS and greater (and you get extra
13066 * credit for larger cum-ack moves).
13070 ac = acked / segsiz;
13071 rack->r_ctl.ack_count += ac;
13072 counter_u64_add(rack_ack_total, ac);
13074 if (rack->r_ctl.ack_count > 0xfff00000) {
13076 * reduce the number to keep us under
13079 rack->r_ctl.ack_count /= 2;
13080 rack->r_ctl.sack_count /= 2;
13082 if (tp->t_flags & TF_NEEDSYN) {
13084 * T/TCP: Connection was half-synchronized, and our SYN has
13085 * been ACK'd (so connection is now fully synchronized). Go
13086 * to non-starred state, increment snd_una for ACK of SYN,
13087 * and check if we can do window scaling.
13089 tp->t_flags &= ~TF_NEEDSYN;
13091 acked_amount = acked = (high_seq - tp->snd_una);
13093 if (acked > sbavail(&so->so_snd))
13094 acked_amount = sbavail(&so->so_snd);
13095 #ifdef NETFLIX_EXP_DETECTION
13097 * We only care on a cum-ack move if we are in a sack-disabled
13098 * state. We have already added in to the ack_count, and we never
13099 * would disable on a cum-ack move, so we only care to do the
13100 * detection if it may "undo" it, i.e. we were in disabled already.
13102 if (rack->sack_attack_disable)
13103 rack_do_detection(tp, rack, acked_amount, segsiz);
13105 if (IN_FASTRECOVERY(tp->t_flags) &&
13106 (rack->rack_no_prr == 0))
13107 rack_update_prr(tp, rack, acked_amount, high_seq);
13108 if (IN_RECOVERY(tp->t_flags)) {
13109 if (SEQ_LT(high_seq, tp->snd_recover) &&
13110 (SEQ_LT(high_seq, tp->snd_max))) {
13111 tcp_rack_partialack(tp);
13113 rack_post_recovery(tp, high_seq);
13117 /* Handle the rack-log-ack part (sendmap) */
13118 if ((sbused(&so->so_snd) == 0) &&
13119 (acked > acked_amount) &&
13120 (tp->t_state >= TCPS_FIN_WAIT_1) &&
13121 (tp->t_flags & TF_SENTFIN)) {
13123 * We must be sure our fin
13124 * was sent and acked (we can be
13125 * in FIN_WAIT_1 without having
13130 * Lets make sure snd_una is updated
13131 * since most likely acked_amount = 0 (it
13134 tp->snd_una = high_seq;
13136 /* Did we make a RTO error? */
13137 if ((tp->t_flags & TF_PREVVALID) &&
13138 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13139 tp->t_flags &= ~TF_PREVVALID;
13140 if (tp->t_rxtshift == 1 &&
13141 (int)(ticks - tp->t_badrxtwin) < 0)
13142 rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13144 /* Handle the data in the socket buffer */
13145 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13146 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13147 if (acked_amount > 0) {
13148 struct mbuf *mfree;
13150 rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13151 SOCKBUF_LOCK(&so->so_snd);
13152 mfree = sbcut_locked(&so->so_snd, acked);
13153 tp->snd_una = high_seq;
13154 /* Note we want to hold the sb lock through the sendmap adjust */
13155 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13156 /* Wake up the socket if we have room to write more */
13157 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13158 SOCKBUF_UNLOCK(&so->so_snd);
13159 tp->t_flags |= TF_WAKESOW;
13162 /* update progress */
13163 tp->t_acktime = ticks;
13164 rack_log_progress_event(rack, tp, tp->t_acktime,
13165 PROGRESS_UPDATE, __LINE__);
13166 /* Clear out shifts and such */
13167 tp->t_rxtshift = 0;
13168 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13169 rack_rto_min, rack_rto_max);
13170 rack->rc_tlp_in_progress = 0;
13171 rack->r_ctl.rc_tlp_cnt_out = 0;
13172 /* Send recover and snd_nxt must be dragged along */
13173 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13174 tp->snd_recover = tp->snd_una;
13175 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13176 tp->snd_nxt = tp->snd_una;
13178 * If the RXT timer is running we want to
13179 * stop it, so we can restart a TLP (or new RXT).
13181 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13182 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13183 #ifdef NETFLIX_HTTP_LOGGING
13184 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13186 tp->snd_wl2 = high_seq;
13188 if (under_pacing &&
13189 (rack->use_fixed_rate == 0) &&
13190 (rack->in_probe_rtt == 0) &&
13191 rack->rc_gp_dyn_mul &&
13192 rack->rc_always_pace) {
13193 /* Check if we are dragging bottom */
13194 rack_check_bottom_drag(tp, rack, so, acked);
13196 if (tp->snd_una == tp->snd_max) {
13197 tp->t_flags &= ~TF_PREVVALID;
13198 rack->r_ctl.retran_during_recovery = 0;
13199 rack->r_ctl.dsack_byte_cnt = 0;
13200 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13201 if (rack->r_ctl.rc_went_idle_time == 0)
13202 rack->r_ctl.rc_went_idle_time = 1;
13203 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13204 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13206 /* Set so we might enter persists... */
13207 rack->r_wanted_output = 1;
13208 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13209 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13210 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13211 (sbavail(&so->so_snd) == 0) &&
13212 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13214 * The socket was gone and the
13215 * peer sent data (not now in the past), time to
13218 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13219 /* tcp_close will kill the inp pre-log the Reset */
13220 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13221 #ifdef TCP_ACCOUNTING
13222 rdstc = get_cyclecount();
13223 if (rdstc > ts_val) {
13224 counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13225 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13226 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13227 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13232 tp = tcp_close(tp);
13234 #ifdef TCP_ACCOUNTING
13240 * We would normally do drop-with-reset which would
13241 * send back a reset. We can't since we don't have
13242 * all the needed bits. Instead lets arrange for
13243 * a call to tcp_output(). That way since we
13244 * are in the closed state we will generate a reset.
13246 * Note if tcp_accounting is on we don't unpin since
13247 * we do that after the goto label.
13249 goto send_out_a_rst;
13251 if ((sbused(&so->so_snd) == 0) &&
13252 (tp->t_state >= TCPS_FIN_WAIT_1) &&
13253 (tp->t_flags & TF_SENTFIN)) {
13255 * If we can't receive any more data, then closing user can
13256 * proceed. Starting the timer is contrary to the
13257 * specification, but if we don't get a FIN we'll hang
13261 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13262 soisdisconnected(so);
13263 tcp_timer_activate(tp, TT_2MSL,
13264 (tcp_fast_finwait2_recycle ?
13265 tcp_finwait2_timeout :
13268 if (ourfinisacked == 0) {
13270 * We don't change to fin-wait-2 if we have our fin acked
13271 * which means we are probably in TCPS_CLOSING.
13273 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13277 /* Wake up the socket if we have room to write more */
13278 if (sbavail(&so->so_snd)) {
13279 rack->r_wanted_output = 1;
13280 if (ctf_progress_timeout_check(tp, true)) {
13281 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13282 tp, tick, PROGRESS_DROP, __LINE__);
13283 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13285 * We cheat here and don't send a RST, we should send one
13286 * when the pacer drops the connection.
13288 #ifdef TCP_ACCOUNTING
13289 rdstc = get_cyclecount();
13290 if (rdstc > ts_val) {
13291 counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13292 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13293 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13294 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13299 INP_WUNLOCK(rack->rc_inp);
13304 if (ourfinisacked) {
13305 switch(tp->t_state) {
13307 #ifdef TCP_ACCOUNTING
13308 rdstc = get_cyclecount();
13309 if (rdstc > ts_val) {
13310 counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13312 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13313 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13314 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13323 case TCPS_LAST_ACK:
13324 #ifdef TCP_ACCOUNTING
13325 rdstc = get_cyclecount();
13326 if (rdstc > ts_val) {
13327 counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13329 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13330 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13331 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13336 tp = tcp_close(tp);
13337 ctf_do_drop(m, tp);
13340 case TCPS_FIN_WAIT_1:
13341 #ifdef TCP_ACCOUNTING
13342 rdstc = get_cyclecount();
13343 if (rdstc > ts_val) {
13344 counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13346 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13347 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13348 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13352 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13353 soisdisconnected(so);
13354 tcp_timer_activate(tp, TT_2MSL,
13355 (tcp_fast_finwait2_recycle ?
13356 tcp_finwait2_timeout :
13359 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13365 if (rack->r_fast_output) {
13367 * We re doing fast output.. can we expand that?
13369 rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13371 #ifdef TCP_ACCOUNTING
13372 rdstc = get_cyclecount();
13373 if (rdstc > ts_val) {
13374 counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13375 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13376 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13377 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13381 } else if (win_up_req) {
13382 rdstc = get_cyclecount();
13383 if (rdstc > ts_val) {
13384 counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13385 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13386 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13391 /* Now is there a next packet, if so we are done */
13395 #ifdef TCP_ACCOUNTING
13398 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13401 rack_handle_might_revert(tp, rack);
13402 ctf_calc_rwin(so, tp);
13403 if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13405 (void)tp->t_fb->tfb_tcp_output(tp);
13408 rack_free_trim(rack);
13409 #ifdef TCP_ACCOUNTING
13412 rack_timer_audit(tp, rack, &so->so_snd);
13413 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13419 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13420 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13421 int32_t nxt_pkt, struct timeval *tv)
13423 #ifdef TCP_ACCOUNTING
13426 int32_t thflags, retval, did_out = 0;
13427 int32_t way_out = 0;
13430 struct timespec ts;
13432 struct tcp_rack *rack;
13433 struct rack_sendmap *rsm;
13434 int32_t prev_state = 0;
13435 #ifdef TCP_ACCOUNTING
13436 int ack_val_set = 0xf;
13440 * tv passed from common code is from either M_TSTMP_LRO or
13441 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13443 if (m->m_flags & M_ACKCMP) {
13444 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13446 if (m->m_flags & M_ACKCMP) {
13447 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13449 counter_u64_add(rack_proc_non_comp_ack, 1);
13450 thflags = th->th_flags;
13451 #ifdef TCP_ACCOUNTING
13453 if (thflags & TH_ACK)
13454 ts_val = get_cyclecount();
13456 cts = tcp_tv_to_usectick(tv);
13457 rack = (struct tcp_rack *)tp->t_fb_ptr;
13459 if ((m->m_flags & M_TSTMP) ||
13460 (m->m_flags & M_TSTMP_LRO)) {
13461 mbuf_tstmp2timespec(m, &ts);
13462 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13463 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13465 rack->r_ctl.act_rcv_time = *tv;
13466 kern_prefetch(rack, &prev_state);
13469 * Unscale the window into a 32-bit value. For the SYN_SENT state
13470 * the scale is zero.
13472 tiwin = th->th_win << tp->snd_scale;
13474 * Parse options on any incoming segment.
13476 memset(&to, 0, sizeof(to));
13477 tcp_dooptions(&to, (u_char *)(th + 1),
13478 (th->th_off << 2) - sizeof(struct tcphdr),
13479 (thflags & TH_SYN) ? TO_SYN : 0);
13480 #ifdef TCP_ACCOUNTING
13481 if (thflags & TH_ACK) {
13483 * We have a tradeoff here. We can either do what we are
13484 * doing i.e. pinning to this CPU and then doing the accounting
13485 * <or> we could do a critical enter, setup the rdtsc and cpu
13486 * as in below, and then validate we are on the same CPU on
13487 * exit. I have choosen to not do the critical enter since
13488 * that often will gain you a context switch, and instead lock
13489 * us (line above this if) to the same CPU with sched_pin(). This
13490 * means we may be context switched out for a higher priority
13491 * interupt but we won't be moved to another CPU.
13493 * If this occurs (which it won't very often since we most likely
13494 * are running this code in interupt context and only a higher
13495 * priority will bump us ... clock?) we will falsely add in
13496 * to the time the interupt processing time plus the ack processing
13497 * time. This is ok since its a rare event.
13499 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13500 ctf_fixed_maxseg(tp));
13503 NET_EPOCH_ASSERT();
13504 INP_WLOCK_ASSERT(tp->t_inpcb);
13505 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13507 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13509 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13510 union tcp_log_stackspecific log;
13511 struct timeval ltv;
13512 #ifdef NETFLIX_HTTP_LOGGING
13513 struct http_sendfile_track *http_req;
13515 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13516 http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13518 http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13521 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13522 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13523 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13524 if (rack->rack_no_prr == 0)
13525 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13527 log.u_bbr.flex1 = 0;
13528 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13529 log.u_bbr.use_lt_bw <<= 1;
13530 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13531 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13532 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13533 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13534 log.u_bbr.flex3 = m->m_flags;
13535 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13536 log.u_bbr.lost = thflags;
13537 log.u_bbr.pacing_gain = 0x1;
13538 #ifdef TCP_ACCOUNTING
13539 log.u_bbr.cwnd_gain = ack_val_set;
13541 log.u_bbr.flex7 = 2;
13542 if (m->m_flags & M_TSTMP) {
13543 /* Record the hardware timestamp if present */
13544 mbuf_tstmp2timespec(m, &ts);
13545 ltv.tv_sec = ts.tv_sec;
13546 ltv.tv_usec = ts.tv_nsec / 1000;
13547 log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v);
13548 } else if (m->m_flags & M_TSTMP_LRO) {
13549 /* Record the LRO the arrival timestamp */
13550 mbuf_tstmp2timespec(m, &ts);
13551 ltv.tv_sec = ts.tv_sec;
13552 ltv.tv_usec = ts.tv_nsec / 1000;
13553 log.u_bbr.flex5 = tcp_tv_to_usectick(<v);
13555 log.u_bbr.timeStamp = tcp_get_usecs(<v);
13556 /* Log the rcv time */
13557 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13558 #ifdef NETFLIX_HTTP_LOGGING
13559 log.u_bbr.applimited = tp->t_http_closed;
13560 log.u_bbr.applimited <<= 8;
13561 log.u_bbr.applimited |= tp->t_http_open;
13562 log.u_bbr.applimited <<= 8;
13563 log.u_bbr.applimited |= tp->t_http_req;
13565 /* Copy out any client req info */
13567 log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13569 log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13570 log.u_bbr.rttProp = http_req->timestamp;
13571 log.u_bbr.cur_del_rate = http_req->start;
13572 if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13573 log.u_bbr.flex8 |= 1;
13575 log.u_bbr.flex8 |= 2;
13576 log.u_bbr.bw_inuse = http_req->end;
13578 log.u_bbr.flex6 = http_req->start_seq;
13579 if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13580 log.u_bbr.flex8 |= 4;
13581 log.u_bbr.epoch = http_req->end_seq;
13585 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13586 tlen, &log, true, <v);
13588 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13591 goto done_with_input;
13594 * If a segment with the ACK-bit set arrives in the SYN-SENT state
13595 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13597 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13598 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13599 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13600 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13601 #ifdef TCP_ACCOUNTING
13608 * Parse options on any incoming segment.
13610 tcp_dooptions(&to, (u_char *)(th + 1),
13611 (th->th_off << 2) - sizeof(struct tcphdr),
13612 (thflags & TH_SYN) ? TO_SYN : 0);
13615 * If timestamps were negotiated during SYN/ACK and a
13616 * segment without a timestamp is received, silently drop
13617 * the segment, unless it is a RST segment or missing timestamps are
13619 * See section 3.2 of RFC 7323.
13621 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13622 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13625 goto done_with_input;
13629 * Segment received on connection. Reset idle time and keep-alive
13630 * timer. XXX: This should be done after segment validation to
13631 * ignore broken/spoofed segs.
13633 if (tp->t_idle_reduce &&
13634 (tp->snd_max == tp->snd_una) &&
13635 ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13636 counter_u64_add(rack_input_idle_reduces, 1);
13637 rack_cc_after_idle(rack, tp);
13639 tp->t_rcvtime = ticks;
13641 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13643 if (tiwin > rack->r_ctl.rc_high_rwnd)
13644 rack->r_ctl.rc_high_rwnd = tiwin;
13646 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13647 * this to occur after we've validated the segment.
13649 if (tp->t_flags2 & TF2_ECN_PERMIT) {
13650 if (thflags & TH_CWR) {
13651 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13652 tp->t_flags |= TF_ACKNOW;
13654 switch (iptos & IPTOS_ECN_MASK) {
13656 tp->t_flags2 |= TF2_ECN_SND_ECE;
13657 KMOD_TCPSTAT_INC(tcps_ecn_ce);
13659 case IPTOS_ECN_ECT0:
13660 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13662 case IPTOS_ECN_ECT1:
13663 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13667 /* Process a packet differently from RFC3168. */
13668 cc_ecnpkt_handler(tp, th, iptos);
13670 /* Congestion experienced. */
13671 if (thflags & TH_ECE) {
13672 rack_cong_signal(tp, CC_ECN, th->th_ack);
13677 * If echoed timestamp is later than the current time, fall back to
13678 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
13679 * were used when this connection was established.
13681 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13682 to.to_tsecr -= tp->ts_offset;
13683 if (TSTMP_GT(to.to_tsecr, cts))
13688 * If its the first time in we need to take care of options and
13689 * verify we can do SACK for rack!
13691 if (rack->r_state == 0) {
13692 /* Should be init'd by rack_init() */
13693 KASSERT(rack->rc_inp != NULL,
13694 ("%s: rack->rc_inp unexpectedly NULL", __func__));
13695 if (rack->rc_inp == NULL) {
13696 rack->rc_inp = tp->t_inpcb;
13700 * Process options only when we get SYN/ACK back. The SYN
13701 * case for incoming connections is handled in tcp_syncache.
13702 * According to RFC1323 the window field in a SYN (i.e., a
13703 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13704 * this is traditional behavior, may need to be cleaned up.
13706 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13707 /* Handle parallel SYN for ECN */
13708 if (!(thflags & TH_ACK) &&
13709 ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13710 ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13711 tp->t_flags2 |= TF2_ECN_PERMIT;
13712 tp->t_flags2 |= TF2_ECN_SND_ECE;
13713 TCPSTAT_INC(tcps_ecn_shs);
13715 if ((to.to_flags & TOF_SCALE) &&
13716 (tp->t_flags & TF_REQ_SCALE)) {
13717 tp->t_flags |= TF_RCVD_SCALE;
13718 tp->snd_scale = to.to_wscale;
13720 tp->t_flags &= ~TF_REQ_SCALE;
13722 * Initial send window. It will be updated with the
13723 * next incoming segment to the scaled value.
13725 tp->snd_wnd = th->th_win;
13726 rack_validate_fo_sendwin_up(tp, rack);
13727 if ((to.to_flags & TOF_TS) &&
13728 (tp->t_flags & TF_REQ_TSTMP)) {
13729 tp->t_flags |= TF_RCVD_TSTMP;
13730 tp->ts_recent = to.to_tsval;
13731 tp->ts_recent_age = cts;
13733 tp->t_flags &= ~TF_REQ_TSTMP;
13734 if (to.to_flags & TOF_MSS) {
13735 tcp_mss(tp, to.to_mss);
13737 if ((tp->t_flags & TF_SACK_PERMIT) &&
13738 (to.to_flags & TOF_SACKPERM) == 0)
13739 tp->t_flags &= ~TF_SACK_PERMIT;
13740 if (IS_FASTOPEN(tp->t_flags)) {
13741 if (to.to_flags & TOF_FASTOPEN) {
13744 if (to.to_flags & TOF_MSS)
13747 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13751 tcp_fastopen_update_cache(tp, mss,
13752 to.to_tfo_len, to.to_tfo_cookie);
13754 tcp_fastopen_disable_path(tp);
13758 * At this point we are at the initial call. Here we decide
13759 * if we are doing RACK or not. We do this by seeing if
13760 * TF_SACK_PERMIT is set and the sack-not-required is clear.
13761 * The code now does do dup-ack counting so if you don't
13762 * switch back you won't get rack & TLP, but you will still
13766 if ((rack_sack_not_required == 0) &&
13767 ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13768 tcp_switch_back_to_default(tp);
13769 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13771 #ifdef TCP_ACCOUNTING
13776 tcp_set_hpts(tp->t_inpcb);
13777 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13779 if (thflags & TH_FIN)
13780 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13781 us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13782 if ((rack->rc_gp_dyn_mul) &&
13783 (rack->use_fixed_rate == 0) &&
13784 (rack->rc_always_pace)) {
13785 /* Check in on probertt */
13786 rack_check_probe_rtt(rack, us_cts);
13788 if (rack->forced_ack) {
13792 * A persist or keep-alive was forced out, update our
13793 * min rtt time. Note we do not worry about lost
13794 * retransmissions since KEEP-ALIVES and persists
13795 * are usually way long on times of sending (though
13796 * if we were really paranoid or worried we could
13797 * at least use timestamps if available to validate).
13799 rack->forced_ack = 0;
13800 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13803 rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
13804 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13807 * This is the one exception case where we set the rack state
13808 * always. All other times (timers etc) we must have a rack-state
13809 * set (so we assure we have done the checks above for SACK).
13811 rack->r_ctl.rc_rcvtime = cts;
13812 if (rack->r_state != tp->t_state)
13813 rack_set_state(tp, rack);
13814 if (SEQ_GT(th->th_ack, tp->snd_una) &&
13815 (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13816 kern_prefetch(rsm, &prev_state);
13817 prev_state = rack->r_state;
13818 rack_clear_rate_sample(rack);
13819 retval = (*rack->r_substate) (m, th, so,
13820 tp, &to, drop_hdrlen,
13821 tlen, tiwin, thflags, nxt_pkt, iptos);
13823 if ((retval == 0) &&
13824 (tp->t_inpcb == NULL)) {
13825 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13826 retval, tp, prev_state);
13831 * If retval is 1 the tcb is unlocked and most likely the tp
13834 INP_WLOCK_ASSERT(tp->t_inpcb);
13835 if ((rack->rc_gp_dyn_mul) &&
13836 (rack->rc_always_pace) &&
13837 (rack->use_fixed_rate == 0) &&
13838 rack->in_probe_rtt &&
13839 (rack->r_ctl.rc_time_probertt_starts == 0)) {
13841 * If we are going for target, lets recheck before
13844 rack_check_probe_rtt(rack, us_cts);
13846 if (rack->set_pacing_done_a_iw == 0) {
13847 /* How much has been acked? */
13848 if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
13849 /* We have enough to set in the pacing segment size */
13850 rack->set_pacing_done_a_iw = 1;
13851 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13854 tcp_rack_xmit_timer_commit(rack, tp);
13855 #ifdef TCP_ACCOUNTING
13857 * If we set the ack_val_se to what ack processing we are doing
13858 * we also want to track how many cycles we burned. Note
13859 * the bits after tcp_output we let be "free". This is because
13860 * we are also tracking the tcp_output times as well. Note the
13861 * use of 0xf here since we only have 11 counter (0 - 0xa) and
13862 * 0xf cannot be returned and is what we initialize it too to
13863 * indicate we are not doing the tabulations.
13865 if (ack_val_set != 0xf) {
13868 crtsc = get_cyclecount();
13869 counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13870 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13871 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
13875 if (nxt_pkt == 0) {
13876 if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13879 (void)tp->t_fb->tfb_tcp_output(tp);
13881 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
13882 rack_free_trim(rack);
13884 if ((nxt_pkt == 0) &&
13885 ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
13886 (SEQ_GT(tp->snd_max, tp->snd_una) ||
13887 (tp->t_flags & TF_DELACK) ||
13888 ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13889 (tp->t_state <= TCPS_CLOSING)))) {
13890 /* We could not send (probably in the hpts but stopped the timer earlier)? */
13891 if ((tp->snd_max == tp->snd_una) &&
13892 ((tp->t_flags & TF_DELACK) == 0) &&
13893 (rack->rc_inp->inp_in_hpts) &&
13894 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13895 /* keep alive not needed if we are hptsi output yet */
13899 if (rack->rc_inp->inp_in_hpts) {
13900 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13901 us_cts = tcp_get_usecs(NULL);
13902 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13904 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13907 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13909 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13911 if (late && (did_out == 0)) {
13913 * We are late in the sending
13914 * and we did not call the output
13915 * (this probably should not happen).
13917 goto do_output_now;
13919 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13922 } else if (nxt_pkt == 0) {
13923 /* Do we have the correct timer running? */
13924 rack_timer_audit(tp, rack, &so->so_snd);
13928 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, m->m_pkthdr.lro_nsegs));
13930 rack->r_wanted_output = 0;
13932 if (tp->t_inpcb == NULL) {
13933 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
13935 retval, tp, prev_state);
13938 #ifdef TCP_ACCOUNTING
13941 * Track the time (see above).
13943 if (ack_val_set != 0xf) {
13946 crtsc = get_cyclecount();
13947 counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13949 * Note we *DO NOT* increment the per-tcb counters since
13950 * in the else the TP may be gone!!
13955 #ifdef TCP_ACCOUNTING
13962 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
13963 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
13967 /* First lets see if we have old packets */
13968 if (tp->t_in_pkt) {
13969 if (ctf_do_queued_segments(so, tp, 1)) {
13974 if (m->m_flags & M_TSTMP_LRO) {
13975 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
13976 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
13978 /* Should not be should we kassert instead? */
13979 tcp_get_usecs(&tv);
13981 if (rack_do_segment_nounlock(m, th, so, tp,
13982 drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
13983 tcp_handle_wakeup(tp, so);
13984 INP_WUNLOCK(tp->t_inpcb);
13988 struct rack_sendmap *
13989 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
13991 struct rack_sendmap *rsm = NULL;
13993 uint32_t srtt = 0, thresh = 0, ts_low = 0;
13995 /* Return the next guy to be re-transmitted */
13996 if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
13999 if (tp->t_flags & TF_SENTFIN) {
14000 /* retran the end FIN? */
14003 /* ok lets look at this one */
14004 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14005 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14008 rsm = rack_find_lowest_rsm(rack);
14013 if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14014 (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14016 * No sack so we automatically do the 3 strikes and
14017 * retransmit (no rack timer would be started).
14022 if (rsm->r_flags & RACK_ACKED) {
14025 if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14026 (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14027 /* Its not yet ready */
14030 srtt = rack_grab_rtt(tp, rack);
14031 idx = rsm->r_rtr_cnt - 1;
14032 ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14033 thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14034 if ((tsused == ts_low) ||
14035 (TSTMP_LT(tsused, ts_low))) {
14036 /* No time since sending */
14039 if ((tsused - ts_low) < thresh) {
14040 /* It has not been long enough yet */
14043 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14044 ((rsm->r_flags & RACK_SACK_PASSED) &&
14045 (rack->sack_attack_disable == 0))) {
14047 * We have passed the dup-ack threshold <or>
14048 * a SACK has indicated this is missing.
14049 * Note that if you are a declared attacker
14050 * it is only the dup-ack threshold that
14051 * will cause retransmits.
14053 /* log retransmit reason */
14054 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14055 rack->r_fast_output = 0;
14062 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14063 uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14064 int line, struct rack_sendmap *rsm)
14066 if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14067 union tcp_log_stackspecific log;
14070 memset(&log, 0, sizeof(log));
14071 log.u_bbr.flex1 = slot;
14072 log.u_bbr.flex2 = len;
14073 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14074 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14075 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14076 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14077 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14078 log.u_bbr.use_lt_bw <<= 1;
14079 log.u_bbr.use_lt_bw |= rack->r_late;
14080 log.u_bbr.use_lt_bw <<= 1;
14081 log.u_bbr.use_lt_bw |= rack->r_early;
14082 log.u_bbr.use_lt_bw <<= 1;
14083 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14084 log.u_bbr.use_lt_bw <<= 1;
14085 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14086 log.u_bbr.use_lt_bw <<= 1;
14087 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14088 log.u_bbr.use_lt_bw <<= 1;
14089 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14090 log.u_bbr.use_lt_bw <<= 1;
14091 log.u_bbr.use_lt_bw |= rack->gp_ready;
14092 log.u_bbr.pkt_epoch = line;
14093 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14094 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14095 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14096 log.u_bbr.bw_inuse = bw_est;
14097 log.u_bbr.delRate = bw;
14098 if (rack->r_ctl.gp_bw == 0)
14099 log.u_bbr.cur_del_rate = 0;
14101 log.u_bbr.cur_del_rate = rack_get_bw(rack);
14102 log.u_bbr.rttProp = len_time;
14103 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14104 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14105 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14106 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14107 /* We are in slow start */
14108 log.u_bbr.flex7 = 1;
14110 /* we are on congestion avoidance */
14111 log.u_bbr.flex7 = 0;
14113 log.u_bbr.flex8 = method;
14114 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14115 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14116 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14117 log.u_bbr.cwnd_gain <<= 1;
14118 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14119 log.u_bbr.cwnd_gain <<= 1;
14120 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14121 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14122 &rack->rc_inp->inp_socket->so_rcv,
14123 &rack->rc_inp->inp_socket->so_snd,
14124 BBR_LOG_HPTSI_CALC, 0,
14125 0, &log, false, &tv);
14130 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14132 uint32_t new_tso, user_max;
14134 user_max = rack->rc_user_set_max_segs * mss;
14135 if (rack->rc_force_max_seg) {
14138 if (rack->use_fixed_rate &&
14139 ((rack->r_ctl.crte == NULL) ||
14140 (bw != rack->r_ctl.crte->rate))) {
14141 /* Use the user mss since we are not exactly matched */
14144 new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14145 if (new_tso > user_max)
14146 new_tso = user_max;
14151 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
14153 uint64_t lentim, fill_bw;
14155 /* Lets first see if we are full, if so continue with normal rate */
14156 rack->r_via_fill_cw = 0;
14157 if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14159 if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14161 if (rack->r_ctl.rc_last_us_rtt == 0)
14163 if (rack->rc_pace_fill_if_rttin_range &&
14164 (rack->r_ctl.rc_last_us_rtt >=
14165 (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14166 /* The rtt is huge, N * smallest, lets not fill */
14170 * first lets calculate the b/w based on the last us-rtt
14173 fill_bw = rack->r_ctl.cwnd_to_use;
14174 /* Take the rwnd if its smaller */
14175 if (fill_bw > rack->rc_tp->snd_wnd)
14176 fill_bw = rack->rc_tp->snd_wnd;
14177 if (rack->r_fill_less_agg) {
14179 * Now take away the inflight (this will reduce our
14180 * aggressiveness and yeah, if we get that much out in 1RTT
14181 * we will have had acks come back and still be behind).
14183 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14185 /* Now lets make it into a b/w */
14186 fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14187 fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14188 /* We are below the min b/w */
14190 *rate_wanted = fill_bw;
14191 if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14193 if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14194 fill_bw = rack->r_ctl.bw_rate_cap;
14195 rack->r_via_fill_cw = 1;
14196 if (rack->r_rack_hw_rate_caps &&
14197 (rack->r_ctl.crte != NULL)) {
14198 uint64_t high_rate;
14200 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14201 if (fill_bw > high_rate) {
14202 /* We are capping bw at the highest rate table entry */
14203 if (*rate_wanted > high_rate) {
14204 /* The original rate was also capped */
14205 rack->r_via_fill_cw = 0;
14207 rack_log_hdwr_pacing(rack,
14208 fill_bw, high_rate, __LINE__,
14210 fill_bw = high_rate;
14214 } else if ((rack->r_ctl.crte == NULL) &&
14215 (rack->rack_hdrw_pacing == 0) &&
14216 (rack->rack_hdw_pace_ena) &&
14217 rack->r_rack_hw_rate_caps &&
14218 (rack->rack_attempt_hdwr_pace == 0) &&
14219 (rack->rc_inp->inp_route.ro_nh != NULL) &&
14220 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14222 * Ok we may have a first attempt that is greater than our top rate
14225 uint64_t high_rate;
14227 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14229 if (fill_bw > high_rate) {
14230 fill_bw = high_rate;
14237 * Ok fill_bw holds our mythical b/w to fill the cwnd
14238 * in a rtt, what does that time wise equate too?
14240 lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14242 *rate_wanted = fill_bw;
14243 if (non_paced || (lentim < slot)) {
14244 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14245 0, lentim, 12, __LINE__, NULL);
14246 return ((int32_t)lentim);
14252 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14254 struct rack_sendmap *lrsm;
14256 int can_start_hw_pacing = 1;
14259 if (rack->rc_always_pace == 0) {
14261 * We use the most optimistic possible cwnd/srtt for
14262 * sending calculations. This will make our
14263 * calculation anticipate getting more through
14264 * quicker then possible. But thats ok we don't want
14265 * the peer to have a gap in data sending.
14267 uint32_t srtt, cwnd, tr_perms = 0;
14268 int32_t reduce = 0;
14272 * We keep no precise pacing with the old method
14273 * instead we use the pacer to mitigate bursts.
14275 if (rack->r_ctl.rc_rack_min_rtt)
14276 srtt = rack->r_ctl.rc_rack_min_rtt;
14278 srtt = max(tp->t_srtt, 1);
14279 if (rack->r_ctl.rc_rack_largest_cwnd)
14280 cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14282 cwnd = rack->r_ctl.cwnd_to_use;
14283 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14284 tr_perms = (cwnd * 1000) / srtt;
14285 if (tr_perms == 0) {
14286 tr_perms = ctf_fixed_maxseg(tp);
14289 * Calculate how long this will take to drain, if
14290 * the calculation comes out to zero, thats ok we
14291 * will use send_a_lot to possibly spin around for
14292 * more increasing tot_len_this_send to the point
14293 * that its going to require a pace, or we hit the
14294 * cwnd. Which in that case we are just waiting for
14297 slot = len / tr_perms;
14298 /* Now do we reduce the time so we don't run dry? */
14299 if (slot && rack_slot_reduction) {
14300 reduce = (slot / rack_slot_reduction);
14301 if (reduce < slot) {
14306 slot *= HPTS_USEC_IN_MSEC;
14309 * We always consider ourselves app limited with old style
14310 * that are not retransmits. This could be the initial
14311 * measurement, but thats ok its all setup and specially
14312 * handled. If another send leaks out, then that too will
14313 * be mark app-limited.
14315 lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14316 if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
14317 rack->r_ctl.rc_first_appl = lrsm;
14318 lrsm->r_flags |= RACK_APP_LIMITED;
14319 rack->r_ctl.rc_app_limited_cnt++;
14322 if (rack->rc_pace_to_cwnd) {
14323 uint64_t rate_wanted = 0;
14325 slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14326 rack->rc_ack_can_sendout_data = 1;
14327 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL);
14329 rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
14331 uint64_t bw_est, res, lentim, rate_wanted;
14332 uint32_t orig_val, srtt, segs, oh;
14336 if ((rack->r_rr_config == 1) && rsm) {
14337 return (rack->r_ctl.rc_min_to);
14339 if (rack->use_fixed_rate) {
14340 rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14341 } else if ((rack->r_ctl.init_rate == 0) &&
14342 #ifdef NETFLIX_PEAKRATE
14343 (rack->rc_tp->t_maxpeakrate == 0) &&
14345 (rack->r_ctl.gp_bw == 0)) {
14346 /* no way to yet do an estimate */
14347 bw_est = rate_wanted = 0;
14349 bw_est = rack_get_bw(rack);
14350 rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14352 if ((bw_est == 0) || (rate_wanted == 0) ||
14353 ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14355 * No way yet to make a b/w estimate or
14356 * our raise is set incorrectly.
14360 /* We need to account for all the overheads */
14361 segs = (len + segsiz - 1) / segsiz;
14363 * We need the diff between 1514 bytes (e-mtu with e-hdr)
14364 * and how much data we put in each packet. Yes this
14365 * means we may be off if we are larger than 1500 bytes
14366 * or smaller. But this just makes us more conservative.
14368 if (rack_hw_rate_min &&
14369 (bw_est < rack_hw_rate_min))
14370 can_start_hw_pacing = 0;
14371 if (ETHERNET_SEGMENT_SIZE > segsiz)
14372 oh = ETHERNET_SEGMENT_SIZE - segsiz;
14376 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14377 res = lentim / rate_wanted;
14378 slot = (uint32_t)res;
14379 orig_val = rack->r_ctl.rc_pace_max_segs;
14380 if (rack->r_ctl.crte == NULL) {
14382 * Only do this if we are not hardware pacing
14383 * since if we are doing hw-pacing below we will
14384 * set make a call after setting up or changing
14387 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14388 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14390 * We lost our rate somehow, this can happen
14391 * if the interface changed underneath us.
14393 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14394 rack->r_ctl.crte = NULL;
14395 /* Lets re-allow attempting to setup pacing */
14396 rack->rack_hdrw_pacing = 0;
14397 rack->rack_attempt_hdwr_pace = 0;
14398 rack_log_hdwr_pacing(rack,
14399 rate_wanted, bw_est, __LINE__,
14402 /* Did we change the TSO size, if so log it */
14403 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14404 rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
14405 prev_fill = rack->r_via_fill_cw;
14406 if ((rack->rc_pace_to_cwnd) &&
14408 (rack->use_fixed_rate == 0) &&
14409 (rack->in_probe_rtt == 0) &&
14410 (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14412 * We want to pace at our rate *or* faster to
14413 * fill the cwnd to the max if its not full.
14415 slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14417 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14418 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14419 if ((rack->rack_hdw_pace_ena) &&
14420 (can_start_hw_pacing > 0) &&
14421 (rack->rack_hdrw_pacing == 0) &&
14422 (rack->rack_attempt_hdwr_pace == 0)) {
14424 * Lets attempt to turn on hardware pacing
14427 rack->rack_attempt_hdwr_pace = 1;
14428 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14429 rack->rc_inp->inp_route.ro_nh->nh_ifp,
14432 &err, &rack->r_ctl.crte_prev_rate);
14433 if (rack->r_ctl.crte) {
14434 rack->rack_hdrw_pacing = 1;
14435 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14436 0, rack->r_ctl.crte,
14438 rack_log_hdwr_pacing(rack,
14439 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14441 rack->r_ctl.last_hw_bw_req = rate_wanted;
14443 counter_u64_add(rack_hw_pace_init_fail, 1);
14445 } else if (rack->rack_hdrw_pacing &&
14446 (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14447 /* Do we need to adjust our rate? */
14448 const struct tcp_hwrate_limit_table *nrte;
14450 if (rack->r_up_only &&
14451 (rate_wanted < rack->r_ctl.crte->rate)) {
14453 * We have four possible states here
14454 * having to do with the previous time
14456 * previous | this-time
14457 * A) 0 | 0 -- fill_cw not in the picture
14458 * B) 1 | 0 -- we were doing a fill-cw but now are not
14459 * C) 1 | 1 -- all rates from fill_cw
14460 * D) 0 | 1 -- we were doing non-fill and now we are filling
14462 * For case A, C and D we don't allow a drop. But for
14463 * case B where we now our on our steady rate we do
14467 if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14470 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14471 (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14472 if (rack_hw_rate_to_low &&
14473 (bw_est < rack_hw_rate_to_low)) {
14475 * The pacing rate is too low for hardware, but
14476 * do allow hardware pacing to be restarted.
14478 rack_log_hdwr_pacing(rack,
14479 bw_est, rack->r_ctl.crte->rate, __LINE__,
14481 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14482 rack->r_ctl.crte = NULL;
14483 rack->rack_attempt_hdwr_pace = 0;
14484 rack->rack_hdrw_pacing = 0;
14485 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14488 nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14490 rack->rc_inp->inp_route.ro_nh->nh_ifp,
14493 &err, &rack->r_ctl.crte_prev_rate);
14494 if (nrte == NULL) {
14495 /* Lost the rate */
14496 rack->rack_hdrw_pacing = 0;
14497 rack->r_ctl.crte = NULL;
14498 rack_log_hdwr_pacing(rack,
14499 rate_wanted, 0, __LINE__,
14501 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14502 counter_u64_add(rack_hw_pace_lost, 1);
14503 } else if (nrte != rack->r_ctl.crte) {
14504 rack->r_ctl.crte = nrte;
14505 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14509 rack_log_hdwr_pacing(rack,
14510 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14512 rack->r_ctl.last_hw_bw_req = rate_wanted;
14515 /* We just need to adjust the segment size */
14516 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14517 rack_log_hdwr_pacing(rack,
14518 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14520 rack->r_ctl.last_hw_bw_req = rate_wanted;
14524 if ((rack->r_ctl.crte != NULL) &&
14525 (rack->r_ctl.crte->rate == rate_wanted)) {
14527 * We need to add a extra if the rates
14528 * are exactly matched. The idea is
14529 * we want the software to make sure the
14530 * queue is empty before adding more, this
14531 * gives us N MSS extra pace times where
14534 slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14537 if (rack_limit_time_with_srtt &&
14538 (rack->use_fixed_rate == 0) &&
14539 #ifdef NETFLIX_PEAKRATE
14540 (rack->rc_tp->t_maxpeakrate == 0) &&
14542 (rack->rack_hdrw_pacing == 0)) {
14544 * Sanity check, we do not allow the pacing delay
14545 * to be longer than the SRTT of the path. If it is
14546 * a slow path, then adding a packet should increase
14547 * the RTT and compensate for this i.e. the srtt will
14548 * be greater so the allowed pacing time will be greater.
14550 * Note this restriction is not for where a peak rate
14551 * is set, we are doing fixed pacing or hardware pacing.
14553 if (rack->rc_tp->t_srtt)
14554 srtt = rack->rc_tp->t_srtt;
14556 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */
14558 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
14562 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
14564 if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14566 * If this rate is seeing enobufs when it
14567 * goes to send then either the nic is out
14568 * of gas or we are mis-estimating the time
14569 * somehow and not letting the queue empty
14570 * completely. Lets add to the pacing time.
14572 int hw_boost_delay;
14574 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14575 if (hw_boost_delay > rack_enobuf_hw_max)
14576 hw_boost_delay = rack_enobuf_hw_max;
14577 else if (hw_boost_delay < rack_enobuf_hw_min)
14578 hw_boost_delay = rack_enobuf_hw_min;
14579 slot += hw_boost_delay;
14582 counter_u64_add(rack_calc_nonzero, 1);
14584 counter_u64_add(rack_calc_zero, 1);
14589 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14590 tcp_seq startseq, uint32_t sb_offset)
14592 struct rack_sendmap *my_rsm = NULL;
14593 struct rack_sendmap fe;
14595 if (tp->t_state < TCPS_ESTABLISHED) {
14597 * We don't start any measurements if we are
14598 * not at least established.
14602 tp->t_flags |= TF_GPUTINPROG;
14603 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14604 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14605 tp->gput_seq = startseq;
14606 rack->app_limited_needs_set = 0;
14607 if (rack->in_probe_rtt)
14608 rack->measure_saw_probe_rtt = 1;
14609 else if ((rack->measure_saw_probe_rtt) &&
14610 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14611 rack->measure_saw_probe_rtt = 0;
14612 if (rack->rc_gp_filled)
14613 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14615 /* Special case initial measurement */
14618 tp->gput_ts = tcp_get_usecs(&tv);
14619 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14622 * We take a guess out into the future,
14623 * if we have no measurement and no
14624 * initial rate, we measure the first
14625 * initial-windows worth of data to
14626 * speed up getting some GP measurement and
14627 * thus start pacing.
14629 if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14630 rack->app_limited_needs_set = 1;
14631 tp->gput_ack = startseq + max(rc_init_window(rack),
14632 (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14633 rack_log_pacing_delay_calc(rack,
14638 rack->r_ctl.rc_app_limited_cnt,
14645 * We are out somewhere in the sb
14646 * can we use the already outstanding data?
14649 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14651 * Yes first one is good and in this case
14652 * the tp->gput_ts is correctly set based on
14653 * the last ack that arrived (no need to
14654 * set things up when an ack comes in).
14656 my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14657 if ((my_rsm == NULL) ||
14658 (my_rsm->r_rtr_cnt != 1)) {
14659 /* retransmission? */
14663 if (rack->r_ctl.rc_first_appl == NULL) {
14665 * If rc_first_appl is NULL
14666 * then the cnt should be 0.
14667 * This is probably an error, maybe
14668 * a KASSERT would be approprate.
14673 * If we have a marker pointer to the last one that is
14674 * app limited we can use that, but we need to set
14675 * things up so that when it gets ack'ed we record
14676 * the ack time (if its not already acked).
14678 rack->app_limited_needs_set = 1;
14680 * We want to get to the rsm that is either
14681 * next with space i.e. over 1 MSS or the one
14682 * after that (after the app-limited).
14684 my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14685 rack->r_ctl.rc_first_appl);
14687 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14688 /* Have to use the next one */
14689 my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14692 /* Use after the first MSS of it is acked */
14693 tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14697 if ((my_rsm == NULL) ||
14698 (my_rsm->r_rtr_cnt != 1)) {
14700 * Either its a retransmit or
14701 * the last is the app-limited one.
14706 tp->gput_seq = my_rsm->r_start;
14708 if (my_rsm->r_flags & RACK_ACKED) {
14710 * This one has been acked use the arrival ack time
14712 tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14713 rack->app_limited_needs_set = 0;
14715 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14716 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14717 rack_log_pacing_delay_calc(rack,
14722 rack->r_ctl.rc_app_limited_cnt,
14730 * We don't know how long we may have been
14731 * idle or if this is the first-send. Lets
14732 * setup the flag so we will trim off
14733 * the first ack'd data so we get a true
14736 rack->app_limited_needs_set = 1;
14737 tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14738 /* Find this guy so we can pull the send time */
14739 fe.r_start = startseq;
14740 my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14742 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14743 if (my_rsm->r_flags & RACK_ACKED) {
14745 * Unlikely since its probably what was
14746 * just transmitted (but I am paranoid).
14748 tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14749 rack->app_limited_needs_set = 0;
14751 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14752 /* This also is unlikely */
14753 tp->gput_seq = my_rsm->r_start;
14757 * TSNH unless we have some send-map limit,
14758 * and even at that it should not be hitting
14759 * that limit (we should have stopped sending).
14764 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14766 rack_log_pacing_delay_calc(rack,
14771 rack->r_ctl.rc_app_limited_cnt,
14772 9, __LINE__, NULL);
14775 static inline uint32_t
14776 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cwnd_to_use,
14777 uint32_t avail, int32_t sb_offset)
14782 if (tp->snd_wnd > cwnd_to_use)
14783 sendwin = cwnd_to_use;
14785 sendwin = tp->snd_wnd;
14786 if (ctf_outstanding(tp) >= tp->snd_wnd) {
14787 /* We never want to go over our peers rcv-window */
14792 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14793 if (flight >= sendwin) {
14795 * We have in flight what we are allowed by cwnd (if
14796 * it was rwnd blocking it would have hit above out
14801 len = sendwin - flight;
14802 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14803 /* We would send too much (beyond the rwnd) */
14804 len = tp->snd_wnd - ctf_outstanding(tp);
14806 if ((len + sb_offset) > avail) {
14808 * We don't have that much in the SB, how much is
14811 len = avail - sb_offset;
14818 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14819 unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14820 int rsm_is_null, int optlen, int line, uint16_t mode)
14822 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14823 union tcp_log_stackspecific log;
14826 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14827 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14828 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14829 log.u_bbr.flex1 = error;
14830 log.u_bbr.flex2 = flags;
14831 log.u_bbr.flex3 = rsm_is_null;
14832 log.u_bbr.flex4 = ipoptlen;
14833 log.u_bbr.flex5 = tp->rcv_numsacks;
14834 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14835 log.u_bbr.flex7 = optlen;
14836 log.u_bbr.flex8 = rack->r_fsb_inited;
14837 log.u_bbr.applimited = rack->r_fast_output;
14838 log.u_bbr.bw_inuse = rack_get_bw(rack);
14839 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
14840 log.u_bbr.cwnd_gain = mode;
14841 log.u_bbr.pkts_out = orig_len;
14842 log.u_bbr.lt_epoch = len;
14843 log.u_bbr.delivered = line;
14844 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14845 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14846 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
14847 len, &log, false, NULL, NULL, 0, &tv);
14852 static struct mbuf *
14853 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
14854 struct rack_fast_send_blk *fsb,
14855 int32_t seglimit, int32_t segsize)
14858 struct ktls_session *tls, *ntls;
14859 struct mbuf *start;
14861 struct mbuf *m, *n, **np, *smb;
14864 int32_t len = *plen;
14866 int32_t len_cp = 0;
14867 uint32_t mlen, frags;
14869 soff = off = the_off;
14874 if (hw_tls && (m->m_flags & M_EXTPG))
14875 tls = m->m_epg_tls;
14887 if (m->m_flags & M_EXTPG)
14888 ntls = m->m_epg_tls;
14893 * Avoid mixing TLS records with handshake
14894 * data or TLS records from different
14904 mlen = min(len, m->m_len - off);
14907 * For M_EXTPG mbufs, add 3 segments
14908 * + 1 in case we are crossing page boundaries
14909 * + 2 in case the TLS hdr/trailer are used
14910 * It is cheaper to just add the segments
14911 * than it is to take the cache miss to look
14912 * at the mbuf ext_pgs state in detail.
14914 if (m->m_flags & M_EXTPG) {
14915 fragsize = min(segsize, PAGE_SIZE);
14918 fragsize = segsize;
14922 /* Break if we really can't fit anymore. */
14923 if ((frags + 1) >= seglimit) {
14929 * Reduce size if you can't copy the whole
14930 * mbuf. If we can't copy the whole mbuf, also
14931 * adjust len so the loop will end after this
14934 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
14935 mlen = (seglimit - frags - 1) * fragsize;
14937 *plen = len_cp + len;
14939 frags += howmany(mlen, fragsize);
14943 KASSERT(seglimit > 0,
14944 ("%s: seglimit went too low", __func__));
14946 n = m_get(M_NOWAIT, m->m_type);
14952 len_cp += n->m_len;
14953 if (m->m_flags & (M_EXT|M_EXTPG)) {
14954 n->m_data = m->m_data + off;
14957 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
14964 if (len || (soff == smb->m_len)) {
14966 * We have more so we move forward or
14967 * we have consumed the entire mbuf and
14968 * len has fell to 0.
14980 * Save off the size of the mbuf. We do
14981 * this so that we can recognize when it
14982 * has been trimmed by sbcut() as acks
14985 fsb->o_m_len = smb->m_len;
14988 * This is the case where the next mbuf went to NULL. This
14989 * means with this copy we have sent everything in the sb.
14990 * In theory we could clear the fast_output flag, but lets
14991 * not since its possible that we could get more added
14992 * and acks that call the extend function which would let
15007 * This is a copy of m_copym(), taking the TSO segment size/limit
15008 * constraints into account, and advancing the sndptr as it goes.
15010 static struct mbuf *
15011 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15012 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15014 struct mbuf *m, *n;
15017 soff = rack->r_ctl.fsb.off;
15018 m = rack->r_ctl.fsb.m;
15019 if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15021 * The mbuf had the front of it chopped off by an ack
15022 * we need to adjust the soff/off by that difference.
15026 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15029 KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15030 KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15031 KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15033 rack, *plen, m, m->m_len));
15034 /* Save off the right location before we copy and advance */
15036 *s_mb = rack->r_ctl.fsb.m;
15037 n = rack_fo_base_copym(m, soff, plen,
15039 seglimit, segsize);
15044 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15045 uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len)
15048 * Enter the fast retransmit path. We are given that a sched_pin is
15049 * in place (if accounting is compliled in) and the cycle count taken
15050 * at the entry is in the ts_val. The concept her is that the rsm
15051 * now holds the mbuf offsets and such so we can directly transmit
15052 * without a lot of overhead, the len field is already set for
15053 * us to prohibit us from sending too much (usually its 1MSS).
15055 struct ip *ip = NULL;
15056 struct udphdr *udp = NULL;
15057 struct tcphdr *th = NULL;
15058 struct mbuf *m = NULL;
15061 struct tcp_log_buffer *lgb;
15062 #ifdef TCP_ACCOUNTING
15068 u_char opt[TCP_MAXOLEN];
15069 uint32_t hdrlen, optlen;
15070 int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15072 uint32_t if_hw_tsomaxsegcount = 0, startseq;
15073 uint32_t if_hw_tsomaxsegsize;
15075 struct ip6_hdr *ip6 = NULL;
15077 if (rack->r_is_v6) {
15078 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15079 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15083 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15084 hdrlen = sizeof(struct tcpiphdr);
15086 if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15089 if (rsm->r_flags & RACK_TLP)
15091 startseq = rsm->r_start;
15092 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15093 inp = rack->rc_inp;
15095 flags = tcp_outflags[tp->t_state];
15096 if (flags & (TH_SYN|TH_RST)) {
15099 if (rsm->r_flags & RACK_HAS_FIN) {
15100 /* We can't send a FIN here */
15103 if (flags & TH_FIN) {
15104 /* We never send a FIN */
15107 if (tp->t_flags & TF_RCVD_TSTMP) {
15108 to.to_tsval = ms_cts + tp->ts_offset;
15109 to.to_tsecr = tp->ts_recent;
15110 to.to_flags = TOF_TS;
15112 optlen = tcp_addoptions(&to, opt);
15114 udp = rack->r_ctl.fsb.udp;
15116 hdrlen += sizeof(struct udphdr);
15117 if (rack->r_ctl.rc_pace_max_segs)
15118 max_val = rack->r_ctl.rc_pace_max_segs;
15119 else if (rack->rc_user_set_max_segs)
15120 max_val = rack->rc_user_set_max_segs * segsiz;
15123 if ((tp->t_flags & TF_TSO) &&
15129 if (MHLEN < hdrlen + max_linkhdr)
15130 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15133 m = m_gethdr(M_NOWAIT, MT_DATA);
15136 m->m_data += max_linkhdr;
15138 th = rack->r_ctl.fsb.th;
15139 /* Establish the len to send */
15142 if ((tso) && (len + optlen > tp->t_maxseg)) {
15143 uint32_t if_hw_tsomax;
15146 /* extract TSO information */
15147 if_hw_tsomax = tp->t_tsomax;
15148 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15149 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15151 * Check if we should limit by maximum payload
15154 if (if_hw_tsomax != 0) {
15155 /* compute maximum TSO length */
15156 max_len = (if_hw_tsomax - hdrlen -
15158 if (max_len <= 0) {
15160 } else if (len > max_len) {
15164 if (len <= segsiz) {
15166 * In case there are too many small fragments don't
15174 if ((tso == 0) && (len > segsiz))
15176 us_cts = tcp_get_usecs(tv);
15178 (len <= MHLEN - hdrlen - max_linkhdr)) {
15181 th->th_seq = htonl(rsm->r_start);
15182 th->th_ack = htonl(tp->rcv_nxt);
15183 if(rsm->r_flags & RACK_HAD_PUSH)
15185 th->th_flags = flags;
15186 th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15187 if (th->th_win == 0) {
15188 tp->t_sndzerowin++;
15189 tp->t_flags |= TF_RXWIN0SENT;
15191 tp->t_flags &= ~TF_RXWIN0SENT;
15192 if (rsm->r_flags & RACK_TLP) {
15194 * TLP should not count in retran count, but
15197 counter_u64_add(rack_tlp_retran, 1);
15198 counter_u64_add(rack_tlp_retran_bytes, len);
15200 tp->t_sndrexmitpack++;
15201 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15202 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15205 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15208 if (rsm->m == NULL)
15210 if (rsm->orig_m_len != rsm->m->m_len) {
15211 /* Fix up the orig_m_len and possibly the mbuf offset */
15212 rack_adjust_orig_mlen(rsm);
15214 m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize);
15215 if (len <= segsiz) {
15217 * Must have ran out of mbufs for the copy
15218 * shorten it to no longer need tso. Lets
15219 * not put on sendalot since we are low on
15224 if ((m->m_next == NULL) || (len <= 0)){
15229 ulen = hdrlen + len - sizeof(struct ip6_hdr);
15231 ulen = hdrlen + len - sizeof(struct ip);
15232 udp->uh_ulen = htons(ulen);
15234 m->m_pkthdr.rcvif = (struct ifnet *)0;
15235 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15237 if (rack->r_is_v6) {
15239 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15240 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15241 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15242 th->th_sum = htons(0);
15243 UDPSTAT_INC(udps_opackets);
15245 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15246 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15247 th->th_sum = in6_cksum_pseudo(ip6,
15248 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15253 #if defined(INET6) && defined(INET)
15259 m->m_pkthdr.csum_flags = CSUM_UDP;
15260 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15261 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15262 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15263 th->th_sum = htons(0);
15264 UDPSTAT_INC(udps_opackets);
15266 m->m_pkthdr.csum_flags = CSUM_TCP;
15267 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15268 th->th_sum = in_pseudo(ip->ip_src.s_addr,
15269 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15270 IPPROTO_TCP + len + optlen));
15272 /* IP version must be set here for ipv4/ipv6 checking later */
15273 KASSERT(ip->ip_v == IPVERSION,
15274 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15278 KASSERT(len > tp->t_maxseg - optlen,
15279 ("%s: len <= tso_segsz tp:%p", __func__, tp));
15280 m->m_pkthdr.csum_flags |= CSUM_TSO;
15281 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15284 if (rack->r_is_v6) {
15285 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15286 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15287 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15288 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15290 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15293 #if defined(INET) && defined(INET6)
15298 ip->ip_len = htons(m->m_pkthdr.len);
15299 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15300 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15301 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15302 if (tp->t_port == 0 || len < V_tcp_minmss) {
15303 ip->ip_off |= htons(IP_DF);
15306 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15310 /* Time to copy in our header */
15311 cpto = mtod(m, uint8_t *);
15312 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15313 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15315 bcopy(opt, th + 1, optlen);
15316 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15318 th->th_off = sizeof(struct tcphdr) >> 2;
15320 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15321 union tcp_log_stackspecific log;
15323 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15324 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15325 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15326 if (rack->rack_no_prr)
15327 log.u_bbr.flex1 = 0;
15329 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15330 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15331 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15332 log.u_bbr.flex4 = max_val;
15333 log.u_bbr.flex5 = 0;
15334 /* Save off the early/late values */
15335 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15336 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15337 log.u_bbr.bw_inuse = rack_get_bw(rack);
15338 log.u_bbr.flex8 = 1;
15339 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15340 log.u_bbr.flex7 = 55;
15341 log.u_bbr.pkts_out = tp->t_maxseg;
15342 log.u_bbr.timeStamp = cts;
15343 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15344 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15345 log.u_bbr.delivered = 0;
15346 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15347 len, &log, false, NULL, NULL, 0, tv);
15351 if (rack->r_is_v6) {
15352 error = ip6_output(m, NULL,
15354 0, NULL, NULL, inp);
15357 #if defined(INET) && defined(INET6)
15362 error = ip_output(m, NULL,
15369 lgb->tlb_errno = error;
15375 rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15376 rsm, RACK_SENT_FP, rsm->m, rsm->soff);
15377 if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15378 rack->rc_tlp_in_progress = 1;
15379 rack->r_ctl.rc_tlp_cnt_out++;
15381 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15382 rack->forced_ack = 0; /* If we send something zap the FA flag */
15383 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15384 rack->r_ctl.retran_during_recovery += len;
15388 idx = (len / segsiz) + 3;
15389 if (idx >= TCP_MSS_ACCT_ATIMER)
15390 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15392 counter_u64_add(rack_out_size[idx], 1);
15394 if (tp->t_rtttime == 0) {
15395 tp->t_rtttime = ticks;
15396 tp->t_rtseq = startseq;
15397 KMOD_TCPSTAT_INC(tcps_segstimed);
15399 counter_u64_add(rack_fto_rsm_send, 1);
15400 if (error && (error == ENOBUFS)) {
15401 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15402 if (rack->rc_enobuf < 0x7f)
15404 if (slot < (10 * HPTS_USEC_IN_MSEC))
15405 slot = 10 * HPTS_USEC_IN_MSEC;
15407 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15409 (rack->rc_always_pace == 0) ||
15410 (rack->r_rr_config == 1)) {
15412 * We have no pacing set or we
15413 * are using old-style rack or
15414 * we are overriden to use the old 1ms pacing.
15416 slot = rack->r_ctl.rc_min_to;
15418 rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15419 if (rack->r_must_retran) {
15420 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15421 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15423 * We have retransmitted all we need.
15425 rack->r_must_retran = 0;
15426 rack->r_ctl.rc_out_at_rto = 0;
15429 #ifdef TCP_ACCOUNTING
15430 crtsc = get_cyclecount();
15431 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15432 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15434 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15435 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15436 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15438 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15439 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15440 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15442 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15453 rack_sndbuf_autoscale(struct tcp_rack *rack)
15456 * Automatic sizing of send socket buffer. Often the send buffer
15457 * size is not optimally adjusted to the actual network conditions
15458 * at hand (delay bandwidth product). Setting the buffer size too
15459 * small limits throughput on links with high bandwidth and high
15460 * delay (eg. trans-continental/oceanic links). Setting the
15461 * buffer size too big consumes too much real kernel memory,
15462 * especially with many connections on busy servers.
15464 * The criteria to step up the send buffer one notch are:
15465 * 1. receive window of remote host is larger than send buffer
15466 * (with a fudge factor of 5/4th);
15467 * 2. send buffer is filled to 7/8th with data (so we actually
15468 * have data to make use of it);
15469 * 3. send buffer fill has not hit maximal automatic size;
15470 * 4. our send window (slow start and cogestion controlled) is
15471 * larger than sent but unacknowledged data in send buffer.
15473 * Note that the rack version moves things much faster since
15474 * we want to avoid hitting cache lines in the rack_fast_output()
15475 * path so this is called much less often and thus moves
15476 * the SB forward by a percentage.
15480 uint32_t sendwin, scaleup;
15483 so = rack->rc_inp->inp_socket;
15484 sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15485 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15486 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15487 sbused(&so->so_snd) >=
15488 (so->so_snd.sb_hiwat / 8 * 7) &&
15489 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15490 sendwin >= (sbused(&so->so_snd) -
15491 (tp->snd_nxt - tp->snd_una))) {
15492 if (rack_autosndbuf_inc)
15493 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15495 scaleup = V_tcp_autosndbuf_inc;
15496 if (scaleup < V_tcp_autosndbuf_inc)
15497 scaleup = V_tcp_autosndbuf_inc;
15498 scaleup += so->so_snd.sb_hiwat;
15499 if (scaleup > V_tcp_autosndbuf_max)
15500 scaleup = V_tcp_autosndbuf_max;
15501 if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15502 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15508 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15509 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15512 * Enter to do fast output. We are given that the sched_pin is
15513 * in place (if accounting is compiled in) and the cycle count taken
15514 * at entry is in place in ts_val. The idea here is that
15515 * we know how many more bytes needs to be sent (presumably either
15516 * during pacing or to fill the cwnd and that was greater than
15517 * the max-burst). We have how much to send and all the info we
15518 * need to just send.
15520 struct ip *ip = NULL;
15521 struct udphdr *udp = NULL;
15522 struct tcphdr *th = NULL;
15523 struct mbuf *m, *s_mb;
15526 struct tcp_log_buffer *lgb;
15527 #ifdef TCP_ACCOUNTING
15531 u_char opt[TCP_MAXOLEN];
15532 uint32_t hdrlen, optlen;
15534 int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
15535 uint32_t us_cts, s_soff;
15536 uint32_t if_hw_tsomaxsegcount = 0, startseq;
15537 uint32_t if_hw_tsomaxsegsize;
15538 uint16_t add_flag = RACK_SENT_FP;
15540 struct ip6_hdr *ip6 = NULL;
15542 if (rack->r_is_v6) {
15543 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15544 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15548 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15549 hdrlen = sizeof(struct tcpiphdr);
15551 if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15555 startseq = tp->snd_max;
15556 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15557 inp = rack->rc_inp;
15558 len = rack->r_ctl.fsb.left_to_send;
15560 flags = rack->r_ctl.fsb.tcp_flags;
15561 if (tp->t_flags & TF_RCVD_TSTMP) {
15562 to.to_tsval = ms_cts + tp->ts_offset;
15563 to.to_tsecr = tp->ts_recent;
15564 to.to_flags = TOF_TS;
15566 optlen = tcp_addoptions(&to, opt);
15568 udp = rack->r_ctl.fsb.udp;
15570 hdrlen += sizeof(struct udphdr);
15571 if (rack->r_ctl.rc_pace_max_segs)
15572 max_val = rack->r_ctl.rc_pace_max_segs;
15573 else if (rack->rc_user_set_max_segs)
15574 max_val = rack->rc_user_set_max_segs * segsiz;
15577 if ((tp->t_flags & TF_TSO) &&
15584 if (MHLEN < hdrlen + max_linkhdr)
15585 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15588 m = m_gethdr(M_NOWAIT, MT_DATA);
15591 m->m_data += max_linkhdr;
15593 th = rack->r_ctl.fsb.th;
15594 /* Establish the len to send */
15597 if ((tso) && (len + optlen > tp->t_maxseg)) {
15598 uint32_t if_hw_tsomax;
15601 /* extract TSO information */
15602 if_hw_tsomax = tp->t_tsomax;
15603 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15604 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15606 * Check if we should limit by maximum payload
15609 if (if_hw_tsomax != 0) {
15610 /* compute maximum TSO length */
15611 max_len = (if_hw_tsomax - hdrlen -
15613 if (max_len <= 0) {
15615 } else if (len > max_len) {
15619 if (len <= segsiz) {
15621 * In case there are too many small fragments don't
15629 if ((tso == 0) && (len > segsiz))
15631 us_cts = tcp_get_usecs(tv);
15633 (len <= MHLEN - hdrlen - max_linkhdr)) {
15636 sb_offset = tp->snd_max - tp->snd_una;
15637 th->th_seq = htonl(tp->snd_max);
15638 th->th_ack = htonl(tp->rcv_nxt);
15639 th->th_flags = flags;
15640 th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15641 if (th->th_win == 0) {
15642 tp->t_sndzerowin++;
15643 tp->t_flags |= TF_RXWIN0SENT;
15645 tp->t_flags &= ~TF_RXWIN0SENT;
15646 tp->snd_up = tp->snd_una; /* drag it along, its deprecated */
15647 KMOD_TCPSTAT_INC(tcps_sndpack);
15648 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15650 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15653 if (rack->r_ctl.fsb.m == NULL)
15656 /* s_mb and s_soff are saved for rack_log_output */
15657 m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff);
15658 if (len <= segsiz) {
15660 * Must have ran out of mbufs for the copy
15661 * shorten it to no longer need tso. Lets
15662 * not put on sendalot since we are low on
15667 if (rack->r_ctl.fsb.rfo_apply_push &&
15668 (len == rack->r_ctl.fsb.left_to_send)) {
15669 th->th_flags |= TH_PUSH;
15670 add_flag |= RACK_HAD_PUSH;
15672 if ((m->m_next == NULL) || (len <= 0)){
15677 ulen = hdrlen + len - sizeof(struct ip6_hdr);
15679 ulen = hdrlen + len - sizeof(struct ip);
15680 udp->uh_ulen = htons(ulen);
15682 m->m_pkthdr.rcvif = (struct ifnet *)0;
15683 if (tp->t_state == TCPS_ESTABLISHED &&
15684 (tp->t_flags2 & TF2_ECN_PERMIT)) {
15686 * If the peer has ECN, mark data packets with ECN capable
15687 * transmission (ECT). Ignore pure ack packets,
15690 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15693 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15696 ip->ip_tos |= IPTOS_ECN_ECT0;
15697 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15699 * Reply with proper ECN notifications.
15700 * Only set CWR on new data segments.
15702 if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15704 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15707 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15710 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15712 if (rack->r_is_v6) {
15714 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15715 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15716 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15717 th->th_sum = htons(0);
15718 UDPSTAT_INC(udps_opackets);
15720 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15721 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15722 th->th_sum = in6_cksum_pseudo(ip6,
15723 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15728 #if defined(INET6) && defined(INET)
15734 m->m_pkthdr.csum_flags = CSUM_UDP;
15735 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15736 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15737 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15738 th->th_sum = htons(0);
15739 UDPSTAT_INC(udps_opackets);
15741 m->m_pkthdr.csum_flags = CSUM_TCP;
15742 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15743 th->th_sum = in_pseudo(ip->ip_src.s_addr,
15744 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15745 IPPROTO_TCP + len + optlen));
15747 /* IP version must be set here for ipv4/ipv6 checking later */
15748 KASSERT(ip->ip_v == IPVERSION,
15749 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15753 KASSERT(len > tp->t_maxseg - optlen,
15754 ("%s: len <= tso_segsz tp:%p", __func__, tp));
15755 m->m_pkthdr.csum_flags |= CSUM_TSO;
15756 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15759 if (rack->r_is_v6) {
15760 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15761 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15762 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15763 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15765 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15768 #if defined(INET) && defined(INET6)
15773 ip->ip_len = htons(m->m_pkthdr.len);
15774 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15775 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15776 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15777 if (tp->t_port == 0 || len < V_tcp_minmss) {
15778 ip->ip_off |= htons(IP_DF);
15781 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15785 /* Time to copy in our header */
15786 cpto = mtod(m, uint8_t *);
15787 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15788 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15790 bcopy(opt, th + 1, optlen);
15791 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15793 th->th_off = sizeof(struct tcphdr) >> 2;
15795 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15796 union tcp_log_stackspecific log;
15798 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15799 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15800 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15801 if (rack->rack_no_prr)
15802 log.u_bbr.flex1 = 0;
15804 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15805 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15806 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15807 log.u_bbr.flex4 = max_val;
15808 log.u_bbr.flex5 = 0;
15809 /* Save off the early/late values */
15810 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15811 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15812 log.u_bbr.bw_inuse = rack_get_bw(rack);
15813 log.u_bbr.flex8 = 0;
15814 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15815 log.u_bbr.flex7 = 44;
15816 log.u_bbr.pkts_out = tp->t_maxseg;
15817 log.u_bbr.timeStamp = cts;
15818 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15819 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15820 log.u_bbr.delivered = 0;
15821 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15822 len, &log, false, NULL, NULL, 0, tv);
15826 if (rack->r_is_v6) {
15827 error = ip6_output(m, NULL,
15829 0, NULL, NULL, inp);
15832 #if defined(INET) && defined(INET6)
15837 error = ip_output(m, NULL,
15843 lgb->tlb_errno = error;
15851 rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
15852 NULL, add_flag, s_mb, s_soff);
15854 if (tp->snd_una == tp->snd_max) {
15855 rack->r_ctl.rc_tlp_rxt_last_time = cts;
15856 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
15857 tp->t_acktime = ticks;
15859 rack->forced_ack = 0; /* If we send something zap the FA flag */
15861 if ((tp->t_flags & TF_GPUTINPROG) == 0)
15862 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
15863 tp->snd_max += len;
15864 tp->snd_nxt = tp->snd_max;
15868 idx = (len / segsiz) + 3;
15869 if (idx >= TCP_MSS_ACCT_ATIMER)
15870 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15872 counter_u64_add(rack_out_size[idx], 1);
15874 if (len <= rack->r_ctl.fsb.left_to_send)
15875 rack->r_ctl.fsb.left_to_send -= len;
15877 rack->r_ctl.fsb.left_to_send = 0;
15878 if (rack->r_ctl.fsb.left_to_send < segsiz) {
15879 rack->r_fast_output = 0;
15880 rack->r_ctl.fsb.left_to_send = 0;
15881 /* At the end of fast_output scale up the sb */
15882 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
15883 rack_sndbuf_autoscale(rack);
15884 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
15886 if (tp->t_rtttime == 0) {
15887 tp->t_rtttime = ticks;
15888 tp->t_rtseq = startseq;
15889 KMOD_TCPSTAT_INC(tcps_segstimed);
15891 if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
15896 th = rack->r_ctl.fsb.th;
15900 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15901 counter_u64_add(rack_fto_send, 1);
15902 slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
15903 rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
15904 #ifdef TCP_ACCOUNTING
15905 crtsc = get_cyclecount();
15906 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15907 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15909 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15910 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15911 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15913 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15914 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15915 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
15917 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
15924 rack->r_fast_output = 0;
15929 rack_output(struct tcpcb *tp)
15933 uint32_t sb_offset, s_moff = 0;
15934 int32_t len, flags, error = 0;
15935 struct mbuf *m, *s_mb = NULL;
15937 uint32_t if_hw_tsomaxsegcount = 0;
15938 uint32_t if_hw_tsomaxsegsize;
15939 int32_t segsiz, minseg;
15940 long tot_len_this_send = 0;
15942 struct ip *ip = NULL;
15945 struct ipovly *ipov = NULL;
15947 struct udphdr *udp = NULL;
15948 struct tcp_rack *rack;
15952 uint8_t wanted_cookie = 0;
15953 u_char opt[TCP_MAXOLEN];
15954 unsigned ipoptlen, optlen, hdrlen, ulen=0;
15957 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
15958 unsigned ipsec_optlen = 0;
15961 int32_t idle, sendalot;
15962 int32_t sub_from_prr = 0;
15963 volatile int32_t sack_rxmit;
15964 struct rack_sendmap *rsm = NULL;
15968 int32_t sup_rack = 0;
15969 uint32_t cts, ms_cts, delayed, early;
15970 uint16_t add_flag = RACK_SENT_SP;
15971 uint8_t hpts_calling, doing_tlp = 0;
15972 uint32_t cwnd_to_use, pace_max_seg;
15973 int32_t do_a_prefetch = 0;
15974 int32_t prefetch_rsm = 0;
15975 int32_t orig_len = 0;
15977 int32_t prefetch_so_done = 0;
15978 struct tcp_log_buffer *lgb;
15980 struct sockbuf *sb;
15981 uint64_t ts_val = 0;
15982 #ifdef TCP_ACCOUNTING
15986 struct ip6_hdr *ip6 = NULL;
15989 uint8_t filled_all = 0;
15990 bool hw_tls = false;
15992 /* setup and take the cache hits here */
15993 rack = (struct tcp_rack *)tp->t_fb_ptr;
15994 #ifdef TCP_ACCOUNTING
15996 ts_val = get_cyclecount();
15998 hpts_calling = rack->rc_inp->inp_hpts_calls;
15999 NET_EPOCH_ASSERT();
16000 INP_WLOCK_ASSERT(rack->rc_inp);
16002 if (tp->t_flags & TF_TOE) {
16003 #ifdef TCP_ACCOUNTING
16006 return (tcp_offload_output(tp));
16010 * For TFO connections in SYN_RECEIVED, only allow the initial
16011 * SYN|ACK and those sent by the retransmit timer.
16013 if (IS_FASTOPEN(tp->t_flags) &&
16014 (tp->t_state == TCPS_SYN_RECEIVED) &&
16015 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
16016 (rack->r_ctl.rc_resend == NULL)) { /* not a retransmit */
16017 #ifdef TCP_ACCOUNTING
16023 if (rack->r_state) {
16024 /* Use the cache line loaded if possible */
16025 isipv6 = rack->r_is_v6;
16027 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16031 cts = tcp_get_usecs(&tv);
16032 ms_cts = tcp_tv_to_mssectick(&tv);
16033 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16034 rack->rc_inp->inp_in_hpts) {
16036 * We are on the hpts for some timer but not hptsi output.
16037 * Remove from the hpts unconditionally.
16039 rack_timer_cancel(tp, rack, cts, __LINE__);
16041 /* Are we pacing and late? */
16042 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16043 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16044 /* We are delayed */
16045 delayed = cts - rack->r_ctl.rc_last_output_to;
16049 /* Do the timers, which may override the pacer */
16050 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16051 if (rack_process_timers(tp, rack, cts, hpts_calling)) {
16052 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16053 #ifdef TCP_ACCOUNTING
16059 if (rack->rc_in_persist) {
16060 if (rack->rc_inp->inp_in_hpts == 0) {
16061 /* Timer is not running */
16062 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16064 #ifdef TCP_ACCOUNTING
16069 if ((rack->r_timer_override) ||
16070 (rack->rc_ack_can_sendout_data) ||
16072 (tp->t_state < TCPS_ESTABLISHED)) {
16073 rack->rc_ack_can_sendout_data = 0;
16074 if (rack->rc_inp->inp_in_hpts)
16075 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16076 } else if (rack->rc_inp->inp_in_hpts) {
16078 * On the hpts you can't pass even if ACKNOW is on, we will
16079 * when the hpts fires.
16081 #ifdef TCP_ACCOUNTING
16082 crtsc = get_cyclecount();
16083 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16084 tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16086 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16087 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16088 tp->tcp_cnt_counters[SND_BLOCKED]++;
16090 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16093 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16096 rack->rc_inp->inp_hpts_calls = 0;
16097 /* Finish out both pacing early and late accounting */
16098 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16099 TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16100 early = rack->r_ctl.rc_last_output_to - cts;
16104 rack->r_ctl.rc_agg_delayed += delayed;
16106 } else if (early) {
16107 rack->r_ctl.rc_agg_early += early;
16110 /* Now that early/late accounting is done turn off the flag */
16111 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16112 rack->r_wanted_output = 0;
16113 rack->r_timer_override = 0;
16114 if ((tp->t_state != rack->r_state) &&
16115 TCPS_HAVEESTABLISHED(tp->t_state)) {
16116 rack_set_state(tp, rack);
16118 if ((rack->r_fast_output) &&
16119 (tp->rcv_numsacks == 0)) {
16123 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16127 inp = rack->rc_inp;
16128 so = inp->inp_socket;
16133 inp = rack->rc_inp;
16135 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16136 * only allow the initial SYN or SYN|ACK and those sent
16137 * by the retransmit timer.
16139 if (IS_FASTOPEN(tp->t_flags) &&
16140 ((tp->t_state == TCPS_SYN_RECEIVED) ||
16141 (tp->t_state == TCPS_SYN_SENT)) &&
16142 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16143 (tp->t_rxtshift == 0)) { /* not a retransmit */
16144 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16145 so = inp->inp_socket;
16147 goto just_return_nolock;
16150 * Determine length of data that should be transmitted, and flags
16151 * that will be used. If there is some data or critical controls
16152 * (SYN, RST) to send, then transmit; otherwise, investigate
16155 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16156 if (tp->t_idle_reduce) {
16157 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16158 rack_cc_after_idle(rack, tp);
16160 tp->t_flags &= ~TF_LASTIDLE;
16162 if (tp->t_flags & TF_MORETOCOME) {
16163 tp->t_flags |= TF_LASTIDLE;
16167 if ((tp->snd_una == tp->snd_max) &&
16168 rack->r_ctl.rc_went_idle_time &&
16169 TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16170 idle = cts - rack->r_ctl.rc_went_idle_time;
16171 if (idle > rack_min_probertt_hold) {
16172 /* Count as a probe rtt */
16173 if (rack->in_probe_rtt == 0) {
16174 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16175 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16176 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16177 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16179 rack_exit_probertt(rack, cts);
16184 if (rack_use_fsb && (rack->r_fsb_inited == 0))
16185 rack_init_fsb_block(tp, rack);
16188 * If we've recently taken a timeout, snd_max will be greater than
16189 * snd_nxt. There may be SACK information that allows us to avoid
16190 * resending already delivered data. Adjust snd_nxt accordingly.
16193 cts = tcp_get_usecs(&tv);
16194 ms_cts = tcp_tv_to_mssectick(&tv);
16197 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16199 if (rack->r_ctl.rc_pace_max_segs == 0)
16200 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16202 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16203 sb_offset = tp->snd_max - tp->snd_una;
16204 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16205 flags = tcp_outflags[tp->t_state];
16206 while (rack->rc_free_cnt < rack_free_cache) {
16207 rsm = rack_alloc(rack);
16209 if (inp->inp_hpts_calls)
16210 /* Retry in a ms */
16211 slot = (1 * HPTS_USEC_IN_MSEC);
16212 so = inp->inp_socket;
16214 goto just_return_nolock;
16216 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16217 rack->rc_free_cnt++;
16220 if (inp->inp_hpts_calls)
16221 inp->inp_hpts_calls = 0;
16225 if (flags & TH_RST) {
16226 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16227 so = inp->inp_socket;
16231 if (rack->r_ctl.rc_resend) {
16232 /* Retransmit timer */
16233 rsm = rack->r_ctl.rc_resend;
16234 rack->r_ctl.rc_resend = NULL;
16235 rsm->r_flags &= ~RACK_TLP;
16236 len = rsm->r_end - rsm->r_start;
16239 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16240 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16241 __func__, __LINE__,
16242 rsm->r_start, tp->snd_una, tp, rack, rsm));
16243 sb_offset = rsm->r_start - tp->snd_una;
16246 } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16247 /* We have a retransmit that takes precedence */
16248 rsm->r_flags &= ~RACK_TLP;
16249 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16250 ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16251 /* Enter recovery if not induced by a time-out */
16252 rack->r_ctl.rc_rsm_start = rsm->r_start;
16253 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16254 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16255 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16258 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16259 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16260 tp, rack, rsm, rsm->r_start, tp->snd_una);
16263 len = rsm->r_end - rsm->r_start;
16264 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16265 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16266 __func__, __LINE__,
16267 rsm->r_start, tp->snd_una, tp, rack, rsm));
16268 sb_offset = rsm->r_start - tp->snd_una;
16274 KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16275 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16277 counter_u64_add(rack_rtm_prr_retran, 1);
16279 } else if (rack->r_ctl.rc_tlpsend) {
16280 /* Tail loss probe */
16286 * Check if we can do a TLP with a RACK'd packet
16287 * this can happen if we are not doing the rack
16288 * cheat and we skipped to a TLP and it
16291 rsm = rack->r_ctl.rc_tlpsend;
16292 rsm->r_flags |= RACK_TLP;
16294 rack->r_ctl.rc_tlpsend = NULL;
16296 tlen = rsm->r_end - rsm->r_start;
16299 tp->t_sndtlppack++;
16300 tp->t_sndtlpbyte += tlen;
16301 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16302 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16303 __func__, __LINE__,
16304 rsm->r_start, tp->snd_una, tp, rack, rsm));
16305 sb_offset = rsm->r_start - tp->snd_una;
16306 cwin = min(tp->snd_wnd, tlen);
16309 if (rack->r_must_retran &&
16312 * Non-Sack and we had a RTO or MTU change, we
16313 * need to retransmit until we reach
16314 * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16316 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16317 int sendwin, flight;
16319 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16320 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16321 if (flight >= sendwin) {
16322 so = inp->inp_socket;
16324 goto just_return_nolock;
16326 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16327 KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16330 rack->r_must_retran = 0;
16331 rack->r_ctl.rc_out_at_rto = 0;
16332 rack->r_must_retran = 0;
16333 so = inp->inp_socket;
16335 goto just_return_nolock;
16338 len = rsm->r_end - rsm->r_start;
16340 sb_offset = rsm->r_start - tp->snd_una;
16344 /* We must be done if there is nothing outstanding */
16345 rack->r_must_retran = 0;
16346 rack->r_ctl.rc_out_at_rto = 0;
16350 * Enforce a connection sendmap count limit if set
16351 * as long as we are not retransmiting.
16353 if ((rsm == NULL) &&
16354 (rack->do_detection == 0) &&
16355 (V_tcp_map_entries_limit > 0) &&
16356 (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16357 counter_u64_add(rack_to_alloc_limited, 1);
16358 if (!rack->alloc_limit_reported) {
16359 rack->alloc_limit_reported = 1;
16360 counter_u64_add(rack_alloc_limited_conns, 1);
16362 so = inp->inp_socket;
16364 goto just_return_nolock;
16366 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16367 /* we are retransmitting the fin */
16371 * When retransmitting data do *not* include the
16372 * FIN. This could happen from a TLP probe.
16378 /* For debugging */
16379 rack->r_ctl.rc_rsm_at_retran = rsm;
16381 if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16382 ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16385 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len);
16389 so = inp->inp_socket;
16391 if (do_a_prefetch == 0) {
16392 kern_prefetch(sb, &do_a_prefetch);
16395 #ifdef NETFLIX_SHARED_CWND
16396 if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16397 rack->rack_enable_scwnd) {
16398 /* We are doing cwnd sharing */
16399 if (rack->gp_ready &&
16400 (rack->rack_attempted_scwnd == 0) &&
16401 (rack->r_ctl.rc_scw == NULL) &&
16403 /* The pcbid is in, lets make an attempt */
16404 counter_u64_add(rack_try_scwnd, 1);
16405 rack->rack_attempted_scwnd = 1;
16406 rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16407 &rack->r_ctl.rc_scw_index,
16410 if (rack->r_ctl.rc_scw &&
16411 (rack->rack_scwnd_is_idle == 1) &&
16412 sbavail(&so->so_snd)) {
16413 /* we are no longer out of data */
16414 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16415 rack->rack_scwnd_is_idle = 0;
16417 if (rack->r_ctl.rc_scw) {
16418 /* First lets update and get the cwnd */
16419 rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16420 rack->r_ctl.rc_scw_index,
16421 tp->snd_cwnd, tp->snd_wnd, segsiz);
16426 * Get standard flags, and add SYN or FIN if requested by 'hidden'
16429 if (tp->t_flags & TF_NEEDFIN)
16431 if (tp->t_flags & TF_NEEDSYN)
16433 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16435 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16437 kern_prefetch(end_rsm, &prefetch_rsm);
16442 * If snd_nxt == snd_max and we have transmitted a FIN, the
16443 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16444 * negative length. This can also occur when TCP opens up its
16445 * congestion window while receiving additional duplicate acks after
16446 * fast-retransmit because TCP will reset snd_nxt to snd_max after
16447 * the fast-retransmit.
16449 * In the normal retransmit-FIN-only case, however, snd_nxt will be
16450 * set to snd_una, the sb_offset will be 0, and the length may wind
16453 * If sack_rxmit is true we are retransmitting from the scoreboard
16454 * in which case len is already set.
16456 if ((sack_rxmit == 0) &&
16457 (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16460 avail = sbavail(sb);
16461 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16462 sb_offset = tp->snd_nxt - tp->snd_una;
16465 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16466 if (rack->r_ctl.rc_tlp_new_data) {
16467 /* TLP is forcing out new data */
16468 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16469 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16471 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16472 if (tp->snd_wnd > sb_offset)
16473 len = tp->snd_wnd - sb_offset;
16477 len = rack->r_ctl.rc_tlp_new_data;
16479 rack->r_ctl.rc_tlp_new_data = 0;
16482 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16484 if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16486 * For prr=off, we need to send only 1 MSS
16487 * at a time. We do this because another sack could
16488 * be arriving that causes us to send retransmits and
16489 * we don't want to be on a long pace due to a larger send
16490 * that keeps us from sending out the retransmit.
16495 uint32_t outstanding;
16497 * We are inside of a Fast recovery episode, this
16498 * is caused by a SACK or 3 dup acks. At this point
16499 * we have sent all the retransmissions and we rely
16500 * on PRR to dictate what we will send in the form of
16504 outstanding = tp->snd_max - tp->snd_una;
16505 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16506 if (tp->snd_wnd > outstanding) {
16507 len = tp->snd_wnd - outstanding;
16508 /* Check to see if we have the data */
16509 if ((sb_offset + len) > avail) {
16510 /* It does not all fit */
16511 if (avail > sb_offset)
16512 len = avail - sb_offset;
16519 } else if (avail > sb_offset) {
16520 len = avail - sb_offset;
16525 if (len > rack->r_ctl.rc_prr_sndcnt) {
16526 len = rack->r_ctl.rc_prr_sndcnt;
16530 counter_u64_add(rack_rtm_prr_newdata, 1);
16533 if (len > segsiz) {
16535 * We should never send more than a MSS when
16536 * retransmitting or sending new data in prr
16537 * mode unless the override flag is on. Most
16538 * likely the PRR algorithm is not going to
16539 * let us send a lot as well :-)
16541 if (rack->r_ctl.rc_prr_sendalot == 0) {
16544 } else if (len < segsiz) {
16546 * Do we send any? The idea here is if the
16547 * send empty's the socket buffer we want to
16548 * do it. However if not then lets just wait
16549 * for our prr_sndcnt to get bigger.
16553 leftinsb = sbavail(sb) - sb_offset;
16554 if (leftinsb > len) {
16555 /* This send does not empty the sb */
16560 } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16562 * If you have not established
16563 * and are not doing FAST OPEN
16566 if ((sack_rxmit == 0) &&
16567 (!IS_FASTOPEN(tp->t_flags))){
16572 if (prefetch_so_done == 0) {
16573 kern_prefetch(so, &prefetch_so_done);
16574 prefetch_so_done = 1;
16577 * Lop off SYN bit if it has already been sent. However, if this is
16578 * SYN-SENT state and if segment contains data and if we don't know
16579 * that foreign host supports TAO, suppress sending segment.
16581 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16582 ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16584 * When sending additional segments following a TFO SYN|ACK,
16585 * do not include the SYN bit.
16587 if (IS_FASTOPEN(tp->t_flags) &&
16588 (tp->t_state == TCPS_SYN_RECEIVED))
16592 * Be careful not to send data and/or FIN on SYN segments. This
16593 * measure is needed to prevent interoperability problems with not
16594 * fully conformant TCP implementations.
16596 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16601 * On TFO sockets, ensure no data is sent in the following cases:
16603 * - When retransmitting SYN|ACK on a passively-created socket
16605 * - When retransmitting SYN on an actively created socket
16607 * - When sending a zero-length cookie (cookie request) on an
16608 * actively created socket
16610 * - When the socket is in the CLOSED state (RST is being sent)
16612 if (IS_FASTOPEN(tp->t_flags) &&
16613 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16614 ((tp->t_state == TCPS_SYN_SENT) &&
16615 (tp->t_tfo_client_cookie_len == 0)) ||
16616 (flags & TH_RST))) {
16620 /* Without fast-open there should never be data sent on a SYN */
16621 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16622 tp->snd_nxt = tp->iss;
16625 if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16626 /* We only send 1 MSS if we have a DSACK block */
16627 add_flag |= RACK_SENT_W_DSACK;
16633 * If FIN has been sent but not acked, but we haven't been
16634 * called to retransmit, len will be < 0. Otherwise, window
16635 * shrank after we sent into it. If window shrank to 0,
16636 * cancel pending retransmit, pull snd_nxt back to (closed)
16637 * window, and set the persist timer if it isn't already
16638 * going. If the window didn't close completely, just wait
16641 * We also do a general check here to ensure that we will
16642 * set the persist timer when we have data to send, but a
16643 * 0-byte window. This makes sure the persist timer is set
16644 * even if the packet hits one of the "goto send" lines
16648 if ((tp->snd_wnd == 0) &&
16649 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16650 (tp->snd_una == tp->snd_max) &&
16651 (sb_offset < (int)sbavail(sb))) {
16652 rack_enter_persist(tp, rack, cts);
16654 } else if ((rsm == NULL) &&
16655 (doing_tlp == 0) &&
16656 (len < pace_max_seg)) {
16658 * We are not sending a maximum sized segment for
16659 * some reason. Should we not send anything (think
16660 * sws or persists)?
16662 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16663 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16665 (len < (int)(sbavail(sb) - sb_offset))) {
16667 * Here the rwnd is less than
16668 * the minimum pacing size, this is not a retransmit,
16669 * we are established and
16670 * the send is not the last in the socket buffer
16671 * we send nothing, and we may enter persists
16672 * if nothing is outstanding.
16675 if (tp->snd_max == tp->snd_una) {
16677 * Nothing out we can
16678 * go into persists.
16680 rack_enter_persist(tp, rack, cts);
16682 } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16683 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16684 (len < (int)(sbavail(sb) - sb_offset)) &&
16687 * Here we are not retransmitting, and
16688 * the cwnd is not so small that we could
16689 * not send at least a min size (rxt timer
16690 * not having gone off), We have 2 segments or
16691 * more already in flight, its not the tail end
16692 * of the socket buffer and the cwnd is blocking
16693 * us from sending out a minimum pacing segment size.
16694 * Lets not send anything.
16697 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16698 min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16699 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16700 (len < (int)(sbavail(sb) - sb_offset)) &&
16701 (TCPS_HAVEESTABLISHED(tp->t_state))) {
16703 * Here we have a send window but we have
16704 * filled it up and we can't send another pacing segment.
16705 * We also have in flight more than 2 segments
16706 * and we are not completing the sb i.e. we allow
16707 * the last bytes of the sb to go out even if
16708 * its not a full pacing segment.
16711 } else if ((rack->r_ctl.crte != NULL) &&
16712 (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16713 (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16714 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16715 (len < (int)(sbavail(sb) - sb_offset))) {
16717 * Here we are doing hardware pacing, this is not a TLP,
16718 * we are not sending a pace max segment size, there is rwnd
16719 * room to send at least N pace_max_seg, the cwnd is greater
16720 * than or equal to a full pacing segments plus 4 mss and we have 2 or
16721 * more segments in flight and its not the tail of the socket buffer.
16723 * We don't want to send instead we need to get more ack's in to
16724 * allow us to send a full pacing segment. Normally, if we are pacing
16725 * about the right speed, we should have finished our pacing
16726 * send as most of the acks have come back if we are at the
16727 * right rate. This is a bit fuzzy since return path delay
16728 * can delay the acks, which is why we want to make sure we
16729 * have cwnd space to have a bit more than a max pace segments in flight.
16731 * If we have not gotten our acks back we are pacing at too high a
16732 * rate delaying will not hurt and will bring our GP estimate down by
16733 * injecting the delay. If we don't do this we will send
16734 * 2 MSS out in response to the acks being clocked in which
16735 * defeats the point of hw-pacing (i.e. to help us get
16736 * larger TSO's out).
16743 /* len will be >= 0 after this point. */
16744 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16745 rack_sndbuf_autoscale(rack);
16747 * Decide if we can use TCP Segmentation Offloading (if supported by
16750 * TSO may only be used if we are in a pure bulk sending state. The
16751 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16752 * options prevent using TSO. With TSO the TCP header is the same
16753 * (except for the sequence number) for all generated packets. This
16754 * makes it impossible to transmit any options which vary per
16755 * generated segment or packet.
16757 * IPv4 handling has a clear separation of ip options and ip header
16758 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16759 * the right thing below to provide length of just ip options and thus
16760 * checking for ipoptlen is enough to decide if ip options are present.
16763 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16765 * Pre-calculate here as we save another lookup into the darknesses
16766 * of IPsec that way and can actually decide if TSO is ok.
16769 if (isipv6 && IPSEC_ENABLED(ipv6))
16770 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16776 if (IPSEC_ENABLED(ipv4))
16777 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16781 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16782 ipoptlen += ipsec_optlen;
16784 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16785 (tp->t_port == 0) &&
16786 ((tp->t_flags & TF_SIGNATURE) == 0) &&
16787 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16791 uint32_t outstanding;
16793 outstanding = tp->snd_max - tp->snd_una;
16794 if (tp->t_flags & TF_SENTFIN) {
16796 * If we sent a fin, snd_max is 1 higher than
16802 if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16805 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16810 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16811 (long)TCP_MAXWIN << tp->rcv_scale);
16814 * Sender silly window avoidance. We transmit under the following
16815 * conditions when len is non-zero:
16817 * - We have a full segment (or more with TSO) - This is the last
16818 * buffer in a write()/send() and we are either idle or running
16819 * NODELAY - we've timed out (e.g. persist timer) - we have more
16820 * then 1/2 the maximum send window's worth of data (receiver may be
16821 * limited the window size) - we need to retransmit
16824 if (len >= segsiz) {
16828 * NOTE! on localhost connections an 'ack' from the remote
16829 * end may occur synchronously with the output and cause us
16830 * to flush a buffer queued with moretocome. XXX
16833 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
16834 (idle || (tp->t_flags & TF_NODELAY)) &&
16835 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
16836 (tp->t_flags & TF_NOPUSH) == 0) {
16840 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */
16844 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
16848 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
16856 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
16857 (ctf_outstanding(tp) < (segsiz * 2))) {
16859 * We have less than two MSS outstanding (delayed ack)
16860 * and our rwnd will not let us send a full sized
16861 * MSS. Lets go ahead and let this small segment
16862 * out because we want to try to have at least two
16863 * packets inflight to not be caught by delayed ack.
16870 * Sending of standalone window updates.
16872 * Window updates are important when we close our window due to a
16873 * full socket buffer and are opening it again after the application
16874 * reads data from it. Once the window has opened again and the
16875 * remote end starts to send again the ACK clock takes over and
16876 * provides the most current window information.
16878 * We must avoid the silly window syndrome whereas every read from
16879 * the receive buffer, no matter how small, causes a window update
16880 * to be sent. We also should avoid sending a flurry of window
16881 * updates when the socket buffer had queued a lot of data and the
16882 * application is doing small reads.
16884 * Prevent a flurry of pointless window updates by only sending an
16885 * update when we can increase the advertized window by more than
16886 * 1/4th of the socket buffer capacity. When the buffer is getting
16887 * full or is very small be more aggressive and send an update
16888 * whenever we can increase by two mss sized segments. In all other
16889 * situations the ACK's to new incoming data will carry further
16890 * window increases.
16892 * Don't send an independent window update if a delayed ACK is
16893 * pending (it will get piggy-backed on it) or the remote side
16894 * already has done a half-close and won't send more data. Skip
16895 * this if the connection is in T/TCP half-open state.
16897 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
16898 !(tp->t_flags & TF_DELACK) &&
16899 !TCPS_HAVERCVDFIN(tp->t_state)) {
16901 * "adv" is the amount we could increase the window, taking
16902 * into account that we are limited by TCP_MAXWIN <<
16909 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
16910 oldwin = (tp->rcv_adv - tp->rcv_nxt);
16914 /* We can't increase the window */
16921 * If the new window size ends up being the same as or less
16922 * than the old size when it is scaled, then don't force
16925 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
16928 if (adv >= (int32_t)(2 * segsiz) &&
16929 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
16930 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
16931 so->so_rcv.sb_hiwat <= 8 * segsiz)) {
16935 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
16943 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
16944 * is also a catch-all for the retransmit timer timeout case.
16946 if (tp->t_flags & TF_ACKNOW) {
16950 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
16955 * If our state indicates that FIN should be sent and we have not
16956 * yet done so, then we need to send.
16958 if ((flags & TH_FIN) &&
16959 (tp->snd_nxt == tp->snd_una)) {
16964 * No reason to send a segment, just return.
16967 SOCKBUF_UNLOCK(sb);
16968 just_return_nolock:
16970 int app_limited = CTF_JR_SENT_DATA;
16972 if (tot_len_this_send > 0) {
16973 /* Make sure snd_nxt is up to max */
16974 rack->r_ctl.fsb.recwin = recwin;
16975 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
16976 if ((error == 0) &&
16978 ((flags & (TH_SYN|TH_FIN)) == 0) &&
16980 (tp->snd_nxt == tp->snd_max) &&
16981 (tp->rcv_numsacks == 0) &&
16982 rack->r_fsb_inited &&
16983 TCPS_HAVEESTABLISHED(tp->t_state) &&
16984 (rack->r_must_retran == 0) &&
16985 ((tp->t_flags & TF_NEEDFIN) == 0) &&
16986 (len > 0) && (orig_len > 0) &&
16987 (orig_len > len) &&
16988 ((orig_len - len) >= segsiz) &&
16990 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
16991 /* We can send at least one more MSS using our fsb */
16993 rack->r_fast_output = 1;
16994 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
16995 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
16996 rack->r_ctl.fsb.tcp_flags = flags;
16997 rack->r_ctl.fsb.left_to_send = orig_len - len;
16998 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
16999 ("rack:%p left_to_send:%u sbavail:%u out:%u",
17000 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17001 (tp->snd_max - tp->snd_una)));
17002 if (rack->r_ctl.fsb.left_to_send < segsiz)
17003 rack->r_fast_output = 0;
17005 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17006 rack->r_ctl.fsb.rfo_apply_push = 1;
17008 rack->r_ctl.fsb.rfo_apply_push = 0;
17011 rack->r_fast_output = 0;
17014 rack_log_fsb(rack, tp, so, flags,
17015 ipoptlen, orig_len, len, 0,
17016 1, optlen, __LINE__, 1);
17017 if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17018 tp->snd_nxt = tp->snd_max;
17020 int end_window = 0;
17021 uint32_t seq = tp->gput_ack;
17023 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17026 * Mark the last sent that we just-returned (hinting
17027 * that delayed ack may play a role in any rtt measurement).
17029 rsm->r_just_ret = 1;
17031 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17032 rack->r_ctl.rc_agg_delayed = 0;
17035 rack->r_ctl.rc_agg_early = 0;
17036 if ((ctf_outstanding(tp) +
17037 min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17038 minseg)) >= tp->snd_wnd) {
17039 /* We are limited by the rwnd */
17040 app_limited = CTF_JR_RWND_LIMITED;
17041 if (IN_FASTRECOVERY(tp->t_flags))
17042 rack->r_ctl.rc_prr_sndcnt = 0;
17043 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17044 /* We are limited by whats available -- app limited */
17045 app_limited = CTF_JR_APP_LIMITED;
17046 if (IN_FASTRECOVERY(tp->t_flags))
17047 rack->r_ctl.rc_prr_sndcnt = 0;
17048 } else if ((idle == 0) &&
17049 ((tp->t_flags & TF_NODELAY) == 0) &&
17050 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17053 * No delay is not on and the
17054 * user is sending less than 1MSS. This
17055 * brings out SWS avoidance so we
17056 * don't send. Another app-limited case.
17058 app_limited = CTF_JR_APP_LIMITED;
17059 } else if (tp->t_flags & TF_NOPUSH) {
17061 * The user has requested no push of
17062 * the last segment and we are
17063 * at the last segment. Another app
17066 app_limited = CTF_JR_APP_LIMITED;
17067 } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17069 app_limited = CTF_JR_CWND_LIMITED;
17070 } else if (IN_FASTRECOVERY(tp->t_flags) &&
17071 (rack->rack_no_prr == 0) &&
17072 (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17073 app_limited = CTF_JR_PRR;
17075 /* Now why here are we not sending? */
17078 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17081 app_limited = CTF_JR_ASSESSING;
17084 * App limited in some fashion, for our pacing GP
17085 * measurements we don't want any gap (even cwnd).
17086 * Close down the measurement window.
17088 if (rack_cwnd_block_ends_measure &&
17089 ((app_limited == CTF_JR_CWND_LIMITED) ||
17090 (app_limited == CTF_JR_PRR))) {
17092 * The reason we are not sending is
17093 * the cwnd (or prr). We have been configured
17094 * to end the measurement window in
17098 } else if (rack_rwnd_block_ends_measure &&
17099 (app_limited == CTF_JR_RWND_LIMITED)) {
17101 * We are rwnd limited and have been
17102 * configured to end the measurement
17103 * window in this case.
17106 } else if (app_limited == CTF_JR_APP_LIMITED) {
17108 * A true application limited period, we have
17112 } else if (app_limited == CTF_JR_ASSESSING) {
17114 * In the assessing case we hit the end of
17115 * the if/else and had no known reason
17116 * This will panic us under invariants..
17118 * If we get this out in logs we need to
17119 * investagate which reason we missed.
17126 if ((tp->t_flags & TF_GPUTINPROG) &&
17127 SEQ_GT(tp->gput_ack, tp->snd_max)) {
17128 /* Mark the last packet has app limited */
17129 tp->gput_ack = tp->snd_max;
17132 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17133 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17134 if (rack->r_ctl.rc_app_limited_cnt == 0)
17135 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17138 * Go out to the end app limited and mark
17139 * this new one as next and move the end_appl up
17142 if (rack->r_ctl.rc_end_appl)
17143 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17144 rack->r_ctl.rc_end_appl = rsm;
17146 rsm->r_flags |= RACK_APP_LIMITED;
17147 rack->r_ctl.rc_app_limited_cnt++;
17150 rack_log_pacing_delay_calc(rack,
17151 rack->r_ctl.rc_app_limited_cnt, seq,
17152 tp->gput_ack, 0, 0, 4, __LINE__, NULL);
17156 /* set the rack tcb into the slot N */
17157 counter_u64_add(rack_paced_segments, 1);
17158 } else if (tot_len_this_send) {
17159 counter_u64_add(rack_unpaced_segments, 1);
17161 /* Check if we need to go into persists or not */
17162 if ((tp->snd_max == tp->snd_una) &&
17163 TCPS_HAVEESTABLISHED(tp->t_state) &&
17165 (sbavail(sb) > tp->snd_wnd) &&
17166 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17167 /* Yes lets make sure to move to persist before timer-start */
17168 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17170 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17171 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17173 #ifdef NETFLIX_SHARED_CWND
17174 if ((sbavail(sb) == 0) &&
17175 rack->r_ctl.rc_scw) {
17176 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17177 rack->rack_scwnd_is_idle = 1;
17180 #ifdef TCP_ACCOUNTING
17181 if (tot_len_this_send > 0) {
17182 crtsc = get_cyclecount();
17183 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17184 tp->tcp_cnt_counters[SND_OUT_DATA]++;
17186 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17187 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17188 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17190 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17191 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17192 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17194 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17196 crtsc = get_cyclecount();
17197 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17198 tp->tcp_cnt_counters[SND_LIMITED]++;
17200 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17201 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17202 tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17204 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17211 if (rsm || sack_rxmit)
17212 counter_u64_add(rack_nfto_resend, 1);
17214 counter_u64_add(rack_non_fto_send, 1);
17215 if ((flags & TH_FIN) &&
17218 * We do not transmit a FIN
17219 * with data outstanding. We
17220 * need to make it so all data
17225 /* Enforce stack imposed max seg size if we have one */
17226 if (rack->r_ctl.rc_pace_max_segs &&
17227 (len > rack->r_ctl.rc_pace_max_segs)) {
17229 len = rack->r_ctl.rc_pace_max_segs;
17231 SOCKBUF_LOCK_ASSERT(sb);
17234 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17236 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17239 * Before ESTABLISHED, force sending of initial options unless TCP
17240 * set not to do any options. NOTE: we assume that the IP/TCP header
17241 * plus TCP options always fit in a single mbuf, leaving room for a
17242 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17243 * + optlen <= MCLBYTES
17248 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17251 hdrlen = sizeof(struct tcpiphdr);
17254 * Compute options for segment. We only have to care about SYN and
17255 * established connection segments. Options for SYN-ACK segments
17256 * are handled in TCP syncache.
17259 if ((tp->t_flags & TF_NOOPT) == 0) {
17260 /* Maximum segment size. */
17261 if (flags & TH_SYN) {
17262 tp->snd_nxt = tp->iss;
17263 to.to_mss = tcp_mssopt(&inp->inp_inc);
17265 to.to_mss -= V_tcp_udp_tunneling_overhead;
17266 to.to_flags |= TOF_MSS;
17269 * On SYN or SYN|ACK transmits on TFO connections,
17270 * only include the TFO option if it is not a
17271 * retransmit, as the presence of the TFO option may
17272 * have caused the original SYN or SYN|ACK to have
17273 * been dropped by a middlebox.
17275 if (IS_FASTOPEN(tp->t_flags) &&
17276 (tp->t_rxtshift == 0)) {
17277 if (tp->t_state == TCPS_SYN_RECEIVED) {
17278 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17280 (u_int8_t *)&tp->t_tfo_cookie.server;
17281 to.to_flags |= TOF_FASTOPEN;
17283 } else if (tp->t_state == TCPS_SYN_SENT) {
17285 tp->t_tfo_client_cookie_len;
17287 tp->t_tfo_cookie.client;
17288 to.to_flags |= TOF_FASTOPEN;
17291 * If we wind up having more data to
17292 * send with the SYN than can fit in
17293 * one segment, don't send any more
17294 * until the SYN|ACK comes back from
17301 /* Window scaling. */
17302 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17303 to.to_wscale = tp->request_r_scale;
17304 to.to_flags |= TOF_SCALE;
17307 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17308 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17309 to.to_tsval = ms_cts + tp->ts_offset;
17310 to.to_tsecr = tp->ts_recent;
17311 to.to_flags |= TOF_TS;
17313 /* Set receive buffer autosizing timestamp. */
17314 if (tp->rfbuf_ts == 0 &&
17315 (so->so_rcv.sb_flags & SB_AUTOSIZE))
17316 tp->rfbuf_ts = tcp_ts_getticks();
17317 /* Selective ACK's. */
17318 if (tp->t_flags & TF_SACK_PERMIT) {
17319 if (flags & TH_SYN)
17320 to.to_flags |= TOF_SACKPERM;
17321 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17322 tp->rcv_numsacks > 0) {
17323 to.to_flags |= TOF_SACK;
17324 to.to_nsacks = tp->rcv_numsacks;
17325 to.to_sacks = (u_char *)tp->sackblks;
17328 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17329 /* TCP-MD5 (RFC2385). */
17330 if (tp->t_flags & TF_SIGNATURE)
17331 to.to_flags |= TOF_SIGNATURE;
17332 #endif /* TCP_SIGNATURE */
17334 /* Processing the options. */
17335 hdrlen += optlen = tcp_addoptions(&to, opt);
17337 * If we wanted a TFO option to be added, but it was unable
17338 * to fit, ensure no data is sent.
17340 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17341 !(to.to_flags & TOF_FASTOPEN))
17345 if (V_tcp_udp_tunneling_port == 0) {
17346 /* The port was removed?? */
17347 SOCKBUF_UNLOCK(&so->so_snd);
17348 #ifdef TCP_ACCOUNTING
17349 crtsc = get_cyclecount();
17350 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17351 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17353 counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17354 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17355 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17357 counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17360 return (EHOSTUNREACH);
17362 hdrlen += sizeof(struct udphdr);
17366 ipoptlen = ip6_optlen(tp->t_inpcb);
17369 if (tp->t_inpcb->inp_options)
17370 ipoptlen = tp->t_inpcb->inp_options->m_len -
17371 offsetof(struct ipoption, ipopt_list);
17374 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17375 ipoptlen += ipsec_optlen;
17379 * Adjust data length if insertion of options will bump the packet
17380 * length beyond the t_maxseg length. Clear the FIN bit because we
17381 * cut off the tail of the segment.
17383 if (len + optlen + ipoptlen > tp->t_maxseg) {
17385 uint32_t if_hw_tsomax;
17389 /* extract TSO information */
17390 if_hw_tsomax = tp->t_tsomax;
17391 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17392 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17393 KASSERT(ipoptlen == 0,
17394 ("%s: TSO can't do IP options", __func__));
17397 * Check if we should limit by maximum payload
17400 if (if_hw_tsomax != 0) {
17401 /* compute maximum TSO length */
17402 max_len = (if_hw_tsomax - hdrlen -
17404 if (max_len <= 0) {
17406 } else if (len > max_len) {
17413 * Prevent the last segment from being fractional
17414 * unless the send sockbuf can be emptied:
17416 max_len = (tp->t_maxseg - optlen);
17417 if ((sb_offset + len) < sbavail(sb)) {
17418 moff = len % (u_int)max_len;
17425 * In case there are too many small fragments don't
17428 if (len <= segsiz) {
17433 * Send the FIN in a separate segment after the bulk
17434 * sending is done. We don't trust the TSO
17435 * implementations to clear the FIN flag on all but
17436 * the last segment.
17438 if (tp->t_flags & TF_NEEDFIN) {
17443 if (optlen + ipoptlen >= tp->t_maxseg) {
17445 * Since we don't have enough space to put
17446 * the IP header chain and the TCP header in
17447 * one packet as required by RFC 7112, don't
17448 * send it. Also ensure that at least one
17449 * byte of the payload can be put into the
17452 SOCKBUF_UNLOCK(&so->so_snd);
17457 len = tp->t_maxseg - optlen - ipoptlen;
17464 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17465 ("%s: len > IP_MAXPACKET", __func__));
17468 if (max_linkhdr + hdrlen > MCLBYTES)
17470 if (max_linkhdr + hdrlen > MHLEN)
17472 panic("tcphdr too big");
17476 * This KASSERT is here to catch edge cases at a well defined place.
17477 * Before, those had triggered (random) panic conditions further
17480 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17482 (flags & TH_FIN) &&
17485 * We have outstanding data, don't send a fin by itself!.
17490 * Grab a header mbuf, attaching a copy of data to be transmitted,
17491 * and initialize the header from the template for sends on this
17494 hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17499 if (rack->r_ctl.rc_pace_max_segs)
17500 max_val = rack->r_ctl.rc_pace_max_segs;
17501 else if (rack->rc_user_set_max_segs)
17502 max_val = rack->rc_user_set_max_segs * segsiz;
17506 * We allow a limit on sending with hptsi.
17508 if (len > max_val) {
17513 if (MHLEN < hdrlen + max_linkhdr)
17514 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17517 m = m_gethdr(M_NOWAIT, MT_DATA);
17520 SOCKBUF_UNLOCK(sb);
17525 m->m_data += max_linkhdr;
17529 * Start the m_copy functions from the closest mbuf to the
17530 * sb_offset in the socket buffer chain.
17532 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17535 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17536 m_copydata(mb, moff, (int)len,
17537 mtod(m, caddr_t)+hdrlen);
17538 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17539 sbsndptr_adv(sb, mb, len);
17542 struct sockbuf *msb;
17544 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17548 m->m_next = tcp_m_copym(
17550 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17551 ((rsm == NULL) ? hw_tls : 0)
17552 #ifdef NETFLIX_COPY_ARGS
17556 if (len <= (tp->t_maxseg - optlen)) {
17558 * Must have ran out of mbufs for the copy
17559 * shorten it to no longer need tso. Lets
17560 * not put on sendalot since we are low on
17565 if (m->m_next == NULL) {
17566 SOCKBUF_UNLOCK(sb);
17573 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17574 if (rsm && (rsm->r_flags & RACK_TLP)) {
17576 * TLP should not count in retran count, but
17579 counter_u64_add(rack_tlp_retran, 1);
17580 counter_u64_add(rack_tlp_retran_bytes, len);
17582 tp->t_sndrexmitpack++;
17583 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17584 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17587 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17591 KMOD_TCPSTAT_INC(tcps_sndpack);
17592 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17594 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17599 * If we're sending everything we've got, set PUSH. (This
17600 * will keep happy those implementations which only give
17601 * data to the user when a buffer fills or a PUSH comes in.)
17603 if (sb_offset + len == sbused(sb) &&
17605 !(flags & TH_SYN)) {
17607 add_flag |= RACK_HAD_PUSH;
17610 SOCKBUF_UNLOCK(sb);
17612 SOCKBUF_UNLOCK(sb);
17613 if (tp->t_flags & TF_ACKNOW)
17614 KMOD_TCPSTAT_INC(tcps_sndacks);
17615 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17616 KMOD_TCPSTAT_INC(tcps_sndctrl);
17618 KMOD_TCPSTAT_INC(tcps_sndwinup);
17620 m = m_gethdr(M_NOWAIT, MT_DATA);
17627 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17629 M_ALIGN(m, hdrlen);
17632 m->m_data += max_linkhdr;
17635 SOCKBUF_UNLOCK_ASSERT(sb);
17636 m->m_pkthdr.rcvif = (struct ifnet *)0;
17638 mac_inpcb_create_mbuf(inp, m);
17640 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17643 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17646 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17647 th = rack->r_ctl.fsb.th;
17648 udp = rack->r_ctl.fsb.udp;
17651 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17653 ulen = hdrlen + len - sizeof(struct ip);
17654 udp->uh_ulen = htons(ulen);
17659 ip6 = mtod(m, struct ip6_hdr *);
17661 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
17662 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17663 udp->uh_dport = tp->t_port;
17664 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17665 udp->uh_ulen = htons(ulen);
17666 th = (struct tcphdr *)(udp + 1);
17668 th = (struct tcphdr *)(ip6 + 1);
17669 tcpip_fillheaders(inp, tp->t_port, ip6, th);
17673 ip = mtod(m, struct ip *);
17675 ipov = (struct ipovly *)ip;
17678 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
17679 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17680 udp->uh_dport = tp->t_port;
17681 ulen = hdrlen + len - sizeof(struct ip);
17682 udp->uh_ulen = htons(ulen);
17683 th = (struct tcphdr *)(udp + 1);
17685 th = (struct tcphdr *)(ip + 1);
17686 tcpip_fillheaders(inp, tp->t_port, ip, th);
17690 * Fill in fields, remembering maximum advertised window for use in
17691 * delaying messages about window sizes. If resending a FIN, be sure
17692 * not to use a new sequence number.
17694 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17695 tp->snd_nxt == tp->snd_max)
17698 * If we are starting a connection, send ECN setup SYN packet. If we
17699 * are on a retransmit, we may resend those bits a number of times
17702 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17703 if (tp->t_rxtshift >= 1) {
17704 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17705 flags |= TH_ECE | TH_CWR;
17707 flags |= TH_ECE | TH_CWR;
17709 /* Handle parallel SYN for ECN */
17710 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17711 (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17713 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17715 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17716 (tp->t_flags2 & TF2_ECN_PERMIT)) {
17718 * If the peer has ECN, mark data packets with ECN capable
17719 * transmission (ECT). Ignore pure ack packets,
17722 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17723 (sack_rxmit == 0)) {
17726 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17729 ip->ip_tos |= IPTOS_ECN_ECT0;
17730 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17732 * Reply with proper ECN notifications.
17733 * Only set CWR on new data segments.
17735 if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17737 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17740 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17744 * If we are doing retransmissions, then snd_nxt will not reflect
17745 * the first unsent octet. For ACK only packets, we do not want the
17746 * sequence number of the retransmitted packet, we want the sequence
17747 * number of the next unsent octet. So, if there is no data (and no
17748 * SYN or FIN), use snd_max instead of snd_nxt when filling in
17749 * ti_seq. But if we are in persist state, snd_max might reflect
17750 * one byte beyond the right edge of the window, so use snd_nxt in
17751 * that case, since we know we aren't doing a retransmission.
17752 * (retransmit and persist are mutually exclusive...)
17754 if (sack_rxmit == 0) {
17755 if (len || (flags & (TH_SYN | TH_FIN))) {
17756 th->th_seq = htonl(tp->snd_nxt);
17757 rack_seq = tp->snd_nxt;
17759 th->th_seq = htonl(tp->snd_max);
17760 rack_seq = tp->snd_max;
17763 th->th_seq = htonl(rsm->r_start);
17764 rack_seq = rsm->r_start;
17766 th->th_ack = htonl(tp->rcv_nxt);
17767 th->th_flags = flags;
17769 * Calculate receive window. Don't shrink window, but avoid silly
17771 * If a RST segment is sent, advertise a window of zero.
17773 if (flags & TH_RST) {
17776 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17777 recwin < (long)segsiz) {
17780 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17781 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17782 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17786 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
17787 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is
17788 * handled in syncache.
17790 if (flags & TH_SYN)
17791 th->th_win = htons((u_short)
17792 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
17794 /* Avoid shrinking window with window scaling. */
17795 recwin = roundup2(recwin, 1 << tp->rcv_scale);
17796 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
17799 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
17800 * window. This may cause the remote transmitter to stall. This
17801 * flag tells soreceive() to disable delayed acknowledgements when
17802 * draining the buffer. This can occur if the receiver is
17803 * attempting to read more data than can be buffered prior to
17804 * transmitting on the connection.
17806 if (th->th_win == 0) {
17807 tp->t_sndzerowin++;
17808 tp->t_flags |= TF_RXWIN0SENT;
17810 tp->t_flags &= ~TF_RXWIN0SENT;
17811 tp->snd_up = tp->snd_una; /* drag it along, its deprecated */
17812 /* Now are we using fsb?, if so copy the template data to the mbuf */
17813 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17816 cpto = mtod(m, uint8_t *);
17817 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
17819 * We have just copied in:
17821 * <optional udphdr>
17822 * tcphdr (no options)
17824 * We need to grab the correct pointers into the mbuf
17825 * for both the tcp header, and possibly the udp header (if tunneling).
17826 * We do this by using the offset in the copy buffer and adding it
17827 * to the mbuf base pointer (cpto).
17831 ip6 = mtod(m, struct ip6_hdr *);
17834 ip = mtod(m, struct ip *);
17835 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
17836 /* If we have a udp header lets set it into the mbuf as well */
17838 udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
17840 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17841 if (to.to_flags & TOF_SIGNATURE) {
17843 * Calculate MD5 signature and put it into the place
17844 * determined before.
17845 * NOTE: since TCP options buffer doesn't point into
17846 * mbuf's data, calculate offset and use it.
17848 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
17849 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
17851 * Do not send segment if the calculation of MD5
17852 * digest has failed.
17859 bcopy(opt, th + 1, optlen);
17860 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
17863 * Put TCP length in extended header, and then checksum extended
17866 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
17870 * ip6_plen is not need to be filled now, and will be filled
17874 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
17875 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17876 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
17877 th->th_sum = htons(0);
17878 UDPSTAT_INC(udps_opackets);
17880 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
17881 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17882 th->th_sum = in6_cksum_pseudo(ip6,
17883 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
17888 #if defined(INET6) && defined(INET)
17894 m->m_pkthdr.csum_flags = CSUM_UDP;
17895 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17896 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
17897 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
17898 th->th_sum = htons(0);
17899 UDPSTAT_INC(udps_opackets);
17901 m->m_pkthdr.csum_flags = CSUM_TCP;
17902 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17903 th->th_sum = in_pseudo(ip->ip_src.s_addr,
17904 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
17905 IPPROTO_TCP + len + optlen));
17907 /* IP version must be set here for ipv4/ipv6 checking later */
17908 KASSERT(ip->ip_v == IPVERSION,
17909 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
17913 * Enable TSO and specify the size of the segments. The TCP pseudo
17914 * header checksum is always provided. XXX: Fixme: This is currently
17915 * not the case for IPv6.
17918 KASSERT(len > tp->t_maxseg - optlen,
17919 ("%s: len <= tso_segsz", __func__));
17920 m->m_pkthdr.csum_flags |= CSUM_TSO;
17921 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
17923 KASSERT(len + hdrlen == m_length(m, NULL),
17924 ("%s: mbuf chain different than expected: %d + %u != %u",
17925 __func__, len, hdrlen, m_length(m, NULL)));
17928 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
17929 hhook_run_tcp_est_out(tp, th, &to, len, tso);
17931 /* We're getting ready to send; log now. */
17932 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
17933 union tcp_log_stackspecific log;
17935 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17936 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
17937 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
17938 if (rack->rack_no_prr)
17939 log.u_bbr.flex1 = 0;
17941 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17942 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
17943 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
17944 log.u_bbr.flex4 = orig_len;
17946 log.u_bbr.flex5 = 0x80000000;
17948 log.u_bbr.flex5 = 0;
17949 /* Save off the early/late values */
17950 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
17951 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
17952 log.u_bbr.bw_inuse = rack_get_bw(rack);
17953 if (rsm || sack_rxmit) {
17955 log.u_bbr.flex8 = 2;
17957 log.u_bbr.flex8 = 1;
17959 log.u_bbr.flex8 = 0;
17961 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17962 log.u_bbr.flex7 = mark;
17963 log.u_bbr.flex7 <<= 8;
17964 log.u_bbr.flex7 |= pass;
17965 log.u_bbr.pkts_out = tp->t_maxseg;
17966 log.u_bbr.timeStamp = cts;
17967 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17968 log.u_bbr.lt_epoch = cwnd_to_use;
17969 log.u_bbr.delivered = sendalot;
17970 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
17971 len, &log, false, NULL, NULL, 0, &tv);
17976 * Fill in IP length and desired time to live and send to IP level.
17977 * There should be a better way to handle ttl and tos; we could keep
17978 * them in the template, but need a way to checksum without them.
17981 * m->m_pkthdr.len should have been set before cksum calcuration,
17982 * because in6_cksum() need it.
17987 * we separately set hoplimit for every segment, since the
17988 * user might want to change the value via setsockopt. Also,
17989 * desired default hop limit might be changed via Neighbor
17992 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
17995 * Set the packet size here for the benefit of DTrace
17996 * probes. ip6_output() will set it properly; it's supposed
17997 * to include the option header lengths as well.
17999 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18001 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18002 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18004 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18006 if (tp->t_state == TCPS_SYN_SENT)
18007 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18009 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18010 /* TODO: IPv6 IP6TOS_ECT bit on */
18011 error = ip6_output(m,
18012 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18013 inp->in6p_outputopts,
18018 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18021 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18022 mtu = inp->inp_route6.ro_nh->nh_mtu;
18025 #if defined(INET) && defined(INET6)
18030 ip->ip_len = htons(m->m_pkthdr.len);
18032 if (inp->inp_vflag & INP_IPV6PROTO)
18033 ip->ip_ttl = in6_selecthlim(inp, NULL);
18035 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18037 * If we do path MTU discovery, then we set DF on every
18038 * packet. This might not be the best thing to do according
18039 * to RFC3390 Section 2. However the tcp hostcache migitates
18040 * the problem so it affects only the first tcp connection
18043 * NB: Don't set DF on small MTU/MSS to have a safe
18046 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18047 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18048 if (tp->t_port == 0 || len < V_tcp_minmss) {
18049 ip->ip_off |= htons(IP_DF);
18052 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18055 if (tp->t_state == TCPS_SYN_SENT)
18056 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18058 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18060 error = ip_output(m,
18061 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18067 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18069 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18070 mtu = inp->inp_route.ro_nh->nh_mtu;
18076 lgb->tlb_errno = error;
18080 * In transmit state, time the transmission and arrange for the
18081 * retransmit. In persist state, just set snd_max.
18084 rack->forced_ack = 0; /* If we send something zap the FA flag */
18085 if (rsm && (doing_tlp == 0)) {
18086 /* Set we retransmitted */
18087 rack->rc_gp_saw_rec = 1;
18089 if (cwnd_to_use > tp->snd_ssthresh) {
18090 /* Set we sent in CA */
18091 rack->rc_gp_saw_ca = 1;
18093 /* Set we sent in SS */
18094 rack->rc_gp_saw_ss = 1;
18097 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18098 (tp->t_flags & TF_SACK_PERMIT) &&
18099 tp->rcv_numsacks > 0)
18100 tcp_clean_dsack_blocks(tp);
18101 tot_len_this_send += len;
18103 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18104 else if (len == 1) {
18105 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18106 } else if (len > 1) {
18109 idx = (len / segsiz) + 3;
18110 if (idx >= TCP_MSS_ACCT_ATIMER)
18111 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18113 counter_u64_add(rack_out_size[idx], 1);
18116 if ((rack->rack_no_prr == 0) &&
18119 if (rack->r_ctl.rc_prr_sndcnt >= len)
18120 rack->r_ctl.rc_prr_sndcnt -= len;
18122 rack->r_ctl.rc_prr_sndcnt = 0;
18125 if (doing_tlp && (rsm == NULL)) {
18126 /* New send doing a TLP */
18127 add_flag |= RACK_TLP;
18128 tp->t_sndtlppack++;
18129 tp->t_sndtlpbyte += len;
18131 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18132 rack_to_usec_ts(&tv),
18133 rsm, add_flag, s_mb, s_moff);
18136 if ((error == 0) &&
18138 (tp->snd_una == tp->snd_max))
18139 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18141 tcp_seq startseq = tp->snd_nxt;
18143 /* Track our lost count */
18144 if (rsm && (doing_tlp == 0))
18145 rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18147 * Advance snd_nxt over sequence space of this segment.
18150 /* We don't log or do anything with errors */
18152 if (doing_tlp == 0) {
18155 * Not a retransmission of some
18156 * sort, new data is going out so
18157 * clear our TLP count and flag.
18159 rack->rc_tlp_in_progress = 0;
18160 rack->r_ctl.rc_tlp_cnt_out = 0;
18164 * We have just sent a TLP, mark that it is true
18165 * and make sure our in progress is set so we
18166 * continue to check the count.
18168 rack->rc_tlp_in_progress = 1;
18169 rack->r_ctl.rc_tlp_cnt_out++;
18171 if (flags & (TH_SYN | TH_FIN)) {
18172 if (flags & TH_SYN)
18174 if (flags & TH_FIN) {
18176 tp->t_flags |= TF_SENTFIN;
18179 /* In the ENOBUFS case we do *not* update snd_max */
18183 tp->snd_nxt += len;
18184 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18185 if (tp->snd_una == tp->snd_max) {
18187 * Update the time we just added data since
18188 * none was outstanding.
18190 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18191 tp->t_acktime = ticks;
18193 tp->snd_max = tp->snd_nxt;
18195 * Time this transmission if not a retransmission and
18196 * not currently timing anything.
18197 * This is only relevant in case of switching back to
18200 if (tp->t_rtttime == 0) {
18201 tp->t_rtttime = ticks;
18202 tp->t_rtseq = startseq;
18203 KMOD_TCPSTAT_INC(tcps_segstimed);
18206 ((tp->t_flags & TF_GPUTINPROG) == 0))
18207 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18210 * If we are doing FO we need to update the mbuf position and subtract
18211 * this happens when the peer sends us duplicate information and
18212 * we thus want to send a DSACK.
18214 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18215 * turned off? If not then we are going to echo multiple DSACK blocks
18216 * out (with the TSO), which we should not be doing.
18218 if (rack->r_fast_output && len) {
18219 if (rack->r_ctl.fsb.left_to_send > len)
18220 rack->r_ctl.fsb.left_to_send -= len;
18222 rack->r_ctl.fsb.left_to_send = 0;
18223 if (rack->r_ctl.fsb.left_to_send < segsiz)
18224 rack->r_fast_output = 0;
18225 if (rack->r_fast_output) {
18226 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18227 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18233 rack->r_ctl.rc_agg_delayed = 0;
18236 rack->r_ctl.rc_agg_early = 0;
18237 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */
18239 * Failures do not advance the seq counter above. For the
18240 * case of ENOBUFS we will fall out and retry in 1ms with
18241 * the hpts. Everything else will just have to retransmit
18244 * In any case, we do not want to loop around for another
18245 * send without a good reason.
18250 tp->t_softerror = error;
18251 #ifdef TCP_ACCOUNTING
18252 crtsc = get_cyclecount();
18253 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18254 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18256 counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18257 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18258 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18260 counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18266 * Pace us right away to retry in a some
18269 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18270 if (rack->rc_enobuf < 0x7f)
18272 if (slot < (10 * HPTS_USEC_IN_MSEC))
18273 slot = 10 * HPTS_USEC_IN_MSEC;
18274 if (rack->r_ctl.crte != NULL) {
18275 counter_u64_add(rack_saw_enobuf_hw, 1);
18276 tcp_rl_log_enobuf(rack->r_ctl.crte);
18278 counter_u64_add(rack_saw_enobuf, 1);
18282 * For some reason the interface we used initially
18283 * to send segments changed to another or lowered
18284 * its MTU. If TSO was active we either got an
18285 * interface without TSO capabilits or TSO was
18286 * turned off. If we obtained mtu from ip_output()
18287 * then update it and try again.
18290 tp->t_flags &= ~TF_TSO;
18292 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18295 slot = 10 * HPTS_USEC_IN_MSEC;
18296 rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18297 #ifdef TCP_ACCOUNTING
18298 crtsc = get_cyclecount();
18299 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18300 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18302 counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18303 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18304 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18306 counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18311 counter_u64_add(rack_saw_enetunreach, 1);
18315 if (TCPS_HAVERCVDSYN(tp->t_state)) {
18316 tp->t_softerror = error;
18320 slot = 10 * HPTS_USEC_IN_MSEC;
18321 rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18322 #ifdef TCP_ACCOUNTING
18323 crtsc = get_cyclecount();
18324 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18325 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18327 counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18328 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18329 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18331 counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18337 rack->rc_enobuf = 0;
18338 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18339 rack->r_ctl.retran_during_recovery += len;
18341 KMOD_TCPSTAT_INC(tcps_sndtotal);
18344 * Data sent (as far as we can tell). If this advertises a larger
18345 * window than any other segment, then remember the size of the
18346 * advertised window. Any pending ACK has now been sent.
18348 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18349 tp->rcv_adv = tp->rcv_nxt + recwin;
18351 tp->last_ack_sent = tp->rcv_nxt;
18352 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18355 /* Do we need to turn off sendalot? */
18356 if (rack->r_ctl.rc_pace_max_segs &&
18357 (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18358 /* We hit our max. */
18360 } else if ((rack->rc_user_set_max_segs) &&
18361 (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18362 /* We hit the user defined max */
18366 if ((error == 0) && (flags & TH_FIN))
18367 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18368 if (flags & TH_RST) {
18370 * We don't send again after sending a RST.
18375 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18376 } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18378 * Get our pacing rate, if an error
18379 * occurred in sending (ENOBUF) we would
18380 * hit the else if with slot preset. Other
18383 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18386 (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18387 rack->use_rack_rr) {
18388 /* Its a retransmit and we use the rack cheat? */
18390 (rack->rc_always_pace == 0) ||
18391 (rack->r_rr_config == 1)) {
18393 * We have no pacing set or we
18394 * are using old-style rack or
18395 * we are overriden to use the old 1ms pacing.
18397 slot = rack->r_ctl.rc_min_to;
18400 /* We have sent clear the flag */
18401 rack->r_ent_rec_ns = 0;
18402 if (rack->r_must_retran) {
18404 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18405 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18407 * We have retransmitted all.
18409 rack->r_must_retran = 0;
18410 rack->r_ctl.rc_out_at_rto = 0;
18412 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18414 * Sending new data will also kill
18417 rack->r_must_retran = 0;
18418 rack->r_ctl.rc_out_at_rto = 0;
18421 rack->r_ctl.fsb.recwin = recwin;
18422 if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18423 SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18425 * We hit an RTO and now have past snd_max at the RTO
18426 * clear all the WAS flags.
18428 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18431 /* set the rack tcb into the slot N */
18432 counter_u64_add(rack_paced_segments, 1);
18433 if ((error == 0) &&
18435 ((flags & (TH_SYN|TH_FIN)) == 0) &&
18437 (tp->snd_nxt == tp->snd_max) &&
18439 (tp->rcv_numsacks == 0) &&
18440 rack->r_fsb_inited &&
18441 TCPS_HAVEESTABLISHED(tp->t_state) &&
18442 (rack->r_must_retran == 0) &&
18443 ((tp->t_flags & TF_NEEDFIN) == 0) &&
18444 (len > 0) && (orig_len > 0) &&
18445 (orig_len > len) &&
18446 ((orig_len - len) >= segsiz) &&
18448 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18449 /* We can send at least one more MSS using our fsb */
18451 rack->r_fast_output = 1;
18452 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18453 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18454 rack->r_ctl.fsb.tcp_flags = flags;
18455 rack->r_ctl.fsb.left_to_send = orig_len - len;
18456 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18457 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18458 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18459 (tp->snd_max - tp->snd_una)));
18460 if (rack->r_ctl.fsb.left_to_send < segsiz)
18461 rack->r_fast_output = 0;
18463 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18464 rack->r_ctl.fsb.rfo_apply_push = 1;
18466 rack->r_ctl.fsb.rfo_apply_push = 0;
18469 rack->r_fast_output = 0;
18470 rack_log_fsb(rack, tp, so, flags,
18471 ipoptlen, orig_len, len, error,
18472 (rsm == NULL), optlen, __LINE__, 2);
18473 } else if (sendalot) {
18477 counter_u64_add(rack_unpaced_segments, 1);
18479 if ((error == 0) &&
18481 ((flags & (TH_SYN|TH_FIN)) == 0) &&
18484 (tp->rcv_numsacks == 0) &&
18485 (tp->snd_nxt == tp->snd_max) &&
18486 (rack->r_must_retran == 0) &&
18487 rack->r_fsb_inited &&
18488 TCPS_HAVEESTABLISHED(tp->t_state) &&
18489 ((tp->t_flags & TF_NEEDFIN) == 0) &&
18490 (len > 0) && (orig_len > 0) &&
18491 (orig_len > len) &&
18492 ((orig_len - len) >= segsiz) &&
18494 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18495 /* we can use fast_output for more */
18497 rack->r_fast_output = 1;
18498 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18499 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18500 rack->r_ctl.fsb.tcp_flags = flags;
18501 rack->r_ctl.fsb.left_to_send = orig_len - len;
18502 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18503 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18504 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18505 (tp->snd_max - tp->snd_una)));
18506 if (rack->r_ctl.fsb.left_to_send < segsiz) {
18507 rack->r_fast_output = 0;
18509 if (rack->r_fast_output) {
18510 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18511 rack->r_ctl.fsb.rfo_apply_push = 1;
18513 rack->r_ctl.fsb.rfo_apply_push = 0;
18514 rack_log_fsb(rack, tp, so, flags,
18515 ipoptlen, orig_len, len, error,
18516 (rsm == NULL), optlen, __LINE__, 3);
18518 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18528 counter_u64_add(rack_unpaced_segments, 1);
18530 /* Assure when we leave that snd_nxt will point to top */
18531 if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18532 tp->snd_nxt = tp->snd_max;
18533 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18534 #ifdef TCP_ACCOUNTING
18535 crtsc = get_cyclecount() - ts_val;
18536 if (tot_len_this_send) {
18537 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18538 tp->tcp_cnt_counters[SND_OUT_DATA]++;
18540 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18541 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18542 tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18544 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18545 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18546 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18548 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18550 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18551 tp->tcp_cnt_counters[SND_OUT_ACK]++;
18553 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18554 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18555 tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18557 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18561 if (error == ENOBUFS)
18567 rack_update_seg(struct tcp_rack *rack)
18571 orig_val = rack->r_ctl.rc_pace_max_segs;
18572 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18573 if (orig_val != rack->r_ctl.rc_pace_max_segs)
18574 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
18578 rack_mtu_change(struct tcpcb *tp)
18581 * The MSS may have changed
18583 struct tcp_rack *rack;
18585 rack = (struct tcp_rack *)tp->t_fb_ptr;
18586 if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18588 * The MTU has changed we need to resend everything
18589 * since all we have sent is lost. We first fix
18590 * up the mtu though.
18592 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18593 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18594 rack_remxt_tmr(tp);
18595 rack->r_fast_output = 0;
18596 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18597 rack->r_ctl.rc_sacked);
18598 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18599 rack->r_must_retran = 1;
18602 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18603 /* We don't use snd_nxt to retransmit */
18604 tp->snd_nxt = tp->snd_max;
18608 rack_set_profile(struct tcp_rack *rack, int prof)
18612 /* pace_always=1 */
18613 if (rack->rc_always_pace == 0) {
18614 if (tcp_can_enable_pacing() == 0)
18617 rack->rc_always_pace = 1;
18618 if (rack->use_fixed_rate || rack->gp_ready)
18619 rack_set_cc_pacing(rack);
18620 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18621 rack->rack_attempt_hdwr_pace = 0;
18623 if (rack_use_cmp_acks)
18624 rack->r_use_cmp_ack = 1;
18625 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18626 rack->r_use_cmp_ack)
18627 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18629 rack->rack_enable_scwnd = 1;
18631 rack->rc_gp_dyn_mul = 1;
18633 rack->r_ctl.rack_per_of_gp_ca = 100;
18635 rack->r_rr_config = 3;
18637 rack->r_ctl.rc_no_push_at_mrtt = 2;
18639 rack->rc_pace_to_cwnd = 1;
18640 rack->rc_pace_fill_if_rttin_range = 0;
18641 rack->rtt_limit_mul = 0;
18643 rack->rack_no_prr = 1;
18645 rack->r_limit_scw = 1;
18647 rack->r_ctl.rack_per_of_gp_rec = 90;
18650 } else if (prof == 3) {
18651 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18652 /* pace_always=1 */
18653 if (rack->rc_always_pace == 0) {
18654 if (tcp_can_enable_pacing() == 0)
18657 rack->rc_always_pace = 1;
18658 if (rack->use_fixed_rate || rack->gp_ready)
18659 rack_set_cc_pacing(rack);
18660 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18661 rack->rack_attempt_hdwr_pace = 0;
18663 if (rack_use_cmp_acks)
18664 rack->r_use_cmp_ack = 1;
18665 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18666 rack->r_use_cmp_ack)
18667 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18669 rack->rack_enable_scwnd = 1;
18671 rack->rc_gp_dyn_mul = 1;
18673 rack->r_ctl.rack_per_of_gp_ca = 100;
18675 rack->r_rr_config = 3;
18677 rack->r_ctl.rc_no_push_at_mrtt = 2;
18679 rack->rc_pace_to_cwnd = 1;
18680 rack->r_fill_less_agg = 1;
18681 rack->rc_pace_fill_if_rttin_range = 0;
18682 rack->rtt_limit_mul = 0;
18684 rack->rack_no_prr = 1;
18686 rack->r_limit_scw = 1;
18688 rack->r_ctl.rack_per_of_gp_rec = 90;
18692 } else if (prof == 2) {
18694 if (rack->rc_always_pace == 0) {
18695 if (tcp_can_enable_pacing() == 0)
18698 rack->rc_always_pace = 1;
18699 if (rack->use_fixed_rate || rack->gp_ready)
18700 rack_set_cc_pacing(rack);
18701 rack->r_use_cmp_ack = 1;
18702 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18703 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18704 /* pace_always=1 */
18705 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18707 rack->rack_enable_scwnd = 1;
18709 rack->rc_gp_dyn_mul = 1;
18710 rack->r_ctl.rack_per_of_gp_ca = 100;
18712 rack->r_rr_config = 3;
18714 rack->r_ctl.rc_no_push_at_mrtt = 2;
18716 rack->rc_pace_to_cwnd = 1;
18717 rack->rc_pace_fill_if_rttin_range = 0;
18718 rack->rtt_limit_mul = 0;
18720 rack->rack_no_prr = 1;
18722 rack->r_limit_scw = 0;
18724 } else if (prof == 0) {
18725 /* This changes things back to the default settings */
18727 if (rack->rc_always_pace) {
18728 tcp_decrement_paced_conn();
18729 rack_undo_cc_pacing(rack);
18730 rack->rc_always_pace = 0;
18732 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18733 rack->rc_always_pace = 1;
18734 if (rack->use_fixed_rate || rack->gp_ready)
18735 rack_set_cc_pacing(rack);
18737 rack->rc_always_pace = 0;
18738 if (rack_use_cmp_acks)
18739 rack->r_use_cmp_ack = 1;
18741 rack->r_use_cmp_ack = 0;
18742 if (rack_disable_prr)
18743 rack->rack_no_prr = 1;
18745 rack->rack_no_prr = 0;
18746 if (rack_gp_no_rec_chg)
18747 rack->rc_gp_no_rec_chg = 1;
18749 rack->rc_gp_no_rec_chg = 0;
18750 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18751 rack->r_mbuf_queue = 1;
18752 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18753 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18754 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18756 rack->r_mbuf_queue = 0;
18757 rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18759 if (rack_enable_shared_cwnd)
18760 rack->rack_enable_scwnd = 1;
18762 rack->rack_enable_scwnd = 0;
18763 if (rack_do_dyn_mul) {
18764 /* When dynamic adjustment is on CA needs to start at 100% */
18765 rack->rc_gp_dyn_mul = 1;
18766 if (rack_do_dyn_mul >= 100)
18767 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18769 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18770 rack->rc_gp_dyn_mul = 0;
18772 rack->r_rr_config = 0;
18773 rack->r_ctl.rc_no_push_at_mrtt = 0;
18774 rack->rc_pace_to_cwnd = 0;
18775 rack->rc_pace_fill_if_rttin_range = 0;
18776 rack->rtt_limit_mul = 0;
18778 if (rack_enable_hw_pacing)
18779 rack->rack_hdw_pace_ena = 1;
18781 rack->rack_hdw_pace_ena = 0;
18782 if (rack_disable_prr)
18783 rack->rack_no_prr = 1;
18785 rack->rack_no_prr = 0;
18786 if (rack_limits_scwnd)
18787 rack->r_limit_scw = 1;
18789 rack->r_limit_scw = 0;
18796 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
18798 struct deferred_opt_list *dol;
18800 dol = malloc(sizeof(struct deferred_opt_list),
18801 M_TCPFSB, M_NOWAIT|M_ZERO);
18804 * No space yikes -- fail out..
18808 dol->optname = sopt_name;
18809 dol->optval = loptval;
18810 TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
18815 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
18816 uint32_t optval, uint64_t loptval)
18818 struct epoch_tracker et;
18819 struct sockopt sopt;
18820 struct cc_newreno_opts opt;
18825 switch (sopt_name) {
18827 case TCP_RACK_PACING_BETA:
18828 RACK_OPTS_INC(tcp_rack_beta);
18829 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18830 /* This only works for newreno. */
18834 if (rack->rc_pacing_cc_set) {
18836 * Set them into the real CC module
18837 * whats in the rack pcb is the old values
18838 * to be used on restoral/
18840 sopt.sopt_dir = SOPT_SET;
18841 opt.name = CC_NEWRENO_BETA;
18843 if (CC_ALGO(tp)->ctl_output != NULL)
18844 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18851 * Not pacing yet so set it into our local
18852 * rack pcb storage.
18854 rack->r_ctl.rc_saved_beta.beta = optval;
18857 case TCP_RACK_PACING_BETA_ECN:
18858 RACK_OPTS_INC(tcp_rack_beta_ecn);
18859 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18860 /* This only works for newreno. */
18864 if (rack->rc_pacing_cc_set) {
18866 * Set them into the real CC module
18867 * whats in the rack pcb is the old values
18868 * to be used on restoral/
18870 sopt.sopt_dir = SOPT_SET;
18871 opt.name = CC_NEWRENO_BETA_ECN;
18873 if (CC_ALGO(tp)->ctl_output != NULL)
18874 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18879 * Not pacing yet so set it into our local
18880 * rack pcb storage.
18882 rack->r_ctl.rc_saved_beta.beta_ecn = optval;
18883 rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
18886 case TCP_DEFER_OPTIONS:
18887 RACK_OPTS_INC(tcp_defer_opt);
18889 if (rack->gp_ready) {
18894 rack->defer_options = 1;
18896 rack->defer_options = 0;
18898 case TCP_RACK_MEASURE_CNT:
18899 RACK_OPTS_INC(tcp_rack_measure_cnt);
18900 if (optval && (optval <= 0xff)) {
18901 rack->r_ctl.req_measurements = optval;
18905 case TCP_REC_ABC_VAL:
18906 RACK_OPTS_INC(tcp_rec_abc_val);
18908 rack->r_use_labc_for_rec = 1;
18910 rack->r_use_labc_for_rec = 0;
18912 case TCP_RACK_ABC_VAL:
18913 RACK_OPTS_INC(tcp_rack_abc_val);
18914 if ((optval > 0) && (optval < 255))
18915 rack->rc_labc = optval;
18919 case TCP_HDWR_UP_ONLY:
18920 RACK_OPTS_INC(tcp_pacing_up_only);
18922 rack->r_up_only = 1;
18924 rack->r_up_only = 0;
18926 case TCP_PACING_RATE_CAP:
18927 RACK_OPTS_INC(tcp_pacing_rate_cap);
18928 rack->r_ctl.bw_rate_cap = loptval;
18930 case TCP_RACK_PROFILE:
18931 RACK_OPTS_INC(tcp_profile);
18932 error = rack_set_profile(rack, optval);
18934 case TCP_USE_CMP_ACKS:
18935 RACK_OPTS_INC(tcp_use_cmp_acks);
18936 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
18937 /* You can't turn it off once its on! */
18939 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
18940 rack->r_use_cmp_ack = 1;
18941 rack->r_mbuf_queue = 1;
18942 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18944 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
18945 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18947 case TCP_SHARED_CWND_TIME_LIMIT:
18948 RACK_OPTS_INC(tcp_lscwnd);
18950 rack->r_limit_scw = 1;
18952 rack->r_limit_scw = 0;
18954 case TCP_RACK_PACE_TO_FILL:
18955 RACK_OPTS_INC(tcp_fillcw);
18957 rack->rc_pace_to_cwnd = 0;
18959 rack->rc_pace_to_cwnd = 1;
18961 rack->r_fill_less_agg = 1;
18963 if ((optval >= rack_gp_rtt_maxmul) &&
18964 rack_gp_rtt_maxmul &&
18966 rack->rc_pace_fill_if_rttin_range = 1;
18967 rack->rtt_limit_mul = optval;
18969 rack->rc_pace_fill_if_rttin_range = 0;
18970 rack->rtt_limit_mul = 0;
18973 case TCP_RACK_NO_PUSH_AT_MAX:
18974 RACK_OPTS_INC(tcp_npush);
18976 rack->r_ctl.rc_no_push_at_mrtt = 0;
18977 else if (optval < 0xff)
18978 rack->r_ctl.rc_no_push_at_mrtt = optval;
18982 case TCP_SHARED_CWND_ENABLE:
18983 RACK_OPTS_INC(tcp_rack_scwnd);
18985 rack->rack_enable_scwnd = 0;
18987 rack->rack_enable_scwnd = 1;
18989 case TCP_RACK_MBUF_QUEUE:
18990 /* Now do we use the LRO mbuf-queue feature */
18991 RACK_OPTS_INC(tcp_rack_mbufq);
18992 if (optval || rack->r_use_cmp_ack)
18993 rack->r_mbuf_queue = 1;
18995 rack->r_mbuf_queue = 0;
18996 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
18997 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18999 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19001 case TCP_RACK_NONRXT_CFG_RATE:
19002 RACK_OPTS_INC(tcp_rack_cfg_rate);
19004 rack->rack_rec_nonrxt_use_cr = 0;
19006 rack->rack_rec_nonrxt_use_cr = 1;
19009 RACK_OPTS_INC(tcp_rack_noprr);
19011 rack->rack_no_prr = 0;
19012 else if (optval == 1)
19013 rack->rack_no_prr = 1;
19014 else if (optval == 2)
19015 rack->no_prr_addback = 1;
19019 case TCP_TIMELY_DYN_ADJ:
19020 RACK_OPTS_INC(tcp_timely_dyn);
19022 rack->rc_gp_dyn_mul = 0;
19024 rack->rc_gp_dyn_mul = 1;
19025 if (optval >= 100) {
19027 * If the user sets something 100 or more
19028 * its the gp_ca value.
19030 rack->r_ctl.rack_per_of_gp_ca = optval;
19034 case TCP_RACK_DO_DETECTION:
19035 RACK_OPTS_INC(tcp_rack_do_detection);
19037 rack->do_detection = 0;
19039 rack->do_detection = 1;
19041 case TCP_RACK_TLP_USE:
19042 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19046 RACK_OPTS_INC(tcp_tlp_use);
19047 rack->rack_tlp_threshold_use = optval;
19049 case TCP_RACK_TLP_REDUCE:
19050 /* RACK TLP cwnd reduction (bool) */
19051 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19052 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19054 /* Pacing related ones */
19055 case TCP_RACK_PACE_ALWAYS:
19057 * zero is old rack method, 1 is new
19058 * method using a pacing rate.
19060 RACK_OPTS_INC(tcp_rack_pace_always);
19062 if (rack->rc_always_pace) {
19065 } else if (tcp_can_enable_pacing()) {
19066 rack->rc_always_pace = 1;
19067 if (rack->use_fixed_rate || rack->gp_ready)
19068 rack_set_cc_pacing(rack);
19075 if (rack->rc_always_pace) {
19076 tcp_decrement_paced_conn();
19077 rack->rc_always_pace = 0;
19078 rack_undo_cc_pacing(rack);
19081 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19082 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19084 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19085 /* A rate may be set irate or other, if so set seg size */
19086 rack_update_seg(rack);
19088 case TCP_BBR_RACK_INIT_RATE:
19089 RACK_OPTS_INC(tcp_initial_rate);
19091 /* Change from kbits per second to bytes per second */
19094 rack->r_ctl.init_rate = val;
19095 if (rack->rc_init_win != rack_default_init_window) {
19099 * Options don't always get applied
19100 * in the order you think. So in order
19101 * to assure we update a cwnd we need
19102 * to check and see if we are still
19103 * where we should raise the cwnd.
19105 win = rc_init_window(rack);
19106 if (SEQ_GT(tp->snd_max, tp->iss))
19107 snt = tp->snd_max - tp->iss;
19111 (tp->snd_cwnd < win))
19112 tp->snd_cwnd = win;
19114 if (rack->rc_always_pace)
19115 rack_update_seg(rack);
19117 case TCP_BBR_IWINTSO:
19118 RACK_OPTS_INC(tcp_initial_win);
19119 if (optval && (optval <= 0xff)) {
19122 rack->rc_init_win = optval;
19123 win = rc_init_window(rack);
19124 if (SEQ_GT(tp->snd_max, tp->iss))
19125 snt = tp->snd_max - tp->iss;
19130 #ifdef NETFLIX_PEAKRATE
19131 tp->t_maxpeakrate |
19133 rack->r_ctl.init_rate)) {
19135 * We are not past the initial window
19136 * and we have some bases for pacing,
19137 * so we need to possibly adjust up
19138 * the cwnd. Note even if we don't set
19139 * the cwnd, its still ok to raise the rc_init_win
19140 * which can be used coming out of idle when we
19141 * would have a rate.
19143 if (tp->snd_cwnd < win)
19144 tp->snd_cwnd = win;
19146 if (rack->rc_always_pace)
19147 rack_update_seg(rack);
19151 case TCP_RACK_FORCE_MSEG:
19152 RACK_OPTS_INC(tcp_rack_force_max_seg);
19154 rack->rc_force_max_seg = 1;
19156 rack->rc_force_max_seg = 0;
19158 case TCP_RACK_PACE_MAX_SEG:
19159 /* Max segments size in a pace in bytes */
19160 RACK_OPTS_INC(tcp_rack_max_seg);
19161 rack->rc_user_set_max_segs = optval;
19162 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19164 case TCP_RACK_PACE_RATE_REC:
19165 /* Set the fixed pacing rate in Bytes per second ca */
19166 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19167 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19168 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19169 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19170 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19171 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19172 rack->use_fixed_rate = 1;
19173 if (rack->rc_always_pace)
19174 rack_set_cc_pacing(rack);
19175 rack_log_pacing_delay_calc(rack,
19176 rack->r_ctl.rc_fixed_pacing_rate_ss,
19177 rack->r_ctl.rc_fixed_pacing_rate_ca,
19178 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19182 case TCP_RACK_PACE_RATE_SS:
19183 /* Set the fixed pacing rate in Bytes per second ca */
19184 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19185 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19186 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19187 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19188 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19189 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19190 rack->use_fixed_rate = 1;
19191 if (rack->rc_always_pace)
19192 rack_set_cc_pacing(rack);
19193 rack_log_pacing_delay_calc(rack,
19194 rack->r_ctl.rc_fixed_pacing_rate_ss,
19195 rack->r_ctl.rc_fixed_pacing_rate_ca,
19196 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19200 case TCP_RACK_PACE_RATE_CA:
19201 /* Set the fixed pacing rate in Bytes per second ca */
19202 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19203 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19204 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19205 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19206 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19207 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19208 rack->use_fixed_rate = 1;
19209 if (rack->rc_always_pace)
19210 rack_set_cc_pacing(rack);
19211 rack_log_pacing_delay_calc(rack,
19212 rack->r_ctl.rc_fixed_pacing_rate_ss,
19213 rack->r_ctl.rc_fixed_pacing_rate_ca,
19214 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19217 case TCP_RACK_GP_INCREASE_REC:
19218 RACK_OPTS_INC(tcp_gp_inc_rec);
19219 rack->r_ctl.rack_per_of_gp_rec = optval;
19220 rack_log_pacing_delay_calc(rack,
19221 rack->r_ctl.rack_per_of_gp_ss,
19222 rack->r_ctl.rack_per_of_gp_ca,
19223 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19226 case TCP_RACK_GP_INCREASE_CA:
19227 RACK_OPTS_INC(tcp_gp_inc_ca);
19231 * We don't allow any reduction
19237 rack->r_ctl.rack_per_of_gp_ca = ca;
19238 rack_log_pacing_delay_calc(rack,
19239 rack->r_ctl.rack_per_of_gp_ss,
19240 rack->r_ctl.rack_per_of_gp_ca,
19241 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19244 case TCP_RACK_GP_INCREASE_SS:
19245 RACK_OPTS_INC(tcp_gp_inc_ss);
19249 * We don't allow any reduction
19255 rack->r_ctl.rack_per_of_gp_ss = ss;
19256 rack_log_pacing_delay_calc(rack,
19257 rack->r_ctl.rack_per_of_gp_ss,
19258 rack->r_ctl.rack_per_of_gp_ca,
19259 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19262 case TCP_RACK_RR_CONF:
19263 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19264 if (optval && optval <= 3)
19265 rack->r_rr_config = optval;
19267 rack->r_rr_config = 0;
19269 case TCP_HDWR_RATE_CAP:
19270 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19272 if (rack->r_rack_hw_rate_caps == 0)
19273 rack->r_rack_hw_rate_caps = 1;
19277 rack->r_rack_hw_rate_caps = 0;
19280 case TCP_BBR_HDWR_PACE:
19281 RACK_OPTS_INC(tcp_hdwr_pacing);
19283 if (rack->rack_hdrw_pacing == 0) {
19284 rack->rack_hdw_pace_ena = 1;
19285 rack->rack_attempt_hdwr_pace = 0;
19289 rack->rack_hdw_pace_ena = 0;
19291 if (rack->r_ctl.crte != NULL) {
19292 rack->rack_hdrw_pacing = 0;
19293 rack->rack_attempt_hdwr_pace = 0;
19294 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19295 rack->r_ctl.crte = NULL;
19300 /* End Pacing related ones */
19301 case TCP_RACK_PRR_SENDALOT:
19302 /* Allow PRR to send more than one seg */
19303 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19304 rack->r_ctl.rc_prr_sendalot = optval;
19306 case TCP_RACK_MIN_TO:
19307 /* Minimum time between rack t-o's in ms */
19308 RACK_OPTS_INC(tcp_rack_min_to);
19309 rack->r_ctl.rc_min_to = optval;
19311 case TCP_RACK_EARLY_SEG:
19312 /* If early recovery max segments */
19313 RACK_OPTS_INC(tcp_rack_early_seg);
19314 rack->r_ctl.rc_early_recovery_segs = optval;
19316 case TCP_RACK_REORD_THRESH:
19317 /* RACK reorder threshold (shift amount) */
19318 RACK_OPTS_INC(tcp_rack_reord_thresh);
19319 if ((optval > 0) && (optval < 31))
19320 rack->r_ctl.rc_reorder_shift = optval;
19324 case TCP_RACK_REORD_FADE:
19325 /* Does reordering fade after ms time */
19326 RACK_OPTS_INC(tcp_rack_reord_fade);
19327 rack->r_ctl.rc_reorder_fade = optval;
19329 case TCP_RACK_TLP_THRESH:
19330 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19331 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19333 rack->r_ctl.rc_tlp_threshold = optval;
19337 case TCP_BBR_USE_RACK_RR:
19338 RACK_OPTS_INC(tcp_rack_rr);
19340 rack->use_rack_rr = 1;
19342 rack->use_rack_rr = 0;
19344 case TCP_FAST_RSM_HACK:
19345 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19347 rack->fast_rsm_hack = 1;
19349 rack->fast_rsm_hack = 0;
19351 case TCP_RACK_PKT_DELAY:
19352 /* RACK added ms i.e. rack-rtt + reord + N */
19353 RACK_OPTS_INC(tcp_rack_pkt_delay);
19354 rack->r_ctl.rc_pkt_delay = optval;
19357 RACK_OPTS_INC(tcp_rack_delayed_ack);
19359 tp->t_delayed_ack = 0;
19361 tp->t_delayed_ack = 1;
19362 if (tp->t_flags & TF_DELACK) {
19363 tp->t_flags &= ~TF_DELACK;
19364 tp->t_flags |= TF_ACKNOW;
19365 NET_EPOCH_ENTER(et);
19367 NET_EPOCH_EXIT(et);
19371 case TCP_BBR_RACK_RTT_USE:
19372 RACK_OPTS_INC(tcp_rack_rtt_use);
19373 if ((optval != USE_RTT_HIGH) &&
19374 (optval != USE_RTT_LOW) &&
19375 (optval != USE_RTT_AVG))
19378 rack->r_ctl.rc_rate_sample_method = optval;
19380 case TCP_DATA_AFTER_CLOSE:
19381 RACK_OPTS_INC(tcp_data_after_close);
19383 rack->rc_allow_data_af_clo = 1;
19385 rack->rc_allow_data_af_clo = 0;
19390 #ifdef NETFLIX_STATS
19391 tcp_log_socket_option(tp, sopt_name, optval, error);
19398 rack_apply_deferred_options(struct tcp_rack *rack)
19400 struct deferred_opt_list *dol, *sdol;
19403 TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19404 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19405 /* Disadvantage of deferal is you loose the error return */
19406 s_optval = (uint32_t)dol->optval;
19407 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19408 free(dol, M_TCPDO);
19413 * rack_ctloutput() must drop the inpcb lock before performing copyin on
19414 * socket option arguments. When it re-acquires the lock after the copy, it
19415 * has to revalidate that the connection is still valid for the socket
19419 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19420 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19423 int32_t error = 0, optval;
19425 switch (sopt->sopt_name) {
19426 case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */
19427 /* Pacing related ones */
19428 case TCP_RACK_PACE_ALWAYS: /* URL:pace_always */
19429 case TCP_BBR_RACK_INIT_RATE: /* URL:irate */
19430 case TCP_BBR_IWINTSO: /* URL:tso_iwin */
19431 case TCP_RACK_PACE_MAX_SEG: /* URL:pace_max_seg */
19432 case TCP_RACK_FORCE_MSEG: /* URL:force_max_seg */
19433 case TCP_RACK_PACE_RATE_CA: /* URL:pr_ca */
19434 case TCP_RACK_PACE_RATE_SS: /* URL:pr_ss*/
19435 case TCP_RACK_PACE_RATE_REC: /* URL:pr_rec */
19436 case TCP_RACK_GP_INCREASE_CA: /* URL:gp_inc_ca */
19437 case TCP_RACK_GP_INCREASE_SS: /* URL:gp_inc_ss */
19438 case TCP_RACK_GP_INCREASE_REC: /* URL:gp_inc_rec */
19439 case TCP_RACK_RR_CONF: /* URL:rrr_conf */
19440 case TCP_BBR_HDWR_PACE: /* URL:hdwrpace */
19441 case TCP_HDWR_RATE_CAP: /* URL: hdwrcap boolean */
19442 case TCP_PACING_RATE_CAP: /* URL:cap-- used by side-channel */
19443 case TCP_HDWR_UP_ONLY: /* URL:uponly -- hardware pacing boolean */
19444 /* End pacing related */
19445 case TCP_FAST_RSM_HACK: /* URL:frsm_hack */
19446 case TCP_DELACK: /* URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19447 case TCP_RACK_PRR_SENDALOT: /* URL:prr_sendalot */
19448 case TCP_RACK_MIN_TO: /* URL:min_to */
19449 case TCP_RACK_EARLY_SEG: /* URL:early_seg */
19450 case TCP_RACK_REORD_THRESH: /* URL:reord_thresh */
19451 case TCP_RACK_REORD_FADE: /* URL:reord_fade */
19452 case TCP_RACK_TLP_THRESH: /* URL:tlp_thresh */
19453 case TCP_RACK_PKT_DELAY: /* URL:pkt_delay */
19454 case TCP_RACK_TLP_USE: /* URL:tlp_use */
19455 case TCP_BBR_RACK_RTT_USE: /* URL:rttuse */
19456 case TCP_BBR_USE_RACK_RR: /* URL:rackrr */
19457 case TCP_RACK_DO_DETECTION: /* URL:detect */
19458 case TCP_NO_PRR: /* URL:noprr */
19459 case TCP_TIMELY_DYN_ADJ: /* URL:dynamic */
19460 case TCP_DATA_AFTER_CLOSE: /* no URL */
19461 case TCP_RACK_NONRXT_CFG_RATE: /* URL:nonrxtcr */
19462 case TCP_SHARED_CWND_ENABLE: /* URL:scwnd */
19463 case TCP_RACK_MBUF_QUEUE: /* URL:mqueue */
19464 case TCP_RACK_NO_PUSH_AT_MAX: /* URL:npush */
19465 case TCP_RACK_PACE_TO_FILL: /* URL:fillcw */
19466 case TCP_SHARED_CWND_TIME_LIMIT: /* URL:lscwnd */
19467 case TCP_RACK_PROFILE: /* URL:profile */
19468 case TCP_USE_CMP_ACKS: /* URL:cmpack */
19469 case TCP_RACK_ABC_VAL: /* URL:labc */
19470 case TCP_REC_ABC_VAL: /* URL:reclabc */
19471 case TCP_RACK_MEASURE_CNT: /* URL:measurecnt */
19472 case TCP_DEFER_OPTIONS: /* URL:defer */
19473 case TCP_RACK_PACING_BETA: /* URL:pacing_beta */
19474 case TCP_RACK_PACING_BETA_ECN: /* URL:pacing_beta_ecn */
19477 /* Filter off all unknown options to the base stack */
19478 return (tcp_default_ctloutput(so, sopt, inp, tp));
19482 if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19483 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19485 * We truncate it down to 32 bits for the socket-option trace this
19486 * means rates > 34Gbps won't show right, but thats probably ok.
19488 optval = (uint32_t)loptval;
19490 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19491 /* Save it in 64 bit form too */
19497 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19499 return (ECONNRESET);
19501 if (rack->defer_options && (rack->gp_ready == 0) &&
19502 (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19503 (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19504 (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19505 (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19506 /* Options are beind deferred */
19507 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19511 /* No memory to defer, fail */
19516 error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19522 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19525 INP_WLOCK_ASSERT(tp->t_inpcb);
19526 bzero(ti, sizeof(*ti));
19528 ti->tcpi_state = tp->t_state;
19529 if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19530 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19531 if (tp->t_flags & TF_SACK_PERMIT)
19532 ti->tcpi_options |= TCPI_OPT_SACK;
19533 if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19534 ti->tcpi_options |= TCPI_OPT_WSCALE;
19535 ti->tcpi_snd_wscale = tp->snd_scale;
19536 ti->tcpi_rcv_wscale = tp->rcv_scale;
19538 if (tp->t_flags2 & TF2_ECN_PERMIT)
19539 ti->tcpi_options |= TCPI_OPT_ECN;
19540 if (tp->t_flags & TF_FASTOPEN)
19541 ti->tcpi_options |= TCPI_OPT_TFO;
19542 /* still kept in ticks is t_rcvtime */
19543 ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19544 /* Since we hold everything in precise useconds this is easy */
19545 ti->tcpi_rtt = tp->t_srtt;
19546 ti->tcpi_rttvar = tp->t_rttvar;
19547 ti->tcpi_rto = tp->t_rxtcur;
19548 ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19549 ti->tcpi_snd_cwnd = tp->snd_cwnd;
19551 * FreeBSD-specific extension fields for tcp_info.
19553 ti->tcpi_rcv_space = tp->rcv_wnd;
19554 ti->tcpi_rcv_nxt = tp->rcv_nxt;
19555 ti->tcpi_snd_wnd = tp->snd_wnd;
19556 ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */
19557 ti->tcpi_snd_nxt = tp->snd_nxt;
19558 ti->tcpi_snd_mss = tp->t_maxseg;
19559 ti->tcpi_rcv_mss = tp->t_maxseg;
19560 ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19561 ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19562 ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19563 #ifdef NETFLIX_STATS
19564 ti->tcpi_total_tlp = tp->t_sndtlppack;
19565 ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19566 memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19569 if (tp->t_flags & TF_TOE) {
19570 ti->tcpi_options |= TCPI_OPT_TOE;
19571 tcp_offload_tcp_info(tp, ti);
19577 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19578 struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19580 int32_t error, optval;
19581 uint64_t val, loptval;
19582 struct tcp_info ti;
19584 * Because all our options are either boolean or an int, we can just
19585 * pull everything into optval and then unlock and copy. If we ever
19586 * add a option that is not a int, then this will have quite an
19587 * impact to this routine.
19590 switch (sopt->sopt_name) {
19592 /* First get the info filled */
19593 rack_fill_info(tp, &ti);
19594 /* Fix up the rtt related fields if needed */
19596 error = sooptcopyout(sopt, &ti, sizeof ti);
19599 * Beta is the congestion control value for NewReno that influences how
19600 * much of a backoff happens when loss is detected. It is normally set
19601 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19602 * when you exit recovery.
19604 case TCP_RACK_PACING_BETA:
19605 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19607 else if (rack->rc_pacing_cc_set == 0)
19608 optval = rack->r_ctl.rc_saved_beta.beta;
19611 * Reach out into the CC data and report back what
19612 * I have previously set. Yeah it looks hackish but
19613 * we don't want to report the saved values.
19615 if (tp->ccv->cc_data)
19616 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19622 * Beta_ecn is the congestion control value for NewReno that influences how
19623 * much of a backoff happens when a ECN mark is detected. It is normally set
19624 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19625 * you exit recovery. Note that classic ECN has a beta of 50, it is only
19626 * ABE Ecn that uses this "less" value, but we do too with pacing :)
19629 case TCP_RACK_PACING_BETA_ECN:
19630 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19632 else if (rack->rc_pacing_cc_set == 0)
19633 optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19636 * Reach out into the CC data and report back what
19637 * I have previously set. Yeah it looks hackish but
19638 * we don't want to report the saved values.
19640 if (tp->ccv->cc_data)
19641 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19646 case TCP_FAST_RSM_HACK:
19647 optval = rack->fast_rsm_hack;
19649 case TCP_DEFER_OPTIONS:
19650 optval = rack->defer_options;
19652 case TCP_RACK_MEASURE_CNT:
19653 optval = rack->r_ctl.req_measurements;
19655 case TCP_REC_ABC_VAL:
19656 optval = rack->r_use_labc_for_rec;
19658 case TCP_RACK_ABC_VAL:
19659 optval = rack->rc_labc;
19661 case TCP_HDWR_UP_ONLY:
19662 optval= rack->r_up_only;
19664 case TCP_PACING_RATE_CAP:
19665 loptval = rack->r_ctl.bw_rate_cap;
19667 case TCP_RACK_PROFILE:
19668 /* You cannot retrieve a profile, its write only */
19671 case TCP_USE_CMP_ACKS:
19672 optval = rack->r_use_cmp_ack;
19674 case TCP_RACK_PACE_TO_FILL:
19675 optval = rack->rc_pace_to_cwnd;
19676 if (optval && rack->r_fill_less_agg)
19679 case TCP_RACK_NO_PUSH_AT_MAX:
19680 optval = rack->r_ctl.rc_no_push_at_mrtt;
19682 case TCP_SHARED_CWND_ENABLE:
19683 optval = rack->rack_enable_scwnd;
19685 case TCP_RACK_NONRXT_CFG_RATE:
19686 optval = rack->rack_rec_nonrxt_use_cr;
19689 if (rack->rack_no_prr == 1)
19691 else if (rack->no_prr_addback == 1)
19696 case TCP_RACK_DO_DETECTION:
19697 optval = rack->do_detection;
19699 case TCP_RACK_MBUF_QUEUE:
19700 /* Now do we use the LRO mbuf-queue feature */
19701 optval = rack->r_mbuf_queue;
19703 case TCP_TIMELY_DYN_ADJ:
19704 optval = rack->rc_gp_dyn_mul;
19706 case TCP_BBR_IWINTSO:
19707 optval = rack->rc_init_win;
19709 case TCP_RACK_TLP_REDUCE:
19710 /* RACK TLP cwnd reduction (bool) */
19711 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
19713 case TCP_BBR_RACK_INIT_RATE:
19714 val = rack->r_ctl.init_rate;
19715 /* convert to kbits per sec */
19718 optval = (uint32_t)val;
19720 case TCP_RACK_FORCE_MSEG:
19721 optval = rack->rc_force_max_seg;
19723 case TCP_RACK_PACE_MAX_SEG:
19724 /* Max segments in a pace */
19725 optval = rack->rc_user_set_max_segs;
19727 case TCP_RACK_PACE_ALWAYS:
19728 /* Use the always pace method */
19729 optval = rack->rc_always_pace;
19731 case TCP_RACK_PRR_SENDALOT:
19732 /* Allow PRR to send more than one seg */
19733 optval = rack->r_ctl.rc_prr_sendalot;
19735 case TCP_RACK_MIN_TO:
19736 /* Minimum time between rack t-o's in ms */
19737 optval = rack->r_ctl.rc_min_to;
19739 case TCP_RACK_EARLY_SEG:
19740 /* If early recovery max segments */
19741 optval = rack->r_ctl.rc_early_recovery_segs;
19743 case TCP_RACK_REORD_THRESH:
19744 /* RACK reorder threshold (shift amount) */
19745 optval = rack->r_ctl.rc_reorder_shift;
19747 case TCP_RACK_REORD_FADE:
19748 /* Does reordering fade after ms time */
19749 optval = rack->r_ctl.rc_reorder_fade;
19751 case TCP_BBR_USE_RACK_RR:
19752 /* Do we use the rack cheat for rxt */
19753 optval = rack->use_rack_rr;
19755 case TCP_RACK_RR_CONF:
19756 optval = rack->r_rr_config;
19758 case TCP_HDWR_RATE_CAP:
19759 optval = rack->r_rack_hw_rate_caps;
19761 case TCP_BBR_HDWR_PACE:
19762 optval = rack->rack_hdw_pace_ena;
19764 case TCP_RACK_TLP_THRESH:
19765 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19766 optval = rack->r_ctl.rc_tlp_threshold;
19768 case TCP_RACK_PKT_DELAY:
19769 /* RACK added ms i.e. rack-rtt + reord + N */
19770 optval = rack->r_ctl.rc_pkt_delay;
19772 case TCP_RACK_TLP_USE:
19773 optval = rack->rack_tlp_threshold_use;
19775 case TCP_RACK_PACE_RATE_CA:
19776 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
19778 case TCP_RACK_PACE_RATE_SS:
19779 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
19781 case TCP_RACK_PACE_RATE_REC:
19782 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
19784 case TCP_RACK_GP_INCREASE_SS:
19785 optval = rack->r_ctl.rack_per_of_gp_ca;
19787 case TCP_RACK_GP_INCREASE_CA:
19788 optval = rack->r_ctl.rack_per_of_gp_ss;
19790 case TCP_BBR_RACK_RTT_USE:
19791 optval = rack->r_ctl.rc_rate_sample_method;
19794 optval = tp->t_delayed_ack;
19796 case TCP_DATA_AFTER_CLOSE:
19797 optval = rack->rc_allow_data_af_clo;
19799 case TCP_SHARED_CWND_TIME_LIMIT:
19800 optval = rack->r_limit_scw;
19803 return (tcp_default_ctloutput(so, sopt, inp, tp));
19808 if (TCP_PACING_RATE_CAP)
19809 error = sooptcopyout(sopt, &loptval, sizeof loptval);
19811 error = sooptcopyout(sopt, &optval, sizeof optval);
19817 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
19819 int32_t error = EINVAL;
19820 struct tcp_rack *rack;
19822 rack = (struct tcp_rack *)tp->t_fb_ptr;
19823 if (rack == NULL) {
19827 if (sopt->sopt_dir == SOPT_SET) {
19828 return (rack_set_sockopt(so, sopt, inp, tp, rack));
19829 } else if (sopt->sopt_dir == SOPT_GET) {
19830 return (rack_get_sockopt(so, sopt, inp, tp, rack));
19838 rack_pru_options(struct tcpcb *tp, int flags)
19840 if (flags & PRUS_OOB)
19841 return (EOPNOTSUPP);
19845 static struct tcp_function_block __tcp_rack = {
19846 .tfb_tcp_block_name = __XSTRING(STACKNAME),
19847 .tfb_tcp_output = rack_output,
19848 .tfb_do_queued_segments = ctf_do_queued_segments,
19849 .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19850 .tfb_tcp_do_segment = rack_do_segment,
19851 .tfb_tcp_ctloutput = rack_ctloutput,
19852 .tfb_tcp_fb_init = rack_init,
19853 .tfb_tcp_fb_fini = rack_fini,
19854 .tfb_tcp_timer_stop_all = rack_stopall,
19855 .tfb_tcp_timer_activate = rack_timer_activate,
19856 .tfb_tcp_timer_active = rack_timer_active,
19857 .tfb_tcp_timer_stop = rack_timer_stop,
19858 .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19859 .tfb_tcp_handoff_ok = rack_handoff_ok,
19860 .tfb_tcp_mtu_chg = rack_mtu_change,
19861 .tfb_pru_options = rack_pru_options,
19865 static const char *rack_stack_names[] = {
19866 __XSTRING(STACKNAME),
19868 __XSTRING(STACKALIAS),
19873 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
19875 memset(mem, 0, size);
19880 rack_dtor(void *mem, int32_t size, void *arg)
19885 static bool rack_mod_inited = false;
19888 tcp_addrack(module_t mod, int32_t type, void *data)
19895 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
19896 sizeof(struct rack_sendmap),
19897 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
19899 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
19900 sizeof(struct tcp_rack),
19901 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
19903 sysctl_ctx_init(&rack_sysctl_ctx);
19904 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
19905 SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
19908 __XSTRING(STACKALIAS),
19910 __XSTRING(STACKNAME),
19912 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
19914 if (rack_sysctl_root == NULL) {
19915 printf("Failed to add sysctl node\n");
19919 rack_init_sysctls();
19920 num_stacks = nitems(rack_stack_names);
19921 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
19922 rack_stack_names, &num_stacks);
19924 printf("Failed to register %s stack name for "
19925 "%s module\n", rack_stack_names[num_stacks],
19926 __XSTRING(MODNAME));
19927 sysctl_ctx_free(&rack_sysctl_ctx);
19929 uma_zdestroy(rack_zone);
19930 uma_zdestroy(rack_pcb_zone);
19931 rack_counter_destroy();
19932 printf("Failed to register rack module -- err:%d\n", err);
19935 tcp_lro_reg_mbufq();
19936 rack_mod_inited = true;
19939 err = deregister_tcp_functions(&__tcp_rack, true, false);
19942 err = deregister_tcp_functions(&__tcp_rack, false, true);
19945 if (rack_mod_inited) {
19946 uma_zdestroy(rack_zone);
19947 uma_zdestroy(rack_pcb_zone);
19948 sysctl_ctx_free(&rack_sysctl_ctx);
19949 rack_counter_destroy();
19950 rack_mod_inited = false;
19952 tcp_lro_dereg_mbufq();
19956 return (EOPNOTSUPP);
19961 static moduledata_t tcp_rack = {
19962 .name = __XSTRING(MODNAME),
19963 .evhand = tcp_addrack,
19967 MODULE_VERSION(MODNAME, 1);
19968 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
19969 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);