1 /*-
2 * Copyright (c) 2016-2020 Netflix, Inc.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
12 *
13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23 * SUCH DAMAGE.
24 *
25 */
26
27 #include <sys/cdefs.h>
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30 #include "opt_ipsec.h"
31 #include "opt_ratelimit.h"
32 #include "opt_kern_tls.h"
33 #if defined(INET) || defined(INET6)
34 #include <sys/param.h>
35 #include <sys/arb.h>
36 #include <sys/module.h>
37 #include <sys/kernel.h>
38 #ifdef TCP_HHOOK
39 #include <sys/hhook.h>
40 #endif
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/mbuf.h>
46 #include <sys/proc.h> /* for proc0 declaration */
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 #ifdef STATS
52 #include <sys/qmath.h>
53 #include <sys/tree.h>
54 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
55 #else
56 #include <sys/tree.h>
57 #endif
58 #include <sys/refcount.h>
59 #include <sys/queue.h>
60 #include <sys/tim_filter.h>
61 #include <sys/smp.h>
62 #include <sys/kthread.h>
63 #include <sys/kern_prefetch.h>
64 #include <sys/protosw.h>
65 #ifdef TCP_ACCOUNTING
66 #include <sys/sched.h>
67 #include <machine/cpu.h>
68 #endif
69 #include <vm/uma.h>
70
71 #include <net/route.h>
72 #include <net/route/nhop.h>
73 #include <net/vnet.h>
74
75 #define TCPSTATES /* for logging */
76
77 #include <netinet/in.h>
78 #include <netinet/in_kdtrace.h>
79 #include <netinet/in_pcb.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
82 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
83 #include <netinet/ip_var.h>
84 #include <netinet/ip6.h>
85 #include <netinet6/in6_pcb.h>
86 #include <netinet6/ip6_var.h>
87 #include <netinet/tcp.h>
88 #define TCPOUTFLAGS
89 #include <netinet/tcp_fsm.h>
90 #include <netinet/tcp_seq.h>
91 #include <netinet/tcp_timer.h>
92 #include <netinet/tcp_var.h>
93 #include <netinet/tcp_log_buf.h>
94 #include <netinet/tcp_syncache.h>
95 #include <netinet/tcp_hpts.h>
96 #include <netinet/tcp_ratelimit.h>
97 #include <netinet/tcp_accounting.h>
98 #include <netinet/tcpip.h>
99 #include <netinet/cc/cc.h>
100 #include <netinet/cc/cc_newreno.h>
101 #include <netinet/tcp_fastopen.h>
102 #include <netinet/tcp_lro.h>
103 #ifdef NETFLIX_SHARED_CWND
104 #include <netinet/tcp_shared_cwnd.h>
105 #endif
106 #ifdef TCP_OFFLOAD
107 #include <netinet/tcp_offload.h>
108 #endif
109 #ifdef INET6
110 #include <netinet6/tcp6_var.h>
111 #endif
112 #include <netinet/tcp_ecn.h>
113
114 #include <netipsec/ipsec_support.h>
115
116 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
117 #include <netipsec/ipsec.h>
118 #include <netipsec/ipsec6.h>
119 #endif /* IPSEC */
120
121 #include <netinet/udp.h>
122 #include <netinet/udp_var.h>
123 #include <machine/in_cksum.h>
124
125 #ifdef MAC
126 #include <security/mac/mac_framework.h>
127 #endif
128 #include "sack_filter.h"
129 #include "tcp_rack.h"
130 #include "tailq_hash.h"
131 #include "rack_bbr_common.h"
132
133 uma_zone_t rack_zone;
134 uma_zone_t rack_pcb_zone;
135
136 #ifndef TICKS2SBT
137 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t)))
138 #endif
139
140 VNET_DECLARE(uint32_t, newreno_beta);
141 VNET_DECLARE(uint32_t, newreno_beta_ecn);
142 #define V_newreno_beta VNET(newreno_beta)
143 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
144
145 #define M_TCPFSB __CONCAT(M_TCPFSB, STACKNAME)
146 #define M_TCPDO __CONCAT(M_TCPDO, STACKNAME)
147
148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb_" __XSTRING(STACKNAME), "TCP fast send block");
149 MALLOC_DEFINE(M_TCPDO, "tcp_do_" __XSTRING(STACKNAME), "TCP deferred options");
150 MALLOC_DEFINE(M_TCPPCM, "tcp_pcm_" __XSTRING(STACKNAME), "TCP PCM measurement information");
151
152 struct sysctl_ctx_list rack_sysctl_ctx;
153 struct sysctl_oid *rack_sysctl_root;
154
155 #define CUM_ACKED 1
156 #define SACKED 2
157
158 /*
159 * The RACK module incorporates a number of
160 * TCP ideas that have been put out into the IETF
161 * over the last few years:
162 * - Matt Mathis's Rate Halving which slowly drops
163 * the congestion window so that the ack clock can
164 * be maintained during a recovery.
165 * - Yuchung Cheng's RACK TCP (for which its named) that
166 * will stop us using the number of dup acks and instead
167 * use time as the gage of when we retransmit.
168 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
169 * of Dukkipati et.al.
170 * RACK depends on SACK, so if an endpoint arrives that
171 * cannot do SACK the state machine below will shuttle the
172 * connection back to using the "default" TCP stack that is
173 * in FreeBSD.
174 *
175 * To implement RACK the original TCP stack was first decomposed
176 * into a functional state machine with individual states
177 * for each of the possible TCP connection states. The do_segment
178 * functions role in life is to mandate the connection supports SACK
179 * initially and then assure that the RACK state matches the conenction
180 * state before calling the states do_segment function. Each
181 * state is simplified due to the fact that the original do_segment
182 * has been decomposed and we *know* what state we are in (no
183 * switches on the state) and all tests for SACK are gone. This
184 * greatly simplifies what each state does.
185 *
186 * TCP output is also over-written with a new version since it
187 * must maintain the new rack scoreboard.
188 *
189 */
190 static int32_t rack_tlp_thresh = 1;
191 static int32_t rack_tlp_limit = 2; /* No more than 2 TLPs w-out new data */
192 static int32_t rack_tlp_use_greater = 1;
193 static int32_t rack_reorder_thresh = 2;
194 static int32_t rack_reorder_fade = 60000000; /* 0 - never fade, def 60,000,000
195 * - 60 seconds */
196 static uint32_t rack_pcm_every_n_rounds = 100;
197 static uint32_t rack_pcm_blast = 0;
198 static uint32_t rack_pcm_is_enabled = 1;
199 static uint8_t rack_ssthresh_rest_rto_rec = 0; /* Do we restore ssthresh when we have rec -> rto -> rec */
200
201 static uint32_t rack_gp_gain_req = 1200; /* Amount percent wise required to gain to record a round has "gaining" */
202 static uint32_t rack_rnd_cnt_req = 0x10005; /* Default number of rounds if we are below rack_gp_gain_req where we exit ss */
203
204
205 static int32_t rack_rxt_scoreboard_clear_thresh = 2;
206 static int32_t rack_dnd_default = 0; /* For rr_conf = 3, what is the default for dnd */
207 static int32_t rack_rxt_controls = 0;
208 static int32_t rack_fill_cw_state = 0;
209 static uint8_t rack_req_measurements = 1;
210 /* Attack threshold detections */
211 static uint32_t rack_highest_sack_thresh_seen = 0;
212 static uint32_t rack_highest_move_thresh_seen = 0;
213 static uint32_t rack_merge_out_sacks_on_attack = 0;
214 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
215 static int32_t rack_hw_rate_caps = 0; /* 1; */
216 static int32_t rack_hw_rate_cap_per = 0; /* 0 -- off */
217 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
218 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
219 static int32_t rack_hw_up_only = 0;
220 static int32_t rack_stats_gets_ms_rtt = 1;
221 static int32_t rack_prr_addbackmax = 2;
222 static int32_t rack_do_hystart = 0;
223 static int32_t rack_apply_rtt_with_reduced_conf = 0;
224 static int32_t rack_hibeta_setting = 0;
225 static int32_t rack_default_pacing_divisor = 250;
226 static uint16_t rack_pacing_min_seg = 0;
227 static int32_t rack_timely_off = 0;
228
229 static uint32_t sad_seg_size_per = 800; /* 80.0 % */
230 static int32_t rack_pkt_delay = 1000;
231 static int32_t rack_send_a_lot_in_prr = 1;
232 static int32_t rack_min_to = 1000; /* Number of microsecond min timeout */
233 static int32_t rack_verbose_logging = 0;
234 static int32_t rack_ignore_data_after_close = 1;
235 static int32_t rack_enable_shared_cwnd = 1;
236 static int32_t rack_use_cmp_acks = 1;
237 static int32_t rack_use_fsb = 1;
238 static int32_t rack_use_rfo = 1;
239 static int32_t rack_use_rsm_rfo = 1;
240 static int32_t rack_max_abc_post_recovery = 2;
241 static int32_t rack_client_low_buf = 0;
242 static int32_t rack_dsack_std_based = 0x3; /* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
243 static int32_t rack_bw_multipler = 0; /* Limit on fill cw's jump up to be this x gp_est */
244 #ifdef TCP_ACCOUNTING
245 static int32_t rack_tcp_accounting = 0;
246 #endif
247 static int32_t rack_limits_scwnd = 1;
248 static int32_t rack_enable_mqueue_for_nonpaced = 0;
249 static int32_t rack_hybrid_allow_set_maxseg = 0;
250 static int32_t rack_disable_prr = 0;
251 static int32_t use_rack_rr = 1;
252 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
253 static int32_t rack_persist_min = 250000; /* 250usec */
254 static int32_t rack_persist_max = 2000000; /* 2 Second in usec's */
255 static int32_t rack_honors_hpts_min_to = 1; /* Do we honor the hpts minimum time out for pacing timers */
256 static uint32_t rack_max_reduce = 10; /* Percent we can reduce slot by */
257 static int32_t rack_sack_not_required = 1; /* set to one to allow non-sack to use rack */
258 static int32_t rack_limit_time_with_srtt = 0;
259 static int32_t rack_autosndbuf_inc = 20; /* In percentage form */
260 static int32_t rack_enobuf_hw_boost_mult = 0; /* How many times the hw rate we boost slot using time_between */
261 static int32_t rack_enobuf_hw_max = 12000; /* 12 ms in usecs */
262 static int32_t rack_enobuf_hw_min = 10000; /* 10 ms in usecs */
263 static int32_t rack_hw_rwnd_factor = 2; /* How many max_segs the rwnd must be before we hold off sending */
264 static int32_t rack_hw_check_queue = 0; /* Do we always pre-check queue depth of a hw queue */
265
266 /*
267 * Currently regular tcp has a rto_min of 30ms
268 * the backoff goes 12 times so that ends up
269 * being a total of 122.850 seconds before a
270 * connection is killed.
271 */
272 static uint32_t rack_def_data_window = 20;
273 static uint32_t rack_goal_bdp = 2;
274 static uint32_t rack_min_srtts = 1;
275 static uint32_t rack_min_measure_usec = 0;
276 static int32_t rack_tlp_min = 10000; /* 10ms */
277 static int32_t rack_rto_min = 30000; /* 30,000 usec same as main freebsd */
278 static int32_t rack_rto_max = 4000000; /* 4 seconds in usec's */
279 static const int32_t rack_free_cache = 2;
280 static int32_t rack_hptsi_segments = 40;
281 static int32_t rack_rate_sample_method = USE_RTT_LOW;
282 static int32_t rack_pace_every_seg = 0;
283 static int32_t rack_delayed_ack_time = 40000; /* 40ms in usecs */
284 static int32_t rack_slot_reduction = 4;
285 static int32_t rack_wma_divisor = 8; /* For WMA calculation */
286 static int32_t rack_cwnd_block_ends_measure = 0;
287 static int32_t rack_rwnd_block_ends_measure = 0;
288 static int32_t rack_def_profile = 0;
289
290 static int32_t rack_lower_cwnd_at_tlp = 0;
291 static int32_t rack_always_send_oldest = 0;
292 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
293
294 static uint16_t rack_per_of_gp_ss = 250; /* 250 % slow-start */
295 static uint16_t rack_per_of_gp_ca = 200; /* 200 % congestion-avoidance */
296 static uint16_t rack_per_of_gp_rec = 200; /* 200 % of bw */
297
298 /* Probertt */
299 static uint16_t rack_per_of_gp_probertt = 60; /* 60% of bw */
300 static uint16_t rack_per_of_gp_lowthresh = 40; /* 40% is bottom */
301 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
302 static uint16_t rack_atexit_prtt_hbp = 130; /* Clamp to 130% on exit prtt if highly buffered path */
303 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
304
305 static uint32_t rack_max_drain_wait = 2; /* How man gp srtt's before we give up draining */
306 static uint32_t rack_must_drain = 1; /* How many GP srtt's we *must* wait */
307 static uint32_t rack_probertt_use_min_rtt_entry = 1; /* Use the min to calculate the goal else gp_srtt */
308 static uint32_t rack_probertt_use_min_rtt_exit = 0;
309 static uint32_t rack_probe_rtt_sets_cwnd = 0;
310 static uint32_t rack_probe_rtt_safety_val = 2000000; /* No more than 2 sec in probe-rtt */
311 static uint32_t rack_time_between_probertt = 9600000; /* 9.6 sec in usecs */
312 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0; /* How many srtt periods does probe-rtt last top fraction */
313 static uint32_t rack_probertt_gpsrtt_cnt_div = 0; /* How many srtt periods does probe-rtt last bottom fraction */
314 static uint32_t rack_min_probertt_hold = 40000; /* Equal to delayed ack time */
315 static uint32_t rack_probertt_filter_life = 10000000;
316 static uint32_t rack_probertt_lower_within = 10;
317 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds) to count as a lowering */
318 static int32_t rack_pace_one_seg = 0; /* Shall we pace for less than 1.4Meg 1MSS at a time */
319 static int32_t rack_probertt_clear_is = 1;
320 static int32_t rack_max_drain_hbp = 1; /* Extra drain times gpsrtt for highly buffered paths */
321 static int32_t rack_hbp_thresh = 3; /* what is the divisor max_rtt/min_rtt to decided a hbp */
322
323 /* Part of pacing */
324 static int32_t rack_max_per_above = 30; /* When we go to increment stop if above 100+this% */
325
326 /* Timely information:
327 *
328 * Here we have various control parameters on how
329 * timely may change the multiplier. rack_gain_p5_ub
330 * is associated with timely but not directly influencing
331 * the rate decision like the other variables. It controls
332 * the way fill-cw interacts with timely and caps how much
333 * timely can boost the fill-cw b/w.
334 *
335 * The other values are various boost/shrink numbers as well
336 * as potential caps when adjustments are made to the timely
337 * gain (returned by rack_get_output_gain(). Remember too that
338 * the gain returned can be overriden by other factors such as
339 * probeRTT as well as fixed-rate-pacing.
340 */
341 static int32_t rack_gain_p5_ub = 250;
342 static int32_t rack_gp_per_bw_mul_up = 2; /* 2% */
343 static int32_t rack_gp_per_bw_mul_down = 4; /* 4% */
344 static int32_t rack_gp_rtt_maxmul = 3; /* 3 x maxmin */
345 static int32_t rack_gp_rtt_minmul = 1; /* minrtt + (minrtt/mindiv) is lower rtt */
346 static int32_t rack_gp_rtt_mindiv = 4; /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
347 static int32_t rack_gp_decrease_per = 80; /* Beta value of timely decrease (.8) = 80 */
348 static int32_t rack_gp_increase_per = 2; /* 2% increase in multiplier */
349 static int32_t rack_per_lower_bound = 50; /* Don't allow to drop below this multiplier */
350 static int32_t rack_per_upper_bound_ss = 0; /* Don't allow SS to grow above this */
351 static int32_t rack_per_upper_bound_ca = 0; /* Don't allow CA to grow above this */
352 static int32_t rack_do_dyn_mul = 0; /* Are the rack gp multipliers dynamic */
353 static int32_t rack_gp_no_rec_chg = 1; /* Prohibit recovery from reducing it's multiplier */
354 static int32_t rack_timely_dec_clear = 6; /* Do we clear decrement count at a value (6)? */
355 static int32_t rack_timely_max_push_rise = 3; /* One round of pushing */
356 static int32_t rack_timely_max_push_drop = 3; /* Three round of pushing */
357 static int32_t rack_timely_min_segs = 4; /* 4 segment minimum */
358 static int32_t rack_timely_no_stopping = 0;
359 static int32_t rack_down_raise_thresh = 100;
360 static int32_t rack_req_segs = 1;
361 static uint64_t rack_bw_rate_cap = 0;
362 static uint64_t rack_fillcw_bw_cap = 3750000; /* Cap fillcw at 30Mbps */
363
364
365 /* Rack specific counters */
366 counter_u64_t rack_saw_enobuf;
367 counter_u64_t rack_saw_enobuf_hw;
368 counter_u64_t rack_saw_enetunreach;
369 counter_u64_t rack_persists_sends;
370 counter_u64_t rack_persists_acks;
371 counter_u64_t rack_persists_loss;
372 counter_u64_t rack_persists_lost_ends;
373 counter_u64_t rack_total_bytes;
374 #ifdef INVARIANTS
375 counter_u64_t rack_adjust_map_bw;
376 #endif
377 /* Tail loss probe counters */
378 counter_u64_t rack_tlp_tot;
379 counter_u64_t rack_tlp_newdata;
380 counter_u64_t rack_tlp_retran;
381 counter_u64_t rack_tlp_retran_bytes;
382 counter_u64_t rack_to_tot;
383 counter_u64_t rack_hot_alloc;
384 counter_u64_t rack_to_alloc;
385 counter_u64_t rack_to_alloc_hard;
386 counter_u64_t rack_to_alloc_emerg;
387 counter_u64_t rack_to_alloc_limited;
388 counter_u64_t rack_alloc_limited_conns;
389 counter_u64_t rack_split_limited;
390 counter_u64_t rack_rxt_clamps_cwnd;
391 counter_u64_t rack_rxt_clamps_cwnd_uniq;
392
393 counter_u64_t rack_multi_single_eq;
394 counter_u64_t rack_proc_non_comp_ack;
395
396 counter_u64_t rack_fto_send;
397 counter_u64_t rack_fto_rsm_send;
398 counter_u64_t rack_nfto_resend;
399 counter_u64_t rack_non_fto_send;
400 counter_u64_t rack_extended_rfo;
401
402 counter_u64_t rack_sack_proc_all;
403 counter_u64_t rack_sack_proc_short;
404 counter_u64_t rack_sack_proc_restart;
405 counter_u64_t rack_sack_attacks_detected;
406 counter_u64_t rack_sack_attacks_reversed;
407 counter_u64_t rack_sack_attacks_suspect;
408 counter_u64_t rack_sack_used_next_merge;
409 counter_u64_t rack_sack_splits;
410 counter_u64_t rack_sack_used_prev_merge;
411 counter_u64_t rack_sack_skipped_acked;
412 counter_u64_t rack_ack_total;
413 counter_u64_t rack_express_sack;
414 counter_u64_t rack_sack_total;
415 counter_u64_t rack_move_none;
416 counter_u64_t rack_move_some;
417
418 counter_u64_t rack_input_idle_reduces;
419 counter_u64_t rack_collapsed_win;
420 counter_u64_t rack_collapsed_win_seen;
421 counter_u64_t rack_collapsed_win_rxt;
422 counter_u64_t rack_collapsed_win_rxt_bytes;
423 counter_u64_t rack_try_scwnd;
424 counter_u64_t rack_hw_pace_init_fail;
425 counter_u64_t rack_hw_pace_lost;
426
427 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
428 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
429
430
431 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
432
433 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do { \
434 (tv) = (value) + slop; \
435 if ((u_long)(tv) < (u_long)(tvmin)) \
436 (tv) = (tvmin); \
437 if ((u_long)(tv) > (u_long)(tvmax)) \
438 (tv) = (tvmax); \
439 } while (0)
440
441 static void
442 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line);
443
444 static int
445 rack_process_ack(struct mbuf *m, struct tcphdr *th,
446 struct socket *so, struct tcpcb *tp, struct tcpopt *to,
447 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val, int32_t orig_tlen);
448 static int
449 rack_process_data(struct mbuf *m, struct tcphdr *th,
450 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
451 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
452 static void
453 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
454 uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
455 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
456 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
457 uint8_t limit_type);
458 static struct rack_sendmap *
459 rack_check_recovery_mode(struct tcpcb *tp,
460 uint32_t tsused);
461 static uint32_t
462 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack);
463 static void
464 rack_cong_signal(struct tcpcb *tp,
465 uint32_t type, uint32_t ack, int );
466 static void rack_counter_destroy(void);
467 static int
468 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
469 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
470 static void
471 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
472 static void
473 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
474 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
475 static void rack_dtor(void *mem, int32_t size, void *arg);
476 static void
477 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
478 uint32_t flex1, uint32_t flex2,
479 uint32_t flex3, uint32_t flex4,
480 uint32_t flex5, uint32_t flex6,
481 uint16_t flex7, uint8_t mod);
482
483 static void
484 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
485 uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
486 struct rack_sendmap *rsm, uint8_t quality);
487 static struct rack_sendmap *
488 rack_find_high_nonack(struct tcp_rack *rack,
489 struct rack_sendmap *rsm);
490 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
491 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
492 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
493 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
494 static void
495 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
496 tcp_seq th_ack, int line, uint8_t quality);
497 static void
498 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
499
500 static uint32_t
501 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
502 static int32_t rack_handoff_ok(struct tcpcb *tp);
503 static int32_t rack_init(struct tcpcb *tp, void **ptr);
504 static void rack_init_sysctls(void);
505
506 static void
507 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
508 struct tcphdr *th, int entered_rec, int dup_ack_struck,
509 int *dsack_seen, int *sacks_seen);
510 static void
511 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
512 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
513 struct rack_sendmap *hintrsm, uint32_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
514
515 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
516
517
518 static void
519 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
520 struct rack_sendmap *rsm, uint32_t cts);
521 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
522 static int32_t rack_output(struct tcpcb *tp);
523
524 static uint32_t
525 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
526 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
527 uint32_t cts, uint32_t segsiz);
528 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
529 static void rack_remxt_tmr(struct tcpcb *tp);
530 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
531 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
532 static int32_t rack_stopall(struct tcpcb *tp);
533 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
534 static uint32_t
535 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
536 struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint32_t add_flag, int segsiz);
537 static void
538 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
539 struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz);
540 static int
541 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
542 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
543 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
544 static int
545 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
546 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
547 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
548
549 static int
550 rack_do_closing(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);
553 static int
554 rack_do_established(struct mbuf *m, struct tcphdr *th,
555 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
556 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
557 static int
558 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
559 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
560 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
561 static int
562 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
563 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
564 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
565 static int
566 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
567 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
568 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
569 static int
570 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
571 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
572 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
573 static int
574 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
575 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
576 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
577 static int
578 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
579 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
580 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
581 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
582 struct rack_sendmap *
583 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
584 uint32_t tsused);
585 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
586 uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
587 static void
588 tcp_rack_partialack(struct tcpcb *tp);
589 static int
590 rack_set_profile(struct tcp_rack *rack, int prof);
591 static void
592 rack_apply_deferred_options(struct tcp_rack *rack);
593
594 int32_t rack_clear_counter=0;
595
596 static uint64_t
rack_get_lt_bw(struct tcp_rack * rack)597 rack_get_lt_bw(struct tcp_rack *rack)
598 {
599 struct timeval tv;
600 uint64_t tim, bytes;
601
602 tim = rack->r_ctl.lt_bw_time;
603 bytes = rack->r_ctl.lt_bw_bytes;
604 if (rack->lt_bw_up) {
605 /* Include all the current bytes too */
606 microuptime(&tv);
607 bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
608 tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
609 }
610 if ((bytes != 0) && (tim != 0))
611 return ((bytes * (uint64_t)1000000) / tim);
612 else
613 return (0);
614 }
615
616 static void
rack_swap_beta_values(struct tcp_rack * rack,uint8_t flex8)617 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
618 {
619 struct sockopt sopt;
620 struct cc_newreno_opts opt;
621 struct tcpcb *tp;
622 uint32_t old_beta;
623 uint32_t old_beta_ecn;
624 int error, failed = 0;
625
626 tp = rack->rc_tp;
627 if (tp->t_cc == NULL) {
628 /* Tcb is leaving */
629 return;
630 }
631 rack->rc_pacing_cc_set = 1;
632 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
633 /* Not new-reno we can't play games with beta! */
634 failed = 1;
635 goto out;
636
637 }
638 if (CC_ALGO(tp)->ctl_output == NULL) {
639 /* Huh, not using new-reno so no swaps.? */
640 failed = 2;
641 goto out;
642 }
643 /* Get the current values out */
644 sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
645 sopt.sopt_dir = SOPT_GET;
646 opt.name = CC_NEWRENO_BETA;
647 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
648 if (error) {
649 failed = 3;
650 goto out;
651 }
652 old_beta = opt.val;
653 opt.name = CC_NEWRENO_BETA_ECN;
654 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
655 if (error) {
656 failed = 4;
657 goto out;
658 }
659 old_beta_ecn = opt.val;
660
661 /* Now lets set in the values we have stored */
662 sopt.sopt_dir = SOPT_SET;
663 opt.name = CC_NEWRENO_BETA;
664 opt.val = rack->r_ctl.rc_saved_beta;
665 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
666 if (error) {
667 failed = 5;
668 goto out;
669 }
670 opt.name = CC_NEWRENO_BETA_ECN;
671 opt.val = rack->r_ctl.rc_saved_beta_ecn;
672 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
673 if (error) {
674 failed = 6;
675 goto out;
676 }
677 /* Save off the values for restoral */
678 rack->r_ctl.rc_saved_beta = old_beta;
679 rack->r_ctl.rc_saved_beta_ecn = old_beta_ecn;
680 out:
681 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
682 union tcp_log_stackspecific log;
683 struct timeval tv;
684 struct newreno *ptr;
685
686 ptr = ((struct newreno *)tp->t_ccv.cc_data);
687 memset(&log, 0, sizeof(log));
688 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
689 log.u_bbr.flex1 = ptr->beta;
690 log.u_bbr.flex2 = ptr->beta_ecn;
691 log.u_bbr.flex3 = ptr->newreno_flags;
692 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta;
693 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta_ecn;
694 log.u_bbr.flex6 = failed;
695 log.u_bbr.flex7 = rack->gp_ready;
696 log.u_bbr.flex7 <<= 1;
697 log.u_bbr.flex7 |= rack->use_fixed_rate;
698 log.u_bbr.flex7 <<= 1;
699 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
700 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
701 log.u_bbr.flex8 = flex8;
702 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
703 0, &log, false, NULL, NULL, 0, &tv);
704 }
705 }
706
707 static void
rack_set_cc_pacing(struct tcp_rack * rack)708 rack_set_cc_pacing(struct tcp_rack *rack)
709 {
710 if (rack->rc_pacing_cc_set)
711 return;
712 /*
713 * Use the swap utility placing in 3 for flex8 to id a
714 * set of a new set of values.
715 */
716 rack->rc_pacing_cc_set = 1;
717 rack_swap_beta_values(rack, 3);
718 }
719
720 static void
rack_undo_cc_pacing(struct tcp_rack * rack)721 rack_undo_cc_pacing(struct tcp_rack *rack)
722 {
723 if (rack->rc_pacing_cc_set == 0)
724 return;
725 /*
726 * Use the swap utility placing in 4 for flex8 to id a
727 * restoral of the old values.
728 */
729 rack->rc_pacing_cc_set = 0;
730 rack_swap_beta_values(rack, 4);
731 }
732
733 static void
rack_remove_pacing(struct tcp_rack * rack)734 rack_remove_pacing(struct tcp_rack *rack)
735 {
736 if (rack->rc_pacing_cc_set)
737 rack_undo_cc_pacing(rack);
738 if (rack->r_ctl.pacing_method & RACK_REG_PACING)
739 tcp_decrement_paced_conn();
740 if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
741 tcp_dec_dgp_pacing_cnt();
742 rack->rc_always_pace = 0;
743 rack->r_ctl.pacing_method = RACK_PACING_NONE;
744 rack->dgp_on = 0;
745 rack->rc_hybrid_mode = 0;
746 rack->use_fixed_rate = 0;
747 }
748
749 static void
rack_log_gpset(struct tcp_rack * rack,uint32_t seq_end,uint32_t ack_end_t,uint32_t send_end_t,int line,uint8_t mode,struct rack_sendmap * rsm)750 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
751 uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
752 {
753 if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
754 union tcp_log_stackspecific log;
755 struct timeval tv;
756
757 memset(&log, 0, sizeof(log));
758 log.u_bbr.flex1 = seq_end;
759 log.u_bbr.flex2 = rack->rc_tp->gput_seq;
760 log.u_bbr.flex3 = ack_end_t;
761 log.u_bbr.flex4 = rack->rc_tp->gput_ts;
762 log.u_bbr.flex5 = send_end_t;
763 log.u_bbr.flex6 = rack->rc_tp->gput_ack;
764 log.u_bbr.flex7 = mode;
765 log.u_bbr.flex8 = 69;
766 log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
767 log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
768 log.u_bbr.pkts_out = line;
769 log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
770 log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
771 log.u_bbr.epoch = rack->r_ctl.current_round;
772 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
773 if (rsm != NULL) {
774 log.u_bbr.applimited = rsm->r_start;
775 log.u_bbr.delivered = rsm->r_end;
776 log.u_bbr.epoch = rsm->r_flags;
777 }
778 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
779 TCP_LOG_EVENTP(rack->rc_tp, NULL,
780 &rack->rc_inp->inp_socket->so_rcv,
781 &rack->rc_inp->inp_socket->so_snd,
782 BBR_LOG_HPTSI_CALC, 0,
783 0, &log, false, &tv);
784 }
785 }
786
787 static int
sysctl_rack_clear(SYSCTL_HANDLER_ARGS)788 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
789 {
790 uint32_t stat;
791 int32_t error;
792
793 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
794 if (error || req->newptr == NULL)
795 return error;
796
797 error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
798 if (error)
799 return (error);
800 if (stat == 1) {
801 #ifdef INVARIANTS
802 printf("Clearing RACK counters\n");
803 #endif
804 counter_u64_zero(rack_tlp_tot);
805 counter_u64_zero(rack_tlp_newdata);
806 counter_u64_zero(rack_tlp_retran);
807 counter_u64_zero(rack_tlp_retran_bytes);
808 counter_u64_zero(rack_to_tot);
809 counter_u64_zero(rack_saw_enobuf);
810 counter_u64_zero(rack_saw_enobuf_hw);
811 counter_u64_zero(rack_saw_enetunreach);
812 counter_u64_zero(rack_persists_sends);
813 counter_u64_zero(rack_total_bytes);
814 counter_u64_zero(rack_persists_acks);
815 counter_u64_zero(rack_persists_loss);
816 counter_u64_zero(rack_persists_lost_ends);
817 #ifdef INVARIANTS
818 counter_u64_zero(rack_adjust_map_bw);
819 #endif
820 counter_u64_zero(rack_to_alloc_hard);
821 counter_u64_zero(rack_to_alloc_emerg);
822 counter_u64_zero(rack_sack_proc_all);
823 counter_u64_zero(rack_fto_send);
824 counter_u64_zero(rack_fto_rsm_send);
825 counter_u64_zero(rack_extended_rfo);
826 counter_u64_zero(rack_hw_pace_init_fail);
827 counter_u64_zero(rack_hw_pace_lost);
828 counter_u64_zero(rack_non_fto_send);
829 counter_u64_zero(rack_nfto_resend);
830 counter_u64_zero(rack_sack_proc_short);
831 counter_u64_zero(rack_sack_proc_restart);
832 counter_u64_zero(rack_to_alloc);
833 counter_u64_zero(rack_to_alloc_limited);
834 counter_u64_zero(rack_alloc_limited_conns);
835 counter_u64_zero(rack_split_limited);
836 counter_u64_zero(rack_rxt_clamps_cwnd);
837 counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
838 counter_u64_zero(rack_multi_single_eq);
839 counter_u64_zero(rack_proc_non_comp_ack);
840 counter_u64_zero(rack_sack_attacks_detected);
841 counter_u64_zero(rack_sack_attacks_reversed);
842 counter_u64_zero(rack_sack_attacks_suspect);
843 counter_u64_zero(rack_sack_used_next_merge);
844 counter_u64_zero(rack_sack_used_prev_merge);
845 counter_u64_zero(rack_sack_splits);
846 counter_u64_zero(rack_sack_skipped_acked);
847 counter_u64_zero(rack_ack_total);
848 counter_u64_zero(rack_express_sack);
849 counter_u64_zero(rack_sack_total);
850 counter_u64_zero(rack_move_none);
851 counter_u64_zero(rack_move_some);
852 counter_u64_zero(rack_try_scwnd);
853 counter_u64_zero(rack_collapsed_win);
854 counter_u64_zero(rack_collapsed_win_rxt);
855 counter_u64_zero(rack_collapsed_win_seen);
856 counter_u64_zero(rack_collapsed_win_rxt_bytes);
857 } else if (stat == 2) {
858 #ifdef INVARIANTS
859 printf("Clearing RACK option array\n");
860 #endif
861 COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
862 } else if (stat == 3) {
863 printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
864 } else if (stat == 4) {
865 #ifdef INVARIANTS
866 printf("Clearing RACK out size array\n");
867 #endif
868 COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
869 }
870 rack_clear_counter = 0;
871 return (0);
872 }
873
874 static void
rack_init_sysctls(void)875 rack_init_sysctls(void)
876 {
877 struct sysctl_oid *rack_counters;
878 struct sysctl_oid *rack_attack;
879 struct sysctl_oid *rack_pacing;
880 struct sysctl_oid *rack_timely;
881 struct sysctl_oid *rack_timers;
882 struct sysctl_oid *rack_tlp;
883 struct sysctl_oid *rack_misc;
884 struct sysctl_oid *rack_features;
885 struct sysctl_oid *rack_measure;
886 struct sysctl_oid *rack_probertt;
887 struct sysctl_oid *rack_hw_pacing;
888
889 rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
890 SYSCTL_CHILDREN(rack_sysctl_root),
891 OID_AUTO,
892 "sack_attack",
893 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
894 "Rack Sack Attack Counters and Controls");
895 rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
896 SYSCTL_CHILDREN(rack_sysctl_root),
897 OID_AUTO,
898 "stats",
899 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
900 "Rack Counters");
901 SYSCTL_ADD_S32(&rack_sysctl_ctx,
902 SYSCTL_CHILDREN(rack_sysctl_root),
903 OID_AUTO, "rate_sample_method", CTLFLAG_RW,
904 &rack_rate_sample_method , USE_RTT_LOW,
905 "What method should we use for rate sampling 0=high, 1=low ");
906 /* Probe rtt related controls */
907 rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
908 SYSCTL_CHILDREN(rack_sysctl_root),
909 OID_AUTO,
910 "probertt",
911 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
912 "ProbeRTT related Controls");
913 SYSCTL_ADD_U16(&rack_sysctl_ctx,
914 SYSCTL_CHILDREN(rack_probertt),
915 OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
916 &rack_atexit_prtt_hbp, 130,
917 "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
918 SYSCTL_ADD_U16(&rack_sysctl_ctx,
919 SYSCTL_CHILDREN(rack_probertt),
920 OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
921 &rack_atexit_prtt, 130,
922 "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
923 SYSCTL_ADD_U16(&rack_sysctl_ctx,
924 SYSCTL_CHILDREN(rack_probertt),
925 OID_AUTO, "gp_per_mul", CTLFLAG_RW,
926 &rack_per_of_gp_probertt, 60,
927 "What percentage of goodput do we pace at in probertt");
928 SYSCTL_ADD_U16(&rack_sysctl_ctx,
929 SYSCTL_CHILDREN(rack_probertt),
930 OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
931 &rack_per_of_gp_probertt_reduce, 10,
932 "What percentage of goodput do we reduce every gp_srtt");
933 SYSCTL_ADD_U16(&rack_sysctl_ctx,
934 SYSCTL_CHILDREN(rack_probertt),
935 OID_AUTO, "gp_per_low", CTLFLAG_RW,
936 &rack_per_of_gp_lowthresh, 40,
937 "What percentage of goodput do we allow the multiplier to fall to");
938 SYSCTL_ADD_U32(&rack_sysctl_ctx,
939 SYSCTL_CHILDREN(rack_probertt),
940 OID_AUTO, "time_between", CTLFLAG_RW,
941 & rack_time_between_probertt, 96000000,
942 "How many useconds between the lowest rtt falling must past before we enter probertt");
943 SYSCTL_ADD_U32(&rack_sysctl_ctx,
944 SYSCTL_CHILDREN(rack_probertt),
945 OID_AUTO, "safety", CTLFLAG_RW,
946 &rack_probe_rtt_safety_val, 2000000,
947 "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
948 SYSCTL_ADD_U32(&rack_sysctl_ctx,
949 SYSCTL_CHILDREN(rack_probertt),
950 OID_AUTO, "sets_cwnd", CTLFLAG_RW,
951 &rack_probe_rtt_sets_cwnd, 0,
952 "Do we set the cwnd too (if always_lower is on)");
953 SYSCTL_ADD_U32(&rack_sysctl_ctx,
954 SYSCTL_CHILDREN(rack_probertt),
955 OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
956 &rack_max_drain_wait, 2,
957 "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
958 SYSCTL_ADD_U32(&rack_sysctl_ctx,
959 SYSCTL_CHILDREN(rack_probertt),
960 OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
961 &rack_must_drain, 1,
962 "We must drain this many gp_srtt's waiting for flight to reach goal");
963 SYSCTL_ADD_U32(&rack_sysctl_ctx,
964 SYSCTL_CHILDREN(rack_probertt),
965 OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
966 &rack_probertt_use_min_rtt_entry, 1,
967 "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
968 SYSCTL_ADD_U32(&rack_sysctl_ctx,
969 SYSCTL_CHILDREN(rack_probertt),
970 OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
971 &rack_probertt_use_min_rtt_exit, 0,
972 "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
973 SYSCTL_ADD_U32(&rack_sysctl_ctx,
974 SYSCTL_CHILDREN(rack_probertt),
975 OID_AUTO, "length_div", CTLFLAG_RW,
976 &rack_probertt_gpsrtt_cnt_div, 0,
977 "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
978 SYSCTL_ADD_U32(&rack_sysctl_ctx,
979 SYSCTL_CHILDREN(rack_probertt),
980 OID_AUTO, "length_mul", CTLFLAG_RW,
981 &rack_probertt_gpsrtt_cnt_mul, 0,
982 "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
983 SYSCTL_ADD_U32(&rack_sysctl_ctx,
984 SYSCTL_CHILDREN(rack_probertt),
985 OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
986 &rack_min_probertt_hold, 200000,
987 "What is the minimum time we hold probertt at target");
988 SYSCTL_ADD_U32(&rack_sysctl_ctx,
989 SYSCTL_CHILDREN(rack_probertt),
990 OID_AUTO, "filter_life", CTLFLAG_RW,
991 &rack_probertt_filter_life, 10000000,
992 "What is the time for the filters life in useconds");
993 SYSCTL_ADD_U32(&rack_sysctl_ctx,
994 SYSCTL_CHILDREN(rack_probertt),
995 OID_AUTO, "lower_within", CTLFLAG_RW,
996 &rack_probertt_lower_within, 10,
997 "If the rtt goes lower within this percentage of the time, go into probe-rtt");
998 SYSCTL_ADD_U32(&rack_sysctl_ctx,
999 SYSCTL_CHILDREN(rack_probertt),
1000 OID_AUTO, "must_move", CTLFLAG_RW,
1001 &rack_min_rtt_movement, 250,
1002 "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1003 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1004 SYSCTL_CHILDREN(rack_probertt),
1005 OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1006 &rack_probertt_clear_is, 1,
1007 "Do we clear I/S counts on exiting probe-rtt");
1008 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1009 SYSCTL_CHILDREN(rack_probertt),
1010 OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1011 &rack_max_drain_hbp, 1,
1012 "How many extra drain gpsrtt's do we get in highly buffered paths");
1013 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1014 SYSCTL_CHILDREN(rack_probertt),
1015 OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1016 &rack_hbp_thresh, 3,
1017 "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1018 /* Pacing related sysctls */
1019 rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1020 SYSCTL_CHILDREN(rack_sysctl_root),
1021 OID_AUTO,
1022 "pacing",
1023 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1024 "Pacing related Controls");
1025 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1026 SYSCTL_CHILDREN(rack_pacing),
1027 OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1028 &rack_pcm_is_enabled, 1,
1029 "Do we by default do PCM measurements?");
1030 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1031 SYSCTL_CHILDREN(rack_pacing),
1032 OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1033 &rack_pcm_every_n_rounds, 100,
1034 "How many rounds before we need to do a PCM measurement");
1035 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1036 SYSCTL_CHILDREN(rack_pacing),
1037 OID_AUTO, "pcm_blast", CTLFLAG_RW,
1038 &rack_pcm_blast, 0,
1039 "Blast out the full cwnd/rwnd when doing a PCM measurement");
1040 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1041 SYSCTL_CHILDREN(rack_pacing),
1042 OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1043 &rack_gp_gain_req, 1200,
1044 "How much do we have to increase the GP to record the round 1200 = 120.0");
1045 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1046 SYSCTL_CHILDREN(rack_pacing),
1047 OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1048 &rack_rnd_cnt_req, 0x10005,
1049 "How many rounds less than rnd_gp_gain will drop us out of SS");
1050 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1051 SYSCTL_CHILDREN(rack_pacing),
1052 OID_AUTO, "no_timely", CTLFLAG_RW,
1053 &rack_timely_off, 0,
1054 "Do we not use timely in DGP?");
1055 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1056 SYSCTL_CHILDREN(rack_pacing),
1057 OID_AUTO, "fillcw", CTLFLAG_RW,
1058 &rack_fill_cw_state, 0,
1059 "Enable fillcw on new connections (default=0 off)?");
1060 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1061 SYSCTL_CHILDREN(rack_pacing),
1062 OID_AUTO, "min_burst", CTLFLAG_RW,
1063 &rack_pacing_min_seg, 0,
1064 "What is the min burst size for pacing (0 disables)?");
1065 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066 SYSCTL_CHILDREN(rack_pacing),
1067 OID_AUTO, "divisor", CTLFLAG_RW,
1068 &rack_default_pacing_divisor, 250,
1069 "What is the default divisor given to the rl code?");
1070 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071 SYSCTL_CHILDREN(rack_pacing),
1072 OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1073 &rack_bw_multipler, 0,
1074 "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1075 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076 SYSCTL_CHILDREN(rack_pacing),
1077 OID_AUTO, "max_pace_over", CTLFLAG_RW,
1078 &rack_max_per_above, 30,
1079 "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1080 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081 SYSCTL_CHILDREN(rack_pacing),
1082 OID_AUTO, "allow1mss", CTLFLAG_RW,
1083 &rack_pace_one_seg, 0,
1084 "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1085 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086 SYSCTL_CHILDREN(rack_pacing),
1087 OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1088 &rack_limit_time_with_srtt, 0,
1089 "Do we limit pacing time based on srtt");
1090 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1091 SYSCTL_CHILDREN(rack_pacing),
1092 OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1093 &rack_per_of_gp_ss, 250,
1094 "If non zero, what percentage of goodput to pace at in slow start");
1095 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1096 SYSCTL_CHILDREN(rack_pacing),
1097 OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1098 &rack_per_of_gp_ca, 150,
1099 "If non zero, what percentage of goodput to pace at in congestion avoidance");
1100 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1101 SYSCTL_CHILDREN(rack_pacing),
1102 OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1103 &rack_per_of_gp_rec, 200,
1104 "If non zero, what percentage of goodput to pace at in recovery");
1105 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106 SYSCTL_CHILDREN(rack_pacing),
1107 OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1108 &rack_hptsi_segments, 40,
1109 "What size is the max for TSO segments in pacing and burst mitigation");
1110 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111 SYSCTL_CHILDREN(rack_pacing),
1112 OID_AUTO, "burst_reduces", CTLFLAG_RW,
1113 &rack_slot_reduction, 4,
1114 "When doing only burst mitigation what is the reduce divisor");
1115 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116 SYSCTL_CHILDREN(rack_sysctl_root),
1117 OID_AUTO, "use_pacing", CTLFLAG_RW,
1118 &rack_pace_every_seg, 0,
1119 "If set we use pacing, if clear we use only the original burst mitigation");
1120 SYSCTL_ADD_U64(&rack_sysctl_ctx,
1121 SYSCTL_CHILDREN(rack_pacing),
1122 OID_AUTO, "rate_cap", CTLFLAG_RW,
1123 &rack_bw_rate_cap, 0,
1124 "If set we apply this value to the absolute rate cap used by pacing");
1125 SYSCTL_ADD_U64(&rack_sysctl_ctx,
1126 SYSCTL_CHILDREN(rack_pacing),
1127 OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1128 &rack_fillcw_bw_cap, 3750000,
1129 "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1130 SYSCTL_ADD_U8(&rack_sysctl_ctx,
1131 SYSCTL_CHILDREN(rack_sysctl_root),
1132 OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1133 &rack_req_measurements, 1,
1134 "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1135 /* Hardware pacing */
1136 rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1137 SYSCTL_CHILDREN(rack_sysctl_root),
1138 OID_AUTO,
1139 "hdwr_pacing",
1140 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1141 "Pacing related Controls");
1142 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1143 SYSCTL_CHILDREN(rack_hw_pacing),
1144 OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1145 &rack_hw_rwnd_factor, 2,
1146 "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1147 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1148 SYSCTL_CHILDREN(rack_hw_pacing),
1149 OID_AUTO, "precheck", CTLFLAG_RW,
1150 &rack_hw_check_queue, 0,
1151 "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1152 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1153 SYSCTL_CHILDREN(rack_hw_pacing),
1154 OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1155 &rack_enobuf_hw_boost_mult, 0,
1156 "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1157 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1158 SYSCTL_CHILDREN(rack_hw_pacing),
1159 OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1160 &rack_enobuf_hw_max, 2,
1161 "What is the max boost the pacing time if we see a ENOBUFS?");
1162 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1163 SYSCTL_CHILDREN(rack_hw_pacing),
1164 OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1165 &rack_enobuf_hw_min, 2,
1166 "What is the min boost the pacing time if we see a ENOBUFS?");
1167 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1168 SYSCTL_CHILDREN(rack_hw_pacing),
1169 OID_AUTO, "enable", CTLFLAG_RW,
1170 &rack_enable_hw_pacing, 0,
1171 "Should RACK attempt to use hw pacing?");
1172 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1173 SYSCTL_CHILDREN(rack_hw_pacing),
1174 OID_AUTO, "rate_cap", CTLFLAG_RW,
1175 &rack_hw_rate_caps, 0,
1176 "Does the highest hardware pacing rate cap the rate we will send at??");
1177 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1178 SYSCTL_CHILDREN(rack_hw_pacing),
1179 OID_AUTO, "uncap_per", CTLFLAG_RW,
1180 &rack_hw_rate_cap_per, 0,
1181 "If you go over b/w by this amount you will be uncapped (0 = never)");
1182 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1183 SYSCTL_CHILDREN(rack_hw_pacing),
1184 OID_AUTO, "rate_min", CTLFLAG_RW,
1185 &rack_hw_rate_min, 0,
1186 "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1187 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1188 SYSCTL_CHILDREN(rack_hw_pacing),
1189 OID_AUTO, "rate_to_low", CTLFLAG_RW,
1190 &rack_hw_rate_to_low, 0,
1191 "If we fall below this rate, dis-engage hw pacing?");
1192 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1193 SYSCTL_CHILDREN(rack_hw_pacing),
1194 OID_AUTO, "up_only", CTLFLAG_RW,
1195 &rack_hw_up_only, 0,
1196 "Do we allow hw pacing to lower the rate selected?");
1197 rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1198 SYSCTL_CHILDREN(rack_sysctl_root),
1199 OID_AUTO,
1200 "timely",
1201 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1202 "Rack Timely RTT Controls");
1203 /* Timely based GP dynmics */
1204 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1205 SYSCTL_CHILDREN(rack_timely),
1206 OID_AUTO, "upper", CTLFLAG_RW,
1207 &rack_gp_per_bw_mul_up, 2,
1208 "Rack timely upper range for equal b/w (in percentage)");
1209 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1210 SYSCTL_CHILDREN(rack_timely),
1211 OID_AUTO, "lower", CTLFLAG_RW,
1212 &rack_gp_per_bw_mul_down, 4,
1213 "Rack timely lower range for equal b/w (in percentage)");
1214 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1215 SYSCTL_CHILDREN(rack_timely),
1216 OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1217 &rack_gp_rtt_maxmul, 3,
1218 "Rack timely multiplier of lowest rtt for rtt_max");
1219 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1220 SYSCTL_CHILDREN(rack_timely),
1221 OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1222 &rack_gp_rtt_mindiv, 4,
1223 "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1224 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1225 SYSCTL_CHILDREN(rack_timely),
1226 OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1227 &rack_gp_rtt_minmul, 1,
1228 "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1229 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1230 SYSCTL_CHILDREN(rack_timely),
1231 OID_AUTO, "decrease", CTLFLAG_RW,
1232 &rack_gp_decrease_per, 80,
1233 "Rack timely Beta value 80 = .8 (scaled by 100)");
1234 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1235 SYSCTL_CHILDREN(rack_timely),
1236 OID_AUTO, "increase", CTLFLAG_RW,
1237 &rack_gp_increase_per, 2,
1238 "Rack timely increase perentage of our GP multiplication factor");
1239 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1240 SYSCTL_CHILDREN(rack_timely),
1241 OID_AUTO, "lowerbound", CTLFLAG_RW,
1242 &rack_per_lower_bound, 50,
1243 "Rack timely lowest percentage we allow GP multiplier to fall to");
1244 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1245 SYSCTL_CHILDREN(rack_timely),
1246 OID_AUTO, "p5_upper", CTLFLAG_RW,
1247 &rack_gain_p5_ub, 250,
1248 "Profile 5 upper bound to timely gain");
1249
1250 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1251 SYSCTL_CHILDREN(rack_timely),
1252 OID_AUTO, "upperboundss", CTLFLAG_RW,
1253 &rack_per_upper_bound_ss, 0,
1254 "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1255 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1256 SYSCTL_CHILDREN(rack_timely),
1257 OID_AUTO, "upperboundca", CTLFLAG_RW,
1258 &rack_per_upper_bound_ca, 0,
1259 "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1260 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1261 SYSCTL_CHILDREN(rack_timely),
1262 OID_AUTO, "dynamicgp", CTLFLAG_RW,
1263 &rack_do_dyn_mul, 0,
1264 "Rack timely do we enable dynmaic timely goodput by default");
1265 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1266 SYSCTL_CHILDREN(rack_timely),
1267 OID_AUTO, "no_rec_red", CTLFLAG_RW,
1268 &rack_gp_no_rec_chg, 1,
1269 "Rack timely do we prohibit the recovery multiplier from being lowered");
1270 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1271 SYSCTL_CHILDREN(rack_timely),
1272 OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1273 &rack_timely_dec_clear, 6,
1274 "Rack timely what threshold do we count to before another boost during b/w decent");
1275 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1276 SYSCTL_CHILDREN(rack_timely),
1277 OID_AUTO, "max_push_rise", CTLFLAG_RW,
1278 &rack_timely_max_push_rise, 3,
1279 "Rack timely how many times do we push up with b/w increase");
1280 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1281 SYSCTL_CHILDREN(rack_timely),
1282 OID_AUTO, "max_push_drop", CTLFLAG_RW,
1283 &rack_timely_max_push_drop, 3,
1284 "Rack timely how many times do we push back on b/w decent");
1285 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1286 SYSCTL_CHILDREN(rack_timely),
1287 OID_AUTO, "min_segs", CTLFLAG_RW,
1288 &rack_timely_min_segs, 4,
1289 "Rack timely when setting the cwnd what is the min num segments");
1290 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1291 SYSCTL_CHILDREN(rack_timely),
1292 OID_AUTO, "nonstop", CTLFLAG_RW,
1293 &rack_timely_no_stopping, 0,
1294 "Rack timely don't stop increase");
1295 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1296 SYSCTL_CHILDREN(rack_timely),
1297 OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1298 &rack_down_raise_thresh, 100,
1299 "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1300 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1301 SYSCTL_CHILDREN(rack_timely),
1302 OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1303 &rack_req_segs, 1,
1304 "Bottom dragging if not these many segments outstanding and room");
1305
1306 /* TLP and Rack related parameters */
1307 rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1308 SYSCTL_CHILDREN(rack_sysctl_root),
1309 OID_AUTO,
1310 "tlp",
1311 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1312 "TLP and Rack related Controls");
1313 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1314 SYSCTL_CHILDREN(rack_tlp),
1315 OID_AUTO, "use_rrr", CTLFLAG_RW,
1316 &use_rack_rr, 1,
1317 "Do we use Rack Rapid Recovery");
1318 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1319 SYSCTL_CHILDREN(rack_tlp),
1320 OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1321 &rack_max_abc_post_recovery, 2,
1322 "Since we do early recovery, do we override the l_abc to a value, if so what?");
1323 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1324 SYSCTL_CHILDREN(rack_tlp),
1325 OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1326 &rack_non_rxt_use_cr, 0,
1327 "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1328 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1329 SYSCTL_CHILDREN(rack_tlp),
1330 OID_AUTO, "tlpmethod", CTLFLAG_RW,
1331 &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1332 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1333 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1334 SYSCTL_CHILDREN(rack_tlp),
1335 OID_AUTO, "limit", CTLFLAG_RW,
1336 &rack_tlp_limit, 2,
1337 "How many TLP's can be sent without sending new data");
1338 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1339 SYSCTL_CHILDREN(rack_tlp),
1340 OID_AUTO, "use_greater", CTLFLAG_RW,
1341 &rack_tlp_use_greater, 1,
1342 "Should we use the rack_rtt time if its greater than srtt");
1343 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1344 SYSCTL_CHILDREN(rack_tlp),
1345 OID_AUTO, "tlpminto", CTLFLAG_RW,
1346 &rack_tlp_min, 10000,
1347 "TLP minimum timeout per the specification (in microseconds)");
1348 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1349 SYSCTL_CHILDREN(rack_tlp),
1350 OID_AUTO, "send_oldest", CTLFLAG_RW,
1351 &rack_always_send_oldest, 0,
1352 "Should we always send the oldest TLP and RACK-TLP");
1353 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1354 SYSCTL_CHILDREN(rack_tlp),
1355 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1356 &rack_lower_cwnd_at_tlp, 0,
1357 "When a TLP completes a retran should we enter recovery");
1358 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1359 SYSCTL_CHILDREN(rack_tlp),
1360 OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1361 &rack_reorder_thresh, 2,
1362 "What factor for rack will be added when seeing reordering (shift right)");
1363 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1364 SYSCTL_CHILDREN(rack_tlp),
1365 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1366 &rack_tlp_thresh, 1,
1367 "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1368 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1369 SYSCTL_CHILDREN(rack_tlp),
1370 OID_AUTO, "reorder_fade", CTLFLAG_RW,
1371 &rack_reorder_fade, 60000000,
1372 "Does reorder detection fade, if so how many microseconds (0 means never)");
1373 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374 SYSCTL_CHILDREN(rack_tlp),
1375 OID_AUTO, "pktdelay", CTLFLAG_RW,
1376 &rack_pkt_delay, 1000,
1377 "Extra RACK time (in microseconds) besides reordering thresh");
1378
1379 /* Timer related controls */
1380 rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1381 SYSCTL_CHILDREN(rack_sysctl_root),
1382 OID_AUTO,
1383 "timers",
1384 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1385 "Timer related controls");
1386 SYSCTL_ADD_U8(&rack_sysctl_ctx,
1387 SYSCTL_CHILDREN(rack_timers),
1388 OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1389 &rack_ssthresh_rest_rto_rec, 0,
1390 "When doing recovery -> rto -> recovery do we reset SSthresh?");
1391 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1392 SYSCTL_CHILDREN(rack_timers),
1393 OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1394 &rack_rxt_scoreboard_clear_thresh, 2,
1395 "How many RTO's are allowed before we clear the scoreboard");
1396 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1397 SYSCTL_CHILDREN(rack_timers),
1398 OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1399 &rack_honors_hpts_min_to, 1,
1400 "Do rack pacing timers honor hpts min timeout");
1401 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1402 SYSCTL_CHILDREN(rack_timers),
1403 OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1404 &rack_max_reduce, 10,
1405 "Max percentage we will reduce slot by for pacing when we are behind");
1406 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1407 SYSCTL_CHILDREN(rack_timers),
1408 OID_AUTO, "persmin", CTLFLAG_RW,
1409 &rack_persist_min, 250000,
1410 "What is the minimum time in microseconds between persists");
1411 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1412 SYSCTL_CHILDREN(rack_timers),
1413 OID_AUTO, "persmax", CTLFLAG_RW,
1414 &rack_persist_max, 2000000,
1415 "What is the largest delay in microseconds between persists");
1416 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1417 SYSCTL_CHILDREN(rack_timers),
1418 OID_AUTO, "delayed_ack", CTLFLAG_RW,
1419 &rack_delayed_ack_time, 40000,
1420 "Delayed ack time (40ms in microseconds)");
1421 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1422 SYSCTL_CHILDREN(rack_timers),
1423 OID_AUTO, "minrto", CTLFLAG_RW,
1424 &rack_rto_min, 30000,
1425 "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1426 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1427 SYSCTL_CHILDREN(rack_timers),
1428 OID_AUTO, "maxrto", CTLFLAG_RW,
1429 &rack_rto_max, 4000000,
1430 "Maximum RTO in microseconds -- should be at least as large as min_rto");
1431 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1432 SYSCTL_CHILDREN(rack_timers),
1433 OID_AUTO, "minto", CTLFLAG_RW,
1434 &rack_min_to, 1000,
1435 "Minimum rack timeout in microseconds");
1436 /* Measure controls */
1437 rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1438 SYSCTL_CHILDREN(rack_sysctl_root),
1439 OID_AUTO,
1440 "measure",
1441 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1442 "Measure related controls");
1443 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1444 SYSCTL_CHILDREN(rack_measure),
1445 OID_AUTO, "wma_divisor", CTLFLAG_RW,
1446 &rack_wma_divisor, 8,
1447 "When doing b/w calculation what is the divisor for the WMA");
1448 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1449 SYSCTL_CHILDREN(rack_measure),
1450 OID_AUTO, "end_cwnd", CTLFLAG_RW,
1451 &rack_cwnd_block_ends_measure, 0,
1452 "Does a cwnd just-return end the measurement window (app limited)");
1453 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1454 SYSCTL_CHILDREN(rack_measure),
1455 OID_AUTO, "end_rwnd", CTLFLAG_RW,
1456 &rack_rwnd_block_ends_measure, 0,
1457 "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1458 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1459 SYSCTL_CHILDREN(rack_measure),
1460 OID_AUTO, "min_target", CTLFLAG_RW,
1461 &rack_def_data_window, 20,
1462 "What is the minimum target window (in mss) for a GP measurements");
1463 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1464 SYSCTL_CHILDREN(rack_measure),
1465 OID_AUTO, "goal_bdp", CTLFLAG_RW,
1466 &rack_goal_bdp, 2,
1467 "What is the goal BDP to measure");
1468 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1469 SYSCTL_CHILDREN(rack_measure),
1470 OID_AUTO, "min_srtts", CTLFLAG_RW,
1471 &rack_min_srtts, 1,
1472 "What is the goal BDP to measure");
1473 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1474 SYSCTL_CHILDREN(rack_measure),
1475 OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1476 &rack_min_measure_usec, 0,
1477 "What is the Minimum time time for a measurement if 0, this is off");
1478 /* Features */
1479 rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1480 SYSCTL_CHILDREN(rack_sysctl_root),
1481 OID_AUTO,
1482 "features",
1483 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1484 "Feature controls");
1485 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1486 SYSCTL_CHILDREN(rack_features),
1487 OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1488 &rack_hybrid_allow_set_maxseg, 0,
1489 "Should hybrid pacing allow the setmss command");
1490 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1491 SYSCTL_CHILDREN(rack_features),
1492 OID_AUTO, "cmpack", CTLFLAG_RW,
1493 &rack_use_cmp_acks, 1,
1494 "Should RACK have LRO send compressed acks");
1495 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1496 SYSCTL_CHILDREN(rack_features),
1497 OID_AUTO, "fsb", CTLFLAG_RW,
1498 &rack_use_fsb, 1,
1499 "Should RACK use the fast send block?");
1500 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1501 SYSCTL_CHILDREN(rack_features),
1502 OID_AUTO, "rfo", CTLFLAG_RW,
1503 &rack_use_rfo, 1,
1504 "Should RACK use rack_fast_output()?");
1505 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1506 SYSCTL_CHILDREN(rack_features),
1507 OID_AUTO, "rsmrfo", CTLFLAG_RW,
1508 &rack_use_rsm_rfo, 1,
1509 "Should RACK use rack_fast_rsm_output()?");
1510 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1511 SYSCTL_CHILDREN(rack_features),
1512 OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1513 &rack_enable_mqueue_for_nonpaced, 0,
1514 "Should RACK use mbuf queuing for non-paced connections");
1515 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1516 SYSCTL_CHILDREN(rack_features),
1517 OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1518 &rack_do_hystart, 0,
1519 "Should RACK enable HyStart++ on connections?");
1520 /* Misc rack controls */
1521 rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1522 SYSCTL_CHILDREN(rack_sysctl_root),
1523 OID_AUTO,
1524 "misc",
1525 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1526 "Misc related controls");
1527 #ifdef TCP_ACCOUNTING
1528 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1529 SYSCTL_CHILDREN(rack_misc),
1530 OID_AUTO, "tcp_acct", CTLFLAG_RW,
1531 &rack_tcp_accounting, 0,
1532 "Should we turn on TCP accounting for all rack sessions?");
1533 #endif
1534 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1535 SYSCTL_CHILDREN(rack_misc),
1536 OID_AUTO, "dnd", CTLFLAG_RW,
1537 &rack_dnd_default, 0,
1538 "Do not disturb default for rack_rrr = 3");
1539 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1540 SYSCTL_CHILDREN(rack_misc),
1541 OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1542 &sad_seg_size_per, 800,
1543 "Percentage of segment size needed in a sack 800 = 80.0?");
1544 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1545 SYSCTL_CHILDREN(rack_misc),
1546 OID_AUTO, "rxt_controls", CTLFLAG_RW,
1547 &rack_rxt_controls, 0,
1548 "Retransmit sending size controls (valid values 0, 1, 2 default=1)?");
1549 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1550 SYSCTL_CHILDREN(rack_misc),
1551 OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1552 &rack_hibeta_setting, 0,
1553 "Do we ue a high beta (80 instead of 50)?");
1554 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1555 SYSCTL_CHILDREN(rack_misc),
1556 OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1557 &rack_apply_rtt_with_reduced_conf, 0,
1558 "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1559 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1560 SYSCTL_CHILDREN(rack_misc),
1561 OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1562 &rack_dsack_std_based, 3,
1563 "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1564 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1565 SYSCTL_CHILDREN(rack_misc),
1566 OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1567 &rack_prr_addbackmax, 2,
1568 "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1569 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1570 SYSCTL_CHILDREN(rack_misc),
1571 OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1572 &rack_stats_gets_ms_rtt, 1,
1573 "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1574 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1575 SYSCTL_CHILDREN(rack_misc),
1576 OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1577 &rack_client_low_buf, 0,
1578 "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1579 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1580 SYSCTL_CHILDREN(rack_misc),
1581 OID_AUTO, "defprofile", CTLFLAG_RW,
1582 &rack_def_profile, 0,
1583 "Should RACK use a default profile (0=no, num == profile num)?");
1584 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1585 SYSCTL_CHILDREN(rack_misc),
1586 OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1587 &rack_enable_shared_cwnd, 1,
1588 "Should RACK try to use the shared cwnd on connections where allowed");
1589 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1590 SYSCTL_CHILDREN(rack_misc),
1591 OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1592 &rack_limits_scwnd, 1,
1593 "Should RACK place low end time limits on the shared cwnd feature");
1594 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1595 SYSCTL_CHILDREN(rack_misc),
1596 OID_AUTO, "no_prr", CTLFLAG_RW,
1597 &rack_disable_prr, 0,
1598 "Should RACK not use prr and only pace (must have pacing on)");
1599 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1600 SYSCTL_CHILDREN(rack_misc),
1601 OID_AUTO, "bb_verbose", CTLFLAG_RW,
1602 &rack_verbose_logging, 0,
1603 "Should RACK black box logging be verbose");
1604 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1605 SYSCTL_CHILDREN(rack_misc),
1606 OID_AUTO, "data_after_close", CTLFLAG_RW,
1607 &rack_ignore_data_after_close, 1,
1608 "Do we hold off sending a RST until all pending data is ack'd");
1609 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1610 SYSCTL_CHILDREN(rack_misc),
1611 OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1612 &rack_sack_not_required, 1,
1613 "Do we allow rack to run on connections not supporting SACK");
1614 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1615 SYSCTL_CHILDREN(rack_misc),
1616 OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1617 &rack_send_a_lot_in_prr, 1,
1618 "Send a lot in prr");
1619 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1620 SYSCTL_CHILDREN(rack_misc),
1621 OID_AUTO, "autoscale", CTLFLAG_RW,
1622 &rack_autosndbuf_inc, 20,
1623 "What percentage should rack scale up its snd buffer by?");
1624
1625
1626 /* Sack Attacker detection stuff */
1627 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1628 SYSCTL_CHILDREN(rack_attack),
1629 OID_AUTO, "merge_out", CTLFLAG_RW,
1630 &rack_merge_out_sacks_on_attack, 0,
1631 "Do we merge the sendmap when we decide we are being attacked?");
1632
1633 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1634 SYSCTL_CHILDREN(rack_attack),
1635 OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1636 &rack_highest_sack_thresh_seen, 0,
1637 "Highest sack to ack ratio seen");
1638 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1639 SYSCTL_CHILDREN(rack_attack),
1640 OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1641 &rack_highest_move_thresh_seen, 0,
1642 "Highest move to non-move ratio seen");
1643 rack_ack_total = counter_u64_alloc(M_WAITOK);
1644 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1645 SYSCTL_CHILDREN(rack_attack),
1646 OID_AUTO, "acktotal", CTLFLAG_RD,
1647 &rack_ack_total,
1648 "Total number of Ack's");
1649 rack_express_sack = counter_u64_alloc(M_WAITOK);
1650 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1651 SYSCTL_CHILDREN(rack_attack),
1652 OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1653 &rack_express_sack,
1654 "Total expresss number of Sack's");
1655 rack_sack_total = counter_u64_alloc(M_WAITOK);
1656 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1657 SYSCTL_CHILDREN(rack_attack),
1658 OID_AUTO, "sacktotal", CTLFLAG_RD,
1659 &rack_sack_total,
1660 "Total number of SACKs");
1661 rack_move_none = counter_u64_alloc(M_WAITOK);
1662 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1663 SYSCTL_CHILDREN(rack_attack),
1664 OID_AUTO, "move_none", CTLFLAG_RD,
1665 &rack_move_none,
1666 "Total number of SACK index reuse of positions under threshold");
1667 rack_move_some = counter_u64_alloc(M_WAITOK);
1668 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1669 SYSCTL_CHILDREN(rack_attack),
1670 OID_AUTO, "move_some", CTLFLAG_RD,
1671 &rack_move_some,
1672 "Total number of SACK index reuse of positions over threshold");
1673 rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1674 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1675 SYSCTL_CHILDREN(rack_attack),
1676 OID_AUTO, "attacks", CTLFLAG_RD,
1677 &rack_sack_attacks_detected,
1678 "Total number of SACK attackers that had sack disabled");
1679 rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1680 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681 SYSCTL_CHILDREN(rack_attack),
1682 OID_AUTO, "reversed", CTLFLAG_RD,
1683 &rack_sack_attacks_reversed,
1684 "Total number of SACK attackers that were later determined false positive");
1685 rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1686 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687 SYSCTL_CHILDREN(rack_attack),
1688 OID_AUTO, "suspect", CTLFLAG_RD,
1689 &rack_sack_attacks_suspect,
1690 "Total number of SACKs that triggered early detection");
1691
1692 rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1693 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1694 SYSCTL_CHILDREN(rack_attack),
1695 OID_AUTO, "nextmerge", CTLFLAG_RD,
1696 &rack_sack_used_next_merge,
1697 "Total number of times we used the next merge");
1698 rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1699 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1700 SYSCTL_CHILDREN(rack_attack),
1701 OID_AUTO, "prevmerge", CTLFLAG_RD,
1702 &rack_sack_used_prev_merge,
1703 "Total number of times we used the prev merge");
1704 /* Counters */
1705 rack_total_bytes = counter_u64_alloc(M_WAITOK);
1706 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1707 SYSCTL_CHILDREN(rack_counters),
1708 OID_AUTO, "totalbytes", CTLFLAG_RD,
1709 &rack_total_bytes,
1710 "Total number of bytes sent");
1711 rack_fto_send = counter_u64_alloc(M_WAITOK);
1712 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1713 SYSCTL_CHILDREN(rack_counters),
1714 OID_AUTO, "fto_send", CTLFLAG_RD,
1715 &rack_fto_send, "Total number of rack_fast_output sends");
1716 rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1717 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1718 SYSCTL_CHILDREN(rack_counters),
1719 OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1720 &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1721 rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1722 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1723 SYSCTL_CHILDREN(rack_counters),
1724 OID_AUTO, "nfto_resend", CTLFLAG_RD,
1725 &rack_nfto_resend, "Total number of rack_output retransmissions");
1726 rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1727 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1728 SYSCTL_CHILDREN(rack_counters),
1729 OID_AUTO, "nfto_send", CTLFLAG_RD,
1730 &rack_non_fto_send, "Total number of rack_output first sends");
1731 rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1732 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1733 SYSCTL_CHILDREN(rack_counters),
1734 OID_AUTO, "rfo_extended", CTLFLAG_RD,
1735 &rack_extended_rfo, "Total number of times we extended rfo");
1736
1737 rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1738 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1739 SYSCTL_CHILDREN(rack_counters),
1740 OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1741 &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1742 rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1743
1744 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1745 SYSCTL_CHILDREN(rack_counters),
1746 OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1747 &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1748 rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1749 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 SYSCTL_CHILDREN(rack_counters),
1751 OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1752 &rack_tlp_tot,
1753 "Total number of tail loss probe expirations");
1754 rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1755 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 SYSCTL_CHILDREN(rack_counters),
1757 OID_AUTO, "tlp_new", CTLFLAG_RD,
1758 &rack_tlp_newdata,
1759 "Total number of tail loss probe sending new data");
1760 rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1761 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762 SYSCTL_CHILDREN(rack_counters),
1763 OID_AUTO, "tlp_retran", CTLFLAG_RD,
1764 &rack_tlp_retran,
1765 "Total number of tail loss probe sending retransmitted data");
1766 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1767 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1768 SYSCTL_CHILDREN(rack_counters),
1769 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1770 &rack_tlp_retran_bytes,
1771 "Total bytes of tail loss probe sending retransmitted data");
1772 rack_to_tot = counter_u64_alloc(M_WAITOK);
1773 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1774 SYSCTL_CHILDREN(rack_counters),
1775 OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1776 &rack_to_tot,
1777 "Total number of times the rack to expired");
1778 rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1779 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1780 SYSCTL_CHILDREN(rack_counters),
1781 OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1782 &rack_saw_enobuf,
1783 "Total number of times a sends returned enobuf for non-hdwr paced connections");
1784 rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1785 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1786 SYSCTL_CHILDREN(rack_counters),
1787 OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1788 &rack_saw_enobuf_hw,
1789 "Total number of times a send returned enobuf for hdwr paced connections");
1790 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1791 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1792 SYSCTL_CHILDREN(rack_counters),
1793 OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1794 &rack_saw_enetunreach,
1795 "Total number of times a send received a enetunreachable");
1796 rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1797 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1798 SYSCTL_CHILDREN(rack_counters),
1799 OID_AUTO, "alloc_hot", CTLFLAG_RD,
1800 &rack_hot_alloc,
1801 "Total allocations from the top of our list");
1802 rack_to_alloc = counter_u64_alloc(M_WAITOK);
1803 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804 SYSCTL_CHILDREN(rack_counters),
1805 OID_AUTO, "allocs", CTLFLAG_RD,
1806 &rack_to_alloc,
1807 "Total allocations of tracking structures");
1808 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1809 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1810 SYSCTL_CHILDREN(rack_counters),
1811 OID_AUTO, "allochard", CTLFLAG_RD,
1812 &rack_to_alloc_hard,
1813 "Total allocations done with sleeping the hard way");
1814 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1815 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1816 SYSCTL_CHILDREN(rack_counters),
1817 OID_AUTO, "allocemerg", CTLFLAG_RD,
1818 &rack_to_alloc_emerg,
1819 "Total allocations done from emergency cache");
1820 rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1821 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1822 SYSCTL_CHILDREN(rack_counters),
1823 OID_AUTO, "alloc_limited", CTLFLAG_RD,
1824 &rack_to_alloc_limited,
1825 "Total allocations dropped due to limit");
1826 rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1827 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1828 SYSCTL_CHILDREN(rack_counters),
1829 OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1830 &rack_alloc_limited_conns,
1831 "Connections with allocations dropped due to limit");
1832 rack_split_limited = counter_u64_alloc(M_WAITOK);
1833 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1834 SYSCTL_CHILDREN(rack_counters),
1835 OID_AUTO, "split_limited", CTLFLAG_RD,
1836 &rack_split_limited,
1837 "Split allocations dropped due to limit");
1838 rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1839 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1840 SYSCTL_CHILDREN(rack_counters),
1841 OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1842 &rack_rxt_clamps_cwnd,
1843 "Number of times that excessive rxt clamped the cwnd down");
1844 rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1845 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1846 SYSCTL_CHILDREN(rack_counters),
1847 OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1848 &rack_rxt_clamps_cwnd_uniq,
1849 "Number of connections that have had excessive rxt clamped the cwnd down");
1850 rack_persists_sends = counter_u64_alloc(M_WAITOK);
1851 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1852 SYSCTL_CHILDREN(rack_counters),
1853 OID_AUTO, "persist_sends", CTLFLAG_RD,
1854 &rack_persists_sends,
1855 "Number of times we sent a persist probe");
1856 rack_persists_acks = counter_u64_alloc(M_WAITOK);
1857 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1858 SYSCTL_CHILDREN(rack_counters),
1859 OID_AUTO, "persist_acks", CTLFLAG_RD,
1860 &rack_persists_acks,
1861 "Number of times a persist probe was acked");
1862 rack_persists_loss = counter_u64_alloc(M_WAITOK);
1863 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1864 SYSCTL_CHILDREN(rack_counters),
1865 OID_AUTO, "persist_loss", CTLFLAG_RD,
1866 &rack_persists_loss,
1867 "Number of times we detected a lost persist probe (no ack)");
1868 rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1869 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1870 SYSCTL_CHILDREN(rack_counters),
1871 OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1872 &rack_persists_lost_ends,
1873 "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1874 #ifdef INVARIANTS
1875 rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1876 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1877 SYSCTL_CHILDREN(rack_counters),
1878 OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1879 &rack_adjust_map_bw,
1880 "Number of times we hit the case where the sb went up and down on a sendmap entry");
1881 #endif
1882 rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1883 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1884 SYSCTL_CHILDREN(rack_counters),
1885 OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1886 &rack_multi_single_eq,
1887 "Number of compressed acks total represented");
1888 rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1889 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1890 SYSCTL_CHILDREN(rack_counters),
1891 OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1892 &rack_proc_non_comp_ack,
1893 "Number of non compresseds acks that we processed");
1894
1895
1896 rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1897 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1898 SYSCTL_CHILDREN(rack_counters),
1899 OID_AUTO, "sack_long", CTLFLAG_RD,
1900 &rack_sack_proc_all,
1901 "Total times we had to walk whole list for sack processing");
1902 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1903 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1904 SYSCTL_CHILDREN(rack_counters),
1905 OID_AUTO, "sack_restart", CTLFLAG_RD,
1906 &rack_sack_proc_restart,
1907 "Total times we had to walk whole list due to a restart");
1908 rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1909 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1910 SYSCTL_CHILDREN(rack_counters),
1911 OID_AUTO, "sack_short", CTLFLAG_RD,
1912 &rack_sack_proc_short,
1913 "Total times we took shortcut for sack processing");
1914 rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1915 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1916 SYSCTL_CHILDREN(rack_attack),
1917 OID_AUTO, "skipacked", CTLFLAG_RD,
1918 &rack_sack_skipped_acked,
1919 "Total number of times we skipped previously sacked");
1920 rack_sack_splits = counter_u64_alloc(M_WAITOK);
1921 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1922 SYSCTL_CHILDREN(rack_attack),
1923 OID_AUTO, "ofsplit", CTLFLAG_RD,
1924 &rack_sack_splits,
1925 "Total number of times we did the old fashion tree split");
1926 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1927 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1928 SYSCTL_CHILDREN(rack_counters),
1929 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1930 &rack_input_idle_reduces,
1931 "Total number of idle reductions on input");
1932 rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1933 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1934 SYSCTL_CHILDREN(rack_counters),
1935 OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1936 &rack_collapsed_win_seen,
1937 "Total number of collapsed window events seen (where our window shrinks)");
1938
1939 rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1940 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1941 SYSCTL_CHILDREN(rack_counters),
1942 OID_AUTO, "collapsed_win", CTLFLAG_RD,
1943 &rack_collapsed_win,
1944 "Total number of collapsed window events where we mark packets");
1945 rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1946 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1947 SYSCTL_CHILDREN(rack_counters),
1948 OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1949 &rack_collapsed_win_rxt,
1950 "Total number of packets that were retransmitted");
1951 rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1952 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1953 SYSCTL_CHILDREN(rack_counters),
1954 OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1955 &rack_collapsed_win_rxt_bytes,
1956 "Total number of bytes that were retransmitted");
1957 rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1958 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1959 SYSCTL_CHILDREN(rack_counters),
1960 OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1961 &rack_try_scwnd,
1962 "Total number of scwnd attempts");
1963 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1964 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1965 OID_AUTO, "outsize", CTLFLAG_RD,
1966 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1967 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1968 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1969 OID_AUTO, "opts", CTLFLAG_RD,
1970 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1971 SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1972 SYSCTL_CHILDREN(rack_sysctl_root),
1973 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1974 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1975 }
1976
1977 static uint32_t
rc_init_window(struct tcp_rack * rack)1978 rc_init_window(struct tcp_rack *rack)
1979 {
1980 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1981
1982 }
1983
1984 static uint64_t
rack_get_fixed_pacing_bw(struct tcp_rack * rack)1985 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1986 {
1987 if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1988 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1989 else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1990 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1991 else
1992 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1993 }
1994
1995 static void
rack_log_hybrid_bw(struct tcp_rack * rack,uint32_t seq,uint64_t cbw,uint64_t tim,uint64_t data,uint8_t mod,uint16_t aux,struct tcp_sendfile_track * cur,int line)1996 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1997 uint64_t data, uint8_t mod, uint16_t aux,
1998 struct tcp_sendfile_track *cur, int line)
1999 {
2000 #ifdef TCP_REQUEST_TRK
2001 int do_log = 0;
2002
2003 /*
2004 * The rate cap one is noisy and only should come out when normal BB logging
2005 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2006 * once per chunk and make up the BBpoint that can be turned on by the client.
2007 */
2008 if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2009 /*
2010 * The very noisy two need to only come out when
2011 * we have verbose logging on.
2012 */
2013 if (rack_verbose_logging != 0)
2014 do_log = tcp_bblogging_on(rack->rc_tp);
2015 else
2016 do_log = 0;
2017 } else if (mod != HYBRID_LOG_BW_MEASURE) {
2018 /*
2019 * All other less noisy logs here except the measure which
2020 * also needs to come out on the point and the log.
2021 */
2022 do_log = tcp_bblogging_on(rack->rc_tp);
2023 } else {
2024 do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2025 }
2026
2027 if (do_log) {
2028 union tcp_log_stackspecific log;
2029 struct timeval tv;
2030 uint64_t lt_bw;
2031
2032 /* Convert our ms to a microsecond */
2033 memset(&log, 0, sizeof(log));
2034
2035 log.u_bbr.cwnd_gain = line;
2036 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2037 log.u_bbr.rttProp = tim;
2038 log.u_bbr.bw_inuse = cbw;
2039 log.u_bbr.delRate = rack_get_gp_est(rack);
2040 lt_bw = rack_get_lt_bw(rack);
2041 log.u_bbr.flex1 = seq;
2042 log.u_bbr.pacing_gain = aux;
2043 /* lt_bw = < flex3 | flex2 > */
2044 log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2045 log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2046 /* Record the last obtained us rtt in inflight */
2047 if (cur == NULL) {
2048 /* Make sure we are looking at the right log if an overide comes in */
2049 cur = rack->r_ctl.rc_last_sft;
2050 }
2051 if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2052 log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2053 else {
2054 /* Use the last known rtt i.e. the rack-rtt */
2055 log.u_bbr.inflight = rack->rc_rack_rtt;
2056 }
2057 if (cur != NULL) {
2058 uint64_t off;
2059
2060 log.u_bbr.cur_del_rate = cur->deadline;
2061 if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2062 /* start = < lost | pkt_epoch > */
2063 log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2064 log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2065 log.u_bbr.flex6 = cur->start_seq;
2066 log.u_bbr.pkts_out = cur->end_seq;
2067 } else {
2068 /* start = < lost | pkt_epoch > */
2069 log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2070 log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2071 /* end = < pkts_out | flex6 > */
2072 log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2073 log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2074 }
2075 /* first_send = <lt_epoch | epoch> */
2076 log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2077 log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2078 /* localtime = <delivered | applimited>*/
2079 log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2080 log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2081 #ifdef TCP_REQUEST_TRK
2082 off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2083 log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2084 #endif
2085 log.u_bbr.inhpts = 1;
2086 log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2087 log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2088 log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2089 } else {
2090 log.u_bbr.flex7 = 0xffff;
2091 log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2092 }
2093 /*
2094 * Compose bbr_state to be a bit wise 0000ADHF
2095 * where A is the always_pace flag
2096 * where D is the dgp_on flag
2097 * where H is the hybrid_mode on flag
2098 * where F is the use_fixed_rate flag.
2099 */
2100 log.u_bbr.bbr_state = rack->rc_always_pace;
2101 log.u_bbr.bbr_state <<= 1;
2102 log.u_bbr.bbr_state |= rack->dgp_on;
2103 log.u_bbr.bbr_state <<= 1;
2104 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2105 log.u_bbr.bbr_state <<= 1;
2106 log.u_bbr.bbr_state |= rack->use_fixed_rate;
2107 log.u_bbr.flex8 = mod;
2108 tcp_log_event(rack->rc_tp, NULL,
2109 &rack->rc_inp->inp_socket->so_rcv,
2110 &rack->rc_inp->inp_socket->so_snd,
2111 TCP_HYBRID_PACING_LOG, 0,
2112 0, &log, false, NULL, __func__, __LINE__, &tv);
2113
2114 }
2115 #endif
2116 }
2117
2118 #ifdef TCP_REQUEST_TRK
2119 static void
rack_log_hybrid_sends(struct tcp_rack * rack,struct tcp_sendfile_track * cur,int line)2120 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2121 {
2122 if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2123 union tcp_log_stackspecific log;
2124 struct timeval tv;
2125 uint64_t off;
2126
2127 /* Convert our ms to a microsecond */
2128 memset(&log, 0, sizeof(log));
2129
2130 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2131 log.u_bbr.delRate = cur->sent_at_fs;
2132
2133 if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2134 /*
2135 * We did not get a new Rules Applied to set so
2136 * no overlapping send occured, this means the
2137 * current byte counts are correct.
2138 */
2139 log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2140 log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2141 } else {
2142 /*
2143 * Overlapping send case, we switched to a new
2144 * send and did a rules applied.
2145 */
2146 log.u_bbr.cur_del_rate = cur->sent_at_ls;
2147 log.u_bbr.rttProp = cur->rxt_at_ls;
2148 }
2149 log.u_bbr.bw_inuse = cur->rxt_at_fs;
2150 log.u_bbr.cwnd_gain = line;
2151 off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2152 log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2153 /* start = < flex1 | flex2 > */
2154 log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2155 log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2156 /* end = < flex3 | flex4 > */
2157 log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2158 log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2159
2160 /* localtime = <delivered | applimited>*/
2161 log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2162 log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2163 /* client timestamp = <lt_epoch | epoch>*/
2164 log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2165 log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2166 /* now set all the flags in */
2167 log.u_bbr.pkts_out = cur->hybrid_flags;
2168 log.u_bbr.lost = cur->playout_ms;
2169 log.u_bbr.flex6 = cur->flags;
2170 /*
2171 * Last send time = <flex5 | pkt_epoch> note we do not distinguish cases
2172 * where a false retransmit occurred so first_send <-> lastsend may
2173 * include longer time then it actually took if we have a false rxt.
2174 */
2175 log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2176 log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2177 /*
2178 * Compose bbr_state to be a bit wise 0000ADHF
2179 * where A is the always_pace flag
2180 * where D is the dgp_on flag
2181 * where H is the hybrid_mode on flag
2182 * where F is the use_fixed_rate flag.
2183 */
2184 log.u_bbr.bbr_state = rack->rc_always_pace;
2185 log.u_bbr.bbr_state <<= 1;
2186 log.u_bbr.bbr_state |= rack->dgp_on;
2187 log.u_bbr.bbr_state <<= 1;
2188 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2189 log.u_bbr.bbr_state <<= 1;
2190 log.u_bbr.bbr_state |= rack->use_fixed_rate;
2191
2192 log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2193 tcp_log_event(rack->rc_tp, NULL,
2194 &rack->rc_inp->inp_socket->so_rcv,
2195 &rack->rc_inp->inp_socket->so_snd,
2196 TCP_HYBRID_PACING_LOG, 0,
2197 0, &log, false, NULL, __func__, __LINE__, &tv);
2198 }
2199 }
2200 #endif
2201
2202 static inline uint64_t
rack_compensate_for_linerate(struct tcp_rack * rack,uint64_t bw)2203 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2204 {
2205 uint64_t ret_bw, ether;
2206 uint64_t u_segsiz;
2207
2208 ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2209 if (rack->r_is_v6){
2210 #ifdef INET6
2211 ether += sizeof(struct ip6_hdr);
2212 #endif
2213 ether += 14; /* eheader size 6+6+2 */
2214 } else {
2215 #ifdef INET
2216 ether += sizeof(struct ip);
2217 #endif
2218 ether += 14; /* eheader size 6+6+2 */
2219 }
2220 u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2221 ret_bw = bw;
2222 ret_bw *= ether;
2223 ret_bw /= u_segsiz;
2224 return (ret_bw);
2225 }
2226
2227 static void
rack_rate_cap_bw(struct tcp_rack * rack,uint64_t * bw,int * capped)2228 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2229 {
2230 #ifdef TCP_REQUEST_TRK
2231 struct timeval tv;
2232 uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2233 #endif
2234
2235 if (rack->r_ctl.bw_rate_cap == 0)
2236 return;
2237 #ifdef TCP_REQUEST_TRK
2238 if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2239 (rack->r_ctl.rc_last_sft != NULL)) {
2240 /*
2241 * We have a dynamic cap. The original target
2242 * is in bw_rate_cap, but we need to look at
2243 * how long it is until we hit the deadline.
2244 */
2245 struct tcp_sendfile_track *ent;
2246
2247 ent = rack->r_ctl.rc_last_sft;
2248 microuptime(&tv);
2249 timenow = tcp_tv_to_lusectick(&tv);
2250 if (timenow >= ent->deadline) {
2251 /* No time left we do DGP only */
2252 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2253 0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2254 rack->r_ctl.bw_rate_cap = 0;
2255 return;
2256 }
2257 /* We have the time */
2258 timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2259 if (timeleft < HPTS_MSEC_IN_SEC) {
2260 /* If there is less than a ms left just use DGPs rate */
2261 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2262 0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2263 rack->r_ctl.bw_rate_cap = 0;
2264 return;
2265 }
2266 /*
2267 * Now lets find the amount of data left to send.
2268 *
2269 * Now ideally we want to use the end_seq to figure out how much more
2270 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2271 */
2272 if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2273 if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2274 lenleft = ent->end_seq - rack->rc_tp->snd_una;
2275 else {
2276 /* TSNH, we should catch it at the send */
2277 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2278 0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2279 rack->r_ctl.bw_rate_cap = 0;
2280 return;
2281 }
2282 } else {
2283 /*
2284 * The hard way, figure out how much is gone and then
2285 * take that away from the total the client asked for
2286 * (thats off by tls overhead if this is tls).
2287 */
2288 if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2289 lengone = rack->rc_tp->snd_una - ent->start_seq;
2290 else
2291 lengone = 0;
2292 if (lengone < (ent->end - ent->start))
2293 lenleft = (ent->end - ent->start) - lengone;
2294 else {
2295 /* TSNH, we should catch it at the send */
2296 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2297 0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2298 rack->r_ctl.bw_rate_cap = 0;
2299 return;
2300 }
2301 }
2302 if (lenleft == 0) {
2303 /* We have it all sent */
2304 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2305 0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2306 if (rack->r_ctl.bw_rate_cap)
2307 goto normal_ratecap;
2308 else
2309 return;
2310 }
2311 calcbw = lenleft * HPTS_USEC_IN_SEC;
2312 calcbw /= timeleft;
2313 /* Now we must compensate for IP/TCP overhead */
2314 calcbw = rack_compensate_for_linerate(rack, calcbw);
2315 /* Update the bit rate cap */
2316 rack->r_ctl.bw_rate_cap = calcbw;
2317 if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2318 (rack_hybrid_allow_set_maxseg == 1) &&
2319 ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2320 /* Lets set in a smaller mss possibly here to match our rate-cap */
2321 uint32_t orig_max;
2322
2323 orig_max = rack->r_ctl.rc_pace_max_segs;
2324 rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2325 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2326 rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2327 }
2328 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2329 calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2330 if ((calcbw > 0) && (*bw > calcbw)) {
2331 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2332 *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2333 *capped = 1;
2334 *bw = calcbw;
2335 }
2336 return;
2337 }
2338 normal_ratecap:
2339 #endif
2340 if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2341 #ifdef TCP_REQUEST_TRK
2342 if (rack->rc_hybrid_mode &&
2343 rack->rc_catch_up &&
2344 (rack->r_ctl.rc_last_sft != NULL) &&
2345 (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2346 (rack_hybrid_allow_set_maxseg == 1) &&
2347 ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2348 /* Lets set in a smaller mss possibly here to match our rate-cap */
2349 uint32_t orig_max;
2350
2351 orig_max = rack->r_ctl.rc_pace_max_segs;
2352 rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2353 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2354 rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2355 }
2356 #endif
2357 *capped = 1;
2358 *bw = rack->r_ctl.bw_rate_cap;
2359 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2360 *bw, 0, 0,
2361 HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2362 }
2363 }
2364
2365 static uint64_t
rack_get_gp_est(struct tcp_rack * rack)2366 rack_get_gp_est(struct tcp_rack *rack)
2367 {
2368 uint64_t bw, lt_bw, ret_bw;
2369
2370 if (rack->rc_gp_filled == 0) {
2371 /*
2372 * We have yet no b/w measurement,
2373 * if we have a user set initial bw
2374 * return it. If we don't have that and
2375 * we have an srtt, use the tcp IW (10) to
2376 * calculate a fictional b/w over the SRTT
2377 * which is more or less a guess. Note
2378 * we don't use our IW from rack on purpose
2379 * so if we have like IW=30, we are not
2380 * calculating a "huge" b/w.
2381 */
2382 uint64_t srtt;
2383
2384 if (rack->dis_lt_bw == 1)
2385 lt_bw = 0;
2386 else
2387 lt_bw = rack_get_lt_bw(rack);
2388 if (lt_bw) {
2389 /*
2390 * No goodput bw but a long-term b/w does exist
2391 * lets use that.
2392 */
2393 ret_bw = lt_bw;
2394 goto compensate;
2395 }
2396 if (rack->r_ctl.init_rate)
2397 return (rack->r_ctl.init_rate);
2398
2399 /* Ok lets come up with the IW guess, if we have a srtt */
2400 if (rack->rc_tp->t_srtt == 0) {
2401 /*
2402 * Go with old pacing method
2403 * i.e. burst mitigation only.
2404 */
2405 return (0);
2406 }
2407 /* Ok lets get the initial TCP win (not racks) */
2408 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2409 srtt = (uint64_t)rack->rc_tp->t_srtt;
2410 bw *= (uint64_t)USECS_IN_SECOND;
2411 bw /= srtt;
2412 ret_bw = bw;
2413 goto compensate;
2414
2415 }
2416 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2417 /* Averaging is done, we can return the value */
2418 bw = rack->r_ctl.gp_bw;
2419 } else {
2420 /* Still doing initial average must calculate */
2421 bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2422 }
2423 if (rack->dis_lt_bw) {
2424 /* We are not using lt-bw */
2425 ret_bw = bw;
2426 goto compensate;
2427 }
2428 lt_bw = rack_get_lt_bw(rack);
2429 if (lt_bw == 0) {
2430 /* If we don't have one then equate it to the gp_bw */
2431 lt_bw = rack->r_ctl.gp_bw;
2432 }
2433 if (rack->use_lesser_lt_bw) {
2434 if (lt_bw < bw)
2435 ret_bw = lt_bw;
2436 else
2437 ret_bw = bw;
2438 } else {
2439 if (lt_bw > bw)
2440 ret_bw = lt_bw;
2441 else
2442 ret_bw = bw;
2443 }
2444 /*
2445 * Now lets compensate based on the TCP/IP overhead. Our
2446 * Goodput estimate does not include this so we must pace out
2447 * a bit faster since our pacing calculations do. The pacing
2448 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2449 * we are using to do this, so we do that here in the opposite
2450 * direction as well. This means that if we are tunneled and the
2451 * segsiz is say 1200 bytes we will get quite a boost, but its
2452 * compensated for in the pacing time the opposite way.
2453 */
2454 compensate:
2455 ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2456 return(ret_bw);
2457 }
2458
2459
2460 static uint64_t
rack_get_bw(struct tcp_rack * rack)2461 rack_get_bw(struct tcp_rack *rack)
2462 {
2463 uint64_t bw;
2464
2465 if (rack->use_fixed_rate) {
2466 /* Return the fixed pacing rate */
2467 return (rack_get_fixed_pacing_bw(rack));
2468 }
2469 bw = rack_get_gp_est(rack);
2470 return (bw);
2471 }
2472
2473 static uint16_t
rack_get_output_gain(struct tcp_rack * rack,struct rack_sendmap * rsm)2474 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2475 {
2476 if (rack->use_fixed_rate) {
2477 return (100);
2478 } else if (rack->in_probe_rtt && (rsm == NULL))
2479 return (rack->r_ctl.rack_per_of_gp_probertt);
2480 else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2481 rack->r_ctl.rack_per_of_gp_rec)) {
2482 if (rsm) {
2483 /* a retransmission always use the recovery rate */
2484 return (rack->r_ctl.rack_per_of_gp_rec);
2485 } else if (rack->rack_rec_nonrxt_use_cr) {
2486 /* Directed to use the configured rate */
2487 goto configured_rate;
2488 } else if (rack->rack_no_prr &&
2489 (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2490 /* No PRR, lets just use the b/w estimate only */
2491 return (100);
2492 } else {
2493 /*
2494 * Here we may have a non-retransmit but we
2495 * have no overrides, so just use the recovery
2496 * rate (prr is in effect).
2497 */
2498 return (rack->r_ctl.rack_per_of_gp_rec);
2499 }
2500 }
2501 configured_rate:
2502 /* For the configured rate we look at our cwnd vs the ssthresh */
2503 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2504 return (rack->r_ctl.rack_per_of_gp_ss);
2505 else
2506 return (rack->r_ctl.rack_per_of_gp_ca);
2507 }
2508
2509 static void
rack_log_dsack_event(struct tcp_rack * rack,uint8_t mod,uint32_t flex4,uint32_t flex5,uint32_t flex6)2510 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2511 {
2512 /*
2513 * Types of logs (mod value)
2514 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2515 * 2 = a dsack round begins, persist is reset to 16.
2516 * 3 = a dsack round ends
2517 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2518 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2519 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2520 */
2521 if (tcp_bblogging_on(rack->rc_tp)) {
2522 union tcp_log_stackspecific log;
2523 struct timeval tv;
2524
2525 memset(&log, 0, sizeof(log));
2526 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2527 log.u_bbr.flex1 <<= 1;
2528 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2529 log.u_bbr.flex1 <<= 1;
2530 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2531 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2532 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2533 log.u_bbr.flex4 = flex4;
2534 log.u_bbr.flex5 = flex5;
2535 log.u_bbr.flex6 = flex6;
2536 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2537 log.u_bbr.flex8 = mod;
2538 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2539 log.u_bbr.epoch = rack->r_ctl.current_round;
2540 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2541 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2542 &rack->rc_inp->inp_socket->so_rcv,
2543 &rack->rc_inp->inp_socket->so_snd,
2544 RACK_DSACK_HANDLING, 0,
2545 0, &log, false, &tv);
2546 }
2547 }
2548
2549 static void
rack_log_hdwr_pacing(struct tcp_rack * rack,uint64_t rate,uint64_t hw_rate,int line,int error,uint16_t mod)2550 rack_log_hdwr_pacing(struct tcp_rack *rack,
2551 uint64_t rate, uint64_t hw_rate, int line,
2552 int error, uint16_t mod)
2553 {
2554 if (tcp_bblogging_on(rack->rc_tp)) {
2555 union tcp_log_stackspecific log;
2556 struct timeval tv;
2557 const struct ifnet *ifp;
2558 uint64_t ifp64;
2559
2560 memset(&log, 0, sizeof(log));
2561 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2562 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2563 if (rack->r_ctl.crte) {
2564 ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2565 } else if (rack->rc_inp->inp_route.ro_nh &&
2566 rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2567 ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2568 } else
2569 ifp = NULL;
2570 if (ifp) {
2571 ifp64 = (uintptr_t)ifp;
2572 log.u_bbr.flex3 = ((ifp64 >> 32) & 0x00000000ffffffff);
2573 log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2574 }
2575 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2576 log.u_bbr.bw_inuse = rate;
2577 log.u_bbr.flex5 = line;
2578 log.u_bbr.flex6 = error;
2579 log.u_bbr.flex7 = mod;
2580 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2581 log.u_bbr.flex8 = rack->use_fixed_rate;
2582 log.u_bbr.flex8 <<= 1;
2583 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2584 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2585 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2586 if (rack->r_ctl.crte)
2587 log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2588 else
2589 log.u_bbr.cur_del_rate = 0;
2590 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2591 log.u_bbr.epoch = rack->r_ctl.current_round;
2592 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2593 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2594 &rack->rc_inp->inp_socket->so_rcv,
2595 &rack->rc_inp->inp_socket->so_snd,
2596 BBR_LOG_HDWR_PACE, 0,
2597 0, &log, false, &tv);
2598 }
2599 }
2600
2601 static uint64_t
rack_get_output_bw(struct tcp_rack * rack,uint64_t bw,struct rack_sendmap * rsm,int * capped)2602 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2603 {
2604 /*
2605 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2606 */
2607 uint64_t bw_est, high_rate;
2608 uint64_t gain;
2609
2610 gain = (uint64_t)rack_get_output_gain(rack, rsm);
2611 bw_est = bw * gain;
2612 bw_est /= (uint64_t)100;
2613 /* Never fall below the minimum (def 64kbps) */
2614 if (bw_est < RACK_MIN_BW)
2615 bw_est = RACK_MIN_BW;
2616 if (rack->r_rack_hw_rate_caps) {
2617 /* Rate caps are in place */
2618 if (rack->r_ctl.crte != NULL) {
2619 /* We have a hdwr rate already */
2620 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2621 if (bw_est >= high_rate) {
2622 /* We are capping bw at the highest rate table entry */
2623 if (rack_hw_rate_cap_per &&
2624 (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2625 rack->r_rack_hw_rate_caps = 0;
2626 goto done;
2627 }
2628 rack_log_hdwr_pacing(rack,
2629 bw_est, high_rate, __LINE__,
2630 0, 3);
2631 bw_est = high_rate;
2632 if (capped)
2633 *capped = 1;
2634 }
2635 } else if ((rack->rack_hdrw_pacing == 0) &&
2636 (rack->rack_hdw_pace_ena) &&
2637 (rack->rack_attempt_hdwr_pace == 0) &&
2638 (rack->rc_inp->inp_route.ro_nh != NULL) &&
2639 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2640 /*
2641 * Special case, we have not yet attempted hardware
2642 * pacing, and yet we may, when we do, find out if we are
2643 * above the highest rate. We need to know the maxbw for the interface
2644 * in question (if it supports ratelimiting). We get back
2645 * a 0, if the interface is not found in the RL lists.
2646 */
2647 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2648 if (high_rate) {
2649 /* Yep, we have a rate is it above this rate? */
2650 if (bw_est > high_rate) {
2651 bw_est = high_rate;
2652 if (capped)
2653 *capped = 1;
2654 }
2655 }
2656 }
2657 }
2658 done:
2659 return (bw_est);
2660 }
2661
2662 static void
rack_log_retran_reason(struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t tsused,uint32_t thresh,int mod)2663 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2664 {
2665 if (tcp_bblogging_on(rack->rc_tp)) {
2666 union tcp_log_stackspecific log;
2667 struct timeval tv;
2668
2669 if ((mod != 1) && (rack_verbose_logging == 0)) {
2670 /*
2671 * We get 3 values currently for mod
2672 * 1 - We are retransmitting and this tells the reason.
2673 * 2 - We are clearing a dup-ack count.
2674 * 3 - We are incrementing a dup-ack count.
2675 *
2676 * The clear/increment are only logged
2677 * if you have BBverbose on.
2678 */
2679 return;
2680 }
2681 memset(&log, 0, sizeof(log));
2682 log.u_bbr.flex1 = tsused;
2683 log.u_bbr.flex2 = thresh;
2684 log.u_bbr.flex3 = rsm->r_flags;
2685 log.u_bbr.flex4 = rsm->r_dupack;
2686 log.u_bbr.flex5 = rsm->r_start;
2687 log.u_bbr.flex6 = rsm->r_end;
2688 log.u_bbr.flex8 = mod;
2689 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2690 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2691 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2692 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2693 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2694 log.u_bbr.pacing_gain = rack->r_must_retran;
2695 log.u_bbr.epoch = rack->r_ctl.current_round;
2696 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2697 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2698 &rack->rc_inp->inp_socket->so_rcv,
2699 &rack->rc_inp->inp_socket->so_snd,
2700 BBR_LOG_SETTINGS_CHG, 0,
2701 0, &log, false, &tv);
2702 }
2703 }
2704
2705 static void
rack_log_to_start(struct tcp_rack * rack,uint32_t cts,uint32_t to,int32_t slot,uint8_t which)2706 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2707 {
2708 if (tcp_bblogging_on(rack->rc_tp)) {
2709 union tcp_log_stackspecific log;
2710 struct timeval tv;
2711
2712 memset(&log, 0, sizeof(log));
2713 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2714 log.u_bbr.flex2 = to;
2715 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2716 log.u_bbr.flex4 = slot;
2717 log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2718 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2719 log.u_bbr.flex7 = rack->rc_in_persist;
2720 log.u_bbr.flex8 = which;
2721 if (rack->rack_no_prr)
2722 log.u_bbr.pkts_out = 0;
2723 else
2724 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2725 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2726 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2727 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2728 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2729 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2730 log.u_bbr.pacing_gain = rack->r_must_retran;
2731 log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2732 log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2733 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2734 log.u_bbr.lost = rack_rto_min;
2735 log.u_bbr.epoch = rack->r_ctl.roundends;
2736 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2737 log.u_bbr.bw_inuse <<= 32;
2738 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2739 log.u_bbr.applimited = rack->rc_tp->t_flags2;
2740 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2741 &rack->rc_inp->inp_socket->so_rcv,
2742 &rack->rc_inp->inp_socket->so_snd,
2743 BBR_LOG_TIMERSTAR, 0,
2744 0, &log, false, &tv);
2745 }
2746 }
2747
2748 static void
rack_log_to_event(struct tcp_rack * rack,int32_t to_num,struct rack_sendmap * rsm)2749 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2750 {
2751 if (tcp_bblogging_on(rack->rc_tp)) {
2752 union tcp_log_stackspecific log;
2753 struct timeval tv;
2754
2755 memset(&log, 0, sizeof(log));
2756 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2757 log.u_bbr.flex8 = to_num;
2758 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2759 log.u_bbr.flex2 = rack->rc_rack_rtt;
2760 if (rsm == NULL)
2761 log.u_bbr.flex3 = 0;
2762 else
2763 log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2764 if (rack->rack_no_prr)
2765 log.u_bbr.flex5 = 0;
2766 else
2767 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2768 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2769 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2770 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2771 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2772 log.u_bbr.pacing_gain = rack->r_must_retran;
2773 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2774 log.u_bbr.bw_inuse <<= 32;
2775 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2776 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2777 &rack->rc_inp->inp_socket->so_rcv,
2778 &rack->rc_inp->inp_socket->so_snd,
2779 BBR_LOG_RTO, 0,
2780 0, &log, false, &tv);
2781 }
2782 }
2783
2784 static void
rack_log_map_chg(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * prev,struct rack_sendmap * rsm,struct rack_sendmap * next,int flag,uint32_t th_ack,int line)2785 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2786 struct rack_sendmap *prev,
2787 struct rack_sendmap *rsm,
2788 struct rack_sendmap *next,
2789 int flag, uint32_t th_ack, int line)
2790 {
2791 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2792 union tcp_log_stackspecific log;
2793 struct timeval tv;
2794
2795 memset(&log, 0, sizeof(log));
2796 log.u_bbr.flex8 = flag;
2797 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2798 log.u_bbr.cur_del_rate = (uintptr_t)prev;
2799 log.u_bbr.delRate = (uintptr_t)rsm;
2800 log.u_bbr.rttProp = (uintptr_t)next;
2801 log.u_bbr.flex7 = 0;
2802 if (prev) {
2803 log.u_bbr.flex1 = prev->r_start;
2804 log.u_bbr.flex2 = prev->r_end;
2805 log.u_bbr.flex7 |= 0x4;
2806 }
2807 if (rsm) {
2808 log.u_bbr.flex3 = rsm->r_start;
2809 log.u_bbr.flex4 = rsm->r_end;
2810 log.u_bbr.flex7 |= 0x2;
2811 }
2812 if (next) {
2813 log.u_bbr.flex5 = next->r_start;
2814 log.u_bbr.flex6 = next->r_end;
2815 log.u_bbr.flex7 |= 0x1;
2816 }
2817 log.u_bbr.applimited = line;
2818 log.u_bbr.pkts_out = th_ack;
2819 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2820 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2821 if (rack->rack_no_prr)
2822 log.u_bbr.lost = 0;
2823 else
2824 log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2825 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2826 log.u_bbr.bw_inuse <<= 32;
2827 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2828 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2829 &rack->rc_inp->inp_socket->so_rcv,
2830 &rack->rc_inp->inp_socket->so_snd,
2831 TCP_LOG_MAPCHG, 0,
2832 0, &log, false, &tv);
2833 }
2834 }
2835
2836 static void
rack_log_rtt_upd(struct tcpcb * tp,struct tcp_rack * rack,uint32_t t,uint32_t len,struct rack_sendmap * rsm,int conf)2837 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2838 struct rack_sendmap *rsm, int conf)
2839 {
2840 if (tcp_bblogging_on(tp)) {
2841 union tcp_log_stackspecific log;
2842 struct timeval tv;
2843 memset(&log, 0, sizeof(log));
2844 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2845 log.u_bbr.flex1 = t;
2846 log.u_bbr.flex2 = len;
2847 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2848 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2849 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2850 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2851 log.u_bbr.flex7 = conf;
2852 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2853 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2854 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2855 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2856 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2857 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2858 if (rsm) {
2859 log.u_bbr.pkt_epoch = rsm->r_start;
2860 log.u_bbr.lost = rsm->r_end;
2861 log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2862 /* We loose any upper of the 24 bits */
2863 log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2864 } else {
2865 /* Its a SYN */
2866 log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2867 log.u_bbr.lost = 0;
2868 log.u_bbr.cwnd_gain = 0;
2869 log.u_bbr.pacing_gain = 0;
2870 }
2871 /* Write out general bits of interest rrs here */
2872 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2873 log.u_bbr.use_lt_bw <<= 1;
2874 log.u_bbr.use_lt_bw |= rack->forced_ack;
2875 log.u_bbr.use_lt_bw <<= 1;
2876 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2877 log.u_bbr.use_lt_bw <<= 1;
2878 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2879 log.u_bbr.use_lt_bw <<= 1;
2880 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2881 log.u_bbr.use_lt_bw <<= 1;
2882 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2883 log.u_bbr.use_lt_bw <<= 1;
2884 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2885 log.u_bbr.use_lt_bw <<= 1;
2886 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2887 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2888 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2889 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2890 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2891 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2892 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2893 log.u_bbr.bw_inuse <<= 32;
2894 if (rsm)
2895 log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2896 TCP_LOG_EVENTP(tp, NULL,
2897 &rack->rc_inp->inp_socket->so_rcv,
2898 &rack->rc_inp->inp_socket->so_snd,
2899 BBR_LOG_BBRRTT, 0,
2900 0, &log, false, &tv);
2901
2902
2903 }
2904 }
2905
2906 static void
rack_log_rtt_sample(struct tcp_rack * rack,uint32_t rtt)2907 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2908 {
2909 /*
2910 * Log the rtt sample we are
2911 * applying to the srtt algorithm in
2912 * useconds.
2913 */
2914 if (tcp_bblogging_on(rack->rc_tp)) {
2915 union tcp_log_stackspecific log;
2916 struct timeval tv;
2917
2918 /* Convert our ms to a microsecond */
2919 memset(&log, 0, sizeof(log));
2920 log.u_bbr.flex1 = rtt;
2921 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2922 log.u_bbr.flex7 = 1;
2923 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2924 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2925 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2926 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2927 log.u_bbr.pacing_gain = rack->r_must_retran;
2928 /*
2929 * We capture in delRate the upper 32 bits as
2930 * the confidence level we had declared, and the
2931 * lower 32 bits as the actual RTT using the arrival
2932 * timestamp.
2933 */
2934 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2935 log.u_bbr.delRate <<= 32;
2936 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2937 /* Lets capture all the things that make up t_rtxcur */
2938 log.u_bbr.applimited = rack_rto_min;
2939 log.u_bbr.epoch = rack_rto_max;
2940 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2941 log.u_bbr.lost = rack_rto_min;
2942 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2943 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2944 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2945 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2946 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2947 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2948 &rack->rc_inp->inp_socket->so_rcv,
2949 &rack->rc_inp->inp_socket->so_snd,
2950 TCP_LOG_RTT, 0,
2951 0, &log, false, &tv);
2952 }
2953 }
2954
2955 static void
rack_log_rtt_sample_calc(struct tcp_rack * rack,uint32_t rtt,uint32_t send_time,uint32_t ack_time,int where)2956 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2957 {
2958 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2959 union tcp_log_stackspecific log;
2960 struct timeval tv;
2961
2962 /* Convert our ms to a microsecond */
2963 memset(&log, 0, sizeof(log));
2964 log.u_bbr.flex1 = rtt;
2965 log.u_bbr.flex2 = send_time;
2966 log.u_bbr.flex3 = ack_time;
2967 log.u_bbr.flex4 = where;
2968 log.u_bbr.flex7 = 2;
2969 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2970 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2971 log.u_bbr.bw_inuse <<= 32;
2972 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2973 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2974 &rack->rc_inp->inp_socket->so_rcv,
2975 &rack->rc_inp->inp_socket->so_snd,
2976 TCP_LOG_RTT, 0,
2977 0, &log, false, &tv);
2978 }
2979 }
2980
2981
2982 static void
rack_log_rtt_sendmap(struct tcp_rack * rack,uint32_t idx,uint64_t tsv,uint32_t tsecho)2983 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2984 {
2985 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2986 union tcp_log_stackspecific log;
2987 struct timeval tv;
2988
2989 /* Convert our ms to a microsecond */
2990 memset(&log, 0, sizeof(log));
2991 log.u_bbr.flex1 = idx;
2992 log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2993 log.u_bbr.flex3 = tsecho;
2994 log.u_bbr.flex7 = 3;
2995 log.u_bbr.rttProp = tsv;
2996 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2997 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2998 log.u_bbr.bw_inuse <<= 32;
2999 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3000 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3001 &rack->rc_inp->inp_socket->so_rcv,
3002 &rack->rc_inp->inp_socket->so_snd,
3003 TCP_LOG_RTT, 0,
3004 0, &log, false, &tv);
3005 }
3006 }
3007
3008
3009 static inline void
rack_log_progress_event(struct tcp_rack * rack,struct tcpcb * tp,uint32_t tick,int event,int line)3010 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line)
3011 {
3012 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3013 union tcp_log_stackspecific log;
3014 struct timeval tv;
3015
3016 memset(&log, 0, sizeof(log));
3017 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3018 log.u_bbr.flex1 = line;
3019 log.u_bbr.flex2 = tick;
3020 log.u_bbr.flex3 = tp->t_maxunacktime;
3021 log.u_bbr.flex4 = tp->t_acktime;
3022 log.u_bbr.flex8 = event;
3023 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3024 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3025 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3026 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3027 log.u_bbr.pacing_gain = rack->r_must_retran;
3028 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3029 log.u_bbr.bw_inuse <<= 32;
3030 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3031 TCP_LOG_EVENTP(tp, NULL,
3032 &rack->rc_inp->inp_socket->so_rcv,
3033 &rack->rc_inp->inp_socket->so_snd,
3034 BBR_LOG_PROGRESS, 0,
3035 0, &log, false, &tv);
3036 }
3037 }
3038
3039 static void
rack_log_type_bbrsnd(struct tcp_rack * rack,uint32_t len,uint32_t slot,uint32_t cts,struct timeval * tv,int line)3040 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3041 {
3042 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3043 union tcp_log_stackspecific log;
3044
3045 memset(&log, 0, sizeof(log));
3046 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3047 log.u_bbr.flex1 = slot;
3048 if (rack->rack_no_prr)
3049 log.u_bbr.flex2 = 0;
3050 else
3051 log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3052 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3053 log.u_bbr.flex6 = line;
3054 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3055 log.u_bbr.flex8 = rack->rc_in_persist;
3056 log.u_bbr.timeStamp = cts;
3057 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3058 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3059 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3060 log.u_bbr.pacing_gain = rack->r_must_retran;
3061 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3062 &rack->rc_inp->inp_socket->so_rcv,
3063 &rack->rc_inp->inp_socket->so_snd,
3064 BBR_LOG_BBRSND, 0,
3065 0, &log, false, tv);
3066 }
3067 }
3068
3069 static void
rack_log_doseg_done(struct tcp_rack * rack,uint32_t cts,int32_t nxt_pkt,int32_t did_out,int way_out,int nsegs)3070 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3071 {
3072 if (tcp_bblogging_on(rack->rc_tp)) {
3073 union tcp_log_stackspecific log;
3074 struct timeval tv;
3075
3076 memset(&log, 0, sizeof(log));
3077 log.u_bbr.flex1 = did_out;
3078 log.u_bbr.flex2 = nxt_pkt;
3079 log.u_bbr.flex3 = way_out;
3080 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3081 if (rack->rack_no_prr)
3082 log.u_bbr.flex5 = 0;
3083 else
3084 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3085 log.u_bbr.flex6 = nsegs;
3086 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3087 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data; /* Do we have ack-can-send set */
3088 log.u_bbr.flex7 <<= 1;
3089 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
3090 log.u_bbr.flex7 <<= 1;
3091 log.u_bbr.flex7 |= rack->r_wanted_output; /* Do we want output */
3092 log.u_bbr.flex8 = rack->rc_in_persist;
3093 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3094 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3095 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3096 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3097 log.u_bbr.use_lt_bw <<= 1;
3098 log.u_bbr.use_lt_bw |= rack->r_might_revert;
3099 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3100 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3101 log.u_bbr.pacing_gain = rack->r_must_retran;
3102 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3103 log.u_bbr.bw_inuse <<= 32;
3104 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3105 log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3106 log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3107 log.u_bbr.lost = rack->rc_tp->t_srtt;
3108 log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3109 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3110 &rack->rc_inp->inp_socket->so_rcv,
3111 &rack->rc_inp->inp_socket->so_snd,
3112 BBR_LOG_DOSEG_DONE, 0,
3113 0, &log, false, &tv);
3114 }
3115 }
3116
3117 static void
rack_log_type_pacing_sizes(struct tcpcb * tp,struct tcp_rack * rack,uint32_t arg1,uint32_t arg2,uint32_t arg3,uint8_t frm)3118 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3119 {
3120 if (tcp_bblogging_on(rack->rc_tp)) {
3121 union tcp_log_stackspecific log;
3122 struct timeval tv;
3123
3124 memset(&log, 0, sizeof(log));
3125 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3126 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3127 log.u_bbr.flex4 = arg1;
3128 log.u_bbr.flex5 = arg2;
3129 log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3130 log.u_bbr.flex6 = arg3;
3131 log.u_bbr.flex8 = frm;
3132 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3133 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3134 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3135 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3136 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3137 log.u_bbr.pacing_gain = rack->r_must_retran;
3138 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3139 &tptosocket(tp)->so_snd,
3140 TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3141 }
3142 }
3143
3144 static void
rack_log_type_just_return(struct tcp_rack * rack,uint32_t cts,uint32_t tlen,uint32_t slot,uint8_t hpts_calling,int reason,uint32_t cwnd_to_use)3145 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3146 uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3147 {
3148 if (tcp_bblogging_on(rack->rc_tp)) {
3149 union tcp_log_stackspecific log;
3150 struct timeval tv;
3151
3152 memset(&log, 0, sizeof(log));
3153 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3154 log.u_bbr.flex1 = slot;
3155 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3156 log.u_bbr.flex4 = reason;
3157 if (rack->rack_no_prr)
3158 log.u_bbr.flex5 = 0;
3159 else
3160 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3161 log.u_bbr.flex7 = hpts_calling;
3162 log.u_bbr.flex8 = rack->rc_in_persist;
3163 log.u_bbr.lt_epoch = cwnd_to_use;
3164 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3165 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3166 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3167 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3168 log.u_bbr.pacing_gain = rack->r_must_retran;
3169 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3170 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3171 log.u_bbr.bw_inuse <<= 32;
3172 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3173 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3174 &rack->rc_inp->inp_socket->so_rcv,
3175 &rack->rc_inp->inp_socket->so_snd,
3176 BBR_LOG_JUSTRET, 0,
3177 tlen, &log, false, &tv);
3178 }
3179 }
3180
3181 static void
rack_log_to_cancel(struct tcp_rack * rack,int32_t hpts_removed,int line,uint32_t us_cts,struct timeval * tv,uint32_t flags_on_entry)3182 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3183 struct timeval *tv, uint32_t flags_on_entry)
3184 {
3185 if (tcp_bblogging_on(rack->rc_tp)) {
3186 union tcp_log_stackspecific log;
3187
3188 memset(&log, 0, sizeof(log));
3189 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3190 log.u_bbr.flex1 = line;
3191 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3192 log.u_bbr.flex3 = flags_on_entry;
3193 log.u_bbr.flex4 = us_cts;
3194 if (rack->rack_no_prr)
3195 log.u_bbr.flex5 = 0;
3196 else
3197 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3198 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3199 log.u_bbr.flex7 = hpts_removed;
3200 log.u_bbr.flex8 = 1;
3201 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3202 log.u_bbr.timeStamp = us_cts;
3203 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3204 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3205 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3206 log.u_bbr.pacing_gain = rack->r_must_retran;
3207 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3208 log.u_bbr.bw_inuse <<= 32;
3209 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3210 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3211 &rack->rc_inp->inp_socket->so_rcv,
3212 &rack->rc_inp->inp_socket->so_snd,
3213 BBR_LOG_TIMERCANC, 0,
3214 0, &log, false, tv);
3215 }
3216 }
3217
3218 static void
rack_log_alt_to_to_cancel(struct tcp_rack * rack,uint32_t flex1,uint32_t flex2,uint32_t flex3,uint32_t flex4,uint32_t flex5,uint32_t flex6,uint16_t flex7,uint8_t mod)3219 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3220 uint32_t flex1, uint32_t flex2,
3221 uint32_t flex3, uint32_t flex4,
3222 uint32_t flex5, uint32_t flex6,
3223 uint16_t flex7, uint8_t mod)
3224 {
3225 if (tcp_bblogging_on(rack->rc_tp)) {
3226 union tcp_log_stackspecific log;
3227 struct timeval tv;
3228
3229 if (mod == 1) {
3230 /* No you can't use 1, its for the real to cancel */
3231 return;
3232 }
3233 memset(&log, 0, sizeof(log));
3234 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3235 log.u_bbr.flex1 = flex1;
3236 log.u_bbr.flex2 = flex2;
3237 log.u_bbr.flex3 = flex3;
3238 log.u_bbr.flex4 = flex4;
3239 log.u_bbr.flex5 = flex5;
3240 log.u_bbr.flex6 = flex6;
3241 log.u_bbr.flex7 = flex7;
3242 log.u_bbr.flex8 = mod;
3243 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3244 &rack->rc_inp->inp_socket->so_rcv,
3245 &rack->rc_inp->inp_socket->so_snd,
3246 BBR_LOG_TIMERCANC, 0,
3247 0, &log, false, &tv);
3248 }
3249 }
3250
3251 static void
rack_log_to_processing(struct tcp_rack * rack,uint32_t cts,int32_t ret,int32_t timers)3252 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3253 {
3254 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3255 union tcp_log_stackspecific log;
3256 struct timeval tv;
3257
3258 memset(&log, 0, sizeof(log));
3259 log.u_bbr.flex1 = timers;
3260 log.u_bbr.flex2 = ret;
3261 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3262 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3263 log.u_bbr.flex5 = cts;
3264 if (rack->rack_no_prr)
3265 log.u_bbr.flex6 = 0;
3266 else
3267 log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3268 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3269 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3270 log.u_bbr.pacing_gain = rack->r_must_retran;
3271 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3272 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3273 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3274 &rack->rc_inp->inp_socket->so_rcv,
3275 &rack->rc_inp->inp_socket->so_snd,
3276 BBR_LOG_TO_PROCESS, 0,
3277 0, &log, false, &tv);
3278 }
3279 }
3280
3281 static void
rack_log_to_prr(struct tcp_rack * rack,int frm,int orig_cwnd,int line)3282 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3283 {
3284 if (tcp_bblogging_on(rack->rc_tp)) {
3285 union tcp_log_stackspecific log;
3286 struct timeval tv;
3287
3288 memset(&log, 0, sizeof(log));
3289 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3290 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3291 if (rack->rack_no_prr)
3292 log.u_bbr.flex3 = 0;
3293 else
3294 log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3295 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3296 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3297 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3298 log.u_bbr.flex7 = line;
3299 log.u_bbr.flex8 = frm;
3300 log.u_bbr.pkts_out = orig_cwnd;
3301 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3302 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3303 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3304 log.u_bbr.use_lt_bw <<= 1;
3305 log.u_bbr.use_lt_bw |= rack->r_might_revert;
3306 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3307 &rack->rc_inp->inp_socket->so_rcv,
3308 &rack->rc_inp->inp_socket->so_snd,
3309 BBR_LOG_BBRUPD, 0,
3310 0, &log, false, &tv);
3311 }
3312 }
3313
3314 static void
rack_counter_destroy(void)3315 rack_counter_destroy(void)
3316 {
3317 counter_u64_free(rack_total_bytes);
3318 counter_u64_free(rack_fto_send);
3319 counter_u64_free(rack_fto_rsm_send);
3320 counter_u64_free(rack_nfto_resend);
3321 counter_u64_free(rack_hw_pace_init_fail);
3322 counter_u64_free(rack_hw_pace_lost);
3323 counter_u64_free(rack_non_fto_send);
3324 counter_u64_free(rack_extended_rfo);
3325 counter_u64_free(rack_ack_total);
3326 counter_u64_free(rack_express_sack);
3327 counter_u64_free(rack_sack_total);
3328 counter_u64_free(rack_move_none);
3329 counter_u64_free(rack_move_some);
3330 counter_u64_free(rack_sack_attacks_detected);
3331 counter_u64_free(rack_sack_attacks_reversed);
3332 counter_u64_free(rack_sack_attacks_suspect);
3333 counter_u64_free(rack_sack_used_next_merge);
3334 counter_u64_free(rack_sack_used_prev_merge);
3335 counter_u64_free(rack_tlp_tot);
3336 counter_u64_free(rack_tlp_newdata);
3337 counter_u64_free(rack_tlp_retran);
3338 counter_u64_free(rack_tlp_retran_bytes);
3339 counter_u64_free(rack_to_tot);
3340 counter_u64_free(rack_saw_enobuf);
3341 counter_u64_free(rack_saw_enobuf_hw);
3342 counter_u64_free(rack_saw_enetunreach);
3343 counter_u64_free(rack_hot_alloc);
3344 counter_u64_free(rack_to_alloc);
3345 counter_u64_free(rack_to_alloc_hard);
3346 counter_u64_free(rack_to_alloc_emerg);
3347 counter_u64_free(rack_to_alloc_limited);
3348 counter_u64_free(rack_alloc_limited_conns);
3349 counter_u64_free(rack_split_limited);
3350 counter_u64_free(rack_multi_single_eq);
3351 counter_u64_free(rack_rxt_clamps_cwnd);
3352 counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3353 counter_u64_free(rack_proc_non_comp_ack);
3354 counter_u64_free(rack_sack_proc_all);
3355 counter_u64_free(rack_sack_proc_restart);
3356 counter_u64_free(rack_sack_proc_short);
3357 counter_u64_free(rack_sack_skipped_acked);
3358 counter_u64_free(rack_sack_splits);
3359 counter_u64_free(rack_input_idle_reduces);
3360 counter_u64_free(rack_collapsed_win);
3361 counter_u64_free(rack_collapsed_win_rxt);
3362 counter_u64_free(rack_collapsed_win_rxt_bytes);
3363 counter_u64_free(rack_collapsed_win_seen);
3364 counter_u64_free(rack_try_scwnd);
3365 counter_u64_free(rack_persists_sends);
3366 counter_u64_free(rack_persists_acks);
3367 counter_u64_free(rack_persists_loss);
3368 counter_u64_free(rack_persists_lost_ends);
3369 #ifdef INVARIANTS
3370 counter_u64_free(rack_adjust_map_bw);
3371 #endif
3372 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3373 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3374 }
3375
3376 static struct rack_sendmap *
rack_alloc(struct tcp_rack * rack)3377 rack_alloc(struct tcp_rack *rack)
3378 {
3379 struct rack_sendmap *rsm;
3380
3381 /*
3382 * First get the top of the list it in
3383 * theory is the "hottest" rsm we have,
3384 * possibly just freed by ack processing.
3385 */
3386 if (rack->rc_free_cnt > rack_free_cache) {
3387 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3388 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3389 counter_u64_add(rack_hot_alloc, 1);
3390 rack->rc_free_cnt--;
3391 return (rsm);
3392 }
3393 /*
3394 * Once we get under our free cache we probably
3395 * no longer have a "hot" one available. Lets
3396 * get one from UMA.
3397 */
3398 rsm = uma_zalloc(rack_zone, M_NOWAIT);
3399 if (rsm) {
3400 rack->r_ctl.rc_num_maps_alloced++;
3401 counter_u64_add(rack_to_alloc, 1);
3402 return (rsm);
3403 }
3404 /*
3405 * Dig in to our aux rsm's (the last two) since
3406 * UMA failed to get us one.
3407 */
3408 if (rack->rc_free_cnt) {
3409 counter_u64_add(rack_to_alloc_emerg, 1);
3410 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3411 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3412 rack->rc_free_cnt--;
3413 return (rsm);
3414 }
3415 return (NULL);
3416 }
3417
3418 static struct rack_sendmap *
rack_alloc_full_limit(struct tcp_rack * rack)3419 rack_alloc_full_limit(struct tcp_rack *rack)
3420 {
3421 if ((V_tcp_map_entries_limit > 0) &&
3422 (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3423 counter_u64_add(rack_to_alloc_limited, 1);
3424 if (!rack->alloc_limit_reported) {
3425 rack->alloc_limit_reported = 1;
3426 counter_u64_add(rack_alloc_limited_conns, 1);
3427 }
3428 return (NULL);
3429 }
3430 return (rack_alloc(rack));
3431 }
3432
3433 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3434 static struct rack_sendmap *
rack_alloc_limit(struct tcp_rack * rack,uint8_t limit_type)3435 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3436 {
3437 struct rack_sendmap *rsm;
3438
3439 if (limit_type) {
3440 /* currently there is only one limit type */
3441 if (rack->r_ctl.rc_split_limit > 0 &&
3442 rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3443 counter_u64_add(rack_split_limited, 1);
3444 if (!rack->alloc_limit_reported) {
3445 rack->alloc_limit_reported = 1;
3446 counter_u64_add(rack_alloc_limited_conns, 1);
3447 }
3448 return (NULL);
3449 }
3450 }
3451
3452 /* allocate and mark in the limit type, if set */
3453 rsm = rack_alloc(rack);
3454 if (rsm != NULL && limit_type) {
3455 rsm->r_limit_type = limit_type;
3456 rack->r_ctl.rc_num_split_allocs++;
3457 }
3458 return (rsm);
3459 }
3460
3461 static void
rack_free_trim(struct tcp_rack * rack)3462 rack_free_trim(struct tcp_rack *rack)
3463 {
3464 struct rack_sendmap *rsm;
3465
3466 /*
3467 * Free up all the tail entries until
3468 * we get our list down to the limit.
3469 */
3470 while (rack->rc_free_cnt > rack_free_cache) {
3471 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3472 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3473 rack->rc_free_cnt--;
3474 rack->r_ctl.rc_num_maps_alloced--;
3475 uma_zfree(rack_zone, rsm);
3476 }
3477 }
3478
3479 static void
rack_free(struct tcp_rack * rack,struct rack_sendmap * rsm)3480 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3481 {
3482 if (rsm->r_flags & RACK_APP_LIMITED) {
3483 if (rack->r_ctl.rc_app_limited_cnt > 0) {
3484 rack->r_ctl.rc_app_limited_cnt--;
3485 }
3486 }
3487 if (rsm->r_limit_type) {
3488 /* currently there is only one limit type */
3489 rack->r_ctl.rc_num_split_allocs--;
3490 }
3491 if (rsm == rack->r_ctl.rc_first_appl) {
3492 rack->r_ctl.cleared_app_ack_seq = rsm->r_start + (rsm->r_end - rsm->r_start);
3493 rack->r_ctl.cleared_app_ack = 1;
3494 if (rack->r_ctl.rc_app_limited_cnt == 0)
3495 rack->r_ctl.rc_first_appl = NULL;
3496 else
3497 rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3498 }
3499 if (rsm == rack->r_ctl.rc_resend)
3500 rack->r_ctl.rc_resend = NULL;
3501 if (rsm == rack->r_ctl.rc_end_appl)
3502 rack->r_ctl.rc_end_appl = NULL;
3503 if (rack->r_ctl.rc_tlpsend == rsm)
3504 rack->r_ctl.rc_tlpsend = NULL;
3505 if (rack->r_ctl.rc_sacklast == rsm)
3506 rack->r_ctl.rc_sacklast = NULL;
3507 memset(rsm, 0, sizeof(struct rack_sendmap));
3508 /* Make sure we are not going to overrun our count limit of 0xff */
3509 if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3510 rack_free_trim(rack);
3511 }
3512 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3513 rack->rc_free_cnt++;
3514 }
3515
3516 static uint32_t
rack_get_measure_window(struct tcpcb * tp,struct tcp_rack * rack)3517 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3518 {
3519 uint64_t srtt, bw, len, tim;
3520 uint32_t segsiz, def_len, minl;
3521
3522 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3523 def_len = rack_def_data_window * segsiz;
3524 if (rack->rc_gp_filled == 0) {
3525 /*
3526 * We have no measurement (IW is in flight?) so
3527 * we can only guess using our data_window sysctl
3528 * value (usually 20MSS).
3529 */
3530 return (def_len);
3531 }
3532 /*
3533 * Now we have a number of factors to consider.
3534 *
3535 * 1) We have a desired BDP which is usually
3536 * at least 2.
3537 * 2) We have a minimum number of rtt's usually 1 SRTT
3538 * but we allow it too to be more.
3539 * 3) We want to make sure a measurement last N useconds (if
3540 * we have set rack_min_measure_usec.
3541 *
3542 * We handle the first concern here by trying to create a data
3543 * window of max(rack_def_data_window, DesiredBDP). The
3544 * second concern we handle in not letting the measurement
3545 * window end normally until at least the required SRTT's
3546 * have gone by which is done further below in
3547 * rack_enough_for_measurement(). Finally the third concern
3548 * we also handle here by calculating how long that time
3549 * would take at the current BW and then return the
3550 * max of our first calculation and that length. Note
3551 * that if rack_min_measure_usec is 0, we don't deal
3552 * with concern 3. Also for both Concern 1 and 3 an
3553 * application limited period could end the measurement
3554 * earlier.
3555 *
3556 * So lets calculate the BDP with the "known" b/w using
3557 * the SRTT has our rtt and then multiply it by the
3558 * goal.
3559 */
3560 bw = rack_get_bw(rack);
3561 srtt = (uint64_t)tp->t_srtt;
3562 len = bw * srtt;
3563 len /= (uint64_t)HPTS_USEC_IN_SEC;
3564 len *= max(1, rack_goal_bdp);
3565 /* Now we need to round up to the nearest MSS */
3566 len = roundup(len, segsiz);
3567 if (rack_min_measure_usec) {
3568 /* Now calculate our min length for this b/w */
3569 tim = rack_min_measure_usec;
3570 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3571 if (minl == 0)
3572 minl = 1;
3573 minl = roundup(minl, segsiz);
3574 if (len < minl)
3575 len = minl;
3576 }
3577 /*
3578 * Now if we have a very small window we want
3579 * to attempt to get the window that is
3580 * as small as possible. This happens on
3581 * low b/w connections and we don't want to
3582 * span huge numbers of rtt's between measurements.
3583 *
3584 * We basically include 2 over our "MIN window" so
3585 * that the measurement can be shortened (possibly) by
3586 * an ack'ed packet.
3587 */
3588 if (len < def_len)
3589 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3590 else
3591 return (max((uint32_t)len, def_len));
3592
3593 }
3594
3595 static int
rack_enough_for_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq th_ack,uint8_t * quality)3596 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3597 {
3598 uint32_t tim, srtts, segsiz;
3599
3600 /*
3601 * Has enough time passed for the GP measurement to be valid?
3602 */
3603 if (SEQ_LT(th_ack, tp->gput_seq)) {
3604 /* Not enough bytes yet */
3605 return (0);
3606 }
3607 if ((tp->snd_max == tp->snd_una) ||
3608 (th_ack == tp->snd_max)){
3609 /*
3610 * All is acked quality of all acked is
3611 * usually low or medium, but we in theory could split
3612 * all acked into two cases, where you got
3613 * a signifigant amount of your window and
3614 * where you did not. For now we leave it
3615 * but it is something to contemplate in the
3616 * future. The danger here is that delayed ack
3617 * is effecting the last byte (which is a 50:50 chance).
3618 */
3619 *quality = RACK_QUALITY_ALLACKED;
3620 return (1);
3621 }
3622 if (SEQ_GEQ(th_ack, tp->gput_ack)) {
3623 /*
3624 * We obtained our entire window of data we wanted
3625 * no matter if we are in recovery or not then
3626 * its ok since expanding the window does not
3627 * make things fuzzy (or at least not as much).
3628 */
3629 *quality = RACK_QUALITY_HIGH;
3630 return (1);
3631 }
3632 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3633 if (SEQ_LT(th_ack, tp->gput_ack) &&
3634 ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3635 /* Not enough bytes yet */
3636 return (0);
3637 }
3638 if (rack->r_ctl.rc_first_appl &&
3639 (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3640 /*
3641 * We are up to the app limited send point
3642 * we have to measure irrespective of the time..
3643 */
3644 *quality = RACK_QUALITY_APPLIMITED;
3645 return (1);
3646 }
3647 /* Now what about time? */
3648 srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3649 tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3650 if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3651 /*
3652 * We do not allow a measurement if we are in recovery
3653 * that would shrink the goodput window we wanted.
3654 * This is to prevent cloudyness of when the last send
3655 * was actually made.
3656 */
3657 *quality = RACK_QUALITY_HIGH;
3658 return (1);
3659 }
3660 /* Nope not even a full SRTT has passed */
3661 return (0);
3662 }
3663
3664 static void
rack_log_timely(struct tcp_rack * rack,uint32_t logged,uint64_t cur_bw,uint64_t low_bnd,uint64_t up_bnd,int line,uint8_t method)3665 rack_log_timely(struct tcp_rack *rack,
3666 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3667 uint64_t up_bnd, int line, uint8_t method)
3668 {
3669 if (tcp_bblogging_on(rack->rc_tp)) {
3670 union tcp_log_stackspecific log;
3671 struct timeval tv;
3672
3673 memset(&log, 0, sizeof(log));
3674 log.u_bbr.flex1 = logged;
3675 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3676 log.u_bbr.flex2 <<= 4;
3677 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3678 log.u_bbr.flex2 <<= 4;
3679 log.u_bbr.flex2 |= rack->rc_gp_incr;
3680 log.u_bbr.flex2 <<= 4;
3681 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3682 log.u_bbr.flex3 = rack->rc_gp_incr;
3683 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3684 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3685 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3686 log.u_bbr.flex7 = rack->rc_gp_bwred;
3687 log.u_bbr.flex8 = method;
3688 log.u_bbr.cur_del_rate = cur_bw;
3689 log.u_bbr.delRate = low_bnd;
3690 log.u_bbr.bw_inuse = up_bnd;
3691 log.u_bbr.rttProp = rack_get_bw(rack);
3692 log.u_bbr.pkt_epoch = line;
3693 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3694 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3695 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3696 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3697 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3698 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3699 log.u_bbr.cwnd_gain <<= 1;
3700 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3701 log.u_bbr.cwnd_gain <<= 1;
3702 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3703 log.u_bbr.cwnd_gain <<= 1;
3704 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3705 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3706 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3707 &rack->rc_inp->inp_socket->so_rcv,
3708 &rack->rc_inp->inp_socket->so_snd,
3709 TCP_TIMELY_WORK, 0,
3710 0, &log, false, &tv);
3711 }
3712 }
3713
3714 static int
rack_bw_can_be_raised(struct tcp_rack * rack,uint64_t cur_bw,uint64_t last_bw_est,uint16_t mult)3715 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3716 {
3717 /*
3718 * Before we increase we need to know if
3719 * the estimate just made was less than
3720 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3721 *
3722 * If we already are pacing at a fast enough
3723 * rate to push us faster there is no sense of
3724 * increasing.
3725 *
3726 * We first caculate our actual pacing rate (ss or ca multiplier
3727 * times our cur_bw).
3728 *
3729 * Then we take the last measured rate and multipy by our
3730 * maximum pacing overage to give us a max allowable rate.
3731 *
3732 * If our act_rate is smaller than our max_allowable rate
3733 * then we should increase. Else we should hold steady.
3734 *
3735 */
3736 uint64_t act_rate, max_allow_rate;
3737
3738 if (rack_timely_no_stopping)
3739 return (1);
3740
3741 if ((cur_bw == 0) || (last_bw_est == 0)) {
3742 /*
3743 * Initial startup case or
3744 * everything is acked case.
3745 */
3746 rack_log_timely(rack, mult, cur_bw, 0, 0,
3747 __LINE__, 9);
3748 return (1);
3749 }
3750 if (mult <= 100) {
3751 /*
3752 * We can always pace at or slightly above our rate.
3753 */
3754 rack_log_timely(rack, mult, cur_bw, 0, 0,
3755 __LINE__, 9);
3756 return (1);
3757 }
3758 act_rate = cur_bw * (uint64_t)mult;
3759 act_rate /= 100;
3760 max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3761 max_allow_rate /= 100;
3762 if (act_rate < max_allow_rate) {
3763 /*
3764 * Here the rate we are actually pacing at
3765 * is smaller than 10% above our last measurement.
3766 * This means we are pacing below what we would
3767 * like to try to achieve (plus some wiggle room).
3768 */
3769 rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate,
3770 __LINE__, 9);
3771 return (1);
3772 } else {
3773 /*
3774 * Here we are already pacing at least rack_max_per_above(10%)
3775 * what we are getting back. This indicates most likely
3776 * that we are being limited (cwnd/rwnd/app) and can't
3777 * get any more b/w. There is no sense of trying to
3778 * raise up the pacing rate its not speeding us up
3779 * and we already are pacing faster than we are getting.
3780 */
3781 rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate,
3782 __LINE__, 8);
3783 return (0);
3784 }
3785 }
3786
3787 static void
rack_validate_multipliers_at_or_above100(struct tcp_rack * rack)3788 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3789 {
3790 /*
3791 * When we drag bottom, we want to assure
3792 * that no multiplier is below 1.0, if so
3793 * we want to restore it to at least that.
3794 */
3795 if (rack->r_ctl.rack_per_of_gp_rec < 100) {
3796 /* This is unlikely we usually do not touch recovery */
3797 rack->r_ctl.rack_per_of_gp_rec = 100;
3798 }
3799 if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3800 rack->r_ctl.rack_per_of_gp_ca = 100;
3801 }
3802 if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3803 rack->r_ctl.rack_per_of_gp_ss = 100;
3804 }
3805 }
3806
3807 static void
rack_validate_multipliers_at_or_below_100(struct tcp_rack * rack)3808 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3809 {
3810 if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3811 rack->r_ctl.rack_per_of_gp_ca = 100;
3812 }
3813 if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3814 rack->r_ctl.rack_per_of_gp_ss = 100;
3815 }
3816 }
3817
3818 static void
rack_increase_bw_mul(struct tcp_rack * rack,int timely_says,uint64_t cur_bw,uint64_t last_bw_est,int override)3819 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3820 {
3821 int32_t calc, logged, plus;
3822
3823 logged = 0;
3824
3825 if (rack->rc_skip_timely)
3826 return;
3827 if (override) {
3828 /*
3829 * override is passed when we are
3830 * loosing b/w and making one last
3831 * gasp at trying to not loose out
3832 * to a new-reno flow.
3833 */
3834 goto extra_boost;
3835 }
3836 /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3837 if (rack->rc_gp_incr &&
3838 ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3839 /*
3840 * Reset and get 5 strokes more before the boost. Note
3841 * that the count is 0 based so we have to add one.
3842 */
3843 extra_boost:
3844 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3845 rack->rc_gp_timely_inc_cnt = 0;
3846 } else
3847 plus = (uint32_t)rack_gp_increase_per;
3848 /* Must be at least 1% increase for true timely increases */
3849 if ((plus < 1) &&
3850 ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3851 plus = 1;
3852 if (rack->rc_gp_saw_rec &&
3853 (rack->rc_gp_no_rec_chg == 0) &&
3854 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3855 rack->r_ctl.rack_per_of_gp_rec)) {
3856 /* We have been in recovery ding it too */
3857 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3858 if (calc > 0xffff)
3859 calc = 0xffff;
3860 logged |= 1;
3861 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3862 if (rack->r_ctl.rack_per_upper_bound_ca &&
3863 (rack->rc_dragged_bottom == 0) &&
3864 (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3865 rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3866 }
3867 if (rack->rc_gp_saw_ca &&
3868 (rack->rc_gp_saw_ss == 0) &&
3869 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3870 rack->r_ctl.rack_per_of_gp_ca)) {
3871 /* In CA */
3872 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3873 if (calc > 0xffff)
3874 calc = 0xffff;
3875 logged |= 2;
3876 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3877 if (rack->r_ctl.rack_per_upper_bound_ca &&
3878 (rack->rc_dragged_bottom == 0) &&
3879 (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3880 rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3881 }
3882 if (rack->rc_gp_saw_ss &&
3883 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3884 rack->r_ctl.rack_per_of_gp_ss)) {
3885 /* In SS */
3886 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3887 if (calc > 0xffff)
3888 calc = 0xffff;
3889 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3890 if (rack->r_ctl.rack_per_upper_bound_ss &&
3891 (rack->rc_dragged_bottom == 0) &&
3892 (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3893 rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3894 logged |= 4;
3895 }
3896 if (logged &&
3897 (rack->rc_gp_incr == 0)){
3898 /* Go into increment mode */
3899 rack->rc_gp_incr = 1;
3900 rack->rc_gp_timely_inc_cnt = 0;
3901 }
3902 if (rack->rc_gp_incr &&
3903 logged &&
3904 (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3905 rack->rc_gp_timely_inc_cnt++;
3906 }
3907 rack_log_timely(rack, logged, plus, 0, 0,
3908 __LINE__, 1);
3909 }
3910
3911 static uint32_t
rack_get_decrease(struct tcp_rack * rack,uint32_t curper,int32_t rtt_diff)3912 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3913 {
3914 /*-
3915 * norm_grad = rtt_diff / minrtt;
3916 * new_per = curper * (1 - B * norm_grad)
3917 *
3918 * B = rack_gp_decrease_per (default 80%)
3919 * rtt_dif = input var current rtt-diff
3920 * curper = input var current percentage
3921 * minrtt = from rack filter
3922 *
3923 * In order to do the floating point calculations above we
3924 * do an integer conversion. The code looks confusing so let me
3925 * translate it into something that use more variables and
3926 * is clearer for us humans :)
3927 *
3928 * uint64_t norm_grad, inverse, reduce_by, final_result;
3929 * uint32_t perf;
3930 *
3931 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3932 * (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3933 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3934 * inverse /= 1000000;
3935 * reduce_by = (1000000 - inverse);
3936 * final_result = (cur_per * reduce_by) / 1000000;
3937 * perf = (uint32_t)final_result;
3938 */
3939 uint64_t perf;
3940
3941 perf = (((uint64_t)curper * ((uint64_t)1000000 -
3942 ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3943 (((uint64_t)rtt_diff * (uint64_t)1000000)/
3944 (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3945 (uint64_t)1000000)) /
3946 (uint64_t)1000000);
3947 if (perf > curper) {
3948 /* TSNH */
3949 perf = curper - 1;
3950 }
3951 return ((uint32_t)perf);
3952 }
3953
3954 static uint32_t
rack_decrease_highrtt(struct tcp_rack * rack,uint32_t curper,uint32_t rtt)3955 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3956 {
3957 /*
3958 * highrttthresh
3959 * result = curper * (1 - (B * ( 1 - ------ ))
3960 * gp_srtt
3961 *
3962 * B = rack_gp_decrease_per (default .8 i.e. 80)
3963 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3964 */
3965 uint64_t perf;
3966 uint32_t highrttthresh;
3967
3968 highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3969
3970 perf = (((uint64_t)curper * ((uint64_t)1000000 -
3971 ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3972 ((uint64_t)highrttthresh * (uint64_t)1000000) /
3973 (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3974 if (tcp_bblogging_on(rack->rc_tp)) {
3975 uint64_t log1;
3976
3977 log1 = rtt;
3978 log1 <<= 32;
3979 log1 |= highrttthresh;
3980 rack_log_timely(rack,
3981 rack_gp_decrease_per,
3982 (uint64_t)curper,
3983 log1,
3984 perf,
3985 __LINE__,
3986 15);
3987 }
3988 return (perf);
3989 }
3990
3991 static void
rack_decrease_bw_mul(struct tcp_rack * rack,int timely_says,uint32_t rtt,int32_t rtt_diff)3992 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3993 {
3994 uint64_t logvar, logvar2, logvar3;
3995 uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3996
3997 if (rack->rc_skip_timely)
3998 return;
3999 if (rack->rc_gp_incr) {
4000 /* Turn off increment counting */
4001 rack->rc_gp_incr = 0;
4002 rack->rc_gp_timely_inc_cnt = 0;
4003 }
4004 ss_red = ca_red = rec_red = 0;
4005 logged = 0;
4006 /* Calculate the reduction value */
4007 if (rtt_diff < 0) {
4008 rtt_diff *= -1;
4009 }
4010 /* Must be at least 1% reduction */
4011 if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4012 /* We have been in recovery ding it too */
4013 if (timely_says == 2) {
4014 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4015 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4016 if (alt < new_per)
4017 val = alt;
4018 else
4019 val = new_per;
4020 } else
4021 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4022 if (rack->r_ctl.rack_per_of_gp_rec > val) {
4023 rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4024 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4025 } else {
4026 rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4027 rec_red = 0;
4028 }
4029 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4030 rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4031 logged |= 1;
4032 }
4033 if (rack->rc_gp_saw_ss) {
4034 /* Sent in SS */
4035 if (timely_says == 2) {
4036 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4037 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4038 if (alt < new_per)
4039 val = alt;
4040 else
4041 val = new_per;
4042 } else
4043 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4044 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4045 ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4046 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4047 } else {
4048 ss_red = new_per;
4049 rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4050 logvar = new_per;
4051 logvar <<= 32;
4052 logvar |= alt;
4053 logvar2 = (uint32_t)rtt;
4054 logvar2 <<= 32;
4055 logvar2 |= (uint32_t)rtt_diff;
4056 logvar3 = rack_gp_rtt_maxmul;
4057 logvar3 <<= 32;
4058 logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4059 rack_log_timely(rack, timely_says,
4060 logvar2, logvar3,
4061 logvar, __LINE__, 10);
4062 }
4063 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4064 rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4065 logged |= 4;
4066 } else if (rack->rc_gp_saw_ca) {
4067 /* Sent in CA */
4068 if (timely_says == 2) {
4069 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4070 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4071 if (alt < new_per)
4072 val = alt;
4073 else
4074 val = new_per;
4075 } else
4076 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4077 if (rack->r_ctl.rack_per_of_gp_ca > val) {
4078 ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4079 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4080 } else {
4081 rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4082 ca_red = 0;
4083 logvar = new_per;
4084 logvar <<= 32;
4085 logvar |= alt;
4086 logvar2 = (uint32_t)rtt;
4087 logvar2 <<= 32;
4088 logvar2 |= (uint32_t)rtt_diff;
4089 logvar3 = rack_gp_rtt_maxmul;
4090 logvar3 <<= 32;
4091 logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4092 rack_log_timely(rack, timely_says,
4093 logvar2, logvar3,
4094 logvar, __LINE__, 10);
4095 }
4096 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4097 rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4098 logged |= 2;
4099 }
4100 if (rack->rc_gp_timely_dec_cnt < 0x7) {
4101 rack->rc_gp_timely_dec_cnt++;
4102 if (rack_timely_dec_clear &&
4103 (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4104 rack->rc_gp_timely_dec_cnt = 0;
4105 }
4106 logvar = ss_red;
4107 logvar <<= 32;
4108 logvar |= ca_red;
4109 rack_log_timely(rack, logged, rec_red, rack_per_lower_bound, logvar,
4110 __LINE__, 2);
4111 }
4112
4113 static void
rack_log_rtt_shrinks(struct tcp_rack * rack,uint32_t us_cts,uint32_t rtt,uint32_t line,uint8_t reas)4114 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4115 uint32_t rtt, uint32_t line, uint8_t reas)
4116 {
4117 if (tcp_bblogging_on(rack->rc_tp)) {
4118 union tcp_log_stackspecific log;
4119 struct timeval tv;
4120
4121 memset(&log, 0, sizeof(log));
4122 log.u_bbr.flex1 = line;
4123 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4124 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4125 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4126 log.u_bbr.flex5 = rtt;
4127 log.u_bbr.flex6 = rack->rc_highly_buffered;
4128 log.u_bbr.flex6 <<= 1;
4129 log.u_bbr.flex6 |= rack->forced_ack;
4130 log.u_bbr.flex6 <<= 1;
4131 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4132 log.u_bbr.flex6 <<= 1;
4133 log.u_bbr.flex6 |= rack->in_probe_rtt;
4134 log.u_bbr.flex6 <<= 1;
4135 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4136 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4137 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4138 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4139 log.u_bbr.flex8 = reas;
4140 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4141 log.u_bbr.delRate = rack_get_bw(rack);
4142 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4143 log.u_bbr.cur_del_rate <<= 32;
4144 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4145 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4146 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4147 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4148 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4149 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4150 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4151 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4152 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4153 log.u_bbr.rttProp = us_cts;
4154 log.u_bbr.rttProp <<= 32;
4155 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4156 TCP_LOG_EVENTP(rack->rc_tp, NULL,
4157 &rack->rc_inp->inp_socket->so_rcv,
4158 &rack->rc_inp->inp_socket->so_snd,
4159 BBR_LOG_RTT_SHRINKS, 0,
4160 0, &log, false, &rack->r_ctl.act_rcv_time);
4161 }
4162 }
4163
4164 static void
rack_set_prtt_target(struct tcp_rack * rack,uint32_t segsiz,uint32_t rtt)4165 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4166 {
4167 uint64_t bwdp;
4168
4169 bwdp = rack_get_bw(rack);
4170 bwdp *= (uint64_t)rtt;
4171 bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4172 rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4173 if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4174 /*
4175 * A window protocol must be able to have 4 packets
4176 * outstanding as the floor in order to function
4177 * (especially considering delayed ack :D).
4178 */
4179 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4180 }
4181 }
4182
4183 static void
rack_enter_probertt(struct tcp_rack * rack,uint32_t us_cts)4184 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4185 {
4186 /**
4187 * ProbeRTT is a bit different in rack_pacing than in
4188 * BBR. It is like BBR in that it uses the lowering of
4189 * the RTT as a signal that we saw something new and
4190 * counts from there for how long between. But it is
4191 * different in that its quite simple. It does not
4192 * play with the cwnd and wait until we get down
4193 * to N segments outstanding and hold that for
4194 * 200ms. Instead it just sets the pacing reduction
4195 * rate to a set percentage (70 by default) and hold
4196 * that for a number of recent GP Srtt's.
4197 */
4198 uint32_t segsiz;
4199
4200 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4201 if (rack->rc_gp_dyn_mul == 0)
4202 return;
4203
4204 if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4205 /* We are idle */
4206 return;
4207 }
4208 if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4209 SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4210 /*
4211 * Stop the goodput now, the idea here is
4212 * that future measurements with in_probe_rtt
4213 * won't register if they are not greater so
4214 * we want to get what info (if any) is available
4215 * now.
4216 */
4217 rack_do_goodput_measurement(rack->rc_tp, rack,
4218 rack->rc_tp->snd_una, __LINE__,
4219 RACK_QUALITY_PROBERTT);
4220 }
4221 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4222 rack->r_ctl.rc_time_probertt_entered = us_cts;
4223 segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4224 rack->r_ctl.rc_pace_min_segs);
4225 rack->in_probe_rtt = 1;
4226 rack->measure_saw_probe_rtt = 1;
4227 rack->r_ctl.rc_time_probertt_starts = 0;
4228 rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4229 if (rack_probertt_use_min_rtt_entry)
4230 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4231 else
4232 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4233 rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4234 __LINE__, RACK_RTTS_ENTERPROBE);
4235 }
4236
4237 static void
rack_exit_probertt(struct tcp_rack * rack,uint32_t us_cts)4238 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4239 {
4240 struct rack_sendmap *rsm;
4241 uint32_t segsiz;
4242
4243 segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4244 rack->r_ctl.rc_pace_min_segs);
4245 rack->in_probe_rtt = 0;
4246 if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4247 SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4248 /*
4249 * Stop the goodput now, the idea here is
4250 * that future measurements with in_probe_rtt
4251 * won't register if they are not greater so
4252 * we want to get what info (if any) is available
4253 * now.
4254 */
4255 rack_do_goodput_measurement(rack->rc_tp, rack,
4256 rack->rc_tp->snd_una, __LINE__,
4257 RACK_QUALITY_PROBERTT);
4258 } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4259 /*
4260 * We don't have enough data to make a measurement.
4261 * So lets just stop and start here after exiting
4262 * probe-rtt. We probably are not interested in
4263 * the results anyway.
4264 */
4265 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4266 }
4267 /*
4268 * Measurements through the current snd_max are going
4269 * to be limited by the slower pacing rate.
4270 *
4271 * We need to mark these as app-limited so we
4272 * don't collapse the b/w.
4273 */
4274 rsm = tqhash_max(rack->r_ctl.tqh);
4275 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4276 if (rack->r_ctl.rc_app_limited_cnt == 0)
4277 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4278 else {
4279 /*
4280 * Go out to the end app limited and mark
4281 * this new one as next and move the end_appl up
4282 * to this guy.
4283 */
4284 if (rack->r_ctl.rc_end_appl)
4285 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4286 rack->r_ctl.rc_end_appl = rsm;
4287 }
4288 rsm->r_flags |= RACK_APP_LIMITED;
4289 rack->r_ctl.rc_app_limited_cnt++;
4290 }
4291 /*
4292 * Now, we need to examine our pacing rate multipliers.
4293 * If its under 100%, we need to kick it back up to
4294 * 100%. We also don't let it be over our "max" above
4295 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4296 * Note setting clamp_atexit_prtt to 0 has the effect
4297 * of setting CA/SS to 100% always at exit (which is
4298 * the default behavior).
4299 */
4300 if (rack_probertt_clear_is) {
4301 rack->rc_gp_incr = 0;
4302 rack->rc_gp_bwred = 0;
4303 rack->rc_gp_timely_inc_cnt = 0;
4304 rack->rc_gp_timely_dec_cnt = 0;
4305 }
4306 /* Do we do any clamping at exit? */
4307 if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4308 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4309 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4310 }
4311 if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4312 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4313 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4314 }
4315 /*
4316 * Lets set rtt_diff to 0, so that we will get a "boost"
4317 * after exiting.
4318 */
4319 rack->r_ctl.rc_rtt_diff = 0;
4320
4321 /* Clear all flags so we start fresh */
4322 rack->rc_tp->t_bytes_acked = 0;
4323 rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4324 /*
4325 * If configured to, set the cwnd and ssthresh to
4326 * our targets.
4327 */
4328 if (rack_probe_rtt_sets_cwnd) {
4329 uint64_t ebdp;
4330 uint32_t setto;
4331
4332 /* Set ssthresh so we get into CA once we hit our target */
4333 if (rack_probertt_use_min_rtt_exit == 1) {
4334 /* Set to min rtt */
4335 rack_set_prtt_target(rack, segsiz,
4336 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4337 } else if (rack_probertt_use_min_rtt_exit == 2) {
4338 /* Set to current gp rtt */
4339 rack_set_prtt_target(rack, segsiz,
4340 rack->r_ctl.rc_gp_srtt);
4341 } else if (rack_probertt_use_min_rtt_exit == 3) {
4342 /* Set to entry gp rtt */
4343 rack_set_prtt_target(rack, segsiz,
4344 rack->r_ctl.rc_entry_gp_rtt);
4345 } else {
4346 uint64_t sum;
4347 uint32_t setval;
4348
4349 sum = rack->r_ctl.rc_entry_gp_rtt;
4350 sum *= 10;
4351 sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4352 if (sum >= 20) {
4353 /*
4354 * A highly buffered path needs
4355 * cwnd space for timely to work.
4356 * Lets set things up as if
4357 * we are heading back here again.
4358 */
4359 setval = rack->r_ctl.rc_entry_gp_rtt;
4360 } else if (sum >= 15) {
4361 /*
4362 * Lets take the smaller of the
4363 * two since we are just somewhat
4364 * buffered.
4365 */
4366 setval = rack->r_ctl.rc_gp_srtt;
4367 if (setval > rack->r_ctl.rc_entry_gp_rtt)
4368 setval = rack->r_ctl.rc_entry_gp_rtt;
4369 } else {
4370 /*
4371 * Here we are not highly buffered
4372 * and should pick the min we can to
4373 * keep from causing loss.
4374 */
4375 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4376 }
4377 rack_set_prtt_target(rack, segsiz,
4378 setval);
4379 }
4380 if (rack_probe_rtt_sets_cwnd > 1) {
4381 /* There is a percentage here to boost */
4382 ebdp = rack->r_ctl.rc_target_probertt_flight;
4383 ebdp *= rack_probe_rtt_sets_cwnd;
4384 ebdp /= 100;
4385 setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4386 } else
4387 setto = rack->r_ctl.rc_target_probertt_flight;
4388 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4389 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4390 /* Enforce a min */
4391 rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4392 }
4393 /* If we set in the cwnd also set the ssthresh point so we are in CA */
4394 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4395 }
4396 rack_log_rtt_shrinks(rack, us_cts,
4397 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4398 __LINE__, RACK_RTTS_EXITPROBE);
4399 /* Clear times last so log has all the info */
4400 rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4401 rack->r_ctl.rc_time_probertt_entered = us_cts;
4402 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4403 rack->r_ctl.rc_time_of_last_probertt = us_cts;
4404 }
4405
4406 static void
rack_check_probe_rtt(struct tcp_rack * rack,uint32_t us_cts)4407 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4408 {
4409 /* Check in on probe-rtt */
4410
4411 if (rack->rc_gp_filled == 0) {
4412 /* We do not do p-rtt unless we have gp measurements */
4413 return;
4414 }
4415 if (rack->in_probe_rtt) {
4416 uint64_t no_overflow;
4417 uint32_t endtime, must_stay;
4418
4419 if (rack->r_ctl.rc_went_idle_time &&
4420 ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4421 /*
4422 * We went idle during prtt, just exit now.
4423 */
4424 rack_exit_probertt(rack, us_cts);
4425 } else if (rack_probe_rtt_safety_val &&
4426 TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4427 ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4428 /*
4429 * Probe RTT safety value triggered!
4430 */
4431 rack_log_rtt_shrinks(rack, us_cts,
4432 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4433 __LINE__, RACK_RTTS_SAFETY);
4434 rack_exit_probertt(rack, us_cts);
4435 }
4436 /* Calculate the max we will wait */
4437 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4438 if (rack->rc_highly_buffered)
4439 endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4440 /* Calculate the min we must wait */
4441 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4442 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4443 TSTMP_LT(us_cts, endtime)) {
4444 uint32_t calc;
4445 /* Do we lower more? */
4446 no_exit:
4447 if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4448 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4449 else
4450 calc = 0;
4451 calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4452 if (calc) {
4453 /* Maybe */
4454 calc *= rack_per_of_gp_probertt_reduce;
4455 if (calc > rack_per_of_gp_probertt)
4456 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4457 else
4458 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4459 /* Limit it too */
4460 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4461 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4462 }
4463 /* We must reach target or the time set */
4464 return;
4465 }
4466 if (rack->r_ctl.rc_time_probertt_starts == 0) {
4467 if ((TSTMP_LT(us_cts, must_stay) &&
4468 rack->rc_highly_buffered) ||
4469 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4470 rack->r_ctl.rc_target_probertt_flight)) {
4471 /* We are not past the must_stay time */
4472 goto no_exit;
4473 }
4474 rack_log_rtt_shrinks(rack, us_cts,
4475 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4476 __LINE__, RACK_RTTS_REACHTARGET);
4477 rack->r_ctl.rc_time_probertt_starts = us_cts;
4478 if (rack->r_ctl.rc_time_probertt_starts == 0)
4479 rack->r_ctl.rc_time_probertt_starts = 1;
4480 /* Restore back to our rate we want to pace at in prtt */
4481 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4482 }
4483 /*
4484 * Setup our end time, some number of gp_srtts plus 200ms.
4485 */
4486 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4487 (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4488 if (rack_probertt_gpsrtt_cnt_div)
4489 endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4490 else
4491 endtime = 0;
4492 endtime += rack_min_probertt_hold;
4493 endtime += rack->r_ctl.rc_time_probertt_starts;
4494 if (TSTMP_GEQ(us_cts, endtime)) {
4495 /* yes, exit probertt */
4496 rack_exit_probertt(rack, us_cts);
4497 }
4498
4499 } else if ((rack->rc_skip_timely == 0) &&
4500 (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4501 ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4502 /* Go into probertt, its been too long since we went lower */
4503 rack_enter_probertt(rack, us_cts);
4504 }
4505 }
4506
4507 static void
rack_update_multiplier(struct tcp_rack * rack,int32_t timely_says,uint64_t last_bw_est,uint32_t rtt,int32_t rtt_diff)4508 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4509 uint32_t rtt, int32_t rtt_diff)
4510 {
4511 uint64_t cur_bw, up_bnd, low_bnd, subfr;
4512 uint32_t losses;
4513
4514 if ((rack->rc_gp_dyn_mul == 0) ||
4515 (rack->use_fixed_rate) ||
4516 (rack->in_probe_rtt) ||
4517 (rack->rc_always_pace == 0)) {
4518 /* No dynamic GP multiplier in play */
4519 return;
4520 }
4521 losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4522 cur_bw = rack_get_bw(rack);
4523 /* Calculate our up and down range */
4524 up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4525 up_bnd /= 100;
4526 up_bnd += rack->r_ctl.last_gp_comp_bw;
4527
4528 subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4529 subfr /= 100;
4530 low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4531 if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4532 /*
4533 * This is the case where our RTT is above
4534 * the max target and we have been configured
4535 * to just do timely no bonus up stuff in that case.
4536 *
4537 * There are two configurations, set to 1, and we
4538 * just do timely if we are over our max. If its
4539 * set above 1 then we slam the multipliers down
4540 * to 100 and then decrement per timely.
4541 */
4542 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4543 __LINE__, 3);
4544 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4545 rack_validate_multipliers_at_or_below_100(rack);
4546 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4547 } else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4548 /*
4549 * We are decreasing this is a bit complicated this
4550 * means we are loosing ground. This could be
4551 * because another flow entered and we are competing
4552 * for b/w with it. This will push the RTT up which
4553 * makes timely unusable unless we want to get shoved
4554 * into a corner and just be backed off (the age
4555 * old problem with delay based CC).
4556 *
4557 * On the other hand if it was a route change we
4558 * would like to stay somewhat contained and not
4559 * blow out the buffers.
4560 */
4561 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4562 __LINE__, 3);
4563 rack->r_ctl.last_gp_comp_bw = cur_bw;
4564 if (rack->rc_gp_bwred == 0) {
4565 /* Go into reduction counting */
4566 rack->rc_gp_bwred = 1;
4567 rack->rc_gp_timely_dec_cnt = 0;
4568 }
4569 if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4570 /*
4571 * Push another time with a faster pacing
4572 * to try to gain back (we include override to
4573 * get a full raise factor).
4574 */
4575 if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4576 (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4577 (timely_says == 0) ||
4578 (rack_down_raise_thresh == 0)) {
4579 /*
4580 * Do an override up in b/w if we were
4581 * below the threshold or if the threshold
4582 * is zero we always do the raise.
4583 */
4584 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4585 } else {
4586 /* Log it stays the same */
4587 rack_log_timely(rack, 0, last_bw_est, low_bnd, 0,
4588 __LINE__, 11);
4589 }
4590 rack->rc_gp_timely_dec_cnt++;
4591 /* We are not incrementing really no-count */
4592 rack->rc_gp_incr = 0;
4593 rack->rc_gp_timely_inc_cnt = 0;
4594 } else {
4595 /*
4596 * Lets just use the RTT
4597 * information and give up
4598 * pushing.
4599 */
4600 goto use_timely;
4601 }
4602 } else if ((timely_says != 2) &&
4603 !losses &&
4604 (last_bw_est > up_bnd)) {
4605 /*
4606 * We are increasing b/w lets keep going, updating
4607 * our b/w and ignoring any timely input, unless
4608 * of course we are at our max raise (if there is one).
4609 */
4610
4611 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4612 __LINE__, 3);
4613 rack->r_ctl.last_gp_comp_bw = cur_bw;
4614 if (rack->rc_gp_saw_ss &&
4615 rack->r_ctl.rack_per_upper_bound_ss &&
4616 (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4617 /*
4618 * In cases where we can't go higher
4619 * we should just use timely.
4620 */
4621 goto use_timely;
4622 }
4623 if (rack->rc_gp_saw_ca &&
4624 rack->r_ctl.rack_per_upper_bound_ca &&
4625 (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4626 /*
4627 * In cases where we can't go higher
4628 * we should just use timely.
4629 */
4630 goto use_timely;
4631 }
4632 rack->rc_gp_bwred = 0;
4633 rack->rc_gp_timely_dec_cnt = 0;
4634 /* You get a set number of pushes if timely is trying to reduce */
4635 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4636 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4637 } else {
4638 /* Log it stays the same */
4639 rack_log_timely(rack, 0, last_bw_est, up_bnd, 0,
4640 __LINE__, 12);
4641 }
4642 return;
4643 } else {
4644 /*
4645 * We are staying between the lower and upper range bounds
4646 * so use timely to decide.
4647 */
4648 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4649 __LINE__, 3);
4650 use_timely:
4651 if (timely_says) {
4652 rack->rc_gp_incr = 0;
4653 rack->rc_gp_timely_inc_cnt = 0;
4654 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4655 !losses &&
4656 (last_bw_est < low_bnd)) {
4657 /* We are loosing ground */
4658 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4659 rack->rc_gp_timely_dec_cnt++;
4660 /* We are not incrementing really no-count */
4661 rack->rc_gp_incr = 0;
4662 rack->rc_gp_timely_inc_cnt = 0;
4663 } else
4664 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4665 } else {
4666 rack->rc_gp_bwred = 0;
4667 rack->rc_gp_timely_dec_cnt = 0;
4668 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4669 }
4670 }
4671 }
4672
4673 static int32_t
rack_make_timely_judgement(struct tcp_rack * rack,uint32_t rtt,int32_t rtt_diff,uint32_t prev_rtt)4674 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4675 {
4676 int32_t timely_says;
4677 uint64_t log_mult, log_rtt_a_diff;
4678
4679 log_rtt_a_diff = rtt;
4680 log_rtt_a_diff <<= 32;
4681 log_rtt_a_diff |= (uint32_t)rtt_diff;
4682 if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4683 rack_gp_rtt_maxmul)) {
4684 /* Reduce the b/w multiplier */
4685 timely_says = 2;
4686 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4687 log_mult <<= 32;
4688 log_mult |= prev_rtt;
4689 rack_log_timely(rack, timely_says, log_mult,
4690 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4691 log_rtt_a_diff, __LINE__, 4);
4692 } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4693 ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4694 max(rack_gp_rtt_mindiv , 1)))) {
4695 /* Increase the b/w multiplier */
4696 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4697 ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4698 max(rack_gp_rtt_mindiv , 1));
4699 log_mult <<= 32;
4700 log_mult |= prev_rtt;
4701 timely_says = 0;
4702 rack_log_timely(rack, timely_says, log_mult ,
4703 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4704 log_rtt_a_diff, __LINE__, 5);
4705 } else {
4706 /*
4707 * Use a gradient to find it the timely gradient
4708 * is:
4709 * grad = rc_rtt_diff / min_rtt;
4710 *
4711 * anything below or equal to 0 will be
4712 * a increase indication. Anything above
4713 * zero is a decrease. Note we take care
4714 * of the actual gradient calculation
4715 * in the reduction (its not needed for
4716 * increase).
4717 */
4718 log_mult = prev_rtt;
4719 if (rtt_diff <= 0) {
4720 /*
4721 * Rttdiff is less than zero, increase the
4722 * b/w multiplier (its 0 or negative)
4723 */
4724 timely_says = 0;
4725 rack_log_timely(rack, timely_says, log_mult,
4726 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4727 } else {
4728 /* Reduce the b/w multiplier */
4729 timely_says = 1;
4730 rack_log_timely(rack, timely_says, log_mult,
4731 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4732 }
4733 }
4734 return (timely_says);
4735 }
4736
4737 static __inline int
rack_in_gp_window(struct tcpcb * tp,struct rack_sendmap * rsm)4738 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4739 {
4740 if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4741 SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4742 /**
4743 * This covers the case that the
4744 * resent is completely inside
4745 * the gp range or up to it.
4746 * |----------------|
4747 * |-----| <or>
4748 * |----|
4749 * <or> |---|
4750 */
4751 return (1);
4752 } else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4753 SEQ_GT(rsm->r_end, tp->gput_seq)){
4754 /**
4755 * This covers the case of
4756 * |--------------|
4757 * |-------->|
4758 */
4759 return (1);
4760 } else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4761 SEQ_LT(rsm->r_start, tp->gput_ack) &&
4762 SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4763
4764 /**
4765 * This covers the case of
4766 * |--------------|
4767 * |-------->|
4768 */
4769 return (1);
4770 }
4771 return (0);
4772 }
4773
4774 static __inline void
rack_mark_in_gp_win(struct tcpcb * tp,struct rack_sendmap * rsm)4775 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4776 {
4777
4778 if ((tp->t_flags & TF_GPUTINPROG) == 0)
4779 return;
4780 /*
4781 * We have a Goodput measurement in progress. Mark
4782 * the send if its within the window. If its not
4783 * in the window make sure it does not have the mark.
4784 */
4785 if (rack_in_gp_window(tp, rsm))
4786 rsm->r_flags |= RACK_IN_GP_WIN;
4787 else
4788 rsm->r_flags &= ~RACK_IN_GP_WIN;
4789 }
4790
4791 static __inline void
rack_clear_gp_marks(struct tcpcb * tp,struct tcp_rack * rack)4792 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4793 {
4794 /* A GP measurement is ending, clear all marks on the send map*/
4795 struct rack_sendmap *rsm = NULL;
4796
4797 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4798 if (rsm == NULL) {
4799 rsm = tqhash_min(rack->r_ctl.tqh);
4800 }
4801 /* Nothing left? */
4802 while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4803 rsm->r_flags &= ~RACK_IN_GP_WIN;
4804 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4805 }
4806 }
4807
4808
4809 static __inline void
rack_tend_gp_marks(struct tcpcb * tp,struct tcp_rack * rack)4810 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4811 {
4812 struct rack_sendmap *rsm = NULL;
4813
4814 if (tp->snd_una == tp->snd_max) {
4815 /* Nothing outstanding yet, nothing to do here */
4816 return;
4817 }
4818 if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4819 /*
4820 * We are measuring ahead of some outstanding
4821 * data. We need to walk through up until we get
4822 * to gp_seq marking so that no rsm is set incorrectly
4823 * with RACK_IN_GP_WIN.
4824 */
4825 rsm = tqhash_min(rack->r_ctl.tqh);
4826 while (rsm != NULL) {
4827 rack_mark_in_gp_win(tp, rsm);
4828 if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4829 break;
4830 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4831 }
4832 }
4833 if (rsm == NULL) {
4834 /*
4835 * Need to find the GP seq, if rsm is
4836 * set we stopped as we hit it.
4837 */
4838 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4839 if (rsm == NULL)
4840 return;
4841 rack_mark_in_gp_win(tp, rsm);
4842 }
4843 /*
4844 * Now we may need to mark already sent rsm, ahead of
4845 * gput_seq in the window since they may have been sent
4846 * *before* we started our measurment. The rsm, if non-null
4847 * has been marked (note if rsm would have been NULL we would have
4848 * returned in the previous block). So we go to the next, and continue
4849 * until we run out of entries or we exceed the gp_ack value.
4850 */
4851 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4852 while (rsm) {
4853 rack_mark_in_gp_win(tp, rsm);
4854 if (SEQ_GT(rsm->r_end, tp->gput_ack))
4855 break;
4856 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4857 }
4858 }
4859
4860 static void
rack_log_gp_calc(struct tcp_rack * rack,uint32_t add_part,uint32_t sub_part,uint32_t srtt,uint64_t meas_bw,uint64_t utim,uint8_t meth,uint32_t line)4861 rack_log_gp_calc(struct tcp_rack *rack, uint32_t add_part, uint32_t sub_part, uint32_t srtt, uint64_t meas_bw, uint64_t utim, uint8_t meth, uint32_t line)
4862 {
4863 if (tcp_bblogging_on(rack->rc_tp)) {
4864 union tcp_log_stackspecific log;
4865 struct timeval tv;
4866
4867 memset(&log, 0, sizeof(log));
4868 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4869 log.u_bbr.flex1 = add_part;
4870 log.u_bbr.flex2 = sub_part;
4871 log.u_bbr.flex3 = rack_wma_divisor;
4872 log.u_bbr.flex4 = srtt;
4873 log.u_bbr.flex7 = (uint16_t)line;
4874 log.u_bbr.flex8 = meth;
4875 log.u_bbr.delRate = rack->r_ctl.gp_bw;
4876 log.u_bbr.cur_del_rate = meas_bw;
4877 log.u_bbr.rttProp = utim;
4878 TCP_LOG_EVENTP(rack->rc_tp, NULL,
4879 &rack->rc_inp->inp_socket->so_rcv,
4880 &rack->rc_inp->inp_socket->so_snd,
4881 BBR_LOG_THRESH_CALC, 0,
4882 0, &log, false, &rack->r_ctl.act_rcv_time);
4883 }
4884 }
4885
4886 static void
rack_do_goodput_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq th_ack,int line,uint8_t quality)4887 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4888 tcp_seq th_ack, int line, uint8_t quality)
4889 {
4890 uint64_t tim, bytes_ps, stim, utim;
4891 uint32_t segsiz, bytes, reqbytes, us_cts;
4892 int32_t gput, new_rtt_diff, timely_says;
4893 uint64_t resid_bw, subpart = 0, addpart = 0, srtt;
4894 int did_add = 0;
4895
4896 us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4897 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4898 if (TSTMP_GEQ(us_cts, tp->gput_ts))
4899 tim = us_cts - tp->gput_ts;
4900 else
4901 tim = 0;
4902 if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4903 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4904 else
4905 stim = 0;
4906 /*
4907 * Use the larger of the send time or ack time. This prevents us
4908 * from being influenced by ack artifacts to come up with too
4909 * high of measurement. Note that since we are spanning over many more
4910 * bytes in most of our measurements hopefully that is less likely to
4911 * occur.
4912 */
4913 if (tim > stim)
4914 utim = max(tim, 1);
4915 else
4916 utim = max(stim, 1);
4917 reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4918 rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4919 if ((tim == 0) && (stim == 0)) {
4920 /*
4921 * Invalid measurement time, maybe
4922 * all on one ack/one send?
4923 */
4924 bytes = 0;
4925 bytes_ps = 0;
4926 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4927 0, 0, 0, 10, __LINE__, NULL, quality);
4928 goto skip_measurement;
4929 }
4930 if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4931 /* We never made a us_rtt measurement? */
4932 bytes = 0;
4933 bytes_ps = 0;
4934 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4935 0, 0, 0, 10, __LINE__, NULL, quality);
4936 goto skip_measurement;
4937 }
4938 /*
4939 * Calculate the maximum possible b/w this connection
4940 * could have. We base our calculation on the lowest
4941 * rtt we have seen during the measurement and the
4942 * largest rwnd the client has given us in that time. This
4943 * forms a BDP that is the maximum that we could ever
4944 * get to the client. Anything larger is not valid.
4945 *
4946 * I originally had code here that rejected measurements
4947 * where the time was less than 1/2 the latest us_rtt.
4948 * But after thinking on that I realized its wrong since
4949 * say you had a 150Mbps or even 1Gbps link, and you
4950 * were a long way away.. example I am in Europe (100ms rtt)
4951 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4952 * bytes my time would be 1.2ms, and yet my rtt would say
4953 * the measurement was invalid the time was < 50ms. The
4954 * same thing is true for 150Mb (8ms of time).
4955 *
4956 * A better way I realized is to look at what the maximum
4957 * the connection could possibly do. This is gated on
4958 * the lowest RTT we have seen and the highest rwnd.
4959 * We should in theory never exceed that, if we are
4960 * then something on the path is storing up packets
4961 * and then feeding them all at once to our endpoint
4962 * messing up our measurement.
4963 */
4964 rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4965 rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4966 rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4967 if (SEQ_LT(th_ack, tp->gput_seq)) {
4968 /* No measurement can be made */
4969 bytes = 0;
4970 bytes_ps = 0;
4971 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4972 0, 0, 0, 10, __LINE__, NULL, quality);
4973 goto skip_measurement;
4974 } else
4975 bytes = (th_ack - tp->gput_seq);
4976 bytes_ps = (uint64_t)bytes;
4977 /*
4978 * Don't measure a b/w for pacing unless we have gotten at least
4979 * an initial windows worth of data in this measurement interval.
4980 *
4981 * Small numbers of bytes get badly influenced by delayed ack and
4982 * other artifacts. Note we take the initial window or our
4983 * defined minimum GP (defaulting to 10 which hopefully is the
4984 * IW).
4985 */
4986 if (rack->rc_gp_filled == 0) {
4987 /*
4988 * The initial estimate is special. We
4989 * have blasted out an IW worth of packets
4990 * without a real valid ack ts results. We
4991 * then setup the app_limited_needs_set flag,
4992 * this should get the first ack in (probably 2
4993 * MSS worth) to be recorded as the timestamp.
4994 * We thus allow a smaller number of bytes i.e.
4995 * IW - 2MSS.
4996 */
4997 reqbytes -= (2 * segsiz);
4998 /* Also lets fill previous for our first measurement to be neutral */
4999 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5000 }
5001 if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
5002 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5003 rack->r_ctl.rc_app_limited_cnt,
5004 0, 0, 10, __LINE__, NULL, quality);
5005 goto skip_measurement;
5006 }
5007 /*
5008 * We now need to calculate the Timely like status so
5009 * we can update (possibly) the b/w multipliers.
5010 */
5011 new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5012 if (rack->rc_gp_filled == 0) {
5013 /* No previous reading */
5014 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5015 } else {
5016 if (rack->measure_saw_probe_rtt == 0) {
5017 /*
5018 * We don't want a probertt to be counted
5019 * since it will be negative incorrectly. We
5020 * expect to be reducing the RTT when we
5021 * pace at a slower rate.
5022 */
5023 rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5024 rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5025 }
5026 }
5027 timely_says = rack_make_timely_judgement(rack,
5028 rack->r_ctl.rc_gp_srtt,
5029 rack->r_ctl.rc_rtt_diff,
5030 rack->r_ctl.rc_prev_gp_srtt
5031 );
5032 bytes_ps *= HPTS_USEC_IN_SEC;
5033 bytes_ps /= utim;
5034 if (bytes_ps > rack->r_ctl.last_max_bw) {
5035 /*
5036 * Something is on path playing
5037 * since this b/w is not possible based
5038 * on our BDP (highest rwnd and lowest rtt
5039 * we saw in the measurement window).
5040 *
5041 * Another option here would be to
5042 * instead skip the measurement.
5043 */
5044 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5045 bytes_ps, rack->r_ctl.last_max_bw, 0,
5046 11, __LINE__, NULL, quality);
5047 bytes_ps = rack->r_ctl.last_max_bw;
5048 }
5049 /* We store gp for b/w in bytes per second */
5050 if (rack->rc_gp_filled == 0) {
5051 /* Initial measurement */
5052 if (bytes_ps) {
5053 rack->r_ctl.gp_bw = bytes_ps;
5054 rack->rc_gp_filled = 1;
5055 rack->r_ctl.num_measurements = 1;
5056 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5057 } else {
5058 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5059 rack->r_ctl.rc_app_limited_cnt,
5060 0, 0, 10, __LINE__, NULL, quality);
5061 }
5062 if (tcp_in_hpts(rack->rc_tp) &&
5063 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5064 /*
5065 * Ok we can't trust the pacer in this case
5066 * where we transition from un-paced to paced.
5067 * Or for that matter when the burst mitigation
5068 * was making a wild guess and got it wrong.
5069 * Stop the pacer and clear up all the aggregate
5070 * delays etc.
5071 */
5072 tcp_hpts_remove(rack->rc_tp);
5073 rack->r_ctl.rc_hpts_flags = 0;
5074 rack->r_ctl.rc_last_output_to = 0;
5075 }
5076 did_add = 2;
5077 } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5078 /* Still a small number run an average */
5079 rack->r_ctl.gp_bw += bytes_ps;
5080 addpart = rack->r_ctl.num_measurements;
5081 rack->r_ctl.num_measurements++;
5082 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5083 /* We have collected enough to move forward */
5084 rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5085 }
5086 rack_set_pace_segments(tp, rack, __LINE__, NULL);
5087 did_add = 3;
5088 } else {
5089 /*
5090 * We want to take 1/wma of the goodput and add in to 7/8th
5091 * of the old value weighted by the srtt. So if your measurement
5092 * period is say 2 SRTT's long you would get 1/4 as the
5093 * value, if it was like 1/2 SRTT then you would get 1/16th.
5094 *
5095 * But we must be careful not to take too much i.e. if the
5096 * srtt is say 20ms and the measurement is taken over
5097 * 400ms our weight would be 400/20 i.e. 20. On the
5098 * other hand if we get a measurement over 1ms with a
5099 * 10ms rtt we only want to take a much smaller portion.
5100 */
5101 uint8_t meth;
5102
5103 if (rack->r_ctl.num_measurements < 0xff) {
5104 rack->r_ctl.num_measurements++;
5105 }
5106 srtt = (uint64_t)tp->t_srtt;
5107 if (srtt == 0) {
5108 /*
5109 * Strange why did t_srtt go back to zero?
5110 */
5111 if (rack->r_ctl.rc_rack_min_rtt)
5112 srtt = rack->r_ctl.rc_rack_min_rtt;
5113 else
5114 srtt = HPTS_USEC_IN_MSEC;
5115 }
5116 /*
5117 * XXXrrs: Note for reviewers, in playing with
5118 * dynamic pacing I discovered this GP calculation
5119 * as done originally leads to some undesired results.
5120 * Basically you can get longer measurements contributing
5121 * too much to the WMA. Thus I changed it if you are doing
5122 * dynamic adjustments to only do the aportioned adjustment
5123 * if we have a very small (time wise) measurement. Longer
5124 * measurements just get there weight (defaulting to 1/8)
5125 * add to the WMA. We may want to think about changing
5126 * this to always do that for both sides i.e. dynamic
5127 * and non-dynamic... but considering lots of folks
5128 * were playing with this I did not want to change the
5129 * calculation per.se. without your thoughts.. Lawerence?
5130 * Peter??
5131 */
5132 if (rack->rc_gp_dyn_mul == 0) {
5133 subpart = rack->r_ctl.gp_bw * utim;
5134 subpart /= (srtt * 8);
5135 if (subpart < (rack->r_ctl.gp_bw / 2)) {
5136 /*
5137 * The b/w update takes no more
5138 * away then 1/2 our running total
5139 * so factor it in.
5140 */
5141 addpart = bytes_ps * utim;
5142 addpart /= (srtt * 8);
5143 meth = 1;
5144 } else {
5145 /*
5146 * Don't allow a single measurement
5147 * to account for more than 1/2 of the
5148 * WMA. This could happen on a retransmission
5149 * where utim becomes huge compared to
5150 * srtt (multiple retransmissions when using
5151 * the sending rate which factors in all the
5152 * transmissions from the first one).
5153 */
5154 subpart = rack->r_ctl.gp_bw / 2;
5155 addpart = bytes_ps / 2;
5156 meth = 2;
5157 }
5158 rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5159 resid_bw = rack->r_ctl.gp_bw - subpart;
5160 rack->r_ctl.gp_bw = resid_bw + addpart;
5161 did_add = 1;
5162 } else {
5163 if ((utim / srtt) <= 1) {
5164 /*
5165 * The b/w update was over a small period
5166 * of time. The idea here is to prevent a small
5167 * measurement time period from counting
5168 * too much. So we scale it based on the
5169 * time so it attributes less than 1/rack_wma_divisor
5170 * of its measurement.
5171 */
5172 subpart = rack->r_ctl.gp_bw * utim;
5173 subpart /= (srtt * rack_wma_divisor);
5174 addpart = bytes_ps * utim;
5175 addpart /= (srtt * rack_wma_divisor);
5176 meth = 3;
5177 } else {
5178 /*
5179 * The scaled measurement was long
5180 * enough so lets just add in the
5181 * portion of the measurement i.e. 1/rack_wma_divisor
5182 */
5183 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5184 addpart = bytes_ps / rack_wma_divisor;
5185 meth = 4;
5186 }
5187 if ((rack->measure_saw_probe_rtt == 0) ||
5188 (bytes_ps > rack->r_ctl.gp_bw)) {
5189 /*
5190 * For probe-rtt we only add it in
5191 * if its larger, all others we just
5192 * add in.
5193 */
5194 did_add = 1;
5195 rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5196 resid_bw = rack->r_ctl.gp_bw - subpart;
5197 rack->r_ctl.gp_bw = resid_bw + addpart;
5198 }
5199 }
5200 rack_set_pace_segments(tp, rack, __LINE__, NULL);
5201 }
5202 /*
5203 * We only watch the growth of the GP during the initial startup
5204 * or first-slowstart that ensues. If we ever needed to watch
5205 * growth of gp outside of that period all we need to do is
5206 * remove the first clause of this if (rc_initial_ss_comp).
5207 */
5208 if ((rack->rc_initial_ss_comp == 0) &&
5209 (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5210 uint64_t gp_est;
5211
5212 gp_est = bytes_ps;
5213 if (tcp_bblogging_on(rack->rc_tp)) {
5214 union tcp_log_stackspecific log;
5215 struct timeval tv;
5216
5217 memset(&log, 0, sizeof(log));
5218 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5219 log.u_bbr.flex1 = rack->r_ctl.current_round;
5220 log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5221 log.u_bbr.delRate = gp_est;
5222 log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5223 log.u_bbr.flex8 = 41;
5224 (void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5225 0, &log, false, NULL, __func__, __LINE__,&tv);
5226 }
5227 if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5228 (rack->r_ctl.last_gpest == 0)) {
5229 /*
5230 * The round we get our measurement averaging going
5231 * is the base round so it always is the source point
5232 * for when we had our first increment. From there on
5233 * we only record the round that had a rise.
5234 */
5235 rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5236 rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5237 } else if (gp_est >= rack->r_ctl.last_gpest) {
5238 /*
5239 * Test to see if its gone up enough
5240 * to set the round count up to now. Note
5241 * that on the seeding of the 4th measurement we
5242 */
5243 gp_est *= 1000;
5244 gp_est /= rack->r_ctl.last_gpest;
5245 if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5246 /*
5247 * We went up enough to record the round.
5248 */
5249 if (tcp_bblogging_on(rack->rc_tp)) {
5250 union tcp_log_stackspecific log;
5251 struct timeval tv;
5252
5253 memset(&log, 0, sizeof(log));
5254 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5255 log.u_bbr.flex1 = rack->r_ctl.current_round;
5256 log.u_bbr.flex2 = (uint32_t)gp_est;
5257 log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5258 log.u_bbr.delRate = gp_est;
5259 log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5260 log.u_bbr.flex8 = 42;
5261 (void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5262 0, &log, false, NULL, __func__, __LINE__,&tv);
5263 }
5264 rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5265 if (rack->r_ctl.use_gp_not_last == 1)
5266 rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5267 else
5268 rack->r_ctl.last_gpest = bytes_ps;
5269 }
5270 }
5271 }
5272 if ((rack->gp_ready == 0) &&
5273 (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5274 /* We have enough measurements now */
5275 rack->gp_ready = 1;
5276 if (rack->dgp_on ||
5277 rack->rack_hibeta)
5278 rack_set_cc_pacing(rack);
5279 if (rack->defer_options)
5280 rack_apply_deferred_options(rack);
5281 }
5282 rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5283 rack_get_bw(rack), 22, did_add, NULL, quality);
5284 /* We do not update any multipliers if we are in or have seen a probe-rtt */
5285
5286 if ((rack->measure_saw_probe_rtt == 0) &&
5287 rack->rc_gp_rtt_set) {
5288 if (rack->rc_skip_timely == 0) {
5289 rack_update_multiplier(rack, timely_says, bytes_ps,
5290 rack->r_ctl.rc_gp_srtt,
5291 rack->r_ctl.rc_rtt_diff);
5292 }
5293 }
5294 rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5295 rack_get_bw(rack), 3, line, NULL, quality);
5296 rack_log_pacing_delay_calc(rack,
5297 bytes, /* flex2 */
5298 tim, /* flex1 */
5299 bytes_ps, /* bw_inuse */
5300 rack->r_ctl.gp_bw, /* delRate */
5301 rack_get_lt_bw(rack), /* rttProp */
5302 20, line, NULL, 0);
5303 /* reset the gp srtt and setup the new prev */
5304 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5305 /* Record the lost count for the next measurement */
5306 rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5307 skip_measurement:
5308 /*
5309 * We restart our diffs based on the gpsrtt in the
5310 * measurement window.
5311 */
5312 rack->rc_gp_rtt_set = 0;
5313 rack->rc_gp_saw_rec = 0;
5314 rack->rc_gp_saw_ca = 0;
5315 rack->rc_gp_saw_ss = 0;
5316 rack->rc_dragged_bottom = 0;
5317 if (quality == RACK_QUALITY_HIGH) {
5318 /*
5319 * Gput in the stats world is in kbps where bytes_ps is
5320 * bytes per second so we do ((x * 8)/ 1000).
5321 */
5322 gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5323 #ifdef STATS
5324 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5325 gput);
5326 /*
5327 * XXXLAS: This is a temporary hack, and should be
5328 * chained off VOI_TCP_GPUT when stats(9) grows an
5329 * API to deal with chained VOIs.
5330 */
5331 if (tp->t_stats_gput_prev > 0)
5332 stats_voi_update_abs_s32(tp->t_stats,
5333 VOI_TCP_GPUT_ND,
5334 ((gput - tp->t_stats_gput_prev) * 100) /
5335 tp->t_stats_gput_prev);
5336 #endif
5337 tp->t_stats_gput_prev = gput;
5338 }
5339 tp->t_flags &= ~TF_GPUTINPROG;
5340 /*
5341 * Now are we app limited now and there is space from where we
5342 * were to where we want to go?
5343 *
5344 * We don't do the other case i.e. non-applimited here since
5345 * the next send will trigger us picking up the missing data.
5346 */
5347 if (rack->r_ctl.rc_first_appl &&
5348 TCPS_HAVEESTABLISHED(tp->t_state) &&
5349 rack->r_ctl.rc_app_limited_cnt &&
5350 (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5351 ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5352 max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5353 /*
5354 * Yep there is enough outstanding to make a measurement here.
5355 */
5356 struct rack_sendmap *rsm;
5357
5358 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5359 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5360 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5361 rack->app_limited_needs_set = 0;
5362 tp->gput_seq = th_ack;
5363 if (rack->in_probe_rtt)
5364 rack->measure_saw_probe_rtt = 1;
5365 else if ((rack->measure_saw_probe_rtt) &&
5366 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5367 rack->measure_saw_probe_rtt = 0;
5368 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5369 /* There is a full window to gain info from */
5370 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5371 } else {
5372 /* We can only measure up to the applimited point */
5373 tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5374 if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5375 /*
5376 * We don't have enough to make a measurement.
5377 */
5378 tp->t_flags &= ~TF_GPUTINPROG;
5379 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5380 0, 0, 0, 6, __LINE__, NULL, quality);
5381 return;
5382 }
5383 }
5384 if (tp->t_state >= TCPS_FIN_WAIT_1) {
5385 /*
5386 * We will get no more data into the SB
5387 * this means we need to have the data available
5388 * before we start a measurement.
5389 */
5390 if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5391 /* Nope not enough data. */
5392 return;
5393 }
5394 }
5395 tp->t_flags |= TF_GPUTINPROG;
5396 /*
5397 * Now we need to find the timestamp of the send at tp->gput_seq
5398 * for the send based measurement.
5399 */
5400 rack->r_ctl.rc_gp_cumack_ts = 0;
5401 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5402 if (rsm) {
5403 /* Ok send-based limit is set */
5404 if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5405 /*
5406 * Move back to include the earlier part
5407 * so our ack time lines up right (this may
5408 * make an overlapping measurement but thats
5409 * ok).
5410 */
5411 tp->gput_seq = rsm->r_start;
5412 }
5413 if (rsm->r_flags & RACK_ACKED) {
5414 struct rack_sendmap *nrsm;
5415
5416 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5417 tp->gput_seq = rsm->r_end;
5418 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5419 if (nrsm)
5420 rsm = nrsm;
5421 else {
5422 rack->app_limited_needs_set = 1;
5423 }
5424 } else
5425 rack->app_limited_needs_set = 1;
5426 /* We always go from the first send */
5427 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5428 } else {
5429 /*
5430 * If we don't find the rsm due to some
5431 * send-limit set the current time, which
5432 * basically disables the send-limit.
5433 */
5434 struct timeval tv;
5435
5436 microuptime(&tv);
5437 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5438 }
5439 rack_tend_gp_marks(tp, rack);
5440 rack_log_pacing_delay_calc(rack,
5441 tp->gput_seq,
5442 tp->gput_ack,
5443 (uintptr_t)rsm,
5444 tp->gput_ts,
5445 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5446 9,
5447 __LINE__, rsm, quality);
5448 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5449 } else {
5450 /*
5451 * To make sure proper timestamp merging occurs, we need to clear
5452 * all GP marks if we don't start a measurement.
5453 */
5454 rack_clear_gp_marks(tp, rack);
5455 }
5456 }
5457
5458 /*
5459 * CC wrapper hook functions
5460 */
5461 static void
rack_ack_received(struct tcpcb * tp,struct tcp_rack * rack,uint32_t th_ack,uint16_t nsegs,uint16_t type,int32_t post_recovery)5462 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5463 uint16_t type, int32_t post_recovery)
5464 {
5465 uint32_t prior_cwnd, acked;
5466 struct tcp_log_buffer *lgb = NULL;
5467 uint8_t labc_to_use, quality;
5468
5469 INP_WLOCK_ASSERT(tptoinpcb(tp));
5470 tp->t_ccv.nsegs = nsegs;
5471 acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5472 if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5473 uint32_t max;
5474
5475 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5476 if (tp->t_ccv.bytes_this_ack > max) {
5477 tp->t_ccv.bytes_this_ack = max;
5478 }
5479 }
5480 #ifdef STATS
5481 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5482 ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5483 #endif
5484 if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5485 /*
5486 * We will ack all the data, time to end any
5487 * lt_bw_up we have running until something
5488 * new is sent. Note we need to use the actual
5489 * ack_rcv_time which with pacing may be different.
5490 */
5491 uint64_t tmark;
5492
5493 rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5494 rack->r_ctl.lt_seq = tp->snd_max;
5495 tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
5496 if (tmark >= rack->r_ctl.lt_timemark) {
5497 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5498 }
5499 rack->r_ctl.lt_timemark = tmark;
5500 rack->lt_bw_up = 0;
5501 }
5502 quality = RACK_QUALITY_NONE;
5503 if ((tp->t_flags & TF_GPUTINPROG) &&
5504 rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5505 /* Measure the Goodput */
5506 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5507 }
5508 /* Which way our we limited, if not cwnd limited no advance in CA */
5509 if (tp->snd_cwnd <= tp->snd_wnd)
5510 tp->t_ccv.flags |= CCF_CWND_LIMITED;
5511 else
5512 tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5513 if (tp->snd_cwnd > tp->snd_ssthresh) {
5514 tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5515 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5516 /* For the setting of a window past use the actual scwnd we are using */
5517 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5518 tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5519 tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5520 }
5521 } else {
5522 tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5523 tp->t_bytes_acked = 0;
5524 }
5525 prior_cwnd = tp->snd_cwnd;
5526 if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5527 (rack_client_low_buf && rack->client_bufferlvl &&
5528 (rack->client_bufferlvl < rack_client_low_buf)))
5529 labc_to_use = rack->rc_labc;
5530 else
5531 labc_to_use = rack_max_abc_post_recovery;
5532 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5533 union tcp_log_stackspecific log;
5534 struct timeval tv;
5535
5536 memset(&log, 0, sizeof(log));
5537 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5538 log.u_bbr.flex1 = th_ack;
5539 log.u_bbr.flex2 = tp->t_ccv.flags;
5540 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5541 log.u_bbr.flex4 = tp->t_ccv.nsegs;
5542 log.u_bbr.flex5 = labc_to_use;
5543 log.u_bbr.flex6 = prior_cwnd;
5544 log.u_bbr.flex7 = V_tcp_do_newsack;
5545 log.u_bbr.flex8 = 1;
5546 lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5547 0, &log, false, NULL, __func__, __LINE__,&tv);
5548 }
5549 if (CC_ALGO(tp)->ack_received != NULL) {
5550 /* XXXLAS: Find a way to live without this */
5551 tp->t_ccv.curack = th_ack;
5552 tp->t_ccv.labc = labc_to_use;
5553 tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5554 CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5555 }
5556 if (lgb) {
5557 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5558 }
5559 if (rack->r_must_retran) {
5560 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5561 /*
5562 * We now are beyond the rxt point so lets disable
5563 * the flag.
5564 */
5565 rack->r_ctl.rc_out_at_rto = 0;
5566 rack->r_must_retran = 0;
5567 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5568 /*
5569 * Only decrement the rc_out_at_rto if the cwnd advances
5570 * at least a whole segment. Otherwise next time the peer
5571 * acks, we won't be able to send this generaly happens
5572 * when we are in Congestion Avoidance.
5573 */
5574 if (acked <= rack->r_ctl.rc_out_at_rto){
5575 rack->r_ctl.rc_out_at_rto -= acked;
5576 } else {
5577 rack->r_ctl.rc_out_at_rto = 0;
5578 }
5579 }
5580 }
5581 #ifdef STATS
5582 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5583 #endif
5584 if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5585 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5586 }
5587 if ((rack->rc_initial_ss_comp == 0) &&
5588 (tp->snd_cwnd >= tp->snd_ssthresh)) {
5589 /*
5590 * The cwnd has grown beyond ssthresh we have
5591 * entered ca and completed our first Slowstart.
5592 */
5593 rack->rc_initial_ss_comp = 1;
5594 }
5595 }
5596
5597 static void
tcp_rack_partialack(struct tcpcb * tp)5598 tcp_rack_partialack(struct tcpcb *tp)
5599 {
5600 struct tcp_rack *rack;
5601
5602 rack = (struct tcp_rack *)tp->t_fb_ptr;
5603 INP_WLOCK_ASSERT(tptoinpcb(tp));
5604 /*
5605 * If we are doing PRR and have enough
5606 * room to send <or> we are pacing and prr
5607 * is disabled we will want to see if we
5608 * can send data (by setting r_wanted_output to
5609 * true).
5610 */
5611 if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5612 rack->rack_no_prr)
5613 rack->r_wanted_output = 1;
5614 }
5615
5616 static void
rack_exit_recovery(struct tcpcb * tp,struct tcp_rack * rack,int how)5617 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
5618 {
5619 /*
5620 * Now exit recovery.
5621 */
5622 EXIT_RECOVERY(tp->t_flags);
5623 }
5624
5625 static void
rack_post_recovery(struct tcpcb * tp,uint32_t th_ack)5626 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5627 {
5628 struct tcp_rack *rack;
5629 uint32_t orig_cwnd;
5630
5631 orig_cwnd = tp->snd_cwnd;
5632 INP_WLOCK_ASSERT(tptoinpcb(tp));
5633 rack = (struct tcp_rack *)tp->t_fb_ptr;
5634 /* only alert CC if we alerted when we entered */
5635 if (CC_ALGO(tp)->post_recovery != NULL) {
5636 tp->t_ccv.curack = th_ack;
5637 CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5638 if (tp->snd_cwnd < tp->snd_ssthresh) {
5639 /*
5640 * Rack has burst control and pacing
5641 * so lets not set this any lower than
5642 * snd_ssthresh per RFC-6582 (option 2).
5643 */
5644 tp->snd_cwnd = tp->snd_ssthresh;
5645 }
5646 }
5647 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5648 union tcp_log_stackspecific log;
5649 struct timeval tv;
5650
5651 memset(&log, 0, sizeof(log));
5652 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5653 log.u_bbr.flex1 = th_ack;
5654 log.u_bbr.flex2 = tp->t_ccv.flags;
5655 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5656 log.u_bbr.flex4 = tp->t_ccv.nsegs;
5657 log.u_bbr.flex5 = V_tcp_abc_l_var;
5658 log.u_bbr.flex6 = orig_cwnd;
5659 log.u_bbr.flex7 = V_tcp_do_newsack;
5660 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5661 log.u_bbr.flex8 = 2;
5662 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5663 0, &log, false, NULL, __func__, __LINE__, &tv);
5664 }
5665 if ((rack->rack_no_prr == 0) &&
5666 (rack->no_prr_addback == 0) &&
5667 (rack->r_ctl.rc_prr_sndcnt > 0)) {
5668 /*
5669 * Suck the next prr cnt back into cwnd, but
5670 * only do that if we are not application limited.
5671 */
5672 if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5673 /*
5674 * We are allowed to add back to the cwnd the amount we did
5675 * not get out if:
5676 * a) no_prr_addback is off.
5677 * b) we are not app limited
5678 * c) we are doing prr
5679 * <and>
5680 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5681 */
5682 tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5683 rack->r_ctl.rc_prr_sndcnt);
5684 }
5685 rack->r_ctl.rc_prr_sndcnt = 0;
5686 rack_log_to_prr(rack, 1, 0, __LINE__);
5687 }
5688 rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5689 tp->snd_recover = tp->snd_una;
5690 if (rack->r_ctl.dsack_persist) {
5691 rack->r_ctl.dsack_persist--;
5692 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5693 rack->r_ctl.num_dsack = 0;
5694 }
5695 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5696 }
5697 if (rack->rto_from_rec == 1) {
5698 rack->rto_from_rec = 0;
5699 if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
5700 tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
5701 }
5702 rack_exit_recovery(tp, rack, 1);
5703 }
5704
5705 static void
rack_cong_signal(struct tcpcb * tp,uint32_t type,uint32_t ack,int line)5706 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5707 {
5708 struct tcp_rack *rack;
5709 uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5710
5711 INP_WLOCK_ASSERT(tptoinpcb(tp));
5712 #ifdef STATS
5713 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5714 #endif
5715 if (IN_RECOVERY(tp->t_flags) == 0) {
5716 in_rec_at_entry = 0;
5717 ssthresh_enter = tp->snd_ssthresh;
5718 cwnd_enter = tp->snd_cwnd;
5719 } else
5720 in_rec_at_entry = 1;
5721 rack = (struct tcp_rack *)tp->t_fb_ptr;
5722 switch (type) {
5723 case CC_NDUPACK:
5724 tp->t_flags &= ~TF_WASFRECOVERY;
5725 tp->t_flags &= ~TF_WASCRECOVERY;
5726 if (!IN_FASTRECOVERY(tp->t_flags)) {
5727 /* Check if this is the end of the initial Start-up i.e. initial slow-start */
5728 if (rack->rc_initial_ss_comp == 0) {
5729 /* Yep it is the end of the initial slowstart */
5730 rack->rc_initial_ss_comp = 1;
5731 }
5732 rack->r_ctl.rc_prr_delivered = 0;
5733 rack->r_ctl.rc_prr_out = 0;
5734 rack->r_fast_output = 0;
5735 if (rack->rack_no_prr == 0) {
5736 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5737 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5738 }
5739 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5740 tp->snd_recover = tp->snd_max;
5741 if (tp->t_flags2 & TF2_ECN_PERMIT)
5742 tp->t_flags2 |= TF2_ECN_SND_CWR;
5743 }
5744 break;
5745 case CC_ECN:
5746 if (!IN_CONGRECOVERY(tp->t_flags) ||
5747 /*
5748 * Allow ECN reaction on ACK to CWR, if
5749 * that data segment was also CE marked.
5750 */
5751 SEQ_GEQ(ack, tp->snd_recover)) {
5752 EXIT_CONGRECOVERY(tp->t_flags);
5753 KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5754 rack->r_fast_output = 0;
5755 tp->snd_recover = tp->snd_max + 1;
5756 if (tp->t_flags2 & TF2_ECN_PERMIT)
5757 tp->t_flags2 |= TF2_ECN_SND_CWR;
5758 }
5759 break;
5760 case CC_RTO:
5761 tp->t_dupacks = 0;
5762 tp->t_bytes_acked = 0;
5763 rack->r_fast_output = 0;
5764 if (IN_RECOVERY(tp->t_flags))
5765 rack_exit_recovery(tp, rack, 2);
5766 orig_cwnd = tp->snd_cwnd;
5767 rack_log_to_prr(rack, 16, orig_cwnd, line);
5768 if (CC_ALGO(tp)->cong_signal == NULL) {
5769 /* TSNH */
5770 tp->snd_ssthresh = max(2,
5771 min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5772 ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5773 tp->snd_cwnd = ctf_fixed_maxseg(tp);
5774 }
5775 if (tp->t_flags2 & TF2_ECN_PERMIT)
5776 tp->t_flags2 |= TF2_ECN_SND_CWR;
5777 break;
5778 case CC_RTO_ERR:
5779 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5780 /* RTO was unnecessary, so reset everything. */
5781 tp->snd_cwnd = tp->snd_cwnd_prev;
5782 tp->snd_ssthresh = tp->snd_ssthresh_prev;
5783 tp->snd_recover = tp->snd_recover_prev;
5784 if (tp->t_flags & TF_WASFRECOVERY) {
5785 ENTER_FASTRECOVERY(tp->t_flags);
5786 tp->t_flags &= ~TF_WASFRECOVERY;
5787 }
5788 if (tp->t_flags & TF_WASCRECOVERY) {
5789 ENTER_CONGRECOVERY(tp->t_flags);
5790 tp->t_flags &= ~TF_WASCRECOVERY;
5791 }
5792 tp->snd_nxt = tp->snd_max;
5793 tp->t_badrxtwin = 0;
5794 break;
5795 }
5796 if ((CC_ALGO(tp)->cong_signal != NULL) &&
5797 (type != CC_RTO)){
5798 tp->t_ccv.curack = ack;
5799 CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5800 }
5801 if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5802 rack_log_to_prr(rack, 15, cwnd_enter, line);
5803 rack->r_ctl.dsack_byte_cnt = 0;
5804 rack->r_ctl.retran_during_recovery = 0;
5805 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5806 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5807 rack->r_ent_rec_ns = 1;
5808 }
5809 }
5810
5811 static inline void
rack_cc_after_idle(struct tcp_rack * rack,struct tcpcb * tp)5812 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5813 {
5814 uint32_t i_cwnd;
5815
5816 INP_WLOCK_ASSERT(tptoinpcb(tp));
5817
5818 if (CC_ALGO(tp)->after_idle != NULL)
5819 CC_ALGO(tp)->after_idle(&tp->t_ccv);
5820
5821 if (tp->snd_cwnd == 1)
5822 i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */
5823 else
5824 i_cwnd = rc_init_window(rack);
5825
5826 /*
5827 * Being idle is no different than the initial window. If the cc
5828 * clamps it down below the initial window raise it to the initial
5829 * window.
5830 */
5831 if (tp->snd_cwnd < i_cwnd) {
5832 tp->snd_cwnd = i_cwnd;
5833 }
5834 }
5835
5836 /*
5837 * Indicate whether this ack should be delayed. We can delay the ack if
5838 * following conditions are met:
5839 * - There is no delayed ack timer in progress.
5840 * - Our last ack wasn't a 0-sized window. We never want to delay
5841 * the ack that opens up a 0-sized window.
5842 * - LRO wasn't used for this segment. We make sure by checking that the
5843 * segment size is not larger than the MSS.
5844 * - Delayed acks are enabled or this is a half-synchronized T/TCP
5845 * connection.
5846 */
5847 #define DELAY_ACK(tp, tlen) \
5848 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5849 ((tp->t_flags & TF_DELACK) == 0) && \
5850 (tlen <= tp->t_maxseg) && \
5851 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5852
5853 static struct rack_sendmap *
rack_find_lowest_rsm(struct tcp_rack * rack)5854 rack_find_lowest_rsm(struct tcp_rack *rack)
5855 {
5856 struct rack_sendmap *rsm;
5857
5858 /*
5859 * Walk the time-order transmitted list looking for an rsm that is
5860 * not acked. This will be the one that was sent the longest time
5861 * ago that is still outstanding.
5862 */
5863 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5864 if (rsm->r_flags & RACK_ACKED) {
5865 continue;
5866 }
5867 goto finish;
5868 }
5869 finish:
5870 return (rsm);
5871 }
5872
5873 static struct rack_sendmap *
rack_find_high_nonack(struct tcp_rack * rack,struct rack_sendmap * rsm)5874 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5875 {
5876 struct rack_sendmap *prsm;
5877
5878 /*
5879 * Walk the sequence order list backward until we hit and arrive at
5880 * the highest seq not acked. In theory when this is called it
5881 * should be the last segment (which it was not).
5882 */
5883 prsm = rsm;
5884
5885 TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
5886 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5887 continue;
5888 }
5889 return (prsm);
5890 }
5891 return (NULL);
5892 }
5893
5894 static uint32_t
rack_calc_thresh_rack(struct tcp_rack * rack,uint32_t srtt,uint32_t cts,int line,int log_allowed)5895 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
5896 {
5897 int32_t lro;
5898 uint32_t thresh;
5899
5900 /*
5901 * lro is the flag we use to determine if we have seen reordering.
5902 * If it gets set we have seen reordering. The reorder logic either
5903 * works in one of two ways:
5904 *
5905 * If reorder-fade is configured, then we track the last time we saw
5906 * re-ordering occur. If we reach the point where enough time as
5907 * passed we no longer consider reordering has occuring.
5908 *
5909 * Or if reorder-face is 0, then once we see reordering we consider
5910 * the connection to alway be subject to reordering and just set lro
5911 * to 1.
5912 *
5913 * In the end if lro is non-zero we add the extra time for
5914 * reordering in.
5915 */
5916 if (srtt == 0)
5917 srtt = 1;
5918 if (rack->r_ctl.rc_reorder_ts) {
5919 if (rack->r_ctl.rc_reorder_fade) {
5920 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5921 lro = cts - rack->r_ctl.rc_reorder_ts;
5922 if (lro == 0) {
5923 /*
5924 * No time as passed since the last
5925 * reorder, mark it as reordering.
5926 */
5927 lro = 1;
5928 }
5929 } else {
5930 /* Negative time? */
5931 lro = 0;
5932 }
5933 if (lro > rack->r_ctl.rc_reorder_fade) {
5934 /* Turn off reordering seen too */
5935 rack->r_ctl.rc_reorder_ts = 0;
5936 lro = 0;
5937 }
5938 } else {
5939 /* Reodering does not fade */
5940 lro = 1;
5941 }
5942 } else {
5943 lro = 0;
5944 }
5945 if (rack->rc_rack_tmr_std_based == 0) {
5946 thresh = srtt + rack->r_ctl.rc_pkt_delay;
5947 } else {
5948 /* Standards based pkt-delay is 1/4 srtt */
5949 thresh = srtt + (srtt >> 2);
5950 }
5951 if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5952 /* It must be set, if not you get 1/4 rtt */
5953 if (rack->r_ctl.rc_reorder_shift)
5954 thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5955 else
5956 thresh += (srtt >> 2);
5957 }
5958 if (rack->rc_rack_use_dsack &&
5959 lro &&
5960 (rack->r_ctl.num_dsack > 0)) {
5961 /*
5962 * We only increase the reordering window if we
5963 * have seen reordering <and> we have a DSACK count.
5964 */
5965 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5966 if (log_allowed)
5967 rack_log_dsack_event(rack, 4, line, srtt, thresh);
5968 }
5969 /* SRTT * 2 is the ceiling */
5970 if (thresh > (srtt * 2)) {
5971 thresh = srtt * 2;
5972 }
5973 /* And we don't want it above the RTO max either */
5974 if (thresh > rack_rto_max) {
5975 thresh = rack_rto_max;
5976 }
5977 if (log_allowed)
5978 rack_log_dsack_event(rack, 6, line, srtt, thresh);
5979 return (thresh);
5980 }
5981
5982 static uint32_t
rack_calc_thresh_tlp(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t srtt)5983 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5984 struct rack_sendmap *rsm, uint32_t srtt)
5985 {
5986 struct rack_sendmap *prsm;
5987 uint32_t thresh, len;
5988 int segsiz;
5989
5990 if (srtt == 0)
5991 srtt = 1;
5992 if (rack->r_ctl.rc_tlp_threshold)
5993 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5994 else
5995 thresh = (srtt * 2);
5996
5997 /* Get the previous sent packet, if any */
5998 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5999 len = rsm->r_end - rsm->r_start;
6000 if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6001 /* Exactly like the ID */
6002 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6003 uint32_t alt_thresh;
6004 /*
6005 * Compensate for delayed-ack with the d-ack time.
6006 */
6007 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6008 if (alt_thresh > thresh)
6009 thresh = alt_thresh;
6010 }
6011 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6012 /* 2.1 behavior */
6013 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6014 if (prsm && (len <= segsiz)) {
6015 /*
6016 * Two packets outstanding, thresh should be (2*srtt) +
6017 * possible inter-packet delay (if any).
6018 */
6019 uint32_t inter_gap = 0;
6020 int idx, nidx;
6021
6022 idx = rsm->r_rtr_cnt - 1;
6023 nidx = prsm->r_rtr_cnt - 1;
6024 if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6025 /* Yes it was sent later (or at the same time) */
6026 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6027 }
6028 thresh += inter_gap;
6029 } else if (len <= segsiz) {
6030 /*
6031 * Possibly compensate for delayed-ack.
6032 */
6033 uint32_t alt_thresh;
6034
6035 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6036 if (alt_thresh > thresh)
6037 thresh = alt_thresh;
6038 }
6039 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6040 /* 2.2 behavior */
6041 if (len <= segsiz) {
6042 uint32_t alt_thresh;
6043 /*
6044 * Compensate for delayed-ack with the d-ack time.
6045 */
6046 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6047 if (alt_thresh > thresh)
6048 thresh = alt_thresh;
6049 }
6050 }
6051 /* Not above an RTO */
6052 if (thresh > tp->t_rxtcur) {
6053 thresh = tp->t_rxtcur;
6054 }
6055 /* Not above a RTO max */
6056 if (thresh > rack_rto_max) {
6057 thresh = rack_rto_max;
6058 }
6059 /* Apply user supplied min TLP */
6060 if (thresh < rack_tlp_min) {
6061 thresh = rack_tlp_min;
6062 }
6063 return (thresh);
6064 }
6065
6066 static uint32_t
rack_grab_rtt(struct tcpcb * tp,struct tcp_rack * rack)6067 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6068 {
6069 /*
6070 * We want the rack_rtt which is the
6071 * last rtt we measured. However if that
6072 * does not exist we fallback to the srtt (which
6073 * we probably will never do) and then as a last
6074 * resort we use RACK_INITIAL_RTO if no srtt is
6075 * yet set.
6076 */
6077 if (rack->rc_rack_rtt)
6078 return (rack->rc_rack_rtt);
6079 else if (tp->t_srtt == 0)
6080 return (RACK_INITIAL_RTO);
6081 return (tp->t_srtt);
6082 }
6083
6084 static struct rack_sendmap *
rack_check_recovery_mode(struct tcpcb * tp,uint32_t tsused)6085 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6086 {
6087 /*
6088 * Check to see that we don't need to fall into recovery. We will
6089 * need to do so if our oldest transmit is past the time we should
6090 * have had an ack.
6091 */
6092 struct tcp_rack *rack;
6093 struct rack_sendmap *rsm;
6094 int32_t idx;
6095 uint32_t srtt, thresh;
6096
6097 rack = (struct tcp_rack *)tp->t_fb_ptr;
6098 if (tqhash_empty(rack->r_ctl.tqh)) {
6099 return (NULL);
6100 }
6101 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6102 if (rsm == NULL)
6103 return (NULL);
6104
6105
6106 if (rsm->r_flags & RACK_ACKED) {
6107 rsm = rack_find_lowest_rsm(rack);
6108 if (rsm == NULL)
6109 return (NULL);
6110 }
6111 idx = rsm->r_rtr_cnt - 1;
6112 srtt = rack_grab_rtt(tp, rack);
6113 thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6114 if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6115 return (NULL);
6116 }
6117 if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6118 return (NULL);
6119 }
6120 /* Ok if we reach here we are over-due and this guy can be sent */
6121 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6122 return (rsm);
6123 }
6124
6125 static uint32_t
rack_get_persists_timer_val(struct tcpcb * tp,struct tcp_rack * rack)6126 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6127 {
6128 int32_t t;
6129 int32_t tt;
6130 uint32_t ret_val;
6131
6132 t = (tp->t_srtt + (tp->t_rttvar << 2));
6133 RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6134 rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6135 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6136 ret_val = (uint32_t)tt;
6137 return (ret_val);
6138 }
6139
6140 static uint32_t
rack_timer_start(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,int sup_rack)6141 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6142 {
6143 /*
6144 * Start the FR timer, we do this based on getting the first one in
6145 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6146 * events we need to stop the running timer (if its running) before
6147 * starting the new one.
6148 */
6149 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6150 uint32_t srtt_cur;
6151 int32_t idx;
6152 int32_t is_tlp_timer = 0;
6153 struct rack_sendmap *rsm;
6154
6155 if (rack->t_timers_stopped) {
6156 /* All timers have been stopped none are to run */
6157 return (0);
6158 }
6159 if (rack->rc_in_persist) {
6160 /* We can't start any timer in persists */
6161 return (rack_get_persists_timer_val(tp, rack));
6162 }
6163 rack->rc_on_min_to = 0;
6164 if ((tp->t_state < TCPS_ESTABLISHED) ||
6165 ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6166 goto activate_rxt;
6167 }
6168 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6169 if ((rsm == NULL) || sup_rack) {
6170 /* Nothing on the send map or no rack */
6171 activate_rxt:
6172 time_since_sent = 0;
6173 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6174 if (rsm) {
6175 /*
6176 * Should we discount the RTX timer any?
6177 *
6178 * We want to discount it the smallest amount.
6179 * If a timer (Rack/TLP or RXT) has gone off more
6180 * recently thats the discount we want to use (now - timer time).
6181 * If the retransmit of the oldest packet was more recent then
6182 * we want to use that (now - oldest-packet-last_transmit_time).
6183 *
6184 */
6185 idx = rsm->r_rtr_cnt - 1;
6186 if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6187 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6188 else
6189 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6190 if (TSTMP_GT(cts, tstmp_touse))
6191 time_since_sent = cts - tstmp_touse;
6192 }
6193 if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6194 sbavail(&tptosocket(tp)->so_snd)) {
6195 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6196 to = tp->t_rxtcur;
6197 if (to > time_since_sent)
6198 to -= time_since_sent;
6199 else
6200 to = rack->r_ctl.rc_min_to;
6201 if (to == 0)
6202 to = 1;
6203 /* Special case for KEEPINIT */
6204 if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6205 (TP_KEEPINIT(tp) != 0) &&
6206 rsm) {
6207 /*
6208 * We have to put a ceiling on the rxt timer
6209 * of the keep-init timeout.
6210 */
6211 uint32_t max_time, red;
6212
6213 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6214 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6215 red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6216 if (red < max_time)
6217 max_time -= red;
6218 else
6219 max_time = 1;
6220 }
6221 /* Reduce timeout to the keep value if needed */
6222 if (max_time < to)
6223 to = max_time;
6224 }
6225 return (to);
6226 }
6227 return (0);
6228 }
6229 if (rsm->r_flags & RACK_ACKED) {
6230 rsm = rack_find_lowest_rsm(rack);
6231 if (rsm == NULL) {
6232 /* No lowest? */
6233 goto activate_rxt;
6234 }
6235 }
6236 /* Convert from ms to usecs */
6237 if ((rsm->r_flags & RACK_SACK_PASSED) ||
6238 (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6239 (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6240 if ((tp->t_flags & TF_SENTFIN) &&
6241 ((tp->snd_max - tp->snd_una) == 1) &&
6242 (rsm->r_flags & RACK_HAS_FIN)) {
6243 /*
6244 * We don't start a rack timer if all we have is a
6245 * FIN outstanding.
6246 */
6247 goto activate_rxt;
6248 }
6249 if ((rack->use_rack_rr == 0) &&
6250 (IN_FASTRECOVERY(tp->t_flags)) &&
6251 (rack->rack_no_prr == 0) &&
6252 (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) {
6253 /*
6254 * We are not cheating, in recovery and
6255 * not enough ack's to yet get our next
6256 * retransmission out.
6257 *
6258 * Note that classified attackers do not
6259 * get to use the rack-cheat.
6260 */
6261 goto activate_tlp;
6262 }
6263 srtt = rack_grab_rtt(tp, rack);
6264 thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6265 idx = rsm->r_rtr_cnt - 1;
6266 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6267 if (SEQ_GEQ(exp, cts)) {
6268 to = exp - cts;
6269 if (to < rack->r_ctl.rc_min_to) {
6270 to = rack->r_ctl.rc_min_to;
6271 if (rack->r_rr_config == 3)
6272 rack->rc_on_min_to = 1;
6273 }
6274 } else {
6275 to = rack->r_ctl.rc_min_to;
6276 if (rack->r_rr_config == 3)
6277 rack->rc_on_min_to = 1;
6278 }
6279 } else {
6280 /* Ok we need to do a TLP not RACK */
6281 activate_tlp:
6282 if ((rack->rc_tlp_in_progress != 0) &&
6283 (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6284 /*
6285 * The previous send was a TLP and we have sent
6286 * N TLP's without sending new data.
6287 */
6288 goto activate_rxt;
6289 }
6290 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6291 if (rsm == NULL) {
6292 /* We found no rsm to TLP with. */
6293 goto activate_rxt;
6294 }
6295 if (rsm->r_flags & RACK_HAS_FIN) {
6296 /* If its a FIN we dont do TLP */
6297 rsm = NULL;
6298 goto activate_rxt;
6299 }
6300 idx = rsm->r_rtr_cnt - 1;
6301 time_since_sent = 0;
6302 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6303 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6304 else
6305 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6306 if (TSTMP_GT(cts, tstmp_touse))
6307 time_since_sent = cts - tstmp_touse;
6308 is_tlp_timer = 1;
6309 if (tp->t_srtt) {
6310 if ((rack->rc_srtt_measure_made == 0) &&
6311 (tp->t_srtt == 1)) {
6312 /*
6313 * If another stack as run and set srtt to 1,
6314 * then the srtt was 0, so lets use the initial.
6315 */
6316 srtt = RACK_INITIAL_RTO;
6317 } else {
6318 srtt_cur = tp->t_srtt;
6319 srtt = srtt_cur;
6320 }
6321 } else
6322 srtt = RACK_INITIAL_RTO;
6323 /*
6324 * If the SRTT is not keeping up and the
6325 * rack RTT has spiked we want to use
6326 * the last RTT not the smoothed one.
6327 */
6328 if (rack_tlp_use_greater &&
6329 tp->t_srtt &&
6330 (srtt < rack_grab_rtt(tp, rack))) {
6331 srtt = rack_grab_rtt(tp, rack);
6332 }
6333 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6334 if (thresh > time_since_sent) {
6335 to = thresh - time_since_sent;
6336 } else {
6337 to = rack->r_ctl.rc_min_to;
6338 rack_log_alt_to_to_cancel(rack,
6339 thresh, /* flex1 */
6340 time_since_sent, /* flex2 */
6341 tstmp_touse, /* flex3 */
6342 rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6343 (uint32_t)rsm->r_tim_lastsent[idx],
6344 srtt,
6345 idx, 99);
6346 }
6347 if (to < rack_tlp_min) {
6348 to = rack_tlp_min;
6349 }
6350 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6351 /*
6352 * If the TLP time works out to larger than the max
6353 * RTO lets not do TLP.. just RTO.
6354 */
6355 goto activate_rxt;
6356 }
6357 }
6358 if (is_tlp_timer == 0) {
6359 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6360 } else {
6361 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6362 }
6363 if (to == 0)
6364 to = 1;
6365 return (to);
6366 }
6367
6368 static void
rack_enter_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,tcp_seq snd_una)6369 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6370 {
6371 if (rack->rc_in_persist == 0) {
6372 if (tp->t_flags & TF_GPUTINPROG) {
6373 /*
6374 * Stop the goodput now, the calling of the
6375 * measurement function clears the flag.
6376 */
6377 rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6378 RACK_QUALITY_PERSIST);
6379 }
6380 #ifdef NETFLIX_SHARED_CWND
6381 if (rack->r_ctl.rc_scw) {
6382 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6383 rack->rack_scwnd_is_idle = 1;
6384 }
6385 #endif
6386 rack->r_ctl.rc_went_idle_time = cts;
6387 if (rack->r_ctl.rc_went_idle_time == 0)
6388 rack->r_ctl.rc_went_idle_time = 1;
6389 if (rack->lt_bw_up) {
6390 /* Suspend our LT BW measurement */
6391 uint64_t tmark;
6392
6393 rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6394 rack->r_ctl.lt_seq = snd_una;
6395 tmark = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
6396 if (tmark >= rack->r_ctl.lt_timemark) {
6397 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6398 }
6399 rack->r_ctl.lt_timemark = tmark;
6400 rack->lt_bw_up = 0;
6401 rack->r_persist_lt_bw_off = 1;
6402 }
6403 rack_timer_cancel(tp, rack, cts, __LINE__);
6404 rack->r_ctl.persist_lost_ends = 0;
6405 rack->probe_not_answered = 0;
6406 rack->forced_ack = 0;
6407 tp->t_rxtshift = 0;
6408 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6409 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6410 rack->rc_in_persist = 1;
6411 }
6412 }
6413
6414 static void
rack_exit_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)6415 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6416 {
6417 if (tcp_in_hpts(rack->rc_tp)) {
6418 tcp_hpts_remove(rack->rc_tp);
6419 rack->r_ctl.rc_hpts_flags = 0;
6420 }
6421 #ifdef NETFLIX_SHARED_CWND
6422 if (rack->r_ctl.rc_scw) {
6423 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6424 rack->rack_scwnd_is_idle = 0;
6425 }
6426 #endif
6427 if (rack->rc_gp_dyn_mul &&
6428 (rack->use_fixed_rate == 0) &&
6429 (rack->rc_always_pace)) {
6430 /*
6431 * Do we count this as if a probe-rtt just
6432 * finished?
6433 */
6434 uint32_t time_idle, idle_min;
6435
6436 time_idle = cts - rack->r_ctl.rc_went_idle_time;
6437 idle_min = rack_min_probertt_hold;
6438 if (rack_probertt_gpsrtt_cnt_div) {
6439 uint64_t extra;
6440 extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6441 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
6442 extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6443 idle_min += (uint32_t)extra;
6444 }
6445 if (time_idle >= idle_min) {
6446 /* Yes, we count it as a probe-rtt. */
6447 uint32_t us_cts;
6448
6449 us_cts = tcp_get_usecs(NULL);
6450 if (rack->in_probe_rtt == 0) {
6451 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6452 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6453 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6454 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6455 } else {
6456 rack_exit_probertt(rack, us_cts);
6457 }
6458 }
6459 }
6460 if (rack->r_persist_lt_bw_off) {
6461 /* Continue where we left off */
6462 rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
6463 rack->lt_bw_up = 1;
6464 rack->r_persist_lt_bw_off = 0;
6465 }
6466 rack->rc_in_persist = 0;
6467 rack->r_ctl.rc_went_idle_time = 0;
6468 tp->t_rxtshift = 0;
6469 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6470 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6471 rack->r_ctl.rc_agg_delayed = 0;
6472 rack->r_early = 0;
6473 rack->r_late = 0;
6474 rack->r_ctl.rc_agg_early = 0;
6475 }
6476
6477 static void
rack_log_hpts_diag(struct tcp_rack * rack,uint32_t cts,struct hpts_diag * diag,struct timeval * tv)6478 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6479 struct hpts_diag *diag, struct timeval *tv)
6480 {
6481 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6482 union tcp_log_stackspecific log;
6483
6484 memset(&log, 0, sizeof(log));
6485 log.u_bbr.flex1 = diag->p_nxt_slot;
6486 log.u_bbr.flex2 = diag->p_cur_slot;
6487 log.u_bbr.flex3 = diag->slot_req;
6488 log.u_bbr.flex4 = diag->inp_hptsslot;
6489 log.u_bbr.flex5 = diag->slot_remaining;
6490 log.u_bbr.flex6 = diag->need_new_to;
6491 log.u_bbr.flex7 = diag->p_hpts_active;
6492 log.u_bbr.flex8 = diag->p_on_min_sleep;
6493 /* Hijack other fields as needed */
6494 log.u_bbr.epoch = diag->have_slept;
6495 log.u_bbr.lt_epoch = diag->yet_to_sleep;
6496 log.u_bbr.pkts_out = diag->co_ret;
6497 log.u_bbr.applimited = diag->hpts_sleep_time;
6498 log.u_bbr.delivered = diag->p_prev_slot;
6499 log.u_bbr.inflight = diag->p_runningslot;
6500 log.u_bbr.bw_inuse = diag->wheel_slot;
6501 log.u_bbr.rttProp = diag->wheel_cts;
6502 log.u_bbr.timeStamp = cts;
6503 log.u_bbr.delRate = diag->maxslots;
6504 log.u_bbr.cur_del_rate = diag->p_curtick;
6505 log.u_bbr.cur_del_rate <<= 32;
6506 log.u_bbr.cur_del_rate |= diag->p_lasttick;
6507 TCP_LOG_EVENTP(rack->rc_tp, NULL,
6508 &rack->rc_inp->inp_socket->so_rcv,
6509 &rack->rc_inp->inp_socket->so_snd,
6510 BBR_LOG_HPTSDIAG, 0,
6511 0, &log, false, tv);
6512 }
6513
6514 }
6515
6516 static void
rack_log_wakeup(struct tcpcb * tp,struct tcp_rack * rack,struct sockbuf * sb,uint32_t len,int type)6517 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6518 {
6519 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6520 union tcp_log_stackspecific log;
6521 struct timeval tv;
6522
6523 memset(&log, 0, sizeof(log));
6524 log.u_bbr.flex1 = sb->sb_flags;
6525 log.u_bbr.flex2 = len;
6526 log.u_bbr.flex3 = sb->sb_state;
6527 log.u_bbr.flex8 = type;
6528 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6529 TCP_LOG_EVENTP(rack->rc_tp, NULL,
6530 &rack->rc_inp->inp_socket->so_rcv,
6531 &rack->rc_inp->inp_socket->so_snd,
6532 TCP_LOG_SB_WAKE, 0,
6533 len, &log, false, &tv);
6534 }
6535 }
6536
6537 static void
rack_start_hpts_timer(struct tcp_rack * rack,struct tcpcb * tp,uint32_t cts,int32_t slot,uint32_t tot_len_this_send,int sup_rack)6538 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6539 int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6540 {
6541 struct hpts_diag diag;
6542 struct inpcb *inp = tptoinpcb(tp);
6543 struct timeval tv;
6544 uint32_t delayed_ack = 0;
6545 uint32_t hpts_timeout;
6546 uint32_t entry_slot = slot;
6547 uint8_t stopped;
6548 uint32_t left = 0;
6549 uint32_t us_cts;
6550
6551 if ((tp->t_state == TCPS_CLOSED) ||
6552 (tp->t_state == TCPS_LISTEN)) {
6553 return;
6554 }
6555 if (tcp_in_hpts(tp)) {
6556 /* Already on the pacer */
6557 return;
6558 }
6559 stopped = rack->rc_tmr_stopped;
6560 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6561 left = rack->r_ctl.rc_timer_exp - cts;
6562 }
6563 rack->r_ctl.rc_timer_exp = 0;
6564 rack->r_ctl.rc_hpts_flags = 0;
6565 us_cts = tcp_get_usecs(&tv);
6566 /* Now early/late accounting */
6567 rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6568 if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6569 /*
6570 * We have a early carry over set,
6571 * we can always add more time so we
6572 * can always make this compensation.
6573 *
6574 * Note if ack's are allowed to wake us do not
6575 * penalize the next timer for being awoke
6576 * by an ack aka the rc_agg_early (non-paced mode).
6577 */
6578 slot += rack->r_ctl.rc_agg_early;
6579 rack->r_early = 0;
6580 rack->r_ctl.rc_agg_early = 0;
6581 }
6582 if ((rack->r_late) &&
6583 ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
6584 /*
6585 * This is harder, we can
6586 * compensate some but it
6587 * really depends on what
6588 * the current pacing time is.
6589 */
6590 if (rack->r_ctl.rc_agg_delayed >= slot) {
6591 /*
6592 * We can't compensate for it all.
6593 * And we have to have some time
6594 * on the clock. We always have a min
6595 * 10 slots (10 x 10 i.e. 100 usecs).
6596 */
6597 if (slot <= HPTS_TICKS_PER_SLOT) {
6598 /* We gain delay */
6599 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6600 slot = HPTS_TICKS_PER_SLOT;
6601 } else {
6602 /* We take off some */
6603 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6604 slot = HPTS_TICKS_PER_SLOT;
6605 }
6606 } else {
6607 slot -= rack->r_ctl.rc_agg_delayed;
6608 rack->r_ctl.rc_agg_delayed = 0;
6609 /* Make sure we have 100 useconds at minimum */
6610 if (slot < HPTS_TICKS_PER_SLOT) {
6611 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6612 slot = HPTS_TICKS_PER_SLOT;
6613 }
6614 if (rack->r_ctl.rc_agg_delayed == 0)
6615 rack->r_late = 0;
6616 }
6617 } else if (rack->r_late) {
6618 /* r_use_hpts_min is on and so is DGP */
6619 uint32_t max_red;
6620
6621 max_red = (slot * rack->r_ctl.max_reduction) / 100;
6622 if (max_red >= rack->r_ctl.rc_agg_delayed) {
6623 slot -= rack->r_ctl.rc_agg_delayed;
6624 rack->r_ctl.rc_agg_delayed = 0;
6625 } else {
6626 slot -= max_red;
6627 rack->r_ctl.rc_agg_delayed -= max_red;
6628 }
6629 }
6630 if ((rack->r_use_hpts_min == 1) &&
6631 (slot > 0) &&
6632 (rack->dgp_on == 1)) {
6633 /*
6634 * We are enforcing a min pacing timer
6635 * based on our hpts min timeout.
6636 */
6637 uint32_t min;
6638
6639 min = get_hpts_min_sleep_time();
6640 if (min > slot) {
6641 slot = min;
6642 }
6643 }
6644 hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6645 if (tp->t_flags & TF_DELACK) {
6646 delayed_ack = TICKS_2_USEC(tcp_delacktime);
6647 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6648 }
6649 if (delayed_ack && ((hpts_timeout == 0) ||
6650 (delayed_ack < hpts_timeout)))
6651 hpts_timeout = delayed_ack;
6652 else
6653 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6654 /*
6655 * If no timers are going to run and we will fall off the hptsi
6656 * wheel, we resort to a keep-alive timer if its configured.
6657 */
6658 if ((hpts_timeout == 0) &&
6659 (slot == 0)) {
6660 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6661 (tp->t_state <= TCPS_CLOSING)) {
6662 /*
6663 * Ok we have no timer (persists, rack, tlp, rxt or
6664 * del-ack), we don't have segments being paced. So
6665 * all that is left is the keepalive timer.
6666 */
6667 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6668 /* Get the established keep-alive time */
6669 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6670 } else {
6671 /*
6672 * Get the initial setup keep-alive time,
6673 * note that this is probably not going to
6674 * happen, since rack will be running a rxt timer
6675 * if a SYN of some sort is outstanding. It is
6676 * actually handled in rack_timeout_rxt().
6677 */
6678 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6679 }
6680 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6681 if (rack->in_probe_rtt) {
6682 /*
6683 * We want to instead not wake up a long time from
6684 * now but to wake up about the time we would
6685 * exit probe-rtt and initiate a keep-alive ack.
6686 * This will get us out of probe-rtt and update
6687 * our min-rtt.
6688 */
6689 hpts_timeout = rack_min_probertt_hold;
6690 }
6691 }
6692 }
6693 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6694 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6695 /*
6696 * RACK, TLP, persists and RXT timers all are restartable
6697 * based on actions input .. i.e we received a packet (ack
6698 * or sack) and that changes things (rw, or snd_una etc).
6699 * Thus we can restart them with a new value. For
6700 * keep-alive, delayed_ack we keep track of what was left
6701 * and restart the timer with a smaller value.
6702 */
6703 if (left < hpts_timeout)
6704 hpts_timeout = left;
6705 }
6706 if (hpts_timeout) {
6707 /*
6708 * Hack alert for now we can't time-out over 2,147,483
6709 * seconds (a bit more than 596 hours), which is probably ok
6710 * :).
6711 */
6712 if (hpts_timeout > 0x7ffffffe)
6713 hpts_timeout = 0x7ffffffe;
6714 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6715 }
6716 rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6717 if ((rack->gp_ready == 0) &&
6718 (rack->use_fixed_rate == 0) &&
6719 (hpts_timeout < slot) &&
6720 (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6721 /*
6722 * We have no good estimate yet for the
6723 * old clunky burst mitigation or the
6724 * real pacing. And the tlp or rxt is smaller
6725 * than the pacing calculation. Lets not
6726 * pace that long since we know the calculation
6727 * so far is not accurate.
6728 */
6729 slot = hpts_timeout;
6730 }
6731 /**
6732 * Turn off all the flags for queuing by default. The
6733 * flags have important meanings to what happens when
6734 * LRO interacts with the transport. Most likely (by default now)
6735 * mbuf_queueing and ack compression are on. So the transport
6736 * has a couple of flags that control what happens (if those
6737 * are not on then these flags won't have any effect since it
6738 * won't go through the queuing LRO path).
6739 *
6740 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6741 * pacing output, so don't disturb. But
6742 * it also means LRO can wake me if there
6743 * is a SACK arrival.
6744 *
6745 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6746 * with the above flag (QUEUE_READY) and
6747 * when present it says don't even wake me
6748 * if a SACK arrives.
6749 *
6750 * The idea behind these flags is that if we are pacing we
6751 * set the MBUF_QUEUE_READY and only get woken up if
6752 * a SACK arrives (which could change things) or if
6753 * our pacing timer expires. If, however, we have a rack
6754 * timer running, then we don't even want a sack to wake
6755 * us since the rack timer has to expire before we can send.
6756 *
6757 * Other cases should usually have none of the flags set
6758 * so LRO can call into us.
6759 */
6760 tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6761 if (slot) {
6762 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6763 rack->r_ctl.rc_last_output_to = us_cts + slot;
6764 /*
6765 * A pacing timer (slot) is being set, in
6766 * such a case we cannot send (we are blocked by
6767 * the timer). So lets tell LRO that it should not
6768 * wake us unless there is a SACK. Note this only
6769 * will be effective if mbuf queueing is on or
6770 * compressed acks are being processed.
6771 */
6772 tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6773 /*
6774 * But wait if we have a Rack timer running
6775 * even a SACK should not disturb us (with
6776 * the exception of r_rr_config 3).
6777 */
6778 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6779 (IN_RECOVERY(tp->t_flags))) {
6780 if (rack->r_rr_config != 3)
6781 tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6782 else if (rack->rc_pace_dnd) {
6783 /*
6784 * When DND is on, we only let a sack
6785 * interrupt us if we are not in recovery.
6786 *
6787 * If DND is off, then we never hit here
6788 * and let all sacks wake us up.
6789 *
6790 */
6791 tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6792 }
6793 }
6794 if (rack->rc_ack_can_sendout_data) {
6795 /*
6796 * Ahh but wait, this is that special case
6797 * where the pacing timer can be disturbed
6798 * backout the changes (used for non-paced
6799 * burst limiting).
6800 */
6801 tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6802 TF2_MBUF_QUEUE_READY);
6803 }
6804 if ((rack->use_rack_rr) &&
6805 (rack->r_rr_config < 2) &&
6806 ((hpts_timeout) && (hpts_timeout < slot))) {
6807 /*
6808 * Arrange for the hpts to kick back in after the
6809 * t-o if the t-o does not cause a send.
6810 */
6811 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6812 __LINE__, &diag);
6813 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6814 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6815 } else {
6816 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
6817 __LINE__, &diag);
6818 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6819 rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6820 }
6821 } else if (hpts_timeout) {
6822 /*
6823 * With respect to t_flags2(?) here, lets let any new acks wake
6824 * us up here. Since we are not pacing (no pacing timer), output
6825 * can happen so we should let it. If its a Rack timer, then any inbound
6826 * packet probably won't change the sending (we will be blocked)
6827 * but it may change the prr stats so letting it in (the set defaults
6828 * at the start of this block) are good enough.
6829 */
6830 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6831 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6832 __LINE__, &diag);
6833 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6834 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6835 } else {
6836 /* No timer starting */
6837 #ifdef INVARIANTS
6838 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6839 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6840 tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6841 }
6842 #endif
6843 }
6844 rack->rc_tmr_stopped = 0;
6845 if (slot)
6846 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6847 }
6848
6849 static void
rack_mark_lost(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t cts)6850 rack_mark_lost(struct tcpcb *tp,
6851 struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
6852 {
6853 struct rack_sendmap *nrsm;
6854 uint32_t thresh, exp;
6855
6856 thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
6857 nrsm = rsm;
6858 TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
6859 if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
6860 /* Got up to all that were marked sack-passed */
6861 break;
6862 }
6863 if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
6864 exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
6865 if (TSTMP_LT(exp, cts) || (exp == cts)) {
6866 /* We now consider it lost */
6867 nrsm->r_flags |= RACK_WAS_LOST;
6868 rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
6869 } else {
6870 /* Past here it won't be lost so stop */
6871 break;
6872 }
6873 }
6874 }
6875 }
6876
6877 /*
6878 * RACK Timer, here we simply do logging and house keeping.
6879 * the normal rack_output() function will call the
6880 * appropriate thing to check if we need to do a RACK retransmit.
6881 * We return 1, saying don't proceed with rack_output only
6882 * when all timers have been stopped (destroyed PCB?).
6883 */
6884 static int
rack_timeout_rack(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)6885 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6886 {
6887 /*
6888 * This timer simply provides an internal trigger to send out data.
6889 * The check_recovery_mode call will see if there are needed
6890 * retransmissions, if so we will enter fast-recovery. The output
6891 * call may or may not do the same thing depending on sysctl
6892 * settings.
6893 */
6894 struct rack_sendmap *rsm;
6895
6896 counter_u64_add(rack_to_tot, 1);
6897 if (rack->r_state && (rack->r_state != tp->t_state))
6898 rack_set_state(tp, rack);
6899 rack->rc_on_min_to = 0;
6900 rsm = rack_check_recovery_mode(tp, cts);
6901 rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6902 if (rsm) {
6903 /* We need to stroke any lost that are now declared as lost */
6904 rack_mark_lost(tp, rack, rsm, cts);
6905 rack->r_ctl.rc_resend = rsm;
6906 rack->r_timer_override = 1;
6907 if (rack->use_rack_rr) {
6908 /*
6909 * Don't accumulate extra pacing delay
6910 * we are allowing the rack timer to
6911 * over-ride pacing i.e. rrr takes precedence
6912 * if the pacing interval is longer than the rrr
6913 * time (in other words we get the min pacing
6914 * time versus rrr pacing time).
6915 */
6916 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6917 }
6918 }
6919 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6920 if (rsm == NULL) {
6921 /* restart a timer and return 1 */
6922 rack_start_hpts_timer(rack, tp, cts,
6923 0, 0, 0);
6924 return (1);
6925 }
6926 return (0);
6927 }
6928
6929
6930
6931 static void
rack_adjust_orig_mlen(struct rack_sendmap * rsm)6932 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6933 {
6934
6935 if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6936 /*
6937 * The trailing space changed, mbufs can grow
6938 * at the tail but they can't shrink from
6939 * it, KASSERT that. Adjust the orig_m_len to
6940 * compensate for this change.
6941 */
6942 KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
6943 ("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
6944 rsm->m,
6945 rsm,
6946 (intmax_t)M_TRAILINGROOM(rsm->m),
6947 rsm->orig_t_space,
6948 rsm->orig_m_len,
6949 rsm->m->m_len));
6950 rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
6951 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
6952 }
6953 if (rsm->m->m_len < rsm->orig_m_len) {
6954 /*
6955 * Mbuf shrank, trimmed off the top by an ack, our
6956 * offset changes.
6957 */
6958 KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
6959 ("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
6960 rsm->m, rsm->m->m_len,
6961 rsm, rsm->orig_m_len,
6962 rsm->soff));
6963 if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
6964 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6965 else
6966 rsm->soff = 0;
6967 rsm->orig_m_len = rsm->m->m_len;
6968 #ifdef INVARIANTS
6969 } else if (rsm->m->m_len > rsm->orig_m_len) {
6970 panic("rsm:%p m:%p m_len grew outside of t_space compensation",
6971 rsm, rsm->m);
6972 #endif
6973 }
6974 }
6975
6976 static void
rack_setup_offset_for_rsm(struct tcp_rack * rack,struct rack_sendmap * src_rsm,struct rack_sendmap * rsm)6977 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6978 {
6979 struct mbuf *m;
6980 uint32_t soff;
6981
6982 if (src_rsm->m &&
6983 ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
6984 (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
6985 /* Fix up the orig_m_len and possibly the mbuf offset */
6986 rack_adjust_orig_mlen(src_rsm);
6987 }
6988 m = src_rsm->m;
6989 soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6990 while (soff >= m->m_len) {
6991 /* Move out past this mbuf */
6992 soff -= m->m_len;
6993 m = m->m_next;
6994 KASSERT((m != NULL),
6995 ("rsm:%p nrsm:%p hit at soff:%u null m",
6996 src_rsm, rsm, soff));
6997 if (m == NULL) {
6998 /* This should *not* happen which is why there is a kassert */
6999 src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7000 (src_rsm->r_start - rack->rc_tp->snd_una),
7001 &src_rsm->soff);
7002 src_rsm->orig_m_len = src_rsm->m->m_len;
7003 src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7004 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7005 (rsm->r_start - rack->rc_tp->snd_una),
7006 &rsm->soff);
7007 rsm->orig_m_len = rsm->m->m_len;
7008 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7009 return;
7010 }
7011 }
7012 rsm->m = m;
7013 rsm->soff = soff;
7014 rsm->orig_m_len = m->m_len;
7015 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7016 }
7017
7018 static __inline void
rack_clone_rsm(struct tcp_rack * rack,struct rack_sendmap * nrsm,struct rack_sendmap * rsm,uint32_t start)7019 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7020 struct rack_sendmap *rsm, uint32_t start)
7021 {
7022 int idx;
7023
7024 nrsm->r_start = start;
7025 nrsm->r_end = rsm->r_end;
7026 nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7027 nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7028 nrsm->r_flags = rsm->r_flags;
7029 nrsm->r_dupack = rsm->r_dupack;
7030 nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7031 nrsm->r_rtr_bytes = 0;
7032 nrsm->r_fas = rsm->r_fas;
7033 nrsm->r_bas = rsm->r_bas;
7034 tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7035 nrsm->r_just_ret = rsm->r_just_ret;
7036 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7037 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7038 }
7039 /* Now if we have SYN flag we keep it on the left edge */
7040 if (nrsm->r_flags & RACK_HAS_SYN)
7041 nrsm->r_flags &= ~RACK_HAS_SYN;
7042 /* Now if we have a FIN flag we keep it on the right edge */
7043 if (rsm->r_flags & RACK_HAS_FIN)
7044 rsm->r_flags &= ~RACK_HAS_FIN;
7045 /* Push bit must go to the right edge as well */
7046 if (rsm->r_flags & RACK_HAD_PUSH)
7047 rsm->r_flags &= ~RACK_HAD_PUSH;
7048 /* Clone over the state of the hw_tls flag */
7049 nrsm->r_hw_tls = rsm->r_hw_tls;
7050 /*
7051 * Now we need to find nrsm's new location in the mbuf chain
7052 * we basically calculate a new offset, which is soff +
7053 * how much is left in original rsm. Then we walk out the mbuf
7054 * chain to find the righ position, it may be the same mbuf
7055 * or maybe not.
7056 */
7057 KASSERT(((rsm->m != NULL) ||
7058 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7059 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7060 if (rsm->m)
7061 rack_setup_offset_for_rsm(rack, rsm, nrsm);
7062 }
7063
7064 static struct rack_sendmap *
rack_merge_rsm(struct tcp_rack * rack,struct rack_sendmap * l_rsm,struct rack_sendmap * r_rsm)7065 rack_merge_rsm(struct tcp_rack *rack,
7066 struct rack_sendmap *l_rsm,
7067 struct rack_sendmap *r_rsm)
7068 {
7069 /*
7070 * We are merging two ack'd RSM's,
7071 * the l_rsm is on the left (lower seq
7072 * values) and the r_rsm is on the right
7073 * (higher seq value). The simplest way
7074 * to merge these is to move the right
7075 * one into the left. I don't think there
7076 * is any reason we need to try to find
7077 * the oldest (or last oldest retransmitted).
7078 */
7079 rack_log_map_chg(rack->rc_tp, rack, NULL,
7080 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7081 tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7082 if (l_rsm->r_dupack < r_rsm->r_dupack)
7083 l_rsm->r_dupack = r_rsm->r_dupack;
7084 if (r_rsm->r_rtr_bytes)
7085 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7086 if (r_rsm->r_in_tmap) {
7087 /* This really should not happen */
7088 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7089 r_rsm->r_in_tmap = 0;
7090 }
7091
7092 /* Now the flags */
7093 if (r_rsm->r_flags & RACK_HAS_FIN)
7094 l_rsm->r_flags |= RACK_HAS_FIN;
7095 if (r_rsm->r_flags & RACK_TLP)
7096 l_rsm->r_flags |= RACK_TLP;
7097 if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7098 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7099 if ((r_rsm->r_flags & RACK_APP_LIMITED) &&
7100 ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7101 /*
7102 * If both are app-limited then let the
7103 * free lower the count. If right is app
7104 * limited and left is not, transfer.
7105 */
7106 l_rsm->r_flags |= RACK_APP_LIMITED;
7107 r_rsm->r_flags &= ~RACK_APP_LIMITED;
7108 if (r_rsm == rack->r_ctl.rc_first_appl)
7109 rack->r_ctl.rc_first_appl = l_rsm;
7110 }
7111 tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7112 /*
7113 * We keep the largest value, which is the newest
7114 * send. We do this in case a segment that is
7115 * joined together and not part of a GP estimate
7116 * later gets expanded into the GP estimate.
7117 *
7118 * We prohibit the merging of unlike kinds i.e.
7119 * all pieces that are in the GP estimate can be
7120 * merged and all pieces that are not in a GP estimate
7121 * can be merged, but not disimilar pieces. Combine
7122 * this with taking the highest here and we should
7123 * be ok unless of course the client reneges. Then
7124 * all bets are off.
7125 */
7126 if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7127 r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7128 l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7129 }
7130 /*
7131 * When merging two RSM's we also need to consider the ack time and keep
7132 * newest. If the ack gets merged into a measurement then that is the
7133 * one we will want to be using.
7134 */
7135 if(l_rsm->r_ack_arrival < r_rsm->r_ack_arrival)
7136 l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7137
7138 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7139 /* Transfer the split limit to the map we free */
7140 r_rsm->r_limit_type = l_rsm->r_limit_type;
7141 l_rsm->r_limit_type = 0;
7142 }
7143 rack_free(rack, r_rsm);
7144 l_rsm->r_flags |= RACK_MERGED;
7145 return (l_rsm);
7146 }
7147
7148 /*
7149 * TLP Timer, here we simply setup what segment we want to
7150 * have the TLP expire on, the normal rack_output() will then
7151 * send it out.
7152 *
7153 * We return 1, saying don't proceed with rack_output only
7154 * when all timers have been stopped (destroyed PCB?).
7155 */
7156 static int
rack_timeout_tlp(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,uint8_t * doing_tlp)7157 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7158 {
7159 /*
7160 * Tail Loss Probe.
7161 */
7162 struct rack_sendmap *rsm = NULL;
7163 int insret __diagused;
7164 struct socket *so = tptosocket(tp);
7165 uint32_t amm;
7166 uint32_t out, avail;
7167 int collapsed_win = 0;
7168
7169 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7170 /* Its not time yet */
7171 return (0);
7172 }
7173 if (ctf_progress_timeout_check(tp, true)) {
7174 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7175 return (-ETIMEDOUT); /* tcp_drop() */
7176 }
7177 /*
7178 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7179 * need to figure out how to force a full MSS segment out.
7180 */
7181 rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7182 rack->r_ctl.retran_during_recovery = 0;
7183 rack->r_might_revert = 0;
7184 rack->r_ctl.dsack_byte_cnt = 0;
7185 counter_u64_add(rack_tlp_tot, 1);
7186 if (rack->r_state && (rack->r_state != tp->t_state))
7187 rack_set_state(tp, rack);
7188 avail = sbavail(&so->so_snd);
7189 out = tp->snd_max - tp->snd_una;
7190 if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7191 /* special case, we need a retransmission */
7192 collapsed_win = 1;
7193 goto need_retran;
7194 }
7195 if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7196 rack->r_ctl.dsack_persist--;
7197 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7198 rack->r_ctl.num_dsack = 0;
7199 }
7200 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7201 }
7202 if ((tp->t_flags & TF_GPUTINPROG) &&
7203 (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7204 /*
7205 * If this is the second in a row
7206 * TLP and we are doing a measurement
7207 * its time to abandon the measurement.
7208 * Something is likely broken on
7209 * the clients network and measuring a
7210 * broken network does us no good.
7211 */
7212 tp->t_flags &= ~TF_GPUTINPROG;
7213 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7214 rack->r_ctl.rc_gp_srtt /*flex1*/,
7215 tp->gput_seq,
7216 0, 0, 18, __LINE__, NULL, 0);
7217 }
7218 /*
7219 * Check our send oldest always settings, and if
7220 * there is an oldest to send jump to the need_retran.
7221 */
7222 if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7223 goto need_retran;
7224
7225 if (avail > out) {
7226 /* New data is available */
7227 amm = avail - out;
7228 if (amm > ctf_fixed_maxseg(tp)) {
7229 amm = ctf_fixed_maxseg(tp);
7230 if ((amm + out) > tp->snd_wnd) {
7231 /* We are rwnd limited */
7232 goto need_retran;
7233 }
7234 } else if (amm < ctf_fixed_maxseg(tp)) {
7235 /* not enough to fill a MTU */
7236 goto need_retran;
7237 }
7238 if (IN_FASTRECOVERY(tp->t_flags)) {
7239 /* Unlikely */
7240 if (rack->rack_no_prr == 0) {
7241 if (out + amm <= tp->snd_wnd) {
7242 rack->r_ctl.rc_prr_sndcnt = amm;
7243 rack->r_ctl.rc_tlp_new_data = amm;
7244 rack_log_to_prr(rack, 4, 0, __LINE__);
7245 }
7246 } else
7247 goto need_retran;
7248 } else {
7249 /* Set the send-new override */
7250 if (out + amm <= tp->snd_wnd)
7251 rack->r_ctl.rc_tlp_new_data = amm;
7252 else
7253 goto need_retran;
7254 }
7255 rack->r_ctl.rc_tlpsend = NULL;
7256 counter_u64_add(rack_tlp_newdata, 1);
7257 goto send;
7258 }
7259 need_retran:
7260 /*
7261 * Ok we need to arrange the last un-acked segment to be re-sent, or
7262 * optionally the first un-acked segment.
7263 */
7264 if (collapsed_win == 0) {
7265 if (rack_always_send_oldest)
7266 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7267 else {
7268 rsm = tqhash_max(rack->r_ctl.tqh);
7269 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7270 rsm = rack_find_high_nonack(rack, rsm);
7271 }
7272 }
7273 if (rsm == NULL) {
7274 #ifdef TCP_BLACKBOX
7275 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7276 #endif
7277 goto out;
7278 }
7279 } else {
7280 /*
7281 * We had a collapsed window, lets find
7282 * the point before the collapse.
7283 */
7284 if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7285 rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7286 else {
7287 rsm = tqhash_min(rack->r_ctl.tqh);
7288 }
7289 if (rsm == NULL) {
7290 /* Huh */
7291 goto out;
7292 }
7293 }
7294 if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7295 /*
7296 * We need to split this the last segment in two.
7297 */
7298 struct rack_sendmap *nrsm;
7299
7300 nrsm = rack_alloc_full_limit(rack);
7301 if (nrsm == NULL) {
7302 /*
7303 * No memory to split, we will just exit and punt
7304 * off to the RXT timer.
7305 */
7306 goto out;
7307 }
7308 rack_clone_rsm(rack, nrsm, rsm,
7309 (rsm->r_end - ctf_fixed_maxseg(tp)));
7310 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7311 #ifndef INVARIANTS
7312 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7313 #else
7314 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7315 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7316 nrsm, insret, rack, rsm);
7317 }
7318 #endif
7319 if (rsm->r_in_tmap) {
7320 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7321 nrsm->r_in_tmap = 1;
7322 }
7323 rsm = nrsm;
7324 }
7325 rack->r_ctl.rc_tlpsend = rsm;
7326 send:
7327 /* Make sure output path knows we are doing a TLP */
7328 *doing_tlp = 1;
7329 rack->r_timer_override = 1;
7330 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7331 return (0);
7332 out:
7333 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7334 return (0);
7335 }
7336
7337 /*
7338 * Delayed ack Timer, here we simply need to setup the
7339 * ACK_NOW flag and remove the DELACK flag. From there
7340 * the output routine will send the ack out.
7341 *
7342 * We only return 1, saying don't proceed, if all timers
7343 * are stopped (destroyed PCB?).
7344 */
7345 static int
rack_timeout_delack(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7346 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7347 {
7348
7349 rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7350 tp->t_flags &= ~TF_DELACK;
7351 tp->t_flags |= TF_ACKNOW;
7352 KMOD_TCPSTAT_INC(tcps_delack);
7353 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7354 return (0);
7355 }
7356
7357 static inline int
rack_send_ack_challange(struct tcp_rack * rack)7358 rack_send_ack_challange(struct tcp_rack *rack)
7359 {
7360 struct tcptemp *t_template;
7361
7362 t_template = tcpip_maketemplate(rack->rc_inp);
7363 if (t_template) {
7364 if (rack->forced_ack == 0) {
7365 rack->forced_ack = 1;
7366 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7367 } else {
7368 rack->probe_not_answered = 1;
7369 }
7370 tcp_respond(rack->rc_tp, t_template->tt_ipgen,
7371 &t_template->tt_t, (struct mbuf *)NULL,
7372 rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
7373 free(t_template, M_TEMP);
7374 /* This does send an ack so kill any D-ack timer */
7375 if (rack->rc_tp->t_flags & TF_DELACK)
7376 rack->rc_tp->t_flags &= ~TF_DELACK;
7377 return(1);
7378 } else
7379 return (0);
7380
7381 }
7382
7383 /*
7384 * Persists timer, here we simply send the
7385 * same thing as a keepalive will.
7386 * the one byte send.
7387 *
7388 * We only return 1, saying don't proceed, if all timers
7389 * are stopped (destroyed PCB?).
7390 */
7391 static int
rack_timeout_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7392 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7393 {
7394 int32_t retval = 1;
7395
7396 if (rack->rc_in_persist == 0)
7397 return (0);
7398 if (ctf_progress_timeout_check(tp, false)) {
7399 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7400 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7401 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7402 return (-ETIMEDOUT); /* tcp_drop() */
7403 }
7404 /*
7405 * Persistence timer into zero window. Force a byte to be output, if
7406 * possible.
7407 */
7408 KMOD_TCPSTAT_INC(tcps_persisttimeo);
7409 /*
7410 * Hack: if the peer is dead/unreachable, we do not time out if the
7411 * window is closed. After a full backoff, drop the connection if
7412 * the idle time (no responses to probes) reaches the maximum
7413 * backoff that we would use if retransmitting.
7414 */
7415 if (tp->t_rxtshift >= V_tcp_retries &&
7416 (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7417 TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7418 KMOD_TCPSTAT_INC(tcps_persistdrop);
7419 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7420 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7421 retval = -ETIMEDOUT; /* tcp_drop() */
7422 goto out;
7423 }
7424 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7425 tp->snd_una == tp->snd_max)
7426 rack_exit_persist(tp, rack, cts);
7427 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7428 /*
7429 * If the user has closed the socket then drop a persisting
7430 * connection after a much reduced timeout.
7431 */
7432 if (tp->t_state > TCPS_CLOSE_WAIT &&
7433 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7434 KMOD_TCPSTAT_INC(tcps_persistdrop);
7435 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7436 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7437 retval = -ETIMEDOUT; /* tcp_drop() */
7438 goto out;
7439 }
7440 if (rack_send_ack_challange(rack)) {
7441 /* only set it if we were answered */
7442 if (rack->probe_not_answered) {
7443 counter_u64_add(rack_persists_loss, 1);
7444 rack->r_ctl.persist_lost_ends++;
7445 }
7446 counter_u64_add(rack_persists_sends, 1);
7447 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7448 }
7449 if (tp->t_rxtshift < V_tcp_retries)
7450 tp->t_rxtshift++;
7451 out:
7452 rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7453 rack_start_hpts_timer(rack, tp, cts,
7454 0, 0, 0);
7455 return (retval);
7456 }
7457
7458 /*
7459 * If a keepalive goes off, we had no other timers
7460 * happening. We always return 1 here since this
7461 * routine either drops the connection or sends
7462 * out a segment with respond.
7463 */
7464 static int
rack_timeout_keepalive(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7465 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7466 {
7467 struct inpcb *inp = tptoinpcb(tp);
7468
7469 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7470 rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7471 /*
7472 * Keep-alive timer went off; send something or drop connection if
7473 * idle for too long.
7474 */
7475 KMOD_TCPSTAT_INC(tcps_keeptimeo);
7476 if (tp->t_state < TCPS_ESTABLISHED)
7477 goto dropit;
7478 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7479 tp->t_state <= TCPS_CLOSING) {
7480 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7481 goto dropit;
7482 /*
7483 * Send a packet designed to force a response if the peer is
7484 * up and reachable: either an ACK if the connection is
7485 * still alive, or an RST if the peer has closed the
7486 * connection due to timeout or reboot. Using sequence
7487 * number tp->snd_una-1 causes the transmitted zero-length
7488 * segment to lie outside the receive window; by the
7489 * protocol spec, this requires the correspondent TCP to
7490 * respond.
7491 */
7492 KMOD_TCPSTAT_INC(tcps_keepprobe);
7493 rack_send_ack_challange(rack);
7494 }
7495 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7496 return (1);
7497 dropit:
7498 KMOD_TCPSTAT_INC(tcps_keepdrops);
7499 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7500 return (-ETIMEDOUT); /* tcp_drop() */
7501 }
7502
7503 /*
7504 * Retransmit helper function, clear up all the ack
7505 * flags and take care of important book keeping.
7506 */
7507 static void
rack_remxt_tmr(struct tcpcb * tp)7508 rack_remxt_tmr(struct tcpcb *tp)
7509 {
7510 /*
7511 * The retransmit timer went off, all sack'd blocks must be
7512 * un-acked.
7513 */
7514 struct rack_sendmap *rsm, *trsm = NULL;
7515 struct tcp_rack *rack;
7516
7517 rack = (struct tcp_rack *)tp->t_fb_ptr;
7518 rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7519 rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7520 rack->r_timer_override = 1;
7521 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7522 rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7523 rack->r_late = 0;
7524 rack->r_early = 0;
7525 rack->r_ctl.rc_agg_delayed = 0;
7526 rack->r_ctl.rc_agg_early = 0;
7527 if (rack->r_state && (rack->r_state != tp->t_state))
7528 rack_set_state(tp, rack);
7529 if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
7530 /*
7531 * We do not clear the scoreboard until we have had
7532 * more than rack_rxt_scoreboard_clear_thresh time-outs.
7533 */
7534 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7535 if (rack->r_ctl.rc_resend != NULL)
7536 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7537
7538 return;
7539 }
7540 /*
7541 * Ideally we would like to be able to
7542 * mark SACK-PASS on anything not acked here.
7543 *
7544 * However, if we do that we would burst out
7545 * all that data 1ms apart. This would be unwise,
7546 * so for now we will just let the normal rxt timer
7547 * and tlp timer take care of it.
7548 *
7549 * Also we really need to stick them back in sequence
7550 * order. This way we send in the proper order and any
7551 * sacks that come floating in will "re-ack" the data.
7552 * To do this we zap the tmap with an INIT and then
7553 * walk through and place every rsm in the tail queue
7554 * hash table back in its seq ordered place.
7555 */
7556 TAILQ_INIT(&rack->r_ctl.rc_tmap);
7557
7558 TQHASH_FOREACH(rsm, rack->r_ctl.tqh) {
7559 rsm->r_dupack = 0;
7560 if (rack_verbose_logging)
7561 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7562 /* We must re-add it back to the tlist */
7563 if (trsm == NULL) {
7564 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7565 } else {
7566 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7567 }
7568 rsm->r_in_tmap = 1;
7569 trsm = rsm;
7570 if (rsm->r_flags & RACK_ACKED)
7571 rsm->r_flags |= RACK_WAS_ACKED;
7572 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
7573 rsm->r_flags |= RACK_MUST_RXT;
7574 }
7575 /* zero the lost since it's all gone */
7576 rack->r_ctl.rc_considered_lost = 0;
7577 /* Clear the count (we just un-acked them) */
7578 rack->r_ctl.rc_sacked = 0;
7579 rack->r_ctl.rc_sacklast = NULL;
7580 /* Clear the tlp rtx mark */
7581 rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7582 if (rack->r_ctl.rc_resend != NULL)
7583 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7584 rack->r_ctl.rc_prr_sndcnt = 0;
7585 rack_log_to_prr(rack, 6, 0, __LINE__);
7586 rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7587 if (rack->r_ctl.rc_resend != NULL)
7588 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7589 if (((tp->t_flags & TF_SACK_PERMIT) == 0) &&
7590 ((tp->t_flags & TF_SENTFIN) == 0)) {
7591 /*
7592 * For non-sack customers new data
7593 * needs to go out as retransmits until
7594 * we retransmit up to snd_max.
7595 */
7596 rack->r_must_retran = 1;
7597 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7598 rack->r_ctl.rc_sacked);
7599 }
7600 }
7601
7602 static void
rack_convert_rtts(struct tcpcb * tp)7603 rack_convert_rtts(struct tcpcb *tp)
7604 {
7605 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7606 tp->t_rxtcur = RACK_REXMTVAL(tp);
7607 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7608 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7609 }
7610 if (tp->t_rxtcur > rack_rto_max) {
7611 tp->t_rxtcur = rack_rto_max;
7612 }
7613 }
7614
7615 static void
rack_cc_conn_init(struct tcpcb * tp)7616 rack_cc_conn_init(struct tcpcb *tp)
7617 {
7618 struct tcp_rack *rack;
7619 uint32_t srtt;
7620
7621 rack = (struct tcp_rack *)tp->t_fb_ptr;
7622 srtt = tp->t_srtt;
7623 cc_conn_init(tp);
7624 /*
7625 * Now convert to rack's internal format,
7626 * if required.
7627 */
7628 if ((srtt == 0) && (tp->t_srtt != 0))
7629 rack_convert_rtts(tp);
7630 /*
7631 * We want a chance to stay in slowstart as
7632 * we create a connection. TCP spec says that
7633 * initially ssthresh is infinite. For our
7634 * purposes that is the snd_wnd.
7635 */
7636 if (tp->snd_ssthresh < tp->snd_wnd) {
7637 tp->snd_ssthresh = tp->snd_wnd;
7638 }
7639 /*
7640 * We also want to assure a IW worth of
7641 * data can get inflight.
7642 */
7643 if (rc_init_window(rack) < tp->snd_cwnd)
7644 tp->snd_cwnd = rc_init_window(rack);
7645 }
7646
7647 /*
7648 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7649 * we will setup to retransmit the lowest seq number outstanding.
7650 */
7651 static int
rack_timeout_rxt(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7652 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7653 {
7654 struct inpcb *inp = tptoinpcb(tp);
7655 int32_t rexmt;
7656 int32_t retval = 0;
7657 bool isipv6;
7658
7659 if ((tp->t_flags & TF_GPUTINPROG) &&
7660 (tp->t_rxtshift)) {
7661 /*
7662 * We have had a second timeout
7663 * measurements on successive rxt's are not profitable.
7664 * It is unlikely to be of any use (the network is
7665 * broken or the client went away).
7666 */
7667 tp->t_flags &= ~TF_GPUTINPROG;
7668 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7669 rack->r_ctl.rc_gp_srtt /*flex1*/,
7670 tp->gput_seq,
7671 0, 0, 18, __LINE__, NULL, 0);
7672 }
7673 if (ctf_progress_timeout_check(tp, false)) {
7674 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7675 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7676 return (-ETIMEDOUT); /* tcp_drop() */
7677 }
7678 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7679 rack->r_ctl.retran_during_recovery = 0;
7680 rack->rc_ack_required = 1;
7681 rack->r_ctl.dsack_byte_cnt = 0;
7682 if (IN_RECOVERY(tp->t_flags) &&
7683 (rack->rto_from_rec == 0)) {
7684 /*
7685 * Mark that we had a rto while in recovery
7686 * and save the ssthresh so if we go back
7687 * into recovery we will have a chance
7688 * to slowstart back to the level.
7689 */
7690 rack->rto_from_rec = 1;
7691 rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
7692 }
7693 if (IN_FASTRECOVERY(tp->t_flags))
7694 tp->t_flags |= TF_WASFRECOVERY;
7695 else
7696 tp->t_flags &= ~TF_WASFRECOVERY;
7697 if (IN_CONGRECOVERY(tp->t_flags))
7698 tp->t_flags |= TF_WASCRECOVERY;
7699 else
7700 tp->t_flags &= ~TF_WASCRECOVERY;
7701 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7702 (tp->snd_una == tp->snd_max)) {
7703 /* Nothing outstanding .. nothing to do */
7704 return (0);
7705 }
7706 if (rack->r_ctl.dsack_persist) {
7707 rack->r_ctl.dsack_persist--;
7708 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7709 rack->r_ctl.num_dsack = 0;
7710 }
7711 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7712 }
7713 /*
7714 * Rack can only run one timer at a time, so we cannot
7715 * run a KEEPINIT (gating SYN sending) and a retransmit
7716 * timer for the SYN. So if we are in a front state and
7717 * have a KEEPINIT timer we need to check the first transmit
7718 * against now to see if we have exceeded the KEEPINIT time
7719 * (if one is set).
7720 */
7721 if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7722 (TP_KEEPINIT(tp) != 0)) {
7723 struct rack_sendmap *rsm;
7724
7725 rsm = tqhash_min(rack->r_ctl.tqh);
7726 if (rsm) {
7727 /* Ok we have something outstanding to test keepinit with */
7728 if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7729 ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7730 /* We have exceeded the KEEPINIT time */
7731 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7732 goto drop_it;
7733 }
7734 }
7735 }
7736 /*
7737 * Retransmission timer went off. Message has not been acked within
7738 * retransmit interval. Back off to a longer retransmit interval
7739 * and retransmit one segment.
7740 */
7741 if ((rack->r_ctl.rc_resend == NULL) ||
7742 ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7743 /*
7744 * If the rwnd collapsed on
7745 * the one we are retransmitting
7746 * it does not count against the
7747 * rxt count.
7748 */
7749 tp->t_rxtshift++;
7750 }
7751 rack_remxt_tmr(tp);
7752 if (tp->t_rxtshift > V_tcp_retries) {
7753 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7754 drop_it:
7755 tp->t_rxtshift = V_tcp_retries;
7756 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7757 /* XXXGL: previously t_softerror was casted to uint16_t */
7758 MPASS(tp->t_softerror >= 0);
7759 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7760 goto out; /* tcp_drop() */
7761 }
7762 if (tp->t_state == TCPS_SYN_SENT) {
7763 /*
7764 * If the SYN was retransmitted, indicate CWND to be limited
7765 * to 1 segment in cc_conn_init().
7766 */
7767 tp->snd_cwnd = 1;
7768 } else if (tp->t_rxtshift == 1) {
7769 /*
7770 * first retransmit; record ssthresh and cwnd so they can be
7771 * recovered if this turns out to be a "bad" retransmit. A
7772 * retransmit is considered "bad" if an ACK for this segment
7773 * is received within RTT/2 interval; the assumption here is
7774 * that the ACK was already in flight. See "On Estimating
7775 * End-to-End Network Path Properties" by Allman and Paxson
7776 * for more details.
7777 */
7778 tp->snd_cwnd_prev = tp->snd_cwnd;
7779 tp->snd_ssthresh_prev = tp->snd_ssthresh;
7780 tp->snd_recover_prev = tp->snd_recover;
7781 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7782 tp->t_flags |= TF_PREVVALID;
7783 } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7784 tp->t_flags &= ~TF_PREVVALID;
7785 KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7786 if ((tp->t_state == TCPS_SYN_SENT) ||
7787 (tp->t_state == TCPS_SYN_RECEIVED))
7788 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7789 else
7790 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7791
7792 RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7793 max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7794 /*
7795 * We enter the path for PLMTUD if connection is established or, if
7796 * connection is FIN_WAIT_1 status, reason for the last is that if
7797 * amount of data we send is very small, we could send it in couple
7798 * of packets and process straight to FIN. In that case we won't
7799 * catch ESTABLISHED state.
7800 */
7801 #ifdef INET6
7802 isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7803 #else
7804 isipv6 = false;
7805 #endif
7806 if (((V_tcp_pmtud_blackhole_detect == 1) ||
7807 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7808 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7809 ((tp->t_state == TCPS_ESTABLISHED) ||
7810 (tp->t_state == TCPS_FIN_WAIT_1))) {
7811 /*
7812 * Idea here is that at each stage of mtu probe (usually,
7813 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7814 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7815 * should take care of that.
7816 */
7817 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7818 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7819 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7820 tp->t_rxtshift % 2 == 0)) {
7821 /*
7822 * Enter Path MTU Black-hole Detection mechanism: -
7823 * Disable Path MTU Discovery (IP "DF" bit). -
7824 * Reduce MTU to lower value than what we negotiated
7825 * with peer.
7826 */
7827 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7828 /* Record that we may have found a black hole. */
7829 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7830 /* Keep track of previous MSS. */
7831 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7832 }
7833
7834 /*
7835 * Reduce the MSS to blackhole value or to the
7836 * default in an attempt to retransmit.
7837 */
7838 #ifdef INET6
7839 if (isipv6 &&
7840 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7841 /* Use the sysctl tuneable blackhole MSS. */
7842 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7843 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7844 } else if (isipv6) {
7845 /* Use the default MSS. */
7846 tp->t_maxseg = V_tcp_v6mssdflt;
7847 /*
7848 * Disable Path MTU Discovery when we switch
7849 * to minmss.
7850 */
7851 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7852 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7853 }
7854 #endif
7855 #if defined(INET6) && defined(INET)
7856 else
7857 #endif
7858 #ifdef INET
7859 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7860 /* Use the sysctl tuneable blackhole MSS. */
7861 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7862 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7863 } else {
7864 /* Use the default MSS. */
7865 tp->t_maxseg = V_tcp_mssdflt;
7866 /*
7867 * Disable Path MTU Discovery when we switch
7868 * to minmss.
7869 */
7870 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7871 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7872 }
7873 #endif
7874 } else {
7875 /*
7876 * If further retransmissions are still unsuccessful
7877 * with a lowered MTU, maybe this isn't a blackhole
7878 * and we restore the previous MSS and blackhole
7879 * detection flags. The limit '6' is determined by
7880 * giving each probe stage (1448, 1188, 524) 2
7881 * chances to recover.
7882 */
7883 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7884 (tp->t_rxtshift >= 6)) {
7885 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7886 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7887 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7888 if (tp->t_maxseg < V_tcp_mssdflt) {
7889 /*
7890 * The MSS is so small we should not
7891 * process incoming SACK's since we are
7892 * subject to attack in such a case.
7893 */
7894 tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
7895 } else {
7896 tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
7897 }
7898 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7899 }
7900 }
7901 }
7902 /*
7903 * Disable RFC1323 and SACK if we haven't got any response to
7904 * our third SYN to work-around some broken terminal servers
7905 * (most of which have hopefully been retired) that have bad VJ
7906 * header compression code which trashes TCP segments containing
7907 * unknown-to-them TCP options.
7908 */
7909 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7910 (tp->t_rxtshift == 3))
7911 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7912 /*
7913 * If we backed off this far, our srtt estimate is probably bogus.
7914 * Clobber it so we'll take the next rtt measurement as our srtt;
7915 * move the current srtt into rttvar to keep the current retransmit
7916 * times until then.
7917 */
7918 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7919 #ifdef INET6
7920 if ((inp->inp_vflag & INP_IPV6) != 0)
7921 in6_losing(inp);
7922 else
7923 #endif
7924 in_losing(inp);
7925 tp->t_rttvar += tp->t_srtt;
7926 tp->t_srtt = 0;
7927 }
7928 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7929 tp->snd_recover = tp->snd_max;
7930 tp->t_flags |= TF_ACKNOW;
7931 tp->t_rtttime = 0;
7932 rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7933 out:
7934 return (retval);
7935 }
7936
7937 static int
rack_process_timers(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,uint8_t hpts_calling,uint8_t * doing_tlp)7938 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7939 {
7940 int32_t ret = 0;
7941 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7942
7943 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7944 (tp->t_flags & TF_GPUTINPROG)) {
7945 /*
7946 * We have a goodput in progress
7947 * and we have entered a late state.
7948 * Do we have enough data in the sb
7949 * to handle the GPUT request?
7950 */
7951 uint32_t bytes;
7952
7953 bytes = tp->gput_ack - tp->gput_seq;
7954 if (SEQ_GT(tp->gput_seq, tp->snd_una))
7955 bytes += tp->gput_seq - tp->snd_una;
7956 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7957 /*
7958 * There are not enough bytes in the socket
7959 * buffer that have been sent to cover this
7960 * measurement. Cancel it.
7961 */
7962 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7963 rack->r_ctl.rc_gp_srtt /*flex1*/,
7964 tp->gput_seq,
7965 0, 0, 18, __LINE__, NULL, 0);
7966 tp->t_flags &= ~TF_GPUTINPROG;
7967 }
7968 }
7969 if (timers == 0) {
7970 return (0);
7971 }
7972 if (tp->t_state == TCPS_LISTEN) {
7973 /* no timers on listen sockets */
7974 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7975 return (0);
7976 return (1);
7977 }
7978 if ((timers & PACE_TMR_RACK) &&
7979 rack->rc_on_min_to) {
7980 /*
7981 * For the rack timer when we
7982 * are on a min-timeout (which means rrr_conf = 3)
7983 * we don't want to check the timer. It may
7984 * be going off for a pace and thats ok we
7985 * want to send the retransmit (if its ready).
7986 *
7987 * If its on a normal rack timer (non-min) then
7988 * we will check if its expired.
7989 */
7990 goto skip_time_check;
7991 }
7992 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7993 uint32_t left;
7994
7995 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
7996 ret = -1;
7997 rack_log_to_processing(rack, cts, ret, 0);
7998 return (0);
7999 }
8000 if (hpts_calling == 0) {
8001 /*
8002 * A user send or queued mbuf (sack) has called us? We
8003 * return 0 and let the pacing guards
8004 * deal with it if they should or
8005 * should not cause a send.
8006 */
8007 ret = -2;
8008 rack_log_to_processing(rack, cts, ret, 0);
8009 return (0);
8010 }
8011 /*
8012 * Ok our timer went off early and we are not paced false
8013 * alarm, go back to sleep. We make sure we don't have
8014 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8015 * flag in place.
8016 */
8017 rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8018 ret = -3;
8019 left = rack->r_ctl.rc_timer_exp - cts;
8020 tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8021 rack_log_to_processing(rack, cts, ret, left);
8022 return (1);
8023 }
8024 skip_time_check:
8025 rack->rc_tmr_stopped = 0;
8026 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8027 if (timers & PACE_TMR_DELACK) {
8028 ret = rack_timeout_delack(tp, rack, cts);
8029 } else if (timers & PACE_TMR_RACK) {
8030 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8031 rack->r_fast_output = 0;
8032 ret = rack_timeout_rack(tp, rack, cts);
8033 } else if (timers & PACE_TMR_TLP) {
8034 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8035 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8036 } else if (timers & PACE_TMR_RXT) {
8037 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8038 rack->r_fast_output = 0;
8039 ret = rack_timeout_rxt(tp, rack, cts);
8040 } else if (timers & PACE_TMR_PERSIT) {
8041 ret = rack_timeout_persist(tp, rack, cts);
8042 } else if (timers & PACE_TMR_KEEP) {
8043 ret = rack_timeout_keepalive(tp, rack, cts);
8044 }
8045 rack_log_to_processing(rack, cts, ret, timers);
8046 return (ret);
8047 }
8048
8049 static void
rack_timer_cancel(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,int line)8050 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8051 {
8052 struct timeval tv;
8053 uint32_t us_cts, flags_on_entry;
8054 uint8_t hpts_removed = 0;
8055
8056 flags_on_entry = rack->r_ctl.rc_hpts_flags;
8057 us_cts = tcp_get_usecs(&tv);
8058 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8059 ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8060 ((tp->snd_max - tp->snd_una) == 0))) {
8061 tcp_hpts_remove(rack->rc_tp);
8062 hpts_removed = 1;
8063 /* If we were not delayed cancel out the flag. */
8064 if ((tp->snd_max - tp->snd_una) == 0)
8065 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8066 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8067 }
8068 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8069 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8070 if (tcp_in_hpts(rack->rc_tp) &&
8071 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8072 /*
8073 * Canceling timer's when we have no output being
8074 * paced. We also must remove ourselves from the
8075 * hpts.
8076 */
8077 tcp_hpts_remove(rack->rc_tp);
8078 hpts_removed = 1;
8079 }
8080 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8081 }
8082 if (hpts_removed == 0)
8083 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8084 }
8085
8086 static int
rack_stopall(struct tcpcb * tp)8087 rack_stopall(struct tcpcb *tp)
8088 {
8089 struct tcp_rack *rack;
8090
8091 rack = (struct tcp_rack *)tp->t_fb_ptr;
8092 rack->t_timers_stopped = 1;
8093
8094 tcp_hpts_remove(tp);
8095
8096 return (0);
8097 }
8098
8099 static void
rack_stop_all_timers(struct tcpcb * tp,struct tcp_rack * rack)8100 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8101 {
8102 /*
8103 * Assure no timers are running.
8104 */
8105 if (tcp_timer_active(tp, TT_PERSIST)) {
8106 /* We enter in persists, set the flag appropriately */
8107 rack->rc_in_persist = 1;
8108 }
8109 if (tcp_in_hpts(rack->rc_tp)) {
8110 tcp_hpts_remove(rack->rc_tp);
8111 }
8112 }
8113
8114 static void
rack_update_rsm(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint64_t ts,uint32_t add_flag,int segsiz)8115 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8116 struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8117 {
8118 int32_t idx;
8119
8120 rsm->r_rtr_cnt++;
8121 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8122 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8123 rsm->r_flags |= RACK_OVERMAX;
8124 }
8125 rsm->r_act_rxt_cnt++;
8126 /* Peg the count/index */
8127 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8128 rsm->r_dupack = 0;
8129 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8130 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8131 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8132 }
8133 if (rsm->r_flags & RACK_WAS_LOST) {
8134 /*
8135 * We retransmitted it putting it back in flight
8136 * remove the lost desgination and reduce the
8137 * bytes considered lost.
8138 */
8139 rsm->r_flags &= ~RACK_WAS_LOST;
8140 KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8141 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
8142 if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8143 rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8144 else
8145 rack->r_ctl.rc_considered_lost = 0;
8146 }
8147 idx = rsm->r_rtr_cnt - 1;
8148 rsm->r_tim_lastsent[idx] = ts;
8149 /*
8150 * Here we don't add in the len of send, since its already
8151 * in snduna <->snd_max.
8152 */
8153 rsm->r_fas = ctf_flight_size(rack->rc_tp,
8154 rack->r_ctl.rc_sacked);
8155 if (rsm->r_flags & RACK_ACKED) {
8156 /* Problably MTU discovery messing with us */
8157 rsm->r_flags &= ~RACK_ACKED;
8158 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8159 }
8160 if (rsm->r_in_tmap) {
8161 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8162 rsm->r_in_tmap = 0;
8163 }
8164 /* Lets make sure it really is in or not the GP window */
8165 rack_mark_in_gp_win(tp, rsm);
8166 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8167 rsm->r_in_tmap = 1;
8168 rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8169 /* Take off the must retransmit flag, if its on */
8170 if (rsm->r_flags & RACK_MUST_RXT) {
8171 if (rack->r_must_retran)
8172 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8173 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8174 /*
8175 * We have retransmitted all we need. Clear
8176 * any must retransmit flags.
8177 */
8178 rack->r_must_retran = 0;
8179 rack->r_ctl.rc_out_at_rto = 0;
8180 }
8181 rsm->r_flags &= ~RACK_MUST_RXT;
8182 }
8183 /* Remove any collapsed flag */
8184 rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8185 if (rsm->r_flags & RACK_SACK_PASSED) {
8186 /* We have retransmitted due to the SACK pass */
8187 rsm->r_flags &= ~RACK_SACK_PASSED;
8188 rsm->r_flags |= RACK_WAS_SACKPASS;
8189 }
8190 }
8191
8192 static uint32_t
rack_update_entry(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint64_t ts,int32_t * lenp,uint32_t add_flag,int segsiz)8193 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8194 struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8195 {
8196 /*
8197 * We (re-)transmitted starting at rsm->r_start for some length
8198 * (possibly less than r_end.
8199 */
8200 struct rack_sendmap *nrsm;
8201 int insret __diagused;
8202 uint32_t c_end;
8203 int32_t len;
8204
8205 len = *lenp;
8206 c_end = rsm->r_start + len;
8207 if (SEQ_GEQ(c_end, rsm->r_end)) {
8208 /*
8209 * We retransmitted the whole piece or more than the whole
8210 * slopping into the next rsm.
8211 */
8212 rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8213 if (c_end == rsm->r_end) {
8214 *lenp = 0;
8215 return (0);
8216 } else {
8217 int32_t act_len;
8218
8219 /* Hangs over the end return whats left */
8220 act_len = rsm->r_end - rsm->r_start;
8221 *lenp = (len - act_len);
8222 return (rsm->r_end);
8223 }
8224 /* We don't get out of this block. */
8225 }
8226 /*
8227 * Here we retransmitted less than the whole thing which means we
8228 * have to split this into what was transmitted and what was not.
8229 */
8230 nrsm = rack_alloc_full_limit(rack);
8231 if (nrsm == NULL) {
8232 /*
8233 * We can't get memory, so lets not proceed.
8234 */
8235 *lenp = 0;
8236 return (0);
8237 }
8238 /*
8239 * So here we are going to take the original rsm and make it what we
8240 * retransmitted. nrsm will be the tail portion we did not
8241 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8242 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8243 * 1, 6 and the new piece will be 6, 11.
8244 */
8245 rack_clone_rsm(rack, nrsm, rsm, c_end);
8246 nrsm->r_dupack = 0;
8247 rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8248 #ifndef INVARIANTS
8249 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8250 #else
8251 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8252 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8253 nrsm, insret, rack, rsm);
8254 }
8255 #endif
8256 if (rsm->r_in_tmap) {
8257 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8258 nrsm->r_in_tmap = 1;
8259 }
8260 rsm->r_flags &= (~RACK_HAS_FIN);
8261 rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8262 /* Log a split of rsm into rsm and nrsm */
8263 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8264 *lenp = 0;
8265 return (0);
8266 }
8267
8268 static void
rack_log_output(struct tcpcb * tp,struct tcpopt * to,int32_t len,uint32_t seq_out,uint16_t th_flags,int32_t err,uint64_t cts,struct rack_sendmap * hintrsm,uint32_t add_flag,struct mbuf * s_mb,uint32_t s_moff,int hw_tls,int segsiz)8269 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8270 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8271 struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
8272 uint32_t s_moff, int hw_tls, int segsiz)
8273 {
8274 struct tcp_rack *rack;
8275 struct rack_sendmap *rsm, *nrsm;
8276 int insret __diagused;
8277
8278 register uint32_t snd_max, snd_una;
8279
8280 /*
8281 * Add to the RACK log of packets in flight or retransmitted. If
8282 * there is a TS option we will use the TS echoed, if not we will
8283 * grab a TS.
8284 *
8285 * Retransmissions will increment the count and move the ts to its
8286 * proper place. Note that if options do not include TS's then we
8287 * won't be able to effectively use the ACK for an RTT on a retran.
8288 *
8289 * Notes about r_start and r_end. Lets consider a send starting at
8290 * sequence 1 for 10 bytes. In such an example the r_start would be
8291 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8292 * This means that r_end is actually the first sequence for the next
8293 * slot (11).
8294 *
8295 */
8296 /*
8297 * If err is set what do we do XXXrrs? should we not add the thing?
8298 * -- i.e. return if err != 0 or should we pretend we sent it? --
8299 * i.e. proceed with add ** do this for now.
8300 */
8301 INP_WLOCK_ASSERT(tptoinpcb(tp));
8302 if (err)
8303 /*
8304 * We don't log errors -- we could but snd_max does not
8305 * advance in this case either.
8306 */
8307 return;
8308
8309 if (th_flags & TH_RST) {
8310 /*
8311 * We don't log resets and we return immediately from
8312 * sending
8313 */
8314 return;
8315 }
8316 rack = (struct tcp_rack *)tp->t_fb_ptr;
8317 snd_una = tp->snd_una;
8318 snd_max = tp->snd_max;
8319 if (th_flags & (TH_SYN | TH_FIN)) {
8320 /*
8321 * The call to rack_log_output is made before bumping
8322 * snd_max. This means we can record one extra byte on a SYN
8323 * or FIN if seq_out is adding more on and a FIN is present
8324 * (and we are not resending).
8325 */
8326 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8327 len++;
8328 if (th_flags & TH_FIN)
8329 len++;
8330 }
8331 if (SEQ_LEQ((seq_out + len), snd_una)) {
8332 /* Are sending an old segment to induce an ack (keep-alive)? */
8333 return;
8334 }
8335 if (SEQ_LT(seq_out, snd_una)) {
8336 /* huh? should we panic? */
8337 uint32_t end;
8338
8339 end = seq_out + len;
8340 seq_out = snd_una;
8341 if (SEQ_GEQ(end, seq_out))
8342 len = end - seq_out;
8343 else
8344 len = 0;
8345 }
8346 if (len == 0) {
8347 /* We don't log zero window probes */
8348 return;
8349 }
8350 if (IN_FASTRECOVERY(tp->t_flags)) {
8351 rack->r_ctl.rc_prr_out += len;
8352 }
8353 /* First question is it a retransmission or new? */
8354 if (seq_out == snd_max) {
8355 /* Its new */
8356 rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8357 again:
8358 rsm = rack_alloc(rack);
8359 if (rsm == NULL) {
8360 /*
8361 * Hmm out of memory and the tcb got destroyed while
8362 * we tried to wait.
8363 */
8364 return;
8365 }
8366 if (th_flags & TH_FIN) {
8367 rsm->r_flags = RACK_HAS_FIN|add_flag;
8368 } else {
8369 rsm->r_flags = add_flag;
8370 }
8371 if (hw_tls)
8372 rsm->r_hw_tls = 1;
8373 rsm->r_tim_lastsent[0] = cts;
8374 rsm->r_rtr_cnt = 1;
8375 rsm->r_act_rxt_cnt = 0;
8376 rsm->r_rtr_bytes = 0;
8377 if (th_flags & TH_SYN) {
8378 /* The data space is one beyond snd_una */
8379 rsm->r_flags |= RACK_HAS_SYN;
8380 }
8381 rsm->r_start = seq_out;
8382 rsm->r_end = rsm->r_start + len;
8383 rack_mark_in_gp_win(tp, rsm);
8384 rsm->r_dupack = 0;
8385 /*
8386 * save off the mbuf location that
8387 * sndmbuf_noadv returned (which is
8388 * where we started copying from)..
8389 */
8390 rsm->m = s_mb;
8391 rsm->soff = s_moff;
8392 /*
8393 * Here we do add in the len of send, since its not yet
8394 * reflected in in snduna <->snd_max
8395 */
8396 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8397 rack->r_ctl.rc_sacked) +
8398 (rsm->r_end - rsm->r_start));
8399 if ((rack->rc_initial_ss_comp == 0) &&
8400 (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
8401 rack->r_ctl.ss_hi_fs = rsm->r_fas;
8402 }
8403 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8404 if (rsm->m) {
8405 if (rsm->m->m_len <= rsm->soff) {
8406 /*
8407 * XXXrrs Question, will this happen?
8408 *
8409 * If sbsndptr is set at the correct place
8410 * then s_moff should always be somewhere
8411 * within rsm->m. But if the sbsndptr was
8412 * off then that won't be true. If it occurs
8413 * we need to walkout to the correct location.
8414 */
8415 struct mbuf *lm;
8416
8417 lm = rsm->m;
8418 while (lm->m_len <= rsm->soff) {
8419 rsm->soff -= lm->m_len;
8420 lm = lm->m_next;
8421 KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8422 __func__, rack, s_moff, s_mb, rsm->soff));
8423 }
8424 rsm->m = lm;
8425 }
8426 rsm->orig_m_len = rsm->m->m_len;
8427 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8428 } else {
8429 rsm->orig_m_len = 0;
8430 rsm->orig_t_space = 0;
8431 }
8432 rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8433 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8434 /* Log a new rsm */
8435 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8436 #ifndef INVARIANTS
8437 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
8438 #else
8439 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8440 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8441 nrsm, insret, rack, rsm);
8442 }
8443 #endif
8444 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8445 rsm->r_in_tmap = 1;
8446 if (rsm->r_flags & RACK_IS_PCM) {
8447 rack->r_ctl.pcm_i.send_time = cts;
8448 rack->r_ctl.pcm_i.eseq = rsm->r_end;
8449 /* First time through we set the start too */
8450 if (rack->pcm_in_progress == 0)
8451 rack->r_ctl.pcm_i.sseq = rsm->r_start;
8452 }
8453 /*
8454 * Special case detection, is there just a single
8455 * packet outstanding when we are not in recovery?
8456 *
8457 * If this is true mark it so.
8458 */
8459 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8460 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8461 struct rack_sendmap *prsm;
8462
8463 prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8464 if (prsm)
8465 prsm->r_one_out_nr = 1;
8466 }
8467 return;
8468 }
8469 /*
8470 * If we reach here its a retransmission and we need to find it.
8471 */
8472 more:
8473 if (hintrsm && (hintrsm->r_start == seq_out)) {
8474 rsm = hintrsm;
8475 hintrsm = NULL;
8476 } else {
8477 /* No hints sorry */
8478 rsm = NULL;
8479 }
8480 if ((rsm) && (rsm->r_start == seq_out)) {
8481 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8482 if (len == 0) {
8483 return;
8484 } else {
8485 goto more;
8486 }
8487 }
8488 /* Ok it was not the last pointer go through it the hard way. */
8489 refind:
8490 rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8491 if (rsm) {
8492 if (rsm->r_start == seq_out) {
8493 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8494 if (len == 0) {
8495 return;
8496 } else {
8497 goto refind;
8498 }
8499 }
8500 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8501 /* Transmitted within this piece */
8502 /*
8503 * Ok we must split off the front and then let the
8504 * update do the rest
8505 */
8506 nrsm = rack_alloc_full_limit(rack);
8507 if (nrsm == NULL) {
8508 rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8509 return;
8510 }
8511 /*
8512 * copy rsm to nrsm and then trim the front of rsm
8513 * to not include this part.
8514 */
8515 rack_clone_rsm(rack, nrsm, rsm, seq_out);
8516 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8517 #ifndef INVARIANTS
8518 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8519 #else
8520 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8521 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8522 nrsm, insret, rack, rsm);
8523 }
8524 #endif
8525 if (rsm->r_in_tmap) {
8526 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8527 nrsm->r_in_tmap = 1;
8528 }
8529 rsm->r_flags &= (~RACK_HAS_FIN);
8530 seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8531 if (len == 0) {
8532 return;
8533 } else if (len > 0)
8534 goto refind;
8535 }
8536 }
8537 /*
8538 * Hmm not found in map did they retransmit both old and on into the
8539 * new?
8540 */
8541 if (seq_out == tp->snd_max) {
8542 goto again;
8543 } else if (SEQ_LT(seq_out, tp->snd_max)) {
8544 #ifdef INVARIANTS
8545 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8546 seq_out, len, tp->snd_una, tp->snd_max);
8547 printf("Starting Dump of all rack entries\n");
8548 TQHASH_FOREACH(rsm, rack->r_ctl.tqh) {
8549 printf("rsm:%p start:%u end:%u\n",
8550 rsm, rsm->r_start, rsm->r_end);
8551 }
8552 printf("Dump complete\n");
8553 panic("seq_out not found rack:%p tp:%p",
8554 rack, tp);
8555 #endif
8556 } else {
8557 #ifdef INVARIANTS
8558 /*
8559 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8560 * flag)
8561 */
8562 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8563 seq_out, len, tp->snd_max, tp);
8564 #endif
8565 }
8566 }
8567
8568 /*
8569 * Record one of the RTT updates from an ack into
8570 * our sample structure.
8571 */
8572
8573 static void
tcp_rack_xmit_timer(struct tcp_rack * rack,int32_t rtt,uint32_t len,uint32_t us_rtt,int confidence,struct rack_sendmap * rsm,uint16_t rtrcnt)8574 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8575 int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8576 {
8577 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8578 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8579 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8580 }
8581 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8582 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8583 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8584 }
8585 if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8586 if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8587 rack->r_ctl.rc_gp_lowrtt = us_rtt;
8588 if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8589 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8590 }
8591 if ((confidence == 1) &&
8592 ((rsm == NULL) ||
8593 (rsm->r_just_ret) ||
8594 (rsm->r_one_out_nr &&
8595 len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8596 /*
8597 * If the rsm had a just return
8598 * hit it then we can't trust the
8599 * rtt measurement for buffer deterimination
8600 * Note that a confidence of 2, indicates
8601 * SACK'd which overrides the r_just_ret or
8602 * the r_one_out_nr. If it was a CUM-ACK and
8603 * we had only two outstanding, but get an
8604 * ack for only 1. Then that also lowers our
8605 * confidence.
8606 */
8607 confidence = 0;
8608 }
8609 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8610 (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8611 if (rack->r_ctl.rack_rs.confidence == 0) {
8612 /*
8613 * We take anything with no current confidence
8614 * saved.
8615 */
8616 rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8617 rack->r_ctl.rack_rs.confidence = confidence;
8618 rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8619 } else if (confidence != 0) {
8620 /*
8621 * Once we have a confident number,
8622 * we can update it with a smaller
8623 * value since this confident number
8624 * may include the DSACK time until
8625 * the next segment (the second one) arrived.
8626 */
8627 rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8628 rack->r_ctl.rack_rs.confidence = confidence;
8629 rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8630 }
8631 }
8632 rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8633 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8634 rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8635 rack->r_ctl.rack_rs.rs_rtt_cnt++;
8636 }
8637
8638 /*
8639 * Collect new round-trip time estimate
8640 * and update averages and current timeout.
8641 */
8642 static void
tcp_rack_xmit_timer_commit(struct tcp_rack * rack,struct tcpcb * tp)8643 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8644 {
8645 int32_t delta;
8646 int32_t rtt;
8647
8648 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8649 /* No valid sample */
8650 return;
8651 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8652 /* We are to use the lowest RTT seen in a single ack */
8653 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8654 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8655 /* We are to use the highest RTT seen in a single ack */
8656 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8657 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8658 /* We are to use the average RTT seen in a single ack */
8659 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8660 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8661 } else {
8662 #ifdef INVARIANTS
8663 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8664 #endif
8665 return;
8666 }
8667 if (rtt == 0)
8668 rtt = 1;
8669 if (rack->rc_gp_rtt_set == 0) {
8670 /*
8671 * With no RTT we have to accept
8672 * even one we are not confident of.
8673 */
8674 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8675 rack->rc_gp_rtt_set = 1;
8676 } else if (rack->r_ctl.rack_rs.confidence) {
8677 /* update the running gp srtt */
8678 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8679 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8680 }
8681 if (rack->r_ctl.rack_rs.confidence) {
8682 /*
8683 * record the low and high for highly buffered path computation,
8684 * we only do this if we are confident (not a retransmission).
8685 */
8686 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8687 rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8688 }
8689 if (rack->rc_highly_buffered == 0) {
8690 /*
8691 * Currently once we declare a path has
8692 * highly buffered there is no going
8693 * back, which may be a problem...
8694 */
8695 if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8696 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8697 rack->r_ctl.rc_highest_us_rtt,
8698 rack->r_ctl.rc_lowest_us_rtt,
8699 RACK_RTTS_SEEHBP);
8700 rack->rc_highly_buffered = 1;
8701 }
8702 }
8703 }
8704 if ((rack->r_ctl.rack_rs.confidence) ||
8705 (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8706 /*
8707 * If we are highly confident of it <or> it was
8708 * never retransmitted we accept it as the last us_rtt.
8709 */
8710 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8711 /* The lowest rtt can be set if its was not retransmited */
8712 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8713 rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8714 if (rack->r_ctl.rc_lowest_us_rtt == 0)
8715 rack->r_ctl.rc_lowest_us_rtt = 1;
8716 }
8717 }
8718 rack = (struct tcp_rack *)tp->t_fb_ptr;
8719 if (tp->t_srtt != 0) {
8720 /*
8721 * We keep a simple srtt in microseconds, like our rtt
8722 * measurement. We don't need to do any tricks with shifting
8723 * etc. Instead we just add in 1/8th of the new measurement
8724 * and subtract out 1/8 of the old srtt. We do the same with
8725 * the variance after finding the absolute value of the
8726 * difference between this sample and the current srtt.
8727 */
8728 delta = tp->t_srtt - rtt;
8729 /* Take off 1/8th of the current sRTT */
8730 tp->t_srtt -= (tp->t_srtt >> 3);
8731 /* Add in 1/8th of the new RTT just measured */
8732 tp->t_srtt += (rtt >> 3);
8733 if (tp->t_srtt <= 0)
8734 tp->t_srtt = 1;
8735 /* Now lets make the absolute value of the variance */
8736 if (delta < 0)
8737 delta = -delta;
8738 /* Subtract out 1/8th */
8739 tp->t_rttvar -= (tp->t_rttvar >> 3);
8740 /* Add in 1/8th of the new variance we just saw */
8741 tp->t_rttvar += (delta >> 3);
8742 if (tp->t_rttvar <= 0)
8743 tp->t_rttvar = 1;
8744 } else {
8745 /*
8746 * No rtt measurement yet - use the unsmoothed rtt. Set the
8747 * variance to half the rtt (so our first retransmit happens
8748 * at 3*rtt).
8749 */
8750 tp->t_srtt = rtt;
8751 tp->t_rttvar = rtt >> 1;
8752 }
8753 rack->rc_srtt_measure_made = 1;
8754 KMOD_TCPSTAT_INC(tcps_rttupdated);
8755 if (tp->t_rttupdated < UCHAR_MAX)
8756 tp->t_rttupdated++;
8757 #ifdef STATS
8758 if (rack_stats_gets_ms_rtt == 0) {
8759 /* Send in the microsecond rtt used for rxt timeout purposes */
8760 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8761 } else if (rack_stats_gets_ms_rtt == 1) {
8762 /* Send in the millisecond rtt used for rxt timeout purposes */
8763 int32_t ms_rtt;
8764
8765 /* Round up */
8766 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8767 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8768 } else if (rack_stats_gets_ms_rtt == 2) {
8769 /* Send in the millisecond rtt has close to the path RTT as we can get */
8770 int32_t ms_rtt;
8771
8772 /* Round up */
8773 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8774 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8775 } else {
8776 /* Send in the microsecond rtt has close to the path RTT as we can get */
8777 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8778 }
8779 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8780 #endif
8781 rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
8782 /*
8783 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8784 * way we do the smoothing, srtt and rttvar will each average +1/2
8785 * tick of bias. When we compute the retransmit timer, we want 1/2
8786 * tick of rounding and 1 extra tick because of +-1/2 tick
8787 * uncertainty in the firing of the timer. The bias will give us
8788 * exactly the 1.5 tick we need. But, because the bias is
8789 * statistical, we have to test that we don't drop below the minimum
8790 * feasible timer (which is 2 ticks).
8791 */
8792 tp->t_rxtshift = 0;
8793 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8794 max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8795 rack_log_rtt_sample(rack, rtt);
8796 tp->t_softerror = 0;
8797 }
8798
8799
8800 static void
rack_apply_updated_usrtt(struct tcp_rack * rack,uint32_t us_rtt,uint32_t us_cts)8801 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8802 {
8803 /*
8804 * Apply to filter the inbound us-rtt at us_cts.
8805 */
8806 uint32_t old_rtt;
8807
8808 old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8809 apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8810 us_rtt, us_cts);
8811 if (old_rtt > us_rtt) {
8812 /* We just hit a new lower rtt time */
8813 rack_log_rtt_shrinks(rack, us_cts, old_rtt,
8814 __LINE__, RACK_RTTS_NEWRTT);
8815 /*
8816 * Only count it if its lower than what we saw within our
8817 * calculated range.
8818 */
8819 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8820 if (rack_probertt_lower_within &&
8821 rack->rc_gp_dyn_mul &&
8822 (rack->use_fixed_rate == 0) &&
8823 (rack->rc_always_pace)) {
8824 /*
8825 * We are seeing a new lower rtt very close
8826 * to the time that we would have entered probe-rtt.
8827 * This is probably due to the fact that a peer flow
8828 * has entered probe-rtt. Lets go in now too.
8829 */
8830 uint32_t val;
8831
8832 val = rack_probertt_lower_within * rack_time_between_probertt;
8833 val /= 100;
8834 if ((rack->in_probe_rtt == 0) &&
8835 (rack->rc_skip_timely == 0) &&
8836 ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
8837 rack_enter_probertt(rack, us_cts);
8838 }
8839 }
8840 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8841 }
8842 }
8843 }
8844
8845 static int
rack_update_rtt(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,struct tcpopt * to,uint32_t cts,int32_t ack_type,tcp_seq th_ack)8846 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8847 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8848 {
8849 uint32_t us_rtt;
8850 int32_t i, all;
8851 uint32_t t, len_acked;
8852
8853 if ((rsm->r_flags & RACK_ACKED) ||
8854 (rsm->r_flags & RACK_WAS_ACKED))
8855 /* Already done */
8856 return (0);
8857 if (rsm->r_no_rtt_allowed) {
8858 /* Not allowed */
8859 return (0);
8860 }
8861 if (ack_type == CUM_ACKED) {
8862 if (SEQ_GT(th_ack, rsm->r_end)) {
8863 len_acked = rsm->r_end - rsm->r_start;
8864 all = 1;
8865 } else {
8866 len_acked = th_ack - rsm->r_start;
8867 all = 0;
8868 }
8869 } else {
8870 len_acked = rsm->r_end - rsm->r_start;
8871 all = 0;
8872 }
8873 if (rsm->r_rtr_cnt == 1) {
8874
8875 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8876 if ((int)t <= 0)
8877 t = 1;
8878 if (!tp->t_rttlow || tp->t_rttlow > t)
8879 tp->t_rttlow = t;
8880 if (!rack->r_ctl.rc_rack_min_rtt ||
8881 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8882 rack->r_ctl.rc_rack_min_rtt = t;
8883 if (rack->r_ctl.rc_rack_min_rtt == 0) {
8884 rack->r_ctl.rc_rack_min_rtt = 1;
8885 }
8886 }
8887 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8888 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8889 else
8890 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8891 if (us_rtt == 0)
8892 us_rtt = 1;
8893 if (CC_ALGO(tp)->rttsample != NULL) {
8894 /* Kick the RTT to the CC */
8895 CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8896 }
8897 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8898 if (ack_type == SACKED) {
8899 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8900 tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8901 } else {
8902 /*
8903 * We need to setup what our confidence
8904 * is in this ack.
8905 *
8906 * If the rsm was app limited and it is
8907 * less than a mss in length (the end
8908 * of the send) then we have a gap. If we
8909 * were app limited but say we were sending
8910 * multiple MSS's then we are more confident
8911 * int it.
8912 *
8913 * When we are not app-limited then we see if
8914 * the rsm is being included in the current
8915 * measurement, we tell this by the app_limited_needs_set
8916 * flag.
8917 *
8918 * Note that being cwnd blocked is not applimited
8919 * as well as the pacing delay between packets which
8920 * are sending only 1 or 2 MSS's also will show up
8921 * in the RTT. We probably need to examine this algorithm
8922 * a bit more and enhance it to account for the delay
8923 * between rsm's. We could do that by saving off the
8924 * pacing delay of each rsm (in an rsm) and then
8925 * factoring that in somehow though for now I am
8926 * not sure how :)
8927 */
8928 int calc_conf = 0;
8929
8930 if (rsm->r_flags & RACK_APP_LIMITED) {
8931 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8932 calc_conf = 0;
8933 else
8934 calc_conf = 1;
8935 } else if (rack->app_limited_needs_set == 0) {
8936 calc_conf = 1;
8937 } else {
8938 calc_conf = 0;
8939 }
8940 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8941 tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8942 calc_conf, rsm, rsm->r_rtr_cnt);
8943 }
8944 if ((rsm->r_flags & RACK_TLP) &&
8945 (!IN_FASTRECOVERY(tp->t_flags))) {
8946 /* Segment was a TLP and our retrans matched */
8947 if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8948 rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
8949 }
8950 }
8951 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8952 (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8953 (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8954 /* New more recent rack_tmit_time */
8955 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8956 if (rack->r_ctl.rc_rack_tmit_time == 0)
8957 rack->r_ctl.rc_rack_tmit_time = 1;
8958 rack->rc_rack_rtt = t;
8959 }
8960 return (1);
8961 }
8962 /*
8963 * We clear the soft/rxtshift since we got an ack.
8964 * There is no assurance we will call the commit() function
8965 * so we need to clear these to avoid incorrect handling.
8966 */
8967 tp->t_rxtshift = 0;
8968 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8969 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8970 tp->t_softerror = 0;
8971 if (to && (to->to_flags & TOF_TS) &&
8972 (ack_type == CUM_ACKED) &&
8973 (to->to_tsecr) &&
8974 ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8975 /*
8976 * Now which timestamp does it match? In this block the ACK
8977 * must be coming from a previous transmission.
8978 */
8979 for (i = 0; i < rsm->r_rtr_cnt; i++) {
8980 if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8981 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8982 if ((int)t <= 0)
8983 t = 1;
8984 if (CC_ALGO(tp)->rttsample != NULL) {
8985 /*
8986 * Kick the RTT to the CC, here
8987 * we lie a bit in that we know the
8988 * retransmission is correct even though
8989 * we retransmitted. This is because
8990 * we match the timestamps.
8991 */
8992 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8993 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8994 else
8995 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8996 CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8997 }
8998 if ((i + 1) < rsm->r_rtr_cnt) {
8999 /*
9000 * The peer ack'd from our previous
9001 * transmission. We have a spurious
9002 * retransmission and thus we dont
9003 * want to update our rack_rtt.
9004 *
9005 * Hmm should there be a CC revert here?
9006 *
9007 */
9008 return (0);
9009 }
9010 if (!tp->t_rttlow || tp->t_rttlow > t)
9011 tp->t_rttlow = t;
9012 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9013 rack->r_ctl.rc_rack_min_rtt = t;
9014 if (rack->r_ctl.rc_rack_min_rtt == 0) {
9015 rack->r_ctl.rc_rack_min_rtt = 1;
9016 }
9017 }
9018 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9019 (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9020 (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9021 /* New more recent rack_tmit_time */
9022 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9023 if (rack->r_ctl.rc_rack_tmit_time == 0)
9024 rack->r_ctl.rc_rack_tmit_time = 1;
9025 rack->rc_rack_rtt = t;
9026 }
9027 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9028 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9029 rsm->r_rtr_cnt);
9030 return (1);
9031 }
9032 }
9033 /* If we are logging log out the sendmap */
9034 if (tcp_bblogging_on(rack->rc_tp)) {
9035 for (i = 0; i < rsm->r_rtr_cnt; i++) {
9036 rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9037 }
9038 }
9039 goto ts_not_found;
9040 } else {
9041 /*
9042 * Ok its a SACK block that we retransmitted. or a windows
9043 * machine without timestamps. We can tell nothing from the
9044 * time-stamp since its not there or the time the peer last
9045 * received a segment that moved forward its cum-ack point.
9046 */
9047 ts_not_found:
9048 i = rsm->r_rtr_cnt - 1;
9049 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9050 if ((int)t <= 0)
9051 t = 1;
9052 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9053 /*
9054 * We retransmitted and the ack came back in less
9055 * than the smallest rtt we have observed. We most
9056 * likely did an improper retransmit as outlined in
9057 * 6.2 Step 2 point 2 in the rack-draft so we
9058 * don't want to update our rack_rtt. We in
9059 * theory (in future) might want to think about reverting our
9060 * cwnd state but we won't for now.
9061 */
9062 return (0);
9063 } else if (rack->r_ctl.rc_rack_min_rtt) {
9064 /*
9065 * We retransmitted it and the retransmit did the
9066 * job.
9067 */
9068 if (!rack->r_ctl.rc_rack_min_rtt ||
9069 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9070 rack->r_ctl.rc_rack_min_rtt = t;
9071 if (rack->r_ctl.rc_rack_min_rtt == 0) {
9072 rack->r_ctl.rc_rack_min_rtt = 1;
9073 }
9074 }
9075 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9076 (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9077 (uint32_t)rsm->r_tim_lastsent[i]))) {
9078 /* New more recent rack_tmit_time */
9079 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9080 if (rack->r_ctl.rc_rack_tmit_time == 0)
9081 rack->r_ctl.rc_rack_tmit_time = 1;
9082 rack->rc_rack_rtt = t;
9083 }
9084 return (1);
9085 }
9086 }
9087 return (0);
9088 }
9089
9090 /*
9091 * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9092 */
9093 static void
rack_log_sack_passed(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t cts)9094 rack_log_sack_passed(struct tcpcb *tp,
9095 struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9096 {
9097 struct rack_sendmap *nrsm;
9098 uint32_t thresh;
9099
9100 /* Get our rxt threshold for lost consideration */
9101 thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9102 /* Now start looking at rsm's */
9103 nrsm = rsm;
9104 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9105 rack_head, r_tnext) {
9106 if (nrsm == rsm) {
9107 /* Skip original segment he is acked */
9108 continue;
9109 }
9110 if (nrsm->r_flags & RACK_ACKED) {
9111 /*
9112 * Skip ack'd segments, though we
9113 * should not see these, since tmap
9114 * should not have ack'd segments.
9115 */
9116 continue;
9117 }
9118 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9119 /*
9120 * If the peer dropped the rwnd on
9121 * these then we don't worry about them.
9122 */
9123 continue;
9124 }
9125 /* Check lost state */
9126 if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9127 uint32_t exp;
9128
9129 exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9130 if (TSTMP_LT(exp, cts) || (exp == cts)) {
9131 /* We consider it lost */
9132 nrsm->r_flags |= RACK_WAS_LOST;
9133 rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9134 }
9135 }
9136 if (nrsm->r_flags & RACK_SACK_PASSED) {
9137 /*
9138 * We found one that is already marked
9139 * passed, we have been here before and
9140 * so all others below this are marked.
9141 */
9142 break;
9143 }
9144 nrsm->r_flags |= RACK_SACK_PASSED;
9145 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9146 }
9147 }
9148
9149 static void
rack_need_set_test(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,tcp_seq th_ack,int line,int use_which)9150 rack_need_set_test(struct tcpcb *tp,
9151 struct tcp_rack *rack,
9152 struct rack_sendmap *rsm,
9153 tcp_seq th_ack,
9154 int line,
9155 int use_which)
9156 {
9157 struct rack_sendmap *s_rsm;
9158
9159 if ((tp->t_flags & TF_GPUTINPROG) &&
9160 SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9161 /*
9162 * We were app limited, and this ack
9163 * butts up or goes beyond the point where we want
9164 * to start our next measurement. We need
9165 * to record the new gput_ts as here and
9166 * possibly update the start sequence.
9167 */
9168 uint32_t seq, ts;
9169
9170 if (rsm->r_rtr_cnt > 1) {
9171 /*
9172 * This is a retransmit, can we
9173 * really make any assessment at this
9174 * point? We are not really sure of
9175 * the timestamp, is it this or the
9176 * previous transmission?
9177 *
9178 * Lets wait for something better that
9179 * is not retransmitted.
9180 */
9181 return;
9182 }
9183 seq = tp->gput_seq;
9184 ts = tp->gput_ts;
9185 rack->app_limited_needs_set = 0;
9186 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9187 /* Do we start at a new end? */
9188 if ((use_which == RACK_USE_BEG) &&
9189 SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9190 /*
9191 * When we get an ACK that just eats
9192 * up some of the rsm, we set RACK_USE_BEG
9193 * since whats at r_start (i.e. th_ack)
9194 * is left unacked and thats where the
9195 * measurement now starts.
9196 */
9197 tp->gput_seq = rsm->r_start;
9198 }
9199 if ((use_which == RACK_USE_END) &&
9200 SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9201 /*
9202 * We use the end when the cumack
9203 * is moving forward and completely
9204 * deleting the rsm passed so basically
9205 * r_end holds th_ack.
9206 *
9207 * For SACK's we also want to use the end
9208 * since this piece just got sacked and
9209 * we want to target anything after that
9210 * in our measurement.
9211 */
9212 tp->gput_seq = rsm->r_end;
9213 }
9214 if (use_which == RACK_USE_END_OR_THACK) {
9215 /*
9216 * special case for ack moving forward,
9217 * not a sack, we need to move all the
9218 * way up to where this ack cum-ack moves
9219 * to.
9220 */
9221 if (SEQ_GT(th_ack, rsm->r_end))
9222 tp->gput_seq = th_ack;
9223 else
9224 tp->gput_seq = rsm->r_end;
9225 }
9226 if (SEQ_LT(tp->gput_seq, tp->snd_max))
9227 s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9228 else
9229 s_rsm = NULL;
9230 /*
9231 * Pick up the correct send time if we can the rsm passed in
9232 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9233 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9234 * find a different seq i.e. the next send up.
9235 *
9236 * If that has not been sent, s_rsm will be NULL and we must
9237 * arrange it so this function will get called again by setting
9238 * app_limited_needs_set.
9239 */
9240 if (s_rsm)
9241 rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9242 else {
9243 /* If we hit here we have to have *not* sent tp->gput_seq */
9244 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9245 /* Set it up so we will go through here again */
9246 rack->app_limited_needs_set = 1;
9247 }
9248 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9249 /*
9250 * We moved beyond this guy's range, re-calculate
9251 * the new end point.
9252 */
9253 if (rack->rc_gp_filled == 0) {
9254 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9255 } else {
9256 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9257 }
9258 }
9259 /*
9260 * We are moving the goal post, we may be able to clear the
9261 * measure_saw_probe_rtt flag.
9262 */
9263 if ((rack->in_probe_rtt == 0) &&
9264 (rack->measure_saw_probe_rtt) &&
9265 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9266 rack->measure_saw_probe_rtt = 0;
9267 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9268 seq, tp->gput_seq,
9269 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9270 (uint64_t)rack->r_ctl.rc_gp_output_ts),
9271 5, line, NULL, 0);
9272 if (rack->rc_gp_filled &&
9273 ((tp->gput_ack - tp->gput_seq) <
9274 max(rc_init_window(rack), (MIN_GP_WIN *
9275 ctf_fixed_maxseg(tp))))) {
9276 uint32_t ideal_amount;
9277
9278 ideal_amount = rack_get_measure_window(tp, rack);
9279 if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9280 /*
9281 * There is no sense of continuing this measurement
9282 * because its too small to gain us anything we
9283 * trust. Skip it and that way we can start a new
9284 * measurement quicker.
9285 */
9286 tp->t_flags &= ~TF_GPUTINPROG;
9287 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9288 0, 0,
9289 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9290 (uint64_t)rack->r_ctl.rc_gp_output_ts),
9291 6, __LINE__, NULL, 0);
9292 } else {
9293 /*
9294 * Reset the window further out.
9295 */
9296 tp->gput_ack = tp->gput_seq + ideal_amount;
9297 }
9298 }
9299 rack_tend_gp_marks(tp, rack);
9300 rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9301 }
9302 }
9303
9304 static inline int
is_rsm_inside_declared_tlp_block(struct tcp_rack * rack,struct rack_sendmap * rsm)9305 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9306 {
9307 if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9308 /* Behind our TLP definition or right at */
9309 return (0);
9310 }
9311 if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9312 /* The start is beyond or right at our end of TLP definition */
9313 return (0);
9314 }
9315 /* It has to be a sub-part of the original TLP recorded */
9316 return (1);
9317 }
9318
9319 static uint32_t
rack_proc_sack_blk(struct tcpcb * tp,struct tcp_rack * rack,struct sackblk * sack,struct tcpopt * to,struct rack_sendmap ** prsm,uint32_t cts,uint32_t segsiz)9320 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9321 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9322 uint32_t segsiz)
9323 {
9324 uint32_t start, end, changed = 0;
9325 struct rack_sendmap stack_map;
9326 struct rack_sendmap *rsm, *nrsm, *prev, *next;
9327 int insret __diagused;
9328 int32_t used_ref = 1;
9329 int can_use_hookery = 0;
9330
9331 start = sack->start;
9332 end = sack->end;
9333 rsm = *prsm;
9334
9335 do_rest_ofb:
9336 if ((rsm == NULL) ||
9337 (SEQ_LT(end, rsm->r_start)) ||
9338 (SEQ_GEQ(start, rsm->r_end)) ||
9339 (SEQ_LT(start, rsm->r_start))) {
9340 /*
9341 * We are not in the right spot,
9342 * find the correct spot in the tree.
9343 */
9344 used_ref = 0;
9345 rsm = tqhash_find(rack->r_ctl.tqh, start);
9346 }
9347 if (rsm == NULL) {
9348 /* TSNH */
9349 goto out;
9350 }
9351 /* Ok we have an ACK for some piece of this rsm */
9352 if (rsm->r_start != start) {
9353 if ((rsm->r_flags & RACK_ACKED) == 0) {
9354 /*
9355 * Before any splitting or hookery is
9356 * done is it a TLP of interest i.e. rxt?
9357 */
9358 if ((rsm->r_flags & RACK_TLP) &&
9359 (rsm->r_rtr_cnt > 1)) {
9360 /*
9361 * We are splitting a rxt TLP, check
9362 * if we need to save off the start/end
9363 */
9364 if (rack->rc_last_tlp_acked_set &&
9365 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9366 /*
9367 * We already turned this on since we are inside
9368 * the previous one was a partially sack now we
9369 * are getting another one (maybe all of it).
9370 *
9371 */
9372 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9373 /*
9374 * Lets make sure we have all of it though.
9375 */
9376 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9377 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9378 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9379 rack->r_ctl.last_tlp_acked_end);
9380 }
9381 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9382 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9383 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9384 rack->r_ctl.last_tlp_acked_end);
9385 }
9386 } else {
9387 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9388 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9389 rack->rc_last_tlp_past_cumack = 0;
9390 rack->rc_last_tlp_acked_set = 1;
9391 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9392 }
9393 }
9394 /**
9395 * Need to split this in two pieces the before and after,
9396 * the before remains in the map, the after must be
9397 * added. In other words we have:
9398 * rsm |--------------|
9399 * sackblk |------->
9400 * rsm will become
9401 * rsm |---|
9402 * and nrsm will be the sacked piece
9403 * nrsm |----------|
9404 *
9405 * But before we start down that path lets
9406 * see if the sack spans over on top of
9407 * the next guy and it is already sacked.
9408 *
9409 */
9410 /*
9411 * Hookery can only be used if the two entries
9412 * are in the same bucket and neither one of
9413 * them staddle the bucket line.
9414 */
9415 next = tqhash_next(rack->r_ctl.tqh, rsm);
9416 if (next &&
9417 (rsm->bindex == next->bindex) &&
9418 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9419 ((next->r_flags & RACK_STRADDLE) == 0) &&
9420 ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9421 ((next->r_flags & RACK_IS_PCM) == 0) &&
9422 (rsm->r_flags & RACK_IN_GP_WIN) &&
9423 (next->r_flags & RACK_IN_GP_WIN))
9424 can_use_hookery = 1;
9425 else
9426 can_use_hookery = 0;
9427 if (next && can_use_hookery &&
9428 (next->r_flags & RACK_ACKED) &&
9429 SEQ_GEQ(end, next->r_start)) {
9430 /**
9431 * So the next one is already acked, and
9432 * we can thus by hookery use our stack_map
9433 * to reflect the piece being sacked and
9434 * then adjust the two tree entries moving
9435 * the start and ends around. So we start like:
9436 * rsm |------------| (not-acked)
9437 * next |-----------| (acked)
9438 * sackblk |-------->
9439 * We want to end like so:
9440 * rsm |------| (not-acked)
9441 * next |-----------------| (acked)
9442 * nrsm |-----|
9443 * Where nrsm is a temporary stack piece we
9444 * use to update all the gizmos.
9445 */
9446 /* Copy up our fudge block */
9447 nrsm = &stack_map;
9448 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9449 /* Now adjust our tree blocks */
9450 tqhash_update_end(rack->r_ctl.tqh, rsm, start);
9451 next->r_start = start;
9452 rsm->r_flags |= RACK_SHUFFLED;
9453 next->r_flags |= RACK_SHUFFLED;
9454 /* Now we must adjust back where next->m is */
9455 rack_setup_offset_for_rsm(rack, rsm, next);
9456 /*
9457 * Which timestamp do we keep? It is rather
9458 * important in GP measurements to have the
9459 * accurate end of the send window.
9460 *
9461 * We keep the largest value, which is the newest
9462 * send. We do this in case a segment that is
9463 * joined together and not part of a GP estimate
9464 * later gets expanded into the GP estimate.
9465 *
9466 * We prohibit the merging of unlike kinds i.e.
9467 * all pieces that are in the GP estimate can be
9468 * merged and all pieces that are not in a GP estimate
9469 * can be merged, but not disimilar pieces. Combine
9470 * this with taking the highest here and we should
9471 * be ok unless of course the client reneges. Then
9472 * all bets are off.
9473 */
9474 if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9475 nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9476 next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9477 /*
9478 * And we must keep the newest ack arrival time.
9479 */
9480 if (next->r_ack_arrival <
9481 rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9482 next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9483
9484
9485 /* We don't need to adjust rsm, it did not change */
9486 /* Clear out the dup ack count of the remainder */
9487 rsm->r_dupack = 0;
9488 rsm->r_just_ret = 0;
9489 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9490 /* Now lets make sure our fudge block is right */
9491 nrsm->r_start = start;
9492 /* Now lets update all the stats and such */
9493 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9494 if (rack->app_limited_needs_set)
9495 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9496 changed += (nrsm->r_end - nrsm->r_start);
9497 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9498 if (rsm->r_flags & RACK_WAS_LOST) {
9499 int my_chg;
9500
9501 my_chg = (nrsm->r_end - nrsm->r_start);
9502 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9503 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9504 if (my_chg <= rack->r_ctl.rc_considered_lost)
9505 rack->r_ctl.rc_considered_lost -= my_chg;
9506 else
9507 rack->r_ctl.rc_considered_lost = 0;
9508 }
9509 if (nrsm->r_flags & RACK_SACK_PASSED) {
9510 rack->r_ctl.rc_reorder_ts = cts;
9511 if (rack->r_ctl.rc_reorder_ts == 0)
9512 rack->r_ctl.rc_reorder_ts = 1;
9513 }
9514 /*
9515 * Now we want to go up from rsm (the
9516 * one left un-acked) to the next one
9517 * in the tmap. We do this so when
9518 * we walk backwards we include marking
9519 * sack-passed on rsm (The one passed in
9520 * is skipped since it is generally called
9521 * on something sacked before removing it
9522 * from the tmap).
9523 */
9524 if (rsm->r_in_tmap) {
9525 nrsm = TAILQ_NEXT(rsm, r_tnext);
9526 /*
9527 * Now that we have the next
9528 * one walk backwards from there.
9529 */
9530 if (nrsm && nrsm->r_in_tmap)
9531 rack_log_sack_passed(tp, rack, nrsm, cts);
9532 }
9533 /* Now are we done? */
9534 if (SEQ_LT(end, next->r_end) ||
9535 (end == next->r_end)) {
9536 /* Done with block */
9537 goto out;
9538 }
9539 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9540 counter_u64_add(rack_sack_used_next_merge, 1);
9541 /* Postion for the next block */
9542 start = next->r_end;
9543 rsm = tqhash_next(rack->r_ctl.tqh, next);
9544 if (rsm == NULL)
9545 goto out;
9546 } else {
9547 /**
9548 * We can't use any hookery here, so we
9549 * need to split the map. We enter like
9550 * so:
9551 * rsm |--------|
9552 * sackblk |----->
9553 * We will add the new block nrsm and
9554 * that will be the new portion, and then
9555 * fall through after reseting rsm. So we
9556 * split and look like this:
9557 * rsm |----|
9558 * sackblk |----->
9559 * nrsm |---|
9560 * We then fall through reseting
9561 * rsm to nrsm, so the next block
9562 * picks it up.
9563 */
9564 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9565 if (nrsm == NULL) {
9566 /*
9567 * failed XXXrrs what can we do but loose the sack
9568 * info?
9569 */
9570 goto out;
9571 }
9572 counter_u64_add(rack_sack_splits, 1);
9573 rack_clone_rsm(rack, nrsm, rsm, start);
9574 rsm->r_just_ret = 0;
9575 #ifndef INVARIANTS
9576 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9577 #else
9578 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9579 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9580 nrsm, insret, rack, rsm);
9581 }
9582 #endif
9583 if (rsm->r_in_tmap) {
9584 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9585 nrsm->r_in_tmap = 1;
9586 }
9587 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9588 rsm->r_flags &= (~RACK_HAS_FIN);
9589 /* Position us to point to the new nrsm that starts the sack blk */
9590 rsm = nrsm;
9591 }
9592 } else {
9593 /* Already sacked this piece */
9594 counter_u64_add(rack_sack_skipped_acked, 1);
9595 if (end == rsm->r_end) {
9596 /* Done with block */
9597 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9598 goto out;
9599 } else if (SEQ_LT(end, rsm->r_end)) {
9600 /* A partial sack to a already sacked block */
9601 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9602 goto out;
9603 } else {
9604 /*
9605 * The end goes beyond this guy
9606 * reposition the start to the
9607 * next block.
9608 */
9609 start = rsm->r_end;
9610 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9611 if (rsm == NULL)
9612 goto out;
9613 }
9614 }
9615 }
9616 if (SEQ_GEQ(end, rsm->r_end)) {
9617 /**
9618 * The end of this block is either beyond this guy or right
9619 * at this guy. I.e.:
9620 * rsm --- |-----|
9621 * end |-----|
9622 * <or>
9623 * end |---------|
9624 */
9625 if ((rsm->r_flags & RACK_ACKED) == 0) {
9626 /*
9627 * Is it a TLP of interest?
9628 */
9629 if ((rsm->r_flags & RACK_TLP) &&
9630 (rsm->r_rtr_cnt > 1)) {
9631 /*
9632 * We are splitting a rxt TLP, check
9633 * if we need to save off the start/end
9634 */
9635 if (rack->rc_last_tlp_acked_set &&
9636 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9637 /*
9638 * We already turned this on since we are inside
9639 * the previous one was a partially sack now we
9640 * are getting another one (maybe all of it).
9641 */
9642 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9643 /*
9644 * Lets make sure we have all of it though.
9645 */
9646 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9647 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9648 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9649 rack->r_ctl.last_tlp_acked_end);
9650 }
9651 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9652 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9653 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9654 rack->r_ctl.last_tlp_acked_end);
9655 }
9656 } else {
9657 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9658 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9659 rack->rc_last_tlp_past_cumack = 0;
9660 rack->rc_last_tlp_acked_set = 1;
9661 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9662 }
9663 }
9664 rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9665 changed += (rsm->r_end - rsm->r_start);
9666 /* You get a count for acking a whole segment or more */
9667 if (rsm->r_flags & RACK_WAS_LOST) {
9668 int my_chg;
9669
9670 my_chg = (rsm->r_end - rsm->r_start);
9671 rsm->r_flags &= ~RACK_WAS_LOST;
9672 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9673 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9674 if (my_chg <= rack->r_ctl.rc_considered_lost)
9675 rack->r_ctl.rc_considered_lost -= my_chg;
9676 else
9677 rack->r_ctl.rc_considered_lost = 0;
9678 }
9679 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9680 if (rsm->r_in_tmap) /* should be true */
9681 rack_log_sack_passed(tp, rack, rsm, cts);
9682 /* Is Reordering occuring? */
9683 if (rsm->r_flags & RACK_SACK_PASSED) {
9684 rsm->r_flags &= ~RACK_SACK_PASSED;
9685 rack->r_ctl.rc_reorder_ts = cts;
9686 if (rack->r_ctl.rc_reorder_ts == 0)
9687 rack->r_ctl.rc_reorder_ts = 1;
9688 }
9689 if (rack->app_limited_needs_set)
9690 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9691 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9692 rsm->r_flags |= RACK_ACKED;
9693 rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9694 if (rsm->r_in_tmap) {
9695 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9696 rsm->r_in_tmap = 0;
9697 }
9698 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9699 } else {
9700 counter_u64_add(rack_sack_skipped_acked, 1);
9701 }
9702 if (end == rsm->r_end) {
9703 /* This block only - done, setup for next */
9704 goto out;
9705 }
9706 /*
9707 * There is more not coverend by this rsm move on
9708 * to the next block in the tail queue hash table.
9709 */
9710 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9711 start = rsm->r_end;
9712 rsm = nrsm;
9713 if (rsm == NULL)
9714 goto out;
9715 goto do_rest_ofb;
9716 }
9717 /**
9718 * The end of this sack block is smaller than
9719 * our rsm i.e.:
9720 * rsm --- |-----|
9721 * end |--|
9722 */
9723 if ((rsm->r_flags & RACK_ACKED) == 0) {
9724 /*
9725 * Is it a TLP of interest?
9726 */
9727 if ((rsm->r_flags & RACK_TLP) &&
9728 (rsm->r_rtr_cnt > 1)) {
9729 /*
9730 * We are splitting a rxt TLP, check
9731 * if we need to save off the start/end
9732 */
9733 if (rack->rc_last_tlp_acked_set &&
9734 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9735 /*
9736 * We already turned this on since we are inside
9737 * the previous one was a partially sack now we
9738 * are getting another one (maybe all of it).
9739 */
9740 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9741 /*
9742 * Lets make sure we have all of it though.
9743 */
9744 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9745 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9746 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9747 rack->r_ctl.last_tlp_acked_end);
9748 }
9749 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9750 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9751 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9752 rack->r_ctl.last_tlp_acked_end);
9753 }
9754 } else {
9755 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9756 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9757 rack->rc_last_tlp_past_cumack = 0;
9758 rack->rc_last_tlp_acked_set = 1;
9759 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9760 }
9761 }
9762 /*
9763 * Hookery can only be used if the two entries
9764 * are in the same bucket and neither one of
9765 * them staddle the bucket line.
9766 */
9767 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9768 if (prev &&
9769 (rsm->bindex == prev->bindex) &&
9770 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9771 ((prev->r_flags & RACK_STRADDLE) == 0) &&
9772 ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9773 ((prev->r_flags & RACK_IS_PCM) == 0) &&
9774 (rsm->r_flags & RACK_IN_GP_WIN) &&
9775 (prev->r_flags & RACK_IN_GP_WIN))
9776 can_use_hookery = 1;
9777 else
9778 can_use_hookery = 0;
9779 if (prev && can_use_hookery &&
9780 (prev->r_flags & RACK_ACKED)) {
9781 /**
9782 * Goal, we want the right remainder of rsm to shrink
9783 * in place and span from (rsm->r_start = end) to rsm->r_end.
9784 * We want to expand prev to go all the way
9785 * to prev->r_end <- end.
9786 * so in the tree we have before:
9787 * prev |--------| (acked)
9788 * rsm |-------| (non-acked)
9789 * sackblk |-|
9790 * We churn it so we end up with
9791 * prev |----------| (acked)
9792 * rsm |-----| (non-acked)
9793 * nrsm |-| (temporary)
9794 *
9795 * Note if either prev/rsm is a TLP we don't
9796 * do this.
9797 */
9798 nrsm = &stack_map;
9799 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9800 tqhash_update_end(rack->r_ctl.tqh, prev, end);
9801 rsm->r_start = end;
9802 rsm->r_flags |= RACK_SHUFFLED;
9803 prev->r_flags |= RACK_SHUFFLED;
9804 /* Now adjust nrsm (stack copy) to be
9805 * the one that is the small
9806 * piece that was "sacked".
9807 */
9808 nrsm->r_end = end;
9809 rsm->r_dupack = 0;
9810 /*
9811 * Which timestamp do we keep? It is rather
9812 * important in GP measurements to have the
9813 * accurate end of the send window.
9814 *
9815 * We keep the largest value, which is the newest
9816 * send. We do this in case a segment that is
9817 * joined together and not part of a GP estimate
9818 * later gets expanded into the GP estimate.
9819 *
9820 * We prohibit the merging of unlike kinds i.e.
9821 * all pieces that are in the GP estimate can be
9822 * merged and all pieces that are not in a GP estimate
9823 * can be merged, but not disimilar pieces. Combine
9824 * this with taking the highest here and we should
9825 * be ok unless of course the client reneges. Then
9826 * all bets are off.
9827 */
9828 if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9829 nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9830 prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9831 }
9832 /*
9833 * And we must keep the newest ack arrival time.
9834 */
9835
9836 if(prev->r_ack_arrival <
9837 rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9838 prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9839
9840 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9841 /*
9842 * Now that the rsm has had its start moved forward
9843 * lets go ahead and get its new place in the world.
9844 */
9845 rack_setup_offset_for_rsm(rack, prev, rsm);
9846 /*
9847 * Now nrsm is our new little piece
9848 * that is acked (which was merged
9849 * to prev). Update the rtt and changed
9850 * based on that. Also check for reordering.
9851 */
9852 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9853 if (rack->app_limited_needs_set)
9854 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9855 changed += (nrsm->r_end - nrsm->r_start);
9856 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9857 if (rsm->r_flags & RACK_WAS_LOST) {
9858 int my_chg;
9859
9860 my_chg = (nrsm->r_end - nrsm->r_start);
9861 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9862 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9863 if (my_chg <= rack->r_ctl.rc_considered_lost)
9864 rack->r_ctl.rc_considered_lost -= my_chg;
9865 else
9866 rack->r_ctl.rc_considered_lost = 0;
9867 }
9868 if (nrsm->r_flags & RACK_SACK_PASSED) {
9869 rack->r_ctl.rc_reorder_ts = cts;
9870 if (rack->r_ctl.rc_reorder_ts == 0)
9871 rack->r_ctl.rc_reorder_ts = 1;
9872 }
9873 rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9874 rsm = prev;
9875 counter_u64_add(rack_sack_used_prev_merge, 1);
9876 } else {
9877 /**
9878 * This is the case where our previous
9879 * block is not acked either, so we must
9880 * split the block in two.
9881 */
9882 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9883 if (nrsm == NULL) {
9884 /* failed rrs what can we do but loose the sack info? */
9885 goto out;
9886 }
9887 if ((rsm->r_flags & RACK_TLP) &&
9888 (rsm->r_rtr_cnt > 1)) {
9889 /*
9890 * We are splitting a rxt TLP, check
9891 * if we need to save off the start/end
9892 */
9893 if (rack->rc_last_tlp_acked_set &&
9894 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9895 /*
9896 * We already turned this on since this block is inside
9897 * the previous one was a partially sack now we
9898 * are getting another one (maybe all of it).
9899 */
9900 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9901 /*
9902 * Lets make sure we have all of it though.
9903 */
9904 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9905 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9906 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9907 rack->r_ctl.last_tlp_acked_end);
9908 }
9909 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9910 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9911 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9912 rack->r_ctl.last_tlp_acked_end);
9913 }
9914 } else {
9915 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9916 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9917 rack->rc_last_tlp_acked_set = 1;
9918 rack->rc_last_tlp_past_cumack = 0;
9919 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9920 }
9921 }
9922 /**
9923 * In this case nrsm becomes
9924 * nrsm->r_start = end;
9925 * nrsm->r_end = rsm->r_end;
9926 * which is un-acked.
9927 * <and>
9928 * rsm->r_end = nrsm->r_start;
9929 * i.e. the remaining un-acked
9930 * piece is left on the left
9931 * hand side.
9932 *
9933 * So we start like this
9934 * rsm |----------| (not acked)
9935 * sackblk |---|
9936 * build it so we have
9937 * rsm |---| (acked)
9938 * nrsm |------| (not acked)
9939 */
9940 counter_u64_add(rack_sack_splits, 1);
9941 rack_clone_rsm(rack, nrsm, rsm, end);
9942 rsm->r_flags &= (~RACK_HAS_FIN);
9943 rsm->r_just_ret = 0;
9944 #ifndef INVARIANTS
9945 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9946 #else
9947 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9948 panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
9949 nrsm, insret, rack, rsm);
9950 }
9951 #endif
9952 if (rsm->r_in_tmap) {
9953 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9954 nrsm->r_in_tmap = 1;
9955 }
9956 nrsm->r_dupack = 0;
9957 rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
9958 rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9959 changed += (rsm->r_end - rsm->r_start);
9960 if (rsm->r_flags & RACK_WAS_LOST) {
9961 int my_chg;
9962
9963 my_chg = (rsm->r_end - rsm->r_start);
9964 rsm->r_flags &= ~RACK_WAS_LOST;
9965 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9966 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9967 if (my_chg <= rack->r_ctl.rc_considered_lost)
9968 rack->r_ctl.rc_considered_lost -= my_chg;
9969 else
9970 rack->r_ctl.rc_considered_lost = 0;
9971 }
9972 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9973
9974 if (rsm->r_in_tmap) /* should be true */
9975 rack_log_sack_passed(tp, rack, rsm, cts);
9976 /* Is Reordering occuring? */
9977 if (rsm->r_flags & RACK_SACK_PASSED) {
9978 rsm->r_flags &= ~RACK_SACK_PASSED;
9979 rack->r_ctl.rc_reorder_ts = cts;
9980 if (rack->r_ctl.rc_reorder_ts == 0)
9981 rack->r_ctl.rc_reorder_ts = 1;
9982 }
9983 if (rack->app_limited_needs_set)
9984 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9985 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9986 rsm->r_flags |= RACK_ACKED;
9987 rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9988 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
9989 if (rsm->r_in_tmap) {
9990 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9991 rsm->r_in_tmap = 0;
9992 }
9993 }
9994 } else if (start != end){
9995 /*
9996 * The block was already acked.
9997 */
9998 counter_u64_add(rack_sack_skipped_acked, 1);
9999 }
10000 out:
10001 if (rsm &&
10002 ((rsm->r_flags & RACK_TLP) == 0) &&
10003 (rsm->r_flags & RACK_ACKED)) {
10004 /*
10005 * Now can we merge where we worked
10006 * with either the previous or
10007 * next block?
10008 */
10009 next = tqhash_next(rack->r_ctl.tqh, rsm);
10010 while (next) {
10011 if (next->r_flags & RACK_TLP)
10012 break;
10013 /* Only allow merges between ones in or out of GP window */
10014 if ((next->r_flags & RACK_IN_GP_WIN) &&
10015 ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10016 break;
10017 }
10018 if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10019 ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10020 break;
10021 }
10022 if (rsm->bindex != next->bindex)
10023 break;
10024 if (rsm->r_flags & RACK_STRADDLE)
10025 break;
10026 if (rsm->r_flags & RACK_IS_PCM)
10027 break;
10028 if (next->r_flags & RACK_STRADDLE)
10029 break;
10030 if (next->r_flags & RACK_IS_PCM)
10031 break;
10032 if (next->r_flags & RACK_ACKED) {
10033 /* yep this and next can be merged */
10034 rsm = rack_merge_rsm(rack, rsm, next);
10035 next = tqhash_next(rack->r_ctl.tqh, rsm);
10036 } else
10037 break;
10038 }
10039 /* Now what about the previous? */
10040 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10041 while (prev) {
10042 if (prev->r_flags & RACK_TLP)
10043 break;
10044 /* Only allow merges between ones in or out of GP window */
10045 if ((prev->r_flags & RACK_IN_GP_WIN) &&
10046 ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10047 break;
10048 }
10049 if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10050 ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10051 break;
10052 }
10053 if (rsm->bindex != prev->bindex)
10054 break;
10055 if (rsm->r_flags & RACK_STRADDLE)
10056 break;
10057 if (rsm->r_flags & RACK_IS_PCM)
10058 break;
10059 if (prev->r_flags & RACK_STRADDLE)
10060 break;
10061 if (prev->r_flags & RACK_IS_PCM)
10062 break;
10063 if (prev->r_flags & RACK_ACKED) {
10064 /* yep the previous and this can be merged */
10065 rsm = rack_merge_rsm(rack, prev, rsm);
10066 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10067 } else
10068 break;
10069 }
10070 }
10071 if (used_ref == 0) {
10072 counter_u64_add(rack_sack_proc_all, 1);
10073 } else {
10074 counter_u64_add(rack_sack_proc_short, 1);
10075 }
10076 /* Save off the next one for quick reference. */
10077 nrsm = tqhash_find(rack->r_ctl.tqh, end);
10078 *prsm = rack->r_ctl.rc_sacklast = nrsm;
10079 return (changed);
10080 }
10081
10082 static void inline
rack_peer_reneges(struct tcp_rack * rack,struct rack_sendmap * rsm,tcp_seq th_ack)10083 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10084 {
10085 struct rack_sendmap *tmap;
10086
10087 tmap = NULL;
10088 while (rsm && (rsm->r_flags & RACK_ACKED)) {
10089 /* Its no longer sacked, mark it so */
10090 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10091 #ifdef INVARIANTS
10092 if (rsm->r_in_tmap) {
10093 panic("rack:%p rsm:%p flags:0x%x in tmap?",
10094 rack, rsm, rsm->r_flags);
10095 }
10096 #endif
10097 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10098 /* Rebuild it into our tmap */
10099 if (tmap == NULL) {
10100 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10101 tmap = rsm;
10102 } else {
10103 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10104 tmap = rsm;
10105 }
10106 tmap->r_in_tmap = 1;
10107 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10108 }
10109 /*
10110 * Now lets possibly clear the sack filter so we start
10111 * recognizing sacks that cover this area.
10112 */
10113 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10114
10115 }
10116
10117
10118 static void inline
rack_rsm_sender_update(struct tcp_rack * rack,struct tcpcb * tp,struct rack_sendmap * rsm,uint8_t from)10119 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10120 {
10121 /*
10122 * We look at advancing the end send time for our GP
10123 * measurement tracking only as the cumulative acknowledgment
10124 * moves forward. You might wonder about this, why not
10125 * at every transmission or retransmission within the
10126 * GP window update the rc_gp_cumack_ts? Well its rather
10127 * nuanced but basically the GP window *may* expand (as
10128 * it does below) or worse and harder to track it may shrink.
10129 *
10130 * This last makes it impossible to track at the time of
10131 * the send, since you may set forward your rc_gp_cumack_ts
10132 * when you send, because that send *is* in your currently
10133 * "guessed" window, but then it shrinks. Now which was
10134 * the send time of the last bytes in the window, by the
10135 * time you ask that question that part of the sendmap
10136 * is freed. So you don't know and you will have too
10137 * long of send window. Instead by updating the time
10138 * marker only when the cumack advances this assures us
10139 * that we will have only the sends in the window of our
10140 * GP measurement.
10141 *
10142 * Another complication from this is the
10143 * merging of sendmap entries. During SACK processing this
10144 * can happen to conserve the sendmap size. That breaks
10145 * everything down in tracking the send window of the GP
10146 * estimate. So to prevent that and keep it working with
10147 * a tiny bit more limited merging, we only allow like
10148 * types to be merged. I.e. if two sends are in the GP window
10149 * then its ok to merge them together. If two sends are not
10150 * in the GP window its ok to merge them together too. Though
10151 * one send in and one send out cannot be merged. We combine
10152 * this with never allowing the shrinking of the GP window when
10153 * we are in recovery so that we can properly calculate the
10154 * sending times.
10155 *
10156 * This all of course seems complicated, because it is.. :)
10157 *
10158 * The cum-ack is being advanced upon the sendmap.
10159 * If we are not doing a GP estimate don't
10160 * proceed.
10161 */
10162 uint64_t ts;
10163
10164 if ((tp->t_flags & TF_GPUTINPROG) == 0)
10165 return;
10166 /*
10167 * If this sendmap entry is going
10168 * beyond the measurement window we had picked,
10169 * expand the measurement window by that much.
10170 */
10171 if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10172 tp->gput_ack = rsm->r_end;
10173 }
10174 /*
10175 * If we have not setup a ack, then we
10176 * have no idea if the newly acked pieces
10177 * will be "in our seq measurement range". If
10178 * it is when we clear the app_limited_needs_set
10179 * flag the timestamp will be updated.
10180 */
10181 if (rack->app_limited_needs_set)
10182 return;
10183 /*
10184 * Finally, we grab out the latest timestamp
10185 * that this packet was sent and then see
10186 * if:
10187 * a) The packet touches are newly defined GP range.
10188 * b) The time is greater than (newer) than the
10189 * one we currently have. If so we update
10190 * our sending end time window.
10191 *
10192 * Note we *do not* do this at send time. The reason
10193 * is that if you do you *may* pick up a newer timestamp
10194 * for a range you are not going to measure. We project
10195 * out how far and then sometimes modify that to be
10196 * smaller. If that occurs then you will have a send
10197 * that does not belong to the range included.
10198 */
10199 if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10200 rack->r_ctl.rc_gp_cumack_ts)
10201 return;
10202 if (rack_in_gp_window(tp, rsm)) {
10203 rack->r_ctl.rc_gp_cumack_ts = ts;
10204 rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10205 __LINE__, from, rsm);
10206 }
10207 }
10208
10209 static void
rack_process_to_cumack(struct tcpcb * tp,struct tcp_rack * rack,register uint32_t th_ack,uint32_t cts,struct tcpopt * to,uint64_t acktime)10210 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to, uint64_t acktime)
10211 {
10212 struct rack_sendmap *rsm;
10213 /*
10214 * The ACK point is advancing to th_ack, we must drop off
10215 * the packets in the rack log and calculate any eligble
10216 * RTT's.
10217 */
10218
10219 if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
10220 /*
10221 * If we have some sack blocks in the filter
10222 * lets prune them out by calling sfb with no blocks.
10223 */
10224 sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
10225 }
10226 if (SEQ_GT(th_ack, tp->snd_una)) {
10227 /* Clear any app ack remembered settings */
10228 rack->r_ctl.cleared_app_ack = 0;
10229 }
10230 rack->r_wanted_output = 1;
10231 if (SEQ_GT(th_ack, tp->snd_una))
10232 rack->r_ctl.last_cumack_advance = acktime;
10233
10234 /* Tend any TLP that has been marked for 1/2 the seq space (its old) */
10235 if ((rack->rc_last_tlp_acked_set == 1)&&
10236 (rack->rc_last_tlp_past_cumack == 1) &&
10237 (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10238 /*
10239 * We have reached the point where our last rack
10240 * tlp retransmit sequence is ahead of the cum-ack.
10241 * This can only happen when the cum-ack moves all
10242 * the way around (its been a full 2^^31+1 bytes
10243 * or more since we sent a retransmitted TLP). Lets
10244 * turn off the valid flag since its not really valid.
10245 *
10246 * Note since sack's also turn on this event we have
10247 * a complication, we have to wait to age it out until
10248 * the cum-ack is by the TLP before checking which is
10249 * what the next else clause does.
10250 */
10251 rack_log_dsack_event(rack, 9, __LINE__,
10252 rack->r_ctl.last_tlp_acked_start,
10253 rack->r_ctl.last_tlp_acked_end);
10254 rack->rc_last_tlp_acked_set = 0;
10255 rack->rc_last_tlp_past_cumack = 0;
10256 } else if ((rack->rc_last_tlp_acked_set == 1) &&
10257 (rack->rc_last_tlp_past_cumack == 0) &&
10258 (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10259 /*
10260 * It is safe to start aging TLP's out.
10261 */
10262 rack->rc_last_tlp_past_cumack = 1;
10263 }
10264 /* We do the same for the tlp send seq as well */
10265 if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10266 (rack->rc_last_sent_tlp_past_cumack == 1) &&
10267 (SEQ_GT(rack->r_ctl.last_sent_tlp_seq, th_ack))) {
10268 rack_log_dsack_event(rack, 9, __LINE__,
10269 rack->r_ctl.last_sent_tlp_seq,
10270 (rack->r_ctl.last_sent_tlp_seq +
10271 rack->r_ctl.last_sent_tlp_len));
10272 rack->rc_last_sent_tlp_seq_valid = 0;
10273 rack->rc_last_sent_tlp_past_cumack = 0;
10274 } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10275 (rack->rc_last_sent_tlp_past_cumack == 0) &&
10276 (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10277 /*
10278 * It is safe to start aging TLP's send.
10279 */
10280 rack->rc_last_sent_tlp_past_cumack = 1;
10281 }
10282 more:
10283 rsm = tqhash_min(rack->r_ctl.tqh);
10284 if (rsm == NULL) {
10285 if ((th_ack - 1) == tp->iss) {
10286 /*
10287 * For the SYN incoming case we will not
10288 * have called tcp_output for the sending of
10289 * the SYN, so there will be no map. All
10290 * other cases should probably be a panic.
10291 */
10292 return;
10293 }
10294 if (tp->t_flags & TF_SENTFIN) {
10295 /* if we sent a FIN we often will not have map */
10296 return;
10297 }
10298 #ifdef INVARIANTS
10299 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
10300 tp,
10301 tp->t_state, th_ack, rack,
10302 tp->snd_una, tp->snd_max);
10303 #endif
10304 return;
10305 }
10306 if (SEQ_LT(th_ack, rsm->r_start)) {
10307 /* Huh map is missing this */
10308 #ifdef INVARIANTS
10309 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10310 rsm->r_start,
10311 th_ack, tp->t_state, rack->r_state);
10312 #endif
10313 return;
10314 }
10315 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10316
10317 /* Now was it a retransmitted TLP? */
10318 if ((rsm->r_flags & RACK_TLP) &&
10319 (rsm->r_rtr_cnt > 1)) {
10320 /*
10321 * Yes, this rsm was a TLP and retransmitted, remember that
10322 * since if a DSACK comes back on this we don't want
10323 * to think of it as a reordered segment. This may
10324 * get updated again with possibly even other TLPs
10325 * in flight, but thats ok. Only when we don't send
10326 * a retransmitted TLP for 1/2 the sequences space
10327 * will it get turned off (above).
10328 */
10329 if (rack->rc_last_tlp_acked_set &&
10330 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10331 /*
10332 * We already turned this on since the end matches,
10333 * the previous one was a partially ack now we
10334 * are getting another one (maybe all of it).
10335 */
10336 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10337 /*
10338 * Lets make sure we have all of it though.
10339 */
10340 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10341 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10342 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10343 rack->r_ctl.last_tlp_acked_end);
10344 }
10345 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10346 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10347 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10348 rack->r_ctl.last_tlp_acked_end);
10349 }
10350 } else {
10351 rack->rc_last_tlp_past_cumack = 1;
10352 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10353 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10354 rack->rc_last_tlp_acked_set = 1;
10355 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10356 }
10357 }
10358 /* Now do we consume the whole thing? */
10359 rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10360 if (SEQ_GEQ(th_ack, rsm->r_end)) {
10361 /* Its all consumed. */
10362 uint32_t left;
10363 uint8_t newly_acked;
10364
10365 if (rsm->r_flags & RACK_WAS_LOST) {
10366 /*
10367 * This can happen when we marked it as lost
10368 * and yet before retransmitting we get an ack
10369 * which can happen due to reordering.
10370 */
10371 rsm->r_flags &= ~RACK_WAS_LOST;
10372 KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
10373 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
10374 if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
10375 rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
10376 else
10377 rack->r_ctl.rc_considered_lost = 0;
10378 }
10379 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10380 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10381 rsm->r_rtr_bytes = 0;
10382 /*
10383 * Record the time of highest cumack sent if its in our measurement
10384 * window and possibly bump out the end.
10385 */
10386 rack_rsm_sender_update(rack, tp, rsm, 4);
10387 tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10388 if (rsm->r_in_tmap) {
10389 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10390 rsm->r_in_tmap = 0;
10391 }
10392 newly_acked = 1;
10393 if (rsm->r_flags & RACK_ACKED) {
10394 /*
10395 * It was acked on the scoreboard -- remove
10396 * it from total
10397 */
10398 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10399 newly_acked = 0;
10400 } else if (rsm->r_flags & RACK_SACK_PASSED) {
10401 /*
10402 * There are segments ACKED on the
10403 * scoreboard further up. We are seeing
10404 * reordering.
10405 */
10406 rsm->r_flags &= ~RACK_SACK_PASSED;
10407 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10408 rsm->r_flags |= RACK_ACKED;
10409 rack->r_ctl.rc_reorder_ts = cts;
10410 if (rack->r_ctl.rc_reorder_ts == 0)
10411 rack->r_ctl.rc_reorder_ts = 1;
10412 if (rack->r_ent_rec_ns) {
10413 /*
10414 * We have sent no more, and we saw an sack
10415 * then ack arrive.
10416 */
10417 rack->r_might_revert = 1;
10418 }
10419 rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10420 } else {
10421 rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10422 }
10423 if ((rsm->r_flags & RACK_TO_REXT) &&
10424 (tp->t_flags & TF_RCVD_TSTMP) &&
10425 (to->to_flags & TOF_TS) &&
10426 (to->to_tsecr != 0) &&
10427 (tp->t_flags & TF_PREVVALID)) {
10428 /*
10429 * We can use the timestamp to see
10430 * if this retransmission was from the
10431 * first transmit. If so we made a mistake.
10432 */
10433 tp->t_flags &= ~TF_PREVVALID;
10434 if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10435 /* The first transmit is what this ack is for */
10436 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10437 }
10438 }
10439 left = th_ack - rsm->r_end;
10440 if (rack->app_limited_needs_set && newly_acked)
10441 rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10442 /* Free back to zone */
10443 rack_free(rack, rsm);
10444 if (left) {
10445 goto more;
10446 }
10447 /* Check for reneging */
10448 rsm = tqhash_min(rack->r_ctl.tqh);
10449 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10450 /*
10451 * The peer has moved snd_una up to
10452 * the edge of this send, i.e. one
10453 * that it had previously acked. The only
10454 * way that can be true if the peer threw
10455 * away data (space issues) that it had
10456 * previously sacked (else it would have
10457 * given us snd_una up to (rsm->r_end).
10458 * We need to undo the acked markings here.
10459 *
10460 * Note we have to look to make sure th_ack is
10461 * our rsm->r_start in case we get an old ack
10462 * where th_ack is behind snd_una.
10463 */
10464 rack_peer_reneges(rack, rsm, th_ack);
10465 }
10466 return;
10467 }
10468 if (rsm->r_flags & RACK_ACKED) {
10469 /*
10470 * It was acked on the scoreboard -- remove it from
10471 * total for the part being cum-acked.
10472 */
10473 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10474 } else {
10475 rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
10476 }
10477 /* And what about the lost flag? */
10478 if (rsm->r_flags & RACK_WAS_LOST) {
10479 /*
10480 * This can happen when we marked it as lost
10481 * and yet before retransmitting we get an ack
10482 * which can happen due to reordering. In this
10483 * case its only a partial ack of the send.
10484 */
10485 KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
10486 ("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm, rack, th_ack));
10487 if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
10488 rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
10489 else
10490 rack->r_ctl.rc_considered_lost = 0;
10491 }
10492 /*
10493 * Clear the dup ack count for
10494 * the piece that remains.
10495 */
10496 rsm->r_dupack = 0;
10497 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10498 if (rsm->r_rtr_bytes) {
10499 /*
10500 * It was retransmitted adjust the
10501 * sack holes for what was acked.
10502 */
10503 int ack_am;
10504
10505 ack_am = (th_ack - rsm->r_start);
10506 if (ack_am >= rsm->r_rtr_bytes) {
10507 rack->r_ctl.rc_holes_rxt -= ack_am;
10508 rsm->r_rtr_bytes -= ack_am;
10509 }
10510 }
10511 /*
10512 * Update where the piece starts and record
10513 * the time of send of highest cumack sent if
10514 * its in our GP range.
10515 */
10516 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10517 /* Now we need to move our offset forward too */
10518 if (rsm->m &&
10519 ((rsm->orig_m_len != rsm->m->m_len) ||
10520 (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10521 /* Fix up the orig_m_len and possibly the mbuf offset */
10522 rack_adjust_orig_mlen(rsm);
10523 }
10524 rsm->soff += (th_ack - rsm->r_start);
10525 rack_rsm_sender_update(rack, tp, rsm, 5);
10526 /* The trim will move th_ack into r_start for us */
10527 tqhash_trim(rack->r_ctl.tqh, th_ack);
10528 /* Now do we need to move the mbuf fwd too? */
10529 {
10530 struct mbuf *m;
10531 uint32_t soff;
10532
10533 m = rsm->m;
10534 soff = rsm->soff;
10535 if (m) {
10536 while (soff >= m->m_len) {
10537 soff -= m->m_len;
10538 KASSERT((m->m_next != NULL),
10539 (" rsm:%p off:%u soff:%u m:%p",
10540 rsm, rsm->soff, soff, m));
10541 m = m->m_next;
10542 if (m == NULL) {
10543 /*
10544 * This is a fall-back that prevents a panic. In reality
10545 * we should be able to walk the mbuf's and find our place.
10546 * At this point snd_una has not been updated with the sbcut() yet
10547 * but tqhash_trim did update rsm->r_start so the offset calcuation
10548 * should work fine. This is undesirable since we will take cache
10549 * hits to access the socket buffer. And even more puzzling is that
10550 * it happens occasionally. It should not :(
10551 */
10552 m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10553 (rsm->r_start - tp->snd_una),
10554 &soff);
10555 break;
10556 }
10557 }
10558 /*
10559 * Now save in our updated values.
10560 */
10561 rsm->m = m;
10562 rsm->soff = soff;
10563 rsm->orig_m_len = rsm->m->m_len;
10564 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10565 }
10566 }
10567 if (rack->app_limited_needs_set &&
10568 SEQ_GEQ(th_ack, tp->gput_seq))
10569 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10570 }
10571
10572 static void
rack_handle_might_revert(struct tcpcb * tp,struct tcp_rack * rack)10573 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10574 {
10575 struct rack_sendmap *rsm;
10576 int sack_pass_fnd = 0;
10577
10578 if (rack->r_might_revert) {
10579 /*
10580 * Ok we have reordering, have not sent anything, we
10581 * might want to revert the congestion state if nothing
10582 * further has SACK_PASSED on it. Lets check.
10583 *
10584 * We also get here when we have DSACKs come in for
10585 * all the data that we FR'd. Note that a rxt or tlp
10586 * timer clears this from happening.
10587 */
10588
10589 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10590 if (rsm->r_flags & RACK_SACK_PASSED) {
10591 sack_pass_fnd = 1;
10592 break;
10593 }
10594 }
10595 if (sack_pass_fnd == 0) {
10596 /*
10597 * We went into recovery
10598 * incorrectly due to reordering!
10599 */
10600 int orig_cwnd;
10601
10602 rack->r_ent_rec_ns = 0;
10603 orig_cwnd = tp->snd_cwnd;
10604 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10605 tp->snd_recover = tp->snd_una;
10606 rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10607 if (IN_RECOVERY(tp->t_flags)) {
10608 rack_exit_recovery(tp, rack, 3);
10609 if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
10610 /*
10611 * We were in recovery, had an RTO
10612 * and then re-entered recovery (more sack's arrived)
10613 * and we have properly recorded the old ssthresh from
10614 * the first recovery. We want to be able to slow-start
10615 * back to this level. The ssthresh from the timeout
10616 * and then back into recovery will end up most likely
10617 * to be min(cwnd=1mss, 2mss). Which makes it basically
10618 * so we get no slow-start after our RTO.
10619 */
10620 rack->rto_from_rec = 0;
10621 if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
10622 tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
10623 }
10624 }
10625 }
10626 rack->r_might_revert = 0;
10627 }
10628 }
10629
10630
10631 static int
rack_note_dsack(struct tcp_rack * rack,tcp_seq start,tcp_seq end)10632 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10633 {
10634
10635 uint32_t am, l_end;
10636 int was_tlp = 0;
10637
10638 if (SEQ_GT(end, start))
10639 am = end - start;
10640 else
10641 am = 0;
10642 if ((rack->rc_last_tlp_acked_set ) &&
10643 (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10644 (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10645 /*
10646 * The DSACK is because of a TLP which we don't
10647 * do anything with the reordering window over since
10648 * it was not reordering that caused the DSACK but
10649 * our previous retransmit TLP.
10650 */
10651 rack_log_dsack_event(rack, 7, __LINE__, start, end);
10652 was_tlp = 1;
10653 goto skip_dsack_round;
10654 }
10655 if (rack->rc_last_sent_tlp_seq_valid) {
10656 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10657 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10658 (SEQ_LEQ(end, l_end))) {
10659 /*
10660 * This dsack is from the last sent TLP, ignore it
10661 * for reordering purposes.
10662 */
10663 rack_log_dsack_event(rack, 7, __LINE__, start, end);
10664 was_tlp = 1;
10665 goto skip_dsack_round;
10666 }
10667 }
10668 if (rack->rc_dsack_round_seen == 0) {
10669 rack->rc_dsack_round_seen = 1;
10670 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10671 rack->r_ctl.num_dsack++;
10672 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
10673 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10674 }
10675 skip_dsack_round:
10676 /*
10677 * We keep track of how many DSACK blocks we get
10678 * after a recovery incident.
10679 */
10680 rack->r_ctl.dsack_byte_cnt += am;
10681 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10682 rack->r_ctl.retran_during_recovery &&
10683 (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10684 /*
10685 * False recovery most likely culprit is reordering. If
10686 * nothing else is missing we need to revert.
10687 */
10688 rack->r_might_revert = 1;
10689 rack_handle_might_revert(rack->rc_tp, rack);
10690 rack->r_might_revert = 0;
10691 rack->r_ctl.retran_during_recovery = 0;
10692 rack->r_ctl.dsack_byte_cnt = 0;
10693 }
10694 return (was_tlp);
10695 }
10696
10697 static uint32_t
do_rack_compute_pipe(struct tcpcb * tp,struct tcp_rack * rack,uint32_t snd_una)10698 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10699 {
10700 return (((tp->snd_max - snd_una) -
10701 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
10702 }
10703
10704 static int32_t
rack_compute_pipe(struct tcpcb * tp)10705 rack_compute_pipe(struct tcpcb *tp)
10706 {
10707 return ((int32_t)do_rack_compute_pipe(tp,
10708 (struct tcp_rack *)tp->t_fb_ptr,
10709 tp->snd_una));
10710 }
10711
10712 static void
rack_update_prr(struct tcpcb * tp,struct tcp_rack * rack,uint32_t changed,tcp_seq th_ack)10713 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10714 {
10715 /* Deal with changed and PRR here (in recovery only) */
10716 uint32_t pipe, snd_una;
10717
10718 rack->r_ctl.rc_prr_delivered += changed;
10719
10720 if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10721 /*
10722 * It is all outstanding, we are application limited
10723 * and thus we don't need more room to send anything.
10724 * Note we use tp->snd_una here and not th_ack because
10725 * the data as yet not been cut from the sb.
10726 */
10727 rack->r_ctl.rc_prr_sndcnt = 0;
10728 return;
10729 }
10730 /* Compute prr_sndcnt */
10731 if (SEQ_GT(tp->snd_una, th_ack)) {
10732 snd_una = tp->snd_una;
10733 } else {
10734 snd_una = th_ack;
10735 }
10736 pipe = do_rack_compute_pipe(tp, rack, snd_una);
10737 if (pipe > tp->snd_ssthresh) {
10738 long sndcnt;
10739
10740 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10741 if (rack->r_ctl.rc_prr_recovery_fs > 0)
10742 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10743 else {
10744 rack->r_ctl.rc_prr_sndcnt = 0;
10745 rack_log_to_prr(rack, 9, 0, __LINE__);
10746 sndcnt = 0;
10747 }
10748 sndcnt++;
10749 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10750 sndcnt -= rack->r_ctl.rc_prr_out;
10751 else
10752 sndcnt = 0;
10753 rack->r_ctl.rc_prr_sndcnt = sndcnt;
10754 rack_log_to_prr(rack, 10, 0, __LINE__);
10755 } else {
10756 uint32_t limit;
10757
10758 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10759 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10760 else
10761 limit = 0;
10762 if (changed > limit)
10763 limit = changed;
10764 limit += ctf_fixed_maxseg(tp);
10765 if (tp->snd_ssthresh > pipe) {
10766 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10767 rack_log_to_prr(rack, 11, 0, __LINE__);
10768 } else {
10769 rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10770 rack_log_to_prr(rack, 12, 0, __LINE__);
10771 }
10772 }
10773 }
10774
10775 static void
rack_log_ack(struct tcpcb * tp,struct tcpopt * to,struct tcphdr * th,int entered_recovery,int dup_ack_struck,int * dsack_seen,int * sacks_seen)10776 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
10777 int *dsack_seen, int *sacks_seen)
10778 {
10779 uint32_t changed;
10780 struct tcp_rack *rack;
10781 struct rack_sendmap *rsm;
10782 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
10783 register uint32_t th_ack;
10784 int32_t i, j, k, num_sack_blks = 0;
10785 uint32_t cts, acked, ack_point;
10786 int loop_start = 0;
10787 uint32_t tsused;
10788 uint32_t segsiz;
10789
10790
10791 INP_WLOCK_ASSERT(tptoinpcb(tp));
10792 if (tcp_get_flags(th) & TH_RST) {
10793 /* We don't log resets */
10794 return;
10795 }
10796 rack = (struct tcp_rack *)tp->t_fb_ptr;
10797 cts = tcp_get_usecs(NULL);
10798 rsm = tqhash_min(rack->r_ctl.tqh);
10799 changed = 0;
10800 th_ack = th->th_ack;
10801 segsiz = ctf_fixed_maxseg(rack->rc_tp);
10802 if (BYTES_THIS_ACK(tp, th) >= segsiz) {
10803 /*
10804 * You only get credit for
10805 * MSS and greater (and you get extra
10806 * credit for larger cum-ack moves).
10807 */
10808 int ac;
10809
10810 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
10811 counter_u64_add(rack_ack_total, ac);
10812 }
10813 if (SEQ_GT(th_ack, tp->snd_una)) {
10814 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
10815 tp->t_acktime = ticks;
10816 }
10817 if (rsm && SEQ_GT(th_ack, rsm->r_start))
10818 changed = th_ack - rsm->r_start;
10819 if (changed) {
10820 rack_process_to_cumack(tp, rack, th_ack, cts, to,
10821 tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
10822 }
10823 if ((to->to_flags & TOF_SACK) == 0) {
10824 /* We are done nothing left and no sack. */
10825 rack_handle_might_revert(tp, rack);
10826 /*
10827 * For cases where we struck a dup-ack
10828 * with no SACK, add to the changes so
10829 * PRR will work right.
10830 */
10831 if (dup_ack_struck && (changed == 0)) {
10832 changed += ctf_fixed_maxseg(rack->rc_tp);
10833 }
10834 goto out;
10835 }
10836 /* Sack block processing */
10837 if (SEQ_GT(th_ack, tp->snd_una))
10838 ack_point = th_ack;
10839 else
10840 ack_point = tp->snd_una;
10841 for (i = 0; i < to->to_nsacks; i++) {
10842 bcopy((to->to_sacks + i * TCPOLEN_SACK),
10843 &sack, sizeof(sack));
10844 sack.start = ntohl(sack.start);
10845 sack.end = ntohl(sack.end);
10846 if (SEQ_GT(sack.end, sack.start) &&
10847 SEQ_GT(sack.start, ack_point) &&
10848 SEQ_LT(sack.start, tp->snd_max) &&
10849 SEQ_GT(sack.end, ack_point) &&
10850 SEQ_LEQ(sack.end, tp->snd_max)) {
10851 sack_blocks[num_sack_blks] = sack;
10852 num_sack_blks++;
10853 } else if (SEQ_LEQ(sack.start, th_ack) &&
10854 SEQ_LEQ(sack.end, th_ack)) {
10855 int was_tlp;
10856
10857 if (dsack_seen != NULL)
10858 *dsack_seen = 1;
10859 was_tlp = rack_note_dsack(rack, sack.start, sack.end);
10860 /*
10861 * Its a D-SACK block.
10862 */
10863 tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
10864 }
10865 }
10866 if (rack->rc_dsack_round_seen) {
10867 /* Is the dsack roound over? */
10868 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
10869 /* Yes it is */
10870 rack->rc_dsack_round_seen = 0;
10871 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
10872 }
10873 }
10874 /*
10875 * Sort the SACK blocks so we can update the rack scoreboard with
10876 * just one pass.
10877 */
10878 num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
10879 num_sack_blks, th->th_ack);
10880 ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
10881 if (sacks_seen != NULL)
10882 *sacks_seen = num_sack_blks;
10883 if (num_sack_blks == 0) {
10884 /* Nothing to sack, but we need to update counts */
10885 goto out_with_totals;
10886 }
10887 /* Its a sack of some sort */
10888 if (num_sack_blks < 2) {
10889 /* Only one, we don't need to sort */
10890 goto do_sack_work;
10891 }
10892 /* Sort the sacks */
10893 for (i = 0; i < num_sack_blks; i++) {
10894 for (j = i + 1; j < num_sack_blks; j++) {
10895 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
10896 sack = sack_blocks[i];
10897 sack_blocks[i] = sack_blocks[j];
10898 sack_blocks[j] = sack;
10899 }
10900 }
10901 }
10902 /*
10903 * Now are any of the sack block ends the same (yes some
10904 * implementations send these)?
10905 */
10906 again:
10907 if (num_sack_blks == 0)
10908 goto out_with_totals;
10909 if (num_sack_blks > 1) {
10910 for (i = 0; i < num_sack_blks; i++) {
10911 for (j = i + 1; j < num_sack_blks; j++) {
10912 if (sack_blocks[i].end == sack_blocks[j].end) {
10913 /*
10914 * Ok these two have the same end we
10915 * want the smallest end and then
10916 * throw away the larger and start
10917 * again.
10918 */
10919 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
10920 /*
10921 * The second block covers
10922 * more area use that
10923 */
10924 sack_blocks[i].start = sack_blocks[j].start;
10925 }
10926 /*
10927 * Now collapse out the dup-sack and
10928 * lower the count
10929 */
10930 for (k = (j + 1); k < num_sack_blks; k++) {
10931 sack_blocks[j].start = sack_blocks[k].start;
10932 sack_blocks[j].end = sack_blocks[k].end;
10933 j++;
10934 }
10935 num_sack_blks--;
10936 goto again;
10937 }
10938 }
10939 }
10940 }
10941 do_sack_work:
10942 /*
10943 * First lets look to see if
10944 * we have retransmitted and
10945 * can use the transmit next?
10946 */
10947 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10948 if (rsm &&
10949 SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
10950 SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
10951 /*
10952 * We probably did the FR and the next
10953 * SACK in continues as we would expect.
10954 */
10955 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, segsiz);
10956 if (acked) {
10957 rack->r_wanted_output = 1;
10958 changed += acked;
10959 }
10960 if (num_sack_blks == 1) {
10961 /*
10962 * This is what we would expect from
10963 * a normal implementation to happen
10964 * after we have retransmitted the FR,
10965 * i.e the sack-filter pushes down
10966 * to 1 block and the next to be retransmitted
10967 * is the sequence in the sack block (has more
10968 * are acked). Count this as ACK'd data to boost
10969 * up the chances of recovering any false positives.
10970 */
10971 counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
10972 counter_u64_add(rack_express_sack, 1);
10973 goto out_with_totals;
10974 } else {
10975 /*
10976 * Start the loop through the
10977 * rest of blocks, past the first block.
10978 */
10979 loop_start = 1;
10980 }
10981 }
10982 counter_u64_add(rack_sack_total, 1);
10983 rsm = rack->r_ctl.rc_sacklast;
10984 for (i = loop_start; i < num_sack_blks; i++) {
10985 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, segsiz);
10986 if (acked) {
10987 rack->r_wanted_output = 1;
10988 changed += acked;
10989 }
10990 }
10991 out_with_totals:
10992 if (num_sack_blks > 1) {
10993 /*
10994 * You get an extra stroke if
10995 * you have more than one sack-blk, this
10996 * could be where we are skipping forward
10997 * and the sack-filter is still working, or
10998 * it could be an attacker constantly
10999 * moving us.
11000 */
11001 counter_u64_add(rack_move_some, 1);
11002 }
11003 out:
11004 if (changed) {
11005 /* Something changed cancel the rack timer */
11006 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11007 }
11008 tsused = tcp_get_usecs(NULL);
11009 rsm = tcp_rack_output(tp, rack, tsused);
11010 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11011 rsm &&
11012 ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11013 /* Enter recovery */
11014 entered_recovery = 1;
11015 rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
11016 /*
11017 * When we enter recovery we need to assure we send
11018 * one packet.
11019 */
11020 if (rack->rack_no_prr == 0) {
11021 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11022 rack_log_to_prr(rack, 8, 0, __LINE__);
11023 }
11024 rack->r_timer_override = 1;
11025 rack->r_early = 0;
11026 rack->r_ctl.rc_agg_early = 0;
11027 } else if (IN_FASTRECOVERY(tp->t_flags) &&
11028 rsm &&
11029 (rack->r_rr_config == 3)) {
11030 /*
11031 * Assure we can output and we get no
11032 * remembered pace time except the retransmit.
11033 */
11034 rack->r_timer_override = 1;
11035 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11036 rack->r_ctl.rc_resend = rsm;
11037 }
11038 if (IN_FASTRECOVERY(tp->t_flags) &&
11039 (rack->rack_no_prr == 0) &&
11040 (entered_recovery == 0)) {
11041 rack_update_prr(tp, rack, changed, th_ack);
11042 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11043 ((tcp_in_hpts(rack->rc_tp) == 0) &&
11044 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11045 /*
11046 * If you are pacing output you don't want
11047 * to override.
11048 */
11049 rack->r_early = 0;
11050 rack->r_ctl.rc_agg_early = 0;
11051 rack->r_timer_override = 1;
11052 }
11053 }
11054 }
11055
11056 static void
rack_strike_dupack(struct tcp_rack * rack,tcp_seq th_ack)11057 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
11058 {
11059 struct rack_sendmap *rsm;
11060
11061 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11062 while (rsm) {
11063 /*
11064 * We need to skip anything already set
11065 * to be retransmitted.
11066 */
11067 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
11068 (rsm->r_flags & RACK_MUST_RXT)) {
11069 rsm = TAILQ_NEXT(rsm, r_tnext);
11070 continue;
11071 }
11072 break;
11073 }
11074 if (rsm && (rsm->r_dupack < 0xff)) {
11075 rsm->r_dupack++;
11076 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11077 struct timeval tv;
11078 uint32_t cts;
11079 /*
11080 * Here we see if we need to retransmit. For
11081 * a SACK type connection if enough time has passed
11082 * we will get a return of the rsm. For a non-sack
11083 * connection we will get the rsm returned if the
11084 * dupack value is 3 or more.
11085 */
11086 cts = tcp_get_usecs(&tv);
11087 rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11088 if (rack->r_ctl.rc_resend != NULL) {
11089 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11090 rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11091 th_ack, __LINE__);
11092 }
11093 rack->r_wanted_output = 1;
11094 rack->r_timer_override = 1;
11095 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11096 }
11097 } else {
11098 rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11099 }
11100 }
11101 }
11102
11103 static void
rack_check_bottom_drag(struct tcpcb * tp,struct tcp_rack * rack,struct socket * so)11104 rack_check_bottom_drag(struct tcpcb *tp,
11105 struct tcp_rack *rack,
11106 struct socket *so)
11107 {
11108 /*
11109 * So what is dragging bottom?
11110 *
11111 * Dragging bottom means you were under pacing and had a
11112 * delay in processing inbound acks waiting on our pacing
11113 * timer to expire. While you were waiting all of the acknowledgments
11114 * for the packets you sent have arrived. This means we are pacing
11115 * way underneath the bottleneck to the point where our Goodput
11116 * measurements stop working, since they require more than one
11117 * ack (usually at least 8 packets worth with multiple acks so we can
11118 * gauge the inter-ack times). If that occurs we have a real problem
11119 * since we are stuck in a hole that we can't get out of without
11120 * something speeding us up.
11121 *
11122 * We also check to see if we are widdling down to just one segment
11123 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
11124 * then we are adding the delayed ack interval into our measurments and
11125 * we need to speed up slightly.
11126 */
11127 uint32_t segsiz, minseg;
11128
11129 segsiz = ctf_fixed_maxseg(tp);
11130 minseg = segsiz;
11131 if (tp->snd_max == tp->snd_una) {
11132 /*
11133 * We are doing dynamic pacing and we are way
11134 * under. Basically everything got acked while
11135 * we were still waiting on the pacer to expire.
11136 *
11137 * This means we need to boost the b/w in
11138 * addition to any earlier boosting of
11139 * the multiplier.
11140 */
11141 uint64_t lt_bw;
11142
11143 tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
11144 lt_bw = rack_get_lt_bw(rack);
11145 rack->rc_dragged_bottom = 1;
11146 rack_validate_multipliers_at_or_above100(rack);
11147 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11148 (rack->dis_lt_bw == 0) &&
11149 (rack->use_lesser_lt_bw == 0) &&
11150 (lt_bw > 0)) {
11151 /*
11152 * Lets use the long-term b/w we have
11153 * been getting as a base.
11154 */
11155 if (rack->rc_gp_filled == 0) {
11156 if (lt_bw > ONE_POINT_TWO_MEG) {
11157 /*
11158 * If we have no measurement
11159 * don't let us set in more than
11160 * 1.2Mbps. If we are still too
11161 * low after pacing with this we
11162 * will hopefully have a max b/w
11163 * available to sanity check things.
11164 */
11165 lt_bw = ONE_POINT_TWO_MEG;
11166 }
11167 rack->r_ctl.rc_rtt_diff = 0;
11168 rack->r_ctl.gp_bw = lt_bw;
11169 rack->rc_gp_filled = 1;
11170 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11171 rack->r_ctl.num_measurements = RACK_REQ_AVG;
11172 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11173 } else if (lt_bw > rack->r_ctl.gp_bw) {
11174 rack->r_ctl.rc_rtt_diff = 0;
11175 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11176 rack->r_ctl.num_measurements = RACK_REQ_AVG;
11177 rack->r_ctl.gp_bw = lt_bw;
11178 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11179 } else
11180 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11181 if ((rack->gp_ready == 0) &&
11182 (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11183 /* We have enough measurements now */
11184 rack->gp_ready = 1;
11185 if (rack->dgp_on ||
11186 rack->rack_hibeta)
11187 rack_set_cc_pacing(rack);
11188 if (rack->defer_options)
11189 rack_apply_deferred_options(rack);
11190 }
11191 } else {
11192 /*
11193 * zero rtt possibly?, settle for just an old increase.
11194 */
11195 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11196 }
11197 } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11198 (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11199 minseg)) &&
11200 (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11201 (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11202 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11203 (segsiz * rack_req_segs))) {
11204 /*
11205 * We are doing dynamic GP pacing and
11206 * we have everything except 1MSS or less
11207 * bytes left out. We are still pacing away.
11208 * And there is data that could be sent, This
11209 * means we are inserting delayed ack time in
11210 * our measurements because we are pacing too slow.
11211 */
11212 rack_validate_multipliers_at_or_above100(rack);
11213 rack->rc_dragged_bottom = 1;
11214 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11215 }
11216 }
11217
11218 #ifdef TCP_REQUEST_TRK
11219 static void
rack_log_hybrid(struct tcp_rack * rack,uint32_t seq,struct tcp_sendfile_track * cur,uint8_t mod,int line,int err)11220 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11221 struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11222 {
11223 int do_log;
11224
11225 do_log = tcp_bblogging_on(rack->rc_tp);
11226 if (do_log == 0) {
11227 if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11228 return;
11229 /* We only allow the three below with point logging on */
11230 if ((mod != HYBRID_LOG_RULES_APP) &&
11231 (mod != HYBRID_LOG_RULES_SET) &&
11232 (mod != HYBRID_LOG_REQ_COMP))
11233 return;
11234
11235 }
11236 if (do_log) {
11237 union tcp_log_stackspecific log;
11238 struct timeval tv;
11239
11240 /* Convert our ms to a microsecond */
11241 memset(&log, 0, sizeof(log));
11242 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11243 log.u_bbr.flex1 = seq;
11244 log.u_bbr.cwnd_gain = line;
11245 if (cur != NULL) {
11246 uint64_t off;
11247
11248 log.u_bbr.flex2 = cur->start_seq;
11249 log.u_bbr.flex3 = cur->end_seq;
11250 log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11251 log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11252 log.u_bbr.flex6 = cur->flags;
11253 log.u_bbr.pkts_out = cur->hybrid_flags;
11254 log.u_bbr.rttProp = cur->timestamp;
11255 log.u_bbr.cur_del_rate = cur->cspr;
11256 log.u_bbr.bw_inuse = cur->start;
11257 log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11258 log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11259 log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11260 log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11261 log.u_bbr.inhpts = 1;
11262 #ifdef TCP_REQUEST_TRK
11263 off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11264 log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11265 #endif
11266 } else {
11267 log.u_bbr.flex2 = err;
11268 }
11269 /*
11270 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11271 */
11272 log.u_bbr.flex7 = rack->rc_catch_up;
11273 log.u_bbr.flex7 <<= 1;
11274 log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11275 log.u_bbr.flex7 <<= 1;
11276 log.u_bbr.flex7 |= rack->dgp_on;
11277 /*
11278 * Compose bbr_state to be a bit wise 0000ADHF
11279 * where A is the always_pace flag
11280 * where D is the dgp_on flag
11281 * where H is the hybrid_mode on flag
11282 * where F is the use_fixed_rate flag.
11283 */
11284 log.u_bbr.bbr_state = rack->rc_always_pace;
11285 log.u_bbr.bbr_state <<= 1;
11286 log.u_bbr.bbr_state |= rack->dgp_on;
11287 log.u_bbr.bbr_state <<= 1;
11288 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
11289 log.u_bbr.bbr_state <<= 1;
11290 log.u_bbr.bbr_state |= rack->use_fixed_rate;
11291 log.u_bbr.flex8 = mod;
11292 log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11293 log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11294 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11295 log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11296 log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11297 log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11298 tcp_log_event(rack->rc_tp, NULL,
11299 &rack->rc_inp->inp_socket->so_rcv,
11300 &rack->rc_inp->inp_socket->so_snd,
11301 TCP_HYBRID_PACING_LOG, 0,
11302 0, &log, false, NULL, __func__, __LINE__, &tv);
11303 }
11304 }
11305 #endif
11306
11307 #ifdef TCP_REQUEST_TRK
11308 static void
rack_set_dgp_hybrid_mode(struct tcp_rack * rack,tcp_seq seq,uint32_t len,uint64_t cts)11309 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11310 {
11311 struct tcp_sendfile_track *rc_cur, *orig_ent;
11312 struct tcpcb *tp;
11313 int err = 0;
11314
11315 orig_ent = rack->r_ctl.rc_last_sft;
11316 rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11317 if (rc_cur == NULL) {
11318 /* If not in the beginning what about the end piece */
11319 if (rack->rc_hybrid_mode)
11320 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11321 rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11322 } else {
11323 err = 12345;
11324 }
11325 /* If we find no parameters we are in straight DGP mode */
11326 if(rc_cur == NULL) {
11327 /* None found for this seq, just DGP for now */
11328 if (rack->rc_hybrid_mode) {
11329 rack->r_ctl.client_suggested_maxseg = 0;
11330 rack->rc_catch_up = 0;
11331 if (rack->cspr_is_fcc == 0)
11332 rack->r_ctl.bw_rate_cap = 0;
11333 else
11334 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11335 }
11336 if (rack->rc_hybrid_mode) {
11337 rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11338 }
11339 if (rack->r_ctl.rc_last_sft) {
11340 rack->r_ctl.rc_last_sft = NULL;
11341 }
11342 return;
11343 }
11344 if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11345 /* This entry was never setup for hybrid pacing on/off etc */
11346 if (rack->rc_hybrid_mode) {
11347 rack->r_ctl.client_suggested_maxseg = 0;
11348 rack->rc_catch_up = 0;
11349 rack->r_ctl.bw_rate_cap = 0;
11350 }
11351 if (rack->r_ctl.rc_last_sft) {
11352 rack->r_ctl.rc_last_sft = NULL;
11353 }
11354 if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11355 rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
11356 rc_cur->first_send = cts;
11357 rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
11358 rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11359 }
11360 return;
11361 }
11362 /*
11363 * Ok if we have a new entry *or* have never
11364 * set up an entry we need to proceed. If
11365 * we have already set it up this entry we
11366 * just continue along with what we already
11367 * setup.
11368 */
11369 tp = rack->rc_tp;
11370 if ((rack->r_ctl.rc_last_sft != NULL) &&
11371 (rack->r_ctl.rc_last_sft == rc_cur)) {
11372 /* Its already in place */
11373 if (rack->rc_hybrid_mode)
11374 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11375 return;
11376 }
11377 if (rack->rc_hybrid_mode == 0) {
11378 rack->r_ctl.rc_last_sft = rc_cur;
11379 if (orig_ent) {
11380 orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11381 orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11382 orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11383 }
11384 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11385 return;
11386 }
11387 if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11388 /* Compensate for all the header overhead's */
11389 if (rack->cspr_is_fcc == 0)
11390 rack->r_ctl.bw_rate_cap = rack_compensate_for_linerate(rack, rc_cur->cspr);
11391 else
11392 rack->r_ctl.fillcw_cap = rack_compensate_for_linerate(rack, rc_cur->cspr);
11393 } else {
11394 if (rack->rc_hybrid_mode) {
11395 if (rack->cspr_is_fcc == 0)
11396 rack->r_ctl.bw_rate_cap = 0;
11397 else
11398 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11399 }
11400 }
11401 if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11402 rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11403 else
11404 rack->r_ctl.client_suggested_maxseg = 0;
11405 if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
11406 /*
11407 * It is the same timestamp as the previous one
11408 * add the hybrid flag that will indicate we use
11409 * sendtime not arrival time for catch-up mode.
11410 */
11411 rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
11412 }
11413 if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11414 (rc_cur->cspr > 0)) {
11415 uint64_t len;
11416
11417 rack->rc_catch_up = 1;
11418 /*
11419 * Calculate the deadline time, first set the
11420 * time to when the request arrived.
11421 */
11422 if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
11423 /*
11424 * For cases where its a duplicate tm (we received more
11425 * than one request for a tm) we want to use now, the point
11426 * where we are just sending the first bit of the request.
11427 */
11428 rc_cur->deadline = cts;
11429 } else {
11430 /*
11431 * Here we have a different tm from the last request
11432 * so we want to use arrival time as our base.
11433 */
11434 rc_cur->deadline = rc_cur->localtime;
11435 }
11436 /*
11437 * Next calculate the length and compensate for
11438 * TLS if need be.
11439 */
11440 len = rc_cur->end - rc_cur->start;
11441 if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11442 /*
11443 * This session is doing TLS. Take a swag guess
11444 * at the overhead.
11445 */
11446 len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11447 }
11448 /*
11449 * Now considering the size, and the cspr, what is the time that
11450 * would be required at the cspr rate. Here we use the raw
11451 * cspr value since the client only looks at the raw data. We
11452 * do use len which includes TLS overhead, but not the TCP/IP etc.
11453 * That will get made up for in the CU pacing rate set.
11454 */
11455 len *= HPTS_USEC_IN_SEC;
11456 len /= rc_cur->cspr;
11457 rc_cur->deadline += len;
11458 } else {
11459 rack->rc_catch_up = 0;
11460 rc_cur->deadline = 0;
11461 }
11462 if (rack->r_ctl.client_suggested_maxseg != 0) {
11463 /*
11464 * We need to reset the max pace segs if we have a
11465 * client_suggested_maxseg.
11466 */
11467 rack_set_pace_segments(tp, rack, __LINE__, NULL);
11468 }
11469 if (orig_ent) {
11470 orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11471 orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11472 orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11473 }
11474 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11475 /* Remember it for next time and for CU mode */
11476 rack->r_ctl.rc_last_sft = rc_cur;
11477 rack->r_ctl.last_tm_mark = rc_cur->timestamp;
11478 }
11479 #endif
11480
11481 static void
rack_chk_req_and_hybrid_on_out(struct tcp_rack * rack,tcp_seq seq,uint32_t len,uint64_t cts)11482 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11483 {
11484 #ifdef TCP_REQUEST_TRK
11485 struct tcp_sendfile_track *ent;
11486
11487 ent = rack->r_ctl.rc_last_sft;
11488 if ((ent == NULL) ||
11489 (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11490 (SEQ_GEQ(seq, ent->end_seq))) {
11491 /* Time to update the track. */
11492 rack_set_dgp_hybrid_mode(rack, seq, len, cts);
11493 ent = rack->r_ctl.rc_last_sft;
11494 }
11495 /* Out of all */
11496 if (ent == NULL) {
11497 return;
11498 }
11499 if (SEQ_LT(ent->end_seq, (seq + len))) {
11500 /*
11501 * This is the case where our end_seq guess
11502 * was wrong. This is usually due to TLS having
11503 * more bytes then our guess. It could also be the
11504 * case that the client sent in two requests closely
11505 * and the SB is full of both so we are sending part
11506 * of each (end|beg). In such a case lets move this
11507 * guys end to match the end of this send. That
11508 * way it will complete when all of it is acked.
11509 */
11510 ent->end_seq = (seq + len);
11511 if (rack->rc_hybrid_mode)
11512 rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11513 }
11514 /* Now validate we have set the send time of this one */
11515 if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11516 ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11517 ent->first_send = cts;
11518 ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11519 ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11520 }
11521 #endif
11522 }
11523
11524 static void
rack_gain_for_fastoutput(struct tcp_rack * rack,struct tcpcb * tp,struct socket * so,uint32_t acked_amount)11525 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11526 {
11527 /*
11528 * The fast output path is enabled and we
11529 * have moved the cumack forward. Lets see if
11530 * we can expand forward the fast path length by
11531 * that amount. What we would ideally like to
11532 * do is increase the number of bytes in the
11533 * fast path block (left_to_send) by the
11534 * acked amount. However we have to gate that
11535 * by two factors:
11536 * 1) The amount outstanding and the rwnd of the peer
11537 * (i.e. we don't want to exceed the rwnd of the peer).
11538 * <and>
11539 * 2) The amount of data left in the socket buffer (i.e.
11540 * we can't send beyond what is in the buffer).
11541 *
11542 * Note that this does not take into account any increase
11543 * in the cwnd. We will only extend the fast path by
11544 * what was acked.
11545 */
11546 uint32_t new_total, gating_val;
11547
11548 new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11549 gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11550 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11551 if (new_total <= gating_val) {
11552 /* We can increase left_to_send by the acked amount */
11553 counter_u64_add(rack_extended_rfo, 1);
11554 rack->r_ctl.fsb.left_to_send = new_total;
11555 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11556 ("rack:%p left_to_send:%u sbavail:%u out:%u",
11557 rack, rack->r_ctl.fsb.left_to_send,
11558 sbavail(&rack->rc_inp->inp_socket->so_snd),
11559 (tp->snd_max - tp->snd_una)));
11560
11561 }
11562 }
11563
11564 static void
rack_adjust_sendmap_head(struct tcp_rack * rack,struct sockbuf * sb)11565 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11566 {
11567 /*
11568 * Here any sendmap entry that points to the
11569 * beginning mbuf must be adjusted to the correct
11570 * offset. This must be called with:
11571 * 1) The socket buffer locked
11572 * 2) snd_una adjusted to its new position.
11573 *
11574 * Note that (2) implies rack_ack_received has also
11575 * been called and all the sbcut's have been done.
11576 *
11577 * We grab the first mbuf in the socket buffer and
11578 * then go through the front of the sendmap, recalculating
11579 * the stored offset for any sendmap entry that has
11580 * that mbuf. We must use the sb functions to do this
11581 * since its possible an add was done has well as
11582 * the subtraction we may have just completed. This should
11583 * not be a penalty though, since we just referenced the sb
11584 * to go in and trim off the mbufs that we freed (of course
11585 * there will be a penalty for the sendmap references though).
11586 *
11587 * Note also with INVARIANT on, we validate with a KASSERT
11588 * that the first sendmap entry has a soff of 0.
11589 *
11590 */
11591 struct mbuf *m;
11592 struct rack_sendmap *rsm;
11593 tcp_seq snd_una;
11594 #ifdef INVARIANTS
11595 int first_processed = 0;
11596 #endif
11597
11598 snd_una = rack->rc_tp->snd_una;
11599 SOCKBUF_LOCK_ASSERT(sb);
11600 m = sb->sb_mb;
11601 rsm = tqhash_min(rack->r_ctl.tqh);
11602 if ((rsm == NULL) || (m == NULL)) {
11603 /* Nothing outstanding */
11604 return;
11605 }
11606 /* The very first RSM's mbuf must point to the head mbuf in the sb */
11607 KASSERT((rsm->m == m),
11608 ("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11609 rack, sb, rsm));
11610 while (rsm->m && (rsm->m == m)) {
11611 /* one to adjust */
11612 #ifdef INVARIANTS
11613 struct mbuf *tm;
11614 uint32_t soff;
11615
11616 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11617 if ((rsm->orig_m_len != m->m_len) ||
11618 (rsm->orig_t_space != M_TRAILINGROOM(m))){
11619 rack_adjust_orig_mlen(rsm);
11620 }
11621 if (first_processed == 0) {
11622 KASSERT((rsm->soff == 0),
11623 ("Rack:%p rsm:%p -- rsm at head but soff not zero",
11624 rack, rsm));
11625 first_processed = 1;
11626 }
11627 if ((rsm->soff != soff) || (rsm->m != tm)) {
11628 /*
11629 * This is not a fatal error, we anticipate it
11630 * might happen (the else code), so we count it here
11631 * so that under invariant we can see that it really
11632 * does happen.
11633 */
11634 counter_u64_add(rack_adjust_map_bw, 1);
11635 }
11636 rsm->m = tm;
11637 rsm->soff = soff;
11638 if (tm) {
11639 rsm->orig_m_len = rsm->m->m_len;
11640 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11641 } else {
11642 rsm->orig_m_len = 0;
11643 rsm->orig_t_space = 0;
11644 }
11645 #else
11646 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11647 if (rsm->m) {
11648 rsm->orig_m_len = rsm->m->m_len;
11649 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11650 } else {
11651 rsm->orig_m_len = 0;
11652 rsm->orig_t_space = 0;
11653 }
11654 #endif
11655 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11656 if (rsm == NULL)
11657 break;
11658 }
11659 }
11660
11661 #ifdef TCP_REQUEST_TRK
11662 static inline void
rack_req_check_for_comp(struct tcp_rack * rack,tcp_seq th_ack)11663 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11664 {
11665 struct tcp_sendfile_track *ent;
11666 int i;
11667
11668 if ((rack->rc_hybrid_mode == 0) &&
11669 (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11670 /*
11671 * Just do normal completions hybrid pacing is not on
11672 * and CLDL is off as well.
11673 */
11674 tcp_req_check_for_comp(rack->rc_tp, th_ack);
11675 return;
11676 }
11677 /*
11678 * Originally I was just going to find the th_ack associated
11679 * with an entry. But then I realized a large strech ack could
11680 * in theory ack two or more requests at once. So instead we
11681 * need to find all entries that are completed by th_ack not
11682 * just a single entry and do our logging.
11683 */
11684 ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11685 while (ent != NULL) {
11686 /*
11687 * We may be doing hybrid pacing or CLDL and need more details possibly
11688 * so we do it manually instead of calling
11689 * tcp_req_check_for_comp()
11690 */
11691 uint64_t laa, tim, data, cbw, ftim;
11692
11693 /* Ok this ack frees it */
11694 rack_log_hybrid(rack, th_ack,
11695 ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11696 rack_log_hybrid_sends(rack, ent, __LINE__);
11697 /* calculate the time based on the ack arrival */
11698 data = ent->end - ent->start;
11699 laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
11700 if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
11701 if (ent->first_send > ent->localtime)
11702 ftim = ent->first_send;
11703 else
11704 ftim = ent->localtime;
11705 } else {
11706 /* TSNH */
11707 ftim = ent->localtime;
11708 }
11709 if (laa > ent->localtime)
11710 tim = laa - ftim;
11711 else
11712 tim = 0;
11713 cbw = data * HPTS_USEC_IN_SEC;
11714 if (tim > 0)
11715 cbw /= tim;
11716 else
11717 cbw = 0;
11718 rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
11719 /*
11720 * Check to see if we are freeing what we are pointing to send wise
11721 * if so be sure to NULL the pointer so we know we are no longer
11722 * set to anything.
11723 */
11724 if (ent == rack->r_ctl.rc_last_sft) {
11725 rack->r_ctl.rc_last_sft = NULL;
11726 if (rack->rc_hybrid_mode) {
11727 rack->rc_catch_up = 0;
11728 if (rack->cspr_is_fcc == 0)
11729 rack->r_ctl.bw_rate_cap = 0;
11730 else
11731 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11732 rack->r_ctl.client_suggested_maxseg = 0;
11733 }
11734 }
11735 /* Generate the log that the tcp_netflix call would have */
11736 tcp_req_log_req_info(rack->rc_tp, ent,
11737 i, TCP_TRK_REQ_LOG_FREED, 0, 0);
11738 /* Free it and see if there is another one */
11739 tcp_req_free_a_slot(rack->rc_tp, ent);
11740 ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11741 }
11742 }
11743 #endif
11744
11745
11746 /*
11747 * Return value of 1, we do not need to call rack_process_data().
11748 * return value of 0, rack_process_data can be called.
11749 * For ret_val if its 0 the TCP is locked, if its non-zero
11750 * its unlocked and probably unsafe to touch the TCB.
11751 */
11752 static int
rack_process_ack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,uint32_t tiwin,int32_t tlen,int32_t * ofia,int32_t thflags,int32_t * ret_val,int32_t orig_tlen)11753 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11754 struct tcpcb *tp, struct tcpopt *to,
11755 uint32_t tiwin, int32_t tlen,
11756 int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
11757 {
11758 int32_t ourfinisacked = 0;
11759 int32_t nsegs, acked_amount;
11760 int32_t acked;
11761 struct mbuf *mfree;
11762 struct tcp_rack *rack;
11763 int32_t under_pacing = 0;
11764 int32_t post_recovery = 0;
11765 uint32_t p_cwnd;
11766
11767 INP_WLOCK_ASSERT(tptoinpcb(tp));
11768
11769 rack = (struct tcp_rack *)tp->t_fb_ptr;
11770 if (SEQ_GEQ(tp->snd_una, tp->iss + (65535 << tp->snd_scale))) {
11771 /* Checking SEG.ACK against ISS is definitely redundant. */
11772 tp->t_flags2 |= TF2_NO_ISS_CHECK;
11773 }
11774 if (!V_tcp_insecure_ack) {
11775 tcp_seq seq_min;
11776 bool ghost_ack_check;
11777
11778 if (tp->t_flags2 & TF2_NO_ISS_CHECK) {
11779 /* Check for too old ACKs (RFC 5961, Section 5.2). */
11780 seq_min = tp->snd_una - tp->max_sndwnd;
11781 ghost_ack_check = false;
11782 } else {
11783 if (SEQ_GT(tp->iss + 1, tp->snd_una - tp->max_sndwnd)) {
11784 /* Checking for ghost ACKs is stricter. */
11785 seq_min = tp->iss + 1;
11786 ghost_ack_check = true;
11787 } else {
11788 /*
11789 * Checking for too old ACKs (RFC 5961,
11790 * Section 5.2) is stricter.
11791 */
11792 seq_min = tp->snd_una - tp->max_sndwnd;
11793 ghost_ack_check = false;
11794 }
11795 }
11796 if (SEQ_LT(th->th_ack, seq_min)) {
11797 if (ghost_ack_check)
11798 TCPSTAT_INC(tcps_rcvghostack);
11799 else
11800 TCPSTAT_INC(tcps_rcvacktooold);
11801 /* Send challenge ACK. */
11802 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11803 rack->r_wanted_output = 1;
11804 return (1);
11805 }
11806 }
11807 if (SEQ_GT(th->th_ack, tp->snd_max)) {
11808 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11809 rack->r_wanted_output = 1;
11810 return (1);
11811 }
11812 if (rack->gp_ready &&
11813 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11814 under_pacing = 1;
11815 }
11816 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
11817 int in_rec, dup_ack_struck = 0;
11818 int dsack_seen = 0, sacks_seen = 0;
11819
11820 in_rec = IN_FASTRECOVERY(tp->t_flags);
11821 if (rack->rc_in_persist) {
11822 tp->t_rxtshift = 0;
11823 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11824 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11825 }
11826
11827 if ((th->th_ack == tp->snd_una) &&
11828 (tiwin == tp->snd_wnd) &&
11829 (orig_tlen == 0) &&
11830 ((to->to_flags & TOF_SACK) == 0)) {
11831 rack_strike_dupack(rack, th->th_ack);
11832 dup_ack_struck = 1;
11833 }
11834 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
11835 dup_ack_struck, &dsack_seen, &sacks_seen);
11836
11837 }
11838 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11839 /*
11840 * Old ack, behind (or duplicate to) the last one rcv'd
11841 * Note: We mark reordering is occuring if its
11842 * less than and we have not closed our window.
11843 */
11844 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
11845 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
11846 if (rack->r_ctl.rc_reorder_ts == 0)
11847 rack->r_ctl.rc_reorder_ts = 1;
11848 }
11849 return (0);
11850 }
11851 /*
11852 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
11853 * something we sent.
11854 */
11855 if (tp->t_flags & TF_NEEDSYN) {
11856 /*
11857 * T/TCP: Connection was half-synchronized, and our SYN has
11858 * been ACK'd (so connection is now fully synchronized). Go
11859 * to non-starred state, increment snd_una for ACK of SYN,
11860 * and check if we can do window scaling.
11861 */
11862 tp->t_flags &= ~TF_NEEDSYN;
11863 tp->snd_una++;
11864 /* Do window scaling? */
11865 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11866 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11867 tp->rcv_scale = tp->request_r_scale;
11868 /* Send window already scaled. */
11869 }
11870 }
11871 nsegs = max(1, m->m_pkthdr.lro_nsegs);
11872
11873 acked = BYTES_THIS_ACK(tp, th);
11874 if (acked) {
11875 /*
11876 * Any time we move the cum-ack forward clear
11877 * keep-alive tied probe-not-answered. The
11878 * persists clears its own on entry.
11879 */
11880 rack->probe_not_answered = 0;
11881 }
11882 KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11883 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11884 /*
11885 * If we just performed our first retransmit, and the ACK arrives
11886 * within our recovery window, then it was a mistake to do the
11887 * retransmit in the first place. Recover our original cwnd and
11888 * ssthresh, and proceed to transmit where we left off.
11889 */
11890 if ((tp->t_flags & TF_PREVVALID) &&
11891 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11892 tp->t_flags &= ~TF_PREVVALID;
11893 if (tp->t_rxtshift == 1 &&
11894 (int)(ticks - tp->t_badrxtwin) < 0)
11895 rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11896 }
11897 if (acked) {
11898 /* assure we are not backed off */
11899 tp->t_rxtshift = 0;
11900 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11901 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11902 rack->rc_tlp_in_progress = 0;
11903 rack->r_ctl.rc_tlp_cnt_out = 0;
11904 /*
11905 * If it is the RXT timer we want to
11906 * stop it, so we can restart a TLP.
11907 */
11908 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11909 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11910 #ifdef TCP_REQUEST_TRK
11911 rack_req_check_for_comp(rack, th->th_ack);
11912 #endif
11913 }
11914 /*
11915 * If we have a timestamp reply, update smoothed round trip time. If
11916 * no timestamp is present but transmit timer is running and timed
11917 * sequence number was acked, update smoothed round trip time. Since
11918 * we now have an rtt measurement, cancel the timer backoff (cf.,
11919 * Phil Karn's retransmit alg.). Recompute the initial retransmit
11920 * timer.
11921 *
11922 * Some boxes send broken timestamp replies during the SYN+ACK
11923 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11924 * and blow up the retransmit timer.
11925 */
11926 /*
11927 * If all outstanding data is acked, stop retransmit timer and
11928 * remember to restart (more output or persist). If there is more
11929 * data to be acked, restart retransmit timer, using current
11930 * (possibly backed-off) value.
11931 */
11932 if (acked == 0) {
11933 if (ofia)
11934 *ofia = ourfinisacked;
11935 return (0);
11936 }
11937 if (IN_RECOVERY(tp->t_flags)) {
11938 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
11939 (SEQ_LT(th->th_ack, tp->snd_max))) {
11940 tcp_rack_partialack(tp);
11941 } else {
11942 rack_post_recovery(tp, th->th_ack);
11943 post_recovery = 1;
11944 /*
11945 * Grab the segsiz, multiply by 2 and add the snd_cwnd
11946 * that is the max the CC should add if we are exiting
11947 * recovery and doing a late add.
11948 */
11949 p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
11950 p_cwnd <<= 1;
11951 p_cwnd += tp->snd_cwnd;
11952 }
11953 } else if ((rack->rto_from_rec == 1) &&
11954 SEQ_GEQ(th->th_ack, tp->snd_recover)) {
11955 /*
11956 * We were in recovery, hit a rxt timeout
11957 * and never re-entered recovery. The timeout(s)
11958 * made up all the lost data. In such a case
11959 * we need to clear the rto_from_rec flag.
11960 */
11961 rack->rto_from_rec = 0;
11962 }
11963 /*
11964 * Let the congestion control algorithm update congestion control
11965 * related information. This typically means increasing the
11966 * congestion window.
11967 */
11968 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
11969 if (post_recovery &&
11970 (tp->snd_cwnd > p_cwnd)) {
11971 /* Must be non-newreno (cubic) getting too ahead of itself */
11972 tp->snd_cwnd = p_cwnd;
11973 }
11974 SOCK_SENDBUF_LOCK(so);
11975 acked_amount = min(acked, (int)sbavail(&so->so_snd));
11976 tp->snd_wnd -= acked_amount;
11977 mfree = sbcut_locked(&so->so_snd, acked_amount);
11978 if ((sbused(&so->so_snd) == 0) &&
11979 (acked > acked_amount) &&
11980 (tp->t_state >= TCPS_FIN_WAIT_1) &&
11981 (tp->t_flags & TF_SENTFIN)) {
11982 /*
11983 * We must be sure our fin
11984 * was sent and acked (we can be
11985 * in FIN_WAIT_1 without having
11986 * sent the fin).
11987 */
11988 ourfinisacked = 1;
11989 }
11990 tp->snd_una = th->th_ack;
11991 /* wakeups? */
11992 if (acked_amount && sbavail(&so->so_snd))
11993 rack_adjust_sendmap_head(rack, &so->so_snd);
11994 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11995 /* NB: sowwakeup_locked() does an implicit unlock. */
11996 sowwakeup_locked(so);
11997 m_freem(mfree);
11998 if (SEQ_GT(tp->snd_una, tp->snd_recover))
11999 tp->snd_recover = tp->snd_una;
12000
12001 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12002 tp->snd_nxt = tp->snd_max;
12003 }
12004 if (under_pacing &&
12005 (rack->use_fixed_rate == 0) &&
12006 (rack->in_probe_rtt == 0) &&
12007 rack->rc_gp_dyn_mul &&
12008 rack->rc_always_pace) {
12009 /* Check if we are dragging bottom */
12010 rack_check_bottom_drag(tp, rack, so);
12011 }
12012 if (tp->snd_una == tp->snd_max) {
12013 /* Nothing left outstanding */
12014 tp->t_flags &= ~TF_PREVVALID;
12015 if (rack->r_ctl.rc_went_idle_time == 0)
12016 rack->r_ctl.rc_went_idle_time = 1;
12017 rack->r_ctl.retran_during_recovery = 0;
12018 rack->r_ctl.dsack_byte_cnt = 0;
12019 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12020 if (sbavail(&tptosocket(tp)->so_snd) == 0)
12021 tp->t_acktime = 0;
12022 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12023 rack->rc_suspicious = 0;
12024 /* Set need output so persist might get set */
12025 rack->r_wanted_output = 1;
12026 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12027 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12028 (sbavail(&so->so_snd) == 0) &&
12029 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12030 /*
12031 * The socket was gone and the
12032 * peer sent data (now or in the past), time to
12033 * reset him.
12034 */
12035 *ret_val = 1;
12036 /* tcp_close will kill the inp pre-log the Reset */
12037 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12038 tp = tcp_close(tp);
12039 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12040 return (1);
12041 }
12042 }
12043 if (ofia)
12044 *ofia = ourfinisacked;
12045 return (0);
12046 }
12047
12048
12049 static void
rack_log_collapse(struct tcp_rack * rack,uint32_t cnt,uint32_t split,uint32_t out,int line,int dir,uint32_t flags,struct rack_sendmap * rsm)12050 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12051 int dir, uint32_t flags, struct rack_sendmap *rsm)
12052 {
12053 if (tcp_bblogging_on(rack->rc_tp)) {
12054 union tcp_log_stackspecific log;
12055 struct timeval tv;
12056
12057 memset(&log, 0, sizeof(log));
12058 log.u_bbr.flex1 = cnt;
12059 log.u_bbr.flex2 = split;
12060 log.u_bbr.flex3 = out;
12061 log.u_bbr.flex4 = line;
12062 log.u_bbr.flex5 = rack->r_must_retran;
12063 log.u_bbr.flex6 = flags;
12064 log.u_bbr.flex7 = rack->rc_has_collapsed;
12065 log.u_bbr.flex8 = dir; /*
12066 * 1 is collapsed, 0 is uncollapsed,
12067 * 2 is log of a rsm being marked, 3 is a split.
12068 */
12069 if (rsm == NULL)
12070 log.u_bbr.rttProp = 0;
12071 else
12072 log.u_bbr.rttProp = (uintptr_t)rsm;
12073 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12074 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12075 TCP_LOG_EVENTP(rack->rc_tp, NULL,
12076 &rack->rc_inp->inp_socket->so_rcv,
12077 &rack->rc_inp->inp_socket->so_snd,
12078 TCP_RACK_LOG_COLLAPSE, 0,
12079 0, &log, false, &tv);
12080 }
12081 }
12082
12083 static void
rack_collapsed_window(struct tcp_rack * rack,uint32_t out,tcp_seq th_ack,int line)12084 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12085 {
12086 /*
12087 * Here all we do is mark the collapsed point and set the flag.
12088 * This may happen again and again, but there is no
12089 * sense splitting our map until we know where the
12090 * peer finally lands in the collapse.
12091 */
12092 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12093 if ((rack->rc_has_collapsed == 0) ||
12094 (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12095 counter_u64_add(rack_collapsed_win_seen, 1);
12096 rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12097 rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12098 rack->rc_has_collapsed = 1;
12099 rack->r_collapse_point_valid = 1;
12100 rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12101 }
12102
12103 static void
rack_un_collapse_window(struct tcp_rack * rack,int line)12104 rack_un_collapse_window(struct tcp_rack *rack, int line)
12105 {
12106 struct rack_sendmap *nrsm, *rsm;
12107 int cnt = 0, split = 0;
12108 int insret __diagused;
12109
12110
12111 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12112 rack->rc_has_collapsed = 0;
12113 rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12114 if (rsm == NULL) {
12115 /* Nothing to do maybe the peer ack'ed it all */
12116 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12117 return;
12118 }
12119 /* Now do we need to split this one? */
12120 if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12121 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12122 rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12123 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12124 if (nrsm == NULL) {
12125 /* We can't get a rsm, mark all? */
12126 nrsm = rsm;
12127 goto no_split;
12128 }
12129 /* Clone it */
12130 split = 1;
12131 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12132 #ifndef INVARIANTS
12133 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12134 #else
12135 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12136 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12137 nrsm, insret, rack, rsm);
12138 }
12139 #endif
12140 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12141 rack->r_ctl.last_collapse_point, __LINE__);
12142 if (rsm->r_in_tmap) {
12143 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12144 nrsm->r_in_tmap = 1;
12145 }
12146 /*
12147 * Set in the new RSM as the
12148 * collapsed starting point
12149 */
12150 rsm = nrsm;
12151 }
12152
12153 no_split:
12154 TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm) {
12155 cnt++;
12156 nrsm->r_flags |= RACK_RWND_COLLAPSED;
12157 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12158 cnt++;
12159 }
12160 if (cnt) {
12161 counter_u64_add(rack_collapsed_win, 1);
12162 }
12163 rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12164 }
12165
12166 static void
rack_handle_delayed_ack(struct tcpcb * tp,struct tcp_rack * rack,int32_t tlen,int32_t tfo_syn)12167 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12168 int32_t tlen, int32_t tfo_syn)
12169 {
12170 if (DELAY_ACK(tp, tlen) || tfo_syn) {
12171 rack_timer_cancel(tp, rack,
12172 rack->r_ctl.rc_rcvtime, __LINE__);
12173 tp->t_flags |= TF_DELACK;
12174 } else {
12175 rack->r_wanted_output = 1;
12176 tp->t_flags |= TF_ACKNOW;
12177 }
12178 }
12179
12180 static void
rack_validate_fo_sendwin_up(struct tcpcb * tp,struct tcp_rack * rack)12181 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12182 {
12183 /*
12184 * If fast output is in progress, lets validate that
12185 * the new window did not shrink on us and make it
12186 * so fast output should end.
12187 */
12188 if (rack->r_fast_output) {
12189 uint32_t out;
12190
12191 /*
12192 * Calculate what we will send if left as is
12193 * and compare that to our send window.
12194 */
12195 out = ctf_outstanding(tp);
12196 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12197 /* ok we have an issue */
12198 if (out >= tp->snd_wnd) {
12199 /* Turn off fast output the window is met or collapsed */
12200 rack->r_fast_output = 0;
12201 } else {
12202 /* we have some room left */
12203 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12204 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12205 /* If not at least 1 full segment never mind */
12206 rack->r_fast_output = 0;
12207 }
12208 }
12209 }
12210 }
12211 }
12212
12213 /*
12214 * Return value of 1, the TCB is unlocked and most
12215 * likely gone, return value of 0, the TCP is still
12216 * locked.
12217 */
12218 static int
rack_process_data(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt)12219 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12220 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12221 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12222 {
12223 /*
12224 * Update window information. Don't look at window if no ACK: TAC's
12225 * send garbage on first SYN.
12226 */
12227 int32_t nsegs;
12228 int32_t tfo_syn;
12229 struct tcp_rack *rack;
12230
12231 INP_WLOCK_ASSERT(tptoinpcb(tp));
12232
12233 rack = (struct tcp_rack *)tp->t_fb_ptr;
12234 nsegs = max(1, m->m_pkthdr.lro_nsegs);
12235 if ((thflags & TH_ACK) &&
12236 (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12237 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12238 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12239 /* keep track of pure window updates */
12240 if (tlen == 0 &&
12241 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12242 KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12243 tp->snd_wnd = tiwin;
12244 rack_validate_fo_sendwin_up(tp, rack);
12245 tp->snd_wl1 = th->th_seq;
12246 tp->snd_wl2 = th->th_ack;
12247 if (tp->snd_wnd > tp->max_sndwnd)
12248 tp->max_sndwnd = tp->snd_wnd;
12249 rack->r_wanted_output = 1;
12250 } else if (thflags & TH_ACK) {
12251 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12252 tp->snd_wnd = tiwin;
12253 rack_validate_fo_sendwin_up(tp, rack);
12254 tp->snd_wl1 = th->th_seq;
12255 tp->snd_wl2 = th->th_ack;
12256 }
12257 }
12258 if (tp->snd_wnd < ctf_outstanding(tp))
12259 /* The peer collapsed the window */
12260 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12261 else if (rack->rc_has_collapsed)
12262 rack_un_collapse_window(rack, __LINE__);
12263 if ((rack->r_collapse_point_valid) &&
12264 (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12265 rack->r_collapse_point_valid = 0;
12266 /* Was persist timer active and now we have window space? */
12267 if ((rack->rc_in_persist != 0) &&
12268 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12269 rack->r_ctl.rc_pace_min_segs))) {
12270 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12271 tp->snd_nxt = tp->snd_max;
12272 /* Make sure we output to start the timer */
12273 rack->r_wanted_output = 1;
12274 }
12275 /* Do we enter persists? */
12276 if ((rack->rc_in_persist == 0) &&
12277 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12278 TCPS_HAVEESTABLISHED(tp->t_state) &&
12279 ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12280 sbavail(&tptosocket(tp)->so_snd) &&
12281 (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12282 /*
12283 * Here the rwnd is less than
12284 * the pacing size, we are established,
12285 * nothing is outstanding, and there is
12286 * data to send. Enter persists.
12287 */
12288 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12289 }
12290 if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12291 m_freem(m);
12292 return (0);
12293 }
12294 /*
12295 * don't process the URG bit, ignore them drag
12296 * along the up.
12297 */
12298 tp->rcv_up = tp->rcv_nxt;
12299
12300 /*
12301 * Process the segment text, merging it into the TCP sequencing
12302 * queue, and arranging for acknowledgment of receipt if necessary.
12303 * This process logically involves adjusting tp->rcv_wnd as data is
12304 * presented to the user (this happens in tcp_usrreq.c, case
12305 * PRU_RCVD). If a FIN has already been received on this connection
12306 * then we just ignore the text.
12307 */
12308 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12309 (tp->t_flags & TF_FASTOPEN));
12310 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12311 TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12312 tcp_seq save_start = th->th_seq;
12313 tcp_seq save_rnxt = tp->rcv_nxt;
12314 int save_tlen = tlen;
12315
12316 m_adj(m, drop_hdrlen); /* delayed header drop */
12317 /*
12318 * Insert segment which includes th into TCP reassembly
12319 * queue with control block tp. Set thflags to whether
12320 * reassembly now includes a segment with FIN. This handles
12321 * the common case inline (segment is the next to be
12322 * received on an established connection, and the queue is
12323 * empty), avoiding linkage into and removal from the queue
12324 * and repetition of various conversions. Set DELACK for
12325 * segments received in order, but ack immediately when
12326 * segments are out of order (so fast retransmit can work).
12327 */
12328 if (th->th_seq == tp->rcv_nxt &&
12329 SEGQ_EMPTY(tp) &&
12330 (TCPS_HAVEESTABLISHED(tp->t_state) ||
12331 tfo_syn)) {
12332 #ifdef NETFLIX_SB_LIMITS
12333 u_int mcnt, appended;
12334
12335 if (so->so_rcv.sb_shlim) {
12336 mcnt = m_memcnt(m);
12337 appended = 0;
12338 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12339 CFO_NOSLEEP, NULL) == false) {
12340 counter_u64_add(tcp_sb_shlim_fails, 1);
12341 m_freem(m);
12342 return (0);
12343 }
12344 }
12345 #endif
12346 rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12347 tp->rcv_nxt += tlen;
12348 if (tlen &&
12349 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12350 (tp->t_fbyte_in == 0)) {
12351 tp->t_fbyte_in = ticks;
12352 if (tp->t_fbyte_in == 0)
12353 tp->t_fbyte_in = 1;
12354 if (tp->t_fbyte_out && tp->t_fbyte_in)
12355 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12356 }
12357 thflags = tcp_get_flags(th) & TH_FIN;
12358 KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12359 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12360 SOCK_RECVBUF_LOCK(so);
12361 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12362 m_freem(m);
12363 } else {
12364 int32_t newsize;
12365
12366 if (tlen > 0) {
12367 newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12368 if (newsize)
12369 if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12370 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12371 }
12372 #ifdef NETFLIX_SB_LIMITS
12373 appended =
12374 #endif
12375 sbappendstream_locked(&so->so_rcv, m, 0);
12376 }
12377 rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12378 /* NB: sorwakeup_locked() does an implicit unlock. */
12379 sorwakeup_locked(so);
12380 #ifdef NETFLIX_SB_LIMITS
12381 if (so->so_rcv.sb_shlim && appended != mcnt)
12382 counter_fo_release(so->so_rcv.sb_shlim,
12383 mcnt - appended);
12384 #endif
12385 } else {
12386 /*
12387 * XXX: Due to the header drop above "th" is
12388 * theoretically invalid by now. Fortunately
12389 * m_adj() doesn't actually frees any mbufs when
12390 * trimming from the head.
12391 */
12392 tcp_seq temp = save_start;
12393
12394 thflags = tcp_reass(tp, th, &temp, &tlen, m);
12395 tp->t_flags |= TF_ACKNOW;
12396 if (tp->t_flags & TF_WAKESOR) {
12397 tp->t_flags &= ~TF_WAKESOR;
12398 /* NB: sorwakeup_locked() does an implicit unlock. */
12399 sorwakeup_locked(so);
12400 }
12401 }
12402 if ((tp->t_flags & TF_SACK_PERMIT) &&
12403 (save_tlen > 0) &&
12404 TCPS_HAVEESTABLISHED(tp->t_state)) {
12405 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12406 /*
12407 * DSACK actually handled in the fastpath
12408 * above.
12409 */
12410 tcp_update_sack_list(tp, save_start,
12411 save_start + save_tlen);
12412 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12413 if ((tp->rcv_numsacks >= 1) &&
12414 (tp->sackblks[0].end == save_start)) {
12415 /*
12416 * Partial overlap, recorded at todrop
12417 * above.
12418 */
12419 tcp_update_sack_list(tp,
12420 tp->sackblks[0].start,
12421 tp->sackblks[0].end);
12422 } else {
12423 tcp_update_dsack_list(tp, save_start,
12424 save_start + save_tlen);
12425 }
12426 } else if (tlen >= save_tlen) {
12427 /* Update of sackblks. */
12428 tcp_update_dsack_list(tp, save_start,
12429 save_start + save_tlen);
12430 } else if (tlen > 0) {
12431 tcp_update_dsack_list(tp, save_start,
12432 save_start + tlen);
12433 }
12434 }
12435 } else {
12436 m_freem(m);
12437 thflags &= ~TH_FIN;
12438 }
12439
12440 /*
12441 * If FIN is received ACK the FIN and let the user know that the
12442 * connection is closing.
12443 */
12444 if (thflags & TH_FIN) {
12445 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12446 /* The socket upcall is handled by socantrcvmore. */
12447 socantrcvmore(so);
12448 /*
12449 * If connection is half-synchronized (ie NEEDSYN
12450 * flag on) then delay ACK, so it may be piggybacked
12451 * when SYN is sent. Otherwise, since we received a
12452 * FIN then no more input can be expected, send ACK
12453 * now.
12454 */
12455 if (tp->t_flags & TF_NEEDSYN) {
12456 rack_timer_cancel(tp, rack,
12457 rack->r_ctl.rc_rcvtime, __LINE__);
12458 tp->t_flags |= TF_DELACK;
12459 } else {
12460 tp->t_flags |= TF_ACKNOW;
12461 }
12462 tp->rcv_nxt++;
12463 }
12464 switch (tp->t_state) {
12465 /*
12466 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12467 * CLOSE_WAIT state.
12468 */
12469 case TCPS_SYN_RECEIVED:
12470 tp->t_starttime = ticks;
12471 /* FALLTHROUGH */
12472 case TCPS_ESTABLISHED:
12473 rack_timer_cancel(tp, rack,
12474 rack->r_ctl.rc_rcvtime, __LINE__);
12475 tcp_state_change(tp, TCPS_CLOSE_WAIT);
12476 break;
12477
12478 /*
12479 * If still in FIN_WAIT_1 STATE FIN has not been
12480 * acked so enter the CLOSING state.
12481 */
12482 case TCPS_FIN_WAIT_1:
12483 rack_timer_cancel(tp, rack,
12484 rack->r_ctl.rc_rcvtime, __LINE__);
12485 tcp_state_change(tp, TCPS_CLOSING);
12486 break;
12487
12488 /*
12489 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12490 * starting the time-wait timer, turning off the
12491 * other standard timers.
12492 */
12493 case TCPS_FIN_WAIT_2:
12494 rack_timer_cancel(tp, rack,
12495 rack->r_ctl.rc_rcvtime, __LINE__);
12496 tcp_twstart(tp);
12497 return (1);
12498 }
12499 }
12500 /*
12501 * Return any desired output.
12502 */
12503 if ((tp->t_flags & TF_ACKNOW) ||
12504 (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12505 rack->r_wanted_output = 1;
12506 }
12507 return (0);
12508 }
12509
12510 /*
12511 * Here nothing is really faster, its just that we
12512 * have broken out the fast-data path also just like
12513 * the fast-ack.
12514 */
12515 static int
rack_do_fastnewdata(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t nxt_pkt,uint8_t iptos)12516 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12517 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12518 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12519 {
12520 int32_t nsegs;
12521 int32_t newsize = 0; /* automatic sockbuf scaling */
12522 struct tcp_rack *rack;
12523 #ifdef NETFLIX_SB_LIMITS
12524 u_int mcnt, appended;
12525 #endif
12526
12527 /*
12528 * If last ACK falls within this segment's sequence numbers, record
12529 * the timestamp. NOTE that the test is modified according to the
12530 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12531 */
12532 if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12533 return (0);
12534 }
12535 if (tiwin && tiwin != tp->snd_wnd) {
12536 return (0);
12537 }
12538 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12539 return (0);
12540 }
12541 if (__predict_false((to->to_flags & TOF_TS) &&
12542 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12543 return (0);
12544 }
12545 if (__predict_false((th->th_ack != tp->snd_una))) {
12546 return (0);
12547 }
12548 if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12549 return (0);
12550 }
12551 if ((to->to_flags & TOF_TS) != 0 &&
12552 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12553 tp->ts_recent_age = tcp_ts_getticks();
12554 tp->ts_recent = to->to_tsval;
12555 }
12556 rack = (struct tcp_rack *)tp->t_fb_ptr;
12557 /*
12558 * This is a pure, in-sequence data packet with nothing on the
12559 * reassembly queue and we have enough buffer space to take it.
12560 */
12561 nsegs = max(1, m->m_pkthdr.lro_nsegs);
12562
12563 #ifdef NETFLIX_SB_LIMITS
12564 if (so->so_rcv.sb_shlim) {
12565 mcnt = m_memcnt(m);
12566 appended = 0;
12567 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12568 CFO_NOSLEEP, NULL) == false) {
12569 counter_u64_add(tcp_sb_shlim_fails, 1);
12570 m_freem(m);
12571 return (1);
12572 }
12573 }
12574 #endif
12575 /* Clean receiver SACK report if present */
12576 if (tp->rcv_numsacks)
12577 tcp_clean_sackreport(tp);
12578 KMOD_TCPSTAT_INC(tcps_preddat);
12579 tp->rcv_nxt += tlen;
12580 if (tlen &&
12581 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12582 (tp->t_fbyte_in == 0)) {
12583 tp->t_fbyte_in = ticks;
12584 if (tp->t_fbyte_in == 0)
12585 tp->t_fbyte_in = 1;
12586 if (tp->t_fbyte_out && tp->t_fbyte_in)
12587 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12588 }
12589 /*
12590 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12591 */
12592 tp->snd_wl1 = th->th_seq;
12593 /*
12594 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12595 */
12596 tp->rcv_up = tp->rcv_nxt;
12597 KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12598 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12599 newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12600
12601 /* Add data to socket buffer. */
12602 SOCK_RECVBUF_LOCK(so);
12603 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12604 m_freem(m);
12605 } else {
12606 /*
12607 * Set new socket buffer size. Give up when limit is
12608 * reached.
12609 */
12610 if (newsize)
12611 if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12612 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12613 m_adj(m, drop_hdrlen); /* delayed header drop */
12614 #ifdef NETFLIX_SB_LIMITS
12615 appended =
12616 #endif
12617 sbappendstream_locked(&so->so_rcv, m, 0);
12618 ctf_calc_rwin(so, tp);
12619 }
12620 rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12621 /* NB: sorwakeup_locked() does an implicit unlock. */
12622 sorwakeup_locked(so);
12623 #ifdef NETFLIX_SB_LIMITS
12624 if (so->so_rcv.sb_shlim && mcnt != appended)
12625 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12626 #endif
12627 rack_handle_delayed_ack(tp, rack, tlen, 0);
12628 if (tp->snd_una == tp->snd_max)
12629 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12630 return (1);
12631 }
12632
12633 /*
12634 * This subfunction is used to try to highly optimize the
12635 * fast path. We again allow window updates that are
12636 * in sequence to remain in the fast-path. We also add
12637 * in the __predict's to attempt to help the compiler.
12638 * Note that if we return a 0, then we can *not* process
12639 * it and the caller should push the packet into the
12640 * slow-path.
12641 */
12642 static int
rack_fastack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t nxt_pkt,uint32_t cts)12643 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12644 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12645 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12646 {
12647 int32_t acked;
12648 int32_t nsegs;
12649 int32_t under_pacing = 0;
12650 struct tcp_rack *rack;
12651
12652 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12653 /* Old ack, behind (or duplicate to) the last one rcv'd */
12654 return (0);
12655 }
12656 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12657 /* Above what we have sent? */
12658 return (0);
12659 }
12660 if (__predict_false(tiwin == 0)) {
12661 /* zero window */
12662 return (0);
12663 }
12664 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12665 /* We need a SYN or a FIN, unlikely.. */
12666 return (0);
12667 }
12668 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12669 /* Timestamp is behind .. old ack with seq wrap? */
12670 return (0);
12671 }
12672 if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12673 /* Still recovering */
12674 return (0);
12675 }
12676 rack = (struct tcp_rack *)tp->t_fb_ptr;
12677 if (rack->r_ctl.rc_sacked) {
12678 /* We have sack holes on our scoreboard */
12679 return (0);
12680 }
12681 /* Ok if we reach here, we can process a fast-ack */
12682 if (rack->gp_ready &&
12683 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12684 under_pacing = 1;
12685 }
12686 nsegs = max(1, m->m_pkthdr.lro_nsegs);
12687 rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12688 /* Did the window get updated? */
12689 if (tiwin != tp->snd_wnd) {
12690 tp->snd_wnd = tiwin;
12691 rack_validate_fo_sendwin_up(tp, rack);
12692 tp->snd_wl1 = th->th_seq;
12693 if (tp->snd_wnd > tp->max_sndwnd)
12694 tp->max_sndwnd = tp->snd_wnd;
12695 }
12696 /* Do we exit persists? */
12697 if ((rack->rc_in_persist != 0) &&
12698 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12699 rack->r_ctl.rc_pace_min_segs))) {
12700 rack_exit_persist(tp, rack, cts);
12701 }
12702 /* Do we enter persists? */
12703 if ((rack->rc_in_persist == 0) &&
12704 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12705 TCPS_HAVEESTABLISHED(tp->t_state) &&
12706 ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12707 sbavail(&tptosocket(tp)->so_snd) &&
12708 (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12709 /*
12710 * Here the rwnd is less than
12711 * the pacing size, we are established,
12712 * nothing is outstanding, and there is
12713 * data to send. Enter persists.
12714 */
12715 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12716 }
12717 /*
12718 * If last ACK falls within this segment's sequence numbers, record
12719 * the timestamp. NOTE that the test is modified according to the
12720 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12721 */
12722 if ((to->to_flags & TOF_TS) != 0 &&
12723 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12724 tp->ts_recent_age = tcp_ts_getticks();
12725 tp->ts_recent = to->to_tsval;
12726 }
12727 /*
12728 * This is a pure ack for outstanding data.
12729 */
12730 KMOD_TCPSTAT_INC(tcps_predack);
12731
12732 /*
12733 * "bad retransmit" recovery.
12734 */
12735 if ((tp->t_flags & TF_PREVVALID) &&
12736 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12737 tp->t_flags &= ~TF_PREVVALID;
12738 if (tp->t_rxtshift == 1 &&
12739 (int)(ticks - tp->t_badrxtwin) < 0)
12740 rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12741 }
12742 /*
12743 * Recalculate the transmit timer / rtt.
12744 *
12745 * Some boxes send broken timestamp replies during the SYN+ACK
12746 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12747 * and blow up the retransmit timer.
12748 */
12749 acked = BYTES_THIS_ACK(tp, th);
12750
12751 #ifdef TCP_HHOOK
12752 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12753 hhook_run_tcp_est_in(tp, th, to);
12754 #endif
12755 KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12756 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12757 if (acked) {
12758 struct mbuf *mfree;
12759
12760 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12761 SOCK_SENDBUF_LOCK(so);
12762 mfree = sbcut_locked(&so->so_snd, acked);
12763 tp->snd_una = th->th_ack;
12764 /* Note we want to hold the sb lock through the sendmap adjust */
12765 rack_adjust_sendmap_head(rack, &so->so_snd);
12766 /* Wake up the socket if we have room to write more */
12767 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12768 sowwakeup_locked(so);
12769 m_freem(mfree);
12770 tp->t_rxtshift = 0;
12771 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12772 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12773 rack->rc_tlp_in_progress = 0;
12774 rack->r_ctl.rc_tlp_cnt_out = 0;
12775 /*
12776 * If it is the RXT timer we want to
12777 * stop it, so we can restart a TLP.
12778 */
12779 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12780 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12781
12782 #ifdef TCP_REQUEST_TRK
12783 rack_req_check_for_comp(rack, th->th_ack);
12784 #endif
12785 }
12786 /*
12787 * Let the congestion control algorithm update congestion control
12788 * related information. This typically means increasing the
12789 * congestion window.
12790 */
12791 if (tp->snd_wnd < ctf_outstanding(tp)) {
12792 /* The peer collapsed the window */
12793 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12794 } else if (rack->rc_has_collapsed)
12795 rack_un_collapse_window(rack, __LINE__);
12796 if ((rack->r_collapse_point_valid) &&
12797 (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
12798 rack->r_collapse_point_valid = 0;
12799 /*
12800 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
12801 */
12802 tp->snd_wl2 = th->th_ack;
12803 tp->t_dupacks = 0;
12804 m_freem(m);
12805 /* ND6_HINT(tp); *//* Some progress has been made. */
12806
12807 /*
12808 * If all outstanding data are acked, stop retransmit timer,
12809 * otherwise restart timer using current (possibly backed-off)
12810 * value. If process is waiting for space, wakeup/selwakeup/signal.
12811 * If data are ready to send, let tcp_output decide between more
12812 * output or persist.
12813 */
12814 if (under_pacing &&
12815 (rack->use_fixed_rate == 0) &&
12816 (rack->in_probe_rtt == 0) &&
12817 rack->rc_gp_dyn_mul &&
12818 rack->rc_always_pace) {
12819 /* Check if we are dragging bottom */
12820 rack_check_bottom_drag(tp, rack, so);
12821 }
12822 if (tp->snd_una == tp->snd_max) {
12823 tp->t_flags &= ~TF_PREVVALID;
12824 rack->r_ctl.retran_during_recovery = 0;
12825 rack->rc_suspicious = 0;
12826 rack->r_ctl.dsack_byte_cnt = 0;
12827 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12828 if (rack->r_ctl.rc_went_idle_time == 0)
12829 rack->r_ctl.rc_went_idle_time = 1;
12830 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12831 if (sbavail(&tptosocket(tp)->so_snd) == 0)
12832 tp->t_acktime = 0;
12833 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12834 }
12835 if (acked && rack->r_fast_output)
12836 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
12837 if (sbavail(&so->so_snd)) {
12838 rack->r_wanted_output = 1;
12839 }
12840 return (1);
12841 }
12842
12843 /*
12844 * Return value of 1, the TCB is unlocked and most
12845 * likely gone, return value of 0, the TCP is still
12846 * locked.
12847 */
12848 static int
rack_do_syn_sent(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)12849 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
12850 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12851 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12852 {
12853 int32_t ret_val = 0;
12854 int32_t orig_tlen = tlen;
12855 int32_t todrop;
12856 int32_t ourfinisacked = 0;
12857 struct tcp_rack *rack;
12858
12859 INP_WLOCK_ASSERT(tptoinpcb(tp));
12860
12861 ctf_calc_rwin(so, tp);
12862 /*
12863 * If the state is SYN_SENT: if seg contains an ACK, but not for our
12864 * SYN, drop the input. if seg contains a RST, then drop the
12865 * connection. if seg does not contain SYN, then drop it. Otherwise
12866 * this is an acceptable SYN segment initialize tp->rcv_nxt and
12867 * tp->irs if seg contains ack then advance tp->snd_una if seg
12868 * contains an ECE and ECN support is enabled, the stream is ECN
12869 * capable. if SYN has been acked change to ESTABLISHED else
12870 * SYN_RCVD state arrange for segment to be acked (eventually)
12871 * continue processing rest of data/controls.
12872 */
12873 if ((thflags & TH_ACK) &&
12874 (SEQ_LEQ(th->th_ack, tp->iss) ||
12875 SEQ_GT(th->th_ack, tp->snd_max))) {
12876 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
12877 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12878 return (1);
12879 }
12880 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
12881 TCP_PROBE5(connect__refused, NULL, tp,
12882 mtod(m, const char *), tp, th);
12883 tp = tcp_drop(tp, ECONNREFUSED);
12884 ctf_do_drop(m, tp);
12885 return (1);
12886 }
12887 if (thflags & TH_RST) {
12888 ctf_do_drop(m, tp);
12889 return (1);
12890 }
12891 if (!(thflags & TH_SYN)) {
12892 ctf_do_drop(m, tp);
12893 return (1);
12894 }
12895 tp->irs = th->th_seq;
12896 tcp_rcvseqinit(tp);
12897 rack = (struct tcp_rack *)tp->t_fb_ptr;
12898 if (thflags & TH_ACK) {
12899 int tfo_partial = 0;
12900
12901 KMOD_TCPSTAT_INC(tcps_connects);
12902 soisconnected(so);
12903 #ifdef MAC
12904 mac_socketpeer_set_from_mbuf(m, so);
12905 #endif
12906 /* Do window scaling on this connection? */
12907 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12908 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12909 tp->rcv_scale = tp->request_r_scale;
12910 }
12911 tp->rcv_adv += min(tp->rcv_wnd,
12912 TCP_MAXWIN << tp->rcv_scale);
12913 /*
12914 * If not all the data that was sent in the TFO SYN
12915 * has been acked, resend the remainder right away.
12916 */
12917 if ((tp->t_flags & TF_FASTOPEN) &&
12918 (tp->snd_una != tp->snd_max)) {
12919 /* Was it a partial ack? */
12920 if (SEQ_LT(th->th_ack, tp->snd_max))
12921 tfo_partial = 1;
12922 }
12923 /*
12924 * If there's data, delay ACK; if there's also a FIN ACKNOW
12925 * will be turned on later.
12926 */
12927 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
12928 rack_timer_cancel(tp, rack,
12929 rack->r_ctl.rc_rcvtime, __LINE__);
12930 tp->t_flags |= TF_DELACK;
12931 } else {
12932 rack->r_wanted_output = 1;
12933 tp->t_flags |= TF_ACKNOW;
12934 }
12935
12936 tcp_ecn_input_syn_sent(tp, thflags, iptos);
12937
12938 if (SEQ_GT(th->th_ack, tp->snd_una)) {
12939 /*
12940 * We advance snd_una for the
12941 * fast open case. If th_ack is
12942 * acknowledging data beyond
12943 * snd_una we can't just call
12944 * ack-processing since the
12945 * data stream in our send-map
12946 * will start at snd_una + 1 (one
12947 * beyond the SYN). If its just
12948 * equal we don't need to do that
12949 * and there is no send_map.
12950 */
12951 tp->snd_una++;
12952 if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
12953 /*
12954 * We sent a SYN with data, and thus have a
12955 * sendmap entry with a SYN set. Lets find it
12956 * and take off the send bit and the byte and
12957 * set it up to be what we send (send it next).
12958 */
12959 struct rack_sendmap *rsm;
12960
12961 rsm = tqhash_min(rack->r_ctl.tqh);
12962 if (rsm) {
12963 if (rsm->r_flags & RACK_HAS_SYN) {
12964 rsm->r_flags &= ~RACK_HAS_SYN;
12965 rsm->r_start++;
12966 }
12967 rack->r_ctl.rc_resend = rsm;
12968 }
12969 }
12970 }
12971 /*
12972 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
12973 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
12974 */
12975 tp->t_starttime = ticks;
12976 if (tp->t_flags & TF_NEEDFIN) {
12977 tcp_state_change(tp, TCPS_FIN_WAIT_1);
12978 tp->t_flags &= ~TF_NEEDFIN;
12979 thflags &= ~TH_SYN;
12980 } else {
12981 tcp_state_change(tp, TCPS_ESTABLISHED);
12982 TCP_PROBE5(connect__established, NULL, tp,
12983 mtod(m, const char *), tp, th);
12984 rack_cc_conn_init(tp);
12985 }
12986 } else {
12987 /*
12988 * Received initial SYN in SYN-SENT[*] state => simultaneous
12989 * open. If segment contains CC option and there is a
12990 * cached CC, apply TAO test. If it succeeds, connection is *
12991 * half-synchronized. Otherwise, do 3-way handshake:
12992 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
12993 * there was no CC option, clear cached CC value.
12994 */
12995 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
12996 tcp_state_change(tp, TCPS_SYN_RECEIVED);
12997 }
12998 /*
12999 * Advance th->th_seq to correspond to first data byte. If data,
13000 * trim to stay within window, dropping FIN if necessary.
13001 */
13002 th->th_seq++;
13003 if (tlen > tp->rcv_wnd) {
13004 todrop = tlen - tp->rcv_wnd;
13005 m_adj(m, -todrop);
13006 tlen = tp->rcv_wnd;
13007 thflags &= ~TH_FIN;
13008 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13009 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13010 }
13011 tp->snd_wl1 = th->th_seq - 1;
13012 tp->rcv_up = th->th_seq;
13013 /*
13014 * Client side of transaction: already sent SYN and data. If the
13015 * remote host used T/TCP to validate the SYN, our data will be
13016 * ACK'd; if so, enter normal data segment processing in the middle
13017 * of step 5, ack processing. Otherwise, goto step 6.
13018 */
13019 if (thflags & TH_ACK) {
13020 /* For syn-sent we need to possibly update the rtt */
13021 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13022 uint32_t t, mcts;
13023
13024 mcts = tcp_ts_getticks();
13025 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13026 if (!tp->t_rttlow || tp->t_rttlow > t)
13027 tp->t_rttlow = t;
13028 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13029 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13030 tcp_rack_xmit_timer_commit(rack, tp);
13031 }
13032 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
13033 return (ret_val);
13034 /* We may have changed to FIN_WAIT_1 above */
13035 if (tp->t_state == TCPS_FIN_WAIT_1) {
13036 /*
13037 * In FIN_WAIT_1 STATE in addition to the processing
13038 * for the ESTABLISHED state if our FIN is now
13039 * acknowledged then enter FIN_WAIT_2.
13040 */
13041 if (ourfinisacked) {
13042 /*
13043 * If we can't receive any more data, then
13044 * closing user can proceed. Starting the
13045 * timer is contrary to the specification,
13046 * but if we don't get a FIN we'll hang
13047 * forever.
13048 *
13049 * XXXjl: we should release the tp also, and
13050 * use a compressed state.
13051 */
13052 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13053 soisdisconnected(so);
13054 tcp_timer_activate(tp, TT_2MSL,
13055 (tcp_fast_finwait2_recycle ?
13056 tcp_finwait2_timeout :
13057 TP_MAXIDLE(tp)));
13058 }
13059 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13060 }
13061 }
13062 }
13063 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13064 tiwin, thflags, nxt_pkt));
13065 }
13066
13067 /*
13068 * Return value of 1, the TCB is unlocked and most
13069 * likely gone, return value of 0, the TCP is still
13070 * locked.
13071 */
13072 static int
rack_do_syn_recv(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13073 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13074 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13075 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13076 {
13077 struct tcp_rack *rack;
13078 int32_t orig_tlen = tlen;
13079 int32_t ret_val = 0;
13080 int32_t ourfinisacked = 0;
13081
13082 rack = (struct tcp_rack *)tp->t_fb_ptr;
13083 ctf_calc_rwin(so, tp);
13084 if ((thflags & TH_RST) ||
13085 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13086 return (ctf_process_rst(m, th, so, tp));
13087 if ((thflags & TH_ACK) &&
13088 (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13089 SEQ_GT(th->th_ack, tp->snd_max))) {
13090 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13091 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13092 return (1);
13093 }
13094 if (tp->t_flags & TF_FASTOPEN) {
13095 /*
13096 * When a TFO connection is in SYN_RECEIVED, the
13097 * only valid packets are the initial SYN, a
13098 * retransmit/copy of the initial SYN (possibly with
13099 * a subset of the original data), a valid ACK, a
13100 * FIN, or a RST.
13101 */
13102 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13103 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13104 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13105 return (1);
13106 } else if (thflags & TH_SYN) {
13107 /* non-initial SYN is ignored */
13108 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13109 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13110 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13111 ctf_do_drop(m, NULL);
13112 return (0);
13113 }
13114 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13115 ctf_do_drop(m, NULL);
13116 return (0);
13117 }
13118 }
13119
13120 /*
13121 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13122 * it's less than ts_recent, drop it.
13123 */
13124 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13125 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13126 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13127 return (ret_val);
13128 }
13129 /*
13130 * In the SYN-RECEIVED state, validate that the packet belongs to
13131 * this connection before trimming the data to fit the receive
13132 * window. Check the sequence number versus IRS since we know the
13133 * sequence numbers haven't wrapped. This is a partial fix for the
13134 * "LAND" DoS attack.
13135 */
13136 if (SEQ_LT(th->th_seq, tp->irs)) {
13137 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13138 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13139 return (1);
13140 }
13141 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13142 return (ret_val);
13143 }
13144 /*
13145 * If last ACK falls within this segment's sequence numbers, record
13146 * its timestamp. NOTE: 1) That the test incorporates suggestions
13147 * from the latest proposal of the tcplw@cray.com list (Braden
13148 * 1993/04/26). 2) That updating only on newer timestamps interferes
13149 * with our earlier PAWS tests, so this check should be solely
13150 * predicated on the sequence space of this segment. 3) That we
13151 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13152 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13153 * SEG.Len, This modified check allows us to overcome RFC1323's
13154 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13155 * p.869. In such cases, we can still calculate the RTT correctly
13156 * when RCV.NXT == Last.ACK.Sent.
13157 */
13158 if ((to->to_flags & TOF_TS) != 0 &&
13159 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13160 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13161 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13162 tp->ts_recent_age = tcp_ts_getticks();
13163 tp->ts_recent = to->to_tsval;
13164 }
13165 tp->snd_wnd = tiwin;
13166 rack_validate_fo_sendwin_up(tp, rack);
13167 /*
13168 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13169 * is on (half-synchronized state), then queue data for later
13170 * processing; else drop segment and return.
13171 */
13172 if ((thflags & TH_ACK) == 0) {
13173 if (tp->t_flags & TF_FASTOPEN) {
13174 rack_cc_conn_init(tp);
13175 }
13176 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13177 tiwin, thflags, nxt_pkt));
13178 }
13179 KMOD_TCPSTAT_INC(tcps_connects);
13180 if (tp->t_flags & TF_SONOTCONN) {
13181 tp->t_flags &= ~TF_SONOTCONN;
13182 soisconnected(so);
13183 }
13184 /* Do window scaling? */
13185 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13186 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13187 tp->rcv_scale = tp->request_r_scale;
13188 }
13189 /*
13190 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* ->
13191 * FIN-WAIT-1
13192 */
13193 tp->t_starttime = ticks;
13194 if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
13195 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13196 tp->t_tfo_pending = NULL;
13197 }
13198 if (tp->t_flags & TF_NEEDFIN) {
13199 tcp_state_change(tp, TCPS_FIN_WAIT_1);
13200 tp->t_flags &= ~TF_NEEDFIN;
13201 } else {
13202 tcp_state_change(tp, TCPS_ESTABLISHED);
13203 TCP_PROBE5(accept__established, NULL, tp,
13204 mtod(m, const char *), tp, th);
13205 /*
13206 * TFO connections call cc_conn_init() during SYN
13207 * processing. Calling it again here for such connections
13208 * is not harmless as it would undo the snd_cwnd reduction
13209 * that occurs when a TFO SYN|ACK is retransmitted.
13210 */
13211 if (!(tp->t_flags & TF_FASTOPEN))
13212 rack_cc_conn_init(tp);
13213 }
13214 /*
13215 * Account for the ACK of our SYN prior to
13216 * regular ACK processing below, except for
13217 * simultaneous SYN, which is handled later.
13218 */
13219 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13220 tp->snd_una++;
13221 /*
13222 * If segment contains data or ACK, will call tcp_reass() later; if
13223 * not, do so now to pass queued data to user.
13224 */
13225 if (tlen == 0 && (thflags & TH_FIN) == 0) {
13226 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13227 (struct mbuf *)0);
13228 if (tp->t_flags & TF_WAKESOR) {
13229 tp->t_flags &= ~TF_WAKESOR;
13230 /* NB: sorwakeup_locked() does an implicit unlock. */
13231 sorwakeup_locked(so);
13232 }
13233 }
13234 tp->snd_wl1 = th->th_seq - 1;
13235 /* For syn-recv we need to possibly update the rtt */
13236 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13237 uint32_t t, mcts;
13238
13239 mcts = tcp_ts_getticks();
13240 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13241 if (!tp->t_rttlow || tp->t_rttlow > t)
13242 tp->t_rttlow = t;
13243 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13244 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13245 tcp_rack_xmit_timer_commit(rack, tp);
13246 }
13247 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13248 return (ret_val);
13249 }
13250 if (tp->t_state == TCPS_FIN_WAIT_1) {
13251 /* We could have went to FIN_WAIT_1 (or EST) above */
13252 /*
13253 * In FIN_WAIT_1 STATE in addition to the processing for the
13254 * ESTABLISHED state if our FIN is now acknowledged then
13255 * enter FIN_WAIT_2.
13256 */
13257 if (ourfinisacked) {
13258 /*
13259 * If we can't receive any more data, then closing
13260 * user can proceed. Starting the timer is contrary
13261 * to the specification, but if we don't get a FIN
13262 * we'll hang forever.
13263 *
13264 * XXXjl: we should release the tp also, and use a
13265 * compressed state.
13266 */
13267 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13268 soisdisconnected(so);
13269 tcp_timer_activate(tp, TT_2MSL,
13270 (tcp_fast_finwait2_recycle ?
13271 tcp_finwait2_timeout :
13272 TP_MAXIDLE(tp)));
13273 }
13274 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13275 }
13276 }
13277 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13278 tiwin, thflags, nxt_pkt));
13279 }
13280
13281 /*
13282 * Return value of 1, the TCB is unlocked and most
13283 * likely gone, return value of 0, the TCP is still
13284 * locked.
13285 */
13286 static int
rack_do_established(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13287 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13288 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13289 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13290 {
13291 int32_t ret_val = 0;
13292 int32_t orig_tlen = tlen;
13293 struct tcp_rack *rack;
13294
13295 /*
13296 * Header prediction: check for the two common cases of a
13297 * uni-directional data xfer. If the packet has no control flags,
13298 * is in-sequence, the window didn't change and we're not
13299 * retransmitting, it's a candidate. If the length is zero and the
13300 * ack moved forward, we're the sender side of the xfer. Just free
13301 * the data acked & wake any higher level process that was blocked
13302 * waiting for space. If the length is non-zero and the ack didn't
13303 * move, we're the receiver side. If we're getting packets in-order
13304 * (the reassembly queue is empty), add the data toc The socket
13305 * buffer and note that we need a delayed ack. Make sure that the
13306 * hidden state-flags are also off. Since we check for
13307 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13308 */
13309 rack = (struct tcp_rack *)tp->t_fb_ptr;
13310 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13311 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13312 __predict_true(SEGQ_EMPTY(tp)) &&
13313 __predict_true(th->th_seq == tp->rcv_nxt)) {
13314 if (tlen == 0) {
13315 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13316 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13317 return (0);
13318 }
13319 } else {
13320 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13321 tiwin, nxt_pkt, iptos)) {
13322 return (0);
13323 }
13324 }
13325 }
13326 ctf_calc_rwin(so, tp);
13327
13328 if ((thflags & TH_RST) ||
13329 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13330 return (ctf_process_rst(m, th, so, tp));
13331
13332 /*
13333 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13334 * synchronized state.
13335 */
13336 if (thflags & TH_SYN) {
13337 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13338 return (ret_val);
13339 }
13340 /*
13341 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13342 * it's less than ts_recent, drop it.
13343 */
13344 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13345 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13346 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13347 return (ret_val);
13348 }
13349 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13350 return (ret_val);
13351 }
13352 /*
13353 * If last ACK falls within this segment's sequence numbers, record
13354 * its timestamp. NOTE: 1) That the test incorporates suggestions
13355 * from the latest proposal of the tcplw@cray.com list (Braden
13356 * 1993/04/26). 2) That updating only on newer timestamps interferes
13357 * with our earlier PAWS tests, so this check should be solely
13358 * predicated on the sequence space of this segment. 3) That we
13359 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13360 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13361 * SEG.Len, This modified check allows us to overcome RFC1323's
13362 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13363 * p.869. In such cases, we can still calculate the RTT correctly
13364 * when RCV.NXT == Last.ACK.Sent.
13365 */
13366 if ((to->to_flags & TOF_TS) != 0 &&
13367 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13368 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13369 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13370 tp->ts_recent_age = tcp_ts_getticks();
13371 tp->ts_recent = to->to_tsval;
13372 }
13373 /*
13374 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13375 * is on (half-synchronized state), then queue data for later
13376 * processing; else drop segment and return.
13377 */
13378 if ((thflags & TH_ACK) == 0) {
13379 if (tp->t_flags & TF_NEEDSYN) {
13380 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13381 tiwin, thflags, nxt_pkt));
13382
13383 } else if (tp->t_flags & TF_ACKNOW) {
13384 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13385 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13386 return (ret_val);
13387 } else {
13388 ctf_do_drop(m, NULL);
13389 return (0);
13390 }
13391 }
13392 /*
13393 * Ack processing.
13394 */
13395 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13396 return (ret_val);
13397 }
13398 if (sbavail(&so->so_snd)) {
13399 if (ctf_progress_timeout_check(tp, true)) {
13400 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13401 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13402 return (1);
13403 }
13404 }
13405 /* State changes only happen in rack_process_data() */
13406 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13407 tiwin, thflags, nxt_pkt));
13408 }
13409
13410 /*
13411 * Return value of 1, the TCB is unlocked and most
13412 * likely gone, return value of 0, the TCP is still
13413 * locked.
13414 */
13415 static int
rack_do_close_wait(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13416 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13417 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13418 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13419 {
13420 int32_t ret_val = 0;
13421 int32_t orig_tlen = tlen;
13422
13423 ctf_calc_rwin(so, tp);
13424 if ((thflags & TH_RST) ||
13425 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13426 return (ctf_process_rst(m, th, so, tp));
13427 /*
13428 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13429 * synchronized state.
13430 */
13431 if (thflags & TH_SYN) {
13432 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13433 return (ret_val);
13434 }
13435 /*
13436 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13437 * it's less than ts_recent, drop it.
13438 */
13439 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13440 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13441 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13442 return (ret_val);
13443 }
13444 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13445 return (ret_val);
13446 }
13447 /*
13448 * If last ACK falls within this segment's sequence numbers, record
13449 * its timestamp. NOTE: 1) That the test incorporates suggestions
13450 * from the latest proposal of the tcplw@cray.com list (Braden
13451 * 1993/04/26). 2) That updating only on newer timestamps interferes
13452 * with our earlier PAWS tests, so this check should be solely
13453 * predicated on the sequence space of this segment. 3) That we
13454 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13455 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13456 * SEG.Len, This modified check allows us to overcome RFC1323's
13457 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13458 * p.869. In such cases, we can still calculate the RTT correctly
13459 * when RCV.NXT == Last.ACK.Sent.
13460 */
13461 if ((to->to_flags & TOF_TS) != 0 &&
13462 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13463 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13464 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13465 tp->ts_recent_age = tcp_ts_getticks();
13466 tp->ts_recent = to->to_tsval;
13467 }
13468 /*
13469 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13470 * is on (half-synchronized state), then queue data for later
13471 * processing; else drop segment and return.
13472 */
13473 if ((thflags & TH_ACK) == 0) {
13474 if (tp->t_flags & TF_NEEDSYN) {
13475 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13476 tiwin, thflags, nxt_pkt));
13477
13478 } else if (tp->t_flags & TF_ACKNOW) {
13479 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13480 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13481 return (ret_val);
13482 } else {
13483 ctf_do_drop(m, NULL);
13484 return (0);
13485 }
13486 }
13487 /*
13488 * Ack processing.
13489 */
13490 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13491 return (ret_val);
13492 }
13493 if (sbavail(&so->so_snd)) {
13494 if (ctf_progress_timeout_check(tp, true)) {
13495 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13496 tp, tick, PROGRESS_DROP, __LINE__);
13497 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13498 return (1);
13499 }
13500 }
13501 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13502 tiwin, thflags, nxt_pkt));
13503 }
13504
13505 static int
rack_check_data_after_close(struct mbuf * m,struct tcpcb * tp,int32_t * tlen,struct tcphdr * th,struct socket * so)13506 rack_check_data_after_close(struct mbuf *m,
13507 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13508 {
13509 struct tcp_rack *rack;
13510
13511 rack = (struct tcp_rack *)tp->t_fb_ptr;
13512 if (rack->rc_allow_data_af_clo == 0) {
13513 close_now:
13514 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13515 /* tcp_close will kill the inp pre-log the Reset */
13516 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13517 tp = tcp_close(tp);
13518 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13519 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13520 return (1);
13521 }
13522 if (sbavail(&so->so_snd) == 0)
13523 goto close_now;
13524 /* Ok we allow data that is ignored and a followup reset */
13525 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13526 tp->rcv_nxt = th->th_seq + *tlen;
13527 tp->t_flags2 |= TF2_DROP_AF_DATA;
13528 rack->r_wanted_output = 1;
13529 *tlen = 0;
13530 return (0);
13531 }
13532
13533 /*
13534 * Return value of 1, the TCB is unlocked and most
13535 * likely gone, return value of 0, the TCP is still
13536 * locked.
13537 */
13538 static int
rack_do_fin_wait_1(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13539 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13540 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13541 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13542 {
13543 int32_t ret_val = 0;
13544 int32_t orig_tlen = tlen;
13545 int32_t ourfinisacked = 0;
13546
13547 ctf_calc_rwin(so, tp);
13548
13549 if ((thflags & TH_RST) ||
13550 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13551 return (ctf_process_rst(m, th, so, tp));
13552 /*
13553 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13554 * synchronized state.
13555 */
13556 if (thflags & TH_SYN) {
13557 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13558 return (ret_val);
13559 }
13560 /*
13561 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13562 * it's less than ts_recent, drop it.
13563 */
13564 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13565 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13566 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13567 return (ret_val);
13568 }
13569 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13570 return (ret_val);
13571 }
13572 /*
13573 * If new data are received on a connection after the user processes
13574 * are gone, then RST the other end.
13575 */
13576 if ((tp->t_flags & TF_CLOSED) && tlen &&
13577 rack_check_data_after_close(m, tp, &tlen, th, so))
13578 return (1);
13579 /*
13580 * If last ACK falls within this segment's sequence numbers, record
13581 * its timestamp. NOTE: 1) That the test incorporates suggestions
13582 * from the latest proposal of the tcplw@cray.com list (Braden
13583 * 1993/04/26). 2) That updating only on newer timestamps interferes
13584 * with our earlier PAWS tests, so this check should be solely
13585 * predicated on the sequence space of this segment. 3) That we
13586 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13587 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13588 * SEG.Len, This modified check allows us to overcome RFC1323's
13589 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13590 * p.869. In such cases, we can still calculate the RTT correctly
13591 * when RCV.NXT == Last.ACK.Sent.
13592 */
13593 if ((to->to_flags & TOF_TS) != 0 &&
13594 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13595 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13596 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13597 tp->ts_recent_age = tcp_ts_getticks();
13598 tp->ts_recent = to->to_tsval;
13599 }
13600 /*
13601 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13602 * is on (half-synchronized state), then queue data for later
13603 * processing; else drop segment and return.
13604 */
13605 if ((thflags & TH_ACK) == 0) {
13606 if (tp->t_flags & TF_NEEDSYN) {
13607 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13608 tiwin, thflags, nxt_pkt));
13609 } else if (tp->t_flags & TF_ACKNOW) {
13610 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13611 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13612 return (ret_val);
13613 } else {
13614 ctf_do_drop(m, NULL);
13615 return (0);
13616 }
13617 }
13618 /*
13619 * Ack processing.
13620 */
13621 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13622 return (ret_val);
13623 }
13624 if (ourfinisacked) {
13625 /*
13626 * If we can't receive any more data, then closing user can
13627 * proceed. Starting the timer is contrary to the
13628 * specification, but if we don't get a FIN we'll hang
13629 * forever.
13630 *
13631 * XXXjl: we should release the tp also, and use a
13632 * compressed state.
13633 */
13634 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13635 soisdisconnected(so);
13636 tcp_timer_activate(tp, TT_2MSL,
13637 (tcp_fast_finwait2_recycle ?
13638 tcp_finwait2_timeout :
13639 TP_MAXIDLE(tp)));
13640 }
13641 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13642 }
13643 if (sbavail(&so->so_snd)) {
13644 if (ctf_progress_timeout_check(tp, true)) {
13645 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13646 tp, tick, PROGRESS_DROP, __LINE__);
13647 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13648 return (1);
13649 }
13650 }
13651 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13652 tiwin, thflags, nxt_pkt));
13653 }
13654
13655 /*
13656 * Return value of 1, the TCB is unlocked and most
13657 * likely gone, return value of 0, the TCP is still
13658 * locked.
13659 */
13660 static int
rack_do_closing(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13661 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13662 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13663 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13664 {
13665 int32_t ret_val = 0;
13666 int32_t orig_tlen = tlen;
13667 int32_t ourfinisacked = 0;
13668
13669 ctf_calc_rwin(so, tp);
13670
13671 if ((thflags & TH_RST) ||
13672 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13673 return (ctf_process_rst(m, th, so, tp));
13674 /*
13675 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13676 * synchronized state.
13677 */
13678 if (thflags & TH_SYN) {
13679 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13680 return (ret_val);
13681 }
13682 /*
13683 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13684 * it's less than ts_recent, drop it.
13685 */
13686 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13687 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13688 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13689 return (ret_val);
13690 }
13691 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13692 return (ret_val);
13693 }
13694 /*
13695 * If last ACK falls within this segment's sequence numbers, record
13696 * its timestamp. NOTE: 1) That the test incorporates suggestions
13697 * from the latest proposal of the tcplw@cray.com list (Braden
13698 * 1993/04/26). 2) That updating only on newer timestamps interferes
13699 * with our earlier PAWS tests, so this check should be solely
13700 * predicated on the sequence space of this segment. 3) That we
13701 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13702 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13703 * SEG.Len, This modified check allows us to overcome RFC1323's
13704 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13705 * p.869. In such cases, we can still calculate the RTT correctly
13706 * when RCV.NXT == Last.ACK.Sent.
13707 */
13708 if ((to->to_flags & TOF_TS) != 0 &&
13709 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13710 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13711 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13712 tp->ts_recent_age = tcp_ts_getticks();
13713 tp->ts_recent = to->to_tsval;
13714 }
13715 /*
13716 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13717 * is on (half-synchronized state), then queue data for later
13718 * processing; else drop segment and return.
13719 */
13720 if ((thflags & TH_ACK) == 0) {
13721 if (tp->t_flags & TF_NEEDSYN) {
13722 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13723 tiwin, thflags, nxt_pkt));
13724 } else if (tp->t_flags & TF_ACKNOW) {
13725 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13726 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13727 return (ret_val);
13728 } else {
13729 ctf_do_drop(m, NULL);
13730 return (0);
13731 }
13732 }
13733 /*
13734 * Ack processing.
13735 */
13736 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13737 return (ret_val);
13738 }
13739 if (ourfinisacked) {
13740 tcp_twstart(tp);
13741 m_freem(m);
13742 return (1);
13743 }
13744 if (sbavail(&so->so_snd)) {
13745 if (ctf_progress_timeout_check(tp, true)) {
13746 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13747 tp, tick, PROGRESS_DROP, __LINE__);
13748 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13749 return (1);
13750 }
13751 }
13752 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13753 tiwin, thflags, nxt_pkt));
13754 }
13755
13756 /*
13757 * Return value of 1, the TCB is unlocked and most
13758 * likely gone, return value of 0, the TCP is still
13759 * locked.
13760 */
13761 static int
rack_do_lastack(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13762 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13763 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13764 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13765 {
13766 int32_t ret_val = 0;
13767 int32_t orig_tlen;
13768 int32_t ourfinisacked = 0;
13769
13770 ctf_calc_rwin(so, tp);
13771
13772 if ((thflags & TH_RST) ||
13773 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13774 return (ctf_process_rst(m, th, so, tp));
13775 /*
13776 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13777 * synchronized state.
13778 */
13779 if (thflags & TH_SYN) {
13780 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13781 return (ret_val);
13782 }
13783 /*
13784 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13785 * it's less than ts_recent, drop it.
13786 */
13787 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13788 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13789 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13790 return (ret_val);
13791 }
13792 orig_tlen = tlen;
13793 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13794 return (ret_val);
13795 }
13796 /*
13797 * If last ACK falls within this segment's sequence numbers, record
13798 * its timestamp. NOTE: 1) That the test incorporates suggestions
13799 * from the latest proposal of the tcplw@cray.com list (Braden
13800 * 1993/04/26). 2) That updating only on newer timestamps interferes
13801 * with our earlier PAWS tests, so this check should be solely
13802 * predicated on the sequence space of this segment. 3) That we
13803 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13804 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13805 * SEG.Len, This modified check allows us to overcome RFC1323's
13806 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13807 * p.869. In such cases, we can still calculate the RTT correctly
13808 * when RCV.NXT == Last.ACK.Sent.
13809 */
13810 if ((to->to_flags & TOF_TS) != 0 &&
13811 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13812 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13813 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13814 tp->ts_recent_age = tcp_ts_getticks();
13815 tp->ts_recent = to->to_tsval;
13816 }
13817 /*
13818 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13819 * is on (half-synchronized state), then queue data for later
13820 * processing; else drop segment and return.
13821 */
13822 if ((thflags & TH_ACK) == 0) {
13823 if (tp->t_flags & TF_NEEDSYN) {
13824 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13825 tiwin, thflags, nxt_pkt));
13826 } else if (tp->t_flags & TF_ACKNOW) {
13827 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13828 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13829 return (ret_val);
13830 } else {
13831 ctf_do_drop(m, NULL);
13832 return (0);
13833 }
13834 }
13835 /*
13836 * case TCPS_LAST_ACK: Ack processing.
13837 */
13838 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13839 return (ret_val);
13840 }
13841 if (ourfinisacked) {
13842 tp = tcp_close(tp);
13843 ctf_do_drop(m, tp);
13844 return (1);
13845 }
13846 if (sbavail(&so->so_snd)) {
13847 if (ctf_progress_timeout_check(tp, true)) {
13848 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13849 tp, tick, PROGRESS_DROP, __LINE__);
13850 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13851 return (1);
13852 }
13853 }
13854 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13855 tiwin, thflags, nxt_pkt));
13856 }
13857
13858 /*
13859 * Return value of 1, the TCB is unlocked and most
13860 * likely gone, return value of 0, the TCP is still
13861 * locked.
13862 */
13863 static int
rack_do_fin_wait_2(struct mbuf * m,struct tcphdr * th,struct socket * so,struct tcpcb * tp,struct tcpopt * to,int32_t drop_hdrlen,int32_t tlen,uint32_t tiwin,int32_t thflags,int32_t nxt_pkt,uint8_t iptos)13864 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
13865 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13866 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13867 {
13868 int32_t ret_val = 0;
13869 int32_t orig_tlen = tlen;
13870 int32_t ourfinisacked = 0;
13871
13872 ctf_calc_rwin(so, tp);
13873
13874 /* Reset receive buffer auto scaling when not in bulk receive mode. */
13875 if ((thflags & TH_RST) ||
13876 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13877 return (ctf_process_rst(m, th, so, tp));
13878 /*
13879 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13880 * synchronized state.
13881 */
13882 if (thflags & TH_SYN) {
13883 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13884 return (ret_val);
13885 }
13886 /*
13887 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13888 * it's less than ts_recent, drop it.
13889 */
13890 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13891 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13892 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13893 return (ret_val);
13894 }
13895 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13896 return (ret_val);
13897 }
13898 /*
13899 * If new data are received on a connection after the user processes
13900 * are gone, then RST the other end.
13901 */
13902 if ((tp->t_flags & TF_CLOSED) && tlen &&
13903 rack_check_data_after_close(m, tp, &tlen, th, so))
13904 return (1);
13905 /*
13906 * If last ACK falls within this segment's sequence numbers, record
13907 * its timestamp. NOTE: 1) That the test incorporates suggestions
13908 * from the latest proposal of the tcplw@cray.com list (Braden
13909 * 1993/04/26). 2) That updating only on newer timestamps interferes
13910 * with our earlier PAWS tests, so this check should be solely
13911 * predicated on the sequence space of this segment. 3) That we
13912 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13913 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13914 * SEG.Len, This modified check allows us to overcome RFC1323's
13915 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13916 * p.869. In such cases, we can still calculate the RTT correctly
13917 * when RCV.NXT == Last.ACK.Sent.
13918 */
13919 if ((to->to_flags & TOF_TS) != 0 &&
13920 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13921 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13922 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13923 tp->ts_recent_age = tcp_ts_getticks();
13924 tp->ts_recent = to->to_tsval;
13925 }
13926 /*
13927 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13928 * is on (half-synchronized state), then queue data for later
13929 * processing; else drop segment and return.
13930 */
13931 if ((thflags & TH_ACK) == 0) {
13932 if (tp->t_flags & TF_NEEDSYN) {
13933 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13934 tiwin, thflags, nxt_pkt));
13935 } else if (tp->t_flags & TF_ACKNOW) {
13936 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13937 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13938 return (ret_val);
13939 } else {
13940 ctf_do_drop(m, NULL);
13941 return (0);
13942 }
13943 }
13944 /*
13945 * Ack processing.
13946 */
13947 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13948 return (ret_val);
13949 }
13950 if (sbavail(&so->so_snd)) {
13951 if (ctf_progress_timeout_check(tp, true)) {
13952 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13953 tp, tick, PROGRESS_DROP, __LINE__);
13954 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13955 return (1);
13956 }
13957 }
13958 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13959 tiwin, thflags, nxt_pkt));
13960 }
13961
13962 static void inline
rack_clear_rate_sample(struct tcp_rack * rack)13963 rack_clear_rate_sample(struct tcp_rack *rack)
13964 {
13965 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
13966 rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
13967 rack->r_ctl.rack_rs.rs_rtt_tot = 0;
13968 }
13969
13970 static void
rack_set_pace_segments(struct tcpcb * tp,struct tcp_rack * rack,uint32_t line,uint64_t * fill_override)13971 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
13972 {
13973 uint64_t bw_est, rate_wanted;
13974 int chged = 0;
13975 uint32_t user_max, orig_min, orig_max;
13976
13977 #ifdef TCP_REQUEST_TRK
13978 if (rack->rc_hybrid_mode &&
13979 (rack->r_ctl.rc_pace_max_segs != 0) &&
13980 (rack_hybrid_allow_set_maxseg == 1) &&
13981 (rack->r_ctl.rc_last_sft != NULL)) {
13982 rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
13983 return;
13984 }
13985 #endif
13986 orig_min = rack->r_ctl.rc_pace_min_segs;
13987 orig_max = rack->r_ctl.rc_pace_max_segs;
13988 user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
13989 if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
13990 chged = 1;
13991 rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
13992 if (rack->use_fixed_rate || rack->rc_force_max_seg) {
13993 if (user_max != rack->r_ctl.rc_pace_max_segs)
13994 chged = 1;
13995 }
13996 if (rack->rc_force_max_seg) {
13997 rack->r_ctl.rc_pace_max_segs = user_max;
13998 } else if (rack->use_fixed_rate) {
13999 bw_est = rack_get_bw(rack);
14000 if ((rack->r_ctl.crte == NULL) ||
14001 (bw_est != rack->r_ctl.crte->rate)) {
14002 rack->r_ctl.rc_pace_max_segs = user_max;
14003 } else {
14004 /* We are pacing right at the hardware rate */
14005 uint32_t segsiz, pace_one;
14006
14007 if (rack_pace_one_seg ||
14008 (rack->r_ctl.rc_user_set_min_segs == 1))
14009 pace_one = 1;
14010 else
14011 pace_one = 0;
14012 segsiz = min(ctf_fixed_maxseg(tp),
14013 rack->r_ctl.rc_pace_min_segs);
14014 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14015 tp, bw_est, segsiz, pace_one,
14016 rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14017 }
14018 } else if (rack->rc_always_pace) {
14019 if (rack->r_ctl.gp_bw ||
14020 rack->r_ctl.init_rate) {
14021 /* We have a rate of some sort set */
14022 uint32_t orig;
14023
14024 bw_est = rack_get_bw(rack);
14025 orig = rack->r_ctl.rc_pace_max_segs;
14026 if (fill_override)
14027 rate_wanted = *fill_override;
14028 else
14029 rate_wanted = rack_get_gp_est(rack);
14030 if (rate_wanted) {
14031 /* We have something */
14032 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14033 rate_wanted,
14034 ctf_fixed_maxseg(rack->rc_tp));
14035 } else
14036 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14037 if (orig != rack->r_ctl.rc_pace_max_segs)
14038 chged = 1;
14039 } else if ((rack->r_ctl.gp_bw == 0) &&
14040 (rack->r_ctl.rc_pace_max_segs == 0)) {
14041 /*
14042 * If we have nothing limit us to bursting
14043 * out IW sized pieces.
14044 */
14045 chged = 1;
14046 rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14047 }
14048 }
14049 if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14050 chged = 1;
14051 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14052 }
14053 if (chged)
14054 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14055 }
14056
14057
14058 static void
rack_init_fsb_block(struct tcpcb * tp,struct tcp_rack * rack,int32_t flags)14059 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14060 {
14061 #ifdef INET6
14062 struct ip6_hdr *ip6 = NULL;
14063 #endif
14064 #ifdef INET
14065 struct ip *ip = NULL;
14066 #endif
14067 struct udphdr *udp = NULL;
14068
14069 /* Ok lets fill in the fast block, it can only be used with no IP options! */
14070 #ifdef INET6
14071 if (rack->r_is_v6) {
14072 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14073 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14074 if (tp->t_port) {
14075 rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14076 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14077 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14078 udp->uh_dport = tp->t_port;
14079 rack->r_ctl.fsb.udp = udp;
14080 rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14081 } else
14082 {
14083 rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14084 rack->r_ctl.fsb.udp = NULL;
14085 }
14086 tcpip_fillheaders(rack->rc_inp,
14087 tp->t_port,
14088 ip6, rack->r_ctl.fsb.th);
14089 rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14090 } else
14091 #endif /* INET6 */
14092 #ifdef INET
14093 {
14094 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14095 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14096 if (tp->t_port) {
14097 rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14098 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14099 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14100 udp->uh_dport = tp->t_port;
14101 rack->r_ctl.fsb.udp = udp;
14102 rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14103 } else
14104 {
14105 rack->r_ctl.fsb.udp = NULL;
14106 rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14107 }
14108 tcpip_fillheaders(rack->rc_inp,
14109 tp->t_port,
14110 ip, rack->r_ctl.fsb.th);
14111 rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14112 }
14113 #endif
14114 rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14115 (long)TCP_MAXWIN << tp->rcv_scale);
14116 rack->r_fsb_inited = 1;
14117 }
14118
14119 static int
rack_init_fsb(struct tcpcb * tp,struct tcp_rack * rack)14120 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14121 {
14122 /*
14123 * Allocate the larger of spaces V6 if available else just
14124 * V4 and include udphdr (overbook)
14125 */
14126 #ifdef INET6
14127 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14128 #else
14129 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14130 #endif
14131 rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14132 M_TCPFSB, M_NOWAIT|M_ZERO);
14133 if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14134 return (ENOMEM);
14135 }
14136 rack->r_fsb_inited = 0;
14137 return (0);
14138 }
14139
14140 static void
rack_log_hystart_event(struct tcp_rack * rack,uint32_t high_seq,uint8_t mod)14141 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14142 {
14143 /*
14144 * Types of logs (mod value)
14145 * 20 - Initial round setup
14146 * 21 - Rack declares a new round.
14147 */
14148 struct tcpcb *tp;
14149
14150 tp = rack->rc_tp;
14151 if (tcp_bblogging_on(tp)) {
14152 union tcp_log_stackspecific log;
14153 struct timeval tv;
14154
14155 memset(&log, 0, sizeof(log));
14156 log.u_bbr.flex1 = rack->r_ctl.current_round;
14157 log.u_bbr.flex2 = rack->r_ctl.roundends;
14158 log.u_bbr.flex3 = high_seq;
14159 log.u_bbr.flex4 = tp->snd_max;
14160 log.u_bbr.flex8 = mod;
14161 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14162 log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14163 log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14164 TCP_LOG_EVENTP(tp, NULL,
14165 &tptosocket(tp)->so_rcv,
14166 &tptosocket(tp)->so_snd,
14167 TCP_HYSTART, 0,
14168 0, &log, false, &tv);
14169 }
14170 }
14171
14172 static void
rack_deferred_init(struct tcpcb * tp,struct tcp_rack * rack)14173 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14174 {
14175 rack->rack_deferred_inited = 1;
14176 rack->r_ctl.roundends = tp->snd_max;
14177 rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14178 rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14179 }
14180
14181 static void
rack_init_retransmit_value(struct tcp_rack * rack,int ctl)14182 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14183 {
14184 /* Retransmit bit controls.
14185 *
14186 * The setting of these values control one of
14187 * three settings you can have and dictate
14188 * how rack does retransmissions. Note this
14189 * is in *any* mode i.e. pacing on or off DGP
14190 * fixed rate pacing, or just bursting rack.
14191 *
14192 * 1 - Use full sized retransmits i.e. limit
14193 * the size to whatever the pace_max_segments
14194 * size is.
14195 *
14196 * 2 - Use pacer min granularity as a guide to
14197 * the size combined with the current calculated
14198 * goodput b/w measurement. So for example if
14199 * the goodput is measured at 20Mbps we would
14200 * calculate 8125 (pacer minimum 250usec in
14201 * that b/w) and then round it up to the next
14202 * MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14203 *
14204 * 0 - The rack default 1 MSS (anything not 0/1/2
14205 * fall here too if we are setting via rack_init()).
14206 *
14207 */
14208 if (ctl == 1) {
14209 rack->full_size_rxt = 1;
14210 rack->shape_rxt_to_pacing_min = 0;
14211 } else if (ctl == 2) {
14212 rack->full_size_rxt = 0;
14213 rack->shape_rxt_to_pacing_min = 1;
14214 } else {
14215 rack->full_size_rxt = 0;
14216 rack->shape_rxt_to_pacing_min = 0;
14217 }
14218 }
14219
14220 static void
rack_log_chg_info(struct tcpcb * tp,struct tcp_rack * rack,uint8_t mod,uint32_t flex1,uint32_t flex2,uint32_t flex3)14221 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14222 uint32_t flex1,
14223 uint32_t flex2,
14224 uint32_t flex3)
14225 {
14226 if (tcp_bblogging_on(rack->rc_tp)) {
14227 union tcp_log_stackspecific log;
14228 struct timeval tv;
14229
14230 memset(&log, 0, sizeof(log));
14231 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14232 log.u_bbr.flex8 = mod;
14233 log.u_bbr.flex1 = flex1;
14234 log.u_bbr.flex2 = flex2;
14235 log.u_bbr.flex3 = flex3;
14236 tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14237 0, &log, false, NULL, __func__, __LINE__, &tv);
14238 }
14239 }
14240
14241 static int
rack_chg_query(struct tcpcb * tp,struct tcp_query_resp * reqr)14242 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14243 {
14244 struct tcp_rack *rack;
14245 struct rack_sendmap *rsm;
14246 int i;
14247
14248
14249 rack = (struct tcp_rack *)tp->t_fb_ptr;
14250 switch (reqr->req) {
14251 case TCP_QUERY_SENDMAP:
14252 if ((reqr->req_param == tp->snd_max) ||
14253 (tp->snd_max == tp->snd_una)){
14254 /* Unlikely */
14255 return (0);
14256 }
14257 rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14258 if (rsm == NULL) {
14259 /* Can't find that seq -- unlikely */
14260 return (0);
14261 }
14262 reqr->sendmap_start = rsm->r_start;
14263 reqr->sendmap_end = rsm->r_end;
14264 reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14265 reqr->sendmap_fas = rsm->r_fas;
14266 if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14267 reqr->sendmap_send_cnt = SNDMAP_NRTX;
14268 for(i=0; i<reqr->sendmap_send_cnt; i++)
14269 reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14270 reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14271 reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14272 reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14273 reqr->sendmap_dupacks = rsm->r_dupack;
14274 rack_log_chg_info(tp, rack, 1,
14275 rsm->r_start,
14276 rsm->r_end,
14277 rsm->r_flags);
14278 return(1);
14279 break;
14280 case TCP_QUERY_TIMERS_UP:
14281 if (rack->r_ctl.rc_hpts_flags == 0) {
14282 /* no timers up */
14283 return (0);
14284 }
14285 reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14286 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14287 reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14288 }
14289 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14290 reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14291 }
14292 rack_log_chg_info(tp, rack, 2,
14293 rack->r_ctl.rc_hpts_flags,
14294 rack->r_ctl.rc_last_output_to,
14295 rack->r_ctl.rc_timer_exp);
14296 return (1);
14297 break;
14298 case TCP_QUERY_RACK_TIMES:
14299 /* Reordering items */
14300 reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14301 reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14302 /* Timerstamps and timers */
14303 reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14304 reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14305 reqr->rack_rtt = rack->rc_rack_rtt;
14306 reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14307 reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14308 /* PRR data */
14309 reqr->rack_sacked = rack->r_ctl.rc_sacked;
14310 reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14311 reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14312 reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14313 reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14314 reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14315 /* TLP and persists info */
14316 reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14317 reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14318 if (rack->rc_in_persist) {
14319 reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14320 reqr->rack_in_persist = 1;
14321 } else {
14322 reqr->rack_time_went_idle = 0;
14323 reqr->rack_in_persist = 0;
14324 }
14325 if (rack->r_wanted_output)
14326 reqr->rack_wanted_output = 1;
14327 else
14328 reqr->rack_wanted_output = 0;
14329 return (1);
14330 break;
14331 default:
14332 return (-EINVAL);
14333 }
14334 }
14335
14336 static void
rack_switch_failed(struct tcpcb * tp)14337 rack_switch_failed(struct tcpcb *tp)
14338 {
14339 /*
14340 * This method gets called if a stack switch was
14341 * attempted and it failed. We are left
14342 * but our hpts timers were stopped and we
14343 * need to validate time units and t_flags2.
14344 */
14345 struct tcp_rack *rack;
14346 struct timeval tv;
14347 uint32_t cts;
14348 uint32_t toval;
14349 struct hpts_diag diag;
14350
14351 rack = (struct tcp_rack *)tp->t_fb_ptr;
14352 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14353 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14354 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14355 else
14356 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14357 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14358 tp->t_flags2 |= TF2_MBUF_ACKCMP;
14359 if (tp->t_in_hpts > IHPTS_NONE) {
14360 /* Strange */
14361 return;
14362 }
14363 cts = tcp_get_usecs(&tv);
14364 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14365 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14366 toval = rack->r_ctl.rc_last_output_to - cts;
14367 } else {
14368 /* one slot please */
14369 toval = HPTS_TICKS_PER_SLOT;
14370 }
14371 } else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14372 if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14373 toval = rack->r_ctl.rc_timer_exp - cts;
14374 } else {
14375 /* one slot please */
14376 toval = HPTS_TICKS_PER_SLOT;
14377 }
14378 } else
14379 toval = HPTS_TICKS_PER_SLOT;
14380 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14381 __LINE__, &diag);
14382 rack_log_hpts_diag(rack, cts, &diag, &tv);
14383 }
14384
14385 static int
rack_init_outstanding(struct tcpcb * tp,struct tcp_rack * rack,uint32_t us_cts,void * ptr)14386 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14387 {
14388 struct rack_sendmap *rsm, *ersm;
14389 int insret __diagused;
14390 /*
14391 * When initing outstanding, we must be quite careful
14392 * to not refer to tp->t_fb_ptr. This has the old rack
14393 * pointer in it, not the "new" one (when we are doing
14394 * a stack switch).
14395 */
14396
14397
14398 if (tp->t_fb->tfb_chg_query == NULL) {
14399 /* Create a send map for the current outstanding data */
14400
14401 rsm = rack_alloc(rack);
14402 if (rsm == NULL) {
14403 uma_zfree(rack_pcb_zone, ptr);
14404 return (ENOMEM);
14405 }
14406 rsm->r_no_rtt_allowed = 1;
14407 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14408 rsm->r_rtr_cnt = 1;
14409 rsm->r_rtr_bytes = 0;
14410 if (tp->t_flags & TF_SENTFIN)
14411 rsm->r_flags |= RACK_HAS_FIN;
14412 rsm->r_end = tp->snd_max;
14413 if (tp->snd_una == tp->iss) {
14414 /* The data space is one beyond snd_una */
14415 rsm->r_flags |= RACK_HAS_SYN;
14416 rsm->r_start = tp->iss;
14417 rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14418 } else
14419 rsm->r_start = tp->snd_una;
14420 rsm->r_dupack = 0;
14421 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14422 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14423 if (rsm->m) {
14424 rsm->orig_m_len = rsm->m->m_len;
14425 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14426 } else {
14427 rsm->orig_m_len = 0;
14428 rsm->orig_t_space = 0;
14429 }
14430 } else {
14431 /*
14432 * This can happen if we have a stand-alone FIN or
14433 * SYN.
14434 */
14435 rsm->m = NULL;
14436 rsm->orig_m_len = 0;
14437 rsm->orig_t_space = 0;
14438 rsm->soff = 0;
14439 }
14440 #ifdef INVARIANTS
14441 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14442 panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14443 insret, rack, rsm);
14444 }
14445 #else
14446 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14447 #endif
14448 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14449 rsm->r_in_tmap = 1;
14450 } else {
14451 /* We have a query mechanism, lets use it */
14452 struct tcp_query_resp qr;
14453 int i;
14454 tcp_seq at;
14455
14456 at = tp->snd_una;
14457 while (at != tp->snd_max) {
14458 memset(&qr, 0, sizeof(qr));
14459 qr.req = TCP_QUERY_SENDMAP;
14460 qr.req_param = at;
14461 if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14462 break;
14463 /* Move forward */
14464 at = qr.sendmap_end;
14465 /* Now lets build the entry for this one */
14466 rsm = rack_alloc(rack);
14467 if (rsm == NULL) {
14468 uma_zfree(rack_pcb_zone, ptr);
14469 return (ENOMEM);
14470 }
14471 memset(rsm, 0, sizeof(struct rack_sendmap));
14472 /* Now configure the rsm and insert it */
14473 rsm->r_dupack = qr.sendmap_dupacks;
14474 rsm->r_start = qr.sendmap_start;
14475 rsm->r_end = qr.sendmap_end;
14476 if (qr.sendmap_fas)
14477 rsm->r_fas = qr.sendmap_end;
14478 else
14479 rsm->r_fas = rsm->r_start - tp->snd_una;
14480 /*
14481 * We have carefully aligned the bits
14482 * so that all we have to do is copy over
14483 * the bits with the mask.
14484 */
14485 rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14486 rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14487 rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14488 rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14489 for (i=0 ; i<rsm->r_rtr_cnt; i++)
14490 rsm->r_tim_lastsent[i] = qr.sendmap_time[i];
14491 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14492 (rsm->r_start - tp->snd_una), &rsm->soff);
14493 if (rsm->m) {
14494 rsm->orig_m_len = rsm->m->m_len;
14495 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14496 } else {
14497 rsm->orig_m_len = 0;
14498 rsm->orig_t_space = 0;
14499 }
14500 #ifdef INVARIANTS
14501 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14502 panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14503 insret, rack, rsm);
14504 }
14505 #else
14506 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14507 #endif
14508 if ((rsm->r_flags & RACK_ACKED) == 0) {
14509 TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14510 if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14511 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14512 /*
14513 * If the existing ersm was sent at
14514 * a later time than the new one, then
14515 * the new one should appear ahead of this
14516 * ersm.
14517 */
14518 rsm->r_in_tmap = 1;
14519 TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14520 break;
14521 }
14522 }
14523 if (rsm->r_in_tmap == 0) {
14524 /*
14525 * Not found so shove it on the tail.
14526 */
14527 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14528 rsm->r_in_tmap = 1;
14529 }
14530 } else {
14531 if ((rack->r_ctl.rc_sacklast == NULL) ||
14532 (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14533 rack->r_ctl.rc_sacklast = rsm;
14534 }
14535 }
14536 rack_log_chg_info(tp, rack, 3,
14537 rsm->r_start,
14538 rsm->r_end,
14539 rsm->r_flags);
14540 }
14541 }
14542 return (0);
14543 }
14544
14545
14546 static int32_t
rack_init(struct tcpcb * tp,void ** ptr)14547 rack_init(struct tcpcb *tp, void **ptr)
14548 {
14549 struct inpcb *inp = tptoinpcb(tp);
14550 struct tcp_rack *rack = NULL;
14551 uint32_t iwin, snt, us_cts;
14552 size_t sz;
14553 int err, no_query;
14554
14555 tcp_hpts_init(tp);
14556
14557 /*
14558 * First are we the initial or are we a switched stack?
14559 * If we are initing via tcp_newtcppcb the ptr passed
14560 * will be tp->t_fb_ptr. If its a stack switch that
14561 * has a previous stack we can query it will be a local
14562 * var that will in the end be set into t_fb_ptr.
14563 */
14564 if (ptr == &tp->t_fb_ptr)
14565 no_query = 1;
14566 else
14567 no_query = 0;
14568 *ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14569 if (*ptr == NULL) {
14570 /*
14571 * We need to allocate memory but cant. The INP and INP_INFO
14572 * locks and they are recursive (happens during setup. So a
14573 * scheme to drop the locks fails :(
14574 *
14575 */
14576 return(ENOMEM);
14577 }
14578 memset(*ptr, 0, sizeof(struct tcp_rack));
14579 rack = (struct tcp_rack *)*ptr;
14580 rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14581 if (rack->r_ctl.tqh == NULL) {
14582 uma_zfree(rack_pcb_zone, rack);
14583 return(ENOMEM);
14584 }
14585 tqhash_init(rack->r_ctl.tqh);
14586 TAILQ_INIT(&rack->r_ctl.rc_free);
14587 TAILQ_INIT(&rack->r_ctl.rc_tmap);
14588 rack->rc_tp = tp;
14589 rack->rc_inp = inp;
14590 /* Set the flag */
14591 rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14592 /* Probably not needed but lets be sure */
14593 rack_clear_rate_sample(rack);
14594 /*
14595 * Save off the default values, socket options will poke
14596 * at these if pacing is not on or we have not yet
14597 * reached where pacing is on (gp_ready/fixed enabled).
14598 * When they get set into the CC module (when gp_ready
14599 * is enabled or we enable fixed) then we will set these
14600 * values into the CC and place in here the old values
14601 * so we have a restoral. Then we will set the flag
14602 * rc_pacing_cc_set. That way whenever we turn off pacing
14603 * or switch off this stack, we will know to go restore
14604 * the saved values.
14605 *
14606 * We specifically put into the beta the ecn value for pacing.
14607 */
14608 rack->rc_new_rnd_needed = 1;
14609 rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14610 /* We want abe like behavior as well */
14611
14612 rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14613 rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14614 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14615 if (rack_fill_cw_state)
14616 rack->rc_pace_to_cwnd = 1;
14617 if (rack_pacing_min_seg)
14618 rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14619 if (use_rack_rr)
14620 rack->use_rack_rr = 1;
14621 if (rack_dnd_default) {
14622 rack->rc_pace_dnd = 1;
14623 }
14624 if (V_tcp_delack_enabled)
14625 tp->t_delayed_ack = 1;
14626 else
14627 tp->t_delayed_ack = 0;
14628 #ifdef TCP_ACCOUNTING
14629 if (rack_tcp_accounting) {
14630 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14631 }
14632 #endif
14633 rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
14634 sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
14635 rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
14636 if (rack->r_ctl.pcm_s == NULL) {
14637 rack->r_ctl.pcm_i.cnt_alloc = 0;
14638 }
14639 #ifdef NETFLIX_STATS
14640 rack->r_ctl.side_chan_dis_mask = tcp_sidechannel_disable_mask;
14641 #endif
14642 rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14643 rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14644 if (rack_enable_shared_cwnd)
14645 rack->rack_enable_scwnd = 1;
14646 rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14647 rack->rc_user_set_max_segs = rack_hptsi_segments;
14648 rack->r_ctl.max_reduction = rack_max_reduce;
14649 rack->rc_force_max_seg = 0;
14650 TAILQ_INIT(&rack->r_ctl.opt_list);
14651 rack->r_ctl.rc_saved_beta = V_newreno_beta_ecn;
14652 rack->r_ctl.rc_saved_beta_ecn = V_newreno_beta_ecn;
14653 if (rack_hibeta_setting) {
14654 rack->rack_hibeta = 1;
14655 if ((rack_hibeta_setting >= 50) &&
14656 (rack_hibeta_setting <= 100)) {
14657 rack->r_ctl.rc_saved_beta = rack_hibeta_setting;
14658 rack->r_ctl.saved_hibeta = rack_hibeta_setting;
14659 }
14660 } else {
14661 rack->r_ctl.saved_hibeta = 50;
14662 }
14663 /*
14664 * We initialize to all ones so we never match 0
14665 * just in case the client sends in 0, it hopefully
14666 * will never have all 1's in ms :-)
14667 */
14668 rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
14669 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14670 rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14671 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14672 rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14673 rack->r_ctl.rc_highest_us_rtt = 0;
14674 rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14675 rack->pcm_enabled = rack_pcm_is_enabled;
14676 if (rack_fillcw_bw_cap)
14677 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
14678 rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14679 if (rack_use_cmp_acks)
14680 rack->r_use_cmp_ack = 1;
14681 if (rack_disable_prr)
14682 rack->rack_no_prr = 1;
14683 if (rack_gp_no_rec_chg)
14684 rack->rc_gp_no_rec_chg = 1;
14685 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14686 rack->r_ctl.pacing_method |= RACK_REG_PACING;
14687 rack->rc_always_pace = 1;
14688 if (rack->rack_hibeta)
14689 rack_set_cc_pacing(rack);
14690 } else
14691 rack->rc_always_pace = 0;
14692 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14693 rack->r_mbuf_queue = 1;
14694 else
14695 rack->r_mbuf_queue = 0;
14696 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14697 if (rack_limits_scwnd)
14698 rack->r_limit_scw = 1;
14699 else
14700 rack->r_limit_scw = 0;
14701 rack_init_retransmit_value(rack, rack_rxt_controls);
14702 rack->rc_labc = V_tcp_abc_l_var;
14703 if (rack_honors_hpts_min_to)
14704 rack->r_use_hpts_min = 1;
14705 if (tp->snd_una != 0) {
14706 rack->rc_sendvars_notset = 0;
14707 /*
14708 * Make sure any TCP timers are not running.
14709 */
14710 tcp_timer_stop(tp);
14711 } else {
14712 /*
14713 * Server side, we are called from the
14714 * syn-cache. This means none of the
14715 * snd_una/max are set yet so we have
14716 * to defer this until the first send.
14717 */
14718 rack->rc_sendvars_notset = 1;
14719 }
14720
14721 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14722 rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14723 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14724 rack->r_ctl.rc_min_to = rack_min_to;
14725 microuptime(&rack->r_ctl.act_rcv_time);
14726 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14727 rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14728 if (rack_hw_up_only)
14729 rack->r_up_only = 1;
14730 if (rack_do_dyn_mul) {
14731 /* When dynamic adjustment is on CA needs to start at 100% */
14732 rack->rc_gp_dyn_mul = 1;
14733 if (rack_do_dyn_mul >= 100)
14734 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14735 } else
14736 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14737 rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14738 if (rack_timely_off) {
14739 rack->rc_skip_timely = 1;
14740 }
14741 if (rack->rc_skip_timely) {
14742 rack->r_ctl.rack_per_of_gp_rec = 90;
14743 rack->r_ctl.rack_per_of_gp_ca = 100;
14744 rack->r_ctl.rack_per_of_gp_ss = 250;
14745 }
14746 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14747 rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14748 rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
14749
14750 setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14751 rack_probertt_filter_life);
14752 us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14753 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14754 rack->r_ctl.rc_time_of_last_probertt = us_cts;
14755 rack->r_ctl.rc_went_idle_time = us_cts;
14756 rack->r_ctl.rc_time_probertt_starts = 0;
14757
14758 rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
14759 if (rack_rnd_cnt_req & 0x10000)
14760 rack->r_ctl.gate_to_fs = 1;
14761 rack->r_ctl.gp_gain_req = rack_gp_gain_req;
14762 if ((rack_rnd_cnt_req & 0x100) > 0) {
14763
14764 }
14765 if (rack_dsack_std_based & 0x1) {
14766 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14767 rack->rc_rack_tmr_std_based = 1;
14768 }
14769 if (rack_dsack_std_based & 0x2) {
14770 /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */
14771 rack->rc_rack_use_dsack = 1;
14772 }
14773 /* We require at least one measurement, even if the sysctl is 0 */
14774 if (rack_req_measurements)
14775 rack->r_ctl.req_measurements = rack_req_measurements;
14776 else
14777 rack->r_ctl.req_measurements = 1;
14778 if (rack_enable_hw_pacing)
14779 rack->rack_hdw_pace_ena = 1;
14780 if (rack_hw_rate_caps)
14781 rack->r_rack_hw_rate_caps = 1;
14782 if (rack_non_rxt_use_cr)
14783 rack->rack_rec_nonrxt_use_cr = 1;
14784 /* Lets setup the fsb block */
14785 err = rack_init_fsb(tp, rack);
14786 if (err) {
14787 uma_zfree(rack_pcb_zone, *ptr);
14788 *ptr = NULL;
14789 return (err);
14790 }
14791 if (rack_do_hystart) {
14792 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
14793 if (rack_do_hystart > 1)
14794 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
14795 if (rack_do_hystart > 2)
14796 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
14797 }
14798 /* Log what we will do with queries */
14799 rack_log_chg_info(tp, rack, 7,
14800 no_query, 0, 0);
14801 if (rack_def_profile)
14802 rack_set_profile(rack, rack_def_profile);
14803 /* Cancel the GP measurement in progress */
14804 tp->t_flags &= ~TF_GPUTINPROG;
14805 if ((tp->t_state != TCPS_CLOSED) &&
14806 (tp->t_state != TCPS_TIME_WAIT)) {
14807 /*
14808 * We are already open, we may
14809 * need to adjust a few things.
14810 */
14811 if (SEQ_GT(tp->snd_max, tp->iss))
14812 snt = tp->snd_max - tp->iss;
14813 else
14814 snt = 0;
14815 iwin = rc_init_window(rack);
14816 if ((snt < iwin) &&
14817 (no_query == 1)) {
14818 /* We are not past the initial window
14819 * on the first init (i.e. a stack switch
14820 * has not yet occured) so we need to make
14821 * sure cwnd and ssthresh is correct.
14822 */
14823 if (tp->snd_cwnd < iwin)
14824 tp->snd_cwnd = iwin;
14825 /*
14826 * If we are within the initial window
14827 * we want ssthresh to be unlimited. Setting
14828 * it to the rwnd (which the default stack does
14829 * and older racks) is not really a good idea
14830 * since we want to be in SS and grow both the
14831 * cwnd and the rwnd (via dynamic rwnd growth). If
14832 * we set it to the rwnd then as the peer grows its
14833 * rwnd we will be stuck in CA and never hit SS.
14834 *
14835 * Its far better to raise it up high (this takes the
14836 * risk that there as been a loss already, probably
14837 * we should have an indicator in all stacks of loss
14838 * but we don't), but considering the normal use this
14839 * is a risk worth taking. The consequences of not
14840 * hitting SS are far worse than going one more time
14841 * into it early on (before we have sent even a IW).
14842 * It is highly unlikely that we will have had a loss
14843 * before getting the IW out.
14844 */
14845 tp->snd_ssthresh = 0xffffffff;
14846 }
14847 /*
14848 * Any init based on sequence numbers
14849 * should be done in the deferred init path
14850 * since we can be CLOSED and not have them
14851 * inited when rack_init() is called. We
14852 * are not closed so lets call it.
14853 */
14854 rack_deferred_init(tp, rack);
14855 }
14856 if ((tp->t_state != TCPS_CLOSED) &&
14857 (tp->t_state != TCPS_TIME_WAIT) &&
14858 (no_query == 0) &&
14859 (tp->snd_una != tp->snd_max)) {
14860 err = rack_init_outstanding(tp, rack, us_cts, *ptr);
14861 if (err) {
14862 *ptr = NULL;
14863 return(err);
14864 }
14865 }
14866 rack_stop_all_timers(tp, rack);
14867 /* Setup all the t_flags2 */
14868 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14869 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14870 else
14871 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14872 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14873 tp->t_flags2 |= TF2_MBUF_ACKCMP;
14874 /*
14875 * Timers in Rack are kept in microseconds so lets
14876 * convert any initial incoming variables
14877 * from ticks into usecs. Note that we
14878 * also change the values of t_srtt and t_rttvar, if
14879 * they are non-zero. They are kept with a 5
14880 * bit decimal so we have to carefully convert
14881 * these to get the full precision.
14882 */
14883 rack_convert_rtts(tp);
14884 rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
14885 if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
14886 /* We do not start any timers on DROPPED connections */
14887 if (tp->t_fb->tfb_chg_query == NULL) {
14888 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14889 } else {
14890 struct tcp_query_resp qr;
14891 int ret;
14892
14893 memset(&qr, 0, sizeof(qr));
14894
14895 /* Get the misc time stamps and such for rack */
14896 qr.req = TCP_QUERY_RACK_TIMES;
14897 ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14898 if (ret == 1) {
14899 rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
14900 rack->r_ctl.num_dsack = qr.rack_num_dsacks;
14901 rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
14902 rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
14903 rack->rc_rack_rtt = qr.rack_rtt;
14904 rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
14905 rack->r_ctl.rc_sacked = qr.rack_sacked;
14906 rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
14907 rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
14908 rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
14909 rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
14910 rack->r_ctl.rc_prr_out = qr.rack_prr_out;
14911 if (qr.rack_tlp_out) {
14912 rack->rc_tlp_in_progress = 1;
14913 rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
14914 } else {
14915 rack->rc_tlp_in_progress = 0;
14916 rack->r_ctl.rc_tlp_cnt_out = 0;
14917 }
14918 if (qr.rack_srtt_measured)
14919 rack->rc_srtt_measure_made = 1;
14920 if (qr.rack_in_persist == 1) {
14921 rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
14922 #ifdef NETFLIX_SHARED_CWND
14923 if (rack->r_ctl.rc_scw) {
14924 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
14925 rack->rack_scwnd_is_idle = 1;
14926 }
14927 #endif
14928 rack->r_ctl.persist_lost_ends = 0;
14929 rack->probe_not_answered = 0;
14930 rack->forced_ack = 0;
14931 tp->t_rxtshift = 0;
14932 rack->rc_in_persist = 1;
14933 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
14934 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
14935 }
14936 if (qr.rack_wanted_output)
14937 rack->r_wanted_output = 1;
14938 rack_log_chg_info(tp, rack, 6,
14939 qr.rack_min_rtt,
14940 qr.rack_rtt,
14941 qr.rack_reorder_ts);
14942 }
14943 /* Get the old stack timers */
14944 qr.req_param = 0;
14945 qr.req = TCP_QUERY_TIMERS_UP;
14946 ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14947 if (ret) {
14948 /*
14949 * non-zero return means we have a timer('s)
14950 * to start. Zero means no timer (no keepalive
14951 * I suppose).
14952 */
14953 uint32_t tov = 0;
14954
14955 rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
14956 if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
14957 rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
14958 if (TSTMP_GT(qr.timer_pacing_to, us_cts))
14959 tov = qr.timer_pacing_to - us_cts;
14960 else
14961 tov = HPTS_TICKS_PER_SLOT;
14962 }
14963 if (qr.timer_hpts_flags & PACE_TMR_MASK) {
14964 rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
14965 if (tov == 0) {
14966 if (TSTMP_GT(qr.timer_timer_exp, us_cts))
14967 tov = qr.timer_timer_exp - us_cts;
14968 else
14969 tov = HPTS_TICKS_PER_SLOT;
14970 }
14971 }
14972 rack_log_chg_info(tp, rack, 4,
14973 rack->r_ctl.rc_hpts_flags,
14974 rack->r_ctl.rc_last_output_to,
14975 rack->r_ctl.rc_timer_exp);
14976 if (tov) {
14977 struct hpts_diag diag;
14978
14979 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
14980 __LINE__, &diag);
14981 rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
14982 }
14983 }
14984 }
14985 rack_log_rtt_shrinks(rack, us_cts, tp->t_rxtcur,
14986 __LINE__, RACK_RTTS_INIT);
14987 }
14988 return (0);
14989 }
14990
14991 static int
rack_handoff_ok(struct tcpcb * tp)14992 rack_handoff_ok(struct tcpcb *tp)
14993 {
14994 if ((tp->t_state == TCPS_CLOSED) ||
14995 (tp->t_state == TCPS_LISTEN)) {
14996 /* Sure no problem though it may not stick */
14997 return (0);
14998 }
14999 if ((tp->t_state == TCPS_SYN_SENT) ||
15000 (tp->t_state == TCPS_SYN_RECEIVED)) {
15001 /*
15002 * We really don't know if you support sack,
15003 * you have to get to ESTAB or beyond to tell.
15004 */
15005 return (EAGAIN);
15006 }
15007 if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15008 /*
15009 * Rack will only send a FIN after all data is acknowledged.
15010 * So in this case we have more data outstanding. We can't
15011 * switch stacks until either all data and only the FIN
15012 * is left (in which case rack_init() now knows how
15013 * to deal with that) <or> all is acknowledged and we
15014 * are only left with incoming data, though why you
15015 * would want to switch to rack after all data is acknowledged
15016 * I have no idea (rrs)!
15017 */
15018 return (EAGAIN);
15019 }
15020 if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15021 return (0);
15022 }
15023 /*
15024 * If we reach here we don't do SACK on this connection so we can
15025 * never do rack.
15026 */
15027 return (EINVAL);
15028 }
15029
15030 static void
rack_fini(struct tcpcb * tp,int32_t tcb_is_purged)15031 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15032 {
15033
15034 if (tp->t_fb_ptr) {
15035 uint32_t cnt_free = 0;
15036 struct tcp_rack *rack;
15037 struct rack_sendmap *rsm;
15038
15039 tcp_handle_orphaned_packets(tp);
15040 tp->t_flags &= ~TF_FORCEDATA;
15041 rack = (struct tcp_rack *)tp->t_fb_ptr;
15042 rack_log_pacing_delay_calc(rack,
15043 0,
15044 0,
15045 0,
15046 rack_get_gp_est(rack), /* delRate */
15047 rack_get_lt_bw(rack), /* rttProp */
15048 20, __LINE__, NULL, 0);
15049 #ifdef NETFLIX_SHARED_CWND
15050 if (rack->r_ctl.rc_scw) {
15051 uint32_t limit;
15052
15053 if (rack->r_limit_scw)
15054 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15055 else
15056 limit = 0;
15057 tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15058 rack->r_ctl.rc_scw_index,
15059 limit);
15060 rack->r_ctl.rc_scw = NULL;
15061 }
15062 #endif
15063 if (rack->r_ctl.fsb.tcp_ip_hdr) {
15064 free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15065 rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15066 rack->r_ctl.fsb.th = NULL;
15067 }
15068 if (rack->rc_always_pace == 1) {
15069 rack_remove_pacing(rack);
15070 }
15071 /* Clean up any options if they were not applied */
15072 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15073 struct deferred_opt_list *dol;
15074
15075 dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15076 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15077 free(dol, M_TCPDO);
15078 }
15079 /* rack does not use force data but other stacks may clear it */
15080 if (rack->r_ctl.crte != NULL) {
15081 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15082 rack->rack_hdrw_pacing = 0;
15083 rack->r_ctl.crte = NULL;
15084 }
15085 #ifdef TCP_BLACKBOX
15086 tcp_log_flowend(tp);
15087 #endif
15088 /*
15089 * Lets take a different approach to purging just
15090 * get each one and free it like a cum-ack would and
15091 * not use a foreach loop.
15092 */
15093 rsm = tqhash_min(rack->r_ctl.tqh);
15094 while (rsm) {
15095 tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15096 rack->r_ctl.rc_num_maps_alloced--;
15097 uma_zfree(rack_zone, rsm);
15098 rsm = tqhash_min(rack->r_ctl.tqh);
15099 }
15100 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15101 while (rsm) {
15102 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15103 rack->r_ctl.rc_num_maps_alloced--;
15104 rack->rc_free_cnt--;
15105 cnt_free++;
15106 uma_zfree(rack_zone, rsm);
15107 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15108 }
15109 if (rack->r_ctl.pcm_s != NULL) {
15110 free(rack->r_ctl.pcm_s, M_TCPPCM);
15111 rack->r_ctl.pcm_s = NULL;
15112 rack->r_ctl.pcm_i.cnt_alloc = 0;
15113 rack->r_ctl.pcm_i.cnt = 0;
15114 }
15115 if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15116 (tcp_bblogging_on(tp))) {
15117 union tcp_log_stackspecific log;
15118 struct timeval tv;
15119
15120 memset(&log, 0, sizeof(log));
15121 log.u_bbr.flex8 = 10;
15122 log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15123 log.u_bbr.flex2 = rack->rc_free_cnt;
15124 log.u_bbr.flex3 = cnt_free;
15125 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15126 rsm = tqhash_min(rack->r_ctl.tqh);
15127 log.u_bbr.delRate = (uintptr_t)rsm;
15128 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15129 log.u_bbr.cur_del_rate = (uintptr_t)rsm;
15130 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15131 log.u_bbr.pkt_epoch = __LINE__;
15132 (void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15133 0, &log, false, NULL, NULL, 0, &tv);
15134 }
15135 KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15136 ("rack:%p num_aloc:%u after freeing all?",
15137 rack,
15138 rack->r_ctl.rc_num_maps_alloced));
15139 rack->rc_free_cnt = 0;
15140 free(rack->r_ctl.tqh, M_TCPFSB);
15141 rack->r_ctl.tqh = NULL;
15142 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15143 tp->t_fb_ptr = NULL;
15144 }
15145 /* Make sure snd_nxt is correctly set */
15146 tp->snd_nxt = tp->snd_max;
15147 }
15148
15149 static void
rack_set_state(struct tcpcb * tp,struct tcp_rack * rack)15150 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15151 {
15152 if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15153 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15154 }
15155 switch (tp->t_state) {
15156 case TCPS_SYN_SENT:
15157 rack->r_state = TCPS_SYN_SENT;
15158 rack->r_substate = rack_do_syn_sent;
15159 break;
15160 case TCPS_SYN_RECEIVED:
15161 rack->r_state = TCPS_SYN_RECEIVED;
15162 rack->r_substate = rack_do_syn_recv;
15163 break;
15164 case TCPS_ESTABLISHED:
15165 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15166 rack->r_state = TCPS_ESTABLISHED;
15167 rack->r_substate = rack_do_established;
15168 break;
15169 case TCPS_CLOSE_WAIT:
15170 rack->r_state = TCPS_CLOSE_WAIT;
15171 rack->r_substate = rack_do_close_wait;
15172 break;
15173 case TCPS_FIN_WAIT_1:
15174 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15175 rack->r_state = TCPS_FIN_WAIT_1;
15176 rack->r_substate = rack_do_fin_wait_1;
15177 break;
15178 case TCPS_CLOSING:
15179 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15180 rack->r_state = TCPS_CLOSING;
15181 rack->r_substate = rack_do_closing;
15182 break;
15183 case TCPS_LAST_ACK:
15184 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15185 rack->r_state = TCPS_LAST_ACK;
15186 rack->r_substate = rack_do_lastack;
15187 break;
15188 case TCPS_FIN_WAIT_2:
15189 rack->r_state = TCPS_FIN_WAIT_2;
15190 rack->r_substate = rack_do_fin_wait_2;
15191 break;
15192 case TCPS_LISTEN:
15193 case TCPS_CLOSED:
15194 case TCPS_TIME_WAIT:
15195 default:
15196 break;
15197 };
15198 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15199 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15200
15201 }
15202
15203 static void
rack_timer_audit(struct tcpcb * tp,struct tcp_rack * rack,struct sockbuf * sb)15204 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15205 {
15206 /*
15207 * We received an ack, and then did not
15208 * call send or were bounced out due to the
15209 * hpts was running. Now a timer is up as well, is
15210 * it the right timer?
15211 */
15212 struct rack_sendmap *rsm;
15213 int tmr_up;
15214
15215 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15216 if (tcp_in_hpts(rack->rc_tp) == 0) {
15217 /*
15218 * Ok we probably need some timer up, but no
15219 * matter what the mask we are not in hpts. We
15220 * may have received an old ack and thus did nothing.
15221 */
15222 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15223 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15224 return;
15225 }
15226 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15227 return;
15228 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15229 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15230 (tmr_up == PACE_TMR_RXT)) {
15231 /* Should be an RXT */
15232 return;
15233 }
15234 if (rsm == NULL) {
15235 /* Nothing outstanding? */
15236 if (tp->t_flags & TF_DELACK) {
15237 if (tmr_up == PACE_TMR_DELACK)
15238 /* We are supposed to have delayed ack up and we do */
15239 return;
15240 } else if (((V_tcp_always_keepalive ||
15241 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15242 (tp->t_state <= TCPS_CLOSING)) &&
15243 (tmr_up == PACE_TMR_KEEP) &&
15244 (tp->snd_max == tp->snd_una)) {
15245 /* We should have keep alive up and we do */
15246 return;
15247 }
15248 }
15249 if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15250 ((tmr_up == PACE_TMR_TLP) ||
15251 (tmr_up == PACE_TMR_RACK) ||
15252 (tmr_up == PACE_TMR_RXT))) {
15253 /*
15254 * Either a Rack, TLP or RXT is fine if we
15255 * have outstanding data.
15256 */
15257 return;
15258 } else if (tmr_up == PACE_TMR_DELACK) {
15259 /*
15260 * If the delayed ack was going to go off
15261 * before the rtx/tlp/rack timer were going to
15262 * expire, then that would be the timer in control.
15263 * Note we don't check the time here trusting the
15264 * code is correct.
15265 */
15266 return;
15267 }
15268 /*
15269 * Ok the timer originally started is not what we want now.
15270 * We will force the hpts to be stopped if any, and restart
15271 * with the slot set to what was in the saved slot.
15272 */
15273 if (tcp_in_hpts(rack->rc_tp)) {
15274 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15275 uint32_t us_cts;
15276
15277 us_cts = tcp_get_usecs(NULL);
15278 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15279 rack->r_early = 1;
15280 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15281 }
15282 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15283 }
15284 tcp_hpts_remove(rack->rc_tp);
15285 }
15286 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15287 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15288 }
15289
15290
15291 static void
rack_do_win_updates(struct tcpcb * tp,struct tcp_rack * rack,uint32_t tiwin,uint32_t seq,uint32_t ack,uint32_t cts)15292 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15293 {
15294 if ((SEQ_LT(tp->snd_wl1, seq) ||
15295 (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15296 (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15297 /* keep track of pure window updates */
15298 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15299 KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15300 tp->snd_wnd = tiwin;
15301 rack_validate_fo_sendwin_up(tp, rack);
15302 tp->snd_wl1 = seq;
15303 tp->snd_wl2 = ack;
15304 if (tp->snd_wnd > tp->max_sndwnd)
15305 tp->max_sndwnd = tp->snd_wnd;
15306 rack->r_wanted_output = 1;
15307 } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15308 tp->snd_wnd = tiwin;
15309 rack_validate_fo_sendwin_up(tp, rack);
15310 tp->snd_wl1 = seq;
15311 tp->snd_wl2 = ack;
15312 } else {
15313 /* Not a valid win update */
15314 return;
15315 }
15316 if (tp->snd_wnd > tp->max_sndwnd)
15317 tp->max_sndwnd = tp->snd_wnd;
15318 /* Do we exit persists? */
15319 if ((rack->rc_in_persist != 0) &&
15320 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15321 rack->r_ctl.rc_pace_min_segs))) {
15322 rack_exit_persist(tp, rack, cts);
15323 }
15324 /* Do we enter persists? */
15325 if ((rack->rc_in_persist == 0) &&
15326 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15327 TCPS_HAVEESTABLISHED(tp->t_state) &&
15328 ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15329 sbavail(&tptosocket(tp)->so_snd) &&
15330 (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15331 /*
15332 * Here the rwnd is less than
15333 * the pacing size, we are established,
15334 * nothing is outstanding, and there is
15335 * data to send. Enter persists.
15336 */
15337 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15338 }
15339 }
15340
15341 static void
rack_log_input_packet(struct tcpcb * tp,struct tcp_rack * rack,struct tcp_ackent * ae,int ackval,uint32_t high_seq)15342 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15343 {
15344
15345 if (tcp_bblogging_on(rack->rc_tp)) {
15346 struct inpcb *inp = tptoinpcb(tp);
15347 union tcp_log_stackspecific log;
15348 struct timeval ltv;
15349 char tcp_hdr_buf[60];
15350 struct tcphdr *th;
15351 struct timespec ts;
15352 uint32_t orig_snd_una;
15353 uint8_t xx = 0;
15354
15355 #ifdef TCP_REQUEST_TRK
15356 struct tcp_sendfile_track *tcp_req;
15357
15358 if (SEQ_GT(ae->ack, tp->snd_una)) {
15359 tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15360 } else {
15361 tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15362 }
15363 #endif
15364 memset(&log, 0, sizeof(log));
15365 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15366 if (rack->rack_no_prr == 0)
15367 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15368 else
15369 log.u_bbr.flex1 = 0;
15370 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15371 log.u_bbr.use_lt_bw <<= 1;
15372 log.u_bbr.use_lt_bw |= rack->r_might_revert;
15373 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15374 log.u_bbr.bbr_state = rack->rc_free_cnt;
15375 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15376 log.u_bbr.pkts_out = tp->t_maxseg;
15377 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15378 log.u_bbr.flex7 = 1;
15379 log.u_bbr.lost = ae->flags;
15380 log.u_bbr.cwnd_gain = ackval;
15381 log.u_bbr.pacing_gain = 0x2;
15382 if (ae->flags & TSTMP_HDWR) {
15383 /* Record the hardware timestamp if present */
15384 log.u_bbr.flex3 = M_TSTMP;
15385 ts.tv_sec = ae->timestamp / 1000000000;
15386 ts.tv_nsec = ae->timestamp % 1000000000;
15387 ltv.tv_sec = ts.tv_sec;
15388 ltv.tv_usec = ts.tv_nsec / 1000;
15389 log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v);
15390 } else if (ae->flags & TSTMP_LRO) {
15391 /* Record the LRO the arrival timestamp */
15392 log.u_bbr.flex3 = M_TSTMP_LRO;
15393 ts.tv_sec = ae->timestamp / 1000000000;
15394 ts.tv_nsec = ae->timestamp % 1000000000;
15395 ltv.tv_sec = ts.tv_sec;
15396 ltv.tv_usec = ts.tv_nsec / 1000;
15397 log.u_bbr.flex5 = tcp_tv_to_usectick(<v);
15398 }
15399 log.u_bbr.timeStamp = tcp_get_usecs(<v);
15400 /* Log the rcv time */
15401 log.u_bbr.delRate = ae->timestamp;
15402 #ifdef TCP_REQUEST_TRK
15403 log.u_bbr.applimited = tp->t_tcpreq_closed;
15404 log.u_bbr.applimited <<= 8;
15405 log.u_bbr.applimited |= tp->t_tcpreq_open;
15406 log.u_bbr.applimited <<= 8;
15407 log.u_bbr.applimited |= tp->t_tcpreq_req;
15408 if (tcp_req) {
15409 /* Copy out any client req info */
15410 /* seconds */
15411 log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15412 /* useconds */
15413 log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15414 log.u_bbr.rttProp = tcp_req->timestamp;
15415 log.u_bbr.cur_del_rate = tcp_req->start;
15416 if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15417 log.u_bbr.flex8 |= 1;
15418 } else {
15419 log.u_bbr.flex8 |= 2;
15420 log.u_bbr.bw_inuse = tcp_req->end;
15421 }
15422 log.u_bbr.flex6 = tcp_req->start_seq;
15423 if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15424 log.u_bbr.flex8 |= 4;
15425 log.u_bbr.epoch = tcp_req->end_seq;
15426 }
15427 }
15428 #endif
15429 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15430 th = (struct tcphdr *)tcp_hdr_buf;
15431 th->th_seq = ae->seq;
15432 th->th_ack = ae->ack;
15433 th->th_win = ae->win;
15434 /* Now fill in the ports */
15435 th->th_sport = inp->inp_fport;
15436 th->th_dport = inp->inp_lport;
15437 tcp_set_flags(th, ae->flags);
15438 /* Now do we have a timestamp option? */
15439 if (ae->flags & HAS_TSTMP) {
15440 u_char *cp;
15441 uint32_t val;
15442
15443 th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15444 cp = (u_char *)(th + 1);
15445 *cp = TCPOPT_NOP;
15446 cp++;
15447 *cp = TCPOPT_NOP;
15448 cp++;
15449 *cp = TCPOPT_TIMESTAMP;
15450 cp++;
15451 *cp = TCPOLEN_TIMESTAMP;
15452 cp++;
15453 val = htonl(ae->ts_value);
15454 bcopy((char *)&val,
15455 (char *)cp, sizeof(uint32_t));
15456 val = htonl(ae->ts_echo);
15457 bcopy((char *)&val,
15458 (char *)(cp + 4), sizeof(uint32_t));
15459 } else
15460 th->th_off = (sizeof(struct tcphdr) >> 2);
15461
15462 /*
15463 * For sane logging we need to play a little trick.
15464 * If the ack were fully processed we would have moved
15465 * snd_una to high_seq, but since compressed acks are
15466 * processed in two phases, at this point (logging) snd_una
15467 * won't be advanced. So we would see multiple acks showing
15468 * the advancement. We can prevent that by "pretending" that
15469 * snd_una was advanced and then un-advancing it so that the
15470 * logging code has the right value for tlb_snd_una.
15471 */
15472 if (tp->snd_una != high_seq) {
15473 orig_snd_una = tp->snd_una;
15474 tp->snd_una = high_seq;
15475 xx = 1;
15476 } else
15477 xx = 0;
15478 TCP_LOG_EVENTP(tp, th,
15479 &tptosocket(tp)->so_rcv,
15480 &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15481 0, &log, true, <v);
15482 if (xx) {
15483 tp->snd_una = orig_snd_una;
15484 }
15485 }
15486
15487 }
15488
15489 static void
rack_handle_probe_response(struct tcp_rack * rack,uint32_t tiwin,uint32_t us_cts)15490 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15491 {
15492 uint32_t us_rtt;
15493 /*
15494 * A persist or keep-alive was forced out, update our
15495 * min rtt time. Note now worry about lost responses.
15496 * When a subsequent keep-alive or persist times out
15497 * and forced_ack is still on, then the last probe
15498 * was not responded to. In such cases we have a
15499 * sysctl that controls the behavior. Either we apply
15500 * the rtt but with reduced confidence (0). Or we just
15501 * plain don't apply the rtt estimate. Having data flow
15502 * will clear the probe_not_answered flag i.e. cum-ack
15503 * move forward <or> exiting and reentering persists.
15504 */
15505
15506 rack->forced_ack = 0;
15507 rack->rc_tp->t_rxtshift = 0;
15508 if ((rack->rc_in_persist &&
15509 (tiwin == rack->rc_tp->snd_wnd)) ||
15510 (rack->rc_in_persist == 0)) {
15511 /*
15512 * In persists only apply the RTT update if this is
15513 * a response to our window probe. And that
15514 * means the rwnd sent must match the current
15515 * snd_wnd. If it does not, then we got a
15516 * window update ack instead. For keepalive
15517 * we allow the answer no matter what the window.
15518 *
15519 * Note that if the probe_not_answered is set then
15520 * the forced_ack_ts is the oldest one i.e. the first
15521 * probe sent that might have been lost. This assures
15522 * us that if we do calculate an RTT it is longer not
15523 * some short thing.
15524 */
15525 if (rack->rc_in_persist)
15526 counter_u64_add(rack_persists_acks, 1);
15527 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15528 if (us_rtt == 0)
15529 us_rtt = 1;
15530 if (rack->probe_not_answered == 0) {
15531 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15532 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15533 } else {
15534 /* We have a retransmitted probe here too */
15535 if (rack_apply_rtt_with_reduced_conf) {
15536 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15537 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15538 }
15539 }
15540 }
15541 }
15542
15543 static void
rack_new_round_starts(struct tcpcb * tp,struct tcp_rack * rack,uint32_t high_seq)15544 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15545 {
15546 /*
15547 * The next send has occurred mark the end of the round
15548 * as when that data gets acknowledged. We can
15549 * also do common things we might need to do when
15550 * a round begins.
15551 */
15552 rack->r_ctl.roundends = tp->snd_max;
15553 rack->rc_new_rnd_needed = 0;
15554 rack_log_hystart_event(rack, tp->snd_max, 4);
15555 }
15556
15557
15558 static void
rack_log_pcm(struct tcp_rack * rack,uint8_t mod,uint32_t flex1,uint32_t flex2,uint32_t flex3)15559 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
15560 uint32_t flex3)
15561 {
15562 if (tcp_bblogging_on(rack->rc_tp)) {
15563 union tcp_log_stackspecific log;
15564 struct timeval tv;
15565
15566 (void)tcp_get_usecs(&tv);
15567 memset(&log, 0, sizeof(log));
15568 log.u_bbr.timeStamp = tcp_tv_to_usectick(&tv);
15569 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15570 log.u_bbr.flex8 = mod;
15571 log.u_bbr.flex1 = flex1;
15572 log.u_bbr.flex2 = flex2;
15573 log.u_bbr.flex3 = flex3;
15574 log.u_bbr.flex4 = rack_pcm_every_n_rounds;
15575 log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
15576 log.u_bbr.bbr_substate = rack->pcm_needed;
15577 log.u_bbr.bbr_substate <<= 1;
15578 log.u_bbr.bbr_substate |= rack->pcm_in_progress;
15579 log.u_bbr.bbr_substate <<= 1;
15580 log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
15581 (void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
15582 0, &log, false, NULL, NULL, 0, &tv);
15583 }
15584 }
15585
15586 static void
rack_new_round_setup(struct tcpcb * tp,struct tcp_rack * rack,uint32_t high_seq)15587 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15588 {
15589 /*
15590 * The round (current_round) has ended. We now
15591 * setup for the next round by incrementing the
15592 * round numnber and doing any round specific
15593 * things.
15594 */
15595 rack_log_hystart_event(rack, high_seq, 21);
15596 rack->r_ctl.current_round++;
15597 /* New round (current_round) begins at next send */
15598 rack->rc_new_rnd_needed = 1;
15599 if ((rack->pcm_enabled == 1) &&
15600 (rack->pcm_needed == 0) &&
15601 (rack->pcm_in_progress == 0)) {
15602 /*
15603 * If we have enabled PCM, then we need to
15604 * check if the round has adanced to the state
15605 * where one is required.
15606 */
15607 int rnds;
15608
15609 rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
15610 if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
15611 rack->pcm_needed = 1;
15612 rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15613 } else if (rack_verbose_logging) {
15614 rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15615 }
15616 }
15617 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
15618 /* We have hystart enabled send the round info in */
15619 if (CC_ALGO(tp)->newround != NULL) {
15620 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
15621 }
15622 }
15623 /*
15624 * For DGP an initial startup check. We want to validate
15625 * that we are not just pushing on slow-start and just
15626 * not gaining.. i.e. filling buffers without getting any
15627 * boost in b/w during the inital slow-start.
15628 */
15629 if (rack->dgp_on &&
15630 (rack->rc_initial_ss_comp == 0) &&
15631 (tp->snd_cwnd < tp->snd_ssthresh) &&
15632 (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
15633 (rack->r_ctl.gp_rnd_thresh > 0) &&
15634 ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
15635
15636 /*
15637 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
15638 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
15639 * exit SS.
15640 *
15641 * Pick up the flight size now as we enter slowstart (not the
15642 * cwnd which may be inflated).
15643 */
15644 rack->rc_initial_ss_comp = 1;
15645
15646 if (tcp_bblogging_on(rack->rc_tp)) {
15647 union tcp_log_stackspecific log;
15648 struct timeval tv;
15649
15650 memset(&log, 0, sizeof(log));
15651 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15652 log.u_bbr.flex1 = rack->r_ctl.current_round;
15653 log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
15654 log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
15655 log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
15656 log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
15657 log.u_bbr.flex8 = 40;
15658 (void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
15659 0, &log, false, NULL, __func__, __LINE__,&tv);
15660 }
15661 if ((rack->r_ctl.gate_to_fs == 1) &&
15662 (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
15663 tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
15664 }
15665 tp->snd_ssthresh = tp->snd_cwnd - 1;
15666 /* Turn off any fast output running */
15667 rack->r_fast_output = 0;
15668 }
15669 }
15670
15671 static int
rack_do_compressed_ack_processing(struct tcpcb * tp,struct socket * so,struct mbuf * m,int nxt_pkt,struct timeval * tv)15672 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15673 {
15674 /*
15675 * Handle a "special" compressed ack mbuf. Each incoming
15676 * ack has only four possible dispositions:
15677 *
15678 * A) It moves the cum-ack forward
15679 * B) It is behind the cum-ack.
15680 * C) It is a window-update ack.
15681 * D) It is a dup-ack.
15682 *
15683 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15684 * in the incoming mbuf. We also need to still pay attention
15685 * to nxt_pkt since there may be another packet after this
15686 * one.
15687 */
15688 #ifdef TCP_ACCOUNTING
15689 uint64_t ts_val;
15690 uint64_t rdstc;
15691 #endif
15692 int segsiz;
15693 struct timespec ts;
15694 struct tcp_rack *rack;
15695 struct tcp_ackent *ae;
15696 uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15697 int cnt, i, did_out, ourfinisacked = 0;
15698 struct tcpopt to_holder, *to = NULL;
15699 #ifdef TCP_ACCOUNTING
15700 int win_up_req = 0;
15701 #endif
15702 int nsegs = 0;
15703 int under_pacing = 0;
15704 int post_recovery = 0;
15705 #ifdef TCP_ACCOUNTING
15706 sched_pin();
15707 #endif
15708 rack = (struct tcp_rack *)tp->t_fb_ptr;
15709 if (rack->gp_ready &&
15710 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15711 under_pacing = 1;
15712
15713 if (rack->r_state != tp->t_state)
15714 rack_set_state(tp, rack);
15715 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15716 (tp->t_flags & TF_GPUTINPROG)) {
15717 /*
15718 * We have a goodput in progress
15719 * and we have entered a late state.
15720 * Do we have enough data in the sb
15721 * to handle the GPUT request?
15722 */
15723 uint32_t bytes;
15724
15725 bytes = tp->gput_ack - tp->gput_seq;
15726 if (SEQ_GT(tp->gput_seq, tp->snd_una))
15727 bytes += tp->gput_seq - tp->snd_una;
15728 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15729 /*
15730 * There are not enough bytes in the socket
15731 * buffer that have been sent to cover this
15732 * measurement. Cancel it.
15733 */
15734 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15735 rack->r_ctl.rc_gp_srtt /*flex1*/,
15736 tp->gput_seq,
15737 0, 0, 18, __LINE__, NULL, 0);
15738 tp->t_flags &= ~TF_GPUTINPROG;
15739 }
15740 }
15741 to = &to_holder;
15742 to->to_flags = 0;
15743 KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15744 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15745 cnt = m->m_len / sizeof(struct tcp_ackent);
15746 counter_u64_add(rack_multi_single_eq, cnt);
15747 high_seq = tp->snd_una;
15748 the_win = tp->snd_wnd;
15749 win_seq = tp->snd_wl1;
15750 win_upd_ack = tp->snd_wl2;
15751 cts = tcp_tv_to_usectick(tv);
15752 ms_cts = tcp_tv_to_mssectick(tv);
15753 rack->r_ctl.rc_rcvtime = cts;
15754 segsiz = ctf_fixed_maxseg(tp);
15755 if ((rack->rc_gp_dyn_mul) &&
15756 (rack->use_fixed_rate == 0) &&
15757 (rack->rc_always_pace)) {
15758 /* Check in on probertt */
15759 rack_check_probe_rtt(rack, cts);
15760 }
15761 for (i = 0; i < cnt; i++) {
15762 #ifdef TCP_ACCOUNTING
15763 ts_val = get_cyclecount();
15764 #endif
15765 rack_clear_rate_sample(rack);
15766 ae = ((mtod(m, struct tcp_ackent *)) + i);
15767 if (ae->flags & TH_FIN)
15768 rack_log_pacing_delay_calc(rack,
15769 0,
15770 0,
15771 0,
15772 rack_get_gp_est(rack), /* delRate */
15773 rack_get_lt_bw(rack), /* rttProp */
15774 20, __LINE__, NULL, 0);
15775 /* Setup the window */
15776 tiwin = ae->win << tp->snd_scale;
15777 if (tiwin > rack->r_ctl.rc_high_rwnd)
15778 rack->r_ctl.rc_high_rwnd = tiwin;
15779 /* figure out the type of ack */
15780 if (SEQ_LT(ae->ack, high_seq)) {
15781 /* Case B*/
15782 ae->ack_val_set = ACK_BEHIND;
15783 } else if (SEQ_GT(ae->ack, high_seq)) {
15784 /* Case A */
15785 ae->ack_val_set = ACK_CUMACK;
15786 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15787 /* Case D */
15788 ae->ack_val_set = ACK_DUPACK;
15789 } else {
15790 /* Case C */
15791 ae->ack_val_set = ACK_RWND;
15792 }
15793 rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15794 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15795 /* Validate timestamp */
15796 if (ae->flags & HAS_TSTMP) {
15797 /* Setup for a timestamp */
15798 to->to_flags = TOF_TS;
15799 ae->ts_echo -= tp->ts_offset;
15800 to->to_tsecr = ae->ts_echo;
15801 to->to_tsval = ae->ts_value;
15802 /*
15803 * If echoed timestamp is later than the current time, fall back to
15804 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
15805 * were used when this connection was established.
15806 */
15807 if (TSTMP_GT(ae->ts_echo, ms_cts))
15808 to->to_tsecr = 0;
15809 if (tp->ts_recent &&
15810 TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15811 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15812 #ifdef TCP_ACCOUNTING
15813 rdstc = get_cyclecount();
15814 if (rdstc > ts_val) {
15815 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15816 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15817 }
15818 }
15819 #endif
15820 continue;
15821 }
15822 }
15823 if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15824 SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15825 tp->ts_recent_age = tcp_ts_getticks();
15826 tp->ts_recent = ae->ts_value;
15827 }
15828 } else {
15829 /* Setup for a no options */
15830 to->to_flags = 0;
15831 }
15832 /* Update the rcv time and perform idle reduction possibly */
15833 if (tp->t_idle_reduce &&
15834 (tp->snd_max == tp->snd_una) &&
15835 (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15836 counter_u64_add(rack_input_idle_reduces, 1);
15837 rack_cc_after_idle(rack, tp);
15838 }
15839 tp->t_rcvtime = ticks;
15840 /* Now what about ECN of a chain of pure ACKs? */
15841 if (tcp_ecn_input_segment(tp, ae->flags, 0,
15842 tcp_packets_this_ack(tp, ae->ack),
15843 ae->codepoint))
15844 rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15845 #ifdef TCP_ACCOUNTING
15846 /* Count for the specific type of ack in */
15847 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15848 tp->tcp_cnt_counters[ae->ack_val_set]++;
15849 }
15850 #endif
15851 /*
15852 * Note how we could move up these in the determination
15853 * above, but we don't so that way the timestamp checks (and ECN)
15854 * is done first before we do any processing on the ACK.
15855 * The non-compressed path through the code has this
15856 * weakness (noted by @jtl) that it actually does some
15857 * processing before verifying the timestamp information.
15858 * We don't take that path here which is why we set
15859 * the ack_val_set first, do the timestamp and ecn
15860 * processing, and then look at what we have setup.
15861 */
15862 if (ae->ack_val_set == ACK_BEHIND) {
15863 /*
15864 * Case B flag reordering, if window is not closed
15865 * or it could be a keep-alive or persists
15866 */
15867 if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15868 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15869 if (rack->r_ctl.rc_reorder_ts == 0)
15870 rack->r_ctl.rc_reorder_ts = 1;
15871 }
15872 } else if (ae->ack_val_set == ACK_DUPACK) {
15873 /* Case D */
15874 rack_strike_dupack(rack, ae->ack);
15875 } else if (ae->ack_val_set == ACK_RWND) {
15876 /* Case C */
15877 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15878 ts.tv_sec = ae->timestamp / 1000000000;
15879 ts.tv_nsec = ae->timestamp % 1000000000;
15880 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15881 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15882 } else {
15883 rack->r_ctl.act_rcv_time = *tv;
15884 }
15885 if (rack->forced_ack) {
15886 rack_handle_probe_response(rack, tiwin,
15887 tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
15888 }
15889 #ifdef TCP_ACCOUNTING
15890 win_up_req = 1;
15891 #endif
15892 win_upd_ack = ae->ack;
15893 win_seq = ae->seq;
15894 the_win = tiwin;
15895 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15896 } else {
15897 /* Case A */
15898 if (SEQ_GT(ae->ack, tp->snd_max)) {
15899 /*
15900 * We just send an ack since the incoming
15901 * ack is beyond the largest seq we sent.
15902 */
15903 if ((tp->t_flags & TF_ACKNOW) == 0) {
15904 ctf_ack_war_checks(tp);
15905 if (tp->t_flags && TF_ACKNOW)
15906 rack->r_wanted_output = 1;
15907 }
15908 } else {
15909 nsegs++;
15910 /* If the window changed setup to update */
15911 if (tiwin != tp->snd_wnd) {
15912 win_upd_ack = ae->ack;
15913 win_seq = ae->seq;
15914 the_win = tiwin;
15915 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15916 }
15917 #ifdef TCP_ACCOUNTING
15918 /* Account for the acks */
15919 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15920 tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
15921 }
15922 #endif
15923 high_seq = ae->ack;
15924 /* Setup our act_rcv_time */
15925 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15926 ts.tv_sec = ae->timestamp / 1000000000;
15927 ts.tv_nsec = ae->timestamp % 1000000000;
15928 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15929 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15930 } else {
15931 rack->r_ctl.act_rcv_time = *tv;
15932 }
15933 rack_process_to_cumack(tp, rack, ae->ack, cts, to,
15934 tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
15935 #ifdef TCP_REQUEST_TRK
15936 rack_req_check_for_comp(rack, high_seq);
15937 #endif
15938 if (rack->rc_dsack_round_seen) {
15939 /* Is the dsack round over? */
15940 if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
15941 /* Yes it is */
15942 rack->rc_dsack_round_seen = 0;
15943 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
15944 }
15945 }
15946 }
15947 }
15948 /* And lets be sure to commit the rtt measurements for this ack */
15949 tcp_rack_xmit_timer_commit(rack, tp);
15950 #ifdef TCP_ACCOUNTING
15951 rdstc = get_cyclecount();
15952 if (rdstc > ts_val) {
15953 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15954 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15955 if (ae->ack_val_set == ACK_CUMACK)
15956 tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
15957 }
15958 }
15959 #endif
15960 }
15961 #ifdef TCP_ACCOUNTING
15962 ts_val = get_cyclecount();
15963 #endif
15964 /* Tend to any collapsed window */
15965 if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
15966 /* The peer collapsed the window */
15967 rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
15968 } else if (rack->rc_has_collapsed)
15969 rack_un_collapse_window(rack, __LINE__);
15970 if ((rack->r_collapse_point_valid) &&
15971 (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
15972 rack->r_collapse_point_valid = 0;
15973 acked_amount = acked = (high_seq - tp->snd_una);
15974 if (acked) {
15975 /*
15976 * The draft (v3) calls for us to use SEQ_GEQ, but that
15977 * causes issues when we are just going app limited. Lets
15978 * instead use SEQ_GT <or> where its equal but more data
15979 * is outstanding.
15980 *
15981 * Also make sure we are on the last ack of a series. We
15982 * have to have all the ack's processed in queue to know
15983 * if there is something left outstanding.
15984 *
15985 */
15986 if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
15987 (rack->rc_new_rnd_needed == 0) &&
15988 (nxt_pkt == 0)) {
15989 /*
15990 * We have crossed into a new round with
15991 * this th_ack value.
15992 */
15993 rack_new_round_setup(tp, rack, high_seq);
15994 }
15995 /*
15996 * Clear the probe not answered flag
15997 * since cum-ack moved forward.
15998 */
15999 rack->probe_not_answered = 0;
16000 if (tp->t_flags & TF_NEEDSYN) {
16001 /*
16002 * T/TCP: Connection was half-synchronized, and our SYN has
16003 * been ACK'd (so connection is now fully synchronized). Go
16004 * to non-starred state, increment snd_una for ACK of SYN,
16005 * and check if we can do window scaling.
16006 */
16007 tp->t_flags &= ~TF_NEEDSYN;
16008 tp->snd_una++;
16009 acked_amount = acked = (high_seq - tp->snd_una);
16010 }
16011 if (acked > sbavail(&so->so_snd))
16012 acked_amount = sbavail(&so->so_snd);
16013 if (IN_FASTRECOVERY(tp->t_flags) &&
16014 (rack->rack_no_prr == 0))
16015 rack_update_prr(tp, rack, acked_amount, high_seq);
16016 if (IN_RECOVERY(tp->t_flags)) {
16017 if (SEQ_LT(high_seq, tp->snd_recover) &&
16018 (SEQ_LT(high_seq, tp->snd_max))) {
16019 tcp_rack_partialack(tp);
16020 } else {
16021 rack_post_recovery(tp, high_seq);
16022 post_recovery = 1;
16023 }
16024 } else if ((rack->rto_from_rec == 1) &&
16025 SEQ_GEQ(high_seq, tp->snd_recover)) {
16026 /*
16027 * We were in recovery, hit a rxt timeout
16028 * and never re-entered recovery. The timeout(s)
16029 * made up all the lost data. In such a case
16030 * we need to clear the rto_from_rec flag.
16031 */
16032 rack->rto_from_rec = 0;
16033 }
16034 /* Handle the rack-log-ack part (sendmap) */
16035 if ((sbused(&so->so_snd) == 0) &&
16036 (acked > acked_amount) &&
16037 (tp->t_state >= TCPS_FIN_WAIT_1) &&
16038 (tp->t_flags & TF_SENTFIN)) {
16039 /*
16040 * We must be sure our fin
16041 * was sent and acked (we can be
16042 * in FIN_WAIT_1 without having
16043 * sent the fin).
16044 */
16045 ourfinisacked = 1;
16046 /*
16047 * Lets make sure snd_una is updated
16048 * since most likely acked_amount = 0 (it
16049 * should be).
16050 */
16051 tp->snd_una = high_seq;
16052 }
16053 /* Did we make a RTO error? */
16054 if ((tp->t_flags & TF_PREVVALID) &&
16055 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16056 tp->t_flags &= ~TF_PREVVALID;
16057 if (tp->t_rxtshift == 1 &&
16058 (int)(ticks - tp->t_badrxtwin) < 0)
16059 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16060 }
16061 /* Handle the data in the socket buffer */
16062 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16063 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16064 if (acked_amount > 0) {
16065 uint32_t p_cwnd;
16066 struct mbuf *mfree;
16067
16068 if (post_recovery) {
16069 /*
16070 * Grab the segsiz, multiply by 2 and add the snd_cwnd
16071 * that is the max the CC should add if we are exiting
16072 * recovery and doing a late add.
16073 */
16074 p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16075 p_cwnd <<= 1;
16076 p_cwnd += tp->snd_cwnd;
16077 }
16078 rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
16079 if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
16080 /* Must be non-newreno (cubic) getting too ahead of itself */
16081 tp->snd_cwnd = p_cwnd;
16082 }
16083 SOCK_SENDBUF_LOCK(so);
16084 mfree = sbcut_locked(&so->so_snd, acked_amount);
16085 tp->snd_una = high_seq;
16086 /* Note we want to hold the sb lock through the sendmap adjust */
16087 rack_adjust_sendmap_head(rack, &so->so_snd);
16088 /* Wake up the socket if we have room to write more */
16089 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16090 sowwakeup_locked(so);
16091 m_freem(mfree);
16092 }
16093 /* update progress */
16094 tp->t_acktime = ticks;
16095 rack_log_progress_event(rack, tp, tp->t_acktime,
16096 PROGRESS_UPDATE, __LINE__);
16097 /* Clear out shifts and such */
16098 tp->t_rxtshift = 0;
16099 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16100 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16101 rack->rc_tlp_in_progress = 0;
16102 rack->r_ctl.rc_tlp_cnt_out = 0;
16103 /* Send recover and snd_nxt must be dragged along */
16104 if (SEQ_GT(tp->snd_una, tp->snd_recover))
16105 tp->snd_recover = tp->snd_una;
16106 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16107 tp->snd_nxt = tp->snd_max;
16108 /*
16109 * If the RXT timer is running we want to
16110 * stop it, so we can restart a TLP (or new RXT).
16111 */
16112 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16113 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16114 tp->snd_wl2 = high_seq;
16115 tp->t_dupacks = 0;
16116 if (under_pacing &&
16117 (rack->use_fixed_rate == 0) &&
16118 (rack->in_probe_rtt == 0) &&
16119 rack->rc_gp_dyn_mul &&
16120 rack->rc_always_pace) {
16121 /* Check if we are dragging bottom */
16122 rack_check_bottom_drag(tp, rack, so);
16123 }
16124 if (tp->snd_una == tp->snd_max) {
16125 tp->t_flags &= ~TF_PREVVALID;
16126 rack->r_ctl.retran_during_recovery = 0;
16127 rack->rc_suspicious = 0;
16128 rack->r_ctl.dsack_byte_cnt = 0;
16129 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16130 if (rack->r_ctl.rc_went_idle_time == 0)
16131 rack->r_ctl.rc_went_idle_time = 1;
16132 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16133 if (sbavail(&tptosocket(tp)->so_snd) == 0)
16134 tp->t_acktime = 0;
16135 /* Set so we might enter persists... */
16136 rack->r_wanted_output = 1;
16137 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16138 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16139 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16140 (sbavail(&so->so_snd) == 0) &&
16141 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16142 /*
16143 * The socket was gone and the
16144 * peer sent data (not now in the past), time to
16145 * reset him.
16146 */
16147 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16148 /* tcp_close will kill the inp pre-log the Reset */
16149 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16150 #ifdef TCP_ACCOUNTING
16151 rdstc = get_cyclecount();
16152 if (rdstc > ts_val) {
16153 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16154 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16155 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16156 }
16157 }
16158 #endif
16159 m_freem(m);
16160 tp = tcp_close(tp);
16161 if (tp == NULL) {
16162 #ifdef TCP_ACCOUNTING
16163 sched_unpin();
16164 #endif
16165 return (1);
16166 }
16167 /*
16168 * We would normally do drop-with-reset which would
16169 * send back a reset. We can't since we don't have
16170 * all the needed bits. Instead lets arrange for
16171 * a call to tcp_output(). That way since we
16172 * are in the closed state we will generate a reset.
16173 *
16174 * Note if tcp_accounting is on we don't unpin since
16175 * we do that after the goto label.
16176 */
16177 goto send_out_a_rst;
16178 }
16179 if ((sbused(&so->so_snd) == 0) &&
16180 (tp->t_state >= TCPS_FIN_WAIT_1) &&
16181 (tp->t_flags & TF_SENTFIN)) {
16182 /*
16183 * If we can't receive any more data, then closing user can
16184 * proceed. Starting the timer is contrary to the
16185 * specification, but if we don't get a FIN we'll hang
16186 * forever.
16187 *
16188 */
16189 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16190 soisdisconnected(so);
16191 tcp_timer_activate(tp, TT_2MSL,
16192 (tcp_fast_finwait2_recycle ?
16193 tcp_finwait2_timeout :
16194 TP_MAXIDLE(tp)));
16195 }
16196 if (ourfinisacked == 0) {
16197 /*
16198 * We don't change to fin-wait-2 if we have our fin acked
16199 * which means we are probably in TCPS_CLOSING.
16200 */
16201 tcp_state_change(tp, TCPS_FIN_WAIT_2);
16202 }
16203 }
16204 }
16205 /* Wake up the socket if we have room to write more */
16206 if (sbavail(&so->so_snd)) {
16207 rack->r_wanted_output = 1;
16208 if (ctf_progress_timeout_check(tp, true)) {
16209 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16210 tp, tick, PROGRESS_DROP, __LINE__);
16211 /*
16212 * We cheat here and don't send a RST, we should send one
16213 * when the pacer drops the connection.
16214 */
16215 #ifdef TCP_ACCOUNTING
16216 rdstc = get_cyclecount();
16217 if (rdstc > ts_val) {
16218 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16219 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16220 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16221 }
16222 }
16223 sched_unpin();
16224 #endif
16225 (void)tcp_drop(tp, ETIMEDOUT);
16226 m_freem(m);
16227 return (1);
16228 }
16229 }
16230 if (ourfinisacked) {
16231 switch(tp->t_state) {
16232 case TCPS_CLOSING:
16233 #ifdef TCP_ACCOUNTING
16234 rdstc = get_cyclecount();
16235 if (rdstc > ts_val) {
16236 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16237 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16238 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16239 }
16240 }
16241 sched_unpin();
16242 #endif
16243 tcp_twstart(tp);
16244 m_freem(m);
16245 return (1);
16246 break;
16247 case TCPS_LAST_ACK:
16248 #ifdef TCP_ACCOUNTING
16249 rdstc = get_cyclecount();
16250 if (rdstc > ts_val) {
16251 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16252 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16253 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16254 }
16255 }
16256 sched_unpin();
16257 #endif
16258 tp = tcp_close(tp);
16259 ctf_do_drop(m, tp);
16260 return (1);
16261 break;
16262 case TCPS_FIN_WAIT_1:
16263 #ifdef TCP_ACCOUNTING
16264 rdstc = get_cyclecount();
16265 if (rdstc > ts_val) {
16266 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16267 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16268 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16269 }
16270 }
16271 #endif
16272 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16273 soisdisconnected(so);
16274 tcp_timer_activate(tp, TT_2MSL,
16275 (tcp_fast_finwait2_recycle ?
16276 tcp_finwait2_timeout :
16277 TP_MAXIDLE(tp)));
16278 }
16279 tcp_state_change(tp, TCPS_FIN_WAIT_2);
16280 break;
16281 default:
16282 break;
16283 }
16284 }
16285 if (rack->r_fast_output) {
16286 /*
16287 * We re doing fast output.. can we expand that?
16288 */
16289 rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16290 }
16291 #ifdef TCP_ACCOUNTING
16292 rdstc = get_cyclecount();
16293 if (rdstc > ts_val) {
16294 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16295 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16296 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16297 }
16298 }
16299
16300 } else if (win_up_req) {
16301 rdstc = get_cyclecount();
16302 if (rdstc > ts_val) {
16303 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16304 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16305 }
16306 }
16307 #endif
16308 }
16309 /* Now is there a next packet, if so we are done */
16310 m_freem(m);
16311 did_out = 0;
16312 if (nxt_pkt) {
16313 #ifdef TCP_ACCOUNTING
16314 sched_unpin();
16315 #endif
16316 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16317 return (0);
16318 }
16319 rack_handle_might_revert(tp, rack);
16320 ctf_calc_rwin(so, tp);
16321 if ((rack->r_wanted_output != 0) ||
16322 (rack->r_fast_output != 0) ||
16323 (tp->t_flags & TF_ACKNOW )) {
16324 send_out_a_rst:
16325 if (tcp_output(tp) < 0) {
16326 #ifdef TCP_ACCOUNTING
16327 sched_unpin();
16328 #endif
16329 return (1);
16330 }
16331 did_out = 1;
16332 }
16333 if (tp->t_flags2 & TF2_HPTS_CALLS)
16334 tp->t_flags2 &= ~TF2_HPTS_CALLS;
16335 rack_free_trim(rack);
16336 #ifdef TCP_ACCOUNTING
16337 sched_unpin();
16338 #endif
16339 rack_timer_audit(tp, rack, &so->so_snd);
16340 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16341 return (0);
16342 }
16343
16344 #define TCP_LRO_TS_OPTION \
16345 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16346 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16347
16348 static int
rack_do_segment_nounlock(struct tcpcb * tp,struct mbuf * m,struct tcphdr * th,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos,int32_t nxt_pkt,struct timeval * tv)16349 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16350 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16351 struct timeval *tv)
16352 {
16353 struct inpcb *inp = tptoinpcb(tp);
16354 struct socket *so = tptosocket(tp);
16355 #ifdef TCP_ACCOUNTING
16356 uint64_t ts_val;
16357 #endif
16358 int32_t thflags, retval, did_out = 0;
16359 int32_t way_out = 0;
16360 /*
16361 * cts - is the current time from tv (caller gets ts) in microseconds.
16362 * ms_cts - is the current time from tv in milliseconds.
16363 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16364 */
16365 uint32_t cts, us_cts, ms_cts;
16366 uint32_t tiwin;
16367 struct timespec ts;
16368 struct tcpopt to;
16369 struct tcp_rack *rack;
16370 struct rack_sendmap *rsm;
16371 int32_t prev_state = 0;
16372 int no_output = 0;
16373 int slot_remaining = 0;
16374 #ifdef TCP_ACCOUNTING
16375 int ack_val_set = 0xf;
16376 #endif
16377 int nsegs;
16378
16379 NET_EPOCH_ASSERT();
16380 INP_WLOCK_ASSERT(inp);
16381
16382 /*
16383 * tv passed from common code is from either M_TSTMP_LRO or
16384 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16385 */
16386 rack = (struct tcp_rack *)tp->t_fb_ptr;
16387 if (rack->rack_deferred_inited == 0) {
16388 /*
16389 * If we are the connecting socket we will
16390 * hit rack_init() when no sequence numbers
16391 * are setup. This makes it so we must defer
16392 * some initialization. Call that now.
16393 */
16394 rack_deferred_init(tp, rack);
16395 }
16396 /*
16397 * Check to see if we need to skip any output plans. This
16398 * can happen in the non-LRO path where we are pacing and
16399 * must process the ack coming in but need to defer sending
16400 * anything becase a pacing timer is running.
16401 */
16402 us_cts = tcp_tv_to_usectick(tv);
16403 if (m->m_flags & M_ACKCMP) {
16404 /*
16405 * All compressed ack's are ack's by definition so
16406 * remove any ack required flag and then do the processing.
16407 */
16408 rack->rc_ack_required = 0;
16409 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16410 }
16411 thflags = tcp_get_flags(th);
16412 if ((rack->rc_always_pace == 1) &&
16413 (rack->rc_ack_can_sendout_data == 0) &&
16414 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16415 (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16416 /*
16417 * Ok conditions are right for queuing the packets
16418 * but we do have to check the flags in the inp, it
16419 * could be, if a sack is present, we want to be awoken and
16420 * so should process the packets.
16421 */
16422 slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16423 if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16424 no_output = 1;
16425 } else {
16426 /*
16427 * If there is no options, or just a
16428 * timestamp option, we will want to queue
16429 * the packets. This is the same that LRO does
16430 * and will need to change with accurate ECN.
16431 */
16432 uint32_t *ts_ptr;
16433 int optlen;
16434
16435 optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16436 ts_ptr = (uint32_t *)(th + 1);
16437 if ((optlen == 0) ||
16438 ((optlen == TCPOLEN_TSTAMP_APPA) &&
16439 (*ts_ptr == TCP_LRO_TS_OPTION)))
16440 no_output = 1;
16441 }
16442 if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16443 /*
16444 * It is unrealistic to think we can pace in less than
16445 * the minimum granularity of the pacer (def:250usec). So
16446 * if we have less than that time remaining we should go
16447 * ahead and allow output to be "early". We will attempt to
16448 * make up for it in any pacing time we try to apply on
16449 * the outbound packet.
16450 */
16451 no_output = 0;
16452 }
16453 }
16454 /*
16455 * If there is a RST or FIN lets dump out the bw
16456 * with a FIN the connection may go on but we
16457 * may not.
16458 */
16459 if ((thflags & TH_FIN) || (thflags & TH_RST))
16460 rack_log_pacing_delay_calc(rack,
16461 rack->r_ctl.gp_bw,
16462 0,
16463 0,
16464 rack_get_gp_est(rack), /* delRate */
16465 rack_get_lt_bw(rack), /* rttProp */
16466 20, __LINE__, NULL, 0);
16467 if (m->m_flags & M_ACKCMP) {
16468 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16469 }
16470 cts = tcp_tv_to_usectick(tv);
16471 ms_cts = tcp_tv_to_mssectick(tv);
16472 nsegs = m->m_pkthdr.lro_nsegs;
16473 counter_u64_add(rack_proc_non_comp_ack, 1);
16474 #ifdef TCP_ACCOUNTING
16475 sched_pin();
16476 if (thflags & TH_ACK)
16477 ts_val = get_cyclecount();
16478 #endif
16479 if ((m->m_flags & M_TSTMP) ||
16480 (m->m_flags & M_TSTMP_LRO)) {
16481 mbuf_tstmp2timespec(m, &ts);
16482 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16483 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16484 } else
16485 rack->r_ctl.act_rcv_time = *tv;
16486 kern_prefetch(rack, &prev_state);
16487 prev_state = 0;
16488 /*
16489 * Unscale the window into a 32-bit value. For the SYN_SENT state
16490 * the scale is zero.
16491 */
16492 tiwin = th->th_win << tp->snd_scale;
16493 #ifdef TCP_ACCOUNTING
16494 if (thflags & TH_ACK) {
16495 /*
16496 * We have a tradeoff here. We can either do what we are
16497 * doing i.e. pinning to this CPU and then doing the accounting
16498 * <or> we could do a critical enter, setup the rdtsc and cpu
16499 * as in below, and then validate we are on the same CPU on
16500 * exit. I have choosen to not do the critical enter since
16501 * that often will gain you a context switch, and instead lock
16502 * us (line above this if) to the same CPU with sched_pin(). This
16503 * means we may be context switched out for a higher priority
16504 * interupt but we won't be moved to another CPU.
16505 *
16506 * If this occurs (which it won't very often since we most likely
16507 * are running this code in interupt context and only a higher
16508 * priority will bump us ... clock?) we will falsely add in
16509 * to the time the interupt processing time plus the ack processing
16510 * time. This is ok since its a rare event.
16511 */
16512 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16513 ctf_fixed_maxseg(tp));
16514 }
16515 #endif
16516 /*
16517 * Parse options on any incoming segment.
16518 */
16519 memset(&to, 0, sizeof(to));
16520 tcp_dooptions(&to, (u_char *)(th + 1),
16521 (th->th_off << 2) - sizeof(struct tcphdr),
16522 (thflags & TH_SYN) ? TO_SYN : 0);
16523 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16524 __func__));
16525 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16526 __func__));
16527 if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
16528 /*
16529 * We don't look at sack's from the
16530 * peer because the MSS is too small which
16531 * can subject us to an attack.
16532 */
16533 to.to_flags &= ~TOF_SACK;
16534 }
16535 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16536 (tp->t_flags & TF_GPUTINPROG)) {
16537 /*
16538 * We have a goodput in progress
16539 * and we have entered a late state.
16540 * Do we have enough data in the sb
16541 * to handle the GPUT request?
16542 */
16543 uint32_t bytes;
16544
16545 bytes = tp->gput_ack - tp->gput_seq;
16546 if (SEQ_GT(tp->gput_seq, tp->snd_una))
16547 bytes += tp->gput_seq - tp->snd_una;
16548 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16549 /*
16550 * There are not enough bytes in the socket
16551 * buffer that have been sent to cover this
16552 * measurement. Cancel it.
16553 */
16554 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16555 rack->r_ctl.rc_gp_srtt /*flex1*/,
16556 tp->gput_seq,
16557 0, 0, 18, __LINE__, NULL, 0);
16558 tp->t_flags &= ~TF_GPUTINPROG;
16559 }
16560 }
16561 if (tcp_bblogging_on(rack->rc_tp)) {
16562 union tcp_log_stackspecific log;
16563 struct timeval ltv;
16564 #ifdef TCP_REQUEST_TRK
16565 struct tcp_sendfile_track *tcp_req;
16566
16567 if (SEQ_GT(th->th_ack, tp->snd_una)) {
16568 tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16569 } else {
16570 tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16571 }
16572 #endif
16573 memset(&log, 0, sizeof(log));
16574 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16575 if (rack->rack_no_prr == 0)
16576 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16577 else
16578 log.u_bbr.flex1 = 0;
16579 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16580 log.u_bbr.use_lt_bw <<= 1;
16581 log.u_bbr.use_lt_bw |= rack->r_might_revert;
16582 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16583 log.u_bbr.bbr_state = rack->rc_free_cnt;
16584 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16585 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16586 log.u_bbr.flex3 = m->m_flags;
16587 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16588 log.u_bbr.lost = thflags;
16589 log.u_bbr.pacing_gain = 0x1;
16590 #ifdef TCP_ACCOUNTING
16591 log.u_bbr.cwnd_gain = ack_val_set;
16592 #endif
16593 log.u_bbr.flex7 = 2;
16594 if (m->m_flags & M_TSTMP) {
16595 /* Record the hardware timestamp if present */
16596 mbuf_tstmp2timespec(m, &ts);
16597 ltv.tv_sec = ts.tv_sec;
16598 ltv.tv_usec = ts.tv_nsec / 1000;
16599 log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v);
16600 } else if (m->m_flags & M_TSTMP_LRO) {
16601 /* Record the LRO the arrival timestamp */
16602 mbuf_tstmp2timespec(m, &ts);
16603 ltv.tv_sec = ts.tv_sec;
16604 ltv.tv_usec = ts.tv_nsec / 1000;
16605 log.u_bbr.flex5 = tcp_tv_to_usectick(<v);
16606 }
16607 log.u_bbr.timeStamp = tcp_get_usecs(<v);
16608 /* Log the rcv time */
16609 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16610 #ifdef TCP_REQUEST_TRK
16611 log.u_bbr.applimited = tp->t_tcpreq_closed;
16612 log.u_bbr.applimited <<= 8;
16613 log.u_bbr.applimited |= tp->t_tcpreq_open;
16614 log.u_bbr.applimited <<= 8;
16615 log.u_bbr.applimited |= tp->t_tcpreq_req;
16616 if (tcp_req) {
16617 /* Copy out any client req info */
16618 /* seconds */
16619 log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16620 /* useconds */
16621 log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16622 log.u_bbr.rttProp = tcp_req->timestamp;
16623 log.u_bbr.cur_del_rate = tcp_req->start;
16624 if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16625 log.u_bbr.flex8 |= 1;
16626 } else {
16627 log.u_bbr.flex8 |= 2;
16628 log.u_bbr.bw_inuse = tcp_req->end;
16629 }
16630 log.u_bbr.flex6 = tcp_req->start_seq;
16631 if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16632 log.u_bbr.flex8 |= 4;
16633 log.u_bbr.epoch = tcp_req->end_seq;
16634 }
16635 }
16636 #endif
16637 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16638 tlen, &log, true, <v);
16639 }
16640 /* Remove ack required flag if set, we have one */
16641 if (thflags & TH_ACK)
16642 rack->rc_ack_required = 0;
16643 rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16644 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16645 way_out = 4;
16646 retval = 0;
16647 m_freem(m);
16648 goto done_with_input;
16649 }
16650 /*
16651 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16652 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16653 */
16654 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16655 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16656 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16657 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16658 #ifdef TCP_ACCOUNTING
16659 sched_unpin();
16660 #endif
16661 return (1);
16662 }
16663 /*
16664 * If timestamps were negotiated during SYN/ACK and a
16665 * segment without a timestamp is received, silently drop
16666 * the segment, unless it is a RST segment or missing timestamps are
16667 * tolerated.
16668 * See section 3.2 of RFC 7323.
16669 */
16670 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16671 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16672 way_out = 5;
16673 retval = 0;
16674 m_freem(m);
16675 goto done_with_input;
16676 }
16677 /*
16678 * Segment received on connection. Reset idle time and keep-alive
16679 * timer. XXX: This should be done after segment validation to
16680 * ignore broken/spoofed segs.
16681 */
16682 if (tp->t_idle_reduce &&
16683 (tp->snd_max == tp->snd_una) &&
16684 (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16685 counter_u64_add(rack_input_idle_reduces, 1);
16686 rack_cc_after_idle(rack, tp);
16687 }
16688 tp->t_rcvtime = ticks;
16689 #ifdef STATS
16690 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16691 #endif
16692 if (tiwin > rack->r_ctl.rc_high_rwnd)
16693 rack->r_ctl.rc_high_rwnd = tiwin;
16694 /*
16695 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16696 * this to occur after we've validated the segment.
16697 */
16698 if (tcp_ecn_input_segment(tp, thflags, tlen,
16699 tcp_packets_this_ack(tp, th->th_ack),
16700 iptos))
16701 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16702
16703 /*
16704 * If echoed timestamp is later than the current time, fall back to
16705 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
16706 * were used when this connection was established.
16707 */
16708 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16709 to.to_tsecr -= tp->ts_offset;
16710 if (TSTMP_GT(to.to_tsecr, ms_cts))
16711 to.to_tsecr = 0;
16712 }
16713 if ((rack->r_rcvpath_rtt_up == 1) &&
16714 (to.to_flags & TOF_TS) &&
16715 (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
16716 uint32_t rtt = 0;
16717
16718 /*
16719 * We are receiving only and thus not sending
16720 * data to do an RTT. We set a flag when we first
16721 * sent this TS to the peer. We now have it back
16722 * and have an RTT to share. We log it as a conf
16723 * 4, we are not so sure about it.. since we
16724 * may have lost an ack.
16725 */
16726 if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
16727 rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
16728 rack->r_rcvpath_rtt_up = 0;
16729 /* Submit and commit the timer */
16730 if (rtt > 0) {
16731 tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
16732 tcp_rack_xmit_timer_commit(rack, tp);
16733 }
16734 }
16735 /*
16736 * If its the first time in we need to take care of options and
16737 * verify we can do SACK for rack!
16738 */
16739 if (rack->r_state == 0) {
16740 /* Should be init'd by rack_init() */
16741 KASSERT(rack->rc_inp != NULL,
16742 ("%s: rack->rc_inp unexpectedly NULL", __func__));
16743 if (rack->rc_inp == NULL) {
16744 rack->rc_inp = inp;
16745 }
16746
16747 /*
16748 * Process options only when we get SYN/ACK back. The SYN
16749 * case for incoming connections is handled in tcp_syncache.
16750 * According to RFC1323 the window field in a SYN (i.e., a
16751 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16752 * this is traditional behavior, may need to be cleaned up.
16753 */
16754 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16755 /* Handle parallel SYN for ECN */
16756 tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16757 if ((to.to_flags & TOF_SCALE) &&
16758 (tp->t_flags & TF_REQ_SCALE)) {
16759 tp->t_flags |= TF_RCVD_SCALE;
16760 tp->snd_scale = to.to_wscale;
16761 } else
16762 tp->t_flags &= ~TF_REQ_SCALE;
16763 /*
16764 * Initial send window. It will be updated with the
16765 * next incoming segment to the scaled value.
16766 */
16767 tp->snd_wnd = th->th_win;
16768 rack_validate_fo_sendwin_up(tp, rack);
16769 if ((to.to_flags & TOF_TS) &&
16770 (tp->t_flags & TF_REQ_TSTMP)) {
16771 tp->t_flags |= TF_RCVD_TSTMP;
16772 tp->ts_recent = to.to_tsval;
16773 tp->ts_recent_age = cts;
16774 } else
16775 tp->t_flags &= ~TF_REQ_TSTMP;
16776 if (to.to_flags & TOF_MSS) {
16777 tcp_mss(tp, to.to_mss);
16778 }
16779 if ((tp->t_flags & TF_SACK_PERMIT) &&
16780 (to.to_flags & TOF_SACKPERM) == 0)
16781 tp->t_flags &= ~TF_SACK_PERMIT;
16782 if (tp->t_flags & TF_FASTOPEN) {
16783 if (to.to_flags & TOF_FASTOPEN) {
16784 uint16_t mss;
16785
16786 if (to.to_flags & TOF_MSS)
16787 mss = to.to_mss;
16788 else
16789 if ((inp->inp_vflag & INP_IPV6) != 0)
16790 mss = TCP6_MSS;
16791 else
16792 mss = TCP_MSS;
16793 tcp_fastopen_update_cache(tp, mss,
16794 to.to_tfo_len, to.to_tfo_cookie);
16795 } else
16796 tcp_fastopen_disable_path(tp);
16797 }
16798 }
16799 /*
16800 * At this point we are at the initial call. Here we decide
16801 * if we are doing RACK or not. We do this by seeing if
16802 * TF_SACK_PERMIT is set and the sack-not-required is clear.
16803 * The code now does do dup-ack counting so if you don't
16804 * switch back you won't get rack & TLP, but you will still
16805 * get this stack.
16806 */
16807
16808 if ((rack_sack_not_required == 0) &&
16809 ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16810 tcp_switch_back_to_default(tp);
16811 (*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16812 tlen, iptos);
16813 #ifdef TCP_ACCOUNTING
16814 sched_unpin();
16815 #endif
16816 return (1);
16817 }
16818 tcp_set_hpts(tp);
16819 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16820 }
16821 if (thflags & TH_FIN)
16822 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16823 us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16824 if ((rack->rc_gp_dyn_mul) &&
16825 (rack->use_fixed_rate == 0) &&
16826 (rack->rc_always_pace)) {
16827 /* Check in on probertt */
16828 rack_check_probe_rtt(rack, cts);
16829 }
16830 rack_clear_rate_sample(rack);
16831 if ((rack->forced_ack) &&
16832 ((tcp_get_flags(th) & TH_RST) == 0)) {
16833 rack_handle_probe_response(rack, tiwin, us_cts);
16834 }
16835 /*
16836 * This is the one exception case where we set the rack state
16837 * always. All other times (timers etc) we must have a rack-state
16838 * set (so we assure we have done the checks above for SACK).
16839 */
16840 rack->r_ctl.rc_rcvtime = cts;
16841 if (rack->r_state != tp->t_state)
16842 rack_set_state(tp, rack);
16843 if (SEQ_GT(th->th_ack, tp->snd_una) &&
16844 (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16845 kern_prefetch(rsm, &prev_state);
16846 prev_state = rack->r_state;
16847 if ((thflags & TH_RST) &&
16848 ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16849 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16850 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16851 /* The connection will be killed by a reset check the tracepoint */
16852 tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16853 }
16854 retval = (*rack->r_substate) (m, th, so,
16855 tp, &to, drop_hdrlen,
16856 tlen, tiwin, thflags, nxt_pkt, iptos);
16857 if (retval == 0) {
16858 /*
16859 * If retval is 1 the tcb is unlocked and most likely the tp
16860 * is gone.
16861 */
16862 INP_WLOCK_ASSERT(inp);
16863 if ((rack->rc_gp_dyn_mul) &&
16864 (rack->rc_always_pace) &&
16865 (rack->use_fixed_rate == 0) &&
16866 rack->in_probe_rtt &&
16867 (rack->r_ctl.rc_time_probertt_starts == 0)) {
16868 /*
16869 * If we are going for target, lets recheck before
16870 * we output.
16871 */
16872 rack_check_probe_rtt(rack, cts);
16873 }
16874 if (rack->set_pacing_done_a_iw == 0) {
16875 /* How much has been acked? */
16876 if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16877 /* We have enough to set in the pacing segment size */
16878 rack->set_pacing_done_a_iw = 1;
16879 rack_set_pace_segments(tp, rack, __LINE__, NULL);
16880 }
16881 }
16882 tcp_rack_xmit_timer_commit(rack, tp);
16883 #ifdef TCP_ACCOUNTING
16884 /*
16885 * If we set the ack_val_se to what ack processing we are doing
16886 * we also want to track how many cycles we burned. Note
16887 * the bits after tcp_output we let be "free". This is because
16888 * we are also tracking the tcp_output times as well. Note the
16889 * use of 0xf here since we only have 11 counter (0 - 0xa) and
16890 * 0xf cannot be returned and is what we initialize it too to
16891 * indicate we are not doing the tabulations.
16892 */
16893 if (ack_val_set != 0xf) {
16894 uint64_t crtsc;
16895
16896 crtsc = get_cyclecount();
16897 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16898 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16899 }
16900 }
16901 #endif
16902 if ((nxt_pkt == 0) && (no_output == 0)) {
16903 if ((rack->r_wanted_output != 0) ||
16904 (tp->t_flags & TF_ACKNOW) ||
16905 (rack->r_fast_output != 0)) {
16906
16907 do_output_now:
16908 if (tcp_output(tp) < 0) {
16909 #ifdef TCP_ACCOUNTING
16910 sched_unpin();
16911 #endif
16912 return (1);
16913 }
16914 did_out = 1;
16915 }
16916 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16917 rack_free_trim(rack);
16918 } else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
16919 goto do_output_now;
16920 } else if ((no_output == 1) &&
16921 (nxt_pkt == 0) &&
16922 (tcp_in_hpts(rack->rc_tp) == 0)) {
16923 /*
16924 * We are not in hpts and we had a pacing timer up. Use
16925 * the remaining time (slot_remaining) to restart the timer.
16926 */
16927 KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
16928 rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
16929 rack_free_trim(rack);
16930 }
16931 /* Clear the flag, it may have been cleared by output but we may not have */
16932 if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
16933 tp->t_flags2 &= ~TF2_HPTS_CALLS;
16934 /*
16935 * The draft (v3) calls for us to use SEQ_GEQ, but that
16936 * causes issues when we are just going app limited. Lets
16937 * instead use SEQ_GT <or> where its equal but more data
16938 * is outstanding.
16939 *
16940 * Also make sure we are on the last ack of a series. We
16941 * have to have all the ack's processed in queue to know
16942 * if there is something left outstanding.
16943 */
16944 if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
16945 (rack->rc_new_rnd_needed == 0) &&
16946 (nxt_pkt == 0)) {
16947 /*
16948 * We have crossed into a new round with
16949 * the new snd_unae.
16950 */
16951 rack_new_round_setup(tp, rack, tp->snd_una);
16952 }
16953 if ((nxt_pkt == 0) &&
16954 ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
16955 (SEQ_GT(tp->snd_max, tp->snd_una) ||
16956 (tp->t_flags & TF_DELACK) ||
16957 ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16958 (tp->t_state <= TCPS_CLOSING)))) {
16959 /* We could not send (probably in the hpts but stopped the timer earlier)? */
16960 if ((tp->snd_max == tp->snd_una) &&
16961 ((tp->t_flags & TF_DELACK) == 0) &&
16962 (tcp_in_hpts(rack->rc_tp)) &&
16963 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
16964 /* keep alive not needed if we are hptsi output yet */
16965 ;
16966 } else {
16967 int late = 0;
16968 if (tcp_in_hpts(tp)) {
16969 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16970 us_cts = tcp_get_usecs(NULL);
16971 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
16972 rack->r_early = 1;
16973 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
16974 } else
16975 late = 1;
16976 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16977 }
16978 tcp_hpts_remove(tp);
16979 }
16980 if (late && (did_out == 0)) {
16981 /*
16982 * We are late in the sending
16983 * and we did not call the output
16984 * (this probably should not happen).
16985 */
16986 goto do_output_now;
16987 }
16988 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16989 }
16990 way_out = 1;
16991 } else if (nxt_pkt == 0) {
16992 /* Do we have the correct timer running? */
16993 rack_timer_audit(tp, rack, &so->so_snd);
16994 way_out = 2;
16995 }
16996 done_with_input:
16997 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
16998 if (did_out)
16999 rack->r_wanted_output = 0;
17000 }
17001
17002 #ifdef TCP_ACCOUNTING
17003 sched_unpin();
17004 #endif
17005 return (retval);
17006 }
17007
17008 static void
rack_do_segment(struct tcpcb * tp,struct mbuf * m,struct tcphdr * th,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos)17009 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17010 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17011 {
17012 struct timeval tv;
17013
17014 /* First lets see if we have old packets */
17015 if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17016 if (ctf_do_queued_segments(tp, 1)) {
17017 m_freem(m);
17018 return;
17019 }
17020 }
17021 if (m->m_flags & M_TSTMP_LRO) {
17022 mbuf_tstmp2timeval(m, &tv);
17023 } else {
17024 /* Should not be should we kassert instead? */
17025 tcp_get_usecs(&tv);
17026 }
17027 if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17028 &tv) == 0) {
17029 INP_WUNLOCK(tptoinpcb(tp));
17030 }
17031 }
17032
17033 struct rack_sendmap *
tcp_rack_output(struct tcpcb * tp,struct tcp_rack * rack,uint32_t tsused)17034 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17035 {
17036 struct rack_sendmap *rsm = NULL;
17037 int32_t idx;
17038 uint32_t srtt = 0, thresh = 0, ts_low = 0;
17039
17040 /* Return the next guy to be re-transmitted */
17041 if (tqhash_empty(rack->r_ctl.tqh)) {
17042 return (NULL);
17043 }
17044 if (tp->t_flags & TF_SENTFIN) {
17045 /* retran the end FIN? */
17046 return (NULL);
17047 }
17048 /* ok lets look at this one */
17049 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17050 if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17051 return (rsm);
17052 }
17053 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17054 goto check_it;
17055 }
17056 rsm = rack_find_lowest_rsm(rack);
17057 if (rsm == NULL) {
17058 return (NULL);
17059 }
17060 check_it:
17061 if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
17062 (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17063 /*
17064 * No sack so we automatically do the 3 strikes and
17065 * retransmit (no rack timer would be started).
17066 */
17067 return (rsm);
17068 }
17069 if (rsm->r_flags & RACK_ACKED) {
17070 return (NULL);
17071 }
17072 if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17073 (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17074 /* Its not yet ready */
17075 return (NULL);
17076 }
17077 srtt = rack_grab_rtt(tp, rack);
17078 idx = rsm->r_rtr_cnt - 1;
17079 ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17080 thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
17081 if ((tsused == ts_low) ||
17082 (TSTMP_LT(tsused, ts_low))) {
17083 /* No time since sending */
17084 return (NULL);
17085 }
17086 if ((tsused - ts_low) < thresh) {
17087 /* It has not been long enough yet */
17088 return (NULL);
17089 }
17090 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17091 ((rsm->r_flags & RACK_SACK_PASSED))) {
17092 /*
17093 * We have passed the dup-ack threshold <or>
17094 * a SACK has indicated this is missing.
17095 * Note that if you are a declared attacker
17096 * it is only the dup-ack threshold that
17097 * will cause retransmits.
17098 */
17099 /* log retransmit reason */
17100 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17101 rack->r_fast_output = 0;
17102 return (rsm);
17103 }
17104 return (NULL);
17105 }
17106
17107 static void
rack_log_pacing_delay_calc(struct tcp_rack * rack,uint32_t len,uint32_t slot,uint64_t bw_est,uint64_t bw,uint64_t len_time,int method,int line,struct rack_sendmap * rsm,uint8_t quality)17108 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
17109 uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17110 int line, struct rack_sendmap *rsm, uint8_t quality)
17111 {
17112 if (tcp_bblogging_on(rack->rc_tp)) {
17113 union tcp_log_stackspecific log;
17114 struct timeval tv;
17115
17116 if (rack_verbose_logging == 0) {
17117 /*
17118 * We are not verbose screen out all but
17119 * ones we always want.
17120 */
17121 if ((method != 2) &&
17122 (method != 3) &&
17123 (method != 7) &&
17124 (method != 89) &&
17125 (method != 14) &&
17126 (method != 20)) {
17127 return;
17128 }
17129 }
17130 memset(&log, 0, sizeof(log));
17131 log.u_bbr.flex1 = slot;
17132 log.u_bbr.flex2 = len;
17133 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17134 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17135 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17136 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17137 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17138 log.u_bbr.use_lt_bw <<= 1;
17139 log.u_bbr.use_lt_bw |= rack->r_late;
17140 log.u_bbr.use_lt_bw <<= 1;
17141 log.u_bbr.use_lt_bw |= rack->r_early;
17142 log.u_bbr.use_lt_bw <<= 1;
17143 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17144 log.u_bbr.use_lt_bw <<= 1;
17145 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17146 log.u_bbr.use_lt_bw <<= 1;
17147 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17148 log.u_bbr.use_lt_bw <<= 1;
17149 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17150 log.u_bbr.use_lt_bw <<= 1;
17151 log.u_bbr.use_lt_bw |= rack->gp_ready;
17152 log.u_bbr.pkt_epoch = line;
17153 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17154 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17155 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17156 log.u_bbr.bw_inuse = bw_est;
17157 log.u_bbr.delRate = bw;
17158 if (rack->r_ctl.gp_bw == 0)
17159 log.u_bbr.cur_del_rate = 0;
17160 else
17161 log.u_bbr.cur_del_rate = rack_get_bw(rack);
17162 log.u_bbr.rttProp = len_time;
17163 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17164 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17165 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17166 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17167 /* We are in slow start */
17168 log.u_bbr.flex7 = 1;
17169 } else {
17170 /* we are on congestion avoidance */
17171 log.u_bbr.flex7 = 0;
17172 }
17173 log.u_bbr.flex8 = method;
17174 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17175 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17176 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17177 log.u_bbr.cwnd_gain <<= 1;
17178 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17179 log.u_bbr.cwnd_gain <<= 1;
17180 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17181 log.u_bbr.bbr_substate = quality;
17182 log.u_bbr.bbr_state = rack->dgp_on;
17183 log.u_bbr.bbr_state <<= 1;
17184 log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17185 log.u_bbr.bbr_state <<= 2;
17186 TCP_LOG_EVENTP(rack->rc_tp, NULL,
17187 &rack->rc_inp->inp_socket->so_rcv,
17188 &rack->rc_inp->inp_socket->so_snd,
17189 BBR_LOG_HPTSI_CALC, 0,
17190 0, &log, false, &tv);
17191 }
17192 }
17193
17194 static uint32_t
rack_get_pacing_len(struct tcp_rack * rack,uint64_t bw,uint32_t mss)17195 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17196 {
17197 uint32_t new_tso, user_max, pace_one;
17198
17199 user_max = rack->rc_user_set_max_segs * mss;
17200 if (rack->rc_force_max_seg) {
17201 return (user_max);
17202 }
17203 if (rack->use_fixed_rate &&
17204 ((rack->r_ctl.crte == NULL) ||
17205 (bw != rack->r_ctl.crte->rate))) {
17206 /* Use the user mss since we are not exactly matched */
17207 return (user_max);
17208 }
17209 if (rack_pace_one_seg ||
17210 (rack->r_ctl.rc_user_set_min_segs == 1))
17211 pace_one = 1;
17212 else
17213 pace_one = 0;
17214
17215 new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17216 pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17217 if (new_tso > user_max)
17218 new_tso = user_max;
17219 if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17220 if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17221 new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17222 }
17223 if (rack->r_ctl.rc_user_set_min_segs &&
17224 ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17225 new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17226 return (new_tso);
17227 }
17228
17229 static uint64_t
rack_arrive_at_discounted_rate(struct tcp_rack * rack,uint64_t window_input,uint32_t * rate_set,uint32_t * gain_b)17230 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17231 {
17232 uint64_t reduced_win;
17233 uint32_t gain;
17234
17235 if (window_input < rc_init_window(rack)) {
17236 /*
17237 * The cwnd is collapsed to
17238 * nearly zero, maybe because of a time-out?
17239 * Lets drop back to the lt-bw.
17240 */
17241 reduced_win = rack_get_lt_bw(rack);
17242 /* Set the flag so the caller knows its a rate and not a reduced window */
17243 *rate_set = 1;
17244 gain = 100;
17245 } else if (IN_RECOVERY(rack->rc_tp->t_flags)) {
17246 /*
17247 * If we are in recover our cwnd needs to be less for
17248 * our pacing consideration.
17249 */
17250 if (rack->rack_hibeta == 0) {
17251 reduced_win = window_input / 2;
17252 gain = 50;
17253 } else {
17254 reduced_win = window_input * rack->r_ctl.saved_hibeta;
17255 reduced_win /= 100;
17256 gain = rack->r_ctl.saved_hibeta;
17257 }
17258 } else {
17259 /*
17260 * Apply Timely factor to increase/decrease the
17261 * amount we are pacing at.
17262 */
17263 gain = rack_get_output_gain(rack, NULL);
17264 if (gain > rack_gain_p5_ub) {
17265 gain = rack_gain_p5_ub;
17266 }
17267 reduced_win = window_input * gain;
17268 reduced_win /= 100;
17269 }
17270 if (gain_b != NULL)
17271 *gain_b = gain;
17272 /*
17273 * What is being returned here is a trimmed down
17274 * window values in all cases where rate_set is left
17275 * at 0. In one case we actually return the rate (lt_bw).
17276 * the "reduced_win" is returned as a slimmed down cwnd that
17277 * is then calculated by the caller into a rate when rate_set
17278 * is 0.
17279 */
17280 return (reduced_win);
17281 }
17282
17283 static int32_t
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)17284 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)
17285 {
17286 uint64_t lentim, fill_bw;
17287
17288 rack->r_via_fill_cw = 0;
17289 if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17290 return (slot);
17291 if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17292 return (slot);
17293 if (rack->r_ctl.rc_last_us_rtt == 0)
17294 return (slot);
17295 if (rack->rc_pace_fill_if_rttin_range &&
17296 (rack->r_ctl.rc_last_us_rtt >=
17297 (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17298 /* The rtt is huge, N * smallest, lets not fill */
17299 return (slot);
17300 }
17301 if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
17302 return (slot);
17303 /*
17304 * first lets calculate the b/w based on the last us-rtt
17305 * and the the smallest send window.
17306 */
17307 fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17308 if (rack->rc_fillcw_apply_discount) {
17309 uint32_t rate_set = 0;
17310
17311 fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17312 if (rate_set) {
17313 goto at_lt_bw;
17314 }
17315 }
17316 /* Take the rwnd if its smaller */
17317 if (fill_bw > rack->rc_tp->snd_wnd)
17318 fill_bw = rack->rc_tp->snd_wnd;
17319 /* Now lets make it into a b/w */
17320 fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17321 fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17322 /* Adjust to any cap */
17323 if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
17324 fill_bw = rack->r_ctl.fillcw_cap;
17325
17326 at_lt_bw:
17327 if (rack_bw_multipler > 0) {
17328 /*
17329 * We want to limit fill-cw to the some multiplier
17330 * of the max(lt_bw, gp_est). The normal default
17331 * is 0 for off, so a sysctl has enabled it.
17332 */
17333 uint64_t lt_bw, gp, rate;
17334
17335 gp = rack_get_gp_est(rack);
17336 lt_bw = rack_get_lt_bw(rack);
17337 if (lt_bw > gp)
17338 rate = lt_bw;
17339 else
17340 rate = gp;
17341 rate *= rack_bw_multipler;
17342 rate /= 100;
17343 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
17344 union tcp_log_stackspecific log;
17345 struct timeval tv;
17346
17347 memset(&log, 0, sizeof(log));
17348 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17349 log.u_bbr.flex1 = rack_bw_multipler;
17350 log.u_bbr.flex2 = len;
17351 log.u_bbr.cur_del_rate = gp;
17352 log.u_bbr.delRate = lt_bw;
17353 log.u_bbr.bw_inuse = rate;
17354 log.u_bbr.rttProp = fill_bw;
17355 log.u_bbr.flex8 = 44;
17356 tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
17357 BBR_LOG_CWND, 0,
17358 0, &log, false, NULL,
17359 __func__, __LINE__, &tv);
17360 }
17361 if (fill_bw > rate)
17362 fill_bw = rate;
17363 }
17364 /* We are below the min b/w */
17365 if (non_paced)
17366 *rate_wanted = fill_bw;
17367 if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17368 return (slot);
17369 rack->r_via_fill_cw = 1;
17370 if (rack->r_rack_hw_rate_caps &&
17371 (rack->r_ctl.crte != NULL)) {
17372 uint64_t high_rate;
17373
17374 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17375 if (fill_bw > high_rate) {
17376 /* We are capping bw at the highest rate table entry */
17377 if (*rate_wanted > high_rate) {
17378 /* The original rate was also capped */
17379 rack->r_via_fill_cw = 0;
17380 }
17381 rack_log_hdwr_pacing(rack,
17382 fill_bw, high_rate, __LINE__,
17383 0, 3);
17384 fill_bw = high_rate;
17385 if (capped)
17386 *capped = 1;
17387 }
17388 } else if ((rack->r_ctl.crte == NULL) &&
17389 (rack->rack_hdrw_pacing == 0) &&
17390 (rack->rack_hdw_pace_ena) &&
17391 rack->r_rack_hw_rate_caps &&
17392 (rack->rack_attempt_hdwr_pace == 0) &&
17393 (rack->rc_inp->inp_route.ro_nh != NULL) &&
17394 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17395 /*
17396 * Ok we may have a first attempt that is greater than our top rate
17397 * lets check.
17398 */
17399 uint64_t high_rate;
17400
17401 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17402 if (high_rate) {
17403 if (fill_bw > high_rate) {
17404 fill_bw = high_rate;
17405 if (capped)
17406 *capped = 1;
17407 }
17408 }
17409 }
17410 if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17411 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17412 fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17413 fill_bw = rack->r_ctl.bw_rate_cap;
17414 }
17415 /*
17416 * Ok fill_bw holds our mythical b/w to fill the cwnd
17417 * in an rtt (unless it was capped), what does that
17418 * time wise equate too?
17419 */
17420 lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17421 lentim /= fill_bw;
17422 *rate_wanted = fill_bw;
17423 if (non_paced || (lentim < slot)) {
17424 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17425 0, lentim, 12, __LINE__, NULL, 0);
17426 return ((int32_t)lentim);
17427 } else
17428 return (slot);
17429 }
17430
17431 static int32_t
rack_get_pacing_delay(struct tcp_rack * rack,struct tcpcb * tp,uint32_t len,struct rack_sendmap * rsm,uint32_t segsiz,int line)17432 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
17433 {
17434 uint64_t srtt;
17435 int32_t slot = 0;
17436 int can_start_hw_pacing = 1;
17437 int err;
17438 int pace_one;
17439
17440 if (rack_pace_one_seg ||
17441 (rack->r_ctl.rc_user_set_min_segs == 1))
17442 pace_one = 1;
17443 else
17444 pace_one = 0;
17445 if (rack->rc_always_pace == 0) {
17446 /*
17447 * We use the most optimistic possible cwnd/srtt for
17448 * sending calculations. This will make our
17449 * calculation anticipate getting more through
17450 * quicker then possible. But thats ok we don't want
17451 * the peer to have a gap in data sending.
17452 */
17453 uint64_t cwnd, tr_perms = 0;
17454 int32_t reduce;
17455
17456 old_method:
17457 /*
17458 * We keep no precise pacing with the old method
17459 * instead we use the pacer to mitigate bursts.
17460 */
17461 if (rack->r_ctl.rc_rack_min_rtt)
17462 srtt = rack->r_ctl.rc_rack_min_rtt;
17463 else
17464 srtt = max(tp->t_srtt, 1);
17465 if (rack->r_ctl.rc_rack_largest_cwnd)
17466 cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17467 else
17468 cwnd = rack->r_ctl.cwnd_to_use;
17469 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
17470 tr_perms = (cwnd * 1000) / srtt;
17471 if (tr_perms == 0) {
17472 tr_perms = ctf_fixed_maxseg(tp);
17473 }
17474 /*
17475 * Calculate how long this will take to drain, if
17476 * the calculation comes out to zero, thats ok we
17477 * will use send_a_lot to possibly spin around for
17478 * more increasing tot_len_this_send to the point
17479 * that its going to require a pace, or we hit the
17480 * cwnd. Which in that case we are just waiting for
17481 * a ACK.
17482 */
17483 slot = len / tr_perms;
17484 /* Now do we reduce the time so we don't run dry? */
17485 if (slot && rack_slot_reduction) {
17486 reduce = (slot / rack_slot_reduction);
17487 if (reduce < slot) {
17488 slot -= reduce;
17489 } else
17490 slot = 0;
17491 } else
17492 reduce = 0;
17493 slot *= HPTS_USEC_IN_MSEC;
17494 if (rack->rc_pace_to_cwnd) {
17495 uint64_t rate_wanted = 0;
17496
17497 slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17498 rack->rc_ack_can_sendout_data = 1;
17499 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17500 } else
17501 rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17502 /*******************************************************/
17503 /* RRS: We insert non-paced call to stats here for len */
17504 /*******************************************************/
17505 } else {
17506 uint64_t bw_est, res, lentim, rate_wanted;
17507 uint32_t segs, oh;
17508 int capped = 0;
17509 int prev_fill;
17510
17511 if ((rack->r_rr_config == 1) && rsm) {
17512 return (rack->r_ctl.rc_min_to);
17513 }
17514 if (rack->use_fixed_rate) {
17515 rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17516 } else if ((rack->r_ctl.init_rate == 0) &&
17517 (rack->r_ctl.gp_bw == 0)) {
17518 /* no way to yet do an estimate */
17519 bw_est = rate_wanted = 0;
17520 } else if (rack->dgp_on) {
17521 bw_est = rack_get_bw(rack);
17522 rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17523 } else {
17524 uint32_t gain, rate_set = 0;
17525
17526 rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17527 rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17528 if (rate_set == 0) {
17529 if (rate_wanted > rack->rc_tp->snd_wnd)
17530 rate_wanted = rack->rc_tp->snd_wnd;
17531 /* Now lets make it into a b/w */
17532 rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17533 rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17534 }
17535 bw_est = rate_wanted;
17536 rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17537 rack->r_ctl.cwnd_to_use,
17538 rate_wanted, bw_est,
17539 rack->r_ctl.rc_last_us_rtt,
17540 88, __LINE__, NULL, gain);
17541 }
17542 if ((bw_est == 0) || (rate_wanted == 0) ||
17543 ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17544 /*
17545 * No way yet to make a b/w estimate or
17546 * our raise is set incorrectly.
17547 */
17548 goto old_method;
17549 }
17550 rack_rate_cap_bw(rack, &rate_wanted, &capped);
17551 /* We need to account for all the overheads */
17552 segs = (len + segsiz - 1) / segsiz;
17553 /*
17554 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17555 * and how much data we put in each packet. Yes this
17556 * means we may be off if we are larger than 1500 bytes
17557 * or smaller. But this just makes us more conservative.
17558 */
17559
17560 oh = (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17561 if (rack->r_is_v6) {
17562 #ifdef INET6
17563 oh += sizeof(struct ip6_hdr);
17564 #endif
17565 } else {
17566 #ifdef INET
17567 oh += sizeof(struct ip);
17568 #endif
17569 }
17570 /* We add a fixed 14 for the ethernet header */
17571 oh += 14;
17572 segs *= oh;
17573 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17574 res = lentim / rate_wanted;
17575 slot = (uint32_t)res;
17576 if (rack_hw_rate_min &&
17577 (rate_wanted < rack_hw_rate_min)) {
17578 can_start_hw_pacing = 0;
17579 if (rack->r_ctl.crte) {
17580 /*
17581 * Ok we need to release it, we
17582 * have fallen too low.
17583 */
17584 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17585 rack->r_ctl.crte = NULL;
17586 rack->rack_attempt_hdwr_pace = 0;
17587 rack->rack_hdrw_pacing = 0;
17588 }
17589 }
17590 if (rack->r_ctl.crte &&
17591 (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17592 /*
17593 * We want more than the hardware can give us,
17594 * don't start any hw pacing.
17595 */
17596 can_start_hw_pacing = 0;
17597 if (rack->r_rack_hw_rate_caps == 0) {
17598 /*
17599 * Ok we need to release it, we
17600 * want more than the card can give us and
17601 * no rate cap is in place. Set it up so
17602 * when we want less we can retry.
17603 */
17604 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17605 rack->r_ctl.crte = NULL;
17606 rack->rack_attempt_hdwr_pace = 0;
17607 rack->rack_hdrw_pacing = 0;
17608 }
17609 }
17610 if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17611 /*
17612 * We lost our rate somehow, this can happen
17613 * if the interface changed underneath us.
17614 */
17615 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17616 rack->r_ctl.crte = NULL;
17617 /* Lets re-allow attempting to setup pacing */
17618 rack->rack_hdrw_pacing = 0;
17619 rack->rack_attempt_hdwr_pace = 0;
17620 rack_log_hdwr_pacing(rack,
17621 rate_wanted, bw_est, __LINE__,
17622 0, 6);
17623 }
17624 prev_fill = rack->r_via_fill_cw;
17625 if ((rack->rc_pace_to_cwnd) &&
17626 (capped == 0) &&
17627 (rack->dgp_on == 1) &&
17628 (rack->use_fixed_rate == 0) &&
17629 (rack->in_probe_rtt == 0) &&
17630 (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17631 /*
17632 * We want to pace at our rate *or* faster to
17633 * fill the cwnd to the max if its not full.
17634 */
17635 slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17636 /* Re-check to make sure we are not exceeding our max b/w */
17637 if ((rack->r_ctl.crte != NULL) &&
17638 (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17639 /*
17640 * We want more than the hardware can give us,
17641 * don't start any hw pacing.
17642 */
17643 can_start_hw_pacing = 0;
17644 if (rack->r_rack_hw_rate_caps == 0) {
17645 /*
17646 * Ok we need to release it, we
17647 * want more than the card can give us and
17648 * no rate cap is in place. Set it up so
17649 * when we want less we can retry.
17650 */
17651 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17652 rack->r_ctl.crte = NULL;
17653 rack->rack_attempt_hdwr_pace = 0;
17654 rack->rack_hdrw_pacing = 0;
17655 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17656 }
17657 }
17658 }
17659 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17660 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17661 if ((rack->rack_hdw_pace_ena) &&
17662 (can_start_hw_pacing > 0) &&
17663 (rack->rack_hdrw_pacing == 0) &&
17664 (rack->rack_attempt_hdwr_pace == 0)) {
17665 /*
17666 * Lets attempt to turn on hardware pacing
17667 * if we can.
17668 */
17669 rack->rack_attempt_hdwr_pace = 1;
17670 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17671 rack->rc_inp->inp_route.ro_nh->nh_ifp,
17672 rate_wanted,
17673 RS_PACING_GEQ,
17674 &err, &rack->r_ctl.crte_prev_rate);
17675 if (rack->r_ctl.crte) {
17676 rack->rack_hdrw_pacing = 1;
17677 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17678 pace_one, rack->r_ctl.crte,
17679 NULL, rack->r_ctl.pace_len_divisor);
17680 rack_log_hdwr_pacing(rack,
17681 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17682 err, 0);
17683 rack->r_ctl.last_hw_bw_req = rate_wanted;
17684 } else {
17685 counter_u64_add(rack_hw_pace_init_fail, 1);
17686 }
17687 } else if (rack->rack_hdrw_pacing &&
17688 (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17689 /* Do we need to adjust our rate? */
17690 const struct tcp_hwrate_limit_table *nrte;
17691
17692 if (rack->r_up_only &&
17693 (rate_wanted < rack->r_ctl.crte->rate)) {
17694 /**
17695 * We have four possible states here
17696 * having to do with the previous time
17697 * and this time.
17698 * previous | this-time
17699 * A) 0 | 0 -- fill_cw not in the picture
17700 * B) 1 | 0 -- we were doing a fill-cw but now are not
17701 * C) 1 | 1 -- all rates from fill_cw
17702 * D) 0 | 1 -- we were doing non-fill and now we are filling
17703 *
17704 * For case A, C and D we don't allow a drop. But for
17705 * case B where we now our on our steady rate we do
17706 * allow a drop.
17707 *
17708 */
17709 if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17710 goto done_w_hdwr;
17711 }
17712 if ((rate_wanted > rack->r_ctl.crte->rate) ||
17713 (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17714 if (rack_hw_rate_to_low &&
17715 (bw_est < rack_hw_rate_to_low)) {
17716 /*
17717 * The pacing rate is too low for hardware, but
17718 * do allow hardware pacing to be restarted.
17719 */
17720 rack_log_hdwr_pacing(rack,
17721 bw_est, rack->r_ctl.crte->rate, __LINE__,
17722 0, 5);
17723 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17724 rack->r_ctl.crte = NULL;
17725 rack->rack_attempt_hdwr_pace = 0;
17726 rack->rack_hdrw_pacing = 0;
17727 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17728 goto done_w_hdwr;
17729 }
17730 nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17731 rack->rc_tp,
17732 rack->rc_inp->inp_route.ro_nh->nh_ifp,
17733 rate_wanted,
17734 RS_PACING_GEQ,
17735 &err, &rack->r_ctl.crte_prev_rate);
17736 if (nrte == NULL) {
17737 /*
17738 * Lost the rate, lets drop hardware pacing
17739 * period.
17740 */
17741 rack->rack_hdrw_pacing = 0;
17742 rack->r_ctl.crte = NULL;
17743 rack_log_hdwr_pacing(rack,
17744 rate_wanted, 0, __LINE__,
17745 err, 1);
17746 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17747 counter_u64_add(rack_hw_pace_lost, 1);
17748 } else if (nrte != rack->r_ctl.crte) {
17749 rack->r_ctl.crte = nrte;
17750 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17751 segsiz, pace_one, rack->r_ctl.crte,
17752 NULL, rack->r_ctl.pace_len_divisor);
17753 rack_log_hdwr_pacing(rack,
17754 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17755 err, 2);
17756 rack->r_ctl.last_hw_bw_req = rate_wanted;
17757 }
17758 } else {
17759 /* We just need to adjust the segment size */
17760 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17761 rack_log_hdwr_pacing(rack,
17762 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17763 0, 4);
17764 rack->r_ctl.last_hw_bw_req = rate_wanted;
17765 }
17766 }
17767 }
17768 done_w_hdwr:
17769 if (rack_limit_time_with_srtt &&
17770 (rack->use_fixed_rate == 0) &&
17771 (rack->rack_hdrw_pacing == 0)) {
17772 /*
17773 * Sanity check, we do not allow the pacing delay
17774 * to be longer than the SRTT of the path. If it is
17775 * a slow path, then adding a packet should increase
17776 * the RTT and compensate for this i.e. the srtt will
17777 * be greater so the allowed pacing time will be greater.
17778 *
17779 * Note this restriction is not for where a peak rate
17780 * is set, we are doing fixed pacing or hardware pacing.
17781 */
17782 if (rack->rc_tp->t_srtt)
17783 srtt = rack->rc_tp->t_srtt;
17784 else
17785 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */
17786 if (srtt < (uint64_t)slot) {
17787 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17788 slot = srtt;
17789 }
17790 }
17791 /*******************************************************************/
17792 /* RRS: We insert paced call to stats here for len and rate_wanted */
17793 /*******************************************************************/
17794 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17795 }
17796 if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17797 /*
17798 * If this rate is seeing enobufs when it
17799 * goes to send then either the nic is out
17800 * of gas or we are mis-estimating the time
17801 * somehow and not letting the queue empty
17802 * completely. Lets add to the pacing time.
17803 */
17804 int hw_boost_delay;
17805
17806 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17807 if (hw_boost_delay > rack_enobuf_hw_max)
17808 hw_boost_delay = rack_enobuf_hw_max;
17809 else if (hw_boost_delay < rack_enobuf_hw_min)
17810 hw_boost_delay = rack_enobuf_hw_min;
17811 slot += hw_boost_delay;
17812 }
17813 return (slot);
17814 }
17815
17816 static void
rack_start_gp_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq startseq,uint32_t sb_offset)17817 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17818 tcp_seq startseq, uint32_t sb_offset)
17819 {
17820 struct rack_sendmap *my_rsm = NULL;
17821
17822 if (tp->t_state < TCPS_ESTABLISHED) {
17823 /*
17824 * We don't start any measurements if we are
17825 * not at least established.
17826 */
17827 return;
17828 }
17829 if (tp->t_state >= TCPS_FIN_WAIT_1) {
17830 /*
17831 * We will get no more data into the SB
17832 * this means we need to have the data available
17833 * before we start a measurement.
17834 */
17835
17836 if (sbavail(&tptosocket(tp)->so_snd) <
17837 max(rc_init_window(rack),
17838 (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17839 /* Nope not enough data */
17840 return;
17841 }
17842 }
17843 tp->t_flags |= TF_GPUTINPROG;
17844 rack->r_ctl.rc_gp_cumack_ts = 0;
17845 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17846 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17847 tp->gput_seq = startseq;
17848 rack->app_limited_needs_set = 0;
17849 if (rack->in_probe_rtt)
17850 rack->measure_saw_probe_rtt = 1;
17851 else if ((rack->measure_saw_probe_rtt) &&
17852 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17853 rack->measure_saw_probe_rtt = 0;
17854 if (rack->rc_gp_filled)
17855 tp->gput_ts = rack->r_ctl.last_cumack_advance;
17856 else {
17857 /* Special case initial measurement */
17858 struct timeval tv;
17859
17860 tp->gput_ts = tcp_get_usecs(&tv);
17861 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17862 }
17863 /*
17864 * We take a guess out into the future,
17865 * if we have no measurement and no
17866 * initial rate, we measure the first
17867 * initial-windows worth of data to
17868 * speed up getting some GP measurement and
17869 * thus start pacing.
17870 */
17871 if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17872 rack->app_limited_needs_set = 1;
17873 tp->gput_ack = startseq + max(rc_init_window(rack),
17874 (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17875 rack_log_pacing_delay_calc(rack,
17876 tp->gput_seq,
17877 tp->gput_ack,
17878 0,
17879 tp->gput_ts,
17880 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17881 9,
17882 __LINE__, NULL, 0);
17883 rack_tend_gp_marks(tp, rack);
17884 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17885 return;
17886 }
17887 if (sb_offset) {
17888 /*
17889 * We are out somewhere in the sb
17890 * can we use the already outstanding data?
17891 */
17892
17893 if (rack->r_ctl.rc_app_limited_cnt == 0) {
17894 /*
17895 * Yes first one is good and in this case
17896 * the tp->gput_ts is correctly set based on
17897 * the last ack that arrived (no need to
17898 * set things up when an ack comes in).
17899 */
17900 my_rsm = tqhash_min(rack->r_ctl.tqh);
17901 if ((my_rsm == NULL) ||
17902 (my_rsm->r_rtr_cnt != 1)) {
17903 /* retransmission? */
17904 goto use_latest;
17905 }
17906 } else {
17907 if (rack->r_ctl.rc_first_appl == NULL) {
17908 /*
17909 * If rc_first_appl is NULL
17910 * then the cnt should be 0.
17911 * This is probably an error, maybe
17912 * a KASSERT would be approprate.
17913 */
17914 goto use_latest;
17915 }
17916 /*
17917 * If we have a marker pointer to the last one that is
17918 * app limited we can use that, but we need to set
17919 * things up so that when it gets ack'ed we record
17920 * the ack time (if its not already acked).
17921 */
17922 rack->app_limited_needs_set = 1;
17923 /*
17924 * We want to get to the rsm that is either
17925 * next with space i.e. over 1 MSS or the one
17926 * after that (after the app-limited).
17927 */
17928 my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17929 if (my_rsm) {
17930 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17931 /* Have to use the next one */
17932 my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17933 else {
17934 /* Use after the first MSS of it is acked */
17935 tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17936 goto start_set;
17937 }
17938 }
17939 if ((my_rsm == NULL) ||
17940 (my_rsm->r_rtr_cnt != 1)) {
17941 /*
17942 * Either its a retransmit or
17943 * the last is the app-limited one.
17944 */
17945 goto use_latest;
17946 }
17947 }
17948 tp->gput_seq = my_rsm->r_start;
17949 start_set:
17950 if (my_rsm->r_flags & RACK_ACKED) {
17951 /*
17952 * This one has been acked use the arrival ack time
17953 */
17954 struct rack_sendmap *nrsm;
17955
17956 tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17957 rack->app_limited_needs_set = 0;
17958 /*
17959 * Ok in this path we need to use the r_end now
17960 * since this guy is the starting ack.
17961 */
17962 tp->gput_seq = my_rsm->r_end;
17963 /*
17964 * We also need to adjust up the sendtime
17965 * to the send of the next data after my_rsm.
17966 */
17967 nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17968 if (nrsm != NULL)
17969 my_rsm = nrsm;
17970 else {
17971 /*
17972 * The next as not been sent, thats the
17973 * case for using the latest.
17974 */
17975 goto use_latest;
17976 }
17977 }
17978 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17979 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
17980 rack->r_ctl.rc_gp_cumack_ts = 0;
17981 if ((rack->r_ctl.cleared_app_ack == 1) &&
17982 (SEQ_GEQ(rack->r_ctl.cleared_app_ack, tp->gput_seq))) {
17983 /*
17984 * We just cleared an application limited period
17985 * so the next seq out needs to skip the first
17986 * ack.
17987 */
17988 rack->app_limited_needs_set = 1;
17989 rack->r_ctl.cleared_app_ack = 0;
17990 }
17991 rack_log_pacing_delay_calc(rack,
17992 tp->gput_seq,
17993 tp->gput_ack,
17994 (uintptr_t)my_rsm,
17995 tp->gput_ts,
17996 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17997 9,
17998 __LINE__, my_rsm, 0);
17999 /* Now lets make sure all are marked as they should be */
18000 rack_tend_gp_marks(tp, rack);
18001 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18002 return;
18003 }
18004
18005 use_latest:
18006 /*
18007 * We don't know how long we may have been
18008 * idle or if this is the first-send. Lets
18009 * setup the flag so we will trim off
18010 * the first ack'd data so we get a true
18011 * measurement.
18012 */
18013 rack->app_limited_needs_set = 1;
18014 tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18015 rack->r_ctl.rc_gp_cumack_ts = 0;
18016 /* Find this guy so we can pull the send time */
18017 my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18018 if (my_rsm) {
18019 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18020 if (my_rsm->r_flags & RACK_ACKED) {
18021 /*
18022 * Unlikely since its probably what was
18023 * just transmitted (but I am paranoid).
18024 */
18025 tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18026 rack->app_limited_needs_set = 0;
18027 }
18028 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18029 /* This also is unlikely */
18030 tp->gput_seq = my_rsm->r_start;
18031 }
18032 } else {
18033 /*
18034 * TSNH unless we have some send-map limit,
18035 * and even at that it should not be hitting
18036 * that limit (we should have stopped sending).
18037 */
18038 struct timeval tv;
18039
18040 microuptime(&tv);
18041 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18042 }
18043 rack_tend_gp_marks(tp, rack);
18044 rack_log_pacing_delay_calc(rack,
18045 tp->gput_seq,
18046 tp->gput_ack,
18047 (uintptr_t)my_rsm,
18048 tp->gput_ts,
18049 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18050 9, __LINE__, NULL, 0);
18051 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18052 }
18053
18054 static inline uint32_t
rack_what_can_we_send(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cwnd_to_use,uint32_t avail,int32_t sb_offset)18055 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cwnd_to_use,
18056 uint32_t avail, int32_t sb_offset)
18057 {
18058 uint32_t len;
18059 uint32_t sendwin;
18060
18061 if (tp->snd_wnd > cwnd_to_use)
18062 sendwin = cwnd_to_use;
18063 else
18064 sendwin = tp->snd_wnd;
18065 if (ctf_outstanding(tp) >= tp->snd_wnd) {
18066 /* We never want to go over our peers rcv-window */
18067 len = 0;
18068 } else {
18069 uint32_t flight;
18070
18071 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18072 if (flight >= sendwin) {
18073 /*
18074 * We have in flight what we are allowed by cwnd (if
18075 * it was rwnd blocking it would have hit above out
18076 * >= tp->snd_wnd).
18077 */
18078 return (0);
18079 }
18080 len = sendwin - flight;
18081 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18082 /* We would send too much (beyond the rwnd) */
18083 len = tp->snd_wnd - ctf_outstanding(tp);
18084 }
18085 if ((len + sb_offset) > avail) {
18086 /*
18087 * We don't have that much in the SB, how much is
18088 * there?
18089 */
18090 len = avail - sb_offset;
18091 }
18092 }
18093 return (len);
18094 }
18095
18096 static void
rack_log_fsb(struct tcp_rack * rack,struct tcpcb * tp,struct socket * so,uint32_t flags,unsigned ipoptlen,int32_t orig_len,int32_t len,int error,int rsm_is_null,int optlen,int line,uint16_t mode)18097 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18098 unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18099 int rsm_is_null, int optlen, int line, uint16_t mode)
18100 {
18101 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18102 union tcp_log_stackspecific log;
18103 struct timeval tv;
18104
18105 memset(&log, 0, sizeof(log));
18106 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18107 log.u_bbr.flex1 = error;
18108 log.u_bbr.flex2 = flags;
18109 log.u_bbr.flex3 = rsm_is_null;
18110 log.u_bbr.flex4 = ipoptlen;
18111 log.u_bbr.flex5 = tp->rcv_numsacks;
18112 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18113 log.u_bbr.flex7 = optlen;
18114 log.u_bbr.flex8 = rack->r_fsb_inited;
18115 log.u_bbr.applimited = rack->r_fast_output;
18116 log.u_bbr.bw_inuse = rack_get_bw(rack);
18117 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18118 log.u_bbr.cwnd_gain = mode;
18119 log.u_bbr.pkts_out = orig_len;
18120 log.u_bbr.lt_epoch = len;
18121 log.u_bbr.delivered = line;
18122 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18123 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18124 tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18125 len, &log, false, NULL, __func__, __LINE__, &tv);
18126 }
18127 }
18128
18129
18130 static struct mbuf *
rack_fo_base_copym(struct mbuf * the_m,uint32_t the_off,int32_t * plen,struct rack_fast_send_blk * fsb,int32_t seglimit,int32_t segsize,int hw_tls)18131 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18132 struct rack_fast_send_blk *fsb,
18133 int32_t seglimit, int32_t segsize, int hw_tls)
18134 {
18135 #ifdef KERN_TLS
18136 struct ktls_session *tls, *ntls;
18137 #ifdef INVARIANTS
18138 struct mbuf *start;
18139 #endif
18140 #endif
18141 struct mbuf *m, *n, **np, *smb;
18142 struct mbuf *top;
18143 int32_t off, soff;
18144 int32_t len = *plen;
18145 int32_t fragsize;
18146 int32_t len_cp = 0;
18147 uint32_t mlen, frags;
18148
18149 soff = off = the_off;
18150 smb = m = the_m;
18151 np = ⊤
18152 top = NULL;
18153 #ifdef KERN_TLS
18154 if (hw_tls && (m->m_flags & M_EXTPG))
18155 tls = m->m_epg_tls;
18156 else
18157 tls = NULL;
18158 #ifdef INVARIANTS
18159 start = m;
18160 #endif
18161 #endif
18162 while (len > 0) {
18163 if (m == NULL) {
18164 *plen = len_cp;
18165 break;
18166 }
18167 #ifdef KERN_TLS
18168 if (hw_tls) {
18169 if (m->m_flags & M_EXTPG)
18170 ntls = m->m_epg_tls;
18171 else
18172 ntls = NULL;
18173
18174 /*
18175 * Avoid mixing TLS records with handshake
18176 * data or TLS records from different
18177 * sessions.
18178 */
18179 if (tls != ntls) {
18180 MPASS(m != start);
18181 *plen = len_cp;
18182 break;
18183 }
18184 }
18185 #endif
18186 mlen = min(len, m->m_len - off);
18187 if (seglimit) {
18188 /*
18189 * For M_EXTPG mbufs, add 3 segments
18190 * + 1 in case we are crossing page boundaries
18191 * + 2 in case the TLS hdr/trailer are used
18192 * It is cheaper to just add the segments
18193 * than it is to take the cache miss to look
18194 * at the mbuf ext_pgs state in detail.
18195 */
18196 if (m->m_flags & M_EXTPG) {
18197 fragsize = min(segsize, PAGE_SIZE);
18198 frags = 3;
18199 } else {
18200 fragsize = segsize;
18201 frags = 0;
18202 }
18203
18204 /* Break if we really can't fit anymore. */
18205 if ((frags + 1) >= seglimit) {
18206 *plen = len_cp;
18207 break;
18208 }
18209
18210 /*
18211 * Reduce size if you can't copy the whole
18212 * mbuf. If we can't copy the whole mbuf, also
18213 * adjust len so the loop will end after this
18214 * mbuf.
18215 */
18216 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18217 mlen = (seglimit - frags - 1) * fragsize;
18218 len = mlen;
18219 *plen = len_cp + len;
18220 }
18221 frags += howmany(mlen, fragsize);
18222 if (frags == 0)
18223 frags++;
18224 seglimit -= frags;
18225 KASSERT(seglimit > 0,
18226 ("%s: seglimit went too low", __func__));
18227 }
18228 n = m_get(M_NOWAIT, m->m_type);
18229 *np = n;
18230 if (n == NULL)
18231 goto nospace;
18232 n->m_len = mlen;
18233 soff += mlen;
18234 len_cp += n->m_len;
18235 if (m->m_flags & (M_EXT | M_EXTPG)) {
18236 n->m_data = m->m_data + off;
18237 mb_dupcl(n, m);
18238 } else {
18239 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18240 (u_int)n->m_len);
18241 }
18242 len -= n->m_len;
18243 off = 0;
18244 m = m->m_next;
18245 np = &n->m_next;
18246 if (len || (soff == smb->m_len)) {
18247 /*
18248 * We have more so we move forward or
18249 * we have consumed the entire mbuf and
18250 * len has fell to 0.
18251 */
18252 soff = 0;
18253 smb = m;
18254 }
18255
18256 }
18257 if (fsb != NULL) {
18258 fsb->m = smb;
18259 fsb->off = soff;
18260 if (smb) {
18261 /*
18262 * Save off the size of the mbuf. We do
18263 * this so that we can recognize when it
18264 * has been trimmed by sbcut() as acks
18265 * come in.
18266 */
18267 fsb->o_m_len = smb->m_len;
18268 fsb->o_t_len = M_TRAILINGROOM(smb);
18269 } else {
18270 /*
18271 * This is the case where the next mbuf went to NULL. This
18272 * means with this copy we have sent everything in the sb.
18273 * In theory we could clear the fast_output flag, but lets
18274 * not since its possible that we could get more added
18275 * and acks that call the extend function which would let
18276 * us send more.
18277 */
18278 fsb->o_m_len = 0;
18279 fsb->o_t_len = 0;
18280 }
18281 }
18282 return (top);
18283 nospace:
18284 if (top)
18285 m_freem(top);
18286 return (NULL);
18287
18288 }
18289
18290 /*
18291 * This is a copy of m_copym(), taking the TSO segment size/limit
18292 * constraints into account, and advancing the sndptr as it goes.
18293 */
18294 static struct mbuf *
rack_fo_m_copym(struct tcp_rack * rack,int32_t * plen,int32_t seglimit,int32_t segsize,struct mbuf ** s_mb,int * s_soff)18295 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18296 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18297 {
18298 struct mbuf *m, *n;
18299 int32_t soff;
18300
18301 m = rack->r_ctl.fsb.m;
18302 if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18303 /*
18304 * The trailing space changed, mbufs can grow
18305 * at the tail but they can't shrink from
18306 * it, KASSERT that. Adjust the orig_m_len to
18307 * compensate for this change.
18308 */
18309 KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18310 ("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18311 m,
18312 rack,
18313 (intmax_t)M_TRAILINGROOM(m),
18314 rack->r_ctl.fsb.o_t_len,
18315 rack->r_ctl.fsb.o_m_len,
18316 m->m_len));
18317 rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18318 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18319 }
18320 if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18321 /*
18322 * Mbuf shrank, trimmed off the top by an ack, our
18323 * offset changes.
18324 */
18325 KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18326 ("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18327 m, m->m_len,
18328 rack, rack->r_ctl.fsb.o_m_len,
18329 rack->r_ctl.fsb.off));
18330
18331 if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18332 rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18333 else
18334 rack->r_ctl.fsb.off = 0;
18335 rack->r_ctl.fsb.o_m_len = m->m_len;
18336 #ifdef INVARIANTS
18337 } else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18338 panic("rack:%p m:%p m_len grew outside of t_space compensation",
18339 rack, m);
18340 #endif
18341 }
18342 soff = rack->r_ctl.fsb.off;
18343 KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18344 KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18345 KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18346 __FUNCTION__,
18347 rack, *plen, m, m->m_len));
18348 /* Save off the right location before we copy and advance */
18349 *s_soff = soff;
18350 *s_mb = rack->r_ctl.fsb.m;
18351 n = rack_fo_base_copym(m, soff, plen,
18352 &rack->r_ctl.fsb,
18353 seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18354 return (n);
18355 }
18356
18357 /* Log the buffer level */
18358 static void
rack_log_queue_level(struct tcpcb * tp,struct tcp_rack * rack,int len,struct timeval * tv,uint32_t cts)18359 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18360 int len, struct timeval *tv,
18361 uint32_t cts)
18362 {
18363 uint32_t p_rate = 0, p_queue = 0, err = 0;
18364 union tcp_log_stackspecific log;
18365
18366 #ifdef RATELIMIT
18367 err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18368 err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18369 #endif
18370 memset(&log, 0, sizeof(log));
18371 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18372 log.u_bbr.flex1 = p_rate;
18373 log.u_bbr.flex2 = p_queue;
18374 log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18375 log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18376 log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18377 log.u_bbr.flex7 = 99;
18378 log.u_bbr.flex8 = 0;
18379 log.u_bbr.pkts_out = err;
18380 log.u_bbr.delRate = rack->r_ctl.crte->rate;
18381 log.u_bbr.timeStamp = cts;
18382 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18383 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18384 len, &log, false, NULL, __func__, __LINE__, tv);
18385
18386 }
18387
18388 static uint32_t
rack_check_queue_level(struct tcp_rack * rack,struct tcpcb * tp,struct timeval * tv,uint32_t cts,int len,uint32_t segsiz)18389 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18390 struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18391 {
18392 uint64_t lentime = 0;
18393 #ifdef RATELIMIT
18394 uint32_t p_rate = 0, p_queue = 0, err;
18395 union tcp_log_stackspecific log;
18396 uint64_t bw;
18397
18398 err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18399 /* Failed or queue is zero */
18400 if (err || (p_queue == 0)) {
18401 lentime = 0;
18402 goto out;
18403 }
18404 err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18405 if (err) {
18406 lentime = 0;
18407 goto out;
18408 }
18409 /*
18410 * If we reach here we have some bytes in
18411 * the queue. The number returned is a value
18412 * between 0 and 0xffff where ffff is full
18413 * and 0 is empty. So how best to make this into
18414 * something usable?
18415 *
18416 * The "safer" way is lets take the b/w gotten
18417 * from the query (which should be our b/w rate)
18418 * and pretend that a full send (our rc_pace_max_segs)
18419 * is outstanding. We factor it so its as if a full
18420 * number of our MSS segment is terms of full
18421 * ethernet segments are outstanding.
18422 */
18423 bw = p_rate / 8;
18424 if (bw) {
18425 lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18426 lentime *= ETHERNET_SEGMENT_SIZE;
18427 lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18428 lentime /= bw;
18429 } else {
18430 /* TSNH -- KASSERT? */
18431 lentime = 0;
18432 }
18433 out:
18434 if (tcp_bblogging_on(tp)) {
18435 memset(&log, 0, sizeof(log));
18436 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18437 log.u_bbr.flex1 = p_rate;
18438 log.u_bbr.flex2 = p_queue;
18439 log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18440 log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18441 log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18442 log.u_bbr.flex7 = 99;
18443 log.u_bbr.flex8 = 0;
18444 log.u_bbr.pkts_out = err;
18445 log.u_bbr.delRate = rack->r_ctl.crte->rate;
18446 log.u_bbr.cur_del_rate = lentime;
18447 log.u_bbr.timeStamp = cts;
18448 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18449 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18450 len, &log, false, NULL, __func__, __LINE__,tv);
18451 }
18452 #endif
18453 return ((uint32_t)lentime);
18454 }
18455
18456 static int
rack_fast_rsm_output(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint64_t ts_val,uint32_t cts,uint32_t ms_cts,struct timeval * tv,int len,uint8_t doing_tlp)18457 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18458 uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18459 {
18460 /*
18461 * Enter the fast retransmit path. We are given that a sched_pin is
18462 * in place (if accounting is compliled in) and the cycle count taken
18463 * at the entry is in the ts_val. The concept her is that the rsm
18464 * now holds the mbuf offsets and such so we can directly transmit
18465 * without a lot of overhead, the len field is already set for
18466 * us to prohibit us from sending too much (usually its 1MSS).
18467 */
18468 struct ip *ip = NULL;
18469 struct udphdr *udp = NULL;
18470 struct tcphdr *th = NULL;
18471 struct mbuf *m = NULL;
18472 struct inpcb *inp;
18473 uint8_t *cpto;
18474 struct tcp_log_buffer *lgb;
18475 #ifdef TCP_ACCOUNTING
18476 uint64_t crtsc;
18477 int cnt_thru = 1;
18478 #endif
18479 struct tcpopt to;
18480 u_char opt[TCP_MAXOLEN];
18481 uint32_t hdrlen, optlen;
18482 int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18483 uint16_t flags;
18484 uint32_t if_hw_tsomaxsegcount = 0, startseq;
18485 uint32_t if_hw_tsomaxsegsize;
18486 int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18487
18488 #ifdef INET6
18489 struct ip6_hdr *ip6 = NULL;
18490
18491 if (rack->r_is_v6) {
18492 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18493 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18494 } else
18495 #endif /* INET6 */
18496 {
18497 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18498 hdrlen = sizeof(struct tcpiphdr);
18499 }
18500 if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18501 goto failed;
18502 }
18503 if (doing_tlp) {
18504 /* Its a TLP add the flag, it may already be there but be sure */
18505 rsm->r_flags |= RACK_TLP;
18506 } else {
18507 /* If it was a TLP it is not not on this retransmit */
18508 rsm->r_flags &= ~RACK_TLP;
18509 }
18510 startseq = rsm->r_start;
18511 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18512 inp = rack->rc_inp;
18513 to.to_flags = 0;
18514 flags = tcp_outflags[tp->t_state];
18515 if (flags & (TH_SYN|TH_RST)) {
18516 goto failed;
18517 }
18518 if (rsm->r_flags & RACK_HAS_FIN) {
18519 /* We can't send a FIN here */
18520 goto failed;
18521 }
18522 if (flags & TH_FIN) {
18523 /* We never send a FIN */
18524 flags &= ~TH_FIN;
18525 }
18526 if (tp->t_flags & TF_RCVD_TSTMP) {
18527 to.to_tsval = ms_cts + tp->ts_offset;
18528 to.to_tsecr = tp->ts_recent;
18529 to.to_flags = TOF_TS;
18530 }
18531 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18532 /* TCP-MD5 (RFC2385). */
18533 if (tp->t_flags & TF_SIGNATURE)
18534 to.to_flags |= TOF_SIGNATURE;
18535 #endif
18536 optlen = tcp_addoptions(&to, opt);
18537 hdrlen += optlen;
18538 udp = rack->r_ctl.fsb.udp;
18539 if (udp)
18540 hdrlen += sizeof(struct udphdr);
18541 if (rack->r_ctl.rc_pace_max_segs)
18542 max_val = rack->r_ctl.rc_pace_max_segs;
18543 else if (rack->rc_user_set_max_segs)
18544 max_val = rack->rc_user_set_max_segs * segsiz;
18545 else
18546 max_val = len;
18547 if ((tp->t_flags & TF_TSO) &&
18548 V_tcp_do_tso &&
18549 (len > segsiz) &&
18550 (tp->t_port == 0))
18551 tso = 1;
18552 #ifdef INET6
18553 if (MHLEN < hdrlen + max_linkhdr)
18554 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18555 else
18556 #endif
18557 m = m_gethdr(M_NOWAIT, MT_DATA);
18558 if (m == NULL)
18559 goto failed;
18560 m->m_data += max_linkhdr;
18561 m->m_len = hdrlen;
18562 th = rack->r_ctl.fsb.th;
18563 /* Establish the len to send */
18564 if (len > max_val)
18565 len = max_val;
18566 if ((tso) && (len + optlen > segsiz)) {
18567 uint32_t if_hw_tsomax;
18568 int32_t max_len;
18569
18570 /* extract TSO information */
18571 if_hw_tsomax = tp->t_tsomax;
18572 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18573 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18574 /*
18575 * Check if we should limit by maximum payload
18576 * length:
18577 */
18578 if (if_hw_tsomax != 0) {
18579 /* compute maximum TSO length */
18580 max_len = (if_hw_tsomax - hdrlen -
18581 max_linkhdr);
18582 if (max_len <= 0) {
18583 goto failed;
18584 } else if (len > max_len) {
18585 len = max_len;
18586 }
18587 }
18588 if (len <= segsiz) {
18589 /*
18590 * In case there are too many small fragments don't
18591 * use TSO:
18592 */
18593 tso = 0;
18594 }
18595 } else {
18596 tso = 0;
18597 }
18598 if ((tso == 0) && (len > segsiz))
18599 len = segsiz;
18600 (void)tcp_get_usecs(tv);
18601 if ((len == 0) ||
18602 (len <= MHLEN - hdrlen - max_linkhdr)) {
18603 goto failed;
18604 }
18605 th->th_seq = htonl(rsm->r_start);
18606 th->th_ack = htonl(tp->rcv_nxt);
18607 /*
18608 * The PUSH bit should only be applied
18609 * if the full retransmission is made. If
18610 * we are sending less than this is the
18611 * left hand edge and should not have
18612 * the PUSH bit.
18613 */
18614 if ((rsm->r_flags & RACK_HAD_PUSH) &&
18615 (len == (rsm->r_end - rsm->r_start)))
18616 flags |= TH_PUSH;
18617 th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18618 if (th->th_win == 0) {
18619 tp->t_sndzerowin++;
18620 tp->t_flags |= TF_RXWIN0SENT;
18621 } else
18622 tp->t_flags &= ~TF_RXWIN0SENT;
18623 if (rsm->r_flags & RACK_TLP) {
18624 /*
18625 * TLP should not count in retran count, but
18626 * in its own bin
18627 */
18628 counter_u64_add(rack_tlp_retran, 1);
18629 counter_u64_add(rack_tlp_retran_bytes, len);
18630 } else {
18631 tp->t_sndrexmitpack++;
18632 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18633 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18634 }
18635 #ifdef STATS
18636 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18637 len);
18638 #endif
18639 if (rsm->m == NULL)
18640 goto failed;
18641 if (rsm->m &&
18642 ((rsm->orig_m_len != rsm->m->m_len) ||
18643 (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18644 /* Fix up the orig_m_len and possibly the mbuf offset */
18645 rack_adjust_orig_mlen(rsm);
18646 }
18647 m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18648 if (len <= segsiz) {
18649 /*
18650 * Must have ran out of mbufs for the copy
18651 * shorten it to no longer need tso. Lets
18652 * not put on sendalot since we are low on
18653 * mbufs.
18654 */
18655 tso = 0;
18656 }
18657 if ((m->m_next == NULL) || (len <= 0)){
18658 goto failed;
18659 }
18660 if (udp) {
18661 if (rack->r_is_v6)
18662 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18663 else
18664 ulen = hdrlen + len - sizeof(struct ip);
18665 udp->uh_ulen = htons(ulen);
18666 }
18667 m->m_pkthdr.rcvif = (struct ifnet *)0;
18668 if (TCPS_HAVERCVDSYN(tp->t_state) &&
18669 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18670 int ect = tcp_ecn_output_established(tp, &flags, len, true);
18671 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18672 (tp->t_flags2 & TF2_ECN_SND_ECE))
18673 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18674 #ifdef INET6
18675 if (rack->r_is_v6) {
18676 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18677 ip6->ip6_flow |= htonl(ect << 20);
18678 }
18679 else
18680 #endif
18681 {
18682 ip->ip_tos &= ~IPTOS_ECN_MASK;
18683 ip->ip_tos |= ect;
18684 }
18685 }
18686 if (rack->r_ctl.crte != NULL) {
18687 /* See if we can send via the hw queue */
18688 slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18689 /* If there is nothing in queue (no pacing time) we can send via the hw queue */
18690 if (slot == 0)
18691 ip_sendflag = 0;
18692 }
18693 tcp_set_flags(th, flags);
18694 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18695 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18696 if (to.to_flags & TOF_SIGNATURE) {
18697 /*
18698 * Calculate MD5 signature and put it into the place
18699 * determined before.
18700 * NOTE: since TCP options buffer doesn't point into
18701 * mbuf's data, calculate offset and use it.
18702 */
18703 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18704 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18705 /*
18706 * Do not send segment if the calculation of MD5
18707 * digest has failed.
18708 */
18709 goto failed;
18710 }
18711 }
18712 #endif
18713 #ifdef INET6
18714 if (rack->r_is_v6) {
18715 if (tp->t_port) {
18716 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18717 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18718 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18719 th->th_sum = htons(0);
18720 UDPSTAT_INC(udps_opackets);
18721 } else {
18722 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18723 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18724 th->th_sum = in6_cksum_pseudo(ip6,
18725 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18726 0);
18727 }
18728 }
18729 #endif
18730 #if defined(INET6) && defined(INET)
18731 else
18732 #endif
18733 #ifdef INET
18734 {
18735 if (tp->t_port) {
18736 m->m_pkthdr.csum_flags = CSUM_UDP;
18737 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18738 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18739 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18740 th->th_sum = htons(0);
18741 UDPSTAT_INC(udps_opackets);
18742 } else {
18743 m->m_pkthdr.csum_flags = CSUM_TCP;
18744 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18745 th->th_sum = in_pseudo(ip->ip_src.s_addr,
18746 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18747 IPPROTO_TCP + len + optlen));
18748 }
18749 /* IP version must be set here for ipv4/ipv6 checking later */
18750 KASSERT(ip->ip_v == IPVERSION,
18751 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18752 }
18753 #endif
18754 if (tso) {
18755 /*
18756 * Here we use segsiz since we have no added options besides
18757 * any standard timestamp options (no DSACKs or SACKS are sent
18758 * via either fast-path).
18759 */
18760 KASSERT(len > segsiz,
18761 ("%s: len <= tso_segsz tp:%p", __func__, tp));
18762 m->m_pkthdr.csum_flags |= CSUM_TSO;
18763 m->m_pkthdr.tso_segsz = segsiz;
18764 }
18765 #ifdef INET6
18766 if (rack->r_is_v6) {
18767 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18768 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18769 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18770 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18771 else
18772 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18773 }
18774 #endif
18775 #if defined(INET) && defined(INET6)
18776 else
18777 #endif
18778 #ifdef INET
18779 {
18780 ip->ip_len = htons(m->m_pkthdr.len);
18781 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18782 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18783 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18784 if (tp->t_port == 0 || len < V_tcp_minmss) {
18785 ip->ip_off |= htons(IP_DF);
18786 }
18787 } else {
18788 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18789 }
18790 }
18791 #endif
18792 if (doing_tlp == 0) {
18793 /* Set we retransmitted */
18794 rack->rc_gp_saw_rec = 1;
18795 } else {
18796 /* Its a TLP set ca or ss */
18797 if (tp->snd_cwnd > tp->snd_ssthresh) {
18798 /* Set we sent in CA */
18799 rack->rc_gp_saw_ca = 1;
18800 } else {
18801 /* Set we sent in SS */
18802 rack->rc_gp_saw_ss = 1;
18803 }
18804 }
18805 /* Time to copy in our header */
18806 cpto = mtod(m, uint8_t *);
18807 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18808 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18809 if (optlen) {
18810 bcopy(opt, th + 1, optlen);
18811 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18812 } else {
18813 th->th_off = sizeof(struct tcphdr) >> 2;
18814 }
18815 if (tcp_bblogging_on(rack->rc_tp)) {
18816 union tcp_log_stackspecific log;
18817
18818 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18819 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18820 counter_u64_add(rack_collapsed_win_rxt, 1);
18821 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18822 }
18823 memset(&log, 0, sizeof(log));
18824 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18825 if (rack->rack_no_prr)
18826 log.u_bbr.flex1 = 0;
18827 else
18828 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18829 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18830 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18831 log.u_bbr.flex4 = max_val;
18832 /* Save off the early/late values */
18833 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18834 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18835 log.u_bbr.bw_inuse = rack_get_bw(rack);
18836 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18837 if (doing_tlp == 0)
18838 log.u_bbr.flex8 = 1;
18839 else
18840 log.u_bbr.flex8 = 2;
18841 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18842 log.u_bbr.flex7 = 55;
18843 log.u_bbr.pkts_out = tp->t_maxseg;
18844 log.u_bbr.timeStamp = cts;
18845 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18846 if (rsm->r_rtr_cnt > 0) {
18847 /*
18848 * When we have a retransmit we want to log the
18849 * burst at send and flight at send from before.
18850 */
18851 log.u_bbr.flex5 = rsm->r_fas;
18852 log.u_bbr.bbr_substate = rsm->r_bas;
18853 } else {
18854 /*
18855 * This is currently unlikely until we do the
18856 * packet pair probes but I will add it for completeness.
18857 */
18858 log.u_bbr.flex5 = log.u_bbr.inflight;
18859 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18860 }
18861 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18862 log.u_bbr.delivered = 0;
18863 log.u_bbr.rttProp = (uintptr_t)rsm;
18864 log.u_bbr.delRate = rsm->r_flags;
18865 log.u_bbr.delRate <<= 31;
18866 log.u_bbr.delRate |= rack->r_must_retran;
18867 log.u_bbr.delRate <<= 1;
18868 log.u_bbr.delRate |= 1;
18869 log.u_bbr.pkt_epoch = __LINE__;
18870 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18871 len, &log, false, NULL, __func__, __LINE__, tv);
18872 } else
18873 lgb = NULL;
18874 if ((rack->r_ctl.crte != NULL) &&
18875 tcp_bblogging_on(tp)) {
18876 rack_log_queue_level(tp, rack, len, tv, cts);
18877 }
18878 #ifdef INET6
18879 if (rack->r_is_v6) {
18880 error = ip6_output(m, inp->in6p_outputopts,
18881 &inp->inp_route6,
18882 ip_sendflag, NULL, NULL, inp);
18883 }
18884 else
18885 #endif
18886 #ifdef INET
18887 {
18888 error = ip_output(m, NULL,
18889 &inp->inp_route,
18890 ip_sendflag, 0, inp);
18891 }
18892 #endif
18893 m = NULL;
18894 if (lgb) {
18895 lgb->tlb_errno = error;
18896 lgb = NULL;
18897 }
18898 /* Move snd_nxt to snd_max so we don't have false retransmissions */
18899 tp->snd_nxt = tp->snd_max;
18900 if (error) {
18901 goto failed;
18902 } else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18903 rack->rc_hw_nobuf = 0;
18904 rack->r_ctl.rc_agg_delayed = 0;
18905 rack->r_early = 0;
18906 rack->r_late = 0;
18907 rack->r_ctl.rc_agg_early = 0;
18908 }
18909 rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18910 rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18911 if (doing_tlp) {
18912 rack->rc_tlp_in_progress = 1;
18913 rack->r_ctl.rc_tlp_cnt_out++;
18914 }
18915 if (error == 0) {
18916 counter_u64_add(rack_total_bytes, len);
18917 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18918 if (doing_tlp) {
18919 rack->rc_last_sent_tlp_past_cumack = 0;
18920 rack->rc_last_sent_tlp_seq_valid = 1;
18921 rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18922 rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18923 }
18924 if (rack->r_ctl.rc_prr_sndcnt >= len)
18925 rack->r_ctl.rc_prr_sndcnt -= len;
18926 else
18927 rack->r_ctl.rc_prr_sndcnt = 0;
18928 }
18929 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18930 rack->forced_ack = 0; /* If we send something zap the FA flag */
18931 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18932 rack->r_ctl.retran_during_recovery += len;
18933 {
18934 int idx;
18935
18936 idx = (len / segsiz) + 3;
18937 if (idx >= TCP_MSS_ACCT_ATIMER)
18938 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18939 else
18940 counter_u64_add(rack_out_size[idx], 1);
18941 }
18942 if (tp->t_rtttime == 0) {
18943 tp->t_rtttime = ticks;
18944 tp->t_rtseq = startseq;
18945 KMOD_TCPSTAT_INC(tcps_segstimed);
18946 }
18947 counter_u64_add(rack_fto_rsm_send, 1);
18948 if (error && (error == ENOBUFS)) {
18949 if (rack->r_ctl.crte != NULL) {
18950 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18951 if (tcp_bblogging_on(rack->rc_tp))
18952 rack_log_queue_level(tp, rack, len, tv, cts);
18953 } else
18954 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18955 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18956 if (rack->rc_enobuf < 0x7f)
18957 rack->rc_enobuf++;
18958 if (slot < (10 * HPTS_USEC_IN_MSEC))
18959 slot = 10 * HPTS_USEC_IN_MSEC;
18960 if (rack->r_ctl.crte != NULL) {
18961 counter_u64_add(rack_saw_enobuf_hw, 1);
18962 tcp_rl_log_enobuf(rack->r_ctl.crte);
18963 }
18964 counter_u64_add(rack_saw_enobuf, 1);
18965 } else {
18966 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
18967 }
18968 rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
18969 #ifdef TCP_ACCOUNTING
18970 crtsc = get_cyclecount();
18971 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18972 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
18973 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18974 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
18975 }
18976 sched_unpin();
18977 #endif
18978 return (0);
18979 failed:
18980 if (m)
18981 m_free(m);
18982 return (-1);
18983 }
18984
18985 static void
rack_sndbuf_autoscale(struct tcp_rack * rack)18986 rack_sndbuf_autoscale(struct tcp_rack *rack)
18987 {
18988 /*
18989 * Automatic sizing of send socket buffer. Often the send buffer
18990 * size is not optimally adjusted to the actual network conditions
18991 * at hand (delay bandwidth product). Setting the buffer size too
18992 * small limits throughput on links with high bandwidth and high
18993 * delay (eg. trans-continental/oceanic links). Setting the
18994 * buffer size too big consumes too much real kernel memory,
18995 * especially with many connections on busy servers.
18996 *
18997 * The criteria to step up the send buffer one notch are:
18998 * 1. receive window of remote host is larger than send buffer
18999 * (with a fudge factor of 5/4th);
19000 * 2. send buffer is filled to 7/8th with data (so we actually
19001 * have data to make use of it);
19002 * 3. send buffer fill has not hit maximal automatic size;
19003 * 4. our send window (slow start and cogestion controlled) is
19004 * larger than sent but unacknowledged data in send buffer.
19005 *
19006 * Note that the rack version moves things much faster since
19007 * we want to avoid hitting cache lines in the rack_fast_output()
19008 * path so this is called much less often and thus moves
19009 * the SB forward by a percentage.
19010 */
19011 struct socket *so;
19012 struct tcpcb *tp;
19013 uint32_t sendwin, scaleup;
19014
19015 tp = rack->rc_tp;
19016 so = rack->rc_inp->inp_socket;
19017 sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19018 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19019 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19020 sbused(&so->so_snd) >=
19021 (so->so_snd.sb_hiwat / 8 * 7) &&
19022 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19023 sendwin >= (sbused(&so->so_snd) -
19024 (tp->snd_max - tp->snd_una))) {
19025 if (rack_autosndbuf_inc)
19026 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19027 else
19028 scaleup = V_tcp_autosndbuf_inc;
19029 if (scaleup < V_tcp_autosndbuf_inc)
19030 scaleup = V_tcp_autosndbuf_inc;
19031 scaleup += so->so_snd.sb_hiwat;
19032 if (scaleup > V_tcp_autosndbuf_max)
19033 scaleup = V_tcp_autosndbuf_max;
19034 if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19035 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19036 }
19037 }
19038 }
19039
19040 static int
rack_fast_output(struct tcpcb * tp,struct tcp_rack * rack,uint64_t ts_val,uint32_t cts,uint32_t ms_cts,struct timeval * tv,long tot_len,int * send_err)19041 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19042 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19043 {
19044 /*
19045 * Enter to do fast output. We are given that the sched_pin is
19046 * in place (if accounting is compiled in) and the cycle count taken
19047 * at entry is in place in ts_val. The idea here is that
19048 * we know how many more bytes needs to be sent (presumably either
19049 * during pacing or to fill the cwnd and that was greater than
19050 * the max-burst). We have how much to send and all the info we
19051 * need to just send.
19052 */
19053 #ifdef INET
19054 struct ip *ip = NULL;
19055 #endif
19056 struct udphdr *udp = NULL;
19057 struct tcphdr *th = NULL;
19058 struct mbuf *m, *s_mb;
19059 struct inpcb *inp;
19060 uint8_t *cpto;
19061 struct tcp_log_buffer *lgb;
19062 #ifdef TCP_ACCOUNTING
19063 uint64_t crtsc;
19064 #endif
19065 struct tcpopt to;
19066 u_char opt[TCP_MAXOLEN];
19067 uint32_t hdrlen, optlen;
19068 #ifdef TCP_ACCOUNTING
19069 int cnt_thru = 1;
19070 #endif
19071 int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19072 uint16_t flags;
19073 uint32_t s_soff;
19074 uint32_t if_hw_tsomaxsegcount = 0, startseq;
19075 uint32_t if_hw_tsomaxsegsize;
19076 uint32_t add_flag = RACK_SENT_FP;
19077 #ifdef INET6
19078 struct ip6_hdr *ip6 = NULL;
19079
19080 if (rack->r_is_v6) {
19081 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19082 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19083 } else
19084 #endif /* INET6 */
19085 {
19086 #ifdef INET
19087 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19088 hdrlen = sizeof(struct tcpiphdr);
19089 #endif
19090 }
19091 if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19092 m = NULL;
19093 goto failed;
19094 }
19095 rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19096 startseq = tp->snd_max;
19097 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19098 inp = rack->rc_inp;
19099 len = rack->r_ctl.fsb.left_to_send;
19100 to.to_flags = 0;
19101 flags = rack->r_ctl.fsb.tcp_flags;
19102 if (tp->t_flags & TF_RCVD_TSTMP) {
19103 to.to_tsval = ms_cts + tp->ts_offset;
19104 to.to_tsecr = tp->ts_recent;
19105 to.to_flags = TOF_TS;
19106 }
19107 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19108 /* TCP-MD5 (RFC2385). */
19109 if (tp->t_flags & TF_SIGNATURE)
19110 to.to_flags |= TOF_SIGNATURE;
19111 #endif
19112 optlen = tcp_addoptions(&to, opt);
19113 hdrlen += optlen;
19114 udp = rack->r_ctl.fsb.udp;
19115 if (udp)
19116 hdrlen += sizeof(struct udphdr);
19117 if (rack->r_ctl.rc_pace_max_segs)
19118 max_val = rack->r_ctl.rc_pace_max_segs;
19119 else if (rack->rc_user_set_max_segs)
19120 max_val = rack->rc_user_set_max_segs * segsiz;
19121 else
19122 max_val = len;
19123 if ((tp->t_flags & TF_TSO) &&
19124 V_tcp_do_tso &&
19125 (len > segsiz) &&
19126 (tp->t_port == 0))
19127 tso = 1;
19128 again:
19129 #ifdef INET6
19130 if (MHLEN < hdrlen + max_linkhdr)
19131 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19132 else
19133 #endif
19134 m = m_gethdr(M_NOWAIT, MT_DATA);
19135 if (m == NULL)
19136 goto failed;
19137 m->m_data += max_linkhdr;
19138 m->m_len = hdrlen;
19139 th = rack->r_ctl.fsb.th;
19140 /* Establish the len to send */
19141 if (len > max_val)
19142 len = max_val;
19143 if ((tso) && (len + optlen > segsiz)) {
19144 uint32_t if_hw_tsomax;
19145 int32_t max_len;
19146
19147 /* extract TSO information */
19148 if_hw_tsomax = tp->t_tsomax;
19149 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19150 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19151 /*
19152 * Check if we should limit by maximum payload
19153 * length:
19154 */
19155 if (if_hw_tsomax != 0) {
19156 /* compute maximum TSO length */
19157 max_len = (if_hw_tsomax - hdrlen -
19158 max_linkhdr);
19159 if (max_len <= 0) {
19160 goto failed;
19161 } else if (len > max_len) {
19162 len = max_len;
19163 }
19164 }
19165 if (len <= segsiz) {
19166 /*
19167 * In case there are too many small fragments don't
19168 * use TSO:
19169 */
19170 tso = 0;
19171 }
19172 } else {
19173 tso = 0;
19174 }
19175 if ((tso == 0) && (len > segsiz))
19176 len = segsiz;
19177 (void)tcp_get_usecs(tv);
19178 if ((len == 0) ||
19179 (len <= MHLEN - hdrlen - max_linkhdr)) {
19180 goto failed;
19181 }
19182 sb_offset = tp->snd_max - tp->snd_una;
19183 th->th_seq = htonl(tp->snd_max);
19184 th->th_ack = htonl(tp->rcv_nxt);
19185 th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19186 if (th->th_win == 0) {
19187 tp->t_sndzerowin++;
19188 tp->t_flags |= TF_RXWIN0SENT;
19189 } else
19190 tp->t_flags &= ~TF_RXWIN0SENT;
19191 tp->snd_up = tp->snd_una; /* drag it along, its deprecated */
19192 KMOD_TCPSTAT_INC(tcps_sndpack);
19193 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19194 #ifdef STATS
19195 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19196 len);
19197 #endif
19198 if (rack->r_ctl.fsb.m == NULL)
19199 goto failed;
19200
19201 /* s_mb and s_soff are saved for rack_log_output */
19202 m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19203 &s_mb, &s_soff);
19204 if (len <= segsiz) {
19205 /*
19206 * Must have ran out of mbufs for the copy
19207 * shorten it to no longer need tso. Lets
19208 * not put on sendalot since we are low on
19209 * mbufs.
19210 */
19211 tso = 0;
19212 }
19213 if (rack->r_ctl.fsb.rfo_apply_push &&
19214 (len == rack->r_ctl.fsb.left_to_send)) {
19215 tcp_set_flags(th, flags | TH_PUSH);
19216 add_flag |= RACK_HAD_PUSH;
19217 }
19218 if ((m->m_next == NULL) || (len <= 0)){
19219 goto failed;
19220 }
19221 if (udp) {
19222 if (rack->r_is_v6)
19223 ulen = hdrlen + len - sizeof(struct ip6_hdr);
19224 else
19225 ulen = hdrlen + len - sizeof(struct ip);
19226 udp->uh_ulen = htons(ulen);
19227 }
19228 m->m_pkthdr.rcvif = (struct ifnet *)0;
19229 if (TCPS_HAVERCVDSYN(tp->t_state) &&
19230 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19231 int ect = tcp_ecn_output_established(tp, &flags, len, false);
19232 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19233 (tp->t_flags2 & TF2_ECN_SND_ECE))
19234 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19235 #ifdef INET6
19236 if (rack->r_is_v6) {
19237 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19238 ip6->ip6_flow |= htonl(ect << 20);
19239 }
19240 else
19241 #endif
19242 {
19243 #ifdef INET
19244 ip->ip_tos &= ~IPTOS_ECN_MASK;
19245 ip->ip_tos |= ect;
19246 #endif
19247 }
19248 }
19249 tcp_set_flags(th, flags);
19250 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
19251 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19252 if (to.to_flags & TOF_SIGNATURE) {
19253 /*
19254 * Calculate MD5 signature and put it into the place
19255 * determined before.
19256 * NOTE: since TCP options buffer doesn't point into
19257 * mbuf's data, calculate offset and use it.
19258 */
19259 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19260 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19261 /*
19262 * Do not send segment if the calculation of MD5
19263 * digest has failed.
19264 */
19265 goto failed;
19266 }
19267 }
19268 #endif
19269 #ifdef INET6
19270 if (rack->r_is_v6) {
19271 if (tp->t_port) {
19272 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19273 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19274 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19275 th->th_sum = htons(0);
19276 UDPSTAT_INC(udps_opackets);
19277 } else {
19278 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19279 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19280 th->th_sum = in6_cksum_pseudo(ip6,
19281 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19282 0);
19283 }
19284 }
19285 #endif
19286 #if defined(INET6) && defined(INET)
19287 else
19288 #endif
19289 #ifdef INET
19290 {
19291 if (tp->t_port) {
19292 m->m_pkthdr.csum_flags = CSUM_UDP;
19293 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19294 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19295 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19296 th->th_sum = htons(0);
19297 UDPSTAT_INC(udps_opackets);
19298 } else {
19299 m->m_pkthdr.csum_flags = CSUM_TCP;
19300 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19301 th->th_sum = in_pseudo(ip->ip_src.s_addr,
19302 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19303 IPPROTO_TCP + len + optlen));
19304 }
19305 /* IP version must be set here for ipv4/ipv6 checking later */
19306 KASSERT(ip->ip_v == IPVERSION,
19307 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
19308 }
19309 #endif
19310 if (tso) {
19311 /*
19312 * Here we use segsiz since we have no added options besides
19313 * any standard timestamp options (no DSACKs or SACKS are sent
19314 * via either fast-path).
19315 */
19316 KASSERT(len > segsiz,
19317 ("%s: len <= tso_segsz tp:%p", __func__, tp));
19318 m->m_pkthdr.csum_flags |= CSUM_TSO;
19319 m->m_pkthdr.tso_segsz = segsiz;
19320 }
19321 #ifdef INET6
19322 if (rack->r_is_v6) {
19323 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19324 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19325 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19326 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19327 else
19328 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19329 }
19330 #endif
19331 #if defined(INET) && defined(INET6)
19332 else
19333 #endif
19334 #ifdef INET
19335 {
19336 ip->ip_len = htons(m->m_pkthdr.len);
19337 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19338 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19339 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19340 if (tp->t_port == 0 || len < V_tcp_minmss) {
19341 ip->ip_off |= htons(IP_DF);
19342 }
19343 } else {
19344 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19345 }
19346 }
19347 #endif
19348 if (tp->snd_cwnd > tp->snd_ssthresh) {
19349 /* Set we sent in CA */
19350 rack->rc_gp_saw_ca = 1;
19351 } else {
19352 /* Set we sent in SS */
19353 rack->rc_gp_saw_ss = 1;
19354 }
19355 /* Time to copy in our header */
19356 cpto = mtod(m, uint8_t *);
19357 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19358 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19359 if (optlen) {
19360 bcopy(opt, th + 1, optlen);
19361 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19362 } else {
19363 th->th_off = sizeof(struct tcphdr) >> 2;
19364 }
19365 if ((rack->r_ctl.crte != NULL) &&
19366 tcp_bblogging_on(tp)) {
19367 rack_log_queue_level(tp, rack, len, tv, cts);
19368 }
19369 if (tcp_bblogging_on(rack->rc_tp)) {
19370 union tcp_log_stackspecific log;
19371
19372 memset(&log, 0, sizeof(log));
19373 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19374 if (rack->rack_no_prr)
19375 log.u_bbr.flex1 = 0;
19376 else
19377 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19378 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19379 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19380 log.u_bbr.flex4 = max_val;
19381 /* Save off the early/late values */
19382 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19383 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19384 log.u_bbr.bw_inuse = rack_get_bw(rack);
19385 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19386 log.u_bbr.flex8 = 0;
19387 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19388 log.u_bbr.flex7 = 44;
19389 log.u_bbr.pkts_out = tp->t_maxseg;
19390 log.u_bbr.timeStamp = cts;
19391 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19392 log.u_bbr.flex5 = log.u_bbr.inflight;
19393 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19394 log.u_bbr.delivered = 0;
19395 log.u_bbr.rttProp = 0;
19396 log.u_bbr.delRate = rack->r_must_retran;
19397 log.u_bbr.delRate <<= 1;
19398 log.u_bbr.pkt_epoch = __LINE__;
19399 /* For fast output no retrans so just inflight and how many mss we send */
19400 log.u_bbr.flex5 = log.u_bbr.inflight;
19401 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19402 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19403 len, &log, false, NULL, __func__, __LINE__, tv);
19404 } else
19405 lgb = NULL;
19406 #ifdef INET6
19407 if (rack->r_is_v6) {
19408 error = ip6_output(m, inp->in6p_outputopts,
19409 &inp->inp_route6,
19410 0, NULL, NULL, inp);
19411 }
19412 #endif
19413 #if defined(INET) && defined(INET6)
19414 else
19415 #endif
19416 #ifdef INET
19417 {
19418 error = ip_output(m, NULL,
19419 &inp->inp_route,
19420 0, 0, inp);
19421 }
19422 #endif
19423 if (lgb) {
19424 lgb->tlb_errno = error;
19425 lgb = NULL;
19426 }
19427 if (error) {
19428 *send_err = error;
19429 m = NULL;
19430 goto failed;
19431 } else if (rack->rc_hw_nobuf) {
19432 rack->rc_hw_nobuf = 0;
19433 rack->r_ctl.rc_agg_delayed = 0;
19434 rack->r_early = 0;
19435 rack->r_late = 0;
19436 rack->r_ctl.rc_agg_early = 0;
19437 }
19438 if ((error == 0) && (rack->lt_bw_up == 0)) {
19439 /* Unlikely */
19440 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19441 rack->r_ctl.lt_seq = tp->snd_una;
19442 rack->lt_bw_up = 1;
19443 } else if ((error == 0) &&
19444 (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
19445 /*
19446 * Need to record what we have since we are
19447 * approaching seq wrap.
19448 */
19449 struct timeval tv;
19450 uint64_t tmark;
19451
19452 rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
19453 rack->r_ctl.lt_seq = tp->snd_una;
19454 tmark = tcp_get_u64_usecs(&tv);
19455 if (tmark > rack->r_ctl.lt_timemark) {
19456 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
19457 rack->r_ctl.lt_timemark = tmark;
19458 }
19459 }
19460 rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19461 NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19462 if (tp->snd_una == tp->snd_max) {
19463 rack->r_ctl.rc_tlp_rxt_last_time = cts;
19464 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19465 tp->t_acktime = ticks;
19466 }
19467 counter_u64_add(rack_total_bytes, len);
19468 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19469
19470 rack->forced_ack = 0; /* If we send something zap the FA flag */
19471 tot_len += len;
19472 if ((tp->t_flags & TF_GPUTINPROG) == 0)
19473 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19474 tp->snd_max += len;
19475 tp->snd_nxt = tp->snd_max;
19476 if (rack->rc_new_rnd_needed) {
19477 rack_new_round_starts(tp, rack, tp->snd_max);
19478 }
19479 {
19480 int idx;
19481
19482 idx = (len / segsiz) + 3;
19483 if (idx >= TCP_MSS_ACCT_ATIMER)
19484 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19485 else
19486 counter_u64_add(rack_out_size[idx], 1);
19487 }
19488 if (len <= rack->r_ctl.fsb.left_to_send)
19489 rack->r_ctl.fsb.left_to_send -= len;
19490 else
19491 rack->r_ctl.fsb.left_to_send = 0;
19492 if (rack->r_ctl.fsb.left_to_send < segsiz) {
19493 rack->r_fast_output = 0;
19494 rack->r_ctl.fsb.left_to_send = 0;
19495 /* At the end of fast_output scale up the sb */
19496 SOCK_SENDBUF_LOCK(rack->rc_inp->inp_socket);
19497 rack_sndbuf_autoscale(rack);
19498 SOCK_SENDBUF_UNLOCK(rack->rc_inp->inp_socket);
19499 }
19500 if (tp->t_rtttime == 0) {
19501 tp->t_rtttime = ticks;
19502 tp->t_rtseq = startseq;
19503 KMOD_TCPSTAT_INC(tcps_segstimed);
19504 }
19505 if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19506 (max_val > len) &&
19507 (tso == 0)) {
19508 max_val -= len;
19509 len = segsiz;
19510 th = rack->r_ctl.fsb.th;
19511 #ifdef TCP_ACCOUNTING
19512 cnt_thru++;
19513 #endif
19514 goto again;
19515 }
19516 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19517 counter_u64_add(rack_fto_send, 1);
19518 slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz, __LINE__);
19519 rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19520 #ifdef TCP_ACCOUNTING
19521 crtsc = get_cyclecount();
19522 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19523 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19524 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19525 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19526 }
19527 sched_unpin();
19528 #endif
19529 return (0);
19530 failed:
19531 if (m)
19532 m_free(m);
19533 rack->r_fast_output = 0;
19534 return (-1);
19535 }
19536
19537 static inline void
rack_setup_fast_output(struct tcpcb * tp,struct tcp_rack * rack,struct sockbuf * sb,int len,int orig_len,int segsiz,uint32_t pace_max_seg,bool hw_tls,uint16_t flags)19538 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19539 struct sockbuf *sb,
19540 int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19541 bool hw_tls,
19542 uint16_t flags)
19543 {
19544 rack->r_fast_output = 1;
19545 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19546 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19547 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19548 rack->r_ctl.fsb.tcp_flags = flags;
19549 rack->r_ctl.fsb.left_to_send = orig_len - len;
19550 if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19551 /* Less than a full sized pace, lets not */
19552 rack->r_fast_output = 0;
19553 return;
19554 } else {
19555 /* Round down to the nearest pace_max_seg */
19556 rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19557 }
19558 if (hw_tls)
19559 rack->r_ctl.fsb.hw_tls = 1;
19560 else
19561 rack->r_ctl.fsb.hw_tls = 0;
19562 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19563 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19564 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19565 (tp->snd_max - tp->snd_una)));
19566 if (rack->r_ctl.fsb.left_to_send < segsiz)
19567 rack->r_fast_output = 0;
19568 else {
19569 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19570 rack->r_ctl.fsb.rfo_apply_push = 1;
19571 else
19572 rack->r_ctl.fsb.rfo_apply_push = 0;
19573 }
19574 }
19575
19576 static uint32_t
rack_get_hpts_pacing_min_for_bw(struct tcp_rack * rack,int32_t segsiz)19577 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19578 {
19579 uint64_t min_time;
19580 uint32_t maxlen;
19581
19582 min_time = (uint64_t)get_hpts_min_sleep_time();
19583 maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19584 maxlen = roundup(maxlen, segsiz);
19585 return (maxlen);
19586 }
19587
19588 static struct rack_sendmap *
rack_check_collapsed(struct tcp_rack * rack,uint32_t cts)19589 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19590 {
19591 struct rack_sendmap *rsm = NULL;
19592 int thresh;
19593
19594 restart:
19595 rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19596 if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19597 /* Nothing, strange turn off validity */
19598 rack->r_collapse_point_valid = 0;
19599 return (NULL);
19600 }
19601 /* Can we send it yet? */
19602 if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19603 /*
19604 * Receiver window has not grown enough for
19605 * the segment to be put on the wire.
19606 */
19607 return (NULL);
19608 }
19609 if (rsm->r_flags & RACK_ACKED) {
19610 /*
19611 * It has been sacked, lets move to the
19612 * next one if possible.
19613 */
19614 rack->r_ctl.last_collapse_point = rsm->r_end;
19615 /* Are we done? */
19616 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19617 rack->r_ctl.high_collapse_point)) {
19618 rack->r_collapse_point_valid = 0;
19619 return (NULL);
19620 }
19621 goto restart;
19622 }
19623 /* Now has it been long enough ? */
19624 thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
19625 if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19626 rack_log_collapse(rack, rsm->r_start,
19627 (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19628 thresh, __LINE__, 6, rsm->r_flags, rsm);
19629 return (rsm);
19630 }
19631 /* Not enough time */
19632 rack_log_collapse(rack, rsm->r_start,
19633 (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19634 thresh, __LINE__, 7, rsm->r_flags, rsm);
19635 return (NULL);
19636 }
19637
19638 static inline void
rack_validate_sizes(struct tcp_rack * rack,int32_t * len,int32_t segsiz,uint32_t pace_max_seg)19639 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19640 {
19641 if ((rack->full_size_rxt == 0) &&
19642 (rack->shape_rxt_to_pacing_min == 0) &&
19643 (*len >= segsiz)) {
19644 *len = segsiz;
19645 } else if (rack->shape_rxt_to_pacing_min &&
19646 rack->gp_ready) {
19647 /* We use pacing min as shaping len req */
19648 uint32_t maxlen;
19649
19650 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19651 if (*len > maxlen)
19652 *len = maxlen;
19653 } else {
19654 /*
19655 * The else is full_size_rxt is on so send it all
19656 * note we do need to check this for exceeding
19657 * our max segment size due to the fact that
19658 * we do sometimes merge chunks together i.e.
19659 * we cannot just assume that we will never have
19660 * a chunk greater than pace_max_seg
19661 */
19662 if (*len > pace_max_seg)
19663 *len = pace_max_seg;
19664 }
19665 }
19666
19667 static int
rack_output(struct tcpcb * tp)19668 rack_output(struct tcpcb *tp)
19669 {
19670 struct socket *so;
19671 uint32_t recwin;
19672 uint32_t sb_offset, s_moff = 0;
19673 int32_t len, error = 0;
19674 uint16_t flags;
19675 struct mbuf *m, *s_mb = NULL;
19676 struct mbuf *mb;
19677 uint32_t if_hw_tsomaxsegcount = 0;
19678 uint32_t if_hw_tsomaxsegsize;
19679 int32_t segsiz, minseg;
19680 long tot_len_this_send = 0;
19681 #ifdef INET
19682 struct ip *ip = NULL;
19683 #endif
19684 struct udphdr *udp = NULL;
19685 struct tcp_rack *rack;
19686 struct tcphdr *th;
19687 uint8_t pass = 0;
19688 uint8_t mark = 0;
19689 uint8_t check_done = 0;
19690 uint8_t wanted_cookie = 0;
19691 u_char opt[TCP_MAXOLEN];
19692 unsigned ipoptlen, optlen, hdrlen, ulen=0;
19693 uint32_t rack_seq;
19694
19695 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19696 unsigned ipsec_optlen = 0;
19697
19698 #endif
19699 int32_t idle, sendalot;
19700 uint32_t tot_idle;
19701 int32_t sub_from_prr = 0;
19702 volatile int32_t sack_rxmit;
19703 struct rack_sendmap *rsm = NULL;
19704 int32_t tso, mtu;
19705 struct tcpopt to;
19706 int32_t slot = 0;
19707 int32_t sup_rack = 0;
19708 uint32_t cts, ms_cts, delayed, early;
19709 uint32_t add_flag = RACK_SENT_SP;
19710 /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19711 uint8_t doing_tlp = 0;
19712 uint32_t cwnd_to_use, pace_max_seg;
19713 int32_t do_a_prefetch = 0;
19714 int32_t prefetch_rsm = 0;
19715 int32_t orig_len = 0;
19716 struct timeval tv;
19717 int32_t prefetch_so_done = 0;
19718 struct tcp_log_buffer *lgb;
19719 struct inpcb *inp = tptoinpcb(tp);
19720 struct sockbuf *sb;
19721 uint64_t ts_val = 0;
19722 #ifdef TCP_ACCOUNTING
19723 uint64_t crtsc;
19724 #endif
19725 #ifdef INET6
19726 struct ip6_hdr *ip6 = NULL;
19727 int32_t isipv6;
19728 #endif
19729 bool hpts_calling, hw_tls = false;
19730
19731 NET_EPOCH_ASSERT();
19732 INP_WLOCK_ASSERT(inp);
19733
19734 /* setup and take the cache hits here */
19735 rack = (struct tcp_rack *)tp->t_fb_ptr;
19736 #ifdef TCP_ACCOUNTING
19737 sched_pin();
19738 ts_val = get_cyclecount();
19739 #endif
19740 hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19741 tp->t_flags2 &= ~TF2_HPTS_CALLS;
19742 #ifdef TCP_OFFLOAD
19743 if (tp->t_flags & TF_TOE) {
19744 #ifdef TCP_ACCOUNTING
19745 sched_unpin();
19746 #endif
19747 return (tcp_offload_output(tp));
19748 }
19749 #endif
19750 if (rack->rack_deferred_inited == 0) {
19751 /*
19752 * If we are the connecting socket we will
19753 * hit rack_init() when no sequence numbers
19754 * are setup. This makes it so we must defer
19755 * some initialization. Call that now.
19756 */
19757 rack_deferred_init(tp, rack);
19758 }
19759 /*
19760 * For TFO connections in SYN_RECEIVED, only allow the initial
19761 * SYN|ACK and those sent by the retransmit timer.
19762 */
19763 if ((tp->t_flags & TF_FASTOPEN) &&
19764 (tp->t_state == TCPS_SYN_RECEIVED) &&
19765 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
19766 (rack->r_ctl.rc_resend == NULL)) { /* not a retransmit */
19767 #ifdef TCP_ACCOUNTING
19768 sched_unpin();
19769 #endif
19770 return (0);
19771 }
19772 #ifdef INET6
19773 if (rack->r_state) {
19774 /* Use the cache line loaded if possible */
19775 isipv6 = rack->r_is_v6;
19776 } else {
19777 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19778 }
19779 #endif
19780 early = 0;
19781 cts = tcp_get_usecs(&tv);
19782 ms_cts = tcp_tv_to_mssectick(&tv);
19783 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19784 tcp_in_hpts(rack->rc_tp)) {
19785 /*
19786 * We are on the hpts for some timer but not hptsi output.
19787 * Remove from the hpts unconditionally.
19788 */
19789 rack_timer_cancel(tp, rack, cts, __LINE__);
19790 }
19791 /* Are we pacing and late? */
19792 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19793 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19794 /* We are delayed */
19795 delayed = cts - rack->r_ctl.rc_last_output_to;
19796 } else {
19797 delayed = 0;
19798 }
19799 /* Do the timers, which may override the pacer */
19800 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19801 int retval;
19802
19803 retval = rack_process_timers(tp, rack, cts, hpts_calling,
19804 &doing_tlp);
19805 if (retval != 0) {
19806 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19807 #ifdef TCP_ACCOUNTING
19808 sched_unpin();
19809 #endif
19810 /*
19811 * If timers want tcp_drop(), then pass error out,
19812 * otherwise suppress it.
19813 */
19814 return (retval < 0 ? retval : 0);
19815 }
19816 }
19817 if (rack->rc_in_persist) {
19818 if (tcp_in_hpts(rack->rc_tp) == 0) {
19819 /* Timer is not running */
19820 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19821 }
19822 #ifdef TCP_ACCOUNTING
19823 sched_unpin();
19824 #endif
19825 return (0);
19826 }
19827 if ((rack->rc_ack_required == 1) &&
19828 (rack->r_timer_override == 0)){
19829 /* A timeout occurred and no ack has arrived */
19830 if (tcp_in_hpts(rack->rc_tp) == 0) {
19831 /* Timer is not running */
19832 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19833 }
19834 #ifdef TCP_ACCOUNTING
19835 sched_unpin();
19836 #endif
19837 return (0);
19838 }
19839 if ((rack->r_timer_override) ||
19840 (rack->rc_ack_can_sendout_data) ||
19841 (delayed) ||
19842 (tp->t_state < TCPS_ESTABLISHED)) {
19843 rack->rc_ack_can_sendout_data = 0;
19844 if (tcp_in_hpts(rack->rc_tp))
19845 tcp_hpts_remove(rack->rc_tp);
19846 } else if (tcp_in_hpts(rack->rc_tp)) {
19847 /*
19848 * On the hpts you can't pass even if ACKNOW is on, we will
19849 * when the hpts fires.
19850 */
19851 #ifdef TCP_ACCOUNTING
19852 crtsc = get_cyclecount();
19853 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19854 tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19855 tp->tcp_cnt_counters[SND_BLOCKED]++;
19856 }
19857 sched_unpin();
19858 #endif
19859 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19860 return (0);
19861 }
19862 /* Finish out both pacing early and late accounting */
19863 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19864 TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19865 early = rack->r_ctl.rc_last_output_to - cts;
19866 } else
19867 early = 0;
19868 if (delayed && (rack->rc_always_pace == 1)) {
19869 rack->r_ctl.rc_agg_delayed += delayed;
19870 rack->r_late = 1;
19871 } else if (early && (rack->rc_always_pace == 1)) {
19872 rack->r_ctl.rc_agg_early += early;
19873 rack->r_early = 1;
19874 } else if (rack->rc_always_pace == 0) {
19875 /* Non-paced we are not late */
19876 rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
19877 rack->r_early = rack->r_late = 0;
19878 }
19879 /* Now that early/late accounting is done turn off the flag */
19880 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19881 rack->r_wanted_output = 0;
19882 rack->r_timer_override = 0;
19883 if ((tp->t_state != rack->r_state) &&
19884 TCPS_HAVEESTABLISHED(tp->t_state)) {
19885 rack_set_state(tp, rack);
19886 }
19887 if ((rack->r_fast_output) &&
19888 (doing_tlp == 0) &&
19889 (tp->rcv_numsacks == 0)) {
19890 int ret;
19891
19892 error = 0;
19893 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19894 if (ret >= 0)
19895 return(ret);
19896 else if (error) {
19897 inp = rack->rc_inp;
19898 so = inp->inp_socket;
19899 sb = &so->so_snd;
19900 goto nomore;
19901 }
19902 }
19903 inp = rack->rc_inp;
19904 /*
19905 * For TFO connections in SYN_SENT or SYN_RECEIVED,
19906 * only allow the initial SYN or SYN|ACK and those sent
19907 * by the retransmit timer.
19908 */
19909 if ((tp->t_flags & TF_FASTOPEN) &&
19910 ((tp->t_state == TCPS_SYN_RECEIVED) ||
19911 (tp->t_state == TCPS_SYN_SENT)) &&
19912 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19913 (tp->t_rxtshift == 0)) { /* not a retransmit */
19914 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19915 #ifdef TCP_ACCOUNTING
19916 sched_unpin();
19917 #endif
19918 return (0);
19919 }
19920 /*
19921 * Determine length of data that should be transmitted, and flags
19922 * that will be used. If there is some data or critical controls
19923 * (SYN, RST) to send, then transmit; otherwise, investigate
19924 * further.
19925 */
19926 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19927 if (tp->t_idle_reduce) {
19928 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19929 rack_cc_after_idle(rack, tp);
19930 }
19931 tp->t_flags &= ~TF_LASTIDLE;
19932 if (idle) {
19933 if (tp->t_flags & TF_MORETOCOME) {
19934 tp->t_flags |= TF_LASTIDLE;
19935 idle = 0;
19936 }
19937 }
19938 if ((tp->snd_una == tp->snd_max) &&
19939 rack->r_ctl.rc_went_idle_time &&
19940 (cts > rack->r_ctl.rc_went_idle_time)) {
19941 tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
19942 if (tot_idle > rack_min_probertt_hold) {
19943 /* Count as a probe rtt */
19944 if (rack->in_probe_rtt == 0) {
19945 rack->r_ctl.rc_lower_rtt_us_cts = cts;
19946 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19947 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19948 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19949 } else {
19950 rack_exit_probertt(rack, cts);
19951 }
19952 }
19953 } else
19954 tot_idle = 0;
19955 if (rack_use_fsb &&
19956 (rack->r_ctl.fsb.tcp_ip_hdr) &&
19957 (rack->r_fsb_inited == 0) &&
19958 (rack->r_state != TCPS_CLOSED))
19959 rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
19960 if (rack->rc_sendvars_notset == 1) {
19961 rack->rc_sendvars_notset = 0;
19962 /*
19963 * Make sure any TCP timers (keep-alive) is not running.
19964 */
19965 tcp_timer_stop(tp);
19966 }
19967 if ((rack->rack_no_prr == 1) &&
19968 (rack->rc_always_pace == 0)) {
19969 /*
19970 * Sanity check before sending, if we have
19971 * no-pacing enabled and prr is turned off that
19972 * is a logistics error. Correct this by turnning
19973 * prr back on. A user *must* set some form of
19974 * pacing in order to turn PRR off. We do this
19975 * in the output path so that we can avoid socket
19976 * option ordering issues that would occur if we
19977 * tried to do it while setting rack_no_prr on.
19978 */
19979 rack->rack_no_prr = 0;
19980 }
19981 if ((rack->pcm_enabled == 1) &&
19982 (rack->pcm_needed == 0) &&
19983 (tot_idle > 0)) {
19984 /*
19985 * We have been idle some micro seconds. We need
19986 * to factor this in to see if a PCM is needed.
19987 */
19988 uint32_t rtts_idle, rnds;
19989
19990 if (tp->t_srtt)
19991 rtts_idle = tot_idle / tp->t_srtt;
19992 else
19993 rtts_idle = 0;
19994 rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
19995 rack->r_ctl.pcm_idle_rounds += rtts_idle;
19996 if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
19997 rack->pcm_needed = 1;
19998 rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
19999 }
20000 }
20001 again:
20002 sendalot = 0;
20003 cts = tcp_get_usecs(&tv);
20004 ms_cts = tcp_tv_to_mssectick(&tv);
20005 tso = 0;
20006 mtu = 0;
20007 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20008 minseg = segsiz;
20009 if (rack->r_ctl.rc_pace_max_segs == 0)
20010 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20011 else
20012 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20013 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
20014 (rack->r_ctl.pcm_max_seg == 0)) {
20015 /*
20016 * We set in our first send so we know that the ctf_fixed_maxseg
20017 * has been fully set. If we do it in rack_init() we most likely
20018 * see 512 bytes so we end up at 5120, not desirable.
20019 */
20020 rack->r_ctl.pcm_max_seg = rc_init_window(rack);
20021 if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
20022 /*
20023 * Assure our initial PCM probe is at least 10 MSS.
20024 */
20025 rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
20026 }
20027 }
20028 if ((rack->r_ctl.pcm_max_seg != 0) && (rack->pcm_needed == 1)) {
20029 uint32_t rw_avail, cwa;
20030
20031 if (tp->snd_wnd > ctf_outstanding(tp))
20032 rw_avail = tp->snd_wnd - ctf_outstanding(tp);
20033 else
20034 rw_avail = 0;
20035 if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
20036 cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20037 else
20038 cwa = 0;
20039 if ((cwa >= rack->r_ctl.pcm_max_seg) &&
20040 (rw_avail > rack->r_ctl.pcm_max_seg)) {
20041 /* Raise up the max seg for this trip through */
20042 pace_max_seg = rack->r_ctl.pcm_max_seg;
20043 /* Disable any fast output */
20044 rack->r_fast_output = 0;
20045 }
20046 if (rack_verbose_logging) {
20047 rack_log_pcm(rack, 4,
20048 cwa, rack->r_ctl.pcm_max_seg, rw_avail);
20049 }
20050 }
20051 sb_offset = tp->snd_max - tp->snd_una;
20052 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20053 flags = tcp_outflags[tp->t_state];
20054 while (rack->rc_free_cnt < rack_free_cache) {
20055 rsm = rack_alloc(rack);
20056 if (rsm == NULL) {
20057 if (hpts_calling)
20058 /* Retry in a ms */
20059 slot = (1 * HPTS_USEC_IN_MSEC);
20060 so = inp->inp_socket;
20061 sb = &so->so_snd;
20062 goto just_return_nolock;
20063 }
20064 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20065 rack->rc_free_cnt++;
20066 rsm = NULL;
20067 }
20068 sack_rxmit = 0;
20069 len = 0;
20070 rsm = NULL;
20071 if (flags & TH_RST) {
20072 SOCK_SENDBUF_LOCK(inp->inp_socket);
20073 so = inp->inp_socket;
20074 sb = &so->so_snd;
20075 goto send;
20076 }
20077 if (rack->r_ctl.rc_resend) {
20078 /* Retransmit timer */
20079 rsm = rack->r_ctl.rc_resend;
20080 rack->r_ctl.rc_resend = NULL;
20081 len = rsm->r_end - rsm->r_start;
20082 sack_rxmit = 1;
20083 sendalot = 0;
20084 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20085 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20086 __func__, __LINE__,
20087 rsm->r_start, tp->snd_una, tp, rack, rsm));
20088 sb_offset = rsm->r_start - tp->snd_una;
20089 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20090 } else if (rack->r_collapse_point_valid &&
20091 ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20092 /*
20093 * If an RSM is returned then enough time has passed
20094 * for us to retransmit it. Move up the collapse point,
20095 * since this rsm has its chance to retransmit now.
20096 */
20097 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20098 rack->r_ctl.last_collapse_point = rsm->r_end;
20099 /* Are we done? */
20100 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20101 rack->r_ctl.high_collapse_point))
20102 rack->r_collapse_point_valid = 0;
20103 sack_rxmit = 1;
20104 /* We are not doing a TLP */
20105 doing_tlp = 0;
20106 len = rsm->r_end - rsm->r_start;
20107 sb_offset = rsm->r_start - tp->snd_una;
20108 sendalot = 0;
20109 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20110 } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20111 /* We have a retransmit that takes precedence */
20112 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20113 ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20114 ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20115 /* Enter recovery if not induced by a time-out */
20116 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20117 }
20118 #ifdef INVARIANTS
20119 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20120 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20121 tp, rack, rsm, rsm->r_start, tp->snd_una);
20122 }
20123 #endif
20124 len = rsm->r_end - rsm->r_start;
20125 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20126 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20127 __func__, __LINE__,
20128 rsm->r_start, tp->snd_una, tp, rack, rsm));
20129 sb_offset = rsm->r_start - tp->snd_una;
20130 sendalot = 0;
20131 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20132 if (len > 0) {
20133 sack_rxmit = 1;
20134 KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20135 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20136 min(len, segsiz));
20137 }
20138 } else if (rack->r_ctl.rc_tlpsend) {
20139 /* Tail loss probe */
20140 long cwin;
20141 long tlen;
20142
20143 /*
20144 * Check if we can do a TLP with a RACK'd packet
20145 * this can happen if we are not doing the rack
20146 * cheat and we skipped to a TLP and it
20147 * went off.
20148 */
20149 rsm = rack->r_ctl.rc_tlpsend;
20150 /* We are doing a TLP make sure the flag is preent */
20151 rsm->r_flags |= RACK_TLP;
20152 rack->r_ctl.rc_tlpsend = NULL;
20153 sack_rxmit = 1;
20154 tlen = rsm->r_end - rsm->r_start;
20155 if (tlen > segsiz)
20156 tlen = segsiz;
20157 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20158 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20159 __func__, __LINE__,
20160 rsm->r_start, tp->snd_una, tp, rack, rsm));
20161 sb_offset = rsm->r_start - tp->snd_una;
20162 cwin = min(tp->snd_wnd, tlen);
20163 len = cwin;
20164 }
20165 if (rack->r_must_retran &&
20166 (doing_tlp == 0) &&
20167 (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20168 (rsm == NULL)) {
20169 /*
20170 * There are two different ways that we
20171 * can get into this block:
20172 * a) This is a non-sack connection, we had a time-out
20173 * and thus r_must_retran was set and everything
20174 * left outstanding as been marked for retransmit.
20175 * b) The MTU of the path shrank, so that everything
20176 * was marked to be retransmitted with the smaller
20177 * mtu and r_must_retran was set.
20178 *
20179 * This means that we expect the sendmap (outstanding)
20180 * to all be marked must. We can use the tmap to
20181 * look at them.
20182 *
20183 */
20184 int sendwin, flight;
20185
20186 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20187 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20188 if (flight >= sendwin) {
20189 /*
20190 * We can't send yet.
20191 */
20192 so = inp->inp_socket;
20193 sb = &so->so_snd;
20194 goto just_return_nolock;
20195 }
20196 /*
20197 * This is the case a/b mentioned above. All
20198 * outstanding/not-acked should be marked.
20199 * We can use the tmap to find them.
20200 */
20201 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20202 if (rsm == NULL) {
20203 /* TSNH */
20204 rack->r_must_retran = 0;
20205 rack->r_ctl.rc_out_at_rto = 0;
20206 so = inp->inp_socket;
20207 sb = &so->so_snd;
20208 goto just_return_nolock;
20209 }
20210 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20211 /*
20212 * The first one does not have the flag, did we collapse
20213 * further up in our list?
20214 */
20215 rack->r_must_retran = 0;
20216 rack->r_ctl.rc_out_at_rto = 0;
20217 rsm = NULL;
20218 sack_rxmit = 0;
20219 } else {
20220 sack_rxmit = 1;
20221 len = rsm->r_end - rsm->r_start;
20222 sb_offset = rsm->r_start - tp->snd_una;
20223 sendalot = 0;
20224 if ((rack->full_size_rxt == 0) &&
20225 (rack->shape_rxt_to_pacing_min == 0) &&
20226 (len >= segsiz))
20227 len = segsiz;
20228 else if (rack->shape_rxt_to_pacing_min &&
20229 rack->gp_ready) {
20230 /* We use pacing min as shaping len req */
20231 uint32_t maxlen;
20232
20233 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20234 if (len > maxlen)
20235 len = maxlen;
20236 }
20237 /*
20238 * Delay removing the flag RACK_MUST_RXT so
20239 * that the fastpath for retransmit will
20240 * work with this rsm.
20241 */
20242 }
20243 }
20244 /*
20245 * Enforce a connection sendmap count limit if set
20246 * as long as we are not retransmiting.
20247 */
20248 if ((rsm == NULL) &&
20249 (V_tcp_map_entries_limit > 0) &&
20250 (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20251 counter_u64_add(rack_to_alloc_limited, 1);
20252 if (!rack->alloc_limit_reported) {
20253 rack->alloc_limit_reported = 1;
20254 counter_u64_add(rack_alloc_limited_conns, 1);
20255 }
20256 so = inp->inp_socket;
20257 sb = &so->so_snd;
20258 goto just_return_nolock;
20259 }
20260 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20261 /* we are retransmitting the fin */
20262 len--;
20263 if (len) {
20264 /*
20265 * When retransmitting data do *not* include the
20266 * FIN. This could happen from a TLP probe.
20267 */
20268 flags &= ~TH_FIN;
20269 }
20270 }
20271 if (rsm && rack->r_fsb_inited &&
20272 rack_use_rsm_rfo &&
20273 ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20274 int ret;
20275
20276 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20277 if (ret == 0)
20278 return (0);
20279 }
20280 so = inp->inp_socket;
20281 sb = &so->so_snd;
20282 if (do_a_prefetch == 0) {
20283 kern_prefetch(sb, &do_a_prefetch);
20284 do_a_prefetch = 1;
20285 }
20286 #ifdef NETFLIX_SHARED_CWND
20287 if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20288 rack->rack_enable_scwnd) {
20289 /* We are doing cwnd sharing */
20290 if (rack->gp_ready &&
20291 (rack->rack_attempted_scwnd == 0) &&
20292 (rack->r_ctl.rc_scw == NULL) &&
20293 tp->t_lib) {
20294 /* The pcbid is in, lets make an attempt */
20295 counter_u64_add(rack_try_scwnd, 1);
20296 rack->rack_attempted_scwnd = 1;
20297 rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20298 &rack->r_ctl.rc_scw_index,
20299 segsiz);
20300 }
20301 if (rack->r_ctl.rc_scw &&
20302 (rack->rack_scwnd_is_idle == 1) &&
20303 sbavail(&so->so_snd)) {
20304 /* we are no longer out of data */
20305 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20306 rack->rack_scwnd_is_idle = 0;
20307 }
20308 if (rack->r_ctl.rc_scw) {
20309 /* First lets update and get the cwnd */
20310 rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20311 rack->r_ctl.rc_scw_index,
20312 tp->snd_cwnd, tp->snd_wnd, segsiz);
20313 }
20314 }
20315 #endif
20316 /*
20317 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20318 * state flags.
20319 */
20320 if (tp->t_flags & TF_NEEDFIN)
20321 flags |= TH_FIN;
20322 if (tp->t_flags & TF_NEEDSYN)
20323 flags |= TH_SYN;
20324 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20325 void *end_rsm;
20326 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20327 if (end_rsm)
20328 kern_prefetch(end_rsm, &prefetch_rsm);
20329 prefetch_rsm = 1;
20330 }
20331 SOCK_SENDBUF_LOCK(so);
20332 if ((sack_rxmit == 0) &&
20333 (TCPS_HAVEESTABLISHED(tp->t_state) ||
20334 (tp->t_flags & TF_FASTOPEN))) {
20335 /*
20336 * We are not retransmitting (sack_rxmit is 0) so we
20337 * are sending new data. This is always based on snd_max.
20338 * Now in theory snd_max may be equal to snd_una, if so
20339 * then nothing is outstanding and the offset would be 0.
20340 */
20341 uint32_t avail;
20342
20343 avail = sbavail(sb);
20344 if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
20345 sb_offset = tp->snd_max - tp->snd_una;
20346 else
20347 sb_offset = 0;
20348 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20349 if (rack->r_ctl.rc_tlp_new_data) {
20350 /* TLP is forcing out new data */
20351 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20352 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20353 }
20354 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20355 if (tp->snd_wnd > sb_offset)
20356 len = tp->snd_wnd - sb_offset;
20357 else
20358 len = 0;
20359 } else {
20360 len = rack->r_ctl.rc_tlp_new_data;
20361 }
20362 rack->r_ctl.rc_tlp_new_data = 0;
20363 } else {
20364 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20365 }
20366 if ((rack->r_ctl.crte == NULL) &&
20367 IN_FASTRECOVERY(tp->t_flags) &&
20368 (rack->full_size_rxt == 0) &&
20369 (rack->shape_rxt_to_pacing_min == 0) &&
20370 (len > segsiz)) {
20371 /*
20372 * For prr=off, we need to send only 1 MSS
20373 * at a time. We do this because another sack could
20374 * be arriving that causes us to send retransmits and
20375 * we don't want to be on a long pace due to a larger send
20376 * that keeps us from sending out the retransmit.
20377 */
20378 len = segsiz;
20379 } else if (rack->shape_rxt_to_pacing_min &&
20380 rack->gp_ready) {
20381 /* We use pacing min as shaping len req */
20382 uint32_t maxlen;
20383
20384 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20385 if (len > maxlen)
20386 len = maxlen;
20387 }/* The else is full_size_rxt is on so send it all */
20388 } else {
20389 uint32_t outstanding;
20390 /*
20391 * We are inside of a Fast recovery episode, this
20392 * is caused by a SACK or 3 dup acks. At this point
20393 * we have sent all the retransmissions and we rely
20394 * on PRR to dictate what we will send in the form of
20395 * new data.
20396 */
20397
20398 outstanding = tp->snd_max - tp->snd_una;
20399 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20400 if (tp->snd_wnd > outstanding) {
20401 len = tp->snd_wnd - outstanding;
20402 /* Check to see if we have the data */
20403 if ((sb_offset + len) > avail) {
20404 /* It does not all fit */
20405 if (avail > sb_offset)
20406 len = avail - sb_offset;
20407 else
20408 len = 0;
20409 }
20410 } else {
20411 len = 0;
20412 }
20413 } else if (avail > sb_offset) {
20414 len = avail - sb_offset;
20415 } else {
20416 len = 0;
20417 }
20418 if (len > 0) {
20419 if (len > rack->r_ctl.rc_prr_sndcnt) {
20420 len = rack->r_ctl.rc_prr_sndcnt;
20421 }
20422 if (len > 0) {
20423 sub_from_prr = 1;
20424 }
20425 }
20426 if (len > segsiz) {
20427 /*
20428 * We should never send more than a MSS when
20429 * retransmitting or sending new data in prr
20430 * mode unless the override flag is on. Most
20431 * likely the PRR algorithm is not going to
20432 * let us send a lot as well :-)
20433 */
20434 if (rack->r_ctl.rc_prr_sendalot == 0) {
20435 len = segsiz;
20436 }
20437 } else if (len < segsiz) {
20438 /*
20439 * Do we send any? The idea here is if the
20440 * send empty's the socket buffer we want to
20441 * do it. However if not then lets just wait
20442 * for our prr_sndcnt to get bigger.
20443 */
20444 long leftinsb;
20445
20446 leftinsb = sbavail(sb) - sb_offset;
20447 if (leftinsb > len) {
20448 /* This send does not empty the sb */
20449 len = 0;
20450 }
20451 }
20452 }
20453 } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20454 /*
20455 * If you have not established
20456 * and are not doing FAST OPEN
20457 * no data please.
20458 */
20459 if ((sack_rxmit == 0) &&
20460 !(tp->t_flags & TF_FASTOPEN)) {
20461 len = 0;
20462 sb_offset = 0;
20463 }
20464 }
20465 if (prefetch_so_done == 0) {
20466 kern_prefetch(so, &prefetch_so_done);
20467 prefetch_so_done = 1;
20468 }
20469 orig_len = len;
20470 /*
20471 * Lop off SYN bit if it has already been sent. However, if this is
20472 * SYN-SENT state and if segment contains data and if we don't know
20473 * that foreign host supports TAO, suppress sending segment.
20474 */
20475 if ((flags & TH_SYN) &&
20476 SEQ_GT(tp->snd_max, tp->snd_una) &&
20477 ((sack_rxmit == 0) &&
20478 (tp->t_rxtshift == 0))) {
20479 /*
20480 * When sending additional segments following a TFO SYN|ACK,
20481 * do not include the SYN bit.
20482 */
20483 if ((tp->t_flags & TF_FASTOPEN) &&
20484 (tp->t_state == TCPS_SYN_RECEIVED))
20485 flags &= ~TH_SYN;
20486 }
20487 /*
20488 * Be careful not to send data and/or FIN on SYN segments. This
20489 * measure is needed to prevent interoperability problems with not
20490 * fully conformant TCP implementations.
20491 */
20492 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20493 len = 0;
20494 flags &= ~TH_FIN;
20495 }
20496 /*
20497 * On TFO sockets, ensure no data is sent in the following cases:
20498 *
20499 * - When retransmitting SYN|ACK on a passively-created socket
20500 *
20501 * - When retransmitting SYN on an actively created socket
20502 *
20503 * - When sending a zero-length cookie (cookie request) on an
20504 * actively created socket
20505 *
20506 * - When the socket is in the CLOSED state (RST is being sent)
20507 */
20508 if ((tp->t_flags & TF_FASTOPEN) &&
20509 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20510 ((tp->t_state == TCPS_SYN_SENT) &&
20511 (tp->t_tfo_client_cookie_len == 0)) ||
20512 (flags & TH_RST))) {
20513 sack_rxmit = 0;
20514 len = 0;
20515 }
20516 /* Without fast-open there should never be data sent on a SYN */
20517 if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
20518 len = 0;
20519 }
20520 if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20521 /* We only send 1 MSS if we have a DSACK block */
20522 add_flag |= RACK_SENT_W_DSACK;
20523 len = segsiz;
20524 }
20525 if (len <= 0) {
20526 /*
20527 * We have nothing to send, or the window shrank, or
20528 * is closed, do we need to go into persists?
20529 */
20530 len = 0;
20531 if ((tp->snd_wnd == 0) &&
20532 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20533 (tp->snd_una == tp->snd_max) &&
20534 (sb_offset < (int)sbavail(sb))) {
20535 rack_enter_persist(tp, rack, cts, tp->snd_una);
20536 }
20537 } else if ((rsm == NULL) &&
20538 (doing_tlp == 0) &&
20539 (len < pace_max_seg)) {
20540 /*
20541 * We are not sending a maximum sized segment for
20542 * some reason. Should we not send anything (think
20543 * sws or persists)?
20544 */
20545 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20546 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20547 (len < minseg) &&
20548 (len < (int)(sbavail(sb) - sb_offset))) {
20549 /*
20550 * Here the rwnd is less than
20551 * the minimum pacing size, this is not a retransmit,
20552 * we are established and
20553 * the send is not the last in the socket buffer
20554 * we send nothing, and we may enter persists
20555 * if nothing is outstanding.
20556 */
20557 len = 0;
20558 if (tp->snd_max == tp->snd_una) {
20559 /*
20560 * Nothing out we can
20561 * go into persists.
20562 */
20563 rack_enter_persist(tp, rack, cts, tp->snd_una);
20564 }
20565 } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20566 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20567 (len < (int)(sbavail(sb) - sb_offset)) &&
20568 (len < minseg)) {
20569 /*
20570 * Here we are not retransmitting, and
20571 * the cwnd is not so small that we could
20572 * not send at least a min size (rxt timer
20573 * not having gone off), We have 2 segments or
20574 * more already in flight, its not the tail end
20575 * of the socket buffer and the cwnd is blocking
20576 * us from sending out a minimum pacing segment size.
20577 * Lets not send anything.
20578 */
20579 len = 0;
20580 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20581 min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20582 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20583 (len < (int)(sbavail(sb) - sb_offset)) &&
20584 (TCPS_HAVEESTABLISHED(tp->t_state))) {
20585 /*
20586 * Here we have a send window but we have
20587 * filled it up and we can't send another pacing segment.
20588 * We also have in flight more than 2 segments
20589 * and we are not completing the sb i.e. we allow
20590 * the last bytes of the sb to go out even if
20591 * its not a full pacing segment.
20592 */
20593 len = 0;
20594 } else if ((rack->r_ctl.crte != NULL) &&
20595 (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20596 (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20597 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20598 (len < (int)(sbavail(sb) - sb_offset))) {
20599 /*
20600 * Here we are doing hardware pacing, this is not a TLP,
20601 * we are not sending a pace max segment size, there is rwnd
20602 * room to send at least N pace_max_seg, the cwnd is greater
20603 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20604 * more segments in flight and its not the tail of the socket buffer.
20605 *
20606 * We don't want to send instead we need to get more ack's in to
20607 * allow us to send a full pacing segment. Normally, if we are pacing
20608 * about the right speed, we should have finished our pacing
20609 * send as most of the acks have come back if we are at the
20610 * right rate. This is a bit fuzzy since return path delay
20611 * can delay the acks, which is why we want to make sure we
20612 * have cwnd space to have a bit more than a max pace segments in flight.
20613 *
20614 * If we have not gotten our acks back we are pacing at too high a
20615 * rate delaying will not hurt and will bring our GP estimate down by
20616 * injecting the delay. If we don't do this we will send
20617 * 2 MSS out in response to the acks being clocked in which
20618 * defeats the point of hw-pacing (i.e. to help us get
20619 * larger TSO's out).
20620 */
20621 len = 0;
20622 }
20623
20624 }
20625 /* len will be >= 0 after this point. */
20626 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20627 rack_sndbuf_autoscale(rack);
20628 /*
20629 * Decide if we can use TCP Segmentation Offloading (if supported by
20630 * hardware).
20631 *
20632 * TSO may only be used if we are in a pure bulk sending state. The
20633 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20634 * options prevent using TSO. With TSO the TCP header is the same
20635 * (except for the sequence number) for all generated packets. This
20636 * makes it impossible to transmit any options which vary per
20637 * generated segment or packet.
20638 *
20639 * IPv4 handling has a clear separation of ip options and ip header
20640 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20641 * the right thing below to provide length of just ip options and thus
20642 * checking for ipoptlen is enough to decide if ip options are present.
20643 */
20644 ipoptlen = 0;
20645 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20646 /*
20647 * Pre-calculate here as we save another lookup into the darknesses
20648 * of IPsec that way and can actually decide if TSO is ok.
20649 */
20650 #ifdef INET6
20651 if (isipv6 && IPSEC_ENABLED(ipv6))
20652 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20653 #ifdef INET
20654 else
20655 #endif
20656 #endif /* INET6 */
20657 #ifdef INET
20658 if (IPSEC_ENABLED(ipv4))
20659 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20660 #endif /* INET */
20661 #endif
20662
20663 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20664 ipoptlen += ipsec_optlen;
20665 #endif
20666 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20667 (tp->t_port == 0) &&
20668 ((tp->t_flags & TF_SIGNATURE) == 0) &&
20669 sack_rxmit == 0 &&
20670 ipoptlen == 0)
20671 tso = 1;
20672 {
20673 uint32_t outstanding __unused;
20674
20675 outstanding = tp->snd_max - tp->snd_una;
20676 if (tp->t_flags & TF_SENTFIN) {
20677 /*
20678 * If we sent a fin, snd_max is 1 higher than
20679 * snd_una
20680 */
20681 outstanding--;
20682 }
20683 if (sack_rxmit) {
20684 if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20685 flags &= ~TH_FIN;
20686 }
20687 }
20688 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20689 (long)TCP_MAXWIN << tp->rcv_scale);
20690
20691 /*
20692 * Sender silly window avoidance. We transmit under the following
20693 * conditions when len is non-zero:
20694 *
20695 * - We have a full segment (or more with TSO) - This is the last
20696 * buffer in a write()/send() and we are either idle or running
20697 * NODELAY - we've timed out (e.g. persist timer) - we have more
20698 * then 1/2 the maximum send window's worth of data (receiver may be
20699 * limited the window size) - we need to retransmit
20700 */
20701 if (len) {
20702 if (len >= segsiz) {
20703 goto send;
20704 }
20705 /*
20706 * NOTE! on localhost connections an 'ack' from the remote
20707 * end may occur synchronously with the output and cause us
20708 * to flush a buffer queued with moretocome. XXX
20709 *
20710 */
20711 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
20712 (idle || (tp->t_flags & TF_NODELAY)) &&
20713 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20714 (tp->t_flags & TF_NOPUSH) == 0) {
20715 pass = 2;
20716 goto send;
20717 }
20718 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */
20719 pass = 22;
20720 goto send;
20721 }
20722 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20723 pass = 4;
20724 goto send;
20725 }
20726 if (sack_rxmit) {
20727 pass = 6;
20728 goto send;
20729 }
20730 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20731 (ctf_outstanding(tp) < (segsiz * 2))) {
20732 /*
20733 * We have less than two MSS outstanding (delayed ack)
20734 * and our rwnd will not let us send a full sized
20735 * MSS. Lets go ahead and let this small segment
20736 * out because we want to try to have at least two
20737 * packets inflight to not be caught by delayed ack.
20738 */
20739 pass = 12;
20740 goto send;
20741 }
20742 }
20743 /*
20744 * Sending of standalone window updates.
20745 *
20746 * Window updates are important when we close our window due to a
20747 * full socket buffer and are opening it again after the application
20748 * reads data from it. Once the window has opened again and the
20749 * remote end starts to send again the ACK clock takes over and
20750 * provides the most current window information.
20751 *
20752 * We must avoid the silly window syndrome whereas every read from
20753 * the receive buffer, no matter how small, causes a window update
20754 * to be sent. We also should avoid sending a flurry of window
20755 * updates when the socket buffer had queued a lot of data and the
20756 * application is doing small reads.
20757 *
20758 * Prevent a flurry of pointless window updates by only sending an
20759 * update when we can increase the advertized window by more than
20760 * 1/4th of the socket buffer capacity. When the buffer is getting
20761 * full or is very small be more aggressive and send an update
20762 * whenever we can increase by two mss sized segments. In all other
20763 * situations the ACK's to new incoming data will carry further
20764 * window increases.
20765 *
20766 * Don't send an independent window update if a delayed ACK is
20767 * pending (it will get piggy-backed on it) or the remote side
20768 * already has done a half-close and won't send more data. Skip
20769 * this if the connection is in T/TCP half-open state.
20770 */
20771 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20772 !(tp->t_flags & TF_DELACK) &&
20773 !TCPS_HAVERCVDFIN(tp->t_state)) {
20774 /*
20775 * "adv" is the amount we could increase the window, taking
20776 * into account that we are limited by TCP_MAXWIN <<
20777 * tp->rcv_scale.
20778 */
20779 int32_t adv;
20780 int oldwin;
20781
20782 adv = recwin;
20783 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20784 oldwin = (tp->rcv_adv - tp->rcv_nxt);
20785 if (adv > oldwin)
20786 adv -= oldwin;
20787 else {
20788 /* We can't increase the window */
20789 adv = 0;
20790 }
20791 } else
20792 oldwin = 0;
20793
20794 /*
20795 * If the new window size ends up being the same as or less
20796 * than the old size when it is scaled, then don't force
20797 * a window update.
20798 */
20799 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20800 goto dontupdate;
20801
20802 if (adv >= (int32_t)(2 * segsiz) &&
20803 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20804 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20805 so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20806 pass = 7;
20807 goto send;
20808 }
20809 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20810 pass = 23;
20811 goto send;
20812 }
20813 }
20814 dontupdate:
20815
20816 /*
20817 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
20818 * is also a catch-all for the retransmit timer timeout case.
20819 */
20820 if (tp->t_flags & TF_ACKNOW) {
20821 pass = 8;
20822 goto send;
20823 }
20824 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20825 pass = 9;
20826 goto send;
20827 }
20828 /*
20829 * If our state indicates that FIN should be sent and we have not
20830 * yet done so, then we need to send.
20831 */
20832 if ((flags & TH_FIN) &&
20833 (tp->snd_max == tp->snd_una)) {
20834 pass = 11;
20835 goto send;
20836 }
20837 /*
20838 * No reason to send a segment, just return.
20839 */
20840 just_return:
20841 SOCK_SENDBUF_UNLOCK(so);
20842 just_return_nolock:
20843 {
20844 int app_limited = CTF_JR_SENT_DATA;
20845
20846 if ((tp->t_flags & TF_FASTOPEN) == 0 &&
20847 (flags & TH_FIN) &&
20848 (len == 0) &&
20849 (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
20850 ((tp->snd_max - tp->snd_una) <= segsiz)) {
20851 /*
20852 * Ok less than or right at a MSS is
20853 * outstanding. The original FreeBSD stack would
20854 * have sent a FIN, which can speed things up for
20855 * a transactional application doing a MSG_WAITALL.
20856 * To speed things up since we do *not* send a FIN
20857 * if data is outstanding, we send a "challenge ack".
20858 * The idea behind that is instead of having to have
20859 * the peer wait for the delayed-ack timer to run off
20860 * we send an ack that makes the peer send us an ack.
20861 */
20862 rack_send_ack_challange(rack);
20863 }
20864 if (tot_len_this_send > 0) {
20865 rack->r_ctl.fsb.recwin = recwin;
20866 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
20867 if ((error == 0) &&
20868 rack_use_rfo &&
20869 ((flags & (TH_SYN|TH_FIN)) == 0) &&
20870 (ipoptlen == 0) &&
20871 rack->r_fsb_inited &&
20872 TCPS_HAVEESTABLISHED(tp->t_state) &&
20873 ((IN_RECOVERY(tp->t_flags)) == 0) &&
20874 (rack->r_must_retran == 0) &&
20875 ((tp->t_flags & TF_NEEDFIN) == 0) &&
20876 (len > 0) && (orig_len > 0) &&
20877 (orig_len > len) &&
20878 ((orig_len - len) >= segsiz) &&
20879 ((optlen == 0) ||
20880 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20881 /* We can send at least one more MSS using our fsb */
20882 rack_setup_fast_output(tp, rack, sb, len, orig_len,
20883 segsiz, pace_max_seg, hw_tls, flags);
20884 } else
20885 rack->r_fast_output = 0;
20886 rack_log_fsb(rack, tp, so, flags,
20887 ipoptlen, orig_len, len, 0,
20888 1, optlen, __LINE__, 1);
20889 /* Assure when we leave that snd_nxt will point to top */
20890 if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20891 tp->snd_nxt = tp->snd_max;
20892 } else {
20893 int end_window = 0;
20894 uint32_t seq = tp->gput_ack;
20895
20896 rsm = tqhash_max(rack->r_ctl.tqh);
20897 if (rsm) {
20898 /*
20899 * Mark the last sent that we just-returned (hinting
20900 * that delayed ack may play a role in any rtt measurement).
20901 */
20902 rsm->r_just_ret = 1;
20903 }
20904 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20905 rack->r_ctl.rc_agg_delayed = 0;
20906 rack->r_early = 0;
20907 rack->r_late = 0;
20908 rack->r_ctl.rc_agg_early = 0;
20909 if ((ctf_outstanding(tp) +
20910 min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20911 minseg)) >= tp->snd_wnd) {
20912 /* We are limited by the rwnd */
20913 app_limited = CTF_JR_RWND_LIMITED;
20914 if (IN_FASTRECOVERY(tp->t_flags))
20915 rack->r_ctl.rc_prr_sndcnt = 0;
20916 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
20917 /* We are limited by whats available -- app limited */
20918 app_limited = CTF_JR_APP_LIMITED;
20919 if (IN_FASTRECOVERY(tp->t_flags))
20920 rack->r_ctl.rc_prr_sndcnt = 0;
20921 } else if ((idle == 0) &&
20922 ((tp->t_flags & TF_NODELAY) == 0) &&
20923 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20924 (len < segsiz)) {
20925 /*
20926 * No delay is not on and the
20927 * user is sending less than 1MSS. This
20928 * brings out SWS avoidance so we
20929 * don't send. Another app-limited case.
20930 */
20931 app_limited = CTF_JR_APP_LIMITED;
20932 } else if (tp->t_flags & TF_NOPUSH) {
20933 /*
20934 * The user has requested no push of
20935 * the last segment and we are
20936 * at the last segment. Another app
20937 * limited case.
20938 */
20939 app_limited = CTF_JR_APP_LIMITED;
20940 } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20941 /* Its the cwnd */
20942 app_limited = CTF_JR_CWND_LIMITED;
20943 } else if (IN_FASTRECOVERY(tp->t_flags) &&
20944 (rack->rack_no_prr == 0) &&
20945 (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20946 app_limited = CTF_JR_PRR;
20947 } else {
20948 /* Now why here are we not sending? */
20949 #ifdef NOW
20950 #ifdef INVARIANTS
20951 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20952 #endif
20953 #endif
20954 app_limited = CTF_JR_ASSESSING;
20955 }
20956 /*
20957 * App limited in some fashion, for our pacing GP
20958 * measurements we don't want any gap (even cwnd).
20959 * Close down the measurement window.
20960 */
20961 if (rack_cwnd_block_ends_measure &&
20962 ((app_limited == CTF_JR_CWND_LIMITED) ||
20963 (app_limited == CTF_JR_PRR))) {
20964 /*
20965 * The reason we are not sending is
20966 * the cwnd (or prr). We have been configured
20967 * to end the measurement window in
20968 * this case.
20969 */
20970 end_window = 1;
20971 } else if (rack_rwnd_block_ends_measure &&
20972 (app_limited == CTF_JR_RWND_LIMITED)) {
20973 /*
20974 * We are rwnd limited and have been
20975 * configured to end the measurement
20976 * window in this case.
20977 */
20978 end_window = 1;
20979 } else if (app_limited == CTF_JR_APP_LIMITED) {
20980 /*
20981 * A true application limited period, we have
20982 * ran out of data.
20983 */
20984 end_window = 1;
20985 } else if (app_limited == CTF_JR_ASSESSING) {
20986 /*
20987 * In the assessing case we hit the end of
20988 * the if/else and had no known reason
20989 * This will panic us under invariants..
20990 *
20991 * If we get this out in logs we need to
20992 * investagate which reason we missed.
20993 */
20994 end_window = 1;
20995 }
20996 if (end_window) {
20997 uint8_t log = 0;
20998
20999 /* Adjust the Gput measurement */
21000 if ((tp->t_flags & TF_GPUTINPROG) &&
21001 SEQ_GT(tp->gput_ack, tp->snd_max)) {
21002 tp->gput_ack = tp->snd_max;
21003 if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21004 /*
21005 * There is not enough to measure.
21006 */
21007 tp->t_flags &= ~TF_GPUTINPROG;
21008 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21009 rack->r_ctl.rc_gp_srtt /*flex1*/,
21010 tp->gput_seq,
21011 0, 0, 18, __LINE__, NULL, 0);
21012 } else
21013 log = 1;
21014 }
21015 /* Mark the last packet has app limited */
21016 rsm = tqhash_max(rack->r_ctl.tqh);
21017 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21018 if (rack->r_ctl.rc_app_limited_cnt == 0)
21019 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21020 else {
21021 /*
21022 * Go out to the end app limited and mark
21023 * this new one as next and move the end_appl up
21024 * to this guy.
21025 */
21026 if (rack->r_ctl.rc_end_appl)
21027 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21028 rack->r_ctl.rc_end_appl = rsm;
21029 }
21030 rsm->r_flags |= RACK_APP_LIMITED;
21031 rack->r_ctl.rc_app_limited_cnt++;
21032 }
21033 if (log)
21034 rack_log_pacing_delay_calc(rack,
21035 rack->r_ctl.rc_app_limited_cnt, seq,
21036 tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21037 }
21038 }
21039 /* Check if we need to go into persists or not */
21040 if ((tp->snd_max == tp->snd_una) &&
21041 TCPS_HAVEESTABLISHED(tp->t_state) &&
21042 sbavail(sb) &&
21043 (sbavail(sb) > tp->snd_wnd) &&
21044 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21045 /* Yes lets make sure to move to persist before timer-start */
21046 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21047 }
21048 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21049 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21050 }
21051 #ifdef NETFLIX_SHARED_CWND
21052 if ((sbavail(sb) == 0) &&
21053 rack->r_ctl.rc_scw) {
21054 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21055 rack->rack_scwnd_is_idle = 1;
21056 }
21057 #endif
21058 #ifdef TCP_ACCOUNTING
21059 if (tot_len_this_send > 0) {
21060 crtsc = get_cyclecount();
21061 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21062 tp->tcp_cnt_counters[SND_OUT_DATA]++;
21063 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21064 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21065 }
21066 } else {
21067 crtsc = get_cyclecount();
21068 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21069 tp->tcp_cnt_counters[SND_LIMITED]++;
21070 tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21071 }
21072 }
21073 sched_unpin();
21074 #endif
21075 return (0);
21076
21077 send:
21078 if ((rack->r_ctl.crte != NULL) &&
21079 (rsm == NULL) &&
21080 ((rack->rc_hw_nobuf == 1) ||
21081 (rack_hw_check_queue && (check_done == 0)))) {
21082 /*
21083 * We only want to do this once with the hw_check_queue,
21084 * for the enobuf case we would only do it once if
21085 * we come around to again, the flag will be clear.
21086 */
21087 check_done = 1;
21088 slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21089 if (slot) {
21090 rack->r_ctl.rc_agg_delayed = 0;
21091 rack->r_ctl.rc_agg_early = 0;
21092 rack->r_early = 0;
21093 rack->r_late = 0;
21094 SOCK_SENDBUF_UNLOCK(so);
21095 goto skip_all_send;
21096 }
21097 }
21098 if (rsm || sack_rxmit)
21099 counter_u64_add(rack_nfto_resend, 1);
21100 else
21101 counter_u64_add(rack_non_fto_send, 1);
21102 if ((flags & TH_FIN) &&
21103 sbavail(sb)) {
21104 /*
21105 * We do not transmit a FIN
21106 * with data outstanding. We
21107 * need to make it so all data
21108 * is acked first.
21109 */
21110 flags &= ~TH_FIN;
21111 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21112 (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21113 ((tp->snd_max - tp->snd_una) <= segsiz)) {
21114 /*
21115 * Ok less than or right at a MSS is
21116 * outstanding. The original FreeBSD stack would
21117 * have sent a FIN, which can speed things up for
21118 * a transactional application doing a MSG_WAITALL.
21119 * To speed things up since we do *not* send a FIN
21120 * if data is outstanding, we send a "challenge ack".
21121 * The idea behind that is instead of having to have
21122 * the peer wait for the delayed-ack timer to run off
21123 * we send an ack that makes the peer send us an ack.
21124 */
21125 rack_send_ack_challange(rack);
21126 }
21127 }
21128 /* Enforce stack imposed max seg size if we have one */
21129 if (pace_max_seg &&
21130 (len > pace_max_seg)) {
21131 mark = 1;
21132 len = pace_max_seg;
21133 }
21134 if ((rsm == NULL) &&
21135 (rack->pcm_in_progress == 0) &&
21136 (rack->r_ctl.pcm_max_seg > 0) &&
21137 (len >= rack->r_ctl.pcm_max_seg)) {
21138 /* It is large enough for a measurement */
21139 add_flag |= RACK_IS_PCM;
21140 rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg, add_flag);
21141 } else if (rack_verbose_logging) {
21142 rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg, add_flag);
21143 }
21144
21145 SOCKBUF_LOCK_ASSERT(sb);
21146 if (len > 0) {
21147 if (len >= segsiz)
21148 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21149 else
21150 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21151 }
21152 /*
21153 * Before ESTABLISHED, force sending of initial options unless TCP
21154 * set not to do any options. NOTE: we assume that the IP/TCP header
21155 * plus TCP options always fit in a single mbuf, leaving room for a
21156 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21157 * + optlen <= MCLBYTES
21158 */
21159 optlen = 0;
21160 #ifdef INET6
21161 if (isipv6)
21162 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21163 else
21164 #endif
21165 hdrlen = sizeof(struct tcpiphdr);
21166
21167 /*
21168 * Ok what seq are we sending from. If we have
21169 * no rsm to use, then we look at various bits,
21170 * if we are putting out a SYN it will be ISS.
21171 * If we are retransmitting a FIN it will
21172 * be snd_max-1 else its snd_max.
21173 */
21174 if (rsm == NULL) {
21175 if (flags & TH_SYN)
21176 rack_seq = tp->iss;
21177 else if ((flags & TH_FIN) &&
21178 (tp->t_flags & TF_SENTFIN))
21179 rack_seq = tp->snd_max - 1;
21180 else
21181 rack_seq = tp->snd_max;
21182 } else {
21183 rack_seq = rsm->r_start;
21184 }
21185 /*
21186 * Compute options for segment. We only have to care about SYN and
21187 * established connection segments. Options for SYN-ACK segments
21188 * are handled in TCP syncache.
21189 */
21190 to.to_flags = 0;
21191 if ((tp->t_flags & TF_NOOPT) == 0) {
21192 /* Maximum segment size. */
21193 if (flags & TH_SYN) {
21194 to.to_mss = tcp_mssopt(&inp->inp_inc);
21195 if (tp->t_port)
21196 to.to_mss -= V_tcp_udp_tunneling_overhead;
21197 to.to_flags |= TOF_MSS;
21198
21199 /*
21200 * On SYN or SYN|ACK transmits on TFO connections,
21201 * only include the TFO option if it is not a
21202 * retransmit, as the presence of the TFO option may
21203 * have caused the original SYN or SYN|ACK to have
21204 * been dropped by a middlebox.
21205 */
21206 if ((tp->t_flags & TF_FASTOPEN) &&
21207 (tp->t_rxtshift == 0)) {
21208 if (tp->t_state == TCPS_SYN_RECEIVED) {
21209 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21210 to.to_tfo_cookie =
21211 (u_int8_t *)&tp->t_tfo_cookie.server;
21212 to.to_flags |= TOF_FASTOPEN;
21213 wanted_cookie = 1;
21214 } else if (tp->t_state == TCPS_SYN_SENT) {
21215 to.to_tfo_len =
21216 tp->t_tfo_client_cookie_len;
21217 to.to_tfo_cookie =
21218 tp->t_tfo_cookie.client;
21219 to.to_flags |= TOF_FASTOPEN;
21220 wanted_cookie = 1;
21221 /*
21222 * If we wind up having more data to
21223 * send with the SYN than can fit in
21224 * one segment, don't send any more
21225 * until the SYN|ACK comes back from
21226 * the other end.
21227 */
21228 sendalot = 0;
21229 }
21230 }
21231 }
21232 /* Window scaling. */
21233 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21234 to.to_wscale = tp->request_r_scale;
21235 to.to_flags |= TOF_SCALE;
21236 }
21237 /* Timestamps. */
21238 if ((tp->t_flags & TF_RCVD_TSTMP) ||
21239 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21240 uint32_t ts_to_use;
21241
21242 if ((rack->r_rcvpath_rtt_up == 1) &&
21243 (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
21244 /*
21245 * When we are doing a rcv_rtt probe all
21246 * other timestamps use the next msec. This
21247 * is safe since our previous ack is in the
21248 * air and we will just have a few more
21249 * on the next ms. This assures that only
21250 * the one ack has the ms_cts that was on
21251 * our ack-probe.
21252 */
21253 ts_to_use = ms_cts + 1;
21254 } else {
21255 ts_to_use = ms_cts;
21256 }
21257 to.to_tsval = ts_to_use + tp->ts_offset;
21258 to.to_tsecr = tp->ts_recent;
21259 to.to_flags |= TOF_TS;
21260 if ((len == 0) &&
21261 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21262 ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
21263 (tp->snd_una == tp->snd_max) &&
21264 (flags & TH_ACK) &&
21265 (sbavail(sb) == 0) &&
21266 (rack->r_ctl.current_round != 0) &&
21267 ((flags & (TH_SYN|TH_FIN)) == 0) &&
21268 (rack->r_rcvpath_rtt_up == 0)) {
21269 rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
21270 rack->r_ctl.last_time_of_arm_rcv = cts;
21271 rack->r_rcvpath_rtt_up = 1;
21272 /* Subtract 1 from seq to force a response */
21273 rack_seq--;
21274 }
21275 }
21276 /* Set receive buffer autosizing timestamp. */
21277 if (tp->rfbuf_ts == 0 &&
21278 (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
21279 tp->rfbuf_ts = ms_cts;
21280 }
21281 /* Selective ACK's. */
21282 if (tp->t_flags & TF_SACK_PERMIT) {
21283 if (flags & TH_SYN)
21284 to.to_flags |= TOF_SACKPERM;
21285 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21286 tp->rcv_numsacks > 0) {
21287 to.to_flags |= TOF_SACK;
21288 to.to_nsacks = tp->rcv_numsacks;
21289 to.to_sacks = (u_char *)tp->sackblks;
21290 }
21291 }
21292 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21293 /* TCP-MD5 (RFC2385). */
21294 if (tp->t_flags & TF_SIGNATURE)
21295 to.to_flags |= TOF_SIGNATURE;
21296 #endif
21297
21298 /* Processing the options. */
21299 hdrlen += optlen = tcp_addoptions(&to, opt);
21300 /*
21301 * If we wanted a TFO option to be added, but it was unable
21302 * to fit, ensure no data is sent.
21303 */
21304 if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
21305 !(to.to_flags & TOF_FASTOPEN))
21306 len = 0;
21307 }
21308 if (tp->t_port) {
21309 if (V_tcp_udp_tunneling_port == 0) {
21310 /* The port was removed?? */
21311 SOCK_SENDBUF_UNLOCK(so);
21312 #ifdef TCP_ACCOUNTING
21313 crtsc = get_cyclecount();
21314 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21315 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21316 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21317 }
21318 sched_unpin();
21319 #endif
21320 return (EHOSTUNREACH);
21321 }
21322 hdrlen += sizeof(struct udphdr);
21323 }
21324 #ifdef INET6
21325 if (isipv6)
21326 ipoptlen = ip6_optlen(inp);
21327 else
21328 #endif
21329 if (inp->inp_options)
21330 ipoptlen = inp->inp_options->m_len -
21331 offsetof(struct ipoption, ipopt_list);
21332 else
21333 ipoptlen = 0;
21334 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21335 ipoptlen += ipsec_optlen;
21336 #endif
21337
21338 /*
21339 * Adjust data length if insertion of options will bump the packet
21340 * length beyond the t_maxseg length. Clear the FIN bit because we
21341 * cut off the tail of the segment.
21342 */
21343 if (len + optlen + ipoptlen > tp->t_maxseg) {
21344 if (tso) {
21345 uint32_t if_hw_tsomax;
21346 uint32_t moff;
21347 int32_t max_len;
21348
21349 /* extract TSO information */
21350 if_hw_tsomax = tp->t_tsomax;
21351 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21352 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21353 KASSERT(ipoptlen == 0,
21354 ("%s: TSO can't do IP options", __func__));
21355
21356 /*
21357 * Check if we should limit by maximum payload
21358 * length:
21359 */
21360 if (if_hw_tsomax != 0) {
21361 /* compute maximum TSO length */
21362 max_len = (if_hw_tsomax - hdrlen -
21363 max_linkhdr);
21364 if (max_len <= 0) {
21365 len = 0;
21366 } else if (len > max_len) {
21367 sendalot = 1;
21368 len = max_len;
21369 mark = 2;
21370 }
21371 }
21372 /*
21373 * Prevent the last segment from being fractional
21374 * unless the send sockbuf can be emptied:
21375 */
21376 max_len = (tp->t_maxseg - optlen);
21377 if ((sb_offset + len) < sbavail(sb)) {
21378 moff = len % (u_int)max_len;
21379 if (moff != 0) {
21380 mark = 3;
21381 len -= moff;
21382 }
21383 }
21384 /*
21385 * In case there are too many small fragments don't
21386 * use TSO:
21387 */
21388 if (len <= max_len) {
21389 mark = 4;
21390 tso = 0;
21391 }
21392 /*
21393 * Send the FIN in a separate segment after the bulk
21394 * sending is done. We don't trust the TSO
21395 * implementations to clear the FIN flag on all but
21396 * the last segment.
21397 */
21398 if (tp->t_flags & TF_NEEDFIN) {
21399 sendalot = 4;
21400 }
21401 } else {
21402 mark = 5;
21403 if (optlen + ipoptlen >= tp->t_maxseg) {
21404 /*
21405 * Since we don't have enough space to put
21406 * the IP header chain and the TCP header in
21407 * one packet as required by RFC 7112, don't
21408 * send it. Also ensure that at least one
21409 * byte of the payload can be put into the
21410 * TCP segment.
21411 */
21412 SOCK_SENDBUF_UNLOCK(so);
21413 error = EMSGSIZE;
21414 sack_rxmit = 0;
21415 goto out;
21416 }
21417 len = tp->t_maxseg - optlen - ipoptlen;
21418 sendalot = 5;
21419 }
21420 } else {
21421 tso = 0;
21422 mark = 6;
21423 }
21424 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21425 ("%s: len > IP_MAXPACKET", __func__));
21426 #ifdef DIAGNOSTIC
21427 #ifdef INET6
21428 if (max_linkhdr + hdrlen > MCLBYTES)
21429 #else
21430 if (max_linkhdr + hdrlen > MHLEN)
21431 #endif
21432 panic("tcphdr too big");
21433 #endif
21434
21435 /*
21436 * This KASSERT is here to catch edge cases at a well defined place.
21437 * Before, those had triggered (random) panic conditions further
21438 * down.
21439 */
21440 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21441 if ((len == 0) &&
21442 (flags & TH_FIN) &&
21443 (sbused(sb))) {
21444 /*
21445 * We have outstanding data, don't send a fin by itself!.
21446 *
21447 * Check to see if we need to send a challenge ack.
21448 */
21449 if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21450 ((tp->snd_max - tp->snd_una) <= segsiz)) {
21451 /*
21452 * Ok less than or right at a MSS is
21453 * outstanding. The original FreeBSD stack would
21454 * have sent a FIN, which can speed things up for
21455 * a transactional application doing a MSG_WAITALL.
21456 * To speed things up since we do *not* send a FIN
21457 * if data is outstanding, we send a "challenge ack".
21458 * The idea behind that is instead of having to have
21459 * the peer wait for the delayed-ack timer to run off
21460 * we send an ack that makes the peer send us an ack.
21461 */
21462 rack_send_ack_challange(rack);
21463 }
21464 goto just_return;
21465 }
21466 /*
21467 * Grab a header mbuf, attaching a copy of data to be transmitted,
21468 * and initialize the header from the template for sends on this
21469 * connection.
21470 */
21471 hw_tls = tp->t_nic_ktls_xmit != 0;
21472 if (len) {
21473 uint32_t max_val;
21474 uint32_t moff;
21475
21476 if (pace_max_seg)
21477 max_val = pace_max_seg;
21478 else
21479 max_val = len;
21480 /*
21481 * We allow a limit on sending with hptsi.
21482 */
21483 if (len > max_val) {
21484 mark = 7;
21485 len = max_val;
21486 }
21487 #ifdef INET6
21488 if (MHLEN < hdrlen + max_linkhdr)
21489 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21490 else
21491 #endif
21492 m = m_gethdr(M_NOWAIT, MT_DATA);
21493
21494 if (m == NULL) {
21495 SOCK_SENDBUF_UNLOCK(so);
21496 error = ENOBUFS;
21497 sack_rxmit = 0;
21498 goto out;
21499 }
21500 m->m_data += max_linkhdr;
21501 m->m_len = hdrlen;
21502
21503 /*
21504 * Start the m_copy functions from the closest mbuf to the
21505 * sb_offset in the socket buffer chain.
21506 */
21507 mb = sbsndptr_noadv(sb, sb_offset, &moff);
21508 s_mb = mb;
21509 s_moff = moff;
21510 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21511 m_copydata(mb, moff, (int)len,
21512 mtod(m, caddr_t)+hdrlen);
21513 /*
21514 * If we are not retransmitting advance the
21515 * sndptr to help remember the next place in
21516 * the sb.
21517 */
21518 if (rsm == NULL)
21519 sbsndptr_adv(sb, mb, len);
21520 m->m_len += len;
21521 } else {
21522 struct sockbuf *msb;
21523
21524 /*
21525 * If we are not retransmitting pass in msb so
21526 * the socket buffer can be advanced. Otherwise
21527 * set it to NULL if its a retransmission since
21528 * we don't want to change the sb remembered
21529 * location.
21530 */
21531 if (rsm == NULL)
21532 msb = sb;
21533 else
21534 msb = NULL;
21535 m->m_next = tcp_m_copym(
21536 mb, moff, &len,
21537 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21538 ((rsm == NULL) ? hw_tls : 0)
21539 #ifdef NETFLIX_COPY_ARGS
21540 , &s_mb, &s_moff
21541 #endif
21542 );
21543 if (len <= (tp->t_maxseg - optlen)) {
21544 /*
21545 * Must have ran out of mbufs for the copy
21546 * shorten it to no longer need tso. Lets
21547 * not put on sendalot since we are low on
21548 * mbufs.
21549 */
21550 tso = 0;
21551 }
21552 if (m->m_next == NULL) {
21553 SOCK_SENDBUF_UNLOCK(so);
21554 (void)m_free(m);
21555 error = ENOBUFS;
21556 sack_rxmit = 0;
21557 goto out;
21558 }
21559 }
21560 if (sack_rxmit) {
21561 if (rsm && (rsm->r_flags & RACK_TLP)) {
21562 /*
21563 * TLP should not count in retran count, but
21564 * in its own bin
21565 */
21566 counter_u64_add(rack_tlp_retran, 1);
21567 counter_u64_add(rack_tlp_retran_bytes, len);
21568 } else {
21569 tp->t_sndrexmitpack++;
21570 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21571 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21572 }
21573 #ifdef STATS
21574 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21575 len);
21576 #endif
21577 } else {
21578 KMOD_TCPSTAT_INC(tcps_sndpack);
21579 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21580 #ifdef STATS
21581 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21582 len);
21583 #endif
21584 }
21585 /*
21586 * If we're sending everything we've got, set PUSH. (This
21587 * will keep happy those implementations which only give
21588 * data to the user when a buffer fills or a PUSH comes in.)
21589 */
21590 if (sb_offset + len == sbused(sb) &&
21591 sbused(sb) &&
21592 !(flags & TH_SYN)) {
21593 flags |= TH_PUSH;
21594 add_flag |= RACK_HAD_PUSH;
21595 }
21596
21597 SOCK_SENDBUF_UNLOCK(so);
21598 } else {
21599 SOCK_SENDBUF_UNLOCK(so);
21600 if (tp->t_flags & TF_ACKNOW)
21601 KMOD_TCPSTAT_INC(tcps_sndacks);
21602 else if (flags & (TH_SYN | TH_FIN | TH_RST))
21603 KMOD_TCPSTAT_INC(tcps_sndctrl);
21604 else
21605 KMOD_TCPSTAT_INC(tcps_sndwinup);
21606
21607 m = m_gethdr(M_NOWAIT, MT_DATA);
21608 if (m == NULL) {
21609 error = ENOBUFS;
21610 sack_rxmit = 0;
21611 goto out;
21612 }
21613 #ifdef INET6
21614 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21615 MHLEN >= hdrlen) {
21616 M_ALIGN(m, hdrlen);
21617 } else
21618 #endif
21619 m->m_data += max_linkhdr;
21620 m->m_len = hdrlen;
21621 }
21622 SOCK_SENDBUF_UNLOCK_ASSERT(so);
21623 m->m_pkthdr.rcvif = (struct ifnet *)0;
21624 #ifdef MAC
21625 mac_inpcb_create_mbuf(inp, m);
21626 #endif
21627 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21628 #ifdef INET6
21629 if (isipv6)
21630 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21631 else
21632 #endif /* INET6 */
21633 #ifdef INET
21634 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21635 #endif
21636 th = rack->r_ctl.fsb.th;
21637 udp = rack->r_ctl.fsb.udp;
21638 if (udp) {
21639 #ifdef INET6
21640 if (isipv6)
21641 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21642 else
21643 #endif /* INET6 */
21644 ulen = hdrlen + len - sizeof(struct ip);
21645 udp->uh_ulen = htons(ulen);
21646 }
21647 } else {
21648 #ifdef INET6
21649 if (isipv6) {
21650 ip6 = mtod(m, struct ip6_hdr *);
21651 if (tp->t_port) {
21652 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21653 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21654 udp->uh_dport = tp->t_port;
21655 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21656 udp->uh_ulen = htons(ulen);
21657 th = (struct tcphdr *)(udp + 1);
21658 } else
21659 th = (struct tcphdr *)(ip6 + 1);
21660 tcpip_fillheaders(inp, tp->t_port, ip6, th);
21661 } else
21662 #endif /* INET6 */
21663 {
21664 #ifdef INET
21665 ip = mtod(m, struct ip *);
21666 if (tp->t_port) {
21667 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21668 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21669 udp->uh_dport = tp->t_port;
21670 ulen = hdrlen + len - sizeof(struct ip);
21671 udp->uh_ulen = htons(ulen);
21672 th = (struct tcphdr *)(udp + 1);
21673 } else
21674 th = (struct tcphdr *)(ip + 1);
21675 tcpip_fillheaders(inp, tp->t_port, ip, th);
21676 #endif
21677 }
21678 }
21679 /*
21680 * If we are starting a connection, send ECN setup SYN packet. If we
21681 * are on a retransmit, we may resend those bits a number of times
21682 * as per RFC 3168.
21683 */
21684 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21685 flags |= tcp_ecn_output_syn_sent(tp);
21686 }
21687 /* Also handle parallel SYN for ECN */
21688 if (TCPS_HAVERCVDSYN(tp->t_state) &&
21689 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21690 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21691 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21692 (tp->t_flags2 & TF2_ECN_SND_ECE))
21693 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21694 #ifdef INET6
21695 if (isipv6) {
21696 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21697 ip6->ip6_flow |= htonl(ect << 20);
21698 }
21699 else
21700 #endif
21701 {
21702 #ifdef INET
21703 ip->ip_tos &= ~IPTOS_ECN_MASK;
21704 ip->ip_tos |= ect;
21705 #endif
21706 }
21707 }
21708 th->th_seq = htonl(rack_seq);
21709 th->th_ack = htonl(tp->rcv_nxt);
21710 tcp_set_flags(th, flags);
21711 /*
21712 * Calculate receive window. Don't shrink window, but avoid silly
21713 * window syndrome.
21714 * If a RST segment is sent, advertise a window of zero.
21715 */
21716 if (flags & TH_RST) {
21717 recwin = 0;
21718 } else {
21719 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21720 recwin < (long)segsiz) {
21721 recwin = 0;
21722 }
21723 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21724 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21725 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21726 }
21727
21728 /*
21729 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21730 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is
21731 * handled in syncache.
21732 */
21733 if (flags & TH_SYN)
21734 th->th_win = htons((u_short)
21735 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21736 else {
21737 /* Avoid shrinking window with window scaling. */
21738 recwin = roundup2(recwin, 1 << tp->rcv_scale);
21739 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21740 }
21741 /*
21742 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21743 * window. This may cause the remote transmitter to stall. This
21744 * flag tells soreceive() to disable delayed acknowledgements when
21745 * draining the buffer. This can occur if the receiver is
21746 * attempting to read more data than can be buffered prior to
21747 * transmitting on the connection.
21748 */
21749 if (th->th_win == 0) {
21750 tp->t_sndzerowin++;
21751 tp->t_flags |= TF_RXWIN0SENT;
21752 } else
21753 tp->t_flags &= ~TF_RXWIN0SENT;
21754 tp->snd_up = tp->snd_una; /* drag it along, its deprecated */
21755 /* Now are we using fsb?, if so copy the template data to the mbuf */
21756 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21757 uint8_t *cpto;
21758
21759 cpto = mtod(m, uint8_t *);
21760 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21761 /*
21762 * We have just copied in:
21763 * IP/IP6
21764 * <optional udphdr>
21765 * tcphdr (no options)
21766 *
21767 * We need to grab the correct pointers into the mbuf
21768 * for both the tcp header, and possibly the udp header (if tunneling).
21769 * We do this by using the offset in the copy buffer and adding it
21770 * to the mbuf base pointer (cpto).
21771 */
21772 #ifdef INET6
21773 if (isipv6)
21774 ip6 = mtod(m, struct ip6_hdr *);
21775 else
21776 #endif /* INET6 */
21777 #ifdef INET
21778 ip = mtod(m, struct ip *);
21779 #endif
21780 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21781 /* If we have a udp header lets set it into the mbuf as well */
21782 if (udp)
21783 udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21784 }
21785 if (optlen) {
21786 bcopy(opt, th + 1, optlen);
21787 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21788 }
21789 /*
21790 * Put TCP length in extended header, and then checksum extended
21791 * header and data.
21792 */
21793 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
21794 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21795 if (to.to_flags & TOF_SIGNATURE) {
21796 /*
21797 * Calculate MD5 signature and put it into the place
21798 * determined before.
21799 * NOTE: since TCP options buffer doesn't point into
21800 * mbuf's data, calculate offset and use it.
21801 */
21802 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21803 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21804 /*
21805 * Do not send segment if the calculation of MD5
21806 * digest has failed.
21807 */
21808 goto out;
21809 }
21810 }
21811 #endif
21812 #ifdef INET6
21813 if (isipv6) {
21814 /*
21815 * ip6_plen is not need to be filled now, and will be filled
21816 * in ip6_output.
21817 */
21818 if (tp->t_port) {
21819 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21820 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21821 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21822 th->th_sum = htons(0);
21823 UDPSTAT_INC(udps_opackets);
21824 } else {
21825 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21826 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21827 th->th_sum = in6_cksum_pseudo(ip6,
21828 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21829 0);
21830 }
21831 }
21832 #endif
21833 #if defined(INET6) && defined(INET)
21834 else
21835 #endif
21836 #ifdef INET
21837 {
21838 if (tp->t_port) {
21839 m->m_pkthdr.csum_flags = CSUM_UDP;
21840 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21841 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21842 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21843 th->th_sum = htons(0);
21844 UDPSTAT_INC(udps_opackets);
21845 } else {
21846 m->m_pkthdr.csum_flags = CSUM_TCP;
21847 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21848 th->th_sum = in_pseudo(ip->ip_src.s_addr,
21849 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21850 IPPROTO_TCP + len + optlen));
21851 }
21852 /* IP version must be set here for ipv4/ipv6 checking later */
21853 KASSERT(ip->ip_v == IPVERSION,
21854 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
21855 }
21856 #endif
21857 /*
21858 * Enable TSO and specify the size of the segments. The TCP pseudo
21859 * header checksum is always provided. XXX: Fixme: This is currently
21860 * not the case for IPv6.
21861 */
21862 if (tso) {
21863 /*
21864 * Here we must use t_maxseg and the optlen since
21865 * the optlen may include SACK's (or DSACK).
21866 */
21867 KASSERT(len > tp->t_maxseg - optlen,
21868 ("%s: len <= tso_segsz", __func__));
21869 m->m_pkthdr.csum_flags |= CSUM_TSO;
21870 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21871 }
21872 KASSERT(len + hdrlen == m_length(m, NULL),
21873 ("%s: mbuf chain different than expected: %d + %u != %u",
21874 __func__, len, hdrlen, m_length(m, NULL)));
21875
21876 #ifdef TCP_HHOOK
21877 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21878 hhook_run_tcp_est_out(tp, th, &to, len, tso);
21879 #endif
21880 if ((rack->r_ctl.crte != NULL) &&
21881 (rack->rc_hw_nobuf == 0) &&
21882 tcp_bblogging_on(tp)) {
21883 rack_log_queue_level(tp, rack, len, &tv, cts);
21884 }
21885 /* We're getting ready to send; log now. */
21886 if (tcp_bblogging_on(rack->rc_tp)) {
21887 union tcp_log_stackspecific log;
21888
21889 memset(&log, 0, sizeof(log));
21890 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21891 if (rack->rack_no_prr)
21892 log.u_bbr.flex1 = 0;
21893 else
21894 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21895 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21896 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21897 log.u_bbr.flex4 = orig_len;
21898 /* Save off the early/late values */
21899 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21900 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21901 log.u_bbr.bw_inuse = rack_get_bw(rack);
21902 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21903 log.u_bbr.flex8 = 0;
21904 if (rsm) {
21905 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21906 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21907 counter_u64_add(rack_collapsed_win_rxt, 1);
21908 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21909 }
21910 if (doing_tlp)
21911 log.u_bbr.flex8 = 2;
21912 else
21913 log.u_bbr.flex8 = 1;
21914 } else {
21915 if (doing_tlp)
21916 log.u_bbr.flex8 = 3;
21917 }
21918 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21919 log.u_bbr.flex7 = mark;
21920 log.u_bbr.flex7 <<= 8;
21921 log.u_bbr.flex7 |= pass;
21922 log.u_bbr.pkts_out = tp->t_maxseg;
21923 log.u_bbr.timeStamp = cts;
21924 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21925 if (rsm && (rsm->r_rtr_cnt > 0)) {
21926 /*
21927 * When we have a retransmit we want to log the
21928 * burst at send and flight at send from before.
21929 */
21930 log.u_bbr.flex5 = rsm->r_fas;
21931 log.u_bbr.bbr_substate = rsm->r_bas;
21932 } else {
21933 /*
21934 * New transmits we log in flex5 the inflight again as
21935 * well as the number of segments in our send in the
21936 * substate field.
21937 */
21938 log.u_bbr.flex5 = log.u_bbr.inflight;
21939 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21940 }
21941 log.u_bbr.lt_epoch = cwnd_to_use;
21942 log.u_bbr.delivered = sendalot;
21943 log.u_bbr.rttProp = (uintptr_t)rsm;
21944 log.u_bbr.pkt_epoch = __LINE__;
21945 if (rsm) {
21946 log.u_bbr.delRate = rsm->r_flags;
21947 log.u_bbr.delRate <<= 31;
21948 log.u_bbr.delRate |= rack->r_must_retran;
21949 log.u_bbr.delRate <<= 1;
21950 log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21951 } else {
21952 log.u_bbr.delRate = rack->r_must_retran;
21953 log.u_bbr.delRate <<= 1;
21954 log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21955 }
21956 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
21957 len, &log, false, NULL, __func__, __LINE__, &tv);
21958 } else
21959 lgb = NULL;
21960
21961 /*
21962 * Fill in IP length and desired time to live and send to IP level.
21963 * There should be a better way to handle ttl and tos; we could keep
21964 * them in the template, but need a way to checksum without them.
21965 */
21966 /*
21967 * m->m_pkthdr.len should have been set before cksum calcuration,
21968 * because in6_cksum() need it.
21969 */
21970 #ifdef INET6
21971 if (isipv6) {
21972 /*
21973 * we separately set hoplimit for every segment, since the
21974 * user might want to change the value via setsockopt. Also,
21975 * desired default hop limit might be changed via Neighbor
21976 * Discovery.
21977 */
21978 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
21979
21980 /*
21981 * Set the packet size here for the benefit of DTrace
21982 * probes. ip6_output() will set it properly; it's supposed
21983 * to include the option header lengths as well.
21984 */
21985 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
21986
21987 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
21988 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
21989 else
21990 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
21991
21992 if (tp->t_state == TCPS_SYN_SENT)
21993 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
21994
21995 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
21996 /* TODO: IPv6 IP6TOS_ECT bit on */
21997 error = ip6_output(m,
21998 inp->in6p_outputopts,
21999 &inp->inp_route6,
22000 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22001 NULL, NULL, inp);
22002
22003 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22004 mtu = inp->inp_route6.ro_nh->nh_mtu;
22005 }
22006 #endif /* INET6 */
22007 #if defined(INET) && defined(INET6)
22008 else
22009 #endif
22010 #ifdef INET
22011 {
22012 ip->ip_len = htons(m->m_pkthdr.len);
22013 #ifdef INET6
22014 if (inp->inp_vflag & INP_IPV6PROTO)
22015 ip->ip_ttl = in6_selecthlim(inp, NULL);
22016 #endif /* INET6 */
22017 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22018 /*
22019 * If we do path MTU discovery, then we set DF on every
22020 * packet. This might not be the best thing to do according
22021 * to RFC3390 Section 2. However the tcp hostcache migitates
22022 * the problem so it affects only the first tcp connection
22023 * with a host.
22024 *
22025 * NB: Don't set DF on small MTU/MSS to have a safe
22026 * fallback.
22027 */
22028 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22029 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22030 if (tp->t_port == 0 || len < V_tcp_minmss) {
22031 ip->ip_off |= htons(IP_DF);
22032 }
22033 } else {
22034 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22035 }
22036
22037 if (tp->t_state == TCPS_SYN_SENT)
22038 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22039
22040 TCP_PROBE5(send, NULL, tp, ip, tp, th);
22041
22042 error = ip_output(m,
22043 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22044 inp->inp_options,
22045 #else
22046 NULL,
22047 #endif
22048 &inp->inp_route,
22049 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22050 inp);
22051 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22052 mtu = inp->inp_route.ro_nh->nh_mtu;
22053 }
22054 #endif /* INET */
22055 if (lgb) {
22056 lgb->tlb_errno = error;
22057 lgb = NULL;
22058 }
22059
22060 out:
22061 /*
22062 * In transmit state, time the transmission and arrange for the
22063 * retransmit. In persist state, just set snd_max.
22064 */
22065 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22066 rack_to_usec_ts(&tv),
22067 rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22068 if (error == 0) {
22069 if (add_flag & RACK_IS_PCM) {
22070 /* We just launched a PCM */
22071 /* rrs here log */
22072 rack->pcm_in_progress = 1;
22073 rack->pcm_needed = 0;
22074 rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg, add_flag);
22075 }
22076 if (rsm == NULL) {
22077 if (rack->lt_bw_up == 0) {
22078 rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
22079 rack->r_ctl.lt_seq = tp->snd_una;
22080 rack->lt_bw_up = 1;
22081 } else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22082 /*
22083 * Need to record what we have since we are
22084 * approaching seq wrap.
22085 */
22086 uint64_t tmark;
22087
22088 rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22089 rack->r_ctl.lt_seq = tp->snd_una;
22090 tmark = tcp_get_u64_usecs(&tv);
22091 if (tmark > rack->r_ctl.lt_timemark) {
22092 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22093 rack->r_ctl.lt_timemark = tmark;
22094 }
22095 }
22096 }
22097 rack->forced_ack = 0; /* If we send something zap the FA flag */
22098 counter_u64_add(rack_total_bytes, len);
22099 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22100 if (rsm && doing_tlp) {
22101 rack->rc_last_sent_tlp_past_cumack = 0;
22102 rack->rc_last_sent_tlp_seq_valid = 1;
22103 rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22104 rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22105 }
22106 if (rack->rc_hw_nobuf) {
22107 rack->rc_hw_nobuf = 0;
22108 rack->r_ctl.rc_agg_delayed = 0;
22109 rack->r_early = 0;
22110 rack->r_late = 0;
22111 rack->r_ctl.rc_agg_early = 0;
22112 }
22113 if (rsm && (doing_tlp == 0)) {
22114 /* Set we retransmitted */
22115 rack->rc_gp_saw_rec = 1;
22116 } else {
22117 if (cwnd_to_use > tp->snd_ssthresh) {
22118 /* Set we sent in CA */
22119 rack->rc_gp_saw_ca = 1;
22120 } else {
22121 /* Set we sent in SS */
22122 rack->rc_gp_saw_ss = 1;
22123 }
22124 }
22125 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22126 (tp->t_flags & TF_SACK_PERMIT) &&
22127 tp->rcv_numsacks > 0)
22128 tcp_clean_dsack_blocks(tp);
22129 tot_len_this_send += len;
22130 if (len == 0) {
22131 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22132 } else {
22133 int idx;
22134
22135 idx = (len / segsiz) + 3;
22136 if (idx >= TCP_MSS_ACCT_ATIMER)
22137 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22138 else
22139 counter_u64_add(rack_out_size[idx], 1);
22140 }
22141 }
22142 if ((rack->rack_no_prr == 0) &&
22143 sub_from_prr &&
22144 (error == 0)) {
22145 if (rack->r_ctl.rc_prr_sndcnt >= len)
22146 rack->r_ctl.rc_prr_sndcnt -= len;
22147 else
22148 rack->r_ctl.rc_prr_sndcnt = 0;
22149 }
22150 sub_from_prr = 0;
22151 if (doing_tlp) {
22152 /* Make sure the TLP is added */
22153 add_flag |= RACK_TLP;
22154 } else if (rsm) {
22155 /* If its a resend without TLP then it must not have the flag */
22156 rsm->r_flags &= ~RACK_TLP;
22157 }
22158
22159
22160 if ((error == 0) &&
22161 (len > 0) &&
22162 (tp->snd_una == tp->snd_max))
22163 rack->r_ctl.rc_tlp_rxt_last_time = cts;
22164
22165 {
22166 /*
22167 * This block is not associated with the above error == 0 test.
22168 * It is used to advance snd_max if we have a new transmit.
22169 */
22170 tcp_seq startseq = tp->snd_max;
22171
22172
22173 if (rsm && (doing_tlp == 0))
22174 rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22175 if (error)
22176 /* We don't log or do anything with errors */
22177 goto nomore;
22178 if (doing_tlp == 0) {
22179 if (rsm == NULL) {
22180 /*
22181 * Not a retransmission of some
22182 * sort, new data is going out so
22183 * clear our TLP count and flag.
22184 */
22185 rack->rc_tlp_in_progress = 0;
22186 rack->r_ctl.rc_tlp_cnt_out = 0;
22187 }
22188 } else {
22189 /*
22190 * We have just sent a TLP, mark that it is true
22191 * and make sure our in progress is set so we
22192 * continue to check the count.
22193 */
22194 rack->rc_tlp_in_progress = 1;
22195 rack->r_ctl.rc_tlp_cnt_out++;
22196 }
22197 /*
22198 * If we are retransmitting we are done, snd_max
22199 * does not get updated.
22200 */
22201 if (sack_rxmit)
22202 goto nomore;
22203 if ((tp->snd_una == tp->snd_max) && (len > 0)) {
22204 /*
22205 * Update the time we just added data since
22206 * nothing was outstanding.
22207 */
22208 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22209 tp->t_acktime = ticks;
22210 }
22211 /*
22212 * Now for special SYN/FIN handling.
22213 */
22214 if (flags & (TH_SYN | TH_FIN)) {
22215 if ((flags & TH_SYN) &&
22216 ((tp->t_flags & TF_SENTSYN) == 0)) {
22217 tp->snd_max++;
22218 tp->t_flags |= TF_SENTSYN;
22219 }
22220 if ((flags & TH_FIN) &&
22221 ((tp->t_flags & TF_SENTFIN) == 0)) {
22222 tp->snd_max++;
22223 tp->t_flags |= TF_SENTFIN;
22224 }
22225 }
22226 tp->snd_max += len;
22227 if (rack->rc_new_rnd_needed) {
22228 rack_new_round_starts(tp, rack, tp->snd_max);
22229 }
22230 /*
22231 * Time this transmission if not a retransmission and
22232 * not currently timing anything.
22233 * This is only relevant in case of switching back to
22234 * the base stack.
22235 */
22236 if (tp->t_rtttime == 0) {
22237 tp->t_rtttime = ticks;
22238 tp->t_rtseq = startseq;
22239 KMOD_TCPSTAT_INC(tcps_segstimed);
22240 }
22241 if (len &&
22242 ((tp->t_flags & TF_GPUTINPROG) == 0))
22243 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22244 /*
22245 * If we are doing FO we need to update the mbuf position and subtract
22246 * this happens when the peer sends us duplicate information and
22247 * we thus want to send a DSACK.
22248 *
22249 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22250 * turned off? If not then we are going to echo multiple DSACK blocks
22251 * out (with the TSO), which we should not be doing.
22252 */
22253 if (rack->r_fast_output && len) {
22254 if (rack->r_ctl.fsb.left_to_send > len)
22255 rack->r_ctl.fsb.left_to_send -= len;
22256 else
22257 rack->r_ctl.fsb.left_to_send = 0;
22258 if (rack->r_ctl.fsb.left_to_send < segsiz)
22259 rack->r_fast_output = 0;
22260 if (rack->r_fast_output) {
22261 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22262 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22263 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22264 }
22265 }
22266 if (rack_pcm_blast == 0) {
22267 if ((orig_len > len) &&
22268 (add_flag & RACK_IS_PCM) &&
22269 (len < pace_max_seg) &&
22270 ((pace_max_seg - len) > segsiz)) {
22271 /*
22272 * We are doing a PCM measurement and we did
22273 * not get enough data in the TSO to meet the
22274 * burst requirement.
22275 */
22276 uint32_t n_len;
22277
22278 n_len = (orig_len - len);
22279 orig_len -= len;
22280 pace_max_seg -= len;
22281 len = n_len;
22282 sb_offset = tp->snd_max - tp->snd_una;
22283 /* Re-lock for the next spin */
22284 SOCK_SENDBUF_LOCK(so);
22285 goto send;
22286 }
22287 } else {
22288 if ((orig_len > len) &&
22289 (add_flag & RACK_IS_PCM) &&
22290 ((orig_len - len) > segsiz)) {
22291 /*
22292 * We are doing a PCM measurement and we did
22293 * not get enough data in the TSO to meet the
22294 * burst requirement.
22295 */
22296 uint32_t n_len;
22297
22298 n_len = (orig_len - len);
22299 orig_len -= len;
22300 len = n_len;
22301 sb_offset = tp->snd_max - tp->snd_una;
22302 /* Re-lock for the next spin */
22303 SOCK_SENDBUF_LOCK(so);
22304 goto send;
22305 }
22306 }
22307 }
22308 nomore:
22309 if (error) {
22310 rack->r_ctl.rc_agg_delayed = 0;
22311 rack->r_early = 0;
22312 rack->r_late = 0;
22313 rack->r_ctl.rc_agg_early = 0;
22314 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */
22315 /*
22316 * Failures do not advance the seq counter above. For the
22317 * case of ENOBUFS we will fall out and retry in 1ms with
22318 * the hpts. Everything else will just have to retransmit
22319 * with the timer.
22320 *
22321 * In any case, we do not want to loop around for another
22322 * send without a good reason.
22323 */
22324 sendalot = 0;
22325 switch (error) {
22326 case EPERM:
22327 case EACCES:
22328 tp->t_softerror = error;
22329 #ifdef TCP_ACCOUNTING
22330 crtsc = get_cyclecount();
22331 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22332 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22333 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22334 }
22335 sched_unpin();
22336 #endif
22337 return (error);
22338 case ENOBUFS:
22339 /*
22340 * Pace us right away to retry in a some
22341 * time
22342 */
22343 if (rack->r_ctl.crte != NULL) {
22344 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22345 if (tcp_bblogging_on(rack->rc_tp))
22346 rack_log_queue_level(tp, rack, len, &tv, cts);
22347 } else
22348 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22349 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22350 if (rack->rc_enobuf < 0x7f)
22351 rack->rc_enobuf++;
22352 if (slot < (10 * HPTS_USEC_IN_MSEC))
22353 slot = 10 * HPTS_USEC_IN_MSEC;
22354 if (rack->r_ctl.crte != NULL) {
22355 counter_u64_add(rack_saw_enobuf_hw, 1);
22356 tcp_rl_log_enobuf(rack->r_ctl.crte);
22357 }
22358 counter_u64_add(rack_saw_enobuf, 1);
22359 goto enobufs;
22360 case EMSGSIZE:
22361 /*
22362 * For some reason the interface we used initially
22363 * to send segments changed to another or lowered
22364 * its MTU. If TSO was active we either got an
22365 * interface without TSO capabilits or TSO was
22366 * turned off. If we obtained mtu from ip_output()
22367 * then update it and try again.
22368 */
22369 if (tso)
22370 tp->t_flags &= ~TF_TSO;
22371 if (mtu != 0) {
22372 int saved_mtu;
22373
22374 saved_mtu = tp->t_maxseg;
22375 tcp_mss_update(tp, -1, mtu, NULL, NULL);
22376 if (saved_mtu > tp->t_maxseg) {
22377 goto again;
22378 }
22379 }
22380 slot = 10 * HPTS_USEC_IN_MSEC;
22381 rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22382 #ifdef TCP_ACCOUNTING
22383 crtsc = get_cyclecount();
22384 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22385 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22386 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22387 }
22388 sched_unpin();
22389 #endif
22390 return (error);
22391 case ENETUNREACH:
22392 counter_u64_add(rack_saw_enetunreach, 1);
22393 /* FALLTHROUGH */
22394 case EHOSTDOWN:
22395 case EHOSTUNREACH:
22396 case ENETDOWN:
22397 if (TCPS_HAVERCVDSYN(tp->t_state)) {
22398 tp->t_softerror = error;
22399 error = 0;
22400 }
22401 /* FALLTHROUGH */
22402 default:
22403 slot = 10 * HPTS_USEC_IN_MSEC;
22404 rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22405 #ifdef TCP_ACCOUNTING
22406 crtsc = get_cyclecount();
22407 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22408 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22409 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22410 }
22411 sched_unpin();
22412 #endif
22413 return (error);
22414 }
22415 } else {
22416 rack->rc_enobuf = 0;
22417 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22418 rack->r_ctl.retran_during_recovery += len;
22419 }
22420 KMOD_TCPSTAT_INC(tcps_sndtotal);
22421
22422 /*
22423 * Data sent (as far as we can tell). If this advertises a larger
22424 * window than any other segment, then remember the size of the
22425 * advertised window. Any pending ACK has now been sent.
22426 */
22427 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22428 tp->rcv_adv = tp->rcv_nxt + recwin;
22429
22430 tp->last_ack_sent = tp->rcv_nxt;
22431 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22432 enobufs:
22433 if (sendalot) {
22434 /* Do we need to turn off sendalot? */
22435 if (pace_max_seg &&
22436 (tot_len_this_send >= pace_max_seg)) {
22437 /* We hit our max. */
22438 sendalot = 0;
22439 }
22440 }
22441 if ((error == 0) && (flags & TH_FIN))
22442 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22443 if (flags & TH_RST) {
22444 /*
22445 * We don't send again after sending a RST.
22446 */
22447 slot = 0;
22448 sendalot = 0;
22449 if (error == 0)
22450 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22451 } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22452 /*
22453 * Get our pacing rate, if an error
22454 * occurred in sending (ENOBUF) we would
22455 * hit the else if with slot preset. Other
22456 * errors return.
22457 */
22458 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
22459 }
22460 /* We have sent clear the flag */
22461 rack->r_ent_rec_ns = 0;
22462 if (rack->r_must_retran) {
22463 if (rsm) {
22464 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22465 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22466 /*
22467 * We have retransmitted all.
22468 */
22469 rack->r_must_retran = 0;
22470 rack->r_ctl.rc_out_at_rto = 0;
22471 }
22472 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22473 /*
22474 * Sending new data will also kill
22475 * the loop.
22476 */
22477 rack->r_must_retran = 0;
22478 rack->r_ctl.rc_out_at_rto = 0;
22479 }
22480 }
22481 rack->r_ctl.fsb.recwin = recwin;
22482 if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22483 SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22484 /*
22485 * We hit an RTO and now have past snd_max at the RTO
22486 * clear all the WAS flags.
22487 */
22488 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22489 }
22490 if (slot) {
22491 /* set the rack tcb into the slot N */
22492 if ((error == 0) &&
22493 rack_use_rfo &&
22494 ((flags & (TH_SYN|TH_FIN)) == 0) &&
22495 (rsm == NULL) &&
22496 (ipoptlen == 0) &&
22497 rack->r_fsb_inited &&
22498 TCPS_HAVEESTABLISHED(tp->t_state) &&
22499 ((IN_RECOVERY(tp->t_flags)) == 0) &&
22500 (rack->r_must_retran == 0) &&
22501 ((tp->t_flags & TF_NEEDFIN) == 0) &&
22502 (len > 0) && (orig_len > 0) &&
22503 (orig_len > len) &&
22504 ((orig_len - len) >= segsiz) &&
22505 ((optlen == 0) ||
22506 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22507 /* We can send at least one more MSS using our fsb */
22508 rack_setup_fast_output(tp, rack, sb, len, orig_len,
22509 segsiz, pace_max_seg, hw_tls, flags);
22510 } else
22511 rack->r_fast_output = 0;
22512 rack_log_fsb(rack, tp, so, flags,
22513 ipoptlen, orig_len, len, error,
22514 (rsm == NULL), optlen, __LINE__, 2);
22515 } else if (sendalot) {
22516 int ret;
22517
22518 sack_rxmit = 0;
22519 if ((error == 0) &&
22520 rack_use_rfo &&
22521 ((flags & (TH_SYN|TH_FIN)) == 0) &&
22522 (rsm == NULL) &&
22523 (ipoptlen == 0) &&
22524 (rack->r_must_retran == 0) &&
22525 rack->r_fsb_inited &&
22526 TCPS_HAVEESTABLISHED(tp->t_state) &&
22527 ((IN_RECOVERY(tp->t_flags)) == 0) &&
22528 ((tp->t_flags & TF_NEEDFIN) == 0) &&
22529 (len > 0) && (orig_len > 0) &&
22530 (orig_len > len) &&
22531 ((orig_len - len) >= segsiz) &&
22532 ((optlen == 0) ||
22533 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22534 /* we can use fast_output for more */
22535 rack_setup_fast_output(tp, rack, sb, len, orig_len,
22536 segsiz, pace_max_seg, hw_tls, flags);
22537 if (rack->r_fast_output) {
22538 error = 0;
22539 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22540 if (ret >= 0)
22541 return (ret);
22542 else if (error)
22543 goto nomore;
22544
22545 }
22546 }
22547 goto again;
22548 }
22549 skip_all_send:
22550 /* Assure when we leave that snd_nxt will point to top */
22551 if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22552 tp->snd_nxt = tp->snd_max;
22553 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22554 #ifdef TCP_ACCOUNTING
22555 crtsc = get_cyclecount() - ts_val;
22556 if (tot_len_this_send) {
22557 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22558 tp->tcp_cnt_counters[SND_OUT_DATA]++;
22559 tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22560 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22561 }
22562 } else {
22563 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22564 tp->tcp_cnt_counters[SND_OUT_ACK]++;
22565 tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22566 }
22567 }
22568 sched_unpin();
22569 #endif
22570 if (error == ENOBUFS)
22571 error = 0;
22572 return (error);
22573 }
22574
22575 static void
rack_update_seg(struct tcp_rack * rack)22576 rack_update_seg(struct tcp_rack *rack)
22577 {
22578 uint32_t orig_val;
22579
22580 orig_val = rack->r_ctl.rc_pace_max_segs;
22581 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22582 if (orig_val != rack->r_ctl.rc_pace_max_segs)
22583 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22584 }
22585
22586 static void
rack_mtu_change(struct tcpcb * tp)22587 rack_mtu_change(struct tcpcb *tp)
22588 {
22589 /*
22590 * The MSS may have changed
22591 */
22592 struct tcp_rack *rack;
22593 struct rack_sendmap *rsm;
22594
22595 rack = (struct tcp_rack *)tp->t_fb_ptr;
22596 if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22597 /*
22598 * The MTU has changed we need to resend everything
22599 * since all we have sent is lost. We first fix
22600 * up the mtu though.
22601 */
22602 rack_set_pace_segments(tp, rack, __LINE__, NULL);
22603 /* We treat this like a full retransmit timeout without the cwnd adjustment */
22604 rack_remxt_tmr(tp);
22605 rack->r_fast_output = 0;
22606 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22607 rack->r_ctl.rc_sacked);
22608 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22609 rack->r_must_retran = 1;
22610 /* Mark all inflight to needing to be rxt'd */
22611 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22612 rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22613 }
22614 }
22615 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22616 /* We don't use snd_nxt to retransmit */
22617 tp->snd_nxt = tp->snd_max;
22618 }
22619
22620 static int
rack_set_dgp(struct tcp_rack * rack)22621 rack_set_dgp(struct tcp_rack *rack)
22622 {
22623 if (rack->dgp_on == 1)
22624 return(0);
22625 if ((rack->use_fixed_rate == 1) &&
22626 (rack->rc_always_pace == 1)) {
22627 /*
22628 * We are already pacing another
22629 * way.
22630 */
22631 return (EBUSY);
22632 }
22633 if (rack->rc_always_pace == 1) {
22634 rack_remove_pacing(rack);
22635 }
22636 if (tcp_incr_dgp_pacing_cnt() == 0)
22637 return (ENOSPC);
22638 rack->r_ctl.pacing_method |= RACK_DGP_PACING;
22639 rack->rc_fillcw_apply_discount = 0;
22640 rack->dgp_on = 1;
22641 rack->rc_always_pace = 1;
22642 rack->rc_pace_dnd = 1;
22643 rack->use_fixed_rate = 0;
22644 if (rack->gp_ready)
22645 rack_set_cc_pacing(rack);
22646 rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22647 rack->rack_attempt_hdwr_pace = 0;
22648 /* rxt settings */
22649 rack->full_size_rxt = 1;
22650 rack->shape_rxt_to_pacing_min = 0;
22651 /* cmpack=1 */
22652 rack->r_use_cmp_ack = 1;
22653 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22654 rack->r_use_cmp_ack)
22655 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22656 /* scwnd=1 */
22657 rack->rack_enable_scwnd = 1;
22658 /* dynamic=100 */
22659 rack->rc_gp_dyn_mul = 1;
22660 /* gp_inc_ca */
22661 rack->r_ctl.rack_per_of_gp_ca = 100;
22662 /* rrr_conf=3 */
22663 rack->r_rr_config = 3;
22664 /* npush=2 */
22665 rack->r_ctl.rc_no_push_at_mrtt = 2;
22666 /* fillcw=1 */
22667 rack->rc_pace_to_cwnd = 1;
22668 rack->rc_pace_fill_if_rttin_range = 0;
22669 rack->rtt_limit_mul = 0;
22670 /* noprr=1 */
22671 rack->rack_no_prr = 1;
22672 /* lscwnd=1 */
22673 rack->r_limit_scw = 1;
22674 /* gp_inc_rec */
22675 rack->r_ctl.rack_per_of_gp_rec = 90;
22676 return (0);
22677 }
22678
22679 static int
rack_set_profile(struct tcp_rack * rack,int prof)22680 rack_set_profile(struct tcp_rack *rack, int prof)
22681 {
22682 int err = EINVAL;
22683 if (prof == 1) {
22684 /*
22685 * Profile 1 is "standard" DGP. It ignores
22686 * client buffer level.
22687 */
22688 err = rack_set_dgp(rack);
22689 if (err)
22690 return (err);
22691 } else if (prof == 6) {
22692 err = rack_set_dgp(rack);
22693 if (err)
22694 return (err);
22695 /*
22696 * Profile 6 tweaks DGP so that it will apply to
22697 * fill-cw the same settings that profile5 does
22698 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22699 */
22700 rack->rc_fillcw_apply_discount = 1;
22701 } else if (prof == 0) {
22702 /* This changes things back to the default settings */
22703 if (rack->rc_always_pace == 1) {
22704 rack_remove_pacing(rack);
22705 } else {
22706 /* Make sure any stray flags are off */
22707 rack->dgp_on = 0;
22708 rack->rc_hybrid_mode = 0;
22709 rack->use_fixed_rate = 0;
22710 }
22711 err = 0;
22712 if (rack_fill_cw_state)
22713 rack->rc_pace_to_cwnd = 1;
22714 else
22715 rack->rc_pace_to_cwnd = 0;
22716
22717 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22718 rack->r_ctl.pacing_method |= RACK_REG_PACING;
22719 rack->rc_always_pace = 1;
22720 if (rack->rack_hibeta)
22721 rack_set_cc_pacing(rack);
22722 } else
22723 rack->rc_always_pace = 0;
22724 if (rack_dsack_std_based & 0x1) {
22725 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22726 rack->rc_rack_tmr_std_based = 1;
22727 }
22728 if (rack_dsack_std_based & 0x2) {
22729 /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */
22730 rack->rc_rack_use_dsack = 1;
22731 }
22732 if (rack_use_cmp_acks)
22733 rack->r_use_cmp_ack = 1;
22734 else
22735 rack->r_use_cmp_ack = 0;
22736 if (rack_disable_prr)
22737 rack->rack_no_prr = 1;
22738 else
22739 rack->rack_no_prr = 0;
22740 if (rack_gp_no_rec_chg)
22741 rack->rc_gp_no_rec_chg = 1;
22742 else
22743 rack->rc_gp_no_rec_chg = 0;
22744 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22745 rack->r_mbuf_queue = 1;
22746 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22747 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22748 rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22749 } else {
22750 rack->r_mbuf_queue = 0;
22751 rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22752 }
22753 if (rack_enable_shared_cwnd)
22754 rack->rack_enable_scwnd = 1;
22755 else
22756 rack->rack_enable_scwnd = 0;
22757 if (rack_do_dyn_mul) {
22758 /* When dynamic adjustment is on CA needs to start at 100% */
22759 rack->rc_gp_dyn_mul = 1;
22760 if (rack_do_dyn_mul >= 100)
22761 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22762 } else {
22763 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22764 rack->rc_gp_dyn_mul = 0;
22765 }
22766 rack->r_rr_config = 0;
22767 rack->r_ctl.rc_no_push_at_mrtt = 0;
22768 rack->rc_pace_fill_if_rttin_range = 0;
22769 rack->rtt_limit_mul = 0;
22770
22771 if (rack_enable_hw_pacing)
22772 rack->rack_hdw_pace_ena = 1;
22773 else
22774 rack->rack_hdw_pace_ena = 0;
22775 if (rack_disable_prr)
22776 rack->rack_no_prr = 1;
22777 else
22778 rack->rack_no_prr = 0;
22779 if (rack_limits_scwnd)
22780 rack->r_limit_scw = 1;
22781 else
22782 rack->r_limit_scw = 0;
22783 rack_init_retransmit_value(rack, rack_rxt_controls);
22784 err = 0;
22785 }
22786 return (err);
22787 }
22788
22789 static int
rack_add_deferred_option(struct tcp_rack * rack,int sopt_name,uint64_t loptval)22790 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22791 {
22792 struct deferred_opt_list *dol;
22793
22794 dol = malloc(sizeof(struct deferred_opt_list),
22795 M_TCPDO, M_NOWAIT|M_ZERO);
22796 if (dol == NULL) {
22797 /*
22798 * No space yikes -- fail out..
22799 */
22800 return (0);
22801 }
22802 dol->optname = sopt_name;
22803 dol->optval = loptval;
22804 TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22805 return (1);
22806 }
22807
22808 static int
process_hybrid_pacing(struct tcp_rack * rack,struct tcp_hybrid_req * hybrid)22809 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22810 {
22811 #ifdef TCP_REQUEST_TRK
22812 struct tcp_sendfile_track *sft;
22813 struct timeval tv;
22814 tcp_seq seq;
22815 int err;
22816
22817 microuptime(&tv);
22818
22819 /* Make sure no fixed rate is on */
22820 rack->use_fixed_rate = 0;
22821 rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22822 rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22823 rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22824 /* Now allocate or find our entry that will have these settings */
22825 sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22826 if (sft == NULL) {
22827 rack->rc_tp->tcp_hybrid_error++;
22828 /* no space, where would it have gone? */
22829 seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22830 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22831 return (ENOSPC);
22832 }
22833 /* mask our internal flags */
22834 hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
22835 /* The seq will be snd_una + everything in the buffer */
22836 seq = sft->start_seq;
22837 if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22838 /* Disabling hybrid pacing */
22839 if (rack->rc_hybrid_mode) {
22840 rack_set_profile(rack, 0);
22841 rack->rc_tp->tcp_hybrid_stop++;
22842 }
22843 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22844 return (0);
22845 }
22846 if (rack->dgp_on == 0) {
22847 /*
22848 * If we have not yet turned DGP on, do so
22849 * now setting pure DGP mode, no buffer level
22850 * response.
22851 */
22852 if ((err = rack_set_profile(rack, 1)) != 0){
22853 /* Failed to turn pacing on */
22854 rack->rc_tp->tcp_hybrid_error++;
22855 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22856 return (err);
22857 }
22858 }
22859 /*
22860 * Now we must switch to hybrid mode as well which also
22861 * means moving to regular pacing.
22862 */
22863 if (rack->rc_hybrid_mode == 0) {
22864 /* First time */
22865 if (tcp_can_enable_pacing()) {
22866 rack->r_ctl.pacing_method |= RACK_REG_PACING;
22867 rack->rc_hybrid_mode = 1;
22868 } else {
22869 return (ENOSPC);
22870 }
22871 if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
22872 /*
22873 * This should be true.
22874 */
22875 tcp_dec_dgp_pacing_cnt();
22876 rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
22877 }
22878 }
22879 /* Now set in our flags */
22880 sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22881 if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22882 sft->cspr = hybrid->cspr;
22883 else
22884 sft->cspr = 0;
22885 if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22886 sft->hint_maxseg = hybrid->hint_maxseg;
22887 else
22888 sft->hint_maxseg = 0;
22889 rack->rc_tp->tcp_hybrid_start++;
22890 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22891 return (0);
22892 #else
22893 return (ENOTSUP);
22894 #endif
22895 }
22896
22897 static int
rack_stack_information(struct tcpcb * tp,struct stack_specific_info * si)22898 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
22899 {
22900 /* We pulled a SSI info log out what was there */
22901 si->bytes_transmitted = tp->t_sndbytes;
22902 si->bytes_retransmitted = tp->t_snd_rxt_bytes;
22903 return (0);
22904 }
22905
22906 static int
rack_process_option(struct tcpcb * tp,struct tcp_rack * rack,int sopt_name,uint32_t optval,uint64_t loptval,struct tcp_hybrid_req * hybrid)22907 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22908 uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22909
22910 {
22911 struct epoch_tracker et;
22912 struct sockopt sopt;
22913 struct cc_newreno_opts opt;
22914 uint64_t val;
22915 int error = 0;
22916 uint16_t ca, ss;
22917
22918 switch (sopt_name) {
22919 case TCP_RACK_SET_RXT_OPTIONS:
22920 if (optval <= 2) {
22921 rack_init_retransmit_value(rack, optval);
22922 } else {
22923 /*
22924 * You must send in 0, 1 or 2 all else is
22925 * invalid.
22926 */
22927 error = EINVAL;
22928 }
22929 break;
22930 case TCP_RACK_DSACK_OPT:
22931 RACK_OPTS_INC(tcp_rack_dsack_opt);
22932 if (optval & 0x1) {
22933 rack->rc_rack_tmr_std_based = 1;
22934 } else {
22935 rack->rc_rack_tmr_std_based = 0;
22936 }
22937 if (optval & 0x2) {
22938 rack->rc_rack_use_dsack = 1;
22939 } else {
22940 rack->rc_rack_use_dsack = 0;
22941 }
22942 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
22943 break;
22944 case TCP_RACK_PACING_DIVISOR:
22945 RACK_OPTS_INC(tcp_rack_pacing_divisor);
22946 if (optval == 0) {
22947 rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
22948 } else {
22949 if (optval < RL_MIN_DIVISOR)
22950 rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
22951 else
22952 rack->r_ctl.pace_len_divisor = optval;
22953 }
22954 break;
22955 case TCP_RACK_HI_BETA:
22956 RACK_OPTS_INC(tcp_rack_hi_beta);
22957 if (optval > 0) {
22958 rack->rack_hibeta = 1;
22959 if ((optval >= 50) &&
22960 (optval <= 100)) {
22961 /*
22962 * User wants to set a custom beta.
22963 */
22964 rack->r_ctl.saved_hibeta = optval;
22965 if (rack->rc_pacing_cc_set)
22966 rack_undo_cc_pacing(rack);
22967 rack->r_ctl.rc_saved_beta = optval;
22968 }
22969 if (rack->rc_pacing_cc_set == 0)
22970 rack_set_cc_pacing(rack);
22971 } else {
22972 rack->rack_hibeta = 0;
22973 if (rack->rc_pacing_cc_set)
22974 rack_undo_cc_pacing(rack);
22975 }
22976 break;
22977 case TCP_RACK_PACING_BETA:
22978 error = EINVAL;
22979 break;
22980 case TCP_RACK_TIMER_SLOP:
22981 RACK_OPTS_INC(tcp_rack_timer_slop);
22982 rack->r_ctl.timer_slop = optval;
22983 if (rack->rc_tp->t_srtt) {
22984 /*
22985 * If we have an SRTT lets update t_rxtcur
22986 * to have the new slop.
22987 */
22988 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
22989 rack_rto_min, rack_rto_max,
22990 rack->r_ctl.timer_slop);
22991 }
22992 break;
22993 case TCP_RACK_PACING_BETA_ECN:
22994 RACK_OPTS_INC(tcp_rack_beta_ecn);
22995 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
22996 /* This only works for newreno. */
22997 error = EINVAL;
22998 break;
22999 }
23000 if (rack->rc_pacing_cc_set) {
23001 /*
23002 * Set them into the real CC module
23003 * whats in the rack pcb is the old values
23004 * to be used on restoral/
23005 */
23006 sopt.sopt_dir = SOPT_SET;
23007 opt.name = CC_NEWRENO_BETA_ECN;
23008 opt.val = optval;
23009 if (CC_ALGO(tp)->ctl_output != NULL)
23010 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23011 else
23012 error = ENOENT;
23013 } else {
23014 /*
23015 * Not pacing yet so set it into our local
23016 * rack pcb storage.
23017 */
23018 rack->r_ctl.rc_saved_beta_ecn = optval;
23019 }
23020 break;
23021 case TCP_DEFER_OPTIONS:
23022 RACK_OPTS_INC(tcp_defer_opt);
23023 if (optval) {
23024 if (rack->gp_ready) {
23025 /* Too late */
23026 error = EINVAL;
23027 break;
23028 }
23029 rack->defer_options = 1;
23030 } else
23031 rack->defer_options = 0;
23032 break;
23033 case TCP_RACK_MEASURE_CNT:
23034 RACK_OPTS_INC(tcp_rack_measure_cnt);
23035 if (optval && (optval <= 0xff)) {
23036 rack->r_ctl.req_measurements = optval;
23037 } else
23038 error = EINVAL;
23039 break;
23040 case TCP_REC_ABC_VAL:
23041 RACK_OPTS_INC(tcp_rec_abc_val);
23042 if (optval > 0)
23043 rack->r_use_labc_for_rec = 1;
23044 else
23045 rack->r_use_labc_for_rec = 0;
23046 break;
23047 case TCP_RACK_ABC_VAL:
23048 RACK_OPTS_INC(tcp_rack_abc_val);
23049 if ((optval > 0) && (optval < 255))
23050 rack->rc_labc = optval;
23051 else
23052 error = EINVAL;
23053 break;
23054 case TCP_HDWR_UP_ONLY:
23055 RACK_OPTS_INC(tcp_pacing_up_only);
23056 if (optval)
23057 rack->r_up_only = 1;
23058 else
23059 rack->r_up_only = 0;
23060 break;
23061 case TCP_FILLCW_RATE_CAP: /* URL:fillcw_cap */
23062 RACK_OPTS_INC(tcp_fillcw_rate_cap);
23063 rack->r_ctl.fillcw_cap = loptval;
23064 break;
23065 case TCP_PACING_RATE_CAP:
23066 RACK_OPTS_INC(tcp_pacing_rate_cap);
23067 if ((rack->dgp_on == 1) &&
23068 (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23069 /*
23070 * If we are doing DGP we need to switch
23071 * to using the pacing limit.
23072 */
23073 if (tcp_can_enable_pacing() == 0) {
23074 error = ENOSPC;
23075 break;
23076 }
23077 /*
23078 * Now change up the flags and counts to be correct.
23079 */
23080 rack->r_ctl.pacing_method |= RACK_REG_PACING;
23081 tcp_dec_dgp_pacing_cnt();
23082 rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23083 }
23084 rack->r_ctl.bw_rate_cap = loptval;
23085 break;
23086 case TCP_HYBRID_PACING:
23087 if (hybrid == NULL) {
23088 error = EINVAL;
23089 break;
23090 }
23091 if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
23092 error = EPERM;
23093 break;
23094 }
23095 error = process_hybrid_pacing(rack, hybrid);
23096 break;
23097 case TCP_SIDECHAN_DIS: /* URL:scodm */
23098 if (optval)
23099 rack->r_ctl.side_chan_dis_mask = optval;
23100 else
23101 rack->r_ctl.side_chan_dis_mask = 0;
23102 break;
23103 case TCP_RACK_PROFILE:
23104 RACK_OPTS_INC(tcp_profile);
23105 error = rack_set_profile(rack, optval);
23106 break;
23107 case TCP_USE_CMP_ACKS:
23108 RACK_OPTS_INC(tcp_use_cmp_acks);
23109 if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23110 /* You can't turn it off once its on! */
23111 error = EINVAL;
23112 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23113 rack->r_use_cmp_ack = 1;
23114 rack->r_mbuf_queue = 1;
23115 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23116 }
23117 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23118 tp->t_flags2 |= TF2_MBUF_ACKCMP;
23119 break;
23120 case TCP_SHARED_CWND_TIME_LIMIT:
23121 RACK_OPTS_INC(tcp_lscwnd);
23122 if (optval)
23123 rack->r_limit_scw = 1;
23124 else
23125 rack->r_limit_scw = 0;
23126 break;
23127 case TCP_RACK_DGP_IN_REC:
23128 error = EINVAL;
23129 break;
23130 case TCP_RACK_PACE_TO_FILL:
23131 RACK_OPTS_INC(tcp_fillcw);
23132 if (optval == 0)
23133 rack->rc_pace_to_cwnd = 0;
23134 else {
23135 rack->rc_pace_to_cwnd = 1;
23136 }
23137 if ((optval >= rack_gp_rtt_maxmul) &&
23138 rack_gp_rtt_maxmul &&
23139 (optval < 0xf)) {
23140 rack->rc_pace_fill_if_rttin_range = 1;
23141 rack->rtt_limit_mul = optval;
23142 } else {
23143 rack->rc_pace_fill_if_rttin_range = 0;
23144 rack->rtt_limit_mul = 0;
23145 }
23146 break;
23147 case TCP_RACK_NO_PUSH_AT_MAX:
23148 RACK_OPTS_INC(tcp_npush);
23149 if (optval == 0)
23150 rack->r_ctl.rc_no_push_at_mrtt = 0;
23151 else if (optval < 0xff)
23152 rack->r_ctl.rc_no_push_at_mrtt = optval;
23153 else
23154 error = EINVAL;
23155 break;
23156 case TCP_SHARED_CWND_ENABLE:
23157 RACK_OPTS_INC(tcp_rack_scwnd);
23158 if (optval == 0)
23159 rack->rack_enable_scwnd = 0;
23160 else
23161 rack->rack_enable_scwnd = 1;
23162 break;
23163 case TCP_RACK_MBUF_QUEUE:
23164 /* Now do we use the LRO mbuf-queue feature */
23165 RACK_OPTS_INC(tcp_rack_mbufq);
23166 if (optval || rack->r_use_cmp_ack)
23167 rack->r_mbuf_queue = 1;
23168 else
23169 rack->r_mbuf_queue = 0;
23170 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23171 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23172 else
23173 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23174 break;
23175 case TCP_RACK_NONRXT_CFG_RATE:
23176 RACK_OPTS_INC(tcp_rack_cfg_rate);
23177 if (optval == 0)
23178 rack->rack_rec_nonrxt_use_cr = 0;
23179 else
23180 rack->rack_rec_nonrxt_use_cr = 1;
23181 break;
23182 case TCP_NO_PRR:
23183 RACK_OPTS_INC(tcp_rack_noprr);
23184 if (optval == 0)
23185 rack->rack_no_prr = 0;
23186 else if (optval == 1)
23187 rack->rack_no_prr = 1;
23188 else if (optval == 2)
23189 rack->no_prr_addback = 1;
23190 else
23191 error = EINVAL;
23192 break;
23193 case RACK_CSPR_IS_FCC: /* URL:csprisfcc */
23194 if (optval > 0)
23195 rack->cspr_is_fcc = 1;
23196 else
23197 rack->cspr_is_fcc = 0;
23198 break;
23199 case TCP_TIMELY_DYN_ADJ:
23200 RACK_OPTS_INC(tcp_timely_dyn);
23201 if (optval == 0)
23202 rack->rc_gp_dyn_mul = 0;
23203 else {
23204 rack->rc_gp_dyn_mul = 1;
23205 if (optval >= 100) {
23206 /*
23207 * If the user sets something 100 or more
23208 * its the gp_ca value.
23209 */
23210 rack->r_ctl.rack_per_of_gp_ca = optval;
23211 }
23212 }
23213 break;
23214 case TCP_RACK_DO_DETECTION:
23215 error = EINVAL;
23216 break;
23217 case TCP_RACK_TLP_USE:
23218 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23219 error = EINVAL;
23220 break;
23221 }
23222 RACK_OPTS_INC(tcp_tlp_use);
23223 rack->rack_tlp_threshold_use = optval;
23224 break;
23225 case TCP_RACK_TLP_REDUCE:
23226 /* RACK TLP cwnd reduction (bool) */
23227 RACK_OPTS_INC(tcp_rack_tlp_reduce);
23228 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23229 break;
23230 /* Pacing related ones */
23231 case TCP_RACK_PACE_ALWAYS:
23232 /*
23233 * zero is old rack method, 1 is new
23234 * method using a pacing rate.
23235 */
23236 RACK_OPTS_INC(tcp_rack_pace_always);
23237 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23238 error = EPERM;
23239 break;
23240 }
23241 if (optval > 0) {
23242 if (rack->rc_always_pace) {
23243 error = EALREADY;
23244 break;
23245 } else if (tcp_can_enable_pacing()) {
23246 rack->r_ctl.pacing_method |= RACK_REG_PACING;
23247 rack->rc_always_pace = 1;
23248 if (rack->rack_hibeta)
23249 rack_set_cc_pacing(rack);
23250 }
23251 else {
23252 error = ENOSPC;
23253 break;
23254 }
23255 } else {
23256 if (rack->rc_always_pace == 1) {
23257 rack_remove_pacing(rack);
23258 }
23259 }
23260 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23261 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23262 else
23263 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23264 /* A rate may be set irate or other, if so set seg size */
23265 rack_update_seg(rack);
23266 break;
23267 case TCP_BBR_RACK_INIT_RATE:
23268 RACK_OPTS_INC(tcp_initial_rate);
23269 val = optval;
23270 /* Change from kbits per second to bytes per second */
23271 val *= 1000;
23272 val /= 8;
23273 rack->r_ctl.init_rate = val;
23274 if (rack->rc_always_pace)
23275 rack_update_seg(rack);
23276 break;
23277 case TCP_BBR_IWINTSO:
23278 error = EINVAL;
23279 break;
23280 case TCP_RACK_FORCE_MSEG:
23281 RACK_OPTS_INC(tcp_rack_force_max_seg);
23282 if (optval)
23283 rack->rc_force_max_seg = 1;
23284 else
23285 rack->rc_force_max_seg = 0;
23286 break;
23287 case TCP_RACK_PACE_MIN_SEG:
23288 RACK_OPTS_INC(tcp_rack_min_seg);
23289 rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23290 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23291 break;
23292 case TCP_RACK_PACE_MAX_SEG:
23293 /* Max segments size in a pace in bytes */
23294 RACK_OPTS_INC(tcp_rack_max_seg);
23295 if ((rack->dgp_on == 1) &&
23296 (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23297 /*
23298 * If we set a max-seg and are doing DGP then
23299 * we now fall under the pacing limits not the
23300 * DGP ones.
23301 */
23302 if (tcp_can_enable_pacing() == 0) {
23303 error = ENOSPC;
23304 break;
23305 }
23306 /*
23307 * Now change up the flags and counts to be correct.
23308 */
23309 rack->r_ctl.pacing_method |= RACK_REG_PACING;
23310 tcp_dec_dgp_pacing_cnt();
23311 rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23312 }
23313 if (optval <= MAX_USER_SET_SEG)
23314 rack->rc_user_set_max_segs = optval;
23315 else
23316 rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23317 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23318 break;
23319 case TCP_RACK_PACE_RATE_REC:
23320 /* Set the fixed pacing rate in Bytes per second ca */
23321 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23322 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23323 error = EPERM;
23324 break;
23325 }
23326 if (rack->dgp_on) {
23327 /*
23328 * We are already pacing another
23329 * way.
23330 */
23331 error = EBUSY;
23332 break;
23333 }
23334 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23335 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23336 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23337 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23338 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23339 rack->use_fixed_rate = 1;
23340 if (rack->rack_hibeta)
23341 rack_set_cc_pacing(rack);
23342 rack_log_pacing_delay_calc(rack,
23343 rack->r_ctl.rc_fixed_pacing_rate_ss,
23344 rack->r_ctl.rc_fixed_pacing_rate_ca,
23345 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23346 __LINE__, NULL,0);
23347 break;
23348
23349 case TCP_RACK_PACE_RATE_SS:
23350 /* Set the fixed pacing rate in Bytes per second ca */
23351 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23352 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23353 error = EPERM;
23354 break;
23355 }
23356 if (rack->dgp_on) {
23357 /*
23358 * We are already pacing another
23359 * way.
23360 */
23361 error = EBUSY;
23362 break;
23363 }
23364 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23365 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23366 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23367 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23368 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23369 rack->use_fixed_rate = 1;
23370 if (rack->rack_hibeta)
23371 rack_set_cc_pacing(rack);
23372 rack_log_pacing_delay_calc(rack,
23373 rack->r_ctl.rc_fixed_pacing_rate_ss,
23374 rack->r_ctl.rc_fixed_pacing_rate_ca,
23375 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23376 __LINE__, NULL, 0);
23377 break;
23378
23379 case TCP_RACK_PACE_RATE_CA:
23380 /* Set the fixed pacing rate in Bytes per second ca */
23381 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23382 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23383 error = EPERM;
23384 break;
23385 }
23386 if (rack->dgp_on) {
23387 /*
23388 * We are already pacing another
23389 * way.
23390 */
23391 error = EBUSY;
23392 break;
23393 }
23394 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23395 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23396 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23397 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23398 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23399 rack->use_fixed_rate = 1;
23400 if (rack->rack_hibeta)
23401 rack_set_cc_pacing(rack);
23402 rack_log_pacing_delay_calc(rack,
23403 rack->r_ctl.rc_fixed_pacing_rate_ss,
23404 rack->r_ctl.rc_fixed_pacing_rate_ca,
23405 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23406 __LINE__, NULL, 0);
23407 break;
23408 case TCP_RACK_GP_INCREASE_REC:
23409 RACK_OPTS_INC(tcp_gp_inc_rec);
23410 rack->r_ctl.rack_per_of_gp_rec = optval;
23411 rack_log_pacing_delay_calc(rack,
23412 rack->r_ctl.rack_per_of_gp_ss,
23413 rack->r_ctl.rack_per_of_gp_ca,
23414 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23415 __LINE__, NULL, 0);
23416 break;
23417 case TCP_RACK_GP_INCREASE_CA:
23418 RACK_OPTS_INC(tcp_gp_inc_ca);
23419 ca = optval;
23420 if (ca < 100) {
23421 /*
23422 * We don't allow any reduction
23423 * over the GP b/w.
23424 */
23425 error = EINVAL;
23426 break;
23427 }
23428 rack->r_ctl.rack_per_of_gp_ca = ca;
23429 rack_log_pacing_delay_calc(rack,
23430 rack->r_ctl.rack_per_of_gp_ss,
23431 rack->r_ctl.rack_per_of_gp_ca,
23432 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23433 __LINE__, NULL, 0);
23434 break;
23435 case TCP_RACK_GP_INCREASE_SS:
23436 RACK_OPTS_INC(tcp_gp_inc_ss);
23437 ss = optval;
23438 if (ss < 100) {
23439 /*
23440 * We don't allow any reduction
23441 * over the GP b/w.
23442 */
23443 error = EINVAL;
23444 break;
23445 }
23446 rack->r_ctl.rack_per_of_gp_ss = ss;
23447 rack_log_pacing_delay_calc(rack,
23448 rack->r_ctl.rack_per_of_gp_ss,
23449 rack->r_ctl.rack_per_of_gp_ca,
23450 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23451 __LINE__, NULL, 0);
23452 break;
23453 case TCP_RACK_RR_CONF:
23454 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23455 if (optval && optval <= 3)
23456 rack->r_rr_config = optval;
23457 else
23458 rack->r_rr_config = 0;
23459 break;
23460 case TCP_PACING_DND: /* URL:dnd */
23461 if (optval > 0)
23462 rack->rc_pace_dnd = 1;
23463 else
23464 rack->rc_pace_dnd = 0;
23465 break;
23466 case TCP_HDWR_RATE_CAP:
23467 RACK_OPTS_INC(tcp_hdwr_rate_cap);
23468 if (optval) {
23469 if (rack->r_rack_hw_rate_caps == 0)
23470 rack->r_rack_hw_rate_caps = 1;
23471 else
23472 error = EALREADY;
23473 } else {
23474 rack->r_rack_hw_rate_caps = 0;
23475 }
23476 break;
23477 case TCP_DGP_UPPER_BOUNDS:
23478 {
23479 uint8_t val;
23480 val = optval & 0x0000ff;
23481 rack->r_ctl.rack_per_upper_bound_ca = val;
23482 val = (optval >> 16) & 0x0000ff;
23483 rack->r_ctl.rack_per_upper_bound_ss = val;
23484 break;
23485 }
23486 case TCP_SS_EEXIT: /* URL:eexit */
23487 if (optval > 0) {
23488 rack->r_ctl.gp_rnd_thresh = optval & 0x0ff;
23489 if (optval & 0x10000) {
23490 rack->r_ctl.gate_to_fs = 1;
23491 } else {
23492 rack->r_ctl.gate_to_fs = 0;
23493 }
23494 if (optval & 0x20000) {
23495 rack->r_ctl.use_gp_not_last = 1;
23496 } else {
23497 rack->r_ctl.use_gp_not_last = 0;
23498 }
23499 if (optval & 0xfffc0000) {
23500 uint32_t v;
23501
23502 v = (optval >> 18) & 0x00003fff;
23503 if (v >= 1000)
23504 rack->r_ctl.gp_gain_req = v;
23505 }
23506 } else {
23507 /* We do not do ss early exit at all */
23508 rack->rc_initial_ss_comp = 1;
23509 rack->r_ctl.gp_rnd_thresh = 0;
23510 }
23511 break;
23512 case TCP_RACK_SPLIT_LIMIT:
23513 RACK_OPTS_INC(tcp_split_limit);
23514 rack->r_ctl.rc_split_limit = optval;
23515 break;
23516 case TCP_BBR_HDWR_PACE:
23517 RACK_OPTS_INC(tcp_hdwr_pacing);
23518 if (optval){
23519 if (rack->rack_hdrw_pacing == 0) {
23520 rack->rack_hdw_pace_ena = 1;
23521 rack->rack_attempt_hdwr_pace = 0;
23522 } else
23523 error = EALREADY;
23524 } else {
23525 rack->rack_hdw_pace_ena = 0;
23526 #ifdef RATELIMIT
23527 if (rack->r_ctl.crte != NULL) {
23528 rack->rack_hdrw_pacing = 0;
23529 rack->rack_attempt_hdwr_pace = 0;
23530 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23531 rack->r_ctl.crte = NULL;
23532 }
23533 #endif
23534 }
23535 break;
23536 /* End Pacing related ones */
23537 case TCP_RACK_PRR_SENDALOT:
23538 /* Allow PRR to send more than one seg */
23539 RACK_OPTS_INC(tcp_rack_prr_sendalot);
23540 rack->r_ctl.rc_prr_sendalot = optval;
23541 break;
23542 case TCP_RACK_MIN_TO:
23543 /* Minimum time between rack t-o's in ms */
23544 RACK_OPTS_INC(tcp_rack_min_to);
23545 rack->r_ctl.rc_min_to = optval;
23546 break;
23547 case TCP_RACK_EARLY_SEG:
23548 /* If early recovery max segments */
23549 RACK_OPTS_INC(tcp_rack_early_seg);
23550 rack->r_ctl.rc_early_recovery_segs = optval;
23551 break;
23552 case TCP_RACK_ENABLE_HYSTART:
23553 {
23554 if (optval) {
23555 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23556 if (rack_do_hystart > RACK_HYSTART_ON)
23557 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23558 if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23559 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23560 } else {
23561 tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23562 }
23563 }
23564 break;
23565 case TCP_RACK_REORD_THRESH:
23566 /* RACK reorder threshold (shift amount) */
23567 RACK_OPTS_INC(tcp_rack_reord_thresh);
23568 if ((optval > 0) && (optval < 31))
23569 rack->r_ctl.rc_reorder_shift = optval;
23570 else
23571 error = EINVAL;
23572 break;
23573 case TCP_RACK_REORD_FADE:
23574 /* Does reordering fade after ms time */
23575 RACK_OPTS_INC(tcp_rack_reord_fade);
23576 rack->r_ctl.rc_reorder_fade = optval;
23577 break;
23578 case TCP_RACK_TLP_THRESH:
23579 /* RACK TLP theshold i.e. srtt+(srtt/N) */
23580 RACK_OPTS_INC(tcp_rack_tlp_thresh);
23581 if (optval)
23582 rack->r_ctl.rc_tlp_threshold = optval;
23583 else
23584 error = EINVAL;
23585 break;
23586 case TCP_BBR_USE_RACK_RR:
23587 RACK_OPTS_INC(tcp_rack_rr);
23588 if (optval)
23589 rack->use_rack_rr = 1;
23590 else
23591 rack->use_rack_rr = 0;
23592 break;
23593 case TCP_RACK_PKT_DELAY:
23594 /* RACK added ms i.e. rack-rtt + reord + N */
23595 RACK_OPTS_INC(tcp_rack_pkt_delay);
23596 rack->r_ctl.rc_pkt_delay = optval;
23597 break;
23598 case TCP_DELACK:
23599 RACK_OPTS_INC(tcp_rack_delayed_ack);
23600 if (optval == 0)
23601 tp->t_delayed_ack = 0;
23602 else
23603 tp->t_delayed_ack = 1;
23604 if (tp->t_flags & TF_DELACK) {
23605 tp->t_flags &= ~TF_DELACK;
23606 tp->t_flags |= TF_ACKNOW;
23607 NET_EPOCH_ENTER(et);
23608 rack_output(tp);
23609 NET_EPOCH_EXIT(et);
23610 }
23611 break;
23612
23613 case TCP_BBR_RACK_RTT_USE:
23614 RACK_OPTS_INC(tcp_rack_rtt_use);
23615 if ((optval != USE_RTT_HIGH) &&
23616 (optval != USE_RTT_LOW) &&
23617 (optval != USE_RTT_AVG))
23618 error = EINVAL;
23619 else
23620 rack->r_ctl.rc_rate_sample_method = optval;
23621 break;
23622 case TCP_HONOR_HPTS_MIN:
23623 RACK_OPTS_INC(tcp_honor_hpts);
23624 if (optval) {
23625 rack->r_use_hpts_min = 1;
23626 /*
23627 * Must be between 2 - 80% to be a reduction else
23628 * we keep the default (10%).
23629 */
23630 if ((optval > 1) && (optval <= 80)) {
23631 rack->r_ctl.max_reduction = optval;
23632 }
23633 } else
23634 rack->r_use_hpts_min = 0;
23635 break;
23636 case TCP_REC_IS_DYN: /* URL:dynrec */
23637 RACK_OPTS_INC(tcp_dyn_rec);
23638 if (optval)
23639 rack->rc_gp_no_rec_chg = 1;
23640 else
23641 rack->rc_gp_no_rec_chg = 0;
23642 break;
23643 case TCP_NO_TIMELY:
23644 RACK_OPTS_INC(tcp_notimely);
23645 if (optval) {
23646 rack->rc_skip_timely = 1;
23647 rack->r_ctl.rack_per_of_gp_rec = 90;
23648 rack->r_ctl.rack_per_of_gp_ca = 100;
23649 rack->r_ctl.rack_per_of_gp_ss = 250;
23650 } else {
23651 rack->rc_skip_timely = 0;
23652 }
23653 break;
23654 case TCP_GP_USE_LTBW:
23655 if (optval == 0) {
23656 rack->use_lesser_lt_bw = 0;
23657 rack->dis_lt_bw = 1;
23658 } else if (optval == 1) {
23659 rack->use_lesser_lt_bw = 1;
23660 rack->dis_lt_bw = 0;
23661 } else if (optval == 2) {
23662 rack->use_lesser_lt_bw = 0;
23663 rack->dis_lt_bw = 0;
23664 }
23665 break;
23666 case TCP_DATA_AFTER_CLOSE:
23667 RACK_OPTS_INC(tcp_data_after_close);
23668 if (optval)
23669 rack->rc_allow_data_af_clo = 1;
23670 else
23671 rack->rc_allow_data_af_clo = 0;
23672 break;
23673 default:
23674 break;
23675 }
23676 tcp_log_socket_option(tp, sopt_name, optval, error);
23677 return (error);
23678 }
23679
23680 static void
rack_inherit(struct tcpcb * tp,struct inpcb * parent)23681 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
23682 {
23683 /*
23684 * A new connection has been created (tp) and
23685 * the parent is the inpcb given. We want to
23686 * apply a read-lock to the parent (we are already
23687 * holding a write lock on the tp) and copy anything
23688 * out of the rack specific data as long as its tfb is
23689 * the same as ours i.e. we are the same stack. Otherwise
23690 * we just return.
23691 */
23692 struct tcpcb *par;
23693 struct tcp_rack *dest, *src;
23694 int cnt = 0;
23695
23696 par = intotcpcb(parent);
23697 if (par->t_fb != tp->t_fb) {
23698 /* Not the same stack */
23699 tcp_log_socket_option(tp, 0, 0, 1);
23700 return;
23701 }
23702 /* Ok if we reach here lets setup the two rack pointers */
23703 dest = (struct tcp_rack *)tp->t_fb_ptr;
23704 src = (struct tcp_rack *)par->t_fb_ptr;
23705 if ((src == NULL) || (dest == NULL)) {
23706 /* Huh? */
23707 tcp_log_socket_option(tp, 0, 0, 2);
23708 return;
23709 }
23710 /* Now copy out anything we wish to inherit i.e. things in socket-options */
23711 /* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
23712 if ((src->dgp_on) && (dest->dgp_on == 0)) {
23713 /* Profile 1 had to be set via sock opt */
23714 rack_set_dgp(dest);
23715 cnt++;
23716 }
23717 /* TCP_RACK_SET_RXT_OPTIONS */
23718 if (dest->full_size_rxt != src->full_size_rxt) {
23719 dest->full_size_rxt = src->full_size_rxt;
23720 cnt++;
23721 }
23722 if (dest->shape_rxt_to_pacing_min != src->shape_rxt_to_pacing_min) {
23723 dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
23724 cnt++;
23725 }
23726 /* TCP_RACK_DSACK_OPT */
23727 if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
23728 dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
23729 cnt++;
23730 }
23731 if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
23732 dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
23733 cnt++;
23734 }
23735 /* TCP_RACK_PACING_DIVISOR */
23736 if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
23737 dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
23738 cnt++;
23739 }
23740 /* TCP_RACK_HI_BETA */
23741 if (src->rack_hibeta != dest->rack_hibeta) {
23742 cnt++;
23743 if (src->rack_hibeta) {
23744 dest->r_ctl.rc_saved_beta = src->r_ctl.rc_saved_beta;
23745 dest->rack_hibeta = 1;
23746 } else {
23747 dest->rack_hibeta = 0;
23748 }
23749 }
23750 /* TCP_RACK_TIMER_SLOP */
23751 if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
23752 dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
23753 cnt++;
23754 }
23755 /* TCP_RACK_PACING_BETA_ECN */
23756 if (dest->r_ctl.rc_saved_beta_ecn != src->r_ctl.rc_saved_beta_ecn) {
23757 dest->r_ctl.rc_saved_beta_ecn = src->r_ctl.rc_saved_beta_ecn;
23758 cnt++;
23759 }
23760 /* We do not do TCP_DEFER_OPTIONS */
23761 /* TCP_RACK_MEASURE_CNT */
23762 if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
23763 dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
23764 cnt++;
23765 }
23766 /* TCP_HDWR_UP_ONLY */
23767 if (dest->r_up_only != src->r_up_only) {
23768 dest->r_up_only = src->r_up_only;
23769 cnt++;
23770 }
23771 /* TCP_FILLCW_RATE_CAP */
23772 if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
23773 dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
23774 cnt++;
23775 }
23776 /* TCP_PACING_RATE_CAP */
23777 if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
23778 dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
23779 cnt++;
23780 }
23781 /* A listener can't set TCP_HYBRID_PACING */
23782 /* TCP_SIDECHAN_DIS */
23783 if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
23784 dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
23785 cnt++;
23786 }
23787 /* TCP_SHARED_CWND_TIME_LIMIT */
23788 if (dest->r_limit_scw != src->r_limit_scw) {
23789 dest->r_limit_scw = src->r_limit_scw;
23790 cnt++;
23791 }
23792 /* TCP_RACK_PACE_TO_FILL */
23793 if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
23794 dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
23795 cnt++;
23796 }
23797 if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
23798 dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
23799 cnt++;
23800 }
23801 if (dest->rtt_limit_mul != src->rtt_limit_mul) {
23802 dest->rtt_limit_mul = src->rtt_limit_mul;
23803 cnt++;
23804 }
23805 /* TCP_RACK_NO_PUSH_AT_MAX */
23806 if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
23807 dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
23808 cnt++;
23809 }
23810 /* TCP_SHARED_CWND_ENABLE */
23811 if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
23812 dest->rack_enable_scwnd = src->rack_enable_scwnd;
23813 cnt++;
23814 }
23815 /* TCP_USE_CMP_ACKS */
23816 if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
23817 dest->r_use_cmp_ack = src->r_use_cmp_ack;
23818 cnt++;
23819 }
23820
23821 if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23822 dest->r_mbuf_queue = src->r_mbuf_queue;
23823 cnt++;
23824 }
23825 /* TCP_RACK_MBUF_QUEUE */
23826 if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23827 dest->r_mbuf_queue = src->r_mbuf_queue;
23828 cnt++;
23829 }
23830 if (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
23831 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23832 } else {
23833 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23834 }
23835 if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
23836 tp->t_flags2 |= TF2_MBUF_ACKCMP;
23837 }
23838 /* TCP_RACK_NONRXT_CFG_RATE */
23839 if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
23840 dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
23841 cnt++;
23842 }
23843 /* TCP_NO_PRR */
23844 if (dest->rack_no_prr != src->rack_no_prr) {
23845 dest->rack_no_prr = src->rack_no_prr;
23846 cnt++;
23847 }
23848 if (dest->no_prr_addback != src->no_prr_addback) {
23849 dest->no_prr_addback = src->no_prr_addback;
23850 cnt++;
23851 }
23852 /* RACK_CSPR_IS_FCC */
23853 if (dest->cspr_is_fcc != src->cspr_is_fcc) {
23854 dest->cspr_is_fcc = src->cspr_is_fcc;
23855 cnt++;
23856 }
23857 /* TCP_TIMELY_DYN_ADJ */
23858 if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
23859 dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
23860 cnt++;
23861 }
23862 if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23863 dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23864 cnt++;
23865 }
23866 /* TCP_RACK_TLP_USE */
23867 if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
23868 dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
23869 cnt++;
23870 }
23871 /* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
23872 /* TCP_BBR_RACK_INIT_RATE */
23873 if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
23874 dest->r_ctl.init_rate = src->r_ctl.init_rate;
23875 cnt++;
23876 }
23877 /* TCP_RACK_FORCE_MSEG */
23878 if (dest->rc_force_max_seg != src->rc_force_max_seg) {
23879 dest->rc_force_max_seg = src->rc_force_max_seg;
23880 cnt++;
23881 }
23882 /* TCP_RACK_PACE_MIN_SEG */
23883 if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
23884 dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
23885 cnt++;
23886 }
23887 /* we don't allow TCP_RACK_PACE_MAX_SEG */
23888 /* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS, TCP_RACK_PACE_RATE_CA */
23889 if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
23890 dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
23891 cnt++;
23892 }
23893 if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
23894 dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
23895 cnt++;
23896 }
23897 if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
23898 dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
23899 cnt++;
23900 }
23901 /* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
23902 if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
23903 dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
23904 cnt++;
23905 }
23906 if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23907 dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23908 cnt++;
23909 }
23910
23911 if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
23912 dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
23913 cnt++;
23914 }
23915 /* TCP_RACK_RR_CONF */
23916 if (dest->r_rr_config != src->r_rr_config) {
23917 dest->r_rr_config = src->r_rr_config;
23918 cnt++;
23919 }
23920 /* TCP_PACING_DND */
23921 if (dest->rc_pace_dnd != src->rc_pace_dnd) {
23922 dest->rc_pace_dnd = src->rc_pace_dnd;
23923 cnt++;
23924 }
23925 /* TCP_HDWR_RATE_CAP */
23926 if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
23927 dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
23928 cnt++;
23929 }
23930 /* TCP_DGP_UPPER_BOUNDS */
23931 if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
23932 dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
23933 cnt++;
23934 }
23935 if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
23936 dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
23937 cnt++;
23938 }
23939 /* TCP_SS_EEXIT */
23940 if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
23941 dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
23942 cnt++;
23943 }
23944 if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
23945 dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
23946 cnt++;
23947 }
23948 if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
23949 dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
23950 cnt++;
23951 }
23952 if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
23953 dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
23954 cnt++;
23955 }
23956 /* TCP_BBR_HDWR_PACE */
23957 if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
23958 dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
23959 cnt++;
23960 }
23961 if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
23962 dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
23963 cnt++;
23964 }
23965 /* TCP_RACK_PRR_SENDALOT */
23966 if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
23967 dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
23968 cnt++;
23969 }
23970 /* TCP_RACK_MIN_TO */
23971 if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
23972 dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
23973 cnt++;
23974 }
23975 /* TCP_RACK_EARLY_SEG */
23976 if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
23977 dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
23978 cnt++;
23979 }
23980 /* TCP_RACK_ENABLE_HYSTART */
23981 if (par->t_ccv.flags != tp->t_ccv.flags) {
23982 cnt++;
23983 if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
23984 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23985 if (rack_do_hystart > RACK_HYSTART_ON)
23986 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23987 if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23988 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23989 } else {
23990 tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23991 }
23992 }
23993 /* TCP_RACK_REORD_THRESH */
23994 if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
23995 dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
23996 cnt++;
23997 }
23998 /* TCP_RACK_REORD_FADE */
23999 if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
24000 dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
24001 cnt++;
24002 }
24003 /* TCP_RACK_TLP_THRESH */
24004 if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
24005 dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
24006 cnt++;
24007 }
24008 /* TCP_BBR_USE_RACK_RR */
24009 if (dest->use_rack_rr != src->use_rack_rr) {
24010 dest->use_rack_rr = src->use_rack_rr;
24011 cnt++;
24012 }
24013 /* TCP_RACK_PKT_DELAY */
24014 if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
24015 dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
24016 cnt++;
24017 }
24018 /* TCP_DELACK will get copied via the main code if applicable */
24019 /* TCP_BBR_RACK_RTT_USE */
24020 if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
24021 dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
24022 cnt++;
24023 }
24024 /* TCP_HONOR_HPTS_MIN */
24025 if (dest->r_use_hpts_min != src->r_use_hpts_min) {
24026 dest->r_use_hpts_min = src->r_use_hpts_min;
24027 cnt++;
24028 }
24029 if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
24030 dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
24031 cnt++;
24032 }
24033 /* TCP_REC_IS_DYN */
24034 if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
24035 dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
24036 cnt++;
24037 }
24038 if (dest->rc_skip_timely != src->rc_skip_timely) {
24039 dest->rc_skip_timely = src->rc_skip_timely;
24040 cnt++;
24041 }
24042 /* TCP_DATA_AFTER_CLOSE */
24043 if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
24044 dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
24045 cnt++;
24046 }
24047 /* TCP_GP_USE_LTBW */
24048 if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
24049 dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
24050 cnt++;
24051 }
24052 if (dest->dis_lt_bw != src->dis_lt_bw) {
24053 dest->dis_lt_bw = src->dis_lt_bw;
24054 cnt++;
24055 }
24056 tcp_log_socket_option(tp, 0, cnt, 0);
24057 }
24058
24059
24060 static void
rack_apply_deferred_options(struct tcp_rack * rack)24061 rack_apply_deferred_options(struct tcp_rack *rack)
24062 {
24063 struct deferred_opt_list *dol, *sdol;
24064 uint32_t s_optval;
24065
24066 TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
24067 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
24068 /* Disadvantage of deferal is you loose the error return */
24069 s_optval = (uint32_t)dol->optval;
24070 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
24071 free(dol, M_TCPDO);
24072 }
24073 }
24074
24075 static void
rack_hw_tls_change(struct tcpcb * tp,int chg)24076 rack_hw_tls_change(struct tcpcb *tp, int chg)
24077 {
24078 /* Update HW tls state */
24079 struct tcp_rack *rack;
24080
24081 rack = (struct tcp_rack *)tp->t_fb_ptr;
24082 if (chg)
24083 rack->r_ctl.fsb.hw_tls = 1;
24084 else
24085 rack->r_ctl.fsb.hw_tls = 0;
24086 }
24087
24088 static int
rack_pru_options(struct tcpcb * tp,int flags)24089 rack_pru_options(struct tcpcb *tp, int flags)
24090 {
24091 if (flags & PRUS_OOB)
24092 return (EOPNOTSUPP);
24093 return (0);
24094 }
24095
24096 static bool
rack_wake_check(struct tcpcb * tp)24097 rack_wake_check(struct tcpcb *tp)
24098 {
24099 struct tcp_rack *rack;
24100 struct timeval tv;
24101 uint32_t cts;
24102
24103 rack = (struct tcp_rack *)tp->t_fb_ptr;
24104 if (rack->r_ctl.rc_hpts_flags) {
24105 cts = tcp_get_usecs(&tv);
24106 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
24107 /*
24108 * Pacing timer is up, check if we are ready.
24109 */
24110 if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
24111 return (true);
24112 } else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
24113 /*
24114 * A timer is up, check if we are ready.
24115 */
24116 if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
24117 return (true);
24118 }
24119 }
24120 return (false);
24121 }
24122
24123 static struct tcp_function_block __tcp_rack = {
24124 .tfb_tcp_block_name = __XSTRING(STACKNAME),
24125 .tfb_tcp_output = rack_output,
24126 .tfb_do_queued_segments = ctf_do_queued_segments,
24127 .tfb_do_segment_nounlock = rack_do_segment_nounlock,
24128 .tfb_tcp_do_segment = rack_do_segment,
24129 .tfb_tcp_ctloutput = rack_ctloutput,
24130 .tfb_tcp_fb_init = rack_init,
24131 .tfb_tcp_fb_fini = rack_fini,
24132 .tfb_tcp_timer_stop_all = rack_stopall,
24133 .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
24134 .tfb_tcp_handoff_ok = rack_handoff_ok,
24135 .tfb_tcp_mtu_chg = rack_mtu_change,
24136 .tfb_pru_options = rack_pru_options,
24137 .tfb_hwtls_change = rack_hw_tls_change,
24138 .tfb_chg_query = rack_chg_query,
24139 .tfb_switch_failed = rack_switch_failed,
24140 .tfb_early_wake_check = rack_wake_check,
24141 .tfb_compute_pipe = rack_compute_pipe,
24142 .tfb_stack_info = rack_stack_information,
24143 .tfb_inherit = rack_inherit,
24144 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP | TCP_FUNC_DEFAULT_OK,
24145
24146 };
24147
24148 /*
24149 * rack_ctloutput() must drop the inpcb lock before performing copyin on
24150 * socket option arguments. When it re-acquires the lock after the copy, it
24151 * has to revalidate that the connection is still valid for the socket
24152 * option.
24153 */
24154 static int
rack_set_sockopt(struct tcpcb * tp,struct sockopt * sopt)24155 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24156 {
24157 struct inpcb *inp = tptoinpcb(tp);
24158 #ifdef INET
24159 struct ip *ip;
24160 #endif
24161 struct tcp_rack *rack;
24162 struct tcp_hybrid_req hybrid;
24163 uint64_t loptval;
24164 int32_t error = 0, optval;
24165
24166 rack = (struct tcp_rack *)tp->t_fb_ptr;
24167 if (rack == NULL) {
24168 INP_WUNLOCK(inp);
24169 return (EINVAL);
24170 }
24171 #ifdef INET
24172 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
24173 #endif
24174
24175 switch (sopt->sopt_level) {
24176 #ifdef INET6
24177 case IPPROTO_IPV6:
24178 MPASS(inp->inp_vflag & INP_IPV6PROTO);
24179 switch (sopt->sopt_name) {
24180 case IPV6_USE_MIN_MTU:
24181 tcp6_use_min_mtu(tp);
24182 break;
24183 }
24184 INP_WUNLOCK(inp);
24185 return (0);
24186 #endif
24187 #ifdef INET
24188 case IPPROTO_IP:
24189 switch (sopt->sopt_name) {
24190 case IP_TOS:
24191 /*
24192 * The DSCP codepoint has changed, update the fsb.
24193 */
24194 ip->ip_tos = rack->rc_inp->inp_ip_tos;
24195 break;
24196 case IP_TTL:
24197 /*
24198 * The TTL has changed, update the fsb.
24199 */
24200 ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
24201 break;
24202 }
24203 INP_WUNLOCK(inp);
24204 return (0);
24205 #endif
24206 #ifdef SO_PEERPRIO
24207 case SOL_SOCKET:
24208 switch (sopt->sopt_name) {
24209 case SO_PEERPRIO: /* SC-URL:bs */
24210 /* Already read in and sanity checked in sosetopt(). */
24211 if (inp->inp_socket) {
24212 rack->client_bufferlvl = inp->inp_socket->so_peerprio;
24213 }
24214 break;
24215 }
24216 INP_WUNLOCK(inp);
24217 return (0);
24218 #endif
24219 case IPPROTO_TCP:
24220 switch (sopt->sopt_name) {
24221 case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */
24222 /* Pacing related ones */
24223 case TCP_RACK_PACE_ALWAYS: /* URL:pace_always */
24224 case TCP_BBR_RACK_INIT_RATE: /* URL:irate */
24225 case TCP_RACK_PACE_MIN_SEG: /* URL:pace_min_seg */
24226 case TCP_RACK_PACE_MAX_SEG: /* URL:pace_max_seg */
24227 case TCP_RACK_FORCE_MSEG: /* URL:force_max_seg */
24228 case TCP_RACK_PACE_RATE_CA: /* URL:pr_ca */
24229 case TCP_RACK_PACE_RATE_SS: /* URL:pr_ss*/
24230 case TCP_RACK_PACE_RATE_REC: /* URL:pr_rec */
24231 case TCP_RACK_GP_INCREASE_CA: /* URL:gp_inc_ca */
24232 case TCP_RACK_GP_INCREASE_SS: /* URL:gp_inc_ss */
24233 case TCP_RACK_GP_INCREASE_REC: /* URL:gp_inc_rec */
24234 case TCP_RACK_RR_CONF: /* URL:rrr_conf */
24235 case TCP_BBR_HDWR_PACE: /* URL:hdwrpace */
24236 case TCP_HDWR_RATE_CAP: /* URL:hdwrcap boolean */
24237 case TCP_PACING_RATE_CAP: /* URL:cap -- used by side-channel */
24238 case TCP_HDWR_UP_ONLY: /* URL:uponly -- hardware pacing boolean */
24239 case TCP_FILLCW_RATE_CAP: /* URL:fillcw_cap */
24240 case TCP_RACK_PACING_BETA_ECN: /* URL:pacing_beta_ecn */
24241 case TCP_RACK_PACE_TO_FILL: /* URL:fillcw */
24242 /* End pacing related */
24243 case TCP_DELACK: /* URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
24244 case TCP_RACK_PRR_SENDALOT: /* URL:prr_sendalot */
24245 case TCP_RACK_MIN_TO: /* URL:min_to */
24246 case TCP_RACK_EARLY_SEG: /* URL:early_seg */
24247 case TCP_RACK_REORD_THRESH: /* URL:reord_thresh */
24248 case TCP_RACK_REORD_FADE: /* URL:reord_fade */
24249 case TCP_RACK_TLP_THRESH: /* URL:tlp_thresh */
24250 case TCP_RACK_PKT_DELAY: /* URL:pkt_delay */
24251 case TCP_RACK_TLP_USE: /* URL:tlp_use */
24252 case TCP_BBR_RACK_RTT_USE: /* URL:rttuse */
24253 case TCP_BBR_USE_RACK_RR: /* URL:rackrr */
24254 case TCP_NO_PRR: /* URL:noprr */
24255 case TCP_TIMELY_DYN_ADJ: /* URL:dynamic */
24256 case TCP_DATA_AFTER_CLOSE: /* no URL */
24257 case TCP_RACK_NONRXT_CFG_RATE: /* URL:nonrxtcr */
24258 case TCP_SHARED_CWND_ENABLE: /* URL:scwnd */
24259 case TCP_RACK_MBUF_QUEUE: /* URL:mqueue */
24260 case TCP_RACK_NO_PUSH_AT_MAX: /* URL:npush */
24261 case TCP_SHARED_CWND_TIME_LIMIT: /* URL:lscwnd */
24262 case TCP_RACK_PROFILE: /* URL:profile */
24263 case TCP_SIDECHAN_DIS: /* URL:scodm */
24264 case TCP_HYBRID_PACING: /* URL:pacing=hybrid */
24265 case TCP_USE_CMP_ACKS: /* URL:cmpack */
24266 case TCP_RACK_ABC_VAL: /* URL:labc */
24267 case TCP_REC_ABC_VAL: /* URL:reclabc */
24268 case TCP_RACK_MEASURE_CNT: /* URL:measurecnt */
24269 case TCP_DEFER_OPTIONS: /* URL:defer */
24270 case TCP_RACK_DSACK_OPT: /* URL:dsack */
24271 case TCP_RACK_TIMER_SLOP: /* URL:timer_slop */
24272 case TCP_RACK_ENABLE_HYSTART: /* URL:hystart */
24273 case TCP_RACK_SET_RXT_OPTIONS: /* URL:rxtsz */
24274 case TCP_RACK_HI_BETA: /* URL:hibeta */
24275 case TCP_RACK_SPLIT_LIMIT: /* URL:split */
24276 case TCP_SS_EEXIT: /* URL:eexit */
24277 case TCP_DGP_UPPER_BOUNDS: /* URL:upper */
24278 case TCP_RACK_PACING_DIVISOR: /* URL:divisor */
24279 case TCP_PACING_DND: /* URL:dnd */
24280 case TCP_NO_TIMELY: /* URL:notimely */
24281 case RACK_CSPR_IS_FCC: /* URL:csprisfcc */
24282 case TCP_HONOR_HPTS_MIN: /* URL:hptsmin */
24283 case TCP_REC_IS_DYN: /* URL:dynrec */
24284 case TCP_GP_USE_LTBW: /* URL:useltbw */
24285 goto process_opt;
24286 break;
24287 default:
24288 /* Filter off all unknown options to the base stack */
24289 return (tcp_default_ctloutput(tp, sopt));
24290 break;
24291 }
24292 default:
24293 INP_WUNLOCK(inp);
24294 return (0);
24295 }
24296 process_opt:
24297 INP_WUNLOCK(inp);
24298 if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24299 (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
24300 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
24301 /*
24302 * We truncate it down to 32 bits for the socket-option trace this
24303 * means rates > 34Gbps won't show right, but thats probably ok.
24304 */
24305 optval = (uint32_t)loptval;
24306 } else if (sopt->sopt_name == TCP_HYBRID_PACING) {
24307 error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
24308 } else {
24309 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
24310 /* Save it in 64 bit form too */
24311 loptval = optval;
24312 }
24313 if (error)
24314 return (error);
24315 INP_WLOCK(inp);
24316 if (tp->t_fb != &__tcp_rack) {
24317 INP_WUNLOCK(inp);
24318 return (ENOPROTOOPT);
24319 }
24320 if (rack->defer_options && (rack->gp_ready == 0) &&
24321 (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
24322 (sopt->sopt_name != TCP_HYBRID_PACING) &&
24323 (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
24324 (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
24325 (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
24326 /* Options are being deferred */
24327 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
24328 INP_WUNLOCK(inp);
24329 return (0);
24330 } else {
24331 /* No memory to defer, fail */
24332 INP_WUNLOCK(inp);
24333 return (ENOMEM);
24334 }
24335 }
24336 error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
24337 INP_WUNLOCK(inp);
24338 return (error);
24339 }
24340
24341 static void
rack_fill_info(struct tcpcb * tp,struct tcp_info * ti)24342 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
24343 {
24344
24345 INP_WLOCK_ASSERT(tptoinpcb(tp));
24346 bzero(ti, sizeof(*ti));
24347
24348 ti->tcpi_state = tp->t_state;
24349 if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
24350 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
24351 if (tp->t_flags & TF_SACK_PERMIT)
24352 ti->tcpi_options |= TCPI_OPT_SACK;
24353 if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
24354 ti->tcpi_options |= TCPI_OPT_WSCALE;
24355 ti->tcpi_snd_wscale = tp->snd_scale;
24356 ti->tcpi_rcv_wscale = tp->rcv_scale;
24357 }
24358 if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
24359 ti->tcpi_options |= TCPI_OPT_ECN;
24360 if (tp->t_flags & TF_FASTOPEN)
24361 ti->tcpi_options |= TCPI_OPT_TFO;
24362 /* still kept in ticks is t_rcvtime */
24363 ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
24364 /* Since we hold everything in precise useconds this is easy */
24365 ti->tcpi_rtt = tp->t_srtt;
24366 ti->tcpi_rttvar = tp->t_rttvar;
24367 ti->tcpi_rto = tp->t_rxtcur;
24368 ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
24369 ti->tcpi_snd_cwnd = tp->snd_cwnd;
24370 /*
24371 * FreeBSD-specific extension fields for tcp_info.
24372 */
24373 ti->tcpi_rcv_space = tp->rcv_wnd;
24374 ti->tcpi_rcv_nxt = tp->rcv_nxt;
24375 ti->tcpi_snd_wnd = tp->snd_wnd;
24376 ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */
24377 ti->tcpi_snd_nxt = tp->snd_nxt;
24378 ti->tcpi_snd_mss = tp->t_maxseg;
24379 ti->tcpi_rcv_mss = tp->t_maxseg;
24380 ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24381 ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24382 ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24383 ti->tcpi_total_tlp = tp->t_sndtlppack;
24384 ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24385 ti->tcpi_rttmin = tp->t_rttlow;
24386 #ifdef NETFLIX_STATS
24387 memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24388 #endif
24389 #ifdef TCP_OFFLOAD
24390 if (tp->t_flags & TF_TOE) {
24391 ti->tcpi_options |= TCPI_OPT_TOE;
24392 tcp_offload_tcp_info(tp, ti);
24393 }
24394 #endif
24395 }
24396
24397 static int
rack_get_sockopt(struct tcpcb * tp,struct sockopt * sopt)24398 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24399 {
24400 struct inpcb *inp = tptoinpcb(tp);
24401 struct tcp_rack *rack;
24402 int32_t error, optval;
24403 uint64_t val, loptval;
24404 struct tcp_info ti;
24405 /*
24406 * Because all our options are either boolean or an int, we can just
24407 * pull everything into optval and then unlock and copy. If we ever
24408 * add a option that is not a int, then this will have quite an
24409 * impact to this routine.
24410 */
24411 error = 0;
24412 rack = (struct tcp_rack *)tp->t_fb_ptr;
24413 if (rack == NULL) {
24414 INP_WUNLOCK(inp);
24415 return (EINVAL);
24416 }
24417 switch (sopt->sopt_name) {
24418 case TCP_INFO:
24419 /* First get the info filled */
24420 rack_fill_info(tp, &ti);
24421 /* Fix up the rtt related fields if needed */
24422 INP_WUNLOCK(inp);
24423 error = sooptcopyout(sopt, &ti, sizeof ti);
24424 return (error);
24425 /*
24426 * Beta is the congestion control value for NewReno that influences how
24427 * much of a backoff happens when loss is detected. It is normally set
24428 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24429 * when you exit recovery.
24430 */
24431 case TCP_RACK_PACING_BETA:
24432 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24433 error = EINVAL;
24434 else if (rack->rc_pacing_cc_set == 0)
24435 optval = rack->r_ctl.rc_saved_beta;
24436 else {
24437 /*
24438 * Reach out into the CC data and report back what
24439 * I have previously set. Yeah it looks hackish but
24440 * we don't want to report the saved values.
24441 */
24442 if (tp->t_ccv.cc_data)
24443 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24444 else
24445 error = EINVAL;
24446 }
24447 break;
24448 /*
24449 * Beta_ecn is the congestion control value for NewReno that influences how
24450 * much of a backoff happens when a ECN mark is detected. It is normally set
24451 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24452 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24453 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24454 */
24455 case TCP_RACK_PACING_BETA_ECN:
24456 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24457 error = EINVAL;
24458 else if (rack->rc_pacing_cc_set == 0)
24459 optval = rack->r_ctl.rc_saved_beta_ecn;
24460 else {
24461 /*
24462 * Reach out into the CC data and report back what
24463 * I have previously set. Yeah it looks hackish but
24464 * we don't want to report the saved values.
24465 */
24466 if (tp->t_ccv.cc_data)
24467 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24468 else
24469 error = EINVAL;
24470 }
24471 break;
24472 case TCP_RACK_DSACK_OPT:
24473 optval = 0;
24474 if (rack->rc_rack_tmr_std_based) {
24475 optval |= 1;
24476 }
24477 if (rack->rc_rack_use_dsack) {
24478 optval |= 2;
24479 }
24480 break;
24481 case TCP_RACK_ENABLE_HYSTART:
24482 {
24483 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24484 optval = RACK_HYSTART_ON;
24485 if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24486 optval = RACK_HYSTART_ON_W_SC;
24487 if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24488 optval = RACK_HYSTART_ON_W_SC_C;
24489 } else {
24490 optval = RACK_HYSTART_OFF;
24491 }
24492 }
24493 break;
24494 case TCP_RACK_DGP_IN_REC:
24495 error = EINVAL;
24496 break;
24497 case TCP_RACK_HI_BETA:
24498 optval = rack->rack_hibeta;
24499 break;
24500 case TCP_DEFER_OPTIONS:
24501 optval = rack->defer_options;
24502 break;
24503 case TCP_RACK_MEASURE_CNT:
24504 optval = rack->r_ctl.req_measurements;
24505 break;
24506 case TCP_REC_ABC_VAL:
24507 optval = rack->r_use_labc_for_rec;
24508 break;
24509 case TCP_RACK_ABC_VAL:
24510 optval = rack->rc_labc;
24511 break;
24512 case TCP_HDWR_UP_ONLY:
24513 optval= rack->r_up_only;
24514 break;
24515 case TCP_FILLCW_RATE_CAP:
24516 loptval = rack->r_ctl.fillcw_cap;
24517 break;
24518 case TCP_PACING_RATE_CAP:
24519 loptval = rack->r_ctl.bw_rate_cap;
24520 break;
24521 case TCP_RACK_PROFILE:
24522 /* You cannot retrieve a profile, its write only */
24523 error = EINVAL;
24524 break;
24525 case TCP_SIDECHAN_DIS:
24526 optval = rack->r_ctl.side_chan_dis_mask;
24527 break;
24528 case TCP_HYBRID_PACING:
24529 /* You cannot retrieve hybrid pacing information, its write only */
24530 error = EINVAL;
24531 break;
24532 case TCP_USE_CMP_ACKS:
24533 optval = rack->r_use_cmp_ack;
24534 break;
24535 case TCP_RACK_PACE_TO_FILL:
24536 optval = rack->rc_pace_to_cwnd;
24537 break;
24538 case TCP_RACK_NO_PUSH_AT_MAX:
24539 optval = rack->r_ctl.rc_no_push_at_mrtt;
24540 break;
24541 case TCP_SHARED_CWND_ENABLE:
24542 optval = rack->rack_enable_scwnd;
24543 break;
24544 case TCP_RACK_NONRXT_CFG_RATE:
24545 optval = rack->rack_rec_nonrxt_use_cr;
24546 break;
24547 case TCP_NO_PRR:
24548 if (rack->rack_no_prr == 1)
24549 optval = 1;
24550 else if (rack->no_prr_addback == 1)
24551 optval = 2;
24552 else
24553 optval = 0;
24554 break;
24555 case TCP_GP_USE_LTBW:
24556 if (rack->dis_lt_bw) {
24557 /* It is not used */
24558 optval = 0;
24559 } else if (rack->use_lesser_lt_bw) {
24560 /* we use min() */
24561 optval = 1;
24562 } else {
24563 /* we use max() */
24564 optval = 2;
24565 }
24566 break;
24567 case TCP_RACK_DO_DETECTION:
24568 error = EINVAL;
24569 break;
24570 case TCP_RACK_MBUF_QUEUE:
24571 /* Now do we use the LRO mbuf-queue feature */
24572 optval = rack->r_mbuf_queue;
24573 break;
24574 case RACK_CSPR_IS_FCC:
24575 optval = rack->cspr_is_fcc;
24576 break;
24577 case TCP_TIMELY_DYN_ADJ:
24578 optval = rack->rc_gp_dyn_mul;
24579 break;
24580 case TCP_BBR_IWINTSO:
24581 error = EINVAL;
24582 break;
24583 case TCP_RACK_TLP_REDUCE:
24584 /* RACK TLP cwnd reduction (bool) */
24585 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24586 break;
24587 case TCP_BBR_RACK_INIT_RATE:
24588 val = rack->r_ctl.init_rate;
24589 /* convert to kbits per sec */
24590 val *= 8;
24591 val /= 1000;
24592 optval = (uint32_t)val;
24593 break;
24594 case TCP_RACK_FORCE_MSEG:
24595 optval = rack->rc_force_max_seg;
24596 break;
24597 case TCP_RACK_PACE_MIN_SEG:
24598 optval = rack->r_ctl.rc_user_set_min_segs;
24599 break;
24600 case TCP_RACK_PACE_MAX_SEG:
24601 /* Max segments in a pace */
24602 optval = rack->rc_user_set_max_segs;
24603 break;
24604 case TCP_RACK_PACE_ALWAYS:
24605 /* Use the always pace method */
24606 optval = rack->rc_always_pace;
24607 break;
24608 case TCP_RACK_PRR_SENDALOT:
24609 /* Allow PRR to send more than one seg */
24610 optval = rack->r_ctl.rc_prr_sendalot;
24611 break;
24612 case TCP_RACK_MIN_TO:
24613 /* Minimum time between rack t-o's in ms */
24614 optval = rack->r_ctl.rc_min_to;
24615 break;
24616 case TCP_RACK_SPLIT_LIMIT:
24617 optval = rack->r_ctl.rc_split_limit;
24618 break;
24619 case TCP_RACK_EARLY_SEG:
24620 /* If early recovery max segments */
24621 optval = rack->r_ctl.rc_early_recovery_segs;
24622 break;
24623 case TCP_RACK_REORD_THRESH:
24624 /* RACK reorder threshold (shift amount) */
24625 optval = rack->r_ctl.rc_reorder_shift;
24626 break;
24627 case TCP_SS_EEXIT:
24628 if (rack->r_ctl.gp_rnd_thresh) {
24629 uint32_t v;
24630
24631 v = rack->r_ctl.gp_gain_req;
24632 v <<= 17;
24633 optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
24634 if (rack->r_ctl.gate_to_fs == 1)
24635 optval |= 0x10000;
24636 } else
24637 optval = 0;
24638 break;
24639 case TCP_RACK_REORD_FADE:
24640 /* Does reordering fade after ms time */
24641 optval = rack->r_ctl.rc_reorder_fade;
24642 break;
24643 case TCP_BBR_USE_RACK_RR:
24644 /* Do we use the rack cheat for rxt */
24645 optval = rack->use_rack_rr;
24646 break;
24647 case TCP_RACK_RR_CONF:
24648 optval = rack->r_rr_config;
24649 break;
24650 case TCP_HDWR_RATE_CAP:
24651 optval = rack->r_rack_hw_rate_caps;
24652 break;
24653 case TCP_BBR_HDWR_PACE:
24654 optval = rack->rack_hdw_pace_ena;
24655 break;
24656 case TCP_RACK_TLP_THRESH:
24657 /* RACK TLP theshold i.e. srtt+(srtt/N) */
24658 optval = rack->r_ctl.rc_tlp_threshold;
24659 break;
24660 case TCP_RACK_PKT_DELAY:
24661 /* RACK added ms i.e. rack-rtt + reord + N */
24662 optval = rack->r_ctl.rc_pkt_delay;
24663 break;
24664 case TCP_RACK_TLP_USE:
24665 optval = rack->rack_tlp_threshold_use;
24666 break;
24667 case TCP_PACING_DND:
24668 optval = rack->rc_pace_dnd;
24669 break;
24670 case TCP_RACK_PACE_RATE_CA:
24671 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24672 break;
24673 case TCP_RACK_PACE_RATE_SS:
24674 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24675 break;
24676 case TCP_RACK_PACE_RATE_REC:
24677 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24678 break;
24679 case TCP_DGP_UPPER_BOUNDS:
24680 optval = rack->r_ctl.rack_per_upper_bound_ss;
24681 optval <<= 16;
24682 optval |= rack->r_ctl.rack_per_upper_bound_ca;
24683 break;
24684 case TCP_RACK_GP_INCREASE_SS:
24685 optval = rack->r_ctl.rack_per_of_gp_ca;
24686 break;
24687 case TCP_RACK_GP_INCREASE_CA:
24688 optval = rack->r_ctl.rack_per_of_gp_ss;
24689 break;
24690 case TCP_RACK_PACING_DIVISOR:
24691 optval = rack->r_ctl.pace_len_divisor;
24692 break;
24693 case TCP_BBR_RACK_RTT_USE:
24694 optval = rack->r_ctl.rc_rate_sample_method;
24695 break;
24696 case TCP_DELACK:
24697 optval = tp->t_delayed_ack;
24698 break;
24699 case TCP_DATA_AFTER_CLOSE:
24700 optval = rack->rc_allow_data_af_clo;
24701 break;
24702 case TCP_SHARED_CWND_TIME_LIMIT:
24703 optval = rack->r_limit_scw;
24704 break;
24705 case TCP_HONOR_HPTS_MIN:
24706 if (rack->r_use_hpts_min)
24707 optval = rack->r_ctl.max_reduction;
24708 else
24709 optval = 0;
24710 break;
24711 case TCP_REC_IS_DYN:
24712 optval = rack->rc_gp_no_rec_chg;
24713 break;
24714 case TCP_NO_TIMELY:
24715 optval = rack->rc_skip_timely;
24716 break;
24717 case TCP_RACK_TIMER_SLOP:
24718 optval = rack->r_ctl.timer_slop;
24719 break;
24720 default:
24721 return (tcp_default_ctloutput(tp, sopt));
24722 break;
24723 }
24724 INP_WUNLOCK(inp);
24725 if (error == 0) {
24726 if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24727 (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
24728 error = sooptcopyout(sopt, &loptval, sizeof loptval);
24729 else
24730 error = sooptcopyout(sopt, &optval, sizeof optval);
24731 }
24732 return (error);
24733 }
24734
24735 static int
rack_ctloutput(struct tcpcb * tp,struct sockopt * sopt)24736 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24737 {
24738 if (sopt->sopt_dir == SOPT_SET) {
24739 return (rack_set_sockopt(tp, sopt));
24740 } else if (sopt->sopt_dir == SOPT_GET) {
24741 return (rack_get_sockopt(tp, sopt));
24742 } else {
24743 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24744 }
24745 }
24746
24747 static const char *rack_stack_names[] = {
24748 __XSTRING(STACKNAME),
24749 #ifdef STACKALIAS
24750 __XSTRING(STACKALIAS),
24751 #endif
24752 };
24753
24754 static int
rack_ctor(void * mem,int32_t size,void * arg,int32_t how)24755 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24756 {
24757 memset(mem, 0, size);
24758 return (0);
24759 }
24760
24761 static void
rack_dtor(void * mem,int32_t size,void * arg)24762 rack_dtor(void *mem, int32_t size, void *arg)
24763 {
24764
24765 }
24766
24767 static bool rack_mod_inited = false;
24768
24769 static int
tcp_addrack(module_t mod,int32_t type,void * data)24770 tcp_addrack(module_t mod, int32_t type, void *data)
24771 {
24772 int32_t err = 0;
24773 int num_stacks;
24774
24775 switch (type) {
24776 case MOD_LOAD:
24777 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24778 sizeof(struct rack_sendmap),
24779 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24780
24781 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24782 sizeof(struct tcp_rack),
24783 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24784
24785 sysctl_ctx_init(&rack_sysctl_ctx);
24786 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24787 SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24788 OID_AUTO,
24789 #ifdef STACKALIAS
24790 __XSTRING(STACKALIAS),
24791 #else
24792 __XSTRING(STACKNAME),
24793 #endif
24794 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24795 "");
24796 if (rack_sysctl_root == NULL) {
24797 printf("Failed to add sysctl node\n");
24798 err = EFAULT;
24799 goto free_uma;
24800 }
24801 rack_init_sysctls();
24802 num_stacks = nitems(rack_stack_names);
24803 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24804 rack_stack_names, &num_stacks);
24805 if (err) {
24806 printf("Failed to register %s stack name for "
24807 "%s module\n", rack_stack_names[num_stacks],
24808 __XSTRING(MODNAME));
24809 sysctl_ctx_free(&rack_sysctl_ctx);
24810 free_uma:
24811 uma_zdestroy(rack_zone);
24812 uma_zdestroy(rack_pcb_zone);
24813 rack_counter_destroy();
24814 printf("Failed to register rack module -- err:%d\n", err);
24815 return (err);
24816 }
24817 tcp_lro_reg_mbufq();
24818 rack_mod_inited = true;
24819 break;
24820 case MOD_QUIESCE:
24821 err = deregister_tcp_functions(&__tcp_rack, true, false);
24822 break;
24823 case MOD_UNLOAD:
24824 err = deregister_tcp_functions(&__tcp_rack, false, true);
24825 if (err == EBUSY)
24826 break;
24827 if (rack_mod_inited) {
24828 uma_zdestroy(rack_zone);
24829 uma_zdestroy(rack_pcb_zone);
24830 sysctl_ctx_free(&rack_sysctl_ctx);
24831 rack_counter_destroy();
24832 rack_mod_inited = false;
24833 }
24834 tcp_lro_dereg_mbufq();
24835 err = 0;
24836 break;
24837 default:
24838 return (EOPNOTSUPP);
24839 }
24840 return (err);
24841 }
24842
24843 static moduledata_t tcp_rack = {
24844 .name = __XSTRING(MODNAME),
24845 .evhand = tcp_addrack,
24846 .priv = 0
24847 };
24848
24849 MODULE_VERSION(MODNAME, 1);
24850 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24851 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24852
24853 #endif /* #if !defined(INET) && !defined(INET6) */
24854