1 /*-
2 * Copyright (c) 2016-2020 Netflix, Inc.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
12 *
13 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23 * SUCH DAMAGE.
24 *
25 */
26
27 #include <sys/cdefs.h>
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30 #include "opt_ipsec.h"
31 #include "opt_ratelimit.h"
32 #include "opt_kern_tls.h"
33 #if defined(INET) || defined(INET6)
34 #include <sys/param.h>
35 #include <sys/arb.h>
36 #include <sys/module.h>
37 #include <sys/kernel.h>
38 #ifdef TCP_HHOOK
39 #include <sys/hhook.h>
40 #endif
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/mutex.h>
44 #include <sys/mbuf.h>
45 #include <sys/proc.h> /* for proc0 declaration */
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 #ifdef STATS
51 #include <sys/qmath.h>
52 #include <sys/tree.h>
53 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
54 #else
55 #include <sys/tree.h>
56 #endif
57 #include <sys/refcount.h>
58 #include <sys/queue.h>
59 #include <sys/tim_filter.h>
60 #include <sys/smp.h>
61 #include <sys/kthread.h>
62 #include <sys/kern_prefetch.h>
63 #include <sys/protosw.h>
64 #ifdef TCP_ACCOUNTING
65 #include <sys/sched.h>
66 #include <machine/cpu.h>
67 #endif
68 #include <vm/uma.h>
69
70 #include <net/route.h>
71 #include <net/route/nhop.h>
72 #include <net/vnet.h>
73
74 #define TCPSTATES /* for logging */
75
76 #include <netinet/in.h>
77 #include <netinet/in_kdtrace.h>
78 #include <netinet/in_pcb.h>
79 #include <netinet/ip.h>
80 #include <netinet/ip_icmp.h> /* required for icmp_var.h */
81 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
82 #include <netinet/ip_var.h>
83 #include <netinet/ip6.h>
84 #include <netinet6/in6_pcb.h>
85 #include <netinet6/ip6_var.h>
86 #include <netinet/tcp.h>
87 #define TCPOUTFLAGS
88 #include <netinet/tcp_fsm.h>
89 #include <netinet/tcp_seq.h>
90 #include <netinet/tcp_timer.h>
91 #include <netinet/tcp_var.h>
92 #include <netinet/tcp_log_buf.h>
93 #include <netinet/tcp_syncache.h>
94 #include <netinet/tcp_hpts.h>
95 #include <netinet/tcp_ratelimit.h>
96 #include <netinet/tcp_accounting.h>
97 #include <netinet/tcpip.h>
98 #include <netinet/cc/cc.h>
99 #include <netinet/cc/cc_newreno.h>
100 #include <netinet/tcp_fastopen.h>
101 #include <netinet/tcp_lro.h>
102 #ifdef NETFLIX_SHARED_CWND
103 #include <netinet/tcp_shared_cwnd.h>
104 #endif
105 #ifdef TCP_OFFLOAD
106 #include <netinet/tcp_offload.h>
107 #endif
108 #ifdef INET6
109 #include <netinet6/tcp6_var.h>
110 #endif
111 #include <netinet/tcp_ecn.h>
112
113 #include <netipsec/ipsec_support.h>
114
115 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
116 #include <netipsec/ipsec.h>
117 #include <netipsec/ipsec6.h>
118 #endif /* IPSEC */
119
120 #include <netinet/udp.h>
121 #include <netinet/udp_var.h>
122 #include <machine/in_cksum.h>
123
124 #ifdef MAC
125 #include <security/mac/mac_framework.h>
126 #endif
127 #include "sack_filter.h"
128 #include "tcp_rack.h"
129 #include "tailq_hash.h"
130 #include "rack_bbr_common.h"
131
132 uma_zone_t rack_zone;
133 uma_zone_t rack_pcb_zone;
134
135 #ifndef TICKS2SBT
136 #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t)))
137 #endif
138
139 VNET_DECLARE(uint32_t, newreno_beta);
140 VNET_DECLARE(uint32_t, newreno_beta_ecn);
141 #define V_newreno_beta VNET(newreno_beta)
142 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
143
144 #define M_TCPFSB __CONCAT(M_TCPFSB, STACKNAME)
145 #define M_TCPDO __CONCAT(M_TCPDO, STACKNAME)
146
147 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb_" __XSTRING(STACKNAME), "TCP fast send block");
148 MALLOC_DEFINE(M_TCPDO, "tcp_do_" __XSTRING(STACKNAME), "TCP deferred options");
149 MALLOC_DEFINE(M_TCPPCM, "tcp_pcm_" __XSTRING(STACKNAME), "TCP PCM measurement information");
150
151 struct sysctl_ctx_list rack_sysctl_ctx;
152 struct sysctl_oid *rack_sysctl_root;
153
154 #define CUM_ACKED 1
155 #define SACKED 2
156
157 /*
158 * The RACK module incorporates a number of
159 * TCP ideas that have been put out into the IETF
160 * over the last few years:
161 * - Matt Mathis's Rate Halving which slowly drops
162 * the congestion window so that the ack clock can
163 * be maintained during a recovery.
164 * - Yuchung Cheng's RACK TCP (for which its named) that
165 * will stop us using the number of dup acks and instead
166 * use time as the gage of when we retransmit.
167 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
168 * of Dukkipati et.al.
169 * RACK depends on SACK, so if an endpoint arrives that
170 * cannot do SACK the state machine below will shuttle the
171 * connection back to using the "default" TCP stack that is
172 * in FreeBSD.
173 *
174 * To implement RACK the original TCP stack was first decomposed
175 * into a functional state machine with individual states
176 * for each of the possible TCP connection states. The do_segment
177 * functions role in life is to mandate the connection supports SACK
178 * initially and then assure that the RACK state matches the conenction
179 * state before calling the states do_segment function. Each
180 * state is simplified due to the fact that the original do_segment
181 * has been decomposed and we *know* what state we are in (no
182 * switches on the state) and all tests for SACK are gone. This
183 * greatly simplifies what each state does.
184 *
185 * TCP output is also over-written with a new version since it
186 * must maintain the new rack scoreboard.
187 *
188 */
189 static int32_t rack_tlp_thresh = 1;
190 static int32_t rack_tlp_limit = 2; /* No more than 2 TLPs w-out new data */
191 static int32_t rack_tlp_use_greater = 1;
192 static int32_t rack_reorder_thresh = 2;
193 static int32_t rack_reorder_fade = 60000000; /* 0 - never fade, def 60,000,000
194 * - 60 seconds */
195 static uint32_t rack_pcm_every_n_rounds = 100;
196 static uint32_t rack_pcm_blast = 0;
197 static uint32_t rack_pcm_is_enabled = 1;
198 static uint8_t rack_ssthresh_rest_rto_rec = 0; /* Do we restore ssthresh when we have rec -> rto -> rec */
199
200 static uint32_t rack_gp_gain_req = 1200; /* Amount percent wise required to gain to record a round as "gaining" */
201 static uint32_t rack_rnd_cnt_req = 0x10005; /* Default number of rounds if we are below rack_gp_gain_req where we exit ss */
202
203
204 static int32_t rack_rxt_scoreboard_clear_thresh = 2;
205 static int32_t rack_dnd_default = 0; /* For rr_conf = 3, what is the default for dnd */
206 static int32_t rack_rxt_controls = 0;
207 static int32_t rack_fill_cw_state = 0;
208 static uint8_t rack_req_measurements = 1;
209 /* Attack threshold detections */
210 static uint32_t rack_highest_sack_thresh_seen = 0;
211 static uint32_t rack_highest_move_thresh_seen = 0;
212 static uint32_t rack_merge_out_sacks_on_attack = 0;
213 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
214 static int32_t rack_hw_rate_caps = 0; /* 1; */
215 static int32_t rack_hw_rate_cap_per = 0; /* 0 -- off */
216 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
217 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
218 static int32_t rack_hw_up_only = 0;
219 static int32_t rack_stats_gets_ms_rtt = 1;
220 static int32_t rack_prr_addbackmax = 2;
221 static int32_t rack_do_hystart = 0;
222 static int32_t rack_apply_rtt_with_reduced_conf = 0;
223 static int32_t rack_hibeta_setting = 0;
224 static int32_t rack_default_pacing_divisor = 250;
225 static uint16_t rack_pacing_min_seg = 0;
226 static int32_t rack_timely_off = 0;
227
228 static uint32_t sad_seg_size_per = 800; /* 80.0 % */
229 static int32_t rack_pkt_delay = 1000;
230 static int32_t rack_send_a_lot_in_prr = 1;
231 static int32_t rack_min_to = 1000; /* Number of microsecond min timeout */
232 static int32_t rack_verbose_logging = 0;
233 static int32_t rack_ignore_data_after_close = 1;
234 static int32_t rack_enable_shared_cwnd = 1;
235 static int32_t rack_use_cmp_acks = 1;
236 static int32_t rack_use_fsb = 1;
237 static int32_t rack_use_rfo = 1;
238 static int32_t rack_use_rsm_rfo = 1;
239 static int32_t rack_max_abc_post_recovery = 2;
240 static int32_t rack_client_low_buf = 0;
241 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 */
242 static int32_t rack_bw_multipler = 0; /* Limit on fill cw's jump up to be this x gp_est */
243 #ifdef TCP_ACCOUNTING
244 static int32_t rack_tcp_accounting = 0;
245 #endif
246 static int32_t rack_limits_scwnd = 1;
247 static int32_t rack_enable_mqueue_for_nonpaced = 0;
248 static int32_t rack_hybrid_allow_set_maxseg = 0;
249 static int32_t rack_disable_prr = 0;
250 static int32_t use_rack_rr = 1;
251 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
252 static int32_t rack_persist_min = 250000; /* 250usec */
253 static int32_t rack_persist_max = 2000000; /* 2 Second in usec's */
254 static int32_t rack_honors_hpts_min_to = 1; /* Do we honor the hpts minimum time out for pacing timers */
255 static uint32_t rack_max_reduce = 10; /* Percent we can reduce slot by */
256 static int32_t rack_sack_not_required = 1; /* set to one to allow non-sack to use rack */
257 static int32_t rack_limit_time_with_srtt = 0;
258 static int32_t rack_autosndbuf_inc = 20; /* In percentage form */
259 static int32_t rack_enobuf_hw_boost_mult = 0; /* How many times the hw rate we boost slot using time_between */
260 static int32_t rack_enobuf_hw_max = 12000; /* 12 ms in usecs */
261 static int32_t rack_enobuf_hw_min = 10000; /* 10 ms in usecs */
262 static int32_t rack_hw_rwnd_factor = 2; /* How many max_segs the rwnd must be before we hold off sending */
263 static int32_t rack_hw_check_queue = 0; /* Do we always pre-check queue depth of a hw queue */
264
265 /*
266 * Currently regular tcp has a rto_min of 30ms
267 * the backoff goes 12 times so that ends up
268 * being a total of 122.850 seconds before a
269 * connection is killed.
270 */
271 static uint32_t rack_def_data_window = 20;
272 static uint32_t rack_goal_bdp = 2;
273 static uint32_t rack_min_srtts = 1;
274 static uint32_t rack_min_measure_usec = 0;
275 static int32_t rack_tlp_min = 10000; /* 10ms */
276 static int32_t rack_rto_min = 30000; /* 30,000 usec same as main freebsd */
277 static int32_t rack_rto_max = 4000000; /* 4 seconds in usec's */
278 static const int32_t rack_free_cache = 2;
279 static int32_t rack_hptsi_segments = 40;
280 static int32_t rack_rate_sample_method = USE_RTT_LOW;
281 static int32_t rack_pace_every_seg = 0;
282 static int32_t rack_delayed_ack_time = 40000; /* 40ms in usecs */
283 static int32_t rack_slot_reduction = 4;
284 static int32_t rack_wma_divisor = 8; /* For WMA calculation */
285 static int32_t rack_cwnd_block_ends_measure = 0;
286 static int32_t rack_rwnd_block_ends_measure = 0;
287 static int32_t rack_def_profile = 0;
288
289 static int32_t rack_lower_cwnd_at_tlp = 0;
290 static int32_t rack_always_send_oldest = 0;
291 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
292
293 static uint16_t rack_per_of_gp_ss = 250; /* 250 % slow-start */
294 static uint16_t rack_per_of_gp_ca = 200; /* 200 % congestion-avoidance */
295 static uint16_t rack_per_of_gp_rec = 200; /* 200 % of bw */
296
297 /* Probertt */
298 static uint16_t rack_per_of_gp_probertt = 60; /* 60% of bw */
299 static uint16_t rack_per_of_gp_lowthresh = 40; /* 40% is bottom */
300 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
301 static uint16_t rack_atexit_prtt_hbp = 130; /* Clamp to 130% on exit prtt if highly buffered path */
302 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
303
304 static uint32_t rack_max_drain_wait = 2; /* How man gp srtt's before we give up draining */
305 static uint32_t rack_must_drain = 1; /* How many GP srtt's we *must* wait */
306 static uint32_t rack_probertt_use_min_rtt_entry = 1; /* Use the min to calculate the goal else gp_srtt */
307 static uint32_t rack_probertt_use_min_rtt_exit = 0;
308 static uint32_t rack_probe_rtt_sets_cwnd = 0;
309 static uint32_t rack_probe_rtt_safety_val = 2000000; /* No more than 2 sec in probe-rtt */
310 static uint32_t rack_time_between_probertt = 9600000; /* 9.6 sec in usecs */
311 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0; /* How many srtt periods does probe-rtt last top fraction */
312 static uint32_t rack_probertt_gpsrtt_cnt_div = 0; /* How many srtt periods does probe-rtt last bottom fraction */
313 static uint32_t rack_min_probertt_hold = 40000; /* Equal to delayed ack time */
314 static uint32_t rack_probertt_filter_life = 10000000;
315 static uint32_t rack_probertt_lower_within = 10;
316 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds) to count as a lowering */
317 static int32_t rack_pace_one_seg = 0; /* Shall we pace for less than 1.4Meg 1MSS at a time */
318 static int32_t rack_probertt_clear_is = 1;
319 static int32_t rack_max_drain_hbp = 1; /* Extra drain times gpsrtt for highly buffered paths */
320 static int32_t rack_hbp_thresh = 3; /* what is the divisor max_rtt/min_rtt to decided a hbp */
321
322 /* Part of pacing */
323 static int32_t rack_max_per_above = 30; /* When we go to increment stop if above 100+this% */
324
325 /* Timely information:
326 *
327 * Here we have various control parameters on how
328 * timely may change the multiplier. rack_gain_p5_ub
329 * is associated with timely but not directly influencing
330 * the rate decision like the other variables. It controls
331 * the way fill-cw interacts with timely and caps how much
332 * timely can boost the fill-cw b/w.
333 *
334 * The other values are various boost/shrink numbers as well
335 * as potential caps when adjustments are made to the timely
336 * gain (returned by rack_get_output_gain(). Remember too that
337 * the gain returned can be overriden by other factors such as
338 * probeRTT as well as fixed-rate-pacing.
339 */
340 static int32_t rack_gain_p5_ub = 250;
341 static int32_t rack_gp_per_bw_mul_up = 2; /* 2% */
342 static int32_t rack_gp_per_bw_mul_down = 4; /* 4% */
343 static int32_t rack_gp_rtt_maxmul = 3; /* 3 x maxmin */
344 static int32_t rack_gp_rtt_minmul = 1; /* minrtt + (minrtt/mindiv) is lower rtt */
345 static int32_t rack_gp_rtt_mindiv = 4; /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
346 static int32_t rack_gp_decrease_per = 80; /* Beta value of timely decrease (.8) = 80 */
347 static int32_t rack_gp_increase_per = 2; /* 2% increase in multiplier */
348 static int32_t rack_per_lower_bound = 50; /* Don't allow to drop below this multiplier */
349 static int32_t rack_per_upper_bound_ss = 0; /* Don't allow SS to grow above this */
350 static int32_t rack_per_upper_bound_ca = 0; /* Don't allow CA to grow above this */
351 static int32_t rack_do_dyn_mul = 0; /* Are the rack gp multipliers dynamic */
352 static int32_t rack_gp_no_rec_chg = 1; /* Prohibit recovery from reducing it's multiplier */
353 static int32_t rack_timely_dec_clear = 6; /* Do we clear decrement count at a value (6)? */
354 static int32_t rack_timely_max_push_rise = 3; /* One round of pushing */
355 static int32_t rack_timely_max_push_drop = 3; /* Three round of pushing */
356 static int32_t rack_timely_min_segs = 4; /* 4 segment minimum */
357 static int32_t rack_timely_no_stopping = 0;
358 static int32_t rack_down_raise_thresh = 100;
359 static int32_t rack_req_segs = 1;
360 static uint64_t rack_bw_rate_cap = 0;
361 static uint64_t rack_fillcw_bw_cap = 3750000; /* Cap fillcw at 30Mbps */
362
363
364 /* Rack specific counters */
365 counter_u64_t rack_saw_enobuf;
366 counter_u64_t rack_saw_enobuf_hw;
367 counter_u64_t rack_saw_enetunreach;
368 counter_u64_t rack_persists_sends;
369 counter_u64_t rack_persists_acks;
370 counter_u64_t rack_persists_loss;
371 counter_u64_t rack_persists_lost_ends;
372 counter_u64_t rack_total_bytes;
373 #ifdef INVARIANTS
374 counter_u64_t rack_adjust_map_bw;
375 #endif
376 /* Tail loss probe counters */
377 counter_u64_t rack_tlp_tot;
378 counter_u64_t rack_tlp_newdata;
379 counter_u64_t rack_tlp_retran;
380 counter_u64_t rack_tlp_retran_bytes;
381 counter_u64_t rack_to_tot;
382 counter_u64_t rack_hot_alloc;
383 counter_u64_t rack_to_alloc;
384 counter_u64_t rack_to_alloc_hard;
385 counter_u64_t rack_to_alloc_emerg;
386 counter_u64_t rack_to_alloc_limited;
387 counter_u64_t rack_alloc_limited_conns;
388 counter_u64_t rack_split_limited;
389 counter_u64_t rack_rxt_clamps_cwnd;
390 counter_u64_t rack_rxt_clamps_cwnd_uniq;
391
392 counter_u64_t rack_multi_single_eq;
393 counter_u64_t rack_proc_non_comp_ack;
394
395 counter_u64_t rack_fto_send;
396 counter_u64_t rack_fto_rsm_send;
397 counter_u64_t rack_nfto_resend;
398 counter_u64_t rack_non_fto_send;
399 counter_u64_t rack_extended_rfo;
400
401 counter_u64_t rack_sack_proc_all;
402 counter_u64_t rack_sack_proc_short;
403 counter_u64_t rack_sack_proc_restart;
404 counter_u64_t rack_sack_attacks_detected;
405 counter_u64_t rack_sack_attacks_reversed;
406 counter_u64_t rack_sack_attacks_suspect;
407 counter_u64_t rack_sack_used_next_merge;
408 counter_u64_t rack_sack_splits;
409 counter_u64_t rack_sack_used_prev_merge;
410 counter_u64_t rack_sack_skipped_acked;
411 counter_u64_t rack_ack_total;
412 counter_u64_t rack_express_sack;
413 counter_u64_t rack_sack_total;
414 counter_u64_t rack_move_none;
415 counter_u64_t rack_move_some;
416
417 counter_u64_t rack_input_idle_reduces;
418 counter_u64_t rack_collapsed_win;
419 counter_u64_t rack_collapsed_win_seen;
420 counter_u64_t rack_collapsed_win_rxt;
421 counter_u64_t rack_collapsed_win_rxt_bytes;
422 counter_u64_t rack_try_scwnd;
423 counter_u64_t rack_hw_pace_init_fail;
424 counter_u64_t rack_hw_pace_lost;
425
426 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
427 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
428
429
430 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
431
432 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do { \
433 (tv) = (value) + slop; \
434 if ((u_long)(tv) < (u_long)(tvmin)) \
435 (tv) = (tvmin); \
436 if ((u_long)(tv) > (u_long)(tvmax)) \
437 (tv) = (tvmax); \
438 } while (0)
439
440 static void
441 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line);
442
443 static int
444 rack_process_ack(struct mbuf *m, struct tcphdr *th,
445 struct socket *so, struct tcpcb *tp, struct tcpopt *to,
446 uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val, int32_t orig_tlen);
447 static int
448 rack_process_data(struct mbuf *m, struct tcphdr *th,
449 struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
450 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
451 static void
452 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
453 uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
454 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
455 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
456 uint8_t limit_type);
457 static struct rack_sendmap *
458 rack_check_recovery_mode(struct tcpcb *tp,
459 uint32_t tsused);
460 static uint32_t
461 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack);
462 static void
463 rack_cong_signal(struct tcpcb *tp,
464 uint32_t type, uint32_t ack, int );
465 static void rack_counter_destroy(void);
466 static int
467 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
468 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
469 static void
470 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
471 static void
472 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
473 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
474 static void rack_dtor(void *mem, int32_t size, void *arg);
475 static void
476 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
477 uint32_t flex1, uint32_t flex2,
478 uint32_t flex3, uint32_t flex4,
479 uint32_t flex5, uint32_t flex6,
480 uint16_t flex7, uint8_t mod);
481
482 static void
483 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
484 uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
485 struct rack_sendmap *rsm, uint8_t quality);
486 static struct rack_sendmap *
487 rack_find_high_nonack(struct tcp_rack *rack,
488 struct rack_sendmap *rsm);
489 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
490 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
491 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
492 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
493 static void
494 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
495 tcp_seq th_ack, int line, uint8_t quality);
496 static void
497 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
498
499 static uint32_t
500 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
501 static int32_t rack_handoff_ok(struct tcpcb *tp);
502 static int32_t rack_init(struct tcpcb *tp, void **ptr);
503 static void rack_init_sysctls(void);
504
505 static void
506 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
507 struct tcphdr *th, int entered_rec, int dup_ack_struck,
508 int *dsack_seen, int *sacks_seen);
509 static void
510 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
511 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
512 struct rack_sendmap *hintrsm, uint32_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
513
514 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
515
516
517 static void
518 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
519 struct rack_sendmap *rsm, uint32_t cts);
520 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
521 static int32_t rack_output(struct tcpcb *tp);
522
523 static uint32_t
524 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
525 struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
526 uint32_t cts, uint32_t segsiz);
527 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
528 static void rack_remxt_tmr(struct tcpcb *tp);
529 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
530 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
531 static int32_t rack_stopall(struct tcpcb *tp);
532 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
533 static uint32_t
534 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
535 struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint32_t add_flag, int segsiz);
536 static void
537 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
538 struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz);
539 static int
540 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
541 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
542 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
543 static int
544 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
545 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
546 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
547
548 static int
549 rack_do_closing(struct mbuf *m, struct tcphdr *th,
550 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
551 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
552 static int
553 rack_do_established(struct mbuf *m, struct tcphdr *th,
554 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
555 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
556 static int
557 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
558 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
559 int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
560 static int
561 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
562 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
563 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
564 static int
565 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
566 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
567 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
568 static int
569 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
570 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
571 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
572 static int
573 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
574 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
575 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
576 static int
577 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
578 struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
579 int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
580 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
581 struct rack_sendmap *
582 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
583 uint32_t tsused);
584 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
585 uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
586 static void
587 tcp_rack_partialack(struct tcpcb *tp);
588 static int
589 rack_set_profile(struct tcp_rack *rack, int prof);
590 static void
591 rack_apply_deferred_options(struct tcp_rack *rack);
592
593 int32_t rack_clear_counter=0;
594
595 static uint64_t
rack_get_lt_bw(struct tcp_rack * rack)596 rack_get_lt_bw(struct tcp_rack *rack)
597 {
598 struct timeval tv;
599 uint64_t tim, bytes;
600
601 tim = rack->r_ctl.lt_bw_time;
602 bytes = rack->r_ctl.lt_bw_bytes;
603 if (rack->lt_bw_up) {
604 /* Include all the current bytes too */
605 microuptime(&tv);
606 bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
607 tim += (tcp_tv_to_lusec(&tv) - rack->r_ctl.lt_timemark);
608 }
609 if ((bytes != 0) && (tim != 0))
610 return ((bytes * (uint64_t)1000000) / tim);
611 else
612 return (0);
613 }
614
615 static void
rack_swap_beta_values(struct tcp_rack * rack,uint8_t flex8)616 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
617 {
618 struct sockopt sopt;
619 struct cc_newreno_opts opt;
620 struct tcpcb *tp;
621 uint32_t old_beta;
622 uint32_t old_beta_ecn;
623 int error = 0, failed = 0;
624
625 tp = rack->rc_tp;
626 if (tp->t_cc == NULL) {
627 /* Tcb is leaving */
628 return;
629 }
630 rack->rc_pacing_cc_set = 1;
631 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
632 /* Not new-reno we can't play games with beta! */
633 failed = 1;
634 goto out;
635
636 }
637 if (CC_ALGO(tp)->ctl_output == NULL) {
638 /* Huh, not using new-reno so no swaps.? */
639 failed = 2;
640 goto out;
641 }
642 /* Get the current values out */
643 sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
644 sopt.sopt_dir = SOPT_GET;
645 opt.name = CC_NEWRENO_BETA;
646 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
647 if (error) {
648 failed = 3;
649 goto out;
650 }
651 old_beta = opt.val;
652 opt.name = CC_NEWRENO_BETA_ECN;
653 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
654 if (error) {
655 failed = 4;
656 goto out;
657 }
658 old_beta_ecn = opt.val;
659
660 /* Now lets set in the values we have stored */
661 sopt.sopt_dir = SOPT_SET;
662 opt.name = CC_NEWRENO_BETA;
663 opt.val = rack->r_ctl.rc_saved_beta;
664 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
665 if (error) {
666 failed = 5;
667 goto out;
668 }
669 opt.name = CC_NEWRENO_BETA_ECN;
670 opt.val = rack->r_ctl.rc_saved_beta_ecn;
671 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
672 if (error) {
673 failed = 6;
674 goto out;
675 }
676 /* Save off the values for restoral */
677 rack->r_ctl.rc_saved_beta = old_beta;
678 rack->r_ctl.rc_saved_beta_ecn = old_beta_ecn;
679 out:
680 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
681 union tcp_log_stackspecific log;
682 struct timeval tv;
683 struct newreno *ptr;
684
685 ptr = ((struct newreno *)tp->t_ccv.cc_data);
686 memset(&log, 0, sizeof(log));
687 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
688 log.u_bbr.flex1 = ptr->beta;
689 log.u_bbr.flex2 = ptr->beta_ecn;
690 log.u_bbr.flex3 = ptr->newreno_flags;
691 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta;
692 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta_ecn;
693 log.u_bbr.flex6 = failed;
694 log.u_bbr.flex7 = rack->gp_ready;
695 log.u_bbr.flex7 <<= 1;
696 log.u_bbr.flex7 |= rack->use_fixed_rate;
697 log.u_bbr.flex7 <<= 1;
698 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
699 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
700 log.u_bbr.flex8 = flex8;
701 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
702 0, &log, false, NULL, NULL, 0, &tv);
703 }
704 }
705
706 static void
rack_set_cc_pacing(struct tcp_rack * rack)707 rack_set_cc_pacing(struct tcp_rack *rack)
708 {
709 if (rack->rc_pacing_cc_set)
710 return;
711 /*
712 * Use the swap utility placing in 3 for flex8 to id a
713 * set of a new set of values.
714 */
715 rack->rc_pacing_cc_set = 1;
716 rack_swap_beta_values(rack, 3);
717 }
718
719 static void
rack_undo_cc_pacing(struct tcp_rack * rack)720 rack_undo_cc_pacing(struct tcp_rack *rack)
721 {
722 if (rack->rc_pacing_cc_set == 0)
723 return;
724 /*
725 * Use the swap utility placing in 4 for flex8 to id a
726 * restoral of the old values.
727 */
728 rack->rc_pacing_cc_set = 0;
729 rack_swap_beta_values(rack, 4);
730 }
731
732 static void
rack_remove_pacing(struct tcp_rack * rack)733 rack_remove_pacing(struct tcp_rack *rack)
734 {
735 if (rack->rc_pacing_cc_set)
736 rack_undo_cc_pacing(rack);
737 if (rack->r_ctl.pacing_method & RACK_REG_PACING)
738 tcp_decrement_paced_conn();
739 if (rack->r_ctl.pacing_method & RACK_DGP_PACING)
740 tcp_dec_dgp_pacing_cnt();
741 rack->rc_always_pace = 0;
742 rack->r_ctl.pacing_method = RACK_PACING_NONE;
743 rack->dgp_on = 0;
744 rack->rc_hybrid_mode = 0;
745 rack->use_fixed_rate = 0;
746 }
747
748 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)749 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
750 uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
751 {
752 if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
753 union tcp_log_stackspecific log;
754 struct timeval tv;
755
756 memset(&log, 0, sizeof(log));
757 log.u_bbr.flex1 = seq_end;
758 log.u_bbr.flex2 = rack->rc_tp->gput_seq;
759 log.u_bbr.flex3 = ack_end_t;
760 log.u_bbr.flex4 = rack->rc_tp->gput_ts;
761 log.u_bbr.flex5 = send_end_t;
762 log.u_bbr.flex6 = rack->rc_tp->gput_ack;
763 log.u_bbr.flex7 = mode;
764 log.u_bbr.flex8 = 69;
765 log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
766 log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
767 log.u_bbr.pkts_out = line;
768 log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
769 log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
770 log.u_bbr.epoch = rack->r_ctl.current_round;
771 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
772 if (rsm != NULL) {
773 log.u_bbr.applimited = rsm->r_start;
774 log.u_bbr.delivered = rsm->r_end;
775 log.u_bbr.epoch = rsm->r_flags;
776 }
777 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
778 TCP_LOG_EVENTP(rack->rc_tp, NULL,
779 &rack->rc_inp->inp_socket->so_rcv,
780 &rack->rc_inp->inp_socket->so_snd,
781 BBR_LOG_HPTSI_CALC, 0,
782 0, &log, false, &tv);
783 }
784 }
785
786 static int
sysctl_rack_clear(SYSCTL_HANDLER_ARGS)787 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
788 {
789 uint32_t stat;
790 int32_t error;
791
792 error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
793 if (error || req->newptr == NULL)
794 return error;
795
796 error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
797 if (error)
798 return (error);
799 if (stat == 1) {
800 #ifdef INVARIANTS
801 printf("Clearing RACK counters\n");
802 #endif
803 counter_u64_zero(rack_tlp_tot);
804 counter_u64_zero(rack_tlp_newdata);
805 counter_u64_zero(rack_tlp_retran);
806 counter_u64_zero(rack_tlp_retran_bytes);
807 counter_u64_zero(rack_to_tot);
808 counter_u64_zero(rack_saw_enobuf);
809 counter_u64_zero(rack_saw_enobuf_hw);
810 counter_u64_zero(rack_saw_enetunreach);
811 counter_u64_zero(rack_persists_sends);
812 counter_u64_zero(rack_total_bytes);
813 counter_u64_zero(rack_persists_acks);
814 counter_u64_zero(rack_persists_loss);
815 counter_u64_zero(rack_persists_lost_ends);
816 #ifdef INVARIANTS
817 counter_u64_zero(rack_adjust_map_bw);
818 #endif
819 counter_u64_zero(rack_to_alloc_hard);
820 counter_u64_zero(rack_to_alloc_emerg);
821 counter_u64_zero(rack_sack_proc_all);
822 counter_u64_zero(rack_fto_send);
823 counter_u64_zero(rack_fto_rsm_send);
824 counter_u64_zero(rack_extended_rfo);
825 counter_u64_zero(rack_hw_pace_init_fail);
826 counter_u64_zero(rack_hw_pace_lost);
827 counter_u64_zero(rack_non_fto_send);
828 counter_u64_zero(rack_nfto_resend);
829 counter_u64_zero(rack_sack_proc_short);
830 counter_u64_zero(rack_sack_proc_restart);
831 counter_u64_zero(rack_to_alloc);
832 counter_u64_zero(rack_to_alloc_limited);
833 counter_u64_zero(rack_alloc_limited_conns);
834 counter_u64_zero(rack_split_limited);
835 counter_u64_zero(rack_rxt_clamps_cwnd);
836 counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
837 counter_u64_zero(rack_multi_single_eq);
838 counter_u64_zero(rack_proc_non_comp_ack);
839 counter_u64_zero(rack_sack_attacks_detected);
840 counter_u64_zero(rack_sack_attacks_reversed);
841 counter_u64_zero(rack_sack_attacks_suspect);
842 counter_u64_zero(rack_sack_used_next_merge);
843 counter_u64_zero(rack_sack_used_prev_merge);
844 counter_u64_zero(rack_sack_splits);
845 counter_u64_zero(rack_sack_skipped_acked);
846 counter_u64_zero(rack_ack_total);
847 counter_u64_zero(rack_express_sack);
848 counter_u64_zero(rack_sack_total);
849 counter_u64_zero(rack_move_none);
850 counter_u64_zero(rack_move_some);
851 counter_u64_zero(rack_try_scwnd);
852 counter_u64_zero(rack_collapsed_win);
853 counter_u64_zero(rack_collapsed_win_rxt);
854 counter_u64_zero(rack_collapsed_win_seen);
855 counter_u64_zero(rack_collapsed_win_rxt_bytes);
856 } else if (stat == 2) {
857 #ifdef INVARIANTS
858 printf("Clearing RACK option array\n");
859 #endif
860 COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
861 } else if (stat == 3) {
862 printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
863 } else if (stat == 4) {
864 #ifdef INVARIANTS
865 printf("Clearing RACK out size array\n");
866 #endif
867 COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
868 }
869 rack_clear_counter = 0;
870 return (0);
871 }
872
873 static void
rack_init_sysctls(void)874 rack_init_sysctls(void)
875 {
876 struct sysctl_oid *rack_counters;
877 struct sysctl_oid *rack_attack;
878 struct sysctl_oid *rack_pacing;
879 struct sysctl_oid *rack_timely;
880 struct sysctl_oid *rack_timers;
881 struct sysctl_oid *rack_tlp;
882 struct sysctl_oid *rack_misc;
883 struct sysctl_oid *rack_features;
884 struct sysctl_oid *rack_measure;
885 struct sysctl_oid *rack_probertt;
886 struct sysctl_oid *rack_hw_pacing;
887
888 rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
889 SYSCTL_CHILDREN(rack_sysctl_root),
890 OID_AUTO,
891 "sack_attack",
892 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
893 "Rack Sack Attack Counters and Controls");
894 rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
895 SYSCTL_CHILDREN(rack_sysctl_root),
896 OID_AUTO,
897 "stats",
898 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
899 "Rack Counters");
900 SYSCTL_ADD_S32(&rack_sysctl_ctx,
901 SYSCTL_CHILDREN(rack_sysctl_root),
902 OID_AUTO, "rate_sample_method", CTLFLAG_RW,
903 &rack_rate_sample_method , USE_RTT_LOW,
904 "What method should we use for rate sampling 0=high, 1=low ");
905 /* Probe rtt related controls */
906 rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
907 SYSCTL_CHILDREN(rack_sysctl_root),
908 OID_AUTO,
909 "probertt",
910 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
911 "ProbeRTT related Controls");
912 SYSCTL_ADD_U16(&rack_sysctl_ctx,
913 SYSCTL_CHILDREN(rack_probertt),
914 OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
915 &rack_atexit_prtt_hbp, 130,
916 "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
917 SYSCTL_ADD_U16(&rack_sysctl_ctx,
918 SYSCTL_CHILDREN(rack_probertt),
919 OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
920 &rack_atexit_prtt, 130,
921 "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
922 SYSCTL_ADD_U16(&rack_sysctl_ctx,
923 SYSCTL_CHILDREN(rack_probertt),
924 OID_AUTO, "gp_per_mul", CTLFLAG_RW,
925 &rack_per_of_gp_probertt, 60,
926 "What percentage of goodput do we pace at in probertt");
927 SYSCTL_ADD_U16(&rack_sysctl_ctx,
928 SYSCTL_CHILDREN(rack_probertt),
929 OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
930 &rack_per_of_gp_probertt_reduce, 10,
931 "What percentage of goodput do we reduce every gp_srtt");
932 SYSCTL_ADD_U16(&rack_sysctl_ctx,
933 SYSCTL_CHILDREN(rack_probertt),
934 OID_AUTO, "gp_per_low", CTLFLAG_RW,
935 &rack_per_of_gp_lowthresh, 40,
936 "What percentage of goodput do we allow the multiplier to fall to");
937 SYSCTL_ADD_U32(&rack_sysctl_ctx,
938 SYSCTL_CHILDREN(rack_probertt),
939 OID_AUTO, "time_between", CTLFLAG_RW,
940 &rack_time_between_probertt, 96000000,
941 "How many useconds between the lowest rtt falling must past before we enter probertt");
942 SYSCTL_ADD_U32(&rack_sysctl_ctx,
943 SYSCTL_CHILDREN(rack_probertt),
944 OID_AUTO, "safety", CTLFLAG_RW,
945 &rack_probe_rtt_safety_val, 2000000,
946 "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
947 SYSCTL_ADD_U32(&rack_sysctl_ctx,
948 SYSCTL_CHILDREN(rack_probertt),
949 OID_AUTO, "sets_cwnd", CTLFLAG_RW,
950 &rack_probe_rtt_sets_cwnd, 0,
951 "Do we set the cwnd too (if always_lower is on)");
952 SYSCTL_ADD_U32(&rack_sysctl_ctx,
953 SYSCTL_CHILDREN(rack_probertt),
954 OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
955 &rack_max_drain_wait, 2,
956 "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
957 SYSCTL_ADD_U32(&rack_sysctl_ctx,
958 SYSCTL_CHILDREN(rack_probertt),
959 OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
960 &rack_must_drain, 1,
961 "We must drain this many gp_srtt's waiting for flight to reach goal");
962 SYSCTL_ADD_U32(&rack_sysctl_ctx,
963 SYSCTL_CHILDREN(rack_probertt),
964 OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
965 &rack_probertt_use_min_rtt_entry, 1,
966 "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
967 SYSCTL_ADD_U32(&rack_sysctl_ctx,
968 SYSCTL_CHILDREN(rack_probertt),
969 OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
970 &rack_probertt_use_min_rtt_exit, 0,
971 "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
972 SYSCTL_ADD_U32(&rack_sysctl_ctx,
973 SYSCTL_CHILDREN(rack_probertt),
974 OID_AUTO, "length_div", CTLFLAG_RW,
975 &rack_probertt_gpsrtt_cnt_div, 0,
976 "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
977 SYSCTL_ADD_U32(&rack_sysctl_ctx,
978 SYSCTL_CHILDREN(rack_probertt),
979 OID_AUTO, "length_mul", CTLFLAG_RW,
980 &rack_probertt_gpsrtt_cnt_mul, 0,
981 "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
982 SYSCTL_ADD_U32(&rack_sysctl_ctx,
983 SYSCTL_CHILDREN(rack_probertt),
984 OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
985 &rack_min_probertt_hold, 200000,
986 "What is the minimum time we hold probertt at target");
987 SYSCTL_ADD_U32(&rack_sysctl_ctx,
988 SYSCTL_CHILDREN(rack_probertt),
989 OID_AUTO, "filter_life", CTLFLAG_RW,
990 &rack_probertt_filter_life, 10000000,
991 "What is the time for the filters life in useconds");
992 SYSCTL_ADD_U32(&rack_sysctl_ctx,
993 SYSCTL_CHILDREN(rack_probertt),
994 OID_AUTO, "lower_within", CTLFLAG_RW,
995 &rack_probertt_lower_within, 10,
996 "If the rtt goes lower within this percentage of the time, go into probe-rtt");
997 SYSCTL_ADD_U32(&rack_sysctl_ctx,
998 SYSCTL_CHILDREN(rack_probertt),
999 OID_AUTO, "must_move", CTLFLAG_RW,
1000 &rack_min_rtt_movement, 250,
1001 "How much is the minimum movement in rtt to count as a drop for probertt purposes");
1002 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1003 SYSCTL_CHILDREN(rack_probertt),
1004 OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
1005 &rack_probertt_clear_is, 1,
1006 "Do we clear I/S counts on exiting probe-rtt");
1007 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 SYSCTL_CHILDREN(rack_probertt),
1009 OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
1010 &rack_max_drain_hbp, 1,
1011 "How many extra drain gpsrtt's do we get in highly buffered paths");
1012 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1013 SYSCTL_CHILDREN(rack_probertt),
1014 OID_AUTO, "hbp_threshold", CTLFLAG_RW,
1015 &rack_hbp_thresh, 3,
1016 "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
1017 /* Pacing related sysctls */
1018 rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1019 SYSCTL_CHILDREN(rack_sysctl_root),
1020 OID_AUTO,
1021 "pacing",
1022 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1023 "Pacing related Controls");
1024 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1025 SYSCTL_CHILDREN(rack_pacing),
1026 OID_AUTO, "pcm_enabled", CTLFLAG_RW,
1027 &rack_pcm_is_enabled, 1,
1028 "Do we by default do PCM measurements?");
1029 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1030 SYSCTL_CHILDREN(rack_pacing),
1031 OID_AUTO, "pcm_rnds", CTLFLAG_RW,
1032 &rack_pcm_every_n_rounds, 100,
1033 "How many rounds before we need to do a PCM measurement");
1034 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1035 SYSCTL_CHILDREN(rack_pacing),
1036 OID_AUTO, "pcm_blast", CTLFLAG_RW,
1037 &rack_pcm_blast, 0,
1038 "Blast out the full cwnd/rwnd when doing a PCM measurement");
1039 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1040 SYSCTL_CHILDREN(rack_pacing),
1041 OID_AUTO, "rnd_gp_gain", CTLFLAG_RW,
1042 &rack_gp_gain_req, 1200,
1043 "How much do we have to increase the GP to record the round 1200 = 120.0");
1044 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1045 SYSCTL_CHILDREN(rack_pacing),
1046 OID_AUTO, "dgp_out_of_ss_at", CTLFLAG_RW,
1047 &rack_rnd_cnt_req, 0x10005,
1048 "How many rounds less than rnd_gp_gain will drop us out of SS");
1049 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1050 SYSCTL_CHILDREN(rack_pacing),
1051 OID_AUTO, "no_timely", CTLFLAG_RW,
1052 &rack_timely_off, 0,
1053 "Do we not use timely in DGP?");
1054 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1055 SYSCTL_CHILDREN(rack_pacing),
1056 OID_AUTO, "fillcw", CTLFLAG_RW,
1057 &rack_fill_cw_state, 0,
1058 "Enable fillcw on new connections (default=0 off)?");
1059 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1060 SYSCTL_CHILDREN(rack_pacing),
1061 OID_AUTO, "min_burst", CTLFLAG_RW,
1062 &rack_pacing_min_seg, 0,
1063 "What is the min burst size for pacing (0 disables)?");
1064 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 SYSCTL_CHILDREN(rack_pacing),
1066 OID_AUTO, "divisor", CTLFLAG_RW,
1067 &rack_default_pacing_divisor, 250,
1068 "What is the default divisor given to the rl code?");
1069 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 SYSCTL_CHILDREN(rack_pacing),
1071 OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1072 &rack_bw_multipler, 0,
1073 "What is the limit multiplier of the current gp_est that fillcw can increase the b/w too, 200 == 200% (0 = off)?");
1074 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 SYSCTL_CHILDREN(rack_pacing),
1076 OID_AUTO, "max_pace_over", CTLFLAG_RW,
1077 &rack_max_per_above, 30,
1078 "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1079 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1080 SYSCTL_CHILDREN(rack_pacing),
1081 OID_AUTO, "allow1mss", CTLFLAG_RW,
1082 &rack_pace_one_seg, 0,
1083 "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1084 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085 SYSCTL_CHILDREN(rack_pacing),
1086 OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1087 &rack_limit_time_with_srtt, 0,
1088 "Do we limit pacing time based on srtt");
1089 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1090 SYSCTL_CHILDREN(rack_pacing),
1091 OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1092 &rack_per_of_gp_ss, 250,
1093 "If non zero, what percentage of goodput to pace at in slow start");
1094 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1095 SYSCTL_CHILDREN(rack_pacing),
1096 OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1097 &rack_per_of_gp_ca, 150,
1098 "If non zero, what percentage of goodput to pace at in congestion avoidance");
1099 SYSCTL_ADD_U16(&rack_sysctl_ctx,
1100 SYSCTL_CHILDREN(rack_pacing),
1101 OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1102 &rack_per_of_gp_rec, 200,
1103 "If non zero, what percentage of goodput to pace at in recovery");
1104 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 SYSCTL_CHILDREN(rack_pacing),
1106 OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1107 &rack_hptsi_segments, 40,
1108 "What size is the max for TSO segments in pacing and burst mitigation");
1109 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1110 SYSCTL_CHILDREN(rack_pacing),
1111 OID_AUTO, "burst_reduces", CTLFLAG_RW,
1112 &rack_slot_reduction, 4,
1113 "When doing only burst mitigation what is the reduce divisor");
1114 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1115 SYSCTL_CHILDREN(rack_sysctl_root),
1116 OID_AUTO, "use_pacing", CTLFLAG_RW,
1117 &rack_pace_every_seg, 0,
1118 "If set we use pacing, if clear we use only the original burst mitigation");
1119 SYSCTL_ADD_U64(&rack_sysctl_ctx,
1120 SYSCTL_CHILDREN(rack_pacing),
1121 OID_AUTO, "rate_cap", CTLFLAG_RW,
1122 &rack_bw_rate_cap, 0,
1123 "If set we apply this value to the absolute rate cap used by pacing");
1124 SYSCTL_ADD_U64(&rack_sysctl_ctx,
1125 SYSCTL_CHILDREN(rack_pacing),
1126 OID_AUTO, "fillcw_cap", CTLFLAG_RW,
1127 &rack_fillcw_bw_cap, 3750000,
1128 "Do we have an absolute cap on the amount of b/w fillcw can specify (0 = no)?");
1129 SYSCTL_ADD_U8(&rack_sysctl_ctx,
1130 SYSCTL_CHILDREN(rack_sysctl_root),
1131 OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1132 &rack_req_measurements, 1,
1133 "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1134 /* Hardware pacing */
1135 rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1136 SYSCTL_CHILDREN(rack_sysctl_root),
1137 OID_AUTO,
1138 "hdwr_pacing",
1139 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1140 "Pacing related Controls");
1141 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1142 SYSCTL_CHILDREN(rack_hw_pacing),
1143 OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1144 &rack_hw_rwnd_factor, 2,
1145 "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1146 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1147 SYSCTL_CHILDREN(rack_hw_pacing),
1148 OID_AUTO, "precheck", CTLFLAG_RW,
1149 &rack_hw_check_queue, 0,
1150 "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1151 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1152 SYSCTL_CHILDREN(rack_hw_pacing),
1153 OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1154 &rack_enobuf_hw_boost_mult, 0,
1155 "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1156 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1157 SYSCTL_CHILDREN(rack_hw_pacing),
1158 OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1159 &rack_enobuf_hw_max, 2,
1160 "What is the max boost the pacing time if we see a ENOBUFS?");
1161 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162 SYSCTL_CHILDREN(rack_hw_pacing),
1163 OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1164 &rack_enobuf_hw_min, 2,
1165 "What is the min boost the pacing time if we see a ENOBUFS?");
1166 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167 SYSCTL_CHILDREN(rack_hw_pacing),
1168 OID_AUTO, "enable", CTLFLAG_RW,
1169 &rack_enable_hw_pacing, 0,
1170 "Should RACK attempt to use hw pacing?");
1171 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172 SYSCTL_CHILDREN(rack_hw_pacing),
1173 OID_AUTO, "rate_cap", CTLFLAG_RW,
1174 &rack_hw_rate_caps, 0,
1175 "Does the highest hardware pacing rate cap the rate we will send at??");
1176 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177 SYSCTL_CHILDREN(rack_hw_pacing),
1178 OID_AUTO, "uncap_per", CTLFLAG_RW,
1179 &rack_hw_rate_cap_per, 0,
1180 "If you go over b/w by this amount you will be uncapped (0 = never)");
1181 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182 SYSCTL_CHILDREN(rack_hw_pacing),
1183 OID_AUTO, "rate_min", CTLFLAG_RW,
1184 &rack_hw_rate_min, 0,
1185 "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1186 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187 SYSCTL_CHILDREN(rack_hw_pacing),
1188 OID_AUTO, "rate_to_low", CTLFLAG_RW,
1189 &rack_hw_rate_to_low, 0,
1190 "If we fall below this rate, dis-engage hw pacing?");
1191 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192 SYSCTL_CHILDREN(rack_hw_pacing),
1193 OID_AUTO, "up_only", CTLFLAG_RW,
1194 &rack_hw_up_only, 0,
1195 "Do we allow hw pacing to lower the rate selected?");
1196 rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1197 SYSCTL_CHILDREN(rack_sysctl_root),
1198 OID_AUTO,
1199 "timely",
1200 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1201 "Rack Timely RTT Controls");
1202 /* Timely based GP dynmics */
1203 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1204 SYSCTL_CHILDREN(rack_timely),
1205 OID_AUTO, "upper", CTLFLAG_RW,
1206 &rack_gp_per_bw_mul_up, 2,
1207 "Rack timely upper range for equal b/w (in percentage)");
1208 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1209 SYSCTL_CHILDREN(rack_timely),
1210 OID_AUTO, "lower", CTLFLAG_RW,
1211 &rack_gp_per_bw_mul_down, 4,
1212 "Rack timely lower range for equal b/w (in percentage)");
1213 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214 SYSCTL_CHILDREN(rack_timely),
1215 OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1216 &rack_gp_rtt_maxmul, 3,
1217 "Rack timely multiplier of lowest rtt for rtt_max");
1218 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219 SYSCTL_CHILDREN(rack_timely),
1220 OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1221 &rack_gp_rtt_mindiv, 4,
1222 "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1223 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224 SYSCTL_CHILDREN(rack_timely),
1225 OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1226 &rack_gp_rtt_minmul, 1,
1227 "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1228 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 SYSCTL_CHILDREN(rack_timely),
1230 OID_AUTO, "decrease", CTLFLAG_RW,
1231 &rack_gp_decrease_per, 80,
1232 "Rack timely Beta value 80 = .8 (scaled by 100)");
1233 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234 SYSCTL_CHILDREN(rack_timely),
1235 OID_AUTO, "increase", CTLFLAG_RW,
1236 &rack_gp_increase_per, 2,
1237 "Rack timely increase perentage of our GP multiplication factor");
1238 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239 SYSCTL_CHILDREN(rack_timely),
1240 OID_AUTO, "lowerbound", CTLFLAG_RW,
1241 &rack_per_lower_bound, 50,
1242 "Rack timely lowest percentage we allow GP multiplier to fall to");
1243 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244 SYSCTL_CHILDREN(rack_timely),
1245 OID_AUTO, "p5_upper", CTLFLAG_RW,
1246 &rack_gain_p5_ub, 250,
1247 "Profile 5 upper bound to timely gain");
1248
1249 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1250 SYSCTL_CHILDREN(rack_timely),
1251 OID_AUTO, "upperboundss", CTLFLAG_RW,
1252 &rack_per_upper_bound_ss, 0,
1253 "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1254 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1255 SYSCTL_CHILDREN(rack_timely),
1256 OID_AUTO, "upperboundca", CTLFLAG_RW,
1257 &rack_per_upper_bound_ca, 0,
1258 "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1259 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1260 SYSCTL_CHILDREN(rack_timely),
1261 OID_AUTO, "dynamicgp", CTLFLAG_RW,
1262 &rack_do_dyn_mul, 0,
1263 "Rack timely do we enable dynmaic timely goodput by default");
1264 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1265 SYSCTL_CHILDREN(rack_timely),
1266 OID_AUTO, "no_rec_red", CTLFLAG_RW,
1267 &rack_gp_no_rec_chg, 1,
1268 "Rack timely do we prohibit the recovery multiplier from being lowered");
1269 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1270 SYSCTL_CHILDREN(rack_timely),
1271 OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1272 &rack_timely_dec_clear, 6,
1273 "Rack timely what threshold do we count to before another boost during b/w decent");
1274 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1275 SYSCTL_CHILDREN(rack_timely),
1276 OID_AUTO, "max_push_rise", CTLFLAG_RW,
1277 &rack_timely_max_push_rise, 3,
1278 "Rack timely how many times do we push up with b/w increase");
1279 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1280 SYSCTL_CHILDREN(rack_timely),
1281 OID_AUTO, "max_push_drop", CTLFLAG_RW,
1282 &rack_timely_max_push_drop, 3,
1283 "Rack timely how many times do we push back on b/w decent");
1284 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1285 SYSCTL_CHILDREN(rack_timely),
1286 OID_AUTO, "min_segs", CTLFLAG_RW,
1287 &rack_timely_min_segs, 4,
1288 "Rack timely when setting the cwnd what is the min num segments");
1289 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1290 SYSCTL_CHILDREN(rack_timely),
1291 OID_AUTO, "nonstop", CTLFLAG_RW,
1292 &rack_timely_no_stopping, 0,
1293 "Rack timely don't stop increase");
1294 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1295 SYSCTL_CHILDREN(rack_timely),
1296 OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1297 &rack_down_raise_thresh, 100,
1298 "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1299 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1300 SYSCTL_CHILDREN(rack_timely),
1301 OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1302 &rack_req_segs, 1,
1303 "Bottom dragging if not these many segments outstanding and room");
1304
1305 /* TLP and Rack related parameters */
1306 rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1307 SYSCTL_CHILDREN(rack_sysctl_root),
1308 OID_AUTO,
1309 "tlp",
1310 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1311 "TLP and Rack related Controls");
1312 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1313 SYSCTL_CHILDREN(rack_tlp),
1314 OID_AUTO, "use_rrr", CTLFLAG_RW,
1315 &use_rack_rr, 1,
1316 "Do we use Rack Rapid Recovery");
1317 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1318 SYSCTL_CHILDREN(rack_tlp),
1319 OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1320 &rack_max_abc_post_recovery, 2,
1321 "Since we do early recovery, do we override the l_abc to a value, if so what?");
1322 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1323 SYSCTL_CHILDREN(rack_tlp),
1324 OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1325 &rack_non_rxt_use_cr, 0,
1326 "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1327 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1328 SYSCTL_CHILDREN(rack_tlp),
1329 OID_AUTO, "tlpmethod", CTLFLAG_RW,
1330 &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1331 "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1332 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1333 SYSCTL_CHILDREN(rack_tlp),
1334 OID_AUTO, "limit", CTLFLAG_RW,
1335 &rack_tlp_limit, 2,
1336 "How many TLP's can be sent without sending new data");
1337 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1338 SYSCTL_CHILDREN(rack_tlp),
1339 OID_AUTO, "use_greater", CTLFLAG_RW,
1340 &rack_tlp_use_greater, 1,
1341 "Should we use the rack_rtt time if its greater than srtt");
1342 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1343 SYSCTL_CHILDREN(rack_tlp),
1344 OID_AUTO, "tlpminto", CTLFLAG_RW,
1345 &rack_tlp_min, 10000,
1346 "TLP minimum timeout per the specification (in microseconds)");
1347 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1348 SYSCTL_CHILDREN(rack_tlp),
1349 OID_AUTO, "send_oldest", CTLFLAG_RW,
1350 &rack_always_send_oldest, 0,
1351 "Should we always send the oldest TLP and RACK-TLP");
1352 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1353 SYSCTL_CHILDREN(rack_tlp),
1354 OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1355 &rack_lower_cwnd_at_tlp, 0,
1356 "When a TLP completes a retran should we enter recovery");
1357 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1358 SYSCTL_CHILDREN(rack_tlp),
1359 OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1360 &rack_reorder_thresh, 2,
1361 "What factor for rack will be added when seeing reordering (shift right)");
1362 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1363 SYSCTL_CHILDREN(rack_tlp),
1364 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1365 &rack_tlp_thresh, 1,
1366 "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1367 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1368 SYSCTL_CHILDREN(rack_tlp),
1369 OID_AUTO, "reorder_fade", CTLFLAG_RW,
1370 &rack_reorder_fade, 60000000,
1371 "Does reorder detection fade, if so how many microseconds (0 means never)");
1372 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373 SYSCTL_CHILDREN(rack_tlp),
1374 OID_AUTO, "pktdelay", CTLFLAG_RW,
1375 &rack_pkt_delay, 1000,
1376 "Extra RACK time (in microseconds) besides reordering thresh");
1377
1378 /* Timer related controls */
1379 rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1380 SYSCTL_CHILDREN(rack_sysctl_root),
1381 OID_AUTO,
1382 "timers",
1383 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1384 "Timer related controls");
1385 SYSCTL_ADD_U8(&rack_sysctl_ctx,
1386 SYSCTL_CHILDREN(rack_timers),
1387 OID_AUTO, "reset_ssth_rec_rto", CTLFLAG_RW,
1388 &rack_ssthresh_rest_rto_rec, 0,
1389 "When doing recovery -> rto -> recovery do we reset SSthresh?");
1390 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1391 SYSCTL_CHILDREN(rack_timers),
1392 OID_AUTO, "scoreboard_thresh", CTLFLAG_RW,
1393 &rack_rxt_scoreboard_clear_thresh, 2,
1394 "How many RTO's are allowed before we clear the scoreboard");
1395 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1396 SYSCTL_CHILDREN(rack_timers),
1397 OID_AUTO, "honor_hpts_min", CTLFLAG_RW,
1398 &rack_honors_hpts_min_to, 1,
1399 "Do rack pacing timers honor hpts min timeout");
1400 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1401 SYSCTL_CHILDREN(rack_timers),
1402 OID_AUTO, "hpts_max_reduce", CTLFLAG_RW,
1403 &rack_max_reduce, 10,
1404 "Max percentage we will reduce slot by for pacing when we are behind");
1405 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1406 SYSCTL_CHILDREN(rack_timers),
1407 OID_AUTO, "persmin", CTLFLAG_RW,
1408 &rack_persist_min, 250000,
1409 "What is the minimum time in microseconds between persists");
1410 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1411 SYSCTL_CHILDREN(rack_timers),
1412 OID_AUTO, "persmax", CTLFLAG_RW,
1413 &rack_persist_max, 2000000,
1414 "What is the largest delay in microseconds between persists");
1415 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1416 SYSCTL_CHILDREN(rack_timers),
1417 OID_AUTO, "delayed_ack", CTLFLAG_RW,
1418 &rack_delayed_ack_time, 40000,
1419 "Delayed ack time (40ms in microseconds)");
1420 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1421 SYSCTL_CHILDREN(rack_timers),
1422 OID_AUTO, "minrto", CTLFLAG_RW,
1423 &rack_rto_min, 30000,
1424 "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1425 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1426 SYSCTL_CHILDREN(rack_timers),
1427 OID_AUTO, "maxrto", CTLFLAG_RW,
1428 &rack_rto_max, 4000000,
1429 "Maximum RTO in microseconds -- should be at least as large as min_rto");
1430 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1431 SYSCTL_CHILDREN(rack_timers),
1432 OID_AUTO, "minto", CTLFLAG_RW,
1433 &rack_min_to, 1000,
1434 "Minimum rack timeout in microseconds");
1435 /* Measure controls */
1436 rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1437 SYSCTL_CHILDREN(rack_sysctl_root),
1438 OID_AUTO,
1439 "measure",
1440 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1441 "Measure related controls");
1442 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443 SYSCTL_CHILDREN(rack_measure),
1444 OID_AUTO, "wma_divisor", CTLFLAG_RW,
1445 &rack_wma_divisor, 8,
1446 "When doing b/w calculation what is the divisor for the WMA");
1447 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448 SYSCTL_CHILDREN(rack_measure),
1449 OID_AUTO, "end_cwnd", CTLFLAG_RW,
1450 &rack_cwnd_block_ends_measure, 0,
1451 "Does a cwnd just-return end the measurement window (app limited)");
1452 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453 SYSCTL_CHILDREN(rack_measure),
1454 OID_AUTO, "end_rwnd", CTLFLAG_RW,
1455 &rack_rwnd_block_ends_measure, 0,
1456 "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1457 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1458 SYSCTL_CHILDREN(rack_measure),
1459 OID_AUTO, "min_target", CTLFLAG_RW,
1460 &rack_def_data_window, 20,
1461 "What is the minimum target window (in mss) for a GP measurements");
1462 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1463 SYSCTL_CHILDREN(rack_measure),
1464 OID_AUTO, "goal_bdp", CTLFLAG_RW,
1465 &rack_goal_bdp, 2,
1466 "What is the goal BDP to measure");
1467 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1468 SYSCTL_CHILDREN(rack_measure),
1469 OID_AUTO, "min_srtts", CTLFLAG_RW,
1470 &rack_min_srtts, 1,
1471 "What is the goal BDP to measure");
1472 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1473 SYSCTL_CHILDREN(rack_measure),
1474 OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1475 &rack_min_measure_usec, 0,
1476 "What is the Minimum time time for a measurement if 0, this is off");
1477 /* Features */
1478 rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1479 SYSCTL_CHILDREN(rack_sysctl_root),
1480 OID_AUTO,
1481 "features",
1482 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1483 "Feature controls");
1484 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1485 SYSCTL_CHILDREN(rack_features),
1486 OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1487 &rack_hybrid_allow_set_maxseg, 0,
1488 "Should hybrid pacing allow the setmss command");
1489 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1490 SYSCTL_CHILDREN(rack_features),
1491 OID_AUTO, "cmpack", CTLFLAG_RW,
1492 &rack_use_cmp_acks, 1,
1493 "Should RACK have LRO send compressed acks");
1494 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1495 SYSCTL_CHILDREN(rack_features),
1496 OID_AUTO, "fsb", CTLFLAG_RW,
1497 &rack_use_fsb, 1,
1498 "Should RACK use the fast send block?");
1499 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1500 SYSCTL_CHILDREN(rack_features),
1501 OID_AUTO, "rfo", CTLFLAG_RW,
1502 &rack_use_rfo, 1,
1503 "Should RACK use rack_fast_output()?");
1504 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1505 SYSCTL_CHILDREN(rack_features),
1506 OID_AUTO, "rsmrfo", CTLFLAG_RW,
1507 &rack_use_rsm_rfo, 1,
1508 "Should RACK use rack_fast_rsm_output()?");
1509 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1510 SYSCTL_CHILDREN(rack_features),
1511 OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1512 &rack_enable_mqueue_for_nonpaced, 0,
1513 "Should RACK use mbuf queuing for non-paced connections");
1514 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1515 SYSCTL_CHILDREN(rack_features),
1516 OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1517 &rack_do_hystart, 0,
1518 "Should RACK enable HyStart++ on connections?");
1519 /* Misc rack controls */
1520 rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1521 SYSCTL_CHILDREN(rack_sysctl_root),
1522 OID_AUTO,
1523 "misc",
1524 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1525 "Misc related controls");
1526 #ifdef TCP_ACCOUNTING
1527 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1528 SYSCTL_CHILDREN(rack_misc),
1529 OID_AUTO, "tcp_acct", CTLFLAG_RW,
1530 &rack_tcp_accounting, 0,
1531 "Should we turn on TCP accounting for all rack sessions?");
1532 #endif
1533 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1534 SYSCTL_CHILDREN(rack_misc),
1535 OID_AUTO, "dnd", CTLFLAG_RW,
1536 &rack_dnd_default, 0,
1537 "Do not disturb default for rack_rrr = 3");
1538 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1539 SYSCTL_CHILDREN(rack_misc),
1540 OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1541 &sad_seg_size_per, 800,
1542 "Percentage of segment size needed in a sack 800 = 80.0?");
1543 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1544 SYSCTL_CHILDREN(rack_misc),
1545 OID_AUTO, "rxt_controls", CTLFLAG_RW,
1546 &rack_rxt_controls, 0,
1547 "Retransmit sending size controls (valid values 0, 1, 2 default=1)?");
1548 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1549 SYSCTL_CHILDREN(rack_misc),
1550 OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1551 &rack_hibeta_setting, 0,
1552 "Do we ue a high beta (80 instead of 50)?");
1553 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1554 SYSCTL_CHILDREN(rack_misc),
1555 OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1556 &rack_apply_rtt_with_reduced_conf, 0,
1557 "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1558 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1559 SYSCTL_CHILDREN(rack_misc),
1560 OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1561 &rack_dsack_std_based, 3,
1562 "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1563 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1564 SYSCTL_CHILDREN(rack_misc),
1565 OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1566 &rack_prr_addbackmax, 2,
1567 "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1568 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1569 SYSCTL_CHILDREN(rack_misc),
1570 OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1571 &rack_stats_gets_ms_rtt, 1,
1572 "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1573 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1574 SYSCTL_CHILDREN(rack_misc),
1575 OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1576 &rack_client_low_buf, 0,
1577 "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1578 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1579 SYSCTL_CHILDREN(rack_misc),
1580 OID_AUTO, "defprofile", CTLFLAG_RW,
1581 &rack_def_profile, 0,
1582 "Should RACK use a default profile (0=no, num == profile num)?");
1583 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1584 SYSCTL_CHILDREN(rack_misc),
1585 OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1586 &rack_enable_shared_cwnd, 1,
1587 "Should RACK try to use the shared cwnd on connections where allowed");
1588 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1589 SYSCTL_CHILDREN(rack_misc),
1590 OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1591 &rack_limits_scwnd, 1,
1592 "Should RACK place low end time limits on the shared cwnd feature");
1593 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1594 SYSCTL_CHILDREN(rack_misc),
1595 OID_AUTO, "no_prr", CTLFLAG_RW,
1596 &rack_disable_prr, 0,
1597 "Should RACK not use prr and only pace (must have pacing on)");
1598 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1599 SYSCTL_CHILDREN(rack_misc),
1600 OID_AUTO, "bb_verbose", CTLFLAG_RW,
1601 &rack_verbose_logging, 0,
1602 "Should RACK black box logging be verbose");
1603 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1604 SYSCTL_CHILDREN(rack_misc),
1605 OID_AUTO, "data_after_close", CTLFLAG_RW,
1606 &rack_ignore_data_after_close, 1,
1607 "Do we hold off sending a RST until all pending data is ack'd");
1608 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1609 SYSCTL_CHILDREN(rack_misc),
1610 OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1611 &rack_sack_not_required, 1,
1612 "Do we allow rack to run on connections not supporting SACK");
1613 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1614 SYSCTL_CHILDREN(rack_misc),
1615 OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1616 &rack_send_a_lot_in_prr, 1,
1617 "Send a lot in prr");
1618 SYSCTL_ADD_S32(&rack_sysctl_ctx,
1619 SYSCTL_CHILDREN(rack_misc),
1620 OID_AUTO, "autoscale", CTLFLAG_RW,
1621 &rack_autosndbuf_inc, 20,
1622 "What percentage should rack scale up its snd buffer by?");
1623
1624
1625 /* Sack Attacker detection stuff */
1626 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1627 SYSCTL_CHILDREN(rack_attack),
1628 OID_AUTO, "merge_out", CTLFLAG_RW,
1629 &rack_merge_out_sacks_on_attack, 0,
1630 "Do we merge the sendmap when we decide we are being attacked?");
1631
1632 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1633 SYSCTL_CHILDREN(rack_attack),
1634 OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1635 &rack_highest_sack_thresh_seen, 0,
1636 "Highest sack to ack ratio seen");
1637 SYSCTL_ADD_U32(&rack_sysctl_ctx,
1638 SYSCTL_CHILDREN(rack_attack),
1639 OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1640 &rack_highest_move_thresh_seen, 0,
1641 "Highest move to non-move ratio seen");
1642 rack_ack_total = counter_u64_alloc(M_WAITOK);
1643 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1644 SYSCTL_CHILDREN(rack_attack),
1645 OID_AUTO, "acktotal", CTLFLAG_RD,
1646 &rack_ack_total,
1647 "Total number of Ack's");
1648 rack_express_sack = counter_u64_alloc(M_WAITOK);
1649 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650 SYSCTL_CHILDREN(rack_attack),
1651 OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1652 &rack_express_sack,
1653 "Total expresss number of Sack's");
1654 rack_sack_total = counter_u64_alloc(M_WAITOK);
1655 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1656 SYSCTL_CHILDREN(rack_attack),
1657 OID_AUTO, "sacktotal", CTLFLAG_RD,
1658 &rack_sack_total,
1659 "Total number of SACKs");
1660 rack_move_none = counter_u64_alloc(M_WAITOK);
1661 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662 SYSCTL_CHILDREN(rack_attack),
1663 OID_AUTO, "move_none", CTLFLAG_RD,
1664 &rack_move_none,
1665 "Total number of SACK index reuse of positions under threshold");
1666 rack_move_some = counter_u64_alloc(M_WAITOK);
1667 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1668 SYSCTL_CHILDREN(rack_attack),
1669 OID_AUTO, "move_some", CTLFLAG_RD,
1670 &rack_move_some,
1671 "Total number of SACK index reuse of positions over threshold");
1672 rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1673 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674 SYSCTL_CHILDREN(rack_attack),
1675 OID_AUTO, "attacks", CTLFLAG_RD,
1676 &rack_sack_attacks_detected,
1677 "Total number of SACK attackers that had sack disabled");
1678 rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1679 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 SYSCTL_CHILDREN(rack_attack),
1681 OID_AUTO, "reversed", CTLFLAG_RD,
1682 &rack_sack_attacks_reversed,
1683 "Total number of SACK attackers that were later determined false positive");
1684 rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1685 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686 SYSCTL_CHILDREN(rack_attack),
1687 OID_AUTO, "suspect", CTLFLAG_RD,
1688 &rack_sack_attacks_suspect,
1689 "Total number of SACKs that triggered early detection");
1690
1691 rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1692 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693 SYSCTL_CHILDREN(rack_attack),
1694 OID_AUTO, "nextmerge", CTLFLAG_RD,
1695 &rack_sack_used_next_merge,
1696 "Total number of times we used the next merge");
1697 rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1698 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1699 SYSCTL_CHILDREN(rack_attack),
1700 OID_AUTO, "prevmerge", CTLFLAG_RD,
1701 &rack_sack_used_prev_merge,
1702 "Total number of times we used the prev merge");
1703 /* Counters */
1704 rack_total_bytes = counter_u64_alloc(M_WAITOK);
1705 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1706 SYSCTL_CHILDREN(rack_counters),
1707 OID_AUTO, "totalbytes", CTLFLAG_RD,
1708 &rack_total_bytes,
1709 "Total number of bytes sent");
1710 rack_fto_send = counter_u64_alloc(M_WAITOK);
1711 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1712 SYSCTL_CHILDREN(rack_counters),
1713 OID_AUTO, "fto_send", CTLFLAG_RD,
1714 &rack_fto_send, "Total number of rack_fast_output sends");
1715 rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1716 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1717 SYSCTL_CHILDREN(rack_counters),
1718 OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1719 &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1720 rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1721 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1722 SYSCTL_CHILDREN(rack_counters),
1723 OID_AUTO, "nfto_resend", CTLFLAG_RD,
1724 &rack_nfto_resend, "Total number of rack_output retransmissions");
1725 rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1726 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1727 SYSCTL_CHILDREN(rack_counters),
1728 OID_AUTO, "nfto_send", CTLFLAG_RD,
1729 &rack_non_fto_send, "Total number of rack_output first sends");
1730 rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1731 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732 SYSCTL_CHILDREN(rack_counters),
1733 OID_AUTO, "rfo_extended", CTLFLAG_RD,
1734 &rack_extended_rfo, "Total number of times we extended rfo");
1735
1736 rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1737 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 SYSCTL_CHILDREN(rack_counters),
1739 OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1740 &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1741 rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1742
1743 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 SYSCTL_CHILDREN(rack_counters),
1745 OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1746 &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1747 rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1748 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1749 SYSCTL_CHILDREN(rack_counters),
1750 OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1751 &rack_tlp_tot,
1752 "Total number of tail loss probe expirations");
1753 rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1754 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1755 SYSCTL_CHILDREN(rack_counters),
1756 OID_AUTO, "tlp_new", CTLFLAG_RD,
1757 &rack_tlp_newdata,
1758 "Total number of tail loss probe sending new data");
1759 rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1760 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1761 SYSCTL_CHILDREN(rack_counters),
1762 OID_AUTO, "tlp_retran", CTLFLAG_RD,
1763 &rack_tlp_retran,
1764 "Total number of tail loss probe sending retransmitted data");
1765 rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1766 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1767 SYSCTL_CHILDREN(rack_counters),
1768 OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1769 &rack_tlp_retran_bytes,
1770 "Total bytes of tail loss probe sending retransmitted data");
1771 rack_to_tot = counter_u64_alloc(M_WAITOK);
1772 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1773 SYSCTL_CHILDREN(rack_counters),
1774 OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1775 &rack_to_tot,
1776 "Total number of times the rack to expired");
1777 rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1778 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1779 SYSCTL_CHILDREN(rack_counters),
1780 OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1781 &rack_saw_enobuf,
1782 "Total number of times a sends returned enobuf for non-hdwr paced connections");
1783 rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1784 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1785 SYSCTL_CHILDREN(rack_counters),
1786 OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1787 &rack_saw_enobuf_hw,
1788 "Total number of times a send returned enobuf for hdwr paced connections");
1789 rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1790 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1791 SYSCTL_CHILDREN(rack_counters),
1792 OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1793 &rack_saw_enetunreach,
1794 "Total number of times a send received a enetunreachable");
1795 rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1796 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1797 SYSCTL_CHILDREN(rack_counters),
1798 OID_AUTO, "alloc_hot", CTLFLAG_RD,
1799 &rack_hot_alloc,
1800 "Total allocations from the top of our list");
1801 rack_to_alloc = counter_u64_alloc(M_WAITOK);
1802 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1803 SYSCTL_CHILDREN(rack_counters),
1804 OID_AUTO, "allocs", CTLFLAG_RD,
1805 &rack_to_alloc,
1806 "Total allocations of tracking structures");
1807 rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1808 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1809 SYSCTL_CHILDREN(rack_counters),
1810 OID_AUTO, "allochard", CTLFLAG_RD,
1811 &rack_to_alloc_hard,
1812 "Total allocations done with sleeping the hard way");
1813 rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1814 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1815 SYSCTL_CHILDREN(rack_counters),
1816 OID_AUTO, "allocemerg", CTLFLAG_RD,
1817 &rack_to_alloc_emerg,
1818 "Total allocations done from emergency cache");
1819 rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1820 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1821 SYSCTL_CHILDREN(rack_counters),
1822 OID_AUTO, "alloc_limited", CTLFLAG_RD,
1823 &rack_to_alloc_limited,
1824 "Total allocations dropped due to limit");
1825 rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1826 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1827 SYSCTL_CHILDREN(rack_counters),
1828 OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1829 &rack_alloc_limited_conns,
1830 "Connections with allocations dropped due to limit");
1831 rack_split_limited = counter_u64_alloc(M_WAITOK);
1832 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1833 SYSCTL_CHILDREN(rack_counters),
1834 OID_AUTO, "split_limited", CTLFLAG_RD,
1835 &rack_split_limited,
1836 "Split allocations dropped due to limit");
1837 rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1838 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1839 SYSCTL_CHILDREN(rack_counters),
1840 OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1841 &rack_rxt_clamps_cwnd,
1842 "Number of times that excessive rxt clamped the cwnd down");
1843 rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1844 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1845 SYSCTL_CHILDREN(rack_counters),
1846 OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1847 &rack_rxt_clamps_cwnd_uniq,
1848 "Number of connections that have had excessive rxt clamped the cwnd down");
1849 rack_persists_sends = counter_u64_alloc(M_WAITOK);
1850 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1851 SYSCTL_CHILDREN(rack_counters),
1852 OID_AUTO, "persist_sends", CTLFLAG_RD,
1853 &rack_persists_sends,
1854 "Number of times we sent a persist probe");
1855 rack_persists_acks = counter_u64_alloc(M_WAITOK);
1856 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1857 SYSCTL_CHILDREN(rack_counters),
1858 OID_AUTO, "persist_acks", CTLFLAG_RD,
1859 &rack_persists_acks,
1860 "Number of times a persist probe was acked");
1861 rack_persists_loss = counter_u64_alloc(M_WAITOK);
1862 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1863 SYSCTL_CHILDREN(rack_counters),
1864 OID_AUTO, "persist_loss", CTLFLAG_RD,
1865 &rack_persists_loss,
1866 "Number of times we detected a lost persist probe (no ack)");
1867 rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1868 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1869 SYSCTL_CHILDREN(rack_counters),
1870 OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1871 &rack_persists_lost_ends,
1872 "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1873 #ifdef INVARIANTS
1874 rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1875 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1876 SYSCTL_CHILDREN(rack_counters),
1877 OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1878 &rack_adjust_map_bw,
1879 "Number of times we hit the case where the sb went up and down on a sendmap entry");
1880 #endif
1881 rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1882 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1883 SYSCTL_CHILDREN(rack_counters),
1884 OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1885 &rack_multi_single_eq,
1886 "Number of compressed acks total represented");
1887 rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1888 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1889 SYSCTL_CHILDREN(rack_counters),
1890 OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1891 &rack_proc_non_comp_ack,
1892 "Number of non compresseds acks that we processed");
1893
1894
1895 rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1896 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1897 SYSCTL_CHILDREN(rack_counters),
1898 OID_AUTO, "sack_long", CTLFLAG_RD,
1899 &rack_sack_proc_all,
1900 "Total times we had to walk whole list for sack processing");
1901 rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1902 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1903 SYSCTL_CHILDREN(rack_counters),
1904 OID_AUTO, "sack_restart", CTLFLAG_RD,
1905 &rack_sack_proc_restart,
1906 "Total times we had to walk whole list due to a restart");
1907 rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1908 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1909 SYSCTL_CHILDREN(rack_counters),
1910 OID_AUTO, "sack_short", CTLFLAG_RD,
1911 &rack_sack_proc_short,
1912 "Total times we took shortcut for sack processing");
1913 rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1914 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1915 SYSCTL_CHILDREN(rack_attack),
1916 OID_AUTO, "skipacked", CTLFLAG_RD,
1917 &rack_sack_skipped_acked,
1918 "Total number of times we skipped previously sacked");
1919 rack_sack_splits = counter_u64_alloc(M_WAITOK);
1920 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1921 SYSCTL_CHILDREN(rack_attack),
1922 OID_AUTO, "ofsplit", CTLFLAG_RD,
1923 &rack_sack_splits,
1924 "Total number of times we did the old fashion tree split");
1925 rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1926 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1927 SYSCTL_CHILDREN(rack_counters),
1928 OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1929 &rack_input_idle_reduces,
1930 "Total number of idle reductions on input");
1931 rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1932 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1933 SYSCTL_CHILDREN(rack_counters),
1934 OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1935 &rack_collapsed_win_seen,
1936 "Total number of collapsed window events seen (where our window shrinks)");
1937
1938 rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1939 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1940 SYSCTL_CHILDREN(rack_counters),
1941 OID_AUTO, "collapsed_win", CTLFLAG_RD,
1942 &rack_collapsed_win,
1943 "Total number of collapsed window events where we mark packets");
1944 rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1945 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1946 SYSCTL_CHILDREN(rack_counters),
1947 OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1948 &rack_collapsed_win_rxt,
1949 "Total number of packets that were retransmitted");
1950 rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1951 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1952 SYSCTL_CHILDREN(rack_counters),
1953 OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1954 &rack_collapsed_win_rxt_bytes,
1955 "Total number of bytes that were retransmitted");
1956 rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1957 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1958 SYSCTL_CHILDREN(rack_counters),
1959 OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1960 &rack_try_scwnd,
1961 "Total number of scwnd attempts");
1962 COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1963 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1964 OID_AUTO, "outsize", CTLFLAG_RD,
1965 rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1966 COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1967 SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1968 OID_AUTO, "opts", CTLFLAG_RD,
1969 rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1970 SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1971 SYSCTL_CHILDREN(rack_sysctl_root),
1972 OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1973 &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1974 }
1975
1976 static uint32_t
rc_init_window(struct tcp_rack * rack)1977 rc_init_window(struct tcp_rack *rack)
1978 {
1979 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1980
1981 }
1982
1983 static uint64_t
rack_get_fixed_pacing_bw(struct tcp_rack * rack)1984 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1985 {
1986 if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1987 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1988 else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1989 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1990 else
1991 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1992 }
1993
1994 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)1995 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1996 uint64_t data, uint8_t mod, uint16_t aux,
1997 struct tcp_sendfile_track *cur, int line)
1998 {
1999 #ifdef TCP_REQUEST_TRK
2000 int do_log = 0;
2001
2002 /*
2003 * The rate cap one is noisy and only should come out when normal BB logging
2004 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
2005 * once per chunk and make up the BBpoint that can be turned on by the client.
2006 */
2007 if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2008 /*
2009 * The very noisy two need to only come out when
2010 * we have verbose logging on.
2011 */
2012 if (rack_verbose_logging != 0)
2013 do_log = tcp_bblogging_on(rack->rc_tp);
2014 else
2015 do_log = 0;
2016 } else if (mod != HYBRID_LOG_BW_MEASURE) {
2017 /*
2018 * All other less noisy logs here except the measure which
2019 * also needs to come out on the point and the log.
2020 */
2021 do_log = tcp_bblogging_on(rack->rc_tp);
2022 } else {
2023 do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2024 }
2025
2026 if (do_log) {
2027 union tcp_log_stackspecific log;
2028 struct timeval tv;
2029 uint64_t lt_bw;
2030
2031 /* Convert our ms to a microsecond */
2032 memset(&log, 0, sizeof(log));
2033
2034 log.u_bbr.cwnd_gain = line;
2035 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2036 log.u_bbr.rttProp = tim;
2037 log.u_bbr.bw_inuse = cbw;
2038 log.u_bbr.delRate = rack_get_gp_est(rack);
2039 lt_bw = rack_get_lt_bw(rack);
2040 log.u_bbr.flex1 = seq;
2041 log.u_bbr.pacing_gain = aux;
2042 /* lt_bw = < flex3 | flex2 > */
2043 log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2044 log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2045 /* Record the last obtained us rtt in inflight */
2046 if (cur == NULL) {
2047 /* Make sure we are looking at the right log if an overide comes in */
2048 cur = rack->r_ctl.rc_last_sft;
2049 }
2050 if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2051 log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2052 else {
2053 /* Use the last known rtt i.e. the rack-rtt */
2054 log.u_bbr.inflight = rack->rc_rack_rtt;
2055 }
2056 if (cur != NULL) {
2057 uint64_t off;
2058
2059 log.u_bbr.cur_del_rate = cur->deadline;
2060 if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2061 /* start = < lost | pkt_epoch > */
2062 log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2063 log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2064 log.u_bbr.flex6 = cur->start_seq;
2065 log.u_bbr.pkts_out = cur->end_seq;
2066 } else {
2067 /* start = < lost | pkt_epoch > */
2068 log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2069 log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2070 /* end = < pkts_out | flex6 > */
2071 log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2072 log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2073 }
2074 /* first_send = <lt_epoch | epoch> */
2075 log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2076 log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2077 /* localtime = <delivered | applimited>*/
2078 log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2079 log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2080 #ifdef TCP_REQUEST_TRK
2081 off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2082 log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2083 #endif
2084 log.u_bbr.inhpts = 1;
2085 log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2086 log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2087 log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2088 } else {
2089 log.u_bbr.flex7 = 0xffff;
2090 log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2091 }
2092 /*
2093 * Compose bbr_state to be a bit wise 0000ADHF
2094 * where A is the always_pace flag
2095 * where D is the dgp_on flag
2096 * where H is the hybrid_mode on flag
2097 * where F is the use_fixed_rate flag.
2098 */
2099 log.u_bbr.bbr_state = rack->rc_always_pace;
2100 log.u_bbr.bbr_state <<= 1;
2101 log.u_bbr.bbr_state |= rack->dgp_on;
2102 log.u_bbr.bbr_state <<= 1;
2103 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2104 log.u_bbr.bbr_state <<= 1;
2105 log.u_bbr.bbr_state |= rack->use_fixed_rate;
2106 log.u_bbr.flex8 = mod;
2107 tcp_log_event(rack->rc_tp, NULL,
2108 &rack->rc_inp->inp_socket->so_rcv,
2109 &rack->rc_inp->inp_socket->so_snd,
2110 TCP_HYBRID_PACING_LOG, 0,
2111 0, &log, false, NULL, __func__, __LINE__, &tv);
2112
2113 }
2114 #endif
2115 }
2116
2117 #ifdef TCP_REQUEST_TRK
2118 static void
rack_log_hybrid_sends(struct tcp_rack * rack,struct tcp_sendfile_track * cur,int line)2119 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2120 {
2121 if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2122 union tcp_log_stackspecific log;
2123 struct timeval tv;
2124 uint64_t off;
2125
2126 /* Convert our ms to a microsecond */
2127 memset(&log, 0, sizeof(log));
2128
2129 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2130 log.u_bbr.delRate = cur->sent_at_fs;
2131
2132 if ((cur->flags & TCP_TRK_TRACK_FLG_LSND) == 0) {
2133 /*
2134 * We did not get a new Rules Applied to set so
2135 * no overlapping send occured, this means the
2136 * current byte counts are correct.
2137 */
2138 log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2139 log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2140 } else {
2141 /*
2142 * Overlapping send case, we switched to a new
2143 * send and did a rules applied.
2144 */
2145 log.u_bbr.cur_del_rate = cur->sent_at_ls;
2146 log.u_bbr.rttProp = cur->rxt_at_ls;
2147 }
2148 log.u_bbr.bw_inuse = cur->rxt_at_fs;
2149 log.u_bbr.cwnd_gain = line;
2150 off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2151 log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2152 /* start = < flex1 | flex2 > */
2153 log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2154 log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2155 /* end = < flex3 | flex4 > */
2156 log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2157 log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2158
2159 /* localtime = <delivered | applimited>*/
2160 log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2161 log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2162 /* client timestamp = <lt_epoch | epoch>*/
2163 log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2164 log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2165 /* now set all the flags in */
2166 log.u_bbr.pkts_out = cur->hybrid_flags;
2167 log.u_bbr.lost = cur->playout_ms;
2168 log.u_bbr.flex6 = cur->flags;
2169 /*
2170 * Last send time = <flex5 | pkt_epoch> note we do not distinguish cases
2171 * where a false retransmit occurred so first_send <-> lastsend may
2172 * include longer time then it actually took if we have a false rxt.
2173 */
2174 log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2175 log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2176 /*
2177 * Compose bbr_state to be a bit wise 0000ADHF
2178 * where A is the always_pace flag
2179 * where D is the dgp_on flag
2180 * where H is the hybrid_mode on flag
2181 * where F is the use_fixed_rate flag.
2182 */
2183 log.u_bbr.bbr_state = rack->rc_always_pace;
2184 log.u_bbr.bbr_state <<= 1;
2185 log.u_bbr.bbr_state |= rack->dgp_on;
2186 log.u_bbr.bbr_state <<= 1;
2187 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2188 log.u_bbr.bbr_state <<= 1;
2189 log.u_bbr.bbr_state |= rack->use_fixed_rate;
2190
2191 log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2192 tcp_log_event(rack->rc_tp, NULL,
2193 &rack->rc_inp->inp_socket->so_rcv,
2194 &rack->rc_inp->inp_socket->so_snd,
2195 TCP_HYBRID_PACING_LOG, 0,
2196 0, &log, false, NULL, __func__, __LINE__, &tv);
2197 }
2198 }
2199 #endif
2200
2201 static inline uint64_t
rack_compensate_for_linerate(struct tcp_rack * rack,uint64_t bw)2202 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2203 {
2204 uint64_t ret_bw, ether;
2205 uint64_t u_segsiz;
2206
2207 ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2208 if (rack->r_is_v6){
2209 #ifdef INET6
2210 ether += sizeof(struct ip6_hdr);
2211 #endif
2212 ether += 14; /* eheader size 6+6+2 */
2213 } else {
2214 #ifdef INET
2215 ether += sizeof(struct ip);
2216 #endif
2217 ether += 14; /* eheader size 6+6+2 */
2218 }
2219 u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2220 ret_bw = bw;
2221 ret_bw *= ether;
2222 ret_bw /= u_segsiz;
2223 return (ret_bw);
2224 }
2225
2226 static void
rack_rate_cap_bw(struct tcp_rack * rack,uint64_t * bw,int * capped)2227 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2228 {
2229 #ifdef TCP_REQUEST_TRK
2230 struct timeval tv;
2231 uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2232 #endif
2233
2234 if (rack->r_ctl.bw_rate_cap == 0)
2235 return;
2236 #ifdef TCP_REQUEST_TRK
2237 if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2238 (rack->r_ctl.rc_last_sft != NULL)) {
2239 /*
2240 * We have a dynamic cap. The original target
2241 * is in bw_rate_cap, but we need to look at
2242 * how long it is until we hit the deadline.
2243 */
2244 struct tcp_sendfile_track *ent;
2245
2246 ent = rack->r_ctl.rc_last_sft;
2247 microuptime(&tv);
2248 timenow = tcp_tv_to_lusec(&tv);
2249 if (timenow >= ent->deadline) {
2250 /* No time left we do DGP only */
2251 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2252 0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2253 rack->r_ctl.bw_rate_cap = 0;
2254 return;
2255 }
2256 /* We have the time */
2257 timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2258 if (timeleft < HPTS_MSEC_IN_SEC) {
2259 /* If there is less than a ms left just use DGPs rate */
2260 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2261 0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2262 rack->r_ctl.bw_rate_cap = 0;
2263 return;
2264 }
2265 /*
2266 * Now lets find the amount of data left to send.
2267 *
2268 * Now ideally we want to use the end_seq to figure out how much more
2269 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2270 */
2271 if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2272 if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2273 lenleft = ent->end_seq - rack->rc_tp->snd_una;
2274 else {
2275 /* TSNH, we should catch it at the send */
2276 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2277 0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2278 rack->r_ctl.bw_rate_cap = 0;
2279 return;
2280 }
2281 } else {
2282 /*
2283 * The hard way, figure out how much is gone and then
2284 * take that away from the total the client asked for
2285 * (thats off by tls overhead if this is tls).
2286 */
2287 if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2288 lengone = rack->rc_tp->snd_una - ent->start_seq;
2289 else
2290 lengone = 0;
2291 if (lengone < (ent->end - ent->start))
2292 lenleft = (ent->end - ent->start) - lengone;
2293 else {
2294 /* TSNH, we should catch it at the send */
2295 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2296 0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2297 rack->r_ctl.bw_rate_cap = 0;
2298 return;
2299 }
2300 }
2301 if (lenleft == 0) {
2302 /* We have it all sent */
2303 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2304 0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2305 if (rack->r_ctl.bw_rate_cap)
2306 goto normal_ratecap;
2307 else
2308 return;
2309 }
2310 calcbw = lenleft * HPTS_USEC_IN_SEC;
2311 calcbw /= timeleft;
2312 /* Now we must compensate for IP/TCP overhead */
2313 calcbw = rack_compensate_for_linerate(rack, calcbw);
2314 /* Update the bit rate cap */
2315 rack->r_ctl.bw_rate_cap = calcbw;
2316 if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2317 (rack_hybrid_allow_set_maxseg == 1) &&
2318 ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2319 /* Lets set in a smaller mss possibly here to match our rate-cap */
2320 uint32_t orig_max;
2321
2322 orig_max = rack->r_ctl.rc_pace_max_segs;
2323 rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2324 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2325 rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2326 }
2327 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2328 calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2329 if ((calcbw > 0) && (*bw > calcbw)) {
2330 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2331 *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2332 *capped = 1;
2333 *bw = calcbw;
2334 }
2335 return;
2336 }
2337 normal_ratecap:
2338 #endif
2339 if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2340 #ifdef TCP_REQUEST_TRK
2341 if (rack->rc_hybrid_mode &&
2342 rack->rc_catch_up &&
2343 (rack->r_ctl.rc_last_sft != NULL) &&
2344 (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2345 (rack_hybrid_allow_set_maxseg == 1) &&
2346 ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2347 /* Lets set in a smaller mss possibly here to match our rate-cap */
2348 uint32_t orig_max;
2349
2350 orig_max = rack->r_ctl.rc_pace_max_segs;
2351 rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2352 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2353 rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2354 }
2355 #endif
2356 *capped = 1;
2357 *bw = rack->r_ctl.bw_rate_cap;
2358 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2359 *bw, 0, 0,
2360 HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2361 }
2362 }
2363
2364 static uint64_t
rack_get_gp_est(struct tcp_rack * rack)2365 rack_get_gp_est(struct tcp_rack *rack)
2366 {
2367 uint64_t bw, lt_bw, ret_bw;
2368
2369 if (rack->rc_gp_filled == 0) {
2370 /*
2371 * We have yet no b/w measurement,
2372 * if we have a user set initial bw
2373 * return it. If we don't have that and
2374 * we have an srtt, use the tcp IW (10) to
2375 * calculate a fictional b/w over the SRTT
2376 * which is more or less a guess. Note
2377 * we don't use our IW from rack on purpose
2378 * so if we have like IW=30, we are not
2379 * calculating a "huge" b/w.
2380 */
2381 uint64_t srtt;
2382
2383 if (rack->dis_lt_bw == 1)
2384 lt_bw = 0;
2385 else
2386 lt_bw = rack_get_lt_bw(rack);
2387 if (lt_bw) {
2388 /*
2389 * No goodput bw but a long-term b/w does exist
2390 * lets use that.
2391 */
2392 ret_bw = lt_bw;
2393 goto compensate;
2394 }
2395 if (rack->r_ctl.init_rate)
2396 return (rack->r_ctl.init_rate);
2397
2398 /* Ok lets come up with the IW guess, if we have a srtt */
2399 if (rack->rc_tp->t_srtt == 0) {
2400 /*
2401 * Go with old pacing method
2402 * i.e. burst mitigation only.
2403 */
2404 return (0);
2405 }
2406 /* Ok lets get the initial TCP win (not racks) */
2407 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2408 srtt = (uint64_t)rack->rc_tp->t_srtt;
2409 bw *= (uint64_t)USECS_IN_SECOND;
2410 bw /= srtt;
2411 ret_bw = bw;
2412 goto compensate;
2413
2414 }
2415 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2416 /* Averaging is done, we can return the value */
2417 bw = rack->r_ctl.gp_bw;
2418 } else {
2419 /* Still doing initial average must calculate */
2420 bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2421 }
2422 if (rack->dis_lt_bw) {
2423 /* We are not using lt-bw */
2424 ret_bw = bw;
2425 goto compensate;
2426 }
2427 lt_bw = rack_get_lt_bw(rack);
2428 if (lt_bw == 0) {
2429 /* If we don't have one then equate it to the gp_bw */
2430 lt_bw = rack->r_ctl.gp_bw;
2431 }
2432 if (rack->use_lesser_lt_bw) {
2433 if (lt_bw < bw)
2434 ret_bw = lt_bw;
2435 else
2436 ret_bw = bw;
2437 } else {
2438 if (lt_bw > bw)
2439 ret_bw = lt_bw;
2440 else
2441 ret_bw = bw;
2442 }
2443 /*
2444 * Now lets compensate based on the TCP/IP overhead. Our
2445 * Goodput estimate does not include this so we must pace out
2446 * a bit faster since our pacing calculations do. The pacing
2447 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2448 * we are using to do this, so we do that here in the opposite
2449 * direction as well. This means that if we are tunneled and the
2450 * segsiz is say 1200 bytes we will get quite a boost, but its
2451 * compensated for in the pacing time the opposite way.
2452 */
2453 compensate:
2454 ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2455 return(ret_bw);
2456 }
2457
2458
2459 static uint64_t
rack_get_bw(struct tcp_rack * rack)2460 rack_get_bw(struct tcp_rack *rack)
2461 {
2462 uint64_t bw;
2463
2464 if (rack->use_fixed_rate) {
2465 /* Return the fixed pacing rate */
2466 return (rack_get_fixed_pacing_bw(rack));
2467 }
2468 bw = rack_get_gp_est(rack);
2469 return (bw);
2470 }
2471
2472 static uint16_t
rack_get_output_gain(struct tcp_rack * rack,struct rack_sendmap * rsm)2473 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2474 {
2475 if (rack->use_fixed_rate) {
2476 return (100);
2477 } else if (rack->in_probe_rtt && (rsm == NULL))
2478 return (rack->r_ctl.rack_per_of_gp_probertt);
2479 else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2480 rack->r_ctl.rack_per_of_gp_rec)) {
2481 if (rsm) {
2482 /* a retransmission always use the recovery rate */
2483 return (rack->r_ctl.rack_per_of_gp_rec);
2484 } else if (rack->rack_rec_nonrxt_use_cr) {
2485 /* Directed to use the configured rate */
2486 goto configured_rate;
2487 } else if (rack->rack_no_prr &&
2488 (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2489 /* No PRR, lets just use the b/w estimate only */
2490 return (100);
2491 } else {
2492 /*
2493 * Here we may have a non-retransmit but we
2494 * have no overrides, so just use the recovery
2495 * rate (prr is in effect).
2496 */
2497 return (rack->r_ctl.rack_per_of_gp_rec);
2498 }
2499 }
2500 configured_rate:
2501 /* For the configured rate we look at our cwnd vs the ssthresh */
2502 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2503 return (rack->r_ctl.rack_per_of_gp_ss);
2504 else
2505 return (rack->r_ctl.rack_per_of_gp_ca);
2506 }
2507
2508 static void
rack_log_dsack_event(struct tcp_rack * rack,uint8_t mod,uint32_t flex4,uint32_t flex5,uint32_t flex6)2509 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2510 {
2511 /*
2512 * Types of logs (mod value)
2513 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2514 * 2 = a dsack round begins, persist is reset to 16.
2515 * 3 = a dsack round ends
2516 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2517 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2518 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2519 */
2520 if (tcp_bblogging_on(rack->rc_tp)) {
2521 union tcp_log_stackspecific log;
2522 struct timeval tv;
2523
2524 memset(&log, 0, sizeof(log));
2525 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2526 log.u_bbr.flex1 <<= 1;
2527 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2528 log.u_bbr.flex1 <<= 1;
2529 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2530 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2531 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2532 log.u_bbr.flex4 = flex4;
2533 log.u_bbr.flex5 = flex5;
2534 log.u_bbr.flex6 = flex6;
2535 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2536 log.u_bbr.flex8 = mod;
2537 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2538 log.u_bbr.epoch = rack->r_ctl.current_round;
2539 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2540 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2541 &rack->rc_inp->inp_socket->so_rcv,
2542 &rack->rc_inp->inp_socket->so_snd,
2543 RACK_DSACK_HANDLING, 0,
2544 0, &log, false, &tv);
2545 }
2546 }
2547
2548 static void
rack_log_hdwr_pacing(struct tcp_rack * rack,uint64_t rate,uint64_t hw_rate,int line,int error,uint16_t mod)2549 rack_log_hdwr_pacing(struct tcp_rack *rack,
2550 uint64_t rate, uint64_t hw_rate, int line,
2551 int error, uint16_t mod)
2552 {
2553 if (tcp_bblogging_on(rack->rc_tp)) {
2554 union tcp_log_stackspecific log;
2555 struct timeval tv;
2556 const struct ifnet *ifp;
2557 uint64_t ifp64;
2558
2559 memset(&log, 0, sizeof(log));
2560 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2561 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2562 if (rack->r_ctl.crte) {
2563 ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2564 } else if (rack->rc_inp->inp_route.ro_nh &&
2565 rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2566 ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2567 } else
2568 ifp = NULL;
2569 if (ifp) {
2570 ifp64 = (uintptr_t)ifp;
2571 log.u_bbr.flex3 = ((ifp64 >> 32) & 0x00000000ffffffff);
2572 log.u_bbr.flex4 = (ifp64 & 0x00000000ffffffff);
2573 }
2574 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2575 log.u_bbr.bw_inuse = rate;
2576 log.u_bbr.flex5 = line;
2577 log.u_bbr.flex6 = error;
2578 log.u_bbr.flex7 = mod;
2579 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2580 log.u_bbr.flex8 = rack->use_fixed_rate;
2581 log.u_bbr.flex8 <<= 1;
2582 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2583 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2584 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2585 if (rack->r_ctl.crte)
2586 log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2587 else
2588 log.u_bbr.cur_del_rate = 0;
2589 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2590 log.u_bbr.epoch = rack->r_ctl.current_round;
2591 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2592 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2593 &rack->rc_inp->inp_socket->so_rcv,
2594 &rack->rc_inp->inp_socket->so_snd,
2595 BBR_LOG_HDWR_PACE, 0,
2596 0, &log, false, &tv);
2597 }
2598 }
2599
2600 static uint64_t
rack_get_output_bw(struct tcp_rack * rack,uint64_t bw,struct rack_sendmap * rsm,int * capped)2601 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2602 {
2603 /*
2604 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2605 */
2606 uint64_t bw_est, high_rate;
2607 uint64_t gain;
2608
2609 gain = (uint64_t)rack_get_output_gain(rack, rsm);
2610 bw_est = bw * gain;
2611 bw_est /= (uint64_t)100;
2612 /* Never fall below the minimum (def 64kbps) */
2613 if (bw_est < RACK_MIN_BW)
2614 bw_est = RACK_MIN_BW;
2615 if (rack->r_rack_hw_rate_caps) {
2616 /* Rate caps are in place */
2617 if (rack->r_ctl.crte != NULL) {
2618 /* We have a hdwr rate already */
2619 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2620 if (bw_est >= high_rate) {
2621 /* We are capping bw at the highest rate table entry */
2622 if (rack_hw_rate_cap_per &&
2623 (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2624 rack->r_rack_hw_rate_caps = 0;
2625 goto done;
2626 }
2627 rack_log_hdwr_pacing(rack,
2628 bw_est, high_rate, __LINE__,
2629 0, 3);
2630 bw_est = high_rate;
2631 if (capped)
2632 *capped = 1;
2633 }
2634 } else if ((rack->rack_hdrw_pacing == 0) &&
2635 (rack->rack_hdw_pace_ena) &&
2636 (rack->rack_attempt_hdwr_pace == 0) &&
2637 (rack->rc_inp->inp_route.ro_nh != NULL) &&
2638 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2639 /*
2640 * Special case, we have not yet attempted hardware
2641 * pacing, and yet we may, when we do, find out if we are
2642 * above the highest rate. We need to know the maxbw for the interface
2643 * in question (if it supports ratelimiting). We get back
2644 * a 0, if the interface is not found in the RL lists.
2645 */
2646 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2647 if (high_rate) {
2648 /* Yep, we have a rate is it above this rate? */
2649 if (bw_est > high_rate) {
2650 bw_est = high_rate;
2651 if (capped)
2652 *capped = 1;
2653 }
2654 }
2655 }
2656 }
2657 done:
2658 return (bw_est);
2659 }
2660
2661 static void
rack_log_retran_reason(struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t tsused,uint32_t thresh,int mod)2662 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2663 {
2664 if (tcp_bblogging_on(rack->rc_tp)) {
2665 union tcp_log_stackspecific log;
2666 struct timeval tv;
2667
2668 if ((mod != 1) && (rack_verbose_logging == 0)) {
2669 /*
2670 * We get 3 values currently for mod
2671 * 1 - We are retransmitting and this tells the reason.
2672 * 2 - We are clearing a dup-ack count.
2673 * 3 - We are incrementing a dup-ack count.
2674 *
2675 * The clear/increment are only logged
2676 * if you have BBverbose on.
2677 */
2678 return;
2679 }
2680 memset(&log, 0, sizeof(log));
2681 log.u_bbr.flex1 = tsused;
2682 log.u_bbr.flex2 = thresh;
2683 log.u_bbr.flex3 = rsm->r_flags;
2684 log.u_bbr.flex4 = rsm->r_dupack;
2685 log.u_bbr.flex5 = rsm->r_start;
2686 log.u_bbr.flex6 = rsm->r_end;
2687 log.u_bbr.flex8 = mod;
2688 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2689 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2690 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2691 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2692 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2693 log.u_bbr.pacing_gain = rack->r_must_retran;
2694 log.u_bbr.epoch = rack->r_ctl.current_round;
2695 log.u_bbr.lt_epoch = rack->r_ctl.rc_considered_lost;
2696 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2697 &rack->rc_inp->inp_socket->so_rcv,
2698 &rack->rc_inp->inp_socket->so_snd,
2699 BBR_LOG_SETTINGS_CHG, 0,
2700 0, &log, false, &tv);
2701 }
2702 }
2703
2704 static void
rack_log_to_start(struct tcp_rack * rack,uint32_t cts,uint32_t to,int32_t slot,uint8_t which)2705 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2706 {
2707 if (tcp_bblogging_on(rack->rc_tp)) {
2708 union tcp_log_stackspecific log;
2709 struct timeval tv;
2710
2711 memset(&log, 0, sizeof(log));
2712 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2713 log.u_bbr.flex2 = to;
2714 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2715 log.u_bbr.flex4 = slot;
2716 log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2717 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2718 log.u_bbr.flex7 = rack->rc_in_persist;
2719 log.u_bbr.flex8 = which;
2720 if (rack->rack_no_prr)
2721 log.u_bbr.pkts_out = 0;
2722 else
2723 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2724 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2725 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2726 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2727 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2728 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2729 log.u_bbr.pacing_gain = rack->r_must_retran;
2730 log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2731 log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2732 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2733 log.u_bbr.lost = rack_rto_min;
2734 log.u_bbr.epoch = rack->r_ctl.roundends;
2735 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2736 log.u_bbr.bw_inuse <<= 32;
2737 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2738 log.u_bbr.applimited = rack->rc_tp->t_flags2;
2739 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2740 &rack->rc_inp->inp_socket->so_rcv,
2741 &rack->rc_inp->inp_socket->so_snd,
2742 BBR_LOG_TIMERSTAR, 0,
2743 0, &log, false, &tv);
2744 }
2745 }
2746
2747 static void
rack_log_to_event(struct tcp_rack * rack,int32_t to_num,struct rack_sendmap * rsm)2748 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2749 {
2750 if (tcp_bblogging_on(rack->rc_tp)) {
2751 union tcp_log_stackspecific log;
2752 struct timeval tv;
2753
2754 memset(&log, 0, sizeof(log));
2755 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2756 log.u_bbr.flex8 = to_num;
2757 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2758 log.u_bbr.flex2 = rack->rc_rack_rtt;
2759 if (rsm == NULL)
2760 log.u_bbr.flex3 = 0;
2761 else
2762 log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2763 if (rack->rack_no_prr)
2764 log.u_bbr.flex5 = 0;
2765 else
2766 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2767 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2768 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2769 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2770 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2771 log.u_bbr.pacing_gain = rack->r_must_retran;
2772 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2773 log.u_bbr.bw_inuse <<= 32;
2774 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2775 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2776 &rack->rc_inp->inp_socket->so_rcv,
2777 &rack->rc_inp->inp_socket->so_snd,
2778 BBR_LOG_RTO, 0,
2779 0, &log, false, &tv);
2780 }
2781 }
2782
2783 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)2784 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2785 struct rack_sendmap *prev,
2786 struct rack_sendmap *rsm,
2787 struct rack_sendmap *next,
2788 int flag, uint32_t th_ack, int line)
2789 {
2790 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2791 union tcp_log_stackspecific log;
2792 struct timeval tv;
2793
2794 memset(&log, 0, sizeof(log));
2795 log.u_bbr.flex8 = flag;
2796 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2797 log.u_bbr.cur_del_rate = (uintptr_t)prev;
2798 log.u_bbr.delRate = (uintptr_t)rsm;
2799 log.u_bbr.rttProp = (uintptr_t)next;
2800 log.u_bbr.flex7 = 0;
2801 if (prev) {
2802 log.u_bbr.flex1 = prev->r_start;
2803 log.u_bbr.flex2 = prev->r_end;
2804 log.u_bbr.flex7 |= 0x4;
2805 }
2806 if (rsm) {
2807 log.u_bbr.flex3 = rsm->r_start;
2808 log.u_bbr.flex4 = rsm->r_end;
2809 log.u_bbr.flex7 |= 0x2;
2810 }
2811 if (next) {
2812 log.u_bbr.flex5 = next->r_start;
2813 log.u_bbr.flex6 = next->r_end;
2814 log.u_bbr.flex7 |= 0x1;
2815 }
2816 log.u_bbr.applimited = line;
2817 log.u_bbr.pkts_out = th_ack;
2818 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2819 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2820 if (rack->rack_no_prr)
2821 log.u_bbr.lost = 0;
2822 else
2823 log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2824 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2825 log.u_bbr.bw_inuse <<= 32;
2826 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2827 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2828 &rack->rc_inp->inp_socket->so_rcv,
2829 &rack->rc_inp->inp_socket->so_snd,
2830 TCP_LOG_MAPCHG, 0,
2831 0, &log, false, &tv);
2832 }
2833 }
2834
2835 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)2836 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2837 struct rack_sendmap *rsm, int conf)
2838 {
2839 if (tcp_bblogging_on(tp)) {
2840 union tcp_log_stackspecific log;
2841 struct timeval tv;
2842 memset(&log, 0, sizeof(log));
2843 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2844 log.u_bbr.flex1 = t;
2845 log.u_bbr.flex2 = len;
2846 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2847 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2848 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2849 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2850 log.u_bbr.flex7 = conf;
2851 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2852 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2853 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2854 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2855 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2856 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2857 if (rsm) {
2858 log.u_bbr.pkt_epoch = rsm->r_start;
2859 log.u_bbr.lost = rsm->r_end;
2860 log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2861 /* We loose any upper of the 24 bits */
2862 log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2863 } else {
2864 /* Its a SYN */
2865 log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2866 log.u_bbr.lost = 0;
2867 log.u_bbr.cwnd_gain = 0;
2868 log.u_bbr.pacing_gain = 0;
2869 }
2870 /* Write out general bits of interest rrs here */
2871 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2872 log.u_bbr.use_lt_bw <<= 1;
2873 log.u_bbr.use_lt_bw |= rack->forced_ack;
2874 log.u_bbr.use_lt_bw <<= 1;
2875 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2876 log.u_bbr.use_lt_bw <<= 1;
2877 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2878 log.u_bbr.use_lt_bw <<= 1;
2879 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2880 log.u_bbr.use_lt_bw <<= 1;
2881 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2882 log.u_bbr.use_lt_bw <<= 1;
2883 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2884 log.u_bbr.use_lt_bw <<= 1;
2885 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2886 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2887 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2888 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2889 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2890 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2891 log.u_bbr.bw_inuse = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
2892 log.u_bbr.bw_inuse <<= 32;
2893 if (rsm)
2894 log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2895 TCP_LOG_EVENTP(tp, NULL,
2896 &rack->rc_inp->inp_socket->so_rcv,
2897 &rack->rc_inp->inp_socket->so_snd,
2898 BBR_LOG_BBRRTT, 0,
2899 0, &log, false, &tv);
2900
2901
2902 }
2903 }
2904
2905 static void
rack_log_rtt_sample(struct tcp_rack * rack,uint32_t rtt)2906 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2907 {
2908 /*
2909 * Log the rtt sample we are
2910 * applying to the srtt algorithm in
2911 * useconds.
2912 */
2913 if (tcp_bblogging_on(rack->rc_tp)) {
2914 union tcp_log_stackspecific log;
2915 struct timeval tv;
2916
2917 /* Convert our ms to a microsecond */
2918 memset(&log, 0, sizeof(log));
2919 log.u_bbr.flex1 = rtt;
2920 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2921 log.u_bbr.flex7 = 1;
2922 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2923 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2924 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2925 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2926 log.u_bbr.pacing_gain = rack->r_must_retran;
2927 /*
2928 * We capture in delRate the upper 32 bits as
2929 * the confidence level we had declared, and the
2930 * lower 32 bits as the actual RTT using the arrival
2931 * timestamp.
2932 */
2933 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2934 log.u_bbr.delRate <<= 32;
2935 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2936 /* Lets capture all the things that make up t_rtxcur */
2937 log.u_bbr.applimited = rack_rto_min;
2938 log.u_bbr.epoch = rack_rto_max;
2939 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2940 log.u_bbr.lost = rack_rto_min;
2941 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2942 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2943 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2944 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2945 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2946 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2947 &rack->rc_inp->inp_socket->so_rcv,
2948 &rack->rc_inp->inp_socket->so_snd,
2949 TCP_LOG_RTT, 0,
2950 0, &log, false, &tv);
2951 }
2952 }
2953
2954 static void
rack_log_rtt_sample_calc(struct tcp_rack * rack,uint32_t rtt,uint32_t send_time,uint32_t ack_time,int where)2955 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2956 {
2957 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2958 union tcp_log_stackspecific log;
2959 struct timeval tv;
2960
2961 /* Convert our ms to a microsecond */
2962 memset(&log, 0, sizeof(log));
2963 log.u_bbr.flex1 = rtt;
2964 log.u_bbr.flex2 = send_time;
2965 log.u_bbr.flex3 = ack_time;
2966 log.u_bbr.flex4 = where;
2967 log.u_bbr.flex7 = 2;
2968 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2969 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2970 log.u_bbr.bw_inuse <<= 32;
2971 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2972 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2973 &rack->rc_inp->inp_socket->so_rcv,
2974 &rack->rc_inp->inp_socket->so_snd,
2975 TCP_LOG_RTT, 0,
2976 0, &log, false, &tv);
2977 }
2978 }
2979
2980
2981 static void
rack_log_rtt_sendmap(struct tcp_rack * rack,uint32_t idx,uint64_t tsv,uint32_t tsecho)2982 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2983 {
2984 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2985 union tcp_log_stackspecific log;
2986 struct timeval tv;
2987
2988 /* Convert our ms to a microsecond */
2989 memset(&log, 0, sizeof(log));
2990 log.u_bbr.flex1 = idx;
2991 log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2992 log.u_bbr.flex3 = tsecho;
2993 log.u_bbr.flex7 = 3;
2994 log.u_bbr.rttProp = tsv;
2995 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2996 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
2997 log.u_bbr.bw_inuse <<= 32;
2998 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
2999 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3000 &rack->rc_inp->inp_socket->so_rcv,
3001 &rack->rc_inp->inp_socket->so_snd,
3002 TCP_LOG_RTT, 0,
3003 0, &log, false, &tv);
3004 }
3005 }
3006
3007
3008 static inline void
rack_log_progress_event(struct tcp_rack * rack,struct tcpcb * tp,uint32_t tick,int event,int line)3009 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line)
3010 {
3011 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3012 union tcp_log_stackspecific log;
3013 struct timeval tv;
3014
3015 memset(&log, 0, sizeof(log));
3016 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3017 log.u_bbr.flex1 = line;
3018 log.u_bbr.flex2 = tick;
3019 log.u_bbr.flex3 = tp->t_maxunacktime;
3020 log.u_bbr.flex4 = tp->t_acktime;
3021 log.u_bbr.flex8 = event;
3022 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3023 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3024 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3025 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3026 log.u_bbr.pacing_gain = rack->r_must_retran;
3027 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3028 log.u_bbr.bw_inuse <<= 32;
3029 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3030 TCP_LOG_EVENTP(tp, NULL,
3031 &rack->rc_inp->inp_socket->so_rcv,
3032 &rack->rc_inp->inp_socket->so_snd,
3033 BBR_LOG_PROGRESS, 0,
3034 0, &log, false, &tv);
3035 }
3036 }
3037
3038 static void
rack_log_type_bbrsnd(struct tcp_rack * rack,uint32_t len,uint32_t slot,uint32_t cts,struct timeval * tv,int line)3039 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
3040 {
3041 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3042 union tcp_log_stackspecific log;
3043
3044 memset(&log, 0, sizeof(log));
3045 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3046 log.u_bbr.flex1 = slot;
3047 if (rack->rack_no_prr)
3048 log.u_bbr.flex2 = 0;
3049 else
3050 log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3051 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3052 log.u_bbr.flex6 = line;
3053 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3054 log.u_bbr.flex8 = rack->rc_in_persist;
3055 log.u_bbr.timeStamp = cts;
3056 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3057 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3058 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3059 log.u_bbr.pacing_gain = rack->r_must_retran;
3060 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3061 &rack->rc_inp->inp_socket->so_rcv,
3062 &rack->rc_inp->inp_socket->so_snd,
3063 BBR_LOG_BBRSND, 0,
3064 0, &log, false, tv);
3065 }
3066 }
3067
3068 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)3069 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 {
3071 if (tcp_bblogging_on(rack->rc_tp)) {
3072 union tcp_log_stackspecific log;
3073 struct timeval tv;
3074
3075 memset(&log, 0, sizeof(log));
3076 log.u_bbr.flex1 = did_out;
3077 log.u_bbr.flex2 = nxt_pkt;
3078 log.u_bbr.flex3 = way_out;
3079 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3080 if (rack->rack_no_prr)
3081 log.u_bbr.flex5 = 0;
3082 else
3083 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3084 log.u_bbr.flex6 = nsegs;
3085 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3086 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data; /* Do we have ack-can-send set */
3087 log.u_bbr.flex7 <<= 1;
3088 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
3089 log.u_bbr.flex7 <<= 1;
3090 log.u_bbr.flex7 |= rack->r_wanted_output; /* Do we want output */
3091 log.u_bbr.flex8 = rack->rc_in_persist;
3092 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3093 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3094 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3095 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3096 log.u_bbr.use_lt_bw <<= 1;
3097 log.u_bbr.use_lt_bw |= rack->r_might_revert;
3098 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3099 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3100 log.u_bbr.pacing_gain = rack->r_must_retran;
3101 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3102 log.u_bbr.bw_inuse <<= 32;
3103 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3104 log.u_bbr.epoch = rack->rc_inp->inp_socket->so_snd.sb_hiwat;
3105 log.u_bbr.lt_epoch = rack->rc_inp->inp_socket->so_rcv.sb_hiwat;
3106 log.u_bbr.lost = rack->rc_tp->t_srtt;
3107 log.u_bbr.pkt_epoch = rack->rc_tp->rfbuf_cnt;
3108 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3109 &rack->rc_inp->inp_socket->so_rcv,
3110 &rack->rc_inp->inp_socket->so_snd,
3111 BBR_LOG_DOSEG_DONE, 0,
3112 0, &log, false, &tv);
3113 }
3114 }
3115
3116 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)3117 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 {
3119 if (tcp_bblogging_on(rack->rc_tp)) {
3120 union tcp_log_stackspecific log;
3121 struct timeval tv;
3122
3123 memset(&log, 0, sizeof(log));
3124 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3125 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3126 log.u_bbr.flex4 = arg1;
3127 log.u_bbr.flex5 = arg2;
3128 log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3129 log.u_bbr.flex6 = arg3;
3130 log.u_bbr.flex8 = frm;
3131 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3132 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3133 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3134 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3135 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3136 log.u_bbr.pacing_gain = rack->r_must_retran;
3137 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3138 &tptosocket(tp)->so_snd,
3139 TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3140 }
3141 }
3142
3143 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)3144 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3145 uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3146 {
3147 if (tcp_bblogging_on(rack->rc_tp)) {
3148 union tcp_log_stackspecific log;
3149 struct timeval tv;
3150
3151 memset(&log, 0, sizeof(log));
3152 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3153 log.u_bbr.flex1 = slot;
3154 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3155 log.u_bbr.flex4 = reason;
3156 if (rack->rack_no_prr)
3157 log.u_bbr.flex5 = 0;
3158 else
3159 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3160 log.u_bbr.flex7 = hpts_calling;
3161 log.u_bbr.flex8 = rack->rc_in_persist;
3162 log.u_bbr.lt_epoch = cwnd_to_use;
3163 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3164 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3165 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3166 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3167 log.u_bbr.pacing_gain = rack->r_must_retran;
3168 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3169 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3170 log.u_bbr.bw_inuse <<= 32;
3171 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3172 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3173 &rack->rc_inp->inp_socket->so_rcv,
3174 &rack->rc_inp->inp_socket->so_snd,
3175 BBR_LOG_JUSTRET, 0,
3176 tlen, &log, false, &tv);
3177 }
3178 }
3179
3180 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)3181 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3182 struct timeval *tv, uint32_t flags_on_entry)
3183 {
3184 if (tcp_bblogging_on(rack->rc_tp)) {
3185 union tcp_log_stackspecific log;
3186
3187 memset(&log, 0, sizeof(log));
3188 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3189 log.u_bbr.flex1 = line;
3190 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3191 log.u_bbr.flex3 = flags_on_entry;
3192 log.u_bbr.flex4 = us_cts;
3193 if (rack->rack_no_prr)
3194 log.u_bbr.flex5 = 0;
3195 else
3196 log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3197 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3198 log.u_bbr.flex7 = hpts_removed;
3199 log.u_bbr.flex8 = 1;
3200 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3201 log.u_bbr.timeStamp = us_cts;
3202 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3203 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3204 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3205 log.u_bbr.pacing_gain = rack->r_must_retran;
3206 log.u_bbr.bw_inuse = rack->r_ctl.current_round;
3207 log.u_bbr.bw_inuse <<= 32;
3208 log.u_bbr.bw_inuse |= rack->r_ctl.rc_considered_lost;
3209 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3210 &rack->rc_inp->inp_socket->so_rcv,
3211 &rack->rc_inp->inp_socket->so_snd,
3212 BBR_LOG_TIMERCANC, 0,
3213 0, &log, false, tv);
3214 }
3215 }
3216
3217 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)3218 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3219 uint32_t flex1, uint32_t flex2,
3220 uint32_t flex3, uint32_t flex4,
3221 uint32_t flex5, uint32_t flex6,
3222 uint16_t flex7, uint8_t mod)
3223 {
3224 if (tcp_bblogging_on(rack->rc_tp)) {
3225 union tcp_log_stackspecific log;
3226 struct timeval tv;
3227
3228 if (mod == 1) {
3229 /* No you can't use 1, its for the real to cancel */
3230 return;
3231 }
3232 memset(&log, 0, sizeof(log));
3233 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3234 log.u_bbr.flex1 = flex1;
3235 log.u_bbr.flex2 = flex2;
3236 log.u_bbr.flex3 = flex3;
3237 log.u_bbr.flex4 = flex4;
3238 log.u_bbr.flex5 = flex5;
3239 log.u_bbr.flex6 = flex6;
3240 log.u_bbr.flex7 = flex7;
3241 log.u_bbr.flex8 = mod;
3242 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3243 &rack->rc_inp->inp_socket->so_rcv,
3244 &rack->rc_inp->inp_socket->so_snd,
3245 BBR_LOG_TIMERCANC, 0,
3246 0, &log, false, &tv);
3247 }
3248 }
3249
3250 static void
rack_log_to_processing(struct tcp_rack * rack,uint32_t cts,int32_t ret,int32_t timers)3251 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3252 {
3253 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3254 union tcp_log_stackspecific log;
3255 struct timeval tv;
3256
3257 memset(&log, 0, sizeof(log));
3258 log.u_bbr.flex1 = timers;
3259 log.u_bbr.flex2 = ret;
3260 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3261 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3262 log.u_bbr.flex5 = cts;
3263 if (rack->rack_no_prr)
3264 log.u_bbr.flex6 = 0;
3265 else
3266 log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3267 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3268 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3269 log.u_bbr.pacing_gain = rack->r_must_retran;
3270 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3271 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3272 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3273 &rack->rc_inp->inp_socket->so_rcv,
3274 &rack->rc_inp->inp_socket->so_snd,
3275 BBR_LOG_TO_PROCESS, 0,
3276 0, &log, false, &tv);
3277 }
3278 }
3279
3280 static void
rack_log_to_prr(struct tcp_rack * rack,int frm,int orig_cwnd,int line)3281 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3282 {
3283 if (tcp_bblogging_on(rack->rc_tp)) {
3284 union tcp_log_stackspecific log;
3285 struct timeval tv;
3286
3287 memset(&log, 0, sizeof(log));
3288 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3289 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3290 if (rack->rack_no_prr)
3291 log.u_bbr.flex3 = 0;
3292 else
3293 log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3294 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3295 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3296 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3297 log.u_bbr.flex7 = line;
3298 log.u_bbr.flex8 = frm;
3299 log.u_bbr.pkts_out = orig_cwnd;
3300 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3301 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3302 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3303 log.u_bbr.use_lt_bw <<= 1;
3304 log.u_bbr.use_lt_bw |= rack->r_might_revert;
3305 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3306 &rack->rc_inp->inp_socket->so_rcv,
3307 &rack->rc_inp->inp_socket->so_snd,
3308 BBR_LOG_BBRUPD, 0,
3309 0, &log, false, &tv);
3310 }
3311 }
3312
3313 static void
rack_counter_destroy(void)3314 rack_counter_destroy(void)
3315 {
3316 counter_u64_free(rack_total_bytes);
3317 counter_u64_free(rack_fto_send);
3318 counter_u64_free(rack_fto_rsm_send);
3319 counter_u64_free(rack_nfto_resend);
3320 counter_u64_free(rack_hw_pace_init_fail);
3321 counter_u64_free(rack_hw_pace_lost);
3322 counter_u64_free(rack_non_fto_send);
3323 counter_u64_free(rack_extended_rfo);
3324 counter_u64_free(rack_ack_total);
3325 counter_u64_free(rack_express_sack);
3326 counter_u64_free(rack_sack_total);
3327 counter_u64_free(rack_move_none);
3328 counter_u64_free(rack_move_some);
3329 counter_u64_free(rack_sack_attacks_detected);
3330 counter_u64_free(rack_sack_attacks_reversed);
3331 counter_u64_free(rack_sack_attacks_suspect);
3332 counter_u64_free(rack_sack_used_next_merge);
3333 counter_u64_free(rack_sack_used_prev_merge);
3334 counter_u64_free(rack_tlp_tot);
3335 counter_u64_free(rack_tlp_newdata);
3336 counter_u64_free(rack_tlp_retran);
3337 counter_u64_free(rack_tlp_retran_bytes);
3338 counter_u64_free(rack_to_tot);
3339 counter_u64_free(rack_saw_enobuf);
3340 counter_u64_free(rack_saw_enobuf_hw);
3341 counter_u64_free(rack_saw_enetunreach);
3342 counter_u64_free(rack_hot_alloc);
3343 counter_u64_free(rack_to_alloc);
3344 counter_u64_free(rack_to_alloc_hard);
3345 counter_u64_free(rack_to_alloc_emerg);
3346 counter_u64_free(rack_to_alloc_limited);
3347 counter_u64_free(rack_alloc_limited_conns);
3348 counter_u64_free(rack_split_limited);
3349 counter_u64_free(rack_multi_single_eq);
3350 counter_u64_free(rack_rxt_clamps_cwnd);
3351 counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3352 counter_u64_free(rack_proc_non_comp_ack);
3353 counter_u64_free(rack_sack_proc_all);
3354 counter_u64_free(rack_sack_proc_restart);
3355 counter_u64_free(rack_sack_proc_short);
3356 counter_u64_free(rack_sack_skipped_acked);
3357 counter_u64_free(rack_sack_splits);
3358 counter_u64_free(rack_input_idle_reduces);
3359 counter_u64_free(rack_collapsed_win);
3360 counter_u64_free(rack_collapsed_win_rxt);
3361 counter_u64_free(rack_collapsed_win_rxt_bytes);
3362 counter_u64_free(rack_collapsed_win_seen);
3363 counter_u64_free(rack_try_scwnd);
3364 counter_u64_free(rack_persists_sends);
3365 counter_u64_free(rack_persists_acks);
3366 counter_u64_free(rack_persists_loss);
3367 counter_u64_free(rack_persists_lost_ends);
3368 #ifdef INVARIANTS
3369 counter_u64_free(rack_adjust_map_bw);
3370 #endif
3371 COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3372 COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3373 }
3374
3375 static struct rack_sendmap *
rack_alloc(struct tcp_rack * rack)3376 rack_alloc(struct tcp_rack *rack)
3377 {
3378 struct rack_sendmap *rsm;
3379
3380 /*
3381 * First get the top of the list it in
3382 * theory is the "hottest" rsm we have,
3383 * possibly just freed by ack processing.
3384 */
3385 if (rack->rc_free_cnt > rack_free_cache) {
3386 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3387 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3388 counter_u64_add(rack_hot_alloc, 1);
3389 rack->rc_free_cnt--;
3390 return (rsm);
3391 }
3392 /*
3393 * Once we get under our free cache we probably
3394 * no longer have a "hot" one available. Lets
3395 * get one from UMA.
3396 */
3397 rsm = uma_zalloc(rack_zone, M_NOWAIT);
3398 if (rsm) {
3399 rack->r_ctl.rc_num_maps_alloced++;
3400 counter_u64_add(rack_to_alloc, 1);
3401 return (rsm);
3402 }
3403 /*
3404 * Dig in to our aux rsm's (the last two) since
3405 * UMA failed to get us one.
3406 */
3407 if (rack->rc_free_cnt) {
3408 counter_u64_add(rack_to_alloc_emerg, 1);
3409 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3410 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3411 rack->rc_free_cnt--;
3412 return (rsm);
3413 }
3414 return (NULL);
3415 }
3416
3417 static struct rack_sendmap *
rack_alloc_full_limit(struct tcp_rack * rack)3418 rack_alloc_full_limit(struct tcp_rack *rack)
3419 {
3420 if ((V_tcp_map_entries_limit > 0) &&
3421 (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3422 counter_u64_add(rack_to_alloc_limited, 1);
3423 if (!rack->alloc_limit_reported) {
3424 rack->alloc_limit_reported = 1;
3425 counter_u64_add(rack_alloc_limited_conns, 1);
3426 }
3427 return (NULL);
3428 }
3429 return (rack_alloc(rack));
3430 }
3431
3432 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3433 static struct rack_sendmap *
rack_alloc_limit(struct tcp_rack * rack,uint8_t limit_type)3434 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3435 {
3436 struct rack_sendmap *rsm;
3437
3438 if (limit_type) {
3439 /* currently there is only one limit type */
3440 if (rack->r_ctl.rc_split_limit > 0 &&
3441 rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3442 counter_u64_add(rack_split_limited, 1);
3443 if (!rack->alloc_limit_reported) {
3444 rack->alloc_limit_reported = 1;
3445 counter_u64_add(rack_alloc_limited_conns, 1);
3446 }
3447 return (NULL);
3448 }
3449 }
3450
3451 /* allocate and mark in the limit type, if set */
3452 rsm = rack_alloc(rack);
3453 if (rsm != NULL && limit_type) {
3454 rsm->r_limit_type = limit_type;
3455 rack->r_ctl.rc_num_split_allocs++;
3456 }
3457 return (rsm);
3458 }
3459
3460 static void
rack_free_trim(struct tcp_rack * rack)3461 rack_free_trim(struct tcp_rack *rack)
3462 {
3463 struct rack_sendmap *rsm;
3464
3465 /*
3466 * Free up all the tail entries until
3467 * we get our list down to the limit.
3468 */
3469 while (rack->rc_free_cnt > rack_free_cache) {
3470 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3471 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3472 rack->rc_free_cnt--;
3473 rack->r_ctl.rc_num_maps_alloced--;
3474 uma_zfree(rack_zone, rsm);
3475 }
3476 }
3477
3478 static void
rack_free(struct tcp_rack * rack,struct rack_sendmap * rsm)3479 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3480 {
3481 if (rsm->r_flags & RACK_APP_LIMITED) {
3482 KASSERT((rack->r_ctl.rc_app_limited_cnt > 0),
3483 ("app_cnt %u, rsm %p", rack->r_ctl.rc_app_limited_cnt, rsm));
3484 rack->r_ctl.rc_app_limited_cnt--;
3485 }
3486 if (rsm->r_limit_type) {
3487 /* currently there is only one limit type */
3488 rack->r_ctl.rc_num_split_allocs--;
3489 }
3490 if (rsm == rack->r_ctl.rc_first_appl) {
3491 rack->r_ctl.cleared_app_ack_seq = rsm->r_end;
3492 rack->r_ctl.cleared_app_ack = 1;
3493 if (rack->r_ctl.rc_app_limited_cnt == 0)
3494 rack->r_ctl.rc_first_appl = NULL;
3495 else
3496 rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3497 }
3498 if (rsm == rack->r_ctl.rc_resend)
3499 rack->r_ctl.rc_resend = NULL;
3500 if (rsm == rack->r_ctl.rc_end_appl)
3501 rack->r_ctl.rc_end_appl = NULL;
3502 if (rack->r_ctl.rc_tlpsend == rsm)
3503 rack->r_ctl.rc_tlpsend = NULL;
3504 if (rack->r_ctl.rc_sacklast == rsm)
3505 rack->r_ctl.rc_sacklast = NULL;
3506 memset(rsm, 0, sizeof(struct rack_sendmap));
3507 /* Make sure we are not going to overrun our count limit of 0xff */
3508 if ((rack->rc_free_cnt + 1) > RACK_FREE_CNT_MAX) {
3509 rack_free_trim(rack);
3510 }
3511 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3512 rack->rc_free_cnt++;
3513 }
3514
3515 static uint32_t
rack_get_measure_window(struct tcpcb * tp,struct tcp_rack * rack)3516 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3517 {
3518 uint64_t srtt, bw, len, tim;
3519 uint32_t segsiz, def_len, minl;
3520
3521 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3522 def_len = rack_def_data_window * segsiz;
3523 if (rack->rc_gp_filled == 0) {
3524 /*
3525 * We have no measurement (IW is in flight?) so
3526 * we can only guess using our data_window sysctl
3527 * value (usually 20MSS).
3528 */
3529 return (def_len);
3530 }
3531 /*
3532 * Now we have a number of factors to consider.
3533 *
3534 * 1) We have a desired BDP which is usually
3535 * at least 2.
3536 * 2) We have a minimum number of rtt's usually 1 SRTT
3537 * but we allow it too to be more.
3538 * 3) We want to make sure a measurement last N useconds (if
3539 * we have set rack_min_measure_usec.
3540 *
3541 * We handle the first concern here by trying to create a data
3542 * window of max(rack_def_data_window, DesiredBDP). The
3543 * second concern we handle in not letting the measurement
3544 * window end normally until at least the required SRTT's
3545 * have gone by which is done further below in
3546 * rack_enough_for_measurement(). Finally the third concern
3547 * we also handle here by calculating how long that time
3548 * would take at the current BW and then return the
3549 * max of our first calculation and that length. Note
3550 * that if rack_min_measure_usec is 0, we don't deal
3551 * with concern 3. Also for both Concern 1 and 3 an
3552 * application limited period could end the measurement
3553 * earlier.
3554 *
3555 * So lets calculate the BDP with the "known" b/w using
3556 * the SRTT as our rtt and then multiply it by the goal.
3557 */
3558 bw = rack_get_bw(rack);
3559 srtt = (uint64_t)tp->t_srtt;
3560 len = bw * srtt;
3561 len /= (uint64_t)HPTS_USEC_IN_SEC;
3562 len *= max(1, rack_goal_bdp);
3563 /* Now we need to round up to the nearest MSS */
3564 len = roundup(len, segsiz);
3565 if (rack_min_measure_usec) {
3566 /* Now calculate our min length for this b/w */
3567 tim = rack_min_measure_usec;
3568 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3569 if (minl == 0)
3570 minl = 1;
3571 minl = roundup(minl, segsiz);
3572 if (len < minl)
3573 len = minl;
3574 }
3575 /*
3576 * Now if we have a very small window we want
3577 * to attempt to get the window that is
3578 * as small as possible. This happens on
3579 * low b/w connections and we don't want to
3580 * span huge numbers of rtt's between measurements.
3581 *
3582 * We basically include 2 over our "MIN window" so
3583 * that the measurement can be shortened (possibly) by
3584 * an ack'ed packet.
3585 */
3586 if (len < def_len)
3587 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3588 else
3589 return (max((uint32_t)len, def_len));
3590
3591 }
3592
3593 static int
rack_enough_for_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq th_ack,uint8_t * quality)3594 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3595 {
3596 uint32_t tim, srtts, segsiz;
3597
3598 /*
3599 * Has enough time passed for the GP measurement to be valid?
3600 */
3601 if (SEQ_LT(th_ack, tp->gput_seq)) {
3602 /* Not enough bytes yet */
3603 return (0);
3604 }
3605 if ((tp->snd_max == tp->snd_una) ||
3606 (th_ack == tp->snd_max)){
3607 /*
3608 * All is acked quality of all acked is
3609 * usually low or medium, but we in theory could split
3610 * all acked into two cases, where you got
3611 * a signifigant amount of your window and
3612 * where you did not. For now we leave it
3613 * but it is something to contemplate in the
3614 * future. The danger here is that delayed ack
3615 * is effecting the last byte (which is a 50:50 chance).
3616 */
3617 *quality = RACK_QUALITY_ALLACKED;
3618 return (1);
3619 }
3620 if (SEQ_GEQ(th_ack, tp->gput_ack)) {
3621 /*
3622 * We obtained our entire window of data we wanted
3623 * no matter if we are in recovery or not then
3624 * its ok since expanding the window does not
3625 * make things fuzzy (or at least not as much).
3626 */
3627 *quality = RACK_QUALITY_HIGH;
3628 return (1);
3629 }
3630 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3631 if (SEQ_LT(th_ack, tp->gput_ack) &&
3632 ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3633 /* Not enough bytes yet */
3634 return (0);
3635 }
3636 if (rack->r_ctl.rc_first_appl &&
3637 (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3638 /*
3639 * We are up to the app limited send point
3640 * we have to measure irrespective of the time..
3641 */
3642 *quality = RACK_QUALITY_APPLIMITED;
3643 return (1);
3644 }
3645 /* Now what about time? */
3646 srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3647 tim = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3648 if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3649 /*
3650 * We do not allow a measurement if we are in recovery
3651 * that would shrink the goodput window we wanted.
3652 * This is to prevent cloudyness of when the last send
3653 * was actually made.
3654 */
3655 *quality = RACK_QUALITY_HIGH;
3656 return (1);
3657 }
3658 /* Nope not even a full SRTT has passed */
3659 return (0);
3660 }
3661
3662 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)3663 rack_log_timely(struct tcp_rack *rack,
3664 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3665 uint64_t up_bnd, int line, uint8_t method)
3666 {
3667 if (tcp_bblogging_on(rack->rc_tp)) {
3668 union tcp_log_stackspecific log;
3669 struct timeval tv;
3670
3671 memset(&log, 0, sizeof(log));
3672 log.u_bbr.flex1 = logged;
3673 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3674 log.u_bbr.flex2 <<= 4;
3675 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3676 log.u_bbr.flex2 <<= 4;
3677 log.u_bbr.flex2 |= rack->rc_gp_incr;
3678 log.u_bbr.flex2 <<= 4;
3679 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3680 log.u_bbr.flex3 = rack->rc_gp_incr;
3681 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3682 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3683 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3684 log.u_bbr.flex7 = rack->rc_gp_bwred;
3685 log.u_bbr.flex8 = method;
3686 log.u_bbr.cur_del_rate = cur_bw;
3687 log.u_bbr.delRate = low_bnd;
3688 log.u_bbr.bw_inuse = up_bnd;
3689 log.u_bbr.rttProp = rack_get_bw(rack);
3690 log.u_bbr.pkt_epoch = line;
3691 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3692 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3693 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3694 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3695 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3696 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3697 log.u_bbr.cwnd_gain <<= 1;
3698 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3699 log.u_bbr.cwnd_gain <<= 1;
3700 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3701 log.u_bbr.cwnd_gain <<= 1;
3702 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3703 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3704 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3705 &rack->rc_inp->inp_socket->so_rcv,
3706 &rack->rc_inp->inp_socket->so_snd,
3707 TCP_TIMELY_WORK, 0,
3708 0, &log, false, &tv);
3709 }
3710 }
3711
3712 static int
rack_bw_can_be_raised(struct tcp_rack * rack,uint64_t cur_bw,uint64_t last_bw_est,uint16_t mult)3713 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3714 {
3715 /*
3716 * Before we increase we need to know if
3717 * the estimate just made was less than
3718 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3719 *
3720 * If we already are pacing at a fast enough
3721 * rate to push us faster there is no sense of
3722 * increasing.
3723 *
3724 * We first caculate our actual pacing rate (ss or ca multiplier
3725 * times our cur_bw).
3726 *
3727 * Then we take the last measured rate and multipy by our
3728 * maximum pacing overage to give us a max allowable rate.
3729 *
3730 * If our act_rate is smaller than our max_allowable rate
3731 * then we should increase. Else we should hold steady.
3732 *
3733 */
3734 uint64_t act_rate, max_allow_rate;
3735
3736 if (rack_timely_no_stopping)
3737 return (1);
3738
3739 if ((cur_bw == 0) || (last_bw_est == 0)) {
3740 /*
3741 * Initial startup case or
3742 * everything is acked case.
3743 */
3744 rack_log_timely(rack, mult, cur_bw, 0, 0,
3745 __LINE__, 9);
3746 return (1);
3747 }
3748 if (mult <= 100) {
3749 /*
3750 * We can always pace at or slightly above our rate.
3751 */
3752 rack_log_timely(rack, mult, cur_bw, 0, 0,
3753 __LINE__, 9);
3754 return (1);
3755 }
3756 act_rate = cur_bw * (uint64_t)mult;
3757 act_rate /= 100;
3758 max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3759 max_allow_rate /= 100;
3760 if (act_rate < max_allow_rate) {
3761 /*
3762 * Here the rate we are actually pacing at
3763 * is smaller than 10% above our last measurement.
3764 * This means we are pacing below what we would
3765 * like to try to achieve (plus some wiggle room).
3766 */
3767 rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate,
3768 __LINE__, 9);
3769 return (1);
3770 } else {
3771 /*
3772 * Here we are already pacing at least rack_max_per_above(10%)
3773 * what we are getting back. This indicates most likely
3774 * that we are being limited (cwnd/rwnd/app) and can't
3775 * get any more b/w. There is no sense of trying to
3776 * raise up the pacing rate its not speeding us up
3777 * and we already are pacing faster than we are getting.
3778 */
3779 rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate,
3780 __LINE__, 8);
3781 return (0);
3782 }
3783 }
3784
3785 static void
rack_validate_multipliers_at_or_above100(struct tcp_rack * rack)3786 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3787 {
3788 /*
3789 * When we drag bottom, we want to assure
3790 * that no multiplier is below 1.0, if so
3791 * we want to restore it to at least that.
3792 */
3793 if (rack->r_ctl.rack_per_of_gp_rec < 100) {
3794 /* This is unlikely we usually do not touch recovery */
3795 rack->r_ctl.rack_per_of_gp_rec = 100;
3796 }
3797 if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3798 rack->r_ctl.rack_per_of_gp_ca = 100;
3799 }
3800 if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3801 rack->r_ctl.rack_per_of_gp_ss = 100;
3802 }
3803 }
3804
3805 static void
rack_validate_multipliers_at_or_below_100(struct tcp_rack * rack)3806 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3807 {
3808 if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3809 rack->r_ctl.rack_per_of_gp_ca = 100;
3810 }
3811 if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3812 rack->r_ctl.rack_per_of_gp_ss = 100;
3813 }
3814 }
3815
3816 static void
rack_increase_bw_mul(struct tcp_rack * rack,int timely_says,uint64_t cur_bw,uint64_t last_bw_est,int override)3817 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3818 {
3819 int32_t calc, logged, plus;
3820
3821 logged = 0;
3822
3823 if (rack->rc_skip_timely)
3824 return;
3825 if (override) {
3826 /*
3827 * override is passed when we are
3828 * loosing b/w and making one last
3829 * gasp at trying to not loose out
3830 * to a new-reno flow.
3831 */
3832 goto extra_boost;
3833 }
3834 /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3835 if (rack->rc_gp_incr &&
3836 ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3837 /*
3838 * Reset and get 5 strokes more before the boost. Note
3839 * that the count is 0 based so we have to add one.
3840 */
3841 extra_boost:
3842 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3843 rack->rc_gp_timely_inc_cnt = 0;
3844 } else
3845 plus = (uint32_t)rack_gp_increase_per;
3846 /* Must be at least 1% increase for true timely increases */
3847 if ((plus < 1) &&
3848 ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3849 plus = 1;
3850 if (rack->rc_gp_saw_rec &&
3851 (rack->rc_gp_no_rec_chg == 0) &&
3852 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3853 rack->r_ctl.rack_per_of_gp_rec)) {
3854 /* We have been in recovery ding it too */
3855 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3856 if (calc > 0xffff)
3857 calc = 0xffff;
3858 logged |= 1;
3859 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3860 if (rack->r_ctl.rack_per_upper_bound_ca &&
3861 (rack->rc_dragged_bottom == 0) &&
3862 (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3863 rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3864 }
3865 if (rack->rc_gp_saw_ca &&
3866 (rack->rc_gp_saw_ss == 0) &&
3867 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3868 rack->r_ctl.rack_per_of_gp_ca)) {
3869 /* In CA */
3870 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3871 if (calc > 0xffff)
3872 calc = 0xffff;
3873 logged |= 2;
3874 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3875 if (rack->r_ctl.rack_per_upper_bound_ca &&
3876 (rack->rc_dragged_bottom == 0) &&
3877 (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3878 rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3879 }
3880 if (rack->rc_gp_saw_ss &&
3881 rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3882 rack->r_ctl.rack_per_of_gp_ss)) {
3883 /* In SS */
3884 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3885 if (calc > 0xffff)
3886 calc = 0xffff;
3887 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3888 if (rack->r_ctl.rack_per_upper_bound_ss &&
3889 (rack->rc_dragged_bottom == 0) &&
3890 (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3891 rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3892 logged |= 4;
3893 }
3894 if (logged &&
3895 (rack->rc_gp_incr == 0)){
3896 /* Go into increment mode */
3897 rack->rc_gp_incr = 1;
3898 rack->rc_gp_timely_inc_cnt = 0;
3899 }
3900 if (rack->rc_gp_incr &&
3901 logged &&
3902 (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3903 rack->rc_gp_timely_inc_cnt++;
3904 }
3905 rack_log_timely(rack, logged, plus, 0, 0,
3906 __LINE__, 1);
3907 }
3908
3909 static uint32_t
rack_get_decrease(struct tcp_rack * rack,uint32_t curper,int32_t rtt_diff)3910 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3911 {
3912 /*-
3913 * norm_grad = rtt_diff / minrtt;
3914 * new_per = curper * (1 - B * norm_grad)
3915 *
3916 * B = rack_gp_decrease_per (default 80%)
3917 * rtt_dif = input var current rtt-diff
3918 * curper = input var current percentage
3919 * minrtt = from rack filter
3920 *
3921 * In order to do the floating point calculations above we
3922 * do an integer conversion. The code looks confusing so let me
3923 * translate it into something that use more variables and
3924 * is clearer for us humans :)
3925 *
3926 * uint64_t norm_grad, inverse, reduce_by, final_result;
3927 * uint32_t perf;
3928 *
3929 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3930 * (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3931 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3932 * inverse /= 1000000;
3933 * reduce_by = (1000000 - inverse);
3934 * final_result = (cur_per * reduce_by) / 1000000;
3935 * perf = (uint32_t)final_result;
3936 */
3937 uint64_t perf;
3938
3939 perf = (((uint64_t)curper * ((uint64_t)1000000 -
3940 ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3941 (((uint64_t)rtt_diff * (uint64_t)1000000)/
3942 (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3943 (uint64_t)1000000)) /
3944 (uint64_t)1000000);
3945 if (perf > curper) {
3946 /* TSNH */
3947 perf = curper - 1;
3948 }
3949 return ((uint32_t)perf);
3950 }
3951
3952 static uint32_t
rack_decrease_highrtt(struct tcp_rack * rack,uint32_t curper,uint32_t rtt)3953 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3954 {
3955 /*
3956 * highrttthresh
3957 * result = curper * (1 - (B * ( 1 - ------ ))
3958 * gp_srtt
3959 *
3960 * B = rack_gp_decrease_per (default .8 i.e. 80)
3961 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3962 */
3963 uint64_t perf;
3964 uint32_t highrttthresh;
3965
3966 highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3967
3968 perf = (((uint64_t)curper * ((uint64_t)1000000 -
3969 ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3970 ((uint64_t)highrttthresh * (uint64_t)1000000) /
3971 (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3972 if (tcp_bblogging_on(rack->rc_tp)) {
3973 uint64_t log1;
3974
3975 log1 = rtt;
3976 log1 <<= 32;
3977 log1 |= highrttthresh;
3978 rack_log_timely(rack,
3979 rack_gp_decrease_per,
3980 (uint64_t)curper,
3981 log1,
3982 perf,
3983 __LINE__,
3984 15);
3985 }
3986 return (perf);
3987 }
3988
3989 static void
rack_decrease_bw_mul(struct tcp_rack * rack,int timely_says,uint32_t rtt,int32_t rtt_diff)3990 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3991 {
3992 uint64_t logvar, logvar2, logvar3;
3993 uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3994
3995 if (rack->rc_skip_timely)
3996 return;
3997 if (rack->rc_gp_incr) {
3998 /* Turn off increment counting */
3999 rack->rc_gp_incr = 0;
4000 rack->rc_gp_timely_inc_cnt = 0;
4001 }
4002 ss_red = ca_red = rec_red = 0;
4003 logged = 0;
4004 /* Calculate the reduction value */
4005 if (rtt_diff < 0) {
4006 rtt_diff *= -1;
4007 }
4008 /* Must be at least 1% reduction */
4009 if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
4010 /* We have been in recovery ding it too */
4011 if (timely_says == 2) {
4012 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
4013 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4014 if (alt < new_per)
4015 val = alt;
4016 else
4017 val = new_per;
4018 } else
4019 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4020 if (rack->r_ctl.rack_per_of_gp_rec > val) {
4021 rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4022 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4023 } else {
4024 rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4025 rec_red = 0;
4026 }
4027 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4028 rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4029 logged |= 1;
4030 }
4031 if (rack->rc_gp_saw_ss) {
4032 /* Sent in SS */
4033 if (timely_says == 2) {
4034 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4035 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4036 if (alt < new_per)
4037 val = alt;
4038 else
4039 val = new_per;
4040 } else
4041 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4042 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4043 ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4044 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4045 } else {
4046 ss_red = new_per;
4047 rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4048 logvar = new_per;
4049 logvar <<= 32;
4050 logvar |= alt;
4051 logvar2 = (uint32_t)rtt;
4052 logvar2 <<= 32;
4053 logvar2 |= (uint32_t)rtt_diff;
4054 logvar3 = rack_gp_rtt_maxmul;
4055 logvar3 <<= 32;
4056 logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4057 rack_log_timely(rack, timely_says,
4058 logvar2, logvar3,
4059 logvar, __LINE__, 10);
4060 }
4061 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4062 rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4063 logged |= 4;
4064 } else if (rack->rc_gp_saw_ca) {
4065 /* Sent in CA */
4066 if (timely_says == 2) {
4067 new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4068 alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4069 if (alt < new_per)
4070 val = alt;
4071 else
4072 val = new_per;
4073 } else
4074 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4075 if (rack->r_ctl.rack_per_of_gp_ca > val) {
4076 ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4077 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4078 } else {
4079 rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4080 ca_red = 0;
4081 logvar = new_per;
4082 logvar <<= 32;
4083 logvar |= alt;
4084 logvar2 = (uint32_t)rtt;
4085 logvar2 <<= 32;
4086 logvar2 |= (uint32_t)rtt_diff;
4087 logvar3 = rack_gp_rtt_maxmul;
4088 logvar3 <<= 32;
4089 logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4090 rack_log_timely(rack, timely_says,
4091 logvar2, logvar3,
4092 logvar, __LINE__, 10);
4093 }
4094 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4095 rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4096 logged |= 2;
4097 }
4098 if (rack->rc_gp_timely_dec_cnt < 0x7) {
4099 rack->rc_gp_timely_dec_cnt++;
4100 if (rack_timely_dec_clear &&
4101 (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4102 rack->rc_gp_timely_dec_cnt = 0;
4103 }
4104 logvar = ss_red;
4105 logvar <<= 32;
4106 logvar |= ca_red;
4107 rack_log_timely(rack, logged, rec_red, rack_per_lower_bound, logvar,
4108 __LINE__, 2);
4109 }
4110
4111 static void
rack_log_rtt_shrinks(struct tcp_rack * rack,uint32_t us_cts,uint32_t rtt,uint32_t line,uint8_t reas)4112 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4113 uint32_t rtt, uint32_t line, uint8_t reas)
4114 {
4115 if (tcp_bblogging_on(rack->rc_tp)) {
4116 union tcp_log_stackspecific log;
4117 struct timeval tv;
4118
4119 memset(&log, 0, sizeof(log));
4120 log.u_bbr.flex1 = line;
4121 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4122 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4123 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4124 log.u_bbr.flex5 = rtt;
4125 log.u_bbr.flex6 = rack->rc_highly_buffered;
4126 log.u_bbr.flex6 <<= 1;
4127 log.u_bbr.flex6 |= rack->forced_ack;
4128 log.u_bbr.flex6 <<= 1;
4129 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4130 log.u_bbr.flex6 <<= 1;
4131 log.u_bbr.flex6 |= rack->in_probe_rtt;
4132 log.u_bbr.flex6 <<= 1;
4133 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4134 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4135 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4136 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4137 log.u_bbr.flex8 = reas;
4138 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4139 log.u_bbr.delRate = rack_get_bw(rack);
4140 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4141 log.u_bbr.cur_del_rate <<= 32;
4142 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4143 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4144 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4145 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4146 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4147 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4148 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4149 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4150 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4151 log.u_bbr.rttProp = us_cts;
4152 log.u_bbr.rttProp <<= 32;
4153 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4154 TCP_LOG_EVENTP(rack->rc_tp, NULL,
4155 &rack->rc_inp->inp_socket->so_rcv,
4156 &rack->rc_inp->inp_socket->so_snd,
4157 BBR_LOG_RTT_SHRINKS, 0,
4158 0, &log, false, &rack->r_ctl.act_rcv_time);
4159 }
4160 }
4161
4162 static void
rack_set_prtt_target(struct tcp_rack * rack,uint32_t segsiz,uint32_t rtt)4163 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4164 {
4165 uint64_t bwdp;
4166
4167 bwdp = rack_get_bw(rack);
4168 bwdp *= (uint64_t)rtt;
4169 bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4170 rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4171 if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4172 /*
4173 * A window protocol must be able to have 4 packets
4174 * outstanding as the floor in order to function
4175 * (especially considering delayed ack :D).
4176 */
4177 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4178 }
4179 }
4180
4181 static void
rack_enter_probertt(struct tcp_rack * rack,uint32_t us_cts)4182 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4183 {
4184 /**
4185 * ProbeRTT is a bit different in rack_pacing than in
4186 * BBR. It is like BBR in that it uses the lowering of
4187 * the RTT as a signal that we saw something new and
4188 * counts from there for how long between. But it is
4189 * different in that its quite simple. It does not
4190 * play with the cwnd and wait until we get down
4191 * to N segments outstanding and hold that for
4192 * 200ms. Instead it just sets the pacing reduction
4193 * rate to a set percentage (70 by default) and hold
4194 * that for a number of recent GP Srtt's.
4195 */
4196 uint32_t segsiz;
4197
4198 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4199 if (rack->rc_gp_dyn_mul == 0)
4200 return;
4201
4202 if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4203 /* We are idle */
4204 return;
4205 }
4206 if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4207 SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4208 /*
4209 * Stop the goodput now, the idea here is
4210 * that future measurements with in_probe_rtt
4211 * won't register if they are not greater so
4212 * we want to get what info (if any) is available
4213 * now.
4214 */
4215 rack_do_goodput_measurement(rack->rc_tp, rack,
4216 rack->rc_tp->snd_una, __LINE__,
4217 RACK_QUALITY_PROBERTT);
4218 }
4219 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4220 rack->r_ctl.rc_time_probertt_entered = us_cts;
4221 segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4222 rack->r_ctl.rc_pace_min_segs);
4223 rack->in_probe_rtt = 1;
4224 rack->measure_saw_probe_rtt = 1;
4225 rack->r_ctl.rc_time_probertt_starts = 0;
4226 rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4227 if (rack_probertt_use_min_rtt_entry)
4228 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4229 else
4230 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4231 rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4232 __LINE__, RACK_RTTS_ENTERPROBE);
4233 }
4234
4235 static void
rack_exit_probertt(struct tcp_rack * rack,uint32_t us_cts)4236 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4237 {
4238 struct rack_sendmap *rsm;
4239 uint32_t segsiz;
4240
4241 segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4242 rack->r_ctl.rc_pace_min_segs);
4243 rack->in_probe_rtt = 0;
4244 if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4245 SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4246 /*
4247 * Stop the goodput now, the idea here is
4248 * that future measurements with in_probe_rtt
4249 * won't register if they are not greater so
4250 * we want to get what info (if any) is available
4251 * now.
4252 */
4253 rack_do_goodput_measurement(rack->rc_tp, rack,
4254 rack->rc_tp->snd_una, __LINE__,
4255 RACK_QUALITY_PROBERTT);
4256 } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4257 /*
4258 * We don't have enough data to make a measurement.
4259 * So lets just stop and start here after exiting
4260 * probe-rtt. We probably are not interested in
4261 * the results anyway.
4262 */
4263 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4264 }
4265 /*
4266 * Measurements through the current snd_max are going
4267 * to be limited by the slower pacing rate.
4268 *
4269 * We need to mark these as app-limited so we
4270 * don't collapse the b/w.
4271 */
4272 rsm = tqhash_max(rack->r_ctl.tqh);
4273 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4274 if (rack->r_ctl.rc_app_limited_cnt == 0)
4275 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4276 else {
4277 /*
4278 * Go out to the end app limited and mark
4279 * this new one as next and move the end_appl up
4280 * to this guy.
4281 */
4282 if (rack->r_ctl.rc_end_appl)
4283 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4284 rack->r_ctl.rc_end_appl = rsm;
4285 }
4286 rsm->r_flags |= RACK_APP_LIMITED;
4287 rack->r_ctl.rc_app_limited_cnt++;
4288 }
4289 /*
4290 * Now, we need to examine our pacing rate multipliers.
4291 * If its under 100%, we need to kick it back up to
4292 * 100%. We also don't let it be over our "max" above
4293 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4294 * Note setting clamp_atexit_prtt to 0 has the effect
4295 * of setting CA/SS to 100% always at exit (which is
4296 * the default behavior).
4297 */
4298 if (rack_probertt_clear_is) {
4299 rack->rc_gp_incr = 0;
4300 rack->rc_gp_bwred = 0;
4301 rack->rc_gp_timely_inc_cnt = 0;
4302 rack->rc_gp_timely_dec_cnt = 0;
4303 }
4304 /* Do we do any clamping at exit? */
4305 if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4306 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4307 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4308 }
4309 if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4310 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4311 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4312 }
4313 /*
4314 * Lets set rtt_diff to 0, so that we will get a "boost"
4315 * after exiting.
4316 */
4317 rack->r_ctl.rc_rtt_diff = 0;
4318
4319 /* Clear all flags so we start fresh */
4320 rack->rc_tp->t_bytes_acked = 0;
4321 rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4322 /*
4323 * If configured to, set the cwnd and ssthresh to
4324 * our targets.
4325 */
4326 if (rack_probe_rtt_sets_cwnd) {
4327 uint64_t ebdp;
4328 uint32_t setto;
4329
4330 /* Set ssthresh so we get into CA once we hit our target */
4331 if (rack_probertt_use_min_rtt_exit == 1) {
4332 /* Set to min rtt */
4333 rack_set_prtt_target(rack, segsiz,
4334 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4335 } else if (rack_probertt_use_min_rtt_exit == 2) {
4336 /* Set to current gp rtt */
4337 rack_set_prtt_target(rack, segsiz,
4338 rack->r_ctl.rc_gp_srtt);
4339 } else if (rack_probertt_use_min_rtt_exit == 3) {
4340 /* Set to entry gp rtt */
4341 rack_set_prtt_target(rack, segsiz,
4342 rack->r_ctl.rc_entry_gp_rtt);
4343 } else {
4344 uint64_t sum;
4345 uint32_t setval;
4346
4347 sum = rack->r_ctl.rc_entry_gp_rtt;
4348 sum *= 10;
4349 sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4350 if (sum >= 20) {
4351 /*
4352 * A highly buffered path needs
4353 * cwnd space for timely to work.
4354 * Lets set things up as if
4355 * we are heading back here again.
4356 */
4357 setval = rack->r_ctl.rc_entry_gp_rtt;
4358 } else if (sum >= 15) {
4359 /*
4360 * Lets take the smaller of the
4361 * two since we are just somewhat
4362 * buffered.
4363 */
4364 setval = rack->r_ctl.rc_gp_srtt;
4365 if (setval > rack->r_ctl.rc_entry_gp_rtt)
4366 setval = rack->r_ctl.rc_entry_gp_rtt;
4367 } else {
4368 /*
4369 * Here we are not highly buffered
4370 * and should pick the min we can to
4371 * keep from causing loss.
4372 */
4373 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4374 }
4375 rack_set_prtt_target(rack, segsiz,
4376 setval);
4377 }
4378 if (rack_probe_rtt_sets_cwnd > 1) {
4379 /* There is a percentage here to boost */
4380 ebdp = rack->r_ctl.rc_target_probertt_flight;
4381 ebdp *= rack_probe_rtt_sets_cwnd;
4382 ebdp /= 100;
4383 setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4384 } else
4385 setto = rack->r_ctl.rc_target_probertt_flight;
4386 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4387 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4388 /* Enforce a min */
4389 rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4390 }
4391 /* If we set in the cwnd also set the ssthresh point so we are in CA */
4392 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4393 }
4394 rack_log_rtt_shrinks(rack, us_cts,
4395 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4396 __LINE__, RACK_RTTS_EXITPROBE);
4397 /* Clear times last so log has all the info */
4398 rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4399 rack->r_ctl.rc_time_probertt_entered = us_cts;
4400 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4401 rack->r_ctl.rc_time_of_last_probertt = us_cts;
4402 }
4403
4404 static void
rack_check_probe_rtt(struct tcp_rack * rack,uint32_t us_cts)4405 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4406 {
4407 /* Check in on probe-rtt */
4408
4409 if (rack->rc_gp_filled == 0) {
4410 /* We do not do p-rtt unless we have gp measurements */
4411 return;
4412 }
4413 if (rack->in_probe_rtt) {
4414 uint64_t no_overflow;
4415 uint32_t endtime, must_stay;
4416
4417 if (rack->r_ctl.rc_went_idle_time &&
4418 ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4419 /*
4420 * We went idle during prtt, just exit now.
4421 */
4422 rack_exit_probertt(rack, us_cts);
4423 } else if (rack_probe_rtt_safety_val &&
4424 TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4425 ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4426 /*
4427 * Probe RTT safety value triggered!
4428 */
4429 rack_log_rtt_shrinks(rack, us_cts,
4430 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4431 __LINE__, RACK_RTTS_SAFETY);
4432 rack_exit_probertt(rack, us_cts);
4433 }
4434 /* Calculate the max we will wait */
4435 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4436 if (rack->rc_highly_buffered)
4437 endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4438 /* Calculate the min we must wait */
4439 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4440 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4441 TSTMP_LT(us_cts, endtime)) {
4442 uint32_t calc;
4443 /* Do we lower more? */
4444 no_exit:
4445 if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4446 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4447 else
4448 calc = 0;
4449 calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4450 if (calc) {
4451 /* Maybe */
4452 calc *= rack_per_of_gp_probertt_reduce;
4453 if (calc > rack_per_of_gp_probertt)
4454 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4455 else
4456 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4457 /* Limit it too */
4458 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4459 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4460 }
4461 /* We must reach target or the time set */
4462 return;
4463 }
4464 if (rack->r_ctl.rc_time_probertt_starts == 0) {
4465 if ((TSTMP_LT(us_cts, must_stay) &&
4466 rack->rc_highly_buffered) ||
4467 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4468 rack->r_ctl.rc_target_probertt_flight)) {
4469 /* We are not past the must_stay time */
4470 goto no_exit;
4471 }
4472 rack_log_rtt_shrinks(rack, us_cts,
4473 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4474 __LINE__, RACK_RTTS_REACHTARGET);
4475 rack->r_ctl.rc_time_probertt_starts = us_cts;
4476 if (rack->r_ctl.rc_time_probertt_starts == 0)
4477 rack->r_ctl.rc_time_probertt_starts = 1;
4478 /* Restore back to our rate we want to pace at in prtt */
4479 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4480 }
4481 /*
4482 * Setup our end time, some number of gp_srtts plus 200ms.
4483 */
4484 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4485 (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4486 if (rack_probertt_gpsrtt_cnt_div)
4487 endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4488 else
4489 endtime = 0;
4490 endtime += rack_min_probertt_hold;
4491 endtime += rack->r_ctl.rc_time_probertt_starts;
4492 if (TSTMP_GEQ(us_cts, endtime)) {
4493 /* yes, exit probertt */
4494 rack_exit_probertt(rack, us_cts);
4495 }
4496
4497 } else if ((rack->rc_skip_timely == 0) &&
4498 (TSTMP_GT(us_cts, rack->r_ctl.rc_lower_rtt_us_cts)) &&
4499 ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt)) {
4500 /* Go into probertt, its been too long since we went lower */
4501 rack_enter_probertt(rack, us_cts);
4502 }
4503 }
4504
4505 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)4506 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4507 uint32_t rtt, int32_t rtt_diff)
4508 {
4509 uint64_t cur_bw, up_bnd, low_bnd, subfr;
4510 uint32_t losses;
4511
4512 if ((rack->rc_gp_dyn_mul == 0) ||
4513 (rack->use_fixed_rate) ||
4514 (rack->in_probe_rtt) ||
4515 (rack->rc_always_pace == 0)) {
4516 /* No dynamic GP multiplier in play */
4517 return;
4518 }
4519 losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4520 cur_bw = rack_get_bw(rack);
4521 /* Calculate our up and down range */
4522 up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4523 up_bnd /= 100;
4524 up_bnd += rack->r_ctl.last_gp_comp_bw;
4525
4526 subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4527 subfr /= 100;
4528 low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4529 if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4530 /*
4531 * This is the case where our RTT is above
4532 * the max target and we have been configured
4533 * to just do timely no bonus up stuff in that case.
4534 *
4535 * There are two configurations, set to 1, and we
4536 * just do timely if we are over our max. If its
4537 * set above 1 then we slam the multipliers down
4538 * to 100 and then decrement per timely.
4539 */
4540 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4541 __LINE__, 3);
4542 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4543 rack_validate_multipliers_at_or_below_100(rack);
4544 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4545 } else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4546 /*
4547 * We are decreasing this is a bit complicated this
4548 * means we are loosing ground. This could be
4549 * because another flow entered and we are competing
4550 * for b/w with it. This will push the RTT up which
4551 * makes timely unusable unless we want to get shoved
4552 * into a corner and just be backed off (the age
4553 * old problem with delay based CC).
4554 *
4555 * On the other hand if it was a route change we
4556 * would like to stay somewhat contained and not
4557 * blow out the buffers.
4558 */
4559 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4560 __LINE__, 3);
4561 rack->r_ctl.last_gp_comp_bw = cur_bw;
4562 if (rack->rc_gp_bwred == 0) {
4563 /* Go into reduction counting */
4564 rack->rc_gp_bwred = 1;
4565 rack->rc_gp_timely_dec_cnt = 0;
4566 }
4567 if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4568 /*
4569 * Push another time with a faster pacing
4570 * to try to gain back (we include override to
4571 * get a full raise factor).
4572 */
4573 if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4574 (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4575 (timely_says == 0) ||
4576 (rack_down_raise_thresh == 0)) {
4577 /*
4578 * Do an override up in b/w if we were
4579 * below the threshold or if the threshold
4580 * is zero we always do the raise.
4581 */
4582 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4583 } else {
4584 /* Log it stays the same */
4585 rack_log_timely(rack, 0, last_bw_est, low_bnd, 0,
4586 __LINE__, 11);
4587 }
4588 rack->rc_gp_timely_dec_cnt++;
4589 /* We are not incrementing really no-count */
4590 rack->rc_gp_incr = 0;
4591 rack->rc_gp_timely_inc_cnt = 0;
4592 } else {
4593 /*
4594 * Lets just use the RTT
4595 * information and give up
4596 * pushing.
4597 */
4598 goto use_timely;
4599 }
4600 } else if ((timely_says != 2) &&
4601 !losses &&
4602 (last_bw_est > up_bnd)) {
4603 /*
4604 * We are increasing b/w lets keep going, updating
4605 * our b/w and ignoring any timely input, unless
4606 * of course we are at our max raise (if there is one).
4607 */
4608
4609 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4610 __LINE__, 3);
4611 rack->r_ctl.last_gp_comp_bw = cur_bw;
4612 if (rack->rc_gp_saw_ss &&
4613 rack->r_ctl.rack_per_upper_bound_ss &&
4614 (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4615 /*
4616 * In cases where we can't go higher
4617 * we should just use timely.
4618 */
4619 goto use_timely;
4620 }
4621 if (rack->rc_gp_saw_ca &&
4622 rack->r_ctl.rack_per_upper_bound_ca &&
4623 (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4624 /*
4625 * In cases where we can't go higher
4626 * we should just use timely.
4627 */
4628 goto use_timely;
4629 }
4630 rack->rc_gp_bwred = 0;
4631 rack->rc_gp_timely_dec_cnt = 0;
4632 /* You get a set number of pushes if timely is trying to reduce */
4633 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4634 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4635 } else {
4636 /* Log it stays the same */
4637 rack_log_timely(rack, 0, last_bw_est, up_bnd, 0,
4638 __LINE__, 12);
4639 }
4640 return;
4641 } else {
4642 /*
4643 * We are staying between the lower and upper range bounds
4644 * so use timely to decide.
4645 */
4646 rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd,
4647 __LINE__, 3);
4648 use_timely:
4649 if (timely_says) {
4650 rack->rc_gp_incr = 0;
4651 rack->rc_gp_timely_inc_cnt = 0;
4652 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4653 !losses &&
4654 (last_bw_est < low_bnd)) {
4655 /* We are loosing ground */
4656 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4657 rack->rc_gp_timely_dec_cnt++;
4658 /* We are not incrementing really no-count */
4659 rack->rc_gp_incr = 0;
4660 rack->rc_gp_timely_inc_cnt = 0;
4661 } else
4662 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4663 } else {
4664 rack->rc_gp_bwred = 0;
4665 rack->rc_gp_timely_dec_cnt = 0;
4666 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4667 }
4668 }
4669 }
4670
4671 static int32_t
rack_make_timely_judgement(struct tcp_rack * rack,uint32_t rtt,int32_t rtt_diff,uint32_t prev_rtt)4672 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4673 {
4674 int32_t timely_says;
4675 uint64_t log_mult, log_rtt_a_diff;
4676
4677 log_rtt_a_diff = rtt;
4678 log_rtt_a_diff <<= 32;
4679 log_rtt_a_diff |= (uint32_t)rtt_diff;
4680 if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4681 rack_gp_rtt_maxmul)) {
4682 /* Reduce the b/w multiplier */
4683 timely_says = 2;
4684 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4685 log_mult <<= 32;
4686 log_mult |= prev_rtt;
4687 rack_log_timely(rack, timely_says, log_mult,
4688 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4689 log_rtt_a_diff, __LINE__, 4);
4690 } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4691 ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4692 max(rack_gp_rtt_mindiv , 1)))) {
4693 /* Increase the b/w multiplier */
4694 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4695 ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4696 max(rack_gp_rtt_mindiv , 1));
4697 log_mult <<= 32;
4698 log_mult |= prev_rtt;
4699 timely_says = 0;
4700 rack_log_timely(rack, timely_says, log_mult ,
4701 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4702 log_rtt_a_diff, __LINE__, 5);
4703 } else {
4704 /*
4705 * Use a gradient to find it the timely gradient
4706 * is:
4707 * grad = rc_rtt_diff / min_rtt;
4708 *
4709 * anything below or equal to 0 will be
4710 * a increase indication. Anything above
4711 * zero is a decrease. Note we take care
4712 * of the actual gradient calculation
4713 * in the reduction (its not needed for
4714 * increase).
4715 */
4716 log_mult = prev_rtt;
4717 if (rtt_diff <= 0) {
4718 /*
4719 * Rttdiff is less than zero, increase the
4720 * b/w multiplier (its 0 or negative)
4721 */
4722 timely_says = 0;
4723 rack_log_timely(rack, timely_says, log_mult,
4724 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4725 } else {
4726 /* Reduce the b/w multiplier */
4727 timely_says = 1;
4728 rack_log_timely(rack, timely_says, log_mult,
4729 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4730 }
4731 }
4732 return (timely_says);
4733 }
4734
4735 static __inline int
rack_in_gp_window(struct tcpcb * tp,struct rack_sendmap * rsm)4736 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4737 {
4738 if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4739 SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4740 /**
4741 * This covers the case that the
4742 * resent is completely inside
4743 * the gp range or up to it.
4744 * |----------------|
4745 * |-----| <or>
4746 * |----|
4747 * <or> |---|
4748 */
4749 return (1);
4750 } else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4751 SEQ_GT(rsm->r_end, tp->gput_seq)){
4752 /**
4753 * This covers the case of
4754 * |--------------|
4755 * |-------->|
4756 */
4757 return (1);
4758 } else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4759 SEQ_LT(rsm->r_start, tp->gput_ack) &&
4760 SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4761
4762 /**
4763 * This covers the case of
4764 * |--------------|
4765 * |-------->|
4766 */
4767 return (1);
4768 }
4769 return (0);
4770 }
4771
4772 static __inline void
rack_mark_in_gp_win(struct tcpcb * tp,struct rack_sendmap * rsm)4773 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4774 {
4775
4776 if ((tp->t_flags & TF_GPUTINPROG) == 0)
4777 return;
4778 /*
4779 * We have a Goodput measurement in progress. Mark
4780 * the send if its within the window. If its not
4781 * in the window make sure it does not have the mark.
4782 */
4783 if (rack_in_gp_window(tp, rsm))
4784 rsm->r_flags |= RACK_IN_GP_WIN;
4785 else
4786 rsm->r_flags &= ~RACK_IN_GP_WIN;
4787 }
4788
4789 static __inline void
rack_clear_gp_marks(struct tcpcb * tp,struct tcp_rack * rack)4790 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4791 {
4792 /* A GP measurement is ending, clear all marks on the send map*/
4793 struct rack_sendmap *rsm = NULL;
4794
4795 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4796 if (rsm == NULL) {
4797 rsm = tqhash_min(rack->r_ctl.tqh);
4798 }
4799 /* Nothing left? */
4800 while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4801 rsm->r_flags &= ~RACK_IN_GP_WIN;
4802 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4803 }
4804 }
4805
4806
4807 static __inline void
rack_tend_gp_marks(struct tcpcb * tp,struct tcp_rack * rack)4808 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4809 {
4810 struct rack_sendmap *rsm = NULL;
4811
4812 if (tp->snd_una == tp->snd_max) {
4813 /* Nothing outstanding yet, nothing to do here */
4814 return;
4815 }
4816 if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4817 /*
4818 * We are measuring ahead of some outstanding
4819 * data. We need to walk through up until we get
4820 * to gp_seq marking so that no rsm is set incorrectly
4821 * with RACK_IN_GP_WIN.
4822 */
4823 rsm = tqhash_min(rack->r_ctl.tqh);
4824 while (rsm != NULL) {
4825 rack_mark_in_gp_win(tp, rsm);
4826 if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4827 break;
4828 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4829 }
4830 }
4831 if (rsm == NULL) {
4832 /*
4833 * Need to find the GP seq, if rsm is
4834 * set we stopped as we hit it.
4835 */
4836 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4837 if (rsm == NULL)
4838 return;
4839 rack_mark_in_gp_win(tp, rsm);
4840 }
4841 /*
4842 * Now we may need to mark already sent rsm, ahead of
4843 * gput_seq in the window since they may have been sent
4844 * *before* we started our measurment. The rsm, if non-null
4845 * has been marked (note if rsm would have been NULL we would have
4846 * returned in the previous block). So we go to the next, and continue
4847 * until we run out of entries or we exceed the gp_ack value.
4848 */
4849 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4850 while (rsm) {
4851 rack_mark_in_gp_win(tp, rsm);
4852 if (SEQ_GT(rsm->r_end, tp->gput_ack))
4853 break;
4854 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4855 }
4856 }
4857
4858 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)4859 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)
4860 {
4861 if (tcp_bblogging_on(rack->rc_tp)) {
4862 union tcp_log_stackspecific log;
4863 struct timeval tv;
4864
4865 memset(&log, 0, sizeof(log));
4866 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4867 log.u_bbr.flex1 = add_part;
4868 log.u_bbr.flex2 = sub_part;
4869 log.u_bbr.flex3 = rack_wma_divisor;
4870 log.u_bbr.flex4 = srtt;
4871 log.u_bbr.flex7 = (uint16_t)line;
4872 log.u_bbr.flex8 = meth;
4873 log.u_bbr.delRate = rack->r_ctl.gp_bw;
4874 log.u_bbr.cur_del_rate = meas_bw;
4875 log.u_bbr.rttProp = utim;
4876 TCP_LOG_EVENTP(rack->rc_tp, NULL,
4877 &rack->rc_inp->inp_socket->so_rcv,
4878 &rack->rc_inp->inp_socket->so_snd,
4879 BBR_LOG_THRESH_CALC, 0,
4880 0, &log, false, &rack->r_ctl.act_rcv_time);
4881 }
4882 }
4883
4884 static void
rack_do_goodput_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq th_ack,int line,uint8_t quality)4885 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4886 tcp_seq th_ack, int line, uint8_t quality)
4887 {
4888 uint64_t tim, bytes_ps, stim, utim;
4889 uint32_t segsiz, bytes, reqbytes, us_cts;
4890 int32_t gput, new_rtt_diff, timely_says;
4891 uint64_t resid_bw, subpart = 0, addpart = 0, srtt;
4892 int did_add = 0;
4893
4894 us_cts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
4895 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4896 if (TSTMP_GEQ(us_cts, tp->gput_ts))
4897 tim = us_cts - tp->gput_ts;
4898 else
4899 tim = 0;
4900 if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4901 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4902 else
4903 stim = 0;
4904 /*
4905 * Use the larger of the send time or ack time. This prevents us
4906 * from being influenced by ack artifacts to come up with too
4907 * high of measurement. Note that since we are spanning over many more
4908 * bytes in most of our measurements hopefully that is less likely to
4909 * occur.
4910 */
4911 if (tim > stim)
4912 utim = max(tim, 1);
4913 else
4914 utim = max(stim, 1);
4915 reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4916 rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4917 if ((tim == 0) && (stim == 0)) {
4918 /*
4919 * Invalid measurement time, maybe
4920 * all on one ack/one send?
4921 */
4922 bytes = 0;
4923 bytes_ps = 0;
4924 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4925 0, 0, 0, 10, __LINE__, NULL, quality);
4926 goto skip_measurement;
4927 }
4928 if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4929 /* We never made a us_rtt measurement? */
4930 bytes = 0;
4931 bytes_ps = 0;
4932 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4933 0, 0, 0, 10, __LINE__, NULL, quality);
4934 goto skip_measurement;
4935 }
4936 /*
4937 * Calculate the maximum possible b/w this connection
4938 * could have. We base our calculation on the lowest
4939 * rtt we have seen during the measurement and the
4940 * largest rwnd the client has given us in that time. This
4941 * forms a BDP that is the maximum that we could ever
4942 * get to the client. Anything larger is not valid.
4943 *
4944 * I originally had code here that rejected measurements
4945 * where the time was less than 1/2 the latest us_rtt.
4946 * But after thinking on that I realized its wrong since
4947 * say you had a 150Mbps or even 1Gbps link, and you
4948 * were a long way away.. example I am in Europe (100ms rtt)
4949 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4950 * bytes my time would be 1.2ms, and yet my rtt would say
4951 * the measurement was invalid the time was < 50ms. The
4952 * same thing is true for 150Mb (8ms of time).
4953 *
4954 * A better way I realized is to look at what the maximum
4955 * the connection could possibly do. This is gated on
4956 * the lowest RTT we have seen and the highest rwnd.
4957 * We should in theory never exceed that, if we are
4958 * then something on the path is storing up packets
4959 * and then feeding them all at once to our endpoint
4960 * messing up our measurement.
4961 */
4962 rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4963 rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4964 rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4965 if (SEQ_LT(th_ack, tp->gput_seq)) {
4966 /* No measurement can be made */
4967 bytes = 0;
4968 bytes_ps = 0;
4969 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4970 0, 0, 0, 10, __LINE__, NULL, quality);
4971 goto skip_measurement;
4972 } else
4973 bytes = (th_ack - tp->gput_seq);
4974 bytes_ps = (uint64_t)bytes;
4975 /*
4976 * Don't measure a b/w for pacing unless we have gotten at least
4977 * an initial windows worth of data in this measurement interval.
4978 *
4979 * Small numbers of bytes get badly influenced by delayed ack and
4980 * other artifacts. Note we take the initial window or our
4981 * defined minimum GP (defaulting to 10 which hopefully is the
4982 * IW).
4983 */
4984 if (rack->rc_gp_filled == 0) {
4985 /*
4986 * The initial estimate is special. We
4987 * have blasted out an IW worth of packets
4988 * without a real valid ack ts results. We
4989 * then setup the app_limited_needs_set flag,
4990 * this should get the first ack in (probably 2
4991 * MSS worth) to be recorded as the timestamp.
4992 * We thus allow a smaller number of bytes i.e.
4993 * IW - 2MSS.
4994 */
4995 reqbytes -= (2 * segsiz);
4996 /* Also lets fill previous for our first measurement to be neutral */
4997 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4998 }
4999 if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
5000 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5001 rack->r_ctl.rc_app_limited_cnt,
5002 0, 0, 10, __LINE__, NULL, quality);
5003 goto skip_measurement;
5004 }
5005 /*
5006 * We now need to calculate the Timely like status so
5007 * we can update (possibly) the b/w multipliers.
5008 */
5009 new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
5010 if (rack->rc_gp_filled == 0) {
5011 /* No previous reading */
5012 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
5013 } else {
5014 if (rack->measure_saw_probe_rtt == 0) {
5015 /*
5016 * We don't want a probertt to be counted
5017 * since it will be negative incorrectly. We
5018 * expect to be reducing the RTT when we
5019 * pace at a slower rate.
5020 */
5021 rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
5022 rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
5023 }
5024 }
5025 timely_says = rack_make_timely_judgement(rack,
5026 rack->r_ctl.rc_gp_srtt,
5027 rack->r_ctl.rc_rtt_diff,
5028 rack->r_ctl.rc_prev_gp_srtt
5029 );
5030 bytes_ps *= HPTS_USEC_IN_SEC;
5031 bytes_ps /= utim;
5032 if (bytes_ps > rack->r_ctl.last_max_bw) {
5033 /*
5034 * Something is on path playing
5035 * since this b/w is not possible based
5036 * on our BDP (highest rwnd and lowest rtt
5037 * we saw in the measurement window).
5038 *
5039 * Another option here would be to
5040 * instead skip the measurement.
5041 */
5042 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
5043 bytes_ps, rack->r_ctl.last_max_bw, 0,
5044 11, __LINE__, NULL, quality);
5045 bytes_ps = rack->r_ctl.last_max_bw;
5046 }
5047 /* We store gp for b/w in bytes per second */
5048 if (rack->rc_gp_filled == 0) {
5049 /* Initial measurement */
5050 if (bytes_ps) {
5051 rack->r_ctl.gp_bw = bytes_ps;
5052 rack->rc_gp_filled = 1;
5053 rack->r_ctl.num_measurements = 1;
5054 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5055 } else {
5056 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5057 rack->r_ctl.rc_app_limited_cnt,
5058 0, 0, 10, __LINE__, NULL, quality);
5059 }
5060 if (tcp_in_hpts(rack->rc_tp) &&
5061 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5062 /*
5063 * Ok we can't trust the pacer in this case
5064 * where we transition from un-paced to paced.
5065 * Or for that matter when the burst mitigation
5066 * was making a wild guess and got it wrong.
5067 * Stop the pacer and clear up all the aggregate
5068 * delays etc.
5069 */
5070 tcp_hpts_remove(rack->rc_tp);
5071 rack->r_ctl.rc_hpts_flags = 0;
5072 rack->r_ctl.rc_last_output_to = 0;
5073 }
5074 did_add = 2;
5075 } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5076 /* Still a small number run an average */
5077 rack->r_ctl.gp_bw += bytes_ps;
5078 addpart = rack->r_ctl.num_measurements;
5079 rack->r_ctl.num_measurements++;
5080 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5081 /* We have collected enough to move forward */
5082 rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5083 }
5084 rack_set_pace_segments(tp, rack, __LINE__, NULL);
5085 did_add = 3;
5086 } else {
5087 /*
5088 * We want to take 1/wma of the goodput and add in to 7/8th
5089 * of the old value weighted by the srtt. So if your measurement
5090 * period is say 2 SRTT's long you would get 1/4 as the
5091 * value, if it was like 1/2 SRTT then you would get 1/16th.
5092 *
5093 * But we must be careful not to take too much i.e. if the
5094 * srtt is say 20ms and the measurement is taken over
5095 * 400ms our weight would be 400/20 i.e. 20. On the
5096 * other hand if we get a measurement over 1ms with a
5097 * 10ms rtt we only want to take a much smaller portion.
5098 */
5099 uint8_t meth;
5100
5101 if (rack->r_ctl.num_measurements < 0xff) {
5102 rack->r_ctl.num_measurements++;
5103 }
5104 srtt = (uint64_t)tp->t_srtt;
5105 if (srtt == 0) {
5106 /*
5107 * Strange why did t_srtt go back to zero?
5108 */
5109 if (rack->r_ctl.rc_rack_min_rtt)
5110 srtt = rack->r_ctl.rc_rack_min_rtt;
5111 else
5112 srtt = HPTS_USEC_IN_MSEC;
5113 }
5114 /*
5115 * XXXrrs: Note for reviewers, in playing with
5116 * dynamic pacing I discovered this GP calculation
5117 * as done originally leads to some undesired results.
5118 * Basically you can get longer measurements contributing
5119 * too much to the WMA. Thus I changed it if you are doing
5120 * dynamic adjustments to only do the aportioned adjustment
5121 * if we have a very small (time wise) measurement. Longer
5122 * measurements just get there weight (defaulting to 1/8)
5123 * add to the WMA. We may want to think about changing
5124 * this to always do that for both sides i.e. dynamic
5125 * and non-dynamic... but considering lots of folks
5126 * were playing with this I did not want to change the
5127 * calculation per.se. without your thoughts.. Lawerence?
5128 * Peter??
5129 */
5130 if (rack->rc_gp_dyn_mul == 0) {
5131 subpart = rack->r_ctl.gp_bw * utim;
5132 subpart /= (srtt * 8);
5133 if (subpart < (rack->r_ctl.gp_bw / 2)) {
5134 /*
5135 * The b/w update takes no more
5136 * away then 1/2 our running total
5137 * so factor it in.
5138 */
5139 addpart = bytes_ps * utim;
5140 addpart /= (srtt * 8);
5141 meth = 1;
5142 } else {
5143 /*
5144 * Don't allow a single measurement
5145 * to account for more than 1/2 of the
5146 * WMA. This could happen on a retransmission
5147 * where utim becomes huge compared to
5148 * srtt (multiple retransmissions when using
5149 * the sending rate which factors in all the
5150 * transmissions from the first one).
5151 */
5152 subpart = rack->r_ctl.gp_bw / 2;
5153 addpart = bytes_ps / 2;
5154 meth = 2;
5155 }
5156 rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5157 resid_bw = rack->r_ctl.gp_bw - subpart;
5158 rack->r_ctl.gp_bw = resid_bw + addpart;
5159 did_add = 1;
5160 } else {
5161 if ((utim / srtt) <= 1) {
5162 /*
5163 * The b/w update was over a small period
5164 * of time. The idea here is to prevent a small
5165 * measurement time period from counting
5166 * too much. So we scale it based on the
5167 * time so it attributes less than 1/rack_wma_divisor
5168 * of its measurement.
5169 */
5170 subpart = rack->r_ctl.gp_bw * utim;
5171 subpart /= (srtt * rack_wma_divisor);
5172 addpart = bytes_ps * utim;
5173 addpart /= (srtt * rack_wma_divisor);
5174 meth = 3;
5175 } else {
5176 /*
5177 * The scaled measurement was long
5178 * enough so lets just add in the
5179 * portion of the measurement i.e. 1/rack_wma_divisor
5180 */
5181 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5182 addpart = bytes_ps / rack_wma_divisor;
5183 meth = 4;
5184 }
5185 if ((rack->measure_saw_probe_rtt == 0) ||
5186 (bytes_ps > rack->r_ctl.gp_bw)) {
5187 /*
5188 * For probe-rtt we only add it in
5189 * if its larger, all others we just
5190 * add in.
5191 */
5192 did_add = 1;
5193 rack_log_gp_calc(rack, addpart, subpart, srtt, bytes_ps, utim, meth, __LINE__);
5194 resid_bw = rack->r_ctl.gp_bw - subpart;
5195 rack->r_ctl.gp_bw = resid_bw + addpart;
5196 }
5197 }
5198 rack_set_pace_segments(tp, rack, __LINE__, NULL);
5199 }
5200 /*
5201 * We only watch the growth of the GP during the initial startup
5202 * or first-slowstart that ensues. If we ever needed to watch
5203 * growth of gp outside of that period all we need to do is
5204 * remove the first clause of this if (rc_initial_ss_comp).
5205 */
5206 if ((rack->rc_initial_ss_comp == 0) &&
5207 (rack->r_ctl.num_measurements >= RACK_REQ_AVG)) {
5208 uint64_t gp_est;
5209
5210 gp_est = bytes_ps;
5211 if (tcp_bblogging_on(rack->rc_tp)) {
5212 union tcp_log_stackspecific log;
5213 struct timeval tv;
5214
5215 memset(&log, 0, sizeof(log));
5216 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5217 log.u_bbr.flex1 = rack->r_ctl.current_round;
5218 log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
5219 log.u_bbr.delRate = gp_est;
5220 log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5221 log.u_bbr.flex8 = 41;
5222 (void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5223 0, &log, false, NULL, __func__, __LINE__,&tv);
5224 }
5225 if ((rack->r_ctl.num_measurements == RACK_REQ_AVG) ||
5226 (rack->r_ctl.last_gpest == 0)) {
5227 /*
5228 * The round we get our measurement averaging going
5229 * is the base round so it always is the source point
5230 * for when we had our first increment. From there on
5231 * we only record the round that had a rise.
5232 */
5233 rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5234 rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5235 } else if (gp_est >= rack->r_ctl.last_gpest) {
5236 /*
5237 * Test to see if its gone up enough
5238 * to set the round count up to now. Note
5239 * that on the seeding of the 4th measurement we
5240 */
5241 gp_est *= 1000;
5242 gp_est /= rack->r_ctl.last_gpest;
5243 if ((uint32_t)gp_est > rack->r_ctl.gp_gain_req) {
5244 /*
5245 * We went up enough to record the round.
5246 */
5247 if (tcp_bblogging_on(rack->rc_tp)) {
5248 union tcp_log_stackspecific log;
5249 struct timeval tv;
5250
5251 memset(&log, 0, sizeof(log));
5252 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5253 log.u_bbr.flex1 = rack->r_ctl.current_round;
5254 log.u_bbr.flex2 = (uint32_t)gp_est;
5255 log.u_bbr.flex3 = rack->r_ctl.gp_gain_req;
5256 log.u_bbr.delRate = gp_est;
5257 log.u_bbr.cur_del_rate = rack->r_ctl.last_gpest;
5258 log.u_bbr.flex8 = 42;
5259 (void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5260 0, &log, false, NULL, __func__, __LINE__,&tv);
5261 }
5262 rack->r_ctl.last_rnd_of_gp_rise = rack->r_ctl.current_round;
5263 if (rack->r_ctl.use_gp_not_last == 1)
5264 rack->r_ctl.last_gpest = rack->r_ctl.gp_bw;
5265 else
5266 rack->r_ctl.last_gpest = bytes_ps;
5267 }
5268 }
5269 }
5270 if ((rack->gp_ready == 0) &&
5271 (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5272 /* We have enough measurements now */
5273 rack->gp_ready = 1;
5274 if (rack->dgp_on ||
5275 rack->rack_hibeta)
5276 rack_set_cc_pacing(rack);
5277 if (rack->defer_options)
5278 rack_apply_deferred_options(rack);
5279 }
5280 rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5281 rack_get_bw(rack), 22, did_add, NULL, quality);
5282 /* We do not update any multipliers if we are in or have seen a probe-rtt */
5283
5284 if ((rack->measure_saw_probe_rtt == 0) &&
5285 rack->rc_gp_rtt_set) {
5286 if (rack->rc_skip_timely == 0) {
5287 rack_update_multiplier(rack, timely_says, bytes_ps,
5288 rack->r_ctl.rc_gp_srtt,
5289 rack->r_ctl.rc_rtt_diff);
5290 }
5291 }
5292 rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5293 rack_get_bw(rack), 3, line, NULL, quality);
5294 rack_log_pacing_delay_calc(rack,
5295 bytes, /* flex2 */
5296 tim, /* flex1 */
5297 bytes_ps, /* bw_inuse */
5298 rack->r_ctl.gp_bw, /* delRate */
5299 rack_get_lt_bw(rack), /* rttProp */
5300 20, line, NULL, 0);
5301 /* reset the gp srtt and setup the new prev */
5302 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5303 /* Record the lost count for the next measurement */
5304 rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5305 skip_measurement:
5306 /*
5307 * We restart our diffs based on the gpsrtt in the
5308 * measurement window.
5309 */
5310 rack->rc_gp_rtt_set = 0;
5311 rack->rc_gp_saw_rec = 0;
5312 rack->rc_gp_saw_ca = 0;
5313 rack->rc_gp_saw_ss = 0;
5314 rack->rc_dragged_bottom = 0;
5315 if (quality == RACK_QUALITY_HIGH) {
5316 /*
5317 * Gput in the stats world is in kbps where bytes_ps is
5318 * bytes per second so we do ((x * 8)/ 1000).
5319 */
5320 gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5321 #ifdef STATS
5322 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5323 gput);
5324 /*
5325 * XXXLAS: This is a temporary hack, and should be
5326 * chained off VOI_TCP_GPUT when stats(9) grows an
5327 * API to deal with chained VOIs.
5328 */
5329 if (tp->t_stats_gput_prev > 0)
5330 stats_voi_update_abs_s32(tp->t_stats,
5331 VOI_TCP_GPUT_ND,
5332 ((gput - tp->t_stats_gput_prev) * 100) /
5333 tp->t_stats_gput_prev);
5334 #endif
5335 tp->t_stats_gput_prev = gput;
5336 }
5337 tp->t_flags &= ~TF_GPUTINPROG;
5338 /*
5339 * Now are we app limited now and there is space from where we
5340 * were to where we want to go?
5341 *
5342 * We don't do the other case i.e. non-applimited here since
5343 * the next send will trigger us picking up the missing data.
5344 */
5345 if (rack->r_ctl.rc_first_appl &&
5346 TCPS_HAVEESTABLISHED(tp->t_state) &&
5347 rack->r_ctl.rc_app_limited_cnt &&
5348 (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5349 ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5350 max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5351 /*
5352 * Yep there is enough outstanding to make a measurement here.
5353 */
5354 struct rack_sendmap *rsm;
5355
5356 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5357 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5358 tp->gput_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
5359 rack->app_limited_needs_set = 0;
5360 tp->gput_seq = th_ack;
5361 if (rack->in_probe_rtt)
5362 rack->measure_saw_probe_rtt = 1;
5363 else if ((rack->measure_saw_probe_rtt) &&
5364 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5365 rack->measure_saw_probe_rtt = 0;
5366 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5367 /* There is a full window to gain info from */
5368 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5369 } else {
5370 /* We can only measure up to the applimited point */
5371 tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5372 if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5373 /*
5374 * We don't have enough to make a measurement.
5375 */
5376 tp->t_flags &= ~TF_GPUTINPROG;
5377 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5378 0, 0, 0, 6, __LINE__, NULL, quality);
5379 return;
5380 }
5381 }
5382 if (tp->t_state >= TCPS_FIN_WAIT_1) {
5383 /*
5384 * We will get no more data into the SB
5385 * this means we need to have the data available
5386 * before we start a measurement.
5387 */
5388 if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5389 /* Nope not enough data. */
5390 return;
5391 }
5392 }
5393 tp->t_flags |= TF_GPUTINPROG;
5394 /*
5395 * Now we need to find the timestamp of the send at tp->gput_seq
5396 * for the send based measurement.
5397 */
5398 rack->r_ctl.rc_gp_cumack_ts = 0;
5399 rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5400 if (rsm) {
5401 /* Ok send-based limit is set */
5402 if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5403 /*
5404 * Move back to include the earlier part
5405 * so our ack time lines up right (this may
5406 * make an overlapping measurement but thats
5407 * ok).
5408 */
5409 tp->gput_seq = rsm->r_start;
5410 }
5411 if (rsm->r_flags & RACK_ACKED) {
5412 struct rack_sendmap *nrsm;
5413
5414 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5415 tp->gput_seq = rsm->r_end;
5416 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5417 if (nrsm)
5418 rsm = nrsm;
5419 else {
5420 rack->app_limited_needs_set = 1;
5421 }
5422 } else
5423 rack->app_limited_needs_set = 1;
5424 /* We always go from the first send */
5425 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5426 } else {
5427 /*
5428 * If we don't find the rsm due to some
5429 * send-limit set the current time, which
5430 * basically disables the send-limit.
5431 */
5432 struct timeval tv;
5433
5434 microuptime(&tv);
5435 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5436 }
5437 rack_tend_gp_marks(tp, rack);
5438 rack_log_pacing_delay_calc(rack,
5439 tp->gput_seq,
5440 tp->gput_ack,
5441 (uintptr_t)rsm,
5442 tp->gput_ts,
5443 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5444 9,
5445 __LINE__, rsm, quality);
5446 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5447 } else {
5448 /*
5449 * To make sure proper timestamp merging occurs, we need to clear
5450 * all GP marks if we don't start a measurement.
5451 */
5452 rack_clear_gp_marks(tp, rack);
5453 }
5454 }
5455
5456 /*
5457 * CC wrapper hook functions
5458 */
5459 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)5460 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5461 uint16_t type, int32_t post_recovery)
5462 {
5463 uint32_t prior_cwnd, acked;
5464 struct tcp_log_buffer *lgb = NULL;
5465 uint8_t labc_to_use, quality;
5466
5467 INP_WLOCK_ASSERT(tptoinpcb(tp));
5468 tp->t_ccv.nsegs = nsegs;
5469 acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5470 if ((post_recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5471 uint32_t max;
5472
5473 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5474 if (tp->t_ccv.bytes_this_ack > max) {
5475 tp->t_ccv.bytes_this_ack = max;
5476 }
5477 }
5478 #ifdef STATS
5479 stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5480 ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5481 #endif
5482 if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5483 /*
5484 * We will ack all the data, time to end any
5485 * lt_bw_up we have running until something
5486 * new is sent. Note we need to use the actual
5487 * ack_rcv_time which with pacing may be different.
5488 */
5489 uint64_t tmark;
5490
5491 rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5492 rack->r_ctl.lt_seq = tp->snd_max;
5493 tmark = tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time);
5494 if (tmark >= rack->r_ctl.lt_timemark) {
5495 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
5496 }
5497 rack->r_ctl.lt_timemark = tmark;
5498 rack->lt_bw_up = 0;
5499 }
5500 quality = RACK_QUALITY_NONE;
5501 if ((tp->t_flags & TF_GPUTINPROG) &&
5502 rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5503 /* Measure the Goodput */
5504 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5505 }
5506 /* Which way our we limited, if not cwnd limited no advance in CA */
5507 if (tp->snd_cwnd <= tp->snd_wnd)
5508 tp->t_ccv.flags |= CCF_CWND_LIMITED;
5509 else
5510 tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5511 if (tp->snd_cwnd > tp->snd_ssthresh) {
5512 tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5513 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5514 /* For the setting of a window past use the actual scwnd we are using */
5515 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5516 tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5517 tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5518 }
5519 } else {
5520 tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5521 tp->t_bytes_acked = 0;
5522 }
5523 prior_cwnd = tp->snd_cwnd;
5524 if ((post_recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5525 (rack_client_low_buf && rack->client_bufferlvl &&
5526 (rack->client_bufferlvl < rack_client_low_buf)))
5527 labc_to_use = rack->rc_labc;
5528 else
5529 labc_to_use = rack_max_abc_post_recovery;
5530 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5531 union tcp_log_stackspecific log;
5532 struct timeval tv;
5533
5534 memset(&log, 0, sizeof(log));
5535 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5536 log.u_bbr.flex1 = th_ack;
5537 log.u_bbr.flex2 = tp->t_ccv.flags;
5538 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5539 log.u_bbr.flex4 = tp->t_ccv.nsegs;
5540 log.u_bbr.flex5 = labc_to_use;
5541 log.u_bbr.flex6 = prior_cwnd;
5542 log.u_bbr.flex7 = V_tcp_do_newsack;
5543 log.u_bbr.flex8 = 1;
5544 lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5545 0, &log, false, NULL, __func__, __LINE__,&tv);
5546 }
5547 if (CC_ALGO(tp)->ack_received != NULL) {
5548 /* XXXLAS: Find a way to live without this */
5549 tp->t_ccv.curack = th_ack;
5550 tp->t_ccv.labc = labc_to_use;
5551 tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5552 CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5553 }
5554 if (lgb) {
5555 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5556 }
5557 if (rack->r_must_retran) {
5558 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5559 /*
5560 * We now are beyond the rxt point so lets disable
5561 * the flag.
5562 */
5563 rack->r_ctl.rc_out_at_rto = 0;
5564 rack->r_must_retran = 0;
5565 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5566 /*
5567 * Only decrement the rc_out_at_rto if the cwnd advances
5568 * at least a whole segment. Otherwise next time the peer
5569 * acks, we won't be able to send this generaly happens
5570 * when we are in Congestion Avoidance.
5571 */
5572 if (acked <= rack->r_ctl.rc_out_at_rto){
5573 rack->r_ctl.rc_out_at_rto -= acked;
5574 } else {
5575 rack->r_ctl.rc_out_at_rto = 0;
5576 }
5577 }
5578 }
5579 #ifdef STATS
5580 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5581 #endif
5582 if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5583 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5584 }
5585 if ((rack->rc_initial_ss_comp == 0) &&
5586 (tp->snd_cwnd >= tp->snd_ssthresh)) {
5587 /*
5588 * The cwnd has grown beyond ssthresh we have
5589 * entered ca and completed our first Slowstart.
5590 */
5591 rack->rc_initial_ss_comp = 1;
5592 }
5593 }
5594
5595 static void
tcp_rack_partialack(struct tcpcb * tp)5596 tcp_rack_partialack(struct tcpcb *tp)
5597 {
5598 struct tcp_rack *rack;
5599
5600 rack = (struct tcp_rack *)tp->t_fb_ptr;
5601 INP_WLOCK_ASSERT(tptoinpcb(tp));
5602 /*
5603 * If we are doing PRR and have enough
5604 * room to send <or> we are pacing and prr
5605 * is disabled we will want to see if we
5606 * can send data (by setting r_wanted_output to
5607 * true).
5608 */
5609 if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5610 rack->rack_no_prr)
5611 rack->r_wanted_output = 1;
5612 }
5613
5614 static void
rack_exit_recovery(struct tcpcb * tp,struct tcp_rack * rack,int how)5615 rack_exit_recovery(struct tcpcb *tp, struct tcp_rack *rack, int how)
5616 {
5617 /*
5618 * Now exit recovery.
5619 */
5620 EXIT_RECOVERY(tp->t_flags);
5621 }
5622
5623 static void
rack_post_recovery(struct tcpcb * tp,uint32_t th_ack)5624 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5625 {
5626 struct tcp_rack *rack;
5627 uint32_t orig_cwnd;
5628
5629 orig_cwnd = tp->snd_cwnd;
5630 INP_WLOCK_ASSERT(tptoinpcb(tp));
5631 rack = (struct tcp_rack *)tp->t_fb_ptr;
5632 /* only alert CC if we alerted when we entered */
5633 if (CC_ALGO(tp)->post_recovery != NULL) {
5634 tp->t_ccv.curack = th_ack;
5635 CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5636 if (tp->snd_cwnd < tp->snd_ssthresh) {
5637 /*
5638 * Rack has burst control and pacing
5639 * so lets not set this any lower than
5640 * snd_ssthresh per RFC-6582 (option 2).
5641 */
5642 tp->snd_cwnd = tp->snd_ssthresh;
5643 }
5644 }
5645 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5646 union tcp_log_stackspecific log;
5647 struct timeval tv;
5648
5649 memset(&log, 0, sizeof(log));
5650 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5651 log.u_bbr.flex1 = th_ack;
5652 log.u_bbr.flex2 = tp->t_ccv.flags;
5653 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5654 log.u_bbr.flex4 = tp->t_ccv.nsegs;
5655 log.u_bbr.flex5 = V_tcp_abc_l_var;
5656 log.u_bbr.flex6 = orig_cwnd;
5657 log.u_bbr.flex7 = V_tcp_do_newsack;
5658 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5659 log.u_bbr.flex8 = 2;
5660 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5661 0, &log, false, NULL, __func__, __LINE__, &tv);
5662 }
5663 if ((rack->rack_no_prr == 0) &&
5664 (rack->no_prr_addback == 0) &&
5665 (rack->r_ctl.rc_prr_sndcnt > 0)) {
5666 /*
5667 * Suck the next prr cnt back into cwnd, but
5668 * only do that if we are not application limited.
5669 */
5670 if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5671 /*
5672 * We are allowed to add back to the cwnd the amount we did
5673 * not get out if:
5674 * a) no_prr_addback is off.
5675 * b) we are not app limited
5676 * c) we are doing prr
5677 * <and>
5678 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5679 */
5680 tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5681 rack->r_ctl.rc_prr_sndcnt);
5682 }
5683 rack->r_ctl.rc_prr_sndcnt = 0;
5684 rack_log_to_prr(rack, 1, 0, __LINE__);
5685 }
5686 rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5687 tp->snd_recover = tp->snd_una;
5688 if (rack->r_ctl.dsack_persist) {
5689 rack->r_ctl.dsack_persist--;
5690 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5691 rack->r_ctl.num_dsack = 0;
5692 }
5693 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5694 }
5695 if (rack->rto_from_rec == 1) {
5696 rack->rto_from_rec = 0;
5697 if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
5698 tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
5699 }
5700 rack_exit_recovery(tp, rack, 1);
5701 }
5702
5703 static void
rack_cong_signal(struct tcpcb * tp,uint32_t type,uint32_t ack,int line)5704 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5705 {
5706 struct tcp_rack *rack;
5707 uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5708
5709 INP_WLOCK_ASSERT(tptoinpcb(tp));
5710 #ifdef STATS
5711 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5712 #endif
5713 if (IN_RECOVERY(tp->t_flags) == 0) {
5714 in_rec_at_entry = 0;
5715 ssthresh_enter = tp->snd_ssthresh;
5716 cwnd_enter = tp->snd_cwnd;
5717 } else
5718 in_rec_at_entry = 1;
5719 rack = (struct tcp_rack *)tp->t_fb_ptr;
5720 switch (type) {
5721 case CC_NDUPACK:
5722 tp->t_flags &= ~TF_WASFRECOVERY;
5723 tp->t_flags &= ~TF_WASCRECOVERY;
5724 if (!IN_FASTRECOVERY(tp->t_flags)) {
5725 /* Check if this is the end of the initial Start-up i.e. initial slow-start */
5726 if (rack->rc_initial_ss_comp == 0) {
5727 /* Yep it is the end of the initial slowstart */
5728 rack->rc_initial_ss_comp = 1;
5729 }
5730 rack->r_ctl.rc_prr_delivered = 0;
5731 rack->r_ctl.rc_prr_out = 0;
5732 rack->r_fast_output = 0;
5733 if (rack->rack_no_prr == 0) {
5734 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5735 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5736 }
5737 rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5738 tp->snd_recover = tp->snd_max;
5739 if (tp->t_flags2 & TF2_ECN_PERMIT)
5740 tp->t_flags2 |= TF2_ECN_SND_CWR;
5741 }
5742 break;
5743 case CC_ECN:
5744 if (!IN_CONGRECOVERY(tp->t_flags) ||
5745 /*
5746 * Allow ECN reaction on ACK to CWR, if
5747 * that data segment was also CE marked.
5748 */
5749 SEQ_GEQ(ack, tp->snd_recover)) {
5750 EXIT_CONGRECOVERY(tp->t_flags);
5751 KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5752 rack->r_fast_output = 0;
5753 tp->snd_recover = tp->snd_max + 1;
5754 if (tp->t_flags2 & TF2_ECN_PERMIT)
5755 tp->t_flags2 |= TF2_ECN_SND_CWR;
5756 }
5757 break;
5758 case CC_RTO:
5759 tp->t_dupacks = 0;
5760 tp->t_bytes_acked = 0;
5761 rack->r_fast_output = 0;
5762 if (IN_RECOVERY(tp->t_flags))
5763 rack_exit_recovery(tp, rack, 2);
5764 orig_cwnd = tp->snd_cwnd;
5765 rack_log_to_prr(rack, 16, orig_cwnd, line);
5766 if (CC_ALGO(tp)->cong_signal == NULL) {
5767 /* TSNH */
5768 tp->snd_ssthresh = max(2,
5769 min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5770 ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5771 tp->snd_cwnd = ctf_fixed_maxseg(tp);
5772 }
5773 if (tp->t_flags2 & TF2_ECN_PERMIT)
5774 tp->t_flags2 |= TF2_ECN_SND_CWR;
5775 break;
5776 case CC_RTO_ERR:
5777 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5778 /* RTO was unnecessary, so reset everything. */
5779 tp->snd_cwnd = tp->snd_cwnd_prev;
5780 tp->snd_ssthresh = tp->snd_ssthresh_prev;
5781 tp->snd_recover = tp->snd_recover_prev;
5782 if (tp->t_flags & TF_WASFRECOVERY) {
5783 ENTER_FASTRECOVERY(tp->t_flags);
5784 tp->t_flags &= ~TF_WASFRECOVERY;
5785 }
5786 if (tp->t_flags & TF_WASCRECOVERY) {
5787 ENTER_CONGRECOVERY(tp->t_flags);
5788 tp->t_flags &= ~TF_WASCRECOVERY;
5789 }
5790 tp->snd_nxt = tp->snd_max;
5791 tp->t_badrxtwin = 0;
5792 break;
5793 }
5794 if ((CC_ALGO(tp)->cong_signal != NULL) &&
5795 (type != CC_RTO)){
5796 tp->t_ccv.curack = ack;
5797 CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5798 }
5799 if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5800 rack_log_to_prr(rack, 15, cwnd_enter, line);
5801 rack->r_ctl.dsack_byte_cnt = 0;
5802 rack->r_ctl.retran_during_recovery = 0;
5803 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5804 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5805 rack->r_ent_rec_ns = 1;
5806 }
5807 }
5808
5809 static inline void
rack_cc_after_idle(struct tcp_rack * rack,struct tcpcb * tp)5810 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5811 {
5812 uint32_t i_cwnd;
5813
5814 INP_WLOCK_ASSERT(tptoinpcb(tp));
5815
5816 if (CC_ALGO(tp)->after_idle != NULL)
5817 CC_ALGO(tp)->after_idle(&tp->t_ccv);
5818
5819 if (tp->snd_cwnd == 1)
5820 i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */
5821 else
5822 i_cwnd = rc_init_window(rack);
5823
5824 /*
5825 * Being idle is no different than the initial window. If the cc
5826 * clamps it down below the initial window raise it to the initial
5827 * window.
5828 */
5829 if (tp->snd_cwnd < i_cwnd) {
5830 tp->snd_cwnd = i_cwnd;
5831 }
5832 }
5833
5834 /*
5835 * Indicate whether this ack should be delayed. We can delay the ack if
5836 * following conditions are met:
5837 * - There is no delayed ack timer in progress.
5838 * - Our last ack wasn't a 0-sized window. We never want to delay
5839 * the ack that opens up a 0-sized window.
5840 * - LRO wasn't used for this segment. We make sure by checking that the
5841 * segment size is not larger than the MSS.
5842 * - Delayed acks are enabled or this is a half-synchronized T/TCP
5843 * connection.
5844 */
5845 #define DELAY_ACK(tp, tlen) \
5846 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5847 ((tp->t_flags & TF_DELACK) == 0) && \
5848 (tlen <= tp->t_maxseg) && \
5849 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5850
5851 static struct rack_sendmap *
rack_find_lowest_rsm(struct tcp_rack * rack)5852 rack_find_lowest_rsm(struct tcp_rack *rack)
5853 {
5854 struct rack_sendmap *rsm;
5855
5856 /*
5857 * Walk the time-order transmitted list looking for an rsm that is
5858 * not acked. This will be the one that was sent the longest time
5859 * ago that is still outstanding.
5860 */
5861 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5862 if (rsm->r_flags & RACK_ACKED) {
5863 continue;
5864 }
5865 goto finish;
5866 }
5867 finish:
5868 return (rsm);
5869 }
5870
5871 static struct rack_sendmap *
rack_find_high_nonack(struct tcp_rack * rack,struct rack_sendmap * rsm)5872 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5873 {
5874 struct rack_sendmap *prsm;
5875
5876 /*
5877 * Walk the sequence order list backward until we hit and arrive at
5878 * the highest seq not acked. In theory when this is called it
5879 * should be the last segment (which it was not).
5880 */
5881 prsm = rsm;
5882
5883 TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
5884 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5885 continue;
5886 }
5887 return (prsm);
5888 }
5889 return (NULL);
5890 }
5891
5892 static uint32_t
rack_calc_thresh_rack(struct tcp_rack * rack,uint32_t srtt,uint32_t cts,int line,int log_allowed)5893 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts, int line, int log_allowed)
5894 {
5895 int32_t lro;
5896 uint32_t thresh;
5897
5898 /*
5899 * lro is the flag we use to determine if we have seen reordering.
5900 * If it gets set we have seen reordering. The reorder logic either
5901 * works in one of two ways:
5902 *
5903 * If reorder-fade is configured, then we track the last time we saw
5904 * re-ordering occur. If we reach the point where enough time as
5905 * passed we no longer consider reordering as occurring.
5906 *
5907 * Or if reorder-face is 0, then once we see reordering we consider
5908 * the connection to alway be subject to reordering and just set lro
5909 * to 1.
5910 *
5911 * In the end if lro is non-zero we add the extra time for
5912 * reordering in.
5913 */
5914 if (srtt == 0)
5915 srtt = 1;
5916 if (rack->r_ctl.rc_reorder_ts) {
5917 if (rack->r_ctl.rc_reorder_fade) {
5918 if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5919 lro = cts - rack->r_ctl.rc_reorder_ts;
5920 if (lro == 0) {
5921 /*
5922 * No time as passed since the last
5923 * reorder, mark it as reordering.
5924 */
5925 lro = 1;
5926 }
5927 } else {
5928 /* Negative time? */
5929 lro = 0;
5930 }
5931 if (lro > rack->r_ctl.rc_reorder_fade) {
5932 /* Turn off reordering seen too */
5933 rack->r_ctl.rc_reorder_ts = 0;
5934 lro = 0;
5935 }
5936 } else {
5937 /* Reodering does not fade */
5938 lro = 1;
5939 }
5940 } else {
5941 lro = 0;
5942 }
5943 if (rack->rc_rack_tmr_std_based == 0) {
5944 thresh = srtt + rack->r_ctl.rc_pkt_delay;
5945 } else {
5946 /* Standards based pkt-delay is 1/4 srtt */
5947 thresh = srtt + (srtt >> 2);
5948 }
5949 if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5950 /* It must be set, if not you get 1/4 rtt */
5951 if (rack->r_ctl.rc_reorder_shift)
5952 thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5953 else
5954 thresh += (srtt >> 2);
5955 }
5956 if (rack->rc_rack_use_dsack &&
5957 lro &&
5958 (rack->r_ctl.num_dsack > 0)) {
5959 /*
5960 * We only increase the reordering window if we
5961 * have seen reordering <and> we have a DSACK count.
5962 */
5963 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5964 if (log_allowed)
5965 rack_log_dsack_event(rack, 4, line, srtt, thresh);
5966 }
5967 /* SRTT * 2 is the ceiling */
5968 if (thresh > (srtt * 2)) {
5969 thresh = srtt * 2;
5970 }
5971 /* And we don't want it above the RTO max either */
5972 if (thresh > rack_rto_max) {
5973 thresh = rack_rto_max;
5974 }
5975 if (log_allowed)
5976 rack_log_dsack_event(rack, 6, line, srtt, thresh);
5977 return (thresh);
5978 }
5979
5980 static uint32_t
rack_calc_thresh_tlp(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t srtt)5981 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5982 struct rack_sendmap *rsm, uint32_t srtt)
5983 {
5984 struct rack_sendmap *prsm;
5985 uint32_t thresh, len;
5986 int segsiz;
5987
5988 if (srtt == 0)
5989 srtt = 1;
5990 if (rack->r_ctl.rc_tlp_threshold)
5991 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5992 else
5993 thresh = (srtt * 2);
5994
5995 /* Get the previous sent packet, if any */
5996 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5997 len = rsm->r_end - rsm->r_start;
5998 if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5999 /* Exactly like the ID */
6000 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6001 uint32_t alt_thresh;
6002 /*
6003 * Compensate for delayed-ack with the d-ack time.
6004 */
6005 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6006 if (alt_thresh > thresh)
6007 thresh = alt_thresh;
6008 }
6009 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6010 /* 2.1 behavior */
6011 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6012 if (prsm && (len <= segsiz)) {
6013 /*
6014 * Two packets outstanding, thresh should be (2*srtt) +
6015 * possible inter-packet delay (if any).
6016 */
6017 uint32_t inter_gap = 0;
6018 int idx, nidx;
6019
6020 idx = rsm->r_rtr_cnt - 1;
6021 nidx = prsm->r_rtr_cnt - 1;
6022 if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6023 /* Yes it was sent later (or at the same time) */
6024 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6025 }
6026 thresh += inter_gap;
6027 } else if (len <= segsiz) {
6028 /*
6029 * Possibly compensate for delayed-ack.
6030 */
6031 uint32_t alt_thresh;
6032
6033 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6034 if (alt_thresh > thresh)
6035 thresh = alt_thresh;
6036 }
6037 } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6038 /* 2.2 behavior */
6039 if (len <= segsiz) {
6040 uint32_t alt_thresh;
6041 /*
6042 * Compensate for delayed-ack with the d-ack time.
6043 */
6044 alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6045 if (alt_thresh > thresh)
6046 thresh = alt_thresh;
6047 }
6048 }
6049 /* Not above an RTO */
6050 if (thresh > tp->t_rxtcur) {
6051 thresh = tp->t_rxtcur;
6052 }
6053 /* Not above a RTO max */
6054 if (thresh > rack_rto_max) {
6055 thresh = rack_rto_max;
6056 }
6057 /* Apply user supplied min TLP */
6058 if (thresh < rack_tlp_min) {
6059 thresh = rack_tlp_min;
6060 }
6061 return (thresh);
6062 }
6063
6064 static uint32_t
rack_grab_rtt(struct tcpcb * tp,struct tcp_rack * rack)6065 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6066 {
6067 /*
6068 * We want the rack_rtt which is the
6069 * last rtt we measured. However if that
6070 * does not exist we fallback to the srtt (which
6071 * we probably will never do) and then as a last
6072 * resort we use RACK_INITIAL_RTO if no srtt is
6073 * yet set.
6074 */
6075 if (rack->rc_rack_rtt)
6076 return (rack->rc_rack_rtt);
6077 else if (tp->t_srtt == 0)
6078 return (RACK_INITIAL_RTO);
6079 return (tp->t_srtt);
6080 }
6081
6082 static struct rack_sendmap *
rack_check_recovery_mode(struct tcpcb * tp,uint32_t tsused)6083 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6084 {
6085 /*
6086 * Check to see that we don't need to fall into recovery. We will
6087 * need to do so if our oldest transmit is past the time we should
6088 * have had an ack.
6089 */
6090 struct tcp_rack *rack;
6091 struct rack_sendmap *rsm;
6092 int32_t idx;
6093 uint32_t srtt, thresh;
6094
6095 rack = (struct tcp_rack *)tp->t_fb_ptr;
6096 if (tqhash_empty(rack->r_ctl.tqh)) {
6097 return (NULL);
6098 }
6099 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6100 if (rsm == NULL)
6101 return (NULL);
6102
6103
6104 if (rsm->r_flags & RACK_ACKED) {
6105 rsm = rack_find_lowest_rsm(rack);
6106 if (rsm == NULL)
6107 return (NULL);
6108 }
6109 idx = rsm->r_rtr_cnt - 1;
6110 srtt = rack_grab_rtt(tp, rack);
6111 thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
6112 if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6113 return (NULL);
6114 }
6115 if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6116 return (NULL);
6117 }
6118 /* Ok if we reach here we are over-due and this guy can be sent */
6119 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6120 return (rsm);
6121 }
6122
6123 static uint32_t
rack_get_persists_timer_val(struct tcpcb * tp,struct tcp_rack * rack)6124 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6125 {
6126 int32_t t;
6127 int32_t tt;
6128 uint32_t ret_val;
6129
6130 t = (tp->t_srtt + (tp->t_rttvar << 2));
6131 RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6132 rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6133 rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6134 ret_val = (uint32_t)tt;
6135 return (ret_val);
6136 }
6137
6138 static uint32_t
rack_timer_start(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,int sup_rack)6139 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6140 {
6141 /*
6142 * Start the FR timer, we do this based on getting the first one in
6143 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6144 * events we need to stop the running timer (if its running) before
6145 * starting the new one.
6146 */
6147 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6148 uint32_t srtt_cur;
6149 int32_t idx;
6150 int32_t is_tlp_timer = 0;
6151 struct rack_sendmap *rsm;
6152
6153 if (rack->t_timers_stopped) {
6154 /* All timers have been stopped none are to run */
6155 return (0);
6156 }
6157 if (rack->rc_in_persist) {
6158 /* We can't start any timer in persists */
6159 return (rack_get_persists_timer_val(tp, rack));
6160 }
6161 rack->rc_on_min_to = 0;
6162 if ((tp->t_state < TCPS_ESTABLISHED) ||
6163 ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6164 goto activate_rxt;
6165 }
6166 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6167 if ((rsm == NULL) || sup_rack) {
6168 /* Nothing on the send map or no rack */
6169 activate_rxt:
6170 time_since_sent = 0;
6171 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6172 if (rsm) {
6173 /*
6174 * Should we discount the RTX timer any?
6175 *
6176 * We want to discount it the smallest amount.
6177 * If a timer (Rack/TLP or RXT) has gone off more
6178 * recently thats the discount we want to use (now - timer time).
6179 * If the retransmit of the oldest packet was more recent then
6180 * we want to use that (now - oldest-packet-last_transmit_time).
6181 *
6182 */
6183 idx = rsm->r_rtr_cnt - 1;
6184 if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6185 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6186 else
6187 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6188 if (TSTMP_GT(cts, tstmp_touse))
6189 time_since_sent = cts - tstmp_touse;
6190 }
6191 if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6192 sbavail(&tptosocket(tp)->so_snd)) {
6193 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6194 to = tp->t_rxtcur;
6195 if (to > time_since_sent)
6196 to -= time_since_sent;
6197 else
6198 to = rack->r_ctl.rc_min_to;
6199 if (to == 0)
6200 to = 1;
6201 /* Special case for KEEPINIT */
6202 if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6203 (TP_KEEPINIT(tp) != 0) &&
6204 rsm) {
6205 /*
6206 * We have to put a ceiling on the rxt timer
6207 * of the keep-init timeout.
6208 */
6209 uint32_t max_time, red;
6210
6211 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6212 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6213 red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6214 if (red < max_time)
6215 max_time -= red;
6216 else
6217 max_time = 1;
6218 }
6219 /* Reduce timeout to the keep value if needed */
6220 if (max_time < to)
6221 to = max_time;
6222 }
6223 return (to);
6224 }
6225 return (0);
6226 }
6227 if (rsm->r_flags & RACK_ACKED) {
6228 rsm = rack_find_lowest_rsm(rack);
6229 if (rsm == NULL) {
6230 /* No lowest? */
6231 goto activate_rxt;
6232 }
6233 }
6234 /* Convert from ms to usecs */
6235 if ((rsm->r_flags & RACK_SACK_PASSED) ||
6236 (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6237 (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6238 if ((tp->t_flags & TF_SENTFIN) &&
6239 ((tp->snd_max - tp->snd_una) == 1) &&
6240 (rsm->r_flags & RACK_HAS_FIN)) {
6241 /*
6242 * We don't start a rack timer if all we have is a
6243 * FIN outstanding.
6244 */
6245 goto activate_rxt;
6246 }
6247 if ((rack->use_rack_rr == 0) &&
6248 (IN_FASTRECOVERY(tp->t_flags)) &&
6249 (rack->rack_no_prr == 0) &&
6250 (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) {
6251 /*
6252 * We are not cheating, in recovery and
6253 * not enough ack's to yet get our next
6254 * retransmission out.
6255 *
6256 * Note that classified attackers do not
6257 * get to use the rack-cheat.
6258 */
6259 goto activate_tlp;
6260 }
6261 srtt = rack_grab_rtt(tp, rack);
6262 thresh = rack_calc_thresh_rack(rack, srtt, cts, __LINE__, 1);
6263 idx = rsm->r_rtr_cnt - 1;
6264 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6265 if (SEQ_GEQ(exp, cts)) {
6266 to = exp - cts;
6267 if (to < rack->r_ctl.rc_min_to) {
6268 to = rack->r_ctl.rc_min_to;
6269 if (rack->r_rr_config == 3)
6270 rack->rc_on_min_to = 1;
6271 }
6272 } else {
6273 to = rack->r_ctl.rc_min_to;
6274 if (rack->r_rr_config == 3)
6275 rack->rc_on_min_to = 1;
6276 }
6277 } else {
6278 /* Ok we need to do a TLP not RACK */
6279 activate_tlp:
6280 if ((rack->rc_tlp_in_progress != 0) &&
6281 (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6282 /*
6283 * The previous send was a TLP and we have sent
6284 * N TLP's without sending new data.
6285 */
6286 goto activate_rxt;
6287 }
6288 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6289 if (rsm == NULL) {
6290 /* We found no rsm to TLP with. */
6291 goto activate_rxt;
6292 }
6293 if (rsm->r_flags & RACK_HAS_FIN) {
6294 /* If its a FIN we dont do TLP */
6295 rsm = NULL;
6296 goto activate_rxt;
6297 }
6298 idx = rsm->r_rtr_cnt - 1;
6299 time_since_sent = 0;
6300 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6301 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6302 else
6303 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6304 if (TSTMP_GT(cts, tstmp_touse))
6305 time_since_sent = cts - tstmp_touse;
6306 is_tlp_timer = 1;
6307 if (tp->t_srtt) {
6308 if ((rack->rc_srtt_measure_made == 0) &&
6309 (tp->t_srtt == 1)) {
6310 /*
6311 * If another stack as run and set srtt to 1,
6312 * then the srtt was 0, so lets use the initial.
6313 */
6314 srtt = RACK_INITIAL_RTO;
6315 } else {
6316 srtt_cur = tp->t_srtt;
6317 srtt = srtt_cur;
6318 }
6319 } else
6320 srtt = RACK_INITIAL_RTO;
6321 /*
6322 * If the SRTT is not keeping up and the
6323 * rack RTT has spiked we want to use
6324 * the last RTT not the smoothed one.
6325 */
6326 if (rack_tlp_use_greater &&
6327 tp->t_srtt &&
6328 (srtt < rack_grab_rtt(tp, rack))) {
6329 srtt = rack_grab_rtt(tp, rack);
6330 }
6331 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6332 if (thresh > time_since_sent) {
6333 to = thresh - time_since_sent;
6334 } else {
6335 to = rack->r_ctl.rc_min_to;
6336 rack_log_alt_to_to_cancel(rack,
6337 thresh, /* flex1 */
6338 time_since_sent, /* flex2 */
6339 tstmp_touse, /* flex3 */
6340 rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6341 (uint32_t)rsm->r_tim_lastsent[idx],
6342 srtt,
6343 idx, 99);
6344 }
6345 if (to < rack_tlp_min) {
6346 to = rack_tlp_min;
6347 }
6348 if (to > TICKS_2_USEC(tcp_rexmit_max)) {
6349 /*
6350 * If the TLP time works out to larger than the max
6351 * RTO lets not do TLP.. just RTO.
6352 */
6353 goto activate_rxt;
6354 }
6355 }
6356 if (is_tlp_timer == 0) {
6357 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6358 } else {
6359 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6360 }
6361 if (to == 0)
6362 to = 1;
6363 return (to);
6364 }
6365
6366 static void
rack_enter_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,tcp_seq snd_una)6367 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6368 {
6369 if (rack->rc_in_persist == 0) {
6370 if (tp->t_flags & TF_GPUTINPROG) {
6371 /*
6372 * Stop the goodput now, the calling of the
6373 * measurement function clears the flag.
6374 */
6375 rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6376 RACK_QUALITY_PERSIST);
6377 }
6378 #ifdef NETFLIX_SHARED_CWND
6379 if (rack->r_ctl.rc_scw) {
6380 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6381 rack->rack_scwnd_is_idle = 1;
6382 }
6383 #endif
6384 rack->r_ctl.rc_went_idle_time = cts;
6385 if (rack->r_ctl.rc_went_idle_time == 0)
6386 rack->r_ctl.rc_went_idle_time = 1;
6387 if (rack->lt_bw_up) {
6388 /* Suspend our LT BW measurement */
6389 uint64_t tmark;
6390
6391 rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6392 rack->r_ctl.lt_seq = snd_una;
6393 tmark = tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time);
6394 if (tmark >= rack->r_ctl.lt_timemark) {
6395 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6396 }
6397 rack->r_ctl.lt_timemark = tmark;
6398 rack->lt_bw_up = 0;
6399 rack->r_persist_lt_bw_off = 1;
6400 }
6401 rack_timer_cancel(tp, rack, cts, __LINE__);
6402 rack->r_ctl.persist_lost_ends = 0;
6403 rack->probe_not_answered = 0;
6404 rack->forced_ack = 0;
6405 tp->t_rxtshift = 0;
6406 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6407 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6408 rack->rc_in_persist = 1;
6409 }
6410 }
6411
6412 static void
rack_exit_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)6413 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6414 {
6415 if (tcp_in_hpts(rack->rc_tp)) {
6416 tcp_hpts_remove(rack->rc_tp);
6417 rack->r_ctl.rc_hpts_flags = 0;
6418 }
6419 #ifdef NETFLIX_SHARED_CWND
6420 if (rack->r_ctl.rc_scw) {
6421 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6422 rack->rack_scwnd_is_idle = 0;
6423 }
6424 #endif
6425 if (rack->rc_gp_dyn_mul &&
6426 (rack->use_fixed_rate == 0) &&
6427 (rack->rc_always_pace)) {
6428 /*
6429 * Do we count this as if a probe-rtt just
6430 * finished?
6431 */
6432 uint32_t time_idle, idle_min;
6433
6434 time_idle = cts - rack->r_ctl.rc_went_idle_time;
6435 idle_min = rack_min_probertt_hold;
6436 if (rack_probertt_gpsrtt_cnt_div) {
6437 uint64_t extra;
6438 extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6439 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
6440 extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6441 idle_min += (uint32_t)extra;
6442 }
6443 if (time_idle >= idle_min) {
6444 /* Yes, we count it as a probe-rtt. */
6445 uint32_t us_cts;
6446
6447 us_cts = tcp_get_usecs(NULL);
6448 if (rack->in_probe_rtt == 0) {
6449 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6450 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6451 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6452 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6453 } else {
6454 rack_exit_probertt(rack, us_cts);
6455 }
6456 }
6457 }
6458 if (rack->r_persist_lt_bw_off) {
6459 /* Continue where we left off */
6460 rack->r_ctl.lt_timemark = tcp_get_u64_usecs(NULL);
6461 rack->lt_bw_up = 1;
6462 rack->r_persist_lt_bw_off = 0;
6463 }
6464 rack->rc_in_persist = 0;
6465 rack->r_ctl.rc_went_idle_time = 0;
6466 tp->t_rxtshift = 0;
6467 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6468 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6469 rack->r_ctl.rc_agg_delayed = 0;
6470 rack->r_early = 0;
6471 rack->r_late = 0;
6472 rack->r_ctl.rc_agg_early = 0;
6473 }
6474
6475 static void
rack_log_hpts_diag(struct tcp_rack * rack,uint32_t cts,struct hpts_diag * diag,struct timeval * tv)6476 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6477 struct hpts_diag *diag, struct timeval *tv)
6478 {
6479 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6480 union tcp_log_stackspecific log;
6481
6482 memset(&log, 0, sizeof(log));
6483 log.u_bbr.flex1 = diag->p_nxt_slot;
6484 log.u_bbr.flex2 = diag->p_cur_slot;
6485 log.u_bbr.flex3 = diag->slot_req;
6486 log.u_bbr.flex4 = diag->inp_hptsslot;
6487 log.u_bbr.flex5 = diag->slot_remaining;
6488 log.u_bbr.flex6 = diag->need_new_to;
6489 log.u_bbr.flex7 = diag->p_hpts_active;
6490 log.u_bbr.flex8 = diag->p_on_min_sleep;
6491 /* Hijack other fields as needed */
6492 log.u_bbr.epoch = diag->have_slept;
6493 log.u_bbr.lt_epoch = diag->yet_to_sleep;
6494 log.u_bbr.pkts_out = diag->co_ret;
6495 log.u_bbr.applimited = diag->hpts_sleep_time;
6496 log.u_bbr.delivered = diag->p_prev_slot;
6497 log.u_bbr.inflight = diag->p_runningslot;
6498 log.u_bbr.bw_inuse = diag->wheel_slot;
6499 log.u_bbr.rttProp = diag->wheel_cts;
6500 log.u_bbr.timeStamp = cts;
6501 log.u_bbr.delRate = diag->maxslots;
6502 log.u_bbr.cur_del_rate = diag->p_curtick;
6503 log.u_bbr.cur_del_rate <<= 32;
6504 log.u_bbr.cur_del_rate |= diag->p_lasttick;
6505 TCP_LOG_EVENTP(rack->rc_tp, NULL,
6506 &rack->rc_inp->inp_socket->so_rcv,
6507 &rack->rc_inp->inp_socket->so_snd,
6508 BBR_LOG_HPTSDIAG, 0,
6509 0, &log, false, tv);
6510 }
6511
6512 }
6513
6514 static void
rack_log_wakeup(struct tcpcb * tp,struct tcp_rack * rack,struct sockbuf * sb,uint32_t len,int type)6515 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6516 {
6517 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6518 union tcp_log_stackspecific log;
6519 struct timeval tv;
6520
6521 memset(&log, 0, sizeof(log));
6522 log.u_bbr.flex1 = sb->sb_flags;
6523 log.u_bbr.flex2 = len;
6524 log.u_bbr.flex3 = sb->sb_state;
6525 log.u_bbr.flex8 = type;
6526 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6527 TCP_LOG_EVENTP(rack->rc_tp, NULL,
6528 &rack->rc_inp->inp_socket->so_rcv,
6529 &rack->rc_inp->inp_socket->so_snd,
6530 TCP_LOG_SB_WAKE, 0,
6531 len, &log, false, &tv);
6532 }
6533 }
6534
6535 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)6536 rack_start_hpts_timer (struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6537 int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6538 {
6539 struct hpts_diag diag;
6540 struct inpcb *inp = tptoinpcb(tp);
6541 struct timeval tv;
6542 uint32_t delayed_ack = 0;
6543 uint32_t hpts_timeout;
6544 uint32_t entry_slot = slot;
6545 uint8_t stopped;
6546 uint32_t left = 0;
6547 uint32_t us_cts;
6548
6549 if ((tp->t_state == TCPS_CLOSED) ||
6550 (tp->t_state == TCPS_LISTEN)) {
6551 return;
6552 }
6553 if (tcp_in_hpts(tp)) {
6554 /* Already on the pacer */
6555 return;
6556 }
6557 stopped = rack->rc_tmr_stopped;
6558 if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6559 left = rack->r_ctl.rc_timer_exp - cts;
6560 }
6561 rack->r_ctl.rc_timer_exp = 0;
6562 rack->r_ctl.rc_hpts_flags = 0;
6563 us_cts = tcp_get_usecs(&tv);
6564 /* Now early/late accounting */
6565 rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6566 if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6567 /*
6568 * We have a early carry over set,
6569 * we can always add more time so we
6570 * can always make this compensation.
6571 *
6572 * Note if ack's are allowed to wake us do not
6573 * penalize the next timer for being awoke
6574 * by an ack aka the rc_agg_early (non-paced mode).
6575 */
6576 slot += rack->r_ctl.rc_agg_early;
6577 rack->r_early = 0;
6578 rack->r_ctl.rc_agg_early = 0;
6579 }
6580 if ((rack->r_late) &&
6581 ((rack->r_use_hpts_min == 0) || (rack->dgp_on == 0))) {
6582 /*
6583 * This is harder, we can
6584 * compensate some but it
6585 * really depends on what
6586 * the current pacing time is.
6587 */
6588 if (rack->r_ctl.rc_agg_delayed >= slot) {
6589 /*
6590 * We can't compensate for it all.
6591 * And we have to have some time
6592 * on the clock. We always have a min
6593 * 10 slots (10 x 10 i.e. 100 usecs).
6594 */
6595 if (slot <= HPTS_USECS_PER_SLOT) {
6596 /* We gain delay */
6597 rack->r_ctl.rc_agg_delayed += (HPTS_USECS_PER_SLOT - slot);
6598 slot = HPTS_USECS_PER_SLOT;
6599 } else {
6600 /* We take off some */
6601 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_USECS_PER_SLOT);
6602 slot = HPTS_USECS_PER_SLOT;
6603 }
6604 } else {
6605 slot -= rack->r_ctl.rc_agg_delayed;
6606 rack->r_ctl.rc_agg_delayed = 0;
6607 /* Make sure we have 100 useconds at minimum */
6608 if (slot < HPTS_USECS_PER_SLOT) {
6609 rack->r_ctl.rc_agg_delayed = HPTS_USECS_PER_SLOT - slot;
6610 slot = HPTS_USECS_PER_SLOT;
6611 }
6612 if (rack->r_ctl.rc_agg_delayed == 0)
6613 rack->r_late = 0;
6614 }
6615 } else if (rack->r_late) {
6616 /* r_use_hpts_min is on and so is DGP */
6617 uint32_t max_red;
6618
6619 max_red = (slot * rack->r_ctl.max_reduction) / 100;
6620 if (max_red >= rack->r_ctl.rc_agg_delayed) {
6621 slot -= rack->r_ctl.rc_agg_delayed;
6622 rack->r_ctl.rc_agg_delayed = 0;
6623 } else {
6624 slot -= max_red;
6625 rack->r_ctl.rc_agg_delayed -= max_red;
6626 }
6627 }
6628 if ((rack->r_use_hpts_min == 1) &&
6629 (slot > 0) &&
6630 (rack->dgp_on == 1)) {
6631 /*
6632 * We are enforcing a min pacing timer
6633 * based on our hpts min timeout.
6634 */
6635 uint32_t min;
6636
6637 min = get_hpts_min_sleep_time();
6638 if (min > slot) {
6639 slot = min;
6640 }
6641 }
6642 hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6643 if (tp->t_flags & TF_DELACK) {
6644 delayed_ack = TICKS_2_USEC(tcp_delacktime);
6645 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6646 }
6647 if (delayed_ack && ((hpts_timeout == 0) ||
6648 (delayed_ack < hpts_timeout)))
6649 hpts_timeout = delayed_ack;
6650 else
6651 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6652 /*
6653 * If no timers are going to run and we will fall off the hptsi
6654 * wheel, we resort to a keep-alive timer if its configured.
6655 */
6656 if ((hpts_timeout == 0) &&
6657 (slot == 0)) {
6658 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6659 (tp->t_state <= TCPS_CLOSING)) {
6660 /*
6661 * Ok we have no timer (persists, rack, tlp, rxt or
6662 * del-ack), we don't have segments being paced. So
6663 * all that is left is the keepalive timer.
6664 */
6665 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6666 /* Get the established keep-alive time */
6667 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6668 } else {
6669 /*
6670 * Get the initial setup keep-alive time,
6671 * note that this is probably not going to
6672 * happen, since rack will be running a rxt timer
6673 * if a SYN of some sort is outstanding. It is
6674 * actually handled in rack_timeout_rxt().
6675 */
6676 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6677 }
6678 rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6679 if (rack->in_probe_rtt) {
6680 /*
6681 * We want to instead not wake up a long time from
6682 * now but to wake up about the time we would
6683 * exit probe-rtt and initiate a keep-alive ack.
6684 * This will get us out of probe-rtt and update
6685 * our min-rtt.
6686 */
6687 hpts_timeout = rack_min_probertt_hold;
6688 }
6689 }
6690 }
6691 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6692 (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6693 /*
6694 * RACK, TLP, persists and RXT timers all are restartable
6695 * based on actions input .. i.e we received a packet (ack
6696 * or sack) and that changes things (rw, or snd_una etc).
6697 * Thus we can restart them with a new value. For
6698 * keep-alive, delayed_ack we keep track of what was left
6699 * and restart the timer with a smaller value.
6700 */
6701 if (left < hpts_timeout)
6702 hpts_timeout = left;
6703 }
6704 if (hpts_timeout) {
6705 /*
6706 * Hack alert for now we can't time-out over 2,147,483
6707 * seconds (a bit more than 596 hours), which is probably ok
6708 * :).
6709 */
6710 if (hpts_timeout > 0x7ffffffe)
6711 hpts_timeout = 0x7ffffffe;
6712 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6713 }
6714 rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6715 if ((rack->gp_ready == 0) &&
6716 (rack->use_fixed_rate == 0) &&
6717 (hpts_timeout < slot) &&
6718 (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6719 /*
6720 * We have no good estimate yet for the
6721 * old clunky burst mitigation or the
6722 * real pacing. And the tlp or rxt is smaller
6723 * than the pacing calculation. Lets not
6724 * pace that long since we know the calculation
6725 * so far is not accurate.
6726 */
6727 slot = hpts_timeout;
6728 }
6729 /**
6730 * Turn off all the flags for queuing by default. The
6731 * flags have important meanings to what happens when
6732 * LRO interacts with the transport. Most likely (by default now)
6733 * mbuf_queueing and ack compression are on. So the transport
6734 * has a couple of flags that control what happens (if those
6735 * are not on then these flags won't have any effect since it
6736 * won't go through the queuing LRO path).
6737 *
6738 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6739 * pacing output, so don't disturb. But
6740 * it also means LRO can wake me if there
6741 * is a SACK arrival.
6742 *
6743 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6744 * with the above flag (QUEUE_READY) and
6745 * when present it says don't even wake me
6746 * if a SACK arrives.
6747 *
6748 * The idea behind these flags is that if we are pacing we
6749 * set the MBUF_QUEUE_READY and only get woken up if
6750 * a SACK arrives (which could change things) or if
6751 * our pacing timer expires. If, however, we have a rack
6752 * timer running, then we don't even want a sack to wake
6753 * us since the rack timer has to expire before we can send.
6754 *
6755 * Other cases should usually have none of the flags set
6756 * so LRO can call into us.
6757 */
6758 tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6759 if (slot) {
6760 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6761 rack->r_ctl.rc_last_output_to = us_cts + slot;
6762 /*
6763 * A pacing timer (slot) is being set, in
6764 * such a case we cannot send (we are blocked by
6765 * the timer). So lets tell LRO that it should not
6766 * wake us unless there is a SACK. Note this only
6767 * will be effective if mbuf queueing is on or
6768 * compressed acks are being processed.
6769 */
6770 tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6771 /*
6772 * But wait if we have a Rack timer running
6773 * even a SACK should not disturb us (with
6774 * the exception of r_rr_config 3).
6775 */
6776 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6777 (IN_RECOVERY(tp->t_flags))) {
6778 if (rack->r_rr_config != 3)
6779 tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6780 else if (rack->rc_pace_dnd) {
6781 /*
6782 * When DND is on, we only let a sack
6783 * interrupt us if we are not in recovery.
6784 *
6785 * If DND is off, then we never hit here
6786 * and let all sacks wake us up.
6787 *
6788 */
6789 tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6790 }
6791 }
6792 if (rack->rc_ack_can_sendout_data) {
6793 /*
6794 * Ahh but wait, this is that special case
6795 * where the pacing timer can be disturbed
6796 * backout the changes (used for non-paced
6797 * burst limiting).
6798 */
6799 tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6800 TF2_MBUF_QUEUE_READY);
6801 }
6802 if ((rack->use_rack_rr) &&
6803 (rack->r_rr_config < 2) &&
6804 ((hpts_timeout) && (hpts_timeout < slot))) {
6805 /*
6806 * Arrange for the hpts to kick back in after the
6807 * t-o if the t-o does not cause a send.
6808 */
6809 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6810 __LINE__, &diag);
6811 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6812 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6813 } else {
6814 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
6815 __LINE__, &diag);
6816 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6817 rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6818 }
6819 } else if (hpts_timeout) {
6820 /*
6821 * With respect to t_flags2(?) here, lets let any new acks wake
6822 * us up here. Since we are not pacing (no pacing timer), output
6823 * can happen so we should let it. If its a Rack timer, then any inbound
6824 * packet probably won't change the sending (we will be blocked)
6825 * but it may change the prr stats so letting it in (the set defaults
6826 * at the start of this block) are good enough.
6827 */
6828 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6829 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
6830 __LINE__, &diag);
6831 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6832 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6833 } else {
6834 /* No timer starting */
6835 #ifdef INVARIANTS
6836 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6837 panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6838 tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6839 }
6840 #endif
6841 }
6842 rack->rc_tmr_stopped = 0;
6843 if (slot)
6844 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6845 }
6846
6847 static void
rack_mark_lost(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t cts)6848 rack_mark_lost(struct tcpcb *tp,
6849 struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
6850 {
6851 struct rack_sendmap *nrsm;
6852 uint32_t thresh, exp;
6853
6854 thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
6855 nrsm = rsm;
6856 TAILQ_FOREACH_FROM(nrsm, &rack->r_ctl.rc_tmap, r_tnext) {
6857 if ((nrsm->r_flags & RACK_SACK_PASSED) == 0) {
6858 /* Got up to all that were marked sack-passed */
6859 break;
6860 }
6861 if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
6862 exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
6863 if (TSTMP_LT(exp, cts) || (exp == cts)) {
6864 /* We now consider it lost */
6865 nrsm->r_flags |= RACK_WAS_LOST;
6866 rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
6867 } else {
6868 /* Past here it won't be lost so stop */
6869 break;
6870 }
6871 }
6872 }
6873 }
6874
6875 /*
6876 * RACK Timer, here we simply do logging and house keeping.
6877 * the normal rack_output() function will call the
6878 * appropriate thing to check if we need to do a RACK retransmit.
6879 * We return 1, saying don't proceed with rack_output only
6880 * when all timers have been stopped (destroyed PCB?).
6881 */
6882 static int
rack_timeout_rack(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)6883 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6884 {
6885 /*
6886 * This timer simply provides an internal trigger to send out data.
6887 * The check_recovery_mode call will see if there are needed
6888 * retransmissions, if so we will enter fast-recovery. The output
6889 * call may or may not do the same thing depending on sysctl
6890 * settings.
6891 */
6892 struct rack_sendmap *rsm;
6893
6894 counter_u64_add(rack_to_tot, 1);
6895 if (rack->r_state && (rack->r_state != tp->t_state))
6896 rack_set_state(tp, rack);
6897 rack->rc_on_min_to = 0;
6898 rsm = rack_check_recovery_mode(tp, cts);
6899 rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6900 if (rsm) {
6901 /* We need to stroke any lost that are now declared as lost */
6902 rack_mark_lost(tp, rack, rsm, cts);
6903 rack->r_ctl.rc_resend = rsm;
6904 rack->r_timer_override = 1;
6905 if (rack->use_rack_rr) {
6906 /*
6907 * Don't accumulate extra pacing delay
6908 * we are allowing the rack timer to
6909 * over-ride pacing i.e. rrr takes precedence
6910 * if the pacing interval is longer than the rrr
6911 * time (in other words we get the min pacing
6912 * time versus rrr pacing time).
6913 */
6914 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6915 }
6916 }
6917 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6918 if (rsm == NULL) {
6919 /* restart a timer and return 1 */
6920 rack_start_hpts_timer(rack, tp, cts,
6921 0, 0, 0);
6922 return (1);
6923 }
6924 return (0);
6925 }
6926
6927
6928
6929 static void
rack_adjust_orig_mlen(struct rack_sendmap * rsm)6930 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6931 {
6932
6933 if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
6934 /*
6935 * The trailing space changed, mbufs can grow
6936 * at the tail but they can't shrink from
6937 * it, KASSERT that. Adjust the orig_m_len to
6938 * compensate for this change.
6939 */
6940 KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
6941 ("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
6942 rsm->m,
6943 rsm,
6944 (intmax_t)M_TRAILINGROOM(rsm->m),
6945 rsm->orig_t_space,
6946 rsm->orig_m_len,
6947 rsm->m->m_len));
6948 rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
6949 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
6950 }
6951 if (rsm->m->m_len < rsm->orig_m_len) {
6952 /*
6953 * Mbuf shrank, trimmed off the top by an ack, our
6954 * offset changes.
6955 */
6956 KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
6957 ("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
6958 rsm->m, rsm->m->m_len,
6959 rsm, rsm->orig_m_len,
6960 rsm->soff));
6961 if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
6962 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6963 else
6964 rsm->soff = 0;
6965 rsm->orig_m_len = rsm->m->m_len;
6966 #ifdef INVARIANTS
6967 } else if (rsm->m->m_len > rsm->orig_m_len) {
6968 panic("rsm:%p m:%p m_len grew outside of t_space compensation",
6969 rsm, rsm->m);
6970 #endif
6971 }
6972 }
6973
6974 static void
rack_setup_offset_for_rsm(struct tcp_rack * rack,struct rack_sendmap * src_rsm,struct rack_sendmap * rsm)6975 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6976 {
6977 struct mbuf *m;
6978 uint32_t soff;
6979
6980 if (src_rsm->m &&
6981 ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
6982 (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
6983 /* Fix up the orig_m_len and possibly the mbuf offset */
6984 rack_adjust_orig_mlen(src_rsm);
6985 }
6986 m = src_rsm->m;
6987 soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6988 while (soff >= m->m_len) {
6989 /* Move out past this mbuf */
6990 soff -= m->m_len;
6991 m = m->m_next;
6992 KASSERT((m != NULL),
6993 ("rsm:%p nrsm:%p hit at soff:%u null m",
6994 src_rsm, rsm, soff));
6995 if (m == NULL) {
6996 /* This should *not* happen which is why there is a kassert */
6997 src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
6998 (src_rsm->r_start - rack->rc_tp->snd_una),
6999 &src_rsm->soff);
7000 src_rsm->orig_m_len = src_rsm->m->m_len;
7001 src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7002 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7003 (rsm->r_start - rack->rc_tp->snd_una),
7004 &rsm->soff);
7005 rsm->orig_m_len = rsm->m->m_len;
7006 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7007 return;
7008 }
7009 }
7010 rsm->m = m;
7011 rsm->soff = soff;
7012 rsm->orig_m_len = m->m_len;
7013 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7014 }
7015
7016 static __inline void
rack_clone_rsm(struct tcp_rack * rack,struct rack_sendmap * nrsm,struct rack_sendmap * rsm,uint32_t start)7017 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7018 struct rack_sendmap *rsm, uint32_t start)
7019 {
7020 int idx;
7021
7022 nrsm->r_start = start;
7023 nrsm->r_end = rsm->r_end;
7024 nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7025 nrsm->r_act_rxt_cnt = rsm->r_act_rxt_cnt;
7026 nrsm->r_flags = rsm->r_flags;
7027 nrsm->r_dupack = rsm->r_dupack;
7028 nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7029 nrsm->r_rtr_bytes = 0;
7030 nrsm->r_fas = rsm->r_fas;
7031 nrsm->r_bas = rsm->r_bas;
7032 tqhash_update_end(rack->r_ctl.tqh, rsm, nrsm->r_start);
7033 nrsm->r_just_ret = rsm->r_just_ret;
7034 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7035 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7036 }
7037 /* Now if we have SYN flag we keep it on the left edge */
7038 if (nrsm->r_flags & RACK_HAS_SYN)
7039 nrsm->r_flags &= ~RACK_HAS_SYN;
7040 /* Now if we have a FIN flag we keep it on the right edge */
7041 if (rsm->r_flags & RACK_HAS_FIN)
7042 rsm->r_flags &= ~RACK_HAS_FIN;
7043 /* Push bit must go to the right edge as well */
7044 if (rsm->r_flags & RACK_HAD_PUSH)
7045 rsm->r_flags &= ~RACK_HAD_PUSH;
7046 /* Update the count if app limited */
7047 if (nrsm->r_flags & RACK_APP_LIMITED)
7048 rack->r_ctl.rc_app_limited_cnt++;
7049 /* Clone over the state of the hw_tls flag */
7050 nrsm->r_hw_tls = rsm->r_hw_tls;
7051 /*
7052 * Now we need to find nrsm's new location in the mbuf chain
7053 * we basically calculate a new offset, which is soff +
7054 * how much is left in original rsm. Then we walk out the mbuf
7055 * chain to find the righ position, it may be the same mbuf
7056 * or maybe not.
7057 */
7058 KASSERT(((rsm->m != NULL) ||
7059 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7060 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7061 if (rsm->m)
7062 rack_setup_offset_for_rsm(rack, rsm, nrsm);
7063 }
7064
7065 static struct rack_sendmap *
rack_merge_rsm(struct tcp_rack * rack,struct rack_sendmap * l_rsm,struct rack_sendmap * r_rsm)7066 rack_merge_rsm(struct tcp_rack *rack,
7067 struct rack_sendmap *l_rsm,
7068 struct rack_sendmap *r_rsm)
7069 {
7070 /*
7071 * We are merging two ack'd RSM's,
7072 * the l_rsm is on the left (lower seq
7073 * values) and the r_rsm is on the right
7074 * (higher seq value). The simplest way
7075 * to merge these is to move the right
7076 * one into the left. I don't think there
7077 * is any reason we need to try to find
7078 * the oldest (or last oldest retransmitted).
7079 */
7080 rack_log_map_chg(rack->rc_tp, rack, NULL,
7081 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7082 tqhash_update_end(rack->r_ctl.tqh, l_rsm, r_rsm->r_end);
7083 if (l_rsm->r_dupack < r_rsm->r_dupack)
7084 l_rsm->r_dupack = r_rsm->r_dupack;
7085 if (r_rsm->r_rtr_bytes)
7086 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7087 if (r_rsm->r_in_tmap) {
7088 /* This really should not happen */
7089 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7090 r_rsm->r_in_tmap = 0;
7091 }
7092
7093 /* Now the flags */
7094 if (r_rsm->r_flags & RACK_HAS_FIN)
7095 l_rsm->r_flags |= RACK_HAS_FIN;
7096 if (r_rsm->r_flags & RACK_TLP)
7097 l_rsm->r_flags |= RACK_TLP;
7098 if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7099 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7100 if ((r_rsm->r_flags & RACK_APP_LIMITED) &&
7101 ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7102 /*
7103 * If both are app-limited then let the
7104 * free lower the count. If right is app
7105 * limited and left is not, transfer.
7106 */
7107 l_rsm->r_flags |= RACK_APP_LIMITED;
7108 r_rsm->r_flags &= ~RACK_APP_LIMITED;
7109 if (r_rsm == rack->r_ctl.rc_first_appl)
7110 rack->r_ctl.rc_first_appl = l_rsm;
7111 }
7112 tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7113 /*
7114 * We keep the largest value, which is the newest
7115 * send. We do this in case a segment that is
7116 * joined together and not part of a GP estimate
7117 * later gets expanded into the GP estimate.
7118 *
7119 * We prohibit the merging of unlike kinds i.e.
7120 * all pieces that are in the GP estimate can be
7121 * merged and all pieces that are not in a GP estimate
7122 * can be merged, but not disimilar pieces. Combine
7123 * this with taking the highest here and we should
7124 * be ok unless of course the client reneges. Then
7125 * all bets are off.
7126 */
7127 if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7128 r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7129 l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7130 }
7131 /*
7132 * When merging two RSM's we also need to consider the ack time and keep
7133 * newest. If the ack gets merged into a measurement then that is the
7134 * one we will want to be using.
7135 */
7136 if(l_rsm->r_ack_arrival < r_rsm->r_ack_arrival)
7137 l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7138
7139 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7140 /* Transfer the split limit to the map we free */
7141 r_rsm->r_limit_type = l_rsm->r_limit_type;
7142 l_rsm->r_limit_type = 0;
7143 }
7144 rack_free(rack, r_rsm);
7145 l_rsm->r_flags |= RACK_MERGED;
7146 return (l_rsm);
7147 }
7148
7149 /*
7150 * TLP Timer, here we simply setup what segment we want to
7151 * have the TLP expire on, the normal rack_output() will then
7152 * send it out.
7153 *
7154 * We return 1, saying don't proceed with rack_output only
7155 * when all timers have been stopped (destroyed PCB?).
7156 */
7157 static int
rack_timeout_tlp(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,uint8_t * doing_tlp)7158 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7159 {
7160 /*
7161 * Tail Loss Probe.
7162 */
7163 struct rack_sendmap *rsm = NULL;
7164 int insret __diagused;
7165 struct socket *so = tptosocket(tp);
7166 uint32_t amm;
7167 uint32_t out, avail;
7168 int collapsed_win = 0;
7169
7170 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7171 /* Its not time yet */
7172 return (0);
7173 }
7174 if (ctf_progress_timeout_check(tp, true)) {
7175 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7176 return (-ETIMEDOUT); /* tcp_drop() */
7177 }
7178 /*
7179 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7180 * need to figure out how to force a full MSS segment out.
7181 */
7182 rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7183 rack->r_ctl.retran_during_recovery = 0;
7184 rack->r_might_revert = 0;
7185 rack->r_ctl.dsack_byte_cnt = 0;
7186 counter_u64_add(rack_tlp_tot, 1);
7187 if (rack->r_state && (rack->r_state != tp->t_state))
7188 rack_set_state(tp, rack);
7189 avail = sbavail(&so->so_snd);
7190 out = tp->snd_max - tp->snd_una;
7191 if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7192 /* special case, we need a retransmission */
7193 collapsed_win = 1;
7194 goto need_retran;
7195 }
7196 if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7197 rack->r_ctl.dsack_persist--;
7198 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7199 rack->r_ctl.num_dsack = 0;
7200 }
7201 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7202 }
7203 if ((tp->t_flags & TF_GPUTINPROG) &&
7204 (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7205 /*
7206 * If this is the second in a row
7207 * TLP and we are doing a measurement
7208 * its time to abandon the measurement.
7209 * Something is likely broken on
7210 * the clients network and measuring a
7211 * broken network does us no good.
7212 */
7213 tp->t_flags &= ~TF_GPUTINPROG;
7214 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7215 rack->r_ctl.rc_gp_srtt /*flex1*/,
7216 tp->gput_seq,
7217 0, 0, 18, __LINE__, NULL, 0);
7218 }
7219 /*
7220 * Check our send oldest always settings, and if
7221 * there is an oldest to send jump to the need_retran.
7222 */
7223 if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7224 goto need_retran;
7225
7226 if (avail > out) {
7227 /* New data is available */
7228 amm = avail - out;
7229 if (amm > ctf_fixed_maxseg(tp)) {
7230 amm = ctf_fixed_maxseg(tp);
7231 if ((amm + out) > tp->snd_wnd) {
7232 /* We are rwnd limited */
7233 goto need_retran;
7234 }
7235 } else if (amm < ctf_fixed_maxseg(tp)) {
7236 /* not enough to fill a MTU */
7237 goto need_retran;
7238 }
7239 if (IN_FASTRECOVERY(tp->t_flags)) {
7240 /* Unlikely */
7241 if (rack->rack_no_prr == 0) {
7242 if (out + amm <= tp->snd_wnd) {
7243 rack->r_ctl.rc_prr_sndcnt = amm;
7244 rack->r_ctl.rc_tlp_new_data = amm;
7245 rack_log_to_prr(rack, 4, 0, __LINE__);
7246 }
7247 } else
7248 goto need_retran;
7249 } else {
7250 /* Set the send-new override */
7251 if (out + amm <= tp->snd_wnd)
7252 rack->r_ctl.rc_tlp_new_data = amm;
7253 else
7254 goto need_retran;
7255 }
7256 rack->r_ctl.rc_tlpsend = NULL;
7257 counter_u64_add(rack_tlp_newdata, 1);
7258 goto send;
7259 }
7260 need_retran:
7261 /*
7262 * Ok we need to arrange the last un-acked segment to be re-sent, or
7263 * optionally the first un-acked segment.
7264 */
7265 if (collapsed_win == 0) {
7266 if (rack_always_send_oldest)
7267 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7268 else {
7269 rsm = tqhash_max(rack->r_ctl.tqh);
7270 if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7271 rsm = rack_find_high_nonack(rack, rsm);
7272 }
7273 }
7274 if (rsm == NULL) {
7275 #ifdef TCP_BLACKBOX
7276 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7277 #endif
7278 goto out;
7279 }
7280 } else {
7281 /*
7282 * We had a collapsed window, lets find
7283 * the point before the collapse.
7284 */
7285 if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7286 rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7287 else {
7288 rsm = tqhash_min(rack->r_ctl.tqh);
7289 }
7290 if (rsm == NULL) {
7291 /* Huh */
7292 goto out;
7293 }
7294 }
7295 if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7296 /*
7297 * We need to split this the last segment in two.
7298 */
7299 struct rack_sendmap *nrsm;
7300
7301 nrsm = rack_alloc_full_limit(rack);
7302 if (nrsm == NULL) {
7303 /*
7304 * No memory to split, we will just exit and punt
7305 * off to the RXT timer.
7306 */
7307 goto out;
7308 }
7309 rack_clone_rsm(rack, nrsm, rsm,
7310 (rsm->r_end - ctf_fixed_maxseg(tp)));
7311 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7312 #ifndef INVARIANTS
7313 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7314 #else
7315 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7316 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7317 nrsm, insret, rack, rsm);
7318 }
7319 #endif
7320 if (rsm->r_in_tmap) {
7321 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7322 nrsm->r_in_tmap = 1;
7323 }
7324 rsm = nrsm;
7325 }
7326 rack->r_ctl.rc_tlpsend = rsm;
7327 send:
7328 /* Make sure output path knows we are doing a TLP */
7329 *doing_tlp = 1;
7330 rack->r_timer_override = 1;
7331 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7332 return (0);
7333 out:
7334 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7335 return (0);
7336 }
7337
7338 /*
7339 * Delayed ack Timer, here we simply need to setup the
7340 * ACK_NOW flag and remove the DELACK flag. From there
7341 * the output routine will send the ack out.
7342 *
7343 * We only return 1, saying don't proceed, if all timers
7344 * are stopped (destroyed PCB?).
7345 */
7346 static int
rack_timeout_delack(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7347 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7348 {
7349
7350 rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7351 tp->t_flags &= ~TF_DELACK;
7352 tp->t_flags |= TF_ACKNOW;
7353 KMOD_TCPSTAT_INC(tcps_delack);
7354 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7355 return (0);
7356 }
7357
7358 static inline int
rack_send_ack_challange(struct tcp_rack * rack)7359 rack_send_ack_challange(struct tcp_rack *rack)
7360 {
7361 struct tcptemp *t_template;
7362
7363 t_template = tcpip_maketemplate(rack->rc_inp);
7364 if (t_template) {
7365 if (rack->forced_ack == 0) {
7366 rack->forced_ack = 1;
7367 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7368 } else {
7369 rack->probe_not_answered = 1;
7370 }
7371 tcp_respond(rack->rc_tp, t_template->tt_ipgen,
7372 &t_template->tt_t, (struct mbuf *)NULL,
7373 rack->rc_tp->rcv_nxt, rack->rc_tp->snd_una - 1, 0);
7374 free(t_template, M_TEMP);
7375 /* This does send an ack so kill any D-ack timer */
7376 if (rack->rc_tp->t_flags & TF_DELACK)
7377 rack->rc_tp->t_flags &= ~TF_DELACK;
7378 return(1);
7379 } else
7380 return (0);
7381
7382 }
7383
7384 /*
7385 * Persists timer, here we simply send the
7386 * same thing as a keepalive will.
7387 * the one byte send.
7388 *
7389 * We only return 1, saying don't proceed, if all timers
7390 * are stopped (destroyed PCB?).
7391 */
7392 static int
rack_timeout_persist(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7393 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7394 {
7395 int32_t retval = 1;
7396
7397 if (rack->rc_in_persist == 0)
7398 return (0);
7399 if (ctf_progress_timeout_check(tp, false)) {
7400 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7401 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7402 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7403 return (-ETIMEDOUT); /* tcp_drop() */
7404 }
7405 /*
7406 * Persistence timer into zero window. Force a byte to be output, if
7407 * possible.
7408 */
7409 KMOD_TCPSTAT_INC(tcps_persisttimeo);
7410 /*
7411 * Hack: if the peer is dead/unreachable, we do not time out if the
7412 * window is closed. After a full backoff, drop the connection if
7413 * the idle time (no responses to probes) reaches the maximum
7414 * backoff that we would use if retransmitting.
7415 */
7416 if (tp->t_rxtshift >= V_tcp_retries &&
7417 (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7418 TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7419 KMOD_TCPSTAT_INC(tcps_persistdrop);
7420 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7421 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7422 retval = -ETIMEDOUT; /* tcp_drop() */
7423 goto out;
7424 }
7425 if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7426 tp->snd_una == tp->snd_max)
7427 rack_exit_persist(tp, rack, cts);
7428 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7429 /*
7430 * If the user has closed the socket then drop a persisting
7431 * connection after a much reduced timeout.
7432 */
7433 if (tp->t_state > TCPS_CLOSE_WAIT &&
7434 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7435 KMOD_TCPSTAT_INC(tcps_persistdrop);
7436 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7437 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7438 retval = -ETIMEDOUT; /* tcp_drop() */
7439 goto out;
7440 }
7441 if (rack_send_ack_challange(rack)) {
7442 /* only set it if we were answered */
7443 if (rack->probe_not_answered) {
7444 counter_u64_add(rack_persists_loss, 1);
7445 rack->r_ctl.persist_lost_ends++;
7446 }
7447 counter_u64_add(rack_persists_sends, 1);
7448 counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7449 }
7450 if (tp->t_rxtshift < V_tcp_retries)
7451 tp->t_rxtshift++;
7452 out:
7453 rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7454 rack_start_hpts_timer(rack, tp, cts,
7455 0, 0, 0);
7456 return (retval);
7457 }
7458
7459 /*
7460 * If a keepalive goes off, we had no other timers
7461 * happening. We always return 1 here since this
7462 * routine either drops the connection or sends
7463 * out a segment with respond.
7464 */
7465 static int
rack_timeout_keepalive(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7466 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7467 {
7468 struct inpcb *inp = tptoinpcb(tp);
7469
7470 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7471 rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7472 /*
7473 * Keep-alive timer went off; send something or drop connection if
7474 * idle for too long.
7475 */
7476 KMOD_TCPSTAT_INC(tcps_keeptimeo);
7477 if (tp->t_state < TCPS_ESTABLISHED)
7478 goto dropit;
7479 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7480 tp->t_state <= TCPS_CLOSING) {
7481 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7482 goto dropit;
7483 /*
7484 * Send a packet designed to force a response if the peer is
7485 * up and reachable: either an ACK if the connection is
7486 * still alive, or an RST if the peer has closed the
7487 * connection due to timeout or reboot. Using sequence
7488 * number tp->snd_una-1 causes the transmitted zero-length
7489 * segment to lie outside the receive window; by the
7490 * protocol spec, this requires the correspondent TCP to
7491 * respond.
7492 */
7493 KMOD_TCPSTAT_INC(tcps_keepprobe);
7494 rack_send_ack_challange(rack);
7495 }
7496 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7497 return (1);
7498 dropit:
7499 KMOD_TCPSTAT_INC(tcps_keepdrops);
7500 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7501 return (-ETIMEDOUT); /* tcp_drop() */
7502 }
7503
7504 /*
7505 * Retransmit helper function, clear up all the ack
7506 * flags and take care of important book keeping.
7507 */
7508 static void
rack_remxt_tmr(struct tcpcb * tp)7509 rack_remxt_tmr(struct tcpcb *tp)
7510 {
7511 /*
7512 * The retransmit timer went off, all sack'd blocks must be
7513 * un-acked.
7514 */
7515 struct rack_sendmap *rsm, *trsm = NULL;
7516 struct tcp_rack *rack;
7517
7518 rack = (struct tcp_rack *)tp->t_fb_ptr;
7519 rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7520 rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7521 rack->r_timer_override = 1;
7522 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7523 rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7524 rack->r_late = 0;
7525 rack->r_early = 0;
7526 rack->r_ctl.rc_agg_delayed = 0;
7527 rack->r_ctl.rc_agg_early = 0;
7528 if (rack->r_state && (rack->r_state != tp->t_state))
7529 rack_set_state(tp, rack);
7530 if (tp->t_rxtshift <= rack_rxt_scoreboard_clear_thresh) {
7531 /*
7532 * We do not clear the scoreboard until we have had
7533 * more than rack_rxt_scoreboard_clear_thresh time-outs.
7534 */
7535 rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7536 if (rack->r_ctl.rc_resend != NULL)
7537 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7538
7539 return;
7540 }
7541 /*
7542 * Ideally we would like to be able to
7543 * mark SACK-PASS on anything not acked here.
7544 *
7545 * However, if we do that we would burst out
7546 * all that data 1ms apart. This would be unwise,
7547 * so for now we will just let the normal rxt timer
7548 * and tlp timer take care of it.
7549 *
7550 * Also we really need to stick them back in sequence
7551 * order. This way we send in the proper order and any
7552 * sacks that come floating in will "re-ack" the data.
7553 * To do this we zap the tmap with an INIT and then
7554 * walk through and place every rsm in the tail queue
7555 * hash table back in its seq ordered place.
7556 */
7557 TAILQ_INIT(&rack->r_ctl.rc_tmap);
7558
7559 TQHASH_FOREACH(rsm, rack->r_ctl.tqh) {
7560 rsm->r_dupack = 0;
7561 if (rack_verbose_logging)
7562 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7563 /* We must re-add it back to the tlist */
7564 if (trsm == NULL) {
7565 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7566 } else {
7567 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7568 }
7569 rsm->r_in_tmap = 1;
7570 trsm = rsm;
7571 if (rsm->r_flags & RACK_ACKED)
7572 rsm->r_flags |= RACK_WAS_ACKED;
7573 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED | RACK_WAS_LOST);
7574 rsm->r_flags |= RACK_MUST_RXT;
7575 }
7576 /* zero the lost since it's all gone */
7577 rack->r_ctl.rc_considered_lost = 0;
7578 /* Clear the count (we just un-acked them) */
7579 rack->r_ctl.rc_sacked = 0;
7580 rack->r_ctl.rc_sacklast = NULL;
7581 /* Clear the tlp rtx mark */
7582 rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7583 if (rack->r_ctl.rc_resend != NULL)
7584 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7585 rack->r_ctl.rc_prr_sndcnt = 0;
7586 rack_log_to_prr(rack, 6, 0, __LINE__);
7587 rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7588 if (rack->r_ctl.rc_resend != NULL)
7589 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7590 if (((tp->t_flags & TF_SACK_PERMIT) == 0) &&
7591 ((tp->t_flags & TF_SENTFIN) == 0)) {
7592 /*
7593 * For non-sack customers new data
7594 * needs to go out as retransmits until
7595 * we retransmit up to snd_max.
7596 */
7597 rack->r_must_retran = 1;
7598 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7599 rack->r_ctl.rc_sacked);
7600 }
7601 }
7602
7603 static void
rack_convert_rtts(struct tcpcb * tp)7604 rack_convert_rtts(struct tcpcb *tp)
7605 {
7606 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7607 tp->t_rxtcur = RACK_REXMTVAL(tp);
7608 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7609 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7610 }
7611 if (tp->t_rxtcur > rack_rto_max) {
7612 tp->t_rxtcur = rack_rto_max;
7613 }
7614 }
7615
7616 static void
rack_cc_conn_init(struct tcpcb * tp)7617 rack_cc_conn_init(struct tcpcb *tp)
7618 {
7619 struct tcp_rack *rack;
7620 uint32_t srtt;
7621
7622 rack = (struct tcp_rack *)tp->t_fb_ptr;
7623 srtt = tp->t_srtt;
7624 cc_conn_init(tp);
7625 /*
7626 * Now convert to rack's internal format,
7627 * if required.
7628 */
7629 if ((srtt == 0) && (tp->t_srtt != 0))
7630 rack_convert_rtts(tp);
7631 /*
7632 * We want a chance to stay in slowstart as
7633 * we create a connection. TCP spec says that
7634 * initially ssthresh is infinite. For our
7635 * purposes that is the snd_wnd.
7636 */
7637 if (tp->snd_ssthresh < tp->snd_wnd) {
7638 tp->snd_ssthresh = tp->snd_wnd;
7639 }
7640 /*
7641 * We also want to assure a IW worth of
7642 * data can get inflight.
7643 */
7644 if (rc_init_window(rack) < tp->snd_cwnd)
7645 tp->snd_cwnd = rc_init_window(rack);
7646 }
7647
7648 /*
7649 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7650 * we will setup to retransmit the lowest seq number outstanding.
7651 */
7652 static int
rack_timeout_rxt(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts)7653 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7654 {
7655 struct inpcb *inp = tptoinpcb(tp);
7656 int32_t rexmt;
7657 int32_t retval = 0;
7658 bool isipv6;
7659
7660 if ((tp->t_flags & TF_GPUTINPROG) &&
7661 (tp->t_rxtshift)) {
7662 /*
7663 * We have had a second timeout
7664 * measurements on successive rxt's are not profitable.
7665 * It is unlikely to be of any use (the network is
7666 * broken or the client went away).
7667 */
7668 tp->t_flags &= ~TF_GPUTINPROG;
7669 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7670 rack->r_ctl.rc_gp_srtt /*flex1*/,
7671 tp->gput_seq,
7672 0, 0, 18, __LINE__, NULL, 0);
7673 }
7674 if (ctf_progress_timeout_check(tp, false)) {
7675 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7676 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7677 return (-ETIMEDOUT); /* tcp_drop() */
7678 }
7679 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7680 rack->r_ctl.retran_during_recovery = 0;
7681 rack->rc_ack_required = 1;
7682 rack->r_ctl.dsack_byte_cnt = 0;
7683 if (IN_RECOVERY(tp->t_flags) &&
7684 (rack->rto_from_rec == 0)) {
7685 /*
7686 * Mark that we had a rto while in recovery
7687 * and save the ssthresh so if we go back
7688 * into recovery we will have a chance
7689 * to slowstart back to the level.
7690 */
7691 rack->rto_from_rec = 1;
7692 rack->r_ctl.rto_ssthresh = tp->snd_ssthresh;
7693 }
7694 if (IN_FASTRECOVERY(tp->t_flags))
7695 tp->t_flags |= TF_WASFRECOVERY;
7696 else
7697 tp->t_flags &= ~TF_WASFRECOVERY;
7698 if (IN_CONGRECOVERY(tp->t_flags))
7699 tp->t_flags |= TF_WASCRECOVERY;
7700 else
7701 tp->t_flags &= ~TF_WASCRECOVERY;
7702 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7703 (tp->snd_una == tp->snd_max)) {
7704 /* Nothing outstanding .. nothing to do */
7705 return (0);
7706 }
7707 if (rack->r_ctl.dsack_persist) {
7708 rack->r_ctl.dsack_persist--;
7709 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7710 rack->r_ctl.num_dsack = 0;
7711 }
7712 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7713 }
7714 /*
7715 * Rack can only run one timer at a time, so we cannot
7716 * run a KEEPINIT (gating SYN sending) and a retransmit
7717 * timer for the SYN. So if we are in a front state and
7718 * have a KEEPINIT timer we need to check the first transmit
7719 * against now to see if we have exceeded the KEEPINIT time
7720 * (if one is set).
7721 */
7722 if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7723 (TP_KEEPINIT(tp) != 0)) {
7724 struct rack_sendmap *rsm;
7725
7726 rsm = tqhash_min(rack->r_ctl.tqh);
7727 if (rsm) {
7728 /* Ok we have something outstanding to test keepinit with */
7729 if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7730 ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7731 /* We have exceeded the KEEPINIT time */
7732 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7733 goto drop_it;
7734 }
7735 }
7736 }
7737 /*
7738 * Retransmission timer went off. Message has not been acked within
7739 * retransmit interval. Back off to a longer retransmit interval
7740 * and retransmit one segment.
7741 */
7742 if ((rack->r_ctl.rc_resend == NULL) ||
7743 ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7744 /*
7745 * If the rwnd collapsed on
7746 * the one we are retransmitting
7747 * it does not count against the
7748 * rxt count.
7749 */
7750 tp->t_rxtshift++;
7751 }
7752 rack_remxt_tmr(tp);
7753 if (tp->t_rxtshift > V_tcp_retries) {
7754 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7755 drop_it:
7756 tp->t_rxtshift = V_tcp_retries;
7757 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7758 /* XXXGL: previously t_softerror was casted to uint16_t */
7759 MPASS(tp->t_softerror >= 0);
7760 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7761 goto out; /* tcp_drop() */
7762 }
7763 if (tp->t_state == TCPS_SYN_SENT) {
7764 /*
7765 * If the SYN was retransmitted, indicate CWND to be limited
7766 * to 1 segment in cc_conn_init().
7767 */
7768 tp->snd_cwnd = 1;
7769 } else if (tp->t_rxtshift == 1) {
7770 /*
7771 * first retransmit; record ssthresh and cwnd so they can be
7772 * recovered if this turns out to be a "bad" retransmit. A
7773 * retransmit is considered "bad" if an ACK for this segment
7774 * is received within RTT/2 interval; the assumption here is
7775 * that the ACK was already in flight. See "On Estimating
7776 * End-to-End Network Path Properties" by Allman and Paxson
7777 * for more details.
7778 */
7779 tp->snd_cwnd_prev = tp->snd_cwnd;
7780 tp->snd_ssthresh_prev = tp->snd_ssthresh;
7781 tp->snd_recover_prev = tp->snd_recover;
7782 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7783 tp->t_flags |= TF_PREVVALID;
7784 } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7785 tp->t_flags &= ~TF_PREVVALID;
7786 KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7787 if ((tp->t_state == TCPS_SYN_SENT) ||
7788 (tp->t_state == TCPS_SYN_RECEIVED))
7789 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7790 else
7791 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7792
7793 RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7794 max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7795 /*
7796 * We enter the path for PLMTUD if connection is established or, if
7797 * connection is FIN_WAIT_1 status, reason for the last is that if
7798 * amount of data we send is very small, we could send it in couple
7799 * of packets and process straight to FIN. In that case we won't
7800 * catch ESTABLISHED state.
7801 */
7802 #ifdef INET6
7803 isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7804 #else
7805 isipv6 = false;
7806 #endif
7807 if (((V_tcp_pmtud_blackhole_detect == 1) ||
7808 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7809 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7810 ((tp->t_state == TCPS_ESTABLISHED) ||
7811 (tp->t_state == TCPS_FIN_WAIT_1))) {
7812 /*
7813 * Idea here is that at each stage of mtu probe (usually,
7814 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7815 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7816 * should take care of that.
7817 */
7818 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7819 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7820 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7821 tp->t_rxtshift % 2 == 0)) {
7822 /*
7823 * Enter Path MTU Black-hole Detection mechanism: -
7824 * Disable Path MTU Discovery (IP "DF" bit). -
7825 * Reduce MTU to lower value than what we negotiated
7826 * with peer.
7827 */
7828 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7829 /* Record that we may have found a black hole. */
7830 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7831 /* Keep track of previous MSS. */
7832 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7833 }
7834
7835 /*
7836 * Reduce the MSS to blackhole value or to the
7837 * default in an attempt to retransmit.
7838 */
7839 #ifdef INET6
7840 if (isipv6 &&
7841 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7842 /* Use the sysctl tuneable blackhole MSS. */
7843 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7844 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7845 } else if (isipv6) {
7846 /* Use the default MSS. */
7847 tp->t_maxseg = V_tcp_v6mssdflt;
7848 /*
7849 * Disable Path MTU Discovery when we switch
7850 * to minmss.
7851 */
7852 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7853 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7854 }
7855 #endif
7856 #if defined(INET6) && defined(INET)
7857 else
7858 #endif
7859 #ifdef INET
7860 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7861 /* Use the sysctl tuneable blackhole MSS. */
7862 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7863 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7864 } else {
7865 /* Use the default MSS. */
7866 tp->t_maxseg = V_tcp_mssdflt;
7867 /*
7868 * Disable Path MTU Discovery when we switch
7869 * to minmss.
7870 */
7871 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7872 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7873 }
7874 #endif
7875 } else {
7876 /*
7877 * If further retransmissions are still unsuccessful
7878 * with a lowered MTU, maybe this isn't a blackhole
7879 * and we restore the previous MSS and blackhole
7880 * detection flags. The limit '6' is determined by
7881 * giving each probe stage (1448, 1188, 524) 2
7882 * chances to recover.
7883 */
7884 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7885 (tp->t_rxtshift >= 6)) {
7886 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7887 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7888 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7889 if (tp->t_maxseg < V_tcp_mssdflt) {
7890 /*
7891 * The MSS is so small we should not
7892 * process incoming SACK's since we are
7893 * subject to attack in such a case.
7894 */
7895 tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
7896 } else {
7897 tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
7898 }
7899 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7900 }
7901 }
7902 }
7903 /*
7904 * Disable RFC1323 and SACK if we haven't got any response to
7905 * our third SYN to work-around some broken terminal servers
7906 * (most of which have hopefully been retired) that have bad VJ
7907 * header compression code which trashes TCP segments containing
7908 * unknown-to-them TCP options.
7909 */
7910 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7911 (tp->t_rxtshift == 3))
7912 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7913 /*
7914 * If we backed off this far, our srtt estimate is probably bogus.
7915 * Clobber it so we'll take the next rtt measurement as our srtt;
7916 * move the current srtt into rttvar to keep the current retransmit
7917 * times until then.
7918 */
7919 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7920 #ifdef INET6
7921 if ((inp->inp_vflag & INP_IPV6) != 0)
7922 in6_losing(inp);
7923 else
7924 #endif
7925 in_losing(inp);
7926 tp->t_rttvar += tp->t_srtt;
7927 tp->t_srtt = 0;
7928 }
7929 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7930 tp->snd_recover = tp->snd_max;
7931 tp->t_flags |= TF_ACKNOW;
7932 tp->t_rtttime = 0;
7933 rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
7934 out:
7935 return (retval);
7936 }
7937
7938 static int
rack_process_timers(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,uint8_t hpts_calling,uint8_t * doing_tlp)7939 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7940 {
7941 int32_t ret = 0;
7942 int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7943
7944 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7945 (tp->t_flags & TF_GPUTINPROG)) {
7946 /*
7947 * We have a goodput in progress
7948 * and we have entered a late state.
7949 * Do we have enough data in the sb
7950 * to handle the GPUT request?
7951 */
7952 uint32_t bytes;
7953
7954 bytes = tp->gput_ack - tp->gput_seq;
7955 if (SEQ_GT(tp->gput_seq, tp->snd_una))
7956 bytes += tp->gput_seq - tp->snd_una;
7957 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
7958 /*
7959 * There are not enough bytes in the socket
7960 * buffer that have been sent to cover this
7961 * measurement. Cancel it.
7962 */
7963 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7964 rack->r_ctl.rc_gp_srtt /*flex1*/,
7965 tp->gput_seq,
7966 0, 0, 18, __LINE__, NULL, 0);
7967 tp->t_flags &= ~TF_GPUTINPROG;
7968 }
7969 }
7970 if (timers == 0) {
7971 return (0);
7972 }
7973 if (tp->t_state == TCPS_LISTEN) {
7974 /* no timers on listen sockets */
7975 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7976 return (0);
7977 return (1);
7978 }
7979 if ((timers & PACE_TMR_RACK) &&
7980 rack->rc_on_min_to) {
7981 /*
7982 * For the rack timer when we
7983 * are on a min-timeout (which means rrr_conf = 3)
7984 * we don't want to check the timer. It may
7985 * be going off for a pace and thats ok we
7986 * want to send the retransmit (if its ready).
7987 *
7988 * If its on a normal rack timer (non-min) then
7989 * we will check if its expired.
7990 */
7991 goto skip_time_check;
7992 }
7993 if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7994 uint32_t left;
7995
7996 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
7997 ret = -1;
7998 rack_log_to_processing(rack, cts, ret, 0);
7999 return (0);
8000 }
8001 if (hpts_calling == 0) {
8002 /*
8003 * A user send or queued mbuf (sack) has called us? We
8004 * return 0 and let the pacing guards
8005 * deal with it if they should or
8006 * should not cause a send.
8007 */
8008 ret = -2;
8009 rack_log_to_processing(rack, cts, ret, 0);
8010 return (0);
8011 }
8012 /*
8013 * Ok our timer went off early and we are not paced false
8014 * alarm, go back to sleep. We make sure we don't have
8015 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8016 * flag in place.
8017 */
8018 rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8019 ret = -3;
8020 left = rack->r_ctl.rc_timer_exp - cts;
8021 tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8022 rack_log_to_processing(rack, cts, ret, left);
8023 return (1);
8024 }
8025 skip_time_check:
8026 rack->rc_tmr_stopped = 0;
8027 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8028 if (timers & PACE_TMR_DELACK) {
8029 ret = rack_timeout_delack(tp, rack, cts);
8030 } else if (timers & PACE_TMR_RACK) {
8031 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8032 rack->r_fast_output = 0;
8033 ret = rack_timeout_rack(tp, rack, cts);
8034 } else if (timers & PACE_TMR_TLP) {
8035 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8036 rack->r_fast_output = 0;
8037 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8038 } else if (timers & PACE_TMR_RXT) {
8039 rack->r_ctl.rc_tlp_rxt_last_time = cts;
8040 rack->r_fast_output = 0;
8041 ret = rack_timeout_rxt(tp, rack, cts);
8042 } else if (timers & PACE_TMR_PERSIT) {
8043 ret = rack_timeout_persist(tp, rack, cts);
8044 } else if (timers & PACE_TMR_KEEP) {
8045 ret = rack_timeout_keepalive(tp, rack, cts);
8046 }
8047 rack_log_to_processing(rack, cts, ret, timers);
8048 return (ret);
8049 }
8050
8051 static void
rack_timer_cancel(struct tcpcb * tp,struct tcp_rack * rack,uint32_t cts,int line)8052 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8053 {
8054 struct timeval tv;
8055 uint32_t us_cts, flags_on_entry;
8056 uint8_t hpts_removed = 0;
8057
8058 flags_on_entry = rack->r_ctl.rc_hpts_flags;
8059 us_cts = tcp_get_usecs(&tv);
8060 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8061 ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8062 ((tp->snd_max - tp->snd_una) == 0))) {
8063 tcp_hpts_remove(rack->rc_tp);
8064 hpts_removed = 1;
8065 /* If we were not delayed cancel out the flag. */
8066 if ((tp->snd_max - tp->snd_una) == 0)
8067 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8068 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8069 }
8070 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8071 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8072 if (tcp_in_hpts(rack->rc_tp) &&
8073 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8074 /*
8075 * Canceling timer's when we have no output being
8076 * paced. We also must remove ourselves from the
8077 * hpts.
8078 */
8079 tcp_hpts_remove(rack->rc_tp);
8080 hpts_removed = 1;
8081 }
8082 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8083 }
8084 if (hpts_removed == 0)
8085 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8086 }
8087
8088 static int
rack_stopall(struct tcpcb * tp)8089 rack_stopall(struct tcpcb *tp)
8090 {
8091 struct tcp_rack *rack;
8092
8093 rack = (struct tcp_rack *)tp->t_fb_ptr;
8094 rack->t_timers_stopped = 1;
8095
8096 tcp_hpts_remove(tp);
8097
8098 return (0);
8099 }
8100
8101 static void
rack_stop_all_timers(struct tcpcb * tp,struct tcp_rack * rack)8102 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8103 {
8104 /*
8105 * Assure no timers are running.
8106 */
8107 if (tcp_timer_active(tp, TT_PERSIST)) {
8108 /* We enter in persists, set the flag appropriately */
8109 rack->rc_in_persist = 1;
8110 }
8111 if (tcp_in_hpts(rack->rc_tp)) {
8112 tcp_hpts_remove(rack->rc_tp);
8113 }
8114 }
8115
8116 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)8117 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8118 struct rack_sendmap *rsm, uint64_t ts, uint32_t add_flag, int segsiz)
8119 {
8120 int32_t idx;
8121
8122 rsm->r_rtr_cnt++;
8123 if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8124 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8125 rsm->r_flags |= RACK_OVERMAX;
8126 }
8127 rsm->r_act_rxt_cnt++;
8128 /* Peg the count/index */
8129 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8130 rsm->r_dupack = 0;
8131 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8132 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8133 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8134 }
8135 if (rsm->r_flags & RACK_WAS_LOST) {
8136 /*
8137 * We retransmitted it putting it back in flight
8138 * remove the lost desgination and reduce the
8139 * bytes considered lost.
8140 */
8141 rsm->r_flags &= ~RACK_WAS_LOST;
8142 KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
8143 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
8144 if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
8145 rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
8146 else
8147 rack->r_ctl.rc_considered_lost = 0;
8148 }
8149 idx = rsm->r_rtr_cnt - 1;
8150 rsm->r_tim_lastsent[idx] = ts;
8151 /*
8152 * Here we don't add in the len of send, since its already
8153 * in snduna <->snd_max.
8154 */
8155 rsm->r_fas = ctf_flight_size(rack->rc_tp,
8156 rack->r_ctl.rc_sacked);
8157 if (rsm->r_flags & RACK_ACKED) {
8158 /* Problably MTU discovery messing with us */
8159 rsm->r_flags &= ~RACK_ACKED;
8160 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8161 }
8162 if (rsm->r_in_tmap) {
8163 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8164 rsm->r_in_tmap = 0;
8165 }
8166 /* Lets make sure it really is in or not the GP window */
8167 rack_mark_in_gp_win(tp, rsm);
8168 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8169 rsm->r_in_tmap = 1;
8170 rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8171 /* Take off the must retransmit flag, if its on */
8172 if (rsm->r_flags & RACK_MUST_RXT) {
8173 if (rack->r_must_retran)
8174 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8175 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8176 /*
8177 * We have retransmitted all we need. Clear
8178 * any must retransmit flags.
8179 */
8180 rack->r_must_retran = 0;
8181 rack->r_ctl.rc_out_at_rto = 0;
8182 }
8183 rsm->r_flags &= ~RACK_MUST_RXT;
8184 }
8185 /* Remove any collapsed flag */
8186 rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8187 if (rsm->r_flags & RACK_SACK_PASSED) {
8188 /* We have retransmitted due to the SACK pass */
8189 rsm->r_flags &= ~RACK_SACK_PASSED;
8190 rsm->r_flags |= RACK_WAS_SACKPASS;
8191 }
8192 }
8193
8194 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)8195 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8196 struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint32_t add_flag, int segsiz)
8197 {
8198 /*
8199 * We (re-)transmitted starting at rsm->r_start for some length
8200 * (possibly less than r_end.
8201 */
8202 struct rack_sendmap *nrsm;
8203 int insret __diagused;
8204 uint32_t c_end;
8205 int32_t len;
8206
8207 len = *lenp;
8208 c_end = rsm->r_start + len;
8209 if (SEQ_GEQ(c_end, rsm->r_end)) {
8210 /*
8211 * We retransmitted the whole piece or more than the whole
8212 * slopping into the next rsm.
8213 */
8214 rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8215 if (c_end == rsm->r_end) {
8216 *lenp = 0;
8217 return (0);
8218 } else {
8219 int32_t act_len;
8220
8221 /* Hangs over the end return whats left */
8222 act_len = rsm->r_end - rsm->r_start;
8223 *lenp = (len - act_len);
8224 return (rsm->r_end);
8225 }
8226 /* We don't get out of this block. */
8227 }
8228 /*
8229 * Here we retransmitted less than the whole thing which means we
8230 * have to split this into what was transmitted and what was not.
8231 */
8232 nrsm = rack_alloc_full_limit(rack);
8233 if (nrsm == NULL) {
8234 /*
8235 * We can't get memory, so lets not proceed.
8236 */
8237 *lenp = 0;
8238 return (0);
8239 }
8240 /*
8241 * So here we are going to take the original rsm and make it what we
8242 * retransmitted. nrsm will be the tail portion we did not
8243 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8244 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8245 * 1, 6 and the new piece will be 6, 11.
8246 */
8247 rack_clone_rsm(rack, nrsm, rsm, c_end);
8248 nrsm->r_dupack = 0;
8249 rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8250 #ifndef INVARIANTS
8251 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8252 #else
8253 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8254 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8255 nrsm, insret, rack, rsm);
8256 }
8257 #endif
8258 if (rsm->r_in_tmap) {
8259 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8260 nrsm->r_in_tmap = 1;
8261 }
8262 rsm->r_flags &= (~RACK_HAS_FIN);
8263 rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8264 /* Log a split of rsm into rsm and nrsm */
8265 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8266 *lenp = 0;
8267 return (0);
8268 }
8269
8270 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)8271 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8272 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8273 struct rack_sendmap *hintrsm, uint32_t add_flag, struct mbuf *s_mb,
8274 uint32_t s_moff, int hw_tls, int segsiz)
8275 {
8276 struct tcp_rack *rack;
8277 struct rack_sendmap *rsm, *nrsm;
8278 int insret __diagused;
8279
8280 register uint32_t snd_max, snd_una;
8281
8282 /*
8283 * Add to the RACK log of packets in flight or retransmitted. If
8284 * there is a TS option we will use the TS echoed, if not we will
8285 * grab a TS.
8286 *
8287 * Retransmissions will increment the count and move the ts to its
8288 * proper place. Note that if options do not include TS's then we
8289 * won't be able to effectively use the ACK for an RTT on a retran.
8290 *
8291 * Notes about r_start and r_end. Lets consider a send starting at
8292 * sequence 1 for 10 bytes. In such an example the r_start would be
8293 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8294 * This means that r_end is actually the first sequence for the next
8295 * slot (11).
8296 *
8297 */
8298 /*
8299 * If err is set what do we do XXXrrs? should we not add the thing?
8300 * -- i.e. return if err != 0 or should we pretend we sent it? --
8301 * i.e. proceed with add ** do this for now.
8302 */
8303 INP_WLOCK_ASSERT(tptoinpcb(tp));
8304 if (err)
8305 /*
8306 * We don't log errors -- we could but snd_max does not
8307 * advance in this case either.
8308 */
8309 return;
8310
8311 if (th_flags & TH_RST) {
8312 /*
8313 * We don't log resets and we return immediately from
8314 * sending
8315 */
8316 return;
8317 }
8318 rack = (struct tcp_rack *)tp->t_fb_ptr;
8319 snd_una = tp->snd_una;
8320 snd_max = tp->snd_max;
8321 if (th_flags & (TH_SYN | TH_FIN)) {
8322 /*
8323 * The call to rack_log_output is made before bumping
8324 * snd_max. This means we can record one extra byte on a SYN
8325 * or FIN if seq_out is adding more on and a FIN is present
8326 * (and we are not resending).
8327 */
8328 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8329 len++;
8330 if (th_flags & TH_FIN)
8331 len++;
8332 }
8333 if (SEQ_LEQ((seq_out + len), snd_una)) {
8334 /* Are sending an old segment to induce an ack (keep-alive)? */
8335 return;
8336 }
8337 if (SEQ_LT(seq_out, snd_una)) {
8338 /* huh? should we panic? */
8339 uint32_t end;
8340
8341 end = seq_out + len;
8342 seq_out = snd_una;
8343 if (SEQ_GEQ(end, seq_out))
8344 len = end - seq_out;
8345 else
8346 len = 0;
8347 }
8348 if (len == 0) {
8349 /* We don't log zero window probes */
8350 return;
8351 }
8352 if (IN_FASTRECOVERY(tp->t_flags)) {
8353 rack->r_ctl.rc_prr_out += len;
8354 }
8355 /* First question is it a retransmission or new? */
8356 if (seq_out == snd_max) {
8357 /* Its new */
8358 rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8359 again:
8360 rsm = rack_alloc(rack);
8361 if (rsm == NULL) {
8362 /*
8363 * Hmm out of memory and the tcb got destroyed while
8364 * we tried to wait.
8365 */
8366 return;
8367 }
8368 if (th_flags & TH_FIN) {
8369 rsm->r_flags = RACK_HAS_FIN|add_flag;
8370 } else {
8371 rsm->r_flags = add_flag;
8372 }
8373 if (hw_tls)
8374 rsm->r_hw_tls = 1;
8375 rsm->r_tim_lastsent[0] = cts;
8376 rsm->r_rtr_cnt = 1;
8377 rsm->r_act_rxt_cnt = 0;
8378 rsm->r_rtr_bytes = 0;
8379 if (th_flags & TH_SYN) {
8380 /* The data space is one beyond snd_una */
8381 rsm->r_flags |= RACK_HAS_SYN;
8382 }
8383 rsm->r_start = seq_out;
8384 rsm->r_end = rsm->r_start + len;
8385 rack_mark_in_gp_win(tp, rsm);
8386 rsm->r_dupack = 0;
8387 /*
8388 * save off the mbuf location that
8389 * sndmbuf_noadv returned (which is
8390 * where we started copying from)..
8391 */
8392 rsm->m = s_mb;
8393 rsm->soff = s_moff;
8394 /*
8395 * Here we do add in the len of send, since its not yet
8396 * reflected in in snduna <->snd_max
8397 */
8398 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8399 rack->r_ctl.rc_sacked) +
8400 (rsm->r_end - rsm->r_start));
8401 if ((rack->rc_initial_ss_comp == 0) &&
8402 (rack->r_ctl.ss_hi_fs < rsm->r_fas)) {
8403 rack->r_ctl.ss_hi_fs = rsm->r_fas;
8404 }
8405 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8406 if (rsm->m) {
8407 if (rsm->m->m_len <= rsm->soff) {
8408 /*
8409 * XXXrrs Question, will this happen?
8410 *
8411 * If sbsndptr is set at the correct place
8412 * then s_moff should always be somewhere
8413 * within rsm->m. But if the sbsndptr was
8414 * off then that won't be true. If it occurs
8415 * we need to walkout to the correct location.
8416 */
8417 struct mbuf *lm;
8418
8419 lm = rsm->m;
8420 while (lm->m_len <= rsm->soff) {
8421 rsm->soff -= lm->m_len;
8422 lm = lm->m_next;
8423 KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8424 __func__, rack, s_moff, s_mb, rsm->soff));
8425 }
8426 rsm->m = lm;
8427 }
8428 rsm->orig_m_len = rsm->m->m_len;
8429 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8430 } else {
8431 rsm->orig_m_len = 0;
8432 rsm->orig_t_space = 0;
8433 }
8434 rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8435 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8436 /* Log a new rsm */
8437 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8438 #ifndef INVARIANTS
8439 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
8440 #else
8441 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8442 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8443 nrsm, insret, rack, rsm);
8444 }
8445 #endif
8446 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8447 rsm->r_in_tmap = 1;
8448 if (rsm->r_flags & RACK_IS_PCM) {
8449 rack->r_ctl.pcm_i.send_time = cts;
8450 rack->r_ctl.pcm_i.eseq = rsm->r_end;
8451 /* First time through we set the start too */
8452 if (rack->pcm_in_progress == 0)
8453 rack->r_ctl.pcm_i.sseq = rsm->r_start;
8454 }
8455 /*
8456 * Special case detection, is there just a single
8457 * packet outstanding when we are not in recovery?
8458 *
8459 * If this is true mark it so.
8460 */
8461 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8462 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8463 struct rack_sendmap *prsm;
8464
8465 prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8466 if (prsm)
8467 prsm->r_one_out_nr = 1;
8468 }
8469 return;
8470 }
8471 /*
8472 * If we reach here its a retransmission and we need to find it.
8473 */
8474 more:
8475 if (hintrsm && (hintrsm->r_start == seq_out)) {
8476 rsm = hintrsm;
8477 hintrsm = NULL;
8478 } else {
8479 /* No hints sorry */
8480 rsm = NULL;
8481 }
8482 if ((rsm) && (rsm->r_start == seq_out)) {
8483 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8484 if (len == 0) {
8485 return;
8486 } else {
8487 goto more;
8488 }
8489 }
8490 /* Ok it was not the last pointer go through it the hard way. */
8491 refind:
8492 rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8493 if (rsm) {
8494 if (rsm->r_start == seq_out) {
8495 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8496 if (len == 0) {
8497 return;
8498 } else {
8499 goto refind;
8500 }
8501 }
8502 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8503 /* Transmitted within this piece */
8504 /*
8505 * Ok we must split off the front and then let the
8506 * update do the rest
8507 */
8508 nrsm = rack_alloc_full_limit(rack);
8509 if (nrsm == NULL) {
8510 rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8511 return;
8512 }
8513 /*
8514 * copy rsm to nrsm and then trim the front of rsm
8515 * to not include this part.
8516 */
8517 rack_clone_rsm(rack, nrsm, rsm, seq_out);
8518 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8519 #ifndef INVARIANTS
8520 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8521 #else
8522 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8523 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8524 nrsm, insret, rack, rsm);
8525 }
8526 #endif
8527 if (rsm->r_in_tmap) {
8528 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8529 nrsm->r_in_tmap = 1;
8530 }
8531 rsm->r_flags &= (~RACK_HAS_FIN);
8532 seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8533 if (len == 0) {
8534 return;
8535 } else if (len > 0)
8536 goto refind;
8537 }
8538 }
8539 /*
8540 * Hmm not found in map did they retransmit both old and on into the
8541 * new?
8542 */
8543 if (seq_out == tp->snd_max) {
8544 goto again;
8545 } else if (SEQ_LT(seq_out, tp->snd_max)) {
8546 #ifdef INVARIANTS
8547 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8548 seq_out, len, tp->snd_una, tp->snd_max);
8549 printf("Starting Dump of all rack entries\n");
8550 TQHASH_FOREACH(rsm, rack->r_ctl.tqh) {
8551 printf("rsm:%p start:%u end:%u\n",
8552 rsm, rsm->r_start, rsm->r_end);
8553 }
8554 printf("Dump complete\n");
8555 panic("seq_out not found rack:%p tp:%p",
8556 rack, tp);
8557 #endif
8558 } else {
8559 #ifdef INVARIANTS
8560 /*
8561 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8562 * flag)
8563 */
8564 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8565 seq_out, len, tp->snd_max, tp);
8566 #endif
8567 }
8568 }
8569
8570 /*
8571 * Record one of the RTT updates from an ack into
8572 * our sample structure.
8573 */
8574
8575 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)8576 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8577 int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8578 {
8579 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8580 (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8581 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8582 }
8583 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8584 (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8585 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8586 }
8587 if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8588 if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8589 rack->r_ctl.rc_gp_lowrtt = us_rtt;
8590 if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8591 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8592 }
8593 if ((confidence == 1) &&
8594 ((rsm == NULL) ||
8595 (rsm->r_just_ret) ||
8596 (rsm->r_one_out_nr &&
8597 len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8598 /*
8599 * If the rsm had a just return
8600 * hit it then we can't trust the
8601 * rtt measurement for buffer deterimination
8602 * Note that a confidence of 2, indicates
8603 * SACK'd which overrides the r_just_ret or
8604 * the r_one_out_nr. If it was a CUM-ACK and
8605 * we had only two outstanding, but get an
8606 * ack for only 1. Then that also lowers our
8607 * confidence.
8608 */
8609 confidence = 0;
8610 }
8611 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8612 (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8613 if (rack->r_ctl.rack_rs.confidence == 0) {
8614 /*
8615 * We take anything with no current confidence
8616 * saved.
8617 */
8618 rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8619 rack->r_ctl.rack_rs.confidence = confidence;
8620 rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8621 } else if (confidence != 0) {
8622 /*
8623 * Once we have a confident number,
8624 * we can update it with a smaller
8625 * value since this confident number
8626 * may include the DSACK time until
8627 * the next segment (the second one) arrived.
8628 */
8629 rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8630 rack->r_ctl.rack_rs.confidence = confidence;
8631 rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8632 }
8633 }
8634 rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8635 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8636 rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8637 rack->r_ctl.rack_rs.rs_rtt_cnt++;
8638 }
8639
8640 /*
8641 * Collect new round-trip time estimate
8642 * and update averages and current timeout.
8643 */
8644 static void
tcp_rack_xmit_timer_commit(struct tcp_rack * rack,struct tcpcb * tp)8645 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8646 {
8647 int32_t delta;
8648 int32_t rtt;
8649
8650 if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8651 /* No valid sample */
8652 return;
8653 if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8654 /* We are to use the lowest RTT seen in a single ack */
8655 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8656 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8657 /* We are to use the highest RTT seen in a single ack */
8658 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8659 } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8660 /* We are to use the average RTT seen in a single ack */
8661 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8662 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8663 } else {
8664 #ifdef INVARIANTS
8665 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8666 #endif
8667 return;
8668 }
8669 if (rtt == 0)
8670 rtt = 1;
8671 if (rack->rc_gp_rtt_set == 0) {
8672 /*
8673 * With no RTT we have to accept
8674 * even one we are not confident of.
8675 */
8676 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8677 rack->rc_gp_rtt_set = 1;
8678 } else if (rack->r_ctl.rack_rs.confidence) {
8679 /* update the running gp srtt */
8680 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8681 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8682 }
8683 if (rack->r_ctl.rack_rs.confidence) {
8684 /*
8685 * record the low and high for highly buffered path computation,
8686 * we only do this if we are confident (not a retransmission).
8687 */
8688 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8689 rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8690 }
8691 if (rack->rc_highly_buffered == 0) {
8692 /*
8693 * Currently once we declare a path has
8694 * highly buffered there is no going
8695 * back, which may be a problem...
8696 */
8697 if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8698 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8699 rack->r_ctl.rc_highest_us_rtt,
8700 rack->r_ctl.rc_lowest_us_rtt,
8701 RACK_RTTS_SEEHBP);
8702 rack->rc_highly_buffered = 1;
8703 }
8704 }
8705 }
8706 if ((rack->r_ctl.rack_rs.confidence) ||
8707 (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8708 /*
8709 * If we are highly confident of it <or> it was
8710 * never retransmitted we accept it as the last us_rtt.
8711 */
8712 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8713 /* The lowest rtt can be set if its was not retransmited */
8714 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8715 rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8716 if (rack->r_ctl.rc_lowest_us_rtt == 0)
8717 rack->r_ctl.rc_lowest_us_rtt = 1;
8718 }
8719 }
8720 rack = (struct tcp_rack *)tp->t_fb_ptr;
8721 if (tp->t_srtt != 0) {
8722 /*
8723 * We keep a simple srtt in microseconds, like our rtt
8724 * measurement. We don't need to do any tricks with shifting
8725 * etc. Instead we just add in 1/8th of the new measurement
8726 * and subtract out 1/8 of the old srtt. We do the same with
8727 * the variance after finding the absolute value of the
8728 * difference between this sample and the current srtt.
8729 */
8730 delta = tp->t_srtt - rtt;
8731 /* Take off 1/8th of the current sRTT */
8732 tp->t_srtt -= (tp->t_srtt >> 3);
8733 /* Add in 1/8th of the new RTT just measured */
8734 tp->t_srtt += (rtt >> 3);
8735 if (tp->t_srtt <= 0)
8736 tp->t_srtt = 1;
8737 /* Now lets make the absolute value of the variance */
8738 if (delta < 0)
8739 delta = -delta;
8740 /* Subtract out 1/8th */
8741 tp->t_rttvar -= (tp->t_rttvar >> 3);
8742 /* Add in 1/8th of the new variance we just saw */
8743 tp->t_rttvar += (delta >> 3);
8744 if (tp->t_rttvar <= 0)
8745 tp->t_rttvar = 1;
8746 } else {
8747 /*
8748 * No rtt measurement yet - use the unsmoothed rtt. Set the
8749 * variance to half the rtt (so our first retransmit happens
8750 * at 3*rtt).
8751 */
8752 tp->t_srtt = rtt;
8753 tp->t_rttvar = rtt >> 1;
8754 }
8755 rack->rc_srtt_measure_made = 1;
8756 KMOD_TCPSTAT_INC(tcps_rttupdated);
8757 if (tp->t_rttupdated < UCHAR_MAX)
8758 tp->t_rttupdated++;
8759 #ifdef STATS
8760 if (rack_stats_gets_ms_rtt == 0) {
8761 /* Send in the microsecond rtt used for rxt timeout purposes */
8762 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8763 } else if (rack_stats_gets_ms_rtt == 1) {
8764 /* Send in the millisecond rtt used for rxt timeout purposes */
8765 int32_t ms_rtt;
8766
8767 /* Round up */
8768 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8769 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8770 } else if (rack_stats_gets_ms_rtt == 2) {
8771 /* Send in the millisecond rtt has close to the path RTT as we can get */
8772 int32_t ms_rtt;
8773
8774 /* Round up */
8775 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8776 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8777 } else {
8778 /* Send in the microsecond rtt has close to the path RTT as we can get */
8779 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8780 }
8781 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8782 #endif
8783 rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_msec(&rack->r_ctl.act_rcv_time);
8784 /*
8785 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8786 * way we do the smoothing, srtt and rttvar will each average +1/2
8787 * tick of bias. When we compute the retransmit timer, we want 1/2
8788 * tick of rounding and 1 extra tick because of +-1/2 tick
8789 * uncertainty in the firing of the timer. The bias will give us
8790 * exactly the 1.5 tick we need. But, because the bias is
8791 * statistical, we have to test that we don't drop below the minimum
8792 * feasible timer (which is 2 ticks).
8793 */
8794 tp->t_rxtshift = 0;
8795 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8796 max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8797 rack_log_rtt_sample(rack, rtt);
8798 tp->t_softerror = 0;
8799 }
8800
8801
8802 static void
rack_apply_updated_usrtt(struct tcp_rack * rack,uint32_t us_rtt,uint32_t us_cts)8803 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8804 {
8805 /*
8806 * Apply to filter the inbound us-rtt at us_cts.
8807 */
8808 uint32_t old_rtt;
8809
8810 old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8811 apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8812 us_rtt, us_cts);
8813 if (old_rtt > us_rtt) {
8814 /* We just hit a new lower rtt time */
8815 rack_log_rtt_shrinks(rack, us_cts, old_rtt,
8816 __LINE__, RACK_RTTS_NEWRTT);
8817 /*
8818 * Only count it if its lower than what we saw within our
8819 * calculated range.
8820 */
8821 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8822 if (rack_probertt_lower_within &&
8823 rack->rc_gp_dyn_mul &&
8824 (rack->use_fixed_rate == 0) &&
8825 (rack->rc_always_pace)) {
8826 /*
8827 * We are seeing a new lower rtt very close
8828 * to the time that we would have entered probe-rtt.
8829 * This is probably due to the fact that a peer flow
8830 * has entered probe-rtt. Lets go in now too.
8831 */
8832 uint32_t val;
8833
8834 val = rack_probertt_lower_within * rack_time_between_probertt;
8835 val /= 100;
8836 if ((rack->in_probe_rtt == 0) &&
8837 (rack->rc_skip_timely == 0) &&
8838 ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
8839 rack_enter_probertt(rack, us_cts);
8840 }
8841 }
8842 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8843 }
8844 }
8845 }
8846
8847 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)8848 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8849 struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8850 {
8851 uint32_t us_rtt;
8852 int32_t i, all;
8853 uint32_t t, len_acked;
8854
8855 if ((rsm->r_flags & RACK_ACKED) ||
8856 (rsm->r_flags & RACK_WAS_ACKED))
8857 /* Already done */
8858 return (0);
8859 if (rsm->r_no_rtt_allowed) {
8860 /* Not allowed */
8861 return (0);
8862 }
8863 if (ack_type == CUM_ACKED) {
8864 if (SEQ_GT(th_ack, rsm->r_end)) {
8865 len_acked = rsm->r_end - rsm->r_start;
8866 all = 1;
8867 } else {
8868 len_acked = th_ack - rsm->r_start;
8869 all = 0;
8870 }
8871 } else {
8872 len_acked = rsm->r_end - rsm->r_start;
8873 all = 0;
8874 }
8875 if (rsm->r_rtr_cnt == 1) {
8876
8877 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8878 if ((int)t <= 0)
8879 t = 1;
8880 if (!tp->t_rttlow || tp->t_rttlow > t)
8881 tp->t_rttlow = t;
8882 if (!rack->r_ctl.rc_rack_min_rtt ||
8883 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8884 rack->r_ctl.rc_rack_min_rtt = t;
8885 if (rack->r_ctl.rc_rack_min_rtt == 0) {
8886 rack->r_ctl.rc_rack_min_rtt = 1;
8887 }
8888 }
8889 if (TSTMP_GT(tcp_tv_to_usec(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8890 us_rtt = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8891 else
8892 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8893 if (us_rtt == 0)
8894 us_rtt = 1;
8895 if (CC_ALGO(tp)->rttsample != NULL) {
8896 /* Kick the RTT to the CC */
8897 CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8898 }
8899 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usec(&rack->r_ctl.act_rcv_time));
8900 if (ack_type == SACKED) {
8901 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8902 tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8903 } else {
8904 /*
8905 * We need to setup what our confidence
8906 * is in this ack.
8907 *
8908 * If the rsm was app limited and it is
8909 * less than a mss in length (the end
8910 * of the send) then we have a gap. If we
8911 * were app limited but say we were sending
8912 * multiple MSS's then we are more confident
8913 * int it.
8914 *
8915 * When we are not app-limited then we see if
8916 * the rsm is being included in the current
8917 * measurement, we tell this by the app_limited_needs_set
8918 * flag.
8919 *
8920 * Note that being cwnd blocked is not applimited
8921 * as well as the pacing delay between packets which
8922 * are sending only 1 or 2 MSS's also will show up
8923 * in the RTT. We probably need to examine this algorithm
8924 * a bit more and enhance it to account for the delay
8925 * between rsm's. We could do that by saving off the
8926 * pacing delay of each rsm (in an rsm) and then
8927 * factoring that in somehow though for now I am
8928 * not sure how :)
8929 */
8930 int calc_conf = 0;
8931
8932 if (rsm->r_flags & RACK_APP_LIMITED) {
8933 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8934 calc_conf = 0;
8935 else
8936 calc_conf = 1;
8937 } else if (rack->app_limited_needs_set == 0) {
8938 calc_conf = 1;
8939 } else {
8940 calc_conf = 0;
8941 }
8942 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8943 tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8944 calc_conf, rsm, rsm->r_rtr_cnt);
8945 }
8946 if ((rsm->r_flags & RACK_TLP) &&
8947 (!IN_FASTRECOVERY(tp->t_flags))) {
8948 /* Segment was a TLP and our retrans matched */
8949 if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8950 rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
8951 }
8952 }
8953 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
8954 (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8955 (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
8956 /* New more recent rack_tmit_time */
8957 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8958 if (rack->r_ctl.rc_rack_tmit_time == 0)
8959 rack->r_ctl.rc_rack_tmit_time = 1;
8960 rack->rc_rack_rtt = t;
8961 }
8962 return (1);
8963 }
8964 /*
8965 * We clear the soft/rxtshift since we got an ack.
8966 * There is no assurance we will call the commit() function
8967 * so we need to clear these to avoid incorrect handling.
8968 */
8969 tp->t_rxtshift = 0;
8970 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8971 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8972 tp->t_softerror = 0;
8973 if (to && (to->to_flags & TOF_TS) &&
8974 (ack_type == CUM_ACKED) &&
8975 (to->to_tsecr) &&
8976 ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8977 /*
8978 * Now which timestamp does it match? In this block the ACK
8979 * must be coming from a previous transmission.
8980 */
8981 for (i = 0; i < rsm->r_rtr_cnt; i++) {
8982 if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8983 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8984 if ((int)t <= 0)
8985 t = 1;
8986 if (CC_ALGO(tp)->rttsample != NULL) {
8987 /*
8988 * Kick the RTT to the CC, here
8989 * we lie a bit in that we know the
8990 * retransmission is correct even though
8991 * we retransmitted. This is because
8992 * we match the timestamps.
8993 */
8994 if (TSTMP_GT(tcp_tv_to_usec(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8995 us_rtt = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8996 else
8997 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8998 CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
8999 }
9000 if ((i + 1) < rsm->r_rtr_cnt) {
9001 /*
9002 * The peer ack'd from our previous
9003 * transmission. We have a spurious
9004 * retransmission and thus we dont
9005 * want to update our rack_rtt.
9006 *
9007 * Hmm should there be a CC revert here?
9008 *
9009 */
9010 return (0);
9011 }
9012 if (!tp->t_rttlow || tp->t_rttlow > t)
9013 tp->t_rttlow = t;
9014 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9015 rack->r_ctl.rc_rack_min_rtt = t;
9016 if (rack->r_ctl.rc_rack_min_rtt == 0) {
9017 rack->r_ctl.rc_rack_min_rtt = 1;
9018 }
9019 }
9020 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9021 (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9022 (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9023 /* New more recent rack_tmit_time */
9024 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9025 if (rack->r_ctl.rc_rack_tmit_time == 0)
9026 rack->r_ctl.rc_rack_tmit_time = 1;
9027 rack->rc_rack_rtt = t;
9028 }
9029 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9030 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9031 rsm->r_rtr_cnt);
9032 return (1);
9033 }
9034 }
9035 /* If we are logging log out the sendmap */
9036 if (tcp_bblogging_on(rack->rc_tp)) {
9037 for (i = 0; i < rsm->r_rtr_cnt; i++) {
9038 rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9039 }
9040 }
9041 goto ts_not_found;
9042 } else {
9043 /*
9044 * Ok its a SACK block that we retransmitted. or a windows
9045 * machine without timestamps. We can tell nothing from the
9046 * time-stamp since its not there or the time the peer last
9047 * received a segment that moved forward its cum-ack point.
9048 */
9049 ts_not_found:
9050 i = rsm->r_rtr_cnt - 1;
9051 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9052 if ((int)t <= 0)
9053 t = 1;
9054 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9055 /*
9056 * We retransmitted and the ack came back in less
9057 * than the smallest rtt we have observed. We most
9058 * likely did an improper retransmit as outlined in
9059 * 6.2 Step 2 point 2 in the rack-draft so we
9060 * don't want to update our rack_rtt. We in
9061 * theory (in future) might want to think about reverting our
9062 * cwnd state but we won't for now.
9063 */
9064 return (0);
9065 } else if (rack->r_ctl.rc_rack_min_rtt) {
9066 /*
9067 * We retransmitted it and the retransmit did the
9068 * job.
9069 */
9070 if (!rack->r_ctl.rc_rack_min_rtt ||
9071 SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9072 rack->r_ctl.rc_rack_min_rtt = t;
9073 if (rack->r_ctl.rc_rack_min_rtt == 0) {
9074 rack->r_ctl.rc_rack_min_rtt = 1;
9075 }
9076 }
9077 if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9078 (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9079 (uint32_t)rsm->r_tim_lastsent[i]))) {
9080 /* New more recent rack_tmit_time */
9081 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9082 if (rack->r_ctl.rc_rack_tmit_time == 0)
9083 rack->r_ctl.rc_rack_tmit_time = 1;
9084 rack->rc_rack_rtt = t;
9085 }
9086 return (1);
9087 }
9088 }
9089 return (0);
9090 }
9091
9092 /*
9093 * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9094 */
9095 static void
rack_log_sack_passed(struct tcpcb * tp,struct tcp_rack * rack,struct rack_sendmap * rsm,uint32_t cts)9096 rack_log_sack_passed(struct tcpcb *tp,
9097 struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t cts)
9098 {
9099 struct rack_sendmap *nrsm;
9100 uint32_t thresh;
9101
9102 /* Get our rxt threshold for lost consideration */
9103 thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(tp, rack), cts, __LINE__, 0);
9104 /* Now start looking at rsm's */
9105 nrsm = rsm;
9106 TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9107 rack_head, r_tnext) {
9108 if (nrsm == rsm) {
9109 /* Skip original segment he is acked */
9110 continue;
9111 }
9112 if (nrsm->r_flags & RACK_ACKED) {
9113 /*
9114 * Skip ack'd segments, though we
9115 * should not see these, since tmap
9116 * should not have ack'd segments.
9117 */
9118 continue;
9119 }
9120 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9121 /*
9122 * If the peer dropped the rwnd on
9123 * these then we don't worry about them.
9124 */
9125 continue;
9126 }
9127 /* Check lost state */
9128 if ((nrsm->r_flags & RACK_WAS_LOST) == 0) {
9129 uint32_t exp;
9130
9131 exp = ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) + thresh;
9132 if (TSTMP_LT(exp, cts) || (exp == cts)) {
9133 /* We consider it lost */
9134 nrsm->r_flags |= RACK_WAS_LOST;
9135 rack->r_ctl.rc_considered_lost += nrsm->r_end - nrsm->r_start;
9136 }
9137 }
9138 if (nrsm->r_flags & RACK_SACK_PASSED) {
9139 /*
9140 * We found one that is already marked
9141 * passed, we have been here before and
9142 * so all others below this are marked.
9143 */
9144 break;
9145 }
9146 nrsm->r_flags |= RACK_SACK_PASSED;
9147 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9148 }
9149 }
9150
9151 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)9152 rack_need_set_test(struct tcpcb *tp,
9153 struct tcp_rack *rack,
9154 struct rack_sendmap *rsm,
9155 tcp_seq th_ack,
9156 int line,
9157 int use_which)
9158 {
9159 struct rack_sendmap *s_rsm;
9160
9161 if ((tp->t_flags & TF_GPUTINPROG) &&
9162 SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9163 /*
9164 * We were app limited, and this ack
9165 * butts up or goes beyond the point where we want
9166 * to start our next measurement. We need
9167 * to record the new gput_ts as here and
9168 * possibly update the start sequence.
9169 */
9170 uint32_t seq, ts;
9171
9172 if (rsm->r_rtr_cnt > 1) {
9173 /*
9174 * This is a retransmit, can we
9175 * really make any assessment at this
9176 * point? We are not really sure of
9177 * the timestamp, is it this or the
9178 * previous transmission?
9179 *
9180 * Lets wait for something better that
9181 * is not retransmitted.
9182 */
9183 return;
9184 }
9185 seq = tp->gput_seq;
9186 ts = tp->gput_ts;
9187 rack->app_limited_needs_set = 0;
9188 tp->gput_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
9189 /* Do we start at a new end? */
9190 if ((use_which == RACK_USE_BEG) &&
9191 SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9192 /*
9193 * When we get an ACK that just eats
9194 * up some of the rsm, we set RACK_USE_BEG
9195 * since whats at r_start (i.e. th_ack)
9196 * is left unacked and thats where the
9197 * measurement now starts.
9198 */
9199 tp->gput_seq = rsm->r_start;
9200 }
9201 if ((use_which == RACK_USE_END) &&
9202 SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9203 /*
9204 * We use the end when the cumack
9205 * is moving forward and completely
9206 * deleting the rsm passed so basically
9207 * r_end holds th_ack.
9208 *
9209 * For SACK's we also want to use the end
9210 * since this piece just got sacked and
9211 * we want to target anything after that
9212 * in our measurement.
9213 */
9214 tp->gput_seq = rsm->r_end;
9215 }
9216 if (use_which == RACK_USE_END_OR_THACK) {
9217 /*
9218 * special case for ack moving forward,
9219 * not a sack, we need to move all the
9220 * way up to where this ack cum-ack moves
9221 * to.
9222 */
9223 if (SEQ_GT(th_ack, rsm->r_end))
9224 tp->gput_seq = th_ack;
9225 else
9226 tp->gput_seq = rsm->r_end;
9227 }
9228 if (SEQ_LT(tp->gput_seq, tp->snd_max))
9229 s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9230 else
9231 s_rsm = NULL;
9232 /*
9233 * Pick up the correct send time if we can the rsm passed in
9234 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9235 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9236 * find a different seq i.e. the next send up.
9237 *
9238 * If that has not been sent, s_rsm will be NULL and we must
9239 * arrange it so this function will get called again by setting
9240 * app_limited_needs_set.
9241 */
9242 if (s_rsm)
9243 rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9244 else {
9245 /* If we hit here we have to have *not* sent tp->gput_seq */
9246 rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9247 /* Set it up so we will go through here again */
9248 rack->app_limited_needs_set = 1;
9249 }
9250 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9251 /*
9252 * We moved beyond this guy's range, re-calculate
9253 * the new end point.
9254 */
9255 if (rack->rc_gp_filled == 0) {
9256 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9257 } else {
9258 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9259 }
9260 }
9261 /*
9262 * We are moving the goal post, we may be able to clear the
9263 * measure_saw_probe_rtt flag.
9264 */
9265 if ((rack->in_probe_rtt == 0) &&
9266 (rack->measure_saw_probe_rtt) &&
9267 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9268 rack->measure_saw_probe_rtt = 0;
9269 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9270 seq, tp->gput_seq,
9271 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9272 (uint64_t)rack->r_ctl.rc_gp_output_ts),
9273 5, line, NULL, 0);
9274 if (rack->rc_gp_filled &&
9275 ((tp->gput_ack - tp->gput_seq) <
9276 max(rc_init_window(rack), (MIN_GP_WIN *
9277 ctf_fixed_maxseg(tp))))) {
9278 uint32_t ideal_amount;
9279
9280 ideal_amount = rack_get_measure_window(tp, rack);
9281 if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9282 /*
9283 * There is no sense of continuing this measurement
9284 * because its too small to gain us anything we
9285 * trust. Skip it and that way we can start a new
9286 * measurement quicker.
9287 */
9288 tp->t_flags &= ~TF_GPUTINPROG;
9289 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9290 0, 0,
9291 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9292 (uint64_t)rack->r_ctl.rc_gp_output_ts),
9293 6, __LINE__, NULL, 0);
9294 } else {
9295 /*
9296 * Reset the window further out.
9297 */
9298 tp->gput_ack = tp->gput_seq + ideal_amount;
9299 }
9300 }
9301 rack_tend_gp_marks(tp, rack);
9302 rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9303 }
9304 }
9305
9306 static inline int
is_rsm_inside_declared_tlp_block(struct tcp_rack * rack,struct rack_sendmap * rsm)9307 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9308 {
9309 if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9310 /* Behind our TLP definition or right at */
9311 return (0);
9312 }
9313 if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9314 /* The start is beyond or right at our end of TLP definition */
9315 return (0);
9316 }
9317 /* It has to be a sub-part of the original TLP recorded */
9318 return (1);
9319 }
9320
9321 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)9322 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9323 struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9324 uint32_t segsiz)
9325 {
9326 uint32_t start, end, changed = 0;
9327 struct rack_sendmap stack_map;
9328 struct rack_sendmap *rsm, *nrsm, *prev, *next;
9329 int insret __diagused;
9330 int32_t used_ref = 1;
9331 int can_use_hookery = 0;
9332
9333 start = sack->start;
9334 end = sack->end;
9335 rsm = *prsm;
9336
9337 do_rest_ofb:
9338 if ((rsm == NULL) ||
9339 (SEQ_LT(end, rsm->r_start)) ||
9340 (SEQ_GEQ(start, rsm->r_end)) ||
9341 (SEQ_LT(start, rsm->r_start))) {
9342 /*
9343 * We are not in the right spot,
9344 * find the correct spot in the tree.
9345 */
9346 used_ref = 0;
9347 rsm = tqhash_find(rack->r_ctl.tqh, start);
9348 }
9349 if (rsm == NULL) {
9350 /* TSNH */
9351 goto out;
9352 }
9353 /* Ok we have an ACK for some piece of this rsm */
9354 if (rsm->r_start != start) {
9355 if ((rsm->r_flags & RACK_ACKED) == 0) {
9356 /*
9357 * Before any splitting or hookery is
9358 * done is it a TLP of interest i.e. rxt?
9359 */
9360 if ((rsm->r_flags & RACK_TLP) &&
9361 (rsm->r_rtr_cnt > 1)) {
9362 /*
9363 * We are splitting a rxt TLP, check
9364 * if we need to save off the start/end
9365 */
9366 if (rack->rc_last_tlp_acked_set &&
9367 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9368 /*
9369 * We already turned this on since we are inside
9370 * the previous one was a partially sack now we
9371 * are getting another one (maybe all of it).
9372 *
9373 */
9374 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9375 /*
9376 * Lets make sure we have all of it though.
9377 */
9378 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9379 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9380 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9381 rack->r_ctl.last_tlp_acked_end);
9382 }
9383 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9384 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9385 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9386 rack->r_ctl.last_tlp_acked_end);
9387 }
9388 } else {
9389 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9390 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9391 rack->rc_last_tlp_past_cumack = 0;
9392 rack->rc_last_tlp_acked_set = 1;
9393 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9394 }
9395 }
9396 /**
9397 * Need to split this in two pieces the before and after,
9398 * the before remains in the map, the after must be
9399 * added. In other words we have:
9400 * rsm |--------------|
9401 * sackblk |------->
9402 * rsm will become
9403 * rsm |---|
9404 * and nrsm will be the sacked piece
9405 * nrsm |----------|
9406 *
9407 * But before we start down that path lets
9408 * see if the sack spans over on top of
9409 * the next guy and it is already sacked.
9410 *
9411 */
9412 /*
9413 * Hookery can only be used if the two entries
9414 * are in the same bucket and neither one of
9415 * them staddle the bucket line.
9416 */
9417 next = tqhash_next(rack->r_ctl.tqh, rsm);
9418 if (next &&
9419 (rsm->bindex == next->bindex) &&
9420 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9421 ((next->r_flags & RACK_STRADDLE) == 0) &&
9422 ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9423 ((next->r_flags & RACK_IS_PCM) == 0) &&
9424 (rsm->r_flags & RACK_IN_GP_WIN) &&
9425 (next->r_flags & RACK_IN_GP_WIN))
9426 can_use_hookery = 1;
9427 else
9428 can_use_hookery = 0;
9429 if (next && can_use_hookery &&
9430 (next->r_flags & RACK_ACKED) &&
9431 SEQ_GEQ(end, next->r_start)) {
9432 /**
9433 * So the next one is already acked, and
9434 * we can thus by hookery use our stack_map
9435 * to reflect the piece being sacked and
9436 * then adjust the two tree entries moving
9437 * the start and ends around. So we start like:
9438 * rsm |------------| (not-acked)
9439 * next |-----------| (acked)
9440 * sackblk |-------->
9441 * We want to end like so:
9442 * rsm |------| (not-acked)
9443 * next |-----------------| (acked)
9444 * nrsm |-----|
9445 * Where nrsm is a temporary stack piece we
9446 * use to update all the gizmos.
9447 */
9448 /* Copy up our fudge block */
9449 nrsm = &stack_map;
9450 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9451 /* Now adjust our tree blocks */
9452 tqhash_update_end(rack->r_ctl.tqh, rsm, start);
9453 next->r_start = start;
9454 rsm->r_flags |= RACK_SHUFFLED;
9455 next->r_flags |= RACK_SHUFFLED;
9456 /* Now we must adjust back where next->m is */
9457 rack_setup_offset_for_rsm(rack, rsm, next);
9458 /*
9459 * Which timestamp do we keep? It is rather
9460 * important in GP measurements to have the
9461 * accurate end of the send window.
9462 *
9463 * We keep the largest value, which is the newest
9464 * send. We do this in case a segment that is
9465 * joined together and not part of a GP estimate
9466 * later gets expanded into the GP estimate.
9467 *
9468 * We prohibit the merging of unlike kinds i.e.
9469 * all pieces that are in the GP estimate can be
9470 * merged and all pieces that are not in a GP estimate
9471 * can be merged, but not disimilar pieces. Combine
9472 * this with taking the highest here and we should
9473 * be ok unless of course the client reneges. Then
9474 * all bets are off.
9475 */
9476 if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9477 nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9478 next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9479 /*
9480 * And we must keep the newest ack arrival time.
9481 */
9482 if (next->r_ack_arrival <
9483 rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9484 next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9485
9486
9487 /* We don't need to adjust rsm, it did not change */
9488 /* Clear out the dup ack count of the remainder */
9489 rsm->r_dupack = 0;
9490 rsm->r_just_ret = 0;
9491 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9492 /* Now lets make sure our fudge block is right */
9493 nrsm->r_start = start;
9494 /* Now lets update all the stats and such */
9495 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9496 if (rack->app_limited_needs_set)
9497 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9498 changed += (nrsm->r_end - nrsm->r_start);
9499 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9500 if (rsm->r_flags & RACK_WAS_LOST) {
9501 int my_chg;
9502
9503 my_chg = (nrsm->r_end - nrsm->r_start);
9504 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9505 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9506 if (my_chg <= rack->r_ctl.rc_considered_lost)
9507 rack->r_ctl.rc_considered_lost -= my_chg;
9508 else
9509 rack->r_ctl.rc_considered_lost = 0;
9510 }
9511 if (nrsm->r_flags & RACK_SACK_PASSED) {
9512 rack->r_ctl.rc_reorder_ts = cts;
9513 if (rack->r_ctl.rc_reorder_ts == 0)
9514 rack->r_ctl.rc_reorder_ts = 1;
9515 }
9516 /*
9517 * Now we want to go up from rsm (the
9518 * one left un-acked) to the next one
9519 * in the tmap. We do this so when
9520 * we walk backwards we include marking
9521 * sack-passed on rsm (The one passed in
9522 * is skipped since it is generally called
9523 * on something sacked before removing it
9524 * from the tmap).
9525 */
9526 if (rsm->r_in_tmap) {
9527 nrsm = TAILQ_NEXT(rsm, r_tnext);
9528 /*
9529 * Now that we have the next
9530 * one walk backwards from there.
9531 */
9532 if (nrsm && nrsm->r_in_tmap)
9533 rack_log_sack_passed(tp, rack, nrsm, cts);
9534 }
9535 /* Now are we done? */
9536 if (SEQ_LT(end, next->r_end) ||
9537 (end == next->r_end)) {
9538 /* Done with block */
9539 goto out;
9540 }
9541 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9542 counter_u64_add(rack_sack_used_next_merge, 1);
9543 /* Postion for the next block */
9544 start = next->r_end;
9545 rsm = tqhash_next(rack->r_ctl.tqh, next);
9546 if (rsm == NULL)
9547 goto out;
9548 } else {
9549 /**
9550 * We can't use any hookery here, so we
9551 * need to split the map. We enter like
9552 * so:
9553 * rsm |--------|
9554 * sackblk |----->
9555 * We will add the new block nrsm and
9556 * that will be the new portion, and then
9557 * fall through after reseting rsm. So we
9558 * split and look like this:
9559 * rsm |----|
9560 * sackblk |----->
9561 * nrsm |---|
9562 * We then fall through reseting
9563 * rsm to nrsm, so the next block
9564 * picks it up.
9565 */
9566 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9567 if (nrsm == NULL) {
9568 /*
9569 * failed XXXrrs what can we do but loose the sack
9570 * info?
9571 */
9572 goto out;
9573 }
9574 counter_u64_add(rack_sack_splits, 1);
9575 rack_clone_rsm(rack, nrsm, rsm, start);
9576 rsm->r_just_ret = 0;
9577 #ifndef INVARIANTS
9578 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9579 #else
9580 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9581 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9582 nrsm, insret, rack, rsm);
9583 }
9584 #endif
9585 if (rsm->r_in_tmap) {
9586 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9587 nrsm->r_in_tmap = 1;
9588 }
9589 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9590 rsm->r_flags &= (~RACK_HAS_FIN);
9591 /* Position us to point to the new nrsm that starts the sack blk */
9592 rsm = nrsm;
9593 }
9594 } else {
9595 /* Already sacked this piece */
9596 counter_u64_add(rack_sack_skipped_acked, 1);
9597 if (end == rsm->r_end) {
9598 /* Done with block */
9599 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9600 goto out;
9601 } else if (SEQ_LT(end, rsm->r_end)) {
9602 /* A partial sack to a already sacked block */
9603 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9604 goto out;
9605 } else {
9606 /*
9607 * The end goes beyond this guy
9608 * reposition the start to the
9609 * next block.
9610 */
9611 start = rsm->r_end;
9612 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9613 if (rsm == NULL)
9614 goto out;
9615 }
9616 }
9617 }
9618 if (SEQ_GEQ(end, rsm->r_end)) {
9619 /**
9620 * The end of this block is either beyond this guy or right
9621 * at this guy. I.e.:
9622 * rsm --- |-----|
9623 * end |-----|
9624 * <or>
9625 * end |---------|
9626 */
9627 if ((rsm->r_flags & RACK_ACKED) == 0) {
9628 /*
9629 * Is it a TLP of interest?
9630 */
9631 if ((rsm->r_flags & RACK_TLP) &&
9632 (rsm->r_rtr_cnt > 1)) {
9633 /*
9634 * We are splitting a rxt TLP, check
9635 * if we need to save off the start/end
9636 */
9637 if (rack->rc_last_tlp_acked_set &&
9638 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9639 /*
9640 * We already turned this on since we are inside
9641 * the previous one was a partially sack now we
9642 * are getting another one (maybe all of it).
9643 */
9644 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9645 /*
9646 * Lets make sure we have all of it though.
9647 */
9648 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9649 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9650 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9651 rack->r_ctl.last_tlp_acked_end);
9652 }
9653 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9654 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9655 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9656 rack->r_ctl.last_tlp_acked_end);
9657 }
9658 } else {
9659 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9660 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9661 rack->rc_last_tlp_past_cumack = 0;
9662 rack->rc_last_tlp_acked_set = 1;
9663 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9664 }
9665 }
9666 rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9667 changed += (rsm->r_end - rsm->r_start);
9668 /* You get a count for acking a whole segment or more */
9669 if (rsm->r_flags & RACK_WAS_LOST) {
9670 int my_chg;
9671
9672 my_chg = (rsm->r_end - rsm->r_start);
9673 rsm->r_flags &= ~RACK_WAS_LOST;
9674 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9675 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9676 if (my_chg <= rack->r_ctl.rc_considered_lost)
9677 rack->r_ctl.rc_considered_lost -= my_chg;
9678 else
9679 rack->r_ctl.rc_considered_lost = 0;
9680 }
9681 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9682 if (rsm->r_in_tmap) /* should be true */
9683 rack_log_sack_passed(tp, rack, rsm, cts);
9684 /* Is Reordering occuring? */
9685 if (rsm->r_flags & RACK_SACK_PASSED) {
9686 rsm->r_flags &= ~RACK_SACK_PASSED;
9687 rack->r_ctl.rc_reorder_ts = cts;
9688 if (rack->r_ctl.rc_reorder_ts == 0)
9689 rack->r_ctl.rc_reorder_ts = 1;
9690 }
9691 if (rack->app_limited_needs_set)
9692 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9693 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9694 rsm->r_flags |= RACK_ACKED;
9695 rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9696 if (rsm->r_in_tmap) {
9697 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9698 rsm->r_in_tmap = 0;
9699 }
9700 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9701 } else {
9702 counter_u64_add(rack_sack_skipped_acked, 1);
9703 }
9704 if (end == rsm->r_end) {
9705 /* This block only - done, setup for next */
9706 goto out;
9707 }
9708 /*
9709 * There is more not coverend by this rsm move on
9710 * to the next block in the tail queue hash table.
9711 */
9712 nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9713 start = rsm->r_end;
9714 rsm = nrsm;
9715 if (rsm == NULL)
9716 goto out;
9717 goto do_rest_ofb;
9718 }
9719 /**
9720 * The end of this sack block is smaller than
9721 * our rsm i.e.:
9722 * rsm --- |-----|
9723 * end |--|
9724 */
9725 if ((rsm->r_flags & RACK_ACKED) == 0) {
9726 /*
9727 * Is it a TLP of interest?
9728 */
9729 if ((rsm->r_flags & RACK_TLP) &&
9730 (rsm->r_rtr_cnt > 1)) {
9731 /*
9732 * We are splitting a rxt TLP, check
9733 * if we need to save off the start/end
9734 */
9735 if (rack->rc_last_tlp_acked_set &&
9736 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9737 /*
9738 * We already turned this on since we are inside
9739 * the previous one was a partially sack now we
9740 * are getting another one (maybe all of it).
9741 */
9742 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9743 /*
9744 * Lets make sure we have all of it though.
9745 */
9746 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9747 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9748 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9749 rack->r_ctl.last_tlp_acked_end);
9750 }
9751 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9752 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9753 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9754 rack->r_ctl.last_tlp_acked_end);
9755 }
9756 } else {
9757 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9758 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9759 rack->rc_last_tlp_past_cumack = 0;
9760 rack->rc_last_tlp_acked_set = 1;
9761 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9762 }
9763 }
9764 /*
9765 * Hookery can only be used if the two entries
9766 * are in the same bucket and neither one of
9767 * them staddle the bucket line.
9768 */
9769 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9770 if (prev &&
9771 (rsm->bindex == prev->bindex) &&
9772 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9773 ((prev->r_flags & RACK_STRADDLE) == 0) &&
9774 ((rsm->r_flags & RACK_IS_PCM) == 0) &&
9775 ((prev->r_flags & RACK_IS_PCM) == 0) &&
9776 (rsm->r_flags & RACK_IN_GP_WIN) &&
9777 (prev->r_flags & RACK_IN_GP_WIN))
9778 can_use_hookery = 1;
9779 else
9780 can_use_hookery = 0;
9781 if (prev && can_use_hookery &&
9782 (prev->r_flags & RACK_ACKED)) {
9783 /**
9784 * Goal, we want the right remainder of rsm to shrink
9785 * in place and span from (rsm->r_start = end) to rsm->r_end.
9786 * We want to expand prev to go all the way
9787 * to prev->r_end <- end.
9788 * so in the tree we have before:
9789 * prev |--------| (acked)
9790 * rsm |-------| (non-acked)
9791 * sackblk |-|
9792 * We churn it so we end up with
9793 * prev |----------| (acked)
9794 * rsm |-----| (non-acked)
9795 * nrsm |-| (temporary)
9796 *
9797 * Note if either prev/rsm is a TLP we don't
9798 * do this.
9799 */
9800 nrsm = &stack_map;
9801 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9802 tqhash_update_end(rack->r_ctl.tqh, prev, end);
9803 rsm->r_start = end;
9804 rsm->r_flags |= RACK_SHUFFLED;
9805 prev->r_flags |= RACK_SHUFFLED;
9806 /* Now adjust nrsm (stack copy) to be
9807 * the one that is the small
9808 * piece that was "sacked".
9809 */
9810 nrsm->r_end = end;
9811 rsm->r_dupack = 0;
9812 /*
9813 * Which timestamp do we keep? It is rather
9814 * important in GP measurements to have the
9815 * accurate end of the send window.
9816 *
9817 * We keep the largest value, which is the newest
9818 * send. We do this in case a segment that is
9819 * joined together and not part of a GP estimate
9820 * later gets expanded into the GP estimate.
9821 *
9822 * We prohibit the merging of unlike kinds i.e.
9823 * all pieces that are in the GP estimate can be
9824 * merged and all pieces that are not in a GP estimate
9825 * can be merged, but not disimilar pieces. Combine
9826 * this with taking the highest here and we should
9827 * be ok unless of course the client reneges. Then
9828 * all bets are off.
9829 */
9830 if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
9831 nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
9832 prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9833 }
9834 /*
9835 * And we must keep the newest ack arrival time.
9836 */
9837
9838 if(prev->r_ack_arrival <
9839 rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9840 prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9841
9842 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9843 /*
9844 * Now that the rsm has had its start moved forward
9845 * lets go ahead and get its new place in the world.
9846 */
9847 rack_setup_offset_for_rsm(rack, prev, rsm);
9848 /*
9849 * Now nrsm is our new little piece
9850 * that is acked (which was merged
9851 * to prev). Update the rtt and changed
9852 * based on that. Also check for reordering.
9853 */
9854 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9855 if (rack->app_limited_needs_set)
9856 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9857 changed += (nrsm->r_end - nrsm->r_start);
9858 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9859 if (rsm->r_flags & RACK_WAS_LOST) {
9860 int my_chg;
9861
9862 my_chg = (nrsm->r_end - nrsm->r_start);
9863 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9864 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9865 if (my_chg <= rack->r_ctl.rc_considered_lost)
9866 rack->r_ctl.rc_considered_lost -= my_chg;
9867 else
9868 rack->r_ctl.rc_considered_lost = 0;
9869 }
9870 if (nrsm->r_flags & RACK_SACK_PASSED) {
9871 rack->r_ctl.rc_reorder_ts = cts;
9872 if (rack->r_ctl.rc_reorder_ts == 0)
9873 rack->r_ctl.rc_reorder_ts = 1;
9874 }
9875 rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
9876 rsm = prev;
9877 counter_u64_add(rack_sack_used_prev_merge, 1);
9878 } else {
9879 /**
9880 * This is the case where our previous
9881 * block is not acked either, so we must
9882 * split the block in two.
9883 */
9884 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9885 if (nrsm == NULL) {
9886 /* failed rrs what can we do but loose the sack info? */
9887 goto out;
9888 }
9889 if ((rsm->r_flags & RACK_TLP) &&
9890 (rsm->r_rtr_cnt > 1)) {
9891 /*
9892 * We are splitting a rxt TLP, check
9893 * if we need to save off the start/end
9894 */
9895 if (rack->rc_last_tlp_acked_set &&
9896 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9897 /*
9898 * We already turned this on since this block is inside
9899 * the previous one was a partially sack now we
9900 * are getting another one (maybe all of it).
9901 */
9902 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9903 /*
9904 * Lets make sure we have all of it though.
9905 */
9906 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9907 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9908 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9909 rack->r_ctl.last_tlp_acked_end);
9910 }
9911 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9912 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9913 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9914 rack->r_ctl.last_tlp_acked_end);
9915 }
9916 } else {
9917 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9918 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9919 rack->rc_last_tlp_acked_set = 1;
9920 rack->rc_last_tlp_past_cumack = 0;
9921 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9922 }
9923 }
9924 /**
9925 * In this case nrsm becomes
9926 * nrsm->r_start = end;
9927 * nrsm->r_end = rsm->r_end;
9928 * which is un-acked.
9929 * <and>
9930 * rsm->r_end = nrsm->r_start;
9931 * i.e. the remaining un-acked
9932 * piece is left on the left
9933 * hand side.
9934 *
9935 * So we start like this
9936 * rsm |----------| (not acked)
9937 * sackblk |---|
9938 * build it so we have
9939 * rsm |---| (acked)
9940 * nrsm |------| (not acked)
9941 */
9942 counter_u64_add(rack_sack_splits, 1);
9943 rack_clone_rsm(rack, nrsm, rsm, end);
9944 rsm->r_flags &= (~RACK_HAS_FIN);
9945 rsm->r_just_ret = 0;
9946 #ifndef INVARIANTS
9947 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9948 #else
9949 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9950 panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
9951 nrsm, insret, rack, rsm);
9952 }
9953 #endif
9954 if (rsm->r_in_tmap) {
9955 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9956 nrsm->r_in_tmap = 1;
9957 }
9958 nrsm->r_dupack = 0;
9959 rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
9960 rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9961 changed += (rsm->r_end - rsm->r_start);
9962 if (rsm->r_flags & RACK_WAS_LOST) {
9963 int my_chg;
9964
9965 my_chg = (rsm->r_end - rsm->r_start);
9966 rsm->r_flags &= ~RACK_WAS_LOST;
9967 KASSERT((rack->r_ctl.rc_considered_lost >= my_chg),
9968 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
9969 if (my_chg <= rack->r_ctl.rc_considered_lost)
9970 rack->r_ctl.rc_considered_lost -= my_chg;
9971 else
9972 rack->r_ctl.rc_considered_lost = 0;
9973 }
9974 rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9975
9976 if (rsm->r_in_tmap) /* should be true */
9977 rack_log_sack_passed(tp, rack, rsm, cts);
9978 /* Is Reordering occuring? */
9979 if (rsm->r_flags & RACK_SACK_PASSED) {
9980 rsm->r_flags &= ~RACK_SACK_PASSED;
9981 rack->r_ctl.rc_reorder_ts = cts;
9982 if (rack->r_ctl.rc_reorder_ts == 0)
9983 rack->r_ctl.rc_reorder_ts = 1;
9984 }
9985 if (rack->app_limited_needs_set)
9986 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9987 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9988 rsm->r_flags |= RACK_ACKED;
9989 rack_update_pcm_ack(rack, 0, rsm->r_start, rsm->r_end);
9990 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
9991 if (rsm->r_in_tmap) {
9992 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9993 rsm->r_in_tmap = 0;
9994 }
9995 }
9996 } else if (start != end){
9997 /*
9998 * The block was already acked.
9999 */
10000 counter_u64_add(rack_sack_skipped_acked, 1);
10001 }
10002 out:
10003 if (rsm &&
10004 ((rsm->r_flags & RACK_TLP) == 0) &&
10005 (rsm->r_flags & RACK_ACKED)) {
10006 /*
10007 * Now can we merge where we worked
10008 * with either the previous or
10009 * next block?
10010 */
10011 next = tqhash_next(rack->r_ctl.tqh, rsm);
10012 while (next) {
10013 if (next->r_flags & RACK_TLP)
10014 break;
10015 /* Only allow merges between ones in or out of GP window */
10016 if ((next->r_flags & RACK_IN_GP_WIN) &&
10017 ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10018 break;
10019 }
10020 if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10021 ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10022 break;
10023 }
10024 if (rsm->bindex != next->bindex)
10025 break;
10026 if (rsm->r_flags & RACK_STRADDLE)
10027 break;
10028 if (rsm->r_flags & RACK_IS_PCM)
10029 break;
10030 if (next->r_flags & RACK_STRADDLE)
10031 break;
10032 if (next->r_flags & RACK_IS_PCM)
10033 break;
10034 if (next->r_flags & RACK_ACKED) {
10035 /* yep this and next can be merged */
10036 rsm = rack_merge_rsm(rack, rsm, next);
10037 next = tqhash_next(rack->r_ctl.tqh, rsm);
10038 } else
10039 break;
10040 }
10041 /* Now what about the previous? */
10042 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10043 while (prev) {
10044 if (prev->r_flags & RACK_TLP)
10045 break;
10046 /* Only allow merges between ones in or out of GP window */
10047 if ((prev->r_flags & RACK_IN_GP_WIN) &&
10048 ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10049 break;
10050 }
10051 if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10052 ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10053 break;
10054 }
10055 if (rsm->bindex != prev->bindex)
10056 break;
10057 if (rsm->r_flags & RACK_STRADDLE)
10058 break;
10059 if (rsm->r_flags & RACK_IS_PCM)
10060 break;
10061 if (prev->r_flags & RACK_STRADDLE)
10062 break;
10063 if (prev->r_flags & RACK_IS_PCM)
10064 break;
10065 if (prev->r_flags & RACK_ACKED) {
10066 /* yep the previous and this can be merged */
10067 rsm = rack_merge_rsm(rack, prev, rsm);
10068 prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10069 } else
10070 break;
10071 }
10072 }
10073 if (used_ref == 0) {
10074 counter_u64_add(rack_sack_proc_all, 1);
10075 } else {
10076 counter_u64_add(rack_sack_proc_short, 1);
10077 }
10078 /* Save off the next one for quick reference. */
10079 nrsm = tqhash_find(rack->r_ctl.tqh, end);
10080 *prsm = rack->r_ctl.rc_sacklast = nrsm;
10081 return (changed);
10082 }
10083
10084 static void inline
rack_peer_reneges(struct tcp_rack * rack,struct rack_sendmap * rsm,tcp_seq th_ack)10085 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10086 {
10087 struct rack_sendmap *tmap;
10088
10089 tmap = NULL;
10090 while (rsm && (rsm->r_flags & RACK_ACKED)) {
10091 /* Its no longer sacked, mark it so */
10092 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10093 #ifdef INVARIANTS
10094 if (rsm->r_in_tmap) {
10095 panic("rack:%p rsm:%p flags:0x%x in tmap?",
10096 rack, rsm, rsm->r_flags);
10097 }
10098 #endif
10099 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10100 /* Rebuild it into our tmap */
10101 if (tmap == NULL) {
10102 TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10103 tmap = rsm;
10104 } else {
10105 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10106 tmap = rsm;
10107 }
10108 tmap->r_in_tmap = 1;
10109 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10110 }
10111 /*
10112 * Now lets possibly clear the sack filter so we start
10113 * recognizing sacks that cover this area.
10114 */
10115 sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10116
10117 }
10118
10119
10120 static void inline
rack_rsm_sender_update(struct tcp_rack * rack,struct tcpcb * tp,struct rack_sendmap * rsm,uint8_t from)10121 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10122 {
10123 /*
10124 * We look at advancing the end send time for our GP
10125 * measurement tracking only as the cumulative acknowledgment
10126 * moves forward. You might wonder about this, why not
10127 * at every transmission or retransmission within the
10128 * GP window update the rc_gp_cumack_ts? Well its rather
10129 * nuanced but basically the GP window *may* expand (as
10130 * it does below) or worse and harder to track it may shrink.
10131 *
10132 * This last makes it impossible to track at the time of
10133 * the send, since you may set forward your rc_gp_cumack_ts
10134 * when you send, because that send *is* in your currently
10135 * "guessed" window, but then it shrinks. Now which was
10136 * the send time of the last bytes in the window, by the
10137 * time you ask that question that part of the sendmap
10138 * is freed. So you don't know and you will have too
10139 * long of send window. Instead by updating the time
10140 * marker only when the cumack advances this assures us
10141 * that we will have only the sends in the window of our
10142 * GP measurement.
10143 *
10144 * Another complication from this is the
10145 * merging of sendmap entries. During SACK processing this
10146 * can happen to conserve the sendmap size. That breaks
10147 * everything down in tracking the send window of the GP
10148 * estimate. So to prevent that and keep it working with
10149 * a tiny bit more limited merging, we only allow like
10150 * types to be merged. I.e. if two sends are in the GP window
10151 * then its ok to merge them together. If two sends are not
10152 * in the GP window its ok to merge them together too. Though
10153 * one send in and one send out cannot be merged. We combine
10154 * this with never allowing the shrinking of the GP window when
10155 * we are in recovery so that we can properly calculate the
10156 * sending times.
10157 *
10158 * This all of course seems complicated, because it is.. :)
10159 *
10160 * The cum-ack is being advanced upon the sendmap.
10161 * If we are not doing a GP estimate don't
10162 * proceed.
10163 */
10164 uint64_t ts;
10165
10166 if ((tp->t_flags & TF_GPUTINPROG) == 0)
10167 return;
10168 /*
10169 * If this sendmap entry is going
10170 * beyond the measurement window we had picked,
10171 * expand the measurement window by that much.
10172 */
10173 if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10174 tp->gput_ack = rsm->r_end;
10175 }
10176 /*
10177 * If we have not setup a ack, then we
10178 * have no idea if the newly acked pieces
10179 * will be "in our seq measurement range". If
10180 * it is when we clear the app_limited_needs_set
10181 * flag the timestamp will be updated.
10182 */
10183 if (rack->app_limited_needs_set)
10184 return;
10185 /*
10186 * Finally, we grab out the latest timestamp
10187 * that this packet was sent and then see
10188 * if:
10189 * a) The packet touches are newly defined GP range.
10190 * b) The time is greater than (newer) than the
10191 * one we currently have. If so we update
10192 * our sending end time window.
10193 *
10194 * Note we *do not* do this at send time. The reason
10195 * is that if you do you *may* pick up a newer timestamp
10196 * for a range you are not going to measure. We project
10197 * out how far and then sometimes modify that to be
10198 * smaller. If that occurs then you will have a send
10199 * that does not belong to the range included.
10200 */
10201 if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10202 rack->r_ctl.rc_gp_cumack_ts)
10203 return;
10204 if (rack_in_gp_window(tp, rsm)) {
10205 rack->r_ctl.rc_gp_cumack_ts = ts;
10206 rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10207 __LINE__, from, rsm);
10208 }
10209 }
10210
10211 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)10212 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)
10213 {
10214 struct rack_sendmap *rsm;
10215 /*
10216 * The ACK point is advancing to th_ack, we must drop off
10217 * the packets in the rack log and calculate any eligble
10218 * RTT's.
10219 */
10220
10221 if (sack_filter_blks_used(&rack->r_ctl.rack_sf)) {
10222 /*
10223 * If we have some sack blocks in the filter
10224 * lets prune them out by calling sfb with no blocks.
10225 */
10226 sack_filter_blks(tp, &rack->r_ctl.rack_sf, NULL, 0, th_ack);
10227 }
10228 if (SEQ_GT(th_ack, tp->snd_una)) {
10229 /* Clear any app ack remembered settings */
10230 rack->r_ctl.cleared_app_ack = 0;
10231 }
10232 rack->r_wanted_output = 1;
10233 if (SEQ_GT(th_ack, tp->snd_una))
10234 rack->r_ctl.last_cumack_advance = acktime;
10235
10236 /* Tend any TLP that has been marked for 1/2 the seq space (its old) */
10237 if ((rack->rc_last_tlp_acked_set == 1)&&
10238 (rack->rc_last_tlp_past_cumack == 1) &&
10239 (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10240 /*
10241 * We have reached the point where our last rack
10242 * tlp retransmit sequence is ahead of the cum-ack.
10243 * This can only happen when the cum-ack moves all
10244 * the way around (its been a full 2^^31+1 bytes
10245 * or more since we sent a retransmitted TLP). Lets
10246 * turn off the valid flag since its not really valid.
10247 *
10248 * Note since sack's also turn on this event we have
10249 * a complication, we have to wait to age it out until
10250 * the cum-ack is by the TLP before checking which is
10251 * what the next else clause does.
10252 */
10253 rack_log_dsack_event(rack, 9, __LINE__,
10254 rack->r_ctl.last_tlp_acked_start,
10255 rack->r_ctl.last_tlp_acked_end);
10256 rack->rc_last_tlp_acked_set = 0;
10257 rack->rc_last_tlp_past_cumack = 0;
10258 } else if ((rack->rc_last_tlp_acked_set == 1) &&
10259 (rack->rc_last_tlp_past_cumack == 0) &&
10260 (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10261 /*
10262 * It is safe to start aging TLP's out.
10263 */
10264 rack->rc_last_tlp_past_cumack = 1;
10265 }
10266 /* We do the same for the tlp send seq as well */
10267 if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10268 (rack->rc_last_sent_tlp_past_cumack == 1) &&
10269 (SEQ_GT(rack->r_ctl.last_sent_tlp_seq, th_ack))) {
10270 rack_log_dsack_event(rack, 9, __LINE__,
10271 rack->r_ctl.last_sent_tlp_seq,
10272 (rack->r_ctl.last_sent_tlp_seq +
10273 rack->r_ctl.last_sent_tlp_len));
10274 rack->rc_last_sent_tlp_seq_valid = 0;
10275 rack->rc_last_sent_tlp_past_cumack = 0;
10276 } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10277 (rack->rc_last_sent_tlp_past_cumack == 0) &&
10278 (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10279 /*
10280 * It is safe to start aging TLP's send.
10281 */
10282 rack->rc_last_sent_tlp_past_cumack = 1;
10283 }
10284 more:
10285 rsm = tqhash_min(rack->r_ctl.tqh);
10286 if (rsm == NULL) {
10287 if ((th_ack - 1) == tp->iss) {
10288 /*
10289 * For the SYN incoming case we will not
10290 * have called tcp_output for the sending of
10291 * the SYN, so there will be no map. All
10292 * other cases should probably be a panic.
10293 */
10294 return;
10295 }
10296 if (tp->t_flags & TF_SENTFIN) {
10297 /* if we sent a FIN we often will not have map */
10298 return;
10299 }
10300 #ifdef INVARIANTS
10301 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u\n",
10302 tp,
10303 tp->t_state, th_ack, rack,
10304 tp->snd_una, tp->snd_max);
10305 #endif
10306 return;
10307 }
10308 if (SEQ_LT(th_ack, rsm->r_start)) {
10309 /* Huh map is missing this */
10310 #ifdef INVARIANTS
10311 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10312 rsm->r_start,
10313 th_ack, tp->t_state, rack->r_state);
10314 #endif
10315 return;
10316 }
10317 rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10318
10319 /* Now was it a retransmitted TLP? */
10320 if ((rsm->r_flags & RACK_TLP) &&
10321 (rsm->r_rtr_cnt > 1)) {
10322 /*
10323 * Yes, this rsm was a TLP and retransmitted, remember that
10324 * since if a DSACK comes back on this we don't want
10325 * to think of it as a reordered segment. This may
10326 * get updated again with possibly even other TLPs
10327 * in flight, but thats ok. Only when we don't send
10328 * a retransmitted TLP for 1/2 the sequences space
10329 * will it get turned off (above).
10330 */
10331 if (rack->rc_last_tlp_acked_set &&
10332 (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10333 /*
10334 * We already turned this on since the end matches,
10335 * the previous one was a partially ack now we
10336 * are getting another one (maybe all of it).
10337 */
10338 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10339 /*
10340 * Lets make sure we have all of it though.
10341 */
10342 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10343 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10344 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10345 rack->r_ctl.last_tlp_acked_end);
10346 }
10347 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10348 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10349 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10350 rack->r_ctl.last_tlp_acked_end);
10351 }
10352 } else {
10353 rack->rc_last_tlp_past_cumack = 1;
10354 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10355 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10356 rack->rc_last_tlp_acked_set = 1;
10357 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10358 }
10359 }
10360 /* Now do we consume the whole thing? */
10361 rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10362 if (SEQ_GEQ(th_ack, rsm->r_end)) {
10363 /* Its all consumed. */
10364 uint32_t left;
10365 uint8_t newly_acked;
10366
10367 if (rsm->r_flags & RACK_WAS_LOST) {
10368 /*
10369 * This can happen when we marked it as lost
10370 * and yet before retransmitting we get an ack
10371 * which can happen due to reordering.
10372 */
10373 rsm->r_flags &= ~RACK_WAS_LOST;
10374 KASSERT((rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start)),
10375 ("rsm:%p rack:%p rc_considered_lost goes negative", rsm, rack));
10376 if (rack->r_ctl.rc_considered_lost >= (rsm->r_end - rsm->r_start))
10377 rack->r_ctl.rc_considered_lost -= rsm->r_end - rsm->r_start;
10378 else
10379 rack->r_ctl.rc_considered_lost = 0;
10380 }
10381 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10382 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10383 rsm->r_rtr_bytes = 0;
10384 /*
10385 * Record the time of highest cumack sent if its in our measurement
10386 * window and possibly bump out the end.
10387 */
10388 rack_rsm_sender_update(rack, tp, rsm, 4);
10389 tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10390 if (rsm->r_in_tmap) {
10391 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10392 rsm->r_in_tmap = 0;
10393 }
10394 newly_acked = 1;
10395 if (rsm->r_flags & RACK_ACKED) {
10396 /*
10397 * It was acked on the scoreboard -- remove
10398 * it from total
10399 */
10400 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10401 newly_acked = 0;
10402 } else if (rsm->r_flags & RACK_SACK_PASSED) {
10403 /*
10404 * There are segments ACKED on the
10405 * scoreboard further up. We are seeing
10406 * reordering.
10407 */
10408 rsm->r_flags &= ~RACK_SACK_PASSED;
10409 rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10410 rsm->r_flags |= RACK_ACKED;
10411 rack->r_ctl.rc_reorder_ts = cts;
10412 if (rack->r_ctl.rc_reorder_ts == 0)
10413 rack->r_ctl.rc_reorder_ts = 1;
10414 if (rack->r_ent_rec_ns) {
10415 /*
10416 * We have sent no more, and we saw an sack
10417 * then ack arrive.
10418 */
10419 rack->r_might_revert = 1;
10420 }
10421 rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10422 } else {
10423 rack_update_pcm_ack(rack, 1, rsm->r_start, rsm->r_end);
10424 }
10425 if ((rsm->r_flags & RACK_TO_REXT) &&
10426 (tp->t_flags & TF_RCVD_TSTMP) &&
10427 (to->to_flags & TOF_TS) &&
10428 (to->to_tsecr != 0) &&
10429 (tp->t_flags & TF_PREVVALID)) {
10430 /*
10431 * We can use the timestamp to see
10432 * if this retransmission was from the
10433 * first transmit. If so we made a mistake.
10434 */
10435 tp->t_flags &= ~TF_PREVVALID;
10436 if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10437 /* The first transmit is what this ack is for */
10438 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10439 }
10440 }
10441 left = th_ack - rsm->r_end;
10442 if (rack->app_limited_needs_set && newly_acked)
10443 rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10444 /* Free back to zone */
10445 rack_free(rack, rsm);
10446 if (left) {
10447 goto more;
10448 }
10449 /* Check for reneging */
10450 rsm = tqhash_min(rack->r_ctl.tqh);
10451 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10452 /*
10453 * The peer has moved snd_una up to
10454 * the edge of this send, i.e. one
10455 * that it had previously acked. The only
10456 * way that can be true if the peer threw
10457 * away data (space issues) that it had
10458 * previously sacked (else it would have
10459 * given us snd_una up to (rsm->r_end).
10460 * We need to undo the acked markings here.
10461 *
10462 * Note we have to look to make sure th_ack is
10463 * our rsm->r_start in case we get an old ack
10464 * where th_ack is behind snd_una.
10465 */
10466 rack_peer_reneges(rack, rsm, th_ack);
10467 }
10468 return;
10469 }
10470 if (rsm->r_flags & RACK_ACKED) {
10471 /*
10472 * It was acked on the scoreboard -- remove it from
10473 * total for the part being cum-acked.
10474 */
10475 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10476 } else {
10477 rack_update_pcm_ack(rack, 1, rsm->r_start, th_ack);
10478 }
10479 /* And what about the lost flag? */
10480 if (rsm->r_flags & RACK_WAS_LOST) {
10481 /*
10482 * This can happen when we marked it as lost
10483 * and yet before retransmitting we get an ack
10484 * which can happen due to reordering. In this
10485 * case its only a partial ack of the send.
10486 */
10487 KASSERT((rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start)),
10488 ("rsm:%p rack:%p rc_considered_lost goes negative th_ack:%u", rsm, rack, th_ack));
10489 if (rack->r_ctl.rc_considered_lost >= (th_ack - rsm->r_start))
10490 rack->r_ctl.rc_considered_lost -= th_ack - rsm->r_start;
10491 else
10492 rack->r_ctl.rc_considered_lost = 0;
10493 }
10494 /*
10495 * Clear the dup ack count for
10496 * the piece that remains.
10497 */
10498 rsm->r_dupack = 0;
10499 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10500 if (rsm->r_rtr_bytes) {
10501 /*
10502 * It was retransmitted adjust the
10503 * sack holes for what was acked.
10504 */
10505 int ack_am;
10506
10507 ack_am = (th_ack - rsm->r_start);
10508 if (ack_am >= rsm->r_rtr_bytes) {
10509 rack->r_ctl.rc_holes_rxt -= ack_am;
10510 rsm->r_rtr_bytes -= ack_am;
10511 }
10512 }
10513 /*
10514 * Update where the piece starts and record
10515 * the time of send of highest cumack sent if
10516 * its in our GP range.
10517 */
10518 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10519 /* Now we need to move our offset forward too */
10520 if (rsm->m &&
10521 ((rsm->orig_m_len != rsm->m->m_len) ||
10522 (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10523 /* Fix up the orig_m_len and possibly the mbuf offset */
10524 rack_adjust_orig_mlen(rsm);
10525 }
10526 rsm->soff += (th_ack - rsm->r_start);
10527 rack_rsm_sender_update(rack, tp, rsm, 5);
10528 /* The trim will move th_ack into r_start for us */
10529 tqhash_trim(rack->r_ctl.tqh, th_ack);
10530 /* Now do we need to move the mbuf fwd too? */
10531 {
10532 struct mbuf *m;
10533 uint32_t soff;
10534
10535 m = rsm->m;
10536 soff = rsm->soff;
10537 if (m) {
10538 while (soff >= m->m_len) {
10539 soff -= m->m_len;
10540 KASSERT((m->m_next != NULL),
10541 (" rsm:%p off:%u soff:%u m:%p",
10542 rsm, rsm->soff, soff, m));
10543 m = m->m_next;
10544 if (m == NULL) {
10545 /*
10546 * This is a fall-back that prevents a panic. In reality
10547 * we should be able to walk the mbuf's and find our place.
10548 * At this point snd_una has not been updated with the sbcut() yet
10549 * but tqhash_trim did update rsm->r_start so the offset calcuation
10550 * should work fine. This is undesirable since we will take cache
10551 * hits to access the socket buffer. And even more puzzling is that
10552 * it happens occasionally. It should not :(
10553 */
10554 m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10555 (rsm->r_start - tp->snd_una),
10556 &soff);
10557 break;
10558 }
10559 }
10560 /*
10561 * Now save in our updated values.
10562 */
10563 rsm->m = m;
10564 rsm->soff = soff;
10565 rsm->orig_m_len = rsm->m->m_len;
10566 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10567 }
10568 }
10569 if (rack->app_limited_needs_set &&
10570 SEQ_GEQ(th_ack, tp->gput_seq))
10571 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10572 }
10573
10574 static void
rack_handle_might_revert(struct tcpcb * tp,struct tcp_rack * rack)10575 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10576 {
10577 struct rack_sendmap *rsm;
10578 int sack_pass_fnd = 0;
10579
10580 if (rack->r_might_revert) {
10581 /*
10582 * Ok we have reordering, have not sent anything, we
10583 * might want to revert the congestion state if nothing
10584 * further has SACK_PASSED on it. Lets check.
10585 *
10586 * We also get here when we have DSACKs come in for
10587 * all the data that we FR'd. Note that a rxt or tlp
10588 * timer clears this from happening.
10589 */
10590
10591 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10592 if (rsm->r_flags & RACK_SACK_PASSED) {
10593 sack_pass_fnd = 1;
10594 break;
10595 }
10596 }
10597 if (sack_pass_fnd == 0) {
10598 /*
10599 * We went into recovery
10600 * incorrectly due to reordering!
10601 */
10602 int orig_cwnd;
10603
10604 rack->r_ent_rec_ns = 0;
10605 orig_cwnd = tp->snd_cwnd;
10606 tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10607 tp->snd_recover = tp->snd_una;
10608 rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10609 if (IN_RECOVERY(tp->t_flags)) {
10610 rack_exit_recovery(tp, rack, 3);
10611 if ((rack->rto_from_rec == 1) && (rack_ssthresh_rest_rto_rec != 0) ){
10612 /*
10613 * We were in recovery, had an RTO
10614 * and then re-entered recovery (more sack's arrived)
10615 * and we have properly recorded the old ssthresh from
10616 * the first recovery. We want to be able to slow-start
10617 * back to this level. The ssthresh from the timeout
10618 * and then back into recovery will end up most likely
10619 * to be min(cwnd=1mss, 2mss). Which makes it basically
10620 * so we get no slow-start after our RTO.
10621 */
10622 rack->rto_from_rec = 0;
10623 if (rack->r_ctl.rto_ssthresh > tp->snd_ssthresh)
10624 tp->snd_ssthresh = rack->r_ctl.rto_ssthresh;
10625 }
10626 }
10627 }
10628 rack->r_might_revert = 0;
10629 }
10630 }
10631
10632
10633 static int
rack_note_dsack(struct tcp_rack * rack,tcp_seq start,tcp_seq end)10634 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10635 {
10636
10637 uint32_t am, l_end;
10638 int was_tlp = 0;
10639
10640 if (SEQ_GT(end, start))
10641 am = end - start;
10642 else
10643 am = 0;
10644 if ((rack->rc_last_tlp_acked_set ) &&
10645 (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
10646 (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
10647 /*
10648 * The DSACK is because of a TLP which we don't
10649 * do anything with the reordering window over since
10650 * it was not reordering that caused the DSACK but
10651 * our previous retransmit TLP.
10652 */
10653 rack_log_dsack_event(rack, 7, __LINE__, start, end);
10654 was_tlp = 1;
10655 goto skip_dsack_round;
10656 }
10657 if (rack->rc_last_sent_tlp_seq_valid) {
10658 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
10659 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
10660 (SEQ_LEQ(end, l_end))) {
10661 /*
10662 * This dsack is from the last sent TLP, ignore it
10663 * for reordering purposes.
10664 */
10665 rack_log_dsack_event(rack, 7, __LINE__, start, end);
10666 was_tlp = 1;
10667 goto skip_dsack_round;
10668 }
10669 }
10670 if (rack->rc_dsack_round_seen == 0) {
10671 rack->rc_dsack_round_seen = 1;
10672 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
10673 rack->r_ctl.num_dsack++;
10674 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
10675 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
10676 }
10677 skip_dsack_round:
10678 /*
10679 * We keep track of how many DSACK blocks we get
10680 * after a recovery incident.
10681 */
10682 rack->r_ctl.dsack_byte_cnt += am;
10683 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
10684 rack->r_ctl.retran_during_recovery &&
10685 (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
10686 /*
10687 * False recovery most likely culprit is reordering. If
10688 * nothing else is missing we need to revert.
10689 */
10690 rack->r_might_revert = 1;
10691 rack_handle_might_revert(rack->rc_tp, rack);
10692 rack->r_might_revert = 0;
10693 rack->r_ctl.retran_during_recovery = 0;
10694 rack->r_ctl.dsack_byte_cnt = 0;
10695 }
10696 return (was_tlp);
10697 }
10698
10699 static uint32_t
do_rack_compute_pipe(struct tcpcb * tp,struct tcp_rack * rack,uint32_t snd_una)10700 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
10701 {
10702 return (((tp->snd_max - snd_una) -
10703 (rack->r_ctl.rc_sacked + rack->r_ctl.rc_considered_lost)) + rack->r_ctl.rc_holes_rxt);
10704 }
10705
10706 static int32_t
rack_compute_pipe(struct tcpcb * tp)10707 rack_compute_pipe(struct tcpcb *tp)
10708 {
10709 return ((int32_t)do_rack_compute_pipe(tp,
10710 (struct tcp_rack *)tp->t_fb_ptr,
10711 tp->snd_una));
10712 }
10713
10714 static void
rack_update_prr(struct tcpcb * tp,struct tcp_rack * rack,uint32_t changed,tcp_seq th_ack)10715 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
10716 {
10717 /* Deal with changed and PRR here (in recovery only) */
10718 uint32_t pipe, snd_una;
10719
10720 rack->r_ctl.rc_prr_delivered += changed;
10721
10722 if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
10723 /*
10724 * It is all outstanding, we are application limited
10725 * and thus we don't need more room to send anything.
10726 * Note we use tp->snd_una here and not th_ack because
10727 * the data as yet not been cut from the sb.
10728 */
10729 rack->r_ctl.rc_prr_sndcnt = 0;
10730 return;
10731 }
10732 /* Compute prr_sndcnt */
10733 if (SEQ_GT(tp->snd_una, th_ack)) {
10734 snd_una = tp->snd_una;
10735 } else {
10736 snd_una = th_ack;
10737 }
10738 pipe = do_rack_compute_pipe(tp, rack, snd_una);
10739 if (pipe > tp->snd_ssthresh) {
10740 long sndcnt;
10741
10742 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
10743 if (rack->r_ctl.rc_prr_recovery_fs > 0)
10744 sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
10745 else {
10746 rack->r_ctl.rc_prr_sndcnt = 0;
10747 rack_log_to_prr(rack, 9, 0, __LINE__);
10748 sndcnt = 0;
10749 }
10750 sndcnt++;
10751 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
10752 sndcnt -= rack->r_ctl.rc_prr_out;
10753 else
10754 sndcnt = 0;
10755 rack->r_ctl.rc_prr_sndcnt = sndcnt;
10756 rack_log_to_prr(rack, 10, 0, __LINE__);
10757 } else {
10758 uint32_t limit;
10759
10760 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
10761 limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
10762 else
10763 limit = 0;
10764 if (changed > limit)
10765 limit = changed;
10766 limit += ctf_fixed_maxseg(tp);
10767 if (tp->snd_ssthresh > pipe) {
10768 rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
10769 rack_log_to_prr(rack, 11, 0, __LINE__);
10770 } else {
10771 rack->r_ctl.rc_prr_sndcnt = min(0, limit);
10772 rack_log_to_prr(rack, 12, 0, __LINE__);
10773 }
10774 }
10775 }
10776
10777 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)10778 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
10779 int *dsack_seen, int *sacks_seen)
10780 {
10781 uint32_t changed;
10782 struct tcp_rack *rack;
10783 struct rack_sendmap *rsm;
10784 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
10785 register uint32_t th_ack;
10786 int32_t i, j, k, num_sack_blks = 0;
10787 uint32_t cts, acked, ack_point;
10788 int loop_start = 0;
10789 uint32_t tsused;
10790 uint32_t segsiz;
10791
10792
10793 INP_WLOCK_ASSERT(tptoinpcb(tp));
10794 if (tcp_get_flags(th) & TH_RST) {
10795 /* We don't log resets */
10796 return;
10797 }
10798 rack = (struct tcp_rack *)tp->t_fb_ptr;
10799 cts = tcp_get_usecs(NULL);
10800 rsm = tqhash_min(rack->r_ctl.tqh);
10801 changed = 0;
10802 th_ack = th->th_ack;
10803 segsiz = ctf_fixed_maxseg(rack->rc_tp);
10804 if (BYTES_THIS_ACK(tp, th) >= segsiz) {
10805 /*
10806 * You only get credit for
10807 * MSS and greater (and you get extra
10808 * credit for larger cum-ack moves).
10809 */
10810 int ac;
10811
10812 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
10813 counter_u64_add(rack_ack_total, ac);
10814 }
10815 if (SEQ_GT(th_ack, tp->snd_una)) {
10816 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
10817 tp->t_acktime = ticks;
10818 }
10819 if (rsm && SEQ_GT(th_ack, rsm->r_start))
10820 changed = th_ack - rsm->r_start;
10821 if (changed) {
10822 rack_process_to_cumack(tp, rack, th_ack, cts, to,
10823 tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time));
10824 }
10825 if ((to->to_flags & TOF_SACK) == 0) {
10826 /* We are done nothing left and no sack. */
10827 rack_handle_might_revert(tp, rack);
10828 /*
10829 * For cases where we struck a dup-ack
10830 * with no SACK, add to the changes so
10831 * PRR will work right.
10832 */
10833 if (dup_ack_struck && (changed == 0)) {
10834 changed += ctf_fixed_maxseg(rack->rc_tp);
10835 }
10836 goto out;
10837 }
10838 /* Sack block processing */
10839 if (SEQ_GT(th_ack, tp->snd_una))
10840 ack_point = th_ack;
10841 else
10842 ack_point = tp->snd_una;
10843 for (i = 0; i < to->to_nsacks; i++) {
10844 bcopy((to->to_sacks + i * TCPOLEN_SACK),
10845 &sack, sizeof(sack));
10846 sack.start = ntohl(sack.start);
10847 sack.end = ntohl(sack.end);
10848 if (SEQ_GT(sack.end, sack.start) &&
10849 SEQ_GT(sack.start, ack_point) &&
10850 SEQ_LT(sack.start, tp->snd_max) &&
10851 SEQ_GT(sack.end, ack_point) &&
10852 SEQ_LEQ(sack.end, tp->snd_max)) {
10853 sack_blocks[num_sack_blks] = sack;
10854 num_sack_blks++;
10855 } else if (SEQ_LEQ(sack.start, th_ack) &&
10856 SEQ_LEQ(sack.end, th_ack)) {
10857 int was_tlp;
10858
10859 if (dsack_seen != NULL)
10860 *dsack_seen = 1;
10861 was_tlp = rack_note_dsack(rack, sack.start, sack.end);
10862 /*
10863 * Its a D-SACK block.
10864 */
10865 tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
10866 }
10867 }
10868 if (rack->rc_dsack_round_seen) {
10869 /* Is the dsack roound over? */
10870 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
10871 /* Yes it is */
10872 rack->rc_dsack_round_seen = 0;
10873 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
10874 }
10875 }
10876 /*
10877 * Sort the SACK blocks so we can update the rack scoreboard with
10878 * just one pass.
10879 */
10880 num_sack_blks = sack_filter_blks(tp, &rack->r_ctl.rack_sf, sack_blocks,
10881 num_sack_blks, th->th_ack);
10882 ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
10883 if (sacks_seen != NULL)
10884 *sacks_seen = num_sack_blks;
10885 if (num_sack_blks == 0) {
10886 /* Nothing to sack, but we need to update counts */
10887 goto out_with_totals;
10888 }
10889 /* Its a sack of some sort */
10890 if (num_sack_blks < 2) {
10891 /* Only one, we don't need to sort */
10892 goto do_sack_work;
10893 }
10894 /* Sort the sacks */
10895 for (i = 0; i < num_sack_blks; i++) {
10896 for (j = i + 1; j < num_sack_blks; j++) {
10897 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
10898 sack = sack_blocks[i];
10899 sack_blocks[i] = sack_blocks[j];
10900 sack_blocks[j] = sack;
10901 }
10902 }
10903 }
10904 /*
10905 * Now are any of the sack block ends the same (yes some
10906 * implementations send these)?
10907 */
10908 again:
10909 if (num_sack_blks == 0)
10910 goto out_with_totals;
10911 if (num_sack_blks > 1) {
10912 for (i = 0; i < num_sack_blks; i++) {
10913 for (j = i + 1; j < num_sack_blks; j++) {
10914 if (sack_blocks[i].end == sack_blocks[j].end) {
10915 /*
10916 * Ok these two have the same end we
10917 * want the smallest end and then
10918 * throw away the larger and start
10919 * again.
10920 */
10921 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
10922 /*
10923 * The second block covers
10924 * more area use that
10925 */
10926 sack_blocks[i].start = sack_blocks[j].start;
10927 }
10928 /*
10929 * Now collapse out the dup-sack and
10930 * lower the count
10931 */
10932 for (k = (j + 1); k < num_sack_blks; k++) {
10933 sack_blocks[j].start = sack_blocks[k].start;
10934 sack_blocks[j].end = sack_blocks[k].end;
10935 j++;
10936 }
10937 num_sack_blks--;
10938 goto again;
10939 }
10940 }
10941 }
10942 }
10943 do_sack_work:
10944 /*
10945 * First lets look to see if
10946 * we have retransmitted and
10947 * can use the transmit next?
10948 */
10949 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
10950 if (rsm &&
10951 SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
10952 SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
10953 /*
10954 * We probably did the FR and the next
10955 * SACK in continues as we would expect.
10956 */
10957 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, segsiz);
10958 if (acked) {
10959 rack->r_wanted_output = 1;
10960 changed += acked;
10961 }
10962 if (num_sack_blks == 1) {
10963 /*
10964 * This is what we would expect from
10965 * a normal implementation to happen
10966 * after we have retransmitted the FR,
10967 * i.e the sack-filter pushes down
10968 * to 1 block and the next to be retransmitted
10969 * is the sequence in the sack block (has more
10970 * are acked). Count this as ACK'd data to boost
10971 * up the chances of recovering any false positives.
10972 */
10973 counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
10974 counter_u64_add(rack_express_sack, 1);
10975 goto out_with_totals;
10976 } else {
10977 /*
10978 * Start the loop through the
10979 * rest of blocks, past the first block.
10980 */
10981 loop_start = 1;
10982 }
10983 }
10984 counter_u64_add(rack_sack_total, 1);
10985 rsm = rack->r_ctl.rc_sacklast;
10986 for (i = loop_start; i < num_sack_blks; i++) {
10987 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, segsiz);
10988 if (acked) {
10989 rack->r_wanted_output = 1;
10990 changed += acked;
10991 }
10992 }
10993 out_with_totals:
10994 if (num_sack_blks > 1) {
10995 /*
10996 * You get an extra stroke if
10997 * you have more than one sack-blk, this
10998 * could be where we are skipping forward
10999 * and the sack-filter is still working, or
11000 * it could be an attacker constantly
11001 * moving us.
11002 */
11003 counter_u64_add(rack_move_some, 1);
11004 }
11005 out:
11006 if (changed) {
11007 /* Something changed cancel the rack timer */
11008 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11009 }
11010 tsused = tcp_get_usecs(NULL);
11011 rsm = tcp_rack_output(tp, rack, tsused);
11012 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11013 rsm &&
11014 ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11015 /* Enter recovery */
11016 entered_recovery = 1;
11017 rack_cong_signal(tp, CC_NDUPACK, th_ack, __LINE__);
11018 /*
11019 * When we enter recovery we need to assure we send
11020 * one packet.
11021 */
11022 if (rack->rack_no_prr == 0) {
11023 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11024 rack_log_to_prr(rack, 8, 0, __LINE__);
11025 }
11026 rack->r_timer_override = 1;
11027 rack->r_early = 0;
11028 rack->r_ctl.rc_agg_early = 0;
11029 } else if (IN_FASTRECOVERY(tp->t_flags) &&
11030 rsm &&
11031 (rack->r_rr_config == 3)) {
11032 /*
11033 * Assure we can output and we get no
11034 * remembered pace time except the retransmit.
11035 */
11036 rack->r_timer_override = 1;
11037 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11038 rack->r_ctl.rc_resend = rsm;
11039 }
11040 if (IN_FASTRECOVERY(tp->t_flags) &&
11041 (rack->rack_no_prr == 0) &&
11042 (entered_recovery == 0)) {
11043 rack_update_prr(tp, rack, changed, th_ack);
11044 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11045 ((tcp_in_hpts(rack->rc_tp) == 0) &&
11046 ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11047 /*
11048 * If you are pacing output you don't want
11049 * to override.
11050 */
11051 rack->r_early = 0;
11052 rack->r_ctl.rc_agg_early = 0;
11053 rack->r_timer_override = 1;
11054 }
11055 }
11056 }
11057
11058 static void
rack_strike_dupack(struct tcp_rack * rack,tcp_seq th_ack)11059 rack_strike_dupack(struct tcp_rack *rack, tcp_seq th_ack)
11060 {
11061 struct rack_sendmap *rsm;
11062
11063 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11064 while (rsm) {
11065 /*
11066 * We need to skip anything already set
11067 * to be retransmitted.
11068 */
11069 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
11070 (rsm->r_flags & RACK_MUST_RXT)) {
11071 rsm = TAILQ_NEXT(rsm, r_tnext);
11072 continue;
11073 }
11074 break;
11075 }
11076 if (rsm && (rsm->r_dupack < 0xff)) {
11077 rsm->r_dupack++;
11078 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11079 struct timeval tv;
11080 uint32_t cts;
11081 /*
11082 * Here we see if we need to retransmit. For
11083 * a SACK type connection if enough time has passed
11084 * we will get a return of the rsm. For a non-sack
11085 * connection we will get the rsm returned if the
11086 * dupack value is 3 or more.
11087 */
11088 cts = tcp_get_usecs(&tv);
11089 rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11090 if (rack->r_ctl.rc_resend != NULL) {
11091 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11092 rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11093 th_ack, __LINE__);
11094 }
11095 rack->r_wanted_output = 1;
11096 rack->r_timer_override = 1;
11097 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11098 }
11099 } else {
11100 rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11101 }
11102 }
11103 }
11104
11105 static void
rack_check_bottom_drag(struct tcpcb * tp,struct tcp_rack * rack,struct socket * so)11106 rack_check_bottom_drag(struct tcpcb *tp,
11107 struct tcp_rack *rack,
11108 struct socket *so)
11109 {
11110 /*
11111 * So what is dragging bottom?
11112 *
11113 * Dragging bottom means you were under pacing and had a
11114 * delay in processing inbound acks waiting on our pacing
11115 * timer to expire. While you were waiting all of the acknowledgments
11116 * for the packets you sent have arrived. This means we are pacing
11117 * way underneath the bottleneck to the point where our Goodput
11118 * measurements stop working, since they require more than one
11119 * ack (usually at least 8 packets worth with multiple acks so we can
11120 * gauge the inter-ack times). If that occurs we have a real problem
11121 * since we are stuck in a hole that we can't get out of without
11122 * something speeding us up.
11123 *
11124 * We also check to see if we are widdling down to just one segment
11125 * outstanding. If this occurs and we have room to send in our cwnd/rwnd
11126 * then we are adding the delayed ack interval into our measurments and
11127 * we need to speed up slightly.
11128 */
11129 uint32_t segsiz, minseg;
11130
11131 segsiz = ctf_fixed_maxseg(tp);
11132 minseg = segsiz;
11133 if (tp->snd_max == tp->snd_una) {
11134 /*
11135 * We are doing dynamic pacing and we are way
11136 * under. Basically everything got acked while
11137 * we were still waiting on the pacer to expire.
11138 *
11139 * This means we need to boost the b/w in
11140 * addition to any earlier boosting of
11141 * the multiplier.
11142 */
11143 uint64_t lt_bw;
11144
11145 tcp_trace_point(rack->rc_tp, TCP_TP_PACED_BOTTOM);
11146 lt_bw = rack_get_lt_bw(rack);
11147 rack->rc_dragged_bottom = 1;
11148 rack_validate_multipliers_at_or_above100(rack);
11149 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11150 (rack->dis_lt_bw == 0) &&
11151 (rack->use_lesser_lt_bw == 0) &&
11152 (lt_bw > 0)) {
11153 /*
11154 * Lets use the long-term b/w we have
11155 * been getting as a base.
11156 */
11157 if (rack->rc_gp_filled == 0) {
11158 if (lt_bw > ONE_POINT_TWO_MEG) {
11159 /*
11160 * If we have no measurement
11161 * don't let us set in more than
11162 * 1.2Mbps. If we are still too
11163 * low after pacing with this we
11164 * will hopefully have a max b/w
11165 * available to sanity check things.
11166 */
11167 lt_bw = ONE_POINT_TWO_MEG;
11168 }
11169 rack->r_ctl.rc_rtt_diff = 0;
11170 rack->r_ctl.gp_bw = lt_bw;
11171 rack->rc_gp_filled = 1;
11172 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11173 rack->r_ctl.num_measurements = RACK_REQ_AVG;
11174 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11175 } else if (lt_bw > rack->r_ctl.gp_bw) {
11176 rack->r_ctl.rc_rtt_diff = 0;
11177 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11178 rack->r_ctl.num_measurements = RACK_REQ_AVG;
11179 rack->r_ctl.gp_bw = lt_bw;
11180 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11181 } else
11182 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11183 if ((rack->gp_ready == 0) &&
11184 (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11185 /* We have enough measurements now */
11186 rack->gp_ready = 1;
11187 if (rack->dgp_on ||
11188 rack->rack_hibeta)
11189 rack_set_cc_pacing(rack);
11190 if (rack->defer_options)
11191 rack_apply_deferred_options(rack);
11192 }
11193 } else {
11194 /*
11195 * zero rtt possibly?, settle for just an old increase.
11196 */
11197 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11198 }
11199 } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11200 (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11201 minseg)) &&
11202 (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11203 (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11204 (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11205 (segsiz * rack_req_segs))) {
11206 /*
11207 * We are doing dynamic GP pacing and
11208 * we have everything except 1MSS or less
11209 * bytes left out. We are still pacing away.
11210 * And there is data that could be sent, This
11211 * means we are inserting delayed ack time in
11212 * our measurements because we are pacing too slow.
11213 */
11214 rack_validate_multipliers_at_or_above100(rack);
11215 rack->rc_dragged_bottom = 1;
11216 rack_increase_bw_mul(rack, -1, 0, 0, 1);
11217 }
11218 }
11219
11220 #ifdef TCP_REQUEST_TRK
11221 static void
rack_log_hybrid(struct tcp_rack * rack,uint32_t seq,struct tcp_sendfile_track * cur,uint8_t mod,int line,int err)11222 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11223 struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11224 {
11225 int do_log;
11226
11227 do_log = tcp_bblogging_on(rack->rc_tp);
11228 if (do_log == 0) {
11229 if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11230 return;
11231 /* We only allow the three below with point logging on */
11232 if ((mod != HYBRID_LOG_RULES_APP) &&
11233 (mod != HYBRID_LOG_RULES_SET) &&
11234 (mod != HYBRID_LOG_REQ_COMP))
11235 return;
11236
11237 }
11238 if (do_log) {
11239 union tcp_log_stackspecific log;
11240 struct timeval tv;
11241
11242 /* Convert our ms to a microsecond */
11243 memset(&log, 0, sizeof(log));
11244 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11245 log.u_bbr.flex1 = seq;
11246 log.u_bbr.cwnd_gain = line;
11247 if (cur != NULL) {
11248 uint64_t off;
11249
11250 log.u_bbr.flex2 = cur->start_seq;
11251 log.u_bbr.flex3 = cur->end_seq;
11252 log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11253 log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11254 log.u_bbr.flex6 = cur->flags;
11255 log.u_bbr.pkts_out = cur->hybrid_flags;
11256 log.u_bbr.rttProp = cur->timestamp;
11257 log.u_bbr.cur_del_rate = cur->cspr;
11258 log.u_bbr.bw_inuse = cur->start;
11259 log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11260 log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11261 log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11262 log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11263 log.u_bbr.inhpts = 1;
11264 #ifdef TCP_REQUEST_TRK
11265 off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11266 log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11267 #endif
11268 } else {
11269 log.u_bbr.flex2 = err;
11270 }
11271 /*
11272 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11273 */
11274 log.u_bbr.flex7 = rack->rc_catch_up;
11275 log.u_bbr.flex7 <<= 1;
11276 log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11277 log.u_bbr.flex7 <<= 1;
11278 log.u_bbr.flex7 |= rack->dgp_on;
11279 /*
11280 * Compose bbr_state to be a bit wise 0000ADHF
11281 * where A is the always_pace flag
11282 * where D is the dgp_on flag
11283 * where H is the hybrid_mode on flag
11284 * where F is the use_fixed_rate flag.
11285 */
11286 log.u_bbr.bbr_state = rack->rc_always_pace;
11287 log.u_bbr.bbr_state <<= 1;
11288 log.u_bbr.bbr_state |= rack->dgp_on;
11289 log.u_bbr.bbr_state <<= 1;
11290 log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
11291 log.u_bbr.bbr_state <<= 1;
11292 log.u_bbr.bbr_state |= rack->use_fixed_rate;
11293 log.u_bbr.flex8 = mod;
11294 log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11295 log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11296 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11297 log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11298 log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11299 log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11300 tcp_log_event(rack->rc_tp, NULL,
11301 &rack->rc_inp->inp_socket->so_rcv,
11302 &rack->rc_inp->inp_socket->so_snd,
11303 TCP_HYBRID_PACING_LOG, 0,
11304 0, &log, false, NULL, __func__, __LINE__, &tv);
11305 }
11306 }
11307 #endif
11308
11309 #ifdef TCP_REQUEST_TRK
11310 static void
rack_set_dgp_hybrid_mode(struct tcp_rack * rack,tcp_seq seq,uint32_t len,uint64_t cts)11311 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11312 {
11313 struct tcp_sendfile_track *rc_cur, *orig_ent;
11314 struct tcpcb *tp;
11315 int err = 0;
11316
11317 orig_ent = rack->r_ctl.rc_last_sft;
11318 rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11319 if (rc_cur == NULL) {
11320 /* If not in the beginning what about the end piece */
11321 if (rack->rc_hybrid_mode)
11322 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11323 rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11324 } else {
11325 err = 12345;
11326 }
11327 /* If we find no parameters we are in straight DGP mode */
11328 if(rc_cur == NULL) {
11329 /* None found for this seq, just DGP for now */
11330 if (rack->rc_hybrid_mode) {
11331 rack->r_ctl.client_suggested_maxseg = 0;
11332 rack->rc_catch_up = 0;
11333 if (rack->cspr_is_fcc == 0)
11334 rack->r_ctl.bw_rate_cap = 0;
11335 else
11336 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11337 }
11338 if (rack->rc_hybrid_mode) {
11339 rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11340 }
11341 if (rack->r_ctl.rc_last_sft) {
11342 rack->r_ctl.rc_last_sft = NULL;
11343 }
11344 return;
11345 }
11346 if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11347 /* This entry was never setup for hybrid pacing on/off etc */
11348 if (rack->rc_hybrid_mode) {
11349 rack->r_ctl.client_suggested_maxseg = 0;
11350 rack->rc_catch_up = 0;
11351 rack->r_ctl.bw_rate_cap = 0;
11352 }
11353 if (rack->r_ctl.rc_last_sft) {
11354 rack->r_ctl.rc_last_sft = NULL;
11355 }
11356 if ((rc_cur->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11357 rc_cur->flags |= TCP_TRK_TRACK_FLG_FSND;
11358 rc_cur->first_send = cts;
11359 rc_cur->sent_at_fs = rack->rc_tp->t_sndbytes;
11360 rc_cur->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11361 }
11362 return;
11363 }
11364 /*
11365 * Ok if we have a new entry *or* have never
11366 * set up an entry we need to proceed. If
11367 * we have already set it up this entry we
11368 * just continue along with what we already
11369 * setup.
11370 */
11371 tp = rack->rc_tp;
11372 if ((rack->r_ctl.rc_last_sft != NULL) &&
11373 (rack->r_ctl.rc_last_sft == rc_cur)) {
11374 /* Its already in place */
11375 if (rack->rc_hybrid_mode)
11376 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11377 return;
11378 }
11379 if (rack->rc_hybrid_mode == 0) {
11380 rack->r_ctl.rc_last_sft = rc_cur;
11381 if (orig_ent) {
11382 orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11383 orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11384 orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11385 }
11386 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11387 return;
11388 }
11389 if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11390 /* Compensate for all the header overhead's */
11391 if (rack->cspr_is_fcc == 0)
11392 rack->r_ctl.bw_rate_cap = rack_compensate_for_linerate(rack, rc_cur->cspr);
11393 else
11394 rack->r_ctl.fillcw_cap = rack_compensate_for_linerate(rack, rc_cur->cspr);
11395 } else {
11396 if (rack->rc_hybrid_mode) {
11397 if (rack->cspr_is_fcc == 0)
11398 rack->r_ctl.bw_rate_cap = 0;
11399 else
11400 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11401 }
11402 }
11403 if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11404 rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11405 else
11406 rack->r_ctl.client_suggested_maxseg = 0;
11407 if (rc_cur->timestamp == rack->r_ctl.last_tm_mark) {
11408 /*
11409 * It is the same timestamp as the previous one
11410 * add the hybrid flag that will indicate we use
11411 * sendtime not arrival time for catch-up mode.
11412 */
11413 rc_cur->hybrid_flags |= TCP_HYBRID_PACING_SENDTIME;
11414 }
11415 if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11416 (rc_cur->cspr > 0)) {
11417 uint64_t len;
11418
11419 rack->rc_catch_up = 1;
11420 /*
11421 * Calculate the deadline time, first set the
11422 * time to when the request arrived.
11423 */
11424 if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_SENDTIME) {
11425 /*
11426 * For cases where its a duplicate tm (we received more
11427 * than one request for a tm) we want to use now, the point
11428 * where we are just sending the first bit of the request.
11429 */
11430 rc_cur->deadline = cts;
11431 } else {
11432 /*
11433 * Here we have a different tm from the last request
11434 * so we want to use arrival time as our base.
11435 */
11436 rc_cur->deadline = rc_cur->localtime;
11437 }
11438 /*
11439 * Next calculate the length and compensate for
11440 * TLS if need be.
11441 */
11442 len = rc_cur->end - rc_cur->start;
11443 if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11444 /*
11445 * This session is doing TLS. Take a swag guess
11446 * at the overhead.
11447 */
11448 len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11449 }
11450 /*
11451 * Now considering the size, and the cspr, what is the time that
11452 * would be required at the cspr rate. Here we use the raw
11453 * cspr value since the client only looks at the raw data. We
11454 * do use len which includes TLS overhead, but not the TCP/IP etc.
11455 * That will get made up for in the CU pacing rate set.
11456 */
11457 len *= HPTS_USEC_IN_SEC;
11458 len /= rc_cur->cspr;
11459 rc_cur->deadline += len;
11460 } else {
11461 rack->rc_catch_up = 0;
11462 rc_cur->deadline = 0;
11463 }
11464 if (rack->r_ctl.client_suggested_maxseg != 0) {
11465 /*
11466 * We need to reset the max pace segs if we have a
11467 * client_suggested_maxseg.
11468 */
11469 rack_set_pace_segments(tp, rack, __LINE__, NULL);
11470 }
11471 if (orig_ent) {
11472 orig_ent->sent_at_ls = rack->rc_tp->t_sndbytes;
11473 orig_ent->rxt_at_ls = rack->rc_tp->t_snd_rxt_bytes;
11474 orig_ent->flags |= TCP_TRK_TRACK_FLG_LSND;
11475 }
11476 rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11477 /* Remember it for next time and for CU mode */
11478 rack->r_ctl.rc_last_sft = rc_cur;
11479 rack->r_ctl.last_tm_mark = rc_cur->timestamp;
11480 }
11481 #endif
11482
11483 static void
rack_chk_req_and_hybrid_on_out(struct tcp_rack * rack,tcp_seq seq,uint32_t len,uint64_t cts)11484 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11485 {
11486 #ifdef TCP_REQUEST_TRK
11487 struct tcp_sendfile_track *ent;
11488
11489 ent = rack->r_ctl.rc_last_sft;
11490 if ((ent == NULL) ||
11491 (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11492 (SEQ_GEQ(seq, ent->end_seq))) {
11493 /* Time to update the track. */
11494 rack_set_dgp_hybrid_mode(rack, seq, len, cts);
11495 ent = rack->r_ctl.rc_last_sft;
11496 }
11497 /* Out of all */
11498 if (ent == NULL) {
11499 return;
11500 }
11501 if (SEQ_LT(ent->end_seq, (seq + len))) {
11502 /*
11503 * This is the case where our end_seq guess
11504 * was wrong. This is usually due to TLS having
11505 * more bytes then our guess. It could also be the
11506 * case that the client sent in two requests closely
11507 * and the SB is full of both so we are sending part
11508 * of each (end|beg). In such a case lets move this
11509 * guys end to match the end of this send. That
11510 * way it will complete when all of it is acked.
11511 */
11512 ent->end_seq = (seq + len);
11513 if (rack->rc_hybrid_mode)
11514 rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11515 }
11516 /* Now validate we have set the send time of this one */
11517 if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11518 ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11519 ent->first_send = cts;
11520 ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11521 ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11522 }
11523 #endif
11524 }
11525
11526 static void
rack_gain_for_fastoutput(struct tcp_rack * rack,struct tcpcb * tp,struct socket * so,uint32_t acked_amount)11527 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11528 {
11529 /*
11530 * The fast output path is enabled and we
11531 * have moved the cumack forward. Lets see if
11532 * we can expand forward the fast path length by
11533 * that amount. What we would ideally like to
11534 * do is increase the number of bytes in the
11535 * fast path block (left_to_send) by the
11536 * acked amount. However we have to gate that
11537 * by two factors:
11538 * 1) The amount outstanding and the rwnd of the peer
11539 * (i.e. we don't want to exceed the rwnd of the peer).
11540 * <and>
11541 * 2) The amount of data left in the socket buffer (i.e.
11542 * we can't send beyond what is in the buffer).
11543 *
11544 * Note that this does not take into account any increase
11545 * in the cwnd. We will only extend the fast path by
11546 * what was acked.
11547 */
11548 uint32_t new_total, gating_val;
11549
11550 new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11551 gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11552 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11553 if (new_total <= gating_val) {
11554 /* We can increase left_to_send by the acked amount */
11555 counter_u64_add(rack_extended_rfo, 1);
11556 rack->r_ctl.fsb.left_to_send = new_total;
11557 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11558 ("rack:%p left_to_send:%u sbavail:%u out:%u",
11559 rack, rack->r_ctl.fsb.left_to_send,
11560 sbavail(&rack->rc_inp->inp_socket->so_snd),
11561 (tp->snd_max - tp->snd_una)));
11562
11563 }
11564 }
11565
11566 static void
rack_adjust_sendmap_head(struct tcp_rack * rack,struct sockbuf * sb)11567 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11568 {
11569 /*
11570 * Here any sendmap entry that points to the
11571 * beginning mbuf must be adjusted to the correct
11572 * offset. This must be called with:
11573 * 1) The socket buffer locked
11574 * 2) snd_una adjusted to its new position.
11575 *
11576 * Note that (2) implies rack_ack_received has also
11577 * been called and all the sbcut's have been done.
11578 *
11579 * We grab the first mbuf in the socket buffer and
11580 * then go through the front of the sendmap, recalculating
11581 * the stored offset for any sendmap entry that has
11582 * that mbuf. We must use the sb functions to do this
11583 * since its possible an add was done has well as
11584 * the subtraction we may have just completed. This should
11585 * not be a penalty though, since we just referenced the sb
11586 * to go in and trim off the mbufs that we freed (of course
11587 * there will be a penalty for the sendmap references though).
11588 *
11589 * Note also with INVARIANT on, we validate with a KASSERT
11590 * that the first sendmap entry has a soff of 0.
11591 *
11592 */
11593 struct mbuf *m;
11594 struct rack_sendmap *rsm;
11595 tcp_seq snd_una;
11596 #ifdef INVARIANTS
11597 int first_processed = 0;
11598 #endif
11599
11600 snd_una = rack->rc_tp->snd_una;
11601 SOCKBUF_LOCK_ASSERT(sb);
11602 m = sb->sb_mb;
11603 rsm = tqhash_min(rack->r_ctl.tqh);
11604 if ((rsm == NULL) || (m == NULL)) {
11605 /* Nothing outstanding */
11606 return;
11607 }
11608 /* The very first RSM's mbuf must point to the head mbuf in the sb */
11609 KASSERT((rsm->m == m),
11610 ("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
11611 rack, sb, rsm));
11612 while (rsm->m && (rsm->m == m)) {
11613 /* one to adjust */
11614 #ifdef INVARIANTS
11615 struct mbuf *tm;
11616 uint32_t soff;
11617
11618 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
11619 if ((rsm->orig_m_len != m->m_len) ||
11620 (rsm->orig_t_space != M_TRAILINGROOM(m))){
11621 rack_adjust_orig_mlen(rsm);
11622 }
11623 if (first_processed == 0) {
11624 KASSERT((rsm->soff == 0),
11625 ("Rack:%p rsm:%p -- rsm at head but soff not zero",
11626 rack, rsm));
11627 first_processed = 1;
11628 }
11629 if ((rsm->soff != soff) || (rsm->m != tm)) {
11630 /*
11631 * This is not a fatal error, we anticipate it
11632 * might happen (the else code), so we count it here
11633 * so that under invariant we can see that it really
11634 * does happen.
11635 */
11636 counter_u64_add(rack_adjust_map_bw, 1);
11637 }
11638 rsm->m = tm;
11639 rsm->soff = soff;
11640 if (tm) {
11641 rsm->orig_m_len = rsm->m->m_len;
11642 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11643 } else {
11644 rsm->orig_m_len = 0;
11645 rsm->orig_t_space = 0;
11646 }
11647 #else
11648 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
11649 if (rsm->m) {
11650 rsm->orig_m_len = rsm->m->m_len;
11651 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
11652 } else {
11653 rsm->orig_m_len = 0;
11654 rsm->orig_t_space = 0;
11655 }
11656 #endif
11657 rsm = tqhash_next(rack->r_ctl.tqh, rsm);
11658 if (rsm == NULL)
11659 break;
11660 }
11661 }
11662
11663 #ifdef TCP_REQUEST_TRK
11664 static inline void
rack_req_check_for_comp(struct tcp_rack * rack,tcp_seq th_ack)11665 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
11666 {
11667 struct tcp_sendfile_track *ent;
11668 int i;
11669
11670 if ((rack->rc_hybrid_mode == 0) &&
11671 (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
11672 /*
11673 * Just do normal completions hybrid pacing is not on
11674 * and CLDL is off as well.
11675 */
11676 tcp_req_check_for_comp(rack->rc_tp, th_ack);
11677 return;
11678 }
11679 /*
11680 * Originally I was just going to find the th_ack associated
11681 * with an entry. But then I realized a large strech ack could
11682 * in theory ack two or more requests at once. So instead we
11683 * need to find all entries that are completed by th_ack not
11684 * just a single entry and do our logging.
11685 */
11686 ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11687 while (ent != NULL) {
11688 /*
11689 * We may be doing hybrid pacing or CLDL and need more details possibly
11690 * so we do it manually instead of calling
11691 * tcp_req_check_for_comp()
11692 */
11693 uint64_t laa, tim, data, cbw, ftim;
11694
11695 /* Ok this ack frees it */
11696 rack_log_hybrid(rack, th_ack,
11697 ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
11698 rack_log_hybrid_sends(rack, ent, __LINE__);
11699 /* calculate the time based on the ack arrival */
11700 data = ent->end - ent->start;
11701 laa = tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time);
11702 if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
11703 if (ent->first_send > ent->localtime)
11704 ftim = ent->first_send;
11705 else
11706 ftim = ent->localtime;
11707 } else {
11708 /* TSNH */
11709 ftim = ent->localtime;
11710 }
11711 if (laa > ent->localtime)
11712 tim = laa - ftim;
11713 else
11714 tim = 0;
11715 cbw = data * HPTS_USEC_IN_SEC;
11716 if (tim > 0)
11717 cbw /= tim;
11718 else
11719 cbw = 0;
11720 rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
11721 /*
11722 * Check to see if we are freeing what we are pointing to send wise
11723 * if so be sure to NULL the pointer so we know we are no longer
11724 * set to anything.
11725 */
11726 if (ent == rack->r_ctl.rc_last_sft) {
11727 rack->r_ctl.rc_last_sft = NULL;
11728 if (rack->rc_hybrid_mode) {
11729 rack->rc_catch_up = 0;
11730 if (rack->cspr_is_fcc == 0)
11731 rack->r_ctl.bw_rate_cap = 0;
11732 else
11733 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
11734 rack->r_ctl.client_suggested_maxseg = 0;
11735 }
11736 }
11737 /* Generate the log that the tcp_netflix call would have */
11738 tcp_req_log_req_info(rack->rc_tp, ent,
11739 i, TCP_TRK_REQ_LOG_FREED, 0, 0);
11740 /* Free it and see if there is another one */
11741 tcp_req_free_a_slot(rack->rc_tp, ent);
11742 ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
11743 }
11744 }
11745 #endif
11746
11747
11748 /*
11749 * Return value of 1, we do not need to call rack_process_data().
11750 * return value of 0, rack_process_data can be called.
11751 * For ret_val if its 0 the TCP is locked, if its non-zero
11752 * its unlocked and probably unsafe to touch the TCB.
11753 */
11754 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)11755 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11756 struct tcpcb *tp, struct tcpopt *to,
11757 uint32_t tiwin, int32_t tlen,
11758 int32_t * ofia, int32_t thflags, int32_t *ret_val, int32_t orig_tlen)
11759 {
11760 int32_t ourfinisacked = 0;
11761 int32_t nsegs, acked_amount;
11762 int32_t acked;
11763 struct mbuf *mfree;
11764 struct tcp_rack *rack;
11765 int32_t under_pacing = 0;
11766 int32_t post_recovery = 0;
11767 uint32_t p_cwnd;
11768
11769 INP_WLOCK_ASSERT(tptoinpcb(tp));
11770
11771 rack = (struct tcp_rack *)tp->t_fb_ptr;
11772 if (SEQ_GEQ(tp->snd_una, tp->iss + (65535 << tp->snd_scale))) {
11773 /* Checking SEG.ACK against ISS is definitely redundant. */
11774 tp->t_flags2 |= TF2_NO_ISS_CHECK;
11775 }
11776 if (!V_tcp_insecure_ack) {
11777 tcp_seq seq_min;
11778 bool ghost_ack_check;
11779
11780 if (tp->t_flags2 & TF2_NO_ISS_CHECK) {
11781 /* Check for too old ACKs (RFC 5961, Section 5.2). */
11782 seq_min = tp->snd_una - tp->max_sndwnd;
11783 ghost_ack_check = false;
11784 } else {
11785 if (SEQ_GT(tp->iss + 1, tp->snd_una - tp->max_sndwnd)) {
11786 /* Checking for ghost ACKs is stricter. */
11787 seq_min = tp->iss + 1;
11788 ghost_ack_check = true;
11789 } else {
11790 /*
11791 * Checking for too old ACKs (RFC 5961,
11792 * Section 5.2) is stricter.
11793 */
11794 seq_min = tp->snd_una - tp->max_sndwnd;
11795 ghost_ack_check = false;
11796 }
11797 }
11798 if (SEQ_LT(th->th_ack, seq_min)) {
11799 if (ghost_ack_check)
11800 TCPSTAT_INC(tcps_rcvghostack);
11801 else
11802 TCPSTAT_INC(tcps_rcvacktooold);
11803 /* Send challenge ACK. */
11804 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11805 rack->r_wanted_output = 1;
11806 return (1);
11807 }
11808 }
11809 if (SEQ_GT(th->th_ack, tp->snd_max)) {
11810 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
11811 rack->r_wanted_output = 1;
11812 return (1);
11813 }
11814 if (rack->gp_ready &&
11815 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11816 under_pacing = 1;
11817 }
11818 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
11819 int in_rec, dup_ack_struck = 0;
11820 int dsack_seen = 0, sacks_seen = 0;
11821
11822 in_rec = IN_FASTRECOVERY(tp->t_flags);
11823 if (rack->rc_in_persist) {
11824 tp->t_rxtshift = 0;
11825 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11826 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11827 }
11828
11829 if ((th->th_ack == tp->snd_una) &&
11830 (tiwin == tp->snd_wnd) &&
11831 (orig_tlen == 0) &&
11832 ((to->to_flags & TOF_SACK) == 0)) {
11833 rack_strike_dupack(rack, th->th_ack);
11834 dup_ack_struck = 1;
11835 }
11836 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
11837 dup_ack_struck, &dsack_seen, &sacks_seen);
11838
11839 }
11840 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11841 /*
11842 * Old ack, behind (or duplicate to) the last one rcv'd
11843 * Note: We mark reordering is occuring if its
11844 * less than and we have not closed our window.
11845 */
11846 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
11847 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
11848 if (rack->r_ctl.rc_reorder_ts == 0)
11849 rack->r_ctl.rc_reorder_ts = 1;
11850 }
11851 return (0);
11852 }
11853 /*
11854 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
11855 * something we sent.
11856 */
11857 if (tp->t_flags & TF_NEEDSYN) {
11858 /*
11859 * T/TCP: Connection was half-synchronized, and our SYN has
11860 * been ACK'd (so connection is now fully synchronized). Go
11861 * to non-starred state, increment snd_una for ACK of SYN,
11862 * and check if we can do window scaling.
11863 */
11864 tp->t_flags &= ~TF_NEEDSYN;
11865 tp->snd_una++;
11866 /* Do window scaling? */
11867 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11868 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11869 tp->rcv_scale = tp->request_r_scale;
11870 /* Send window already scaled. */
11871 }
11872 }
11873 nsegs = max(1, m->m_pkthdr.lro_nsegs);
11874
11875 acked = BYTES_THIS_ACK(tp, th);
11876 if (acked) {
11877 /*
11878 * Any time we move the cum-ack forward clear
11879 * keep-alive tied probe-not-answered. The
11880 * persists clears its own on entry.
11881 */
11882 rack->probe_not_answered = 0;
11883 }
11884 KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11885 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11886 /*
11887 * If we just performed our first retransmit, and the ACK arrives
11888 * within our recovery window, then it was a mistake to do the
11889 * retransmit in the first place. Recover our original cwnd and
11890 * ssthresh, and proceed to transmit where we left off.
11891 */
11892 if ((tp->t_flags & TF_PREVVALID) &&
11893 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11894 tp->t_flags &= ~TF_PREVVALID;
11895 if (tp->t_rxtshift == 1 &&
11896 (int)(ticks - tp->t_badrxtwin) < 0)
11897 rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11898 }
11899 if (acked) {
11900 /* assure we are not backed off */
11901 tp->t_rxtshift = 0;
11902 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11903 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11904 rack->rc_tlp_in_progress = 0;
11905 rack->r_ctl.rc_tlp_cnt_out = 0;
11906 /*
11907 * If it is the RXT timer we want to
11908 * stop it, so we can restart a TLP.
11909 */
11910 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11911 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11912 #ifdef TCP_REQUEST_TRK
11913 rack_req_check_for_comp(rack, th->th_ack);
11914 #endif
11915 }
11916 /*
11917 * If we have a timestamp reply, update smoothed round trip time. If
11918 * no timestamp is present but transmit timer is running and timed
11919 * sequence number was acked, update smoothed round trip time. Since
11920 * we now have an rtt measurement, cancel the timer backoff (cf.,
11921 * Phil Karn's retransmit alg.). Recompute the initial retransmit
11922 * timer.
11923 *
11924 * Some boxes send broken timestamp replies during the SYN+ACK
11925 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11926 * and blow up the retransmit timer.
11927 */
11928 /*
11929 * If all outstanding data is acked, stop retransmit timer and
11930 * remember to restart (more output or persist). If there is more
11931 * data to be acked, restart retransmit timer, using current
11932 * (possibly backed-off) value.
11933 */
11934 if (acked == 0) {
11935 if (ofia)
11936 *ofia = ourfinisacked;
11937 return (0);
11938 }
11939 if (IN_RECOVERY(tp->t_flags)) {
11940 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
11941 (SEQ_LT(th->th_ack, tp->snd_max))) {
11942 tcp_rack_partialack(tp);
11943 } else {
11944 rack_post_recovery(tp, th->th_ack);
11945 post_recovery = 1;
11946 /*
11947 * Grab the segsiz, multiply by 2 and add the snd_cwnd
11948 * that is the max the CC should add if we are exiting
11949 * recovery and doing a late add.
11950 */
11951 p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
11952 p_cwnd <<= 1;
11953 p_cwnd += tp->snd_cwnd;
11954 }
11955 } else if ((rack->rto_from_rec == 1) &&
11956 SEQ_GEQ(th->th_ack, tp->snd_recover)) {
11957 /*
11958 * We were in recovery, hit a rxt timeout
11959 * and never re-entered recovery. The timeout(s)
11960 * made up all the lost data. In such a case
11961 * we need to clear the rto_from_rec flag.
11962 */
11963 rack->rto_from_rec = 0;
11964 }
11965 /*
11966 * Let the congestion control algorithm update congestion control
11967 * related information. This typically means increasing the
11968 * congestion window.
11969 */
11970 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, post_recovery);
11971 if (post_recovery &&
11972 (tp->snd_cwnd > p_cwnd)) {
11973 /* Must be non-newreno (cubic) getting too ahead of itself */
11974 tp->snd_cwnd = p_cwnd;
11975 }
11976 SOCK_SENDBUF_LOCK(so);
11977 acked_amount = min(acked, (int)sbavail(&so->so_snd));
11978 tp->snd_wnd -= acked_amount;
11979 mfree = sbcut_locked(&so->so_snd, acked_amount);
11980 if ((sbused(&so->so_snd) == 0) &&
11981 (acked > acked_amount) &&
11982 (tp->t_state >= TCPS_FIN_WAIT_1) &&
11983 (tp->t_flags & TF_SENTFIN)) {
11984 /*
11985 * We must be sure our fin
11986 * was sent and acked (we can be
11987 * in FIN_WAIT_1 without having
11988 * sent the fin).
11989 */
11990 ourfinisacked = 1;
11991 }
11992 tp->snd_una = th->th_ack;
11993 /* wakeups? */
11994 if (acked_amount && sbavail(&so->so_snd))
11995 rack_adjust_sendmap_head(rack, &so->so_snd);
11996 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11997 /* NB: sowwakeup_locked() does an implicit unlock. */
11998 sowwakeup_locked(so);
11999 m_freem(mfree);
12000 if (SEQ_GT(tp->snd_una, tp->snd_recover))
12001 tp->snd_recover = tp->snd_una;
12002
12003 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12004 tp->snd_nxt = tp->snd_max;
12005 }
12006 if (under_pacing &&
12007 (rack->use_fixed_rate == 0) &&
12008 (rack->in_probe_rtt == 0) &&
12009 rack->rc_gp_dyn_mul &&
12010 rack->rc_always_pace) {
12011 /* Check if we are dragging bottom */
12012 rack_check_bottom_drag(tp, rack, so);
12013 }
12014 if (tp->snd_una == tp->snd_max) {
12015 /* Nothing left outstanding */
12016 tp->t_flags &= ~TF_PREVVALID;
12017 if (rack->r_ctl.rc_went_idle_time == 0)
12018 rack->r_ctl.rc_went_idle_time = 1;
12019 rack->r_ctl.retran_during_recovery = 0;
12020 rack->r_ctl.dsack_byte_cnt = 0;
12021 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12022 if (sbavail(&tptosocket(tp)->so_snd) == 0)
12023 tp->t_acktime = 0;
12024 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12025 rack->rc_suspicious = 0;
12026 /* Set need output so persist might get set */
12027 rack->r_wanted_output = 1;
12028 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12029 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12030 (sbavail(&so->so_snd) == 0) &&
12031 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12032 /*
12033 * The socket was gone and the
12034 * peer sent data (now or in the past), time to
12035 * reset him.
12036 */
12037 *ret_val = 1;
12038 /* tcp_close will kill the inp pre-log the Reset */
12039 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12040 tp = tcp_close(tp);
12041 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12042 return (1);
12043 }
12044 }
12045 if (ofia)
12046 *ofia = ourfinisacked;
12047 return (0);
12048 }
12049
12050
12051 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)12052 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12053 int dir, uint32_t flags, struct rack_sendmap *rsm)
12054 {
12055 if (tcp_bblogging_on(rack->rc_tp)) {
12056 union tcp_log_stackspecific log;
12057 struct timeval tv;
12058
12059 memset(&log, 0, sizeof(log));
12060 log.u_bbr.flex1 = cnt;
12061 log.u_bbr.flex2 = split;
12062 log.u_bbr.flex3 = out;
12063 log.u_bbr.flex4 = line;
12064 log.u_bbr.flex5 = rack->r_must_retran;
12065 log.u_bbr.flex6 = flags;
12066 log.u_bbr.flex7 = rack->rc_has_collapsed;
12067 log.u_bbr.flex8 = dir; /*
12068 * 1 is collapsed, 0 is uncollapsed,
12069 * 2 is log of a rsm being marked, 3 is a split.
12070 */
12071 if (rsm == NULL)
12072 log.u_bbr.rttProp = 0;
12073 else
12074 log.u_bbr.rttProp = (uintptr_t)rsm;
12075 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12076 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12077 TCP_LOG_EVENTP(rack->rc_tp, NULL,
12078 &rack->rc_inp->inp_socket->so_rcv,
12079 &rack->rc_inp->inp_socket->so_snd,
12080 TCP_RACK_LOG_COLLAPSE, 0,
12081 0, &log, false, &tv);
12082 }
12083 }
12084
12085 static void
rack_collapsed_window(struct tcp_rack * rack,uint32_t out,tcp_seq th_ack,int line)12086 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12087 {
12088 /*
12089 * Here all we do is mark the collapsed point and set the flag.
12090 * This may happen again and again, but there is no
12091 * sense splitting our map until we know where the
12092 * peer finally lands in the collapse.
12093 */
12094 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12095 if ((rack->rc_has_collapsed == 0) ||
12096 (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12097 counter_u64_add(rack_collapsed_win_seen, 1);
12098 rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12099 rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12100 rack->rc_has_collapsed = 1;
12101 rack->r_collapse_point_valid = 1;
12102 rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12103 }
12104
12105 static void
rack_un_collapse_window(struct tcp_rack * rack,int line)12106 rack_un_collapse_window(struct tcp_rack *rack, int line)
12107 {
12108 struct rack_sendmap *nrsm, *rsm;
12109 int cnt = 0, split = 0;
12110 int insret __diagused;
12111
12112
12113 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12114 rack->rc_has_collapsed = 0;
12115 rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12116 if (rsm == NULL) {
12117 /* Nothing to do maybe the peer ack'ed it all */
12118 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12119 return;
12120 }
12121 /* Now do we need to split this one? */
12122 if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12123 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12124 rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12125 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12126 if (nrsm == NULL) {
12127 /* We can't get a rsm, mark all? */
12128 nrsm = rsm;
12129 goto no_split;
12130 }
12131 /* Clone it */
12132 split = 1;
12133 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12134 #ifndef INVARIANTS
12135 (void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12136 #else
12137 if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12138 panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12139 nrsm, insret, rack, rsm);
12140 }
12141 #endif
12142 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12143 rack->r_ctl.last_collapse_point, __LINE__);
12144 if (rsm->r_in_tmap) {
12145 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12146 nrsm->r_in_tmap = 1;
12147 }
12148 /*
12149 * Set in the new RSM as the
12150 * collapsed starting point
12151 */
12152 rsm = nrsm;
12153 }
12154
12155 no_split:
12156 TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm) {
12157 cnt++;
12158 nrsm->r_flags |= RACK_RWND_COLLAPSED;
12159 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12160 cnt++;
12161 }
12162 if (cnt) {
12163 counter_u64_add(rack_collapsed_win, 1);
12164 }
12165 rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12166 }
12167
12168 static void
rack_handle_delayed_ack(struct tcpcb * tp,struct tcp_rack * rack,int32_t tlen,int32_t tfo_syn)12169 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12170 int32_t tlen, int32_t tfo_syn)
12171 {
12172 if (DELAY_ACK(tp, tlen) || tfo_syn) {
12173 rack_timer_cancel(tp, rack,
12174 rack->r_ctl.rc_rcvtime, __LINE__);
12175 tp->t_flags |= TF_DELACK;
12176 } else {
12177 rack->r_wanted_output = 1;
12178 tp->t_flags |= TF_ACKNOW;
12179 }
12180 }
12181
12182 static void
rack_validate_fo_sendwin_up(struct tcpcb * tp,struct tcp_rack * rack)12183 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12184 {
12185 /*
12186 * If fast output is in progress, lets validate that
12187 * the new window did not shrink on us and make it
12188 * so fast output should end.
12189 */
12190 if (rack->r_fast_output) {
12191 uint32_t out;
12192
12193 /*
12194 * Calculate what we will send if left as is
12195 * and compare that to our send window.
12196 */
12197 out = ctf_outstanding(tp);
12198 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12199 /* ok we have an issue */
12200 if (out >= tp->snd_wnd) {
12201 /* Turn off fast output the window is met or collapsed */
12202 rack->r_fast_output = 0;
12203 } else {
12204 /* we have some room left */
12205 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12206 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12207 /* If not at least 1 full segment never mind */
12208 rack->r_fast_output = 0;
12209 }
12210 }
12211 }
12212 }
12213 }
12214
12215 /*
12216 * Return value of 1, the TCB is unlocked and most
12217 * likely gone, return value of 0, the TCP is still
12218 * locked.
12219 */
12220 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)12221 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12222 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12223 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12224 {
12225 /*
12226 * Update window information. Don't look at window if no ACK: TAC's
12227 * send garbage on first SYN.
12228 */
12229 int32_t nsegs;
12230 int32_t tfo_syn;
12231 struct tcp_rack *rack;
12232
12233 INP_WLOCK_ASSERT(tptoinpcb(tp));
12234
12235 rack = (struct tcp_rack *)tp->t_fb_ptr;
12236 nsegs = max(1, m->m_pkthdr.lro_nsegs);
12237 if ((thflags & TH_ACK) &&
12238 (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12239 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12240 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12241 /* keep track of pure window updates */
12242 if (tlen == 0 &&
12243 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12244 KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12245 tp->snd_wnd = tiwin;
12246 rack_validate_fo_sendwin_up(tp, rack);
12247 tp->snd_wl1 = th->th_seq;
12248 tp->snd_wl2 = th->th_ack;
12249 if (tp->snd_wnd > tp->max_sndwnd)
12250 tp->max_sndwnd = tp->snd_wnd;
12251 rack->r_wanted_output = 1;
12252 } else if (thflags & TH_ACK) {
12253 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12254 tp->snd_wnd = tiwin;
12255 rack_validate_fo_sendwin_up(tp, rack);
12256 tp->snd_wl1 = th->th_seq;
12257 tp->snd_wl2 = th->th_ack;
12258 }
12259 }
12260 if (tp->snd_wnd < ctf_outstanding(tp))
12261 /* The peer collapsed the window */
12262 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12263 else if (rack->rc_has_collapsed)
12264 rack_un_collapse_window(rack, __LINE__);
12265 if ((rack->r_collapse_point_valid) &&
12266 (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12267 rack->r_collapse_point_valid = 0;
12268 /* Was persist timer active and now we have window space? */
12269 if ((rack->rc_in_persist != 0) &&
12270 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12271 rack->r_ctl.rc_pace_min_segs))) {
12272 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12273 tp->snd_nxt = tp->snd_max;
12274 /* Make sure we output to start the timer */
12275 rack->r_wanted_output = 1;
12276 }
12277 /* Do we enter persists? */
12278 if ((rack->rc_in_persist == 0) &&
12279 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12280 TCPS_HAVEESTABLISHED(tp->t_state) &&
12281 ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12282 sbavail(&tptosocket(tp)->so_snd) &&
12283 (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12284 /*
12285 * Here the rwnd is less than
12286 * the pacing size, we are established,
12287 * nothing is outstanding, and there is
12288 * data to send. Enter persists.
12289 */
12290 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12291 }
12292 if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12293 m_freem(m);
12294 return (0);
12295 }
12296 /*
12297 * don't process the URG bit, ignore them drag
12298 * along the up.
12299 */
12300 tp->rcv_up = tp->rcv_nxt;
12301
12302 /*
12303 * Process the segment text, merging it into the TCP sequencing
12304 * queue, and arranging for acknowledgment of receipt if necessary.
12305 * This process logically involves adjusting tp->rcv_wnd as data is
12306 * presented to the user (this happens in tcp_usrreq.c, case
12307 * PRU_RCVD). If a FIN has already been received on this connection
12308 * then we just ignore the text.
12309 */
12310 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12311 (tp->t_flags & TF_FASTOPEN));
12312 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12313 TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12314 tcp_seq save_start = th->th_seq;
12315 tcp_seq save_rnxt = tp->rcv_nxt;
12316 int save_tlen = tlen;
12317
12318 m_adj(m, drop_hdrlen); /* delayed header drop */
12319 /*
12320 * Insert segment which includes th into TCP reassembly
12321 * queue with control block tp. Set thflags to whether
12322 * reassembly now includes a segment with FIN. This handles
12323 * the common case inline (segment is the next to be
12324 * received on an established connection, and the queue is
12325 * empty), avoiding linkage into and removal from the queue
12326 * and repetition of various conversions. Set DELACK for
12327 * segments received in order, but ack immediately when
12328 * segments are out of order (so fast retransmit can work).
12329 */
12330 if (th->th_seq == tp->rcv_nxt &&
12331 SEGQ_EMPTY(tp) &&
12332 (TCPS_HAVEESTABLISHED(tp->t_state) ||
12333 tfo_syn)) {
12334 #ifdef NETFLIX_SB_LIMITS
12335 u_int mcnt, appended;
12336
12337 if (so->so_rcv.sb_shlim) {
12338 mcnt = m_memcnt(m);
12339 appended = 0;
12340 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12341 CFO_NOSLEEP, NULL) == false) {
12342 counter_u64_add(tcp_sb_shlim_fails, 1);
12343 m_freem(m);
12344 return (0);
12345 }
12346 }
12347 #endif
12348 rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12349 tp->rcv_nxt += tlen;
12350 if (tlen &&
12351 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12352 (tp->t_fbyte_in == 0)) {
12353 tp->t_fbyte_in = ticks;
12354 if (tp->t_fbyte_in == 0)
12355 tp->t_fbyte_in = 1;
12356 if (tp->t_fbyte_out && tp->t_fbyte_in)
12357 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12358 }
12359 thflags = tcp_get_flags(th) & TH_FIN;
12360 KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12361 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12362 SOCK_RECVBUF_LOCK(so);
12363 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12364 m_freem(m);
12365 } else {
12366 int32_t newsize;
12367
12368 if (tlen > 0) {
12369 newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12370 if (newsize)
12371 if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12372 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12373 }
12374 #ifdef NETFLIX_SB_LIMITS
12375 appended =
12376 #endif
12377 sbappendstream_locked(&so->so_rcv, m, 0);
12378 }
12379 rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12380 /* NB: sorwakeup_locked() does an implicit unlock. */
12381 sorwakeup_locked(so);
12382 #ifdef NETFLIX_SB_LIMITS
12383 if (so->so_rcv.sb_shlim && appended != mcnt)
12384 counter_fo_release(so->so_rcv.sb_shlim,
12385 mcnt - appended);
12386 #endif
12387 } else {
12388 /*
12389 * XXX: Due to the header drop above "th" is
12390 * theoretically invalid by now. Fortunately
12391 * m_adj() doesn't actually frees any mbufs when
12392 * trimming from the head.
12393 */
12394 tcp_seq temp = save_start;
12395
12396 thflags = tcp_reass(tp, th, &temp, &tlen, m);
12397 tp->t_flags |= TF_ACKNOW;
12398 if (tp->t_flags & TF_WAKESOR) {
12399 tp->t_flags &= ~TF_WAKESOR;
12400 /* NB: sorwakeup_locked() does an implicit unlock. */
12401 sorwakeup_locked(so);
12402 }
12403 }
12404 if ((tp->t_flags & TF_SACK_PERMIT) &&
12405 (save_tlen > 0) &&
12406 TCPS_HAVEESTABLISHED(tp->t_state)) {
12407 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12408 /*
12409 * DSACK actually handled in the fastpath
12410 * above.
12411 */
12412 tcp_update_sack_list(tp, save_start,
12413 save_start + save_tlen);
12414 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12415 if ((tp->rcv_numsacks >= 1) &&
12416 (tp->sackblks[0].end == save_start)) {
12417 /*
12418 * Partial overlap, recorded at todrop
12419 * above.
12420 */
12421 tcp_update_sack_list(tp,
12422 tp->sackblks[0].start,
12423 tp->sackblks[0].end);
12424 } else {
12425 tcp_update_dsack_list(tp, save_start,
12426 save_start + save_tlen);
12427 }
12428 } else if (tlen >= save_tlen) {
12429 /* Update of sackblks. */
12430 tcp_update_dsack_list(tp, save_start,
12431 save_start + save_tlen);
12432 } else if (tlen > 0) {
12433 tcp_update_dsack_list(tp, save_start,
12434 save_start + tlen);
12435 }
12436 }
12437 } else {
12438 m_freem(m);
12439 thflags &= ~TH_FIN;
12440 }
12441
12442 /*
12443 * If FIN is received ACK the FIN and let the user know that the
12444 * connection is closing.
12445 */
12446 if (thflags & TH_FIN) {
12447 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12448 /* The socket upcall is handled by socantrcvmore. */
12449 socantrcvmore(so);
12450 /*
12451 * If connection is half-synchronized (ie NEEDSYN
12452 * flag on) then delay ACK, so it may be piggybacked
12453 * when SYN is sent. Otherwise, since we received a
12454 * FIN then no more input can be expected, send ACK
12455 * now.
12456 */
12457 if (tp->t_flags & TF_NEEDSYN) {
12458 rack_timer_cancel(tp, rack,
12459 rack->r_ctl.rc_rcvtime, __LINE__);
12460 tp->t_flags |= TF_DELACK;
12461 } else {
12462 tp->t_flags |= TF_ACKNOW;
12463 }
12464 tp->rcv_nxt++;
12465 }
12466 switch (tp->t_state) {
12467 /*
12468 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12469 * CLOSE_WAIT state.
12470 */
12471 case TCPS_SYN_RECEIVED:
12472 tp->t_starttime = ticks;
12473 /* FALLTHROUGH */
12474 case TCPS_ESTABLISHED:
12475 rack_timer_cancel(tp, rack,
12476 rack->r_ctl.rc_rcvtime, __LINE__);
12477 tcp_state_change(tp, TCPS_CLOSE_WAIT);
12478 break;
12479
12480 /*
12481 * If still in FIN_WAIT_1 STATE FIN has not been
12482 * acked so enter the CLOSING state.
12483 */
12484 case TCPS_FIN_WAIT_1:
12485 rack_timer_cancel(tp, rack,
12486 rack->r_ctl.rc_rcvtime, __LINE__);
12487 tcp_state_change(tp, TCPS_CLOSING);
12488 break;
12489
12490 /*
12491 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12492 * starting the time-wait timer, turning off the
12493 * other standard timers.
12494 */
12495 case TCPS_FIN_WAIT_2:
12496 rack_timer_cancel(tp, rack,
12497 rack->r_ctl.rc_rcvtime, __LINE__);
12498 tcp_twstart(tp);
12499 return (1);
12500 }
12501 }
12502 /*
12503 * Return any desired output.
12504 */
12505 if ((tp->t_flags & TF_ACKNOW) ||
12506 (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12507 rack->r_wanted_output = 1;
12508 }
12509 return (0);
12510 }
12511
12512 /*
12513 * Here nothing is really faster, its just that we
12514 * have broken out the fast-data path also just like
12515 * the fast-ack.
12516 */
12517 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)12518 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12519 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12520 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12521 {
12522 int32_t nsegs;
12523 int32_t newsize = 0; /* automatic sockbuf scaling */
12524 struct tcp_rack *rack;
12525 #ifdef NETFLIX_SB_LIMITS
12526 u_int mcnt, appended;
12527 #endif
12528
12529 /*
12530 * If last ACK falls within this segment's sequence numbers, record
12531 * the timestamp. NOTE that the test is modified according to the
12532 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12533 */
12534 if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12535 return (0);
12536 }
12537 if (tiwin && tiwin != tp->snd_wnd) {
12538 return (0);
12539 }
12540 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12541 return (0);
12542 }
12543 if (__predict_false((to->to_flags & TOF_TS) &&
12544 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12545 return (0);
12546 }
12547 if (__predict_false((th->th_ack != tp->snd_una))) {
12548 return (0);
12549 }
12550 if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12551 return (0);
12552 }
12553 if ((to->to_flags & TOF_TS) != 0 &&
12554 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12555 tp->ts_recent_age = tcp_ts_getticks();
12556 tp->ts_recent = to->to_tsval;
12557 }
12558 rack = (struct tcp_rack *)tp->t_fb_ptr;
12559 /*
12560 * This is a pure, in-sequence data packet with nothing on the
12561 * reassembly queue and we have enough buffer space to take it.
12562 */
12563 nsegs = max(1, m->m_pkthdr.lro_nsegs);
12564
12565 #ifdef NETFLIX_SB_LIMITS
12566 if (so->so_rcv.sb_shlim) {
12567 mcnt = m_memcnt(m);
12568 appended = 0;
12569 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12570 CFO_NOSLEEP, NULL) == false) {
12571 counter_u64_add(tcp_sb_shlim_fails, 1);
12572 m_freem(m);
12573 return (1);
12574 }
12575 }
12576 #endif
12577 /* Clean receiver SACK report if present */
12578 if (tp->rcv_numsacks)
12579 tcp_clean_sackreport(tp);
12580 KMOD_TCPSTAT_INC(tcps_preddat);
12581 tp->rcv_nxt += tlen;
12582 if (tlen &&
12583 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12584 (tp->t_fbyte_in == 0)) {
12585 tp->t_fbyte_in = ticks;
12586 if (tp->t_fbyte_in == 0)
12587 tp->t_fbyte_in = 1;
12588 if (tp->t_fbyte_out && tp->t_fbyte_in)
12589 tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12590 }
12591 /*
12592 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12593 */
12594 tp->snd_wl1 = th->th_seq;
12595 /*
12596 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12597 */
12598 tp->rcv_up = tp->rcv_nxt;
12599 KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12600 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12601 newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12602
12603 /* Add data to socket buffer. */
12604 SOCK_RECVBUF_LOCK(so);
12605 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12606 m_freem(m);
12607 } else {
12608 /*
12609 * Set new socket buffer size. Give up when limit is
12610 * reached.
12611 */
12612 if (newsize)
12613 if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12614 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12615 m_adj(m, drop_hdrlen); /* delayed header drop */
12616 #ifdef NETFLIX_SB_LIMITS
12617 appended =
12618 #endif
12619 sbappendstream_locked(&so->so_rcv, m, 0);
12620 ctf_calc_rwin(so, tp);
12621 }
12622 rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12623 /* NB: sorwakeup_locked() does an implicit unlock. */
12624 sorwakeup_locked(so);
12625 #ifdef NETFLIX_SB_LIMITS
12626 if (so->so_rcv.sb_shlim && mcnt != appended)
12627 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12628 #endif
12629 rack_handle_delayed_ack(tp, rack, tlen, 0);
12630 if (tp->snd_una == tp->snd_max)
12631 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12632 return (1);
12633 }
12634
12635 /*
12636 * This subfunction is used to try to highly optimize the
12637 * fast path. We again allow window updates that are
12638 * in sequence to remain in the fast-path. We also add
12639 * in the __predict's to attempt to help the compiler.
12640 * Note that if we return a 0, then we can *not* process
12641 * it and the caller should push the packet into the
12642 * slow-path.
12643 */
12644 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)12645 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12646 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12647 uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12648 {
12649 int32_t acked;
12650 int32_t nsegs;
12651 int32_t under_pacing = 0;
12652 struct tcp_rack *rack;
12653
12654 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12655 /* Old ack, behind (or duplicate to) the last one rcv'd */
12656 return (0);
12657 }
12658 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
12659 /* Above what we have sent? */
12660 return (0);
12661 }
12662 if (__predict_false(tiwin == 0)) {
12663 /* zero window */
12664 return (0);
12665 }
12666 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
12667 /* We need a SYN or a FIN, unlikely.. */
12668 return (0);
12669 }
12670 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
12671 /* Timestamp is behind .. old ack with seq wrap? */
12672 return (0);
12673 }
12674 if (__predict_false(IN_RECOVERY(tp->t_flags))) {
12675 /* Still recovering */
12676 return (0);
12677 }
12678 rack = (struct tcp_rack *)tp->t_fb_ptr;
12679 if (rack->r_ctl.rc_sacked) {
12680 /* We have sack holes on our scoreboard */
12681 return (0);
12682 }
12683 /* Ok if we reach here, we can process a fast-ack */
12684 if (rack->gp_ready &&
12685 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12686 under_pacing = 1;
12687 }
12688 nsegs = max(1, m->m_pkthdr.lro_nsegs);
12689 rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
12690 /* Did the window get updated? */
12691 if (tiwin != tp->snd_wnd) {
12692 tp->snd_wnd = tiwin;
12693 rack_validate_fo_sendwin_up(tp, rack);
12694 tp->snd_wl1 = th->th_seq;
12695 if (tp->snd_wnd > tp->max_sndwnd)
12696 tp->max_sndwnd = tp->snd_wnd;
12697 }
12698 /* Do we exit persists? */
12699 if ((rack->rc_in_persist != 0) &&
12700 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12701 rack->r_ctl.rc_pace_min_segs))) {
12702 rack_exit_persist(tp, rack, cts);
12703 }
12704 /* Do we enter persists? */
12705 if ((rack->rc_in_persist == 0) &&
12706 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12707 TCPS_HAVEESTABLISHED(tp->t_state) &&
12708 ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12709 sbavail(&tptosocket(tp)->so_snd) &&
12710 (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12711 /*
12712 * Here the rwnd is less than
12713 * the pacing size, we are established,
12714 * nothing is outstanding, and there is
12715 * data to send. Enter persists.
12716 */
12717 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
12718 }
12719 /*
12720 * If last ACK falls within this segment's sequence numbers, record
12721 * the timestamp. NOTE that the test is modified according to the
12722 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12723 */
12724 if ((to->to_flags & TOF_TS) != 0 &&
12725 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12726 tp->ts_recent_age = tcp_ts_getticks();
12727 tp->ts_recent = to->to_tsval;
12728 }
12729 /*
12730 * This is a pure ack for outstanding data.
12731 */
12732 KMOD_TCPSTAT_INC(tcps_predack);
12733
12734 /*
12735 * "bad retransmit" recovery.
12736 */
12737 if ((tp->t_flags & TF_PREVVALID) &&
12738 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12739 tp->t_flags &= ~TF_PREVVALID;
12740 if (tp->t_rxtshift == 1 &&
12741 (int)(ticks - tp->t_badrxtwin) < 0)
12742 rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12743 }
12744 /*
12745 * Recalculate the transmit timer / rtt.
12746 *
12747 * Some boxes send broken timestamp replies during the SYN+ACK
12748 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12749 * and blow up the retransmit timer.
12750 */
12751 acked = BYTES_THIS_ACK(tp, th);
12752
12753 #ifdef TCP_HHOOK
12754 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
12755 hhook_run_tcp_est_in(tp, th, to);
12756 #endif
12757 KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12758 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12759 if (acked) {
12760 struct mbuf *mfree;
12761
12762 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
12763 SOCK_SENDBUF_LOCK(so);
12764 mfree = sbcut_locked(&so->so_snd, acked);
12765 tp->snd_una = th->th_ack;
12766 /* Note we want to hold the sb lock through the sendmap adjust */
12767 rack_adjust_sendmap_head(rack, &so->so_snd);
12768 /* Wake up the socket if we have room to write more */
12769 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12770 sowwakeup_locked(so);
12771 m_freem(mfree);
12772 tp->t_rxtshift = 0;
12773 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12774 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12775 rack->rc_tlp_in_progress = 0;
12776 rack->r_ctl.rc_tlp_cnt_out = 0;
12777 /*
12778 * If it is the RXT timer we want to
12779 * stop it, so we can restart a TLP.
12780 */
12781 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12782 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12783
12784 #ifdef TCP_REQUEST_TRK
12785 rack_req_check_for_comp(rack, th->th_ack);
12786 #endif
12787 }
12788 /*
12789 * Let the congestion control algorithm update congestion control
12790 * related information. This typically means increasing the
12791 * congestion window.
12792 */
12793 if (tp->snd_wnd < ctf_outstanding(tp)) {
12794 /* The peer collapsed the window */
12795 rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12796 } else if (rack->rc_has_collapsed)
12797 rack_un_collapse_window(rack, __LINE__);
12798 if ((rack->r_collapse_point_valid) &&
12799 (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
12800 rack->r_collapse_point_valid = 0;
12801 /*
12802 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
12803 */
12804 tp->snd_wl2 = th->th_ack;
12805 tp->t_dupacks = 0;
12806 m_freem(m);
12807 /* ND6_HINT(tp); *//* Some progress has been made. */
12808
12809 /*
12810 * If all outstanding data are acked, stop retransmit timer,
12811 * otherwise restart timer using current (possibly backed-off)
12812 * value. If process is waiting for space, wakeup/selwakeup/signal.
12813 * If data are ready to send, let tcp_output decide between more
12814 * output or persist.
12815 */
12816 if (under_pacing &&
12817 (rack->use_fixed_rate == 0) &&
12818 (rack->in_probe_rtt == 0) &&
12819 rack->rc_gp_dyn_mul &&
12820 rack->rc_always_pace) {
12821 /* Check if we are dragging bottom */
12822 rack_check_bottom_drag(tp, rack, so);
12823 }
12824 if (tp->snd_una == tp->snd_max) {
12825 tp->t_flags &= ~TF_PREVVALID;
12826 rack->r_ctl.retran_during_recovery = 0;
12827 rack->rc_suspicious = 0;
12828 rack->r_ctl.dsack_byte_cnt = 0;
12829 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12830 if (rack->r_ctl.rc_went_idle_time == 0)
12831 rack->r_ctl.rc_went_idle_time = 1;
12832 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12833 if (sbavail(&tptosocket(tp)->so_snd) == 0)
12834 tp->t_acktime = 0;
12835 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12836 }
12837 if (acked && rack->r_fast_output)
12838 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
12839 if (sbavail(&so->so_snd)) {
12840 rack->r_wanted_output = 1;
12841 }
12842 return (1);
12843 }
12844
12845 /*
12846 * Return value of 1, the TCB is unlocked and most
12847 * likely gone, return value of 0, the TCP is still
12848 * locked.
12849 */
12850 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)12851 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
12852 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12853 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12854 {
12855 int32_t ret_val = 0;
12856 int32_t orig_tlen = tlen;
12857 int32_t todrop;
12858 int32_t ourfinisacked = 0;
12859 struct tcp_rack *rack;
12860
12861 INP_WLOCK_ASSERT(tptoinpcb(tp));
12862
12863 ctf_calc_rwin(so, tp);
12864 /*
12865 * If the state is SYN_SENT: if seg contains an ACK, but not for our
12866 * SYN, drop the input. if seg contains a RST, then drop the
12867 * connection. if seg does not contain SYN, then drop it. Otherwise
12868 * this is an acceptable SYN segment initialize tp->rcv_nxt and
12869 * tp->irs if seg contains ack then advance tp->snd_una if seg
12870 * contains an ECE and ECN support is enabled, the stream is ECN
12871 * capable. if SYN has been acked change to ESTABLISHED else
12872 * SYN_RCVD state arrange for segment to be acked (eventually)
12873 * continue processing rest of data/controls.
12874 */
12875 if ((thflags & TH_ACK) &&
12876 (SEQ_LEQ(th->th_ack, tp->iss) ||
12877 SEQ_GT(th->th_ack, tp->snd_max))) {
12878 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
12879 ctf_do_dropwithreset(m, tp, th, BANDLIM_TCP_RST, tlen);
12880 return (1);
12881 }
12882 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
12883 TCP_PROBE5(connect__refused, NULL, tp,
12884 mtod(m, const char *), tp, th);
12885 tp = tcp_drop(tp, ECONNREFUSED);
12886 ctf_do_drop(m, tp);
12887 return (1);
12888 }
12889 if (thflags & TH_RST) {
12890 ctf_do_drop(m, tp);
12891 return (1);
12892 }
12893 if (!(thflags & TH_SYN)) {
12894 ctf_do_drop(m, tp);
12895 return (1);
12896 }
12897 tp->irs = th->th_seq;
12898 tcp_rcvseqinit(tp);
12899 rack = (struct tcp_rack *)tp->t_fb_ptr;
12900 if (thflags & TH_ACK) {
12901 int tfo_partial = 0;
12902
12903 KMOD_TCPSTAT_INC(tcps_connects);
12904 soisconnected(so);
12905 #ifdef MAC
12906 mac_socketpeer_set_from_mbuf(m, so);
12907 #endif
12908 /* Do window scaling on this connection? */
12909 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12910 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12911 tp->rcv_scale = tp->request_r_scale;
12912 }
12913 tp->rcv_adv += min(tp->rcv_wnd,
12914 TCP_MAXWIN << tp->rcv_scale);
12915 /*
12916 * If not all the data that was sent in the TFO SYN
12917 * has been acked, resend the remainder right away.
12918 */
12919 if ((tp->t_flags & TF_FASTOPEN) &&
12920 (tp->snd_una != tp->snd_max)) {
12921 /* Was it a partial ack? */
12922 if (SEQ_LT(th->th_ack, tp->snd_max))
12923 tfo_partial = 1;
12924 }
12925 /*
12926 * If there's data, delay ACK; if there's also a FIN ACKNOW
12927 * will be turned on later.
12928 */
12929 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
12930 rack_timer_cancel(tp, rack,
12931 rack->r_ctl.rc_rcvtime, __LINE__);
12932 tp->t_flags |= TF_DELACK;
12933 } else {
12934 rack->r_wanted_output = 1;
12935 tp->t_flags |= TF_ACKNOW;
12936 }
12937
12938 tcp_ecn_input_syn_sent(tp, thflags, iptos);
12939
12940 if (SEQ_GT(th->th_ack, tp->snd_una)) {
12941 /*
12942 * We advance snd_una for the
12943 * fast open case. If th_ack is
12944 * acknowledging data beyond
12945 * snd_una we can't just call
12946 * ack-processing since the
12947 * data stream in our send-map
12948 * will start at snd_una + 1 (one
12949 * beyond the SYN). If its just
12950 * equal we don't need to do that
12951 * and there is no send_map.
12952 */
12953 tp->snd_una++;
12954 if (tfo_partial && (SEQ_GT(tp->snd_max, tp->snd_una))) {
12955 /*
12956 * We sent a SYN with data, and thus have a
12957 * sendmap entry with a SYN set. Lets find it
12958 * and take off the send bit and the byte and
12959 * set it up to be what we send (send it next).
12960 */
12961 struct rack_sendmap *rsm;
12962
12963 rsm = tqhash_min(rack->r_ctl.tqh);
12964 if (rsm) {
12965 if (rsm->r_flags & RACK_HAS_SYN) {
12966 rsm->r_flags &= ~RACK_HAS_SYN;
12967 rsm->r_start++;
12968 }
12969 rack->r_ctl.rc_resend = rsm;
12970 }
12971 }
12972 }
12973 /*
12974 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
12975 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
12976 */
12977 tp->t_starttime = ticks;
12978 if (tp->t_flags & TF_NEEDFIN) {
12979 tcp_state_change(tp, TCPS_FIN_WAIT_1);
12980 tp->t_flags &= ~TF_NEEDFIN;
12981 thflags &= ~TH_SYN;
12982 } else {
12983 tcp_state_change(tp, TCPS_ESTABLISHED);
12984 TCP_PROBE5(connect__established, NULL, tp,
12985 mtod(m, const char *), tp, th);
12986 rack_cc_conn_init(tp);
12987 }
12988 } else {
12989 /*
12990 * Received initial SYN in SYN-SENT[*] state => simultaneous
12991 * open. If segment contains CC option and there is a
12992 * cached CC, apply TAO test. If it succeeds, connection is *
12993 * half-synchronized. Otherwise, do 3-way handshake:
12994 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
12995 * there was no CC option, clear cached CC value.
12996 */
12997 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
12998 tcp_state_change(tp, TCPS_SYN_RECEIVED);
12999 }
13000 /*
13001 * Advance th->th_seq to correspond to first data byte. If data,
13002 * trim to stay within window, dropping FIN if necessary.
13003 */
13004 th->th_seq++;
13005 if (tlen > tp->rcv_wnd) {
13006 todrop = tlen - tp->rcv_wnd;
13007 m_adj(m, -todrop);
13008 tlen = tp->rcv_wnd;
13009 thflags &= ~TH_FIN;
13010 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13011 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13012 }
13013 tp->snd_wl1 = th->th_seq - 1;
13014 tp->rcv_up = th->th_seq;
13015 /*
13016 * Client side of transaction: already sent SYN and data. If the
13017 * remote host used T/TCP to validate the SYN, our data will be
13018 * ACK'd; if so, enter normal data segment processing in the middle
13019 * of step 5, ack processing. Otherwise, goto step 6.
13020 */
13021 if (thflags & TH_ACK) {
13022 /* For syn-sent we need to possibly update the rtt */
13023 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13024 uint32_t t, mcts;
13025
13026 mcts = tcp_ts_getticks();
13027 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13028 if (!tp->t_rttlow || tp->t_rttlow > t)
13029 tp->t_rttlow = t;
13030 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13031 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13032 tcp_rack_xmit_timer_commit(rack, tp);
13033 }
13034 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen))
13035 return (ret_val);
13036 /* We may have changed to FIN_WAIT_1 above */
13037 if (tp->t_state == TCPS_FIN_WAIT_1) {
13038 /*
13039 * In FIN_WAIT_1 STATE in addition to the processing
13040 * for the ESTABLISHED state if our FIN is now
13041 * acknowledged then enter FIN_WAIT_2.
13042 */
13043 if (ourfinisacked) {
13044 /*
13045 * If we can't receive any more data, then
13046 * closing user can proceed. Starting the
13047 * timer is contrary to the specification,
13048 * but if we don't get a FIN we'll hang
13049 * forever.
13050 *
13051 * XXXjl: we should release the tp also, and
13052 * use a compressed state.
13053 */
13054 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13055 soisdisconnected(so);
13056 tcp_timer_activate(tp, TT_2MSL,
13057 (tcp_fast_finwait2_recycle ?
13058 tcp_finwait2_timeout :
13059 TP_MAXIDLE(tp)));
13060 }
13061 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13062 }
13063 }
13064 }
13065 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13066 tiwin, thflags, nxt_pkt));
13067 }
13068
13069 /*
13070 * Return value of 1, the TCB is unlocked and most
13071 * likely gone, return value of 0, the TCP is still
13072 * locked.
13073 */
13074 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)13075 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13076 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13077 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13078 {
13079 struct tcp_rack *rack;
13080 int32_t orig_tlen = tlen;
13081 int32_t ret_val = 0;
13082 int32_t ourfinisacked = 0;
13083
13084 rack = (struct tcp_rack *)tp->t_fb_ptr;
13085 ctf_calc_rwin(so, tp);
13086 if ((thflags & TH_RST) ||
13087 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13088 return (ctf_process_rst(m, th, so, tp));
13089 if ((thflags & TH_ACK) &&
13090 (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13091 SEQ_GT(th->th_ack, tp->snd_max))) {
13092 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13093 ctf_do_dropwithreset(m, tp, th, BANDLIM_TCP_RST, tlen);
13094 return (1);
13095 }
13096 if (tp->t_flags & TF_FASTOPEN) {
13097 /*
13098 * When a TFO connection is in SYN_RECEIVED, the
13099 * only valid packets are the initial SYN, a
13100 * retransmit/copy of the initial SYN (possibly with
13101 * a subset of the original data), a valid ACK, a
13102 * FIN, or a RST.
13103 */
13104 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13105 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13106 ctf_do_dropwithreset(m, tp, th, BANDLIM_TCP_RST, tlen);
13107 return (1);
13108 } else if (thflags & TH_SYN) {
13109 /* non-initial SYN is ignored */
13110 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13111 (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13112 (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13113 ctf_do_drop(m, NULL);
13114 return (0);
13115 }
13116 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13117 ctf_do_drop(m, NULL);
13118 return (0);
13119 }
13120 }
13121
13122 /*
13123 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13124 * it's less than ts_recent, drop it.
13125 */
13126 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13127 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13128 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13129 return (ret_val);
13130 }
13131 /*
13132 * In the SYN-RECEIVED state, validate that the packet belongs to
13133 * this connection before trimming the data to fit the receive
13134 * window. Check the sequence number versus IRS since we know the
13135 * sequence numbers haven't wrapped. This is a partial fix for the
13136 * "LAND" DoS attack.
13137 */
13138 if (SEQ_LT(th->th_seq, tp->irs)) {
13139 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13140 ctf_do_dropwithreset(m, tp, th, BANDLIM_TCP_RST, tlen);
13141 return (1);
13142 }
13143 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13144 return (ret_val);
13145 }
13146 /*
13147 * If last ACK falls within this segment's sequence numbers, record
13148 * its timestamp. NOTE: 1) That the test incorporates suggestions
13149 * from the latest proposal of the tcplw@cray.com list (Braden
13150 * 1993/04/26). 2) That updating only on newer timestamps interferes
13151 * with our earlier PAWS tests, so this check should be solely
13152 * predicated on the sequence space of this segment. 3) That we
13153 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13154 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13155 * SEG.Len, This modified check allows us to overcome RFC1323's
13156 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13157 * p.869. In such cases, we can still calculate the RTT correctly
13158 * when RCV.NXT == Last.ACK.Sent.
13159 */
13160 if ((to->to_flags & TOF_TS) != 0 &&
13161 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13162 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13163 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13164 tp->ts_recent_age = tcp_ts_getticks();
13165 tp->ts_recent = to->to_tsval;
13166 }
13167 tp->snd_wnd = tiwin;
13168 rack_validate_fo_sendwin_up(tp, rack);
13169 /*
13170 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13171 * is on (half-synchronized state), then queue data for later
13172 * processing; else drop segment and return.
13173 */
13174 if ((thflags & TH_ACK) == 0) {
13175 if (tp->t_flags & TF_FASTOPEN) {
13176 rack_cc_conn_init(tp);
13177 }
13178 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13179 tiwin, thflags, nxt_pkt));
13180 }
13181 KMOD_TCPSTAT_INC(tcps_connects);
13182 if (tp->t_flags & TF_SONOTCONN) {
13183 tp->t_flags &= ~TF_SONOTCONN;
13184 soisconnected(so);
13185 }
13186 /* Do window scaling? */
13187 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13188 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13189 tp->rcv_scale = tp->request_r_scale;
13190 }
13191 /*
13192 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* ->
13193 * FIN-WAIT-1
13194 */
13195 tp->t_starttime = ticks;
13196 if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) {
13197 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13198 tp->t_tfo_pending = NULL;
13199 }
13200 if (tp->t_flags & TF_NEEDFIN) {
13201 tcp_state_change(tp, TCPS_FIN_WAIT_1);
13202 tp->t_flags &= ~TF_NEEDFIN;
13203 } else {
13204 tcp_state_change(tp, TCPS_ESTABLISHED);
13205 TCP_PROBE5(accept__established, NULL, tp,
13206 mtod(m, const char *), tp, th);
13207 /*
13208 * TFO connections call cc_conn_init() during SYN
13209 * processing. Calling it again here for such connections
13210 * is not harmless as it would undo the snd_cwnd reduction
13211 * that occurs when a TFO SYN|ACK is retransmitted.
13212 */
13213 if (!(tp->t_flags & TF_FASTOPEN))
13214 rack_cc_conn_init(tp);
13215 }
13216 /*
13217 * Account for the ACK of our SYN prior to
13218 * regular ACK processing below, except for
13219 * simultaneous SYN, which is handled later.
13220 */
13221 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13222 tp->snd_una++;
13223 /*
13224 * If segment contains data or ACK, will call tcp_reass() later; if
13225 * not, do so now to pass queued data to user.
13226 */
13227 if (tlen == 0 && (thflags & TH_FIN) == 0) {
13228 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13229 (struct mbuf *)0);
13230 if (tp->t_flags & TF_WAKESOR) {
13231 tp->t_flags &= ~TF_WAKESOR;
13232 /* NB: sorwakeup_locked() does an implicit unlock. */
13233 sorwakeup_locked(so);
13234 }
13235 }
13236 tp->snd_wl1 = th->th_seq - 1;
13237 /* For syn-recv we need to possibly update the rtt */
13238 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13239 uint32_t t, mcts;
13240
13241 mcts = tcp_ts_getticks();
13242 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13243 if (!tp->t_rttlow || tp->t_rttlow > t)
13244 tp->t_rttlow = t;
13245 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13246 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13247 tcp_rack_xmit_timer_commit(rack, tp);
13248 }
13249 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13250 return (ret_val);
13251 }
13252 if (tp->t_state == TCPS_FIN_WAIT_1) {
13253 /* We could have went to FIN_WAIT_1 (or EST) above */
13254 /*
13255 * In FIN_WAIT_1 STATE in addition to the processing for the
13256 * ESTABLISHED state if our FIN is now acknowledged then
13257 * enter FIN_WAIT_2.
13258 */
13259 if (ourfinisacked) {
13260 /*
13261 * If we can't receive any more data, then closing
13262 * user can proceed. Starting the timer is contrary
13263 * to the specification, but if we don't get a FIN
13264 * we'll hang forever.
13265 *
13266 * XXXjl: we should release the tp also, and use a
13267 * compressed state.
13268 */
13269 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13270 soisdisconnected(so);
13271 tcp_timer_activate(tp, TT_2MSL,
13272 (tcp_fast_finwait2_recycle ?
13273 tcp_finwait2_timeout :
13274 TP_MAXIDLE(tp)));
13275 }
13276 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13277 }
13278 }
13279 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13280 tiwin, thflags, nxt_pkt));
13281 }
13282
13283 /*
13284 * Return value of 1, the TCB is unlocked and most
13285 * likely gone, return value of 0, the TCP is still
13286 * locked.
13287 */
13288 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)13289 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13290 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13291 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13292 {
13293 int32_t ret_val = 0;
13294 int32_t orig_tlen = tlen;
13295 struct tcp_rack *rack;
13296
13297 /*
13298 * Header prediction: check for the two common cases of a
13299 * uni-directional data xfer. If the packet has no control flags,
13300 * is in-sequence, the window didn't change and we're not
13301 * retransmitting, it's a candidate. If the length is zero and the
13302 * ack moved forward, we're the sender side of the xfer. Just free
13303 * the data acked & wake any higher level process that was blocked
13304 * waiting for space. If the length is non-zero and the ack didn't
13305 * move, we're the receiver side. If we're getting packets in-order
13306 * (the reassembly queue is empty), add the data toc The socket
13307 * buffer and note that we need a delayed ack. Make sure that the
13308 * hidden state-flags are also off. Since we check for
13309 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13310 */
13311 rack = (struct tcp_rack *)tp->t_fb_ptr;
13312 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13313 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13314 __predict_true(SEGQ_EMPTY(tp)) &&
13315 __predict_true(th->th_seq == tp->rcv_nxt)) {
13316 if (tlen == 0) {
13317 if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13318 tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13319 return (0);
13320 }
13321 } else {
13322 if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13323 tiwin, nxt_pkt, iptos)) {
13324 return (0);
13325 }
13326 }
13327 }
13328 ctf_calc_rwin(so, tp);
13329
13330 if ((thflags & TH_RST) ||
13331 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13332 return (ctf_process_rst(m, th, so, tp));
13333
13334 /*
13335 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13336 * synchronized state.
13337 */
13338 if (thflags & TH_SYN) {
13339 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13340 return (ret_val);
13341 }
13342 /*
13343 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13344 * it's less than ts_recent, drop it.
13345 */
13346 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13347 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13348 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13349 return (ret_val);
13350 }
13351 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13352 return (ret_val);
13353 }
13354 /*
13355 * If last ACK falls within this segment's sequence numbers, record
13356 * its timestamp. NOTE: 1) That the test incorporates suggestions
13357 * from the latest proposal of the tcplw@cray.com list (Braden
13358 * 1993/04/26). 2) That updating only on newer timestamps interferes
13359 * with our earlier PAWS tests, so this check should be solely
13360 * predicated on the sequence space of this segment. 3) That we
13361 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13362 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13363 * SEG.Len, This modified check allows us to overcome RFC1323's
13364 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13365 * p.869. In such cases, we can still calculate the RTT correctly
13366 * when RCV.NXT == Last.ACK.Sent.
13367 */
13368 if ((to->to_flags & TOF_TS) != 0 &&
13369 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13370 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13371 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13372 tp->ts_recent_age = tcp_ts_getticks();
13373 tp->ts_recent = to->to_tsval;
13374 }
13375 /*
13376 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13377 * is on (half-synchronized state), then queue data for later
13378 * processing; else drop segment and return.
13379 */
13380 if ((thflags & TH_ACK) == 0) {
13381 if (tp->t_flags & TF_NEEDSYN) {
13382 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13383 tiwin, thflags, nxt_pkt));
13384
13385 } else if (tp->t_flags & TF_ACKNOW) {
13386 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13387 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13388 return (ret_val);
13389 } else {
13390 ctf_do_drop(m, NULL);
13391 return (0);
13392 }
13393 }
13394 /*
13395 * Ack processing.
13396 */
13397 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13398 return (ret_val);
13399 }
13400 if (sbavail(&so->so_snd)) {
13401 if (ctf_progress_timeout_check(tp, true)) {
13402 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13403 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_TCP_RST, tlen);
13404 return (1);
13405 }
13406 }
13407 /* State changes only happen in rack_process_data() */
13408 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13409 tiwin, thflags, nxt_pkt));
13410 }
13411
13412 /*
13413 * Return value of 1, the TCB is unlocked and most
13414 * likely gone, return value of 0, the TCP is still
13415 * locked.
13416 */
13417 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)13418 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13419 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13420 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13421 {
13422 int32_t ret_val = 0;
13423 int32_t orig_tlen = tlen;
13424
13425 ctf_calc_rwin(so, tp);
13426 if ((thflags & TH_RST) ||
13427 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13428 return (ctf_process_rst(m, th, so, tp));
13429 /*
13430 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13431 * synchronized state.
13432 */
13433 if (thflags & TH_SYN) {
13434 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13435 return (ret_val);
13436 }
13437 /*
13438 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13439 * it's less than ts_recent, drop it.
13440 */
13441 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13442 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13443 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13444 return (ret_val);
13445 }
13446 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13447 return (ret_val);
13448 }
13449 /*
13450 * If last ACK falls within this segment's sequence numbers, record
13451 * its timestamp. NOTE: 1) That the test incorporates suggestions
13452 * from the latest proposal of the tcplw@cray.com list (Braden
13453 * 1993/04/26). 2) That updating only on newer timestamps interferes
13454 * with our earlier PAWS tests, so this check should be solely
13455 * predicated on the sequence space of this segment. 3) That we
13456 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13457 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13458 * SEG.Len, This modified check allows us to overcome RFC1323's
13459 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13460 * p.869. In such cases, we can still calculate the RTT correctly
13461 * when RCV.NXT == Last.ACK.Sent.
13462 */
13463 if ((to->to_flags & TOF_TS) != 0 &&
13464 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13465 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13466 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13467 tp->ts_recent_age = tcp_ts_getticks();
13468 tp->ts_recent = to->to_tsval;
13469 }
13470 /*
13471 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13472 * is on (half-synchronized state), then queue data for later
13473 * processing; else drop segment and return.
13474 */
13475 if ((thflags & TH_ACK) == 0) {
13476 if (tp->t_flags & TF_NEEDSYN) {
13477 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13478 tiwin, thflags, nxt_pkt));
13479
13480 } else if (tp->t_flags & TF_ACKNOW) {
13481 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13482 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13483 return (ret_val);
13484 } else {
13485 ctf_do_drop(m, NULL);
13486 return (0);
13487 }
13488 }
13489 /*
13490 * Ack processing.
13491 */
13492 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val, orig_tlen)) {
13493 return (ret_val);
13494 }
13495 if (sbavail(&so->so_snd)) {
13496 if (ctf_progress_timeout_check(tp, true)) {
13497 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13498 tp, tick, PROGRESS_DROP, __LINE__);
13499 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_TCP_RST, tlen);
13500 return (1);
13501 }
13502 }
13503 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13504 tiwin, thflags, nxt_pkt));
13505 }
13506
13507 static int
rack_check_data_after_close(struct mbuf * m,struct tcpcb * tp,int32_t * tlen,struct tcphdr * th,struct socket * so)13508 rack_check_data_after_close(struct mbuf *m,
13509 struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13510 {
13511 struct tcp_rack *rack;
13512
13513 rack = (struct tcp_rack *)tp->t_fb_ptr;
13514 if (rack->rc_allow_data_af_clo == 0) {
13515 close_now:
13516 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13517 /* tcp_close will kill the inp pre-log the Reset */
13518 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13519 tp = tcp_close(tp);
13520 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13521 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13522 return (1);
13523 }
13524 if (sbavail(&so->so_snd) == 0)
13525 goto close_now;
13526 /* Ok we allow data that is ignored and a followup reset */
13527 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13528 tp->rcv_nxt = th->th_seq + *tlen;
13529 tp->t_flags2 |= TF2_DROP_AF_DATA;
13530 rack->r_wanted_output = 1;
13531 *tlen = 0;
13532 return (0);
13533 }
13534
13535 /*
13536 * Return value of 1, the TCB is unlocked and most
13537 * likely gone, return value of 0, the TCP is still
13538 * locked.
13539 */
13540 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)13541 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13542 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13543 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13544 {
13545 int32_t ret_val = 0;
13546 int32_t orig_tlen = tlen;
13547 int32_t ourfinisacked = 0;
13548
13549 ctf_calc_rwin(so, tp);
13550
13551 if ((thflags & TH_RST) ||
13552 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13553 return (ctf_process_rst(m, th, so, tp));
13554 /*
13555 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13556 * synchronized state.
13557 */
13558 if (thflags & TH_SYN) {
13559 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13560 return (ret_val);
13561 }
13562 /*
13563 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13564 * it's less than ts_recent, drop it.
13565 */
13566 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13567 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13568 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13569 return (ret_val);
13570 }
13571 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13572 return (ret_val);
13573 }
13574 /*
13575 * If new data are received on a connection after the user processes
13576 * are gone, then RST the other end.
13577 */
13578 if ((tp->t_flags & TF_CLOSED) && tlen &&
13579 rack_check_data_after_close(m, tp, &tlen, th, so))
13580 return (1);
13581 /*
13582 * If last ACK falls within this segment's sequence numbers, record
13583 * its timestamp. NOTE: 1) That the test incorporates suggestions
13584 * from the latest proposal of the tcplw@cray.com list (Braden
13585 * 1993/04/26). 2) That updating only on newer timestamps interferes
13586 * with our earlier PAWS tests, so this check should be solely
13587 * predicated on the sequence space of this segment. 3) That we
13588 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13589 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13590 * SEG.Len, This modified check allows us to overcome RFC1323's
13591 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13592 * p.869. In such cases, we can still calculate the RTT correctly
13593 * when RCV.NXT == Last.ACK.Sent.
13594 */
13595 if ((to->to_flags & TOF_TS) != 0 &&
13596 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13597 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13598 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13599 tp->ts_recent_age = tcp_ts_getticks();
13600 tp->ts_recent = to->to_tsval;
13601 }
13602 /*
13603 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13604 * is on (half-synchronized state), then queue data for later
13605 * processing; else drop segment and return.
13606 */
13607 if ((thflags & TH_ACK) == 0) {
13608 if (tp->t_flags & TF_NEEDSYN) {
13609 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13610 tiwin, thflags, nxt_pkt));
13611 } else if (tp->t_flags & TF_ACKNOW) {
13612 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13613 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13614 return (ret_val);
13615 } else {
13616 ctf_do_drop(m, NULL);
13617 return (0);
13618 }
13619 }
13620 /*
13621 * Ack processing.
13622 */
13623 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13624 return (ret_val);
13625 }
13626 if (ourfinisacked) {
13627 /*
13628 * If we can't receive any more data, then closing user can
13629 * proceed. Starting the timer is contrary to the
13630 * specification, but if we don't get a FIN we'll hang
13631 * forever.
13632 *
13633 * XXXjl: we should release the tp also, and use a
13634 * compressed state.
13635 */
13636 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13637 soisdisconnected(so);
13638 tcp_timer_activate(tp, TT_2MSL,
13639 (tcp_fast_finwait2_recycle ?
13640 tcp_finwait2_timeout :
13641 TP_MAXIDLE(tp)));
13642 }
13643 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13644 }
13645 if (sbavail(&so->so_snd)) {
13646 if (ctf_progress_timeout_check(tp, true)) {
13647 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13648 tp, tick, PROGRESS_DROP, __LINE__);
13649 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_TCP_RST, tlen);
13650 return (1);
13651 }
13652 }
13653 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13654 tiwin, thflags, nxt_pkt));
13655 }
13656
13657 /*
13658 * Return value of 1, the TCB is unlocked and most
13659 * likely gone, return value of 0, the TCP is still
13660 * locked.
13661 */
13662 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)13663 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
13664 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13665 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13666 {
13667 int32_t ret_val = 0;
13668 int32_t orig_tlen = tlen;
13669 int32_t ourfinisacked = 0;
13670
13671 ctf_calc_rwin(so, tp);
13672
13673 if ((thflags & TH_RST) ||
13674 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13675 return (ctf_process_rst(m, th, so, tp));
13676 /*
13677 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13678 * synchronized state.
13679 */
13680 if (thflags & TH_SYN) {
13681 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13682 return (ret_val);
13683 }
13684 /*
13685 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13686 * it's less than ts_recent, drop it.
13687 */
13688 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13689 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13690 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13691 return (ret_val);
13692 }
13693 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13694 return (ret_val);
13695 }
13696 /*
13697 * If last ACK falls within this segment's sequence numbers, record
13698 * its timestamp. NOTE: 1) That the test incorporates suggestions
13699 * from the latest proposal of the tcplw@cray.com list (Braden
13700 * 1993/04/26). 2) That updating only on newer timestamps interferes
13701 * with our earlier PAWS tests, so this check should be solely
13702 * predicated on the sequence space of this segment. 3) That we
13703 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13704 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13705 * SEG.Len, This modified check allows us to overcome RFC1323's
13706 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13707 * p.869. In such cases, we can still calculate the RTT correctly
13708 * when RCV.NXT == Last.ACK.Sent.
13709 */
13710 if ((to->to_flags & TOF_TS) != 0 &&
13711 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13712 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13713 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13714 tp->ts_recent_age = tcp_ts_getticks();
13715 tp->ts_recent = to->to_tsval;
13716 }
13717 /*
13718 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13719 * is on (half-synchronized state), then queue data for later
13720 * processing; else drop segment and return.
13721 */
13722 if ((thflags & TH_ACK) == 0) {
13723 if (tp->t_flags & TF_NEEDSYN) {
13724 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13725 tiwin, thflags, nxt_pkt));
13726 } else if (tp->t_flags & TF_ACKNOW) {
13727 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13728 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13729 return (ret_val);
13730 } else {
13731 ctf_do_drop(m, NULL);
13732 return (0);
13733 }
13734 }
13735 /*
13736 * Ack processing.
13737 */
13738 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13739 return (ret_val);
13740 }
13741 if (ourfinisacked) {
13742 tcp_twstart(tp);
13743 m_freem(m);
13744 return (1);
13745 }
13746 if (sbavail(&so->so_snd)) {
13747 if (ctf_progress_timeout_check(tp, true)) {
13748 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13749 tp, tick, PROGRESS_DROP, __LINE__);
13750 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_TCP_RST, tlen);
13751 return (1);
13752 }
13753 }
13754 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13755 tiwin, thflags, nxt_pkt));
13756 }
13757
13758 /*
13759 * Return value of 1, the TCB is unlocked and most
13760 * likely gone, return value of 0, the TCP is still
13761 * locked.
13762 */
13763 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)13764 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
13765 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13766 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13767 {
13768 int32_t ret_val = 0;
13769 int32_t orig_tlen;
13770 int32_t ourfinisacked = 0;
13771
13772 ctf_calc_rwin(so, tp);
13773
13774 if ((thflags & TH_RST) ||
13775 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13776 return (ctf_process_rst(m, th, so, tp));
13777 /*
13778 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13779 * synchronized state.
13780 */
13781 if (thflags & TH_SYN) {
13782 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13783 return (ret_val);
13784 }
13785 /*
13786 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13787 * it's less than ts_recent, drop it.
13788 */
13789 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13790 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13791 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13792 return (ret_val);
13793 }
13794 orig_tlen = tlen;
13795 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13796 return (ret_val);
13797 }
13798 /*
13799 * If last ACK falls within this segment's sequence numbers, record
13800 * its timestamp. NOTE: 1) That the test incorporates suggestions
13801 * from the latest proposal of the tcplw@cray.com list (Braden
13802 * 1993/04/26). 2) That updating only on newer timestamps interferes
13803 * with our earlier PAWS tests, so this check should be solely
13804 * predicated on the sequence space of this segment. 3) That we
13805 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13806 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13807 * SEG.Len, This modified check allows us to overcome RFC1323's
13808 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13809 * p.869. In such cases, we can still calculate the RTT correctly
13810 * when RCV.NXT == Last.ACK.Sent.
13811 */
13812 if ((to->to_flags & TOF_TS) != 0 &&
13813 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13814 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13815 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13816 tp->ts_recent_age = tcp_ts_getticks();
13817 tp->ts_recent = to->to_tsval;
13818 }
13819 /*
13820 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13821 * is on (half-synchronized state), then queue data for later
13822 * processing; else drop segment and return.
13823 */
13824 if ((thflags & TH_ACK) == 0) {
13825 if (tp->t_flags & TF_NEEDSYN) {
13826 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13827 tiwin, thflags, nxt_pkt));
13828 } else if (tp->t_flags & TF_ACKNOW) {
13829 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13830 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13831 return (ret_val);
13832 } else {
13833 ctf_do_drop(m, NULL);
13834 return (0);
13835 }
13836 }
13837 /*
13838 * case TCPS_LAST_ACK: Ack processing.
13839 */
13840 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13841 return (ret_val);
13842 }
13843 if (ourfinisacked) {
13844 tp = tcp_close(tp);
13845 ctf_do_drop(m, tp);
13846 return (1);
13847 }
13848 if (sbavail(&so->so_snd)) {
13849 if (ctf_progress_timeout_check(tp, true)) {
13850 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13851 tp, tick, PROGRESS_DROP, __LINE__);
13852 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_TCP_RST, tlen);
13853 return (1);
13854 }
13855 }
13856 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13857 tiwin, thflags, nxt_pkt));
13858 }
13859
13860 /*
13861 * Return value of 1, the TCB is unlocked and most
13862 * likely gone, return value of 0, the TCP is still
13863 * locked.
13864 */
13865 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)13866 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
13867 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13868 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13869 {
13870 int32_t ret_val = 0;
13871 int32_t orig_tlen = tlen;
13872 int32_t ourfinisacked = 0;
13873
13874 ctf_calc_rwin(so, tp);
13875
13876 /* Reset receive buffer auto scaling when not in bulk receive mode. */
13877 if ((thflags & TH_RST) ||
13878 (tp->t_fin_is_rst && (thflags & TH_FIN)))
13879 return (ctf_process_rst(m, th, so, tp));
13880 /*
13881 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13882 * synchronized state.
13883 */
13884 if (thflags & TH_SYN) {
13885 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13886 return (ret_val);
13887 }
13888 /*
13889 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13890 * it's less than ts_recent, drop it.
13891 */
13892 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13893 TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13894 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13895 return (ret_val);
13896 }
13897 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
13898 return (ret_val);
13899 }
13900 /*
13901 * If new data are received on a connection after the user processes
13902 * are gone, then RST the other end.
13903 */
13904 if ((tp->t_flags & TF_CLOSED) && tlen &&
13905 rack_check_data_after_close(m, tp, &tlen, th, so))
13906 return (1);
13907 /*
13908 * If last ACK falls within this segment's sequence numbers, record
13909 * its timestamp. NOTE: 1) That the test incorporates suggestions
13910 * from the latest proposal of the tcplw@cray.com list (Braden
13911 * 1993/04/26). 2) That updating only on newer timestamps interferes
13912 * with our earlier PAWS tests, so this check should be solely
13913 * predicated on the sequence space of this segment. 3) That we
13914 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13915 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13916 * SEG.Len, This modified check allows us to overcome RFC1323's
13917 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13918 * p.869. In such cases, we can still calculate the RTT correctly
13919 * when RCV.NXT == Last.ACK.Sent.
13920 */
13921 if ((to->to_flags & TOF_TS) != 0 &&
13922 SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13923 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13924 ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13925 tp->ts_recent_age = tcp_ts_getticks();
13926 tp->ts_recent = to->to_tsval;
13927 }
13928 /*
13929 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag
13930 * is on (half-synchronized state), then queue data for later
13931 * processing; else drop segment and return.
13932 */
13933 if ((thflags & TH_ACK) == 0) {
13934 if (tp->t_flags & TF_NEEDSYN) {
13935 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13936 tiwin, thflags, nxt_pkt));
13937 } else if (tp->t_flags & TF_ACKNOW) {
13938 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13939 ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13940 return (ret_val);
13941 } else {
13942 ctf_do_drop(m, NULL);
13943 return (0);
13944 }
13945 }
13946 /*
13947 * Ack processing.
13948 */
13949 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val, orig_tlen)) {
13950 return (ret_val);
13951 }
13952 if (sbavail(&so->so_snd)) {
13953 if (ctf_progress_timeout_check(tp, true)) {
13954 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13955 tp, tick, PROGRESS_DROP, __LINE__);
13956 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_TCP_RST, tlen);
13957 return (1);
13958 }
13959 }
13960 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13961 tiwin, thflags, nxt_pkt));
13962 }
13963
13964 static void inline
rack_clear_rate_sample(struct tcp_rack * rack)13965 rack_clear_rate_sample(struct tcp_rack *rack)
13966 {
13967 rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
13968 rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
13969 rack->r_ctl.rack_rs.rs_rtt_tot = 0;
13970 }
13971
13972 static void
rack_set_pace_segments(struct tcpcb * tp,struct tcp_rack * rack,uint32_t line,uint64_t * fill_override)13973 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
13974 {
13975 uint64_t bw_est, rate_wanted;
13976 int chged = 0;
13977 uint32_t user_max, orig_min, orig_max;
13978
13979 #ifdef TCP_REQUEST_TRK
13980 if (rack->rc_hybrid_mode &&
13981 (rack->r_ctl.rc_pace_max_segs != 0) &&
13982 (rack_hybrid_allow_set_maxseg == 1) &&
13983 (rack->r_ctl.rc_last_sft != NULL)) {
13984 rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
13985 return;
13986 }
13987 #endif
13988 orig_min = rack->r_ctl.rc_pace_min_segs;
13989 orig_max = rack->r_ctl.rc_pace_max_segs;
13990 user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
13991 if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
13992 chged = 1;
13993 rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
13994 if (rack->use_fixed_rate || rack->rc_force_max_seg) {
13995 if (user_max != rack->r_ctl.rc_pace_max_segs)
13996 chged = 1;
13997 }
13998 if (rack->rc_force_max_seg) {
13999 rack->r_ctl.rc_pace_max_segs = user_max;
14000 } else if (rack->use_fixed_rate) {
14001 bw_est = rack_get_bw(rack);
14002 if ((rack->r_ctl.crte == NULL) ||
14003 (bw_est != rack->r_ctl.crte->rate)) {
14004 rack->r_ctl.rc_pace_max_segs = user_max;
14005 } else {
14006 /* We are pacing right at the hardware rate */
14007 uint32_t segsiz, pace_one;
14008
14009 if (rack_pace_one_seg ||
14010 (rack->r_ctl.rc_user_set_min_segs == 1))
14011 pace_one = 1;
14012 else
14013 pace_one = 0;
14014 segsiz = min(ctf_fixed_maxseg(tp),
14015 rack->r_ctl.rc_pace_min_segs);
14016 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14017 tp, bw_est, segsiz, pace_one,
14018 rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14019 }
14020 } else if (rack->rc_always_pace) {
14021 if (rack->r_ctl.gp_bw ||
14022 rack->r_ctl.init_rate) {
14023 /* We have a rate of some sort set */
14024 uint32_t orig;
14025
14026 bw_est = rack_get_bw(rack);
14027 orig = rack->r_ctl.rc_pace_max_segs;
14028 if (fill_override)
14029 rate_wanted = *fill_override;
14030 else
14031 rate_wanted = rack_get_gp_est(rack);
14032 if (rate_wanted) {
14033 /* We have something */
14034 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14035 rate_wanted,
14036 ctf_fixed_maxseg(rack->rc_tp));
14037 } else
14038 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14039 if (orig != rack->r_ctl.rc_pace_max_segs)
14040 chged = 1;
14041 } else if ((rack->r_ctl.gp_bw == 0) &&
14042 (rack->r_ctl.rc_pace_max_segs == 0)) {
14043 /*
14044 * If we have nothing limit us to bursting
14045 * out IW sized pieces.
14046 */
14047 chged = 1;
14048 rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14049 }
14050 }
14051 if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14052 chged = 1;
14053 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14054 }
14055 if (chged)
14056 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14057 }
14058
14059
14060 static void
rack_init_fsb_block(struct tcpcb * tp,struct tcp_rack * rack,int32_t flags)14061 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14062 {
14063 #ifdef INET6
14064 struct ip6_hdr *ip6 = NULL;
14065 #endif
14066 #ifdef INET
14067 struct ip *ip = NULL;
14068 #endif
14069 struct udphdr *udp = NULL;
14070
14071 /* Ok lets fill in the fast block, it can only be used with no IP options! */
14072 #ifdef INET6
14073 if (rack->r_is_v6) {
14074 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14075 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14076 if (tp->t_port) {
14077 rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14078 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14079 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14080 udp->uh_dport = tp->t_port;
14081 rack->r_ctl.fsb.udp = udp;
14082 rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14083 } else
14084 {
14085 rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14086 rack->r_ctl.fsb.udp = NULL;
14087 }
14088 tcpip_fillheaders(rack->rc_inp,
14089 tp->t_port,
14090 ip6, rack->r_ctl.fsb.th);
14091 rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14092 } else
14093 #endif /* INET6 */
14094 #ifdef INET
14095 {
14096 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14097 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14098 if (tp->t_port) {
14099 rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14100 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14101 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14102 udp->uh_dport = tp->t_port;
14103 rack->r_ctl.fsb.udp = udp;
14104 rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14105 } else
14106 {
14107 rack->r_ctl.fsb.udp = NULL;
14108 rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14109 }
14110 tcpip_fillheaders(rack->rc_inp,
14111 tp->t_port,
14112 ip, rack->r_ctl.fsb.th);
14113 rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14114 }
14115 #endif
14116 rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14117 (long)TCP_MAXWIN << tp->rcv_scale);
14118 rack->r_fsb_inited = 1;
14119 }
14120
14121 static int
rack_init_fsb(struct tcpcb * tp,struct tcp_rack * rack)14122 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14123 {
14124 /*
14125 * Allocate the larger of spaces V6 if available else just
14126 * V4 and include udphdr (overbook)
14127 */
14128 #ifdef INET6
14129 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14130 #else
14131 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14132 #endif
14133 rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14134 M_TCPFSB, M_NOWAIT|M_ZERO);
14135 if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14136 return (ENOMEM);
14137 }
14138 rack->r_fsb_inited = 0;
14139 return (0);
14140 }
14141
14142 static void
rack_log_hystart_event(struct tcp_rack * rack,uint32_t high_seq,uint8_t mod)14143 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14144 {
14145 /*
14146 * Types of logs (mod value)
14147 * 20 - Initial round setup
14148 * 21 - Rack declares a new round.
14149 */
14150 struct tcpcb *tp;
14151
14152 tp = rack->rc_tp;
14153 if (tcp_bblogging_on(tp)) {
14154 union tcp_log_stackspecific log;
14155 struct timeval tv;
14156
14157 memset(&log, 0, sizeof(log));
14158 log.u_bbr.flex1 = rack->r_ctl.current_round;
14159 log.u_bbr.flex2 = rack->r_ctl.roundends;
14160 log.u_bbr.flex3 = high_seq;
14161 log.u_bbr.flex4 = tp->snd_max;
14162 log.u_bbr.flex8 = mod;
14163 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14164 log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14165 log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14166 TCP_LOG_EVENTP(tp, NULL,
14167 &tptosocket(tp)->so_rcv,
14168 &tptosocket(tp)->so_snd,
14169 TCP_HYSTART, 0,
14170 0, &log, false, &tv);
14171 }
14172 }
14173
14174 static void
rack_deferred_init(struct tcpcb * tp,struct tcp_rack * rack)14175 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14176 {
14177 rack->rack_deferred_inited = 1;
14178 rack->r_ctl.roundends = tp->snd_max;
14179 rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14180 rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14181 }
14182
14183 static void
rack_init_retransmit_value(struct tcp_rack * rack,int ctl)14184 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14185 {
14186 /* Retransmit bit controls.
14187 *
14188 * The setting of these values control one of
14189 * three settings you can have and dictate
14190 * how rack does retransmissions. Note this
14191 * is in *any* mode i.e. pacing on or off DGP
14192 * fixed rate pacing, or just bursting rack.
14193 *
14194 * 1 - Use full sized retransmits i.e. limit
14195 * the size to whatever the pace_max_segments
14196 * size is.
14197 *
14198 * 2 - Use pacer min granularity as a guide to
14199 * the size combined with the current calculated
14200 * goodput b/w measurement. So for example if
14201 * the goodput is measured at 20Mbps we would
14202 * calculate 8125 (pacer minimum 250usec in
14203 * that b/w) and then round it up to the next
14204 * MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14205 *
14206 * 0 - The rack default 1 MSS (anything not 0/1/2
14207 * fall here too if we are setting via rack_init()).
14208 *
14209 */
14210 if (ctl == 1) {
14211 rack->full_size_rxt = 1;
14212 rack->shape_rxt_to_pacing_min = 0;
14213 } else if (ctl == 2) {
14214 rack->full_size_rxt = 0;
14215 rack->shape_rxt_to_pacing_min = 1;
14216 } else {
14217 rack->full_size_rxt = 0;
14218 rack->shape_rxt_to_pacing_min = 0;
14219 }
14220 }
14221
14222 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)14223 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14224 uint32_t flex1,
14225 uint32_t flex2,
14226 uint32_t flex3)
14227 {
14228 if (tcp_bblogging_on(rack->rc_tp)) {
14229 union tcp_log_stackspecific log;
14230 struct timeval tv;
14231
14232 memset(&log, 0, sizeof(log));
14233 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14234 log.u_bbr.flex8 = mod;
14235 log.u_bbr.flex1 = flex1;
14236 log.u_bbr.flex2 = flex2;
14237 log.u_bbr.flex3 = flex3;
14238 tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14239 0, &log, false, NULL, __func__, __LINE__, &tv);
14240 }
14241 }
14242
14243 static int
rack_chg_query(struct tcpcb * tp,struct tcp_query_resp * reqr)14244 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14245 {
14246 struct tcp_rack *rack;
14247 struct rack_sendmap *rsm;
14248 int i;
14249
14250
14251 rack = (struct tcp_rack *)tp->t_fb_ptr;
14252 switch (reqr->req) {
14253 case TCP_QUERY_SENDMAP:
14254 if ((reqr->req_param == tp->snd_max) ||
14255 (tp->snd_max == tp->snd_una)){
14256 /* Unlikely */
14257 return (0);
14258 }
14259 rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14260 if (rsm == NULL) {
14261 /* Can't find that seq -- unlikely */
14262 return (0);
14263 }
14264 reqr->sendmap_start = rsm->r_start;
14265 reqr->sendmap_end = rsm->r_end;
14266 reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14267 reqr->sendmap_fas = rsm->r_fas;
14268 if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14269 reqr->sendmap_send_cnt = SNDMAP_NRTX;
14270 for(i=0; i<reqr->sendmap_send_cnt; i++)
14271 reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14272 reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14273 reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14274 reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14275 reqr->sendmap_dupacks = rsm->r_dupack;
14276 rack_log_chg_info(tp, rack, 1,
14277 rsm->r_start,
14278 rsm->r_end,
14279 rsm->r_flags);
14280 return(1);
14281 break;
14282 case TCP_QUERY_TIMERS_UP:
14283 if (rack->r_ctl.rc_hpts_flags == 0) {
14284 /* no timers up */
14285 return (0);
14286 }
14287 reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14288 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14289 reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14290 }
14291 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14292 reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14293 }
14294 rack_log_chg_info(tp, rack, 2,
14295 rack->r_ctl.rc_hpts_flags,
14296 rack->r_ctl.rc_last_output_to,
14297 rack->r_ctl.rc_timer_exp);
14298 return (1);
14299 break;
14300 case TCP_QUERY_RACK_TIMES:
14301 /* Reordering items */
14302 reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14303 reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14304 /* Timerstamps and timers */
14305 reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14306 reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14307 reqr->rack_rtt = rack->rc_rack_rtt;
14308 reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14309 reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14310 /* PRR data */
14311 reqr->rack_sacked = rack->r_ctl.rc_sacked;
14312 reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14313 reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14314 reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14315 reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14316 reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14317 /* TLP and persists info */
14318 reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14319 reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14320 if (rack->rc_in_persist) {
14321 reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14322 reqr->rack_in_persist = 1;
14323 } else {
14324 reqr->rack_time_went_idle = 0;
14325 reqr->rack_in_persist = 0;
14326 }
14327 if (rack->r_wanted_output)
14328 reqr->rack_wanted_output = 1;
14329 else
14330 reqr->rack_wanted_output = 0;
14331 return (1);
14332 break;
14333 default:
14334 return (-EINVAL);
14335 }
14336 }
14337
14338 static void
rack_switch_failed(struct tcpcb * tp)14339 rack_switch_failed(struct tcpcb *tp)
14340 {
14341 /*
14342 * This method gets called if a stack switch was
14343 * attempted and it failed. We are left
14344 * but our hpts timers were stopped and we
14345 * need to validate time units and t_flags2.
14346 */
14347 struct tcp_rack *rack;
14348 struct timeval tv;
14349 uint32_t cts;
14350 uint32_t toval;
14351 struct hpts_diag diag;
14352
14353 rack = (struct tcp_rack *)tp->t_fb_ptr;
14354 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14355 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14356 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14357 else
14358 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14359 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14360 tp->t_flags2 |= TF2_MBUF_ACKCMP;
14361 if (tp->t_in_hpts > IHPTS_NONE) {
14362 /* Strange */
14363 return;
14364 }
14365 cts = tcp_get_usecs(&tv);
14366 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14367 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14368 toval = rack->r_ctl.rc_last_output_to - cts;
14369 } else {
14370 /* one slot please */
14371 toval = HPTS_USECS_PER_SLOT;
14372 }
14373 } else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14374 if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14375 toval = rack->r_ctl.rc_timer_exp - cts;
14376 } else {
14377 /* one slot please */
14378 toval = HPTS_USECS_PER_SLOT;
14379 }
14380 } else
14381 toval = HPTS_USECS_PER_SLOT;
14382 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14383 __LINE__, &diag);
14384 rack_log_hpts_diag(rack, cts, &diag, &tv);
14385 }
14386
14387 static int
rack_init_outstanding(struct tcpcb * tp,struct tcp_rack * rack,uint32_t us_cts,void * ptr)14388 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14389 {
14390 struct rack_sendmap *rsm, *ersm;
14391 int insret __diagused;
14392 /*
14393 * When initing outstanding, we must be quite careful
14394 * to not refer to tp->t_fb_ptr. This has the old rack
14395 * pointer in it, not the "new" one (when we are doing
14396 * a stack switch).
14397 */
14398
14399
14400 if (tp->t_fb->tfb_chg_query == NULL) {
14401 /* Create a send map for the current outstanding data */
14402
14403 rsm = rack_alloc(rack);
14404 if (rsm == NULL) {
14405 uma_zfree(rack_pcb_zone, ptr);
14406 return (ENOMEM);
14407 }
14408 rsm->r_no_rtt_allowed = 1;
14409 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14410 rsm->r_rtr_cnt = 1;
14411 rsm->r_rtr_bytes = 0;
14412 if (tp->t_flags & TF_SENTFIN)
14413 rsm->r_flags |= RACK_HAS_FIN;
14414 rsm->r_end = tp->snd_max;
14415 if (tp->snd_una == tp->iss) {
14416 /* The data space is one beyond snd_una */
14417 rsm->r_flags |= RACK_HAS_SYN;
14418 rsm->r_start = tp->iss;
14419 rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14420 } else
14421 rsm->r_start = tp->snd_una;
14422 rsm->r_dupack = 0;
14423 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14424 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14425 if (rsm->m) {
14426 rsm->orig_m_len = rsm->m->m_len;
14427 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14428 } else {
14429 rsm->orig_m_len = 0;
14430 rsm->orig_t_space = 0;
14431 }
14432 } else {
14433 /*
14434 * This can happen if we have a stand-alone FIN or
14435 * SYN.
14436 */
14437 rsm->m = NULL;
14438 rsm->orig_m_len = 0;
14439 rsm->orig_t_space = 0;
14440 rsm->soff = 0;
14441 }
14442 #ifdef INVARIANTS
14443 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14444 panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14445 insret, rack, rsm);
14446 }
14447 #else
14448 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14449 #endif
14450 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14451 rsm->r_in_tmap = 1;
14452 } else {
14453 /* We have a query mechanism, lets use it */
14454 struct tcp_query_resp qr;
14455 int i;
14456 tcp_seq at;
14457
14458 at = tp->snd_una;
14459 while (at != tp->snd_max) {
14460 memset(&qr, 0, sizeof(qr));
14461 qr.req = TCP_QUERY_SENDMAP;
14462 qr.req_param = at;
14463 if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14464 break;
14465 /* Move forward */
14466 at = qr.sendmap_end;
14467 /* Now lets build the entry for this one */
14468 rsm = rack_alloc(rack);
14469 if (rsm == NULL) {
14470 uma_zfree(rack_pcb_zone, ptr);
14471 return (ENOMEM);
14472 }
14473 memset(rsm, 0, sizeof(struct rack_sendmap));
14474 /* Now configure the rsm and insert it */
14475 rsm->r_dupack = qr.sendmap_dupacks;
14476 rsm->r_start = qr.sendmap_start;
14477 rsm->r_end = qr.sendmap_end;
14478 if (qr.sendmap_fas)
14479 rsm->r_fas = qr.sendmap_end;
14480 else
14481 rsm->r_fas = rsm->r_start - tp->snd_una;
14482 /*
14483 * We have carefully aligned the bits
14484 * so that all we have to do is copy over
14485 * the bits with the mask.
14486 */
14487 rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14488 rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14489 rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14490 rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14491 for (i=0 ; i<rsm->r_rtr_cnt; i++)
14492 rsm->r_tim_lastsent[i] = qr.sendmap_time[i];
14493 rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14494 (rsm->r_start - tp->snd_una), &rsm->soff);
14495 if (rsm->m) {
14496 rsm->orig_m_len = rsm->m->m_len;
14497 rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14498 } else {
14499 rsm->orig_m_len = 0;
14500 rsm->orig_t_space = 0;
14501 }
14502 #ifdef INVARIANTS
14503 if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14504 panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14505 insret, rack, rsm);
14506 }
14507 #else
14508 (void)tqhash_insert(rack->r_ctl.tqh, rsm);
14509 #endif
14510 if ((rsm->r_flags & RACK_ACKED) == 0) {
14511 TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14512 if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14513 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14514 /*
14515 * If the existing ersm was sent at
14516 * a later time than the new one, then
14517 * the new one should appear ahead of this
14518 * ersm.
14519 */
14520 rsm->r_in_tmap = 1;
14521 TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14522 break;
14523 }
14524 }
14525 if (rsm->r_in_tmap == 0) {
14526 /*
14527 * Not found so shove it on the tail.
14528 */
14529 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14530 rsm->r_in_tmap = 1;
14531 }
14532 } else {
14533 if ((rack->r_ctl.rc_sacklast == NULL) ||
14534 (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14535 rack->r_ctl.rc_sacklast = rsm;
14536 }
14537 }
14538 rack_log_chg_info(tp, rack, 3,
14539 rsm->r_start,
14540 rsm->r_end,
14541 rsm->r_flags);
14542 }
14543 }
14544 return (0);
14545 }
14546
14547
14548 static int32_t
rack_init(struct tcpcb * tp,void ** ptr)14549 rack_init(struct tcpcb *tp, void **ptr)
14550 {
14551 struct inpcb *inp = tptoinpcb(tp);
14552 struct tcp_rack *rack = NULL;
14553 uint32_t iwin, snt, us_cts;
14554 size_t sz;
14555 int err, no_query;
14556
14557 tcp_hpts_init(tp);
14558
14559 /*
14560 * First are we the initial or are we a switched stack?
14561 * If we are initing via tcp_newtcppcb the ptr passed
14562 * will be tp->t_fb_ptr. If its a stack switch that
14563 * has a previous stack we can query it will be a local
14564 * var that will in the end be set into t_fb_ptr.
14565 */
14566 if (ptr == &tp->t_fb_ptr)
14567 no_query = 1;
14568 else
14569 no_query = 0;
14570 *ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14571 if (*ptr == NULL) {
14572 /*
14573 * We need to allocate memory but cant. The INP and INP_INFO
14574 * locks and they are recursive (happens during setup. So a
14575 * scheme to drop the locks fails :(
14576 *
14577 */
14578 return(ENOMEM);
14579 }
14580 memset(*ptr, 0, sizeof(struct tcp_rack));
14581 rack = (struct tcp_rack *)*ptr;
14582 rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
14583 if (rack->r_ctl.tqh == NULL) {
14584 uma_zfree(rack_pcb_zone, rack);
14585 return(ENOMEM);
14586 }
14587 tqhash_init(rack->r_ctl.tqh);
14588 TAILQ_INIT(&rack->r_ctl.rc_free);
14589 TAILQ_INIT(&rack->r_ctl.rc_tmap);
14590 rack->rc_tp = tp;
14591 rack->rc_inp = inp;
14592 /* Set the flag */
14593 rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
14594 /* Probably not needed but lets be sure */
14595 rack_clear_rate_sample(rack);
14596 /*
14597 * Save off the default values, socket options will poke
14598 * at these if pacing is not on or we have not yet
14599 * reached where pacing is on (gp_ready/fixed enabled).
14600 * When they get set into the CC module (when gp_ready
14601 * is enabled or we enable fixed) then we will set these
14602 * values into the CC and place in here the old values
14603 * so we have a restoral. Then we will set the flag
14604 * rc_pacing_cc_set. That way whenever we turn off pacing
14605 * or switch off this stack, we will know to go restore
14606 * the saved values.
14607 *
14608 * We specifically put into the beta the ecn value for pacing.
14609 */
14610 rack->rc_new_rnd_needed = 1;
14611 rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
14612 /* We want abe like behavior as well */
14613
14614 rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
14615 rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
14616 rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
14617 if (rack_fill_cw_state)
14618 rack->rc_pace_to_cwnd = 1;
14619 if (rack_pacing_min_seg)
14620 rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
14621 if (use_rack_rr)
14622 rack->use_rack_rr = 1;
14623 if (rack_dnd_default) {
14624 rack->rc_pace_dnd = 1;
14625 }
14626 if (V_tcp_delack_enabled)
14627 tp->t_delayed_ack = 1;
14628 else
14629 tp->t_delayed_ack = 0;
14630 #ifdef TCP_ACCOUNTING
14631 if (rack_tcp_accounting) {
14632 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
14633 }
14634 #endif
14635 rack->r_ctl.pcm_i.cnt_alloc = RACK_DEFAULT_PCM_ARRAY;
14636 sz = (sizeof(struct rack_pcm_stats) * rack->r_ctl.pcm_i.cnt_alloc);
14637 rack->r_ctl.pcm_s = malloc(sz,M_TCPPCM, M_NOWAIT);
14638 if (rack->r_ctl.pcm_s == NULL) {
14639 rack->r_ctl.pcm_i.cnt_alloc = 0;
14640 }
14641 rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
14642 rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
14643 if (rack_enable_shared_cwnd)
14644 rack->rack_enable_scwnd = 1;
14645 rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
14646 rack->rc_user_set_max_segs = rack_hptsi_segments;
14647 rack->r_ctl.max_reduction = rack_max_reduce;
14648 rack->rc_force_max_seg = 0;
14649 TAILQ_INIT(&rack->r_ctl.opt_list);
14650 rack->r_ctl.rc_saved_beta = V_newreno_beta_ecn;
14651 rack->r_ctl.rc_saved_beta_ecn = V_newreno_beta_ecn;
14652 if (rack_hibeta_setting) {
14653 rack->rack_hibeta = 1;
14654 if ((rack_hibeta_setting >= 50) &&
14655 (rack_hibeta_setting <= 100)) {
14656 rack->r_ctl.rc_saved_beta = rack_hibeta_setting;
14657 rack->r_ctl.saved_hibeta = rack_hibeta_setting;
14658 }
14659 } else {
14660 rack->r_ctl.saved_hibeta = 50;
14661 }
14662 /*
14663 * We initialize to all ones so we never match 0
14664 * just in case the client sends in 0, it hopefully
14665 * will never have all 1's in ms :-)
14666 */
14667 rack->r_ctl.last_tm_mark = 0xffffffffffffffff;
14668 rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
14669 rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
14670 rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
14671 rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
14672 rack->r_ctl.rc_highest_us_rtt = 0;
14673 rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
14674 rack->pcm_enabled = rack_pcm_is_enabled;
14675 if (rack_fillcw_bw_cap)
14676 rack->r_ctl.fillcw_cap = rack_fillcw_bw_cap;
14677 rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
14678 if (rack_use_cmp_acks)
14679 rack->r_use_cmp_ack = 1;
14680 if (rack_disable_prr)
14681 rack->rack_no_prr = 1;
14682 if (rack_gp_no_rec_chg)
14683 rack->rc_gp_no_rec_chg = 1;
14684 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
14685 rack->r_ctl.pacing_method |= RACK_REG_PACING;
14686 rack->rc_always_pace = 1;
14687 if (rack->rack_hibeta)
14688 rack_set_cc_pacing(rack);
14689 } else
14690 rack->rc_always_pace = 0;
14691 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
14692 rack->r_mbuf_queue = 1;
14693 else
14694 rack->r_mbuf_queue = 0;
14695 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14696 if (rack_limits_scwnd)
14697 rack->r_limit_scw = 1;
14698 else
14699 rack->r_limit_scw = 0;
14700 rack_init_retransmit_value(rack, rack_rxt_controls);
14701 rack->rc_labc = V_tcp_abc_l_var;
14702 if (rack_honors_hpts_min_to)
14703 rack->r_use_hpts_min = 1;
14704 if (tp->snd_una != 0) {
14705 rack->rc_sendvars_notset = 0;
14706 /*
14707 * Make sure any TCP timers are not running.
14708 */
14709 tcp_timer_stop(tp);
14710 } else {
14711 /*
14712 * Server side, we are called from the
14713 * syn-cache. This means none of the
14714 * snd_una/max are set yet so we have
14715 * to defer this until the first send.
14716 */
14717 rack->rc_sendvars_notset = 1;
14718 }
14719
14720 rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
14721 rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
14722 rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
14723 rack->r_ctl.rc_min_to = rack_min_to;
14724 microuptime(&rack->r_ctl.act_rcv_time);
14725 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
14726 rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
14727 if (rack_hw_up_only)
14728 rack->r_up_only = 1;
14729 if (rack_do_dyn_mul) {
14730 /* When dynamic adjustment is on CA needs to start at 100% */
14731 rack->rc_gp_dyn_mul = 1;
14732 if (rack_do_dyn_mul >= 100)
14733 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
14734 } else
14735 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
14736 rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
14737 if (rack_timely_off) {
14738 rack->rc_skip_timely = 1;
14739 }
14740 if (rack->rc_skip_timely) {
14741 rack->r_ctl.rack_per_of_gp_rec = 90;
14742 rack->r_ctl.rack_per_of_gp_ca = 100;
14743 rack->r_ctl.rack_per_of_gp_ss = 250;
14744 }
14745 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
14746 rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_msec(&rack->r_ctl.act_rcv_time);
14747 rack->r_ctl.last_rcv_tstmp_for_rtt = tcp_tv_to_msec(&rack->r_ctl.act_rcv_time);
14748
14749 setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
14750 rack_probertt_filter_life);
14751 us_cts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
14752 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
14753 rack->r_ctl.rc_time_of_last_probertt = us_cts;
14754 rack->r_ctl.rc_went_idle_time = us_cts;
14755 rack->r_ctl.rc_time_probertt_starts = 0;
14756
14757 rack->r_ctl.gp_rnd_thresh = rack_rnd_cnt_req & 0xff;
14758 if (rack_rnd_cnt_req & 0x10000)
14759 rack->r_ctl.gate_to_fs = 1;
14760 rack->r_ctl.gp_gain_req = rack_gp_gain_req;
14761 if ((rack_rnd_cnt_req & 0x100) > 0) {
14762
14763 }
14764 if (rack_dsack_std_based & 0x1) {
14765 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
14766 rack->rc_rack_tmr_std_based = 1;
14767 }
14768 if (rack_dsack_std_based & 0x2) {
14769 /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */
14770 rack->rc_rack_use_dsack = 1;
14771 }
14772 /* We require at least one measurement, even if the sysctl is 0 */
14773 if (rack_req_measurements)
14774 rack->r_ctl.req_measurements = rack_req_measurements;
14775 else
14776 rack->r_ctl.req_measurements = 1;
14777 if (rack_enable_hw_pacing)
14778 rack->rack_hdw_pace_ena = 1;
14779 if (rack_hw_rate_caps)
14780 rack->r_rack_hw_rate_caps = 1;
14781 if (rack_non_rxt_use_cr)
14782 rack->rack_rec_nonrxt_use_cr = 1;
14783 /* Lets setup the fsb block */
14784 err = rack_init_fsb(tp, rack);
14785 if (err) {
14786 uma_zfree(rack_pcb_zone, *ptr);
14787 *ptr = NULL;
14788 return (err);
14789 }
14790 if (rack_do_hystart) {
14791 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
14792 if (rack_do_hystart > 1)
14793 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
14794 if (rack_do_hystart > 2)
14795 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
14796 }
14797 /* Log what we will do with queries */
14798 rack_log_chg_info(tp, rack, 7,
14799 no_query, 0, 0);
14800 if (rack_def_profile)
14801 rack_set_profile(rack, rack_def_profile);
14802 /* Cancel the GP measurement in progress */
14803 tp->t_flags &= ~TF_GPUTINPROG;
14804 if ((tp->t_state != TCPS_CLOSED) &&
14805 (tp->t_state != TCPS_TIME_WAIT)) {
14806 /*
14807 * We are already open, we may
14808 * need to adjust a few things.
14809 */
14810 if (SEQ_GT(tp->snd_max, tp->iss))
14811 snt = tp->snd_max - tp->iss;
14812 else
14813 snt = 0;
14814 iwin = rc_init_window(rack);
14815 if ((snt < iwin) &&
14816 (no_query == 1)) {
14817 /* We are not past the initial window
14818 * on the first init (i.e. a stack switch
14819 * has not yet occured) so we need to make
14820 * sure cwnd and ssthresh is correct.
14821 */
14822 if (tp->snd_cwnd < iwin)
14823 tp->snd_cwnd = iwin;
14824 /*
14825 * If we are within the initial window
14826 * we want ssthresh to be unlimited. Setting
14827 * it to the rwnd (which the default stack does
14828 * and older racks) is not really a good idea
14829 * since we want to be in SS and grow both the
14830 * cwnd and the rwnd (via dynamic rwnd growth). If
14831 * we set it to the rwnd then as the peer grows its
14832 * rwnd we will be stuck in CA and never hit SS.
14833 *
14834 * Its far better to raise it up high (this takes the
14835 * risk that there as been a loss already, probably
14836 * we should have an indicator in all stacks of loss
14837 * but we don't), but considering the normal use this
14838 * is a risk worth taking. The consequences of not
14839 * hitting SS are far worse than going one more time
14840 * into it early on (before we have sent even a IW).
14841 * It is highly unlikely that we will have had a loss
14842 * before getting the IW out.
14843 */
14844 tp->snd_ssthresh = 0xffffffff;
14845 }
14846 /*
14847 * Any init based on sequence numbers
14848 * should be done in the deferred init path
14849 * since we can be CLOSED and not have them
14850 * inited when rack_init() is called. We
14851 * are not closed so lets call it.
14852 */
14853 rack_deferred_init(tp, rack);
14854 }
14855 if ((tp->t_state != TCPS_CLOSED) &&
14856 (tp->t_state != TCPS_TIME_WAIT) &&
14857 (no_query == 0) &&
14858 (tp->snd_una != tp->snd_max)) {
14859 err = rack_init_outstanding(tp, rack, us_cts, *ptr);
14860 if (err) {
14861 *ptr = NULL;
14862 return(err);
14863 }
14864 }
14865 rack_stop_all_timers(tp, rack);
14866 /* Setup all the t_flags2 */
14867 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14868 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14869 else
14870 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14871 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14872 tp->t_flags2 |= TF2_MBUF_ACKCMP;
14873 /*
14874 * Timers in Rack are kept in microseconds so lets
14875 * convert any initial incoming variables
14876 * from ticks into usecs. Note that we
14877 * also change the values of t_srtt and t_rttvar, if
14878 * they are non-zero. They are kept with a 5
14879 * bit decimal so we have to carefully convert
14880 * these to get the full precision.
14881 */
14882 rack_convert_rtts(tp);
14883 rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
14884 if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
14885 /* We do not start any timers on DROPPED connections */
14886 if (tp->t_fb->tfb_chg_query == NULL) {
14887 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14888 } else {
14889 struct tcp_query_resp qr;
14890 int ret;
14891
14892 memset(&qr, 0, sizeof(qr));
14893
14894 /* Get the misc time stamps and such for rack */
14895 qr.req = TCP_QUERY_RACK_TIMES;
14896 ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14897 if (ret == 1) {
14898 rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
14899 rack->r_ctl.num_dsack = qr.rack_num_dsacks;
14900 rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
14901 rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
14902 rack->rc_rack_rtt = qr.rack_rtt;
14903 rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
14904 rack->r_ctl.rc_sacked = qr.rack_sacked;
14905 rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
14906 rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
14907 rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
14908 rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
14909 rack->r_ctl.rc_prr_out = qr.rack_prr_out;
14910 if (qr.rack_tlp_out) {
14911 rack->rc_tlp_in_progress = 1;
14912 rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
14913 } else {
14914 rack->rc_tlp_in_progress = 0;
14915 rack->r_ctl.rc_tlp_cnt_out = 0;
14916 }
14917 if (qr.rack_srtt_measured)
14918 rack->rc_srtt_measure_made = 1;
14919 if (qr.rack_in_persist == 1) {
14920 rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
14921 #ifdef NETFLIX_SHARED_CWND
14922 if (rack->r_ctl.rc_scw) {
14923 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
14924 rack->rack_scwnd_is_idle = 1;
14925 }
14926 #endif
14927 rack->r_ctl.persist_lost_ends = 0;
14928 rack->probe_not_answered = 0;
14929 rack->forced_ack = 0;
14930 tp->t_rxtshift = 0;
14931 rack->rc_in_persist = 1;
14932 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
14933 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
14934 }
14935 if (qr.rack_wanted_output)
14936 rack->r_wanted_output = 1;
14937 rack_log_chg_info(tp, rack, 6,
14938 qr.rack_min_rtt,
14939 qr.rack_rtt,
14940 qr.rack_reorder_ts);
14941 }
14942 /* Get the old stack timers */
14943 qr.req_param = 0;
14944 qr.req = TCP_QUERY_TIMERS_UP;
14945 ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
14946 if (ret) {
14947 /*
14948 * non-zero return means we have a timer('s)
14949 * to start. Zero means no timer (no keepalive
14950 * I suppose).
14951 */
14952 uint32_t tov = 0;
14953
14954 rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
14955 if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
14956 rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
14957 if (TSTMP_GT(qr.timer_pacing_to, us_cts))
14958 tov = qr.timer_pacing_to - us_cts;
14959 else
14960 tov = HPTS_USECS_PER_SLOT;
14961 }
14962 if (qr.timer_hpts_flags & PACE_TMR_MASK) {
14963 rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
14964 if (tov == 0) {
14965 if (TSTMP_GT(qr.timer_timer_exp, us_cts))
14966 tov = qr.timer_timer_exp - us_cts;
14967 else
14968 tov = HPTS_USECS_PER_SLOT;
14969 }
14970 }
14971 rack_log_chg_info(tp, rack, 4,
14972 rack->r_ctl.rc_hpts_flags,
14973 rack->r_ctl.rc_last_output_to,
14974 rack->r_ctl.rc_timer_exp);
14975 if (tov) {
14976 struct hpts_diag diag;
14977
14978 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
14979 __LINE__, &diag);
14980 rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
14981 }
14982 }
14983 }
14984 rack_log_rtt_shrinks(rack, us_cts, tp->t_rxtcur,
14985 __LINE__, RACK_RTTS_INIT);
14986 }
14987 return (0);
14988 }
14989
14990 static int
rack_handoff_ok(struct tcpcb * tp)14991 rack_handoff_ok(struct tcpcb *tp)
14992 {
14993 if ((tp->t_state == TCPS_CLOSED) ||
14994 (tp->t_state == TCPS_LISTEN)) {
14995 /* Sure no problem though it may not stick */
14996 return (0);
14997 }
14998 if ((tp->t_state == TCPS_SYN_SENT) ||
14999 (tp->t_state == TCPS_SYN_RECEIVED)) {
15000 /*
15001 * We really don't know if you support sack,
15002 * you have to get to ESTAB or beyond to tell.
15003 */
15004 return (EAGAIN);
15005 }
15006 if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15007 /*
15008 * Rack will only send a FIN after all data is acknowledged.
15009 * So in this case we have more data outstanding. We can't
15010 * switch stacks until either all data and only the FIN
15011 * is left (in which case rack_init() now knows how
15012 * to deal with that) <or> all is acknowledged and we
15013 * are only left with incoming data, though why you
15014 * would want to switch to rack after all data is acknowledged
15015 * I have no idea (rrs)!
15016 */
15017 return (EAGAIN);
15018 }
15019 if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15020 return (0);
15021 }
15022 /*
15023 * If we reach here we don't do SACK on this connection so we can
15024 * never do rack.
15025 */
15026 return (EINVAL);
15027 }
15028
15029 static void
rack_fini(struct tcpcb * tp,int32_t tcb_is_purged)15030 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15031 {
15032
15033 if (tp->t_fb_ptr) {
15034 uint32_t cnt_free = 0;
15035 struct tcp_rack *rack;
15036 struct rack_sendmap *rsm;
15037
15038 tcp_handle_orphaned_packets(tp);
15039 tp->t_flags &= ~TF_FORCEDATA;
15040 rack = (struct tcp_rack *)tp->t_fb_ptr;
15041 rack_log_pacing_delay_calc(rack,
15042 0,
15043 0,
15044 0,
15045 rack_get_gp_est(rack), /* delRate */
15046 rack_get_lt_bw(rack), /* rttProp */
15047 20, __LINE__, NULL, 0);
15048 #ifdef NETFLIX_SHARED_CWND
15049 if (rack->r_ctl.rc_scw) {
15050 uint32_t limit;
15051
15052 if (rack->r_limit_scw)
15053 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15054 else
15055 limit = 0;
15056 tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15057 rack->r_ctl.rc_scw_index,
15058 limit);
15059 rack->r_ctl.rc_scw = NULL;
15060 }
15061 #endif
15062 if (rack->r_ctl.fsb.tcp_ip_hdr) {
15063 free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15064 rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15065 rack->r_ctl.fsb.th = NULL;
15066 }
15067 if (rack->rc_always_pace == 1) {
15068 rack_remove_pacing(rack);
15069 }
15070 /* Clean up any options if they were not applied */
15071 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15072 struct deferred_opt_list *dol;
15073
15074 dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15075 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15076 free(dol, M_TCPDO);
15077 }
15078 /* rack does not use force data but other stacks may clear it */
15079 if (rack->r_ctl.crte != NULL) {
15080 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15081 rack->rack_hdrw_pacing = 0;
15082 rack->r_ctl.crte = NULL;
15083 }
15084 #ifdef TCP_BLACKBOX
15085 tcp_log_flowend(tp);
15086 #endif
15087 /*
15088 * Lets take a different approach to purging just
15089 * get each one and free it like a cum-ack would and
15090 * not use a foreach loop.
15091 */
15092 rsm = tqhash_min(rack->r_ctl.tqh);
15093 while (rsm) {
15094 tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15095 rack->r_ctl.rc_num_maps_alloced--;
15096 uma_zfree(rack_zone, rsm);
15097 rsm = tqhash_min(rack->r_ctl.tqh);
15098 }
15099 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15100 while (rsm) {
15101 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15102 rack->r_ctl.rc_num_maps_alloced--;
15103 rack->rc_free_cnt--;
15104 cnt_free++;
15105 uma_zfree(rack_zone, rsm);
15106 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15107 }
15108 if (rack->r_ctl.pcm_s != NULL) {
15109 free(rack->r_ctl.pcm_s, M_TCPPCM);
15110 rack->r_ctl.pcm_s = NULL;
15111 rack->r_ctl.pcm_i.cnt_alloc = 0;
15112 rack->r_ctl.pcm_i.cnt = 0;
15113 }
15114 if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15115 (tcp_bblogging_on(tp))) {
15116 union tcp_log_stackspecific log;
15117 struct timeval tv;
15118
15119 memset(&log, 0, sizeof(log));
15120 log.u_bbr.flex8 = 10;
15121 log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15122 log.u_bbr.flex2 = rack->rc_free_cnt;
15123 log.u_bbr.flex3 = cnt_free;
15124 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15125 rsm = tqhash_min(rack->r_ctl.tqh);
15126 log.u_bbr.delRate = (uintptr_t)rsm;
15127 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15128 log.u_bbr.cur_del_rate = (uintptr_t)rsm;
15129 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15130 log.u_bbr.pkt_epoch = __LINE__;
15131 (void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15132 0, &log, false, NULL, NULL, 0, &tv);
15133 }
15134 KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15135 ("rack:%p num_aloc:%u after freeing all?",
15136 rack,
15137 rack->r_ctl.rc_num_maps_alloced));
15138 rack->rc_free_cnt = 0;
15139 free(rack->r_ctl.tqh, M_TCPFSB);
15140 rack->r_ctl.tqh = NULL;
15141 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15142 tp->t_fb_ptr = NULL;
15143 }
15144 /* Make sure snd_nxt is correctly set */
15145 tp->snd_nxt = tp->snd_max;
15146 }
15147
15148 static void
rack_set_state(struct tcpcb * tp,struct tcp_rack * rack)15149 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15150 {
15151 if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15152 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15153 }
15154 switch (tp->t_state) {
15155 case TCPS_SYN_SENT:
15156 rack->r_state = TCPS_SYN_SENT;
15157 rack->r_substate = rack_do_syn_sent;
15158 break;
15159 case TCPS_SYN_RECEIVED:
15160 rack->r_state = TCPS_SYN_RECEIVED;
15161 rack->r_substate = rack_do_syn_recv;
15162 break;
15163 case TCPS_ESTABLISHED:
15164 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15165 rack->r_state = TCPS_ESTABLISHED;
15166 rack->r_substate = rack_do_established;
15167 break;
15168 case TCPS_CLOSE_WAIT:
15169 rack->r_state = TCPS_CLOSE_WAIT;
15170 rack->r_substate = rack_do_close_wait;
15171 break;
15172 case TCPS_FIN_WAIT_1:
15173 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15174 rack->r_state = TCPS_FIN_WAIT_1;
15175 rack->r_substate = rack_do_fin_wait_1;
15176 break;
15177 case TCPS_CLOSING:
15178 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15179 rack->r_state = TCPS_CLOSING;
15180 rack->r_substate = rack_do_closing;
15181 break;
15182 case TCPS_LAST_ACK:
15183 rack_set_pace_segments(tp, rack, __LINE__, NULL);
15184 rack->r_state = TCPS_LAST_ACK;
15185 rack->r_substate = rack_do_lastack;
15186 break;
15187 case TCPS_FIN_WAIT_2:
15188 rack->r_state = TCPS_FIN_WAIT_2;
15189 rack->r_substate = rack_do_fin_wait_2;
15190 break;
15191 case TCPS_LISTEN:
15192 case TCPS_CLOSED:
15193 case TCPS_TIME_WAIT:
15194 default:
15195 break;
15196 };
15197 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15198 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15199
15200 }
15201
15202 static void
rack_timer_audit(struct tcpcb * tp,struct tcp_rack * rack,struct sockbuf * sb)15203 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15204 {
15205 /*
15206 * We received an ack, and then did not
15207 * call send or were bounced out due to the
15208 * hpts was running. Now a timer is up as well, is
15209 * it the right timer?
15210 */
15211 struct rack_sendmap *rsm;
15212 int tmr_up;
15213
15214 tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15215 if (tcp_in_hpts(rack->rc_tp) == 0) {
15216 /*
15217 * Ok we probably need some timer up, but no
15218 * matter what the mask we are not in hpts. We
15219 * may have received an old ack and thus did nothing.
15220 */
15221 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15222 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15223 return;
15224 }
15225 if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15226 return;
15227 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15228 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15229 (tmr_up == PACE_TMR_RXT)) {
15230 /* Should be an RXT */
15231 return;
15232 }
15233 if (rsm == NULL) {
15234 /* Nothing outstanding? */
15235 if (tp->t_flags & TF_DELACK) {
15236 if (tmr_up == PACE_TMR_DELACK)
15237 /* We are supposed to have delayed ack up and we do */
15238 return;
15239 } else if (((V_tcp_always_keepalive ||
15240 rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15241 (tp->t_state <= TCPS_CLOSING)) &&
15242 (tmr_up == PACE_TMR_KEEP) &&
15243 (tp->snd_max == tp->snd_una)) {
15244 /* We should have keep alive up and we do */
15245 return;
15246 }
15247 }
15248 if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15249 ((tmr_up == PACE_TMR_TLP) ||
15250 (tmr_up == PACE_TMR_RACK) ||
15251 (tmr_up == PACE_TMR_RXT))) {
15252 /*
15253 * Either a Rack, TLP or RXT is fine if we
15254 * have outstanding data.
15255 */
15256 return;
15257 } else if (tmr_up == PACE_TMR_DELACK) {
15258 /*
15259 * If the delayed ack was going to go off
15260 * before the rtx/tlp/rack timer were going to
15261 * expire, then that would be the timer in control.
15262 * Note we don't check the time here trusting the
15263 * code is correct.
15264 */
15265 return;
15266 }
15267 /*
15268 * Ok the timer originally started is not what we want now.
15269 * We will force the hpts to be stopped if any, and restart
15270 * with the slot set to what was in the saved slot.
15271 */
15272 if (tcp_in_hpts(rack->rc_tp)) {
15273 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15274 uint32_t us_cts;
15275
15276 us_cts = tcp_get_usecs(NULL);
15277 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15278 rack->r_early = 1;
15279 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15280 }
15281 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15282 }
15283 tcp_hpts_remove(rack->rc_tp);
15284 }
15285 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15286 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15287 }
15288
15289
15290 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)15291 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15292 {
15293 if ((SEQ_LT(tp->snd_wl1, seq) ||
15294 (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15295 (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15296 /* keep track of pure window updates */
15297 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15298 KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15299 tp->snd_wnd = tiwin;
15300 rack_validate_fo_sendwin_up(tp, rack);
15301 tp->snd_wl1 = seq;
15302 tp->snd_wl2 = ack;
15303 if (tp->snd_wnd > tp->max_sndwnd)
15304 tp->max_sndwnd = tp->snd_wnd;
15305 rack->r_wanted_output = 1;
15306 } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15307 tp->snd_wnd = tiwin;
15308 rack_validate_fo_sendwin_up(tp, rack);
15309 tp->snd_wl1 = seq;
15310 tp->snd_wl2 = ack;
15311 } else {
15312 /* Not a valid win update */
15313 return;
15314 }
15315 if (tp->snd_wnd > tp->max_sndwnd)
15316 tp->max_sndwnd = tp->snd_wnd;
15317 /* Do we exit persists? */
15318 if ((rack->rc_in_persist != 0) &&
15319 (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15320 rack->r_ctl.rc_pace_min_segs))) {
15321 rack_exit_persist(tp, rack, cts);
15322 }
15323 /* Do we enter persists? */
15324 if ((rack->rc_in_persist == 0) &&
15325 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15326 TCPS_HAVEESTABLISHED(tp->t_state) &&
15327 ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15328 sbavail(&tptosocket(tp)->so_snd) &&
15329 (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15330 /*
15331 * Here the rwnd is less than
15332 * the pacing size, we are established,
15333 * nothing is outstanding, and there is
15334 * data to send. Enter persists.
15335 */
15336 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15337 }
15338 }
15339
15340 static void
rack_log_input_packet(struct tcpcb * tp,struct tcp_rack * rack,struct tcp_ackent * ae,int ackval,uint32_t high_seq)15341 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15342 {
15343
15344 if (tcp_bblogging_on(rack->rc_tp)) {
15345 struct inpcb *inp = tptoinpcb(tp);
15346 union tcp_log_stackspecific log;
15347 struct timeval ltv;
15348 char tcp_hdr_buf[60];
15349 struct tcphdr *th;
15350 struct timespec ts;
15351 uint32_t orig_snd_una;
15352 uint8_t xx = 0;
15353
15354 #ifdef TCP_REQUEST_TRK
15355 struct tcp_sendfile_track *tcp_req;
15356
15357 if (SEQ_GT(ae->ack, tp->snd_una)) {
15358 tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15359 } else {
15360 tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15361 }
15362 #endif
15363 memset(&log, 0, sizeof(log));
15364 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15365 if (rack->rack_no_prr == 0)
15366 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15367 else
15368 log.u_bbr.flex1 = 0;
15369 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15370 log.u_bbr.use_lt_bw <<= 1;
15371 log.u_bbr.use_lt_bw |= rack->r_might_revert;
15372 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15373 log.u_bbr.bbr_state = rack->rc_free_cnt;
15374 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15375 log.u_bbr.pkts_out = tp->t_maxseg;
15376 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15377 log.u_bbr.flex7 = 1;
15378 log.u_bbr.lost = ae->flags;
15379 log.u_bbr.cwnd_gain = ackval;
15380 log.u_bbr.pacing_gain = 0x2;
15381 if (ae->flags & TSTMP_HDWR) {
15382 /* Record the hardware timestamp if present */
15383 log.u_bbr.flex3 = M_TSTMP;
15384 ts.tv_sec = ae->timestamp / 1000000000;
15385 ts.tv_nsec = ae->timestamp % 1000000000;
15386 ltv.tv_sec = ts.tv_sec;
15387 ltv.tv_usec = ts.tv_nsec / 1000;
15388 log.u_bbr.lt_epoch = tcp_tv_to_usec(<v);
15389 } else if (ae->flags & TSTMP_LRO) {
15390 /* Record the LRO the arrival timestamp */
15391 log.u_bbr.flex3 = M_TSTMP_LRO;
15392 ts.tv_sec = ae->timestamp / 1000000000;
15393 ts.tv_nsec = ae->timestamp % 1000000000;
15394 ltv.tv_sec = ts.tv_sec;
15395 ltv.tv_usec = ts.tv_nsec / 1000;
15396 log.u_bbr.flex5 = tcp_tv_to_usec(<v);
15397 }
15398 log.u_bbr.timeStamp = tcp_get_usecs(<v);
15399 /* Log the rcv time */
15400 log.u_bbr.delRate = ae->timestamp;
15401 #ifdef TCP_REQUEST_TRK
15402 log.u_bbr.applimited = tp->t_tcpreq_closed;
15403 log.u_bbr.applimited <<= 8;
15404 log.u_bbr.applimited |= tp->t_tcpreq_open;
15405 log.u_bbr.applimited <<= 8;
15406 log.u_bbr.applimited |= tp->t_tcpreq_req;
15407 if (tcp_req) {
15408 /* Copy out any client req info */
15409 /* seconds */
15410 log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15411 /* useconds */
15412 log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15413 log.u_bbr.rttProp = tcp_req->timestamp;
15414 log.u_bbr.cur_del_rate = tcp_req->start;
15415 if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15416 log.u_bbr.flex8 |= 1;
15417 } else {
15418 log.u_bbr.flex8 |= 2;
15419 log.u_bbr.bw_inuse = tcp_req->end;
15420 }
15421 log.u_bbr.flex6 = tcp_req->start_seq;
15422 if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15423 log.u_bbr.flex8 |= 4;
15424 log.u_bbr.epoch = tcp_req->end_seq;
15425 }
15426 }
15427 #endif
15428 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15429 th = (struct tcphdr *)tcp_hdr_buf;
15430 th->th_seq = ae->seq;
15431 th->th_ack = ae->ack;
15432 th->th_win = ae->win;
15433 /* Now fill in the ports */
15434 th->th_sport = inp->inp_fport;
15435 th->th_dport = inp->inp_lport;
15436 tcp_set_flags(th, ae->flags);
15437 /* Now do we have a timestamp option? */
15438 if (ae->flags & HAS_TSTMP) {
15439 u_char *cp;
15440 uint32_t val;
15441
15442 th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15443 cp = (u_char *)(th + 1);
15444 *cp = TCPOPT_NOP;
15445 cp++;
15446 *cp = TCPOPT_NOP;
15447 cp++;
15448 *cp = TCPOPT_TIMESTAMP;
15449 cp++;
15450 *cp = TCPOLEN_TIMESTAMP;
15451 cp++;
15452 val = htonl(ae->ts_value);
15453 bcopy((char *)&val,
15454 (char *)cp, sizeof(uint32_t));
15455 val = htonl(ae->ts_echo);
15456 bcopy((char *)&val,
15457 (char *)(cp + 4), sizeof(uint32_t));
15458 } else
15459 th->th_off = (sizeof(struct tcphdr) >> 2);
15460
15461 /*
15462 * For sane logging we need to play a little trick.
15463 * If the ack were fully processed we would have moved
15464 * snd_una to high_seq, but since compressed acks are
15465 * processed in two phases, at this point (logging) snd_una
15466 * won't be advanced. So we would see multiple acks showing
15467 * the advancement. We can prevent that by "pretending" that
15468 * snd_una was advanced and then un-advancing it so that the
15469 * logging code has the right value for tlb_snd_una.
15470 */
15471 if (tp->snd_una != high_seq) {
15472 orig_snd_una = tp->snd_una;
15473 tp->snd_una = high_seq;
15474 xx = 1;
15475 } else
15476 xx = 0;
15477 TCP_LOG_EVENTP(tp, th,
15478 &tptosocket(tp)->so_rcv,
15479 &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15480 0, &log, true, <v);
15481 if (xx) {
15482 tp->snd_una = orig_snd_una;
15483 }
15484 }
15485
15486 }
15487
15488 static void
rack_handle_probe_response(struct tcp_rack * rack,uint32_t tiwin,uint32_t us_cts)15489 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15490 {
15491 uint32_t us_rtt;
15492 /*
15493 * A persist or keep-alive was forced out, update our
15494 * min rtt time. Note now worry about lost responses.
15495 * When a subsequent keep-alive or persist times out
15496 * and forced_ack is still on, then the last probe
15497 * was not responded to. In such cases we have a
15498 * sysctl that controls the behavior. Either we apply
15499 * the rtt but with reduced confidence (0). Or we just
15500 * plain don't apply the rtt estimate. Having data flow
15501 * will clear the probe_not_answered flag i.e. cum-ack
15502 * move forward <or> exiting and reentering persists.
15503 */
15504
15505 rack->forced_ack = 0;
15506 rack->rc_tp->t_rxtshift = 0;
15507 if ((rack->rc_in_persist &&
15508 (tiwin == rack->rc_tp->snd_wnd)) ||
15509 (rack->rc_in_persist == 0)) {
15510 /*
15511 * In persists only apply the RTT update if this is
15512 * a response to our window probe. And that
15513 * means the rwnd sent must match the current
15514 * snd_wnd. If it does not, then we got a
15515 * window update ack instead. For keepalive
15516 * we allow the answer no matter what the window.
15517 *
15518 * Note that if the probe_not_answered is set then
15519 * the forced_ack_ts is the oldest one i.e. the first
15520 * probe sent that might have been lost. This assures
15521 * us that if we do calculate an RTT it is longer not
15522 * some short thing.
15523 */
15524 if (rack->rc_in_persist)
15525 counter_u64_add(rack_persists_acks, 1);
15526 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15527 if (us_rtt == 0)
15528 us_rtt = 1;
15529 if (rack->probe_not_answered == 0) {
15530 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15531 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15532 } else {
15533 /* We have a retransmitted probe here too */
15534 if (rack_apply_rtt_with_reduced_conf) {
15535 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15536 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15537 }
15538 }
15539 }
15540 }
15541
15542 static void
rack_new_round_starts(struct tcpcb * tp,struct tcp_rack * rack,uint32_t high_seq)15543 rack_new_round_starts(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15544 {
15545 /*
15546 * The next send has occurred mark the end of the round
15547 * as when that data gets acknowledged. We can
15548 * also do common things we might need to do when
15549 * a round begins.
15550 */
15551 rack->r_ctl.roundends = tp->snd_max;
15552 rack->rc_new_rnd_needed = 0;
15553 rack_log_hystart_event(rack, tp->snd_max, 4);
15554 }
15555
15556
15557 static void
rack_log_pcm(struct tcp_rack * rack,uint8_t mod,uint32_t flex1,uint32_t flex2,uint32_t flex3)15558 rack_log_pcm(struct tcp_rack *rack, uint8_t mod, uint32_t flex1, uint32_t flex2,
15559 uint32_t flex3)
15560 {
15561 if (tcp_bblogging_on(rack->rc_tp)) {
15562 union tcp_log_stackspecific log;
15563 struct timeval tv;
15564
15565 (void)tcp_get_usecs(&tv);
15566 memset(&log, 0, sizeof(log));
15567 log.u_bbr.timeStamp = tcp_tv_to_usec(&tv);
15568 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15569 log.u_bbr.flex8 = mod;
15570 log.u_bbr.flex1 = flex1;
15571 log.u_bbr.flex2 = flex2;
15572 log.u_bbr.flex3 = flex3;
15573 log.u_bbr.flex4 = rack_pcm_every_n_rounds;
15574 log.u_bbr.flex5 = rack->r_ctl.pcm_idle_rounds;
15575 log.u_bbr.bbr_substate = rack->pcm_needed;
15576 log.u_bbr.bbr_substate <<= 1;
15577 log.u_bbr.bbr_substate |= rack->pcm_in_progress;
15578 log.u_bbr.bbr_substate <<= 1;
15579 log.u_bbr.bbr_substate |= rack->pcm_enabled; /* bits are NIE for Needed, Inprogress, Enabled */
15580 (void)tcp_log_event(rack->rc_tp, NULL, NULL, NULL, TCP_PCM_MEASURE, ERRNO_UNK,
15581 0, &log, false, NULL, NULL, 0, &tv);
15582 }
15583 }
15584
15585 static void
rack_new_round_setup(struct tcpcb * tp,struct tcp_rack * rack,uint32_t high_seq)15586 rack_new_round_setup(struct tcpcb *tp, struct tcp_rack *rack, uint32_t high_seq)
15587 {
15588 /*
15589 * The round (current_round) has ended. We now
15590 * setup for the next round by incrementing the
15591 * round numnber and doing any round specific
15592 * things.
15593 */
15594 rack_log_hystart_event(rack, high_seq, 21);
15595 rack->r_ctl.current_round++;
15596 /* New round (current_round) begins at next send */
15597 rack->rc_new_rnd_needed = 1;
15598 if ((rack->pcm_enabled == 1) &&
15599 (rack->pcm_needed == 0) &&
15600 (rack->pcm_in_progress == 0)) {
15601 /*
15602 * If we have enabled PCM, then we need to
15603 * check if the round has adanced to the state
15604 * where one is required.
15605 */
15606 int rnds;
15607
15608 rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
15609 if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
15610 rack->pcm_needed = 1;
15611 rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15612 } else if (rack_verbose_logging) {
15613 rack_log_pcm(rack, 3, rack->r_ctl.last_pcm_round, rack_pcm_every_n_rounds, rack->r_ctl.current_round );
15614 }
15615 }
15616 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
15617 /* We have hystart enabled send the round info in */
15618 if (CC_ALGO(tp)->newround != NULL) {
15619 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
15620 }
15621 }
15622 /*
15623 * For DGP an initial startup check. We want to validate
15624 * that we are not just pushing on slow-start and just
15625 * not gaining.. i.e. filling buffers without getting any
15626 * boost in b/w during the inital slow-start.
15627 */
15628 if (rack->dgp_on &&
15629 (rack->rc_initial_ss_comp == 0) &&
15630 (tp->snd_cwnd < tp->snd_ssthresh) &&
15631 (rack->r_ctl.num_measurements >= RACK_REQ_AVG) &&
15632 (rack->r_ctl.gp_rnd_thresh > 0) &&
15633 ((rack->r_ctl.current_round - rack->r_ctl.last_rnd_of_gp_rise) >= rack->r_ctl.gp_rnd_thresh)) {
15634
15635 /*
15636 * We are in the initial SS and we have hd rack_rnd_cnt_req rounds(def:5) where
15637 * we have not gained the required amount in the gp_est (120.0% aka 1200). Lets
15638 * exit SS.
15639 *
15640 * Pick up the flight size now as we enter slowstart (not the
15641 * cwnd which may be inflated).
15642 */
15643 rack->rc_initial_ss_comp = 1;
15644
15645 if (tcp_bblogging_on(rack->rc_tp)) {
15646 union tcp_log_stackspecific log;
15647 struct timeval tv;
15648
15649 memset(&log, 0, sizeof(log));
15650 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15651 log.u_bbr.flex1 = rack->r_ctl.current_round;
15652 log.u_bbr.flex2 = rack->r_ctl.last_rnd_of_gp_rise;
15653 log.u_bbr.flex3 = rack->r_ctl.gp_rnd_thresh;
15654 log.u_bbr.flex4 = rack->r_ctl.gate_to_fs;
15655 log.u_bbr.flex5 = rack->r_ctl.ss_hi_fs;
15656 log.u_bbr.flex8 = 40;
15657 (void)tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
15658 0, &log, false, NULL, __func__, __LINE__,&tv);
15659 }
15660 if ((rack->r_ctl.gate_to_fs == 1) &&
15661 (tp->snd_cwnd > rack->r_ctl.ss_hi_fs)) {
15662 tp->snd_cwnd = rack->r_ctl.ss_hi_fs;
15663 }
15664 tp->snd_ssthresh = tp->snd_cwnd - 1;
15665 /* Turn off any fast output running */
15666 rack->r_fast_output = 0;
15667 }
15668 }
15669
15670 static int
rack_do_compressed_ack_processing(struct tcpcb * tp,struct socket * so,struct mbuf * m,int nxt_pkt,struct timeval * tv)15671 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15672 {
15673 /*
15674 * Handle a "special" compressed ack mbuf. Each incoming
15675 * ack has only four possible dispositions:
15676 *
15677 * A) It moves the cum-ack forward
15678 * B) It is behind the cum-ack.
15679 * C) It is a window-update ack.
15680 * D) It is a dup-ack.
15681 *
15682 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15683 * in the incoming mbuf. We also need to still pay attention
15684 * to nxt_pkt since there may be another packet after this
15685 * one.
15686 */
15687 #ifdef TCP_ACCOUNTING
15688 uint64_t ts_val;
15689 uint64_t rdstc;
15690 #endif
15691 int segsiz;
15692 struct timespec ts;
15693 struct tcp_rack *rack;
15694 struct tcp_ackent *ae;
15695 uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15696 int cnt, i, did_out, ourfinisacked = 0;
15697 struct tcpopt to_holder, *to = NULL;
15698 #ifdef TCP_ACCOUNTING
15699 int win_up_req = 0;
15700 #endif
15701 int nsegs = 0;
15702 int under_pacing = 0;
15703 int post_recovery = 0;
15704 #ifdef TCP_ACCOUNTING
15705 sched_pin();
15706 #endif
15707 rack = (struct tcp_rack *)tp->t_fb_ptr;
15708 if (rack->gp_ready &&
15709 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15710 under_pacing = 1;
15711
15712 if (rack->r_state != tp->t_state)
15713 rack_set_state(tp, rack);
15714 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15715 (tp->t_flags & TF_GPUTINPROG)) {
15716 /*
15717 * We have a goodput in progress
15718 * and we have entered a late state.
15719 * Do we have enough data in the sb
15720 * to handle the GPUT request?
15721 */
15722 uint32_t bytes;
15723
15724 bytes = tp->gput_ack - tp->gput_seq;
15725 if (SEQ_GT(tp->gput_seq, tp->snd_una))
15726 bytes += tp->gput_seq - tp->snd_una;
15727 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15728 /*
15729 * There are not enough bytes in the socket
15730 * buffer that have been sent to cover this
15731 * measurement. Cancel it.
15732 */
15733 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15734 rack->r_ctl.rc_gp_srtt /*flex1*/,
15735 tp->gput_seq,
15736 0, 0, 18, __LINE__, NULL, 0);
15737 tp->t_flags &= ~TF_GPUTINPROG;
15738 }
15739 }
15740 to = &to_holder;
15741 to->to_flags = 0;
15742 KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15743 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15744 cnt = m->m_len / sizeof(struct tcp_ackent);
15745 counter_u64_add(rack_multi_single_eq, cnt);
15746 high_seq = tp->snd_una;
15747 the_win = tp->snd_wnd;
15748 win_seq = tp->snd_wl1;
15749 win_upd_ack = tp->snd_wl2;
15750 cts = tcp_tv_to_usec(tv);
15751 ms_cts = tcp_tv_to_msec(tv);
15752 rack->r_ctl.rc_rcvtime = cts;
15753 segsiz = ctf_fixed_maxseg(tp);
15754 if ((rack->rc_gp_dyn_mul) &&
15755 (rack->use_fixed_rate == 0) &&
15756 (rack->rc_always_pace)) {
15757 /* Check in on probertt */
15758 rack_check_probe_rtt(rack, cts);
15759 }
15760 for (i = 0; i < cnt; i++) {
15761 #ifdef TCP_ACCOUNTING
15762 ts_val = get_cyclecount();
15763 #endif
15764 rack_clear_rate_sample(rack);
15765 ae = ((mtod(m, struct tcp_ackent *)) + i);
15766 if (ae->flags & TH_FIN)
15767 rack_log_pacing_delay_calc(rack,
15768 0,
15769 0,
15770 0,
15771 rack_get_gp_est(rack), /* delRate */
15772 rack_get_lt_bw(rack), /* rttProp */
15773 20, __LINE__, NULL, 0);
15774 /* Setup the window */
15775 tiwin = ae->win << tp->snd_scale;
15776 if (tiwin > rack->r_ctl.rc_high_rwnd)
15777 rack->r_ctl.rc_high_rwnd = tiwin;
15778 /* figure out the type of ack */
15779 if (SEQ_LT(ae->ack, high_seq)) {
15780 /* Case B*/
15781 ae->ack_val_set = ACK_BEHIND;
15782 } else if (SEQ_GT(ae->ack, high_seq)) {
15783 /* Case A */
15784 ae->ack_val_set = ACK_CUMACK;
15785 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
15786 /* Case D */
15787 ae->ack_val_set = ACK_DUPACK;
15788 } else {
15789 /* Case C */
15790 ae->ack_val_set = ACK_RWND;
15791 }
15792 rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
15793 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
15794 /* Validate timestamp */
15795 if (ae->flags & HAS_TSTMP) {
15796 /* Setup for a timestamp */
15797 to->to_flags = TOF_TS;
15798 ae->ts_echo -= tp->ts_offset;
15799 to->to_tsecr = ae->ts_echo;
15800 to->to_tsval = ae->ts_value;
15801 /*
15802 * If echoed timestamp is later than the current time, fall back to
15803 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
15804 * were used when this connection was established.
15805 */
15806 if (TSTMP_GT(ae->ts_echo, ms_cts))
15807 to->to_tsecr = 0;
15808 if (tp->ts_recent &&
15809 TSTMP_LT(ae->ts_value, tp->ts_recent)) {
15810 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
15811 #ifdef TCP_ACCOUNTING
15812 rdstc = get_cyclecount();
15813 if (rdstc > ts_val) {
15814 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15815 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15816 }
15817 }
15818 #endif
15819 continue;
15820 }
15821 }
15822 if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
15823 SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
15824 tp->ts_recent_age = tcp_ts_getticks();
15825 tp->ts_recent = ae->ts_value;
15826 }
15827 } else {
15828 /* Setup for a no options */
15829 to->to_flags = 0;
15830 }
15831 /* Update the rcv time and perform idle reduction possibly */
15832 if (tp->t_idle_reduce &&
15833 (tp->snd_max == tp->snd_una) &&
15834 (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
15835 counter_u64_add(rack_input_idle_reduces, 1);
15836 rack_cc_after_idle(rack, tp);
15837 }
15838 tp->t_rcvtime = ticks;
15839 /* Now what about ECN of a chain of pure ACKs? */
15840 if (tcp_ecn_input_segment(tp, ae->flags, 0,
15841 tcp_packets_this_ack(tp, ae->ack),
15842 ae->codepoint))
15843 rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
15844 #ifdef TCP_ACCOUNTING
15845 /* Count for the specific type of ack in */
15846 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15847 tp->tcp_cnt_counters[ae->ack_val_set]++;
15848 }
15849 #endif
15850 /*
15851 * Note how we could move up these in the determination
15852 * above, but we don't so that way the timestamp checks (and ECN)
15853 * is done first before we do any processing on the ACK.
15854 * The non-compressed path through the code has this
15855 * weakness (noted by @jtl) that it actually does some
15856 * processing before verifying the timestamp information.
15857 * We don't take that path here which is why we set
15858 * the ack_val_set first, do the timestamp and ecn
15859 * processing, and then look at what we have setup.
15860 */
15861 if (ae->ack_val_set == ACK_BEHIND) {
15862 /*
15863 * Case B flag reordering, if window is not closed
15864 * or it could be a keep-alive or persists
15865 */
15866 if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
15867 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
15868 if (rack->r_ctl.rc_reorder_ts == 0)
15869 rack->r_ctl.rc_reorder_ts = 1;
15870 }
15871 } else if (ae->ack_val_set == ACK_DUPACK) {
15872 /* Case D */
15873 rack_strike_dupack(rack, ae->ack);
15874 } else if (ae->ack_val_set == ACK_RWND) {
15875 /* Case C */
15876 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15877 ts.tv_sec = ae->timestamp / 1000000000;
15878 ts.tv_nsec = ae->timestamp % 1000000000;
15879 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15880 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15881 } else {
15882 rack->r_ctl.act_rcv_time = *tv;
15883 }
15884 if (rack->forced_ack) {
15885 rack_handle_probe_response(rack, tiwin,
15886 tcp_tv_to_usec(&rack->r_ctl.act_rcv_time));
15887 }
15888 #ifdef TCP_ACCOUNTING
15889 win_up_req = 1;
15890 #endif
15891 win_upd_ack = ae->ack;
15892 win_seq = ae->seq;
15893 the_win = tiwin;
15894 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15895 } else {
15896 /* Case A */
15897 if (SEQ_GT(ae->ack, tp->snd_max)) {
15898 /*
15899 * We just send an ack since the incoming
15900 * ack is beyond the largest seq we sent.
15901 */
15902 if ((tp->t_flags & TF_ACKNOW) == 0) {
15903 ctf_ack_war_checks(tp);
15904 if (tp->t_flags && TF_ACKNOW)
15905 rack->r_wanted_output = 1;
15906 }
15907 } else {
15908 nsegs++;
15909 /* If the window changed setup to update */
15910 if (tiwin != tp->snd_wnd) {
15911 win_upd_ack = ae->ack;
15912 win_seq = ae->seq;
15913 the_win = tiwin;
15914 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
15915 }
15916 #ifdef TCP_ACCOUNTING
15917 /* Account for the acks */
15918 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15919 tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
15920 }
15921 #endif
15922 high_seq = ae->ack;
15923 /* Setup our act_rcv_time */
15924 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
15925 ts.tv_sec = ae->timestamp / 1000000000;
15926 ts.tv_nsec = ae->timestamp % 1000000000;
15927 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
15928 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
15929 } else {
15930 rack->r_ctl.act_rcv_time = *tv;
15931 }
15932 rack_process_to_cumack(tp, rack, ae->ack, cts, to,
15933 tcp_tv_to_lusec(&rack->r_ctl.act_rcv_time));
15934 #ifdef TCP_REQUEST_TRK
15935 rack_req_check_for_comp(rack, high_seq);
15936 #endif
15937 if (rack->rc_dsack_round_seen) {
15938 /* Is the dsack round over? */
15939 if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
15940 /* Yes it is */
15941 rack->rc_dsack_round_seen = 0;
15942 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
15943 }
15944 }
15945 }
15946 }
15947 /* And lets be sure to commit the rtt measurements for this ack */
15948 tcp_rack_xmit_timer_commit(rack, tp);
15949 #ifdef TCP_ACCOUNTING
15950 rdstc = get_cyclecount();
15951 if (rdstc > ts_val) {
15952 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15953 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
15954 if (ae->ack_val_set == ACK_CUMACK)
15955 tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
15956 }
15957 }
15958 #endif
15959 }
15960 #ifdef TCP_ACCOUNTING
15961 ts_val = get_cyclecount();
15962 #endif
15963 /* Tend to any collapsed window */
15964 if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
15965 /* The peer collapsed the window */
15966 rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
15967 } else if (rack->rc_has_collapsed)
15968 rack_un_collapse_window(rack, __LINE__);
15969 if ((rack->r_collapse_point_valid) &&
15970 (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
15971 rack->r_collapse_point_valid = 0;
15972 acked_amount = acked = (high_seq - tp->snd_una);
15973 if (acked) {
15974 /*
15975 * The draft (v3) calls for us to use SEQ_GEQ, but that
15976 * causes issues when we are just going app limited. Lets
15977 * instead use SEQ_GT <or> where its equal but more data
15978 * is outstanding.
15979 *
15980 * Also make sure we are on the last ack of a series. We
15981 * have to have all the ack's processed in queue to know
15982 * if there is something left outstanding.
15983 *
15984 */
15985 if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
15986 (rack->rc_new_rnd_needed == 0) &&
15987 (nxt_pkt == 0)) {
15988 /*
15989 * We have crossed into a new round with
15990 * this th_ack value.
15991 */
15992 rack_new_round_setup(tp, rack, high_seq);
15993 }
15994 /*
15995 * Clear the probe not answered flag
15996 * since cum-ack moved forward.
15997 */
15998 rack->probe_not_answered = 0;
15999 if (tp->t_flags & TF_NEEDSYN) {
16000 /*
16001 * T/TCP: Connection was half-synchronized, and our SYN has
16002 * been ACK'd (so connection is now fully synchronized). Go
16003 * to non-starred state, increment snd_una for ACK of SYN,
16004 * and check if we can do window scaling.
16005 */
16006 tp->t_flags &= ~TF_NEEDSYN;
16007 tp->snd_una++;
16008 acked_amount = acked = (high_seq - tp->snd_una);
16009 }
16010 if (acked > sbavail(&so->so_snd))
16011 acked_amount = sbavail(&so->so_snd);
16012 if (IN_FASTRECOVERY(tp->t_flags) &&
16013 (rack->rack_no_prr == 0))
16014 rack_update_prr(tp, rack, acked_amount, high_seq);
16015 if (IN_RECOVERY(tp->t_flags)) {
16016 if (SEQ_LT(high_seq, tp->snd_recover) &&
16017 (SEQ_LT(high_seq, tp->snd_max))) {
16018 tcp_rack_partialack(tp);
16019 } else {
16020 rack_post_recovery(tp, high_seq);
16021 post_recovery = 1;
16022 }
16023 } else if ((rack->rto_from_rec == 1) &&
16024 SEQ_GEQ(high_seq, tp->snd_recover)) {
16025 /*
16026 * We were in recovery, hit a rxt timeout
16027 * and never re-entered recovery. The timeout(s)
16028 * made up all the lost data. In such a case
16029 * we need to clear the rto_from_rec flag.
16030 */
16031 rack->rto_from_rec = 0;
16032 }
16033 /* Handle the rack-log-ack part (sendmap) */
16034 if ((sbused(&so->so_snd) == 0) &&
16035 (acked > acked_amount) &&
16036 (tp->t_state >= TCPS_FIN_WAIT_1) &&
16037 (tp->t_flags & TF_SENTFIN)) {
16038 /*
16039 * We must be sure our fin
16040 * was sent and acked (we can be
16041 * in FIN_WAIT_1 without having
16042 * sent the fin).
16043 */
16044 ourfinisacked = 1;
16045 /*
16046 * Lets make sure snd_una is updated
16047 * since most likely acked_amount = 0 (it
16048 * should be).
16049 */
16050 tp->snd_una = high_seq;
16051 }
16052 /* Did we make a RTO error? */
16053 if ((tp->t_flags & TF_PREVVALID) &&
16054 ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16055 tp->t_flags &= ~TF_PREVVALID;
16056 if (tp->t_rxtshift == 1 &&
16057 (int)(ticks - tp->t_badrxtwin) < 0)
16058 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16059 }
16060 /* Handle the data in the socket buffer */
16061 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16062 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16063 if (acked_amount > 0) {
16064 uint32_t p_cwnd;
16065 struct mbuf *mfree;
16066
16067 if (post_recovery) {
16068 /*
16069 * Grab the segsiz, multiply by 2 and add the snd_cwnd
16070 * that is the max the CC should add if we are exiting
16071 * recovery and doing a late add.
16072 */
16073 p_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16074 p_cwnd <<= 1;
16075 p_cwnd += tp->snd_cwnd;
16076 }
16077 rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, post_recovery);
16078 if (post_recovery && (tp->snd_cwnd > p_cwnd)) {
16079 /* Must be non-newreno (cubic) getting too ahead of itself */
16080 tp->snd_cwnd = p_cwnd;
16081 }
16082 SOCK_SENDBUF_LOCK(so);
16083 mfree = sbcut_locked(&so->so_snd, acked_amount);
16084 tp->snd_una = high_seq;
16085 /* Note we want to hold the sb lock through the sendmap adjust */
16086 rack_adjust_sendmap_head(rack, &so->so_snd);
16087 /* Wake up the socket if we have room to write more */
16088 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16089 sowwakeup_locked(so);
16090 m_freem(mfree);
16091 }
16092 /* update progress */
16093 tp->t_acktime = ticks;
16094 rack_log_progress_event(rack, tp, tp->t_acktime,
16095 PROGRESS_UPDATE, __LINE__);
16096 /* Clear out shifts and such */
16097 tp->t_rxtshift = 0;
16098 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16099 rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16100 rack->rc_tlp_in_progress = 0;
16101 rack->r_ctl.rc_tlp_cnt_out = 0;
16102 /* Send recover and snd_nxt must be dragged along */
16103 if (SEQ_GT(tp->snd_una, tp->snd_recover))
16104 tp->snd_recover = tp->snd_una;
16105 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16106 tp->snd_nxt = tp->snd_max;
16107 /*
16108 * If the RXT timer is running we want to
16109 * stop it, so we can restart a TLP (or new RXT).
16110 */
16111 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16112 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16113 tp->snd_wl2 = high_seq;
16114 tp->t_dupacks = 0;
16115 if (under_pacing &&
16116 (rack->use_fixed_rate == 0) &&
16117 (rack->in_probe_rtt == 0) &&
16118 rack->rc_gp_dyn_mul &&
16119 rack->rc_always_pace) {
16120 /* Check if we are dragging bottom */
16121 rack_check_bottom_drag(tp, rack, so);
16122 }
16123 if (tp->snd_una == tp->snd_max) {
16124 tp->t_flags &= ~TF_PREVVALID;
16125 rack->r_ctl.retran_during_recovery = 0;
16126 rack->rc_suspicious = 0;
16127 rack->r_ctl.dsack_byte_cnt = 0;
16128 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16129 if (rack->r_ctl.rc_went_idle_time == 0)
16130 rack->r_ctl.rc_went_idle_time = 1;
16131 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16132 if (sbavail(&tptosocket(tp)->so_snd) == 0)
16133 tp->t_acktime = 0;
16134 /* Set so we might enter persists... */
16135 rack->r_wanted_output = 1;
16136 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16137 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16138 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16139 (sbavail(&so->so_snd) == 0) &&
16140 (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16141 /*
16142 * The socket was gone and the
16143 * peer sent data (not now in the past), time to
16144 * reset him.
16145 */
16146 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16147 /* tcp_close will kill the inp pre-log the Reset */
16148 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16149 #ifdef TCP_ACCOUNTING
16150 rdstc = get_cyclecount();
16151 if (rdstc > ts_val) {
16152 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16153 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16154 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16155 }
16156 }
16157 #endif
16158 m_freem(m);
16159 tp = tcp_close(tp);
16160 if (tp == NULL) {
16161 #ifdef TCP_ACCOUNTING
16162 sched_unpin();
16163 #endif
16164 return (1);
16165 }
16166 /*
16167 * We would normally do drop-with-reset which would
16168 * send back a reset. We can't since we don't have
16169 * all the needed bits. Instead lets arrange for
16170 * a call to tcp_output(). That way since we
16171 * are in the closed state we will generate a reset.
16172 *
16173 * Note if tcp_accounting is on we don't unpin since
16174 * we do that after the goto label.
16175 */
16176 goto send_out_a_rst;
16177 }
16178 if ((sbused(&so->so_snd) == 0) &&
16179 (tp->t_state >= TCPS_FIN_WAIT_1) &&
16180 (tp->t_flags & TF_SENTFIN)) {
16181 /*
16182 * If we can't receive any more data, then closing user can
16183 * proceed. Starting the timer is contrary to the
16184 * specification, but if we don't get a FIN we'll hang
16185 * forever.
16186 *
16187 */
16188 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16189 soisdisconnected(so);
16190 tcp_timer_activate(tp, TT_2MSL,
16191 (tcp_fast_finwait2_recycle ?
16192 tcp_finwait2_timeout :
16193 TP_MAXIDLE(tp)));
16194 }
16195 if (ourfinisacked == 0) {
16196 /*
16197 * We don't change to fin-wait-2 if we have our fin acked
16198 * which means we are probably in TCPS_CLOSING.
16199 */
16200 tcp_state_change(tp, TCPS_FIN_WAIT_2);
16201 }
16202 }
16203 }
16204 /* Wake up the socket if we have room to write more */
16205 if (sbavail(&so->so_snd)) {
16206 rack->r_wanted_output = 1;
16207 if (ctf_progress_timeout_check(tp, true)) {
16208 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16209 tp, tick, PROGRESS_DROP, __LINE__);
16210 /*
16211 * We cheat here and don't send a RST, we should send one
16212 * when the pacer drops the connection.
16213 */
16214 #ifdef TCP_ACCOUNTING
16215 rdstc = get_cyclecount();
16216 if (rdstc > ts_val) {
16217 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16218 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16219 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16220 }
16221 }
16222 sched_unpin();
16223 #endif
16224 (void)tcp_drop(tp, ETIMEDOUT);
16225 m_freem(m);
16226 return (1);
16227 }
16228 }
16229 if (ourfinisacked) {
16230 switch(tp->t_state) {
16231 case TCPS_CLOSING:
16232 #ifdef TCP_ACCOUNTING
16233 rdstc = get_cyclecount();
16234 if (rdstc > ts_val) {
16235 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16236 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16237 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16238 }
16239 }
16240 sched_unpin();
16241 #endif
16242 tcp_twstart(tp);
16243 m_freem(m);
16244 return (1);
16245 break;
16246 case TCPS_LAST_ACK:
16247 #ifdef TCP_ACCOUNTING
16248 rdstc = get_cyclecount();
16249 if (rdstc > ts_val) {
16250 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16251 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16252 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16253 }
16254 }
16255 sched_unpin();
16256 #endif
16257 tp = tcp_close(tp);
16258 ctf_do_drop(m, tp);
16259 return (1);
16260 break;
16261 case TCPS_FIN_WAIT_1:
16262 #ifdef TCP_ACCOUNTING
16263 rdstc = get_cyclecount();
16264 if (rdstc > ts_val) {
16265 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16266 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16267 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16268 }
16269 }
16270 #endif
16271 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16272 soisdisconnected(so);
16273 tcp_timer_activate(tp, TT_2MSL,
16274 (tcp_fast_finwait2_recycle ?
16275 tcp_finwait2_timeout :
16276 TP_MAXIDLE(tp)));
16277 }
16278 tcp_state_change(tp, TCPS_FIN_WAIT_2);
16279 break;
16280 default:
16281 break;
16282 }
16283 }
16284 if (rack->r_fast_output) {
16285 /*
16286 * We re doing fast output.. can we expand that?
16287 */
16288 rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16289 }
16290 #ifdef TCP_ACCOUNTING
16291 rdstc = get_cyclecount();
16292 if (rdstc > ts_val) {
16293 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16294 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16295 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16296 }
16297 }
16298
16299 } else if (win_up_req) {
16300 rdstc = get_cyclecount();
16301 if (rdstc > ts_val) {
16302 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16303 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16304 }
16305 }
16306 #endif
16307 }
16308 /* Now is there a next packet, if so we are done */
16309 m_freem(m);
16310 did_out = 0;
16311 if (nxt_pkt) {
16312 #ifdef TCP_ACCOUNTING
16313 sched_unpin();
16314 #endif
16315 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16316 return (0);
16317 }
16318 rack_handle_might_revert(tp, rack);
16319 ctf_calc_rwin(so, tp);
16320 if ((rack->r_wanted_output != 0) ||
16321 (rack->r_fast_output != 0) ||
16322 (tp->t_flags & TF_ACKNOW )) {
16323 send_out_a_rst:
16324 if (tcp_output(tp) < 0) {
16325 #ifdef TCP_ACCOUNTING
16326 sched_unpin();
16327 #endif
16328 return (1);
16329 }
16330 did_out = 1;
16331 }
16332 if (tp->t_flags2 & TF2_HPTS_CALLS)
16333 tp->t_flags2 &= ~TF2_HPTS_CALLS;
16334 rack_free_trim(rack);
16335 #ifdef TCP_ACCOUNTING
16336 sched_unpin();
16337 #endif
16338 rack_timer_audit(tp, rack, &so->so_snd);
16339 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16340 return (0);
16341 }
16342
16343 #define TCP_LRO_TS_OPTION \
16344 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16345 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16346
16347 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)16348 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16349 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16350 struct timeval *tv)
16351 {
16352 struct inpcb *inp = tptoinpcb(tp);
16353 struct socket *so = tptosocket(tp);
16354 #ifdef TCP_ACCOUNTING
16355 uint64_t ts_val;
16356 #endif
16357 int32_t thflags, retval, did_out = 0;
16358 int32_t way_out = 0;
16359 /*
16360 * cts - is the current time from tv (caller gets ts) in microseconds.
16361 * ms_cts - is the current time from tv in milliseconds.
16362 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16363 */
16364 uint32_t cts, us_cts, ms_cts;
16365 uint32_t tiwin;
16366 struct timespec ts;
16367 struct tcpopt to;
16368 struct tcp_rack *rack;
16369 struct rack_sendmap *rsm;
16370 int32_t prev_state = 0;
16371 int no_output = 0;
16372 int slot_remaining = 0;
16373 #ifdef TCP_ACCOUNTING
16374 int ack_val_set = 0xf;
16375 #endif
16376 int nsegs;
16377
16378 NET_EPOCH_ASSERT();
16379 INP_WLOCK_ASSERT(inp);
16380
16381 /*
16382 * tv passed from common code is from either M_TSTMP_LRO or
16383 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16384 */
16385 rack = (struct tcp_rack *)tp->t_fb_ptr;
16386 if (rack->rack_deferred_inited == 0) {
16387 /*
16388 * If we are the connecting socket we will
16389 * hit rack_init() when no sequence numbers
16390 * are setup. This makes it so we must defer
16391 * some initialization. Call that now.
16392 */
16393 rack_deferred_init(tp, rack);
16394 }
16395 /*
16396 * Check to see if we need to skip any output plans. This
16397 * can happen in the non-LRO path where we are pacing and
16398 * must process the ack coming in but need to defer sending
16399 * anything becase a pacing timer is running.
16400 */
16401 us_cts = tcp_tv_to_usec(tv);
16402 if (m->m_flags & M_ACKCMP) {
16403 /*
16404 * All compressed ack's are ack's by definition so
16405 * remove any ack required flag and then do the processing.
16406 */
16407 rack->rc_ack_required = 0;
16408 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16409 }
16410 thflags = tcp_get_flags(th);
16411 if ((rack->rc_always_pace == 1) &&
16412 (rack->rc_ack_can_sendout_data == 0) &&
16413 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16414 (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16415 /*
16416 * Ok conditions are right for queuing the packets
16417 * but we do have to check the flags in the inp, it
16418 * could be, if a sack is present, we want to be awoken and
16419 * so should process the packets.
16420 */
16421 slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16422 if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16423 no_output = 1;
16424 } else {
16425 /*
16426 * If there is no options, or just a
16427 * timestamp option, we will want to queue
16428 * the packets. This is the same that LRO does
16429 * and will need to change with accurate ECN.
16430 */
16431 uint32_t *ts_ptr;
16432 int optlen;
16433
16434 optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16435 ts_ptr = (uint32_t *)(th + 1);
16436 if ((optlen == 0) ||
16437 ((optlen == TCPOLEN_TSTAMP_APPA) &&
16438 (*ts_ptr == TCP_LRO_TS_OPTION)))
16439 no_output = 1;
16440 }
16441 if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16442 /*
16443 * It is unrealistic to think we can pace in less than
16444 * the minimum granularity of the pacer (def:250usec). So
16445 * if we have less than that time remaining we should go
16446 * ahead and allow output to be "early". We will attempt to
16447 * make up for it in any pacing time we try to apply on
16448 * the outbound packet.
16449 */
16450 no_output = 0;
16451 }
16452 }
16453 /*
16454 * If there is a RST or FIN lets dump out the bw
16455 * with a FIN the connection may go on but we
16456 * may not.
16457 */
16458 if ((thflags & TH_FIN) || (thflags & TH_RST))
16459 rack_log_pacing_delay_calc(rack,
16460 rack->r_ctl.gp_bw,
16461 0,
16462 0,
16463 rack_get_gp_est(rack), /* delRate */
16464 rack_get_lt_bw(rack), /* rttProp */
16465 20, __LINE__, NULL, 0);
16466 if (m->m_flags & M_ACKCMP) {
16467 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16468 }
16469 cts = tcp_tv_to_usec(tv);
16470 ms_cts = tcp_tv_to_msec(tv);
16471 nsegs = m->m_pkthdr.lro_nsegs;
16472 counter_u64_add(rack_proc_non_comp_ack, 1);
16473 #ifdef TCP_ACCOUNTING
16474 sched_pin();
16475 if (thflags & TH_ACK)
16476 ts_val = get_cyclecount();
16477 #endif
16478 if ((m->m_flags & M_TSTMP) ||
16479 (m->m_flags & M_TSTMP_LRO)) {
16480 mbuf_tstmp2timespec(m, &ts);
16481 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16482 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16483 } else
16484 rack->r_ctl.act_rcv_time = *tv;
16485 kern_prefetch(rack, &prev_state);
16486 prev_state = 0;
16487 /*
16488 * Unscale the window into a 32-bit value. For the SYN_SENT state
16489 * the scale is zero.
16490 */
16491 tiwin = th->th_win << tp->snd_scale;
16492 #ifdef TCP_ACCOUNTING
16493 if (thflags & TH_ACK) {
16494 /*
16495 * We have a tradeoff here. We can either do what we are
16496 * doing i.e. pinning to this CPU and then doing the accounting
16497 * <or> we could do a critical enter, setup the rdtsc and cpu
16498 * as in below, and then validate we are on the same CPU on
16499 * exit. I have choosen to not do the critical enter since
16500 * that often will gain you a context switch, and instead lock
16501 * us (line above this if) to the same CPU with sched_pin(). This
16502 * means we may be context switched out for a higher priority
16503 * interupt but we won't be moved to another CPU.
16504 *
16505 * If this occurs (which it won't very often since we most likely
16506 * are running this code in interupt context and only a higher
16507 * priority will bump us ... clock?) we will falsely add in
16508 * to the time the interupt processing time plus the ack processing
16509 * time. This is ok since its a rare event.
16510 */
16511 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16512 ctf_fixed_maxseg(tp));
16513 }
16514 #endif
16515 /*
16516 * Parse options on any incoming segment.
16517 */
16518 memset(&to, 0, sizeof(to));
16519 tcp_dooptions(&to, (u_char *)(th + 1),
16520 (th->th_off << 2) - sizeof(struct tcphdr),
16521 (thflags & TH_SYN) ? TO_SYN : 0);
16522 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16523 __func__));
16524 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16525 __func__));
16526 if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) {
16527 /*
16528 * We don't look at sack's from the
16529 * peer because the MSS is too small which
16530 * can subject us to an attack.
16531 */
16532 to.to_flags &= ~TOF_SACK;
16533 }
16534 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16535 (tp->t_flags & TF_GPUTINPROG)) {
16536 /*
16537 * We have a goodput in progress
16538 * and we have entered a late state.
16539 * Do we have enough data in the sb
16540 * to handle the GPUT request?
16541 */
16542 uint32_t bytes;
16543
16544 bytes = tp->gput_ack - tp->gput_seq;
16545 if (SEQ_GT(tp->gput_seq, tp->snd_una))
16546 bytes += tp->gput_seq - tp->snd_una;
16547 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16548 /*
16549 * There are not enough bytes in the socket
16550 * buffer that have been sent to cover this
16551 * measurement. Cancel it.
16552 */
16553 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16554 rack->r_ctl.rc_gp_srtt /*flex1*/,
16555 tp->gput_seq,
16556 0, 0, 18, __LINE__, NULL, 0);
16557 tp->t_flags &= ~TF_GPUTINPROG;
16558 }
16559 }
16560 if (tcp_bblogging_on(rack->rc_tp)) {
16561 union tcp_log_stackspecific log;
16562 struct timeval ltv;
16563 #ifdef TCP_REQUEST_TRK
16564 struct tcp_sendfile_track *tcp_req;
16565
16566 if (SEQ_GT(th->th_ack, tp->snd_una)) {
16567 tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16568 } else {
16569 tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16570 }
16571 #endif
16572 memset(&log, 0, sizeof(log));
16573 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16574 if (rack->rack_no_prr == 0)
16575 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16576 else
16577 log.u_bbr.flex1 = 0;
16578 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16579 log.u_bbr.use_lt_bw <<= 1;
16580 log.u_bbr.use_lt_bw |= rack->r_might_revert;
16581 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16582 log.u_bbr.bbr_state = rack->rc_free_cnt;
16583 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16584 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16585 log.u_bbr.flex3 = m->m_flags;
16586 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16587 log.u_bbr.lost = thflags;
16588 log.u_bbr.pacing_gain = 0x1;
16589 #ifdef TCP_ACCOUNTING
16590 log.u_bbr.cwnd_gain = ack_val_set;
16591 #endif
16592 log.u_bbr.flex7 = 2;
16593 if (m->m_flags & M_TSTMP) {
16594 /* Record the hardware timestamp if present */
16595 mbuf_tstmp2timespec(m, &ts);
16596 ltv.tv_sec = ts.tv_sec;
16597 ltv.tv_usec = ts.tv_nsec / 1000;
16598 log.u_bbr.lt_epoch = tcp_tv_to_usec(<v);
16599 } else if (m->m_flags & M_TSTMP_LRO) {
16600 /* Record the LRO the arrival timestamp */
16601 mbuf_tstmp2timespec(m, &ts);
16602 ltv.tv_sec = ts.tv_sec;
16603 ltv.tv_usec = ts.tv_nsec / 1000;
16604 log.u_bbr.flex5 = tcp_tv_to_usec(<v);
16605 }
16606 log.u_bbr.timeStamp = tcp_get_usecs(<v);
16607 /* Log the rcv time */
16608 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16609 #ifdef TCP_REQUEST_TRK
16610 log.u_bbr.applimited = tp->t_tcpreq_closed;
16611 log.u_bbr.applimited <<= 8;
16612 log.u_bbr.applimited |= tp->t_tcpreq_open;
16613 log.u_bbr.applimited <<= 8;
16614 log.u_bbr.applimited |= tp->t_tcpreq_req;
16615 if (tcp_req) {
16616 /* Copy out any client req info */
16617 /* seconds */
16618 log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16619 /* useconds */
16620 log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16621 log.u_bbr.rttProp = tcp_req->timestamp;
16622 log.u_bbr.cur_del_rate = tcp_req->start;
16623 if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16624 log.u_bbr.flex8 |= 1;
16625 } else {
16626 log.u_bbr.flex8 |= 2;
16627 log.u_bbr.bw_inuse = tcp_req->end;
16628 }
16629 log.u_bbr.flex6 = tcp_req->start_seq;
16630 if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16631 log.u_bbr.flex8 |= 4;
16632 log.u_bbr.epoch = tcp_req->end_seq;
16633 }
16634 }
16635 #endif
16636 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16637 tlen, &log, true, <v);
16638 }
16639 /* Remove ack required flag if set, we have one */
16640 if (thflags & TH_ACK)
16641 rack->rc_ack_required = 0;
16642 rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16643 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16644 way_out = 4;
16645 retval = 0;
16646 m_freem(m);
16647 goto done_with_input;
16648 }
16649 /*
16650 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16651 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16652 */
16653 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16654 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16655 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16656 ctf_do_dropwithreset(m, tp, th, BANDLIM_TCP_RST, tlen);
16657 #ifdef TCP_ACCOUNTING
16658 sched_unpin();
16659 #endif
16660 return (1);
16661 }
16662 /*
16663 * If timestamps were negotiated during SYN/ACK and a
16664 * segment without a timestamp is received, silently drop
16665 * the segment, unless it is a RST segment or missing timestamps are
16666 * tolerated.
16667 * See section 3.2 of RFC 7323.
16668 */
16669 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16670 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16671 way_out = 5;
16672 retval = 0;
16673 m_freem(m);
16674 goto done_with_input;
16675 }
16676 /*
16677 * Segment received on connection. Reset idle time and keep-alive
16678 * timer. XXX: This should be done after segment validation to
16679 * ignore broken/spoofed segs.
16680 */
16681 if (tp->t_idle_reduce &&
16682 (tp->snd_max == tp->snd_una) &&
16683 (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16684 counter_u64_add(rack_input_idle_reduces, 1);
16685 rack_cc_after_idle(rack, tp);
16686 }
16687 tp->t_rcvtime = ticks;
16688 #ifdef STATS
16689 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16690 #endif
16691 if (tiwin > rack->r_ctl.rc_high_rwnd)
16692 rack->r_ctl.rc_high_rwnd = tiwin;
16693 /*
16694 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16695 * this to occur after we've validated the segment.
16696 */
16697 if (tcp_ecn_input_segment(tp, thflags, tlen,
16698 tcp_packets_this_ack(tp, th->th_ack),
16699 iptos))
16700 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16701
16702 /*
16703 * If echoed timestamp is later than the current time, fall back to
16704 * non RFC1323 RTT calculation. Normalize timestamp if syncookies
16705 * were used when this connection was established.
16706 */
16707 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16708 to.to_tsecr -= tp->ts_offset;
16709 if (TSTMP_GT(to.to_tsecr, ms_cts))
16710 to.to_tsecr = 0;
16711 }
16712 if ((rack->r_rcvpath_rtt_up == 1) &&
16713 (to.to_flags & TOF_TS) &&
16714 (TSTMP_GEQ(to.to_tsecr, rack->r_ctl.last_rcv_tstmp_for_rtt))) {
16715 uint32_t rtt = 0;
16716
16717 /*
16718 * We are receiving only and thus not sending
16719 * data to do an RTT. We set a flag when we first
16720 * sent this TS to the peer. We now have it back
16721 * and have an RTT to share. We log it as a conf
16722 * 4, we are not so sure about it.. since we
16723 * may have lost an ack.
16724 */
16725 if (TSTMP_GT(cts, rack->r_ctl.last_time_of_arm_rcv))
16726 rtt = (cts - rack->r_ctl.last_time_of_arm_rcv);
16727 rack->r_rcvpath_rtt_up = 0;
16728 /* Submit and commit the timer */
16729 if (rtt > 0) {
16730 tcp_rack_xmit_timer(rack, rtt, 0, rtt, 4, NULL, 1);
16731 tcp_rack_xmit_timer_commit(rack, tp);
16732 }
16733 }
16734 /*
16735 * If its the first time in we need to take care of options and
16736 * verify we can do SACK for rack!
16737 */
16738 if (rack->r_state == 0) {
16739 /* Should be init'd by rack_init() */
16740 KASSERT(rack->rc_inp != NULL,
16741 ("%s: rack->rc_inp unexpectedly NULL", __func__));
16742 if (rack->rc_inp == NULL) {
16743 rack->rc_inp = inp;
16744 }
16745
16746 /*
16747 * Process options only when we get SYN/ACK back. The SYN
16748 * case for incoming connections is handled in tcp_syncache.
16749 * According to RFC1323 the window field in a SYN (i.e., a
16750 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16751 * this is traditional behavior, may need to be cleaned up.
16752 */
16753 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16754 /* Handle parallel SYN for ECN */
16755 tcp_ecn_input_parallel_syn(tp, thflags, iptos);
16756 if ((to.to_flags & TOF_SCALE) &&
16757 (tp->t_flags & TF_REQ_SCALE)) {
16758 tp->t_flags |= TF_RCVD_SCALE;
16759 tp->snd_scale = to.to_wscale;
16760 } else
16761 tp->t_flags &= ~TF_REQ_SCALE;
16762 /*
16763 * Initial send window. It will be updated with the
16764 * next incoming segment to the scaled value.
16765 */
16766 tp->snd_wnd = th->th_win;
16767 rack_validate_fo_sendwin_up(tp, rack);
16768 if ((to.to_flags & TOF_TS) &&
16769 (tp->t_flags & TF_REQ_TSTMP)) {
16770 tp->t_flags |= TF_RCVD_TSTMP;
16771 tp->ts_recent = to.to_tsval;
16772 tp->ts_recent_age = cts;
16773 } else
16774 tp->t_flags &= ~TF_REQ_TSTMP;
16775 if (to.to_flags & TOF_MSS) {
16776 tcp_mss(tp, to.to_mss);
16777 }
16778 if ((tp->t_flags & TF_SACK_PERMIT) &&
16779 (to.to_flags & TOF_SACKPERM) == 0)
16780 tp->t_flags &= ~TF_SACK_PERMIT;
16781 if (tp->t_flags & TF_FASTOPEN) {
16782 if (to.to_flags & TOF_FASTOPEN) {
16783 uint16_t mss;
16784
16785 if (to.to_flags & TOF_MSS)
16786 mss = to.to_mss;
16787 else
16788 if ((inp->inp_vflag & INP_IPV6) != 0)
16789 mss = TCP6_MSS;
16790 else
16791 mss = TCP_MSS;
16792 tcp_fastopen_update_cache(tp, mss,
16793 to.to_tfo_len, to.to_tfo_cookie);
16794 } else
16795 tcp_fastopen_disable_path(tp);
16796 }
16797 }
16798 /*
16799 * At this point we are at the initial call. Here we decide
16800 * if we are doing RACK or not. We do this by seeing if
16801 * TF_SACK_PERMIT is set and the sack-not-required is clear.
16802 * The code now does do dup-ack counting so if you don't
16803 * switch back you won't get rack & TLP, but you will still
16804 * get this stack.
16805 */
16806
16807 if ((rack_sack_not_required == 0) &&
16808 ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
16809 tcp_switch_back_to_default(tp);
16810 (*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
16811 tlen, iptos);
16812 #ifdef TCP_ACCOUNTING
16813 sched_unpin();
16814 #endif
16815 return (1);
16816 }
16817 tcp_set_hpts(tp);
16818 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
16819 }
16820 if (thflags & TH_FIN)
16821 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
16822 us_cts = tcp_tv_to_usec(&rack->r_ctl.act_rcv_time);
16823 if ((rack->rc_gp_dyn_mul) &&
16824 (rack->use_fixed_rate == 0) &&
16825 (rack->rc_always_pace)) {
16826 /* Check in on probertt */
16827 rack_check_probe_rtt(rack, cts);
16828 }
16829 rack_clear_rate_sample(rack);
16830 if ((rack->forced_ack) &&
16831 ((tcp_get_flags(th) & TH_RST) == 0)) {
16832 rack_handle_probe_response(rack, tiwin, us_cts);
16833 }
16834 /*
16835 * This is the one exception case where we set the rack state
16836 * always. All other times (timers etc) we must have a rack-state
16837 * set (so we assure we have done the checks above for SACK).
16838 */
16839 rack->r_ctl.rc_rcvtime = cts;
16840 if (rack->r_state != tp->t_state)
16841 rack_set_state(tp, rack);
16842 if (SEQ_GT(th->th_ack, tp->snd_una) &&
16843 (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
16844 kern_prefetch(rsm, &prev_state);
16845 prev_state = rack->r_state;
16846 if ((thflags & TH_RST) &&
16847 ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
16848 SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
16849 (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
16850 /* The connection will be killed by a reset check the tracepoint */
16851 tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
16852 }
16853 retval = (*rack->r_substate) (m, th, so,
16854 tp, &to, drop_hdrlen,
16855 tlen, tiwin, thflags, nxt_pkt, iptos);
16856 if (retval == 0) {
16857 /*
16858 * If retval is 1 the tcb is unlocked and most likely the tp
16859 * is gone.
16860 */
16861 INP_WLOCK_ASSERT(inp);
16862 if ((rack->rc_gp_dyn_mul) &&
16863 (rack->rc_always_pace) &&
16864 (rack->use_fixed_rate == 0) &&
16865 rack->in_probe_rtt &&
16866 (rack->r_ctl.rc_time_probertt_starts == 0)) {
16867 /*
16868 * If we are going for target, lets recheck before
16869 * we output.
16870 */
16871 rack_check_probe_rtt(rack, cts);
16872 }
16873 if (rack->set_pacing_done_a_iw == 0) {
16874 /* How much has been acked? */
16875 if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
16876 /* We have enough to set in the pacing segment size */
16877 rack->set_pacing_done_a_iw = 1;
16878 rack_set_pace_segments(tp, rack, __LINE__, NULL);
16879 }
16880 }
16881 tcp_rack_xmit_timer_commit(rack, tp);
16882 #ifdef TCP_ACCOUNTING
16883 /*
16884 * If we set the ack_val_se to what ack processing we are doing
16885 * we also want to track how many cycles we burned. Note
16886 * the bits after tcp_output we let be "free". This is because
16887 * we are also tracking the tcp_output times as well. Note the
16888 * use of 0xf here since we only have 11 counter (0 - 0xa) and
16889 * 0xf cannot be returned and is what we initialize it too to
16890 * indicate we are not doing the tabulations.
16891 */
16892 if (ack_val_set != 0xf) {
16893 uint64_t crtsc;
16894
16895 crtsc = get_cyclecount();
16896 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16897 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
16898 }
16899 }
16900 #endif
16901 if ((nxt_pkt == 0) && (no_output == 0)) {
16902 if ((rack->r_wanted_output != 0) ||
16903 (tp->t_flags & TF_ACKNOW) ||
16904 (rack->r_fast_output != 0)) {
16905
16906 do_output_now:
16907 if (tcp_output(tp) < 0) {
16908 #ifdef TCP_ACCOUNTING
16909 sched_unpin();
16910 #endif
16911 return (1);
16912 }
16913 did_out = 1;
16914 }
16915 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16916 rack_free_trim(rack);
16917 } else if ((nxt_pkt == 0) && (tp->t_flags & TF_ACKNOW)) {
16918 goto do_output_now;
16919 } else if ((no_output == 1) &&
16920 (nxt_pkt == 0) &&
16921 (tcp_in_hpts(rack->rc_tp) == 0)) {
16922 /*
16923 * We are not in hpts and we had a pacing timer up. Use
16924 * the remaining time (slot_remaining) to restart the timer.
16925 */
16926 KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
16927 rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
16928 rack_free_trim(rack);
16929 }
16930 /* Clear the flag, it may have been cleared by output but we may not have */
16931 if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
16932 tp->t_flags2 &= ~TF2_HPTS_CALLS;
16933 /*
16934 * The draft (v3) calls for us to use SEQ_GEQ, but that
16935 * causes issues when we are just going app limited. Lets
16936 * instead use SEQ_GT <or> where its equal but more data
16937 * is outstanding.
16938 *
16939 * Also make sure we are on the last ack of a series. We
16940 * have to have all the ack's processed in queue to know
16941 * if there is something left outstanding.
16942 */
16943 if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
16944 (rack->rc_new_rnd_needed == 0) &&
16945 (nxt_pkt == 0)) {
16946 /*
16947 * We have crossed into a new round with
16948 * the new snd_unae.
16949 */
16950 rack_new_round_setup(tp, rack, tp->snd_una);
16951 }
16952 if ((nxt_pkt == 0) &&
16953 ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
16954 (SEQ_GT(tp->snd_max, tp->snd_una) ||
16955 (tp->t_flags & TF_DELACK) ||
16956 ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
16957 (tp->t_state <= TCPS_CLOSING)))) {
16958 /* We could not send (probably in the hpts but stopped the timer earlier)? */
16959 if ((tp->snd_max == tp->snd_una) &&
16960 ((tp->t_flags & TF_DELACK) == 0) &&
16961 (tcp_in_hpts(rack->rc_tp)) &&
16962 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
16963 /* keep alive not needed if we are hptsi output yet */
16964 ;
16965 } else {
16966 int late = 0;
16967 if (tcp_in_hpts(tp)) {
16968 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
16969 us_cts = tcp_get_usecs(NULL);
16970 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
16971 rack->r_early = 1;
16972 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
16973 } else
16974 late = 1;
16975 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16976 }
16977 tcp_hpts_remove(tp);
16978 }
16979 if (late && (did_out == 0)) {
16980 /*
16981 * We are late in the sending
16982 * and we did not call the output
16983 * (this probably should not happen).
16984 */
16985 goto do_output_now;
16986 }
16987 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
16988 }
16989 way_out = 1;
16990 } else if (nxt_pkt == 0) {
16991 /* Do we have the correct timer running? */
16992 rack_timer_audit(tp, rack, &so->so_snd);
16993 way_out = 2;
16994 }
16995 done_with_input:
16996 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
16997 if (did_out)
16998 rack->r_wanted_output = 0;
16999 }
17000
17001 #ifdef TCP_ACCOUNTING
17002 sched_unpin();
17003 #endif
17004 return (retval);
17005 }
17006
17007 static void
rack_do_segment(struct tcpcb * tp,struct mbuf * m,struct tcphdr * th,int32_t drop_hdrlen,int32_t tlen,uint8_t iptos)17008 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17009 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17010 {
17011 struct timeval tv;
17012
17013 /* First lets see if we have old packets */
17014 if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17015 if (ctf_do_queued_segments(tp, 1)) {
17016 m_freem(m);
17017 return;
17018 }
17019 }
17020 if (m->m_flags & M_TSTMP_LRO) {
17021 mbuf_tstmp2timeval(m, &tv);
17022 } else {
17023 /* Should not be should we kassert instead? */
17024 tcp_get_usecs(&tv);
17025 }
17026 if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17027 &tv) == 0) {
17028 INP_WUNLOCK(tptoinpcb(tp));
17029 }
17030 }
17031
17032 struct rack_sendmap *
tcp_rack_output(struct tcpcb * tp,struct tcp_rack * rack,uint32_t tsused)17033 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17034 {
17035 struct rack_sendmap *rsm = NULL;
17036 int32_t idx;
17037 uint32_t srtt = 0, thresh = 0, ts_low = 0;
17038
17039 /* Return the next guy to be re-transmitted */
17040 if (tqhash_empty(rack->r_ctl.tqh)) {
17041 return (NULL);
17042 }
17043 if (tp->t_flags & TF_SENTFIN) {
17044 /* retran the end FIN? */
17045 return (NULL);
17046 }
17047 /* ok lets look at this one */
17048 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17049 if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17050 return (rsm);
17051 }
17052 if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17053 goto check_it;
17054 }
17055 rsm = rack_find_lowest_rsm(rack);
17056 if (rsm == NULL) {
17057 return (NULL);
17058 }
17059 check_it:
17060 if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
17061 (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17062 /*
17063 * No sack so we automatically do the 3 strikes and
17064 * retransmit (no rack timer would be started).
17065 */
17066 return (rsm);
17067 }
17068 if (rsm->r_flags & RACK_ACKED) {
17069 return (NULL);
17070 }
17071 if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17072 (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17073 /* Its not yet ready */
17074 return (NULL);
17075 }
17076 srtt = rack_grab_rtt(tp, rack);
17077 idx = rsm->r_rtr_cnt - 1;
17078 ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17079 thresh = rack_calc_thresh_rack(rack, srtt, tsused, __LINE__, 1);
17080 if ((tsused == ts_low) ||
17081 (TSTMP_LT(tsused, ts_low))) {
17082 /* No time since sending */
17083 return (NULL);
17084 }
17085 if ((tsused - ts_low) < thresh) {
17086 /* It has not been long enough yet */
17087 return (NULL);
17088 }
17089 if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17090 ((rsm->r_flags & RACK_SACK_PASSED))) {
17091 /*
17092 * We have passed the dup-ack threshold <or>
17093 * a SACK has indicated this is missing.
17094 * Note that if you are a declared attacker
17095 * it is only the dup-ack threshold that
17096 * will cause retransmits.
17097 */
17098 /* log retransmit reason */
17099 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17100 rack->r_fast_output = 0;
17101 return (rsm);
17102 }
17103 return (NULL);
17104 }
17105
17106 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)17107 rack_log_pacing_delay_calc (struct tcp_rack *rack, uint32_t len, uint32_t slot,
17108 uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17109 int line, struct rack_sendmap *rsm, uint8_t quality)
17110 {
17111 if (tcp_bblogging_on(rack->rc_tp)) {
17112 union tcp_log_stackspecific log;
17113 struct timeval tv;
17114
17115 if (rack_verbose_logging == 0) {
17116 /*
17117 * We are not verbose screen out all but
17118 * ones we always want.
17119 */
17120 if ((method != 2) &&
17121 (method != 3) &&
17122 (method != 7) &&
17123 (method != 89) &&
17124 (method != 14) &&
17125 (method != 20)) {
17126 return;
17127 }
17128 }
17129 memset(&log, 0, sizeof(log));
17130 log.u_bbr.flex1 = slot;
17131 log.u_bbr.flex2 = len;
17132 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17133 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17134 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17135 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17136 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17137 log.u_bbr.use_lt_bw <<= 1;
17138 log.u_bbr.use_lt_bw |= rack->r_late;
17139 log.u_bbr.use_lt_bw <<= 1;
17140 log.u_bbr.use_lt_bw |= rack->r_early;
17141 log.u_bbr.use_lt_bw <<= 1;
17142 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17143 log.u_bbr.use_lt_bw <<= 1;
17144 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17145 log.u_bbr.use_lt_bw <<= 1;
17146 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17147 log.u_bbr.use_lt_bw <<= 1;
17148 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17149 log.u_bbr.use_lt_bw <<= 1;
17150 log.u_bbr.use_lt_bw |= rack->gp_ready;
17151 log.u_bbr.pkt_epoch = line;
17152 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17153 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17154 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17155 log.u_bbr.bw_inuse = bw_est;
17156 log.u_bbr.delRate = bw;
17157 if (rack->r_ctl.gp_bw == 0)
17158 log.u_bbr.cur_del_rate = 0;
17159 else
17160 log.u_bbr.cur_del_rate = rack_get_bw(rack);
17161 log.u_bbr.rttProp = len_time;
17162 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17163 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17164 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17165 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17166 /* We are in slow start */
17167 log.u_bbr.flex7 = 1;
17168 } else {
17169 /* we are on congestion avoidance */
17170 log.u_bbr.flex7 = 0;
17171 }
17172 log.u_bbr.flex8 = method;
17173 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17174 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17175 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17176 log.u_bbr.cwnd_gain <<= 1;
17177 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17178 log.u_bbr.cwnd_gain <<= 1;
17179 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17180 log.u_bbr.cwnd_gain <<= 1;
17181 log.u_bbr.cwnd_gain |= rack->use_fixed_rate;
17182 log.u_bbr.cwnd_gain <<= 1;
17183 log.u_bbr.cwnd_gain |= rack->rc_always_pace;
17184 log.u_bbr.cwnd_gain <<= 1;
17185 log.u_bbr.cwnd_gain |= rack->gp_ready;
17186 log.u_bbr.bbr_substate = quality;
17187 log.u_bbr.bbr_state = rack->dgp_on;
17188 log.u_bbr.bbr_state <<= 1;
17189 log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17190 log.u_bbr.bbr_state <<= 2;
17191 TCP_LOG_EVENTP(rack->rc_tp, NULL,
17192 &rack->rc_inp->inp_socket->so_rcv,
17193 &rack->rc_inp->inp_socket->so_snd,
17194 BBR_LOG_HPTSI_CALC, 0,
17195 0, &log, false, &tv);
17196 }
17197 }
17198
17199 static uint32_t
rack_get_pacing_len(struct tcp_rack * rack,uint64_t bw,uint32_t mss)17200 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17201 {
17202 uint32_t new_tso, user_max, pace_one;
17203
17204 user_max = rack->rc_user_set_max_segs * mss;
17205 if (rack->rc_force_max_seg) {
17206 return (user_max);
17207 }
17208 if (rack->use_fixed_rate &&
17209 ((rack->r_ctl.crte == NULL) ||
17210 (bw != rack->r_ctl.crte->rate))) {
17211 /* Use the user mss since we are not exactly matched */
17212 return (user_max);
17213 }
17214 if (rack_pace_one_seg ||
17215 (rack->r_ctl.rc_user_set_min_segs == 1))
17216 pace_one = 1;
17217 else
17218 pace_one = 0;
17219
17220 new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17221 pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17222 if (new_tso > user_max)
17223 new_tso = user_max;
17224 if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17225 if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17226 new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17227 }
17228 if (rack->r_ctl.rc_user_set_min_segs &&
17229 ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17230 new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17231 return (new_tso);
17232 }
17233
17234 static uint64_t
rack_arrive_at_discounted_rate(struct tcp_rack * rack,uint64_t window_input,uint32_t * rate_set,uint32_t * gain_b)17235 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17236 {
17237 uint64_t reduced_win;
17238 uint32_t gain;
17239
17240 if (window_input < rc_init_window(rack)) {
17241 /*
17242 * The cwnd is collapsed to
17243 * nearly zero, maybe because of a time-out?
17244 * Lets drop back to the lt-bw.
17245 */
17246 reduced_win = rack_get_lt_bw(rack);
17247 /* Set the flag so the caller knows its a rate and not a reduced window */
17248 *rate_set = 1;
17249 gain = 100;
17250 } else if (IN_RECOVERY(rack->rc_tp->t_flags)) {
17251 /*
17252 * If we are in recover our cwnd needs to be less for
17253 * our pacing consideration.
17254 */
17255 if (rack->rack_hibeta == 0) {
17256 reduced_win = window_input / 2;
17257 gain = 50;
17258 } else {
17259 reduced_win = window_input * rack->r_ctl.saved_hibeta;
17260 reduced_win /= 100;
17261 gain = rack->r_ctl.saved_hibeta;
17262 }
17263 } else {
17264 /*
17265 * Apply Timely factor to increase/decrease the
17266 * amount we are pacing at.
17267 */
17268 gain = rack_get_output_gain(rack, NULL);
17269 if (gain > rack_gain_p5_ub) {
17270 gain = rack_gain_p5_ub;
17271 }
17272 reduced_win = window_input * gain;
17273 reduced_win /= 100;
17274 }
17275 if (gain_b != NULL)
17276 *gain_b = gain;
17277 /*
17278 * What is being returned here is a trimmed down
17279 * window values in all cases where rate_set is left
17280 * at 0. In one case we actually return the rate (lt_bw).
17281 * the "reduced_win" is returned as a slimmed down cwnd that
17282 * is then calculated by the caller into a rate when rate_set
17283 * is 0.
17284 */
17285 return (reduced_win);
17286 }
17287
17288 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)17289 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)
17290 {
17291 uint64_t lentim, fill_bw;
17292
17293 rack->r_via_fill_cw = 0;
17294 if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17295 return (slot);
17296 if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17297 return (slot);
17298 if (rack->r_ctl.rc_last_us_rtt == 0)
17299 return (slot);
17300 if (rack->rc_pace_fill_if_rttin_range &&
17301 (rack->r_ctl.rc_last_us_rtt >=
17302 (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17303 /* The rtt is huge, N * smallest, lets not fill */
17304 return (slot);
17305 }
17306 if (rack->r_ctl.fillcw_cap && *rate_wanted >= rack->r_ctl.fillcw_cap)
17307 return (slot);
17308 /*
17309 * first lets calculate the b/w based on the last us-rtt
17310 * and the the smallest send window.
17311 */
17312 fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17313 if (rack->rc_fillcw_apply_discount) {
17314 uint32_t rate_set = 0;
17315
17316 fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17317 if (rate_set) {
17318 goto at_lt_bw;
17319 }
17320 }
17321 /* Take the rwnd if its smaller */
17322 if (fill_bw > rack->rc_tp->snd_wnd)
17323 fill_bw = rack->rc_tp->snd_wnd;
17324 /* Now lets make it into a b/w */
17325 fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17326 fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17327 /* Adjust to any cap */
17328 if (rack->r_ctl.fillcw_cap && fill_bw >= rack->r_ctl.fillcw_cap)
17329 fill_bw = rack->r_ctl.fillcw_cap;
17330
17331 at_lt_bw:
17332 if (rack_bw_multipler > 0) {
17333 /*
17334 * We want to limit fill-cw to the some multiplier
17335 * of the max(lt_bw, gp_est). The normal default
17336 * is 0 for off, so a sysctl has enabled it.
17337 */
17338 uint64_t lt_bw, gp, rate;
17339
17340 gp = rack_get_gp_est(rack);
17341 lt_bw = rack_get_lt_bw(rack);
17342 if (lt_bw > gp)
17343 rate = lt_bw;
17344 else
17345 rate = gp;
17346 rate *= rack_bw_multipler;
17347 rate /= 100;
17348 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
17349 union tcp_log_stackspecific log;
17350 struct timeval tv;
17351
17352 memset(&log, 0, sizeof(log));
17353 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17354 log.u_bbr.flex1 = rack_bw_multipler;
17355 log.u_bbr.flex2 = len;
17356 log.u_bbr.cur_del_rate = gp;
17357 log.u_bbr.delRate = lt_bw;
17358 log.u_bbr.bw_inuse = rate;
17359 log.u_bbr.rttProp = fill_bw;
17360 log.u_bbr.flex8 = 44;
17361 tcp_log_event(rack->rc_tp, NULL, NULL, NULL,
17362 BBR_LOG_CWND, 0,
17363 0, &log, false, NULL,
17364 __func__, __LINE__, &tv);
17365 }
17366 if (fill_bw > rate)
17367 fill_bw = rate;
17368 }
17369 /* We are below the min b/w */
17370 if (non_paced)
17371 *rate_wanted = fill_bw;
17372 if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17373 return (slot);
17374 rack->r_via_fill_cw = 1;
17375 if (rack->r_rack_hw_rate_caps &&
17376 (rack->r_ctl.crte != NULL)) {
17377 uint64_t high_rate;
17378
17379 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17380 if (fill_bw > high_rate) {
17381 /* We are capping bw at the highest rate table entry */
17382 if (*rate_wanted > high_rate) {
17383 /* The original rate was also capped */
17384 rack->r_via_fill_cw = 0;
17385 }
17386 rack_log_hdwr_pacing(rack,
17387 fill_bw, high_rate, __LINE__,
17388 0, 3);
17389 fill_bw = high_rate;
17390 if (capped)
17391 *capped = 1;
17392 }
17393 } else if ((rack->r_ctl.crte == NULL) &&
17394 (rack->rack_hdrw_pacing == 0) &&
17395 (rack->rack_hdw_pace_ena) &&
17396 rack->r_rack_hw_rate_caps &&
17397 (rack->rack_attempt_hdwr_pace == 0) &&
17398 (rack->rc_inp->inp_route.ro_nh != NULL) &&
17399 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17400 /*
17401 * Ok we may have a first attempt that is greater than our top rate
17402 * lets check.
17403 */
17404 uint64_t high_rate;
17405
17406 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17407 if (high_rate) {
17408 if (fill_bw > high_rate) {
17409 fill_bw = high_rate;
17410 if (capped)
17411 *capped = 1;
17412 }
17413 }
17414 }
17415 if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17416 rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17417 fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17418 fill_bw = rack->r_ctl.bw_rate_cap;
17419 }
17420 /*
17421 * Ok fill_bw holds our mythical b/w to fill the cwnd
17422 * in an rtt (unless it was capped), what does that
17423 * time wise equate too?
17424 */
17425 lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17426 lentim /= fill_bw;
17427 *rate_wanted = fill_bw;
17428 if (non_paced || (lentim < slot)) {
17429 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17430 0, lentim, 12, __LINE__, NULL, 0);
17431 return ((int32_t)lentim);
17432 } else
17433 return (slot);
17434 }
17435
17436 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)17437 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz, int line)
17438 {
17439 uint64_t srtt;
17440 int32_t slot = 0;
17441 int can_start_hw_pacing = 1;
17442 int err;
17443 int pace_one;
17444
17445 if (rack_pace_one_seg ||
17446 (rack->r_ctl.rc_user_set_min_segs == 1))
17447 pace_one = 1;
17448 else
17449 pace_one = 0;
17450 if (rack->rc_always_pace == 0) {
17451 /*
17452 * We use the most optimistic possible cwnd/srtt for
17453 * sending calculations. This will make our
17454 * calculation anticipate getting more through
17455 * quicker then possible. But thats ok we don't want
17456 * the peer to have a gap in data sending.
17457 */
17458 uint64_t cwnd, tr_perms = 0;
17459 int32_t reduce;
17460
17461 old_method:
17462 /*
17463 * We keep no precise pacing with the old method
17464 * instead we use the pacer to mitigate bursts.
17465 */
17466 if (rack->r_ctl.rc_rack_min_rtt)
17467 srtt = rack->r_ctl.rc_rack_min_rtt;
17468 else
17469 srtt = max(tp->t_srtt, 1);
17470 if (rack->r_ctl.rc_rack_largest_cwnd)
17471 cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17472 else
17473 cwnd = rack->r_ctl.cwnd_to_use;
17474 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
17475 tr_perms = (cwnd * 1000) / srtt;
17476 if (tr_perms == 0) {
17477 tr_perms = ctf_fixed_maxseg(tp);
17478 }
17479 /*
17480 * Calculate how long this will take to drain, if
17481 * the calculation comes out to zero, thats ok we
17482 * will use send_a_lot to possibly spin around for
17483 * more increasing tot_len_this_send to the point
17484 * that its going to require a pace, or we hit the
17485 * cwnd. Which in that case we are just waiting for
17486 * a ACK.
17487 */
17488 slot = len / tr_perms;
17489 /* Now do we reduce the time so we don't run dry? */
17490 if (slot && rack_slot_reduction) {
17491 reduce = (slot / rack_slot_reduction);
17492 if (reduce < slot) {
17493 slot -= reduce;
17494 } else
17495 slot = 0;
17496 } else
17497 reduce = 0;
17498 slot *= HPTS_USEC_IN_MSEC;
17499 if (rack->rc_pace_to_cwnd) {
17500 uint64_t rate_wanted = 0;
17501
17502 slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17503 rack->rc_ack_can_sendout_data = 1;
17504 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17505 } else
17506 rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17507 /*******************************************************/
17508 /* RRS: We insert non-paced call to stats here for len */
17509 /*******************************************************/
17510 } else {
17511 uint64_t bw_est, res, lentim, rate_wanted;
17512 uint32_t segs, oh;
17513 int capped = 0;
17514 int prev_fill;
17515
17516 if ((rack->r_rr_config == 1) && rsm) {
17517 return (rack->r_ctl.rc_min_to);
17518 }
17519 if (rack->use_fixed_rate) {
17520 rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17521 } else if ((rack->r_ctl.init_rate == 0) &&
17522 (rack->r_ctl.gp_bw == 0)) {
17523 /* no way to yet do an estimate */
17524 bw_est = rate_wanted = 0;
17525 } else if (rack->dgp_on) {
17526 bw_est = rack_get_bw(rack);
17527 rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17528 } else {
17529 uint32_t gain, rate_set = 0;
17530
17531 rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17532 rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17533 if (rate_set == 0) {
17534 if (rate_wanted > rack->rc_tp->snd_wnd)
17535 rate_wanted = rack->rc_tp->snd_wnd;
17536 /* Now lets make it into a b/w */
17537 rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17538 rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17539 }
17540 bw_est = rate_wanted;
17541 rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17542 rack->r_ctl.cwnd_to_use,
17543 rate_wanted, bw_est,
17544 rack->r_ctl.rc_last_us_rtt,
17545 88, __LINE__, NULL, gain);
17546 }
17547 if (((bw_est == 0) || (rate_wanted == 0) || (rack->gp_ready == 0)) &&
17548 (rack->use_fixed_rate == 0)) {
17549 /*
17550 * No way yet to make a b/w estimate or
17551 * our raise is set incorrectly.
17552 */
17553 goto old_method;
17554 }
17555 rack_rate_cap_bw(rack, &rate_wanted, &capped);
17556 /* We need to account for all the overheads */
17557 segs = (len + segsiz - 1) / segsiz;
17558 /*
17559 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17560 * and how much data we put in each packet. Yes this
17561 * means we may be off if we are larger than 1500 bytes
17562 * or smaller. But this just makes us more conservative.
17563 */
17564
17565 oh = (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17566 if (rack->r_is_v6) {
17567 #ifdef INET6
17568 oh += sizeof(struct ip6_hdr);
17569 #endif
17570 } else {
17571 #ifdef INET
17572 oh += sizeof(struct ip);
17573 #endif
17574 }
17575 /* We add a fixed 14 for the ethernet header */
17576 oh += 14;
17577 segs *= oh;
17578 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17579 res = lentim / rate_wanted;
17580 slot = (uint32_t)res;
17581 if (rack_hw_rate_min &&
17582 (rate_wanted < rack_hw_rate_min)) {
17583 can_start_hw_pacing = 0;
17584 if (rack->r_ctl.crte) {
17585 /*
17586 * Ok we need to release it, we
17587 * have fallen too low.
17588 */
17589 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17590 rack->r_ctl.crte = NULL;
17591 rack->rack_attempt_hdwr_pace = 0;
17592 rack->rack_hdrw_pacing = 0;
17593 }
17594 }
17595 if (rack->r_ctl.crte &&
17596 (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17597 /*
17598 * We want more than the hardware can give us,
17599 * don't start any hw pacing.
17600 */
17601 can_start_hw_pacing = 0;
17602 if (rack->r_rack_hw_rate_caps == 0) {
17603 /*
17604 * Ok we need to release it, we
17605 * want more than the card can give us and
17606 * no rate cap is in place. Set it up so
17607 * when we want less we can retry.
17608 */
17609 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17610 rack->r_ctl.crte = NULL;
17611 rack->rack_attempt_hdwr_pace = 0;
17612 rack->rack_hdrw_pacing = 0;
17613 }
17614 }
17615 if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17616 /*
17617 * We lost our rate somehow, this can happen
17618 * if the interface changed underneath us.
17619 */
17620 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17621 rack->r_ctl.crte = NULL;
17622 /* Lets re-allow attempting to setup pacing */
17623 rack->rack_hdrw_pacing = 0;
17624 rack->rack_attempt_hdwr_pace = 0;
17625 rack_log_hdwr_pacing(rack,
17626 rate_wanted, bw_est, __LINE__,
17627 0, 6);
17628 }
17629 prev_fill = rack->r_via_fill_cw;
17630 if ((rack->rc_pace_to_cwnd) &&
17631 (capped == 0) &&
17632 (rack->dgp_on == 1) &&
17633 (rack->use_fixed_rate == 0) &&
17634 (rack->in_probe_rtt == 0) &&
17635 (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17636 /*
17637 * We want to pace at our rate *or* faster to
17638 * fill the cwnd to the max if its not full.
17639 */
17640 slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17641 /* Re-check to make sure we are not exceeding our max b/w */
17642 if ((rack->r_ctl.crte != NULL) &&
17643 (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17644 /*
17645 * We want more than the hardware can give us,
17646 * don't start any hw pacing.
17647 */
17648 can_start_hw_pacing = 0;
17649 if (rack->r_rack_hw_rate_caps == 0) {
17650 /*
17651 * Ok we need to release it, we
17652 * want more than the card can give us and
17653 * no rate cap is in place. Set it up so
17654 * when we want less we can retry.
17655 */
17656 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17657 rack->r_ctl.crte = NULL;
17658 rack->rack_attempt_hdwr_pace = 0;
17659 rack->rack_hdrw_pacing = 0;
17660 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17661 }
17662 }
17663 }
17664 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17665 (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17666 if ((rack->rack_hdw_pace_ena) &&
17667 (can_start_hw_pacing > 0) &&
17668 (rack->rack_hdrw_pacing == 0) &&
17669 (rack->rack_attempt_hdwr_pace == 0)) {
17670 /*
17671 * Lets attempt to turn on hardware pacing
17672 * if we can.
17673 */
17674 rack->rack_attempt_hdwr_pace = 1;
17675 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17676 rack->rc_inp->inp_route.ro_nh->nh_ifp,
17677 rate_wanted,
17678 RS_PACING_GEQ,
17679 &err, &rack->r_ctl.crte_prev_rate);
17680 if (rack->r_ctl.crte) {
17681 rack->rack_hdrw_pacing = 1;
17682 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17683 pace_one, rack->r_ctl.crte,
17684 NULL, rack->r_ctl.pace_len_divisor);
17685 rack_log_hdwr_pacing(rack,
17686 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17687 err, 0);
17688 rack->r_ctl.last_hw_bw_req = rate_wanted;
17689 } else {
17690 counter_u64_add(rack_hw_pace_init_fail, 1);
17691 }
17692 } else if (rack->rack_hdrw_pacing &&
17693 (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17694 /* Do we need to adjust our rate? */
17695 const struct tcp_hwrate_limit_table *nrte;
17696
17697 if (rack->r_up_only &&
17698 (rate_wanted < rack->r_ctl.crte->rate)) {
17699 /**
17700 * We have four possible states here
17701 * having to do with the previous time
17702 * and this time.
17703 * previous | this-time
17704 * A) 0 | 0 -- fill_cw not in the picture
17705 * B) 1 | 0 -- we were doing a fill-cw but now are not
17706 * C) 1 | 1 -- all rates from fill_cw
17707 * D) 0 | 1 -- we were doing non-fill and now we are filling
17708 *
17709 * For case A, C and D we don't allow a drop. But for
17710 * case B where we now our on our steady rate we do
17711 * allow a drop.
17712 *
17713 */
17714 if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17715 goto done_w_hdwr;
17716 }
17717 if ((rate_wanted > rack->r_ctl.crte->rate) ||
17718 (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17719 if (rack_hw_rate_to_low &&
17720 (bw_est < rack_hw_rate_to_low)) {
17721 /*
17722 * The pacing rate is too low for hardware, but
17723 * do allow hardware pacing to be restarted.
17724 */
17725 rack_log_hdwr_pacing(rack,
17726 bw_est, rack->r_ctl.crte->rate, __LINE__,
17727 0, 5);
17728 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17729 rack->r_ctl.crte = NULL;
17730 rack->rack_attempt_hdwr_pace = 0;
17731 rack->rack_hdrw_pacing = 0;
17732 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17733 goto done_w_hdwr;
17734 }
17735 nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17736 rack->rc_tp,
17737 rack->rc_inp->inp_route.ro_nh->nh_ifp,
17738 rate_wanted,
17739 RS_PACING_GEQ,
17740 &err, &rack->r_ctl.crte_prev_rate);
17741 if (nrte == NULL) {
17742 /*
17743 * Lost the rate, lets drop hardware pacing
17744 * period.
17745 */
17746 rack->rack_hdrw_pacing = 0;
17747 rack->r_ctl.crte = NULL;
17748 rack_log_hdwr_pacing(rack,
17749 rate_wanted, 0, __LINE__,
17750 err, 1);
17751 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17752 counter_u64_add(rack_hw_pace_lost, 1);
17753 } else if (nrte != rack->r_ctl.crte) {
17754 rack->r_ctl.crte = nrte;
17755 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17756 segsiz, pace_one, rack->r_ctl.crte,
17757 NULL, rack->r_ctl.pace_len_divisor);
17758 rack_log_hdwr_pacing(rack,
17759 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17760 err, 2);
17761 rack->r_ctl.last_hw_bw_req = rate_wanted;
17762 }
17763 } else {
17764 /* We just need to adjust the segment size */
17765 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17766 rack_log_hdwr_pacing(rack,
17767 rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17768 0, 4);
17769 rack->r_ctl.last_hw_bw_req = rate_wanted;
17770 }
17771 }
17772 }
17773 done_w_hdwr:
17774 if (rack_limit_time_with_srtt &&
17775 (rack->use_fixed_rate == 0) &&
17776 (rack->rack_hdrw_pacing == 0)) {
17777 /*
17778 * Sanity check, we do not allow the pacing delay
17779 * to be longer than the SRTT of the path. If it is
17780 * a slow path, then adding a packet should increase
17781 * the RTT and compensate for this i.e. the srtt will
17782 * be greater so the allowed pacing time will be greater.
17783 *
17784 * Note this restriction is not for where a peak rate
17785 * is set, we are doing fixed pacing or hardware pacing.
17786 */
17787 if (rack->rc_tp->t_srtt)
17788 srtt = rack->rc_tp->t_srtt;
17789 else
17790 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */
17791 if (srtt < (uint64_t)slot) {
17792 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
17793 slot = srtt;
17794 }
17795 }
17796 /*******************************************************************/
17797 /* RRS: We insert paced call to stats here for len and rate_wanted */
17798 /*******************************************************************/
17799 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
17800 }
17801 if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
17802 /*
17803 * If this rate is seeing enobufs when it
17804 * goes to send then either the nic is out
17805 * of gas or we are mis-estimating the time
17806 * somehow and not letting the queue empty
17807 * completely. Lets add to the pacing time.
17808 */
17809 int hw_boost_delay;
17810
17811 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
17812 if (hw_boost_delay > rack_enobuf_hw_max)
17813 hw_boost_delay = rack_enobuf_hw_max;
17814 else if (hw_boost_delay < rack_enobuf_hw_min)
17815 hw_boost_delay = rack_enobuf_hw_min;
17816 slot += hw_boost_delay;
17817 }
17818 return (slot);
17819 }
17820
17821 static void
rack_start_gp_measurement(struct tcpcb * tp,struct tcp_rack * rack,tcp_seq startseq,uint32_t sb_offset)17822 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
17823 tcp_seq startseq, uint32_t sb_offset)
17824 {
17825 struct rack_sendmap *my_rsm = NULL;
17826
17827 if (tp->t_state < TCPS_ESTABLISHED) {
17828 /*
17829 * We don't start any measurements if we are
17830 * not at least established.
17831 */
17832 return;
17833 }
17834 if (tp->t_state >= TCPS_FIN_WAIT_1) {
17835 /*
17836 * We will get no more data into the SB
17837 * this means we need to have the data available
17838 * before we start a measurement.
17839 */
17840
17841 if (sbavail(&tptosocket(tp)->so_snd) <
17842 max(rc_init_window(rack),
17843 (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
17844 /* Nope not enough data */
17845 return;
17846 }
17847 }
17848 tp->t_flags |= TF_GPUTINPROG;
17849 rack->r_ctl.rc_gp_cumack_ts = 0;
17850 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
17851 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
17852 tp->gput_seq = startseq;
17853 rack->app_limited_needs_set = 0;
17854 if (rack->in_probe_rtt)
17855 rack->measure_saw_probe_rtt = 1;
17856 else if ((rack->measure_saw_probe_rtt) &&
17857 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
17858 rack->measure_saw_probe_rtt = 0;
17859 if (rack->rc_gp_filled)
17860 tp->gput_ts = rack->r_ctl.last_cumack_advance;
17861 else {
17862 /* Special case initial measurement */
17863 struct timeval tv;
17864
17865 tp->gput_ts = tcp_get_usecs(&tv);
17866 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
17867 }
17868 /*
17869 * We take a guess out into the future,
17870 * if we have no measurement and no
17871 * initial rate, we measure the first
17872 * initial-windows worth of data to
17873 * speed up getting some GP measurement and
17874 * thus start pacing.
17875 */
17876 if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
17877 rack->app_limited_needs_set = 1;
17878 tp->gput_ack = startseq + max(rc_init_window(rack),
17879 (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
17880 rack_log_pacing_delay_calc(rack,
17881 tp->gput_seq,
17882 tp->gput_ack,
17883 0,
17884 tp->gput_ts,
17885 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
17886 9,
17887 __LINE__, NULL, 0);
17888 rack_tend_gp_marks(tp, rack);
17889 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
17890 return;
17891 }
17892 if (sb_offset) {
17893 /*
17894 * We are out somewhere in the sb
17895 * can we use the already outstanding data?
17896 */
17897
17898 if (rack->r_ctl.rc_app_limited_cnt == 0) {
17899 /*
17900 * Yes first one is good and in this case
17901 * the tp->gput_ts is correctly set based on
17902 * the last ack that arrived (no need to
17903 * set things up when an ack comes in).
17904 */
17905 my_rsm = tqhash_min(rack->r_ctl.tqh);
17906 if ((my_rsm == NULL) ||
17907 (my_rsm->r_rtr_cnt != 1)) {
17908 /* retransmission? */
17909 goto use_latest;
17910 }
17911 } else {
17912 if (rack->r_ctl.rc_first_appl == NULL) {
17913 /*
17914 * If rc_first_appl is NULL
17915 * then the cnt should be 0.
17916 * This is probably an error, maybe
17917 * a KASSERT would be approprate.
17918 */
17919 goto use_latest;
17920 }
17921 /*
17922 * If we have a marker pointer to the last one that is
17923 * app limited we can use that, but we need to set
17924 * things up so that when it gets ack'ed we record
17925 * the ack time (if its not already acked).
17926 */
17927 rack->app_limited_needs_set = 1;
17928 /*
17929 * We want to get to the rsm that is either
17930 * next with space i.e. over 1 MSS or the one
17931 * after that (after the app-limited).
17932 */
17933 my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
17934 if (my_rsm) {
17935 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
17936 /* Have to use the next one */
17937 my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17938 else {
17939 /* Use after the first MSS of it is acked */
17940 tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
17941 goto start_set;
17942 }
17943 }
17944 if ((my_rsm == NULL) ||
17945 (my_rsm->r_rtr_cnt != 1)) {
17946 /*
17947 * Either its a retransmit or
17948 * the last is the app-limited one.
17949 */
17950 goto use_latest;
17951 }
17952 }
17953 tp->gput_seq = my_rsm->r_start;
17954 start_set:
17955 if (my_rsm->r_flags & RACK_ACKED) {
17956 /*
17957 * This one has been acked use the arrival ack time
17958 */
17959 struct rack_sendmap *nrsm;
17960
17961 tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
17962 rack->app_limited_needs_set = 0;
17963 /*
17964 * Ok in this path we need to use the r_end now
17965 * since this guy is the starting ack.
17966 */
17967 tp->gput_seq = my_rsm->r_end;
17968 /*
17969 * We also need to adjust up the sendtime
17970 * to the send of the next data after my_rsm.
17971 */
17972 nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
17973 if (nrsm != NULL)
17974 my_rsm = nrsm;
17975 else {
17976 /*
17977 * The next as not been sent, thats the
17978 * case for using the latest.
17979 */
17980 goto use_latest;
17981 }
17982 }
17983 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
17984 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
17985 rack->r_ctl.rc_gp_cumack_ts = 0;
17986 if ((rack->r_ctl.cleared_app_ack == 1) &&
17987 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.cleared_app_ack_seq))) {
17988 /*
17989 * We just cleared an application limited period
17990 * so the next seq out needs to skip the first
17991 * ack.
17992 */
17993 rack->app_limited_needs_set = 1;
17994 rack->r_ctl.cleared_app_ack = 0;
17995 }
17996 rack_log_pacing_delay_calc(rack,
17997 tp->gput_seq,
17998 tp->gput_ack,
17999 (uintptr_t)my_rsm,
18000 tp->gput_ts,
18001 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18002 9,
18003 __LINE__, my_rsm, 0);
18004 /* Now lets make sure all are marked as they should be */
18005 rack_tend_gp_marks(tp, rack);
18006 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18007 return;
18008 }
18009
18010 use_latest:
18011 /*
18012 * We don't know how long we may have been
18013 * idle or if this is the first-send. Lets
18014 * setup the flag so we will trim off
18015 * the first ack'd data so we get a true
18016 * measurement.
18017 */
18018 rack->app_limited_needs_set = 1;
18019 tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18020 rack->r_ctl.rc_gp_cumack_ts = 0;
18021 /* Find this guy so we can pull the send time */
18022 my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18023 if (my_rsm) {
18024 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18025 if (my_rsm->r_flags & RACK_ACKED) {
18026 /*
18027 * Unlikely since its probably what was
18028 * just transmitted (but I am paranoid).
18029 */
18030 tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18031 rack->app_limited_needs_set = 0;
18032 }
18033 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18034 /* This also is unlikely */
18035 tp->gput_seq = my_rsm->r_start;
18036 }
18037 } else {
18038 /*
18039 * TSNH unless we have some send-map limit,
18040 * and even at that it should not be hitting
18041 * that limit (we should have stopped sending).
18042 */
18043 struct timeval tv;
18044
18045 microuptime(&tv);
18046 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18047 }
18048 rack_tend_gp_marks(tp, rack);
18049 rack_log_pacing_delay_calc(rack,
18050 tp->gput_seq,
18051 tp->gput_ack,
18052 (uintptr_t)my_rsm,
18053 tp->gput_ts,
18054 (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18055 9, __LINE__, NULL, 0);
18056 rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18057 }
18058
18059 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)18060 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cwnd_to_use,
18061 uint32_t avail, int32_t sb_offset)
18062 {
18063 uint32_t len;
18064 uint32_t sendwin;
18065
18066 if (tp->snd_wnd > cwnd_to_use)
18067 sendwin = cwnd_to_use;
18068 else
18069 sendwin = tp->snd_wnd;
18070 if (ctf_outstanding(tp) >= tp->snd_wnd) {
18071 /* We never want to go over our peers rcv-window */
18072 len = 0;
18073 } else {
18074 uint32_t flight;
18075
18076 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18077 if (flight >= sendwin) {
18078 /*
18079 * We have in flight what we are allowed by cwnd (if
18080 * it was rwnd blocking it would have hit above out
18081 * >= tp->snd_wnd).
18082 */
18083 return (0);
18084 }
18085 len = sendwin - flight;
18086 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18087 /* We would send too much (beyond the rwnd) */
18088 len = tp->snd_wnd - ctf_outstanding(tp);
18089 }
18090 if ((len + sb_offset) > avail) {
18091 /*
18092 * We don't have that much in the SB, how much is
18093 * there?
18094 */
18095 len = avail - sb_offset;
18096 }
18097 }
18098 return (len);
18099 }
18100
18101 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)18102 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18103 unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18104 int rsm_is_null, int optlen, int line, uint16_t mode)
18105 {
18106 if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18107 union tcp_log_stackspecific log;
18108 struct timeval tv;
18109
18110 memset(&log, 0, sizeof(log));
18111 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18112 log.u_bbr.flex1 = error;
18113 log.u_bbr.flex2 = flags;
18114 log.u_bbr.flex3 = rsm_is_null;
18115 log.u_bbr.flex4 = ipoptlen;
18116 log.u_bbr.flex5 = tp->rcv_numsacks;
18117 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18118 log.u_bbr.flex7 = optlen;
18119 log.u_bbr.flex8 = rack->r_fsb_inited;
18120 log.u_bbr.applimited = rack->r_fast_output;
18121 log.u_bbr.bw_inuse = rack_get_bw(rack);
18122 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18123 log.u_bbr.cwnd_gain = mode;
18124 log.u_bbr.pkts_out = orig_len;
18125 log.u_bbr.lt_epoch = len;
18126 log.u_bbr.delivered = line;
18127 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18128 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18129 tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18130 len, &log, false, NULL, __func__, __LINE__, &tv);
18131 }
18132 }
18133
18134
18135 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)18136 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18137 struct rack_fast_send_blk *fsb,
18138 int32_t seglimit, int32_t segsize, int hw_tls)
18139 {
18140 #ifdef KERN_TLS
18141 struct ktls_session *tls, *ntls;
18142 #ifdef INVARIANTS
18143 struct mbuf *start;
18144 #endif
18145 #endif
18146 struct mbuf *m, *n, **np, *smb;
18147 struct mbuf *top;
18148 int32_t off, soff;
18149 int32_t len = *plen;
18150 int32_t fragsize;
18151 int32_t len_cp = 0;
18152 uint32_t mlen, frags;
18153
18154 soff = off = the_off;
18155 smb = m = the_m;
18156 np = ⊤
18157 top = NULL;
18158 #ifdef KERN_TLS
18159 if (hw_tls && (m->m_flags & M_EXTPG))
18160 tls = m->m_epg_tls;
18161 else
18162 tls = NULL;
18163 #ifdef INVARIANTS
18164 start = m;
18165 #endif
18166 #endif
18167 while (len > 0) {
18168 if (m == NULL) {
18169 *plen = len_cp;
18170 break;
18171 }
18172 #ifdef KERN_TLS
18173 if (hw_tls) {
18174 if (m->m_flags & M_EXTPG)
18175 ntls = m->m_epg_tls;
18176 else
18177 ntls = NULL;
18178
18179 /*
18180 * Avoid mixing TLS records with handshake
18181 * data or TLS records from different
18182 * sessions.
18183 */
18184 if (tls != ntls) {
18185 MPASS(m != start);
18186 *plen = len_cp;
18187 break;
18188 }
18189 }
18190 #endif
18191 mlen = min(len, m->m_len - off);
18192 if (seglimit) {
18193 /*
18194 * For M_EXTPG mbufs, add 3 segments
18195 * + 1 in case we are crossing page boundaries
18196 * + 2 in case the TLS hdr/trailer are used
18197 * It is cheaper to just add the segments
18198 * than it is to take the cache miss to look
18199 * at the mbuf ext_pgs state in detail.
18200 */
18201 if (m->m_flags & M_EXTPG) {
18202 fragsize = min(segsize, PAGE_SIZE);
18203 frags = 3;
18204 } else {
18205 fragsize = segsize;
18206 frags = 0;
18207 }
18208
18209 /* Break if we really can't fit anymore. */
18210 if ((frags + 1) >= seglimit) {
18211 *plen = len_cp;
18212 break;
18213 }
18214
18215 /*
18216 * Reduce size if you can't copy the whole
18217 * mbuf. If we can't copy the whole mbuf, also
18218 * adjust len so the loop will end after this
18219 * mbuf.
18220 */
18221 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18222 mlen = (seglimit - frags - 1) * fragsize;
18223 len = mlen;
18224 *plen = len_cp + len;
18225 }
18226 frags += howmany(mlen, fragsize);
18227 if (frags == 0)
18228 frags++;
18229 seglimit -= frags;
18230 KASSERT(seglimit > 0,
18231 ("%s: seglimit went too low", __func__));
18232 }
18233 n = m_get(M_NOWAIT, m->m_type);
18234 *np = n;
18235 if (n == NULL)
18236 goto nospace;
18237 n->m_len = mlen;
18238 soff += mlen;
18239 len_cp += n->m_len;
18240 if (m->m_flags & (M_EXT | M_EXTPG)) {
18241 n->m_data = m->m_data + off;
18242 mb_dupcl(n, m);
18243 } else {
18244 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18245 (u_int)n->m_len);
18246 }
18247 len -= n->m_len;
18248 off = 0;
18249 m = m->m_next;
18250 np = &n->m_next;
18251 if (len || (soff == smb->m_len)) {
18252 /*
18253 * We have more so we move forward or
18254 * we have consumed the entire mbuf and
18255 * len has fell to 0.
18256 */
18257 soff = 0;
18258 smb = m;
18259 }
18260
18261 }
18262 if (fsb != NULL) {
18263 fsb->m = smb;
18264 fsb->off = soff;
18265 if (smb) {
18266 /*
18267 * Save off the size of the mbuf. We do
18268 * this so that we can recognize when it
18269 * has been trimmed by sbcut() as acks
18270 * come in.
18271 */
18272 fsb->o_m_len = smb->m_len;
18273 fsb->o_t_len = M_TRAILINGROOM(smb);
18274 } else {
18275 /*
18276 * This is the case where the next mbuf went to NULL. This
18277 * means with this copy we have sent everything in the sb.
18278 * In theory we could clear the fast_output flag, but lets
18279 * not since its possible that we could get more added
18280 * and acks that call the extend function which would let
18281 * us send more.
18282 */
18283 fsb->o_m_len = 0;
18284 fsb->o_t_len = 0;
18285 }
18286 }
18287 return (top);
18288 nospace:
18289 if (top)
18290 m_freem(top);
18291 return (NULL);
18292
18293 }
18294
18295 /*
18296 * This is a copy of m_copym(), taking the TSO segment size/limit
18297 * constraints into account, and advancing the sndptr as it goes.
18298 */
18299 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)18300 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18301 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18302 {
18303 struct mbuf *m, *n;
18304 int32_t soff;
18305
18306 m = rack->r_ctl.fsb.m;
18307 if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18308 /*
18309 * The trailing space changed, mbufs can grow
18310 * at the tail but they can't shrink from
18311 * it, KASSERT that. Adjust the orig_m_len to
18312 * compensate for this change.
18313 */
18314 KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18315 ("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18316 m,
18317 rack,
18318 (intmax_t)M_TRAILINGROOM(m),
18319 rack->r_ctl.fsb.o_t_len,
18320 rack->r_ctl.fsb.o_m_len,
18321 m->m_len));
18322 rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18323 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18324 }
18325 if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18326 /*
18327 * Mbuf shrank, trimmed off the top by an ack, our
18328 * offset changes.
18329 */
18330 KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18331 ("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18332 m, m->m_len,
18333 rack, rack->r_ctl.fsb.o_m_len,
18334 rack->r_ctl.fsb.off));
18335
18336 if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18337 rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18338 else
18339 rack->r_ctl.fsb.off = 0;
18340 rack->r_ctl.fsb.o_m_len = m->m_len;
18341 #ifdef INVARIANTS
18342 } else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18343 panic("rack:%p m:%p m_len grew outside of t_space compensation",
18344 rack, m);
18345 #endif
18346 }
18347 soff = rack->r_ctl.fsb.off;
18348 KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18349 KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18350 KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18351 __FUNCTION__,
18352 rack, *plen, m, m->m_len));
18353 /* Save off the right location before we copy and advance */
18354 *s_soff = soff;
18355 *s_mb = rack->r_ctl.fsb.m;
18356 n = rack_fo_base_copym(m, soff, plen,
18357 &rack->r_ctl.fsb,
18358 seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18359 return (n);
18360 }
18361
18362 /* Log the buffer level */
18363 static void
rack_log_queue_level(struct tcpcb * tp,struct tcp_rack * rack,int len,struct timeval * tv,uint32_t cts)18364 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18365 int len, struct timeval *tv,
18366 uint32_t cts)
18367 {
18368 uint32_t p_rate = 0, p_queue = 0, err = 0;
18369 union tcp_log_stackspecific log;
18370
18371 #ifdef RATELIMIT
18372 err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18373 err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18374 #endif
18375 memset(&log, 0, sizeof(log));
18376 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18377 log.u_bbr.flex1 = p_rate;
18378 log.u_bbr.flex2 = p_queue;
18379 log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18380 log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18381 log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18382 log.u_bbr.flex7 = 99;
18383 log.u_bbr.flex8 = 0;
18384 log.u_bbr.pkts_out = err;
18385 log.u_bbr.delRate = rack->r_ctl.crte->rate;
18386 log.u_bbr.timeStamp = cts;
18387 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18388 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18389 len, &log, false, NULL, __func__, __LINE__, tv);
18390
18391 }
18392
18393 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)18394 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18395 struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18396 {
18397 uint64_t lentime = 0;
18398 #ifdef RATELIMIT
18399 uint32_t p_rate = 0, p_queue = 0, err;
18400 union tcp_log_stackspecific log;
18401 uint64_t bw;
18402
18403 err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18404 /* Failed or queue is zero */
18405 if (err || (p_queue == 0)) {
18406 lentime = 0;
18407 goto out;
18408 }
18409 err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18410 if (err) {
18411 lentime = 0;
18412 goto out;
18413 }
18414 /*
18415 * If we reach here we have some bytes in
18416 * the queue. The number returned is a value
18417 * between 0 and 0xffff where ffff is full
18418 * and 0 is empty. So how best to make this into
18419 * something usable?
18420 *
18421 * The "safer" way is lets take the b/w gotten
18422 * from the query (which should be our b/w rate)
18423 * and pretend that a full send (our rc_pace_max_segs)
18424 * is outstanding. We factor it so its as if a full
18425 * number of our MSS segment is terms of full
18426 * ethernet segments are outstanding.
18427 */
18428 bw = p_rate / 8;
18429 if (bw) {
18430 lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18431 lentime *= ETHERNET_SEGMENT_SIZE;
18432 lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18433 lentime /= bw;
18434 } else {
18435 /* TSNH -- KASSERT? */
18436 lentime = 0;
18437 }
18438 out:
18439 if (tcp_bblogging_on(tp)) {
18440 memset(&log, 0, sizeof(log));
18441 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18442 log.u_bbr.flex1 = p_rate;
18443 log.u_bbr.flex2 = p_queue;
18444 log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18445 log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18446 log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18447 log.u_bbr.flex7 = 99;
18448 log.u_bbr.flex8 = 0;
18449 log.u_bbr.pkts_out = err;
18450 log.u_bbr.delRate = rack->r_ctl.crte->rate;
18451 log.u_bbr.cur_del_rate = lentime;
18452 log.u_bbr.timeStamp = cts;
18453 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18454 tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18455 len, &log, false, NULL, __func__, __LINE__,tv);
18456 }
18457 #endif
18458 return ((uint32_t)lentime);
18459 }
18460
18461 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)18462 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18463 uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18464 {
18465 /*
18466 * Enter the fast retransmit path. We are given that a sched_pin is
18467 * in place (if accounting is compliled in) and the cycle count taken
18468 * at the entry is in the ts_val. The concept her is that the rsm
18469 * now holds the mbuf offsets and such so we can directly transmit
18470 * without a lot of overhead, the len field is already set for
18471 * us to prohibit us from sending too much (usually its 1MSS).
18472 */
18473 struct ip *ip = NULL;
18474 struct udphdr *udp = NULL;
18475 struct tcphdr *th = NULL;
18476 struct mbuf *m = NULL;
18477 struct inpcb *inp;
18478 uint8_t *cpto;
18479 struct tcp_log_buffer *lgb;
18480 #ifdef TCP_ACCOUNTING
18481 uint64_t crtsc;
18482 int cnt_thru = 1;
18483 #endif
18484 struct tcpopt to;
18485 u_char opt[TCP_MAXOLEN];
18486 uint32_t hdrlen, optlen;
18487 int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18488 uint16_t flags;
18489 uint32_t if_hw_tsomaxsegcount = 0, startseq;
18490 uint32_t if_hw_tsomaxsegsize;
18491 int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18492
18493 #ifdef INET6
18494 struct ip6_hdr *ip6 = NULL;
18495
18496 if (rack->r_is_v6) {
18497 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18498 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18499 } else
18500 #endif /* INET6 */
18501 {
18502 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18503 hdrlen = sizeof(struct tcpiphdr);
18504 }
18505 if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18506 goto failed;
18507 }
18508 if (doing_tlp) {
18509 /* Its a TLP add the flag, it may already be there but be sure */
18510 rsm->r_flags |= RACK_TLP;
18511 } else {
18512 /* If it was a TLP it is not not on this retransmit */
18513 rsm->r_flags &= ~RACK_TLP;
18514 }
18515 startseq = rsm->r_start;
18516 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18517 inp = rack->rc_inp;
18518 to.to_flags = 0;
18519 flags = tcp_outflags[tp->t_state];
18520 if (flags & (TH_SYN|TH_RST)) {
18521 goto failed;
18522 }
18523 if (rsm->r_flags & RACK_HAS_FIN) {
18524 /* We can't send a FIN here */
18525 goto failed;
18526 }
18527 if (flags & TH_FIN) {
18528 /* We never send a FIN */
18529 flags &= ~TH_FIN;
18530 }
18531 if (tp->t_flags & TF_RCVD_TSTMP) {
18532 to.to_tsval = ms_cts + tp->ts_offset;
18533 to.to_tsecr = tp->ts_recent;
18534 to.to_flags = TOF_TS;
18535 }
18536 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18537 /* TCP-MD5 (RFC2385). */
18538 if (tp->t_flags & TF_SIGNATURE)
18539 to.to_flags |= TOF_SIGNATURE;
18540 #endif
18541 optlen = tcp_addoptions(&to, opt);
18542 hdrlen += optlen;
18543 udp = rack->r_ctl.fsb.udp;
18544 if (udp)
18545 hdrlen += sizeof(struct udphdr);
18546 if (rack->r_ctl.rc_pace_max_segs)
18547 max_val = rack->r_ctl.rc_pace_max_segs;
18548 else if (rack->rc_user_set_max_segs)
18549 max_val = rack->rc_user_set_max_segs * segsiz;
18550 else
18551 max_val = len;
18552 if ((tp->t_flags & TF_TSO) &&
18553 V_tcp_do_tso &&
18554 (len > segsiz) &&
18555 (tp->t_port == 0))
18556 tso = 1;
18557 #ifdef INET6
18558 if (MHLEN < hdrlen + max_linkhdr)
18559 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18560 else
18561 #endif
18562 m = m_gethdr(M_NOWAIT, MT_DATA);
18563 if (m == NULL)
18564 goto failed;
18565 m->m_data += max_linkhdr;
18566 m->m_len = hdrlen;
18567 th = rack->r_ctl.fsb.th;
18568 /* Establish the len to send */
18569 if (len > max_val)
18570 len = max_val;
18571 if ((tso) && (len + optlen > segsiz)) {
18572 uint32_t if_hw_tsomax;
18573 int32_t max_len;
18574
18575 /* extract TSO information */
18576 if_hw_tsomax = tp->t_tsomax;
18577 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18578 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18579 /*
18580 * Check if we should limit by maximum payload
18581 * length:
18582 */
18583 if (if_hw_tsomax != 0) {
18584 /* compute maximum TSO length */
18585 max_len = (if_hw_tsomax - hdrlen -
18586 max_linkhdr);
18587 if (max_len <= 0) {
18588 goto failed;
18589 } else if (len > max_len) {
18590 len = max_len;
18591 }
18592 }
18593 if (len <= segsiz) {
18594 /*
18595 * In case there are too many small fragments don't
18596 * use TSO:
18597 */
18598 tso = 0;
18599 }
18600 } else {
18601 tso = 0;
18602 }
18603 if ((tso == 0) && (len > segsiz))
18604 len = segsiz;
18605 (void)tcp_get_usecs(tv);
18606 if ((len == 0) ||
18607 (len <= MHLEN - hdrlen - max_linkhdr)) {
18608 goto failed;
18609 }
18610 th->th_seq = htonl(rsm->r_start);
18611 th->th_ack = htonl(tp->rcv_nxt);
18612 /*
18613 * The PUSH bit should only be applied
18614 * if the full retransmission is made. If
18615 * we are sending less than this is the
18616 * left hand edge and should not have
18617 * the PUSH bit.
18618 */
18619 if ((rsm->r_flags & RACK_HAD_PUSH) &&
18620 (len == (rsm->r_end - rsm->r_start)))
18621 flags |= TH_PUSH;
18622 th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18623 if (th->th_win == 0) {
18624 tp->t_sndzerowin++;
18625 tp->t_flags |= TF_RXWIN0SENT;
18626 } else
18627 tp->t_flags &= ~TF_RXWIN0SENT;
18628 if (rsm->r_flags & RACK_TLP) {
18629 /*
18630 * TLP should not count in retran count, but
18631 * in its own bin
18632 */
18633 counter_u64_add(rack_tlp_retran, 1);
18634 counter_u64_add(rack_tlp_retran_bytes, len);
18635 } else {
18636 tp->t_sndrexmitpack++;
18637 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18638 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18639 }
18640 #ifdef STATS
18641 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18642 len);
18643 #endif
18644 if (rsm->m == NULL)
18645 goto failed;
18646 if (rsm->m &&
18647 ((rsm->orig_m_len != rsm->m->m_len) ||
18648 (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18649 /* Fix up the orig_m_len and possibly the mbuf offset */
18650 rack_adjust_orig_mlen(rsm);
18651 }
18652 m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18653 if (len <= segsiz) {
18654 /*
18655 * Must have ran out of mbufs for the copy
18656 * shorten it to no longer need tso. Lets
18657 * not put on sendalot since we are low on
18658 * mbufs.
18659 */
18660 tso = 0;
18661 }
18662 if ((m->m_next == NULL) || (len <= 0)){
18663 goto failed;
18664 }
18665 if (udp) {
18666 if (rack->r_is_v6)
18667 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18668 else
18669 ulen = hdrlen + len - sizeof(struct ip);
18670 udp->uh_ulen = htons(ulen);
18671 }
18672 m->m_pkthdr.rcvif = (struct ifnet *)0;
18673 if (TCPS_HAVERCVDSYN(tp->t_state) &&
18674 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18675 int ect = tcp_ecn_output_established(tp, &flags, len, true);
18676 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18677 (tp->t_flags2 & TF2_ECN_SND_ECE))
18678 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18679 #ifdef INET6
18680 if (rack->r_is_v6) {
18681 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18682 ip6->ip6_flow |= htonl(ect << 20);
18683 }
18684 else
18685 #endif
18686 {
18687 ip->ip_tos &= ~IPTOS_ECN_MASK;
18688 ip->ip_tos |= ect;
18689 }
18690 }
18691 if (rack->r_ctl.crte != NULL) {
18692 /* See if we can send via the hw queue */
18693 slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18694 /* If there is nothing in queue (no pacing time) we can send via the hw queue */
18695 if (slot == 0)
18696 ip_sendflag = 0;
18697 }
18698 tcp_set_flags(th, flags);
18699 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18700 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18701 if (to.to_flags & TOF_SIGNATURE) {
18702 /*
18703 * Calculate MD5 signature and put it into the place
18704 * determined before.
18705 * NOTE: since TCP options buffer doesn't point into
18706 * mbuf's data, calculate offset and use it.
18707 */
18708 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18709 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18710 /*
18711 * Do not send segment if the calculation of MD5
18712 * digest has failed.
18713 */
18714 goto failed;
18715 }
18716 }
18717 #endif
18718 #ifdef INET6
18719 if (rack->r_is_v6) {
18720 if (tp->t_port) {
18721 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18722 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18723 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18724 th->th_sum = htons(0);
18725 UDPSTAT_INC(udps_opackets);
18726 } else {
18727 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18728 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18729 th->th_sum = in6_cksum_pseudo(ip6,
18730 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18731 0);
18732 }
18733 }
18734 #endif
18735 #if defined(INET6) && defined(INET)
18736 else
18737 #endif
18738 #ifdef INET
18739 {
18740 if (tp->t_port) {
18741 m->m_pkthdr.csum_flags = CSUM_UDP;
18742 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18743 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18744 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18745 th->th_sum = htons(0);
18746 UDPSTAT_INC(udps_opackets);
18747 } else {
18748 m->m_pkthdr.csum_flags = CSUM_TCP;
18749 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18750 th->th_sum = in_pseudo(ip->ip_src.s_addr,
18751 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18752 IPPROTO_TCP + len + optlen));
18753 }
18754 /* IP version must be set here for ipv4/ipv6 checking later */
18755 KASSERT(ip->ip_v == IPVERSION,
18756 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18757 }
18758 #endif
18759 if (tso) {
18760 /*
18761 * Here we use segsiz since we have no added options besides
18762 * any standard timestamp options (no DSACKs or SACKS are sent
18763 * via either fast-path).
18764 */
18765 KASSERT(len > segsiz,
18766 ("%s: len <= tso_segsz tp:%p", __func__, tp));
18767 m->m_pkthdr.csum_flags |= CSUM_TSO;
18768 m->m_pkthdr.tso_segsz = segsiz;
18769 }
18770 #ifdef INET6
18771 if (rack->r_is_v6) {
18772 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18773 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18774 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18775 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18776 else
18777 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18778 }
18779 #endif
18780 #if defined(INET) && defined(INET6)
18781 else
18782 #endif
18783 #ifdef INET
18784 {
18785 ip->ip_len = htons(m->m_pkthdr.len);
18786 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
18787 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18788 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18789 if (tp->t_port == 0 || len < V_tcp_minmss) {
18790 ip->ip_off |= htons(IP_DF);
18791 }
18792 } else {
18793 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18794 }
18795 }
18796 #endif
18797 if (doing_tlp == 0) {
18798 /* Set we retransmitted */
18799 rack->rc_gp_saw_rec = 1;
18800 } else {
18801 /* Its a TLP set ca or ss */
18802 if (tp->snd_cwnd > tp->snd_ssthresh) {
18803 /* Set we sent in CA */
18804 rack->rc_gp_saw_ca = 1;
18805 } else {
18806 /* Set we sent in SS */
18807 rack->rc_gp_saw_ss = 1;
18808 }
18809 }
18810 /* Time to copy in our header */
18811 cpto = mtod(m, uint8_t *);
18812 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18813 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18814 if (optlen) {
18815 bcopy(opt, th + 1, optlen);
18816 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18817 } else {
18818 th->th_off = sizeof(struct tcphdr) >> 2;
18819 }
18820 if (tcp_bblogging_on(rack->rc_tp)) {
18821 union tcp_log_stackspecific log;
18822
18823 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18824 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18825 counter_u64_add(rack_collapsed_win_rxt, 1);
18826 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18827 }
18828 memset(&log, 0, sizeof(log));
18829 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18830 if (rack->rack_no_prr)
18831 log.u_bbr.flex1 = 0;
18832 else
18833 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18834 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18835 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18836 log.u_bbr.flex4 = max_val;
18837 /* Save off the early/late values */
18838 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18839 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18840 log.u_bbr.bw_inuse = rack_get_bw(rack);
18841 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
18842 if (doing_tlp == 0)
18843 log.u_bbr.flex8 = 1;
18844 else
18845 log.u_bbr.flex8 = 2;
18846 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18847 log.u_bbr.flex7 = 55;
18848 log.u_bbr.pkts_out = tp->t_maxseg;
18849 log.u_bbr.timeStamp = cts;
18850 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18851 if (rsm->r_rtr_cnt > 0) {
18852 /*
18853 * When we have a retransmit we want to log the
18854 * burst at send and flight at send from before.
18855 */
18856 log.u_bbr.flex5 = rsm->r_fas;
18857 log.u_bbr.bbr_substate = rsm->r_bas;
18858 } else {
18859 /*
18860 * This is currently unlikely until we do the
18861 * packet pair probes but I will add it for completeness.
18862 */
18863 log.u_bbr.flex5 = log.u_bbr.inflight;
18864 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
18865 }
18866 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
18867 log.u_bbr.delivered = 0;
18868 log.u_bbr.rttProp = (uintptr_t)rsm;
18869 log.u_bbr.delRate = rsm->r_flags;
18870 log.u_bbr.delRate <<= 31;
18871 log.u_bbr.delRate |= rack->r_must_retran;
18872 log.u_bbr.delRate <<= 1;
18873 log.u_bbr.delRate |= 1;
18874 log.u_bbr.pkt_epoch = __LINE__;
18875 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
18876 len, &log, false, NULL, __func__, __LINE__, tv);
18877 } else
18878 lgb = NULL;
18879 if ((rack->r_ctl.crte != NULL) &&
18880 tcp_bblogging_on(tp)) {
18881 rack_log_queue_level(tp, rack, len, tv, cts);
18882 }
18883 #ifdef INET6
18884 if (rack->r_is_v6) {
18885 error = ip6_output(m, inp->in6p_outputopts,
18886 &inp->inp_route6,
18887 ip_sendflag, NULL, NULL, inp);
18888 }
18889 else
18890 #endif
18891 #ifdef INET
18892 {
18893 error = ip_output(m, NULL,
18894 &inp->inp_route,
18895 ip_sendflag, 0, inp);
18896 }
18897 #endif
18898 m = NULL;
18899 if (lgb) {
18900 lgb->tlb_errno = error;
18901 lgb = NULL;
18902 }
18903 /* Move snd_nxt to snd_max so we don't have false retransmissions */
18904 tp->snd_nxt = tp->snd_max;
18905 if (error) {
18906 goto failed;
18907 } else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
18908 rack->rc_hw_nobuf = 0;
18909 rack->r_ctl.rc_agg_delayed = 0;
18910 rack->r_early = 0;
18911 rack->r_late = 0;
18912 rack->r_ctl.rc_agg_early = 0;
18913 }
18914 rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
18915 rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
18916 if (doing_tlp) {
18917 rack->rc_tlp_in_progress = 1;
18918 rack->r_ctl.rc_tlp_cnt_out++;
18919 }
18920 if (error == 0) {
18921 counter_u64_add(rack_total_bytes, len);
18922 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
18923 if (doing_tlp) {
18924 rack->rc_last_sent_tlp_past_cumack = 0;
18925 rack->rc_last_sent_tlp_seq_valid = 1;
18926 rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18927 rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18928 }
18929 if (rack->r_ctl.rc_prr_sndcnt >= len)
18930 rack->r_ctl.rc_prr_sndcnt -= len;
18931 else
18932 rack->r_ctl.rc_prr_sndcnt = 0;
18933 }
18934 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18935 rack->forced_ack = 0; /* If we send something zap the FA flag */
18936 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18937 rack->r_ctl.retran_during_recovery += len;
18938 {
18939 int idx;
18940
18941 idx = (len / segsiz) + 3;
18942 if (idx >= TCP_MSS_ACCT_ATIMER)
18943 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18944 else
18945 counter_u64_add(rack_out_size[idx], 1);
18946 }
18947 if (tp->t_rtttime == 0) {
18948 tp->t_rtttime = ticks;
18949 tp->t_rtseq = startseq;
18950 KMOD_TCPSTAT_INC(tcps_segstimed);
18951 }
18952 counter_u64_add(rack_fto_rsm_send, 1);
18953 if (error && (error == ENOBUFS)) {
18954 if (rack->r_ctl.crte != NULL) {
18955 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
18956 if (tcp_bblogging_on(rack->rc_tp))
18957 rack_log_queue_level(tp, rack, len, tv, cts);
18958 } else
18959 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
18960 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18961 if (rack->rc_enobuf < 0x7f)
18962 rack->rc_enobuf++;
18963 if (slot < (10 * HPTS_USEC_IN_MSEC))
18964 slot = 10 * HPTS_USEC_IN_MSEC;
18965 if (rack->r_ctl.crte != NULL) {
18966 counter_u64_add(rack_saw_enobuf_hw, 1);
18967 tcp_rl_log_enobuf(rack->r_ctl.crte);
18968 }
18969 counter_u64_add(rack_saw_enobuf, 1);
18970 } else {
18971 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz, __LINE__);
18972 }
18973 rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
18974 #ifdef TCP_ACCOUNTING
18975 crtsc = get_cyclecount();
18976 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18977 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
18978 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18979 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
18980 }
18981 sched_unpin();
18982 #endif
18983 return (0);
18984 failed:
18985 if (m)
18986 m_free(m);
18987 return (-1);
18988 }
18989
18990 static void
rack_sndbuf_autoscale(struct tcp_rack * rack)18991 rack_sndbuf_autoscale(struct tcp_rack *rack)
18992 {
18993 /*
18994 * Automatic sizing of send socket buffer. Often the send buffer
18995 * size is not optimally adjusted to the actual network conditions
18996 * at hand (delay bandwidth product). Setting the buffer size too
18997 * small limits throughput on links with high bandwidth and high
18998 * delay (eg. trans-continental/oceanic links). Setting the
18999 * buffer size too big consumes too much real kernel memory,
19000 * especially with many connections on busy servers.
19001 *
19002 * The criteria to step up the send buffer one notch are:
19003 * 1. receive window of remote host is larger than send buffer
19004 * (with a fudge factor of 5/4th);
19005 * 2. send buffer is filled to 7/8th with data (so we actually
19006 * have data to make use of it);
19007 * 3. send buffer fill has not hit maximal automatic size;
19008 * 4. our send window (slow start and cogestion controlled) is
19009 * larger than sent but unacknowledged data in send buffer.
19010 *
19011 * Note that the rack version moves things much faster since
19012 * we want to avoid hitting cache lines in the rack_fast_output()
19013 * path so this is called much less often and thus moves
19014 * the SB forward by a percentage.
19015 */
19016 struct socket *so;
19017 struct tcpcb *tp;
19018 uint32_t sendwin, scaleup;
19019
19020 tp = rack->rc_tp;
19021 so = rack->rc_inp->inp_socket;
19022 sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19023 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19024 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19025 sbused(&so->so_snd) >=
19026 (so->so_snd.sb_hiwat / 8 * 7) &&
19027 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19028 sendwin >= (sbused(&so->so_snd) -
19029 (tp->snd_max - tp->snd_una))) {
19030 if (rack_autosndbuf_inc)
19031 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19032 else
19033 scaleup = V_tcp_autosndbuf_inc;
19034 if (scaleup < V_tcp_autosndbuf_inc)
19035 scaleup = V_tcp_autosndbuf_inc;
19036 scaleup += so->so_snd.sb_hiwat;
19037 if (scaleup > V_tcp_autosndbuf_max)
19038 scaleup = V_tcp_autosndbuf_max;
19039 if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19040 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19041 }
19042 }
19043 }
19044
19045 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,int line)19046 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19047 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long *tot_len, int *send_err, int line)
19048 {
19049 /*
19050 * Enter to do fast output. We are given that the sched_pin is
19051 * in place (if accounting is compiled in) and the cycle count taken
19052 * at entry is in place in ts_val. The idea here is that
19053 * we know how many more bytes needs to be sent (presumably either
19054 * during pacing or to fill the cwnd and that was greater than
19055 * the max-burst). We have how much to send and all the info we
19056 * need to just send.
19057 */
19058 #ifdef INET
19059 struct ip *ip = NULL;
19060 #endif
19061 struct udphdr *udp = NULL;
19062 struct tcphdr *th = NULL;
19063 struct mbuf *m, *s_mb;
19064 struct inpcb *inp;
19065 uint8_t *cpto;
19066 struct tcp_log_buffer *lgb;
19067 #ifdef TCP_ACCOUNTING
19068 uint64_t crtsc;
19069 #endif
19070 struct tcpopt to;
19071 u_char opt[TCP_MAXOLEN];
19072 uint32_t hdrlen, optlen;
19073 #ifdef TCP_ACCOUNTING
19074 int cnt_thru = 1;
19075 #endif
19076 int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19077 uint16_t flags;
19078 uint32_t s_soff;
19079 uint32_t if_hw_tsomaxsegcount = 0, startseq;
19080 uint32_t if_hw_tsomaxsegsize;
19081 uint32_t add_flag = RACK_SENT_FP;
19082 #ifdef INET6
19083 struct ip6_hdr *ip6 = NULL;
19084
19085 if (rack->r_is_v6) {
19086 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19087 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19088 } else
19089 #endif /* INET6 */
19090 {
19091 #ifdef INET
19092 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19093 hdrlen = sizeof(struct tcpiphdr);
19094 #endif
19095 }
19096 if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19097 m = NULL;
19098 goto failed;
19099 }
19100 rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19101 startseq = tp->snd_max;
19102 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19103 inp = rack->rc_inp;
19104 len = rack->r_ctl.fsb.left_to_send;
19105 to.to_flags = 0;
19106 flags = rack->r_ctl.fsb.tcp_flags;
19107 if (tp->t_flags & TF_RCVD_TSTMP) {
19108 to.to_tsval = ms_cts + tp->ts_offset;
19109 to.to_tsecr = tp->ts_recent;
19110 to.to_flags = TOF_TS;
19111 }
19112 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19113 /* TCP-MD5 (RFC2385). */
19114 if (tp->t_flags & TF_SIGNATURE)
19115 to.to_flags |= TOF_SIGNATURE;
19116 #endif
19117 optlen = tcp_addoptions(&to, opt);
19118 hdrlen += optlen;
19119 udp = rack->r_ctl.fsb.udp;
19120 if (udp)
19121 hdrlen += sizeof(struct udphdr);
19122 if (rack->r_ctl.rc_pace_max_segs)
19123 max_val = rack->r_ctl.rc_pace_max_segs;
19124 else if (rack->rc_user_set_max_segs)
19125 max_val = rack->rc_user_set_max_segs * segsiz;
19126 else
19127 max_val = len;
19128 if ((tp->t_flags & TF_TSO) &&
19129 V_tcp_do_tso &&
19130 (len > segsiz) &&
19131 (tp->t_port == 0))
19132 tso = 1;
19133 again:
19134 #ifdef INET6
19135 if (MHLEN < hdrlen + max_linkhdr)
19136 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19137 else
19138 #endif
19139 m = m_gethdr(M_NOWAIT, MT_DATA);
19140 if (m == NULL)
19141 goto failed;
19142 m->m_data += max_linkhdr;
19143 m->m_len = hdrlen;
19144 th = rack->r_ctl.fsb.th;
19145 /* Establish the len to send */
19146 if (len > max_val)
19147 len = max_val;
19148 if ((tso) && (len + optlen > segsiz)) {
19149 uint32_t if_hw_tsomax;
19150 int32_t max_len;
19151
19152 /* extract TSO information */
19153 if_hw_tsomax = tp->t_tsomax;
19154 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19155 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19156 /*
19157 * Check if we should limit by maximum payload
19158 * length:
19159 */
19160 if (if_hw_tsomax != 0) {
19161 /* compute maximum TSO length */
19162 max_len = (if_hw_tsomax - hdrlen -
19163 max_linkhdr);
19164 if (max_len <= 0) {
19165 goto failed;
19166 } else if (len > max_len) {
19167 len = max_len;
19168 }
19169 }
19170 if (len <= segsiz) {
19171 /*
19172 * In case there are too many small fragments don't
19173 * use TSO:
19174 */
19175 tso = 0;
19176 }
19177 } else {
19178 tso = 0;
19179 }
19180 if ((tso == 0) && (len > segsiz))
19181 len = segsiz;
19182 (void)tcp_get_usecs(tv);
19183 if ((len == 0) ||
19184 (len <= MHLEN - hdrlen - max_linkhdr)) {
19185 goto failed;
19186 }
19187 sb_offset = tp->snd_max - tp->snd_una;
19188 th->th_seq = htonl(tp->snd_max);
19189 th->th_ack = htonl(tp->rcv_nxt);
19190 th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19191 if (th->th_win == 0) {
19192 tp->t_sndzerowin++;
19193 tp->t_flags |= TF_RXWIN0SENT;
19194 } else
19195 tp->t_flags &= ~TF_RXWIN0SENT;
19196 tp->snd_up = tp->snd_una; /* drag it along, its deprecated */
19197 KMOD_TCPSTAT_INC(tcps_sndpack);
19198 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19199 #ifdef STATS
19200 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19201 len);
19202 #endif
19203 if (rack->r_ctl.fsb.m == NULL)
19204 goto failed;
19205
19206 /* s_mb and s_soff are saved for rack_log_output */
19207 m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19208 &s_mb, &s_soff);
19209 if (len <= segsiz) {
19210 /*
19211 * Must have ran out of mbufs for the copy
19212 * shorten it to no longer need tso. Lets
19213 * not put on sendalot since we are low on
19214 * mbufs.
19215 */
19216 tso = 0;
19217 }
19218 if (rack->r_ctl.fsb.rfo_apply_push &&
19219 (len == rack->r_ctl.fsb.left_to_send)) {
19220 flags |= TH_PUSH;
19221 add_flag |= RACK_HAD_PUSH;
19222 }
19223 if ((m->m_next == NULL) || (len <= 0)){
19224 goto failed;
19225 }
19226 if (udp) {
19227 if (rack->r_is_v6)
19228 ulen = hdrlen + len - sizeof(struct ip6_hdr);
19229 else
19230 ulen = hdrlen + len - sizeof(struct ip);
19231 udp->uh_ulen = htons(ulen);
19232 }
19233 m->m_pkthdr.rcvif = (struct ifnet *)0;
19234 if (TCPS_HAVERCVDSYN(tp->t_state) &&
19235 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19236 int ect = tcp_ecn_output_established(tp, &flags, len, false);
19237 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19238 (tp->t_flags2 & TF2_ECN_SND_ECE))
19239 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19240 #ifdef INET6
19241 if (rack->r_is_v6) {
19242 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19243 ip6->ip6_flow |= htonl(ect << 20);
19244 }
19245 else
19246 #endif
19247 {
19248 #ifdef INET
19249 ip->ip_tos &= ~IPTOS_ECN_MASK;
19250 ip->ip_tos |= ect;
19251 #endif
19252 }
19253 }
19254 tcp_set_flags(th, flags);
19255 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
19256 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19257 if (to.to_flags & TOF_SIGNATURE) {
19258 /*
19259 * Calculate MD5 signature and put it into the place
19260 * determined before.
19261 * NOTE: since TCP options buffer doesn't point into
19262 * mbuf's data, calculate offset and use it.
19263 */
19264 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19265 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19266 /*
19267 * Do not send segment if the calculation of MD5
19268 * digest has failed.
19269 */
19270 goto failed;
19271 }
19272 }
19273 #endif
19274 #ifdef INET6
19275 if (rack->r_is_v6) {
19276 if (tp->t_port) {
19277 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19278 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19279 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19280 th->th_sum = htons(0);
19281 UDPSTAT_INC(udps_opackets);
19282 } else {
19283 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19284 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19285 th->th_sum = in6_cksum_pseudo(ip6,
19286 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19287 0);
19288 }
19289 }
19290 #endif
19291 #if defined(INET6) && defined(INET)
19292 else
19293 #endif
19294 #ifdef INET
19295 {
19296 if (tp->t_port) {
19297 m->m_pkthdr.csum_flags = CSUM_UDP;
19298 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19299 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19300 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19301 th->th_sum = htons(0);
19302 UDPSTAT_INC(udps_opackets);
19303 } else {
19304 m->m_pkthdr.csum_flags = CSUM_TCP;
19305 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19306 th->th_sum = in_pseudo(ip->ip_src.s_addr,
19307 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19308 IPPROTO_TCP + len + optlen));
19309 }
19310 /* IP version must be set here for ipv4/ipv6 checking later */
19311 KASSERT(ip->ip_v == IPVERSION,
19312 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
19313 }
19314 #endif
19315 if (tso) {
19316 /*
19317 * Here we use segsiz since we have no added options besides
19318 * any standard timestamp options (no DSACKs or SACKS are sent
19319 * via either fast-path).
19320 */
19321 KASSERT(len > segsiz,
19322 ("%s: len <= tso_segsz tp:%p", __func__, tp));
19323 m->m_pkthdr.csum_flags |= CSUM_TSO;
19324 m->m_pkthdr.tso_segsz = segsiz;
19325 }
19326 #ifdef INET6
19327 if (rack->r_is_v6) {
19328 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19329 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19330 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19331 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19332 else
19333 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19334 }
19335 #endif
19336 #if defined(INET) && defined(INET6)
19337 else
19338 #endif
19339 #ifdef INET
19340 {
19341 ip->ip_len = htons(m->m_pkthdr.len);
19342 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19343 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19344 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19345 if (tp->t_port == 0 || len < V_tcp_minmss) {
19346 ip->ip_off |= htons(IP_DF);
19347 }
19348 } else {
19349 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19350 }
19351 }
19352 #endif
19353 if (tp->snd_cwnd > tp->snd_ssthresh) {
19354 /* Set we sent in CA */
19355 rack->rc_gp_saw_ca = 1;
19356 } else {
19357 /* Set we sent in SS */
19358 rack->rc_gp_saw_ss = 1;
19359 }
19360 /* Time to copy in our header */
19361 cpto = mtod(m, uint8_t *);
19362 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19363 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19364 if (optlen) {
19365 bcopy(opt, th + 1, optlen);
19366 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19367 } else {
19368 th->th_off = sizeof(struct tcphdr) >> 2;
19369 }
19370 if ((rack->r_ctl.crte != NULL) &&
19371 tcp_bblogging_on(tp)) {
19372 rack_log_queue_level(tp, rack, len, tv, cts);
19373 }
19374 if (tcp_bblogging_on(rack->rc_tp)) {
19375 union tcp_log_stackspecific log;
19376
19377 memset(&log, 0, sizeof(log));
19378 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19379 if (rack->rack_no_prr)
19380 log.u_bbr.flex1 = 0;
19381 else
19382 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19383 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19384 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19385 log.u_bbr.flex4 = max_val;
19386 /* Save off the early/late values */
19387 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19388 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19389 log.u_bbr.bw_inuse = rack_get_bw(rack);
19390 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19391 log.u_bbr.flex8 = 0;
19392 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19393 log.u_bbr.flex7 = 44;
19394 log.u_bbr.pkts_out = tp->t_maxseg;
19395 log.u_bbr.timeStamp = cts;
19396 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19397 log.u_bbr.flex5 = log.u_bbr.inflight;
19398 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19399 log.u_bbr.delivered = rack->r_ctl.fsb.left_to_send;
19400 log.u_bbr.rttProp = 0;
19401 log.u_bbr.delRate = rack->r_must_retran;
19402 log.u_bbr.delRate <<= 1;
19403 log.u_bbr.pkt_epoch = line;
19404 /* For fast output no retrans so just inflight and how many mss we send */
19405 log.u_bbr.flex5 = log.u_bbr.inflight;
19406 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19407 lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19408 len, &log, false, NULL, __func__, __LINE__, tv);
19409 } else
19410 lgb = NULL;
19411 #ifdef INET6
19412 if (rack->r_is_v6) {
19413 error = ip6_output(m, inp->in6p_outputopts,
19414 &inp->inp_route6,
19415 0, NULL, NULL, inp);
19416 }
19417 #endif
19418 #if defined(INET) && defined(INET6)
19419 else
19420 #endif
19421 #ifdef INET
19422 {
19423 error = ip_output(m, NULL,
19424 &inp->inp_route,
19425 0, 0, inp);
19426 }
19427 #endif
19428 if (lgb) {
19429 lgb->tlb_errno = error;
19430 lgb = NULL;
19431 }
19432 if (error) {
19433 *send_err = error;
19434 m = NULL;
19435 goto failed;
19436 } else if (rack->rc_hw_nobuf) {
19437 rack->rc_hw_nobuf = 0;
19438 rack->r_ctl.rc_agg_delayed = 0;
19439 rack->r_early = 0;
19440 rack->r_late = 0;
19441 rack->r_ctl.rc_agg_early = 0;
19442 }
19443 if ((error == 0) && (rack->lt_bw_up == 0)) {
19444 /* Unlikely */
19445 rack->r_ctl.lt_timemark = tcp_tv_to_lusec(tv);
19446 rack->r_ctl.lt_seq = tp->snd_una;
19447 rack->lt_bw_up = 1;
19448 } else if ((error == 0) &&
19449 (((tp->snd_max + len) - rack->r_ctl.lt_seq) > 0x7fffffff)) {
19450 /*
19451 * Need to record what we have since we are
19452 * approaching seq wrap.
19453 */
19454 struct timeval tv;
19455 uint64_t tmark;
19456
19457 rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
19458 rack->r_ctl.lt_seq = tp->snd_una;
19459 tmark = tcp_get_u64_usecs(&tv);
19460 if (tmark > rack->r_ctl.lt_timemark) {
19461 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
19462 rack->r_ctl.lt_timemark = tmark;
19463 }
19464 }
19465 rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19466 NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19467 if (tp->snd_una == tp->snd_max) {
19468 rack->r_ctl.rc_tlp_rxt_last_time = cts;
19469 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19470 tp->t_acktime = ticks;
19471 }
19472 counter_u64_add(rack_total_bytes, len);
19473 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19474
19475 rack->forced_ack = 0; /* If we send something zap the FA flag */
19476 *tot_len += len;
19477 if ((tp->t_flags & TF_GPUTINPROG) == 0)
19478 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19479 tp->snd_max += len;
19480 tp->snd_nxt = tp->snd_max;
19481 if (rack->rc_new_rnd_needed) {
19482 rack_new_round_starts(tp, rack, tp->snd_max);
19483 }
19484 {
19485 int idx;
19486
19487 idx = (len / segsiz) + 3;
19488 if (idx >= TCP_MSS_ACCT_ATIMER)
19489 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19490 else
19491 counter_u64_add(rack_out_size[idx], 1);
19492 }
19493 if (len <= rack->r_ctl.fsb.left_to_send)
19494 rack->r_ctl.fsb.left_to_send -= len;
19495 else
19496 rack->r_ctl.fsb.left_to_send = 0;
19497 if (rack->r_ctl.fsb.left_to_send < segsiz) {
19498 rack->r_fast_output = 0;
19499 rack->r_ctl.fsb.left_to_send = 0;
19500 /* At the end of fast_output scale up the sb */
19501 SOCK_SENDBUF_LOCK(rack->rc_inp->inp_socket);
19502 rack_sndbuf_autoscale(rack);
19503 SOCK_SENDBUF_UNLOCK(rack->rc_inp->inp_socket);
19504 }
19505 if (tp->t_rtttime == 0) {
19506 tp->t_rtttime = ticks;
19507 tp->t_rtseq = startseq;
19508 KMOD_TCPSTAT_INC(tcps_segstimed);
19509 }
19510 if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19511 (max_val > len) &&
19512 (*tot_len < rack->r_ctl.rc_pace_max_segs) &&
19513 (tso == 0)) {
19514 max_val -= len;
19515 len = segsiz;
19516 th = rack->r_ctl.fsb.th;
19517 #ifdef TCP_ACCOUNTING
19518 cnt_thru++;
19519 #endif
19520 goto again;
19521 }
19522 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19523 counter_u64_add(rack_fto_send, 1);
19524 slot = rack_get_pacing_delay(rack, tp, *tot_len, NULL, segsiz, __LINE__);
19525 rack_start_hpts_timer(rack, tp, cts, slot, *tot_len, 0);
19526 #ifdef TCP_ACCOUNTING
19527 crtsc = get_cyclecount();
19528 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19529 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19530 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19531 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((*tot_len + segsiz - 1) / segsiz);
19532 }
19533 sched_unpin();
19534 #endif
19535 return (0);
19536 failed:
19537 if (m)
19538 m_free(m);
19539 rack->r_fast_output = 0;
19540 return (-1);
19541 }
19542
19543 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)19544 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19545 struct sockbuf *sb,
19546 int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19547 bool hw_tls,
19548 uint16_t flags)
19549 {
19550 rack->r_fast_output = 1;
19551 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19552 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19553 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19554 rack->r_ctl.fsb.tcp_flags = flags;
19555 rack->r_ctl.fsb.left_to_send = orig_len - len;
19556 if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19557 /* Less than a full sized pace, lets not */
19558 rack->r_fast_output = 0;
19559 return;
19560 } else {
19561 /* Round down to the nearest pace_max_seg */
19562 rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19563 }
19564 if (hw_tls)
19565 rack->r_ctl.fsb.hw_tls = 1;
19566 else
19567 rack->r_ctl.fsb.hw_tls = 0;
19568 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19569 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19570 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19571 (tp->snd_max - tp->snd_una)));
19572 if (rack->r_ctl.fsb.left_to_send < segsiz)
19573 rack->r_fast_output = 0;
19574 else {
19575 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19576 rack->r_ctl.fsb.rfo_apply_push = 1;
19577 else
19578 rack->r_ctl.fsb.rfo_apply_push = 0;
19579 }
19580 }
19581
19582 static uint32_t
rack_get_hpts_pacing_min_for_bw(struct tcp_rack * rack,int32_t segsiz)19583 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19584 {
19585 uint64_t min_time;
19586 uint32_t maxlen;
19587
19588 min_time = (uint64_t)get_hpts_min_sleep_time();
19589 maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19590 maxlen = roundup(maxlen, segsiz);
19591 return (maxlen);
19592 }
19593
19594 static struct rack_sendmap *
rack_check_collapsed(struct tcp_rack * rack,uint32_t cts)19595 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19596 {
19597 struct rack_sendmap *rsm = NULL;
19598 int thresh;
19599
19600 restart:
19601 rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19602 if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19603 /* Nothing, strange turn off validity */
19604 rack->r_collapse_point_valid = 0;
19605 return (NULL);
19606 }
19607 /* Can we send it yet? */
19608 if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19609 /*
19610 * Receiver window has not grown enough for
19611 * the segment to be put on the wire.
19612 */
19613 return (NULL);
19614 }
19615 if (rsm->r_flags & RACK_ACKED) {
19616 /*
19617 * It has been sacked, lets move to the
19618 * next one if possible.
19619 */
19620 rack->r_ctl.last_collapse_point = rsm->r_end;
19621 /* Are we done? */
19622 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19623 rack->r_ctl.high_collapse_point)) {
19624 rack->r_collapse_point_valid = 0;
19625 return (NULL);
19626 }
19627 goto restart;
19628 }
19629 /* Now has it been long enough ? */
19630 thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts, __LINE__, 1);
19631 if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19632 rack_log_collapse(rack, rsm->r_start,
19633 (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19634 thresh, __LINE__, 6, rsm->r_flags, rsm);
19635 return (rsm);
19636 }
19637 /* Not enough time */
19638 rack_log_collapse(rack, rsm->r_start,
19639 (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19640 thresh, __LINE__, 7, rsm->r_flags, rsm);
19641 return (NULL);
19642 }
19643
19644 static inline void
rack_validate_sizes(struct tcp_rack * rack,int32_t * len,int32_t segsiz,uint32_t pace_max_seg)19645 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19646 {
19647 if ((rack->full_size_rxt == 0) &&
19648 (rack->shape_rxt_to_pacing_min == 0) &&
19649 (*len >= segsiz)) {
19650 *len = segsiz;
19651 } else if (rack->shape_rxt_to_pacing_min &&
19652 rack->gp_ready) {
19653 /* We use pacing min as shaping len req */
19654 uint32_t maxlen;
19655
19656 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19657 if (*len > maxlen)
19658 *len = maxlen;
19659 } else {
19660 /*
19661 * The else is full_size_rxt is on so send it all
19662 * note we do need to check this for exceeding
19663 * our max segment size due to the fact that
19664 * we do sometimes merge chunks together i.e.
19665 * we cannot just assume that we will never have
19666 * a chunk greater than pace_max_seg
19667 */
19668 if (*len > pace_max_seg)
19669 *len = pace_max_seg;
19670 }
19671 }
19672
19673 static int
rack_output(struct tcpcb * tp)19674 rack_output(struct tcpcb *tp)
19675 {
19676 struct socket *so;
19677 uint32_t recwin;
19678 uint32_t sb_offset, s_moff = 0;
19679 int32_t len, error = 0;
19680 uint16_t flags;
19681 struct mbuf *m, *s_mb = NULL;
19682 struct mbuf *mb;
19683 uint32_t if_hw_tsomaxsegcount = 0;
19684 uint32_t if_hw_tsomaxsegsize;
19685 int32_t segsiz, minseg;
19686 long tot_len_this_send = 0;
19687 #ifdef INET
19688 struct ip *ip = NULL;
19689 #endif
19690 struct udphdr *udp = NULL;
19691 struct tcp_rack *rack;
19692 struct tcphdr *th;
19693 uint8_t pass = 0;
19694 uint8_t mark = 0;
19695 uint8_t check_done = 0;
19696 uint8_t wanted_cookie = 0;
19697 u_char opt[TCP_MAXOLEN];
19698 unsigned ipoptlen, optlen, hdrlen, ulen=0;
19699 uint32_t rack_seq;
19700
19701 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19702 unsigned ipsec_optlen = 0;
19703
19704 #endif
19705 int32_t idle, sendalot;
19706 uint32_t tot_idle;
19707 int32_t sub_from_prr = 0;
19708 volatile int32_t sack_rxmit;
19709 struct rack_sendmap *rsm = NULL;
19710 int32_t tso, mtu;
19711 struct tcpopt to;
19712 int32_t slot = 0;
19713 int32_t sup_rack = 0;
19714 uint32_t cts, ms_cts, delayed, early;
19715 uint32_t add_flag = RACK_SENT_SP;
19716 /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19717 uint8_t doing_tlp = 0;
19718 uint32_t cwnd_to_use, pace_max_seg;
19719 int32_t do_a_prefetch = 0;
19720 int32_t prefetch_rsm = 0;
19721 int32_t orig_len = 0;
19722 struct timeval tv;
19723 int32_t prefetch_so_done = 0;
19724 struct tcp_log_buffer *lgb;
19725 struct inpcb *inp = tptoinpcb(tp);
19726 struct sockbuf *sb;
19727 uint64_t ts_val = 0;
19728 #ifdef TCP_ACCOUNTING
19729 uint64_t crtsc;
19730 #endif
19731 #ifdef INET6
19732 struct ip6_hdr *ip6 = NULL;
19733 int32_t isipv6;
19734 #endif
19735 bool hpts_calling, hw_tls = false;
19736
19737 NET_EPOCH_ASSERT();
19738 INP_WLOCK_ASSERT(inp);
19739
19740 /* setup and take the cache hits here */
19741 rack = (struct tcp_rack *)tp->t_fb_ptr;
19742 #ifdef TCP_ACCOUNTING
19743 sched_pin();
19744 ts_val = get_cyclecount();
19745 #endif
19746 hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19747 tp->t_flags2 &= ~TF2_HPTS_CALLS;
19748 #ifdef TCP_OFFLOAD
19749 if (tp->t_flags & TF_TOE) {
19750 #ifdef TCP_ACCOUNTING
19751 sched_unpin();
19752 #endif
19753 return (tcp_offload_output(tp));
19754 }
19755 #endif
19756 if (rack->rack_deferred_inited == 0) {
19757 /*
19758 * If we are the connecting socket we will
19759 * hit rack_init() when no sequence numbers
19760 * are setup. This makes it so we must defer
19761 * some initialization. Call that now.
19762 */
19763 rack_deferred_init(tp, rack);
19764 }
19765 /*
19766 * For TFO connections in SYN_RECEIVED, only allow the initial
19767 * SYN|ACK and those sent by the retransmit timer.
19768 */
19769 if ((tp->t_flags & TF_FASTOPEN) &&
19770 (tp->t_state == TCPS_SYN_RECEIVED) &&
19771 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
19772 (rack->r_ctl.rc_resend == NULL)) { /* not a retransmit */
19773 #ifdef TCP_ACCOUNTING
19774 sched_unpin();
19775 #endif
19776 return (0);
19777 }
19778 #ifdef INET6
19779 if (rack->r_state) {
19780 /* Use the cache line loaded if possible */
19781 isipv6 = rack->r_is_v6;
19782 } else {
19783 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
19784 }
19785 #endif
19786 early = 0;
19787 cts = tcp_get_usecs(&tv);
19788 ms_cts = tcp_tv_to_msec(&tv);
19789 if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
19790 tcp_in_hpts(rack->rc_tp)) {
19791 /*
19792 * We are on the hpts for some timer but not hptsi output.
19793 * Remove from the hpts unconditionally.
19794 */
19795 rack_timer_cancel(tp, rack, cts, __LINE__);
19796 }
19797 /* Are we pacing and late? */
19798 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19799 TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
19800 /* We are delayed */
19801 delayed = cts - rack->r_ctl.rc_last_output_to;
19802 } else {
19803 delayed = 0;
19804 }
19805 /* Do the timers, which may override the pacer */
19806 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
19807 int retval;
19808
19809 retval = rack_process_timers(tp, rack, cts, hpts_calling,
19810 &doing_tlp);
19811 if (retval != 0) {
19812 counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
19813 #ifdef TCP_ACCOUNTING
19814 sched_unpin();
19815 #endif
19816 /*
19817 * If timers want tcp_drop(), then pass error out,
19818 * otherwise suppress it.
19819 */
19820 return (retval < 0 ? retval : 0);
19821 }
19822 }
19823 if (rack->rc_in_persist) {
19824 if (tcp_in_hpts(rack->rc_tp) == 0) {
19825 /* Timer is not running */
19826 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19827 }
19828 #ifdef TCP_ACCOUNTING
19829 sched_unpin();
19830 #endif
19831 return (0);
19832 }
19833 if ((rack->rc_ack_required == 1) &&
19834 (rack->r_timer_override == 0)){
19835 /* A timeout occurred and no ack has arrived */
19836 if (tcp_in_hpts(rack->rc_tp) == 0) {
19837 /* Timer is not running */
19838 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19839 }
19840 #ifdef TCP_ACCOUNTING
19841 sched_unpin();
19842 #endif
19843 return (0);
19844 }
19845 if ((rack->r_timer_override) ||
19846 (rack->rc_ack_can_sendout_data) ||
19847 (delayed) ||
19848 (tp->t_state < TCPS_ESTABLISHED)) {
19849 rack->rc_ack_can_sendout_data = 0;
19850 if (tcp_in_hpts(rack->rc_tp))
19851 tcp_hpts_remove(rack->rc_tp);
19852 } else if (tcp_in_hpts(rack->rc_tp)) {
19853 /*
19854 * On the hpts you can't pass even if ACKNOW is on, we will
19855 * when the hpts fires.
19856 */
19857 #ifdef TCP_ACCOUNTING
19858 crtsc = get_cyclecount();
19859 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19860 tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
19861 tp->tcp_cnt_counters[SND_BLOCKED]++;
19862 }
19863 sched_unpin();
19864 #endif
19865 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
19866 return (0);
19867 }
19868 /* Finish out both pacing early and late accounting */
19869 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
19870 TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
19871 early = rack->r_ctl.rc_last_output_to - cts;
19872 } else
19873 early = 0;
19874 if (delayed && (rack->rc_always_pace == 1)) {
19875 rack->r_ctl.rc_agg_delayed += delayed;
19876 rack->r_late = 1;
19877 } else if (early && (rack->rc_always_pace == 1)) {
19878 rack->r_ctl.rc_agg_early += early;
19879 rack->r_early = 1;
19880 } else if (rack->rc_always_pace == 0) {
19881 /* Non-paced we are not late */
19882 rack->r_ctl.rc_agg_delayed = rack->r_ctl.rc_agg_early = 0;
19883 rack->r_early = rack->r_late = 0;
19884 }
19885 /* Now that early/late accounting is done turn off the flag */
19886 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
19887 rack->r_wanted_output = 0;
19888 rack->r_timer_override = 0;
19889 if ((tp->t_state != rack->r_state) &&
19890 TCPS_HAVEESTABLISHED(tp->t_state)) {
19891 rack_set_state(tp, rack);
19892 }
19893 segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19894 minseg = segsiz;
19895 if (rack->r_ctl.rc_pace_max_segs == 0)
19896 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
19897 else
19898 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
19899 if ((rack->r_fast_output) &&
19900 (doing_tlp == 0) &&
19901 (tp->rcv_numsacks == 0)) {
19902 int ret;
19903
19904 error = 0;
19905 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, &tot_len_this_send, &error, __LINE__);
19906 if (ret > 0)
19907 return(ret);
19908 else if (error) {
19909 inp = rack->rc_inp;
19910 so = inp->inp_socket;
19911 sb = &so->so_snd;
19912 goto nomore;
19913 } else {
19914 /* Return == 0, if there is more we can send tot_len wise fall through and send */
19915 if (tot_len_this_send >= pace_max_seg)
19916 return (ret);
19917 #ifdef TCP_ACCOUNTING
19918 /* We need to re-pin since fast_output un-pined */
19919 sched_pin();
19920 ts_val = get_cyclecount();
19921 #endif
19922 /* Fall back out so we can send any more that may bring us to pace_max_seg */
19923 }
19924 }
19925 inp = rack->rc_inp;
19926 /*
19927 * For TFO connections in SYN_SENT or SYN_RECEIVED,
19928 * only allow the initial SYN or SYN|ACK and those sent
19929 * by the retransmit timer.
19930 */
19931 if ((tp->t_flags & TF_FASTOPEN) &&
19932 ((tp->t_state == TCPS_SYN_RECEIVED) ||
19933 (tp->t_state == TCPS_SYN_SENT)) &&
19934 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
19935 (tp->t_rxtshift == 0)) { /* not a retransmit */
19936 rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
19937 #ifdef TCP_ACCOUNTING
19938 sched_unpin();
19939 #endif
19940 return (0);
19941 }
19942 /*
19943 * Determine length of data that should be transmitted, and flags
19944 * that will be used. If there is some data or critical controls
19945 * (SYN, RST) to send, then transmit; otherwise, investigate
19946 * further.
19947 */
19948 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
19949 if (tp->t_idle_reduce) {
19950 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
19951 rack_cc_after_idle(rack, tp);
19952 }
19953 tp->t_flags &= ~TF_LASTIDLE;
19954 if (idle) {
19955 if (tp->t_flags & TF_MORETOCOME) {
19956 tp->t_flags |= TF_LASTIDLE;
19957 idle = 0;
19958 }
19959 }
19960 if ((tp->snd_una == tp->snd_max) &&
19961 rack->r_ctl.rc_went_idle_time &&
19962 (cts > rack->r_ctl.rc_went_idle_time)) {
19963 tot_idle = (cts - rack->r_ctl.rc_went_idle_time);
19964 if (tot_idle > rack_min_probertt_hold) {
19965 /* Count as a probe rtt */
19966 if (rack->in_probe_rtt == 0) {
19967 rack->r_ctl.rc_lower_rtt_us_cts = cts;
19968 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
19969 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
19970 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
19971 } else {
19972 rack_exit_probertt(rack, cts);
19973 }
19974 }
19975 } else
19976 tot_idle = 0;
19977 if (rack_use_fsb &&
19978 (rack->r_ctl.fsb.tcp_ip_hdr) &&
19979 (rack->r_fsb_inited == 0) &&
19980 (rack->r_state != TCPS_CLOSED))
19981 rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
19982 if (rack->rc_sendvars_notset == 1) {
19983 rack->rc_sendvars_notset = 0;
19984 /*
19985 * Make sure any TCP timers (keep-alive) is not running.
19986 */
19987 tcp_timer_stop(tp);
19988 }
19989 if ((rack->rack_no_prr == 1) &&
19990 (rack->rc_always_pace == 0)) {
19991 /*
19992 * Sanity check before sending, if we have
19993 * no-pacing enabled and prr is turned off that
19994 * is a logistics error. Correct this by turnning
19995 * prr back on. A user *must* set some form of
19996 * pacing in order to turn PRR off. We do this
19997 * in the output path so that we can avoid socket
19998 * option ordering issues that would occur if we
19999 * tried to do it while setting rack_no_prr on.
20000 */
20001 rack->rack_no_prr = 0;
20002 }
20003 if ((rack->pcm_enabled == 1) &&
20004 (rack->pcm_needed == 0) &&
20005 (tot_idle > 0)) {
20006 /*
20007 * We have been idle some micro seconds. We need
20008 * to factor this in to see if a PCM is needed.
20009 */
20010 uint32_t rtts_idle, rnds;
20011
20012 if (tp->t_srtt)
20013 rtts_idle = tot_idle / tp->t_srtt;
20014 else
20015 rtts_idle = 0;
20016 rnds = rack->r_ctl.current_round - rack->r_ctl.last_pcm_round;
20017 rack->r_ctl.pcm_idle_rounds += rtts_idle;
20018 if ((rnds + rack->r_ctl.pcm_idle_rounds) >= rack_pcm_every_n_rounds) {
20019 rack->pcm_needed = 1;
20020 rack_log_pcm(rack, 8, rack->r_ctl.last_pcm_round, rtts_idle, rack->r_ctl.current_round );
20021 }
20022 }
20023 again:
20024 sendalot = 0;
20025 cts = tcp_get_usecs(&tv);
20026 ms_cts = tcp_tv_to_msec(&tv);
20027 tso = 0;
20028 mtu = 0;
20029 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
20030 (rack->r_ctl.pcm_max_seg == 0)) {
20031 /*
20032 * We set in our first send so we know that the ctf_fixed_maxseg
20033 * has been fully set. If we do it in rack_init() we most likely
20034 * see 512 bytes so we end up at 5120, not desirable.
20035 */
20036 rack->r_ctl.pcm_max_seg = rc_init_window(rack);
20037 if (rack->r_ctl.pcm_max_seg < (ctf_fixed_maxseg(tp) * 10)) {
20038 /*
20039 * Assure our initial PCM probe is at least 10 MSS.
20040 */
20041 rack->r_ctl.pcm_max_seg = ctf_fixed_maxseg(tp) * 10;
20042 }
20043 }
20044 if ((rack->r_ctl.pcm_max_seg != 0) && (rack->pcm_needed == 1)) {
20045 uint32_t rw_avail, cwa;
20046
20047 if (tp->snd_wnd > ctf_outstanding(tp))
20048 rw_avail = tp->snd_wnd - ctf_outstanding(tp);
20049 else
20050 rw_avail = 0;
20051 if (tp->snd_cwnd > ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked))
20052 cwa = tp->snd_cwnd -ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
20053 else
20054 cwa = 0;
20055 if ((cwa >= rack->r_ctl.pcm_max_seg) &&
20056 (rw_avail > rack->r_ctl.pcm_max_seg)) {
20057 /* Raise up the max seg for this trip through */
20058 pace_max_seg = rack->r_ctl.pcm_max_seg;
20059 /* Disable any fast output */
20060 rack->r_fast_output = 0;
20061 }
20062 if (rack_verbose_logging) {
20063 rack_log_pcm(rack, 4,
20064 cwa, rack->r_ctl.pcm_max_seg, rw_avail);
20065 }
20066 }
20067 sb_offset = tp->snd_max - tp->snd_una;
20068 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20069 flags = tcp_outflags[tp->t_state];
20070 while (rack->rc_free_cnt < rack_free_cache) {
20071 rsm = rack_alloc(rack);
20072 if (rsm == NULL) {
20073 if (hpts_calling)
20074 /* Retry in a ms */
20075 slot = (1 * HPTS_USEC_IN_MSEC);
20076 so = inp->inp_socket;
20077 sb = &so->so_snd;
20078 goto just_return_nolock;
20079 }
20080 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20081 rack->rc_free_cnt++;
20082 rsm = NULL;
20083 }
20084 sack_rxmit = 0;
20085 len = 0;
20086 rsm = NULL;
20087 if (flags & TH_RST) {
20088 SOCK_SENDBUF_LOCK(inp->inp_socket);
20089 so = inp->inp_socket;
20090 sb = &so->so_snd;
20091 goto send;
20092 }
20093 if (rack->r_ctl.rc_resend) {
20094 /* Retransmit timer */
20095 rsm = rack->r_ctl.rc_resend;
20096 rack->r_ctl.rc_resend = NULL;
20097 len = rsm->r_end - rsm->r_start;
20098 sack_rxmit = 1;
20099 sendalot = 0;
20100 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20101 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20102 __func__, __LINE__,
20103 rsm->r_start, tp->snd_una, tp, rack, rsm));
20104 sb_offset = rsm->r_start - tp->snd_una;
20105 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20106 } else if (rack->r_collapse_point_valid &&
20107 ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20108 /*
20109 * If an RSM is returned then enough time has passed
20110 * for us to retransmit it. Move up the collapse point,
20111 * since this rsm has its chance to retransmit now.
20112 */
20113 tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20114 rack->r_ctl.last_collapse_point = rsm->r_end;
20115 /* Are we done? */
20116 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20117 rack->r_ctl.high_collapse_point))
20118 rack->r_collapse_point_valid = 0;
20119 sack_rxmit = 1;
20120 /* We are not doing a TLP */
20121 doing_tlp = 0;
20122 len = rsm->r_end - rsm->r_start;
20123 sb_offset = rsm->r_start - tp->snd_una;
20124 sendalot = 0;
20125 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20126 } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20127 /* We have a retransmit that takes precedence */
20128 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20129 ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20130 ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20131 /* Enter recovery if not induced by a time-out */
20132 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20133 }
20134 #ifdef INVARIANTS
20135 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20136 panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20137 tp, rack, rsm, rsm->r_start, tp->snd_una);
20138 }
20139 #endif
20140 len = rsm->r_end - rsm->r_start;
20141 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20142 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20143 __func__, __LINE__,
20144 rsm->r_start, tp->snd_una, tp, rack, rsm));
20145 sb_offset = rsm->r_start - tp->snd_una;
20146 sendalot = 0;
20147 rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20148 if (len > 0) {
20149 sack_rxmit = 1;
20150 KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20151 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20152 min(len, segsiz));
20153 }
20154 } else if (rack->r_ctl.rc_tlpsend) {
20155 /* Tail loss probe */
20156 long cwin;
20157 long tlen;
20158
20159 /*
20160 * Check if we can do a TLP with a RACK'd packet
20161 * this can happen if we are not doing the rack
20162 * cheat and we skipped to a TLP and it
20163 * went off.
20164 */
20165 rsm = rack->r_ctl.rc_tlpsend;
20166 /* We are doing a TLP make sure the flag is preent */
20167 rsm->r_flags |= RACK_TLP;
20168 rack->r_ctl.rc_tlpsend = NULL;
20169 sack_rxmit = 1;
20170 tlen = rsm->r_end - rsm->r_start;
20171 if (tlen > segsiz)
20172 tlen = segsiz;
20173 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20174 ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20175 __func__, __LINE__,
20176 rsm->r_start, tp->snd_una, tp, rack, rsm));
20177 sb_offset = rsm->r_start - tp->snd_una;
20178 cwin = min(tp->snd_wnd, tlen);
20179 len = cwin;
20180 }
20181 if (rack->r_must_retran &&
20182 (doing_tlp == 0) &&
20183 (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20184 (rsm == NULL)) {
20185 /*
20186 * There are two different ways that we
20187 * can get into this block:
20188 * a) This is a non-sack connection, we had a time-out
20189 * and thus r_must_retran was set and everything
20190 * left outstanding as been marked for retransmit.
20191 * b) The MTU of the path shrank, so that everything
20192 * was marked to be retransmitted with the smaller
20193 * mtu and r_must_retran was set.
20194 *
20195 * This means that we expect the sendmap (outstanding)
20196 * to all be marked must. We can use the tmap to
20197 * look at them.
20198 *
20199 */
20200 int sendwin, flight;
20201
20202 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20203 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20204 if (flight >= sendwin) {
20205 /*
20206 * We can't send yet.
20207 */
20208 so = inp->inp_socket;
20209 sb = &so->so_snd;
20210 goto just_return_nolock;
20211 }
20212 /*
20213 * This is the case a/b mentioned above. All
20214 * outstanding/not-acked should be marked.
20215 * We can use the tmap to find them.
20216 */
20217 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20218 if (rsm == NULL) {
20219 /* TSNH */
20220 rack->r_must_retran = 0;
20221 rack->r_ctl.rc_out_at_rto = 0;
20222 so = inp->inp_socket;
20223 sb = &so->so_snd;
20224 goto just_return_nolock;
20225 }
20226 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20227 /*
20228 * The first one does not have the flag, did we collapse
20229 * further up in our list?
20230 */
20231 rack->r_must_retran = 0;
20232 rack->r_ctl.rc_out_at_rto = 0;
20233 rsm = NULL;
20234 sack_rxmit = 0;
20235 } else {
20236 sack_rxmit = 1;
20237 len = rsm->r_end - rsm->r_start;
20238 sb_offset = rsm->r_start - tp->snd_una;
20239 sendalot = 0;
20240 if ((rack->full_size_rxt == 0) &&
20241 (rack->shape_rxt_to_pacing_min == 0) &&
20242 (len >= segsiz))
20243 len = segsiz;
20244 else if (rack->shape_rxt_to_pacing_min &&
20245 rack->gp_ready) {
20246 /* We use pacing min as shaping len req */
20247 uint32_t maxlen;
20248
20249 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20250 if (len > maxlen)
20251 len = maxlen;
20252 }
20253 /*
20254 * Delay removing the flag RACK_MUST_RXT so
20255 * that the fastpath for retransmit will
20256 * work with this rsm.
20257 */
20258 }
20259 }
20260 /*
20261 * Enforce a connection sendmap count limit if set
20262 * as long as we are not retransmiting.
20263 */
20264 if ((rsm == NULL) &&
20265 (V_tcp_map_entries_limit > 0) &&
20266 (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20267 counter_u64_add(rack_to_alloc_limited, 1);
20268 if (!rack->alloc_limit_reported) {
20269 rack->alloc_limit_reported = 1;
20270 counter_u64_add(rack_alloc_limited_conns, 1);
20271 }
20272 so = inp->inp_socket;
20273 sb = &so->so_snd;
20274 goto just_return_nolock;
20275 }
20276 if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20277 /* we are retransmitting the fin */
20278 len--;
20279 if (len) {
20280 /*
20281 * When retransmitting data do *not* include the
20282 * FIN. This could happen from a TLP probe.
20283 */
20284 flags &= ~TH_FIN;
20285 }
20286 }
20287 if (rsm && rack->r_fsb_inited &&
20288 rack_use_rsm_rfo &&
20289 ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20290 int ret;
20291
20292 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20293 if (ret == 0)
20294 return (0);
20295 }
20296 so = inp->inp_socket;
20297 sb = &so->so_snd;
20298 if (do_a_prefetch == 0) {
20299 kern_prefetch(sb, &do_a_prefetch);
20300 do_a_prefetch = 1;
20301 }
20302 #ifdef NETFLIX_SHARED_CWND
20303 if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20304 rack->rack_enable_scwnd) {
20305 /* We are doing cwnd sharing */
20306 if (rack->gp_ready &&
20307 (rack->rack_attempted_scwnd == 0) &&
20308 (rack->r_ctl.rc_scw == NULL) &&
20309 tp->t_lib) {
20310 /* The pcbid is in, lets make an attempt */
20311 counter_u64_add(rack_try_scwnd, 1);
20312 rack->rack_attempted_scwnd = 1;
20313 rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20314 &rack->r_ctl.rc_scw_index,
20315 segsiz);
20316 }
20317 if (rack->r_ctl.rc_scw &&
20318 (rack->rack_scwnd_is_idle == 1) &&
20319 sbavail(&so->so_snd)) {
20320 /* we are no longer out of data */
20321 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20322 rack->rack_scwnd_is_idle = 0;
20323 }
20324 if (rack->r_ctl.rc_scw) {
20325 /* First lets update and get the cwnd */
20326 rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20327 rack->r_ctl.rc_scw_index,
20328 tp->snd_cwnd, tp->snd_wnd, segsiz);
20329 }
20330 }
20331 #endif
20332 /*
20333 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20334 * state flags.
20335 */
20336 if (tp->t_flags & TF_NEEDFIN)
20337 flags |= TH_FIN;
20338 if (tp->t_flags & TF_NEEDSYN)
20339 flags |= TH_SYN;
20340 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20341 void *end_rsm;
20342 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20343 if (end_rsm)
20344 kern_prefetch(end_rsm, &prefetch_rsm);
20345 prefetch_rsm = 1;
20346 }
20347 SOCK_SENDBUF_LOCK(so);
20348 if ((sack_rxmit == 0) &&
20349 (TCPS_HAVEESTABLISHED(tp->t_state) ||
20350 (tp->t_flags & TF_FASTOPEN))) {
20351 /*
20352 * We are not retransmitting (sack_rxmit is 0) so we
20353 * are sending new data. This is always based on snd_max.
20354 * Now in theory snd_max may be equal to snd_una, if so
20355 * then nothing is outstanding and the offset would be 0.
20356 */
20357 uint32_t avail;
20358
20359 avail = sbavail(sb);
20360 if (SEQ_GT(tp->snd_max, tp->snd_una) && avail)
20361 sb_offset = tp->snd_max - tp->snd_una;
20362 else
20363 sb_offset = 0;
20364 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20365 if (rack->r_ctl.rc_tlp_new_data) {
20366 /* TLP is forcing out new data */
20367 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20368 rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20369 }
20370 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20371 if (tp->snd_wnd > sb_offset)
20372 len = tp->snd_wnd - sb_offset;
20373 else
20374 len = 0;
20375 } else {
20376 len = rack->r_ctl.rc_tlp_new_data;
20377 }
20378 rack->r_ctl.rc_tlp_new_data = 0;
20379 } else {
20380 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20381 }
20382 if ((rack->r_ctl.crte == NULL) &&
20383 IN_FASTRECOVERY(tp->t_flags) &&
20384 (rack->full_size_rxt == 0) &&
20385 (rack->shape_rxt_to_pacing_min == 0) &&
20386 (len > segsiz)) {
20387 /*
20388 * For prr=off, we need to send only 1 MSS
20389 * at a time. We do this because another sack could
20390 * be arriving that causes us to send retransmits and
20391 * we don't want to be on a long pace due to a larger send
20392 * that keeps us from sending out the retransmit.
20393 */
20394 len = segsiz;
20395 } else if (rack->shape_rxt_to_pacing_min &&
20396 rack->gp_ready) {
20397 /* We use pacing min as shaping len req */
20398 uint32_t maxlen;
20399
20400 maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20401 if (len > maxlen)
20402 len = maxlen;
20403 }/* The else is full_size_rxt is on so send it all */
20404 } else {
20405 uint32_t outstanding;
20406 /*
20407 * We are inside of a Fast recovery episode, this
20408 * is caused by a SACK or 3 dup acks. At this point
20409 * we have sent all the retransmissions and we rely
20410 * on PRR to dictate what we will send in the form of
20411 * new data.
20412 */
20413
20414 outstanding = tp->snd_max - tp->snd_una;
20415 if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20416 if (tp->snd_wnd > outstanding) {
20417 len = tp->snd_wnd - outstanding;
20418 /* Check to see if we have the data */
20419 if ((sb_offset + len) > avail) {
20420 /* It does not all fit */
20421 if (avail > sb_offset)
20422 len = avail - sb_offset;
20423 else
20424 len = 0;
20425 }
20426 } else {
20427 len = 0;
20428 }
20429 } else if (avail > sb_offset) {
20430 len = avail - sb_offset;
20431 } else {
20432 len = 0;
20433 }
20434 if (len > 0) {
20435 if (len > rack->r_ctl.rc_prr_sndcnt) {
20436 len = rack->r_ctl.rc_prr_sndcnt;
20437 }
20438 if (len > 0) {
20439 sub_from_prr = 1;
20440 }
20441 }
20442 if (len > segsiz) {
20443 /*
20444 * We should never send more than a MSS when
20445 * retransmitting or sending new data in prr
20446 * mode unless the override flag is on. Most
20447 * likely the PRR algorithm is not going to
20448 * let us send a lot as well :-)
20449 */
20450 if (rack->r_ctl.rc_prr_sendalot == 0) {
20451 len = segsiz;
20452 }
20453 } else if (len < segsiz) {
20454 /*
20455 * Do we send any? The idea here is if the
20456 * send empty's the socket buffer we want to
20457 * do it. However if not then lets just wait
20458 * for our prr_sndcnt to get bigger.
20459 */
20460 long leftinsb;
20461
20462 leftinsb = sbavail(sb) - sb_offset;
20463 if (leftinsb > len) {
20464 /* This send does not empty the sb */
20465 len = 0;
20466 }
20467 }
20468 }
20469 } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20470 /*
20471 * If you have not established
20472 * and are not doing FAST OPEN
20473 * no data please.
20474 */
20475 if ((sack_rxmit == 0) &&
20476 !(tp->t_flags & TF_FASTOPEN)) {
20477 len = 0;
20478 sb_offset = 0;
20479 }
20480 }
20481 if (prefetch_so_done == 0) {
20482 kern_prefetch(so, &prefetch_so_done);
20483 prefetch_so_done = 1;
20484 }
20485 orig_len = len;
20486 /*
20487 * Lop off SYN bit if it has already been sent. However, if this is
20488 * SYN-SENT state and if segment contains data and if we don't know
20489 * that foreign host supports TAO, suppress sending segment.
20490 */
20491 if ((flags & TH_SYN) &&
20492 SEQ_GT(tp->snd_max, tp->snd_una) &&
20493 ((sack_rxmit == 0) &&
20494 (tp->t_rxtshift == 0))) {
20495 /*
20496 * When sending additional segments following a TFO SYN|ACK,
20497 * do not include the SYN bit.
20498 */
20499 if ((tp->t_flags & TF_FASTOPEN) &&
20500 (tp->t_state == TCPS_SYN_RECEIVED))
20501 flags &= ~TH_SYN;
20502 }
20503 /*
20504 * Be careful not to send data and/or FIN on SYN segments. This
20505 * measure is needed to prevent interoperability problems with not
20506 * fully conformant TCP implementations.
20507 */
20508 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20509 len = 0;
20510 flags &= ~TH_FIN;
20511 }
20512 /*
20513 * On TFO sockets, ensure no data is sent in the following cases:
20514 *
20515 * - When retransmitting SYN|ACK on a passively-created socket
20516 *
20517 * - When retransmitting SYN on an actively created socket
20518 *
20519 * - When sending a zero-length cookie (cookie request) on an
20520 * actively created socket
20521 *
20522 * - When the socket is in the CLOSED state (RST is being sent)
20523 */
20524 if ((tp->t_flags & TF_FASTOPEN) &&
20525 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20526 ((tp->t_state == TCPS_SYN_SENT) &&
20527 (tp->t_tfo_client_cookie_len == 0)) ||
20528 (flags & TH_RST))) {
20529 sack_rxmit = 0;
20530 len = 0;
20531 }
20532 /* Without fast-open there should never be data sent on a SYN */
20533 if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) {
20534 len = 0;
20535 }
20536 if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20537 /* We only send 1 MSS if we have a DSACK block */
20538 add_flag |= RACK_SENT_W_DSACK;
20539 len = segsiz;
20540 }
20541 if (len <= 0) {
20542 /*
20543 * We have nothing to send, or the window shrank, or
20544 * is closed, do we need to go into persists?
20545 */
20546 len = 0;
20547 if ((tp->snd_wnd == 0) &&
20548 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20549 (tp->snd_una == tp->snd_max) &&
20550 (sb_offset < (int)sbavail(sb))) {
20551 rack_enter_persist(tp, rack, cts, tp->snd_una);
20552 }
20553 } else if ((rsm == NULL) &&
20554 (doing_tlp == 0) &&
20555 (len < pace_max_seg)) {
20556 /*
20557 * We are not sending a maximum sized segment for
20558 * some reason. Should we not send anything (think
20559 * sws or persists)?
20560 */
20561 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20562 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20563 (len < minseg) &&
20564 (len < (int)(sbavail(sb) - sb_offset))) {
20565 /*
20566 * Here the rwnd is less than
20567 * the minimum pacing size, this is not a retransmit,
20568 * we are established and
20569 * the send is not the last in the socket buffer
20570 * we send nothing, and we may enter persists
20571 * if nothing is outstanding.
20572 */
20573 len = 0;
20574 if (tp->snd_max == tp->snd_una) {
20575 /*
20576 * Nothing out we can
20577 * go into persists.
20578 */
20579 rack_enter_persist(tp, rack, cts, tp->snd_una);
20580 }
20581 } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20582 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20583 (len < (int)(sbavail(sb) - sb_offset)) &&
20584 (len < minseg)) {
20585 /*
20586 * Here we are not retransmitting, and
20587 * the cwnd is not so small that we could
20588 * not send at least a min size (rxt timer
20589 * not having gone off), We have 2 segments or
20590 * more already in flight, its not the tail end
20591 * of the socket buffer and the cwnd is blocking
20592 * us from sending out a minimum pacing segment size.
20593 * Lets not send anything.
20594 */
20595 len = 0;
20596 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20597 min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20598 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20599 (len < (int)(sbavail(sb) - sb_offset)) &&
20600 (TCPS_HAVEESTABLISHED(tp->t_state))) {
20601 /*
20602 * Here we have a send window but we have
20603 * filled it up and we can't send another pacing segment.
20604 * We also have in flight more than 2 segments
20605 * and we are not completing the sb i.e. we allow
20606 * the last bytes of the sb to go out even if
20607 * its not a full pacing segment.
20608 */
20609 len = 0;
20610 } else if ((rack->r_ctl.crte != NULL) &&
20611 (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20612 (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20613 (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20614 (len < (int)(sbavail(sb) - sb_offset))) {
20615 /*
20616 * Here we are doing hardware pacing, this is not a TLP,
20617 * we are not sending a pace max segment size, there is rwnd
20618 * room to send at least N pace_max_seg, the cwnd is greater
20619 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20620 * more segments in flight and its not the tail of the socket buffer.
20621 *
20622 * We don't want to send instead we need to get more ack's in to
20623 * allow us to send a full pacing segment. Normally, if we are pacing
20624 * about the right speed, we should have finished our pacing
20625 * send as most of the acks have come back if we are at the
20626 * right rate. This is a bit fuzzy since return path delay
20627 * can delay the acks, which is why we want to make sure we
20628 * have cwnd space to have a bit more than a max pace segments in flight.
20629 *
20630 * If we have not gotten our acks back we are pacing at too high a
20631 * rate delaying will not hurt and will bring our GP estimate down by
20632 * injecting the delay. If we don't do this we will send
20633 * 2 MSS out in response to the acks being clocked in which
20634 * defeats the point of hw-pacing (i.e. to help us get
20635 * larger TSO's out).
20636 */
20637 len = 0;
20638 }
20639
20640 }
20641 /* len will be >= 0 after this point. */
20642 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20643 rack_sndbuf_autoscale(rack);
20644 /*
20645 * Decide if we can use TCP Segmentation Offloading (if supported by
20646 * hardware).
20647 *
20648 * TSO may only be used if we are in a pure bulk sending state. The
20649 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20650 * options prevent using TSO. With TSO the TCP header is the same
20651 * (except for the sequence number) for all generated packets. This
20652 * makes it impossible to transmit any options which vary per
20653 * generated segment or packet.
20654 *
20655 * IPv4 handling has a clear separation of ip options and ip header
20656 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20657 * the right thing below to provide length of just ip options and thus
20658 * checking for ipoptlen is enough to decide if ip options are present.
20659 */
20660 ipoptlen = 0;
20661 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20662 /*
20663 * Pre-calculate here as we save another lookup into the darknesses
20664 * of IPsec that way and can actually decide if TSO is ok.
20665 */
20666 #ifdef INET6
20667 if (isipv6 && IPSEC_ENABLED(ipv6))
20668 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20669 #ifdef INET
20670 else
20671 #endif
20672 #endif /* INET6 */
20673 #ifdef INET
20674 if (IPSEC_ENABLED(ipv4))
20675 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20676 #endif /* INET */
20677 #endif
20678
20679 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20680 ipoptlen += ipsec_optlen;
20681 #endif
20682 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20683 (tp->t_port == 0) &&
20684 ((tp->t_flags & TF_SIGNATURE) == 0) &&
20685 sack_rxmit == 0 &&
20686 ipoptlen == 0)
20687 tso = 1;
20688 {
20689 uint32_t outstanding __unused;
20690
20691 outstanding = tp->snd_max - tp->snd_una;
20692 if (tp->t_flags & TF_SENTFIN) {
20693 /*
20694 * If we sent a fin, snd_max is 1 higher than
20695 * snd_una
20696 */
20697 outstanding--;
20698 }
20699 if (sack_rxmit) {
20700 if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20701 flags &= ~TH_FIN;
20702 }
20703 }
20704 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20705 (long)TCP_MAXWIN << tp->rcv_scale);
20706
20707 /*
20708 * Sender silly window avoidance. We transmit under the following
20709 * conditions when len is non-zero:
20710 *
20711 * - We have a full segment (or more with TSO) - This is the last
20712 * buffer in a write()/send() and we are either idle or running
20713 * NODELAY - we've timed out (e.g. persist timer) - we have more
20714 * then 1/2 the maximum send window's worth of data (receiver may be
20715 * limited the window size) - we need to retransmit
20716 */
20717 if (len) {
20718 if (len >= segsiz) {
20719 goto send;
20720 }
20721 /*
20722 * NOTE! on localhost connections an 'ack' from the remote
20723 * end may occur synchronously with the output and cause us
20724 * to flush a buffer queued with moretocome. XXX
20725 *
20726 */
20727 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
20728 (idle || (tp->t_flags & TF_NODELAY)) &&
20729 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20730 (tp->t_flags & TF_NOPUSH) == 0) {
20731 pass = 2;
20732 goto send;
20733 }
20734 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */
20735 pass = 22;
20736 goto send;
20737 }
20738 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20739 pass = 4;
20740 goto send;
20741 }
20742 if (sack_rxmit) {
20743 pass = 6;
20744 goto send;
20745 }
20746 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20747 (ctf_outstanding(tp) < (segsiz * 2))) {
20748 /*
20749 * We have less than two MSS outstanding (delayed ack)
20750 * and our rwnd will not let us send a full sized
20751 * MSS. Lets go ahead and let this small segment
20752 * out because we want to try to have at least two
20753 * packets inflight to not be caught by delayed ack.
20754 */
20755 pass = 12;
20756 goto send;
20757 }
20758 }
20759 /*
20760 * Sending of standalone window updates.
20761 *
20762 * Window updates are important when we close our window due to a
20763 * full socket buffer and are opening it again after the application
20764 * reads data from it. Once the window has opened again and the
20765 * remote end starts to send again the ACK clock takes over and
20766 * provides the most current window information.
20767 *
20768 * We must avoid the silly window syndrome whereas every read from
20769 * the receive buffer, no matter how small, causes a window update
20770 * to be sent. We also should avoid sending a flurry of window
20771 * updates when the socket buffer had queued a lot of data and the
20772 * application is doing small reads.
20773 *
20774 * Prevent a flurry of pointless window updates by only sending an
20775 * update when we can increase the advertized window by more than
20776 * 1/4th of the socket buffer capacity. When the buffer is getting
20777 * full or is very small be more aggressive and send an update
20778 * whenever we can increase by two mss sized segments. In all other
20779 * situations the ACK's to new incoming data will carry further
20780 * window increases.
20781 *
20782 * Don't send an independent window update if a delayed ACK is
20783 * pending (it will get piggy-backed on it) or the remote side
20784 * already has done a half-close and won't send more data. Skip
20785 * this if the connection is in T/TCP half-open state.
20786 */
20787 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20788 !(tp->t_flags & TF_DELACK) &&
20789 !TCPS_HAVERCVDFIN(tp->t_state)) {
20790 /*
20791 * "adv" is the amount we could increase the window, taking
20792 * into account that we are limited by TCP_MAXWIN <<
20793 * tp->rcv_scale.
20794 */
20795 int32_t adv;
20796 int oldwin;
20797
20798 adv = recwin;
20799 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20800 oldwin = (tp->rcv_adv - tp->rcv_nxt);
20801 if (adv > oldwin)
20802 adv -= oldwin;
20803 else {
20804 /* We can't increase the window */
20805 adv = 0;
20806 }
20807 } else
20808 oldwin = 0;
20809
20810 /*
20811 * If the new window size ends up being the same as or less
20812 * than the old size when it is scaled, then don't force
20813 * a window update.
20814 */
20815 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20816 goto dontupdate;
20817
20818 if (adv >= (int32_t)(2 * segsiz) &&
20819 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20820 recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20821 so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20822 pass = 7;
20823 goto send;
20824 }
20825 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20826 pass = 23;
20827 goto send;
20828 }
20829 }
20830 dontupdate:
20831
20832 /*
20833 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
20834 * is also a catch-all for the retransmit timer timeout case.
20835 */
20836 if (tp->t_flags & TF_ACKNOW) {
20837 pass = 8;
20838 goto send;
20839 }
20840 if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
20841 pass = 9;
20842 goto send;
20843 }
20844 /*
20845 * If our state indicates that FIN should be sent and we have not
20846 * yet done so, then we need to send.
20847 */
20848 if ((flags & TH_FIN) &&
20849 (tp->snd_max == tp->snd_una)) {
20850 pass = 11;
20851 goto send;
20852 }
20853 /*
20854 * No reason to send a segment, just return.
20855 */
20856 just_return:
20857 SOCK_SENDBUF_UNLOCK(so);
20858 just_return_nolock:
20859 {
20860 int app_limited = CTF_JR_SENT_DATA;
20861
20862 if ((tp->t_flags & TF_FASTOPEN) == 0 &&
20863 (flags & TH_FIN) &&
20864 (len == 0) &&
20865 (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
20866 ((tp->snd_max - tp->snd_una) <= segsiz)) {
20867 /*
20868 * Ok less than or right at a MSS is
20869 * outstanding. The original FreeBSD stack would
20870 * have sent a FIN, which can speed things up for
20871 * a transactional application doing a MSG_WAITALL.
20872 * To speed things up since we do *not* send a FIN
20873 * if data is outstanding, we send a "challenge ack".
20874 * The idea behind that is instead of having to have
20875 * the peer wait for the delayed-ack timer to run off
20876 * we send an ack that makes the peer send us an ack.
20877 */
20878 rack_send_ack_challange(rack);
20879 }
20880 if (tot_len_this_send > 0) {
20881 rack->r_ctl.fsb.recwin = recwin;
20882 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz, __LINE__);
20883 if ((error == 0) &&
20884 rack_use_rfo &&
20885 ((flags & (TH_SYN|TH_FIN)) == 0) &&
20886 (ipoptlen == 0) &&
20887 rack->r_fsb_inited &&
20888 TCPS_HAVEESTABLISHED(tp->t_state) &&
20889 ((IN_RECOVERY(tp->t_flags)) == 0) &&
20890 (doing_tlp == 0) &&
20891 (rack->r_must_retran == 0) &&
20892 ((tp->t_flags & TF_NEEDFIN) == 0) &&
20893 (len > 0) && (orig_len > 0) &&
20894 (orig_len > len) &&
20895 ((orig_len - len) >= segsiz) &&
20896 ((optlen == 0) ||
20897 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
20898 /* We can send at least one more MSS using our fsb */
20899 rack_setup_fast_output(tp, rack, sb, len, orig_len,
20900 segsiz, pace_max_seg, hw_tls, flags);
20901 } else
20902 rack->r_fast_output = 0;
20903 rack_log_fsb(rack, tp, so, flags,
20904 ipoptlen, orig_len, len, 0,
20905 1, optlen, __LINE__, 1);
20906 /* Assure when we leave that snd_nxt will point to top */
20907 if (SEQ_GT(tp->snd_max, tp->snd_nxt))
20908 tp->snd_nxt = tp->snd_max;
20909 } else {
20910 int end_window = 0;
20911 uint32_t seq = tp->gput_ack;
20912
20913 rsm = tqhash_max(rack->r_ctl.tqh);
20914 if (rsm) {
20915 /*
20916 * Mark the last sent that we just-returned (hinting
20917 * that delayed ack may play a role in any rtt measurement).
20918 */
20919 rsm->r_just_ret = 1;
20920 }
20921 counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
20922 rack->r_ctl.rc_agg_delayed = 0;
20923 rack->r_early = 0;
20924 rack->r_late = 0;
20925 rack->r_ctl.rc_agg_early = 0;
20926 if ((ctf_outstanding(tp) +
20927 min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
20928 minseg)) >= tp->snd_wnd) {
20929 /* We are limited by the rwnd */
20930 app_limited = CTF_JR_RWND_LIMITED;
20931 if (IN_FASTRECOVERY(tp->t_flags))
20932 rack->r_ctl.rc_prr_sndcnt = 0;
20933 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
20934 /* We are limited by whats available -- app limited */
20935 app_limited = CTF_JR_APP_LIMITED;
20936 if (IN_FASTRECOVERY(tp->t_flags))
20937 rack->r_ctl.rc_prr_sndcnt = 0;
20938 } else if ((idle == 0) &&
20939 ((tp->t_flags & TF_NODELAY) == 0) &&
20940 ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20941 (len < segsiz)) {
20942 /*
20943 * No delay is not on and the
20944 * user is sending less than 1MSS. This
20945 * brings out SWS avoidance so we
20946 * don't send. Another app-limited case.
20947 */
20948 app_limited = CTF_JR_APP_LIMITED;
20949 } else if (tp->t_flags & TF_NOPUSH) {
20950 /*
20951 * The user has requested no push of
20952 * the last segment and we are
20953 * at the last segment. Another app
20954 * limited case.
20955 */
20956 app_limited = CTF_JR_APP_LIMITED;
20957 } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
20958 /* Its the cwnd */
20959 app_limited = CTF_JR_CWND_LIMITED;
20960 } else if (IN_FASTRECOVERY(tp->t_flags) &&
20961 (rack->rack_no_prr == 0) &&
20962 (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
20963 app_limited = CTF_JR_PRR;
20964 } else {
20965 /* Now why here are we not sending? */
20966 #ifdef NOW
20967 #ifdef INVARIANTS
20968 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
20969 #endif
20970 #endif
20971 app_limited = CTF_JR_ASSESSING;
20972 }
20973 /*
20974 * App limited in some fashion, for our pacing GP
20975 * measurements we don't want any gap (even cwnd).
20976 * Close down the measurement window.
20977 */
20978 if (rack_cwnd_block_ends_measure &&
20979 ((app_limited == CTF_JR_CWND_LIMITED) ||
20980 (app_limited == CTF_JR_PRR))) {
20981 /*
20982 * The reason we are not sending is
20983 * the cwnd (or prr). We have been configured
20984 * to end the measurement window in
20985 * this case.
20986 */
20987 end_window = 1;
20988 } else if (rack_rwnd_block_ends_measure &&
20989 (app_limited == CTF_JR_RWND_LIMITED)) {
20990 /*
20991 * We are rwnd limited and have been
20992 * configured to end the measurement
20993 * window in this case.
20994 */
20995 end_window = 1;
20996 } else if (app_limited == CTF_JR_APP_LIMITED) {
20997 /*
20998 * A true application limited period, we have
20999 * ran out of data.
21000 */
21001 end_window = 1;
21002 } else if (app_limited == CTF_JR_ASSESSING) {
21003 /*
21004 * In the assessing case we hit the end of
21005 * the if/else and had no known reason
21006 * This will panic us under invariants..
21007 *
21008 * If we get this out in logs we need to
21009 * investagate which reason we missed.
21010 */
21011 end_window = 1;
21012 }
21013 if (end_window) {
21014 uint8_t log = 0;
21015
21016 /* Adjust the Gput measurement */
21017 if ((tp->t_flags & TF_GPUTINPROG) &&
21018 SEQ_GT(tp->gput_ack, tp->snd_max)) {
21019 tp->gput_ack = tp->snd_max;
21020 if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21021 /*
21022 * There is not enough to measure.
21023 */
21024 tp->t_flags &= ~TF_GPUTINPROG;
21025 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21026 rack->r_ctl.rc_gp_srtt /*flex1*/,
21027 tp->gput_seq,
21028 0, 0, 18, __LINE__, NULL, 0);
21029 } else
21030 log = 1;
21031 }
21032 /* Mark the last packet as app limited */
21033 rsm = tqhash_max(rack->r_ctl.tqh);
21034 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21035 if (rack->r_ctl.rc_app_limited_cnt == 0)
21036 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21037 else {
21038 /*
21039 * Go out to the end app limited and mark
21040 * this new one as next and move the end_appl up
21041 * to this guy.
21042 */
21043 if (rack->r_ctl.rc_end_appl)
21044 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21045 rack->r_ctl.rc_end_appl = rsm;
21046 }
21047 rsm->r_flags |= RACK_APP_LIMITED;
21048 rack->r_ctl.rc_app_limited_cnt++;
21049 }
21050 if (log)
21051 rack_log_pacing_delay_calc(rack,
21052 rack->r_ctl.rc_app_limited_cnt, seq,
21053 tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21054 }
21055 }
21056 /* Check if we need to go into persists or not */
21057 if ((tp->snd_max == tp->snd_una) &&
21058 TCPS_HAVEESTABLISHED(tp->t_state) &&
21059 sbavail(sb) &&
21060 (sbavail(sb) > tp->snd_wnd) &&
21061 (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21062 /* Yes lets make sure to move to persist before timer-start */
21063 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21064 }
21065 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21066 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21067 }
21068 #ifdef NETFLIX_SHARED_CWND
21069 if ((sbavail(sb) == 0) &&
21070 rack->r_ctl.rc_scw) {
21071 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21072 rack->rack_scwnd_is_idle = 1;
21073 }
21074 #endif
21075 #ifdef TCP_ACCOUNTING
21076 if (tot_len_this_send > 0) {
21077 crtsc = get_cyclecount();
21078 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21079 tp->tcp_cnt_counters[SND_OUT_DATA]++;
21080 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21081 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21082 }
21083 } else {
21084 crtsc = get_cyclecount();
21085 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21086 tp->tcp_cnt_counters[SND_LIMITED]++;
21087 tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21088 }
21089 }
21090 sched_unpin();
21091 #endif
21092 return (0);
21093
21094 send:
21095 if ((rack->r_ctl.crte != NULL) &&
21096 (rsm == NULL) &&
21097 ((rack->rc_hw_nobuf == 1) ||
21098 (rack_hw_check_queue && (check_done == 0)))) {
21099 /*
21100 * We only want to do this once with the hw_check_queue,
21101 * for the enobuf case we would only do it once if
21102 * we come around to again, the flag will be clear.
21103 */
21104 check_done = 1;
21105 slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21106 if (slot) {
21107 rack->r_ctl.rc_agg_delayed = 0;
21108 rack->r_ctl.rc_agg_early = 0;
21109 rack->r_early = 0;
21110 rack->r_late = 0;
21111 SOCK_SENDBUF_UNLOCK(so);
21112 goto skip_all_send;
21113 }
21114 }
21115 if (rsm || sack_rxmit)
21116 counter_u64_add(rack_nfto_resend, 1);
21117 else
21118 counter_u64_add(rack_non_fto_send, 1);
21119 if ((flags & TH_FIN) &&
21120 sbavail(sb)) {
21121 /*
21122 * We do not transmit a FIN
21123 * with data outstanding. We
21124 * need to make it so all data
21125 * is acked first.
21126 */
21127 flags &= ~TH_FIN;
21128 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21129 (sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21130 ((tp->snd_max - tp->snd_una) <= segsiz)) {
21131 /*
21132 * Ok less than or right at a MSS is
21133 * outstanding. The original FreeBSD stack would
21134 * have sent a FIN, which can speed things up for
21135 * a transactional application doing a MSG_WAITALL.
21136 * To speed things up since we do *not* send a FIN
21137 * if data is outstanding, we send a "challenge ack".
21138 * The idea behind that is instead of having to have
21139 * the peer wait for the delayed-ack timer to run off
21140 * we send an ack that makes the peer send us an ack.
21141 */
21142 rack_send_ack_challange(rack);
21143 }
21144 }
21145 /* Enforce stack imposed max seg size if we have one */
21146 if (pace_max_seg &&
21147 (len > pace_max_seg)) {
21148 mark = 1;
21149 len = pace_max_seg;
21150 }
21151 if ((rsm == NULL) &&
21152 (rack->pcm_in_progress == 0) &&
21153 (rack->r_ctl.pcm_max_seg > 0) &&
21154 (len >= rack->r_ctl.pcm_max_seg)) {
21155 /* It is large enough for a measurement */
21156 add_flag |= RACK_IS_PCM;
21157 rack_log_pcm(rack, 5, len, rack->r_ctl.pcm_max_seg, add_flag);
21158 } else if (rack_verbose_logging) {
21159 rack_log_pcm(rack, 6, len, rack->r_ctl.pcm_max_seg, add_flag);
21160 }
21161
21162 SOCKBUF_LOCK_ASSERT(sb);
21163 if (len > 0) {
21164 if (len >= segsiz)
21165 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21166 else
21167 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21168 }
21169 /*
21170 * Before ESTABLISHED, force sending of initial options unless TCP
21171 * set not to do any options. NOTE: we assume that the IP/TCP header
21172 * plus TCP options always fit in a single mbuf, leaving room for a
21173 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21174 * + optlen <= MCLBYTES
21175 */
21176 optlen = 0;
21177 #ifdef INET6
21178 if (isipv6)
21179 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21180 else
21181 #endif
21182 hdrlen = sizeof(struct tcpiphdr);
21183
21184 /*
21185 * Ok what seq are we sending from. If we have
21186 * no rsm to use, then we look at various bits,
21187 * if we are putting out a SYN it will be ISS.
21188 * If we are retransmitting a FIN it will
21189 * be snd_max-1 else its snd_max.
21190 */
21191 if (rsm == NULL) {
21192 if (flags & TH_SYN)
21193 rack_seq = tp->iss;
21194 else if ((flags & TH_FIN) &&
21195 (tp->t_flags & TF_SENTFIN))
21196 rack_seq = tp->snd_max - 1;
21197 else
21198 rack_seq = tp->snd_max;
21199 } else {
21200 rack_seq = rsm->r_start;
21201 }
21202 /*
21203 * Compute options for segment. We only have to care about SYN and
21204 * established connection segments. Options for SYN-ACK segments
21205 * are handled in TCP syncache.
21206 */
21207 to.to_flags = 0;
21208 if ((tp->t_flags & TF_NOOPT) == 0) {
21209 /* Maximum segment size. */
21210 if (flags & TH_SYN) {
21211 to.to_mss = tcp_mssopt(&inp->inp_inc);
21212 if (tp->t_port)
21213 to.to_mss -= V_tcp_udp_tunneling_overhead;
21214 to.to_flags |= TOF_MSS;
21215
21216 /*
21217 * On SYN or SYN|ACK transmits on TFO connections,
21218 * only include the TFO option if it is not a
21219 * retransmit, as the presence of the TFO option may
21220 * have caused the original SYN or SYN|ACK to have
21221 * been dropped by a middlebox.
21222 */
21223 if ((tp->t_flags & TF_FASTOPEN) &&
21224 (tp->t_rxtshift == 0)) {
21225 if (tp->t_state == TCPS_SYN_RECEIVED) {
21226 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21227 to.to_tfo_cookie =
21228 (u_int8_t *)&tp->t_tfo_cookie.server;
21229 to.to_flags |= TOF_FASTOPEN;
21230 wanted_cookie = 1;
21231 } else if (tp->t_state == TCPS_SYN_SENT) {
21232 to.to_tfo_len =
21233 tp->t_tfo_client_cookie_len;
21234 to.to_tfo_cookie =
21235 tp->t_tfo_cookie.client;
21236 to.to_flags |= TOF_FASTOPEN;
21237 wanted_cookie = 1;
21238 /*
21239 * If we wind up having more data to
21240 * send with the SYN than can fit in
21241 * one segment, don't send any more
21242 * until the SYN|ACK comes back from
21243 * the other end.
21244 */
21245 sendalot = 0;
21246 }
21247 }
21248 }
21249 /* Window scaling. */
21250 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21251 to.to_wscale = tp->request_r_scale;
21252 to.to_flags |= TOF_SCALE;
21253 }
21254 /* Timestamps. */
21255 if ((tp->t_flags & TF_RCVD_TSTMP) ||
21256 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21257 uint32_t ts_to_use;
21258
21259 if ((rack->r_rcvpath_rtt_up == 1) &&
21260 (ms_cts == rack->r_ctl.last_rcv_tstmp_for_rtt)) {
21261 /*
21262 * When we are doing a rcv_rtt probe all
21263 * other timestamps use the next msec. This
21264 * is safe since our previous ack is in the
21265 * air and we will just have a few more
21266 * on the next ms. This assures that only
21267 * the one ack has the ms_cts that was on
21268 * our ack-probe.
21269 */
21270 ts_to_use = ms_cts + 1;
21271 } else {
21272 ts_to_use = ms_cts;
21273 }
21274 to.to_tsval = ts_to_use + tp->ts_offset;
21275 to.to_tsecr = tp->ts_recent;
21276 to.to_flags |= TOF_TS;
21277 if ((len == 0) &&
21278 (TCPS_HAVEESTABLISHED(tp->t_state)) &&
21279 ((ms_cts - rack->r_ctl.last_rcv_tstmp_for_rtt) > RCV_PATH_RTT_MS) &&
21280 (tp->snd_una == tp->snd_max) &&
21281 (flags & TH_ACK) &&
21282 (sbavail(sb) == 0) &&
21283 (rack->r_ctl.current_round != 0) &&
21284 ((flags & (TH_SYN|TH_FIN)) == 0) &&
21285 (rack->r_rcvpath_rtt_up == 0)) {
21286 rack->r_ctl.last_rcv_tstmp_for_rtt = ms_cts;
21287 rack->r_ctl.last_time_of_arm_rcv = cts;
21288 rack->r_rcvpath_rtt_up = 1;
21289 /* Subtract 1 from seq to force a response */
21290 rack_seq--;
21291 }
21292 }
21293 /* Set receive buffer autosizing timestamp. */
21294 if (tp->rfbuf_ts == 0 &&
21295 (so->so_rcv.sb_flags & SB_AUTOSIZE)) {
21296 tp->rfbuf_ts = ms_cts;
21297 }
21298 /* Selective ACK's. */
21299 if (tp->t_flags & TF_SACK_PERMIT) {
21300 if (flags & TH_SYN)
21301 to.to_flags |= TOF_SACKPERM;
21302 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21303 tp->rcv_numsacks > 0) {
21304 to.to_flags |= TOF_SACK;
21305 to.to_nsacks = tp->rcv_numsacks;
21306 to.to_sacks = (u_char *)tp->sackblks;
21307 }
21308 }
21309 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21310 /* TCP-MD5 (RFC2385). */
21311 if (tp->t_flags & TF_SIGNATURE)
21312 to.to_flags |= TOF_SIGNATURE;
21313 #endif
21314
21315 /* Processing the options. */
21316 hdrlen += optlen = tcp_addoptions(&to, opt);
21317 /*
21318 * If we wanted a TFO option to be added, but it was unable
21319 * to fit, ensure no data is sent.
21320 */
21321 if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie &&
21322 !(to.to_flags & TOF_FASTOPEN))
21323 len = 0;
21324 }
21325 if (tp->t_port) {
21326 if (V_tcp_udp_tunneling_port == 0) {
21327 /* The port was removed?? */
21328 SOCK_SENDBUF_UNLOCK(so);
21329 #ifdef TCP_ACCOUNTING
21330 crtsc = get_cyclecount();
21331 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21332 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21333 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21334 }
21335 sched_unpin();
21336 #endif
21337 return (EHOSTUNREACH);
21338 }
21339 hdrlen += sizeof(struct udphdr);
21340 }
21341 #ifdef INET6
21342 if (isipv6)
21343 ipoptlen = ip6_optlen(inp);
21344 else
21345 #endif
21346 if (inp->inp_options)
21347 ipoptlen = inp->inp_options->m_len -
21348 offsetof(struct ipoption, ipopt_list);
21349 else
21350 ipoptlen = 0;
21351 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21352 ipoptlen += ipsec_optlen;
21353 #endif
21354
21355 /*
21356 * Adjust data length if insertion of options will bump the packet
21357 * length beyond the t_maxseg length. Clear the FIN bit because we
21358 * cut off the tail of the segment.
21359 */
21360 if (len + optlen + ipoptlen > tp->t_maxseg) {
21361 if (tso) {
21362 uint32_t if_hw_tsomax;
21363 uint32_t moff;
21364 int32_t max_len;
21365
21366 /* extract TSO information */
21367 if_hw_tsomax = tp->t_tsomax;
21368 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21369 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21370 KASSERT(ipoptlen == 0,
21371 ("%s: TSO can't do IP options", __func__));
21372
21373 /*
21374 * Check if we should limit by maximum payload
21375 * length:
21376 */
21377 if (if_hw_tsomax != 0) {
21378 /* compute maximum TSO length */
21379 max_len = (if_hw_tsomax - hdrlen -
21380 max_linkhdr);
21381 if (max_len <= 0) {
21382 len = 0;
21383 } else if (len > max_len) {
21384 if (doing_tlp == 0)
21385 sendalot = 1;
21386 len = max_len;
21387 mark = 2;
21388 }
21389 }
21390 /*
21391 * Prevent the last segment from being fractional
21392 * unless the send sockbuf can be emptied:
21393 */
21394 max_len = (tp->t_maxseg - optlen);
21395 if ((sb_offset + len) < sbavail(sb)) {
21396 moff = len % (u_int)max_len;
21397 if (moff != 0) {
21398 mark = 3;
21399 len -= moff;
21400 }
21401 }
21402 /*
21403 * In case there are too many small fragments don't
21404 * use TSO:
21405 */
21406 if (len <= max_len) {
21407 mark = 4;
21408 tso = 0;
21409 }
21410 /*
21411 * Send the FIN in a separate segment after the bulk
21412 * sending is done. We don't trust the TSO
21413 * implementations to clear the FIN flag on all but
21414 * the last segment.
21415 */
21416 if (tp->t_flags & TF_NEEDFIN) {
21417 sendalot = 4;
21418 }
21419 } else {
21420 mark = 5;
21421 if (optlen + ipoptlen >= tp->t_maxseg) {
21422 /*
21423 * Since we don't have enough space to put
21424 * the IP header chain and the TCP header in
21425 * one packet as required by RFC 7112, don't
21426 * send it. Also ensure that at least one
21427 * byte of the payload can be put into the
21428 * TCP segment.
21429 */
21430 SOCK_SENDBUF_UNLOCK(so);
21431 error = EMSGSIZE;
21432 sack_rxmit = 0;
21433 goto out;
21434 }
21435 len = tp->t_maxseg - optlen - ipoptlen;
21436 sendalot = 5;
21437 }
21438 } else {
21439 tso = 0;
21440 mark = 6;
21441 }
21442 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21443 ("%s: len > IP_MAXPACKET", __func__));
21444 #ifdef DIAGNOSTIC
21445 #ifdef INET6
21446 if (max_linkhdr + hdrlen > MCLBYTES)
21447 #else
21448 if (max_linkhdr + hdrlen > MHLEN)
21449 #endif
21450 panic("tcphdr too big");
21451 #endif
21452
21453 /*
21454 * This KASSERT is here to catch edge cases at a well defined place.
21455 * Before, those had triggered (random) panic conditions further
21456 * down.
21457 */
21458 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21459 if ((len == 0) &&
21460 (flags & TH_FIN) &&
21461 (sbused(sb))) {
21462 /*
21463 * We have outstanding data, don't send a fin by itself!.
21464 *
21465 * Check to see if we need to send a challenge ack.
21466 */
21467 if ((sbused(sb) == (tp->snd_max - tp->snd_una)) &&
21468 ((tp->snd_max - tp->snd_una) <= segsiz)) {
21469 /*
21470 * Ok less than or right at a MSS is
21471 * outstanding. The original FreeBSD stack would
21472 * have sent a FIN, which can speed things up for
21473 * a transactional application doing a MSG_WAITALL.
21474 * To speed things up since we do *not* send a FIN
21475 * if data is outstanding, we send a "challenge ack".
21476 * The idea behind that is instead of having to have
21477 * the peer wait for the delayed-ack timer to run off
21478 * we send an ack that makes the peer send us an ack.
21479 */
21480 rack_send_ack_challange(rack);
21481 }
21482 goto just_return;
21483 }
21484 /*
21485 * Grab a header mbuf, attaching a copy of data to be transmitted,
21486 * and initialize the header from the template for sends on this
21487 * connection.
21488 */
21489 hw_tls = tp->t_nic_ktls_xmit != 0;
21490 if (len) {
21491 uint32_t max_val;
21492 uint32_t moff;
21493
21494 if (pace_max_seg)
21495 max_val = pace_max_seg;
21496 else
21497 max_val = len;
21498 /*
21499 * We allow a limit on sending with hptsi.
21500 */
21501 if (len > max_val) {
21502 mark = 7;
21503 len = max_val;
21504 }
21505 #ifdef INET6
21506 if (MHLEN < hdrlen + max_linkhdr)
21507 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21508 else
21509 #endif
21510 m = m_gethdr(M_NOWAIT, MT_DATA);
21511
21512 if (m == NULL) {
21513 SOCK_SENDBUF_UNLOCK(so);
21514 error = ENOBUFS;
21515 sack_rxmit = 0;
21516 goto out;
21517 }
21518 m->m_data += max_linkhdr;
21519 m->m_len = hdrlen;
21520
21521 /*
21522 * Start the m_copy functions from the closest mbuf to the
21523 * sb_offset in the socket buffer chain.
21524 */
21525 mb = sbsndptr_noadv(sb, sb_offset, &moff);
21526 s_mb = mb;
21527 s_moff = moff;
21528 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21529 m_copydata(mb, moff, (int)len,
21530 mtod(m, caddr_t)+hdrlen);
21531 /*
21532 * If we are not retransmitting advance the
21533 * sndptr to help remember the next place in
21534 * the sb.
21535 */
21536 if (rsm == NULL)
21537 sbsndptr_adv(sb, mb, len);
21538 m->m_len += len;
21539 } else {
21540 struct sockbuf *msb;
21541
21542 /*
21543 * If we are not retransmitting pass in msb so
21544 * the socket buffer can be advanced. Otherwise
21545 * set it to NULL if its a retransmission since
21546 * we don't want to change the sb remembered
21547 * location.
21548 */
21549 if (rsm == NULL)
21550 msb = sb;
21551 else
21552 msb = NULL;
21553 m->m_next = tcp_m_copym(
21554 mb, moff, &len,
21555 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21556 ((rsm == NULL) ? hw_tls : 0));
21557 if (len <= (tp->t_maxseg - optlen)) {
21558 /*
21559 * Must have ran out of mbufs for the copy
21560 * shorten it to no longer need tso. Lets
21561 * not put on sendalot since we are low on
21562 * mbufs.
21563 */
21564 tso = 0;
21565 }
21566 if (m->m_next == NULL) {
21567 SOCK_SENDBUF_UNLOCK(so);
21568 (void)m_free(m);
21569 error = ENOBUFS;
21570 sack_rxmit = 0;
21571 goto out;
21572 }
21573 }
21574 if (sack_rxmit) {
21575 if (rsm && (rsm->r_flags & RACK_TLP)) {
21576 /*
21577 * TLP should not count in retran count, but
21578 * in its own bin
21579 */
21580 counter_u64_add(rack_tlp_retran, 1);
21581 counter_u64_add(rack_tlp_retran_bytes, len);
21582 } else {
21583 tp->t_sndrexmitpack++;
21584 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21585 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21586 }
21587 #ifdef STATS
21588 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21589 len);
21590 #endif
21591 } else {
21592 KMOD_TCPSTAT_INC(tcps_sndpack);
21593 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21594 #ifdef STATS
21595 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21596 len);
21597 #endif
21598 }
21599 /*
21600 * If we're sending everything we've got, set PUSH. (This
21601 * will keep happy those implementations which only give
21602 * data to the user when a buffer fills or a PUSH comes in.)
21603 */
21604 if (sb_offset + len == sbused(sb) &&
21605 sbused(sb) &&
21606 !(flags & TH_SYN)) {
21607 flags |= TH_PUSH;
21608 add_flag |= RACK_HAD_PUSH;
21609 }
21610 SOCK_SENDBUF_UNLOCK(so);
21611 } else {
21612 SOCK_SENDBUF_UNLOCK(so);
21613 if (tp->t_flags & TF_ACKNOW)
21614 KMOD_TCPSTAT_INC(tcps_sndacks);
21615 else if (flags & (TH_SYN | TH_FIN | TH_RST))
21616 KMOD_TCPSTAT_INC(tcps_sndctrl);
21617 else
21618 KMOD_TCPSTAT_INC(tcps_sndwinup);
21619
21620 m = m_gethdr(M_NOWAIT, MT_DATA);
21621 if (m == NULL) {
21622 error = ENOBUFS;
21623 sack_rxmit = 0;
21624 goto out;
21625 }
21626 #ifdef INET6
21627 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21628 MHLEN >= hdrlen) {
21629 M_ALIGN(m, hdrlen);
21630 } else
21631 #endif
21632 m->m_data += max_linkhdr;
21633 m->m_len = hdrlen;
21634 }
21635 SOCK_SENDBUF_UNLOCK_ASSERT(so);
21636 m->m_pkthdr.rcvif = (struct ifnet *)0;
21637 #ifdef MAC
21638 mac_inpcb_create_mbuf(inp, m);
21639 #endif
21640 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21641 #ifdef INET6
21642 if (isipv6)
21643 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21644 else
21645 #endif /* INET6 */
21646 #ifdef INET
21647 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21648 #endif
21649 th = rack->r_ctl.fsb.th;
21650 udp = rack->r_ctl.fsb.udp;
21651 if (udp) {
21652 #ifdef INET6
21653 if (isipv6)
21654 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21655 else
21656 #endif /* INET6 */
21657 ulen = hdrlen + len - sizeof(struct ip);
21658 udp->uh_ulen = htons(ulen);
21659 }
21660 } else {
21661 #ifdef INET6
21662 if (isipv6) {
21663 ip6 = mtod(m, struct ip6_hdr *);
21664 if (tp->t_port) {
21665 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21666 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21667 udp->uh_dport = tp->t_port;
21668 ulen = hdrlen + len - sizeof(struct ip6_hdr);
21669 udp->uh_ulen = htons(ulen);
21670 th = (struct tcphdr *)(udp + 1);
21671 } else
21672 th = (struct tcphdr *)(ip6 + 1);
21673 tcpip_fillheaders(inp, tp->t_port, ip6, th);
21674 } else
21675 #endif /* INET6 */
21676 {
21677 #ifdef INET
21678 ip = mtod(m, struct ip *);
21679 if (tp->t_port) {
21680 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21681 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21682 udp->uh_dport = tp->t_port;
21683 ulen = hdrlen + len - sizeof(struct ip);
21684 udp->uh_ulen = htons(ulen);
21685 th = (struct tcphdr *)(udp + 1);
21686 } else
21687 th = (struct tcphdr *)(ip + 1);
21688 tcpip_fillheaders(inp, tp->t_port, ip, th);
21689 #endif
21690 }
21691 }
21692 /*
21693 * If we are starting a connection, send ECN setup SYN packet. If we
21694 * are on a retransmit, we may resend those bits a number of times
21695 * as per RFC 3168.
21696 */
21697 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21698 flags |= tcp_ecn_output_syn_sent(tp);
21699 }
21700 /* Also handle parallel SYN for ECN */
21701 if (TCPS_HAVERCVDSYN(tp->t_state) &&
21702 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21703 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21704 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21705 (tp->t_flags2 & TF2_ECN_SND_ECE))
21706 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21707 #ifdef INET6
21708 if (isipv6) {
21709 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21710 ip6->ip6_flow |= htonl(ect << 20);
21711 }
21712 else
21713 #endif
21714 {
21715 #ifdef INET
21716 ip->ip_tos &= ~IPTOS_ECN_MASK;
21717 ip->ip_tos |= ect;
21718 #endif
21719 }
21720 }
21721 th->th_seq = htonl(rack_seq);
21722 th->th_ack = htonl(tp->rcv_nxt);
21723 tcp_set_flags(th, flags);
21724 /*
21725 * Calculate receive window. Don't shrink window, but avoid silly
21726 * window syndrome.
21727 * If a RST segment is sent, advertise a window of zero.
21728 */
21729 if (flags & TH_RST) {
21730 recwin = 0;
21731 } else {
21732 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21733 recwin < (long)segsiz) {
21734 recwin = 0;
21735 }
21736 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21737 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21738 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21739 }
21740
21741 /*
21742 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21743 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is
21744 * handled in syncache.
21745 */
21746 if (flags & TH_SYN)
21747 th->th_win = htons((u_short)
21748 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21749 else {
21750 /* Avoid shrinking window with window scaling. */
21751 recwin = roundup2(recwin, 1 << tp->rcv_scale);
21752 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21753 }
21754 /*
21755 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21756 * window. This may cause the remote transmitter to stall. This
21757 * flag tells soreceive() to disable delayed acknowledgements when
21758 * draining the buffer. This can occur if the receiver is
21759 * attempting to read more data than can be buffered prior to
21760 * transmitting on the connection.
21761 */
21762 if (th->th_win == 0) {
21763 tp->t_sndzerowin++;
21764 tp->t_flags |= TF_RXWIN0SENT;
21765 } else
21766 tp->t_flags &= ~TF_RXWIN0SENT;
21767 tp->snd_up = tp->snd_una; /* drag it along, its deprecated */
21768 /* Now are we using fsb?, if so copy the template data to the mbuf */
21769 if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21770 uint8_t *cpto;
21771
21772 cpto = mtod(m, uint8_t *);
21773 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21774 /*
21775 * We have just copied in:
21776 * IP/IP6
21777 * <optional udphdr>
21778 * tcphdr (no options)
21779 *
21780 * We need to grab the correct pointers into the mbuf
21781 * for both the tcp header, and possibly the udp header (if tunneling).
21782 * We do this by using the offset in the copy buffer and adding it
21783 * to the mbuf base pointer (cpto).
21784 */
21785 #ifdef INET6
21786 if (isipv6)
21787 ip6 = mtod(m, struct ip6_hdr *);
21788 else
21789 #endif /* INET6 */
21790 #ifdef INET
21791 ip = mtod(m, struct ip *);
21792 #endif
21793 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21794 /* If we have a udp header lets set it into the mbuf as well */
21795 if (udp)
21796 udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21797 }
21798 if (optlen) {
21799 bcopy(opt, th + 1, optlen);
21800 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21801 }
21802 /*
21803 * Put TCP length in extended header, and then checksum extended
21804 * header and data.
21805 */
21806 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
21807 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21808 if (to.to_flags & TOF_SIGNATURE) {
21809 /*
21810 * Calculate MD5 signature and put it into the place
21811 * determined before.
21812 * NOTE: since TCP options buffer doesn't point into
21813 * mbuf's data, calculate offset and use it.
21814 */
21815 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21816 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21817 /*
21818 * Do not send segment if the calculation of MD5
21819 * digest has failed.
21820 */
21821 goto out;
21822 }
21823 }
21824 #endif
21825 #ifdef INET6
21826 if (isipv6) {
21827 /*
21828 * ip6_plen is not need to be filled now, and will be filled
21829 * in ip6_output.
21830 */
21831 if (tp->t_port) {
21832 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21833 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21834 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21835 th->th_sum = htons(0);
21836 UDPSTAT_INC(udps_opackets);
21837 } else {
21838 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21839 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21840 th->th_sum = in6_cksum_pseudo(ip6,
21841 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21842 0);
21843 }
21844 }
21845 #endif
21846 #if defined(INET6) && defined(INET)
21847 else
21848 #endif
21849 #ifdef INET
21850 {
21851 if (tp->t_port) {
21852 m->m_pkthdr.csum_flags = CSUM_UDP;
21853 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21854 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21855 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21856 th->th_sum = htons(0);
21857 UDPSTAT_INC(udps_opackets);
21858 } else {
21859 m->m_pkthdr.csum_flags = CSUM_TCP;
21860 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21861 th->th_sum = in_pseudo(ip->ip_src.s_addr,
21862 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21863 IPPROTO_TCP + len + optlen));
21864 }
21865 /* IP version must be set here for ipv4/ipv6 checking later */
21866 KASSERT(ip->ip_v == IPVERSION,
21867 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
21868 }
21869 #endif
21870 /*
21871 * Enable TSO and specify the size of the segments. The TCP pseudo
21872 * header checksum is always provided. XXX: Fixme: This is currently
21873 * not the case for IPv6.
21874 */
21875 if (tso) {
21876 /*
21877 * Here we must use t_maxseg and the optlen since
21878 * the optlen may include SACK's (or DSACK).
21879 */
21880 KASSERT(len > tp->t_maxseg - optlen,
21881 ("%s: len <= tso_segsz", __func__));
21882 m->m_pkthdr.csum_flags |= CSUM_TSO;
21883 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21884 }
21885 KASSERT(len + hdrlen == m_length(m, NULL),
21886 ("%s: mbuf chain different than expected: %d + %u != %u",
21887 __func__, len, hdrlen, m_length(m, NULL)));
21888
21889 #ifdef TCP_HHOOK
21890 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21891 hhook_run_tcp_est_out(tp, th, &to, len, tso);
21892 #endif
21893 if ((rack->r_ctl.crte != NULL) &&
21894 (rack->rc_hw_nobuf == 0) &&
21895 tcp_bblogging_on(tp)) {
21896 rack_log_queue_level(tp, rack, len, &tv, cts);
21897 }
21898 /* We're getting ready to send; log now. */
21899 if (tcp_bblogging_on(rack->rc_tp)) {
21900 union tcp_log_stackspecific log;
21901
21902 memset(&log, 0, sizeof(log));
21903 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21904 if (rack->rack_no_prr)
21905 log.u_bbr.flex1 = 0;
21906 else
21907 log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21908 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
21909 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
21910 log.u_bbr.flex4 = orig_len;
21911 /* Save off the early/late values */
21912 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
21913 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
21914 log.u_bbr.bw_inuse = rack_get_bw(rack);
21915 log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
21916 log.u_bbr.flex8 = 0;
21917 if (rsm) {
21918 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
21919 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
21920 counter_u64_add(rack_collapsed_win_rxt, 1);
21921 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
21922 }
21923 if (doing_tlp)
21924 log.u_bbr.flex8 = 2;
21925 else
21926 log.u_bbr.flex8 = 1;
21927 } else {
21928 if (doing_tlp)
21929 log.u_bbr.flex8 = 3;
21930 }
21931 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
21932 log.u_bbr.flex7 = mark;
21933 log.u_bbr.flex7 <<= 8;
21934 log.u_bbr.flex7 |= pass;
21935 log.u_bbr.pkts_out = tp->t_maxseg;
21936 log.u_bbr.timeStamp = cts;
21937 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
21938 if (rsm && (rsm->r_rtr_cnt > 0)) {
21939 /*
21940 * When we have a retransmit we want to log the
21941 * burst at send and flight at send from before.
21942 */
21943 log.u_bbr.flex5 = rsm->r_fas;
21944 log.u_bbr.bbr_substate = rsm->r_bas;
21945 } else {
21946 /*
21947 * New transmits we log in flex5 the inflight again as
21948 * well as the number of segments in our send in the
21949 * substate field.
21950 */
21951 log.u_bbr.flex5 = log.u_bbr.inflight;
21952 log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
21953 }
21954 log.u_bbr.lt_epoch = cwnd_to_use;
21955 log.u_bbr.delivered = sendalot;
21956 log.u_bbr.rttProp = (uintptr_t)rsm;
21957 log.u_bbr.pkt_epoch = __LINE__;
21958 if (rsm) {
21959 log.u_bbr.delRate = rsm->r_flags;
21960 log.u_bbr.delRate <<= 31;
21961 log.u_bbr.delRate |= rack->r_must_retran;
21962 log.u_bbr.delRate <<= 1;
21963 log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21964 } else {
21965 log.u_bbr.delRate = rack->r_must_retran;
21966 log.u_bbr.delRate <<= 1;
21967 log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
21968 }
21969 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
21970 len, &log, false, NULL, __func__, __LINE__, &tv);
21971 } else
21972 lgb = NULL;
21973
21974 /*
21975 * Fill in IP length and desired time to live and send to IP level.
21976 * There should be a better way to handle ttl and tos; we could keep
21977 * them in the template, but need a way to checksum without them.
21978 */
21979 /*
21980 * m->m_pkthdr.len should have been set before cksum calcuration,
21981 * because in6_cksum() need it.
21982 */
21983 #ifdef INET6
21984 if (isipv6) {
21985 /*
21986 * we separately set hoplimit for every segment, since the
21987 * user might want to change the value via setsockopt. Also,
21988 * desired default hop limit might be changed via Neighbor
21989 * Discovery.
21990 */
21991 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
21992
21993 /*
21994 * Set the packet size here for the benefit of DTrace
21995 * probes. ip6_output() will set it properly; it's supposed
21996 * to include the option header lengths as well.
21997 */
21998 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
21999
22000 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22001 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22002 else
22003 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22004
22005 if (tp->t_state == TCPS_SYN_SENT)
22006 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22007
22008 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22009 /* TODO: IPv6 IP6TOS_ECT bit on */
22010 error = ip6_output(m,
22011 inp->in6p_outputopts,
22012 &inp->inp_route6,
22013 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22014 NULL, NULL, inp);
22015
22016 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22017 mtu = inp->inp_route6.ro_nh->nh_mtu;
22018 }
22019 #endif /* INET6 */
22020 #if defined(INET) && defined(INET6)
22021 else
22022 #endif
22023 #ifdef INET
22024 {
22025 ip->ip_len = htons(m->m_pkthdr.len);
22026 #ifdef INET6
22027 if (inp->inp_vflag & INP_IPV6PROTO)
22028 ip->ip_ttl = in6_selecthlim(inp, NULL);
22029 #endif /* INET6 */
22030 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22031 /*
22032 * If we do path MTU discovery, then we set DF on every
22033 * packet. This might not be the best thing to do according
22034 * to RFC3390 Section 2. However the tcp hostcache migitates
22035 * the problem so it affects only the first tcp connection
22036 * with a host.
22037 *
22038 * NB: Don't set DF on small MTU/MSS to have a safe
22039 * fallback.
22040 */
22041 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22042 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22043 if (tp->t_port == 0 || len < V_tcp_minmss) {
22044 ip->ip_off |= htons(IP_DF);
22045 }
22046 } else {
22047 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22048 }
22049
22050 if (tp->t_state == TCPS_SYN_SENT)
22051 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22052
22053 TCP_PROBE5(send, NULL, tp, ip, tp, th);
22054
22055 error = ip_output(m,
22056 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22057 inp->inp_options,
22058 #else
22059 NULL,
22060 #endif
22061 &inp->inp_route,
22062 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22063 inp);
22064 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22065 mtu = inp->inp_route.ro_nh->nh_mtu;
22066 }
22067 #endif /* INET */
22068 if (lgb) {
22069 lgb->tlb_errno = error;
22070 lgb = NULL;
22071 }
22072
22073 out:
22074 /*
22075 * In transmit state, time the transmission and arrange for the
22076 * retransmit. In persist state, just set snd_max.
22077 */
22078 if ((rsm == NULL) && doing_tlp)
22079 add_flag |= RACK_TLP;
22080 rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22081 rack_to_usec_ts(&tv),
22082 rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22083 if (error == 0) {
22084 if (add_flag & RACK_IS_PCM) {
22085 /* We just launched a PCM */
22086 /* rrs here log */
22087 rack->pcm_in_progress = 1;
22088 rack->pcm_needed = 0;
22089 rack_log_pcm(rack, 7, len, rack->r_ctl.pcm_max_seg, add_flag);
22090 }
22091 if (rsm == NULL) {
22092 if (rack->lt_bw_up == 0) {
22093 rack->r_ctl.lt_timemark = tcp_tv_to_lusec(&tv);
22094 rack->r_ctl.lt_seq = tp->snd_una;
22095 rack->lt_bw_up = 1;
22096 } else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22097 /*
22098 * Need to record what we have since we are
22099 * approaching seq wrap.
22100 */
22101 uint64_t tmark;
22102
22103 rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22104 rack->r_ctl.lt_seq = tp->snd_una;
22105 tmark = tcp_get_u64_usecs(&tv);
22106 if (tmark > rack->r_ctl.lt_timemark) {
22107 rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22108 rack->r_ctl.lt_timemark = tmark;
22109 }
22110 }
22111 }
22112 rack->forced_ack = 0; /* If we send something zap the FA flag */
22113 counter_u64_add(rack_total_bytes, len);
22114 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22115 if (rsm && doing_tlp) {
22116 rack->rc_last_sent_tlp_past_cumack = 0;
22117 rack->rc_last_sent_tlp_seq_valid = 1;
22118 rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22119 rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22120 }
22121 if (rack->rc_hw_nobuf) {
22122 rack->rc_hw_nobuf = 0;
22123 rack->r_ctl.rc_agg_delayed = 0;
22124 rack->r_early = 0;
22125 rack->r_late = 0;
22126 rack->r_ctl.rc_agg_early = 0;
22127 }
22128 if (rsm && (doing_tlp == 0)) {
22129 /* Set we retransmitted */
22130 rack->rc_gp_saw_rec = 1;
22131 } else {
22132 if (cwnd_to_use > tp->snd_ssthresh) {
22133 /* Set we sent in CA */
22134 rack->rc_gp_saw_ca = 1;
22135 } else {
22136 /* Set we sent in SS */
22137 rack->rc_gp_saw_ss = 1;
22138 }
22139 }
22140 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22141 (tp->t_flags & TF_SACK_PERMIT) &&
22142 tp->rcv_numsacks > 0)
22143 tcp_clean_dsack_blocks(tp);
22144 tot_len_this_send += len;
22145 if (len == 0) {
22146 counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22147 } else {
22148 int idx;
22149
22150 idx = (len / segsiz) + 3;
22151 if (idx >= TCP_MSS_ACCT_ATIMER)
22152 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22153 else
22154 counter_u64_add(rack_out_size[idx], 1);
22155 }
22156 }
22157 if ((rack->rack_no_prr == 0) &&
22158 sub_from_prr &&
22159 (error == 0)) {
22160 if (rack->r_ctl.rc_prr_sndcnt >= len)
22161 rack->r_ctl.rc_prr_sndcnt -= len;
22162 else
22163 rack->r_ctl.rc_prr_sndcnt = 0;
22164 }
22165 sub_from_prr = 0;
22166 if (rsm != NULL) {
22167 if (doing_tlp)
22168 /* Make sure the TLP is added */
22169 rsm->r_flags |= RACK_TLP;
22170 else
22171 /* If its a resend without TLP then it must not have the flag */
22172 rsm->r_flags &= ~RACK_TLP;
22173 }
22174 if ((error == 0) &&
22175 (len > 0) &&
22176 (tp->snd_una == tp->snd_max))
22177 rack->r_ctl.rc_tlp_rxt_last_time = cts;
22178
22179 {
22180 /*
22181 * This block is not associated with the above error == 0 test.
22182 * It is used to advance snd_max if we have a new transmit.
22183 */
22184 tcp_seq startseq = tp->snd_max;
22185
22186
22187 if (rsm && (doing_tlp == 0))
22188 rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22189 if (error)
22190 /* We don't log or do anything with errors */
22191 goto nomore;
22192 if (doing_tlp == 0) {
22193 if (rsm == NULL) {
22194 /*
22195 * Not a retransmission of some
22196 * sort, new data is going out so
22197 * clear our TLP count and flag.
22198 */
22199 rack->rc_tlp_in_progress = 0;
22200 rack->r_ctl.rc_tlp_cnt_out = 0;
22201 }
22202 } else {
22203 /*
22204 * We have just sent a TLP, mark that it is true
22205 * and make sure our in progress is set so we
22206 * continue to check the count.
22207 */
22208 rack->rc_tlp_in_progress = 1;
22209 rack->r_ctl.rc_tlp_cnt_out++;
22210 }
22211 /*
22212 * If we are retransmitting we are done, snd_max
22213 * does not get updated.
22214 */
22215 if (sack_rxmit)
22216 goto nomore;
22217 if ((tp->snd_una == tp->snd_max) && (len > 0)) {
22218 /*
22219 * Update the time we just added data since
22220 * nothing was outstanding.
22221 */
22222 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22223 tp->t_acktime = ticks;
22224 }
22225 /*
22226 * Now for special SYN/FIN handling.
22227 */
22228 if (flags & (TH_SYN | TH_FIN)) {
22229 if ((flags & TH_SYN) &&
22230 ((tp->t_flags & TF_SENTSYN) == 0)) {
22231 tp->snd_max++;
22232 tp->t_flags |= TF_SENTSYN;
22233 }
22234 if ((flags & TH_FIN) &&
22235 ((tp->t_flags & TF_SENTFIN) == 0)) {
22236 tp->snd_max++;
22237 tp->t_flags |= TF_SENTFIN;
22238 }
22239 }
22240 tp->snd_max += len;
22241 if (rack->rc_new_rnd_needed) {
22242 rack_new_round_starts(tp, rack, tp->snd_max);
22243 }
22244 /*
22245 * Time this transmission if not a retransmission and
22246 * not currently timing anything.
22247 * This is only relevant in case of switching back to
22248 * the base stack.
22249 */
22250 if (tp->t_rtttime == 0) {
22251 tp->t_rtttime = ticks;
22252 tp->t_rtseq = startseq;
22253 KMOD_TCPSTAT_INC(tcps_segstimed);
22254 }
22255 if (len &&
22256 ((tp->t_flags & TF_GPUTINPROG) == 0))
22257 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22258 /*
22259 * If we are doing FO we need to update the mbuf position and subtract
22260 * this happens when the peer sends us duplicate information and
22261 * we thus want to send a DSACK.
22262 *
22263 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22264 * turned off? If not then we are going to echo multiple DSACK blocks
22265 * out (with the TSO), which we should not be doing.
22266 */
22267 if (rack->r_fast_output && len) {
22268 if (rack->r_ctl.fsb.left_to_send > len)
22269 rack->r_ctl.fsb.left_to_send -= len;
22270 else
22271 rack->r_ctl.fsb.left_to_send = 0;
22272 if (rack->r_ctl.fsb.left_to_send < segsiz)
22273 rack->r_fast_output = 0;
22274 if (rack->r_fast_output) {
22275 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22276 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22277 rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22278 }
22279 }
22280 if (rack_pcm_blast == 0) {
22281 if ((orig_len > len) &&
22282 (add_flag & RACK_IS_PCM) &&
22283 (len < pace_max_seg) &&
22284 ((pace_max_seg - len) > segsiz)) {
22285 /*
22286 * We are doing a PCM measurement and we did
22287 * not get enough data in the TSO to meet the
22288 * burst requirement.
22289 */
22290 uint32_t n_len;
22291
22292 n_len = (orig_len - len);
22293 orig_len -= len;
22294 pace_max_seg -= len;
22295 len = n_len;
22296 sb_offset = tp->snd_max - tp->snd_una;
22297 /* Re-lock for the next spin */
22298 SOCK_SENDBUF_LOCK(so);
22299 goto send;
22300 }
22301 } else {
22302 if ((orig_len > len) &&
22303 (add_flag & RACK_IS_PCM) &&
22304 ((orig_len - len) > segsiz)) {
22305 /*
22306 * We are doing a PCM measurement and we did
22307 * not get enough data in the TSO to meet the
22308 * burst requirement.
22309 */
22310 uint32_t n_len;
22311
22312 n_len = (orig_len - len);
22313 orig_len -= len;
22314 len = n_len;
22315 sb_offset = tp->snd_max - tp->snd_una;
22316 /* Re-lock for the next spin */
22317 SOCK_SENDBUF_LOCK(so);
22318 goto send;
22319 }
22320 }
22321 }
22322 nomore:
22323 if (error) {
22324 rack->r_ctl.rc_agg_delayed = 0;
22325 rack->r_early = 0;
22326 rack->r_late = 0;
22327 rack->r_ctl.rc_agg_early = 0;
22328 SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */
22329 /*
22330 * Failures do not advance the seq counter above. For the
22331 * case of ENOBUFS we will fall out and retry in 1ms with
22332 * the hpts. Everything else will just have to retransmit
22333 * with the timer.
22334 *
22335 * In any case, we do not want to loop around for another
22336 * send without a good reason.
22337 */
22338 sendalot = 0;
22339 switch (error) {
22340 case EPERM:
22341 case EACCES:
22342 tp->t_softerror = error;
22343 #ifdef TCP_ACCOUNTING
22344 crtsc = get_cyclecount();
22345 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22346 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22347 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22348 }
22349 sched_unpin();
22350 #endif
22351 return (error);
22352 case ENOBUFS:
22353 /*
22354 * Pace us right away to retry in a some
22355 * time
22356 */
22357 if (rack->r_ctl.crte != NULL) {
22358 tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22359 if (tcp_bblogging_on(rack->rc_tp))
22360 rack_log_queue_level(tp, rack, len, &tv, cts);
22361 } else
22362 tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22363 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22364 if (rack->rc_enobuf < 0x7f)
22365 rack->rc_enobuf++;
22366 if (slot < (10 * HPTS_USEC_IN_MSEC))
22367 slot = 10 * HPTS_USEC_IN_MSEC;
22368 if (rack->r_ctl.crte != NULL) {
22369 counter_u64_add(rack_saw_enobuf_hw, 1);
22370 tcp_rl_log_enobuf(rack->r_ctl.crte);
22371 }
22372 counter_u64_add(rack_saw_enobuf, 1);
22373 goto enobufs;
22374 case EMSGSIZE:
22375 /*
22376 * For some reason the interface we used initially
22377 * to send segments changed to another or lowered
22378 * its MTU. If TSO was active we either got an
22379 * interface without TSO capabilits or TSO was
22380 * turned off. If we obtained mtu from ip_output()
22381 * then update it and try again.
22382 */
22383 if (tso)
22384 tp->t_flags &= ~TF_TSO;
22385 if (mtu != 0) {
22386 int saved_mtu;
22387
22388 saved_mtu = tp->t_maxseg;
22389 tcp_mss_update(tp, -1, mtu, NULL, NULL);
22390 if (saved_mtu > tp->t_maxseg) {
22391 goto again;
22392 }
22393 }
22394 slot = 10 * HPTS_USEC_IN_MSEC;
22395 rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22396 #ifdef TCP_ACCOUNTING
22397 crtsc = get_cyclecount();
22398 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22399 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22400 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22401 }
22402 sched_unpin();
22403 #endif
22404 return (error);
22405 case ENETUNREACH:
22406 counter_u64_add(rack_saw_enetunreach, 1);
22407 /* FALLTHROUGH */
22408 case EHOSTDOWN:
22409 case EHOSTUNREACH:
22410 case ENETDOWN:
22411 if (TCPS_HAVERCVDSYN(tp->t_state)) {
22412 tp->t_softerror = error;
22413 error = 0;
22414 }
22415 /* FALLTHROUGH */
22416 default:
22417 slot = 10 * HPTS_USEC_IN_MSEC;
22418 rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22419 #ifdef TCP_ACCOUNTING
22420 crtsc = get_cyclecount();
22421 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22422 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22423 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22424 }
22425 sched_unpin();
22426 #endif
22427 return (error);
22428 }
22429 } else {
22430 rack->rc_enobuf = 0;
22431 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22432 rack->r_ctl.retran_during_recovery += len;
22433 }
22434 KMOD_TCPSTAT_INC(tcps_sndtotal);
22435
22436 /*
22437 * Data sent (as far as we can tell). If this advertises a larger
22438 * window than any other segment, then remember the size of the
22439 * advertised window. Any pending ACK has now been sent.
22440 */
22441 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22442 tp->rcv_adv = tp->rcv_nxt + recwin;
22443
22444 tp->last_ack_sent = tp->rcv_nxt;
22445 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22446 enobufs:
22447 if (sendalot) {
22448 /* Do we need to turn off sendalot? */
22449 if (pace_max_seg &&
22450 (tot_len_this_send >= pace_max_seg)) {
22451 /* We hit our max. */
22452 sendalot = 0;
22453 }
22454 }
22455 if ((error == 0) && (flags & TH_FIN))
22456 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22457 if (flags & TH_RST) {
22458 /*
22459 * We don't send again after sending a RST.
22460 */
22461 slot = 0;
22462 sendalot = 0;
22463 if (error == 0)
22464 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22465 } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22466 /*
22467 * Get our pacing rate, if an error
22468 * occurred in sending (ENOBUF) we would
22469 * hit the else if with slot preset. Other
22470 * errors return.
22471 */
22472 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz, __LINE__);
22473 }
22474 /* We have sent clear the flag */
22475 rack->r_ent_rec_ns = 0;
22476 if (rack->r_must_retran) {
22477 if (rsm) {
22478 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22479 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22480 /*
22481 * We have retransmitted all.
22482 */
22483 rack->r_must_retran = 0;
22484 rack->r_ctl.rc_out_at_rto = 0;
22485 }
22486 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22487 /*
22488 * Sending new data will also kill
22489 * the loop.
22490 */
22491 rack->r_must_retran = 0;
22492 rack->r_ctl.rc_out_at_rto = 0;
22493 }
22494 }
22495 rack->r_ctl.fsb.recwin = recwin;
22496 if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22497 SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22498 /*
22499 * We hit an RTO and now have past snd_max at the RTO
22500 * clear all the WAS flags.
22501 */
22502 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22503 }
22504 if (slot) {
22505 /* set the rack tcb into the slot N */
22506 if ((error == 0) &&
22507 rack_use_rfo &&
22508 ((flags & (TH_SYN|TH_FIN)) == 0) &&
22509 (rsm == NULL) &&
22510 (ipoptlen == 0) &&
22511 (doing_tlp == 0) &&
22512 rack->r_fsb_inited &&
22513 TCPS_HAVEESTABLISHED(tp->t_state) &&
22514 ((IN_RECOVERY(tp->t_flags)) == 0) &&
22515 (rack->r_must_retran == 0) &&
22516 ((tp->t_flags & TF_NEEDFIN) == 0) &&
22517 (len > 0) && (orig_len > 0) &&
22518 (orig_len > len) &&
22519 ((orig_len - len) >= segsiz) &&
22520 ((optlen == 0) ||
22521 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22522 /* We can send at least one more MSS using our fsb */
22523 rack_setup_fast_output(tp, rack, sb, len, orig_len,
22524 segsiz, pace_max_seg, hw_tls, flags);
22525 } else
22526 rack->r_fast_output = 0;
22527 rack_log_fsb(rack, tp, so, flags,
22528 ipoptlen, orig_len, len, error,
22529 (rsm == NULL), optlen, __LINE__, 2);
22530 } else if (sendalot) {
22531 int ret;
22532
22533 sack_rxmit = 0;
22534 if ((error == 0) &&
22535 rack_use_rfo &&
22536 ((flags & (TH_SYN|TH_FIN)) == 0) &&
22537 (rsm == NULL) &&
22538 (doing_tlp == 0) &&
22539 (ipoptlen == 0) &&
22540 (rack->r_must_retran == 0) &&
22541 rack->r_fsb_inited &&
22542 TCPS_HAVEESTABLISHED(tp->t_state) &&
22543 ((IN_RECOVERY(tp->t_flags)) == 0) &&
22544 ((tp->t_flags & TF_NEEDFIN) == 0) &&
22545 (len > 0) && (orig_len > 0) &&
22546 (orig_len > len) &&
22547 ((orig_len - len) >= segsiz) &&
22548 ((optlen == 0) ||
22549 ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22550 /* we can use fast_output for more */
22551 rack_setup_fast_output(tp, rack, sb, len, orig_len,
22552 segsiz, pace_max_seg, hw_tls, flags);
22553 if (rack->r_fast_output) {
22554 error = 0;
22555 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, &tot_len_this_send, &error, __LINE__);
22556 if (ret >= 0)
22557 return (ret);
22558 else if (error)
22559 goto nomore;
22560
22561 }
22562 }
22563 goto again;
22564 }
22565 skip_all_send:
22566 /* Assure when we leave that snd_nxt will point to top */
22567 if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22568 tp->snd_nxt = tp->snd_max;
22569 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22570 #ifdef TCP_ACCOUNTING
22571 crtsc = get_cyclecount() - ts_val;
22572 if (tot_len_this_send) {
22573 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22574 tp->tcp_cnt_counters[SND_OUT_DATA]++;
22575 tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22576 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22577 }
22578 } else {
22579 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22580 tp->tcp_cnt_counters[SND_OUT_ACK]++;
22581 tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22582 }
22583 }
22584 sched_unpin();
22585 #endif
22586 if (error == ENOBUFS)
22587 error = 0;
22588 return (error);
22589 }
22590
22591 static void
rack_update_seg(struct tcp_rack * rack)22592 rack_update_seg(struct tcp_rack *rack)
22593 {
22594 uint32_t orig_val;
22595
22596 orig_val = rack->r_ctl.rc_pace_max_segs;
22597 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22598 if (orig_val != rack->r_ctl.rc_pace_max_segs)
22599 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22600 }
22601
22602 static void
rack_mtu_change(struct tcpcb * tp)22603 rack_mtu_change(struct tcpcb *tp)
22604 {
22605 /*
22606 * The MSS may have changed
22607 */
22608 struct tcp_rack *rack;
22609 struct rack_sendmap *rsm;
22610
22611 rack = (struct tcp_rack *)tp->t_fb_ptr;
22612 if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22613 /*
22614 * The MTU has changed we need to resend everything
22615 * since all we have sent is lost. We first fix
22616 * up the mtu though.
22617 */
22618 rack_set_pace_segments(tp, rack, __LINE__, NULL);
22619 /* We treat this like a full retransmit timeout without the cwnd adjustment */
22620 rack_remxt_tmr(tp);
22621 rack->r_fast_output = 0;
22622 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22623 rack->r_ctl.rc_sacked);
22624 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22625 rack->r_must_retran = 1;
22626 /* Mark all inflight to needing to be rxt'd */
22627 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22628 rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22629 }
22630 }
22631 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22632 /* We don't use snd_nxt to retransmit */
22633 tp->snd_nxt = tp->snd_max;
22634 }
22635
22636 static int
rack_set_dgp(struct tcp_rack * rack)22637 rack_set_dgp(struct tcp_rack *rack)
22638 {
22639 if (rack->dgp_on == 1)
22640 return(0);
22641 if ((rack->use_fixed_rate == 1) &&
22642 (rack->rc_always_pace == 1)) {
22643 /*
22644 * We are already pacing another
22645 * way.
22646 */
22647 return (EBUSY);
22648 }
22649 if (rack->rc_always_pace == 1) {
22650 rack_remove_pacing(rack);
22651 }
22652 if (tcp_incr_dgp_pacing_cnt() == 0)
22653 return (ENOSPC);
22654 rack->r_ctl.pacing_method |= RACK_DGP_PACING;
22655 rack->rc_fillcw_apply_discount = 0;
22656 rack->dgp_on = 1;
22657 rack->rc_always_pace = 1;
22658 rack->rc_pace_dnd = 1;
22659 rack->use_fixed_rate = 0;
22660 if (rack->gp_ready)
22661 rack_set_cc_pacing(rack);
22662 rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22663 rack->rack_attempt_hdwr_pace = 0;
22664 /* rxt settings */
22665 rack->full_size_rxt = 1;
22666 rack->shape_rxt_to_pacing_min = 0;
22667 /* cmpack=1 */
22668 rack->r_use_cmp_ack = 1;
22669 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22670 rack->r_use_cmp_ack)
22671 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22672 /* scwnd=1 */
22673 rack->rack_enable_scwnd = 1;
22674 /* dynamic=100 */
22675 rack->rc_gp_dyn_mul = 1;
22676 /* gp_inc_ca */
22677 rack->r_ctl.rack_per_of_gp_ca = 100;
22678 /* rrr_conf=3 */
22679 rack->r_rr_config = 3;
22680 /* npush=2 */
22681 rack->r_ctl.rc_no_push_at_mrtt = 2;
22682 /* fillcw=1 */
22683 rack->rc_pace_to_cwnd = 1;
22684 rack->rc_pace_fill_if_rttin_range = 0;
22685 rack->rtt_limit_mul = 0;
22686 /* noprr=1 */
22687 rack->rack_no_prr = 1;
22688 /* lscwnd=1 */
22689 rack->r_limit_scw = 1;
22690 /* gp_inc_rec */
22691 rack->r_ctl.rack_per_of_gp_rec = 90;
22692 return (0);
22693 }
22694
22695 static int
rack_set_profile(struct tcp_rack * rack,int prof)22696 rack_set_profile(struct tcp_rack *rack, int prof)
22697 {
22698 int err = EINVAL;
22699 if (prof == 1) {
22700 /*
22701 * Profile 1 is "standard" DGP. It ignores
22702 * client buffer level.
22703 */
22704 err = rack_set_dgp(rack);
22705 if (err)
22706 return (err);
22707 } else if (prof == 6) {
22708 err = rack_set_dgp(rack);
22709 if (err)
22710 return (err);
22711 /*
22712 * Profile 6 tweaks DGP so that it will apply to
22713 * fill-cw the same settings that profile5 does
22714 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22715 */
22716 rack->rc_fillcw_apply_discount = 1;
22717 } else if (prof == 0) {
22718 /* This changes things back to the default settings */
22719 if (rack->rc_always_pace == 1) {
22720 rack_remove_pacing(rack);
22721 } else {
22722 /* Make sure any stray flags are off */
22723 rack->dgp_on = 0;
22724 rack->rc_hybrid_mode = 0;
22725 rack->use_fixed_rate = 0;
22726 }
22727 err = 0;
22728 if (rack_fill_cw_state)
22729 rack->rc_pace_to_cwnd = 1;
22730 else
22731 rack->rc_pace_to_cwnd = 0;
22732
22733 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22734 rack->r_ctl.pacing_method |= RACK_REG_PACING;
22735 rack->rc_always_pace = 1;
22736 if (rack->rack_hibeta)
22737 rack_set_cc_pacing(rack);
22738 } else
22739 rack->rc_always_pace = 0;
22740 if (rack_dsack_std_based & 0x1) {
22741 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22742 rack->rc_rack_tmr_std_based = 1;
22743 }
22744 if (rack_dsack_std_based & 0x2) {
22745 /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */
22746 rack->rc_rack_use_dsack = 1;
22747 }
22748 if (rack_use_cmp_acks)
22749 rack->r_use_cmp_ack = 1;
22750 else
22751 rack->r_use_cmp_ack = 0;
22752 if (rack_disable_prr)
22753 rack->rack_no_prr = 1;
22754 else
22755 rack->rack_no_prr = 0;
22756 if (rack_gp_no_rec_chg)
22757 rack->rc_gp_no_rec_chg = 1;
22758 else
22759 rack->rc_gp_no_rec_chg = 0;
22760 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22761 rack->r_mbuf_queue = 1;
22762 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22763 rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22764 rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22765 } else {
22766 rack->r_mbuf_queue = 0;
22767 rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22768 }
22769 if (rack_enable_shared_cwnd)
22770 rack->rack_enable_scwnd = 1;
22771 else
22772 rack->rack_enable_scwnd = 0;
22773 if (rack_do_dyn_mul) {
22774 /* When dynamic adjustment is on CA needs to start at 100% */
22775 rack->rc_gp_dyn_mul = 1;
22776 if (rack_do_dyn_mul >= 100)
22777 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22778 } else {
22779 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22780 rack->rc_gp_dyn_mul = 0;
22781 }
22782 rack->r_rr_config = 0;
22783 rack->r_ctl.rc_no_push_at_mrtt = 0;
22784 rack->rc_pace_fill_if_rttin_range = 0;
22785 rack->rtt_limit_mul = 0;
22786
22787 if (rack_enable_hw_pacing)
22788 rack->rack_hdw_pace_ena = 1;
22789 else
22790 rack->rack_hdw_pace_ena = 0;
22791 if (rack_disable_prr)
22792 rack->rack_no_prr = 1;
22793 else
22794 rack->rack_no_prr = 0;
22795 if (rack_limits_scwnd)
22796 rack->r_limit_scw = 1;
22797 else
22798 rack->r_limit_scw = 0;
22799 rack_init_retransmit_value(rack, rack_rxt_controls);
22800 err = 0;
22801 }
22802 return (err);
22803 }
22804
22805 static int
rack_add_deferred_option(struct tcp_rack * rack,int sopt_name,uint64_t loptval)22806 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22807 {
22808 struct deferred_opt_list *dol;
22809
22810 dol = malloc(sizeof(struct deferred_opt_list),
22811 M_TCPDO, M_NOWAIT|M_ZERO);
22812 if (dol == NULL) {
22813 /*
22814 * No space yikes -- fail out..
22815 */
22816 return (0);
22817 }
22818 dol->optname = sopt_name;
22819 dol->optval = loptval;
22820 TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22821 return (1);
22822 }
22823
22824 static int
process_hybrid_pacing(struct tcp_rack * rack,struct tcp_hybrid_req * hybrid)22825 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22826 {
22827 #ifdef TCP_REQUEST_TRK
22828 struct tcp_sendfile_track *sft;
22829 struct timeval tv;
22830 tcp_seq seq;
22831 int err;
22832
22833 microuptime(&tv);
22834
22835 /* Make sure no fixed rate is on */
22836 rack->use_fixed_rate = 0;
22837 rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22838 rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22839 rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22840 /* Now allocate or find our entry that will have these settings */
22841 sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusec(&tv), 0);
22842 if (sft == NULL) {
22843 rack->rc_tp->tcp_hybrid_error++;
22844 /* no space, where would it have gone? */
22845 seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22846 rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22847 return (ENOSPC);
22848 }
22849 /* mask our internal flags */
22850 hybrid->hybrid_flags &= TCP_HYBRID_PACING_USER_MASK;
22851 /* The seq will be snd_una + everything in the buffer */
22852 seq = sft->start_seq;
22853 if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22854 /* Disabling hybrid pacing */
22855 if (rack->rc_hybrid_mode) {
22856 rack_set_profile(rack, 0);
22857 rack->rc_tp->tcp_hybrid_stop++;
22858 }
22859 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22860 return (0);
22861 }
22862 if (rack->dgp_on == 0) {
22863 /*
22864 * If we have not yet turned DGP on, do so
22865 * now setting pure DGP mode, no buffer level
22866 * response.
22867 */
22868 if ((err = rack_set_profile(rack, 1)) != 0){
22869 /* Failed to turn pacing on */
22870 rack->rc_tp->tcp_hybrid_error++;
22871 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22872 return (err);
22873 }
22874 }
22875 /*
22876 * Now we must switch to hybrid mode as well which also
22877 * means moving to regular pacing.
22878 */
22879 if (rack->rc_hybrid_mode == 0) {
22880 /* First time */
22881 if (tcp_can_enable_pacing()) {
22882 rack->r_ctl.pacing_method |= RACK_REG_PACING;
22883 rack->rc_hybrid_mode = 1;
22884 } else {
22885 return (ENOSPC);
22886 }
22887 if (rack->r_ctl.pacing_method & RACK_DGP_PACING) {
22888 /*
22889 * This should be true.
22890 */
22891 tcp_dec_dgp_pacing_cnt();
22892 rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
22893 }
22894 }
22895 /* Now set in our flags */
22896 sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22897 if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22898 sft->cspr = hybrid->cspr;
22899 else
22900 sft->cspr = 0;
22901 if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22902 sft->hint_maxseg = hybrid->hint_maxseg;
22903 else
22904 sft->hint_maxseg = 0;
22905 rack->rc_tp->tcp_hybrid_start++;
22906 rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
22907 return (0);
22908 #else
22909 return (ENOTSUP);
22910 #endif
22911 }
22912
22913 static int
rack_stack_information(struct tcpcb * tp,struct stack_specific_info * si)22914 rack_stack_information(struct tcpcb *tp, struct stack_specific_info *si)
22915 {
22916 /* We pulled a SSI info log out what was there */
22917 si->bytes_transmitted = tp->t_sndbytes;
22918 si->bytes_retransmitted = tp->t_snd_rxt_bytes;
22919 return (0);
22920 }
22921
22922 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)22923 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
22924 uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
22925
22926 {
22927 struct epoch_tracker et;
22928 struct sockopt sopt;
22929 struct cc_newreno_opts opt;
22930 uint64_t val;
22931 int error = 0;
22932 uint16_t ca, ss;
22933
22934 switch (sopt_name) {
22935 case TCP_RACK_SET_RXT_OPTIONS:
22936 if (optval <= 2) {
22937 rack_init_retransmit_value(rack, optval);
22938 } else {
22939 /*
22940 * You must send in 0, 1 or 2 all else is
22941 * invalid.
22942 */
22943 error = EINVAL;
22944 }
22945 break;
22946 case TCP_RACK_DSACK_OPT:
22947 RACK_OPTS_INC(tcp_rack_dsack_opt);
22948 if (optval & 0x1) {
22949 rack->rc_rack_tmr_std_based = 1;
22950 } else {
22951 rack->rc_rack_tmr_std_based = 0;
22952 }
22953 if (optval & 0x2) {
22954 rack->rc_rack_use_dsack = 1;
22955 } else {
22956 rack->rc_rack_use_dsack = 0;
22957 }
22958 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
22959 break;
22960 case TCP_RACK_PACING_DIVISOR:
22961 RACK_OPTS_INC(tcp_rack_pacing_divisor);
22962 if (optval == 0) {
22963 rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
22964 } else {
22965 if (optval < RL_MIN_DIVISOR)
22966 rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
22967 else
22968 rack->r_ctl.pace_len_divisor = optval;
22969 }
22970 break;
22971 case TCP_RACK_HI_BETA:
22972 RACK_OPTS_INC(tcp_rack_hi_beta);
22973 if (optval > 0) {
22974 rack->rack_hibeta = 1;
22975 if ((optval >= 50) &&
22976 (optval <= 100)) {
22977 /*
22978 * User wants to set a custom beta.
22979 */
22980 rack->r_ctl.saved_hibeta = optval;
22981 if (rack->rc_pacing_cc_set)
22982 rack_undo_cc_pacing(rack);
22983 rack->r_ctl.rc_saved_beta = optval;
22984 }
22985 if (rack->rc_pacing_cc_set == 0)
22986 rack_set_cc_pacing(rack);
22987 } else {
22988 rack->rack_hibeta = 0;
22989 if (rack->rc_pacing_cc_set)
22990 rack_undo_cc_pacing(rack);
22991 }
22992 break;
22993 case TCP_RACK_PACING_BETA:
22994 error = EINVAL;
22995 break;
22996 case TCP_RACK_TIMER_SLOP:
22997 RACK_OPTS_INC(tcp_rack_timer_slop);
22998 rack->r_ctl.timer_slop = optval;
22999 if (rack->rc_tp->t_srtt) {
23000 /*
23001 * If we have an SRTT lets update t_rxtcur
23002 * to have the new slop.
23003 */
23004 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23005 rack_rto_min, rack_rto_max,
23006 rack->r_ctl.timer_slop);
23007 }
23008 break;
23009 case TCP_RACK_PACING_BETA_ECN:
23010 RACK_OPTS_INC(tcp_rack_beta_ecn);
23011 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23012 /* This only works for newreno. */
23013 error = EINVAL;
23014 break;
23015 }
23016 if (rack->rc_pacing_cc_set) {
23017 /*
23018 * Set them into the real CC module
23019 * whats in the rack pcb is the old values
23020 * to be used on restoral/
23021 */
23022 sopt.sopt_dir = SOPT_SET;
23023 opt.name = CC_NEWRENO_BETA_ECN;
23024 opt.val = optval;
23025 if (CC_ALGO(tp)->ctl_output != NULL)
23026 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23027 else
23028 error = ENOENT;
23029 } else {
23030 /*
23031 * Not pacing yet so set it into our local
23032 * rack pcb storage.
23033 */
23034 rack->r_ctl.rc_saved_beta_ecn = optval;
23035 }
23036 break;
23037 case TCP_DEFER_OPTIONS:
23038 RACK_OPTS_INC(tcp_defer_opt);
23039 if (optval) {
23040 if (rack->gp_ready) {
23041 /* Too late */
23042 error = EINVAL;
23043 break;
23044 }
23045 rack->defer_options = 1;
23046 } else
23047 rack->defer_options = 0;
23048 break;
23049 case TCP_RACK_MEASURE_CNT:
23050 RACK_OPTS_INC(tcp_rack_measure_cnt);
23051 if (optval && (optval <= 0xff)) {
23052 rack->r_ctl.req_measurements = optval;
23053 } else
23054 error = EINVAL;
23055 break;
23056 case TCP_REC_ABC_VAL:
23057 RACK_OPTS_INC(tcp_rec_abc_val);
23058 if (optval > 0)
23059 rack->r_use_labc_for_rec = 1;
23060 else
23061 rack->r_use_labc_for_rec = 0;
23062 break;
23063 case TCP_RACK_ABC_VAL:
23064 RACK_OPTS_INC(tcp_rack_abc_val);
23065 if ((optval > 0) && (optval < 255))
23066 rack->rc_labc = optval;
23067 else
23068 error = EINVAL;
23069 break;
23070 case TCP_HDWR_UP_ONLY:
23071 RACK_OPTS_INC(tcp_pacing_up_only);
23072 if (optval)
23073 rack->r_up_only = 1;
23074 else
23075 rack->r_up_only = 0;
23076 break;
23077 case TCP_FILLCW_RATE_CAP: /* URL:fillcw_cap */
23078 RACK_OPTS_INC(tcp_fillcw_rate_cap);
23079 rack->r_ctl.fillcw_cap = loptval;
23080 break;
23081 case TCP_PACING_RATE_CAP:
23082 RACK_OPTS_INC(tcp_pacing_rate_cap);
23083 if ((rack->dgp_on == 1) &&
23084 (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23085 /*
23086 * If we are doing DGP we need to switch
23087 * to using the pacing limit.
23088 */
23089 if (tcp_can_enable_pacing() == 0) {
23090 error = ENOSPC;
23091 break;
23092 }
23093 /*
23094 * Now change up the flags and counts to be correct.
23095 */
23096 rack->r_ctl.pacing_method |= RACK_REG_PACING;
23097 tcp_dec_dgp_pacing_cnt();
23098 rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23099 }
23100 rack->r_ctl.bw_rate_cap = loptval;
23101 break;
23102 case TCP_HYBRID_PACING:
23103 if (hybrid == NULL) {
23104 error = EINVAL;
23105 break;
23106 }
23107 if (rack->r_ctl.side_chan_dis_mask & HYBRID_DIS_MASK) {
23108 error = EPERM;
23109 break;
23110 }
23111 error = process_hybrid_pacing(rack, hybrid);
23112 break;
23113 case TCP_SIDECHAN_DIS: /* URL:scodm */
23114 if (optval)
23115 rack->r_ctl.side_chan_dis_mask = optval;
23116 else
23117 rack->r_ctl.side_chan_dis_mask = 0;
23118 break;
23119 case TCP_RACK_PROFILE:
23120 RACK_OPTS_INC(tcp_profile);
23121 error = rack_set_profile(rack, optval);
23122 break;
23123 case TCP_USE_CMP_ACKS:
23124 RACK_OPTS_INC(tcp_use_cmp_acks);
23125 if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23126 /* You can't turn it off once its on! */
23127 error = EINVAL;
23128 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23129 rack->r_use_cmp_ack = 1;
23130 rack->r_mbuf_queue = 1;
23131 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23132 }
23133 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23134 tp->t_flags2 |= TF2_MBUF_ACKCMP;
23135 break;
23136 case TCP_SHARED_CWND_TIME_LIMIT:
23137 RACK_OPTS_INC(tcp_lscwnd);
23138 if (optval)
23139 rack->r_limit_scw = 1;
23140 else
23141 rack->r_limit_scw = 0;
23142 break;
23143 case TCP_RACK_DGP_IN_REC:
23144 error = EINVAL;
23145 break;
23146 case TCP_RACK_PACE_TO_FILL:
23147 RACK_OPTS_INC(tcp_fillcw);
23148 if (optval == 0)
23149 rack->rc_pace_to_cwnd = 0;
23150 else {
23151 rack->rc_pace_to_cwnd = 1;
23152 }
23153 if ((optval >= rack_gp_rtt_maxmul) &&
23154 rack_gp_rtt_maxmul &&
23155 (optval < 0xf)) {
23156 rack->rc_pace_fill_if_rttin_range = 1;
23157 rack->rtt_limit_mul = optval;
23158 } else {
23159 rack->rc_pace_fill_if_rttin_range = 0;
23160 rack->rtt_limit_mul = 0;
23161 }
23162 break;
23163 case TCP_RACK_NO_PUSH_AT_MAX:
23164 RACK_OPTS_INC(tcp_npush);
23165 if (optval == 0)
23166 rack->r_ctl.rc_no_push_at_mrtt = 0;
23167 else if (optval < 0xff)
23168 rack->r_ctl.rc_no_push_at_mrtt = optval;
23169 else
23170 error = EINVAL;
23171 break;
23172 case TCP_SHARED_CWND_ENABLE:
23173 RACK_OPTS_INC(tcp_rack_scwnd);
23174 if (optval == 0)
23175 rack->rack_enable_scwnd = 0;
23176 else
23177 rack->rack_enable_scwnd = 1;
23178 break;
23179 case TCP_RACK_MBUF_QUEUE:
23180 /* Now do we use the LRO mbuf-queue feature */
23181 RACK_OPTS_INC(tcp_rack_mbufq);
23182 if (optval || rack->r_use_cmp_ack)
23183 rack->r_mbuf_queue = 1;
23184 else
23185 rack->r_mbuf_queue = 0;
23186 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23187 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23188 else
23189 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23190 break;
23191 case TCP_RACK_NONRXT_CFG_RATE:
23192 RACK_OPTS_INC(tcp_rack_cfg_rate);
23193 if (optval == 0)
23194 rack->rack_rec_nonrxt_use_cr = 0;
23195 else
23196 rack->rack_rec_nonrxt_use_cr = 1;
23197 break;
23198 case TCP_NO_PRR:
23199 RACK_OPTS_INC(tcp_rack_noprr);
23200 if (optval == 0)
23201 rack->rack_no_prr = 0;
23202 else if (optval == 1)
23203 rack->rack_no_prr = 1;
23204 else if (optval == 2)
23205 rack->no_prr_addback = 1;
23206 else
23207 error = EINVAL;
23208 break;
23209 case RACK_CSPR_IS_FCC: /* URL:csprisfcc */
23210 if (optval > 0)
23211 rack->cspr_is_fcc = 1;
23212 else
23213 rack->cspr_is_fcc = 0;
23214 break;
23215 case TCP_TIMELY_DYN_ADJ:
23216 RACK_OPTS_INC(tcp_timely_dyn);
23217 if (optval == 0)
23218 rack->rc_gp_dyn_mul = 0;
23219 else {
23220 rack->rc_gp_dyn_mul = 1;
23221 if (optval >= 100) {
23222 /*
23223 * If the user sets something 100 or more
23224 * its the gp_ca value.
23225 */
23226 rack->r_ctl.rack_per_of_gp_ca = optval;
23227 }
23228 }
23229 break;
23230 case TCP_RACK_DO_DETECTION:
23231 error = EINVAL;
23232 break;
23233 case TCP_RACK_TLP_USE:
23234 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23235 error = EINVAL;
23236 break;
23237 }
23238 RACK_OPTS_INC(tcp_tlp_use);
23239 rack->rack_tlp_threshold_use = optval;
23240 break;
23241 case TCP_RACK_TLP_REDUCE:
23242 /* RACK TLP cwnd reduction (bool) */
23243 RACK_OPTS_INC(tcp_rack_tlp_reduce);
23244 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23245 break;
23246 /* Pacing related ones */
23247 case TCP_RACK_PACE_ALWAYS:
23248 /*
23249 * zero is old rack method, 1 is new
23250 * method using a pacing rate.
23251 */
23252 RACK_OPTS_INC(tcp_rack_pace_always);
23253 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23254 error = EPERM;
23255 break;
23256 }
23257 if (optval > 0) {
23258 if (rack->rc_always_pace) {
23259 error = EALREADY;
23260 break;
23261 } else if (tcp_can_enable_pacing()) {
23262 rack->r_ctl.pacing_method |= RACK_REG_PACING;
23263 rack->rc_always_pace = 1;
23264 if (rack->rack_hibeta)
23265 rack_set_cc_pacing(rack);
23266 }
23267 else {
23268 error = ENOSPC;
23269 break;
23270 }
23271 } else {
23272 if (rack->rc_always_pace == 1) {
23273 rack_remove_pacing(rack);
23274 }
23275 }
23276 if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23277 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23278 else
23279 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23280 /* A rate may be set irate or other, if so set seg size */
23281 rack_update_seg(rack);
23282 break;
23283 case TCP_BBR_RACK_INIT_RATE:
23284 RACK_OPTS_INC(tcp_initial_rate);
23285 val = optval;
23286 /* Change from kbits per second to bytes per second */
23287 val *= 1000;
23288 val /= 8;
23289 rack->r_ctl.init_rate = val;
23290 if (rack->rc_always_pace)
23291 rack_update_seg(rack);
23292 break;
23293 case TCP_BBR_IWINTSO:
23294 error = EINVAL;
23295 break;
23296 case TCP_RACK_FORCE_MSEG:
23297 RACK_OPTS_INC(tcp_rack_force_max_seg);
23298 if (optval)
23299 rack->rc_force_max_seg = 1;
23300 else
23301 rack->rc_force_max_seg = 0;
23302 break;
23303 case TCP_RACK_PACE_MIN_SEG:
23304 RACK_OPTS_INC(tcp_rack_min_seg);
23305 rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23306 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23307 break;
23308 case TCP_RACK_PACE_MAX_SEG:
23309 /* Max segments size in a pace in bytes */
23310 RACK_OPTS_INC(tcp_rack_max_seg);
23311 if ((rack->dgp_on == 1) &&
23312 (rack->r_ctl.pacing_method & RACK_DGP_PACING)) {
23313 /*
23314 * If we set a max-seg and are doing DGP then
23315 * we now fall under the pacing limits not the
23316 * DGP ones.
23317 */
23318 if (tcp_can_enable_pacing() == 0) {
23319 error = ENOSPC;
23320 break;
23321 }
23322 /*
23323 * Now change up the flags and counts to be correct.
23324 */
23325 rack->r_ctl.pacing_method |= RACK_REG_PACING;
23326 tcp_dec_dgp_pacing_cnt();
23327 rack->r_ctl.pacing_method &= ~RACK_DGP_PACING;
23328 }
23329 if (optval <= MAX_USER_SET_SEG)
23330 rack->rc_user_set_max_segs = optval;
23331 else
23332 rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23333 rack_set_pace_segments(tp, rack, __LINE__, NULL);
23334 break;
23335 case TCP_RACK_PACE_RATE_REC:
23336 /* Set the fixed pacing rate in Bytes per second ca */
23337 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23338 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23339 error = EPERM;
23340 break;
23341 }
23342 if (rack->dgp_on) {
23343 /*
23344 * We are already pacing another
23345 * way.
23346 */
23347 error = EBUSY;
23348 break;
23349 }
23350 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23351 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23352 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23353 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23354 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23355 rack->use_fixed_rate = 1;
23356 if (rack->rack_hibeta)
23357 rack_set_cc_pacing(rack);
23358 rack_log_pacing_delay_calc(rack,
23359 rack->r_ctl.rc_fixed_pacing_rate_ss,
23360 rack->r_ctl.rc_fixed_pacing_rate_ca,
23361 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23362 __LINE__, NULL,0);
23363 break;
23364
23365 case TCP_RACK_PACE_RATE_SS:
23366 /* Set the fixed pacing rate in Bytes per second ca */
23367 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23368 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23369 error = EPERM;
23370 break;
23371 }
23372 if (rack->dgp_on) {
23373 /*
23374 * We are already pacing another
23375 * way.
23376 */
23377 error = EBUSY;
23378 break;
23379 }
23380 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23381 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23382 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23383 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23384 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23385 rack->use_fixed_rate = 1;
23386 if (rack->rack_hibeta)
23387 rack_set_cc_pacing(rack);
23388 rack_log_pacing_delay_calc(rack,
23389 rack->r_ctl.rc_fixed_pacing_rate_ss,
23390 rack->r_ctl.rc_fixed_pacing_rate_ca,
23391 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23392 __LINE__, NULL, 0);
23393 break;
23394
23395 case TCP_RACK_PACE_RATE_CA:
23396 /* Set the fixed pacing rate in Bytes per second ca */
23397 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23398 if (rack->r_ctl.side_chan_dis_mask & CCSP_DIS_MASK) {
23399 error = EPERM;
23400 break;
23401 }
23402 if (rack->dgp_on) {
23403 /*
23404 * We are already pacing another
23405 * way.
23406 */
23407 error = EBUSY;
23408 break;
23409 }
23410 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23411 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23412 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23413 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23414 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23415 rack->use_fixed_rate = 1;
23416 if (rack->rack_hibeta)
23417 rack_set_cc_pacing(rack);
23418 rack_log_pacing_delay_calc(rack,
23419 rack->r_ctl.rc_fixed_pacing_rate_ss,
23420 rack->r_ctl.rc_fixed_pacing_rate_ca,
23421 rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23422 __LINE__, NULL, 0);
23423 break;
23424 case TCP_RACK_GP_INCREASE_REC:
23425 RACK_OPTS_INC(tcp_gp_inc_rec);
23426 rack->r_ctl.rack_per_of_gp_rec = optval;
23427 rack_log_pacing_delay_calc(rack,
23428 rack->r_ctl.rack_per_of_gp_ss,
23429 rack->r_ctl.rack_per_of_gp_ca,
23430 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23431 __LINE__, NULL, 0);
23432 break;
23433 case TCP_RACK_GP_INCREASE_CA:
23434 RACK_OPTS_INC(tcp_gp_inc_ca);
23435 ca = optval;
23436 if (ca < 100) {
23437 /*
23438 * We don't allow any reduction
23439 * over the GP b/w.
23440 */
23441 error = EINVAL;
23442 break;
23443 }
23444 rack->r_ctl.rack_per_of_gp_ca = ca;
23445 rack_log_pacing_delay_calc(rack,
23446 rack->r_ctl.rack_per_of_gp_ss,
23447 rack->r_ctl.rack_per_of_gp_ca,
23448 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23449 __LINE__, NULL, 0);
23450 break;
23451 case TCP_RACK_GP_INCREASE_SS:
23452 RACK_OPTS_INC(tcp_gp_inc_ss);
23453 ss = optval;
23454 if (ss < 100) {
23455 /*
23456 * We don't allow any reduction
23457 * over the GP b/w.
23458 */
23459 error = EINVAL;
23460 break;
23461 }
23462 rack->r_ctl.rack_per_of_gp_ss = ss;
23463 rack_log_pacing_delay_calc(rack,
23464 rack->r_ctl.rack_per_of_gp_ss,
23465 rack->r_ctl.rack_per_of_gp_ca,
23466 rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23467 __LINE__, NULL, 0);
23468 break;
23469 case TCP_RACK_RR_CONF:
23470 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23471 if (optval && optval <= 3)
23472 rack->r_rr_config = optval;
23473 else
23474 rack->r_rr_config = 0;
23475 break;
23476 case TCP_PACING_DND: /* URL:dnd */
23477 if (optval > 0)
23478 rack->rc_pace_dnd = 1;
23479 else
23480 rack->rc_pace_dnd = 0;
23481 break;
23482 case TCP_HDWR_RATE_CAP:
23483 RACK_OPTS_INC(tcp_hdwr_rate_cap);
23484 if (optval) {
23485 if (rack->r_rack_hw_rate_caps == 0)
23486 rack->r_rack_hw_rate_caps = 1;
23487 else
23488 error = EALREADY;
23489 } else {
23490 rack->r_rack_hw_rate_caps = 0;
23491 }
23492 break;
23493 case TCP_DGP_UPPER_BOUNDS:
23494 {
23495 uint8_t val;
23496 val = optval & 0x0000ff;
23497 rack->r_ctl.rack_per_upper_bound_ca = val;
23498 val = (optval >> 16) & 0x0000ff;
23499 rack->r_ctl.rack_per_upper_bound_ss = val;
23500 break;
23501 }
23502 case TCP_SS_EEXIT: /* URL:eexit */
23503 if (optval > 0) {
23504 rack->r_ctl.gp_rnd_thresh = optval & 0x0ff;
23505 if (optval & 0x10000) {
23506 rack->r_ctl.gate_to_fs = 1;
23507 } else {
23508 rack->r_ctl.gate_to_fs = 0;
23509 }
23510 if (optval & 0x20000) {
23511 rack->r_ctl.use_gp_not_last = 1;
23512 } else {
23513 rack->r_ctl.use_gp_not_last = 0;
23514 }
23515 if (optval & 0xfffc0000) {
23516 uint32_t v;
23517
23518 v = (optval >> 18) & 0x00003fff;
23519 if (v >= 1000)
23520 rack->r_ctl.gp_gain_req = v;
23521 }
23522 } else {
23523 /* We do not do ss early exit at all */
23524 rack->rc_initial_ss_comp = 1;
23525 rack->r_ctl.gp_rnd_thresh = 0;
23526 }
23527 break;
23528 case TCP_RACK_SPLIT_LIMIT:
23529 RACK_OPTS_INC(tcp_split_limit);
23530 rack->r_ctl.rc_split_limit = optval;
23531 break;
23532 case TCP_BBR_HDWR_PACE:
23533 RACK_OPTS_INC(tcp_hdwr_pacing);
23534 if (optval){
23535 if (rack->rack_hdrw_pacing == 0) {
23536 rack->rack_hdw_pace_ena = 1;
23537 rack->rack_attempt_hdwr_pace = 0;
23538 } else
23539 error = EALREADY;
23540 } else {
23541 rack->rack_hdw_pace_ena = 0;
23542 #ifdef RATELIMIT
23543 if (rack->r_ctl.crte != NULL) {
23544 rack->rack_hdrw_pacing = 0;
23545 rack->rack_attempt_hdwr_pace = 0;
23546 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23547 rack->r_ctl.crte = NULL;
23548 }
23549 #endif
23550 }
23551 break;
23552 /* End Pacing related ones */
23553 case TCP_RACK_PRR_SENDALOT:
23554 /* Allow PRR to send more than one seg */
23555 RACK_OPTS_INC(tcp_rack_prr_sendalot);
23556 rack->r_ctl.rc_prr_sendalot = optval;
23557 break;
23558 case TCP_RACK_MIN_TO:
23559 /* Minimum time between rack t-o's in ms */
23560 RACK_OPTS_INC(tcp_rack_min_to);
23561 rack->r_ctl.rc_min_to = optval;
23562 break;
23563 case TCP_RACK_EARLY_SEG:
23564 /* If early recovery max segments */
23565 RACK_OPTS_INC(tcp_rack_early_seg);
23566 rack->r_ctl.rc_early_recovery_segs = optval;
23567 break;
23568 case TCP_RACK_ENABLE_HYSTART:
23569 {
23570 if (optval) {
23571 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23572 if (rack_do_hystart > RACK_HYSTART_ON)
23573 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23574 if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23575 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23576 } else {
23577 tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23578 }
23579 }
23580 break;
23581 case TCP_RACK_REORD_THRESH:
23582 /* RACK reorder threshold (shift amount) */
23583 RACK_OPTS_INC(tcp_rack_reord_thresh);
23584 if ((optval > 0) && (optval < 31))
23585 rack->r_ctl.rc_reorder_shift = optval;
23586 else
23587 error = EINVAL;
23588 break;
23589 case TCP_RACK_REORD_FADE:
23590 /* Does reordering fade after ms time */
23591 RACK_OPTS_INC(tcp_rack_reord_fade);
23592 rack->r_ctl.rc_reorder_fade = optval;
23593 break;
23594 case TCP_RACK_TLP_THRESH:
23595 /* RACK TLP theshold i.e. srtt+(srtt/N) */
23596 RACK_OPTS_INC(tcp_rack_tlp_thresh);
23597 if (optval)
23598 rack->r_ctl.rc_tlp_threshold = optval;
23599 else
23600 error = EINVAL;
23601 break;
23602 case TCP_BBR_USE_RACK_RR:
23603 RACK_OPTS_INC(tcp_rack_rr);
23604 if (optval)
23605 rack->use_rack_rr = 1;
23606 else
23607 rack->use_rack_rr = 0;
23608 break;
23609 case TCP_RACK_PKT_DELAY:
23610 /* RACK added ms i.e. rack-rtt + reord + N */
23611 RACK_OPTS_INC(tcp_rack_pkt_delay);
23612 rack->r_ctl.rc_pkt_delay = optval;
23613 break;
23614 case TCP_DELACK:
23615 RACK_OPTS_INC(tcp_rack_delayed_ack);
23616 if (optval == 0)
23617 tp->t_delayed_ack = 0;
23618 else
23619 tp->t_delayed_ack = 1;
23620 if (tp->t_flags & TF_DELACK) {
23621 tp->t_flags &= ~TF_DELACK;
23622 tp->t_flags |= TF_ACKNOW;
23623 NET_EPOCH_ENTER(et);
23624 rack_output(tp);
23625 NET_EPOCH_EXIT(et);
23626 }
23627 break;
23628
23629 case TCP_BBR_RACK_RTT_USE:
23630 RACK_OPTS_INC(tcp_rack_rtt_use);
23631 if ((optval != USE_RTT_HIGH) &&
23632 (optval != USE_RTT_LOW) &&
23633 (optval != USE_RTT_AVG))
23634 error = EINVAL;
23635 else
23636 rack->r_ctl.rc_rate_sample_method = optval;
23637 break;
23638 case TCP_HONOR_HPTS_MIN:
23639 RACK_OPTS_INC(tcp_honor_hpts);
23640 if (optval) {
23641 rack->r_use_hpts_min = 1;
23642 /*
23643 * Must be between 2 - 80% to be a reduction else
23644 * we keep the default (10%).
23645 */
23646 if ((optval > 1) && (optval <= 80)) {
23647 rack->r_ctl.max_reduction = optval;
23648 }
23649 } else
23650 rack->r_use_hpts_min = 0;
23651 break;
23652 case TCP_REC_IS_DYN: /* URL:dynrec */
23653 RACK_OPTS_INC(tcp_dyn_rec);
23654 if (optval)
23655 rack->rc_gp_no_rec_chg = 1;
23656 else
23657 rack->rc_gp_no_rec_chg = 0;
23658 break;
23659 case TCP_NO_TIMELY:
23660 RACK_OPTS_INC(tcp_notimely);
23661 if (optval) {
23662 rack->rc_skip_timely = 1;
23663 rack->r_ctl.rack_per_of_gp_rec = 90;
23664 rack->r_ctl.rack_per_of_gp_ca = 100;
23665 rack->r_ctl.rack_per_of_gp_ss = 250;
23666 } else {
23667 rack->rc_skip_timely = 0;
23668 }
23669 break;
23670 case TCP_GP_USE_LTBW:
23671 if (optval == 0) {
23672 rack->use_lesser_lt_bw = 0;
23673 rack->dis_lt_bw = 1;
23674 } else if (optval == 1) {
23675 rack->use_lesser_lt_bw = 1;
23676 rack->dis_lt_bw = 0;
23677 } else if (optval == 2) {
23678 rack->use_lesser_lt_bw = 0;
23679 rack->dis_lt_bw = 0;
23680 }
23681 break;
23682 case TCP_DATA_AFTER_CLOSE:
23683 RACK_OPTS_INC(tcp_data_after_close);
23684 if (optval)
23685 rack->rc_allow_data_af_clo = 1;
23686 else
23687 rack->rc_allow_data_af_clo = 0;
23688 break;
23689 default:
23690 break;
23691 }
23692 tcp_log_socket_option(tp, sopt_name, optval, error);
23693 return (error);
23694 }
23695
23696 static void
rack_inherit(struct tcpcb * tp,struct inpcb * parent)23697 rack_inherit(struct tcpcb *tp, struct inpcb *parent)
23698 {
23699 /*
23700 * A new connection has been created (tp) and
23701 * the parent is the inpcb given. We want to
23702 * apply a read-lock to the parent (we are already
23703 * holding a write lock on the tp) and copy anything
23704 * out of the rack specific data as long as its tfb is
23705 * the same as ours i.e. we are the same stack. Otherwise
23706 * we just return.
23707 */
23708 struct tcpcb *par;
23709 struct tcp_rack *dest, *src;
23710 int cnt = 0;
23711
23712 par = intotcpcb(parent);
23713 if (par->t_fb != tp->t_fb) {
23714 /* Not the same stack */
23715 tcp_log_socket_option(tp, 0, 0, 1);
23716 return;
23717 }
23718 /* Ok if we reach here lets setup the two rack pointers */
23719 dest = (struct tcp_rack *)tp->t_fb_ptr;
23720 src = (struct tcp_rack *)par->t_fb_ptr;
23721 if ((src == NULL) || (dest == NULL)) {
23722 /* Huh? */
23723 tcp_log_socket_option(tp, 0, 0, 2);
23724 return;
23725 }
23726 /* Now copy out anything we wish to inherit i.e. things in socket-options */
23727 /* TCP_RACK_PROFILE we can't know but we can set DGP if its on */
23728 if ((src->dgp_on) && (dest->dgp_on == 0)) {
23729 /* Profile 1 had to be set via sock opt */
23730 rack_set_dgp(dest);
23731 cnt++;
23732 }
23733 /* TCP_RACK_SET_RXT_OPTIONS */
23734 if (dest->full_size_rxt != src->full_size_rxt) {
23735 dest->full_size_rxt = src->full_size_rxt;
23736 cnt++;
23737 }
23738 if (dest->shape_rxt_to_pacing_min != src->shape_rxt_to_pacing_min) {
23739 dest->shape_rxt_to_pacing_min = src->shape_rxt_to_pacing_min;
23740 cnt++;
23741 }
23742 /* TCP_RACK_DSACK_OPT */
23743 if (dest->rc_rack_tmr_std_based != src->rc_rack_tmr_std_based) {
23744 dest->rc_rack_tmr_std_based = src->rc_rack_tmr_std_based;
23745 cnt++;
23746 }
23747 if (dest->rc_rack_use_dsack != src->rc_rack_use_dsack) {
23748 dest->rc_rack_use_dsack = src->rc_rack_use_dsack;
23749 cnt++;
23750 }
23751 /* TCP_RACK_PACING_DIVISOR */
23752 if (dest->r_ctl.pace_len_divisor != src->r_ctl.pace_len_divisor) {
23753 dest->r_ctl.pace_len_divisor = src->r_ctl.pace_len_divisor;
23754 cnt++;
23755 }
23756 /* TCP_RACK_HI_BETA */
23757 if (src->rack_hibeta != dest->rack_hibeta) {
23758 cnt++;
23759 if (src->rack_hibeta) {
23760 dest->r_ctl.rc_saved_beta = src->r_ctl.rc_saved_beta;
23761 dest->rack_hibeta = 1;
23762 } else {
23763 dest->rack_hibeta = 0;
23764 }
23765 }
23766 /* TCP_RACK_TIMER_SLOP */
23767 if (dest->r_ctl.timer_slop != src->r_ctl.timer_slop) {
23768 dest->r_ctl.timer_slop = src->r_ctl.timer_slop;
23769 cnt++;
23770 }
23771 /* TCP_RACK_PACING_BETA_ECN */
23772 if (dest->r_ctl.rc_saved_beta_ecn != src->r_ctl.rc_saved_beta_ecn) {
23773 dest->r_ctl.rc_saved_beta_ecn = src->r_ctl.rc_saved_beta_ecn;
23774 cnt++;
23775 }
23776 /* We do not do TCP_DEFER_OPTIONS */
23777 /* TCP_RACK_MEASURE_CNT */
23778 if (dest->r_ctl.req_measurements != src->r_ctl.req_measurements) {
23779 dest->r_ctl.req_measurements = src->r_ctl.req_measurements;
23780 cnt++;
23781 }
23782 /* TCP_HDWR_UP_ONLY */
23783 if (dest->r_up_only != src->r_up_only) {
23784 dest->r_up_only = src->r_up_only;
23785 cnt++;
23786 }
23787 /* TCP_FILLCW_RATE_CAP */
23788 if (dest->r_ctl.fillcw_cap != src->r_ctl.fillcw_cap) {
23789 dest->r_ctl.fillcw_cap = src->r_ctl.fillcw_cap;
23790 cnt++;
23791 }
23792 /* TCP_PACING_RATE_CAP */
23793 if (dest->r_ctl.bw_rate_cap != src->r_ctl.bw_rate_cap) {
23794 dest->r_ctl.bw_rate_cap = src->r_ctl.bw_rate_cap;
23795 cnt++;
23796 }
23797 /* A listener can't set TCP_HYBRID_PACING */
23798 /* TCP_SIDECHAN_DIS */
23799 if (dest->r_ctl.side_chan_dis_mask != src->r_ctl.side_chan_dis_mask) {
23800 dest->r_ctl.side_chan_dis_mask = src->r_ctl.side_chan_dis_mask;
23801 cnt++;
23802 }
23803 /* TCP_SHARED_CWND_TIME_LIMIT */
23804 if (dest->r_limit_scw != src->r_limit_scw) {
23805 dest->r_limit_scw = src->r_limit_scw;
23806 cnt++;
23807 }
23808 /* TCP_RACK_PACE_TO_FILL */
23809 if (dest->rc_pace_to_cwnd != src->rc_pace_to_cwnd) {
23810 dest->rc_pace_to_cwnd = src->rc_pace_to_cwnd;
23811 cnt++;
23812 }
23813 if (dest->rc_pace_fill_if_rttin_range != src->rc_pace_fill_if_rttin_range) {
23814 dest->rc_pace_fill_if_rttin_range = src->rc_pace_fill_if_rttin_range;
23815 cnt++;
23816 }
23817 if (dest->rtt_limit_mul != src->rtt_limit_mul) {
23818 dest->rtt_limit_mul = src->rtt_limit_mul;
23819 cnt++;
23820 }
23821 /* TCP_RACK_NO_PUSH_AT_MAX */
23822 if (dest->r_ctl.rc_no_push_at_mrtt != src->r_ctl.rc_no_push_at_mrtt) {
23823 dest->r_ctl.rc_no_push_at_mrtt = src->r_ctl.rc_no_push_at_mrtt;
23824 cnt++;
23825 }
23826 /* TCP_SHARED_CWND_ENABLE */
23827 if (dest->rack_enable_scwnd != src->rack_enable_scwnd) {
23828 dest->rack_enable_scwnd = src->rack_enable_scwnd;
23829 cnt++;
23830 }
23831 /* TCP_USE_CMP_ACKS */
23832 if (dest->r_use_cmp_ack != src->r_use_cmp_ack) {
23833 dest->r_use_cmp_ack = src->r_use_cmp_ack;
23834 cnt++;
23835 }
23836
23837 if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23838 dest->r_mbuf_queue = src->r_mbuf_queue;
23839 cnt++;
23840 }
23841 /* TCP_RACK_MBUF_QUEUE */
23842 if (dest->r_mbuf_queue != src->r_mbuf_queue) {
23843 dest->r_mbuf_queue = src->r_mbuf_queue;
23844 cnt++;
23845 }
23846 if (dest->r_mbuf_queue || dest->rc_always_pace || dest->r_use_cmp_ack) {
23847 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23848 } else {
23849 tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23850 }
23851 if (dest->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) {
23852 tp->t_flags2 |= TF2_MBUF_ACKCMP;
23853 }
23854 /* TCP_RACK_NONRXT_CFG_RATE */
23855 if (dest->rack_rec_nonrxt_use_cr != src->rack_rec_nonrxt_use_cr) {
23856 dest->rack_rec_nonrxt_use_cr = src->rack_rec_nonrxt_use_cr;
23857 cnt++;
23858 }
23859 /* TCP_NO_PRR */
23860 if (dest->rack_no_prr != src->rack_no_prr) {
23861 dest->rack_no_prr = src->rack_no_prr;
23862 cnt++;
23863 }
23864 if (dest->no_prr_addback != src->no_prr_addback) {
23865 dest->no_prr_addback = src->no_prr_addback;
23866 cnt++;
23867 }
23868 /* RACK_CSPR_IS_FCC */
23869 if (dest->cspr_is_fcc != src->cspr_is_fcc) {
23870 dest->cspr_is_fcc = src->cspr_is_fcc;
23871 cnt++;
23872 }
23873 /* TCP_TIMELY_DYN_ADJ */
23874 if (dest->rc_gp_dyn_mul != src->rc_gp_dyn_mul) {
23875 dest->rc_gp_dyn_mul = src->rc_gp_dyn_mul;
23876 cnt++;
23877 }
23878 if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23879 dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23880 cnt++;
23881 }
23882 /* TCP_RACK_TLP_USE */
23883 if (dest->rack_tlp_threshold_use != src->rack_tlp_threshold_use) {
23884 dest->rack_tlp_threshold_use = src->rack_tlp_threshold_use;
23885 cnt++;
23886 }
23887 /* we don't allow inheritence of TCP_RACK_PACE_ALWAYS */
23888 /* TCP_BBR_RACK_INIT_RATE */
23889 if (dest->r_ctl.init_rate != src->r_ctl.init_rate) {
23890 dest->r_ctl.init_rate = src->r_ctl.init_rate;
23891 cnt++;
23892 }
23893 /* TCP_RACK_FORCE_MSEG */
23894 if (dest->rc_force_max_seg != src->rc_force_max_seg) {
23895 dest->rc_force_max_seg = src->rc_force_max_seg;
23896 cnt++;
23897 }
23898 /* TCP_RACK_PACE_MIN_SEG */
23899 if (dest->r_ctl.rc_user_set_min_segs != src->r_ctl.rc_user_set_min_segs) {
23900 dest->r_ctl.rc_user_set_min_segs = src->r_ctl.rc_user_set_min_segs;
23901 cnt++;
23902 }
23903 /* we don't allow TCP_RACK_PACE_MAX_SEG */
23904 /* TCP_RACK_PACE_RATE_REC, TCP_RACK_PACE_RATE_SS, TCP_RACK_PACE_RATE_CA */
23905 if (dest->r_ctl.rc_fixed_pacing_rate_ca != src->r_ctl.rc_fixed_pacing_rate_ca) {
23906 dest->r_ctl.rc_fixed_pacing_rate_ca = src->r_ctl.rc_fixed_pacing_rate_ca;
23907 cnt++;
23908 }
23909 if (dest->r_ctl.rc_fixed_pacing_rate_ss != src->r_ctl.rc_fixed_pacing_rate_ss) {
23910 dest->r_ctl.rc_fixed_pacing_rate_ss = src->r_ctl.rc_fixed_pacing_rate_ss;
23911 cnt++;
23912 }
23913 if (dest->r_ctl.rc_fixed_pacing_rate_rec != src->r_ctl.rc_fixed_pacing_rate_rec) {
23914 dest->r_ctl.rc_fixed_pacing_rate_rec = src->r_ctl.rc_fixed_pacing_rate_rec;
23915 cnt++;
23916 }
23917 /* TCP_RACK_GP_INCREASE_REC, TCP_RACK_GP_INCREASE_CA, TCP_RACK_GP_INCREASE_SS */
23918 if (dest->r_ctl.rack_per_of_gp_rec != src->r_ctl.rack_per_of_gp_rec) {
23919 dest->r_ctl.rack_per_of_gp_rec = src->r_ctl.rack_per_of_gp_rec;
23920 cnt++;
23921 }
23922 if (dest->r_ctl.rack_per_of_gp_ca != src->r_ctl.rack_per_of_gp_ca) {
23923 dest->r_ctl.rack_per_of_gp_ca = src->r_ctl.rack_per_of_gp_ca;
23924 cnt++;
23925 }
23926
23927 if (dest->r_ctl.rack_per_of_gp_ss != src->r_ctl.rack_per_of_gp_ss) {
23928 dest->r_ctl.rack_per_of_gp_ss = src->r_ctl.rack_per_of_gp_ss;
23929 cnt++;
23930 }
23931 /* TCP_RACK_RR_CONF */
23932 if (dest->r_rr_config != src->r_rr_config) {
23933 dest->r_rr_config = src->r_rr_config;
23934 cnt++;
23935 }
23936 /* TCP_PACING_DND */
23937 if (dest->rc_pace_dnd != src->rc_pace_dnd) {
23938 dest->rc_pace_dnd = src->rc_pace_dnd;
23939 cnt++;
23940 }
23941 /* TCP_HDWR_RATE_CAP */
23942 if (dest->r_rack_hw_rate_caps != src->r_rack_hw_rate_caps) {
23943 dest->r_rack_hw_rate_caps = src->r_rack_hw_rate_caps;
23944 cnt++;
23945 }
23946 /* TCP_DGP_UPPER_BOUNDS */
23947 if (dest->r_ctl.rack_per_upper_bound_ca != src->r_ctl.rack_per_upper_bound_ca) {
23948 dest->r_ctl.rack_per_upper_bound_ca = src->r_ctl.rack_per_upper_bound_ca;
23949 cnt++;
23950 }
23951 if (dest->r_ctl.rack_per_upper_bound_ss != src->r_ctl.rack_per_upper_bound_ss) {
23952 dest->r_ctl.rack_per_upper_bound_ss = src->r_ctl.rack_per_upper_bound_ss;
23953 cnt++;
23954 }
23955 /* TCP_SS_EEXIT */
23956 if (dest->r_ctl.gp_rnd_thresh != src->r_ctl.gp_rnd_thresh) {
23957 dest->r_ctl.gp_rnd_thresh = src->r_ctl.gp_rnd_thresh;
23958 cnt++;
23959 }
23960 if (dest->r_ctl.gate_to_fs != src->r_ctl.gate_to_fs) {
23961 dest->r_ctl.gate_to_fs = src->r_ctl.gate_to_fs;
23962 cnt++;
23963 }
23964 if (dest->r_ctl.use_gp_not_last != src->r_ctl.use_gp_not_last) {
23965 dest->r_ctl.use_gp_not_last = src->r_ctl.use_gp_not_last;
23966 cnt++;
23967 }
23968 if (dest->r_ctl.gp_gain_req != src->r_ctl.gp_gain_req) {
23969 dest->r_ctl.gp_gain_req = src->r_ctl.gp_gain_req;
23970 cnt++;
23971 }
23972 /* TCP_BBR_HDWR_PACE */
23973 if (dest->rack_hdw_pace_ena != src->rack_hdw_pace_ena) {
23974 dest->rack_hdw_pace_ena = src->rack_hdw_pace_ena;
23975 cnt++;
23976 }
23977 if (dest->rack_attempt_hdwr_pace != src->rack_attempt_hdwr_pace) {
23978 dest->rack_attempt_hdwr_pace = src->rack_attempt_hdwr_pace;
23979 cnt++;
23980 }
23981 /* TCP_RACK_PRR_SENDALOT */
23982 if (dest->r_ctl.rc_prr_sendalot != src->r_ctl.rc_prr_sendalot) {
23983 dest->r_ctl.rc_prr_sendalot = src->r_ctl.rc_prr_sendalot;
23984 cnt++;
23985 }
23986 /* TCP_RACK_MIN_TO */
23987 if (dest->r_ctl.rc_min_to != src->r_ctl.rc_min_to) {
23988 dest->r_ctl.rc_min_to = src->r_ctl.rc_min_to;
23989 cnt++;
23990 }
23991 /* TCP_RACK_EARLY_SEG */
23992 if (dest->r_ctl.rc_early_recovery_segs != src->r_ctl.rc_early_recovery_segs) {
23993 dest->r_ctl.rc_early_recovery_segs = src->r_ctl.rc_early_recovery_segs;
23994 cnt++;
23995 }
23996 /* TCP_RACK_ENABLE_HYSTART */
23997 if (par->t_ccv.flags != tp->t_ccv.flags) {
23998 cnt++;
23999 if (par->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24000 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
24001 if (rack_do_hystart > RACK_HYSTART_ON)
24002 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
24003 if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
24004 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
24005 } else {
24006 tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
24007 }
24008 }
24009 /* TCP_RACK_REORD_THRESH */
24010 if (dest->r_ctl.rc_reorder_shift != src->r_ctl.rc_reorder_shift) {
24011 dest->r_ctl.rc_reorder_shift = src->r_ctl.rc_reorder_shift;
24012 cnt++;
24013 }
24014 /* TCP_RACK_REORD_FADE */
24015 if (dest->r_ctl.rc_reorder_fade != src->r_ctl.rc_reorder_fade) {
24016 dest->r_ctl.rc_reorder_fade = src->r_ctl.rc_reorder_fade;
24017 cnt++;
24018 }
24019 /* TCP_RACK_TLP_THRESH */
24020 if (dest->r_ctl.rc_tlp_threshold != src->r_ctl.rc_tlp_threshold) {
24021 dest->r_ctl.rc_tlp_threshold = src->r_ctl.rc_tlp_threshold;
24022 cnt++;
24023 }
24024 /* TCP_BBR_USE_RACK_RR */
24025 if (dest->use_rack_rr != src->use_rack_rr) {
24026 dest->use_rack_rr = src->use_rack_rr;
24027 cnt++;
24028 }
24029 /* TCP_RACK_PKT_DELAY */
24030 if (dest->r_ctl.rc_pkt_delay != src->r_ctl.rc_pkt_delay) {
24031 dest->r_ctl.rc_pkt_delay = src->r_ctl.rc_pkt_delay;
24032 cnt++;
24033 }
24034 /* TCP_DELACK will get copied via the main code if applicable */
24035 /* TCP_BBR_RACK_RTT_USE */
24036 if (dest->r_ctl.rc_rate_sample_method != src->r_ctl.rc_rate_sample_method) {
24037 dest->r_ctl.rc_rate_sample_method = src->r_ctl.rc_rate_sample_method;
24038 cnt++;
24039 }
24040 /* TCP_HONOR_HPTS_MIN */
24041 if (dest->r_use_hpts_min != src->r_use_hpts_min) {
24042 dest->r_use_hpts_min = src->r_use_hpts_min;
24043 cnt++;
24044 }
24045 if (dest->r_ctl.max_reduction != src->r_ctl.max_reduction) {
24046 dest->r_ctl.max_reduction = src->r_ctl.max_reduction;
24047 cnt++;
24048 }
24049 /* TCP_REC_IS_DYN */
24050 if (dest->rc_gp_no_rec_chg != src->rc_gp_no_rec_chg) {
24051 dest->rc_gp_no_rec_chg = src->rc_gp_no_rec_chg;
24052 cnt++;
24053 }
24054 if (dest->rc_skip_timely != src->rc_skip_timely) {
24055 dest->rc_skip_timely = src->rc_skip_timely;
24056 cnt++;
24057 }
24058 /* TCP_DATA_AFTER_CLOSE */
24059 if (dest->rc_allow_data_af_clo != src->rc_allow_data_af_clo) {
24060 dest->rc_allow_data_af_clo = src->rc_allow_data_af_clo;
24061 cnt++;
24062 }
24063 /* TCP_GP_USE_LTBW */
24064 if (src->use_lesser_lt_bw != dest->use_lesser_lt_bw) {
24065 dest->use_lesser_lt_bw = src->use_lesser_lt_bw;
24066 cnt++;
24067 }
24068 if (dest->dis_lt_bw != src->dis_lt_bw) {
24069 dest->dis_lt_bw = src->dis_lt_bw;
24070 cnt++;
24071 }
24072 tcp_log_socket_option(tp, 0, cnt, 0);
24073 }
24074
24075
24076 static void
rack_apply_deferred_options(struct tcp_rack * rack)24077 rack_apply_deferred_options(struct tcp_rack *rack)
24078 {
24079 struct deferred_opt_list *dol, *sdol;
24080 uint32_t s_optval;
24081
24082 TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
24083 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
24084 /* Disadvantage of deferal is you loose the error return */
24085 s_optval = (uint32_t)dol->optval;
24086 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
24087 free(dol, M_TCPDO);
24088 }
24089 }
24090
24091 static void
rack_hw_tls_change(struct tcpcb * tp,int chg)24092 rack_hw_tls_change(struct tcpcb *tp, int chg)
24093 {
24094 /* Update HW tls state */
24095 struct tcp_rack *rack;
24096
24097 rack = (struct tcp_rack *)tp->t_fb_ptr;
24098 if (chg)
24099 rack->r_ctl.fsb.hw_tls = 1;
24100 else
24101 rack->r_ctl.fsb.hw_tls = 0;
24102 }
24103
24104 static int
rack_pru_options(struct tcpcb * tp,int flags)24105 rack_pru_options(struct tcpcb *tp, int flags)
24106 {
24107 if (flags & PRUS_OOB)
24108 return (EOPNOTSUPP);
24109 return (0);
24110 }
24111
24112 static bool
rack_wake_check(struct tcpcb * tp)24113 rack_wake_check(struct tcpcb *tp)
24114 {
24115 struct tcp_rack *rack;
24116 struct timeval tv;
24117 uint32_t cts;
24118
24119 rack = (struct tcp_rack *)tp->t_fb_ptr;
24120 if (rack->r_ctl.rc_hpts_flags) {
24121 cts = tcp_get_usecs(&tv);
24122 if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
24123 /*
24124 * Pacing timer is up, check if we are ready.
24125 */
24126 if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
24127 return (true);
24128 } else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
24129 /*
24130 * A timer is up, check if we are ready.
24131 */
24132 if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
24133 return (true);
24134 }
24135 }
24136 return (false);
24137 }
24138
24139 static struct tcp_function_block __tcp_rack = {
24140 .tfb_tcp_block_name = __XSTRING(STACKNAME),
24141 .tfb_tcp_output = rack_output,
24142 .tfb_do_queued_segments = ctf_do_queued_segments,
24143 .tfb_do_segment_nounlock = rack_do_segment_nounlock,
24144 .tfb_tcp_do_segment = rack_do_segment,
24145 .tfb_tcp_ctloutput = rack_ctloutput,
24146 .tfb_tcp_fb_init = rack_init,
24147 .tfb_tcp_fb_fini = rack_fini,
24148 .tfb_tcp_timer_stop_all = rack_stopall,
24149 .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
24150 .tfb_tcp_handoff_ok = rack_handoff_ok,
24151 .tfb_tcp_mtu_chg = rack_mtu_change,
24152 .tfb_pru_options = rack_pru_options,
24153 .tfb_hwtls_change = rack_hw_tls_change,
24154 .tfb_chg_query = rack_chg_query,
24155 .tfb_switch_failed = rack_switch_failed,
24156 .tfb_early_wake_check = rack_wake_check,
24157 .tfb_compute_pipe = rack_compute_pipe,
24158 .tfb_stack_info = rack_stack_information,
24159 .tfb_inherit = rack_inherit,
24160 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP | TCP_FUNC_DEFAULT_OK,
24161
24162 };
24163
24164 /*
24165 * rack_ctloutput() must drop the inpcb lock before performing copyin on
24166 * socket option arguments. When it re-acquires the lock after the copy, it
24167 * has to revalidate that the connection is still valid for the socket
24168 * option.
24169 */
24170 static int
rack_set_sockopt(struct tcpcb * tp,struct sockopt * sopt)24171 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24172 {
24173 struct inpcb *inp = tptoinpcb(tp);
24174 #ifdef INET
24175 struct ip *ip;
24176 #endif
24177 struct tcp_rack *rack;
24178 struct tcp_hybrid_req hybrid;
24179 uint64_t loptval;
24180 int32_t error = 0, optval;
24181
24182 rack = (struct tcp_rack *)tp->t_fb_ptr;
24183 if (rack == NULL) {
24184 INP_WUNLOCK(inp);
24185 return (EINVAL);
24186 }
24187 #ifdef INET
24188 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
24189 #endif
24190
24191 switch (sopt->sopt_level) {
24192 #ifdef INET6
24193 case IPPROTO_IPV6:
24194 MPASS(inp->inp_vflag & INP_IPV6PROTO);
24195 switch (sopt->sopt_name) {
24196 case IPV6_USE_MIN_MTU:
24197 tcp6_use_min_mtu(tp);
24198 break;
24199 }
24200 INP_WUNLOCK(inp);
24201 return (0);
24202 #endif
24203 #ifdef INET
24204 case IPPROTO_IP:
24205 switch (sopt->sopt_name) {
24206 case IP_TOS:
24207 /*
24208 * The DSCP codepoint has changed, update the fsb.
24209 */
24210 ip->ip_tos = rack->rc_inp->inp_ip_tos;
24211 break;
24212 case IP_TTL:
24213 /*
24214 * The TTL has changed, update the fsb.
24215 */
24216 ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
24217 break;
24218 }
24219 INP_WUNLOCK(inp);
24220 return (0);
24221 #endif
24222 #ifdef SO_PEERPRIO
24223 case SOL_SOCKET:
24224 switch (sopt->sopt_name) {
24225 case SO_PEERPRIO: /* SC-URL:bs */
24226 /* Already read in and sanity checked in sosetopt(). */
24227 if (inp->inp_socket) {
24228 rack->client_bufferlvl = inp->inp_socket->so_peerprio;
24229 }
24230 break;
24231 }
24232 INP_WUNLOCK(inp);
24233 return (0);
24234 #endif
24235 case IPPROTO_TCP:
24236 switch (sopt->sopt_name) {
24237 case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */
24238 /* Pacing related ones */
24239 case TCP_RACK_PACE_ALWAYS: /* URL:pace_always */
24240 case TCP_BBR_RACK_INIT_RATE: /* URL:irate */
24241 case TCP_RACK_PACE_MIN_SEG: /* URL:pace_min_seg */
24242 case TCP_RACK_PACE_MAX_SEG: /* URL:pace_max_seg */
24243 case TCP_RACK_FORCE_MSEG: /* URL:force_max_seg */
24244 case TCP_RACK_PACE_RATE_CA: /* URL:pr_ca */
24245 case TCP_RACK_PACE_RATE_SS: /* URL:pr_ss*/
24246 case TCP_RACK_PACE_RATE_REC: /* URL:pr_rec */
24247 case TCP_RACK_GP_INCREASE_CA: /* URL:gp_inc_ca */
24248 case TCP_RACK_GP_INCREASE_SS: /* URL:gp_inc_ss */
24249 case TCP_RACK_GP_INCREASE_REC: /* URL:gp_inc_rec */
24250 case TCP_RACK_RR_CONF: /* URL:rrr_conf */
24251 case TCP_BBR_HDWR_PACE: /* URL:hdwrpace */
24252 case TCP_HDWR_RATE_CAP: /* URL:hdwrcap boolean */
24253 case TCP_PACING_RATE_CAP: /* URL:cap -- used by side-channel */
24254 case TCP_HDWR_UP_ONLY: /* URL:uponly -- hardware pacing boolean */
24255 case TCP_FILLCW_RATE_CAP: /* URL:fillcw_cap */
24256 case TCP_RACK_PACING_BETA_ECN: /* URL:pacing_beta_ecn */
24257 case TCP_RACK_PACE_TO_FILL: /* URL:fillcw */
24258 /* End pacing related */
24259 case TCP_DELACK: /* URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
24260 case TCP_RACK_PRR_SENDALOT: /* URL:prr_sendalot */
24261 case TCP_RACK_MIN_TO: /* URL:min_to */
24262 case TCP_RACK_EARLY_SEG: /* URL:early_seg */
24263 case TCP_RACK_REORD_THRESH: /* URL:reord_thresh */
24264 case TCP_RACK_REORD_FADE: /* URL:reord_fade */
24265 case TCP_RACK_TLP_THRESH: /* URL:tlp_thresh */
24266 case TCP_RACK_PKT_DELAY: /* URL:pkt_delay */
24267 case TCP_RACK_TLP_USE: /* URL:tlp_use */
24268 case TCP_BBR_RACK_RTT_USE: /* URL:rttuse */
24269 case TCP_BBR_USE_RACK_RR: /* URL:rackrr */
24270 case TCP_NO_PRR: /* URL:noprr */
24271 case TCP_TIMELY_DYN_ADJ: /* URL:dynamic */
24272 case TCP_DATA_AFTER_CLOSE: /* no URL */
24273 case TCP_RACK_NONRXT_CFG_RATE: /* URL:nonrxtcr */
24274 case TCP_SHARED_CWND_ENABLE: /* URL:scwnd */
24275 case TCP_RACK_MBUF_QUEUE: /* URL:mqueue */
24276 case TCP_RACK_NO_PUSH_AT_MAX: /* URL:npush */
24277 case TCP_SHARED_CWND_TIME_LIMIT: /* URL:lscwnd */
24278 case TCP_RACK_PROFILE: /* URL:profile */
24279 case TCP_SIDECHAN_DIS: /* URL:scodm */
24280 case TCP_HYBRID_PACING: /* URL:pacing=hybrid */
24281 case TCP_USE_CMP_ACKS: /* URL:cmpack */
24282 case TCP_RACK_ABC_VAL: /* URL:labc */
24283 case TCP_REC_ABC_VAL: /* URL:reclabc */
24284 case TCP_RACK_MEASURE_CNT: /* URL:measurecnt */
24285 case TCP_DEFER_OPTIONS: /* URL:defer */
24286 case TCP_RACK_DSACK_OPT: /* URL:dsack */
24287 case TCP_RACK_TIMER_SLOP: /* URL:timer_slop */
24288 case TCP_RACK_ENABLE_HYSTART: /* URL:hystart */
24289 case TCP_RACK_SET_RXT_OPTIONS: /* URL:rxtsz */
24290 case TCP_RACK_HI_BETA: /* URL:hibeta */
24291 case TCP_RACK_SPLIT_LIMIT: /* URL:split */
24292 case TCP_SS_EEXIT: /* URL:eexit */
24293 case TCP_DGP_UPPER_BOUNDS: /* URL:upper */
24294 case TCP_RACK_PACING_DIVISOR: /* URL:divisor */
24295 case TCP_PACING_DND: /* URL:dnd */
24296 case TCP_NO_TIMELY: /* URL:notimely */
24297 case RACK_CSPR_IS_FCC: /* URL:csprisfcc */
24298 case TCP_HONOR_HPTS_MIN: /* URL:hptsmin */
24299 case TCP_REC_IS_DYN: /* URL:dynrec */
24300 case TCP_GP_USE_LTBW: /* URL:useltbw */
24301 goto process_opt;
24302 break;
24303 default:
24304 /* Filter off all unknown options to the base stack */
24305 return (tcp_default_ctloutput(tp, sopt));
24306 break;
24307 }
24308 default:
24309 INP_WUNLOCK(inp);
24310 return (0);
24311 }
24312 process_opt:
24313 INP_WUNLOCK(inp);
24314 if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24315 (sopt->sopt_name == TCP_FILLCW_RATE_CAP)) {
24316 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
24317 /*
24318 * We truncate it down to 32 bits for the socket-option trace this
24319 * means rates > 34Gbps won't show right, but thats probably ok.
24320 */
24321 optval = (uint32_t)loptval;
24322 } else if (sopt->sopt_name == TCP_HYBRID_PACING) {
24323 error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
24324 } else {
24325 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
24326 /* Save it in 64 bit form too */
24327 loptval = optval;
24328 }
24329 if (error)
24330 return (error);
24331 INP_WLOCK(inp);
24332 if (tp->t_fb != &__tcp_rack) {
24333 INP_WUNLOCK(inp);
24334 return (ENOPROTOOPT);
24335 }
24336 if (rack->defer_options && (rack->gp_ready == 0) &&
24337 (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
24338 (sopt->sopt_name != TCP_HYBRID_PACING) &&
24339 (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
24340 (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
24341 (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
24342 /* Options are being deferred */
24343 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
24344 INP_WUNLOCK(inp);
24345 return (0);
24346 } else {
24347 /* No memory to defer, fail */
24348 INP_WUNLOCK(inp);
24349 return (ENOMEM);
24350 }
24351 }
24352 error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
24353 INP_WUNLOCK(inp);
24354 return (error);
24355 }
24356
24357 static void
rack_fill_info(struct tcpcb * tp,struct tcp_info * ti)24358 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
24359 {
24360
24361 INP_WLOCK_ASSERT(tptoinpcb(tp));
24362 bzero(ti, sizeof(*ti));
24363
24364 ti->tcpi_state = tp->t_state;
24365 if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
24366 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
24367 if (tp->t_flags & TF_SACK_PERMIT)
24368 ti->tcpi_options |= TCPI_OPT_SACK;
24369 if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
24370 ti->tcpi_options |= TCPI_OPT_WSCALE;
24371 ti->tcpi_snd_wscale = tp->snd_scale;
24372 ti->tcpi_rcv_wscale = tp->rcv_scale;
24373 }
24374 if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
24375 ti->tcpi_options |= TCPI_OPT_ECN;
24376 if (tp->t_flags & TF_FASTOPEN)
24377 ti->tcpi_options |= TCPI_OPT_TFO;
24378 /* still kept in ticks is t_rcvtime */
24379 ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
24380 /* Since we hold everything in precise useconds this is easy */
24381 ti->tcpi_rtt = tp->t_srtt;
24382 ti->tcpi_rttvar = tp->t_rttvar;
24383 ti->tcpi_rto = tp->t_rxtcur;
24384 ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
24385 ti->tcpi_snd_cwnd = tp->snd_cwnd;
24386 /*
24387 * FreeBSD-specific extension fields for tcp_info.
24388 */
24389 ti->tcpi_rcv_space = tp->rcv_wnd;
24390 ti->tcpi_rcv_nxt = tp->rcv_nxt;
24391 ti->tcpi_snd_wnd = tp->snd_wnd;
24392 ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */
24393 ti->tcpi_snd_nxt = tp->snd_nxt;
24394 ti->tcpi_snd_mss = tp->t_maxseg;
24395 ti->tcpi_rcv_mss = tp->t_maxseg;
24396 ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24397 ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24398 ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24399 ti->tcpi_total_tlp = tp->t_sndtlppack;
24400 ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24401 ti->tcpi_rttmin = tp->t_rttlow;
24402 #ifdef NETFLIX_STATS
24403 memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24404 #endif
24405 #ifdef TCP_OFFLOAD
24406 if (tp->t_flags & TF_TOE) {
24407 ti->tcpi_options |= TCPI_OPT_TOE;
24408 tcp_offload_tcp_info(tp, ti);
24409 }
24410 #endif
24411 }
24412
24413 static int
rack_get_sockopt(struct tcpcb * tp,struct sockopt * sopt)24414 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24415 {
24416 struct inpcb *inp = tptoinpcb(tp);
24417 struct tcp_rack *rack;
24418 int32_t error, optval;
24419 uint64_t val, loptval;
24420 struct tcp_info ti;
24421 /*
24422 * Because all our options are either boolean or an int, we can just
24423 * pull everything into optval and then unlock and copy. If we ever
24424 * add a option that is not a int, then this will have quite an
24425 * impact to this routine.
24426 */
24427 error = 0;
24428 rack = (struct tcp_rack *)tp->t_fb_ptr;
24429 if (rack == NULL) {
24430 INP_WUNLOCK(inp);
24431 return (EINVAL);
24432 }
24433 switch (sopt->sopt_name) {
24434 case TCP_INFO:
24435 /* First get the info filled */
24436 rack_fill_info(tp, &ti);
24437 /* Fix up the rtt related fields if needed */
24438 INP_WUNLOCK(inp);
24439 error = sooptcopyout(sopt, &ti, sizeof ti);
24440 return (error);
24441 /*
24442 * Beta is the congestion control value for NewReno that influences how
24443 * much of a backoff happens when loss is detected. It is normally set
24444 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24445 * when you exit recovery.
24446 */
24447 case TCP_RACK_PACING_BETA:
24448 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24449 error = EINVAL;
24450 else if (rack->rc_pacing_cc_set == 0)
24451 optval = rack->r_ctl.rc_saved_beta;
24452 else {
24453 /*
24454 * Reach out into the CC data and report back what
24455 * I have previously set. Yeah it looks hackish but
24456 * we don't want to report the saved values.
24457 */
24458 if (tp->t_ccv.cc_data)
24459 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24460 else
24461 error = EINVAL;
24462 }
24463 break;
24464 /*
24465 * Beta_ecn is the congestion control value for NewReno that influences how
24466 * much of a backoff happens when a ECN mark is detected. It is normally set
24467 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24468 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24469 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24470 */
24471 case TCP_RACK_PACING_BETA_ECN:
24472 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24473 error = EINVAL;
24474 else if (rack->rc_pacing_cc_set == 0)
24475 optval = rack->r_ctl.rc_saved_beta_ecn;
24476 else {
24477 /*
24478 * Reach out into the CC data and report back what
24479 * I have previously set. Yeah it looks hackish but
24480 * we don't want to report the saved values.
24481 */
24482 if (tp->t_ccv.cc_data)
24483 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24484 else
24485 error = EINVAL;
24486 }
24487 break;
24488 case TCP_RACK_DSACK_OPT:
24489 optval = 0;
24490 if (rack->rc_rack_tmr_std_based) {
24491 optval |= 1;
24492 }
24493 if (rack->rc_rack_use_dsack) {
24494 optval |= 2;
24495 }
24496 break;
24497 case TCP_RACK_ENABLE_HYSTART:
24498 {
24499 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24500 optval = RACK_HYSTART_ON;
24501 if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24502 optval = RACK_HYSTART_ON_W_SC;
24503 if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24504 optval = RACK_HYSTART_ON_W_SC_C;
24505 } else {
24506 optval = RACK_HYSTART_OFF;
24507 }
24508 }
24509 break;
24510 case TCP_RACK_DGP_IN_REC:
24511 error = EINVAL;
24512 break;
24513 case TCP_RACK_HI_BETA:
24514 optval = rack->rack_hibeta;
24515 break;
24516 case TCP_DEFER_OPTIONS:
24517 optval = rack->defer_options;
24518 break;
24519 case TCP_RACK_MEASURE_CNT:
24520 optval = rack->r_ctl.req_measurements;
24521 break;
24522 case TCP_REC_ABC_VAL:
24523 optval = rack->r_use_labc_for_rec;
24524 break;
24525 case TCP_RACK_ABC_VAL:
24526 optval = rack->rc_labc;
24527 break;
24528 case TCP_HDWR_UP_ONLY:
24529 optval= rack->r_up_only;
24530 break;
24531 case TCP_FILLCW_RATE_CAP:
24532 loptval = rack->r_ctl.fillcw_cap;
24533 break;
24534 case TCP_PACING_RATE_CAP:
24535 loptval = rack->r_ctl.bw_rate_cap;
24536 break;
24537 case TCP_RACK_PROFILE:
24538 /* You cannot retrieve a profile, its write only */
24539 error = EINVAL;
24540 break;
24541 case TCP_SIDECHAN_DIS:
24542 optval = rack->r_ctl.side_chan_dis_mask;
24543 break;
24544 case TCP_HYBRID_PACING:
24545 /* You cannot retrieve hybrid pacing information, its write only */
24546 error = EINVAL;
24547 break;
24548 case TCP_USE_CMP_ACKS:
24549 optval = rack->r_use_cmp_ack;
24550 break;
24551 case TCP_RACK_PACE_TO_FILL:
24552 optval = rack->rc_pace_to_cwnd;
24553 break;
24554 case TCP_RACK_NO_PUSH_AT_MAX:
24555 optval = rack->r_ctl.rc_no_push_at_mrtt;
24556 break;
24557 case TCP_SHARED_CWND_ENABLE:
24558 optval = rack->rack_enable_scwnd;
24559 break;
24560 case TCP_RACK_NONRXT_CFG_RATE:
24561 optval = rack->rack_rec_nonrxt_use_cr;
24562 break;
24563 case TCP_NO_PRR:
24564 if (rack->rack_no_prr == 1)
24565 optval = 1;
24566 else if (rack->no_prr_addback == 1)
24567 optval = 2;
24568 else
24569 optval = 0;
24570 break;
24571 case TCP_GP_USE_LTBW:
24572 if (rack->dis_lt_bw) {
24573 /* It is not used */
24574 optval = 0;
24575 } else if (rack->use_lesser_lt_bw) {
24576 /* we use min() */
24577 optval = 1;
24578 } else {
24579 /* we use max() */
24580 optval = 2;
24581 }
24582 break;
24583 case TCP_RACK_DO_DETECTION:
24584 error = EINVAL;
24585 break;
24586 case TCP_RACK_MBUF_QUEUE:
24587 /* Now do we use the LRO mbuf-queue feature */
24588 optval = rack->r_mbuf_queue;
24589 break;
24590 case RACK_CSPR_IS_FCC:
24591 optval = rack->cspr_is_fcc;
24592 break;
24593 case TCP_TIMELY_DYN_ADJ:
24594 optval = rack->rc_gp_dyn_mul;
24595 break;
24596 case TCP_BBR_IWINTSO:
24597 error = EINVAL;
24598 break;
24599 case TCP_RACK_TLP_REDUCE:
24600 /* RACK TLP cwnd reduction (bool) */
24601 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24602 break;
24603 case TCP_BBR_RACK_INIT_RATE:
24604 val = rack->r_ctl.init_rate;
24605 /* convert to kbits per sec */
24606 val *= 8;
24607 val /= 1000;
24608 optval = (uint32_t)val;
24609 break;
24610 case TCP_RACK_FORCE_MSEG:
24611 optval = rack->rc_force_max_seg;
24612 break;
24613 case TCP_RACK_PACE_MIN_SEG:
24614 optval = rack->r_ctl.rc_user_set_min_segs;
24615 break;
24616 case TCP_RACK_PACE_MAX_SEG:
24617 /* Max segments in a pace */
24618 optval = rack->rc_user_set_max_segs;
24619 break;
24620 case TCP_RACK_PACE_ALWAYS:
24621 /* Use the always pace method */
24622 optval = rack->rc_always_pace;
24623 break;
24624 case TCP_RACK_PRR_SENDALOT:
24625 /* Allow PRR to send more than one seg */
24626 optval = rack->r_ctl.rc_prr_sendalot;
24627 break;
24628 case TCP_RACK_MIN_TO:
24629 /* Minimum time between rack t-o's in ms */
24630 optval = rack->r_ctl.rc_min_to;
24631 break;
24632 case TCP_RACK_SPLIT_LIMIT:
24633 optval = rack->r_ctl.rc_split_limit;
24634 break;
24635 case TCP_RACK_EARLY_SEG:
24636 /* If early recovery max segments */
24637 optval = rack->r_ctl.rc_early_recovery_segs;
24638 break;
24639 case TCP_RACK_REORD_THRESH:
24640 /* RACK reorder threshold (shift amount) */
24641 optval = rack->r_ctl.rc_reorder_shift;
24642 break;
24643 case TCP_SS_EEXIT:
24644 if (rack->r_ctl.gp_rnd_thresh) {
24645 uint32_t v;
24646
24647 v = rack->r_ctl.gp_gain_req;
24648 v <<= 17;
24649 optval = v | (rack->r_ctl.gp_rnd_thresh & 0xff);
24650 if (rack->r_ctl.gate_to_fs == 1)
24651 optval |= 0x10000;
24652 } else
24653 optval = 0;
24654 break;
24655 case TCP_RACK_REORD_FADE:
24656 /* Does reordering fade after ms time */
24657 optval = rack->r_ctl.rc_reorder_fade;
24658 break;
24659 case TCP_BBR_USE_RACK_RR:
24660 /* Do we use the rack cheat for rxt */
24661 optval = rack->use_rack_rr;
24662 break;
24663 case TCP_RACK_RR_CONF:
24664 optval = rack->r_rr_config;
24665 break;
24666 case TCP_HDWR_RATE_CAP:
24667 optval = rack->r_rack_hw_rate_caps;
24668 break;
24669 case TCP_BBR_HDWR_PACE:
24670 optval = rack->rack_hdw_pace_ena;
24671 break;
24672 case TCP_RACK_TLP_THRESH:
24673 /* RACK TLP theshold i.e. srtt+(srtt/N) */
24674 optval = rack->r_ctl.rc_tlp_threshold;
24675 break;
24676 case TCP_RACK_PKT_DELAY:
24677 /* RACK added ms i.e. rack-rtt + reord + N */
24678 optval = rack->r_ctl.rc_pkt_delay;
24679 break;
24680 case TCP_RACK_TLP_USE:
24681 optval = rack->rack_tlp_threshold_use;
24682 break;
24683 case TCP_PACING_DND:
24684 optval = rack->rc_pace_dnd;
24685 break;
24686 case TCP_RACK_PACE_RATE_CA:
24687 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24688 break;
24689 case TCP_RACK_PACE_RATE_SS:
24690 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24691 break;
24692 case TCP_RACK_PACE_RATE_REC:
24693 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24694 break;
24695 case TCP_DGP_UPPER_BOUNDS:
24696 optval = rack->r_ctl.rack_per_upper_bound_ss;
24697 optval <<= 16;
24698 optval |= rack->r_ctl.rack_per_upper_bound_ca;
24699 break;
24700 case TCP_RACK_GP_INCREASE_SS:
24701 optval = rack->r_ctl.rack_per_of_gp_ca;
24702 break;
24703 case TCP_RACK_GP_INCREASE_CA:
24704 optval = rack->r_ctl.rack_per_of_gp_ss;
24705 break;
24706 case TCP_RACK_PACING_DIVISOR:
24707 optval = rack->r_ctl.pace_len_divisor;
24708 break;
24709 case TCP_BBR_RACK_RTT_USE:
24710 optval = rack->r_ctl.rc_rate_sample_method;
24711 break;
24712 case TCP_DELACK:
24713 optval = tp->t_delayed_ack;
24714 break;
24715 case TCP_DATA_AFTER_CLOSE:
24716 optval = rack->rc_allow_data_af_clo;
24717 break;
24718 case TCP_SHARED_CWND_TIME_LIMIT:
24719 optval = rack->r_limit_scw;
24720 break;
24721 case TCP_HONOR_HPTS_MIN:
24722 if (rack->r_use_hpts_min)
24723 optval = rack->r_ctl.max_reduction;
24724 else
24725 optval = 0;
24726 break;
24727 case TCP_REC_IS_DYN:
24728 optval = rack->rc_gp_no_rec_chg;
24729 break;
24730 case TCP_NO_TIMELY:
24731 optval = rack->rc_skip_timely;
24732 break;
24733 case TCP_RACK_TIMER_SLOP:
24734 optval = rack->r_ctl.timer_slop;
24735 break;
24736 default:
24737 return (tcp_default_ctloutput(tp, sopt));
24738 break;
24739 }
24740 INP_WUNLOCK(inp);
24741 if (error == 0) {
24742 if ((sopt->sopt_name == TCP_PACING_RATE_CAP) ||
24743 (sopt->sopt_name == TCP_FILLCW_RATE_CAP))
24744 error = sooptcopyout(sopt, &loptval, sizeof loptval);
24745 else
24746 error = sooptcopyout(sopt, &optval, sizeof optval);
24747 }
24748 return (error);
24749 }
24750
24751 static int
rack_ctloutput(struct tcpcb * tp,struct sockopt * sopt)24752 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24753 {
24754 if (sopt->sopt_dir == SOPT_SET) {
24755 return (rack_set_sockopt(tp, sopt));
24756 } else if (sopt->sopt_dir == SOPT_GET) {
24757 return (rack_get_sockopt(tp, sopt));
24758 } else {
24759 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24760 }
24761 }
24762
24763 static const char *rack_stack_names[] = {
24764 __XSTRING(STACKNAME),
24765 #ifdef STACKALIAS
24766 __XSTRING(STACKALIAS),
24767 #endif
24768 };
24769
24770 static int
rack_ctor(void * mem,int32_t size,void * arg,int32_t how)24771 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24772 {
24773 memset(mem, 0, size);
24774 return (0);
24775 }
24776
24777 static void
rack_dtor(void * mem,int32_t size,void * arg)24778 rack_dtor(void *mem, int32_t size, void *arg)
24779 {
24780
24781 }
24782
24783 static bool rack_mod_inited = false;
24784
24785 static int
tcp_addrack(module_t mod,int32_t type,void * data)24786 tcp_addrack(module_t mod, int32_t type, void *data)
24787 {
24788 int32_t err = 0;
24789 int num_stacks;
24790
24791 switch (type) {
24792 case MOD_LOAD:
24793 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24794 sizeof(struct rack_sendmap),
24795 rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24796
24797 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24798 sizeof(struct tcp_rack),
24799 rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24800
24801 sysctl_ctx_init(&rack_sysctl_ctx);
24802 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24803 SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24804 OID_AUTO,
24805 #ifdef STACKALIAS
24806 __XSTRING(STACKALIAS),
24807 #else
24808 __XSTRING(STACKNAME),
24809 #endif
24810 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24811 "");
24812 if (rack_sysctl_root == NULL) {
24813 printf("Failed to add sysctl node\n");
24814 err = EFAULT;
24815 goto free_uma;
24816 }
24817 rack_init_sysctls();
24818 num_stacks = nitems(rack_stack_names);
24819 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24820 rack_stack_names, &num_stacks);
24821 if (err) {
24822 printf("Failed to register %s stack name for "
24823 "%s module\n", rack_stack_names[num_stacks],
24824 __XSTRING(MODNAME));
24825 sysctl_ctx_free(&rack_sysctl_ctx);
24826 free_uma:
24827 uma_zdestroy(rack_zone);
24828 uma_zdestroy(rack_pcb_zone);
24829 rack_counter_destroy();
24830 printf("Failed to register rack module -- err:%d\n", err);
24831 return (err);
24832 }
24833 tcp_lro_reg_mbufq();
24834 rack_mod_inited = true;
24835 break;
24836 case MOD_QUIESCE:
24837 err = deregister_tcp_functions(&__tcp_rack, true, false);
24838 break;
24839 case MOD_UNLOAD:
24840 err = deregister_tcp_functions(&__tcp_rack, false, true);
24841 if (err == EBUSY)
24842 break;
24843 if (rack_mod_inited) {
24844 uma_zdestroy(rack_zone);
24845 uma_zdestroy(rack_pcb_zone);
24846 sysctl_ctx_free(&rack_sysctl_ctx);
24847 rack_counter_destroy();
24848 rack_mod_inited = false;
24849 }
24850 tcp_lro_dereg_mbufq();
24851 err = 0;
24852 break;
24853 default:
24854 return (EOPNOTSUPP);
24855 }
24856 return (err);
24857 }
24858
24859 static moduledata_t tcp_rack = {
24860 .name = __XSTRING(MODNAME),
24861 .evhand = tcp_addrack,
24862 .priv = 0
24863 };
24864
24865 MODULE_VERSION(MODNAME, 1);
24866 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24867 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24868
24869 #endif /* #if !defined(INET) && !defined(INET6) */
24870