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 * Author: Randall Stewart <rrs@netflix.com> 28 * This work is based on the ACM Queue paper 29 * BBR - Congestion Based Congestion Control 30 * and also numerous discussions with Neal, Yuchung and Van. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include "opt_inet.h" 37 #include "opt_inet6.h" 38 #include "opt_ipsec.h" 39 #include "opt_ratelimit.h" 40 #include <sys/param.h> 41 #include <sys/arb.h> 42 #include <sys/module.h> 43 #include <sys/kernel.h> 44 #include <sys/libkern.h> 45 #ifdef TCP_HHOOK 46 #include <sys/hhook.h> 47 #endif 48 #include <sys/malloc.h> 49 #include <sys/mbuf.h> 50 #include <sys/proc.h> 51 #include <sys/socket.h> 52 #include <sys/socketvar.h> 53 #include <sys/sysctl.h> 54 #include <sys/systm.h> 55 #ifdef STATS 56 #include <sys/qmath.h> 57 #include <sys/tree.h> 58 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 59 #endif 60 #include <sys/refcount.h> 61 #include <sys/queue.h> 62 #include <sys/eventhandler.h> 63 #include <sys/smp.h> 64 #include <sys/kthread.h> 65 #include <sys/lock.h> 66 #include <sys/mutex.h> 67 #include <sys/tim_filter.h> 68 #include <sys/time.h> 69 #include <sys/protosw.h> 70 #include <vm/uma.h> 71 #include <sys/kern_prefetch.h> 72 73 #include <net/route.h> 74 #include <net/route/nhop.h> 75 #include <net/vnet.h> 76 77 #define TCPSTATES /* for logging */ 78 79 #include <netinet/in.h> 80 #include <netinet/in_kdtrace.h> 81 #include <netinet/in_pcb.h> 82 #include <netinet/ip.h> 83 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 84 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 85 #include <netinet/ip_var.h> 86 #include <netinet/ip6.h> 87 #include <netinet6/in6_pcb.h> 88 #include <netinet6/ip6_var.h> 89 #define TCPOUTFLAGS 90 #include <netinet/tcp.h> 91 #include <netinet/tcp_fsm.h> 92 #include <netinet/tcp_seq.h> 93 #include <netinet/tcp_timer.h> 94 #include <netinet/tcp_var.h> 95 #include <netinet/tcpip.h> 96 #include <netinet/tcp_hpts.h> 97 #include <netinet/cc/cc.h> 98 #include <netinet/tcp_log_buf.h> 99 #include <netinet/tcp_ratelimit.h> 100 #include <netinet/tcp_lro.h> 101 #ifdef TCP_OFFLOAD 102 #include <netinet/tcp_offload.h> 103 #endif 104 #ifdef INET6 105 #include <netinet6/tcp6_var.h> 106 #endif 107 #include <netinet/tcp_fastopen.h> 108 109 #include <netipsec/ipsec_support.h> 110 #include <net/if.h> 111 #include <net/if_var.h> 112 #include <net/ethernet.h> 113 114 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 115 #include <netipsec/ipsec.h> 116 #include <netipsec/ipsec6.h> 117 #endif /* IPSEC */ 118 119 #include <netinet/udp.h> 120 #include <netinet/udp_var.h> 121 #include <machine/in_cksum.h> 122 123 #ifdef MAC 124 #include <security/mac/mac_framework.h> 125 #endif 126 127 #include "sack_filter.h" 128 #include "tcp_bbr.h" 129 #include "rack_bbr_common.h" 130 uma_zone_t bbr_zone; 131 uma_zone_t bbr_pcb_zone; 132 133 struct sysctl_ctx_list bbr_sysctl_ctx; 134 struct sysctl_oid *bbr_sysctl_root; 135 136 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ 137 (tv) = (value); \ 138 if ((u_long)(tv) < (u_long)(tvmin)) \ 139 (tv) = (tvmin); \ 140 if ((u_long)(tv) > (u_long)(tvmax)) \ 141 (tv) = (tvmax); \ 142 } while(0) 143 144 /*#define BBR_INVARIANT 1*/ 145 146 /* 147 * initial window 148 */ 149 static uint32_t bbr_def_init_win = 10; 150 static int32_t bbr_persist_min = 250000; /* 250ms */ 151 static int32_t bbr_persist_max = 1000000; /* 1 Second */ 152 static int32_t bbr_cwnd_may_shrink = 0; 153 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; 154 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; 155 static int32_t bbr_hardware_pacing_limit = 8000; 156 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ 157 static int32_t bbr_no_retran = 0; 158 159 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ 160 static int32_t bbr_max_net_error_cnt = 10; 161 /* Should the following be dynamic too -- loss wise */ 162 static int32_t bbr_rtt_gain_thresh = 0; 163 /* Measurement controls */ 164 static int32_t bbr_use_google_algo = 1; 165 static int32_t bbr_ts_limiting = 1; 166 static int32_t bbr_ts_can_raise = 0; 167 static int32_t bbr_do_red = 600; 168 static int32_t bbr_red_scale = 20000; 169 static int32_t bbr_red_mul = 1; 170 static int32_t bbr_red_div = 2; 171 static int32_t bbr_red_growth_restrict = 1; 172 static int32_t bbr_target_is_bbunit = 0; 173 static int32_t bbr_drop_limit = 0; 174 /* 175 * How much gain do we need to see to 176 * stay in startup? 177 */ 178 static int32_t bbr_marks_rxt_sack_passed = 0; 179 static int32_t bbr_start_exit = 25; 180 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ 181 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ 182 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ 183 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this 184 * if we go back ever to where the pacer 185 * has priority over timers. 186 */ 187 static int32_t bbr_policer_call_from_rack_to = 0; 188 static int32_t bbr_policer_detection_enabled = 1; 189 static int32_t bbr_min_measurements_req = 1; /* We need at least 2 190 * measurements before we are 191 * "good" note that 2 == 1. 192 * This is because we use a > 193 * comparison. This means if 194 * min_measure was 0, it takes 195 * num-measures > min(0) and 196 * you get 1 measurement and 197 * you are good. Set to 1, you 198 * have to have two 199 * measurements (this is done 200 * to prevent it from being ok 201 * to have no measurements). */ 202 static int32_t bbr_no_pacing_until = 4; 203 204 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ 205 static int32_t bbr_min_peer_delta = 20; /* 20 units */ 206 static int32_t bbr_delta_percent = 150; /* 15.0 % */ 207 208 static int32_t bbr_target_cwnd_mult_limit = 8; 209 /* 210 * bbr_cwnd_min_val is the number of 211 * segments we hold to in the RTT probe 212 * state typically 4. 213 */ 214 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; 215 216 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; 217 218 static int32_t bbr_gain_to_target = 1; 219 static int32_t bbr_gain_gets_extra_too = 1; 220 /* 221 * bbr_high_gain is the 2/ln(2) value we need 222 * to double the sending rate in startup. This 223 * is used for both cwnd and hptsi gain's. 224 */ 225 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 226 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; 227 static int32_t bbr_use_lower_gain_in_startup = 1; 228 229 /* thresholds for reduction on drain in sub-states/drain */ 230 static int32_t bbr_drain_rtt = BBR_SRTT; 231 static int32_t bbr_drain_floor = 88; 232 static int32_t google_allow_early_out = 1; 233 static int32_t google_consider_lost = 1; 234 static int32_t bbr_drain_drop_mul = 4; 235 static int32_t bbr_drain_drop_div = 5; 236 static int32_t bbr_rand_ot = 50; 237 static int32_t bbr_can_force_probertt = 0; 238 static int32_t bbr_can_adjust_probertt = 1; 239 static int32_t bbr_probertt_sets_rtt = 0; 240 static int32_t bbr_can_use_ts_for_rtt = 1; 241 static int32_t bbr_is_ratio = 0; 242 static int32_t bbr_sub_drain_app_limit = 1; 243 static int32_t bbr_prtt_slam_cwnd = 1; 244 static int32_t bbr_sub_drain_slam_cwnd = 1; 245 static int32_t bbr_slam_cwnd_in_main_drain = 1; 246 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter 247 * hold */ 248 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); 249 /* 250 * bbr_drain_gain is the reverse of the high_gain 251 * designed to drain back out the standing queue 252 * that is formed in startup by causing a larger 253 * hptsi gain and thus drainging the packets 254 * in flight. 255 */ 256 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; 257 static int32_t bbr_rttprobe_gain = 192; 258 259 /* 260 * The cwnd_gain is the default cwnd gain applied when 261 * calculating a target cwnd. Note that the cwnd is 262 * a secondary factor in the way BBR works (see the 263 * paper and think about it, it will take some time). 264 * Basically the hptsi_gain spreads the packets out 265 * so you never get more than BDP to the peer even 266 * if the cwnd is high. In our implemenation that 267 * means in non-recovery/retransmission scenarios 268 * cwnd will never be reached by the flight-size. 269 */ 270 static int32_t bbr_cwnd_gain = BBR_UNIT * 2; 271 static int32_t bbr_tlp_type_to_use = BBR_SRTT; 272 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ 273 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ 274 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ 275 static int32_t bbr_ignore_data_after_close = 1; 276 static int16_t bbr_hptsi_gain[] = { 277 (BBR_UNIT *5 / 4), 278 (BBR_UNIT * 3 / 4), 279 BBR_UNIT, 280 BBR_UNIT, 281 BBR_UNIT, 282 BBR_UNIT, 283 BBR_UNIT, 284 BBR_UNIT 285 }; 286 int32_t bbr_use_rack_resend_cheat = 1; 287 int32_t bbr_sends_full_iwnd = 1; 288 289 #define BBR_HPTSI_GAIN_MAX 8 290 /* 291 * The BBR module incorporates a number of 292 * TCP ideas that have been put out into the IETF 293 * over the last few years: 294 * - Yuchung Cheng's RACK TCP (for which its named) that 295 * will stop us using the number of dup acks and instead 296 * use time as the gage of when we retransmit. 297 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 298 * of Dukkipati et.al. 299 * - Van Jacobson's et.al BBR. 300 * 301 * RACK depends on SACK, so if an endpoint arrives that 302 * cannot do SACK the state machine below will shuttle the 303 * connection back to using the "default" TCP stack that is 304 * in FreeBSD. 305 * 306 * To implement BBR and RACK the original TCP stack was first decomposed 307 * into a functional state machine with individual states 308 * for each of the possible TCP connection states. The do_segment 309 * functions role in life is to mandate the connection supports SACK 310 * initially and then assure that the RACK state matches the conenction 311 * state before calling the states do_segment function. Data processing 312 * of inbound segments also now happens in the hpts_do_segment in general 313 * with only one exception. This is so we can keep the connection on 314 * a single CPU. 315 * 316 * Each state is simplified due to the fact that the original do_segment 317 * has been decomposed and we *know* what state we are in (no 318 * switches on the state) and all tests for SACK are gone. This 319 * greatly simplifies what each state does. 320 * 321 * TCP output is also over-written with a new version since it 322 * must maintain the new rack scoreboard and has had hptsi 323 * integrated as a requirment. Still todo is to eliminate the 324 * use of the callout_() system and use the hpts for all 325 * timers as well. 326 */ 327 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ 328 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ 329 static const int32_t bbr_min_req_free = 2; /* The min we must have on the 330 * free list */ 331 static int32_t bbr_tlp_thresh = 1; 332 static int32_t bbr_reorder_thresh = 2; 333 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def 334 * 60,000,000 - 60 seconds */ 335 static int32_t bbr_pkt_delay = 1000; 336 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ 337 static int32_t bbr_incr_timers = 1; 338 339 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ 340 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ 341 static int32_t bbr_exit_startup_at_loss = 1; 342 343 /* 344 * bbr_lt_bw_ratio is 1/8th 345 * bbr_lt_bw_diff is < 4 Kbit/sec 346 */ 347 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ 348 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ 349 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use 350 * the lt_bw for */ 351 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure 352 * lt_bw */ 353 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ 354 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ 355 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ 356 357 static int32_t bbr_verbose_logging = 0; 358 /* 359 * Currently regular tcp has a rto_min of 30ms 360 * the backoff goes 12 times so that ends up 361 * being a total of 122.850 seconds before a 362 * connection is killed. 363 */ 364 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ 365 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ 366 367 /****************************************************/ 368 /* DEFAULT TSO SIZING (cpu performance impacting) */ 369 /****************************************************/ 370 /* What amount is our formula using to get TSO size */ 371 static int32_t bbr_hptsi_per_second = 1000; 372 373 /* 374 * For hptsi under bbr_cross_over connections what is delay 375 * target 7ms (in usec) combined with a seg_max of 2 376 * gets us close to identical google behavior in 377 * TSO size selection (possibly more 1MSS sends). 378 */ 379 static int32_t bbr_hptsi_segments_delay_tar = 7000; 380 381 /* Does pacing delay include overhead's in its time calculations? */ 382 static int32_t bbr_include_enet_oh = 0; 383 static int32_t bbr_include_ip_oh = 1; 384 static int32_t bbr_include_tcp_oh = 1; 385 static int32_t bbr_google_discount = 10; 386 387 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ 388 static int32_t bbr_state_is_pkt_epoch = 0; 389 static int32_t bbr_state_drain_2_tar = 1; 390 /* What is the max the 0 - bbr_cross_over MBPS TSO target 391 * can reach using our delay target. Note that this 392 * value becomes the floor for the cross over 393 * algorithm. 394 */ 395 static int32_t bbr_hptsi_segments_max = 2; 396 static int32_t bbr_hptsi_segments_floor = 1; 397 static int32_t bbr_hptsi_utter_max = 0; 398 399 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ 400 static int32_t bbr_hptsi_bytes_min = 1460; 401 static int32_t bbr_all_get_min = 0; 402 403 /* Cross over point from algo-a to algo-b */ 404 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; 405 406 /* Do we deal with our restart state? */ 407 static int32_t bbr_uses_idle_restart = 0; 408 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ 409 410 /* Do we allow hardware pacing? */ 411 static int32_t bbr_allow_hdwr_pacing = 0; 412 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ 413 static int32_t bbr_hdwr_pace_floor = 1; 414 static int32_t bbr_hdwr_pacing_delay_cnt = 10; 415 416 /****************************************************/ 417 static int32_t bbr_resends_use_tso = 0; 418 static int32_t bbr_tlp_max_resend = 2; 419 static int32_t bbr_sack_block_limit = 128; 420 421 #define BBR_MAX_STAT 19 422 counter_u64_t bbr_state_time[BBR_MAX_STAT]; 423 counter_u64_t bbr_state_lost[BBR_MAX_STAT]; 424 counter_u64_t bbr_state_resend[BBR_MAX_STAT]; 425 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 426 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 427 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; 428 counter_u64_t bbr_flows_whdwr_pacing; 429 counter_u64_t bbr_flows_nohdwr_pacing; 430 431 counter_u64_t bbr_nohdwr_pacing_enobuf; 432 counter_u64_t bbr_hdwr_pacing_enobuf; 433 434 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); 435 436 /* 437 * Static defintions we need for forward declarations. 438 */ 439 static uint32_t 440 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, 441 uint32_t useconds_time, uint64_t bw); 442 static uint32_t 443 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); 444 static void 445 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); 446 static void 447 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); 448 static void 449 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, 450 int dolog); 451 static uint32_t 452 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); 453 static void 454 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, 455 int32_t pkt_epoch, uint32_t losses); 456 static uint32_t 457 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, 458 struct bbr_sendmap *rsm); 459 static uint32_t 460 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); 461 static uint32_t 462 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 463 struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); 464 static void 465 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 466 int32_t line); 467 static void 468 bbr_set_state_target(struct tcp_bbr *bbr, int line); 469 static void 470 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); 471 static void 472 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, 473 int event, int line); 474 static void 475 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); 476 static void 477 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); 478 static void 479 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 480 uint32_t rtt, uint32_t line, uint8_t is_start, 481 uint16_t set); 482 static struct bbr_sendmap * 483 bbr_find_lowest_rsm(struct tcp_bbr *bbr); 484 static __inline uint32_t 485 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); 486 static void 487 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, 488 uint8_t which); 489 static void 490 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, 491 uint32_t time_since_sent, uint32_t srtt, 492 uint32_t thresh, uint32_t to); 493 static void 494 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); 495 static void 496 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, 497 uint32_t del_by, uint32_t cts, uint32_t sloton, 498 uint32_t prev_delay); 499 static void 500 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 501 int32_t line); 502 static void 503 bbr_stop_all_timers(struct tcpcb *tp); 504 static void 505 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); 506 static void 507 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); 508 static void 509 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); 510 static void 511 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 512 uint32_t cts, uint32_t usecs, uint64_t bw, 513 uint32_t override, int mod); 514 static int 515 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt); 516 517 static inline uint8_t 518 bbr_state_val(struct tcp_bbr *bbr) 519 { 520 return(bbr->rc_bbr_substate); 521 } 522 523 static inline uint32_t 524 get_min_cwnd(struct tcp_bbr *bbr) 525 { 526 int mss; 527 528 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 529 bbr->r_ctl.rc_pace_max_segs); 530 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) 531 return (bbr_cwnd_min_val_hs * mss); 532 else 533 return (bbr_cwnd_min_val * mss); 534 } 535 536 static uint32_t 537 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) 538 { 539 uint64_t srtt, var; 540 uint64_t ret_val; 541 542 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 543 if (tp->t_srtt == 0) { 544 srtt = (uint64_t)BBR_INITIAL_RTO; 545 var = 0; 546 } else { 547 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 548 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); 549 } 550 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), 551 bbr_persist_min, bbr_persist_max); 552 return ((uint32_t)ret_val); 553 } 554 555 static uint32_t 556 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 557 { 558 /* 559 * Start the FR timer, we do this based on getting the first one in 560 * the rc_tmap. Note that if its NULL we must stop the timer. in all 561 * events we need to stop the running timer (if its running) before 562 * starting the new one. 563 */ 564 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; 565 int32_t idx; 566 int32_t is_tlp_timer = 0; 567 struct bbr_sendmap *rsm; 568 569 if (bbr->rc_all_timers_stopped) { 570 /* All timers have been stopped none are to run */ 571 return (0); 572 } 573 if (bbr->rc_in_persist) { 574 /* We can't start any timer in persists */ 575 return (bbr_get_persists_timer_val(tp, bbr)); 576 } 577 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 578 if ((rsm == NULL) || 579 ((tp->t_flags & TF_SACK_PERMIT) == 0) || 580 (tp->t_state < TCPS_ESTABLISHED)) { 581 /* Nothing on the send map */ 582 activate_rxt: 583 if (SEQ_LT(tp->snd_una, tp->snd_max) || 584 sbavail(&tptosocket(tp)->so_snd)) { 585 uint64_t tov; 586 587 time_since_sent = 0; 588 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 589 if (rsm) { 590 idx = rsm->r_rtr_cnt - 1; 591 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 592 tstmp_touse = rsm->r_tim_lastsent[idx]; 593 else 594 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 595 if (TSTMP_GT(tstmp_touse, cts)) 596 time_since_sent = cts - tstmp_touse; 597 } 598 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 599 if (tp->t_srtt == 0) 600 tov = BBR_INITIAL_RTO; 601 else 602 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + 603 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); 604 if (tp->t_rxtshift) 605 tov *= tcp_backoff[tp->t_rxtshift]; 606 if (tov > time_since_sent) 607 tov -= time_since_sent; 608 else 609 tov = bbr->r_ctl.rc_min_to; 610 TCPT_RANGESET_NOSLOP(to, tov, 611 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), 612 (bbr->rc_max_rto_sec * USECS_IN_SECOND)); 613 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); 614 return (to); 615 } 616 return (0); 617 } 618 if (rsm->r_flags & BBR_ACKED) { 619 rsm = bbr_find_lowest_rsm(bbr); 620 if (rsm == NULL) { 621 /* No lowest? */ 622 goto activate_rxt; 623 } 624 } 625 /* Convert from ms to usecs */ 626 if (rsm->r_flags & BBR_SACK_PASSED) { 627 if ((tp->t_flags & TF_SENTFIN) && 628 ((tp->snd_max - tp->snd_una) == 1) && 629 (rsm->r_flags & BBR_HAS_FIN)) { 630 /* 631 * We don't start a bbr rack timer if all we have is 632 * a FIN outstanding. 633 */ 634 goto activate_rxt; 635 } 636 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); 637 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); 638 idx = rsm->r_rtr_cnt - 1; 639 exp = rsm->r_tim_lastsent[idx] + thresh; 640 if (SEQ_GEQ(exp, cts)) { 641 to = exp - cts; 642 if (to < bbr->r_ctl.rc_min_to) { 643 to = bbr->r_ctl.rc_min_to; 644 } 645 } else { 646 to = bbr->r_ctl.rc_min_to; 647 } 648 } else { 649 /* Ok we need to do a TLP not RACK */ 650 if (bbr->rc_tlp_in_progress != 0) { 651 /* 652 * The previous send was a TLP. 653 */ 654 goto activate_rxt; 655 } 656 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 657 if (rsm == NULL) { 658 /* We found no rsm to TLP with. */ 659 goto activate_rxt; 660 } 661 if (rsm->r_flags & BBR_HAS_FIN) { 662 /* If its a FIN we don't do TLP */ 663 rsm = NULL; 664 goto activate_rxt; 665 } 666 time_since_sent = 0; 667 idx = rsm->r_rtr_cnt - 1; 668 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 669 tstmp_touse = rsm->r_tim_lastsent[idx]; 670 else 671 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 672 if (TSTMP_GT(tstmp_touse, cts)) 673 time_since_sent = cts - tstmp_touse; 674 is_tlp_timer = 1; 675 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); 676 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); 677 if (thresh > time_since_sent) 678 to = thresh - time_since_sent; 679 else 680 to = bbr->r_ctl.rc_min_to; 681 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 682 /* 683 * If the TLP time works out to larger than the max 684 * RTO lets not do TLP.. just RTO. 685 */ 686 goto activate_rxt; 687 } 688 if ((bbr->rc_tlp_rtx_out == 1) && 689 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { 690 /* 691 * Second retransmit of the same TLP 692 * lets not. 693 */ 694 bbr->rc_tlp_rtx_out = 0; 695 goto activate_rxt; 696 } 697 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { 698 /* 699 * The tail is no longer the last one I did a probe 700 * on 701 */ 702 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 703 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 704 } 705 } 706 if (is_tlp_timer == 0) { 707 BBR_STAT_INC(bbr_to_arm_rack); 708 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 709 } else { 710 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); 711 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 712 /* 713 * We have exceeded how many times we can retran the 714 * current TLP timer, switch to the RTO timer. 715 */ 716 goto activate_rxt; 717 } else { 718 BBR_STAT_INC(bbr_to_arm_tlp); 719 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 720 } 721 } 722 return (to); 723 } 724 725 static inline int32_t 726 bbr_minseg(struct tcp_bbr *bbr) 727 { 728 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); 729 } 730 731 static void 732 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) 733 { 734 struct inpcb *inp = tptoinpcb(tp); 735 struct hpts_diag diag; 736 uint32_t delayed_ack = 0; 737 uint32_t left = 0; 738 uint32_t hpts_timeout; 739 uint8_t stopped; 740 int32_t delay_calc = 0; 741 uint32_t prev_delay = 0; 742 743 if (tcp_in_hpts(inp)) { 744 /* A previous call is already set up */ 745 return; 746 } 747 if ((tp->t_state == TCPS_CLOSED) || 748 (tp->t_state == TCPS_LISTEN)) { 749 return; 750 } 751 stopped = bbr->rc_tmr_stopped; 752 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 753 left = bbr->r_ctl.rc_timer_exp - cts; 754 } 755 bbr->r_ctl.rc_hpts_flags = 0; 756 bbr->r_ctl.rc_timer_exp = 0; 757 prev_delay = bbr->r_ctl.rc_last_delay_val; 758 if (bbr->r_ctl.rc_last_delay_val && 759 (slot == 0)) { 760 /* 761 * If a previous pacer delay was in place we 762 * are not coming from the output side (where 763 * we calculate a delay, more likely a timer). 764 */ 765 slot = bbr->r_ctl.rc_last_delay_val; 766 if (TSTMP_GT(cts, bbr->rc_pacer_started)) { 767 /* Compensate for time passed */ 768 delay_calc = cts - bbr->rc_pacer_started; 769 if (delay_calc <= slot) 770 slot -= delay_calc; 771 } 772 } 773 /* Do we have early to make up for by pushing out the pacing time? */ 774 if (bbr->r_agg_early_set) { 775 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); 776 slot += bbr->r_ctl.rc_agg_early; 777 bbr->r_ctl.rc_agg_early = 0; 778 bbr->r_agg_early_set = 0; 779 } 780 /* Are we running a total debt that needs to be compensated for? */ 781 if (bbr->r_ctl.rc_hptsi_agg_delay) { 782 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { 783 /* We nuke the delay */ 784 slot -= bbr->r_ctl.rc_hptsi_agg_delay; 785 bbr->r_ctl.rc_hptsi_agg_delay = 0; 786 } else { 787 /* We nuke some of the delay, put in a minimal 100usecs */ 788 bbr->r_ctl.rc_hptsi_agg_delay -= slot; 789 bbr->r_ctl.rc_last_delay_val = slot = 100; 790 } 791 } 792 bbr->r_ctl.rc_last_delay_val = slot; 793 hpts_timeout = bbr_timer_start(tp, bbr, cts); 794 if (tp->t_flags & TF_DELACK) { 795 if (bbr->rc_in_persist == 0) { 796 delayed_ack = bbr_delack_time; 797 } else { 798 /* 799 * We are in persists and have 800 * gotten a new data element. 801 */ 802 if (hpts_timeout > bbr_delack_time) { 803 /* 804 * Lets make the persists timer (which acks) 805 * be the smaller of hpts_timeout and bbr_delack_time. 806 */ 807 hpts_timeout = bbr_delack_time; 808 } 809 } 810 } 811 if (delayed_ack && 812 ((hpts_timeout == 0) || 813 (delayed_ack < hpts_timeout))) { 814 /* We need a Delayed ack timer */ 815 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 816 hpts_timeout = delayed_ack; 817 } 818 if (slot) { 819 /* Mark that we have a pacing timer up */ 820 BBR_STAT_INC(bbr_paced_segments); 821 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 822 } 823 /* 824 * If no timers are going to run and we will fall off thfe hptsi 825 * wheel, we resort to a keep-alive timer if its configured. 826 */ 827 if ((hpts_timeout == 0) && 828 (slot == 0)) { 829 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 830 (tp->t_state <= TCPS_CLOSING)) { 831 /* 832 * Ok we have no timer (persists, rack, tlp, rxt or 833 * del-ack), we don't have segments being paced. So 834 * all that is left is the keepalive timer. 835 */ 836 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 837 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); 838 } else { 839 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); 840 } 841 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 842 } 843 } 844 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 845 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 846 /* 847 * RACK, TLP, persists and RXT timers all are restartable 848 * based on actions input .. i.e we received a packet (ack 849 * or sack) and that changes things (rw, or snd_una etc). 850 * Thus we can restart them with a new value. For 851 * keep-alive, delayed_ack we keep track of what was left 852 * and restart the timer with a smaller value. 853 */ 854 if (left < hpts_timeout) 855 hpts_timeout = left; 856 } 857 if (bbr->r_ctl.rc_incr_tmrs && slot && 858 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { 859 /* 860 * If configured to do so, and the timer is either 861 * the TLP or RXT timer, we need to increase the timeout 862 * by the pacing time. Consider the bottleneck at my 863 * machine as an example, we are sending something 864 * to start a TLP on. The last packet won't be emitted 865 * fully until the pacing time (the bottleneck will hold 866 * the data in place). Once the packet is emitted that 867 * is when we want to start waiting for the TLP. This 868 * is most evident with hardware pacing (where the nic 869 * is holding the packet(s) before emitting). But it 870 * can also show up in the network so we do it for all 871 * cases. Technically we would take off one packet from 872 * this extra delay but this is easier and being more 873 * conservative is probably better. 874 */ 875 hpts_timeout += slot; 876 } 877 if (hpts_timeout) { 878 /* 879 * Hack alert for now we can't time-out over 2147 seconds (a 880 * bit more than 35min) 881 */ 882 if (hpts_timeout > 0x7ffffffe) 883 hpts_timeout = 0x7ffffffe; 884 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 885 } else 886 bbr->r_ctl.rc_timer_exp = 0; 887 if ((slot) && 888 (bbr->rc_use_google || 889 bbr->output_error_seen || 890 (slot <= hpts_timeout)) ) { 891 /* 892 * Tell LRO that it can queue packets while 893 * we pace. 894 */ 895 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 896 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 897 (bbr->rc_cwnd_limited == 0)) { 898 /* 899 * If we are not cwnd limited and we 900 * are running a rack timer we put on 901 * the do not disturbe even for sack. 902 */ 903 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 904 } else 905 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 906 bbr->rc_pacer_started = cts; 907 908 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot), 909 __LINE__, &diag); 910 bbr->rc_timer_first = 0; 911 bbr->bbr_timer_src = frm; 912 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); 913 bbr_log_hpts_diag(bbr, cts, &diag); 914 } else if (hpts_timeout) { 915 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout), 916 __LINE__, &diag); 917 /* 918 * We add the flag here as well if the slot is set, 919 * since hpts will call in to clear the queue first before 920 * calling the output routine (which does our timers). 921 * We don't want to set the flag if its just a timer 922 * else the arrival of data might (that causes us 923 * to send more) might get delayed. Imagine being 924 * on a keep-alive timer and a request comes in for 925 * more data. 926 */ 927 if (slot) 928 bbr->rc_pacer_started = cts; 929 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 930 (bbr->rc_cwnd_limited == 0)) { 931 /* 932 * For a rack timer, don't wake us even 933 * if a sack arrives as long as we are 934 * not cwnd limited. 935 */ 936 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 937 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 938 } else { 939 /* All other timers wake us up */ 940 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 941 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 942 } 943 bbr->bbr_timer_src = frm; 944 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); 945 bbr_log_hpts_diag(bbr, cts, &diag); 946 bbr->rc_timer_first = 1; 947 } 948 bbr->rc_tmr_stopped = 0; 949 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); 950 } 951 952 static void 953 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) 954 { 955 /* 956 * We received an ack, and then did not call send or were bounced 957 * out due to the hpts was running. Now a timer is up as well, is it 958 * the right timer? 959 */ 960 struct inpcb *inp; 961 struct bbr_sendmap *rsm; 962 uint32_t hpts_timeout; 963 int tmr_up; 964 965 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 966 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 967 return; 968 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 969 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 970 (tmr_up == PACE_TMR_RXT)) { 971 /* Should be an RXT */ 972 return; 973 } 974 inp = bbr->rc_inp; 975 if (rsm == NULL) { 976 /* Nothing outstanding? */ 977 if (tp->t_flags & TF_DELACK) { 978 if (tmr_up == PACE_TMR_DELACK) 979 /* 980 * We are supposed to have delayed ack up 981 * and we do 982 */ 983 return; 984 } else if (sbavail(&inp->inp_socket->so_snd) && 985 (tmr_up == PACE_TMR_RXT)) { 986 /* 987 * if we hit enobufs then we would expect the 988 * possibility of nothing outstanding and the RXT up 989 * (and the hptsi timer). 990 */ 991 return; 992 } else if (((V_tcp_always_keepalive || 993 inp->inp_socket->so_options & SO_KEEPALIVE) && 994 (tp->t_state <= TCPS_CLOSING)) && 995 (tmr_up == PACE_TMR_KEEP) && 996 (tp->snd_max == tp->snd_una)) { 997 /* We should have keep alive up and we do */ 998 return; 999 } 1000 } 1001 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { 1002 if ((tp->t_flags & TF_SENTFIN) && 1003 ((tp->snd_max - tp->snd_una) == 1) && 1004 (rsm->r_flags & BBR_HAS_FIN)) { 1005 /* needs to be a RXT */ 1006 if (tmr_up == PACE_TMR_RXT) 1007 return; 1008 else 1009 goto wrong_timer; 1010 } else if (tmr_up == PACE_TMR_RACK) 1011 return; 1012 else 1013 goto wrong_timer; 1014 } else if (rsm && (tmr_up == PACE_TMR_RACK)) { 1015 /* Rack timer has priority if we have data out */ 1016 return; 1017 } else if (SEQ_GT(tp->snd_max, tp->snd_una) && 1018 ((tmr_up == PACE_TMR_TLP) || 1019 (tmr_up == PACE_TMR_RXT))) { 1020 /* 1021 * Either a TLP or RXT is fine if no sack-passed is in place 1022 * and data is outstanding. 1023 */ 1024 return; 1025 } else if (tmr_up == PACE_TMR_DELACK) { 1026 /* 1027 * If the delayed ack was going to go off before the 1028 * rtx/tlp/rack timer were going to expire, then that would 1029 * be the timer in control. Note we don't check the time 1030 * here trusting the code is correct. 1031 */ 1032 return; 1033 } 1034 if (SEQ_GT(tp->snd_max, tp->snd_una) && 1035 ((tmr_up == PACE_TMR_RXT) || 1036 (tmr_up == PACE_TMR_TLP) || 1037 (tmr_up == PACE_TMR_RACK))) { 1038 /* 1039 * We have outstanding data and 1040 * we *do* have a RACK, TLP or RXT 1041 * timer running. We won't restart 1042 * anything here since thats probably ok we 1043 * will get called with some timer here shortly. 1044 */ 1045 return; 1046 } 1047 /* 1048 * Ok the timer originally started is not what we want now. We will 1049 * force the hpts to be stopped if any, and restart with the slot 1050 * set to what was in the saved slot. 1051 */ 1052 wrong_timer: 1053 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { 1054 if (tcp_in_hpts(inp)) 1055 tcp_hpts_remove(inp); 1056 bbr_timer_cancel(bbr, __LINE__, cts); 1057 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 1058 0); 1059 } else { 1060 /* 1061 * Output is hptsi so we just need to switch the type of 1062 * timer. We don't bother with keep-alive, since when we 1063 * jump through the output, it will start the keep-alive if 1064 * nothing is sent. 1065 * 1066 * We only need a delayed-ack added and or the hpts_timeout. 1067 */ 1068 hpts_timeout = bbr_timer_start(tp, bbr, cts); 1069 if (tp->t_flags & TF_DELACK) { 1070 if (hpts_timeout == 0) { 1071 hpts_timeout = bbr_delack_time; 1072 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1073 } 1074 else if (hpts_timeout > bbr_delack_time) { 1075 hpts_timeout = bbr_delack_time; 1076 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1077 } 1078 } 1079 if (hpts_timeout) { 1080 if (hpts_timeout > 0x7ffffffe) 1081 hpts_timeout = 0x7ffffffe; 1082 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 1083 } 1084 } 1085 } 1086 1087 int32_t bbr_clear_lost = 0; 1088 1089 /* 1090 * Considers the two time values now (cts) and earlier. 1091 * If cts is smaller than earlier, we could have 1092 * had a sequence wrap (our counter wraps every 1093 * 70 min or so) or it could be just clock skew 1094 * getting us two different time values. Clock skew 1095 * will show up within 10ms or so. So in such 1096 * a case (where cts is behind earlier time by 1097 * less than 10ms) we return 0. Otherwise we 1098 * return the true difference between them. 1099 */ 1100 static inline uint32_t 1101 bbr_calc_time(uint32_t cts, uint32_t earlier_time) { 1102 /* 1103 * Given two timestamps, the current time stamp cts, and some other 1104 * time-stamp taken in theory earlier return the difference. The 1105 * trick is here sometimes locking will get the other timestamp 1106 * after the cts. If this occurs we need to return 0. 1107 */ 1108 if (TSTMP_GEQ(cts, earlier_time)) 1109 return (cts - earlier_time); 1110 /* 1111 * cts is behind earlier_time if its less than 10ms consider it 0. 1112 * If its more than 10ms difference then we had a time wrap. Else 1113 * its just the normal locking foo. I wonder if we should not go to 1114 * 64bit TS and get rid of this issue. 1115 */ 1116 if (TSTMP_GEQ((cts + 10000), earlier_time)) 1117 return (0); 1118 /* 1119 * Ok the time must have wrapped. So we need to answer a large 1120 * amount of time, which the normal subtraction should do. 1121 */ 1122 return (cts - earlier_time); 1123 } 1124 1125 static int 1126 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) 1127 { 1128 uint32_t stat; 1129 int32_t error; 1130 1131 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); 1132 if (error || req->newptr == NULL) 1133 return error; 1134 1135 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 1136 if (error) 1137 return (error); 1138 if (stat == 1) { 1139 #ifdef BBR_INVARIANTS 1140 printf("Clearing BBR lost counters\n"); 1141 #endif 1142 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); 1143 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); 1144 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); 1145 } else if (stat == 2) { 1146 #ifdef BBR_INVARIANTS 1147 printf("Clearing BBR option counters\n"); 1148 #endif 1149 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); 1150 } else if (stat == 3) { 1151 #ifdef BBR_INVARIANTS 1152 printf("Clearing BBR stats counters\n"); 1153 #endif 1154 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); 1155 } else if (stat == 4) { 1156 #ifdef BBR_INVARIANTS 1157 printf("Clearing BBR out-size counters\n"); 1158 #endif 1159 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); 1160 } 1161 bbr_clear_lost = 0; 1162 return (0); 1163 } 1164 1165 static void 1166 bbr_init_sysctls(void) 1167 { 1168 struct sysctl_oid *bbr_probertt; 1169 struct sysctl_oid *bbr_hptsi; 1170 struct sysctl_oid *bbr_measure; 1171 struct sysctl_oid *bbr_cwnd; 1172 struct sysctl_oid *bbr_timeout; 1173 struct sysctl_oid *bbr_states; 1174 struct sysctl_oid *bbr_startup; 1175 struct sysctl_oid *bbr_policer; 1176 1177 /* Probe rtt controls */ 1178 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1179 SYSCTL_CHILDREN(bbr_sysctl_root), 1180 OID_AUTO, 1181 "probertt", 1182 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1183 ""); 1184 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1185 SYSCTL_CHILDREN(bbr_probertt), 1186 OID_AUTO, "gain", CTLFLAG_RW, 1187 &bbr_rttprobe_gain, 192, 1188 "What is the filter gain drop in probe_rtt (0=disable)?"); 1189 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1190 SYSCTL_CHILDREN(bbr_probertt), 1191 OID_AUTO, "cwnd", CTLFLAG_RW, 1192 &bbr_rtt_probe_cwndtarg, 4, 1193 "How many mss's are outstanding during probe-rtt"); 1194 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1195 SYSCTL_CHILDREN(bbr_probertt), 1196 OID_AUTO, "int", CTLFLAG_RW, 1197 &bbr_rtt_probe_limit, 4000000, 1198 "If RTT has not shrank in this many micro-seconds enter probe-rtt"); 1199 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1200 SYSCTL_CHILDREN(bbr_probertt), 1201 OID_AUTO, "mintime", CTLFLAG_RW, 1202 &bbr_rtt_probe_time, 200000, 1203 "How many microseconds in probe-rtt"); 1204 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1205 SYSCTL_CHILDREN(bbr_probertt), 1206 OID_AUTO, "filter_len_sec", CTLFLAG_RW, 1207 &bbr_filter_len_sec, 6, 1208 "How long in seconds does the rttProp filter run?"); 1209 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1210 SYSCTL_CHILDREN(bbr_probertt), 1211 OID_AUTO, "drain_rtt", CTLFLAG_RW, 1212 &bbr_drain_rtt, BBR_SRTT, 1213 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); 1214 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1215 SYSCTL_CHILDREN(bbr_probertt), 1216 OID_AUTO, "can_force", CTLFLAG_RW, 1217 &bbr_can_force_probertt, 0, 1218 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); 1219 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1220 SYSCTL_CHILDREN(bbr_probertt), 1221 OID_AUTO, "enter_sets_force", CTLFLAG_RW, 1222 &bbr_probertt_sets_rtt, 0, 1223 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); 1224 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1225 SYSCTL_CHILDREN(bbr_probertt), 1226 OID_AUTO, "can_adjust", CTLFLAG_RW, 1227 &bbr_can_adjust_probertt, 1, 1228 "Can we dynamically adjust the probe-rtt limits and times?"); 1229 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1230 SYSCTL_CHILDREN(bbr_probertt), 1231 OID_AUTO, "is_ratio", CTLFLAG_RW, 1232 &bbr_is_ratio, 0, 1233 "is the limit to filter a ratio?"); 1234 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1235 SYSCTL_CHILDREN(bbr_probertt), 1236 OID_AUTO, "use_cwnd", CTLFLAG_RW, 1237 &bbr_prtt_slam_cwnd, 0, 1238 "Should we set/recover cwnd?"); 1239 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1240 SYSCTL_CHILDREN(bbr_probertt), 1241 OID_AUTO, "can_use_ts", CTLFLAG_RW, 1242 &bbr_can_use_ts_for_rtt, 1, 1243 "Can we use the ms timestamp if available for retransmistted rtt calculations?"); 1244 1245 /* Pacing controls */ 1246 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1247 SYSCTL_CHILDREN(bbr_sysctl_root), 1248 OID_AUTO, 1249 "pacing", 1250 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1251 ""); 1252 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1253 SYSCTL_CHILDREN(bbr_hptsi), 1254 OID_AUTO, "hw_pacing", CTLFLAG_RW, 1255 &bbr_allow_hdwr_pacing, 1, 1256 "Do we allow hardware pacing?"); 1257 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1258 SYSCTL_CHILDREN(bbr_hptsi), 1259 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, 1260 &bbr_hardware_pacing_limit, 4000, 1261 "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); 1262 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1263 SYSCTL_CHILDREN(bbr_hptsi), 1264 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, 1265 &bbr_hdwr_pace_adjust, 2, 1266 "Multiplier to calculated tso size?"); 1267 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1268 SYSCTL_CHILDREN(bbr_hptsi), 1269 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, 1270 &bbr_hdwr_pace_floor, 1, 1271 "Do we invoke the hardware pacing floor?"); 1272 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1273 SYSCTL_CHILDREN(bbr_hptsi), 1274 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, 1275 &bbr_hdwr_pacing_delay_cnt, 10, 1276 "How many packets must be sent after hdwr pacing is enabled"); 1277 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1278 SYSCTL_CHILDREN(bbr_hptsi), 1279 OID_AUTO, "bw_cross", CTLFLAG_RW, 1280 &bbr_cross_over, 3000000, 1281 "What is the point where we cross over to linux like TSO size set"); 1282 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1283 SYSCTL_CHILDREN(bbr_hptsi), 1284 OID_AUTO, "seg_deltarg", CTLFLAG_RW, 1285 &bbr_hptsi_segments_delay_tar, 7000, 1286 "What is the worse case delay target for hptsi < 48Mbp connections"); 1287 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1288 SYSCTL_CHILDREN(bbr_hptsi), 1289 OID_AUTO, "enet_oh", CTLFLAG_RW, 1290 &bbr_include_enet_oh, 0, 1291 "Do we include the ethernet overhead in calculating pacing delay?"); 1292 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1293 SYSCTL_CHILDREN(bbr_hptsi), 1294 OID_AUTO, "ip_oh", CTLFLAG_RW, 1295 &bbr_include_ip_oh, 1, 1296 "Do we include the IP overhead in calculating pacing delay?"); 1297 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1298 SYSCTL_CHILDREN(bbr_hptsi), 1299 OID_AUTO, "tcp_oh", CTLFLAG_RW, 1300 &bbr_include_tcp_oh, 0, 1301 "Do we include the TCP overhead in calculating pacing delay?"); 1302 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1303 SYSCTL_CHILDREN(bbr_hptsi), 1304 OID_AUTO, "google_discount", CTLFLAG_RW, 1305 &bbr_google_discount, 10, 1306 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); 1307 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1308 SYSCTL_CHILDREN(bbr_hptsi), 1309 OID_AUTO, "all_get_min", CTLFLAG_RW, 1310 &bbr_all_get_min, 0, 1311 "If you are less than a MSS do you just get the min?"); 1312 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1313 SYSCTL_CHILDREN(bbr_hptsi), 1314 OID_AUTO, "tso_min", CTLFLAG_RW, 1315 &bbr_hptsi_bytes_min, 1460, 1316 "For 0 -> 24Mbps what is floor number of segments for TSO"); 1317 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1318 SYSCTL_CHILDREN(bbr_hptsi), 1319 OID_AUTO, "seg_tso_max", CTLFLAG_RW, 1320 &bbr_hptsi_segments_max, 6, 1321 "For 0 -> 24Mbps what is top number of segments for TSO"); 1322 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1323 SYSCTL_CHILDREN(bbr_hptsi), 1324 OID_AUTO, "seg_floor", CTLFLAG_RW, 1325 &bbr_hptsi_segments_floor, 1, 1326 "Minimum TSO size we will fall too in segments"); 1327 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1328 SYSCTL_CHILDREN(bbr_hptsi), 1329 OID_AUTO, "utter_max", CTLFLAG_RW, 1330 &bbr_hptsi_utter_max, 0, 1331 "The absolute maximum that any pacing (outside of hardware) can be"); 1332 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1333 SYSCTL_CHILDREN(bbr_hptsi), 1334 OID_AUTO, "seg_divisor", CTLFLAG_RW, 1335 &bbr_hptsi_per_second, 100, 1336 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); 1337 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1338 SYSCTL_CHILDREN(bbr_hptsi), 1339 OID_AUTO, "srtt_mul", CTLFLAG_RW, 1340 &bbr_hptsi_max_mul, 1, 1341 "The multiplier for pace len max"); 1342 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1343 SYSCTL_CHILDREN(bbr_hptsi), 1344 OID_AUTO, "srtt_div", CTLFLAG_RW, 1345 &bbr_hptsi_max_div, 2, 1346 "The divisor for pace len max"); 1347 /* Measurement controls */ 1348 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1349 SYSCTL_CHILDREN(bbr_sysctl_root), 1350 OID_AUTO, 1351 "measure", 1352 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1353 "Measurement controls"); 1354 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1355 SYSCTL_CHILDREN(bbr_measure), 1356 OID_AUTO, "min_i_bw", CTLFLAG_RW, 1357 &bbr_initial_bw_bps, 62500, 1358 "Minimum initial b/w in bytes per second"); 1359 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1360 SYSCTL_CHILDREN(bbr_measure), 1361 OID_AUTO, "no_sack_needed", CTLFLAG_RW, 1362 &bbr_sack_not_required, 0, 1363 "Do we allow bbr to run on connections not supporting SACK?"); 1364 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1365 SYSCTL_CHILDREN(bbr_measure), 1366 OID_AUTO, "use_google", CTLFLAG_RW, 1367 &bbr_use_google_algo, 0, 1368 "Use has close to google V1.0 has possible?"); 1369 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1370 SYSCTL_CHILDREN(bbr_measure), 1371 OID_AUTO, "ts_limiting", CTLFLAG_RW, 1372 &bbr_ts_limiting, 1, 1373 "Do we attempt to use the peers timestamp to limit b/w caculations?"); 1374 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1375 SYSCTL_CHILDREN(bbr_measure), 1376 OID_AUTO, "ts_can_raise", CTLFLAG_RW, 1377 &bbr_ts_can_raise, 0, 1378 "Can we raise the b/w via timestamp b/w calculation?"); 1379 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1380 SYSCTL_CHILDREN(bbr_measure), 1381 OID_AUTO, "ts_delta", CTLFLAG_RW, 1382 &bbr_min_usec_delta, 20000, 1383 "How long in usec between ts of our sends in ts validation code?"); 1384 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1385 SYSCTL_CHILDREN(bbr_measure), 1386 OID_AUTO, "ts_peer_delta", CTLFLAG_RW, 1387 &bbr_min_peer_delta, 20, 1388 "What min numerical value should be between the peer deltas?"); 1389 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1390 SYSCTL_CHILDREN(bbr_measure), 1391 OID_AUTO, "ts_delta_percent", CTLFLAG_RW, 1392 &bbr_delta_percent, 150, 1393 "What percentage (150 = 15.0) do we allow variance for?"); 1394 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1395 SYSCTL_CHILDREN(bbr_measure), 1396 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, 1397 &bbr_min_measurements_req, 1, 1398 "What is the minimum measurement count we need before we switch to our b/w estimate"); 1399 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1400 SYSCTL_CHILDREN(bbr_measure), 1401 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, 1402 &bbr_no_pacing_until, 4, 1403 "How many pkt-epoch's (0 is off) do we need before pacing is on?"); 1404 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1405 SYSCTL_CHILDREN(bbr_measure), 1406 OID_AUTO, "quanta", CTLFLAG_RW, 1407 &bbr_quanta, 2, 1408 "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); 1409 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1410 SYSCTL_CHILDREN(bbr_measure), 1411 OID_AUTO, "noretran", CTLFLAG_RW, 1412 &bbr_no_retran, 0, 1413 "Should google mode not use retransmission measurements for the b/w estimation?"); 1414 /* State controls */ 1415 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1416 SYSCTL_CHILDREN(bbr_sysctl_root), 1417 OID_AUTO, 1418 "states", 1419 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1420 "State controls"); 1421 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1422 SYSCTL_CHILDREN(bbr_states), 1423 OID_AUTO, "idle_restart", CTLFLAG_RW, 1424 &bbr_uses_idle_restart, 0, 1425 "Do we use a new special idle_restart state to ramp back up quickly?"); 1426 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1427 SYSCTL_CHILDREN(bbr_states), 1428 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, 1429 &bbr_idle_restart_threshold, 100000, 1430 "How long must we be idle before we restart??"); 1431 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1432 SYSCTL_CHILDREN(bbr_states), 1433 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, 1434 &bbr_state_is_pkt_epoch, 0, 1435 "Do we use a pkt-epoch for substate if 0 rttProp?"); 1436 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1437 SYSCTL_CHILDREN(bbr_states), 1438 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, 1439 &bbr_rtt_gain_thresh, 0, 1440 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); 1441 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1442 SYSCTL_CHILDREN(bbr_states), 1443 OID_AUTO, "drain_floor", CTLFLAG_RW, 1444 &bbr_drain_floor, 88, 1445 "What is the lowest we can drain (pg) too?"); 1446 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1447 SYSCTL_CHILDREN(bbr_states), 1448 OID_AUTO, "drain_2_target", CTLFLAG_RW, 1449 &bbr_state_drain_2_tar, 1, 1450 "Do we drain to target in drain substate?"); 1451 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1452 SYSCTL_CHILDREN(bbr_states), 1453 OID_AUTO, "gain_2_target", CTLFLAG_RW, 1454 &bbr_gain_to_target, 1, 1455 "Does probe bw gain to target??"); 1456 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1457 SYSCTL_CHILDREN(bbr_states), 1458 OID_AUTO, "gain_extra_time", CTLFLAG_RW, 1459 &bbr_gain_gets_extra_too, 1, 1460 "Does probe bw gain get the extra time too?"); 1461 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1462 SYSCTL_CHILDREN(bbr_states), 1463 OID_AUTO, "ld_div", CTLFLAG_RW, 1464 &bbr_drain_drop_div, 5, 1465 "Long drain drop divider?"); 1466 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1467 SYSCTL_CHILDREN(bbr_states), 1468 OID_AUTO, "ld_mul", CTLFLAG_RW, 1469 &bbr_drain_drop_mul, 4, 1470 "Long drain drop multiplier?"); 1471 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1472 SYSCTL_CHILDREN(bbr_states), 1473 OID_AUTO, "rand_ot_disc", CTLFLAG_RW, 1474 &bbr_rand_ot, 50, 1475 "Random discount of the ot?"); 1476 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1477 SYSCTL_CHILDREN(bbr_states), 1478 OID_AUTO, "dr_filter_life", CTLFLAG_RW, 1479 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, 1480 "How many packet-epochs does the b/w delivery rate last?"); 1481 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1482 SYSCTL_CHILDREN(bbr_states), 1483 OID_AUTO, "subdrain_applimited", CTLFLAG_RW, 1484 &bbr_sub_drain_app_limit, 0, 1485 "Does our sub-state drain invoke app limited if its long?"); 1486 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1487 SYSCTL_CHILDREN(bbr_states), 1488 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, 1489 &bbr_sub_drain_slam_cwnd, 0, 1490 "Should we set/recover cwnd for sub-state drain?"); 1491 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1492 SYSCTL_CHILDREN(bbr_states), 1493 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, 1494 &bbr_slam_cwnd_in_main_drain, 0, 1495 "Should we set/recover cwnd for main-state drain?"); 1496 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1497 SYSCTL_CHILDREN(bbr_states), 1498 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, 1499 &google_allow_early_out, 1, 1500 "Should we allow google probe-bw/drain to exit early at flight target?"); 1501 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1502 SYSCTL_CHILDREN(bbr_states), 1503 OID_AUTO, "google_exit_loss", CTLFLAG_RW, 1504 &google_consider_lost, 1, 1505 "Should we have losses exit gain of probebw in google mode??"); 1506 /* Startup controls */ 1507 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1508 SYSCTL_CHILDREN(bbr_sysctl_root), 1509 OID_AUTO, 1510 "startup", 1511 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1512 "Startup controls"); 1513 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1514 SYSCTL_CHILDREN(bbr_startup), 1515 OID_AUTO, "cheat_iwnd", CTLFLAG_RW, 1516 &bbr_sends_full_iwnd, 1, 1517 "Do we not pace but burst out initial windows has our TSO size?"); 1518 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1519 SYSCTL_CHILDREN(bbr_startup), 1520 OID_AUTO, "loss_threshold", CTLFLAG_RW, 1521 &bbr_startup_loss_thresh, 2000, 1522 "In startup what is the loss threshold in a pe that will exit us from startup?"); 1523 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1524 SYSCTL_CHILDREN(bbr_startup), 1525 OID_AUTO, "use_lowerpg", CTLFLAG_RW, 1526 &bbr_use_lower_gain_in_startup, 1, 1527 "Should we use a lower hptsi gain if we see loss in startup?"); 1528 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1529 SYSCTL_CHILDREN(bbr_startup), 1530 OID_AUTO, "gain", CTLFLAG_RW, 1531 &bbr_start_exit, 25, 1532 "What gain percent do we need to see to stay in startup??"); 1533 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1534 SYSCTL_CHILDREN(bbr_startup), 1535 OID_AUTO, "low_gain", CTLFLAG_RW, 1536 &bbr_low_start_exit, 15, 1537 "What gain percent do we need to see to stay in the lower gain startup??"); 1538 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1539 SYSCTL_CHILDREN(bbr_startup), 1540 OID_AUTO, "loss_exit", CTLFLAG_RW, 1541 &bbr_exit_startup_at_loss, 1, 1542 "Should we exit startup at loss in an epoch if we are not gaining?"); 1543 /* CWND controls */ 1544 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1545 SYSCTL_CHILDREN(bbr_sysctl_root), 1546 OID_AUTO, 1547 "cwnd", 1548 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1549 "Cwnd controls"); 1550 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1551 SYSCTL_CHILDREN(bbr_cwnd), 1552 OID_AUTO, "tar_rtt", CTLFLAG_RW, 1553 &bbr_cwndtarget_rtt_touse, 0, 1554 "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); 1555 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1556 SYSCTL_CHILDREN(bbr_cwnd), 1557 OID_AUTO, "may_shrink", CTLFLAG_RW, 1558 &bbr_cwnd_may_shrink, 0, 1559 "Can the cwnd shrink if it would grow to more than the target?"); 1560 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1561 SYSCTL_CHILDREN(bbr_cwnd), 1562 OID_AUTO, "max_target_limit", CTLFLAG_RW, 1563 &bbr_target_cwnd_mult_limit, 8, 1564 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); 1565 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1566 SYSCTL_CHILDREN(bbr_cwnd), 1567 OID_AUTO, "highspeed_min", CTLFLAG_RW, 1568 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, 1569 "What is the high-speed min cwnd (rttProp under 1ms)"); 1570 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1571 SYSCTL_CHILDREN(bbr_cwnd), 1572 OID_AUTO, "lowspeed_min", CTLFLAG_RW, 1573 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, 1574 "What is the min cwnd (rttProp > 1ms)"); 1575 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1576 SYSCTL_CHILDREN(bbr_cwnd), 1577 OID_AUTO, "initwin", CTLFLAG_RW, 1578 &bbr_def_init_win, 10, 1579 "What is the BBR initial window, if 0 use tcp version"); 1580 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1581 SYSCTL_CHILDREN(bbr_cwnd), 1582 OID_AUTO, "do_loss_red", CTLFLAG_RW, 1583 &bbr_do_red, 600, 1584 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); 1585 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1586 SYSCTL_CHILDREN(bbr_cwnd), 1587 OID_AUTO, "red_scale", CTLFLAG_RW, 1588 &bbr_red_scale, 20000, 1589 "What RTT do we scale with?"); 1590 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1591 SYSCTL_CHILDREN(bbr_cwnd), 1592 OID_AUTO, "red_growslow", CTLFLAG_RW, 1593 &bbr_red_growth_restrict, 1, 1594 "Do we restrict cwnd growth for whats in flight?"); 1595 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1596 SYSCTL_CHILDREN(bbr_cwnd), 1597 OID_AUTO, "red_div", CTLFLAG_RW, 1598 &bbr_red_div, 2, 1599 "If we reduce whats the divisor?"); 1600 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1601 SYSCTL_CHILDREN(bbr_cwnd), 1602 OID_AUTO, "red_mul", CTLFLAG_RW, 1603 &bbr_red_mul, 1, 1604 "If we reduce whats the mulitiplier?"); 1605 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1606 SYSCTL_CHILDREN(bbr_cwnd), 1607 OID_AUTO, "target_is_unit", CTLFLAG_RW, 1608 &bbr_target_is_bbunit, 0, 1609 "Is the state target the pacing_gain or BBR_UNIT?"); 1610 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1611 SYSCTL_CHILDREN(bbr_cwnd), 1612 OID_AUTO, "drop_limit", CTLFLAG_RW, 1613 &bbr_drop_limit, 0, 1614 "Number of segments limit for drop (0=use min_cwnd w/flight)?"); 1615 1616 /* Timeout controls */ 1617 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1618 SYSCTL_CHILDREN(bbr_sysctl_root), 1619 OID_AUTO, 1620 "timeout", 1621 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1622 "Time out controls"); 1623 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1624 SYSCTL_CHILDREN(bbr_timeout), 1625 OID_AUTO, "delack", CTLFLAG_RW, 1626 &bbr_delack_time, 100000, 1627 "BBR's delayed ack time"); 1628 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1629 SYSCTL_CHILDREN(bbr_timeout), 1630 OID_AUTO, "tlp_uses", CTLFLAG_RW, 1631 &bbr_tlp_type_to_use, 3, 1632 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); 1633 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1634 SYSCTL_CHILDREN(bbr_timeout), 1635 OID_AUTO, "persmin", CTLFLAG_RW, 1636 &bbr_persist_min, 250000, 1637 "What is the minimum time in microseconds between persists"); 1638 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1639 SYSCTL_CHILDREN(bbr_timeout), 1640 OID_AUTO, "persmax", CTLFLAG_RW, 1641 &bbr_persist_max, 1000000, 1642 "What is the largest delay in microseconds between persists"); 1643 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1644 SYSCTL_CHILDREN(bbr_timeout), 1645 OID_AUTO, "tlp_minto", CTLFLAG_RW, 1646 &bbr_tlp_min, 10000, 1647 "TLP Min timeout in usecs"); 1648 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1649 SYSCTL_CHILDREN(bbr_timeout), 1650 OID_AUTO, "tlp_dack_time", CTLFLAG_RW, 1651 &bbr_delayed_ack_time, 200000, 1652 "TLP delayed ack compensation value"); 1653 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1654 SYSCTL_CHILDREN(bbr_sysctl_root), 1655 OID_AUTO, "minrto", CTLFLAG_RW, 1656 &bbr_rto_min_ms, 30, 1657 "Minimum RTO in ms"); 1658 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1659 SYSCTL_CHILDREN(bbr_timeout), 1660 OID_AUTO, "maxrto", CTLFLAG_RW, 1661 &bbr_rto_max_sec, 4, 1662 "Maximum RTO in seconds -- should be at least as large as min_rto"); 1663 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1664 SYSCTL_CHILDREN(bbr_timeout), 1665 OID_AUTO, "tlp_retry", CTLFLAG_RW, 1666 &bbr_tlp_max_resend, 2, 1667 "How many times does TLP retry a single segment or multiple with no ACK"); 1668 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1669 SYSCTL_CHILDREN(bbr_timeout), 1670 OID_AUTO, "minto", CTLFLAG_RW, 1671 &bbr_min_to, 1000, 1672 "Minimum rack timeout in useconds"); 1673 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1674 SYSCTL_CHILDREN(bbr_timeout), 1675 OID_AUTO, "pktdelay", CTLFLAG_RW, 1676 &bbr_pkt_delay, 1000, 1677 "Extra RACK time (in useconds) besides reordering thresh"); 1678 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1679 SYSCTL_CHILDREN(bbr_timeout), 1680 OID_AUTO, "incr_tmrs", CTLFLAG_RW, 1681 &bbr_incr_timers, 1, 1682 "Increase the RXT/TLP timer by the pacing time used?"); 1683 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1684 SYSCTL_CHILDREN(bbr_timeout), 1685 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, 1686 &bbr_marks_rxt_sack_passed, 0, 1687 "Mark sack passed on all those not ack'd when a RXT hits?"); 1688 /* Policer controls */ 1689 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1690 SYSCTL_CHILDREN(bbr_sysctl_root), 1691 OID_AUTO, 1692 "policer", 1693 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1694 "Policer controls"); 1695 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1696 SYSCTL_CHILDREN(bbr_policer), 1697 OID_AUTO, "detect_enable", CTLFLAG_RW, 1698 &bbr_policer_detection_enabled, 1, 1699 "Is policer detection enabled??"); 1700 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1701 SYSCTL_CHILDREN(bbr_policer), 1702 OID_AUTO, "min_pes", CTLFLAG_RW, 1703 &bbr_lt_intvl_min_rtts, 4, 1704 "Minimum number of PE's?"); 1705 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1706 SYSCTL_CHILDREN(bbr_policer), 1707 OID_AUTO, "bwdiff", CTLFLAG_RW, 1708 &bbr_lt_bw_diff, (4000/8), 1709 "Minimal bw diff?"); 1710 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1711 SYSCTL_CHILDREN(bbr_policer), 1712 OID_AUTO, "bwratio", CTLFLAG_RW, 1713 &bbr_lt_bw_ratio, 8, 1714 "Minimal bw diff?"); 1715 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1716 SYSCTL_CHILDREN(bbr_policer), 1717 OID_AUTO, "from_rack_rxt", CTLFLAG_RW, 1718 &bbr_policer_call_from_rack_to, 0, 1719 "Do we call the policer detection code from a rack-timeout?"); 1720 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1721 SYSCTL_CHILDREN(bbr_policer), 1722 OID_AUTO, "false_postive", CTLFLAG_RW, 1723 &bbr_lt_intvl_fp, 0, 1724 "What packet epoch do we do false-positive detection at (0=no)?"); 1725 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1726 SYSCTL_CHILDREN(bbr_policer), 1727 OID_AUTO, "loss_thresh", CTLFLAG_RW, 1728 &bbr_lt_loss_thresh, 196, 1729 "Loss threshold 196 = 19.6%?"); 1730 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1731 SYSCTL_CHILDREN(bbr_policer), 1732 OID_AUTO, "false_postive_thresh", CTLFLAG_RW, 1733 &bbr_lt_fd_thresh, 100, 1734 "What percentage is the false detection threshold (150=15.0)?"); 1735 /* All the rest */ 1736 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1737 SYSCTL_CHILDREN(bbr_sysctl_root), 1738 OID_AUTO, "cheat_rxt", CTLFLAG_RW, 1739 &bbr_use_rack_resend_cheat, 0, 1740 "Do we burst 1ms between sends on retransmissions (like rack)?"); 1741 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1742 SYSCTL_CHILDREN(bbr_sysctl_root), 1743 OID_AUTO, "error_paceout", CTLFLAG_RW, 1744 &bbr_error_base_paceout, 10000, 1745 "When we hit an error what is the min to pace out in usec's?"); 1746 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1747 SYSCTL_CHILDREN(bbr_sysctl_root), 1748 OID_AUTO, "kill_paceout", CTLFLAG_RW, 1749 &bbr_max_net_error_cnt, 10, 1750 "When we hit this many errors in a row, kill the session?"); 1751 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1752 SYSCTL_CHILDREN(bbr_sysctl_root), 1753 OID_AUTO, "data_after_close", CTLFLAG_RW, 1754 &bbr_ignore_data_after_close, 1, 1755 "Do we hold off sending a RST until all pending data is ack'd"); 1756 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1757 SYSCTL_CHILDREN(bbr_sysctl_root), 1758 OID_AUTO, "resend_use_tso", CTLFLAG_RW, 1759 &bbr_resends_use_tso, 0, 1760 "Can resends use TSO?"); 1761 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1762 SYSCTL_CHILDREN(bbr_sysctl_root), 1763 OID_AUTO, "sblklimit", CTLFLAG_RW, 1764 &bbr_sack_block_limit, 128, 1765 "When do we start ignoring small sack blocks"); 1766 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1767 SYSCTL_CHILDREN(bbr_sysctl_root), 1768 OID_AUTO, "bb_verbose", CTLFLAG_RW, 1769 &bbr_verbose_logging, 0, 1770 "Should BBR black box logging be verbose"); 1771 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1772 SYSCTL_CHILDREN(bbr_sysctl_root), 1773 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 1774 &bbr_reorder_thresh, 2, 1775 "What factor for rack will be added when seeing reordering (shift right)"); 1776 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1777 SYSCTL_CHILDREN(bbr_sysctl_root), 1778 OID_AUTO, "reorder_fade", CTLFLAG_RW, 1779 &bbr_reorder_fade, 0, 1780 "Does reorder detection fade, if so how many ms (0 means never)"); 1781 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1782 SYSCTL_CHILDREN(bbr_sysctl_root), 1783 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 1784 &bbr_tlp_thresh, 1, 1785 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 1786 /* Stats and counters */ 1787 /* The pacing counters for hdwr/software can't be in the array */ 1788 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1789 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1790 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1791 SYSCTL_CHILDREN(bbr_sysctl_root), 1792 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, 1793 &bbr_hdwr_pacing_enobuf, 1794 "Total number of enobufs for hardware paced flows"); 1795 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1796 SYSCTL_CHILDREN(bbr_sysctl_root), 1797 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, 1798 &bbr_nohdwr_pacing_enobuf, 1799 "Total number of enobufs for non-hardware paced flows"); 1800 1801 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); 1802 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1803 SYSCTL_CHILDREN(bbr_sysctl_root), 1804 OID_AUTO, "hdwr_pacing", CTLFLAG_RD, 1805 &bbr_flows_whdwr_pacing, 1806 "Total number of hardware paced flows"); 1807 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); 1808 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1809 SYSCTL_CHILDREN(bbr_sysctl_root), 1810 OID_AUTO, "software_pacing", CTLFLAG_RD, 1811 &bbr_flows_nohdwr_pacing, 1812 "Total number of software paced flows"); 1813 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); 1814 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1815 OID_AUTO, "stats", CTLFLAG_RD, 1816 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); 1817 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); 1818 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1819 OID_AUTO, "opts", CTLFLAG_RD, 1820 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); 1821 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); 1822 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1823 OID_AUTO, "lost", CTLFLAG_RD, 1824 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); 1825 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); 1826 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1827 OID_AUTO, "stateresend", CTLFLAG_RD, 1828 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); 1829 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); 1830 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1831 OID_AUTO, "statetime", CTLFLAG_RD, 1832 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); 1833 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 1834 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1835 OID_AUTO, "outsize", CTLFLAG_RD, 1836 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); 1837 SYSCTL_ADD_PROC(&bbr_sysctl_ctx, 1838 SYSCTL_CHILDREN(bbr_sysctl_root), 1839 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1840 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); 1841 } 1842 1843 static void 1844 bbr_counter_destroy(void) 1845 { 1846 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); 1847 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); 1848 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); 1849 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); 1850 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); 1851 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); 1852 counter_u64_free(bbr_nohdwr_pacing_enobuf); 1853 counter_u64_free(bbr_hdwr_pacing_enobuf); 1854 counter_u64_free(bbr_flows_whdwr_pacing); 1855 counter_u64_free(bbr_flows_nohdwr_pacing); 1856 1857 } 1858 1859 static __inline void 1860 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) 1861 { 1862 memset(l, 0, sizeof(union tcp_log_stackspecific)); 1863 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; 1864 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); 1865 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 1866 l->bw_inuse = bbr_get_bw(bbr); 1867 l->inflight = ctf_flight_size(bbr->rc_tp, 1868 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 1869 l->applimited = bbr->r_ctl.r_app_limited_until; 1870 l->delivered = bbr->r_ctl.rc_delivered; 1871 l->timeStamp = cts; 1872 l->lost = bbr->r_ctl.rc_lost; 1873 l->bbr_state = bbr->rc_bbr_state; 1874 l->bbr_substate = bbr_state_val(bbr); 1875 l->epoch = bbr->r_ctl.rc_rtt_epoch; 1876 l->lt_epoch = bbr->r_ctl.rc_lt_epoch; 1877 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 1878 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; 1879 l->inhpts = tcp_in_hpts(bbr->rc_inp); 1880 l->use_lt_bw = bbr->rc_lt_use_bw; 1881 l->pkts_out = bbr->r_ctl.rc_flight_at_input; 1882 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; 1883 } 1884 1885 static void 1886 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) 1887 { 1888 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1889 union tcp_log_stackspecific log; 1890 1891 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1892 log.u_bbr.flex1 = 0; 1893 log.u_bbr.flex2 = 0; 1894 log.u_bbr.flex5 = 0; 1895 log.u_bbr.flex3 = 0; 1896 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; 1897 log.u_bbr.flex7 = reason; 1898 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; 1899 log.u_bbr.flex8 = 0; 1900 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1901 &bbr->rc_inp->inp_socket->so_rcv, 1902 &bbr->rc_inp->inp_socket->so_snd, 1903 BBR_LOG_BW_RED_EV, 0, 1904 0, &log, false, &bbr->rc_tv); 1905 } 1906 } 1907 1908 static void 1909 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) 1910 { 1911 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1912 union tcp_log_stackspecific log; 1913 1914 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1915 log.u_bbr.flex1 = seq; 1916 log.u_bbr.flex2 = count; 1917 log.u_bbr.flex8 = mode; 1918 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1919 &bbr->rc_inp->inp_socket->so_rcv, 1920 &bbr->rc_inp->inp_socket->so_snd, 1921 BBR_LOG_LOWGAIN, 0, 1922 0, &log, false, &bbr->rc_tv); 1923 } 1924 } 1925 1926 static void 1927 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, 1928 uint8_t reason, uint32_t p_maxseg, int len) 1929 { 1930 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1931 union tcp_log_stackspecific log; 1932 1933 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1934 log.u_bbr.flex1 = p_maxseg; 1935 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; 1936 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 1937 log.u_bbr.flex4 = reason; 1938 log.u_bbr.flex5 = bbr->rc_in_persist; 1939 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; 1940 log.u_bbr.flex7 = p_maxseg; 1941 log.u_bbr.flex8 = bbr->rc_in_persist; 1942 log.u_bbr.pkts_out = 0; 1943 log.u_bbr.applimited = len; 1944 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1945 &bbr->rc_inp->inp_socket->so_rcv, 1946 &bbr->rc_inp->inp_socket->so_snd, 1947 BBR_LOG_JUSTRET, 0, 1948 tlen, &log, false, &bbr->rc_tv); 1949 } 1950 } 1951 1952 static void 1953 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) 1954 { 1955 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1956 union tcp_log_stackspecific log; 1957 1958 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1959 log.u_bbr.flex1 = seq; 1960 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 1961 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; 1962 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1963 &bbr->rc_inp->inp_socket->so_rcv, 1964 &bbr->rc_inp->inp_socket->so_snd, 1965 BBR_LOG_ENTREC, 0, 1966 0, &log, false, &bbr->rc_tv); 1967 } 1968 } 1969 1970 static void 1971 bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts) 1972 { 1973 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 1974 union tcp_log_stackspecific log; 1975 1976 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1977 log.u_bbr.flex1 = tso; 1978 log.u_bbr.flex2 = maxseg; 1979 log.u_bbr.flex3 = mtu; 1980 log.u_bbr.flex4 = csum_flags; 1981 TCP_LOG_EVENTP(tp, NULL, 1982 &bbr->rc_inp->inp_socket->so_rcv, 1983 &bbr->rc_inp->inp_socket->so_snd, 1984 BBR_LOG_MSGSIZE, 0, 1985 0, &log, false, &bbr->rc_tv); 1986 } 1987 } 1988 1989 static void 1990 bbr_log_flowend(struct tcp_bbr *bbr) 1991 { 1992 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1993 union tcp_log_stackspecific log; 1994 struct sockbuf *r, *s; 1995 struct timeval tv; 1996 1997 if (bbr->rc_inp->inp_socket) { 1998 r = &bbr->rc_inp->inp_socket->so_rcv; 1999 s = &bbr->rc_inp->inp_socket->so_snd; 2000 } else { 2001 r = s = NULL; 2002 } 2003 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); 2004 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2005 r, s, 2006 TCP_LOG_FLOWEND, 0, 2007 0, &log, false, &tv); 2008 } 2009 } 2010 2011 static void 2012 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, 2013 uint32_t lost, uint32_t del) 2014 { 2015 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2016 union tcp_log_stackspecific log; 2017 2018 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2019 log.u_bbr.flex1 = lost; 2020 log.u_bbr.flex2 = del; 2021 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; 2022 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; 2023 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2024 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2025 log.u_bbr.flex7 = line; 2026 log.u_bbr.flex8 = 0; 2027 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; 2028 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2029 &bbr->rc_inp->inp_socket->so_rcv, 2030 &bbr->rc_inp->inp_socket->so_snd, 2031 BBR_LOG_PKT_EPOCH, 0, 2032 0, &log, false, &bbr->rc_tv); 2033 } 2034 } 2035 2036 static void 2037 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) 2038 { 2039 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2040 union tcp_log_stackspecific log; 2041 2042 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2043 log.u_bbr.flex1 = bbr->r_ctl.rc_lost; 2044 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; 2045 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; 2046 log.u_bbr.flex7 = line; 2047 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2048 &bbr->rc_inp->inp_socket->so_rcv, 2049 &bbr->rc_inp->inp_socket->so_snd, 2050 BBR_LOG_TIME_EPOCH, 0, 2051 0, &log, false, &bbr->rc_tv); 2052 } 2053 } 2054 2055 static void 2056 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) 2057 { 2058 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2059 union tcp_log_stackspecific log; 2060 2061 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2062 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2063 log.u_bbr.flex2 = new_tar; 2064 log.u_bbr.flex3 = line; 2065 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2066 log.u_bbr.flex5 = bbr_quanta; 2067 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; 2068 log.u_bbr.flex7 = bbr->rc_last_options; 2069 log.u_bbr.flex8 = meth; 2070 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2071 &bbr->rc_inp->inp_socket->so_rcv, 2072 &bbr->rc_inp->inp_socket->so_snd, 2073 BBR_LOG_STATE_TARGET, 0, 2074 0, &log, false, &bbr->rc_tv); 2075 } 2076 2077 } 2078 2079 static void 2080 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2081 { 2082 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2083 union tcp_log_stackspecific log; 2084 2085 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2086 log.u_bbr.flex1 = line; 2087 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2088 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2089 if (bbr_state_is_pkt_epoch) 2090 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 2091 else 2092 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); 2093 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2094 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2095 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); 2096 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; 2097 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; 2098 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2099 &bbr->rc_inp->inp_socket->so_rcv, 2100 &bbr->rc_inp->inp_socket->so_snd, 2101 BBR_LOG_STATE, 0, 2102 0, &log, false, &bbr->rc_tv); 2103 } 2104 } 2105 2106 static void 2107 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 2108 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) 2109 { 2110 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2111 union tcp_log_stackspecific log; 2112 2113 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2114 log.u_bbr.flex1 = line; 2115 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2116 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; 2117 log.u_bbr.flex4 = applied; 2118 log.u_bbr.flex5 = rtt; 2119 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2120 log.u_bbr.flex7 = cond; 2121 log.u_bbr.flex8 = reas; 2122 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2123 &bbr->rc_inp->inp_socket->so_rcv, 2124 &bbr->rc_inp->inp_socket->so_snd, 2125 BBR_LOG_RTT_SHRINKS, 0, 2126 0, &log, false, &bbr->rc_tv); 2127 } 2128 } 2129 2130 static void 2131 bbr_log_type_exit_rec(struct tcp_bbr *bbr) 2132 { 2133 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2134 union tcp_log_stackspecific log; 2135 2136 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2137 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; 2138 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 2139 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2140 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2141 &bbr->rc_inp->inp_socket->so_rcv, 2142 &bbr->rc_inp->inp_socket->so_snd, 2143 BBR_LOG_EXITREC, 0, 2144 0, &log, false, &bbr->rc_tv); 2145 } 2146 } 2147 2148 static void 2149 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, 2150 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) 2151 { 2152 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2153 union tcp_log_stackspecific log; 2154 2155 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2156 log.u_bbr.flex1 = line; 2157 log.u_bbr.flex2 = prev_acked; 2158 log.u_bbr.flex3 = bytes_this_ack; 2159 log.u_bbr.flex4 = chg; 2160 log.u_bbr.flex5 = th_ack; 2161 log.u_bbr.flex6 = target; 2162 log.u_bbr.flex8 = meth; 2163 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2164 &bbr->rc_inp->inp_socket->so_rcv, 2165 &bbr->rc_inp->inp_socket->so_snd, 2166 BBR_LOG_CWND, 0, 2167 0, &log, false, &bbr->rc_tv); 2168 } 2169 } 2170 2171 static void 2172 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) 2173 { 2174 /* 2175 * Log the rtt sample we are applying to the srtt algorithm in 2176 * useconds. 2177 */ 2178 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2179 union tcp_log_stackspecific log; 2180 2181 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2182 log.u_bbr.flex1 = rtt; 2183 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; 2184 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; 2185 log.u_bbr.flex4 = bbr->rc_tp->ts_offset; 2186 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2187 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); 2188 log.u_bbr.flex6 = tsin; 2189 log.u_bbr.flex7 = 0; 2190 log.u_bbr.flex8 = bbr->rc_ack_was_delayed; 2191 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2192 &bbr->rc_inp->inp_socket->so_rcv, 2193 &bbr->rc_inp->inp_socket->so_snd, 2194 TCP_LOG_RTT, 0, 2195 0, &log, false, &bbr->rc_tv); 2196 } 2197 } 2198 2199 static void 2200 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) 2201 { 2202 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2203 union tcp_log_stackspecific log; 2204 2205 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2206 log.u_bbr.flex1 = time_in; 2207 log.u_bbr.flex2 = line; 2208 log.u_bbr.flex8 = enter_exit; 2209 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2210 &bbr->rc_inp->inp_socket->so_rcv, 2211 &bbr->rc_inp->inp_socket->so_snd, 2212 BBR_LOG_PERSIST, 0, 2213 0, &log, false, &bbr->rc_tv); 2214 } 2215 } 2216 static void 2217 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) 2218 { 2219 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2220 union tcp_log_stackspecific log; 2221 2222 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2223 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; 2224 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2225 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2226 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; 2227 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2228 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2229 &bbr->rc_inp->inp_socket->so_rcv, 2230 &bbr->rc_inp->inp_socket->so_snd, 2231 BBR_LOG_ACKCLEAR, 0, 2232 0, &log, false, &bbr->rc_tv); 2233 } 2234 } 2235 2236 static void 2237 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, 2238 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) 2239 { 2240 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2241 union tcp_log_stackspecific log; 2242 struct timeval tv; 2243 2244 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2245 log.u_bbr.flex1 = nsegs; 2246 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; 2247 if (m) { 2248 struct timespec ts; 2249 2250 log.u_bbr.flex3 = m->m_flags; 2251 if (m->m_flags & M_TSTMP) { 2252 mbuf_tstmp2timespec(m, &ts); 2253 tv.tv_sec = ts.tv_sec; 2254 tv.tv_usec = ts.tv_nsec / 1000; 2255 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); 2256 } else { 2257 log.u_bbr.lt_epoch = 0; 2258 } 2259 if (m->m_flags & M_TSTMP_LRO) { 2260 mbuf_tstmp2timeval(m, &tv); 2261 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); 2262 } else { 2263 /* No arrival timestamp */ 2264 log.u_bbr.flex5 = 0; 2265 } 2266 2267 log.u_bbr.pkts_out = tcp_get_usecs(&tv); 2268 } else { 2269 log.u_bbr.flex3 = 0; 2270 log.u_bbr.flex5 = 0; 2271 log.u_bbr.flex6 = 0; 2272 log.u_bbr.pkts_out = 0; 2273 } 2274 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; 2275 log.u_bbr.flex7 = bbr->r_wanted_output; 2276 log.u_bbr.flex8 = bbr->rc_in_persist; 2277 TCP_LOG_EVENTP(bbr->rc_tp, th, 2278 &bbr->rc_inp->inp_socket->so_rcv, 2279 &bbr->rc_inp->inp_socket->so_snd, 2280 TCP_LOG_IN, 0, 2281 tlen, &log, true, &bbr->rc_tv); 2282 } 2283 } 2284 2285 static void 2286 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) 2287 { 2288 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2289 union tcp_log_stackspecific log; 2290 2291 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2292 log.u_bbr.flex1 = did_out; 2293 log.u_bbr.flex2 = nxt_pkt; 2294 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; 2295 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2296 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; 2297 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; 2298 log.u_bbr.flex7 = bbr->r_wanted_output; 2299 log.u_bbr.flex8 = bbr->rc_in_persist; 2300 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; 2301 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2302 &bbr->rc_inp->inp_socket->so_rcv, 2303 &bbr->rc_inp->inp_socket->so_snd, 2304 BBR_LOG_DOSEG_DONE, 0, 2305 0, &log, true, &bbr->rc_tv); 2306 } 2307 } 2308 2309 static void 2310 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, 2311 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) 2312 { 2313 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2314 union tcp_log_stackspecific log; 2315 2316 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2317 log.u_bbr.flex1 = line; 2318 log.u_bbr.flex2 = o_len; 2319 log.u_bbr.flex3 = segcnt; 2320 log.u_bbr.flex4 = segsiz; 2321 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2322 &bbr->rc_inp->inp_socket->so_rcv, 2323 &bbr->rc_inp->inp_socket->so_snd, 2324 BBR_LOG_ENOBUF_JMP, ENOBUFS, 2325 len, &log, true, &bbr->rc_tv); 2326 } 2327 } 2328 2329 static void 2330 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) 2331 { 2332 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2333 union tcp_log_stackspecific log; 2334 2335 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2336 log.u_bbr.flex1 = timers; 2337 log.u_bbr.flex2 = ret; 2338 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 2339 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2340 log.u_bbr.flex5 = cts; 2341 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2342 log.u_bbr.flex8 = hpts_calling; 2343 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2344 &bbr->rc_inp->inp_socket->so_rcv, 2345 &bbr->rc_inp->inp_socket->so_snd, 2346 BBR_LOG_TO_PROCESS, 0, 2347 0, &log, false, &bbr->rc_tv); 2348 } 2349 } 2350 2351 static void 2352 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) 2353 { 2354 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2355 union tcp_log_stackspecific log; 2356 uint64_t ar; 2357 2358 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2359 log.u_bbr.flex1 = bbr->bbr_timer_src; 2360 log.u_bbr.flex2 = 0; 2361 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2362 ar = (uint64_t)(bbr->r_ctl.rc_resend); 2363 ar >>= 32; 2364 ar &= 0x00000000ffffffff; 2365 log.u_bbr.flex4 = (uint32_t)ar; 2366 ar = (uint64_t)bbr->r_ctl.rc_resend; 2367 ar &= 0x00000000ffffffff; 2368 log.u_bbr.flex5 = (uint32_t)ar; 2369 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2370 log.u_bbr.flex8 = to_num; 2371 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2372 &bbr->rc_inp->inp_socket->so_rcv, 2373 &bbr->rc_inp->inp_socket->so_snd, 2374 BBR_LOG_RTO, 0, 2375 0, &log, false, &bbr->rc_tv); 2376 } 2377 } 2378 2379 static void 2380 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) 2381 { 2382 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2383 union tcp_log_stackspecific log; 2384 2385 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2386 log.u_bbr.flex1 = flex1; 2387 log.u_bbr.flex2 = flex2; 2388 log.u_bbr.flex3 = flex3; 2389 log.u_bbr.flex4 = 0; 2390 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2391 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2392 log.u_bbr.flex8 = reason; 2393 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; 2394 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2395 &bbr->rc_inp->inp_socket->so_rcv, 2396 &bbr->rc_inp->inp_socket->so_snd, 2397 BBR_LOG_REDUCE, 0, 2398 0, &log, false, &bbr->rc_tv); 2399 } 2400 } 2401 2402 static void 2403 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) 2404 { 2405 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2406 union tcp_log_stackspecific log; 2407 2408 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2409 log.u_bbr.flex1 = diag->p_nxt_slot; 2410 log.u_bbr.flex2 = diag->p_cur_slot; 2411 log.u_bbr.flex3 = diag->slot_req; 2412 log.u_bbr.flex4 = diag->inp_hptsslot; 2413 log.u_bbr.flex5 = diag->slot_remaining; 2414 log.u_bbr.flex6 = diag->need_new_to; 2415 log.u_bbr.flex7 = diag->p_hpts_active; 2416 log.u_bbr.flex8 = diag->p_on_min_sleep; 2417 /* Hijack other fields as needed */ 2418 log.u_bbr.epoch = diag->have_slept; 2419 log.u_bbr.lt_epoch = diag->yet_to_sleep; 2420 log.u_bbr.pkts_out = diag->co_ret; 2421 log.u_bbr.applimited = diag->hpts_sleep_time; 2422 log.u_bbr.delivered = diag->p_prev_slot; 2423 log.u_bbr.inflight = diag->p_runningslot; 2424 log.u_bbr.bw_inuse = diag->wheel_slot; 2425 log.u_bbr.rttProp = diag->wheel_cts; 2426 log.u_bbr.delRate = diag->maxslots; 2427 log.u_bbr.cur_del_rate = diag->p_curtick; 2428 log.u_bbr.cur_del_rate <<= 32; 2429 log.u_bbr.cur_del_rate |= diag->p_lasttick; 2430 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2431 &bbr->rc_inp->inp_socket->so_rcv, 2432 &bbr->rc_inp->inp_socket->so_snd, 2433 BBR_LOG_HPTSDIAG, 0, 2434 0, &log, false, &bbr->rc_tv); 2435 } 2436 } 2437 2438 static void 2439 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 2440 uint32_t thresh, uint32_t to) 2441 { 2442 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2443 union tcp_log_stackspecific log; 2444 2445 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2446 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; 2447 log.u_bbr.flex2 = time_since_sent; 2448 log.u_bbr.flex3 = srtt; 2449 log.u_bbr.flex4 = thresh; 2450 log.u_bbr.flex5 = to; 2451 log.u_bbr.flex6 = bbr->rc_tp->t_srtt; 2452 log.u_bbr.flex8 = mode; 2453 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2454 &bbr->rc_inp->inp_socket->so_rcv, 2455 &bbr->rc_inp->inp_socket->so_snd, 2456 BBR_LOG_TIMERPREP, 0, 2457 0, &log, false, &bbr->rc_tv); 2458 } 2459 } 2460 2461 static void 2462 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 2463 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) 2464 { 2465 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2466 union tcp_log_stackspecific log; 2467 2468 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2469 log.u_bbr.flex1 = usecs; 2470 log.u_bbr.flex2 = len; 2471 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); 2472 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); 2473 if (override) 2474 log.u_bbr.flex5 = (1 << 2); 2475 else 2476 log.u_bbr.flex5 = 0; 2477 log.u_bbr.flex6 = override; 2478 log.u_bbr.flex7 = gain; 2479 log.u_bbr.flex8 = mod; 2480 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2481 &bbr->rc_inp->inp_socket->so_rcv, 2482 &bbr->rc_inp->inp_socket->so_snd, 2483 BBR_LOG_HPTSI_CALC, 0, 2484 len, &log, false, &bbr->rc_tv); 2485 } 2486 } 2487 2488 static void 2489 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 2490 { 2491 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2492 union tcp_log_stackspecific log; 2493 2494 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2495 2496 log.u_bbr.flex1 = bbr->bbr_timer_src; 2497 log.u_bbr.flex2 = to; 2498 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2499 log.u_bbr.flex4 = slot; 2500 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; 2501 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2502 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; 2503 log.u_bbr.flex8 = which; 2504 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2505 &bbr->rc_inp->inp_socket->so_rcv, 2506 &bbr->rc_inp->inp_socket->so_snd, 2507 BBR_LOG_TIMERSTAR, 0, 2508 0, &log, false, &bbr->rc_tv); 2509 } 2510 } 2511 2512 static void 2513 bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm) 2514 { 2515 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2516 union tcp_log_stackspecific log; 2517 2518 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2519 log.u_bbr.flex1 = thresh; 2520 log.u_bbr.flex2 = lro; 2521 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; 2522 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 2523 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2524 log.u_bbr.flex6 = srtt; 2525 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; 2526 log.u_bbr.flex8 = frm; 2527 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2528 &bbr->rc_inp->inp_socket->so_rcv, 2529 &bbr->rc_inp->inp_socket->so_snd, 2530 BBR_LOG_THRESH_CALC, 0, 2531 0, &log, false, &bbr->rc_tv); 2532 } 2533 } 2534 2535 static void 2536 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) 2537 { 2538 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2539 union tcp_log_stackspecific log; 2540 2541 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2542 log.u_bbr.flex1 = line; 2543 log.u_bbr.flex2 = bbr->bbr_timer_src; 2544 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2545 log.u_bbr.flex4 = bbr->rc_in_persist; 2546 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2547 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2548 log.u_bbr.flex8 = hpts_removed; 2549 log.u_bbr.pkts_out = bbr->rc_pacer_started; 2550 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2551 &bbr->rc_inp->inp_socket->so_rcv, 2552 &bbr->rc_inp->inp_socket->so_snd, 2553 BBR_LOG_TIMERCANC, 0, 2554 0, &log, false, &bbr->rc_tv); 2555 } 2556 } 2557 2558 static void 2559 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) 2560 { 2561 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2562 union tcp_log_stackspecific log; 2563 2564 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2565 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; 2566 log.u_bbr.flex2 = (peer_delta >> 32); 2567 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); 2568 log.u_bbr.flex4 = (delta >> 32); 2569 log.u_bbr.flex5 = (delta & 0x00000000ffffffff); 2570 log.u_bbr.flex7 = bbr->rc_ts_clock_set; 2571 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; 2572 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2573 &bbr->rc_inp->inp_socket->so_rcv, 2574 &bbr->rc_inp->inp_socket->so_snd, 2575 BBR_LOG_TSTMP_VAL, 0, 2576 0, &log, false, &bbr->rc_tv); 2577 } 2578 } 2579 2580 static void 2581 bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr) 2582 { 2583 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2584 union tcp_log_stackspecific log; 2585 2586 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2587 log.u_bbr.flex1 = tsosz; 2588 log.u_bbr.flex2 = tls; 2589 log.u_bbr.flex3 = tcp_min_hptsi_time; 2590 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; 2591 log.u_bbr.flex5 = old_val; 2592 log.u_bbr.flex6 = maxseg; 2593 log.u_bbr.flex7 = bbr->rc_no_pacing; 2594 log.u_bbr.flex7 <<= 1; 2595 log.u_bbr.flex7 |= bbr->rc_past_init_win; 2596 if (hdwr) 2597 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; 2598 else 2599 log.u_bbr.flex8 = bbr->rc_use_google; 2600 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2601 &bbr->rc_inp->inp_socket->so_rcv, 2602 &bbr->rc_inp->inp_socket->so_snd, 2603 BBR_LOG_BBRTSO, 0, 2604 0, &log, false, &bbr->rc_tv); 2605 } 2606 } 2607 2608 static void 2609 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, 2610 uint32_t flags, uint32_t line) 2611 { 2612 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2613 union tcp_log_stackspecific log; 2614 2615 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2616 log.u_bbr.flex1 = line; 2617 log.u_bbr.flex2 = rsm->r_start; 2618 log.u_bbr.flex3 = rsm->r_end; 2619 log.u_bbr.flex4 = rsm->r_delivered; 2620 log.u_bbr.flex5 = rsm->r_rtr_cnt; 2621 log.u_bbr.flex6 = rsm->r_dupack; 2622 log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; 2623 log.u_bbr.flex8 = rsm->r_flags; 2624 /* Hijack the pkts_out fids */ 2625 log.u_bbr.applimited = flags; 2626 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2627 &bbr->rc_inp->inp_socket->so_rcv, 2628 &bbr->rc_inp->inp_socket->so_snd, 2629 BBR_RSM_CLEARED, 0, 2630 0, &log, false, &bbr->rc_tv); 2631 } 2632 } 2633 2634 static void 2635 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, 2636 uint32_t flex3, uint32_t flex2, uint32_t flex5, 2637 uint32_t flex6, uint32_t pkts_out, int flex7, 2638 uint32_t flex4, uint32_t flex1) 2639 { 2640 2641 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2642 union tcp_log_stackspecific log; 2643 2644 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2645 log.u_bbr.flex1 = flex1; 2646 log.u_bbr.flex2 = flex2; 2647 log.u_bbr.flex3 = flex3; 2648 log.u_bbr.flex4 = flex4; 2649 log.u_bbr.flex5 = flex5; 2650 log.u_bbr.flex6 = flex6; 2651 log.u_bbr.flex7 = flex7; 2652 /* Hijack the pkts_out fids */ 2653 log.u_bbr.pkts_out = pkts_out; 2654 log.u_bbr.flex8 = flex8; 2655 if (bbr->rc_ack_was_delayed) 2656 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; 2657 else 2658 log.u_bbr.epoch = 0; 2659 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2660 &bbr->rc_inp->inp_socket->so_rcv, 2661 &bbr->rc_inp->inp_socket->so_snd, 2662 BBR_LOG_BBRUPD, 0, 2663 flex2, &log, false, &bbr->rc_tv); 2664 } 2665 } 2666 2667 static void 2668 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, 2669 uint32_t newbw, uint32_t obw, uint32_t diff, 2670 uint32_t tim) 2671 { 2672 if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2673 union tcp_log_stackspecific log; 2674 2675 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2676 log.u_bbr.flex1 = reason; 2677 log.u_bbr.flex2 = newbw; 2678 log.u_bbr.flex3 = obw; 2679 log.u_bbr.flex4 = diff; 2680 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; 2681 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; 2682 log.u_bbr.flex7 = bbr->rc_lt_is_sampling; 2683 log.u_bbr.pkts_out = tim; 2684 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; 2685 if (bbr->rc_lt_use_bw == 0) 2686 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 2687 else 2688 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 2689 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2690 &bbr->rc_inp->inp_socket->so_rcv, 2691 &bbr->rc_inp->inp_socket->so_snd, 2692 BBR_LOG_BWSAMP, 0, 2693 0, &log, false, &bbr->rc_tv); 2694 } 2695 } 2696 2697 static inline void 2698 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) 2699 { 2700 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2701 union tcp_log_stackspecific log; 2702 2703 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2704 log.u_bbr.flex1 = line; 2705 log.u_bbr.flex2 = tick; 2706 log.u_bbr.flex3 = tp->t_maxunacktime; 2707 log.u_bbr.flex4 = tp->t_acktime; 2708 log.u_bbr.flex8 = event; 2709 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2710 &bbr->rc_inp->inp_socket->so_rcv, 2711 &bbr->rc_inp->inp_socket->so_snd, 2712 BBR_LOG_PROGRESS, 0, 2713 0, &log, false, &bbr->rc_tv); 2714 } 2715 } 2716 2717 static void 2718 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, 2719 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, 2720 int error) 2721 { 2722 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2723 union tcp_log_stackspecific log; 2724 2725 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2726 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); 2727 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); 2728 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); 2729 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); 2730 log.u_bbr.bw_inuse = rate; 2731 log.u_bbr.flex5 = line; 2732 log.u_bbr.flex6 = error; 2733 log.u_bbr.flex8 = bbr->skip_gain; 2734 log.u_bbr.flex8 <<= 1; 2735 log.u_bbr.flex8 |= bbr->gain_is_limited; 2736 log.u_bbr.flex8 <<= 1; 2737 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; 2738 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 2739 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2740 &bbr->rc_inp->inp_socket->so_rcv, 2741 &bbr->rc_inp->inp_socket->so_snd, 2742 BBR_LOG_HDWR_PACE, 0, 2743 0, &log, false, &bbr->rc_tv); 2744 } 2745 } 2746 2747 static void 2748 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay) 2749 { 2750 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2751 union tcp_log_stackspecific log; 2752 2753 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2754 log.u_bbr.flex1 = slot; 2755 log.u_bbr.flex2 = del_by; 2756 log.u_bbr.flex3 = prev_delay; 2757 log.u_bbr.flex4 = line; 2758 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; 2759 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; 2760 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); 2761 log.u_bbr.flex8 = bbr->rc_in_persist; 2762 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2763 &bbr->rc_inp->inp_socket->so_rcv, 2764 &bbr->rc_inp->inp_socket->so_snd, 2765 BBR_LOG_BBRSND, 0, 2766 len, &log, false, &bbr->rc_tv); 2767 } 2768 } 2769 2770 static void 2771 bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags) 2772 { 2773 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2774 union tcp_log_stackspecific log; 2775 2776 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2777 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; 2778 log.u_bbr.flex2 = 0; 2779 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; 2780 log.u_bbr.flex4 = end; 2781 log.u_bbr.flex5 = seq; 2782 log.u_bbr.flex6 = t; 2783 log.u_bbr.flex7 = match; 2784 log.u_bbr.flex8 = flags; 2785 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2786 &bbr->rc_inp->inp_socket->so_rcv, 2787 &bbr->rc_inp->inp_socket->so_snd, 2788 BBR_LOG_BBRRTT, 0, 2789 0, &log, false, &bbr->rc_tv); 2790 } 2791 } 2792 2793 static void 2794 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) 2795 { 2796 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2797 union tcp_log_stackspecific log; 2798 2799 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2800 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2801 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 2802 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; 2803 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2804 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; 2805 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; 2806 log.u_bbr.flex7 = 0; 2807 log.u_bbr.flex8 = entry_method; 2808 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2809 &bbr->rc_inp->inp_socket->so_rcv, 2810 &bbr->rc_inp->inp_socket->so_snd, 2811 BBR_LOG_EXIT_GAIN, 0, 2812 0, &log, false, &bbr->rc_tv); 2813 } 2814 } 2815 2816 static void 2817 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) 2818 { 2819 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2820 union tcp_log_stackspecific log; 2821 2822 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2823 /* R-HU */ 2824 log.u_bbr.flex1 = 0; 2825 log.u_bbr.flex2 = 0; 2826 log.u_bbr.flex3 = 0; 2827 log.u_bbr.flex4 = 0; 2828 log.u_bbr.flex7 = 0; 2829 log.u_bbr.flex8 = settings_desired; 2830 2831 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2832 &bbr->rc_inp->inp_socket->so_rcv, 2833 &bbr->rc_inp->inp_socket->so_snd, 2834 BBR_LOG_SETTINGS_CHG, 0, 2835 0, &log, false, &bbr->rc_tv); 2836 } 2837 } 2838 2839 /* 2840 * Returns the bw from the our filter. 2841 */ 2842 static inline uint64_t 2843 bbr_get_full_bw(struct tcp_bbr *bbr) 2844 { 2845 uint64_t bw; 2846 2847 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2848 2849 return (bw); 2850 } 2851 2852 static inline void 2853 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2854 { 2855 uint64_t calclr; 2856 uint32_t lost, del; 2857 2858 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) 2859 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; 2860 else 2861 lost = 0; 2862 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; 2863 if (lost == 0) { 2864 calclr = 0; 2865 } else if (del) { 2866 calclr = lost; 2867 calclr *= (uint64_t)1000; 2868 calclr /= (uint64_t)del; 2869 } else { 2870 /* Nothing delivered? 100.0% loss */ 2871 calclr = 1000; 2872 } 2873 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; 2874 if (IN_RECOVERY(bbr->rc_tp->t_flags)) 2875 bbr->r_ctl.recovery_lr += (uint32_t)calclr; 2876 bbr->r_ctl.rc_pkt_epoch++; 2877 if (bbr->rc_no_pacing && 2878 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { 2879 bbr->rc_no_pacing = 0; 2880 tcp_bbr_tso_size_check(bbr, cts); 2881 } 2882 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); 2883 bbr->r_ctl.rc_pkt_epoch_time = cts; 2884 /* What was our loss rate */ 2885 bbr_log_pkt_epoch(bbr, cts, line, lost, del); 2886 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; 2887 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; 2888 } 2889 2890 static inline void 2891 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2892 { 2893 uint32_t epoch_time; 2894 2895 /* Tick the RTT clock */ 2896 bbr->r_ctl.rc_rtt_epoch++; 2897 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; 2898 bbr_log_time_epoch(bbr, cts, line, epoch_time); 2899 bbr->r_ctl.rc_rcv_epoch_start = cts; 2900 } 2901 2902 static inline void 2903 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) 2904 { 2905 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { 2906 bbr->rc_is_pkt_epoch_now = 1; 2907 } 2908 } 2909 2910 /* 2911 * Returns the bw from either the b/w filter 2912 * or from the lt_bw (if the connection is being 2913 * policed). 2914 */ 2915 static inline uint64_t 2916 __bbr_get_bw(struct tcp_bbr *bbr) 2917 { 2918 uint64_t bw, min_bw; 2919 uint64_t rtt; 2920 int gm_measure_cnt = 1; 2921 2922 /* 2923 * For startup we make, like google, a 2924 * minimum b/w. This is generated from the 2925 * IW and the rttProp. We do fall back to srtt 2926 * if for some reason (initial handshake) we don't 2927 * have a rttProp. We, in the worst case, fall back 2928 * to the configured min_bw (rc_initial_hptsi_bw). 2929 */ 2930 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 2931 /* Attempt first to use rttProp */ 2932 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2933 if (rtt && (rtt < 0xffffffff)) { 2934 measure: 2935 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2936 ((uint64_t)1000000); 2937 min_bw /= rtt; 2938 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2939 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2940 } 2941 2942 } else if (bbr->rc_tp->t_srtt != 0) { 2943 /* No rttProp, use srtt? */ 2944 rtt = bbr_get_rtt(bbr, BBR_SRTT); 2945 goto measure; 2946 } else { 2947 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2948 } 2949 } else 2950 min_bw = 0; 2951 2952 if ((bbr->rc_past_init_win == 0) && 2953 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) 2954 bbr->rc_past_init_win = 1; 2955 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) 2956 gm_measure_cnt = 0; 2957 if (gm_measure_cnt && 2958 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || 2959 (bbr->rc_past_init_win == 0))) { 2960 /* For google we use our guess rate until we get 1 measurement */ 2961 2962 use_initial_window: 2963 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2964 if (rtt && (rtt < 0xffffffff)) { 2965 /* 2966 * We have an RTT measurement. Use that in 2967 * combination with our initial window to calculate 2968 * a b/w. 2969 */ 2970 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2971 ((uint64_t)1000000); 2972 bw /= rtt; 2973 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2974 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2975 } 2976 } else { 2977 /* Drop back to the 40 and punt to a default */ 2978 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2979 } 2980 if (bw < 1) 2981 /* Probably should panic */ 2982 bw = 1; 2983 if (bw > min_bw) 2984 return (bw); 2985 else 2986 return (min_bw); 2987 } 2988 if (bbr->rc_lt_use_bw) 2989 bw = bbr->r_ctl.rc_lt_bw; 2990 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) 2991 bw = bbr->r_ctl.red_bw; 2992 else 2993 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2994 if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { 2995 /* 2996 * Enforce user set rate limit, keep in mind that 2997 * t_peakrate_thr is in B/s already 2998 */ 2999 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); 3000 } 3001 if (bw == 0) { 3002 /* We should not be at 0, go to the initial window then */ 3003 goto use_initial_window; 3004 } 3005 if (bw < 1) 3006 /* Probably should panic */ 3007 bw = 1; 3008 if (bw < min_bw) 3009 bw = min_bw; 3010 return (bw); 3011 } 3012 3013 static inline uint64_t 3014 bbr_get_bw(struct tcp_bbr *bbr) 3015 { 3016 uint64_t bw; 3017 3018 bw = __bbr_get_bw(bbr); 3019 return (bw); 3020 } 3021 3022 static inline void 3023 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) 3024 { 3025 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; 3026 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; 3027 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3028 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3029 } 3030 3031 static inline void 3032 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) 3033 { 3034 bbr->rc_lt_is_sampling = 0; 3035 bbr->rc_lt_use_bw = 0; 3036 bbr->r_ctl.rc_lt_bw = 0; 3037 bbr_reset_lt_bw_interval(bbr, cts); 3038 } 3039 3040 static inline void 3041 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) 3042 { 3043 uint64_t diff; 3044 3045 /* Do we have a previous sample? */ 3046 if (bbr->r_ctl.rc_lt_bw) { 3047 /* Get the diff in bytes per second */ 3048 if (bbr->r_ctl.rc_lt_bw > bw) 3049 diff = bbr->r_ctl.rc_lt_bw - bw; 3050 else 3051 diff = bw - bbr->r_ctl.rc_lt_bw; 3052 if ((diff <= bbr_lt_bw_diff) || 3053 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { 3054 /* Consider us policed */ 3055 uint32_t saved_bw; 3056 3057 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; 3058 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ 3059 bbr->rc_lt_use_bw = 1; 3060 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 3061 /* 3062 * Use pkt based epoch for measuring length of 3063 * policer up 3064 */ 3065 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; 3066 /* 3067 * reason 4 is we need to start consider being 3068 * policed 3069 */ 3070 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); 3071 return; 3072 } 3073 } 3074 bbr->r_ctl.rc_lt_bw = bw; 3075 bbr_reset_lt_bw_interval(bbr, cts); 3076 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); 3077 } 3078 3079 static void 3080 bbr_randomize_extra_state_time(struct tcp_bbr *bbr) 3081 { 3082 uint32_t ran, deduct; 3083 3084 ran = arc4random_uniform(bbr_rand_ot); 3085 if (ran) { 3086 deduct = bbr->r_ctl.rc_level_state_extra / ran; 3087 bbr->r_ctl.rc_level_state_extra -= deduct; 3088 } 3089 } 3090 /* 3091 * Return randomly the starting state 3092 * to use in probebw. 3093 */ 3094 static uint8_t 3095 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) 3096 { 3097 uint32_t ran; 3098 uint8_t ret_val; 3099 3100 /* Initialize the offset to 0 */ 3101 bbr->r_ctl.rc_exta_time_gd = 0; 3102 bbr->rc_hit_state_1 = 0; 3103 bbr->r_ctl.rc_level_state_extra = 0; 3104 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); 3105 /* 3106 * The math works funny here :) the return value is used to set the 3107 * substate and then the state change is called which increments by 3108 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when 3109 * we fully enter the state. Note that the (8 - 1 - ran) assures that 3110 * we return 1 - 7, so we dont return 0 and end up starting in 3111 * state 1 (DRAIN). 3112 */ 3113 ret_val = BBR_SUBSTATE_COUNT - 1 - ran; 3114 /* Set an epoch */ 3115 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) 3116 bbr_set_epoch(bbr, cts, __LINE__); 3117 3118 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 3119 return (ret_val); 3120 } 3121 3122 static void 3123 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) 3124 { 3125 uint32_t diff, d_time; 3126 uint64_t del_time, bw, lost, delivered; 3127 3128 if (bbr->r_use_policer == 0) 3129 return; 3130 if (bbr->rc_lt_use_bw) { 3131 /* We are using lt bw do we stop yet? */ 3132 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 3133 if (diff > bbr_lt_bw_max_rtts) { 3134 /* Reset it all */ 3135 reset_all: 3136 bbr_reset_lt_bw_sampling(bbr, cts); 3137 if (bbr->rc_filled_pipe) { 3138 bbr_set_epoch(bbr, cts, __LINE__); 3139 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 3140 bbr_substate_change(bbr, cts, __LINE__, 0); 3141 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 3142 bbr_log_type_statechange(bbr, cts, __LINE__); 3143 } else { 3144 /* 3145 * This should not happen really 3146 * unless we remove the startup/drain 3147 * restrictions above. 3148 */ 3149 bbr->rc_bbr_state = BBR_STATE_STARTUP; 3150 bbr_set_epoch(bbr, cts, __LINE__); 3151 bbr->r_ctl.rc_bbr_state_time = cts; 3152 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 3153 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 3154 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 3155 bbr_set_state_target(bbr, __LINE__); 3156 bbr_log_type_statechange(bbr, cts, __LINE__); 3157 } 3158 /* reason 0 is to stop using lt-bw */ 3159 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); 3160 return; 3161 } 3162 if (bbr_lt_intvl_fp == 0) { 3163 /* Not doing false-positive detection */ 3164 return; 3165 } 3166 /* False positive detection */ 3167 if (diff == bbr_lt_intvl_fp) { 3168 /* At bbr_lt_intvl_fp we record the lost */ 3169 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3170 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3171 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { 3172 /* Now is our loss rate still high? */ 3173 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3174 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3175 if ((delivered == 0) || 3176 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { 3177 /* No still below our threshold */ 3178 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); 3179 } else { 3180 /* Yikes its still high, it must be a false positive */ 3181 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); 3182 goto reset_all; 3183 } 3184 } 3185 return; 3186 } 3187 /* 3188 * Wait for the first loss before sampling, to let the policer 3189 * exhaust its tokens and estimate the steady-state rate allowed by 3190 * the policer. Starting samples earlier includes bursts that 3191 * over-estimate the bw. 3192 */ 3193 if (bbr->rc_lt_is_sampling == 0) { 3194 /* reason 1 is to begin doing the sampling */ 3195 if (loss_detected == 0) 3196 return; 3197 bbr_reset_lt_bw_interval(bbr, cts); 3198 bbr->rc_lt_is_sampling = 1; 3199 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); 3200 return; 3201 } 3202 /* Now how long were we delivering long term last> */ 3203 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) 3204 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; 3205 else 3206 d_time = 0; 3207 3208 /* To avoid underestimates, reset sampling if we run out of data. */ 3209 if (bbr->r_ctl.r_app_limited_until) { 3210 /* Can not measure in app-limited state */ 3211 bbr_reset_lt_bw_sampling(bbr, cts); 3212 /* reason 2 is to reset sampling due to app limits */ 3213 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); 3214 return; 3215 } 3216 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 3217 if (diff < bbr_lt_intvl_min_rtts) { 3218 /* 3219 * need more samples (we don't 3220 * start on a round like linux so 3221 * we need 1 more). 3222 */ 3223 /* 6 is not_enough time or no-loss */ 3224 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3225 return; 3226 } 3227 if (diff > (4 * bbr_lt_intvl_min_rtts)) { 3228 /* 3229 * For now if we wait too long, reset all sampling. We need 3230 * to do some research here, its possible that we should 3231 * base this on how much loss as occurred.. something like 3232 * if its under 10% (or some thresh) reset all otherwise 3233 * don't. Thats for phase II I guess. 3234 */ 3235 bbr_reset_lt_bw_sampling(bbr, cts); 3236 /* reason 3 is to reset sampling due too long of sampling */ 3237 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3238 return; 3239 } 3240 /* 3241 * End sampling interval when a packet is lost, so we estimate the 3242 * policer tokens were exhausted. Stopping the sampling before the 3243 * tokens are exhausted under-estimates the policed rate. 3244 */ 3245 if (loss_detected == 0) { 3246 /* 6 is not_enough time or no-loss */ 3247 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3248 return; 3249 } 3250 /* Calculate packets lost and delivered in sampling interval. */ 3251 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3252 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3253 if ((delivered == 0) || 3254 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { 3255 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); 3256 return; 3257 } 3258 if (d_time < 1000) { 3259 /* Not enough time. wait */ 3260 /* 6 is not_enough time or no-loss */ 3261 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3262 return; 3263 } 3264 if (d_time >= (0xffffffff / USECS_IN_MSEC)) { 3265 /* Too long */ 3266 bbr_reset_lt_bw_sampling(bbr, cts); 3267 /* reason 3 is to reset sampling due too long of sampling */ 3268 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3269 return; 3270 } 3271 del_time = d_time; 3272 bw = delivered; 3273 bw *= (uint64_t)USECS_IN_SECOND; 3274 bw /= del_time; 3275 bbr_lt_bw_samp_done(bbr, bw, cts, d_time); 3276 } 3277 3278 /* 3279 * Allocate a sendmap from our zone. 3280 */ 3281 static struct bbr_sendmap * 3282 bbr_alloc(struct tcp_bbr *bbr) 3283 { 3284 struct bbr_sendmap *rsm; 3285 3286 BBR_STAT_INC(bbr_to_alloc); 3287 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); 3288 if (rsm) { 3289 bbr->r_ctl.rc_num_maps_alloced++; 3290 return (rsm); 3291 } 3292 if (bbr->r_ctl.rc_free_cnt) { 3293 BBR_STAT_INC(bbr_to_alloc_emerg); 3294 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 3295 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 3296 bbr->r_ctl.rc_free_cnt--; 3297 return (rsm); 3298 } 3299 BBR_STAT_INC(bbr_to_alloc_failed); 3300 return (NULL); 3301 } 3302 3303 static struct bbr_sendmap * 3304 bbr_alloc_full_limit(struct tcp_bbr *bbr) 3305 { 3306 if ((V_tcp_map_entries_limit > 0) && 3307 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 3308 BBR_STAT_INC(bbr_alloc_limited); 3309 if (!bbr->alloc_limit_reported) { 3310 bbr->alloc_limit_reported = 1; 3311 BBR_STAT_INC(bbr_alloc_limited_conns); 3312 } 3313 return (NULL); 3314 } 3315 return (bbr_alloc(bbr)); 3316 } 3317 3318 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 3319 static struct bbr_sendmap * 3320 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) 3321 { 3322 struct bbr_sendmap *rsm; 3323 3324 if (limit_type) { 3325 /* currently there is only one limit type */ 3326 if (V_tcp_map_split_limit > 0 && 3327 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { 3328 BBR_STAT_INC(bbr_split_limited); 3329 if (!bbr->alloc_limit_reported) { 3330 bbr->alloc_limit_reported = 1; 3331 BBR_STAT_INC(bbr_alloc_limited_conns); 3332 } 3333 return (NULL); 3334 } 3335 } 3336 3337 /* allocate and mark in the limit type, if set */ 3338 rsm = bbr_alloc(bbr); 3339 if (rsm != NULL && limit_type) { 3340 rsm->r_limit_type = limit_type; 3341 bbr->r_ctl.rc_num_split_allocs++; 3342 } 3343 return (rsm); 3344 } 3345 3346 static void 3347 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 3348 { 3349 if (rsm->r_limit_type) { 3350 /* currently there is only one limit type */ 3351 bbr->r_ctl.rc_num_split_allocs--; 3352 } 3353 if (rsm->r_is_smallmap) 3354 bbr->r_ctl.rc_num_small_maps_alloced--; 3355 if (bbr->r_ctl.rc_tlp_send == rsm) 3356 bbr->r_ctl.rc_tlp_send = NULL; 3357 if (bbr->r_ctl.rc_resend == rsm) { 3358 bbr->r_ctl.rc_resend = NULL; 3359 } 3360 if (bbr->r_ctl.rc_next == rsm) 3361 bbr->r_ctl.rc_next = NULL; 3362 if (bbr->r_ctl.rc_sacklast == rsm) 3363 bbr->r_ctl.rc_sacklast = NULL; 3364 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 3365 memset(rsm, 0, sizeof(struct bbr_sendmap)); 3366 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 3367 rsm->r_limit_type = 0; 3368 bbr->r_ctl.rc_free_cnt++; 3369 return; 3370 } 3371 bbr->r_ctl.rc_num_maps_alloced--; 3372 uma_zfree(bbr_zone, rsm); 3373 } 3374 3375 /* 3376 * Returns the BDP. 3377 */ 3378 static uint64_t 3379 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { 3380 /* 3381 * Calculate the bytes in flight needed given the bw (in bytes per 3382 * second) and the specifyed rtt in useconds. We need to put out the 3383 * returned value per RTT to match that rate. Gain will normally 3384 * raise it up from there. 3385 * 3386 * This should not overflow as long as the bandwidth is below 1 3387 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller 3388 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). 3389 */ 3390 uint64_t usec_per_sec; 3391 3392 usec_per_sec = USECS_IN_SECOND; 3393 return ((rtt * bw) / usec_per_sec); 3394 } 3395 3396 /* 3397 * Return the initial cwnd. 3398 */ 3399 static uint32_t 3400 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) 3401 { 3402 uint32_t i_cwnd; 3403 3404 if (bbr->rc_init_win) { 3405 i_cwnd = bbr->rc_init_win * tp->t_maxseg; 3406 } else if (V_tcp_initcwnd_segments) 3407 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 3408 max(2 * tp->t_maxseg, 14600)); 3409 else if (V_tcp_do_rfc3390) 3410 i_cwnd = min(4 * tp->t_maxseg, 3411 max(2 * tp->t_maxseg, 4380)); 3412 else { 3413 /* Per RFC5681 Section 3.1 */ 3414 if (tp->t_maxseg > 2190) 3415 i_cwnd = 2 * tp->t_maxseg; 3416 else if (tp->t_maxseg > 1095) 3417 i_cwnd = 3 * tp->t_maxseg; 3418 else 3419 i_cwnd = 4 * tp->t_maxseg; 3420 } 3421 return (i_cwnd); 3422 } 3423 3424 /* 3425 * Given a specified gain, return the target 3426 * cwnd based on that gain. 3427 */ 3428 static uint32_t 3429 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) 3430 { 3431 uint64_t bdp, rtt; 3432 uint32_t cwnd; 3433 3434 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || 3435 (bbr_get_full_bw(bbr) == 0)) { 3436 /* No measurements yet */ 3437 return (bbr_initial_cwnd(bbr, bbr->rc_tp)); 3438 } 3439 /* 3440 * Get bytes per RTT needed (rttProp is normally in 3441 * bbr_cwndtarget_rtt_touse) 3442 */ 3443 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); 3444 /* Get the bdp from the two values */ 3445 bdp = bbr_get_bw_delay_prod(rtt, bw); 3446 /* Now apply the gain */ 3447 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); 3448 3449 return (cwnd); 3450 } 3451 3452 static uint32_t 3453 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) 3454 { 3455 uint32_t cwnd, mss; 3456 3457 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 3458 /* Get the base cwnd with gain rounded to a mss */ 3459 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); 3460 /* 3461 * Add in N (2 default since we do not have a 3462 * fq layer to trap packets in) quanta's per the I-D 3463 * section 4.2.3.2 quanta adjust. 3464 */ 3465 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); 3466 if (bbr->rc_use_google) { 3467 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3468 (bbr_state_val(bbr) == BBR_SUB_GAIN)) { 3469 /* 3470 * The linux implementation adds 3471 * an extra 2 x mss in gain cycle which 3472 * is documented no-where except in the code. 3473 * so we add more for Neal undocumented feature 3474 */ 3475 cwnd += 2 * mss; 3476 } 3477 if ((cwnd / mss) & 0x1) { 3478 /* Round up for odd num mss */ 3479 cwnd += mss; 3480 } 3481 } 3482 /* Are we below the min cwnd? */ 3483 if (cwnd < get_min_cwnd(bbr)) 3484 return (get_min_cwnd(bbr)); 3485 return (cwnd); 3486 } 3487 3488 static uint16_t 3489 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) 3490 { 3491 if (gain < 1) 3492 gain = 1; 3493 return (gain); 3494 } 3495 3496 static uint32_t 3497 bbr_get_header_oh(struct tcp_bbr *bbr) 3498 { 3499 int seg_oh; 3500 3501 seg_oh = 0; 3502 if (bbr->r_ctl.rc_inc_tcp_oh) { 3503 /* Do we include TCP overhead? */ 3504 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); 3505 } 3506 if (bbr->r_ctl.rc_inc_ip_oh) { 3507 /* Do we include IP overhead? */ 3508 #ifdef INET6 3509 if (bbr->r_is_v6) { 3510 seg_oh += sizeof(struct ip6_hdr); 3511 } else 3512 #endif 3513 { 3514 3515 #ifdef INET 3516 seg_oh += sizeof(struct ip); 3517 #endif 3518 } 3519 } 3520 if (bbr->r_ctl.rc_inc_enet_oh) { 3521 /* Do we include the ethernet overhead? */ 3522 seg_oh += sizeof(struct ether_header); 3523 } 3524 return(seg_oh); 3525 } 3526 3527 static uint32_t 3528 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) 3529 { 3530 uint64_t divor, res, tim; 3531 3532 if (useconds_time == 0) 3533 return (0); 3534 gain = bbr_gain_adjust(bbr, gain); 3535 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; 3536 tim = useconds_time; 3537 res = (tim * bw * gain) / divor; 3538 if (res == 0) 3539 res = 1; 3540 return ((uint32_t)res); 3541 } 3542 3543 /* 3544 * Given a gain and a length return the delay in useconds that 3545 * should be used to evenly space out packets 3546 * on the connection (based on the gain factor). 3547 */ 3548 static uint32_t 3549 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) 3550 { 3551 uint64_t bw, lentim, res; 3552 uint32_t usecs, srtt, over = 0; 3553 uint32_t seg_oh, num_segs, maxseg; 3554 3555 if (len == 0) 3556 return (0); 3557 3558 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 3559 num_segs = (len + maxseg - 1) / maxseg; 3560 if (bbr->rc_use_google == 0) { 3561 seg_oh = bbr_get_header_oh(bbr); 3562 len += (num_segs * seg_oh); 3563 } 3564 gain = bbr_gain_adjust(bbr, gain); 3565 bw = bbr_get_bw(bbr); 3566 if (bbr->rc_use_google) { 3567 uint64_t cbw; 3568 3569 /* 3570 * Reduce the b/w by the google discount 3571 * factor 10 = 1%. 3572 */ 3573 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); 3574 cbw /= (uint64_t)1000; 3575 /* We don't apply a discount if it results in 0 */ 3576 if (cbw > 0) 3577 bw = cbw; 3578 } 3579 lentim = ((uint64_t)len * 3580 (uint64_t)USECS_IN_SECOND * 3581 (uint64_t)BBR_UNIT); 3582 res = lentim / ((uint64_t)gain * bw); 3583 if (res == 0) 3584 res = 1; 3585 usecs = (uint32_t)res; 3586 srtt = bbr_get_rtt(bbr, BBR_SRTT); 3587 if (bbr_hptsi_max_mul && bbr_hptsi_max_div && 3588 (bbr->rc_use_google == 0) && 3589 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { 3590 /* 3591 * We cannot let the delay be more than 1/2 the srtt time. 3592 * Otherwise we cannot pace out or send properly. 3593 */ 3594 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; 3595 BBR_STAT_INC(bbr_hpts_min_time); 3596 } 3597 if (!nolog) 3598 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); 3599 return (usecs); 3600 } 3601 3602 static void 3603 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, 3604 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) 3605 { 3606 uint64_t bw; 3607 uint32_t cwnd, target_cwnd, saved_bytes, maxseg; 3608 int32_t meth; 3609 3610 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3611 3612 #ifdef STATS 3613 if ((tp->t_flags & TF_GPUTINPROG) && 3614 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 3615 /* 3616 * Strech acks and compressed acks will cause this to 3617 * oscillate but we are doing it the same way as the main 3618 * stack so it will be compariable (though possibly not 3619 * ideal). 3620 */ 3621 int32_t cgput; 3622 int64_t gput, time_stamp; 3623 3624 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; 3625 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); 3626 cgput = gput / time_stamp; 3627 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 3628 cgput); 3629 if (tp->t_stats_gput_prev > 0) 3630 stats_voi_update_abs_s32(tp->t_stats, 3631 VOI_TCP_GPUT_ND, 3632 ((gput - tp->t_stats_gput_prev) * 100) / 3633 tp->t_stats_gput_prev); 3634 tp->t_flags &= ~TF_GPUTINPROG; 3635 tp->t_stats_gput_prev = cgput; 3636 } 3637 #endif 3638 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3639 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 3640 /* We don't change anything in probe-rtt */ 3641 return; 3642 } 3643 maxseg = tp->t_maxseg - bbr->rc_last_options; 3644 saved_bytes = bytes_this_ack; 3645 bytes_this_ack += sack_changed; 3646 if (bytes_this_ack > prev_acked) { 3647 bytes_this_ack -= prev_acked; 3648 /* 3649 * A byte ack'd gives us a full mss 3650 * to be like linux i.e. they count packets. 3651 */ 3652 if ((bytes_this_ack < maxseg) && bbr->rc_use_google) 3653 bytes_this_ack = maxseg; 3654 } else { 3655 /* Unlikely */ 3656 bytes_this_ack = 0; 3657 } 3658 cwnd = tp->snd_cwnd; 3659 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3660 if (bw) 3661 target_cwnd = bbr_get_target_cwnd(bbr, 3662 bw, 3663 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); 3664 else 3665 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); 3666 if (IN_RECOVERY(tp->t_flags) && 3667 (bbr->bbr_prev_in_rec == 0)) { 3668 /* 3669 * We are entering recovery and 3670 * thus packet conservation. 3671 */ 3672 bbr->pkt_conservation = 1; 3673 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; 3674 cwnd = ctf_flight_size(tp, 3675 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3676 bytes_this_ack; 3677 } 3678 if (IN_RECOVERY(tp->t_flags)) { 3679 uint32_t flight; 3680 3681 bbr->bbr_prev_in_rec = 1; 3682 if (cwnd > losses) { 3683 cwnd -= losses; 3684 if (cwnd < maxseg) 3685 cwnd = maxseg; 3686 } else 3687 cwnd = maxseg; 3688 flight = ctf_flight_size(tp, 3689 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3690 bbr_log_type_cwndupd(bbr, flight, 0, 3691 losses, 10, 0, 0, line); 3692 if (bbr->pkt_conservation) { 3693 uint32_t time_in; 3694 3695 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) 3696 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; 3697 else 3698 time_in = 0; 3699 3700 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 3701 /* Clear packet conservation after an rttProp */ 3702 bbr->pkt_conservation = 0; 3703 } else { 3704 if ((flight + bytes_this_ack) > cwnd) 3705 cwnd = flight + bytes_this_ack; 3706 if (cwnd < get_min_cwnd(bbr)) 3707 cwnd = get_min_cwnd(bbr); 3708 tp->snd_cwnd = cwnd; 3709 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, 3710 prev_acked, 1, target_cwnd, th->th_ack, line); 3711 return; 3712 } 3713 } 3714 } else 3715 bbr->bbr_prev_in_rec = 0; 3716 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { 3717 bbr->r_ctl.restrict_growth--; 3718 if (bytes_this_ack > maxseg) 3719 bytes_this_ack = maxseg; 3720 } 3721 if (bbr->rc_filled_pipe) { 3722 /* 3723 * Here we have exited startup and filled the pipe. We will 3724 * thus allow the cwnd to shrink to the target. We hit here 3725 * mostly. 3726 */ 3727 uint32_t s_cwnd; 3728 3729 meth = 2; 3730 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); 3731 if (s_cwnd > cwnd) 3732 cwnd = s_cwnd; 3733 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) 3734 cwnd = s_cwnd; 3735 } else { 3736 /* 3737 * Here we are still in startup, we increase cwnd by what 3738 * has been acked. 3739 */ 3740 if ((cwnd < target_cwnd) || 3741 (bbr->rc_past_init_win == 0)) { 3742 meth = 3; 3743 cwnd += bytes_this_ack; 3744 } else { 3745 /* 3746 * Method 4 means we are at target so no gain in 3747 * startup and past the initial window. 3748 */ 3749 meth = 4; 3750 } 3751 } 3752 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); 3753 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); 3754 } 3755 3756 static void 3757 tcp_bbr_partialack(struct tcpcb *tp) 3758 { 3759 struct tcp_bbr *bbr; 3760 3761 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3762 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3763 if (ctf_flight_size(tp, 3764 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 3765 tp->snd_cwnd) { 3766 bbr->r_wanted_output = 1; 3767 } 3768 } 3769 3770 static void 3771 bbr_post_recovery(struct tcpcb *tp) 3772 { 3773 struct tcp_bbr *bbr; 3774 uint32_t flight; 3775 3776 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3777 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3778 /* 3779 * Here we just exit recovery. 3780 */ 3781 EXIT_RECOVERY(tp->t_flags); 3782 /* Lock in our b/w reduction for the specified number of pkt-epochs */ 3783 bbr->r_recovery_bw = 0; 3784 tp->snd_recover = tp->snd_una; 3785 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3786 bbr->pkt_conservation = 0; 3787 if (bbr->rc_use_google == 0) { 3788 /* 3789 * For non-google mode lets 3790 * go ahead and make sure we clear 3791 * the recovery state so if we 3792 * bounce back in to recovery we 3793 * will do PC. 3794 */ 3795 bbr->bbr_prev_in_rec = 0; 3796 } 3797 bbr_log_type_exit_rec(bbr); 3798 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3799 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3800 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); 3801 } else { 3802 /* For probe-rtt case lets fix up its saved_cwnd */ 3803 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { 3804 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; 3805 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); 3806 } 3807 } 3808 flight = ctf_flight_size(tp, 3809 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3810 if ((bbr->rc_use_google == 0) && 3811 bbr_do_red) { 3812 uint64_t val, lr2use; 3813 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; 3814 uint32_t *cwnd_p; 3815 3816 if (bbr_get_rtt(bbr, BBR_SRTT)) { 3817 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); 3818 val /= bbr_get_rtt(bbr, BBR_SRTT); 3819 ratio = (uint32_t)val; 3820 } else 3821 ratio = 1000; 3822 3823 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, 3824 bbr->r_ctl.recovery_lr, 21, 3825 ratio, 3826 bbr->r_ctl.rc_red_cwnd_pe, 3827 __LINE__); 3828 if ((ratio < bbr_do_red) || (bbr_do_red == 0)) 3829 goto done; 3830 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3831 bbr_prtt_slam_cwnd) || 3832 (bbr_sub_drain_slam_cwnd && 3833 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3834 bbr->rc_hit_state_1 && 3835 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || 3836 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && 3837 bbr_slam_cwnd_in_main_drain)) { 3838 /* 3839 * Here we must poke at the saved cwnd 3840 * as well as the cwnd. 3841 */ 3842 cwnd = bbr->r_ctl.rc_saved_cwnd; 3843 cwnd_p = &bbr->r_ctl.rc_saved_cwnd; 3844 } else { 3845 cwnd = tp->snd_cwnd; 3846 cwnd_p = &tp->snd_cwnd; 3847 } 3848 maxseg = tp->t_maxseg - bbr->rc_last_options; 3849 /* Add the overall lr with the recovery lr */ 3850 if (bbr->r_ctl.rc_lost == 0) 3851 lr2use = 0; 3852 else if (bbr->r_ctl.rc_delivered == 0) 3853 lr2use = 1000; 3854 else { 3855 lr2use = bbr->r_ctl.rc_lost * 1000; 3856 lr2use /= bbr->r_ctl.rc_delivered; 3857 } 3858 lr2use += bbr->r_ctl.recovery_lr; 3859 acks_inflight = (flight / (maxseg * 2)); 3860 if (bbr_red_scale) { 3861 lr2use *= bbr_get_rtt(bbr, BBR_SRTT); 3862 lr2use /= bbr_red_scale; 3863 if ((bbr_red_growth_restrict) && 3864 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) 3865 bbr->r_ctl.restrict_growth += acks_inflight; 3866 } 3867 if (lr2use) { 3868 val = (uint64_t)cwnd * lr2use; 3869 val /= 1000; 3870 if (cwnd > val) 3871 newcwnd = roundup((cwnd - val), maxseg); 3872 else 3873 newcwnd = maxseg; 3874 } else { 3875 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; 3876 val /= (uint64_t)bbr_red_div; 3877 newcwnd = roundup((uint32_t)val, maxseg); 3878 } 3879 /* with standard delayed acks how many acks can I expect? */ 3880 if (bbr_drop_limit == 0) { 3881 /* 3882 * Anticpate how much we will 3883 * raise the cwnd based on the acks. 3884 */ 3885 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { 3886 /* We do enforce the min (with the acks) */ 3887 newcwnd = (get_min_cwnd(bbr) - acks_inflight); 3888 } 3889 } else { 3890 /* 3891 * A strict drop limit of N is inplace 3892 */ 3893 if (newcwnd < (bbr_drop_limit * maxseg)) { 3894 newcwnd = bbr_drop_limit * maxseg; 3895 } 3896 } 3897 /* For the next N acks do we restrict the growth */ 3898 *cwnd_p = newcwnd; 3899 if (tp->snd_cwnd > newcwnd) 3900 tp->snd_cwnd = newcwnd; 3901 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, 3902 (uint32_t)lr2use, 3903 bbr_get_rtt(bbr, BBR_SRTT), __LINE__); 3904 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; 3905 } 3906 done: 3907 bbr->r_ctl.recovery_lr = 0; 3908 if (flight <= tp->snd_cwnd) { 3909 bbr->r_wanted_output = 1; 3910 } 3911 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3912 } 3913 3914 static void 3915 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) 3916 { 3917 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3918 /* Limit the drop in b/w to 1/2 our current filter. */ 3919 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) 3920 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; 3921 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) 3922 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; 3923 tcp_bbr_tso_size_check(bbr, cts); 3924 } 3925 3926 static void 3927 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) 3928 { 3929 struct tcp_bbr *bbr; 3930 3931 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3932 #ifdef STATS 3933 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); 3934 #endif 3935 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3936 switch (type) { 3937 case CC_NDUPACK: 3938 if (!IN_RECOVERY(tp->t_flags)) { 3939 tp->snd_recover = tp->snd_max; 3940 /* Start a new epoch */ 3941 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 3942 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { 3943 /* 3944 * Move forward the lt epoch 3945 * so it won't count the truncated 3946 * epoch. 3947 */ 3948 bbr->r_ctl.rc_lt_epoch++; 3949 } 3950 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 3951 /* 3952 * Just like the policer detection code 3953 * if we are in startup we must push 3954 * forward the last startup epoch 3955 * to hide the truncated PE. 3956 */ 3957 bbr->r_ctl.rc_bbr_last_startup_epoch++; 3958 } 3959 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; 3960 ENTER_RECOVERY(tp->t_flags); 3961 bbr->rc_tlp_rtx_out = 0; 3962 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; 3963 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3964 if (tcp_in_hpts(bbr->rc_inp) && 3965 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { 3966 /* 3967 * When we enter recovery, we need to restart 3968 * any timers. This may mean we gain an agg 3969 * early, which will be made up for at the last 3970 * rxt out. 3971 */ 3972 bbr->rc_timer_first = 1; 3973 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 3974 } 3975 /* 3976 * Calculate a new cwnd based on to the current 3977 * delivery rate with no gain. We get the bdp 3978 * without gaining it up like we normally would and 3979 * we use the last cur_del_rate. 3980 */ 3981 if ((bbr->rc_use_google == 0) && 3982 (bbr->r_ctl.bbr_rttprobe_gain_val || 3983 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { 3984 tp->snd_cwnd = ctf_flight_size(tp, 3985 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3986 (tp->t_maxseg - bbr->rc_last_options); 3987 if (tp->snd_cwnd < get_min_cwnd(bbr)) { 3988 /* We always gate to min cwnd */ 3989 tp->snd_cwnd = get_min_cwnd(bbr); 3990 } 3991 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); 3992 } 3993 bbr_log_type_enter_rec(bbr, rsm->r_start); 3994 } 3995 break; 3996 case CC_RTO_ERR: 3997 KMOD_TCPSTAT_INC(tcps_sndrexmitbad); 3998 /* RTO was unnecessary, so reset everything. */ 3999 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); 4000 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 4001 tp->snd_cwnd = tp->snd_cwnd_prev; 4002 tp->snd_ssthresh = tp->snd_ssthresh_prev; 4003 tp->snd_recover = tp->snd_recover_prev; 4004 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 4005 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); 4006 } 4007 tp->t_badrxtwin = 0; 4008 break; 4009 } 4010 } 4011 4012 /* 4013 * Indicate whether this ack should be delayed. We can delay the ack if 4014 * following conditions are met: 4015 * - There is no delayed ack timer in progress. 4016 * - Our last ack wasn't a 0-sized window. We never want to delay 4017 * the ack that opens up a 0-sized window. 4018 * - LRO wasn't used for this segment. We make sure by checking that the 4019 * segment size is not larger than the MSS. 4020 * - Delayed acks are enabled or this is a half-synchronized T/TCP 4021 * connection. 4022 * - The data being acked is less than a full segment (a stretch ack 4023 * of more than a segment we should ack. 4024 * - nsegs is 1 (if its more than that we received more than 1 ack). 4025 */ 4026 #define DELAY_ACK(tp, bbr, nsegs) \ 4027 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 4028 ((tp->t_flags & TF_DELACK) == 0) && \ 4029 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ 4030 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 4031 4032 /* 4033 * Return the lowest RSM in the map of 4034 * packets still in flight that is not acked. 4035 * This should normally find on the first one 4036 * since we remove packets from the send 4037 * map after they are marked ACKED. 4038 */ 4039 static struct bbr_sendmap * 4040 bbr_find_lowest_rsm(struct tcp_bbr *bbr) 4041 { 4042 struct bbr_sendmap *rsm; 4043 4044 /* 4045 * Walk the time-order transmitted list looking for an rsm that is 4046 * not acked. This will be the one that was sent the longest time 4047 * ago that is still outstanding. 4048 */ 4049 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { 4050 if (rsm->r_flags & BBR_ACKED) { 4051 continue; 4052 } 4053 goto finish; 4054 } 4055 finish: 4056 return (rsm); 4057 } 4058 4059 static struct bbr_sendmap * 4060 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 4061 { 4062 struct bbr_sendmap *prsm; 4063 4064 /* 4065 * Walk the sequence order list backward until we hit and arrive at 4066 * the highest seq not acked. In theory when this is called it 4067 * should be the last segment (which it was not). 4068 */ 4069 prsm = rsm; 4070 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4071 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { 4072 continue; 4073 } 4074 return (prsm); 4075 } 4076 return (NULL); 4077 } 4078 4079 /* 4080 * Returns to the caller the number of microseconds that 4081 * the packet can be outstanding before we think we 4082 * should have had an ack returned. 4083 */ 4084 static uint32_t 4085 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) 4086 { 4087 /* 4088 * lro is the flag we use to determine if we have seen reordering. 4089 * If it gets set we have seen reordering. The reorder logic either 4090 * works in one of two ways: 4091 * 4092 * If reorder-fade is configured, then we track the last time we saw 4093 * re-ordering occur. If we reach the point where enough time as 4094 * passed we no longer consider reordering has occuring. 4095 * 4096 * Or if reorder-face is 0, then once we see reordering we consider 4097 * the connection to alway be subject to reordering and just set lro 4098 * to 1. 4099 * 4100 * In the end if lro is non-zero we add the extra time for 4101 * reordering in. 4102 */ 4103 int32_t lro; 4104 uint32_t thresh, t_rxtcur; 4105 4106 if (srtt == 0) 4107 srtt = 1; 4108 if (bbr->r_ctl.rc_reorder_ts) { 4109 if (bbr->r_ctl.rc_reorder_fade) { 4110 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { 4111 lro = cts - bbr->r_ctl.rc_reorder_ts; 4112 if (lro == 0) { 4113 /* 4114 * No time as passed since the last 4115 * reorder, mark it as reordering. 4116 */ 4117 lro = 1; 4118 } 4119 } else { 4120 /* Negative time? */ 4121 lro = 0; 4122 } 4123 if (lro > bbr->r_ctl.rc_reorder_fade) { 4124 /* Turn off reordering seen too */ 4125 bbr->r_ctl.rc_reorder_ts = 0; 4126 lro = 0; 4127 } 4128 } else { 4129 /* Reodering does not fade */ 4130 lro = 1; 4131 } 4132 } else { 4133 lro = 0; 4134 } 4135 thresh = srtt + bbr->r_ctl.rc_pkt_delay; 4136 if (lro) { 4137 /* It must be set, if not you get 1/4 rtt */ 4138 if (bbr->r_ctl.rc_reorder_shift) 4139 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); 4140 else 4141 thresh += (srtt >> 2); 4142 } else { 4143 thresh += 1000; 4144 } 4145 /* We don't let the rack timeout be above a RTO */ 4146 if ((bbr->rc_tp)->t_srtt == 0) 4147 t_rxtcur = BBR_INITIAL_RTO; 4148 else 4149 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 4150 if (thresh > t_rxtcur) { 4151 thresh = t_rxtcur; 4152 } 4153 /* And we don't want it above the RTO max either */ 4154 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4155 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4156 } 4157 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); 4158 return (thresh); 4159 } 4160 4161 /* 4162 * Return to the caller the amount of time in mico-seconds 4163 * that should be used for the TLP timer from the last 4164 * send time of this packet. 4165 */ 4166 static uint32_t 4167 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 4168 struct bbr_sendmap *rsm, uint32_t srtt, 4169 uint32_t cts) 4170 { 4171 uint32_t thresh, len, maxseg, t_rxtcur; 4172 struct bbr_sendmap *prsm; 4173 4174 if (srtt == 0) 4175 srtt = 1; 4176 if (bbr->rc_tlp_threshold) 4177 thresh = srtt + (srtt / bbr->rc_tlp_threshold); 4178 else 4179 thresh = (srtt * 2); 4180 maxseg = tp->t_maxseg - bbr->rc_last_options; 4181 /* Get the previous sent packet, if any */ 4182 len = rsm->r_end - rsm->r_start; 4183 4184 /* 2.1 behavior */ 4185 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); 4186 if (prsm && (len <= maxseg)) { 4187 /* 4188 * Two packets outstanding, thresh should be (2*srtt) + 4189 * possible inter-packet delay (if any). 4190 */ 4191 uint32_t inter_gap = 0; 4192 int idx, nidx; 4193 4194 idx = rsm->r_rtr_cnt - 1; 4195 nidx = prsm->r_rtr_cnt - 1; 4196 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 4197 /* Yes it was sent later (or at the same time) */ 4198 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 4199 } 4200 thresh += inter_gap; 4201 } else if (len <= maxseg) { 4202 /* 4203 * Possibly compensate for delayed-ack. 4204 */ 4205 uint32_t alt_thresh; 4206 4207 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; 4208 if (alt_thresh > thresh) 4209 thresh = alt_thresh; 4210 } 4211 /* Not above the current RTO */ 4212 if (tp->t_srtt == 0) 4213 t_rxtcur = BBR_INITIAL_RTO; 4214 else 4215 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); 4216 4217 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); 4218 /* Not above an RTO */ 4219 if (thresh > t_rxtcur) { 4220 thresh = t_rxtcur; 4221 } 4222 /* Not above a RTO max */ 4223 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4224 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4225 } 4226 /* And now apply the user TLP min */ 4227 if (thresh < bbr_tlp_min) { 4228 thresh = bbr_tlp_min; 4229 } 4230 return (thresh); 4231 } 4232 4233 /* 4234 * Return one of three RTTs to use (in microseconds). 4235 */ 4236 static __inline uint32_t 4237 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) 4238 { 4239 uint32_t f_rtt; 4240 uint32_t srtt; 4241 4242 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 4243 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { 4244 /* We have no rtt at all */ 4245 if (bbr->rc_tp->t_srtt == 0) 4246 f_rtt = BBR_INITIAL_RTO; 4247 else 4248 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4249 /* 4250 * Since we don't know how good the rtt is apply a 4251 * delayed-ack min 4252 */ 4253 if (f_rtt < bbr_delayed_ack_time) { 4254 f_rtt = bbr_delayed_ack_time; 4255 } 4256 } 4257 /* Take the filter version or last measured pkt-rtt */ 4258 if (rtt_type == BBR_RTT_PROP) { 4259 srtt = f_rtt; 4260 } else if (rtt_type == BBR_RTT_PKTRTT) { 4261 if (bbr->r_ctl.rc_pkt_epoch_rtt) { 4262 srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 4263 } else { 4264 /* No pkt rtt yet */ 4265 srtt = f_rtt; 4266 } 4267 } else if (rtt_type == BBR_RTT_RACK) { 4268 srtt = bbr->r_ctl.rc_last_rtt; 4269 /* We need to add in any internal delay for our timer */ 4270 if (bbr->rc_ack_was_delayed) 4271 srtt += bbr->r_ctl.rc_ack_hdwr_delay; 4272 } else if (rtt_type == BBR_SRTT) { 4273 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4274 } else { 4275 /* TSNH */ 4276 srtt = f_rtt; 4277 #ifdef BBR_INVARIANTS 4278 panic("Unknown rtt request type %d", rtt_type); 4279 #endif 4280 } 4281 return (srtt); 4282 } 4283 4284 static int 4285 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) 4286 { 4287 uint32_t thresh; 4288 4289 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), 4290 cts, rsm); 4291 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { 4292 /* It is lost (past time) */ 4293 return (1); 4294 } 4295 return (0); 4296 } 4297 4298 /* 4299 * Return a sendmap if we need to retransmit something. 4300 */ 4301 static struct bbr_sendmap * 4302 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4303 { 4304 /* 4305 * Check to see that we don't need to fall into recovery. We will 4306 * need to do so if our oldest transmit is past the time we should 4307 * have had an ack. 4308 */ 4309 4310 struct bbr_sendmap *rsm; 4311 int32_t idx; 4312 4313 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { 4314 /* Nothing outstanding that we know of */ 4315 return (NULL); 4316 } 4317 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 4318 if (rsm == NULL) { 4319 /* Nothing in the transmit map */ 4320 return (NULL); 4321 } 4322 if (tp->t_flags & TF_SENTFIN) { 4323 /* Fin restricted, don't find anything once a fin is sent */ 4324 return (NULL); 4325 } 4326 if (rsm->r_flags & BBR_ACKED) { 4327 /* 4328 * Ok the first one is acked (this really should not happen 4329 * since we remove the from the tmap once they are acked) 4330 */ 4331 rsm = bbr_find_lowest_rsm(bbr); 4332 if (rsm == NULL) 4333 return (NULL); 4334 } 4335 idx = rsm->r_rtr_cnt - 1; 4336 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { 4337 /* Send timestamp is the same or less? can't be ready */ 4338 return (NULL); 4339 } 4340 /* Get our RTT time */ 4341 if (bbr_is_lost(bbr, rsm, cts) && 4342 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || 4343 (rsm->r_flags & BBR_SACK_PASSED))) { 4344 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4345 rsm->r_flags |= BBR_MARKED_LOST; 4346 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4347 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4348 } 4349 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); 4350 #ifdef BBR_INVARIANTS 4351 if ((rsm->r_end - rsm->r_start) == 0) 4352 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); 4353 #endif 4354 return (rsm); 4355 } 4356 return (NULL); 4357 } 4358 4359 /* 4360 * RACK Timer, here we simply do logging and house keeping. 4361 * the normal bbr_output_wtime() function will call the 4362 * appropriate thing to check if we need to do a RACK retransmit. 4363 * We return 1, saying don't proceed with bbr_output_wtime only 4364 * when all timers have been stopped (destroyed PCB?). 4365 */ 4366 static int 4367 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4368 { 4369 /* 4370 * This timer simply provides an internal trigger to send out data. 4371 * The check_recovery_mode call will see if there are needed 4372 * retransmissions, if so we will enter fast-recovery. The output 4373 * call may or may not do the same thing depending on sysctl 4374 * settings. 4375 */ 4376 uint32_t lost; 4377 4378 if (bbr->rc_all_timers_stopped) { 4379 return (1); 4380 } 4381 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4382 /* Its not time yet */ 4383 return (0); 4384 } 4385 BBR_STAT_INC(bbr_to_tot); 4386 lost = bbr->r_ctl.rc_lost; 4387 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4388 bbr_set_state(tp, bbr, 0); 4389 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); 4390 if (bbr->r_ctl.rc_resend == NULL) { 4391 /* Lets do the check here */ 4392 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 4393 } 4394 if (bbr_policer_call_from_rack_to) 4395 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4396 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 4397 return (0); 4398 } 4399 4400 static __inline void 4401 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) 4402 { 4403 int idx; 4404 4405 nrsm->r_start = start; 4406 nrsm->r_end = rsm->r_end; 4407 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4408 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; 4409 nrsm->r_flags = rsm->r_flags; 4410 /* We don't transfer forward the SYN flag */ 4411 nrsm->r_flags &= ~BBR_HAS_SYN; 4412 /* We move forward the FIN flag, not that this should happen */ 4413 rsm->r_flags &= ~BBR_HAS_FIN; 4414 nrsm->r_dupack = rsm->r_dupack; 4415 nrsm->r_rtr_bytes = 0; 4416 nrsm->r_is_gain = rsm->r_is_gain; 4417 nrsm->r_is_drain = rsm->r_is_drain; 4418 nrsm->r_delivered = rsm->r_delivered; 4419 nrsm->r_ts_valid = rsm->r_ts_valid; 4420 nrsm->r_del_ack_ts = rsm->r_del_ack_ts; 4421 nrsm->r_del_time = rsm->r_del_time; 4422 nrsm->r_app_limited = rsm->r_app_limited; 4423 nrsm->r_first_sent_time = rsm->r_first_sent_time; 4424 nrsm->r_flight_at_send = rsm->r_flight_at_send; 4425 /* We split a piece the lower section looses any just_ret flag. */ 4426 nrsm->r_bbr_state = rsm->r_bbr_state; 4427 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4428 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4429 } 4430 rsm->r_end = nrsm->r_start; 4431 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 4432 idx /= 8; 4433 /* Check if we got too small */ 4434 if ((rsm->r_is_smallmap == 0) && 4435 ((rsm->r_end - rsm->r_start) <= idx)) { 4436 bbr->r_ctl.rc_num_small_maps_alloced++; 4437 rsm->r_is_smallmap = 1; 4438 } 4439 /* Check the new one as well */ 4440 if ((nrsm->r_end - nrsm->r_start) <= idx) { 4441 bbr->r_ctl.rc_num_small_maps_alloced++; 4442 nrsm->r_is_smallmap = 1; 4443 } 4444 } 4445 4446 static int 4447 bbr_sack_mergable(struct bbr_sendmap *at, 4448 uint32_t start, uint32_t end) 4449 { 4450 /* 4451 * Given a sack block defined by 4452 * start and end, and a current position 4453 * at. Return 1 if either side of at 4454 * would show that the block is mergable 4455 * to that side. A block to be mergable 4456 * must have overlap with the start/end 4457 * and be in the SACK'd state. 4458 */ 4459 struct bbr_sendmap *l_rsm; 4460 struct bbr_sendmap *r_rsm; 4461 4462 /* first get the either side blocks */ 4463 l_rsm = TAILQ_PREV(at, bbr_head, r_next); 4464 r_rsm = TAILQ_NEXT(at, r_next); 4465 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { 4466 /* Potentially mergeable */ 4467 if ((l_rsm->r_end == start) || 4468 (SEQ_LT(start, l_rsm->r_end) && 4469 SEQ_GT(end, l_rsm->r_end))) { 4470 /* 4471 * map blk |------| 4472 * sack blk |------| 4473 * <or> 4474 * map blk |------| 4475 * sack blk |------| 4476 */ 4477 return (1); 4478 } 4479 } 4480 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { 4481 /* Potentially mergeable */ 4482 if ((r_rsm->r_start == end) || 4483 (SEQ_LT(start, r_rsm->r_start) && 4484 SEQ_GT(end, r_rsm->r_start))) { 4485 /* 4486 * map blk |---------| 4487 * sack blk |----| 4488 * <or> 4489 * map blk |---------| 4490 * sack blk |-------| 4491 */ 4492 return (1); 4493 } 4494 } 4495 return (0); 4496 } 4497 4498 static struct bbr_sendmap * 4499 bbr_merge_rsm(struct tcp_bbr *bbr, 4500 struct bbr_sendmap *l_rsm, 4501 struct bbr_sendmap *r_rsm) 4502 { 4503 /* 4504 * We are merging two ack'd RSM's, 4505 * the l_rsm is on the left (lower seq 4506 * values) and the r_rsm is on the right 4507 * (higher seq value). The simplest way 4508 * to merge these is to move the right 4509 * one into the left. I don't think there 4510 * is any reason we need to try to find 4511 * the oldest (or last oldest retransmitted). 4512 */ 4513 l_rsm->r_end = r_rsm->r_end; 4514 if (l_rsm->r_dupack < r_rsm->r_dupack) 4515 l_rsm->r_dupack = r_rsm->r_dupack; 4516 if (r_rsm->r_rtr_bytes) 4517 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 4518 if (r_rsm->r_in_tmap) { 4519 /* This really should not happen */ 4520 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); 4521 } 4522 if (r_rsm->r_app_limited) 4523 l_rsm->r_app_limited = r_rsm->r_app_limited; 4524 /* Now the flags */ 4525 if (r_rsm->r_flags & BBR_HAS_FIN) 4526 l_rsm->r_flags |= BBR_HAS_FIN; 4527 if (r_rsm->r_flags & BBR_TLP) 4528 l_rsm->r_flags |= BBR_TLP; 4529 if (r_rsm->r_flags & BBR_RWND_COLLAPSED) 4530 l_rsm->r_flags |= BBR_RWND_COLLAPSED; 4531 if (r_rsm->r_flags & BBR_MARKED_LOST) { 4532 /* This really should not happen */ 4533 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; 4534 } 4535 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); 4536 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 4537 /* Transfer the split limit to the map we free */ 4538 r_rsm->r_limit_type = l_rsm->r_limit_type; 4539 l_rsm->r_limit_type = 0; 4540 } 4541 bbr_free(bbr, r_rsm); 4542 return(l_rsm); 4543 } 4544 4545 /* 4546 * TLP Timer, here we simply setup what segment we want to 4547 * have the TLP expire on, the normal bbr_output_wtime() will then 4548 * send it out. 4549 * 4550 * We return 1, saying don't proceed with bbr_output_wtime only 4551 * when all timers have been stopped (destroyed PCB?). 4552 */ 4553 static int 4554 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4555 { 4556 /* 4557 * Tail Loss Probe. 4558 */ 4559 struct bbr_sendmap *rsm = NULL; 4560 struct socket *so; 4561 uint32_t amm; 4562 uint32_t out, avail; 4563 uint32_t maxseg; 4564 int collapsed_win = 0; 4565 4566 if (bbr->rc_all_timers_stopped) { 4567 return (1); 4568 } 4569 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4570 /* Its not time yet */ 4571 return (0); 4572 } 4573 if (ctf_progress_timeout_check(tp, true)) { 4574 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4575 return (-ETIMEDOUT); /* tcp_drop() */ 4576 } 4577 /* Did we somehow get into persists? */ 4578 if (bbr->rc_in_persist) { 4579 return (0); 4580 } 4581 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4582 bbr_set_state(tp, bbr, 0); 4583 BBR_STAT_INC(bbr_tlp_tot); 4584 maxseg = tp->t_maxseg - bbr->rc_last_options; 4585 /* 4586 * A TLP timer has expired. We have been idle for 2 rtts. So we now 4587 * need to figure out how to force a full MSS segment out. 4588 */ 4589 so = tptosocket(tp); 4590 avail = sbavail(&so->so_snd); 4591 out = ctf_outstanding(tp); 4592 if (out > tp->snd_wnd) { 4593 /* special case, we need a retransmission */ 4594 collapsed_win = 1; 4595 goto need_retran; 4596 } 4597 if (avail > out) { 4598 /* New data is available */ 4599 amm = avail - out; 4600 if (amm > maxseg) { 4601 amm = maxseg; 4602 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 4603 /* not enough to fill a MTU and no-delay is off */ 4604 goto need_retran; 4605 } 4606 /* Set the send-new override */ 4607 if ((out + amm) <= tp->snd_wnd) { 4608 bbr->rc_tlp_new_data = 1; 4609 } else { 4610 goto need_retran; 4611 } 4612 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4613 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; 4614 bbr->r_ctl.rc_tlp_send = NULL; 4615 /* cap any slots */ 4616 BBR_STAT_INC(bbr_tlp_newdata); 4617 goto send; 4618 } 4619 need_retran: 4620 /* 4621 * Ok we need to arrange the last un-acked segment to be re-sent, or 4622 * optionally the first un-acked segment. 4623 */ 4624 if (collapsed_win == 0) { 4625 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 4626 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { 4627 rsm = bbr_find_high_nonack(bbr, rsm); 4628 } 4629 if (rsm == NULL) { 4630 goto restore; 4631 } 4632 } else { 4633 /* 4634 * We must find the last segment 4635 * that was acceptable by the client. 4636 */ 4637 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4638 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { 4639 /* Found one */ 4640 break; 4641 } 4642 } 4643 if (rsm == NULL) { 4644 /* None? if so send the first */ 4645 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4646 if (rsm == NULL) 4647 goto restore; 4648 } 4649 } 4650 if ((rsm->r_end - rsm->r_start) > maxseg) { 4651 /* 4652 * We need to split this the last segment in two. 4653 */ 4654 struct bbr_sendmap *nrsm; 4655 4656 nrsm = bbr_alloc_full_limit(bbr); 4657 if (nrsm == NULL) { 4658 /* 4659 * We can't get memory to split, we can either just 4660 * not split it. Or retransmit the whole piece, lets 4661 * do the large send (BTLP :-) ). 4662 */ 4663 goto go_for_it; 4664 } 4665 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); 4666 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 4667 if (rsm->r_in_tmap) { 4668 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4669 nrsm->r_in_tmap = 1; 4670 } 4671 rsm->r_flags &= (~BBR_HAS_FIN); 4672 rsm = nrsm; 4673 } 4674 go_for_it: 4675 bbr->r_ctl.rc_tlp_send = rsm; 4676 bbr->rc_tlp_rtx_out = 1; 4677 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { 4678 bbr->r_ctl.rc_tlp_seg_send_cnt++; 4679 tp->t_rxtshift++; 4680 } else { 4681 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 4682 bbr->r_ctl.rc_tlp_seg_send_cnt = 1; 4683 } 4684 send: 4685 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 4686 /* 4687 * Can't [re]/transmit a segment we have retransmitted the 4688 * max times. We need the retransmit timer to take over. 4689 */ 4690 restore: 4691 bbr->rc_tlp_new_data = 0; 4692 bbr->r_ctl.rc_tlp_send = NULL; 4693 if (rsm) 4694 rsm->r_flags &= ~BBR_TLP; 4695 BBR_STAT_INC(bbr_tlp_retran_fail); 4696 return (0); 4697 } else if (rsm) { 4698 rsm->r_flags |= BBR_TLP; 4699 } 4700 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && 4701 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { 4702 /* 4703 * We have retransmitted to many times for TLP. Switch to 4704 * the regular RTO timer 4705 */ 4706 goto restore; 4707 } 4708 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); 4709 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 4710 return (0); 4711 } 4712 4713 /* 4714 * Delayed ack Timer, here we simply need to setup the 4715 * ACK_NOW flag and remove the DELACK flag. From there 4716 * the output routine will send the ack out. 4717 * 4718 * We only return 1, saying don't proceed, if all timers 4719 * are stopped (destroyed PCB?). 4720 */ 4721 static int 4722 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4723 { 4724 if (bbr->rc_all_timers_stopped) { 4725 return (1); 4726 } 4727 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); 4728 tp->t_flags &= ~TF_DELACK; 4729 tp->t_flags |= TF_ACKNOW; 4730 KMOD_TCPSTAT_INC(tcps_delack); 4731 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 4732 return (0); 4733 } 4734 4735 /* 4736 * Here we send a KEEP-ALIVE like probe to the 4737 * peer, we do not send data. 4738 * 4739 * We only return 1, saying don't proceed, if all timers 4740 * are stopped (destroyed PCB?). 4741 */ 4742 static int 4743 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4744 { 4745 struct tcptemp *t_template; 4746 int32_t retval = 1; 4747 4748 if (bbr->rc_all_timers_stopped) { 4749 return (1); 4750 } 4751 if (bbr->rc_in_persist == 0) 4752 return (0); 4753 4754 /* 4755 * Persistence timer into zero window. Force a byte to be output, if 4756 * possible. 4757 */ 4758 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); 4759 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 4760 KMOD_TCPSTAT_INC(tcps_persisttimeo); 4761 /* 4762 * Have we exceeded the user specified progress time? 4763 */ 4764 if (ctf_progress_timeout_check(tp, true)) { 4765 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4766 return (-ETIMEDOUT); /* tcp_drop() */ 4767 } 4768 /* 4769 * Hack: if the peer is dead/unreachable, we do not time out if the 4770 * window is closed. After a full backoff, drop the connection if 4771 * the idle time (no responses to probes) reaches the maximum 4772 * backoff that we would use if retransmitting. 4773 */ 4774 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 4775 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 4776 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 4777 KMOD_TCPSTAT_INC(tcps_persistdrop); 4778 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4779 return (-ETIMEDOUT); /* tcp_drop() */ 4780 } 4781 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && 4782 tp->snd_una == tp->snd_max) { 4783 bbr_exit_persist(tp, bbr, cts, __LINE__); 4784 retval = 0; 4785 goto out; 4786 } 4787 /* 4788 * If the user has closed the socket then drop a persisting 4789 * connection after a much reduced timeout. 4790 */ 4791 if (tp->t_state > TCPS_CLOSE_WAIT && 4792 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 4793 KMOD_TCPSTAT_INC(tcps_persistdrop); 4794 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4795 return (-ETIMEDOUT); /* tcp_drop() */ 4796 } 4797 t_template = tcpip_maketemplate(bbr->rc_inp); 4798 if (t_template) { 4799 tcp_respond(tp, t_template->tt_ipgen, 4800 &t_template->tt_t, (struct mbuf *)NULL, 4801 tp->rcv_nxt, tp->snd_una - 1, 0); 4802 /* This sends an ack */ 4803 if (tp->t_flags & TF_DELACK) 4804 tp->t_flags &= ~TF_DELACK; 4805 free(t_template, M_TEMP); 4806 } 4807 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 4808 tp->t_rxtshift++; 4809 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); 4810 out: 4811 return (retval); 4812 } 4813 4814 /* 4815 * If a keepalive goes off, we had no other timers 4816 * happening. We always return 1 here since this 4817 * routine either drops the connection or sends 4818 * out a segment with respond. 4819 */ 4820 static int 4821 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4822 { 4823 struct tcptemp *t_template; 4824 struct inpcb *inp = tptoinpcb(tp); 4825 4826 if (bbr->rc_all_timers_stopped) { 4827 return (1); 4828 } 4829 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 4830 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); 4831 /* 4832 * Keep-alive timer went off; send something or drop connection if 4833 * idle for too long. 4834 */ 4835 KMOD_TCPSTAT_INC(tcps_keeptimeo); 4836 if (tp->t_state < TCPS_ESTABLISHED) 4837 goto dropit; 4838 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 4839 tp->t_state <= TCPS_CLOSING) { 4840 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 4841 goto dropit; 4842 /* 4843 * Send a packet designed to force a response if the peer is 4844 * up and reachable: either an ACK if the connection is 4845 * still alive, or an RST if the peer has closed the 4846 * connection due to timeout or reboot. Using sequence 4847 * number tp->snd_una-1 causes the transmitted zero-length 4848 * segment to lie outside the receive window; by the 4849 * protocol spec, this requires the correspondent TCP to 4850 * respond. 4851 */ 4852 KMOD_TCPSTAT_INC(tcps_keepprobe); 4853 t_template = tcpip_maketemplate(inp); 4854 if (t_template) { 4855 tcp_respond(tp, t_template->tt_ipgen, 4856 &t_template->tt_t, (struct mbuf *)NULL, 4857 tp->rcv_nxt, tp->snd_una - 1, 0); 4858 free(t_template, M_TEMP); 4859 } 4860 } 4861 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); 4862 return (1); 4863 dropit: 4864 KMOD_TCPSTAT_INC(tcps_keepdrops); 4865 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); 4866 return (-ETIMEDOUT); /* tcp_drop() */ 4867 } 4868 4869 /* 4870 * Retransmit helper function, clear up all the ack 4871 * flags and take care of important book keeping. 4872 */ 4873 static void 4874 bbr_remxt_tmr(struct tcpcb *tp) 4875 { 4876 /* 4877 * The retransmit timer went off, all sack'd blocks must be 4878 * un-acked. 4879 */ 4880 struct bbr_sendmap *rsm, *trsm = NULL; 4881 struct tcp_bbr *bbr; 4882 uint32_t cts, lost; 4883 4884 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 4885 cts = tcp_get_usecs(&bbr->rc_tv); 4886 lost = bbr->r_ctl.rc_lost; 4887 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4888 bbr_set_state(tp, bbr, 0); 4889 4890 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 4891 if (rsm->r_flags & BBR_ACKED) { 4892 uint32_t old_flags; 4893 4894 rsm->r_dupack = 0; 4895 if (rsm->r_in_tmap == 0) { 4896 /* We must re-add it back to the tlist */ 4897 if (trsm == NULL) { 4898 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 4899 } else { 4900 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); 4901 } 4902 rsm->r_in_tmap = 1; 4903 } 4904 old_flags = rsm->r_flags; 4905 rsm->r_flags |= BBR_RXT_CLEARED; 4906 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); 4907 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 4908 } else { 4909 if ((tp->t_state < TCPS_ESTABLISHED) && 4910 (rsm->r_start == tp->snd_una)) { 4911 /* 4912 * Special case for TCP FO. Where 4913 * we sent more data beyond the snd_max. 4914 * We don't mark that as lost and stop here. 4915 */ 4916 break; 4917 } 4918 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4919 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4920 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4921 } 4922 if (bbr_marks_rxt_sack_passed) { 4923 /* 4924 * With this option, we will rack out 4925 * in 1ms increments the rest of the packets. 4926 */ 4927 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; 4928 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4929 } else { 4930 /* 4931 * With this option we only mark them lost 4932 * and remove all sack'd markings. We will run 4933 * another RXT or a TLP. This will cause 4934 * us to eventually send more based on what 4935 * ack's come in. 4936 */ 4937 rsm->r_flags |= BBR_MARKED_LOST; 4938 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4939 rsm->r_flags &= ~BBR_SACK_PASSED; 4940 } 4941 } 4942 trsm = rsm; 4943 } 4944 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4945 /* Clear the count (we just un-acked them) */ 4946 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); 4947 bbr->rc_tlp_new_data = 0; 4948 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4949 /* zap the behindness on a rxt */ 4950 bbr->r_ctl.rc_hptsi_agg_delay = 0; 4951 bbr->r_agg_early_set = 0; 4952 bbr->r_ctl.rc_agg_early = 0; 4953 bbr->rc_tlp_rtx_out = 0; 4954 bbr->r_ctl.rc_sacked = 0; 4955 bbr->r_ctl.rc_sacklast = NULL; 4956 bbr->r_timer_override = 1; 4957 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4958 } 4959 4960 /* 4961 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 4962 * we will setup to retransmit the lowest seq number outstanding. 4963 */ 4964 static int 4965 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4966 { 4967 struct inpcb *inp = tptoinpcb(tp); 4968 int32_t rexmt; 4969 int32_t retval = 0; 4970 bool isipv6; 4971 4972 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 4973 if (bbr->rc_all_timers_stopped) { 4974 return (1); 4975 } 4976 if (TCPS_HAVEESTABLISHED(tp->t_state) && 4977 (tp->snd_una == tp->snd_max)) { 4978 /* Nothing outstanding .. nothing to do */ 4979 return (0); 4980 } 4981 /* 4982 * Retransmission timer went off. Message has not been acked within 4983 * retransmit interval. Back off to a longer retransmit interval 4984 * and retransmit one segment. 4985 */ 4986 if (ctf_progress_timeout_check(tp, true)) { 4987 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4988 return (-ETIMEDOUT); /* tcp_drop() */ 4989 } 4990 bbr_remxt_tmr(tp); 4991 if ((bbr->r_ctl.rc_resend == NULL) || 4992 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { 4993 /* 4994 * If the rwnd collapsed on 4995 * the one we are retransmitting 4996 * it does not count against the 4997 * rxt count. 4998 */ 4999 tp->t_rxtshift++; 5000 } 5001 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { 5002 tp->t_rxtshift = TCP_MAXRXTSHIFT; 5003 KMOD_TCPSTAT_INC(tcps_timeoutdrop); 5004 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); 5005 /* XXXGL: previously t_softerror was casted to uint16_t */ 5006 MPASS(tp->t_softerror >= 0); 5007 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; 5008 return (retval); /* tcp_drop() */ 5009 } 5010 if (tp->t_state == TCPS_SYN_SENT) { 5011 /* 5012 * If the SYN was retransmitted, indicate CWND to be limited 5013 * to 1 segment in cc_conn_init(). 5014 */ 5015 tp->snd_cwnd = 1; 5016 } else if (tp->t_rxtshift == 1) { 5017 /* 5018 * first retransmit; record ssthresh and cwnd so they can be 5019 * recovered if this turns out to be a "bad" retransmit. A 5020 * retransmit is considered "bad" if an ACK for this segment 5021 * is received within RTT/2 interval; the assumption here is 5022 * that the ACK was already in flight. See "On Estimating 5023 * End-to-End Network Path Properties" by Allman and Paxson 5024 * for more details. 5025 */ 5026 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5027 if (!IN_RECOVERY(tp->t_flags)) { 5028 tp->snd_cwnd_prev = tp->snd_cwnd; 5029 tp->snd_ssthresh_prev = tp->snd_ssthresh; 5030 tp->snd_recover_prev = tp->snd_recover; 5031 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 5032 tp->t_flags |= TF_PREVVALID; 5033 } else { 5034 tp->t_flags &= ~TF_PREVVALID; 5035 } 5036 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5037 } else { 5038 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5039 tp->t_flags &= ~TF_PREVVALID; 5040 } 5041 KMOD_TCPSTAT_INC(tcps_rexmttimeo); 5042 if ((tp->t_state == TCPS_SYN_SENT) || 5043 (tp->t_state == TCPS_SYN_RECEIVED)) 5044 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; 5045 else 5046 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 5047 TCPT_RANGESET(tp->t_rxtcur, rexmt, 5048 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), 5049 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 5050 /* 5051 * We enter the path for PLMTUD if connection is established or, if 5052 * connection is FIN_WAIT_1 status, reason for the last is that if 5053 * amount of data we send is very small, we could send it in couple 5054 * of packets and process straight to FIN. In that case we won't 5055 * catch ESTABLISHED state. 5056 */ 5057 #ifdef INET6 5058 isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false; 5059 #else 5060 isipv6 = false; 5061 #endif 5062 if (((V_tcp_pmtud_blackhole_detect == 1) || 5063 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || 5064 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && 5065 ((tp->t_state == TCPS_ESTABLISHED) || 5066 (tp->t_state == TCPS_FIN_WAIT_1))) { 5067 /* 5068 * Idea here is that at each stage of mtu probe (usually, 5069 * 1448 -> 1188 -> 524) should be given 2 chances to recover 5070 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 5071 * should take care of that. 5072 */ 5073 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 5074 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 5075 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 5076 tp->t_rxtshift % 2 == 0)) { 5077 /* 5078 * Enter Path MTU Black-hole Detection mechanism: - 5079 * Disable Path MTU Discovery (IP "DF" bit). - 5080 * Reduce MTU to lower value than what we negotiated 5081 * with peer. 5082 */ 5083 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 5084 /* 5085 * Record that we may have found a black 5086 * hole. 5087 */ 5088 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 5089 /* Keep track of previous MSS. */ 5090 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 5091 } 5092 /* 5093 * Reduce the MSS to blackhole value or to the 5094 * default in an attempt to retransmit. 5095 */ 5096 #ifdef INET6 5097 isipv6 = bbr->r_is_v6; 5098 if (isipv6 && 5099 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 5100 /* Use the sysctl tuneable blackhole MSS. */ 5101 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 5102 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5103 } else if (isipv6) { 5104 /* Use the default MSS. */ 5105 tp->t_maxseg = V_tcp_v6mssdflt; 5106 /* 5107 * Disable Path MTU Discovery when we switch 5108 * to minmss. 5109 */ 5110 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5111 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5112 } 5113 #endif 5114 #if defined(INET6) && defined(INET) 5115 else 5116 #endif 5117 #ifdef INET 5118 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 5119 /* Use the sysctl tuneable blackhole MSS. */ 5120 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 5121 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5122 } else { 5123 /* Use the default MSS. */ 5124 tp->t_maxseg = V_tcp_mssdflt; 5125 /* 5126 * Disable Path MTU Discovery when we switch 5127 * to minmss. 5128 */ 5129 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5130 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5131 } 5132 #endif 5133 } else { 5134 /* 5135 * If further retransmissions are still unsuccessful 5136 * with a lowered MTU, maybe this isn't a blackhole 5137 * and we restore the previous MSS and blackhole 5138 * detection flags. The limit '6' is determined by 5139 * giving each probe stage (1448, 1188, 524) 2 5140 * chances to recover. 5141 */ 5142 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 5143 (tp->t_rxtshift >= 6)) { 5144 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 5145 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 5146 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 5147 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); 5148 } 5149 } 5150 } 5151 /* 5152 * Disable RFC1323 and SACK if we haven't got any response to our 5153 * third SYN to work-around some broken terminal servers (most of 5154 * which have hopefully been retired) that have bad VJ header 5155 * compression code which trashes TCP segments containing 5156 * unknown-to-them TCP options. 5157 */ 5158 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 5159 (tp->t_rxtshift == 3)) 5160 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 5161 /* 5162 * If we backed off this far, our srtt estimate is probably bogus. 5163 * Clobber it so we'll take the next rtt measurement as our srtt; 5164 * move the current srtt into rttvar to keep the current retransmit 5165 * times until then. 5166 */ 5167 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 5168 #ifdef INET6 5169 if (bbr->r_is_v6) 5170 in6_losing(inp); 5171 else 5172 #endif 5173 in_losing(inp); 5174 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 5175 tp->t_srtt = 0; 5176 } 5177 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 5178 tp->snd_recover = tp->snd_max; 5179 tp->t_flags |= TF_ACKNOW; 5180 tp->t_rtttime = 0; 5181 5182 return (retval); 5183 } 5184 5185 static int 5186 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) 5187 { 5188 int32_t ret = 0; 5189 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 5190 5191 if (timers == 0) { 5192 return (0); 5193 } 5194 if (tp->t_state == TCPS_LISTEN) { 5195 /* no timers on listen sockets */ 5196 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 5197 return (0); 5198 return (1); 5199 } 5200 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 5201 uint32_t left; 5202 5203 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 5204 ret = -1; 5205 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5206 return (0); 5207 } 5208 if (hpts_calling == 0) { 5209 ret = -2; 5210 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5211 return (0); 5212 } 5213 /* 5214 * Ok our timer went off early and we are not paced false 5215 * alarm, go back to sleep. 5216 */ 5217 left = bbr->r_ctl.rc_timer_exp - cts; 5218 ret = -3; 5219 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); 5220 tcp_hpts_insert(tptoinpcb(tp), HPTS_USEC_TO_SLOTS(left)); 5221 return (1); 5222 } 5223 bbr->rc_tmr_stopped = 0; 5224 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 5225 if (timers & PACE_TMR_DELACK) { 5226 ret = bbr_timeout_delack(tp, bbr, cts); 5227 } else if (timers & PACE_TMR_PERSIT) { 5228 ret = bbr_timeout_persist(tp, bbr, cts); 5229 } else if (timers & PACE_TMR_RACK) { 5230 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5231 ret = bbr_timeout_rack(tp, bbr, cts); 5232 } else if (timers & PACE_TMR_TLP) { 5233 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5234 ret = bbr_timeout_tlp(tp, bbr, cts); 5235 } else if (timers & PACE_TMR_RXT) { 5236 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5237 ret = bbr_timeout_rxt(tp, bbr, cts); 5238 } else if (timers & PACE_TMR_KEEP) { 5239 ret = bbr_timeout_keepalive(tp, bbr, cts); 5240 } 5241 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); 5242 return (ret); 5243 } 5244 5245 static void 5246 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) 5247 { 5248 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 5249 uint8_t hpts_removed = 0; 5250 5251 if (tcp_in_hpts(bbr->rc_inp) && 5252 (bbr->rc_timer_first == 1)) { 5253 /* 5254 * If we are canceling timer's when we have the 5255 * timer ahead of the output being paced. We also 5256 * must remove ourselves from the hpts. 5257 */ 5258 hpts_removed = 1; 5259 tcp_hpts_remove(bbr->rc_inp); 5260 if (bbr->r_ctl.rc_last_delay_val) { 5261 /* Update the last hptsi delay too */ 5262 uint32_t time_since_send; 5263 5264 if (TSTMP_GT(cts, bbr->rc_pacer_started)) 5265 time_since_send = cts - bbr->rc_pacer_started; 5266 else 5267 time_since_send = 0; 5268 if (bbr->r_ctl.rc_last_delay_val > time_since_send) { 5269 /* Cut down our slot time */ 5270 bbr->r_ctl.rc_last_delay_val -= time_since_send; 5271 } else { 5272 bbr->r_ctl.rc_last_delay_val = 0; 5273 } 5274 bbr->rc_pacer_started = cts; 5275 } 5276 } 5277 bbr->rc_timer_first = 0; 5278 bbr_log_to_cancel(bbr, line, cts, hpts_removed); 5279 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 5280 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 5281 } 5282 } 5283 5284 static int 5285 bbr_stopall(struct tcpcb *tp) 5286 { 5287 struct tcp_bbr *bbr; 5288 5289 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 5290 bbr->rc_all_timers_stopped = 1; 5291 return (0); 5292 } 5293 5294 static uint32_t 5295 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) 5296 { 5297 struct bbr_sendmap *rsm; 5298 5299 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 5300 if ((rsm == NULL) || (u_rsm == rsm)) 5301 return (cts); 5302 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 5303 } 5304 5305 static void 5306 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, 5307 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) 5308 { 5309 int32_t idx; 5310 5311 rsm->r_rtr_cnt++; 5312 rsm->r_dupack = 0; 5313 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { 5314 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; 5315 rsm->r_flags |= BBR_OVERMAX; 5316 } 5317 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 5318 /* Take off the collapsed flag at rxt */ 5319 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 5320 } 5321 if (rsm->r_flags & BBR_MARKED_LOST) { 5322 /* We have retransmitted, its no longer lost */ 5323 rsm->r_flags &= ~BBR_MARKED_LOST; 5324 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 5325 } 5326 if (rsm->r_flags & BBR_RXT_CLEARED) { 5327 /* 5328 * We hit a RXT timer on it and 5329 * we cleared the "acked" flag. 5330 * We now have it going back into 5331 * flight, we can remove the cleared 5332 * flag and possibly do accounting on 5333 * this piece. 5334 */ 5335 rsm->r_flags &= ~BBR_RXT_CLEARED; 5336 } 5337 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { 5338 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 5339 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 5340 } 5341 idx = rsm->r_rtr_cnt - 1; 5342 rsm->r_tim_lastsent[idx] = cts; 5343 rsm->r_pacing_delay = pacing_time; 5344 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5345 rsm->r_ts_valid = bbr->rc_ts_valid; 5346 if (bbr->rc_ts_valid) 5347 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5348 if (bbr->r_ctl.r_app_limited_until) 5349 rsm->r_app_limited = 1; 5350 else 5351 rsm->r_app_limited = 0; 5352 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5353 rsm->r_bbr_state = bbr_state_val(bbr); 5354 else 5355 rsm->r_bbr_state = 8; 5356 if (rsm->r_flags & BBR_ACKED) { 5357 /* Problably MTU discovery messing with us */ 5358 uint32_t old_flags; 5359 5360 old_flags = rsm->r_flags; 5361 rsm->r_flags &= ~BBR_ACKED; 5362 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 5363 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 5364 if (bbr->r_ctl.rc_sacked == 0) 5365 bbr->r_ctl.rc_sacklast = NULL; 5366 } 5367 if (rsm->r_in_tmap) { 5368 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5369 } 5370 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5371 rsm->r_in_tmap = 1; 5372 if (rsm->r_flags & BBR_SACK_PASSED) { 5373 /* We have retransmitted due to the SACK pass */ 5374 rsm->r_flags &= ~BBR_SACK_PASSED; 5375 rsm->r_flags |= BBR_WAS_SACKPASS; 5376 } 5377 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5378 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5379 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5380 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 5381 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5382 rsm->r_is_gain = 1; 5383 rsm->r_is_drain = 0; 5384 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5385 rsm->r_is_drain = 1; 5386 rsm->r_is_gain = 0; 5387 } else { 5388 rsm->r_is_drain = 0; 5389 rsm->r_is_gain = 0; 5390 } 5391 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ 5392 } 5393 5394 /* 5395 * Returns 0, or the sequence where we stopped 5396 * updating. We also update the lenp to be the amount 5397 * of data left. 5398 */ 5399 5400 static uint32_t 5401 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, 5402 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) 5403 { 5404 /* 5405 * We (re-)transmitted starting at rsm->r_start for some length 5406 * (possibly less than r_end. 5407 */ 5408 struct bbr_sendmap *nrsm; 5409 uint32_t c_end; 5410 int32_t len; 5411 5412 len = *lenp; 5413 c_end = rsm->r_start + len; 5414 if (SEQ_GEQ(c_end, rsm->r_end)) { 5415 /* 5416 * We retransmitted the whole piece or more than the whole 5417 * slopping into the next rsm. 5418 */ 5419 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5420 if (c_end == rsm->r_end) { 5421 *lenp = 0; 5422 return (0); 5423 } else { 5424 int32_t act_len; 5425 5426 /* Hangs over the end return whats left */ 5427 act_len = rsm->r_end - rsm->r_start; 5428 *lenp = (len - act_len); 5429 return (rsm->r_end); 5430 } 5431 /* We don't get out of this block. */ 5432 } 5433 /* 5434 * Here we retransmitted less than the whole thing which means we 5435 * have to split this into what was transmitted and what was not. 5436 */ 5437 nrsm = bbr_alloc_full_limit(bbr); 5438 if (nrsm == NULL) { 5439 *lenp = 0; 5440 return (0); 5441 } 5442 /* 5443 * So here we are going to take the original rsm and make it what we 5444 * retransmitted. nrsm will be the tail portion we did not 5445 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 5446 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 5447 * 1, 6 and the new piece will be 6, 11. 5448 */ 5449 bbr_clone_rsm(bbr, nrsm, rsm, c_end); 5450 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 5451 nrsm->r_dupack = 0; 5452 if (rsm->r_in_tmap) { 5453 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 5454 nrsm->r_in_tmap = 1; 5455 } 5456 rsm->r_flags &= (~BBR_HAS_FIN); 5457 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5458 *lenp = 0; 5459 return (0); 5460 } 5461 5462 static uint64_t 5463 bbr_get_hardware_rate(struct tcp_bbr *bbr) 5464 { 5465 uint64_t bw; 5466 5467 bw = bbr_get_bw(bbr); 5468 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; 5469 bw /= (uint64_t)BBR_UNIT; 5470 return(bw); 5471 } 5472 5473 static void 5474 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, 5475 uint64_t act_rate, uint64_t rate_wanted) 5476 { 5477 /* 5478 * We could not get a full gains worth 5479 * of rate. 5480 */ 5481 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { 5482 /* we can't even get the real rate */ 5483 uint64_t red; 5484 5485 bbr->skip_gain = 1; 5486 bbr->gain_is_limited = 0; 5487 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; 5488 if (red) 5489 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); 5490 } else { 5491 /* We can use a lower gain */ 5492 bbr->skip_gain = 0; 5493 bbr->gain_is_limited = 1; 5494 } 5495 } 5496 5497 static void 5498 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) 5499 { 5500 const struct tcp_hwrate_limit_table *nrte; 5501 int error, rate = -1; 5502 5503 if (bbr->r_ctl.crte == NULL) 5504 return; 5505 if ((bbr->rc_inp->inp_route.ro_nh == NULL) || 5506 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { 5507 /* Lost our routes? */ 5508 /* Clear the way for a re-attempt */ 5509 bbr->bbr_attempt_hdwr_pace = 0; 5510 lost_rate: 5511 bbr->gain_is_limited = 0; 5512 bbr->skip_gain = 0; 5513 bbr->bbr_hdrw_pacing = 0; 5514 counter_u64_add(bbr_flows_whdwr_pacing, -1); 5515 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 5516 tcp_bbr_tso_size_check(bbr, cts); 5517 return; 5518 } 5519 rate = bbr_get_hardware_rate(bbr); 5520 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, 5521 bbr->rc_tp, 5522 bbr->rc_inp->inp_route.ro_nh->nh_ifp, 5523 rate, 5524 (RS_PACING_GEQ|RS_PACING_SUB_OK), 5525 &error, NULL); 5526 if (nrte == NULL) { 5527 goto lost_rate; 5528 } 5529 if (nrte != bbr->r_ctl.crte) { 5530 bbr->r_ctl.crte = nrte; 5531 if (error == 0) { 5532 BBR_STAT_INC(bbr_hdwr_rl_mod_ok); 5533 if (bbr->r_ctl.crte->rate < rate) { 5534 /* We have a problem */ 5535 bbr_setup_less_of_rate(bbr, cts, 5536 bbr->r_ctl.crte->rate, rate); 5537 } else { 5538 /* We are good */ 5539 bbr->gain_is_limited = 0; 5540 bbr->skip_gain = 0; 5541 } 5542 } else { 5543 /* A failure should release the tag */ 5544 BBR_STAT_INC(bbr_hdwr_rl_mod_fail); 5545 bbr->gain_is_limited = 0; 5546 bbr->skip_gain = 0; 5547 bbr->bbr_hdrw_pacing = 0; 5548 } 5549 bbr_type_log_hdwr_pacing(bbr, 5550 bbr->r_ctl.crte->ptbl->rs_ifp, 5551 rate, 5552 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), 5553 __LINE__, 5554 cts, 5555 error); 5556 } 5557 } 5558 5559 static void 5560 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) 5561 { 5562 /* 5563 * If we have hardware pacing support 5564 * we need to factor that in for our 5565 * TSO size. 5566 */ 5567 const struct tcp_hwrate_limit_table *rlp; 5568 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; 5569 5570 if ((bbr->bbr_hdrw_pacing == 0) || 5571 (IN_RECOVERY(bbr->rc_tp->t_flags)) || 5572 (bbr->r_ctl.crte == NULL)) 5573 return; 5574 if (bbr->hw_pacing_set == 0) { 5575 /* Not yet by the hdwr pacing count delay */ 5576 return; 5577 } 5578 if (bbr_hdwr_pace_adjust == 0) { 5579 /* No adjustment */ 5580 return; 5581 } 5582 rlp = bbr->r_ctl.crte; 5583 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) 5584 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5585 else 5586 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5587 /* 5588 * So lets first get the 5589 * time we will take between 5590 * TSO sized sends currently without 5591 * hardware help. 5592 */ 5593 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, 5594 bbr->r_ctl.rc_pace_max_segs, cts, 1); 5595 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; 5596 hdwr_delay *= rlp->time_between; 5597 if (cur_delay > hdwr_delay) 5598 delta = cur_delay - hdwr_delay; 5599 else 5600 delta = 0; 5601 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, 5602 (bbr->r_ctl.rc_pace_max_segs / maxseg), 5603 1); 5604 if (delta && 5605 (delta < (max(rlp->time_between, 5606 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { 5607 /* 5608 * Now lets divide by the pacing 5609 * time between each segment the 5610 * hardware sends rounding up and 5611 * derive a bytes from that. We multiply 5612 * that by bbr_hdwr_pace_adjust to get 5613 * more bang for our buck. 5614 * 5615 * The goal is to have the software pacer 5616 * waiting no more than an additional 5617 * pacing delay if we can (without the 5618 * compensation i.e. x bbr_hdwr_pace_adjust). 5619 */ 5620 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), 5621 (bbr->r_ctl.rc_pace_max_segs/maxseg)); 5622 seg_sz *= bbr_hdwr_pace_adjust; 5623 if (bbr_hdwr_pace_floor && 5624 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5625 /* Currently hardware paces 5626 * out rs_min_seg segments at a time. 5627 * We need to make sure we always send at least 5628 * a full burst of bbr_hdwr_pace_floor down. 5629 */ 5630 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5631 } 5632 seg_sz *= maxseg; 5633 } else if (delta == 0) { 5634 /* 5635 * The highest pacing rate is 5636 * above our b/w gained. This means 5637 * we probably are going quite fast at 5638 * the hardware highest rate. Lets just multiply 5639 * the calculated TSO size by the 5640 * multiplier factor (its probably 5641 * 4 segments in the default config for 5642 * mlx). 5643 */ 5644 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; 5645 if (bbr_hdwr_pace_floor && 5646 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5647 /* Currently hardware paces 5648 * out rs_min_seg segments at a time. 5649 * We need to make sure we always send at least 5650 * a full burst of bbr_hdwr_pace_floor down. 5651 */ 5652 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5653 } 5654 } else { 5655 /* 5656 * The pacing time difference is so 5657 * big that the hardware will 5658 * pace out more rapidly then we 5659 * really want and then we 5660 * will have a long delay. Lets just keep 5661 * the same TSO size so its as if 5662 * we were not using hdwr pacing (we 5663 * just gain a bit of spacing from the 5664 * hardware if seg_sz > 1). 5665 */ 5666 seg_sz = bbr->r_ctl.rc_pace_max_segs; 5667 } 5668 if (seg_sz > bbr->r_ctl.rc_pace_max_segs) 5669 new_tso = seg_sz; 5670 else 5671 new_tso = bbr->r_ctl.rc_pace_max_segs; 5672 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) 5673 new_tso = PACE_MAX_IP_BYTES - maxseg; 5674 5675 if (new_tso != bbr->r_ctl.rc_pace_max_segs) { 5676 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); 5677 bbr->r_ctl.rc_pace_max_segs = new_tso; 5678 } 5679 } 5680 5681 static void 5682 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) 5683 { 5684 uint64_t bw; 5685 uint32_t old_tso = 0, new_tso; 5686 uint32_t maxseg, bytes; 5687 uint32_t tls_seg=0; 5688 /* 5689 * Google/linux uses the following algorithm to determine 5690 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): 5691 * 5692 * bytes = bw_in_bytes_per_second / 1000 5693 * bytes = min(bytes, 64k) 5694 * tso_segs = bytes / MSS 5695 * if (bw < 1.2Mbs) 5696 * min_tso_segs = 1 5697 * else 5698 * min_tso_segs = 2 5699 * tso_segs = max(tso_segs, min_tso_segs) 5700 * 5701 * * Note apply a device specific limit (we apply this in the 5702 * tcp_m_copym). 5703 * Note that before the initial measurement is made google bursts out 5704 * a full iwnd just like new-reno/cubic. 5705 * 5706 * We do not use this algorithm. Instead we 5707 * use a two phased approach: 5708 * 5709 * if ( bw <= per-tcb-cross-over) 5710 * goal_tso = calculate how much with this bw we 5711 * can send in goal-time seconds. 5712 * if (goal_tso > mss) 5713 * seg = goal_tso / mss 5714 * tso = seg * mss 5715 * else 5716 * tso = mss 5717 * if (tso > per-tcb-max) 5718 * tso = per-tcb-max 5719 * else if ( bw > 512Mbps) 5720 * tso = max-tso (64k/mss) 5721 * else 5722 * goal_tso = bw / per-tcb-divsor 5723 * seg = (goal_tso + mss-1)/mss 5724 * tso = seg * mss 5725 * 5726 * if (tso < per-tcb-floor) 5727 * tso = per-tcb-floor 5728 * if (tso > per-tcb-utter_max) 5729 * tso = per-tcb-utter_max 5730 * 5731 * Note the default per-tcb-divisor is 1000 (same as google). 5732 * the goal cross over is 30Mbps however. To recreate googles 5733 * algorithm you need to set: 5734 * 5735 * cross-over = 23,168,000 bps 5736 * goal-time = 18000 5737 * per-tcb-max = 2 5738 * per-tcb-divisor = 1000 5739 * per-tcb-floor = 1 5740 * 5741 * This will get you "google bbr" behavior with respect to tso size. 5742 * 5743 * Note we do set anything TSO size until we are past the initial 5744 * window. Before that we gnerally use either a single MSS 5745 * or we use the full IW size (so we burst a IW at a time) 5746 */ 5747 5748 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { 5749 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5750 } else { 5751 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5752 } 5753 old_tso = bbr->r_ctl.rc_pace_max_segs; 5754 if (bbr->rc_past_init_win == 0) { 5755 /* 5756 * Not enough data has been acknowledged to make a 5757 * judgement. Set up the initial TSO based on if we 5758 * are sending a full IW at once or not. 5759 */ 5760 if (bbr->rc_use_google) 5761 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); 5762 else if (bbr->bbr_init_win_cheat) 5763 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); 5764 else 5765 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5766 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) 5767 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; 5768 if (bbr->r_ctl.rc_pace_max_segs == 0) { 5769 bbr->r_ctl.rc_pace_max_segs = maxseg; 5770 } 5771 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); 5772 bbr_adjust_for_hw_pacing(bbr, cts); 5773 return; 5774 } 5775 /** 5776 * Now lets set the TSO goal based on our delivery rate in 5777 * bytes per second. Note we only do this if 5778 * we have acked at least the initial cwnd worth of data. 5779 */ 5780 bw = bbr_get_bw(bbr); 5781 if (IN_RECOVERY(bbr->rc_tp->t_flags) && 5782 (bbr->rc_use_google == 0)) { 5783 /* We clamp to one MSS in recovery */ 5784 new_tso = maxseg; 5785 } else if (bbr->rc_use_google) { 5786 int min_tso_segs; 5787 5788 /* Google considers the gain too */ 5789 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { 5790 bw *= bbr->r_ctl.rc_bbr_hptsi_gain; 5791 bw /= BBR_UNIT; 5792 } 5793 bytes = bw / 1024; 5794 if (bytes > (64 * 1024)) 5795 bytes = 64 * 1024; 5796 new_tso = bytes / maxseg; 5797 if (bw < ONE_POINT_TWO_MEG) 5798 min_tso_segs = 1; 5799 else 5800 min_tso_segs = 2; 5801 if (new_tso < min_tso_segs) 5802 new_tso = min_tso_segs; 5803 new_tso *= maxseg; 5804 } else if (bbr->rc_no_pacing) { 5805 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; 5806 } else if (bw <= bbr->r_ctl.bbr_cross_over) { 5807 /* 5808 * Calculate the worse case b/w TSO if we are inserting no 5809 * more than a delay_target number of TSO's. 5810 */ 5811 uint32_t tso_len, min_tso; 5812 5813 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); 5814 if (tso_len > maxseg) { 5815 new_tso = tso_len / maxseg; 5816 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) 5817 new_tso = bbr->r_ctl.bbr_hptsi_segments_max; 5818 new_tso *= maxseg; 5819 } else { 5820 /* 5821 * less than a full sized frame yikes.. long rtt or 5822 * low bw? 5823 */ 5824 min_tso = bbr_minseg(bbr); 5825 if ((tso_len > min_tso) && (bbr_all_get_min == 0)) 5826 new_tso = rounddown(tso_len, min_tso); 5827 else 5828 new_tso = min_tso; 5829 } 5830 } else if (bw > FIVETWELVE_MBPS) { 5831 /* 5832 * This guy is so fast b/w wise that we can TSO as large as 5833 * possible of segments that the NIC will allow. 5834 */ 5835 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5836 } else { 5837 /* 5838 * This formula is based on attempting to send a segment or 5839 * more every bbr_hptsi_per_second. The default is 1000 5840 * which means you are targeting what you can send every 1ms 5841 * based on the peers bw. 5842 * 5843 * If the number drops to say 500, then you are looking more 5844 * at 2ms and you will raise how much we send in a single 5845 * TSO thus saving CPU (less bbr_output_wtime() calls). The 5846 * trade off of course is you will send more at once and 5847 * thus tend to clump up the sends into larger "bursts" 5848 * building a queue. 5849 */ 5850 bw /= bbr->r_ctl.bbr_hptsi_per_second; 5851 new_tso = roundup(bw, (uint64_t)maxseg); 5852 /* 5853 * Gate the floor to match what our lower than 48Mbps 5854 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus 5855 * becomes the floor for this calculation. 5856 */ 5857 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) 5858 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); 5859 } 5860 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) 5861 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; 5862 if (new_tso > PACE_MAX_IP_BYTES) 5863 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5864 /* Enforce an utter maximum. */ 5865 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { 5866 new_tso = bbr->r_ctl.bbr_utter_max * maxseg; 5867 } 5868 if (old_tso != new_tso) { 5869 /* Only log changes */ 5870 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); 5871 bbr->r_ctl.rc_pace_max_segs = new_tso; 5872 } 5873 /* We have hardware pacing! */ 5874 bbr_adjust_for_hw_pacing(bbr, cts); 5875 } 5876 5877 static void 5878 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, 5879 uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, 5880 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, 5881 struct sockbuf *sb) 5882 { 5883 5884 struct bbr_sendmap *rsm, *nrsm; 5885 register uint32_t snd_max, snd_una; 5886 uint32_t pacing_time; 5887 /* 5888 * Add to the RACK log of packets in flight or retransmitted. If 5889 * there is a TS option we will use the TS echoed, if not we will 5890 * grab a TS. 5891 * 5892 * Retransmissions will increment the count and move the ts to its 5893 * proper place. Note that if options do not include TS's then we 5894 * won't be able to effectively use the ACK for an RTT on a retran. 5895 * 5896 * Notes about r_start and r_end. Lets consider a send starting at 5897 * sequence 1 for 10 bytes. In such an example the r_start would be 5898 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 5899 * This means that r_end is actually the first sequence for the next 5900 * slot (11). 5901 * 5902 */ 5903 INP_WLOCK_ASSERT(tptoinpcb(tp)); 5904 if (err) { 5905 /* 5906 * We don't log errors -- we could but snd_max does not 5907 * advance in this case either. 5908 */ 5909 return; 5910 } 5911 if (th_flags & TH_RST) { 5912 /* 5913 * We don't log resets and we return immediately from 5914 * sending 5915 */ 5916 *abandon = 1; 5917 return; 5918 } 5919 snd_una = tp->snd_una; 5920 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { 5921 /* 5922 * The call to bbr_log_output is made before bumping 5923 * snd_max. This means we can record one extra byte on a SYN 5924 * or FIN if seq_out is adding more on and a FIN is present 5925 * (and we are not resending). 5926 */ 5927 if ((th_flags & TH_SYN) && (tp->iss == seq_out)) 5928 len++; 5929 if (th_flags & TH_FIN) 5930 len++; 5931 } 5932 if (SEQ_LEQ((seq_out + len), snd_una)) { 5933 /* Are sending an old segment to induce an ack (keep-alive)? */ 5934 return; 5935 } 5936 if (SEQ_LT(seq_out, snd_una)) { 5937 /* huh? should we panic? */ 5938 uint32_t end; 5939 5940 end = seq_out + len; 5941 seq_out = snd_una; 5942 len = end - seq_out; 5943 } 5944 snd_max = tp->snd_max; 5945 if (len == 0) { 5946 /* We don't log zero window probes */ 5947 return; 5948 } 5949 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); 5950 /* First question is it a retransmission? */ 5951 if (seq_out == snd_max) { 5952 again: 5953 rsm = bbr_alloc(bbr); 5954 if (rsm == NULL) { 5955 return; 5956 } 5957 rsm->r_flags = 0; 5958 if (th_flags & TH_SYN) 5959 rsm->r_flags |= BBR_HAS_SYN; 5960 if (th_flags & TH_FIN) 5961 rsm->r_flags |= BBR_HAS_FIN; 5962 rsm->r_tim_lastsent[0] = cts; 5963 rsm->r_rtr_cnt = 1; 5964 rsm->r_rtr_bytes = 0; 5965 rsm->r_start = seq_out; 5966 rsm->r_end = rsm->r_start + len; 5967 rsm->r_dupack = 0; 5968 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5969 rsm->r_pacing_delay = pacing_time; 5970 rsm->r_ts_valid = bbr->rc_ts_valid; 5971 if (bbr->rc_ts_valid) 5972 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5973 rsm->r_del_time = bbr->r_ctl.rc_del_time; 5974 if (bbr->r_ctl.r_app_limited_until) 5975 rsm->r_app_limited = 1; 5976 else 5977 rsm->r_app_limited = 0; 5978 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5979 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5980 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5981 /* 5982 * Here we must also add in this rsm since snd_max 5983 * is updated after we return from a new send. 5984 */ 5985 rsm->r_flight_at_send += len; 5986 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 5987 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5988 rsm->r_in_tmap = 1; 5989 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5990 rsm->r_bbr_state = bbr_state_val(bbr); 5991 else 5992 rsm->r_bbr_state = 8; 5993 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5994 rsm->r_is_gain = 1; 5995 rsm->r_is_drain = 0; 5996 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5997 rsm->r_is_drain = 1; 5998 rsm->r_is_gain = 0; 5999 } else { 6000 rsm->r_is_drain = 0; 6001 rsm->r_is_gain = 0; 6002 } 6003 return; 6004 } 6005 /* 6006 * If we reach here its a retransmission and we need to find it. 6007 */ 6008 more: 6009 if (hintrsm && (hintrsm->r_start == seq_out)) { 6010 rsm = hintrsm; 6011 hintrsm = NULL; 6012 } else if (bbr->r_ctl.rc_next) { 6013 /* We have a hint from a previous run */ 6014 rsm = bbr->r_ctl.rc_next; 6015 } else { 6016 /* No hints sorry */ 6017 rsm = NULL; 6018 } 6019 if ((rsm) && (rsm->r_start == seq_out)) { 6020 /* 6021 * We used rc_next or hintrsm to retransmit, hopefully the 6022 * likely case. 6023 */ 6024 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6025 if (len == 0) { 6026 return; 6027 } else { 6028 goto more; 6029 } 6030 } 6031 /* Ok it was not the last pointer go through it the hard way. */ 6032 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6033 if (rsm->r_start == seq_out) { 6034 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6035 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 6036 if (len == 0) { 6037 return; 6038 } else { 6039 continue; 6040 } 6041 } 6042 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 6043 /* Transmitted within this piece */ 6044 /* 6045 * Ok we must split off the front and then let the 6046 * update do the rest 6047 */ 6048 nrsm = bbr_alloc_full_limit(bbr); 6049 if (nrsm == NULL) { 6050 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 6051 return; 6052 } 6053 /* 6054 * copy rsm to nrsm and then trim the front of rsm 6055 * to not include this part. 6056 */ 6057 bbr_clone_rsm(bbr, nrsm, rsm, seq_out); 6058 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 6059 if (rsm->r_in_tmap) { 6060 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 6061 nrsm->r_in_tmap = 1; 6062 } 6063 rsm->r_flags &= (~BBR_HAS_FIN); 6064 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); 6065 if (len == 0) { 6066 return; 6067 } 6068 } 6069 } 6070 /* 6071 * Hmm not found in map did they retransmit both old and on into the 6072 * new? 6073 */ 6074 if (seq_out == tp->snd_max) { 6075 goto again; 6076 } else if (SEQ_LT(seq_out, tp->snd_max)) { 6077 #ifdef BBR_INVARIANTS 6078 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 6079 seq_out, len, tp->snd_una, tp->snd_max); 6080 printf("Starting Dump of all rack entries\n"); 6081 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6082 printf("rsm:%p start:%u end:%u\n", 6083 rsm, rsm->r_start, rsm->r_end); 6084 } 6085 printf("Dump complete\n"); 6086 panic("seq_out not found rack:%p tp:%p", 6087 bbr, tp); 6088 #endif 6089 } else { 6090 #ifdef BBR_INVARIANTS 6091 /* 6092 * Hmm beyond sndmax? (only if we are using the new rtt-pack 6093 * flag) 6094 */ 6095 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 6096 seq_out, len, tp->snd_max, tp); 6097 #endif 6098 } 6099 } 6100 6101 static void 6102 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) 6103 { 6104 /* 6105 * Collapse timeout back the cum-ack moved. 6106 */ 6107 tp->t_rxtshift = 0; 6108 tp->t_softerror = 0; 6109 } 6110 6111 static void 6112 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) 6113 { 6114 bbr->rtt_valid = 1; 6115 bbr->r_ctl.cur_rtt = rtt_usecs; 6116 bbr->r_ctl.ts_in = tsin; 6117 if (rsm_send_time) 6118 bbr->r_ctl.cur_rtt_send_time = rsm_send_time; 6119 } 6120 6121 static void 6122 bbr_make_timestamp_determination(struct tcp_bbr *bbr) 6123 { 6124 /** 6125 * We have in our bbr control: 6126 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). 6127 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). 6128 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) 6129 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) 6130 * 6131 * Now we can calculate the time between the sends by doing: 6132 * 6133 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts 6134 * 6135 * And the peer's time between receiving them by doing: 6136 * 6137 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp 6138 * 6139 * We want to figure out if the timestamp values are in msec, 10msec or usec. 6140 * We also may find that we can't use the timestamps if say we see 6141 * that the peer_delta indicates that though we may have taken 10ms to 6142 * pace out the data, it only saw 1ms between the two packets. This would 6143 * indicate that somewhere on the path is a batching entity that is giving 6144 * out time-slices of the actual b/w. This would mean we could not use 6145 * reliably the peers timestamps. 6146 * 6147 * We expect delta > peer_delta initially. Until we figure out the 6148 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. 6149 * If we place 1000 there then its a ms vs our usec. If we place 10000 there 6150 * then its 10ms vs our usec. If the peer is running a usec clock we would 6151 * put a 1 there. If the value is faster then ours, we will disable the 6152 * use of timestamps (though we could revist this later if we find it to be not 6153 * just an isolated one or two flows)). 6154 * 6155 * To detect the batching middle boxes we will come up with our compensation and 6156 * if with it in place, we find the peer is drastically off (by some margin) in 6157 * the smaller direction, then we will assume the worst case and disable use of timestamps. 6158 * 6159 */ 6160 uint64_t delta, peer_delta, delta_up; 6161 6162 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; 6163 if (delta < bbr_min_usec_delta) { 6164 /* 6165 * Have not seen a min amount of time 6166 * between our send times so we can 6167 * make a determination of the timestamp 6168 * yet. 6169 */ 6170 return; 6171 } 6172 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; 6173 if (peer_delta < bbr_min_peer_delta) { 6174 /* 6175 * We may have enough in the form of 6176 * our delta but the peers number 6177 * has not changed that much. It could 6178 * be its clock ratio is such that 6179 * we need more data (10ms tick) or 6180 * there may be other compression scenarios 6181 * going on. In any event we need the 6182 * spread to be larger. 6183 */ 6184 return; 6185 } 6186 /* Ok lets first see which way our delta is going */ 6187 if (peer_delta > delta) { 6188 /* Very unlikely, the peer without 6189 * compensation shows that it saw 6190 * the two sends arrive further apart 6191 * then we saw then in micro-seconds. 6192 */ 6193 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { 6194 /* well it looks like the peer is a micro-second clock. */ 6195 bbr->rc_ts_clock_set = 1; 6196 bbr->r_ctl.bbr_peer_tsratio = 1; 6197 } else { 6198 bbr->rc_ts_cant_be_used = 1; 6199 bbr->rc_ts_clock_set = 1; 6200 } 6201 return; 6202 } 6203 /* Ok we know that the peer_delta is smaller than our send distance */ 6204 bbr->rc_ts_clock_set = 1; 6205 /* First question is it within the percentage that they are using usec time? */ 6206 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; 6207 if ((peer_delta + delta_up) >= delta) { 6208 /* Its a usec clock */ 6209 bbr->r_ctl.bbr_peer_tsratio = 1; 6210 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6211 return; 6212 } 6213 /* Ok if not usec, what about 10usec (though unlikely)? */ 6214 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; 6215 if (((peer_delta * 10) + delta_up) >= delta) { 6216 bbr->r_ctl.bbr_peer_tsratio = 10; 6217 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6218 return; 6219 } 6220 /* And what about 100usec (though again unlikely)? */ 6221 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; 6222 if (((peer_delta * 100) + delta_up) >= delta) { 6223 bbr->r_ctl.bbr_peer_tsratio = 100; 6224 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6225 return; 6226 } 6227 /* And how about 1 msec (the most likely one)? */ 6228 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; 6229 if (((peer_delta * 1000) + delta_up) >= delta) { 6230 bbr->r_ctl.bbr_peer_tsratio = 1000; 6231 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6232 return; 6233 } 6234 /* Ok if not msec could it be 10 msec? */ 6235 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; 6236 if (((peer_delta * 10000) + delta_up) >= delta) { 6237 bbr->r_ctl.bbr_peer_tsratio = 10000; 6238 return; 6239 } 6240 /* If we fall down here the clock tick so slowly we can't use it */ 6241 bbr->rc_ts_cant_be_used = 1; 6242 bbr->r_ctl.bbr_peer_tsratio = 0; 6243 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6244 } 6245 6246 /* 6247 * Collect new round-trip time estimate 6248 * and update averages and current timeout. 6249 */ 6250 static void 6251 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) 6252 { 6253 int32_t delta; 6254 uint32_t rtt, tsin; 6255 int32_t rtt_ticks; 6256 6257 if (bbr->rtt_valid == 0) 6258 /* No valid sample */ 6259 return; 6260 6261 rtt = bbr->r_ctl.cur_rtt; 6262 tsin = bbr->r_ctl.ts_in; 6263 if (bbr->rc_prtt_set_ts) { 6264 /* 6265 * We are to force feed the rttProp filter due 6266 * to an entry into PROBE_RTT. This assures 6267 * that the times are sync'd between when we 6268 * go into PROBE_RTT and the filter expiration. 6269 * 6270 * Google does not use a true filter, so they do 6271 * this implicitly since they only keep one value 6272 * and when they enter probe-rtt they update the 6273 * value to the newest rtt. 6274 */ 6275 uint32_t rtt_prop; 6276 6277 bbr->rc_prtt_set_ts = 0; 6278 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 6279 if (rtt > rtt_prop) 6280 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); 6281 else 6282 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6283 } 6284 if (bbr->rc_ack_was_delayed) 6285 rtt += bbr->r_ctl.rc_ack_hdwr_delay; 6286 6287 if (rtt < bbr->r_ctl.rc_lowest_rtt) 6288 bbr->r_ctl.rc_lowest_rtt = rtt; 6289 bbr_log_rtt_sample(bbr, rtt, tsin); 6290 if (bbr->r_init_rtt) { 6291 /* 6292 * The initial rtt is not-trusted, nuke it and lets get 6293 * our first valid measurement in. 6294 */ 6295 bbr->r_init_rtt = 0; 6296 tp->t_srtt = 0; 6297 } 6298 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { 6299 /* 6300 * So we have not yet figured out 6301 * what the peers TSTMP value is 6302 * in (most likely ms). We need a 6303 * series of cum-ack's to determine 6304 * this reliably. 6305 */ 6306 if (bbr->rc_ack_is_cumack) { 6307 if (bbr->rc_ts_data_set) { 6308 /* Lets attempt to determine the timestamp granularity. */ 6309 bbr_make_timestamp_determination(bbr); 6310 } else { 6311 bbr->rc_ts_data_set = 1; 6312 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; 6313 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; 6314 } 6315 } else { 6316 /* 6317 * We have to have consecutive acks 6318 * reset any "filled" state to none. 6319 */ 6320 bbr->rc_ts_data_set = 0; 6321 } 6322 } 6323 /* Round it up */ 6324 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); 6325 if (rtt_ticks == 0) 6326 rtt_ticks = 1; 6327 if (tp->t_srtt != 0) { 6328 /* 6329 * srtt is stored as fixed point with 5 bits after the 6330 * binary point (i.e., scaled by 8). The following magic is 6331 * equivalent to the smoothing algorithm in rfc793 with an 6332 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 6333 * Adjust rtt to origin 0. 6334 */ 6335 6336 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) 6337 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 6338 6339 tp->t_srtt += delta; 6340 if (tp->t_srtt <= 0) 6341 tp->t_srtt = 1; 6342 6343 /* 6344 * We accumulate a smoothed rtt variance (actually, a 6345 * smoothed mean difference), then set the retransmit timer 6346 * to smoothed rtt + 4 times the smoothed variance. rttvar 6347 * is stored as fixed point with 4 bits after the binary 6348 * point (scaled by 16). The following is equivalent to 6349 * rfc793 smoothing with an alpha of .75 (rttvar = 6350 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 6351 * wired-in beta. 6352 */ 6353 if (delta < 0) 6354 delta = -delta; 6355 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 6356 tp->t_rttvar += delta; 6357 if (tp->t_rttvar <= 0) 6358 tp->t_rttvar = 1; 6359 } else { 6360 /* 6361 * No rtt measurement yet - use the unsmoothed rtt. Set the 6362 * variance to half the rtt (so our first retransmit happens 6363 * at 3*rtt). 6364 */ 6365 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; 6366 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); 6367 } 6368 KMOD_TCPSTAT_INC(tcps_rttupdated); 6369 tp->t_rttupdated++; 6370 #ifdef STATS 6371 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); 6372 #endif 6373 /* 6374 * the retransmit should happen at rtt + 4 * rttvar. Because of the 6375 * way we do the smoothing, srtt and rttvar will each average +1/2 6376 * tick of bias. When we compute the retransmit timer, we want 1/2 6377 * tick of rounding and 1 extra tick because of +-1/2 tick 6378 * uncertainty in the firing of the timer. The bias will give us 6379 * exactly the 1.5 tick we need. But, because the bias is 6380 * statistical, we have to test that we don't drop below the minimum 6381 * feasible timer (which is 2 ticks). 6382 */ 6383 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 6384 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), 6385 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 6386 6387 /* 6388 * We received an ack for a packet that wasn't retransmitted; it is 6389 * probably safe to discard any error indications we've received 6390 * recently. This isn't quite right, but close enough for now (a 6391 * route might have failed after we sent a segment, and the return 6392 * path might not be symmetrical). 6393 */ 6394 tp->t_softerror = 0; 6395 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 6396 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) 6397 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; 6398 } 6399 6400 static void 6401 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) 6402 { 6403 bbr->r_ctl.rc_rtt_shrinks = cts; 6404 if (bbr_can_force_probertt && 6405 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 6406 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 6407 /* 6408 * We should enter probe-rtt its been too long 6409 * since we have been there. 6410 */ 6411 bbr_enter_probe_rtt(bbr, cts, __LINE__); 6412 } else 6413 bbr_check_probe_rtt_limits(bbr, cts); 6414 } 6415 6416 static void 6417 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) 6418 { 6419 uint64_t orig_bw; 6420 6421 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { 6422 /* We never apply a zero measurement */ 6423 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 6424 0, 0, 0, 0, 0, 0); 6425 return; 6426 } 6427 if (bbr->r_ctl.r_measurement_count < 0xffffffff) 6428 bbr->r_ctl.r_measurement_count++; 6429 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 6430 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); 6431 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, 6432 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 6433 0, 0, 0, 0, 0, 0); 6434 if (orig_bw && 6435 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { 6436 if (bbr->bbr_hdrw_pacing) { 6437 /* 6438 * Apply a new rate to the hardware 6439 * possibly. 6440 */ 6441 bbr_update_hardware_pacing_rate(bbr, cts); 6442 } 6443 bbr_set_state_target(bbr, __LINE__); 6444 tcp_bbr_tso_size_check(bbr, cts); 6445 if (bbr->r_recovery_bw) { 6446 bbr_setup_red_bw(bbr, cts); 6447 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); 6448 } 6449 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) 6450 tcp_bbr_tso_size_check(bbr, cts); 6451 } 6452 6453 static void 6454 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6455 { 6456 if (bbr->rc_in_persist == 0) { 6457 /* We log only when not in persist */ 6458 /* Translate to a Bytes Per Second */ 6459 uint64_t tim, bw, ts_diff, ts_bw; 6460 uint32_t delivered; 6461 6462 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6463 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6464 else 6465 tim = 1; 6466 /* 6467 * Now that we have processed the tim (skipping the sample 6468 * or possibly updating the time, go ahead and 6469 * calculate the cdr. 6470 */ 6471 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6472 bw = (uint64_t)delivered; 6473 bw *= (uint64_t)USECS_IN_SECOND; 6474 bw /= tim; 6475 if (bw == 0) { 6476 /* We must have a calculatable amount */ 6477 return; 6478 } 6479 /* 6480 * If we are using this b/w shove it in now so we 6481 * can see in the trace viewer if it gets over-ridden. 6482 */ 6483 if (rsm->r_ts_valid && 6484 bbr->rc_ts_valid && 6485 bbr->rc_ts_clock_set && 6486 (bbr->rc_ts_cant_be_used == 0) && 6487 bbr->rc_use_ts_limit) { 6488 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); 6489 ts_diff *= bbr->r_ctl.bbr_peer_tsratio; 6490 if ((delivered == 0) || 6491 (rtt < 1000)) { 6492 /* Can't use the ts */ 6493 bbr_log_type_bbrupd(bbr, 61, cts, 6494 ts_diff, 6495 bbr->r_ctl.last_inbound_ts, 6496 rsm->r_del_ack_ts, 0, 6497 0, 0, 0, delivered); 6498 } else { 6499 ts_bw = (uint64_t)delivered; 6500 ts_bw *= (uint64_t)USECS_IN_SECOND; 6501 ts_bw /= ts_diff; 6502 bbr_log_type_bbrupd(bbr, 62, cts, 6503 (ts_bw >> 32), 6504 (ts_bw & 0xffffffff), 0, 0, 6505 0, 0, ts_diff, delivered); 6506 if ((bbr->ts_can_raise) && 6507 (ts_bw > bw)) { 6508 bbr_log_type_bbrupd(bbr, 8, cts, 6509 delivered, 6510 ts_diff, 6511 (bw >> 32), 6512 (bw & 0x00000000ffffffff), 6513 0, 0, 0, 0); 6514 bw = ts_bw; 6515 } else if (ts_bw && (ts_bw < bw)) { 6516 bbr_log_type_bbrupd(bbr, 7, cts, 6517 delivered, 6518 ts_diff, 6519 (bw >> 32), 6520 (bw & 0x00000000ffffffff), 6521 0, 0, 0, 0); 6522 bw = ts_bw; 6523 } 6524 } 6525 } 6526 if (rsm->r_first_sent_time && 6527 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6528 uint64_t sbw, sti; 6529 /* 6530 * We use what was in flight at the time of our 6531 * send and the size of this send to figure 6532 * out what we have been sending at (amount). 6533 * For the time we take from the time of 6534 * the send of the first send outstanding 6535 * until this send plus this sends pacing 6536 * time. This gives us a good calculation 6537 * as to the rate we have been sending at. 6538 */ 6539 6540 sbw = (uint64_t)(rsm->r_flight_at_send); 6541 sbw *= (uint64_t)USECS_IN_SECOND; 6542 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6543 sti += rsm->r_pacing_delay; 6544 sbw /= sti; 6545 if (sbw < bw) { 6546 bbr_log_type_bbrupd(bbr, 6, cts, 6547 delivered, 6548 (uint32_t)sti, 6549 (bw >> 32), 6550 (uint32_t)bw, 6551 rsm->r_first_sent_time, 0, (sbw >> 32), 6552 (uint32_t)sbw); 6553 bw = sbw; 6554 } 6555 } 6556 /* Use the google algorithm for b/w measurements */ 6557 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6558 if ((rsm->r_app_limited == 0) || 6559 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { 6560 tcp_bbr_commit_bw(bbr, cts); 6561 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6562 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6563 } 6564 } 6565 } 6566 6567 static void 6568 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6569 { 6570 if (bbr->rc_in_persist == 0) { 6571 /* We log only when not in persist */ 6572 /* Translate to a Bytes Per Second */ 6573 uint64_t tim, bw; 6574 uint32_t delivered; 6575 int no_apply = 0; 6576 6577 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6578 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6579 else 6580 tim = 1; 6581 /* 6582 * Now that we have processed the tim (skipping the sample 6583 * or possibly updating the time, go ahead and 6584 * calculate the cdr. 6585 */ 6586 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6587 bw = (uint64_t)delivered; 6588 bw *= (uint64_t)USECS_IN_SECOND; 6589 bw /= tim; 6590 if (tim < bbr->r_ctl.rc_lowest_rtt) { 6591 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6592 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6593 6594 no_apply = 1; 6595 } 6596 /* 6597 * If we are using this b/w shove it in now so we 6598 * can see in the trace viewer if it gets over-ridden. 6599 */ 6600 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6601 /* Gate by the sending rate */ 6602 if (rsm->r_first_sent_time && 6603 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6604 uint64_t sbw, sti; 6605 /* 6606 * We use what was in flight at the time of our 6607 * send and the size of this send to figure 6608 * out what we have been sending at (amount). 6609 * For the time we take from the time of 6610 * the send of the first send outstanding 6611 * until this send plus this sends pacing 6612 * time. This gives us a good calculation 6613 * as to the rate we have been sending at. 6614 */ 6615 6616 sbw = (uint64_t)(rsm->r_flight_at_send); 6617 sbw *= (uint64_t)USECS_IN_SECOND; 6618 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6619 sti += rsm->r_pacing_delay; 6620 sbw /= sti; 6621 if (sbw < bw) { 6622 bbr_log_type_bbrupd(bbr, 6, cts, 6623 delivered, 6624 (uint32_t)sti, 6625 (bw >> 32), 6626 (uint32_t)bw, 6627 rsm->r_first_sent_time, 0, (sbw >> 32), 6628 (uint32_t)sbw); 6629 bw = sbw; 6630 } 6631 if ((sti > tim) && 6632 (sti < bbr->r_ctl.rc_lowest_rtt)) { 6633 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6634 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6635 no_apply = 1; 6636 } else 6637 no_apply = 0; 6638 } 6639 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6640 if ((no_apply == 0) && 6641 ((rsm->r_app_limited == 0) || 6642 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { 6643 tcp_bbr_commit_bw(bbr, cts); 6644 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6645 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6646 } 6647 } 6648 } 6649 6650 static void 6651 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, 6652 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) 6653 { 6654 uint64_t old_rttprop; 6655 6656 /* Update our delivery time and amount */ 6657 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); 6658 bbr->r_ctl.rc_del_time = cts; 6659 if (rtt == 0) { 6660 /* 6661 * 0 means its a retransmit, for now we don't use these for 6662 * the rest of BBR. 6663 */ 6664 return; 6665 } 6666 if ((bbr->rc_use_google == 0) && 6667 (match != BBR_RTT_BY_EXACTMATCH) && 6668 (match != BBR_RTT_BY_TIMESTAMP)){ 6669 /* 6670 * We get a lot of rtt updates, lets not pay attention to 6671 * any that are not an exact match. That way we don't have 6672 * to worry about timestamps and the whole nonsense of 6673 * unsure if its a retransmission etc (if we ever had the 6674 * timestamp fixed to always have the last thing sent this 6675 * would not be a issue). 6676 */ 6677 return; 6678 } 6679 if ((bbr_no_retran && bbr->rc_use_google) && 6680 (match != BBR_RTT_BY_EXACTMATCH) && 6681 (match != BBR_RTT_BY_TIMESTAMP)){ 6682 /* 6683 * We only do measurements in google mode 6684 * with bbr_no_retran on for sure things. 6685 */ 6686 return; 6687 } 6688 /* Only update srtt if we know by exact match */ 6689 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); 6690 if (ack_type == BBR_CUM_ACKED) 6691 bbr->rc_ack_is_cumack = 1; 6692 else 6693 bbr->rc_ack_is_cumack = 0; 6694 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); 6695 /* 6696 * Note the following code differs to the original 6697 * BBR spec. It calls for <= not <. However after a 6698 * long discussion in email with Neal, he acknowledged 6699 * that it should be < than so that we will have flows 6700 * going into probe-rtt (we were seeing cases where that 6701 * did not happen and caused ugly things to occur). We 6702 * have added this agreed upon fix to our code base. 6703 */ 6704 if (rtt < old_rttprop) { 6705 /* Update when we last saw a rtt drop */ 6706 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); 6707 bbr_set_reduced_rtt(bbr, cts, __LINE__); 6708 } 6709 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, 6710 match, rsm->r_start, rsm->r_flags); 6711 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6712 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { 6713 /* 6714 * The RTT-prop moved, reset the target (may be a 6715 * nop for some states). 6716 */ 6717 bbr_set_state_target(bbr, __LINE__); 6718 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) 6719 bbr_log_rtt_shrinks(bbr, cts, 0, 0, 6720 __LINE__, BBR_RTTS_NEW_TARGET, 0); 6721 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) 6722 /* It went up */ 6723 bbr_check_probe_rtt_limits(bbr, cts); 6724 } 6725 if ((bbr->rc_use_google == 0) && 6726 (match == BBR_RTT_BY_TIMESTAMP)) { 6727 /* 6728 * We don't do b/w update with 6729 * these since they are not really 6730 * reliable. 6731 */ 6732 return; 6733 } 6734 if (bbr->r_ctl.r_app_limited_until && 6735 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { 6736 /* We are no longer app-limited */ 6737 bbr->r_ctl.r_app_limited_until = 0; 6738 } 6739 if (bbr->rc_use_google) { 6740 bbr_google_measurement(bbr, rsm, rtt, cts); 6741 } else { 6742 bbr_nf_measurement(bbr, rsm, rtt, cts); 6743 } 6744 } 6745 6746 /* 6747 * Convert a timestamp that the main stack 6748 * uses (milliseconds) into one that bbr uses 6749 * (microseconds). Return that converted timestamp. 6750 */ 6751 static uint32_t 6752 bbr_ts_convert(uint32_t cts) { 6753 uint32_t sec, msec; 6754 6755 sec = cts / MS_IN_USEC; 6756 msec = cts - (MS_IN_USEC * sec); 6757 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); 6758 } 6759 6760 /* 6761 * Return 0 if we did not update the RTT time, return 6762 * 1 if we did. 6763 */ 6764 static int 6765 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, 6766 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) 6767 { 6768 int32_t i; 6769 uint32_t t, uts = 0; 6770 6771 if ((rsm->r_flags & BBR_ACKED) || 6772 (rsm->r_flags & BBR_WAS_RENEGED) || 6773 (rsm->r_flags & BBR_RXT_CLEARED)) { 6774 /* Already done */ 6775 return (0); 6776 } 6777 if (rsm->r_rtt_not_allowed) { 6778 /* Not allowed */ 6779 return (0); 6780 } 6781 if (rsm->r_rtr_cnt == 1) { 6782 /* 6783 * Only one transmit. Hopefully the normal case. 6784 */ 6785 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) 6786 t = cts - rsm->r_tim_lastsent[0]; 6787 else 6788 t = 1; 6789 if ((int)t <= 0) 6790 t = 1; 6791 bbr->r_ctl.rc_last_rtt = t; 6792 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6793 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); 6794 return (1); 6795 } 6796 /* Convert to usecs */ 6797 if ((bbr_can_use_ts_for_rtt == 1) && 6798 (bbr->rc_use_google == 1) && 6799 (ack_type == BBR_CUM_ACKED) && 6800 (to->to_flags & TOF_TS) && 6801 (to->to_tsecr != 0)) { 6802 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; 6803 if (t < 1) 6804 t = 1; 6805 t *= MS_IN_USEC; 6806 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6807 BBR_RTT_BY_TIMESTAMP, 6808 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], 6809 ack_type, to); 6810 return (1); 6811 } 6812 uts = bbr_ts_convert(to->to_tsecr); 6813 if ((to->to_flags & TOF_TS) && 6814 (to->to_tsecr != 0) && 6815 (ack_type == BBR_CUM_ACKED) && 6816 ((rsm->r_flags & BBR_OVERMAX) == 0)) { 6817 /* 6818 * Now which timestamp does it match? In this block the ACK 6819 * may be coming from a previous transmission. 6820 */ 6821 uint32_t fudge; 6822 6823 fudge = BBR_TIMER_FUDGE; 6824 for (i = 0; i < rsm->r_rtr_cnt; i++) { 6825 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && 6826 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { 6827 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6828 t = cts - rsm->r_tim_lastsent[i]; 6829 else 6830 t = 1; 6831 if ((int)t <= 0) 6832 t = 1; 6833 bbr->r_ctl.rc_last_rtt = t; 6834 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, 6835 rsm->r_tim_lastsent[i], ack_type, to); 6836 if ((i + 1) < rsm->r_rtr_cnt) { 6837 /* Likely */ 6838 return (0); 6839 } else if (rsm->r_flags & BBR_TLP) { 6840 bbr->rc_tlp_rtx_out = 0; 6841 } 6842 return (1); 6843 } 6844 } 6845 /* Fall through if we can't find a matching timestamp */ 6846 } 6847 /* 6848 * Ok its a SACK block that we retransmitted. or a windows 6849 * machine without timestamps. We can tell nothing from the 6850 * time-stamp since its not there or the time the peer last 6851 * recieved a segment that moved forward its cum-ack point. 6852 * 6853 * Lets look at the last retransmit and see what we can tell 6854 * (with BBR for space we only keep 2 note we have to keep 6855 * at least 2 so the map can not be condensed more). 6856 */ 6857 i = rsm->r_rtr_cnt - 1; 6858 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6859 t = cts - rsm->r_tim_lastsent[i]; 6860 else 6861 goto not_sure; 6862 if (t < bbr->r_ctl.rc_lowest_rtt) { 6863 /* 6864 * We retransmitted and the ack came back in less 6865 * than the smallest rtt we have observed in the 6866 * windowed rtt. We most likey did an improper 6867 * retransmit as outlined in 4.2 Step 3 point 2 in 6868 * the rack-draft. 6869 * 6870 * Use the prior transmission to update all the 6871 * information as long as there is only one prior 6872 * transmission. 6873 */ 6874 if ((rsm->r_flags & BBR_OVERMAX) == 0) { 6875 #ifdef BBR_INVARIANTS 6876 if (rsm->r_rtr_cnt == 1) 6877 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); 6878 #endif 6879 i = rsm->r_rtr_cnt - 2; 6880 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6881 t = cts - rsm->r_tim_lastsent[i]; 6882 else 6883 t = 1; 6884 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, 6885 rsm->r_tim_lastsent[i], ack_type, to); 6886 return (0); 6887 } else { 6888 /* 6889 * Too many prior transmissions, just 6890 * updated BBR delivered 6891 */ 6892 not_sure: 6893 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6894 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6895 } 6896 } else { 6897 /* 6898 * We retransmitted it and the retransmit did the 6899 * job. 6900 */ 6901 if (rsm->r_flags & BBR_TLP) 6902 bbr->rc_tlp_rtx_out = 0; 6903 if ((rsm->r_flags & BBR_OVERMAX) == 0) 6904 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, 6905 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); 6906 else 6907 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6908 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6909 return (1); 6910 } 6911 return (0); 6912 } 6913 6914 /* 6915 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 6916 */ 6917 static void 6918 bbr_log_sack_passed(struct tcpcb *tp, 6919 struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 6920 { 6921 struct bbr_sendmap *nrsm; 6922 6923 nrsm = rsm; 6924 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, 6925 bbr_head, r_tnext) { 6926 if (nrsm == rsm) { 6927 /* Skip orginal segment he is acked */ 6928 continue; 6929 } 6930 if (nrsm->r_flags & BBR_ACKED) { 6931 /* Skip ack'd segments */ 6932 continue; 6933 } 6934 if (nrsm->r_flags & BBR_SACK_PASSED) { 6935 /* 6936 * We found one that is already marked 6937 * passed, we have been here before and 6938 * so all others below this are marked. 6939 */ 6940 break; 6941 } 6942 BBR_STAT_INC(bbr_sack_passed); 6943 nrsm->r_flags |= BBR_SACK_PASSED; 6944 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && 6945 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { 6946 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; 6947 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; 6948 nrsm->r_flags |= BBR_MARKED_LOST; 6949 } 6950 nrsm->r_flags &= ~BBR_WAS_SACKPASS; 6951 } 6952 } 6953 6954 /* 6955 * Returns the number of bytes that were 6956 * newly ack'd by sack blocks. 6957 */ 6958 static uint32_t 6959 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, 6960 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) 6961 { 6962 int32_t times = 0; 6963 uint32_t start, end, changed = 0; 6964 struct bbr_sendmap *rsm, *nrsm; 6965 int32_t used_ref = 1; 6966 uint8_t went_back = 0, went_fwd = 0; 6967 6968 start = sack->start; 6969 end = sack->end; 6970 rsm = *prsm; 6971 if (rsm == NULL) 6972 used_ref = 0; 6973 6974 /* Do we locate the block behind where we last were? */ 6975 if (rsm && SEQ_LT(start, rsm->r_start)) { 6976 went_back = 1; 6977 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 6978 if (SEQ_GEQ(start, rsm->r_start) && 6979 SEQ_LT(start, rsm->r_end)) { 6980 goto do_rest_ofb; 6981 } 6982 } 6983 } 6984 start_at_beginning: 6985 went_fwd = 1; 6986 /* 6987 * Ok lets locate the block where this guy is fwd from rsm (if its 6988 * set) 6989 */ 6990 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { 6991 if (SEQ_GEQ(start, rsm->r_start) && 6992 SEQ_LT(start, rsm->r_end)) { 6993 break; 6994 } 6995 } 6996 do_rest_ofb: 6997 if (rsm == NULL) { 6998 /* 6999 * This happens when we get duplicate sack blocks with the 7000 * same end. For example SACK 4: 100 SACK 3: 100 The sort 7001 * will not change there location so we would just start at 7002 * the end of the first one and get lost. 7003 */ 7004 if (tp->t_flags & TF_SENTFIN) { 7005 /* 7006 * Check to see if we have not logged the FIN that 7007 * went out. 7008 */ 7009 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7010 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 7011 /* 7012 * Ok we did not get the FIN logged. 7013 */ 7014 nrsm->r_end++; 7015 rsm = nrsm; 7016 goto do_rest_ofb; 7017 } 7018 } 7019 if (times == 1) { 7020 #ifdef BBR_INVARIANTS 7021 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", 7022 tp, bbr, sack, to, prsm); 7023 #else 7024 goto out; 7025 #endif 7026 } 7027 times++; 7028 BBR_STAT_INC(bbr_sack_proc_restart); 7029 rsm = NULL; 7030 goto start_at_beginning; 7031 } 7032 /* Ok we have an ACK for some piece of rsm */ 7033 if (rsm->r_start != start) { 7034 /* 7035 * Need to split this in two pieces the before and after. 7036 */ 7037 if (bbr_sack_mergable(rsm, start, end)) 7038 nrsm = bbr_alloc_full_limit(bbr); 7039 else 7040 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7041 if (nrsm == NULL) { 7042 /* We could not allocate ignore the sack */ 7043 struct sackblk blk; 7044 7045 blk.start = start; 7046 blk.end = end; 7047 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7048 goto out; 7049 } 7050 bbr_clone_rsm(bbr, nrsm, rsm, start); 7051 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7052 if (rsm->r_in_tmap) { 7053 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7054 nrsm->r_in_tmap = 1; 7055 } 7056 rsm->r_flags &= (~BBR_HAS_FIN); 7057 rsm = nrsm; 7058 } 7059 if (SEQ_GEQ(end, rsm->r_end)) { 7060 /* 7061 * The end of this block is either beyond this guy or right 7062 * at this guy. 7063 */ 7064 if ((rsm->r_flags & BBR_ACKED) == 0) { 7065 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7066 changed += (rsm->r_end - rsm->r_start); 7067 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7068 bbr_log_sack_passed(tp, bbr, rsm); 7069 if (rsm->r_flags & BBR_MARKED_LOST) { 7070 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7071 } 7072 /* Is Reordering occuring? */ 7073 if (rsm->r_flags & BBR_SACK_PASSED) { 7074 BBR_STAT_INC(bbr_reorder_seen); 7075 bbr->r_ctl.rc_reorder_ts = cts; 7076 if (rsm->r_flags & BBR_MARKED_LOST) { 7077 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7078 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7079 /* LT sampling also needs adjustment */ 7080 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7081 } 7082 } 7083 rsm->r_flags |= BBR_ACKED; 7084 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7085 if (rsm->r_in_tmap) { 7086 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7087 rsm->r_in_tmap = 0; 7088 } 7089 } 7090 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7091 if (end == rsm->r_end) { 7092 /* This block only - done */ 7093 goto out; 7094 } 7095 /* There is more not coverend by this rsm move on */ 7096 start = rsm->r_end; 7097 nrsm = TAILQ_NEXT(rsm, r_next); 7098 rsm = nrsm; 7099 times = 0; 7100 goto do_rest_ofb; 7101 } 7102 if (rsm->r_flags & BBR_ACKED) { 7103 /* Been here done that */ 7104 goto out; 7105 } 7106 /* Ok we need to split off this one at the tail */ 7107 if (bbr_sack_mergable(rsm, start, end)) 7108 nrsm = bbr_alloc_full_limit(bbr); 7109 else 7110 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7111 if (nrsm == NULL) { 7112 /* failed XXXrrs what can we do but loose the sack info? */ 7113 struct sackblk blk; 7114 7115 blk.start = start; 7116 blk.end = end; 7117 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7118 goto out; 7119 } 7120 /* Clone it */ 7121 bbr_clone_rsm(bbr, nrsm, rsm, end); 7122 /* The sack block does not cover this guy fully */ 7123 rsm->r_flags &= (~BBR_HAS_FIN); 7124 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7125 if (rsm->r_in_tmap) { 7126 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7127 nrsm->r_in_tmap = 1; 7128 } 7129 nrsm->r_dupack = 0; 7130 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7131 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7132 changed += (rsm->r_end - rsm->r_start); 7133 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7134 bbr_log_sack_passed(tp, bbr, rsm); 7135 /* Is Reordering occuring? */ 7136 if (rsm->r_flags & BBR_MARKED_LOST) { 7137 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7138 } 7139 if (rsm->r_flags & BBR_SACK_PASSED) { 7140 BBR_STAT_INC(bbr_reorder_seen); 7141 bbr->r_ctl.rc_reorder_ts = cts; 7142 if (rsm->r_flags & BBR_MARKED_LOST) { 7143 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7144 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7145 /* LT sampling also needs adjustment */ 7146 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7147 } 7148 } 7149 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7150 rsm->r_flags |= BBR_ACKED; 7151 if (rsm->r_in_tmap) { 7152 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7153 rsm->r_in_tmap = 0; 7154 } 7155 out: 7156 if (rsm && (rsm->r_flags & BBR_ACKED)) { 7157 /* 7158 * Now can we merge this newly acked 7159 * block with either the previous or 7160 * next block? 7161 */ 7162 nrsm = TAILQ_NEXT(rsm, r_next); 7163 if (nrsm && 7164 (nrsm->r_flags & BBR_ACKED)) { 7165 /* yep this and next can be merged */ 7166 rsm = bbr_merge_rsm(bbr, rsm, nrsm); 7167 } 7168 /* Now what about the previous? */ 7169 nrsm = TAILQ_PREV(rsm, bbr_head, r_next); 7170 if (nrsm && 7171 (nrsm->r_flags & BBR_ACKED)) { 7172 /* yep the previous and this can be merged */ 7173 rsm = bbr_merge_rsm(bbr, nrsm, rsm); 7174 } 7175 } 7176 if (used_ref == 0) { 7177 BBR_STAT_INC(bbr_sack_proc_all); 7178 } else { 7179 BBR_STAT_INC(bbr_sack_proc_short); 7180 } 7181 if (went_fwd && went_back) { 7182 BBR_STAT_INC(bbr_sack_search_both); 7183 } else if (went_fwd) { 7184 BBR_STAT_INC(bbr_sack_search_fwd); 7185 } else if (went_back) { 7186 BBR_STAT_INC(bbr_sack_search_back); 7187 } 7188 /* Save off where the next seq is */ 7189 if (rsm) 7190 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 7191 else 7192 bbr->r_ctl.rc_sacklast = NULL; 7193 *prsm = rsm; 7194 return (changed); 7195 } 7196 7197 static void inline 7198 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) 7199 { 7200 struct bbr_sendmap *tmap; 7201 7202 BBR_STAT_INC(bbr_reneges_seen); 7203 tmap = NULL; 7204 while (rsm && (rsm->r_flags & BBR_ACKED)) { 7205 /* Its no longer sacked, mark it so */ 7206 uint32_t oflags; 7207 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7208 #ifdef BBR_INVARIANTS 7209 if (rsm->r_in_tmap) { 7210 panic("bbr:%p rsm:%p flags:0x%x in tmap?", 7211 bbr, rsm, rsm->r_flags); 7212 } 7213 #endif 7214 oflags = rsm->r_flags; 7215 if (rsm->r_flags & BBR_MARKED_LOST) { 7216 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7217 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7218 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7219 /* LT sampling also needs adjustment */ 7220 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7221 } 7222 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); 7223 rsm->r_flags |= BBR_WAS_RENEGED; 7224 rsm->r_flags |= BBR_RXT_CLEARED; 7225 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); 7226 /* Rebuild it into our tmap */ 7227 if (tmap == NULL) { 7228 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7229 tmap = rsm; 7230 } else { 7231 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); 7232 tmap = rsm; 7233 } 7234 tmap->r_in_tmap = 1; 7235 /* 7236 * XXXrrs Delivered? Should we do anything here? 7237 * 7238 * Of course we don't on a rxt timeout so maybe its ok that 7239 * we don't? 7240 * 7241 * For now lets not. 7242 */ 7243 rsm = TAILQ_NEXT(rsm, r_next); 7244 } 7245 /* 7246 * Now lets possibly clear the sack filter so we start recognizing 7247 * sacks that cover this area. 7248 */ 7249 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); 7250 } 7251 7252 static void 7253 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) 7254 { 7255 struct tcp_bbr *bbr; 7256 struct bbr_sendmap *rsm; 7257 uint32_t cts; 7258 7259 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7260 cts = bbr->r_ctl.rc_rcvtime; 7261 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7262 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { 7263 if ((rsm->r_end - rsm->r_start) <= 1) { 7264 /* Log out the SYN completely */ 7265 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7266 rsm->r_rtr_bytes = 0; 7267 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7268 if (rsm->r_in_tmap) { 7269 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7270 rsm->r_in_tmap = 0; 7271 } 7272 if (bbr->r_ctl.rc_next == rsm) { 7273 /* scoot along the marker */ 7274 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7275 } 7276 if (to != NULL) 7277 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); 7278 bbr_free(bbr, rsm); 7279 } else { 7280 /* There is more (Fast open)? strip out SYN. */ 7281 rsm->r_flags &= ~BBR_HAS_SYN; 7282 rsm->r_start++; 7283 } 7284 } 7285 } 7286 7287 /* 7288 * Returns the number of bytes that were 7289 * acknowledged by SACK blocks. 7290 */ 7291 7292 static uint32_t 7293 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, 7294 uint32_t *prev_acked) 7295 { 7296 uint32_t changed, last_seq, entered_recovery = 0; 7297 struct tcp_bbr *bbr; 7298 struct bbr_sendmap *rsm; 7299 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 7300 register uint32_t th_ack; 7301 int32_t i, j, k, new_sb, num_sack_blks = 0; 7302 uint32_t cts, acked, ack_point, sack_changed = 0; 7303 uint32_t p_maxseg, maxseg, p_acked = 0; 7304 7305 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7306 if (tcp_get_flags(th) & TH_RST) { 7307 /* We don't log resets */ 7308 return (0); 7309 } 7310 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7311 cts = bbr->r_ctl.rc_rcvtime; 7312 7313 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7314 changed = 0; 7315 maxseg = tp->t_maxseg - bbr->rc_last_options; 7316 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); 7317 th_ack = th->th_ack; 7318 if (SEQ_GT(th_ack, tp->snd_una)) { 7319 acked = th_ack - tp->snd_una; 7320 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); 7321 bbr->rc_tp->t_acktime = ticks; 7322 } else 7323 acked = 0; 7324 if (SEQ_LEQ(th_ack, tp->snd_una)) { 7325 /* Only sent here for sack processing */ 7326 goto proc_sack; 7327 } 7328 if (rsm && SEQ_GT(th_ack, rsm->r_start)) { 7329 changed = th_ack - rsm->r_start; 7330 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { 7331 /* 7332 * For the SYN incoming case we will not have called 7333 * tcp_output for the sending of the SYN, so there will be 7334 * no map. All other cases should probably be a panic. 7335 */ 7336 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { 7337 /* 7338 * We have a timestamp that can be used to generate 7339 * an initial RTT. 7340 */ 7341 uint32_t ts, now, rtt; 7342 7343 ts = bbr_ts_convert(to->to_tsecr); 7344 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); 7345 rtt = now - ts; 7346 if (rtt < 1) 7347 rtt = 1; 7348 bbr_log_type_bbrrttprop(bbr, rtt, 7349 tp->iss, 0, cts, 7350 BBR_RTT_BY_TIMESTAMP, tp->iss, 0); 7351 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 7352 changed = 1; 7353 bbr->r_wanted_output = 1; 7354 goto out; 7355 } 7356 goto proc_sack; 7357 } else if (rsm == NULL) { 7358 goto out; 7359 } 7360 if (changed) { 7361 /* 7362 * The ACK point is advancing to th_ack, we must drop off 7363 * the packets in the rack log and calculate any eligble 7364 * RTT's. 7365 */ 7366 bbr->r_wanted_output = 1; 7367 more: 7368 if (rsm == NULL) { 7369 if (tp->t_flags & TF_SENTFIN) { 7370 /* if we send a FIN we will not hav a map */ 7371 goto proc_sack; 7372 } 7373 #ifdef BBR_INVARIANTS 7374 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", 7375 tp, 7376 th, tp->t_state, bbr, 7377 tp->snd_una, tp->snd_max, changed); 7378 #endif 7379 goto proc_sack; 7380 } 7381 } 7382 if (SEQ_LT(th_ack, rsm->r_start)) { 7383 /* Huh map is missing this */ 7384 #ifdef BBR_INVARIANTS 7385 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", 7386 rsm->r_start, 7387 th_ack, tp->t_state, 7388 bbr->r_state, bbr); 7389 panic("th-ack is bad bbr:%p tp:%p", bbr, tp); 7390 #endif 7391 goto proc_sack; 7392 } else if (th_ack == rsm->r_start) { 7393 /* None here to ack */ 7394 goto proc_sack; 7395 } 7396 /* 7397 * Clear the dup ack counter, it will 7398 * either be freed or if there is some 7399 * remaining we need to start it at zero. 7400 */ 7401 rsm->r_dupack = 0; 7402 /* Now do we consume the whole thing? */ 7403 if (SEQ_GEQ(th_ack, rsm->r_end)) { 7404 /* Its all consumed. */ 7405 uint32_t left; 7406 7407 if (rsm->r_flags & BBR_ACKED) { 7408 /* 7409 * It was acked on the scoreboard -- remove it from 7410 * total 7411 */ 7412 p_acked += (rsm->r_end - rsm->r_start); 7413 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7414 if (bbr->r_ctl.rc_sacked == 0) 7415 bbr->r_ctl.rc_sacklast = NULL; 7416 } else { 7417 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); 7418 if (rsm->r_flags & BBR_MARKED_LOST) { 7419 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7420 } 7421 if (rsm->r_flags & BBR_SACK_PASSED) { 7422 /* 7423 * There are acked segments ACKED on the 7424 * scoreboard further up. We are seeing 7425 * reordering. 7426 */ 7427 BBR_STAT_INC(bbr_reorder_seen); 7428 bbr->r_ctl.rc_reorder_ts = cts; 7429 if (rsm->r_flags & BBR_MARKED_LOST) { 7430 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7431 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7432 /* LT sampling also needs adjustment */ 7433 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7434 } 7435 } 7436 rsm->r_flags &= ~BBR_MARKED_LOST; 7437 } 7438 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7439 rsm->r_rtr_bytes = 0; 7440 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7441 if (rsm->r_in_tmap) { 7442 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7443 rsm->r_in_tmap = 0; 7444 } 7445 if (bbr->r_ctl.rc_next == rsm) { 7446 /* scoot along the marker */ 7447 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7448 } 7449 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7450 /* Adjust the packet counts */ 7451 left = th_ack - rsm->r_end; 7452 /* Free back to zone */ 7453 bbr_free(bbr, rsm); 7454 if (left) { 7455 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7456 goto more; 7457 } 7458 goto proc_sack; 7459 } 7460 if (rsm->r_flags & BBR_ACKED) { 7461 /* 7462 * It was acked on the scoreboard -- remove it from total 7463 * for the part being cum-acked. 7464 */ 7465 p_acked += (rsm->r_end - rsm->r_start); 7466 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 7467 if (bbr->r_ctl.rc_sacked == 0) 7468 bbr->r_ctl.rc_sacklast = NULL; 7469 } else { 7470 /* 7471 * It was acked up to th_ack point for the first time 7472 */ 7473 struct bbr_sendmap lrsm; 7474 7475 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); 7476 lrsm.r_end = th_ack; 7477 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); 7478 } 7479 if ((rsm->r_flags & BBR_MARKED_LOST) && 7480 ((rsm->r_flags & BBR_ACKED) == 0)) { 7481 /* 7482 * It was marked lost and partly ack'd now 7483 * for the first time. We lower the rc_lost_bytes 7484 * and still leave it MARKED. 7485 */ 7486 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; 7487 } 7488 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7489 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7490 rsm->r_rtr_bytes = 0; 7491 /* adjust packet count */ 7492 rsm->r_start = th_ack; 7493 proc_sack: 7494 /* Check for reneging */ 7495 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7496 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { 7497 /* 7498 * The peer has moved snd_una up to the edge of this send, 7499 * i.e. one that it had previously acked. The only way that 7500 * can be true if the peer threw away data (space issues) 7501 * that it had previously sacked (else it would have given 7502 * us snd_una up to (rsm->r_end). We need to undo the acked 7503 * markings here. 7504 * 7505 * Note we have to look to make sure th_ack is our 7506 * rsm->r_start in case we get an old ack where th_ack is 7507 * behind snd_una. 7508 */ 7509 bbr_peer_reneges(bbr, rsm, th->th_ack); 7510 } 7511 if ((to->to_flags & TOF_SACK) == 0) { 7512 /* We are done nothing left to log */ 7513 goto out; 7514 } 7515 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7516 if (rsm) { 7517 last_seq = rsm->r_end; 7518 } else { 7519 last_seq = tp->snd_max; 7520 } 7521 /* Sack block processing */ 7522 if (SEQ_GT(th_ack, tp->snd_una)) 7523 ack_point = th_ack; 7524 else 7525 ack_point = tp->snd_una; 7526 for (i = 0; i < to->to_nsacks; i++) { 7527 bcopy((to->to_sacks + i * TCPOLEN_SACK), 7528 &sack, sizeof(sack)); 7529 sack.start = ntohl(sack.start); 7530 sack.end = ntohl(sack.end); 7531 if (SEQ_GT(sack.end, sack.start) && 7532 SEQ_GT(sack.start, ack_point) && 7533 SEQ_LT(sack.start, tp->snd_max) && 7534 SEQ_GT(sack.end, ack_point) && 7535 SEQ_LEQ(sack.end, tp->snd_max)) { 7536 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && 7537 (SEQ_LT(sack.end, last_seq)) && 7538 ((sack.end - sack.start) < (p_maxseg / 8))) { 7539 /* 7540 * Not the last piece and its smaller than 7541 * 1/8th of a p_maxseg. We ignore this. 7542 */ 7543 BBR_STAT_INC(bbr_runt_sacks); 7544 continue; 7545 } 7546 sack_blocks[num_sack_blks] = sack; 7547 num_sack_blks++; 7548 } else if (SEQ_LEQ(sack.start, th_ack) && 7549 SEQ_LEQ(sack.end, th_ack)) { 7550 /* 7551 * Its a D-SACK block. 7552 */ 7553 tcp_record_dsack(tp, sack.start, sack.end, 0); 7554 } 7555 } 7556 if (num_sack_blks == 0) 7557 goto out; 7558 /* 7559 * Sort the SACK blocks so we can update the rack scoreboard with 7560 * just one pass. 7561 */ 7562 new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, 7563 num_sack_blks, th->th_ack); 7564 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); 7565 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); 7566 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); 7567 num_sack_blks = new_sb; 7568 if (num_sack_blks < 2) { 7569 goto do_sack_work; 7570 } 7571 /* Sort the sacks */ 7572 for (i = 0; i < num_sack_blks; i++) { 7573 for (j = i + 1; j < num_sack_blks; j++) { 7574 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 7575 sack = sack_blocks[i]; 7576 sack_blocks[i] = sack_blocks[j]; 7577 sack_blocks[j] = sack; 7578 } 7579 } 7580 } 7581 /* 7582 * Now are any of the sack block ends the same (yes some 7583 * implememtations send these)? 7584 */ 7585 again: 7586 if (num_sack_blks > 1) { 7587 for (i = 0; i < num_sack_blks; i++) { 7588 for (j = i + 1; j < num_sack_blks; j++) { 7589 if (sack_blocks[i].end == sack_blocks[j].end) { 7590 /* 7591 * Ok these two have the same end we 7592 * want the smallest end and then 7593 * throw away the larger and start 7594 * again. 7595 */ 7596 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 7597 /* 7598 * The second block covers 7599 * more area use that 7600 */ 7601 sack_blocks[i].start = sack_blocks[j].start; 7602 } 7603 /* 7604 * Now collapse out the dup-sack and 7605 * lower the count 7606 */ 7607 for (k = (j + 1); k < num_sack_blks; k++) { 7608 sack_blocks[j].start = sack_blocks[k].start; 7609 sack_blocks[j].end = sack_blocks[k].end; 7610 j++; 7611 } 7612 num_sack_blks--; 7613 goto again; 7614 } 7615 } 7616 } 7617 } 7618 do_sack_work: 7619 rsm = bbr->r_ctl.rc_sacklast; 7620 for (i = 0; i < num_sack_blks; i++) { 7621 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); 7622 if (acked) { 7623 bbr->r_wanted_output = 1; 7624 changed += acked; 7625 sack_changed += acked; 7626 } 7627 } 7628 out: 7629 *prev_acked = p_acked; 7630 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 7631 /* 7632 * Ok we have a high probability that we need to go in to 7633 * recovery since we have data sack'd 7634 */ 7635 struct bbr_sendmap *rsm; 7636 7637 rsm = bbr_check_recovery_mode(tp, bbr, cts); 7638 if (rsm) { 7639 /* Enter recovery */ 7640 entered_recovery = 1; 7641 bbr->r_wanted_output = 1; 7642 /* 7643 * When we enter recovery we need to assure we send 7644 * one packet. 7645 */ 7646 if (bbr->r_ctl.rc_resend == NULL) { 7647 bbr->r_ctl.rc_resend = rsm; 7648 } 7649 } 7650 } 7651 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 7652 /* 7653 * See if we need to rack-retransmit anything if so set it 7654 * up as the thing to resend assuming something else is not 7655 * already in that position. 7656 */ 7657 if (bbr->r_ctl.rc_resend == NULL) { 7658 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 7659 } 7660 } 7661 /* 7662 * We return the amount that changed via sack, this is used by the 7663 * ack-received code to augment what was changed between th_ack <-> 7664 * snd_una. 7665 */ 7666 return (sack_changed); 7667 } 7668 7669 static void 7670 bbr_strike_dupack(struct tcp_bbr *bbr) 7671 { 7672 struct bbr_sendmap *rsm; 7673 7674 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 7675 if (rsm && (rsm->r_dupack < 0xff)) { 7676 rsm->r_dupack++; 7677 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) 7678 bbr->r_wanted_output = 1; 7679 } 7680 } 7681 7682 /* 7683 * Return value of 1, we do not need to call bbr_process_data(). 7684 * return value of 0, bbr_process_data can be called. 7685 * For ret_val if its 0 the TCB is locked and valid, if its non-zero 7686 * its unlocked and probably unsafe to touch the TCB. 7687 */ 7688 static int 7689 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 7690 struct tcpcb *tp, struct tcpopt *to, 7691 uint32_t tiwin, int32_t tlen, 7692 int32_t * ofia, int32_t thflags, int32_t * ret_val) 7693 { 7694 int32_t ourfinisacked = 0; 7695 int32_t acked_amount; 7696 uint16_t nsegs; 7697 int32_t acked; 7698 uint32_t lost, sack_changed = 0; 7699 struct mbuf *mfree; 7700 struct tcp_bbr *bbr; 7701 uint32_t prev_acked = 0; 7702 7703 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7704 lost = bbr->r_ctl.rc_lost; 7705 nsegs = max(1, m->m_pkthdr.lro_nsegs); 7706 if (SEQ_GT(th->th_ack, tp->snd_max)) { 7707 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 7708 bbr->r_wanted_output = 1; 7709 return (1); 7710 } 7711 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 7712 /* Process the ack */ 7713 if (bbr->rc_in_persist) 7714 tp->t_rxtshift = 0; 7715 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) 7716 bbr_strike_dupack(bbr); 7717 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 7718 } 7719 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); 7720 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 7721 /* 7722 * Old ack, behind the last one rcv'd or a duplicate ack 7723 * with SACK info. 7724 */ 7725 if (th->th_ack == tp->snd_una) { 7726 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); 7727 if (bbr->r_state == TCPS_SYN_SENT) { 7728 /* 7729 * Special case on where we sent SYN. When 7730 * the SYN-ACK is processed in syn_sent 7731 * state it bumps the snd_una. This causes 7732 * us to hit here even though we did ack 1 7733 * byte. 7734 * 7735 * Go through the nothing left case so we 7736 * send data. 7737 */ 7738 goto nothing_left; 7739 } 7740 } 7741 return (0); 7742 } 7743 /* 7744 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 7745 * something we sent. 7746 */ 7747 if (tp->t_flags & TF_NEEDSYN) { 7748 /* 7749 * T/TCP: Connection was half-synchronized, and our SYN has 7750 * been ACK'd (so connection is now fully synchronized). Go 7751 * to non-starred state, increment snd_una for ACK of SYN, 7752 * and check if we can do window scaling. 7753 */ 7754 tp->t_flags &= ~TF_NEEDSYN; 7755 tp->snd_una++; 7756 /* Do window scaling? */ 7757 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 7758 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 7759 tp->rcv_scale = tp->request_r_scale; 7760 /* Send window already scaled. */ 7761 } 7762 } 7763 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7764 7765 acked = BYTES_THIS_ACK(tp, th); 7766 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 7767 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 7768 7769 /* 7770 * If we just performed our first retransmit, and the ACK arrives 7771 * within our recovery window, then it was a mistake to do the 7772 * retransmit in the first place. Recover our original cwnd and 7773 * ssthresh, and proceed to transmit where we left off. 7774 */ 7775 if (tp->t_flags & TF_PREVVALID) { 7776 tp->t_flags &= ~TF_PREVVALID; 7777 if (tp->t_rxtshift == 1 && 7778 (int)(ticks - tp->t_badrxtwin) < 0) 7779 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 7780 } 7781 SOCKBUF_LOCK(&so->so_snd); 7782 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 7783 tp->snd_wnd -= acked_amount; 7784 mfree = sbcut_locked(&so->so_snd, acked_amount); 7785 /* NB: sowwakeup_locked() does an implicit unlock. */ 7786 sowwakeup_locked(so); 7787 m_freem(mfree); 7788 if (SEQ_GT(th->th_ack, tp->snd_una)) { 7789 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 7790 } 7791 tp->snd_una = th->th_ack; 7792 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); 7793 if (IN_RECOVERY(tp->t_flags)) { 7794 if (SEQ_LT(th->th_ack, tp->snd_recover) && 7795 (SEQ_LT(th->th_ack, tp->snd_max))) { 7796 tcp_bbr_partialack(tp); 7797 } else { 7798 bbr_post_recovery(tp); 7799 } 7800 } 7801 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 7802 tp->snd_recover = tp->snd_una; 7803 } 7804 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 7805 tp->snd_nxt = tp->snd_max; 7806 } 7807 if (tp->snd_una == tp->snd_max) { 7808 /* Nothing left outstanding */ 7809 nothing_left: 7810 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 7811 if (sbavail(&so->so_snd) == 0) 7812 bbr->rc_tp->t_acktime = 0; 7813 if ((sbused(&so->so_snd) == 0) && 7814 (tp->t_flags & TF_SENTFIN)) { 7815 ourfinisacked = 1; 7816 } 7817 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 7818 if (bbr->rc_in_persist == 0) { 7819 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 7820 } 7821 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 7822 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 7823 /* 7824 * We invalidate the last ack here since we 7825 * don't want to transfer forward the time 7826 * for our sum's calculations. 7827 */ 7828 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 7829 (sbavail(&so->so_snd) == 0) && 7830 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 7831 /* 7832 * The socket was gone and the peer sent data, time 7833 * to reset him. 7834 */ 7835 *ret_val = 1; 7836 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 7837 /* tcp_close will kill the inp pre-log the Reset */ 7838 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 7839 tp = tcp_close(tp); 7840 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 7841 BBR_STAT_INC(bbr_dropped_af_data); 7842 return (1); 7843 } 7844 /* Set need output so persist might get set */ 7845 bbr->r_wanted_output = 1; 7846 } 7847 if (ofia) 7848 *ofia = ourfinisacked; 7849 return (0); 7850 } 7851 7852 static void 7853 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7854 { 7855 if (bbr->rc_in_persist == 0) { 7856 bbr_timer_cancel(bbr, __LINE__, cts); 7857 bbr->r_ctl.rc_last_delay_val = 0; 7858 tp->t_rxtshift = 0; 7859 bbr->rc_in_persist = 1; 7860 bbr->r_ctl.rc_went_idle_time = cts; 7861 /* We should be capped when rw went to 0 but just in case */ 7862 bbr_log_type_pesist(bbr, cts, 0, line, 1); 7863 /* Time freezes for the state, so do the accounting now */ 7864 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 7865 uint32_t time_in; 7866 7867 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 7868 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7869 int32_t idx; 7870 7871 idx = bbr_state_val(bbr); 7872 counter_u64_add(bbr_state_time[(idx + 5)], time_in); 7873 } else { 7874 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 7875 } 7876 } 7877 bbr->r_ctl.rc_bbr_state_time = cts; 7878 } 7879 } 7880 7881 static void 7882 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) 7883 { 7884 /* 7885 * Note that if idle time does not exceed our 7886 * threshold, we do nothing continuing the state 7887 * transitions we were last walking through. 7888 */ 7889 if (idle_time >= bbr_idle_restart_threshold) { 7890 if (bbr->rc_use_idle_restart) { 7891 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; 7892 /* 7893 * Set our target using BBR_UNIT, so 7894 * we increase at a dramatic rate but 7895 * we stop when we get the pipe 7896 * full again for our current b/w estimate. 7897 */ 7898 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 7899 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 7900 bbr_set_state_target(bbr, __LINE__); 7901 /* Now setup our gains to ramp up */ 7902 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 7903 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 7904 bbr_log_type_statechange(bbr, cts, __LINE__); 7905 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7906 bbr_substate_change(bbr, cts, __LINE__, 1); 7907 } 7908 } 7909 } 7910 7911 static void 7912 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7913 { 7914 uint32_t idle_time; 7915 7916 if (bbr->rc_in_persist == 0) 7917 return; 7918 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); 7919 bbr->rc_in_persist = 0; 7920 bbr->rc_hit_state_1 = 0; 7921 bbr->r_ctl.rc_del_time = cts; 7922 /* 7923 * We invalidate the last ack here since we 7924 * don't want to transfer forward the time 7925 * for our sum's calculations. 7926 */ 7927 if (tcp_in_hpts(bbr->rc_inp)) { 7928 tcp_hpts_remove(bbr->rc_inp); 7929 bbr->rc_timer_first = 0; 7930 bbr->r_ctl.rc_hpts_flags = 0; 7931 bbr->r_ctl.rc_last_delay_val = 0; 7932 bbr->r_ctl.rc_hptsi_agg_delay = 0; 7933 bbr->r_agg_early_set = 0; 7934 bbr->r_ctl.rc_agg_early = 0; 7935 } 7936 bbr_log_type_pesist(bbr, cts, idle_time, line, 0); 7937 if (idle_time >= bbr_rtt_probe_time) { 7938 /* 7939 * This qualifies as a RTT_PROBE session since we drop the 7940 * data outstanding to nothing and waited more than 7941 * bbr_rtt_probe_time. 7942 */ 7943 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); 7944 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; 7945 } 7946 tp->t_rxtshift = 0; 7947 /* 7948 * If in probeBW and we have persisted more than an RTT lets do 7949 * special handling. 7950 */ 7951 /* Force a time based epoch */ 7952 bbr_set_epoch(bbr, cts, __LINE__); 7953 /* 7954 * Setup the lost so we don't count anything against the guy 7955 * we have been stuck with during persists. 7956 */ 7957 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 7958 /* Time un-freezes for the state */ 7959 bbr->r_ctl.rc_bbr_state_time = cts; 7960 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || 7961 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { 7962 /* 7963 * If we are going back to probe-bw 7964 * or probe_rtt, we may need to possibly 7965 * do a fast restart. 7966 */ 7967 bbr_restart_after_idle(bbr, cts, idle_time); 7968 } 7969 } 7970 7971 static void 7972 bbr_collapsed_window(struct tcp_bbr *bbr) 7973 { 7974 /* 7975 * Now we must walk the 7976 * send map and divide the 7977 * ones left stranded. These 7978 * guys can't cause us to abort 7979 * the connection and are really 7980 * "unsent". However if a buggy 7981 * client actually did keep some 7982 * of the data i.e. collapsed the win 7983 * and refused to ack and then opened 7984 * the win and acked that data. We would 7985 * get into an ack war, the simplier 7986 * method then of just pretending we 7987 * did not send those segments something 7988 * won't work. 7989 */ 7990 struct bbr_sendmap *rsm, *nrsm; 7991 tcp_seq max_seq; 7992 uint32_t maxseg; 7993 int can_split = 0; 7994 int fnd = 0; 7995 7996 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 7997 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; 7998 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); 7999 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 8000 /* Find the first seq past or at maxseq */ 8001 if (rsm->r_flags & BBR_RWND_COLLAPSED) 8002 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8003 if (SEQ_GEQ(max_seq, rsm->r_start) && 8004 SEQ_GEQ(rsm->r_end, max_seq)) { 8005 fnd = 1; 8006 break; 8007 } 8008 } 8009 bbr->rc_has_collapsed = 0; 8010 if (!fnd) { 8011 /* Nothing to do strange */ 8012 return; 8013 } 8014 /* 8015 * Now can we split? 8016 * 8017 * We don't want to split if splitting 8018 * would generate too many small segments 8019 * less we let an attacker fragment our 8020 * send_map and leave us out of memory. 8021 */ 8022 if ((max_seq != rsm->r_start) && 8023 (max_seq != rsm->r_end)){ 8024 /* can we split? */ 8025 int res1, res2; 8026 8027 res1 = max_seq - rsm->r_start; 8028 res2 = rsm->r_end - max_seq; 8029 if ((res1 >= (maxseg/8)) && 8030 (res2 >= (maxseg/8))) { 8031 /* No small pieces here */ 8032 can_split = 1; 8033 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { 8034 /* We are under the limit */ 8035 can_split = 1; 8036 } 8037 } 8038 /* Ok do we need to split this rsm? */ 8039 if (max_seq == rsm->r_start) { 8040 /* It's this guy no split required */ 8041 nrsm = rsm; 8042 } else if (max_seq == rsm->r_end) { 8043 /* It's the next one no split required. */ 8044 nrsm = TAILQ_NEXT(rsm, r_next); 8045 if (nrsm == NULL) { 8046 /* Huh? */ 8047 return; 8048 } 8049 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { 8050 /* yep we need to split it */ 8051 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 8052 if (nrsm == NULL) { 8053 /* failed XXXrrs what can we do mark the whole? */ 8054 nrsm = rsm; 8055 goto no_split; 8056 } 8057 /* Clone it */ 8058 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); 8059 bbr_clone_rsm(bbr, nrsm, rsm, max_seq); 8060 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 8061 if (rsm->r_in_tmap) { 8062 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 8063 nrsm->r_in_tmap = 1; 8064 } 8065 } else { 8066 /* 8067 * Split not allowed just start here just 8068 * use this guy. 8069 */ 8070 nrsm = rsm; 8071 } 8072 no_split: 8073 BBR_STAT_INC(bbr_collapsed_win); 8074 /* reuse fnd as a count */ 8075 fnd = 0; 8076 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { 8077 nrsm->r_flags |= BBR_RWND_COLLAPSED; 8078 fnd++; 8079 bbr->rc_has_collapsed = 1; 8080 } 8081 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); 8082 } 8083 8084 static void 8085 bbr_un_collapse_window(struct tcp_bbr *bbr) 8086 { 8087 struct bbr_sendmap *rsm; 8088 int cleared = 0; 8089 8090 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 8091 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 8092 /* Clear the flag */ 8093 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8094 cleared++; 8095 } else 8096 break; 8097 } 8098 bbr_log_type_rwnd_collapse(bbr, 8099 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); 8100 bbr->rc_has_collapsed = 0; 8101 } 8102 8103 /* 8104 * Return value of 1, the TCB is unlocked and most 8105 * likely gone, return value of 0, the TCB is still 8106 * locked. 8107 */ 8108 static int 8109 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 8110 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 8111 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 8112 { 8113 /* 8114 * Update window information. Don't look at window if no ACK: TAC's 8115 * send garbage on first SYN. 8116 */ 8117 uint16_t nsegs; 8118 int32_t tfo_syn; 8119 struct tcp_bbr *bbr; 8120 8121 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8122 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8123 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8124 if ((thflags & TH_ACK) && 8125 (SEQ_LT(tp->snd_wl1, th->th_seq) || 8126 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 8127 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 8128 /* keep track of pure window updates */ 8129 if (tlen == 0 && 8130 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 8131 KMOD_TCPSTAT_INC(tcps_rcvwinupd); 8132 tp->snd_wnd = tiwin; 8133 tp->snd_wl1 = th->th_seq; 8134 tp->snd_wl2 = th->th_ack; 8135 if (tp->snd_wnd > tp->max_sndwnd) 8136 tp->max_sndwnd = tp->snd_wnd; 8137 bbr->r_wanted_output = 1; 8138 } else if (thflags & TH_ACK) { 8139 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 8140 tp->snd_wnd = tiwin; 8141 tp->snd_wl1 = th->th_seq; 8142 tp->snd_wl2 = th->th_ack; 8143 } 8144 } 8145 if (tp->snd_wnd < ctf_outstanding(tp)) 8146 /* The peer collapsed its window on us */ 8147 bbr_collapsed_window(bbr); 8148 else if (bbr->rc_has_collapsed) 8149 bbr_un_collapse_window(bbr); 8150 /* Was persist timer active and now we have window space? */ 8151 if ((bbr->rc_in_persist != 0) && 8152 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8153 bbr_minseg(bbr)))) { 8154 /* 8155 * Make the rate persist at end of persist mode if idle long 8156 * enough 8157 */ 8158 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8159 8160 /* Make sure we output to start the timer */ 8161 bbr->r_wanted_output = 1; 8162 } 8163 /* Do we need to enter persist? */ 8164 if ((bbr->rc_in_persist == 0) && 8165 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8166 TCPS_HAVEESTABLISHED(tp->t_state) && 8167 (tp->snd_max == tp->snd_una) && 8168 sbavail(&so->so_snd) && 8169 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8170 /* No send window.. we must enter persist */ 8171 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8172 } 8173 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 8174 m_freem(m); 8175 return (0); 8176 } 8177 /* 8178 * We don't support urgent data but 8179 * drag along the up just to make sure 8180 * if there is a stack switch no one 8181 * is surprised. 8182 */ 8183 tp->rcv_up = tp->rcv_nxt; 8184 8185 /* 8186 * Process the segment text, merging it into the TCP sequencing 8187 * queue, and arranging for acknowledgment of receipt if necessary. 8188 * This process logically involves adjusting tp->rcv_wnd as data is 8189 * presented to the user (this happens in tcp_usrreq.c, case 8190 * PRU_RCVD). If a FIN has already been received on this connection 8191 * then we just ignore the text. 8192 */ 8193 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 8194 IS_FASTOPEN(tp->t_flags)); 8195 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && 8196 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8197 tcp_seq save_start = th->th_seq; 8198 tcp_seq save_rnxt = tp->rcv_nxt; 8199 int save_tlen = tlen; 8200 8201 m_adj(m, drop_hdrlen); /* delayed header drop */ 8202 /* 8203 * Insert segment which includes th into TCP reassembly 8204 * queue with control block tp. Set thflags to whether 8205 * reassembly now includes a segment with FIN. This handles 8206 * the common case inline (segment is the next to be 8207 * received on an established connection, and the queue is 8208 * empty), avoiding linkage into and removal from the queue 8209 * and repetition of various conversions. Set DELACK for 8210 * segments received in order, but ack immediately when 8211 * segments are out of order (so fast retransmit can work). 8212 */ 8213 if (th->th_seq == tp->rcv_nxt && 8214 SEGQ_EMPTY(tp) && 8215 (TCPS_HAVEESTABLISHED(tp->t_state) || 8216 tfo_syn)) { 8217 #ifdef NETFLIX_SB_LIMITS 8218 u_int mcnt, appended; 8219 8220 if (so->so_rcv.sb_shlim) { 8221 mcnt = m_memcnt(m); 8222 appended = 0; 8223 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8224 CFO_NOSLEEP, NULL) == false) { 8225 counter_u64_add(tcp_sb_shlim_fails, 1); 8226 m_freem(m); 8227 return (0); 8228 } 8229 } 8230 8231 #endif 8232 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { 8233 bbr->bbr_segs_rcvd += max(1, nsegs); 8234 tp->t_flags |= TF_DELACK; 8235 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8236 } else { 8237 bbr->r_wanted_output = 1; 8238 tp->t_flags |= TF_ACKNOW; 8239 } 8240 tp->rcv_nxt += tlen; 8241 if (tlen && 8242 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8243 (tp->t_fbyte_in == 0)) { 8244 tp->t_fbyte_in = ticks; 8245 if (tp->t_fbyte_in == 0) 8246 tp->t_fbyte_in = 1; 8247 if (tp->t_fbyte_out && tp->t_fbyte_in) 8248 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8249 } 8250 thflags = tcp_get_flags(th) & TH_FIN; 8251 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8252 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8253 SOCKBUF_LOCK(&so->so_rcv); 8254 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 8255 m_freem(m); 8256 else 8257 #ifdef NETFLIX_SB_LIMITS 8258 appended = 8259 #endif 8260 sbappendstream_locked(&so->so_rcv, m, 0); 8261 /* NB: sorwakeup_locked() does an implicit unlock. */ 8262 sorwakeup_locked(so); 8263 #ifdef NETFLIX_SB_LIMITS 8264 if (so->so_rcv.sb_shlim && appended != mcnt) 8265 counter_fo_release(so->so_rcv.sb_shlim, 8266 mcnt - appended); 8267 #endif 8268 8269 } else { 8270 /* 8271 * XXX: Due to the header drop above "th" is 8272 * theoretically invalid by now. Fortunately 8273 * m_adj() doesn't actually frees any mbufs when 8274 * trimming from the head. 8275 */ 8276 tcp_seq temp = save_start; 8277 8278 thflags = tcp_reass(tp, th, &temp, &tlen, m); 8279 tp->t_flags |= TF_ACKNOW; 8280 if (tp->t_flags & TF_WAKESOR) { 8281 tp->t_flags &= ~TF_WAKESOR; 8282 /* NB: sorwakeup_locked() does an implicit unlock. */ 8283 sorwakeup_locked(so); 8284 } 8285 } 8286 if ((tp->t_flags & TF_SACK_PERMIT) && 8287 (save_tlen > 0) && 8288 TCPS_HAVEESTABLISHED(tp->t_state)) { 8289 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { 8290 /* 8291 * DSACK actually handled in the fastpath 8292 * above. 8293 */ 8294 tcp_update_sack_list(tp, save_start, 8295 save_start + save_tlen); 8296 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 8297 if ((tp->rcv_numsacks >= 1) && 8298 (tp->sackblks[0].end == save_start)) { 8299 /* 8300 * Partial overlap, recorded at todrop 8301 * above. 8302 */ 8303 tcp_update_sack_list(tp, 8304 tp->sackblks[0].start, 8305 tp->sackblks[0].end); 8306 } else { 8307 tcp_update_dsack_list(tp, save_start, 8308 save_start + save_tlen); 8309 } 8310 } else if (tlen >= save_tlen) { 8311 /* Update of sackblks. */ 8312 tcp_update_dsack_list(tp, save_start, 8313 save_start + save_tlen); 8314 } else if (tlen > 0) { 8315 tcp_update_dsack_list(tp, save_start, 8316 save_start + tlen); 8317 } 8318 } 8319 } else { 8320 m_freem(m); 8321 thflags &= ~TH_FIN; 8322 } 8323 8324 /* 8325 * If FIN is received ACK the FIN and let the user know that the 8326 * connection is closing. 8327 */ 8328 if (thflags & TH_FIN) { 8329 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8330 /* The socket upcall is handled by socantrcvmore. */ 8331 socantrcvmore(so); 8332 /* 8333 * If connection is half-synchronized (ie NEEDSYN 8334 * flag on) then delay ACK, so it may be piggybacked 8335 * when SYN is sent. Otherwise, since we received a 8336 * FIN then no more input can be expected, send ACK 8337 * now. 8338 */ 8339 if (tp->t_flags & TF_NEEDSYN) { 8340 tp->t_flags |= TF_DELACK; 8341 bbr_timer_cancel(bbr, 8342 __LINE__, bbr->r_ctl.rc_rcvtime); 8343 } else { 8344 tp->t_flags |= TF_ACKNOW; 8345 } 8346 tp->rcv_nxt++; 8347 } 8348 switch (tp->t_state) { 8349 /* 8350 * In SYN_RECEIVED and ESTABLISHED STATES enter the 8351 * CLOSE_WAIT state. 8352 */ 8353 case TCPS_SYN_RECEIVED: 8354 tp->t_starttime = ticks; 8355 /* FALLTHROUGH */ 8356 case TCPS_ESTABLISHED: 8357 tcp_state_change(tp, TCPS_CLOSE_WAIT); 8358 break; 8359 8360 /* 8361 * If still in FIN_WAIT_1 STATE FIN has not been 8362 * acked so enter the CLOSING state. 8363 */ 8364 case TCPS_FIN_WAIT_1: 8365 tcp_state_change(tp, TCPS_CLOSING); 8366 break; 8367 8368 /* 8369 * In FIN_WAIT_2 state enter the TIME_WAIT state, 8370 * starting the time-wait timer, turning off the 8371 * other standard timers. 8372 */ 8373 case TCPS_FIN_WAIT_2: 8374 bbr->rc_timer_first = 1; 8375 bbr_timer_cancel(bbr, 8376 __LINE__, bbr->r_ctl.rc_rcvtime); 8377 tcp_twstart(tp); 8378 return (1); 8379 } 8380 } 8381 /* 8382 * Return any desired output. 8383 */ 8384 if ((tp->t_flags & TF_ACKNOW) || 8385 (sbavail(&so->so_snd) > ctf_outstanding(tp))) { 8386 bbr->r_wanted_output = 1; 8387 } 8388 return (0); 8389 } 8390 8391 /* 8392 * Here nothing is really faster, its just that we 8393 * have broken out the fast-data path also just like 8394 * the fast-ack. Return 1 if we processed the packet 8395 * return 0 if you need to take the "slow-path". 8396 */ 8397 static int 8398 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 8399 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8400 uint32_t tiwin, int32_t nxt_pkt) 8401 { 8402 uint16_t nsegs; 8403 int32_t newsize = 0; /* automatic sockbuf scaling */ 8404 struct tcp_bbr *bbr; 8405 #ifdef NETFLIX_SB_LIMITS 8406 u_int mcnt, appended; 8407 #endif 8408 8409 /* On the hpts and we would have called output */ 8410 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8411 8412 /* 8413 * If last ACK falls within this segment's sequence numbers, record 8414 * the timestamp. NOTE that the test is modified according to the 8415 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8416 */ 8417 if (bbr->r_ctl.rc_resend != NULL) { 8418 return (0); 8419 } 8420 if (tiwin && tiwin != tp->snd_wnd) { 8421 return (0); 8422 } 8423 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 8424 return (0); 8425 } 8426 if (__predict_false((to->to_flags & TOF_TS) && 8427 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 8428 return (0); 8429 } 8430 if (__predict_false((th->th_ack != tp->snd_una))) { 8431 return (0); 8432 } 8433 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 8434 return (0); 8435 } 8436 if ((to->to_flags & TOF_TS) != 0 && 8437 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8438 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 8439 tp->ts_recent = to->to_tsval; 8440 } 8441 /* 8442 * This is a pure, in-sequence data packet with nothing on the 8443 * reassembly queue and we have enough buffer space to take it. 8444 */ 8445 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8446 8447 #ifdef NETFLIX_SB_LIMITS 8448 if (so->so_rcv.sb_shlim) { 8449 mcnt = m_memcnt(m); 8450 appended = 0; 8451 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8452 CFO_NOSLEEP, NULL) == false) { 8453 counter_u64_add(tcp_sb_shlim_fails, 1); 8454 m_freem(m); 8455 return (1); 8456 } 8457 } 8458 #endif 8459 /* Clean receiver SACK report if present */ 8460 if (tp->rcv_numsacks) 8461 tcp_clean_sackreport(tp); 8462 KMOD_TCPSTAT_INC(tcps_preddat); 8463 tp->rcv_nxt += tlen; 8464 if (tlen && 8465 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8466 (tp->t_fbyte_in == 0)) { 8467 tp->t_fbyte_in = ticks; 8468 if (tp->t_fbyte_in == 0) 8469 tp->t_fbyte_in = 1; 8470 if (tp->t_fbyte_out && tp->t_fbyte_in) 8471 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8472 } 8473 /* 8474 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 8475 */ 8476 tp->snd_wl1 = th->th_seq; 8477 /* 8478 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 8479 */ 8480 tp->rcv_up = tp->rcv_nxt; 8481 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8482 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8483 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 8484 8485 /* Add data to socket buffer. */ 8486 SOCKBUF_LOCK(&so->so_rcv); 8487 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8488 m_freem(m); 8489 } else { 8490 /* 8491 * Set new socket buffer size. Give up when limit is 8492 * reached. 8493 */ 8494 if (newsize) 8495 if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) 8496 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 8497 m_adj(m, drop_hdrlen); /* delayed header drop */ 8498 8499 #ifdef NETFLIX_SB_LIMITS 8500 appended = 8501 #endif 8502 sbappendstream_locked(&so->so_rcv, m, 0); 8503 ctf_calc_rwin(so, tp); 8504 } 8505 /* NB: sorwakeup_locked() does an implicit unlock. */ 8506 sorwakeup_locked(so); 8507 #ifdef NETFLIX_SB_LIMITS 8508 if (so->so_rcv.sb_shlim && mcnt != appended) 8509 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); 8510 #endif 8511 if (DELAY_ACK(tp, bbr, nsegs)) { 8512 bbr->bbr_segs_rcvd += max(1, nsegs); 8513 tp->t_flags |= TF_DELACK; 8514 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8515 } else { 8516 bbr->r_wanted_output = 1; 8517 tp->t_flags |= TF_ACKNOW; 8518 } 8519 return (1); 8520 } 8521 8522 /* 8523 * This subfunction is used to try to highly optimize the 8524 * fast path. We again allow window updates that are 8525 * in sequence to remain in the fast-path. We also add 8526 * in the __predict's to attempt to help the compiler. 8527 * Note that if we return a 0, then we can *not* process 8528 * it and the caller should push the packet into the 8529 * slow-path. If we return 1, then all is well and 8530 * the packet is fully processed. 8531 */ 8532 static int 8533 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 8534 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8535 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) 8536 { 8537 int32_t acked; 8538 uint16_t nsegs; 8539 uint32_t sack_changed; 8540 uint32_t prev_acked = 0; 8541 struct tcp_bbr *bbr; 8542 8543 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 8544 /* Old ack, behind (or duplicate to) the last one rcv'd */ 8545 return (0); 8546 } 8547 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 8548 /* Above what we have sent? */ 8549 return (0); 8550 } 8551 if (__predict_false(tiwin == 0)) { 8552 /* zero window */ 8553 return (0); 8554 } 8555 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 8556 /* We need a SYN or a FIN, unlikely.. */ 8557 return (0); 8558 } 8559 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 8560 /* Timestamp is behind .. old ack with seq wrap? */ 8561 return (0); 8562 } 8563 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 8564 /* Still recovering */ 8565 return (0); 8566 } 8567 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8568 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { 8569 /* We are retransmitting */ 8570 return (0); 8571 } 8572 if (__predict_false(bbr->rc_in_persist != 0)) { 8573 /* In persist mode */ 8574 return (0); 8575 } 8576 if (bbr->r_ctl.rc_sacked) { 8577 /* We have sack holes on our scoreboard */ 8578 return (0); 8579 } 8580 /* Ok if we reach here, we can process a fast-ack */ 8581 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8582 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 8583 /* 8584 * We never detect loss in fast ack [we can't 8585 * have a sack and can't be in recovery so 8586 * we always pass 0 (nothing detected)]. 8587 */ 8588 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); 8589 /* Did the window get updated? */ 8590 if (tiwin != tp->snd_wnd) { 8591 tp->snd_wnd = tiwin; 8592 tp->snd_wl1 = th->th_seq; 8593 if (tp->snd_wnd > tp->max_sndwnd) 8594 tp->max_sndwnd = tp->snd_wnd; 8595 } 8596 /* Do we need to exit persists? */ 8597 if ((bbr->rc_in_persist != 0) && 8598 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8599 bbr_minseg(bbr)))) { 8600 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8601 bbr->r_wanted_output = 1; 8602 } 8603 /* Do we need to enter persists? */ 8604 if ((bbr->rc_in_persist == 0) && 8605 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8606 TCPS_HAVEESTABLISHED(tp->t_state) && 8607 (tp->snd_max == tp->snd_una) && 8608 sbavail(&so->so_snd) && 8609 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8610 /* No send window.. we must enter persist */ 8611 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8612 } 8613 /* 8614 * If last ACK falls within this segment's sequence numbers, record 8615 * the timestamp. NOTE that the test is modified according to the 8616 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8617 */ 8618 if ((to->to_flags & TOF_TS) != 0 && 8619 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8620 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; 8621 tp->ts_recent = to->to_tsval; 8622 } 8623 /* 8624 * This is a pure ack for outstanding data. 8625 */ 8626 KMOD_TCPSTAT_INC(tcps_predack); 8627 8628 /* 8629 * "bad retransmit" recovery. 8630 */ 8631 if (tp->t_flags & TF_PREVVALID) { 8632 tp->t_flags &= ~TF_PREVVALID; 8633 if (tp->t_rxtshift == 1 && 8634 (int)(ticks - tp->t_badrxtwin) < 0) 8635 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 8636 } 8637 /* 8638 * Recalculate the transmit timer / rtt. 8639 * 8640 * Some boxes send broken timestamp replies during the SYN+ACK 8641 * phase, ignore timestamps of 0 or we could calculate a huge RTT 8642 * and blow up the retransmit timer. 8643 */ 8644 acked = BYTES_THIS_ACK(tp, th); 8645 8646 #ifdef TCP_HHOOK 8647 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 8648 hhook_run_tcp_est_in(tp, th, to); 8649 #endif 8650 8651 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 8652 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 8653 sbdrop(&so->so_snd, acked); 8654 8655 if (SEQ_GT(th->th_ack, tp->snd_una)) 8656 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 8657 tp->snd_una = th->th_ack; 8658 if (tp->snd_wnd < ctf_outstanding(tp)) 8659 /* The peer collapsed its window on us */ 8660 bbr_collapsed_window(bbr); 8661 else if (bbr->rc_has_collapsed) 8662 bbr_un_collapse_window(bbr); 8663 8664 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 8665 tp->snd_recover = tp->snd_una; 8666 } 8667 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); 8668 /* 8669 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 8670 */ 8671 tp->snd_wl2 = th->th_ack; 8672 m_freem(m); 8673 /* 8674 * If all outstanding data are acked, stop retransmit timer, 8675 * otherwise restart timer using current (possibly backed-off) 8676 * value. If process is waiting for space, wakeup/selwakeup/signal. 8677 * If data are ready to send, let tcp_output decide between more 8678 * output or persist. 8679 * Wake up the socket if we have room to write more. 8680 */ 8681 sowwakeup(so); 8682 if (tp->snd_una == tp->snd_max) { 8683 /* Nothing left outstanding */ 8684 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 8685 if (sbavail(&so->so_snd) == 0) 8686 bbr->rc_tp->t_acktime = 0; 8687 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8688 if (bbr->rc_in_persist == 0) { 8689 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 8690 } 8691 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 8692 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 8693 /* 8694 * We invalidate the last ack here since we 8695 * don't want to transfer forward the time 8696 * for our sum's calculations. 8697 */ 8698 bbr->r_wanted_output = 1; 8699 } 8700 if (sbavail(&so->so_snd)) { 8701 bbr->r_wanted_output = 1; 8702 } 8703 return (1); 8704 } 8705 8706 /* 8707 * Return value of 1, the TCB is unlocked and most 8708 * likely gone, return value of 0, the TCB is still 8709 * locked. 8710 */ 8711 static int 8712 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 8713 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8714 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8715 { 8716 int32_t todrop; 8717 int32_t ourfinisacked = 0; 8718 struct tcp_bbr *bbr; 8719 int32_t ret_val = 0; 8720 8721 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8722 8723 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8724 ctf_calc_rwin(so, tp); 8725 /* 8726 * If the state is SYN_SENT: if seg contains an ACK, but not for our 8727 * SYN, drop the input. if seg contains a RST, then drop the 8728 * connection. if seg does not contain SYN, then drop it. Otherwise 8729 * this is an acceptable SYN segment initialize tp->rcv_nxt and 8730 * tp->irs if seg contains ack then advance tp->snd_una. BRR does 8731 * not support ECN so we will not say we are capable. if SYN has 8732 * been acked change to ESTABLISHED else SYN_RCVD state arrange for 8733 * segment to be acked (eventually) continue processing rest of 8734 * data/controls, beginning with URG 8735 */ 8736 if ((thflags & TH_ACK) && 8737 (SEQ_LEQ(th->th_ack, tp->iss) || 8738 SEQ_GT(th->th_ack, tp->snd_max))) { 8739 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8740 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8741 return (1); 8742 } 8743 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 8744 TCP_PROBE5(connect__refused, NULL, tp, 8745 mtod(m, const char *), tp, th); 8746 tp = tcp_drop(tp, ECONNREFUSED); 8747 ctf_do_drop(m, tp); 8748 return (1); 8749 } 8750 if (thflags & TH_RST) { 8751 ctf_do_drop(m, tp); 8752 return (1); 8753 } 8754 if (!(thflags & TH_SYN)) { 8755 ctf_do_drop(m, tp); 8756 return (1); 8757 } 8758 tp->irs = th->th_seq; 8759 tcp_rcvseqinit(tp); 8760 if (thflags & TH_ACK) { 8761 int tfo_partial = 0; 8762 8763 KMOD_TCPSTAT_INC(tcps_connects); 8764 soisconnected(so); 8765 #ifdef MAC 8766 mac_socketpeer_set_from_mbuf(m, so); 8767 #endif 8768 /* Do window scaling on this connection? */ 8769 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 8770 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 8771 tp->rcv_scale = tp->request_r_scale; 8772 } 8773 tp->rcv_adv += min(tp->rcv_wnd, 8774 TCP_MAXWIN << tp->rcv_scale); 8775 /* 8776 * If not all the data that was sent in the TFO SYN 8777 * has been acked, resend the remainder right away. 8778 */ 8779 if (IS_FASTOPEN(tp->t_flags) && 8780 (tp->snd_una != tp->snd_max)) { 8781 tp->snd_nxt = th->th_ack; 8782 tfo_partial = 1; 8783 } 8784 /* 8785 * If there's data, delay ACK; if there's also a FIN ACKNOW 8786 * will be turned on later. 8787 */ 8788 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { 8789 bbr->bbr_segs_rcvd += 1; 8790 tp->t_flags |= TF_DELACK; 8791 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8792 } else { 8793 bbr->r_wanted_output = 1; 8794 tp->t_flags |= TF_ACKNOW; 8795 } 8796 if (SEQ_GT(th->th_ack, tp->iss)) { 8797 /* 8798 * The SYN is acked 8799 * handle it specially. 8800 */ 8801 bbr_log_syn(tp, to); 8802 } 8803 if (SEQ_GT(th->th_ack, tp->snd_una)) { 8804 /* 8805 * We advance snd_una for the 8806 * fast open case. If th_ack is 8807 * acknowledging data beyond 8808 * snd_una we can't just call 8809 * ack-processing since the 8810 * data stream in our send-map 8811 * will start at snd_una + 1 (one 8812 * beyond the SYN). If its just 8813 * equal we don't need to do that 8814 * and there is no send_map. 8815 */ 8816 tp->snd_una++; 8817 } 8818 /* 8819 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 8820 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 8821 */ 8822 tp->t_starttime = ticks; 8823 if (tp->t_flags & TF_NEEDFIN) { 8824 tcp_state_change(tp, TCPS_FIN_WAIT_1); 8825 tp->t_flags &= ~TF_NEEDFIN; 8826 thflags &= ~TH_SYN; 8827 } else { 8828 tcp_state_change(tp, TCPS_ESTABLISHED); 8829 TCP_PROBE5(connect__established, NULL, tp, 8830 mtod(m, const char *), tp, th); 8831 cc_conn_init(tp); 8832 } 8833 } else { 8834 /* 8835 * Received initial SYN in SYN-SENT[*] state => simultaneous 8836 * open. If segment contains CC option and there is a 8837 * cached CC, apply TAO test. If it succeeds, connection is * 8838 * half-synchronized. Otherwise, do 3-way handshake: 8839 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 8840 * there was no CC option, clear cached CC value. 8841 */ 8842 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN); 8843 tcp_state_change(tp, TCPS_SYN_RECEIVED); 8844 } 8845 /* 8846 * Advance th->th_seq to correspond to first data byte. If data, 8847 * trim to stay within window, dropping FIN if necessary. 8848 */ 8849 th->th_seq++; 8850 if (tlen > tp->rcv_wnd) { 8851 todrop = tlen - tp->rcv_wnd; 8852 m_adj(m, -todrop); 8853 tlen = tp->rcv_wnd; 8854 thflags &= ~TH_FIN; 8855 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 8856 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 8857 } 8858 tp->snd_wl1 = th->th_seq - 1; 8859 tp->rcv_up = th->th_seq; 8860 /* 8861 * Client side of transaction: already sent SYN and data. If the 8862 * remote host used T/TCP to validate the SYN, our data will be 8863 * ACK'd; if so, enter normal data segment processing in the middle 8864 * of step 5, ack processing. Otherwise, goto step 6. 8865 */ 8866 if (thflags & TH_ACK) { 8867 if ((to->to_flags & TOF_TS) != 0) { 8868 uint32_t t, rtt; 8869 8870 t = tcp_tv_to_mssectick(&bbr->rc_tv); 8871 if (TSTMP_GEQ(t, to->to_tsecr)) { 8872 rtt = t - to->to_tsecr; 8873 if (rtt == 0) { 8874 rtt = 1; 8875 } 8876 rtt *= MS_IN_USEC; 8877 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 8878 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, 8879 rtt, bbr->r_ctl.rc_rcvtime); 8880 } 8881 } 8882 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 8883 return (ret_val); 8884 /* We may have changed to FIN_WAIT_1 above */ 8885 if (tp->t_state == TCPS_FIN_WAIT_1) { 8886 /* 8887 * In FIN_WAIT_1 STATE in addition to the processing 8888 * for the ESTABLISHED state if our FIN is now 8889 * acknowledged then enter FIN_WAIT_2. 8890 */ 8891 if (ourfinisacked) { 8892 /* 8893 * If we can't receive any more data, then 8894 * closing user can proceed. Starting the 8895 * timer is contrary to the specification, 8896 * but if we don't get a FIN we'll hang 8897 * forever. 8898 * 8899 * XXXjl: we should release the tp also, and 8900 * use a compressed state. 8901 */ 8902 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8903 soisdisconnected(so); 8904 tcp_timer_activate(tp, TT_2MSL, 8905 (tcp_fast_finwait2_recycle ? 8906 tcp_finwait2_timeout : 8907 TP_MAXIDLE(tp))); 8908 } 8909 tcp_state_change(tp, TCPS_FIN_WAIT_2); 8910 } 8911 } 8912 } 8913 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 8914 tiwin, thflags, nxt_pkt)); 8915 } 8916 8917 /* 8918 * Return value of 1, the TCB is unlocked and most 8919 * likely gone, return value of 0, the TCB is still 8920 * locked. 8921 */ 8922 static int 8923 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 8924 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8925 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8926 { 8927 int32_t ourfinisacked = 0; 8928 int32_t ret_val; 8929 struct tcp_bbr *bbr; 8930 8931 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8932 8933 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8934 ctf_calc_rwin(so, tp); 8935 if ((thflags & TH_ACK) && 8936 (SEQ_LEQ(th->th_ack, tp->snd_una) || 8937 SEQ_GT(th->th_ack, tp->snd_max))) { 8938 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8939 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8940 return (1); 8941 } 8942 if (IS_FASTOPEN(tp->t_flags)) { 8943 /* 8944 * When a TFO connection is in SYN_RECEIVED, the only valid 8945 * packets are the initial SYN, a retransmit/copy of the 8946 * initial SYN (possibly with a subset of the original 8947 * data), a valid ACK, a FIN, or a RST. 8948 */ 8949 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 8950 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8951 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8952 return (1); 8953 } else if (thflags & TH_SYN) { 8954 /* non-initial SYN is ignored */ 8955 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 8956 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 8957 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 8958 ctf_do_drop(m, NULL); 8959 return (0); 8960 } 8961 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 8962 ctf_do_drop(m, NULL); 8963 return (0); 8964 } 8965 } 8966 if ((thflags & TH_RST) || 8967 (tp->t_fin_is_rst && (thflags & TH_FIN))) 8968 return (ctf_process_rst(m, th, so, tp)); 8969 /* 8970 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 8971 * it's less than ts_recent, drop it. 8972 */ 8973 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 8974 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 8975 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 8976 return (ret_val); 8977 } 8978 /* 8979 * In the SYN-RECEIVED state, validate that the packet belongs to 8980 * this connection before trimming the data to fit the receive 8981 * window. Check the sequence number versus IRS since we know the 8982 * sequence numbers haven't wrapped. This is a partial fix for the 8983 * "LAND" DoS attack. 8984 */ 8985 if (SEQ_LT(th->th_seq, tp->irs)) { 8986 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8987 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8988 return (1); 8989 } 8990 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 8991 return (ret_val); 8992 } 8993 /* 8994 * If last ACK falls within this segment's sequence numbers, record 8995 * its timestamp. NOTE: 1) That the test incorporates suggestions 8996 * from the latest proposal of the tcplw@cray.com list (Braden 8997 * 1993/04/26). 2) That updating only on newer timestamps interferes 8998 * with our earlier PAWS tests, so this check should be solely 8999 * predicated on the sequence space of this segment. 3) That we 9000 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9001 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9002 * SEG.Len, This modified check allows us to overcome RFC1323's 9003 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9004 * p.869. In such cases, we can still calculate the RTT correctly 9005 * when RCV.NXT == Last.ACK.Sent. 9006 */ 9007 if ((to->to_flags & TOF_TS) != 0 && 9008 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9009 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9010 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9011 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9012 tp->ts_recent = to->to_tsval; 9013 } 9014 tp->snd_wnd = tiwin; 9015 /* 9016 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9017 * is on (half-synchronized state), then queue data for later 9018 * processing; else drop segment and return. 9019 */ 9020 if ((thflags & TH_ACK) == 0) { 9021 if (IS_FASTOPEN(tp->t_flags)) { 9022 cc_conn_init(tp); 9023 } 9024 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9025 tiwin, thflags, nxt_pkt)); 9026 } 9027 KMOD_TCPSTAT_INC(tcps_connects); 9028 if (tp->t_flags & TF_SONOTCONN) { 9029 tp->t_flags &= ~TF_SONOTCONN; 9030 soisconnected(so); 9031 } 9032 /* Do window scaling? */ 9033 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 9034 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 9035 tp->rcv_scale = tp->request_r_scale; 9036 } 9037 /* 9038 * ok for the first time in lets see if we can use the ts to figure 9039 * out what the initial RTT was. 9040 */ 9041 if ((to->to_flags & TOF_TS) != 0) { 9042 uint32_t t, rtt; 9043 9044 t = tcp_tv_to_mssectick(&bbr->rc_tv); 9045 if (TSTMP_GEQ(t, to->to_tsecr)) { 9046 rtt = t - to->to_tsecr; 9047 if (rtt == 0) { 9048 rtt = 1; 9049 } 9050 rtt *= MS_IN_USEC; 9051 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 9052 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); 9053 } 9054 } 9055 /* Drop off any SYN in the send map (probably not there) */ 9056 if (thflags & TH_ACK) 9057 bbr_log_syn(tp, to); 9058 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 9059 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 9060 tp->t_tfo_pending = NULL; 9061 } 9062 /* 9063 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 9064 * FIN-WAIT-1 9065 */ 9066 tp->t_starttime = ticks; 9067 if (tp->t_flags & TF_NEEDFIN) { 9068 tcp_state_change(tp, TCPS_FIN_WAIT_1); 9069 tp->t_flags &= ~TF_NEEDFIN; 9070 } else { 9071 tcp_state_change(tp, TCPS_ESTABLISHED); 9072 TCP_PROBE5(accept__established, NULL, tp, 9073 mtod(m, const char *), tp, th); 9074 /* 9075 * TFO connections call cc_conn_init() during SYN 9076 * processing. Calling it again here for such connections 9077 * is not harmless as it would undo the snd_cwnd reduction 9078 * that occurs when a TFO SYN|ACK is retransmitted. 9079 */ 9080 if (!IS_FASTOPEN(tp->t_flags)) 9081 cc_conn_init(tp); 9082 } 9083 /* 9084 * Account for the ACK of our SYN prior to 9085 * regular ACK processing below, except for 9086 * simultaneous SYN, which is handled later. 9087 */ 9088 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) 9089 tp->snd_una++; 9090 /* 9091 * If segment contains data or ACK, will call tcp_reass() later; if 9092 * not, do so now to pass queued data to user. 9093 */ 9094 if (tlen == 0 && (thflags & TH_FIN) == 0) { 9095 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 9096 (struct mbuf *)0); 9097 if (tp->t_flags & TF_WAKESOR) { 9098 tp->t_flags &= ~TF_WAKESOR; 9099 /* NB: sorwakeup_locked() does an implicit unlock. */ 9100 sorwakeup_locked(so); 9101 } 9102 } 9103 tp->snd_wl1 = th->th_seq - 1; 9104 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9105 return (ret_val); 9106 } 9107 if (tp->t_state == TCPS_FIN_WAIT_1) { 9108 /* We could have went to FIN_WAIT_1 (or EST) above */ 9109 /* 9110 * In FIN_WAIT_1 STATE in addition to the processing for the 9111 * ESTABLISHED state if our FIN is now acknowledged then 9112 * enter FIN_WAIT_2. 9113 */ 9114 if (ourfinisacked) { 9115 /* 9116 * If we can't receive any more data, then closing 9117 * user can proceed. Starting the timer is contrary 9118 * to the specification, but if we don't get a FIN 9119 * we'll hang forever. 9120 * 9121 * XXXjl: we should release the tp also, and use a 9122 * compressed state. 9123 */ 9124 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9125 soisdisconnected(so); 9126 tcp_timer_activate(tp, TT_2MSL, 9127 (tcp_fast_finwait2_recycle ? 9128 tcp_finwait2_timeout : 9129 TP_MAXIDLE(tp))); 9130 } 9131 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9132 } 9133 } 9134 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9135 tiwin, thflags, nxt_pkt)); 9136 } 9137 9138 /* 9139 * Return value of 1, the TCB is unlocked and most 9140 * likely gone, return value of 0, the TCB is still 9141 * locked. 9142 */ 9143 static int 9144 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 9145 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9146 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9147 { 9148 struct tcp_bbr *bbr; 9149 int32_t ret_val; 9150 9151 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9152 9153 /* 9154 * Header prediction: check for the two common cases of a 9155 * uni-directional data xfer. If the packet has no control flags, 9156 * is in-sequence, the window didn't change and we're not 9157 * retransmitting, it's a candidate. If the length is zero and the 9158 * ack moved forward, we're the sender side of the xfer. Just free 9159 * the data acked & wake any higher level process that was blocked 9160 * waiting for space. If the length is non-zero and the ack didn't 9161 * move, we're the receiver side. If we're getting packets in-order 9162 * (the reassembly queue is empty), add the data toc The socket 9163 * buffer and note that we need a delayed ack. Make sure that the 9164 * hidden state-flags are also off. Since we check for 9165 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 9166 */ 9167 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9168 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { 9169 /* 9170 * If we have delived under 4 segments increase the initial 9171 * window if raised by the peer. We use this to determine 9172 * dynamic and static rwnd's at the end of a connection. 9173 */ 9174 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); 9175 } 9176 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 9177 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 9178 __predict_true(SEGQ_EMPTY(tp)) && 9179 __predict_true(th->th_seq == tp->rcv_nxt)) { 9180 if (tlen == 0) { 9181 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 9182 tiwin, nxt_pkt, iptos)) { 9183 return (0); 9184 } 9185 } else { 9186 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 9187 tiwin, nxt_pkt)) { 9188 return (0); 9189 } 9190 } 9191 } 9192 ctf_calc_rwin(so, tp); 9193 9194 if ((thflags & TH_RST) || 9195 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9196 return (ctf_process_rst(m, th, so, tp)); 9197 /* 9198 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9199 * synchronized state. 9200 */ 9201 if (thflags & TH_SYN) { 9202 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9203 return (ret_val); 9204 } 9205 /* 9206 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9207 * it's less than ts_recent, drop it. 9208 */ 9209 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9210 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9211 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9212 return (ret_val); 9213 } 9214 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9215 return (ret_val); 9216 } 9217 /* 9218 * If last ACK falls within this segment's sequence numbers, record 9219 * its timestamp. NOTE: 1) That the test incorporates suggestions 9220 * from the latest proposal of the tcplw@cray.com list (Braden 9221 * 1993/04/26). 2) That updating only on newer timestamps interferes 9222 * with our earlier PAWS tests, so this check should be solely 9223 * predicated on the sequence space of this segment. 3) That we 9224 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9225 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9226 * SEG.Len, This modified check allows us to overcome RFC1323's 9227 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9228 * p.869. In such cases, we can still calculate the RTT correctly 9229 * when RCV.NXT == Last.ACK.Sent. 9230 */ 9231 if ((to->to_flags & TOF_TS) != 0 && 9232 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9233 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9234 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9235 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9236 tp->ts_recent = to->to_tsval; 9237 } 9238 /* 9239 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9240 * is on (half-synchronized state), then queue data for later 9241 * processing; else drop segment and return. 9242 */ 9243 if ((thflags & TH_ACK) == 0) { 9244 if (tp->t_flags & TF_NEEDSYN) { 9245 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9246 tiwin, thflags, nxt_pkt)); 9247 } else if (tp->t_flags & TF_ACKNOW) { 9248 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9249 bbr->r_wanted_output = 1; 9250 return (ret_val); 9251 } else { 9252 ctf_do_drop(m, NULL); 9253 return (0); 9254 } 9255 } 9256 /* 9257 * Ack processing. 9258 */ 9259 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9260 return (ret_val); 9261 } 9262 if (sbavail(&so->so_snd)) { 9263 if (ctf_progress_timeout_check(tp, true)) { 9264 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9265 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9266 return (1); 9267 } 9268 } 9269 /* State changes only happen in bbr_process_data() */ 9270 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9271 tiwin, thflags, nxt_pkt)); 9272 } 9273 9274 /* 9275 * Return value of 1, the TCB is unlocked and most 9276 * likely gone, return value of 0, the TCB is still 9277 * locked. 9278 */ 9279 static int 9280 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 9281 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9282 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9283 { 9284 struct tcp_bbr *bbr; 9285 int32_t ret_val; 9286 9287 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9288 9289 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9290 ctf_calc_rwin(so, tp); 9291 if ((thflags & TH_RST) || 9292 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9293 return (ctf_process_rst(m, th, so, tp)); 9294 /* 9295 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9296 * synchronized state. 9297 */ 9298 if (thflags & TH_SYN) { 9299 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9300 return (ret_val); 9301 } 9302 /* 9303 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9304 * it's less than ts_recent, drop it. 9305 */ 9306 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9307 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9308 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9309 return (ret_val); 9310 } 9311 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9312 return (ret_val); 9313 } 9314 /* 9315 * If last ACK falls within this segment's sequence numbers, record 9316 * its timestamp. NOTE: 1) That the test incorporates suggestions 9317 * from the latest proposal of the tcplw@cray.com list (Braden 9318 * 1993/04/26). 2) That updating only on newer timestamps interferes 9319 * with our earlier PAWS tests, so this check should be solely 9320 * predicated on the sequence space of this segment. 3) That we 9321 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9322 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9323 * SEG.Len, This modified check allows us to overcome RFC1323's 9324 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9325 * p.869. In such cases, we can still calculate the RTT correctly 9326 * when RCV.NXT == Last.ACK.Sent. 9327 */ 9328 if ((to->to_flags & TOF_TS) != 0 && 9329 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9330 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9331 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9332 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9333 tp->ts_recent = to->to_tsval; 9334 } 9335 /* 9336 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9337 * is on (half-synchronized state), then queue data for later 9338 * processing; else drop segment and return. 9339 */ 9340 if ((thflags & TH_ACK) == 0) { 9341 if (tp->t_flags & TF_NEEDSYN) { 9342 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9343 tiwin, thflags, nxt_pkt)); 9344 } else if (tp->t_flags & TF_ACKNOW) { 9345 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9346 bbr->r_wanted_output = 1; 9347 return (ret_val); 9348 } else { 9349 ctf_do_drop(m, NULL); 9350 return (0); 9351 } 9352 } 9353 /* 9354 * Ack processing. 9355 */ 9356 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9357 return (ret_val); 9358 } 9359 if (sbavail(&so->so_snd)) { 9360 if (ctf_progress_timeout_check(tp, true)) { 9361 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9362 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9363 return (1); 9364 } 9365 } 9366 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9367 tiwin, thflags, nxt_pkt)); 9368 } 9369 9370 static int 9371 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, 9372 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) 9373 { 9374 9375 if (bbr->rc_allow_data_af_clo == 0) { 9376 close_now: 9377 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 9378 /* tcp_close will kill the inp pre-log the Reset */ 9379 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 9380 tp = tcp_close(tp); 9381 KMOD_TCPSTAT_INC(tcps_rcvafterclose); 9382 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 9383 return (1); 9384 } 9385 if (sbavail(&so->so_snd) == 0) 9386 goto close_now; 9387 /* Ok we allow data that is ignored and a followup reset */ 9388 tp->rcv_nxt = th->th_seq + *tlen; 9389 tp->t_flags2 |= TF2_DROP_AF_DATA; 9390 bbr->r_wanted_output = 1; 9391 *tlen = 0; 9392 return (0); 9393 } 9394 9395 /* 9396 * Return value of 1, the TCB is unlocked and most 9397 * likely gone, return value of 0, the TCB is still 9398 * locked. 9399 */ 9400 static int 9401 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 9402 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9403 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9404 { 9405 int32_t ourfinisacked = 0; 9406 int32_t ret_val; 9407 struct tcp_bbr *bbr; 9408 9409 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9410 9411 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9412 ctf_calc_rwin(so, tp); 9413 if ((thflags & TH_RST) || 9414 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9415 return (ctf_process_rst(m, th, so, tp)); 9416 /* 9417 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9418 * synchronized state. 9419 */ 9420 if (thflags & TH_SYN) { 9421 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9422 return (ret_val); 9423 } 9424 /* 9425 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9426 * it's less than ts_recent, drop it. 9427 */ 9428 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9429 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9430 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9431 return (ret_val); 9432 } 9433 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9434 return (ret_val); 9435 } 9436 /* 9437 * If new data are received on a connection after the user processes 9438 * are gone, then RST the other end. 9439 * We call a new function now so we might continue and setup 9440 * to reset at all data being ack'd. 9441 */ 9442 if ((tp->t_flags & TF_CLOSED) && tlen && 9443 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9444 return (1); 9445 /* 9446 * If last ACK falls within this segment's sequence numbers, record 9447 * its timestamp. NOTE: 1) That the test incorporates suggestions 9448 * from the latest proposal of the tcplw@cray.com list (Braden 9449 * 1993/04/26). 2) That updating only on newer timestamps interferes 9450 * with our earlier PAWS tests, so this check should be solely 9451 * predicated on the sequence space of this segment. 3) That we 9452 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9453 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9454 * SEG.Len, This modified check allows us to overcome RFC1323's 9455 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9456 * p.869. In such cases, we can still calculate the RTT correctly 9457 * when RCV.NXT == Last.ACK.Sent. 9458 */ 9459 if ((to->to_flags & TOF_TS) != 0 && 9460 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9461 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9462 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9463 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9464 tp->ts_recent = to->to_tsval; 9465 } 9466 /* 9467 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9468 * is on (half-synchronized state), then queue data for later 9469 * processing; else drop segment and return. 9470 */ 9471 if ((thflags & TH_ACK) == 0) { 9472 if (tp->t_flags & TF_NEEDSYN) { 9473 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9474 tiwin, thflags, nxt_pkt)); 9475 } else if (tp->t_flags & TF_ACKNOW) { 9476 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9477 bbr->r_wanted_output = 1; 9478 return (ret_val); 9479 } else { 9480 ctf_do_drop(m, NULL); 9481 return (0); 9482 } 9483 } 9484 /* 9485 * Ack processing. 9486 */ 9487 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9488 return (ret_val); 9489 } 9490 if (ourfinisacked) { 9491 /* 9492 * If we can't receive any more data, then closing user can 9493 * proceed. Starting the timer is contrary to the 9494 * specification, but if we don't get a FIN we'll hang 9495 * forever. 9496 * 9497 * XXXjl: we should release the tp also, and use a 9498 * compressed state. 9499 */ 9500 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9501 soisdisconnected(so); 9502 tcp_timer_activate(tp, TT_2MSL, 9503 (tcp_fast_finwait2_recycle ? 9504 tcp_finwait2_timeout : 9505 TP_MAXIDLE(tp))); 9506 } 9507 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9508 } 9509 if (sbavail(&so->so_snd)) { 9510 if (ctf_progress_timeout_check(tp, true)) { 9511 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9512 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9513 return (1); 9514 } 9515 } 9516 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9517 tiwin, thflags, nxt_pkt)); 9518 } 9519 9520 /* 9521 * Return value of 1, the TCB is unlocked and most 9522 * likely gone, return value of 0, the TCB is still 9523 * locked. 9524 */ 9525 static int 9526 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 9527 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9528 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9529 { 9530 int32_t ourfinisacked = 0; 9531 int32_t ret_val; 9532 struct tcp_bbr *bbr; 9533 9534 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9535 9536 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9537 ctf_calc_rwin(so, tp); 9538 if ((thflags & TH_RST) || 9539 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9540 return (ctf_process_rst(m, th, so, tp)); 9541 /* 9542 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9543 * synchronized state. 9544 */ 9545 if (thflags & TH_SYN) { 9546 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9547 return (ret_val); 9548 } 9549 /* 9550 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9551 * it's less than ts_recent, drop it. 9552 */ 9553 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9554 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9555 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9556 return (ret_val); 9557 } 9558 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9559 return (ret_val); 9560 } 9561 /* 9562 * If new data are received on a connection after the user processes 9563 * are gone, then RST the other end. 9564 * We call a new function now so we might continue and setup 9565 * to reset at all data being ack'd. 9566 */ 9567 if ((tp->t_flags & TF_CLOSED) && tlen && 9568 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9569 return (1); 9570 /* 9571 * If last ACK falls within this segment's sequence numbers, record 9572 * its timestamp. NOTE: 1) That the test incorporates suggestions 9573 * from the latest proposal of the tcplw@cray.com list (Braden 9574 * 1993/04/26). 2) That updating only on newer timestamps interferes 9575 * with our earlier PAWS tests, so this check should be solely 9576 * predicated on the sequence space of this segment. 3) That we 9577 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9578 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9579 * SEG.Len, This modified check allows us to overcome RFC1323's 9580 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9581 * p.869. In such cases, we can still calculate the RTT correctly 9582 * when RCV.NXT == Last.ACK.Sent. 9583 */ 9584 if ((to->to_flags & TOF_TS) != 0 && 9585 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9586 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9587 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9588 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9589 tp->ts_recent = to->to_tsval; 9590 } 9591 /* 9592 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9593 * is on (half-synchronized state), then queue data for later 9594 * processing; else drop segment and return. 9595 */ 9596 if ((thflags & TH_ACK) == 0) { 9597 if (tp->t_flags & TF_NEEDSYN) { 9598 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9599 tiwin, thflags, nxt_pkt)); 9600 } else if (tp->t_flags & TF_ACKNOW) { 9601 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9602 bbr->r_wanted_output = 1; 9603 return (ret_val); 9604 } else { 9605 ctf_do_drop(m, NULL); 9606 return (0); 9607 } 9608 } 9609 /* 9610 * Ack processing. 9611 */ 9612 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9613 return (ret_val); 9614 } 9615 if (ourfinisacked) { 9616 tcp_twstart(tp); 9617 m_freem(m); 9618 return (1); 9619 } 9620 if (sbavail(&so->so_snd)) { 9621 if (ctf_progress_timeout_check(tp, true)) { 9622 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9623 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9624 return (1); 9625 } 9626 } 9627 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9628 tiwin, thflags, nxt_pkt)); 9629 } 9630 9631 /* 9632 * Return value of 1, the TCB is unlocked and most 9633 * likely gone, return value of 0, the TCB is still 9634 * locked. 9635 */ 9636 static int 9637 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 9638 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9639 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9640 { 9641 int32_t ourfinisacked = 0; 9642 int32_t ret_val; 9643 struct tcp_bbr *bbr; 9644 9645 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9646 9647 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9648 ctf_calc_rwin(so, tp); 9649 if ((thflags & TH_RST) || 9650 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9651 return (ctf_process_rst(m, th, so, tp)); 9652 /* 9653 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9654 * synchronized state. 9655 */ 9656 if (thflags & TH_SYN) { 9657 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9658 return (ret_val); 9659 } 9660 /* 9661 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9662 * it's less than ts_recent, drop it. 9663 */ 9664 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9665 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9666 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9667 return (ret_val); 9668 } 9669 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9670 return (ret_val); 9671 } 9672 /* 9673 * If new data are received on a connection after the user processes 9674 * are gone, then RST the other end. 9675 * We call a new function now so we might continue and setup 9676 * to reset at all data being ack'd. 9677 */ 9678 if ((tp->t_flags & TF_CLOSED) && tlen && 9679 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9680 return (1); 9681 /* 9682 * If last ACK falls within this segment's sequence numbers, record 9683 * its timestamp. NOTE: 1) That the test incorporates suggestions 9684 * from the latest proposal of the tcplw@cray.com list (Braden 9685 * 1993/04/26). 2) That updating only on newer timestamps interferes 9686 * with our earlier PAWS tests, so this check should be solely 9687 * predicated on the sequence space of this segment. 3) That we 9688 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9689 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9690 * SEG.Len, This modified check allows us to overcome RFC1323's 9691 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9692 * p.869. In such cases, we can still calculate the RTT correctly 9693 * when RCV.NXT == Last.ACK.Sent. 9694 */ 9695 if ((to->to_flags & TOF_TS) != 0 && 9696 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9697 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9698 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9699 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9700 tp->ts_recent = to->to_tsval; 9701 } 9702 /* 9703 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9704 * is on (half-synchronized state), then queue data for later 9705 * processing; else drop segment and return. 9706 */ 9707 if ((thflags & TH_ACK) == 0) { 9708 if (tp->t_flags & TF_NEEDSYN) { 9709 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9710 tiwin, thflags, nxt_pkt)); 9711 } else if (tp->t_flags & TF_ACKNOW) { 9712 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9713 bbr->r_wanted_output = 1; 9714 return (ret_val); 9715 } else { 9716 ctf_do_drop(m, NULL); 9717 return (0); 9718 } 9719 } 9720 /* 9721 * case TCPS_LAST_ACK: Ack processing. 9722 */ 9723 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9724 return (ret_val); 9725 } 9726 if (ourfinisacked) { 9727 tp = tcp_close(tp); 9728 ctf_do_drop(m, tp); 9729 return (1); 9730 } 9731 if (sbavail(&so->so_snd)) { 9732 if (ctf_progress_timeout_check(tp, true)) { 9733 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9734 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9735 return (1); 9736 } 9737 } 9738 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9739 tiwin, thflags, nxt_pkt)); 9740 } 9741 9742 /* 9743 * Return value of 1, the TCB is unlocked and most 9744 * likely gone, return value of 0, the TCB is still 9745 * locked. 9746 */ 9747 static int 9748 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 9749 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9750 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9751 { 9752 int32_t ourfinisacked = 0; 9753 int32_t ret_val; 9754 struct tcp_bbr *bbr; 9755 9756 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9757 9758 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9759 ctf_calc_rwin(so, tp); 9760 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 9761 if ((thflags & TH_RST) || 9762 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9763 return (ctf_process_rst(m, th, so, tp)); 9764 9765 /* 9766 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9767 * synchronized state. 9768 */ 9769 if (thflags & TH_SYN) { 9770 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9771 return (ret_val); 9772 } 9773 /* 9774 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9775 * it's less than ts_recent, drop it. 9776 */ 9777 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9778 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9779 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9780 return (ret_val); 9781 } 9782 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9783 return (ret_val); 9784 } 9785 /* 9786 * If new data are received on a connection after the user processes 9787 * are gone, then we may RST the other end depending on the outcome 9788 * of bbr_check_data_after_close. 9789 * We call a new function now so we might continue and setup 9790 * to reset at all data being ack'd. 9791 */ 9792 if ((tp->t_flags & TF_CLOSED) && tlen && 9793 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9794 return (1); 9795 /* 9796 * If last ACK falls within this segment's sequence numbers, record 9797 * its timestamp. NOTE: 1) That the test incorporates suggestions 9798 * from the latest proposal of the tcplw@cray.com list (Braden 9799 * 1993/04/26). 2) That updating only on newer timestamps interferes 9800 * with our earlier PAWS tests, so this check should be solely 9801 * predicated on the sequence space of this segment. 3) That we 9802 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9803 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9804 * SEG.Len, This modified check allows us to overcome RFC1323's 9805 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9806 * p.869. In such cases, we can still calculate the RTT correctly 9807 * when RCV.NXT == Last.ACK.Sent. 9808 */ 9809 if ((to->to_flags & TOF_TS) != 0 && 9810 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9811 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9812 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9813 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9814 tp->ts_recent = to->to_tsval; 9815 } 9816 /* 9817 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9818 * is on (half-synchronized state), then queue data for later 9819 * processing; else drop segment and return. 9820 */ 9821 if ((thflags & TH_ACK) == 0) { 9822 if (tp->t_flags & TF_NEEDSYN) { 9823 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9824 tiwin, thflags, nxt_pkt)); 9825 } else if (tp->t_flags & TF_ACKNOW) { 9826 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9827 bbr->r_wanted_output = 1; 9828 return (ret_val); 9829 } else { 9830 ctf_do_drop(m, NULL); 9831 return (0); 9832 } 9833 } 9834 /* 9835 * Ack processing. 9836 */ 9837 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9838 return (ret_val); 9839 } 9840 if (sbavail(&so->so_snd)) { 9841 if (ctf_progress_timeout_check(tp, true)) { 9842 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9843 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9844 return (1); 9845 } 9846 } 9847 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9848 tiwin, thflags, nxt_pkt)); 9849 } 9850 9851 static void 9852 bbr_stop_all_timers(struct tcpcb *tp) 9853 { 9854 struct tcp_bbr *bbr; 9855 9856 /* 9857 * Assure no timers are running. 9858 */ 9859 if (tcp_timer_active(tp, TT_PERSIST)) { 9860 /* We enter in persists, set the flag appropriately */ 9861 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9862 bbr->rc_in_persist = 1; 9863 } 9864 } 9865 9866 static void 9867 bbr_google_mode_on(struct tcp_bbr *bbr) 9868 { 9869 bbr->rc_use_google = 1; 9870 bbr->rc_no_pacing = 0; 9871 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9872 bbr->r_use_policer = bbr_policer_detection_enabled; 9873 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 9874 bbr->bbr_use_rack_cheat = 0; 9875 bbr->r_ctl.rc_incr_tmrs = 0; 9876 bbr->r_ctl.rc_inc_tcp_oh = 0; 9877 bbr->r_ctl.rc_inc_ip_oh = 0; 9878 bbr->r_ctl.rc_inc_enet_oh = 0; 9879 reset_time(&bbr->r_ctl.rc_delrate, 9880 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 9881 reset_time_small(&bbr->r_ctl.rc_rttprop, 9882 (11 * USECS_IN_SECOND)); 9883 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9884 } 9885 9886 static void 9887 bbr_google_mode_off(struct tcp_bbr *bbr) 9888 { 9889 bbr->rc_use_google = 0; 9890 bbr->r_ctl.bbr_google_discount = 0; 9891 bbr->no_pacing_until = bbr_no_pacing_until; 9892 bbr->r_use_policer = 0; 9893 if (bbr->no_pacing_until) 9894 bbr->rc_no_pacing = 1; 9895 else 9896 bbr->rc_no_pacing = 0; 9897 if (bbr_use_rack_resend_cheat) 9898 bbr->bbr_use_rack_cheat = 1; 9899 else 9900 bbr->bbr_use_rack_cheat = 0; 9901 if (bbr_incr_timers) 9902 bbr->r_ctl.rc_incr_tmrs = 1; 9903 else 9904 bbr->r_ctl.rc_incr_tmrs = 0; 9905 if (bbr_include_tcp_oh) 9906 bbr->r_ctl.rc_inc_tcp_oh = 1; 9907 else 9908 bbr->r_ctl.rc_inc_tcp_oh = 0; 9909 if (bbr_include_ip_oh) 9910 bbr->r_ctl.rc_inc_ip_oh = 1; 9911 else 9912 bbr->r_ctl.rc_inc_ip_oh = 0; 9913 if (bbr_include_enet_oh) 9914 bbr->r_ctl.rc_inc_enet_oh = 1; 9915 else 9916 bbr->r_ctl.rc_inc_enet_oh = 0; 9917 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 9918 reset_time(&bbr->r_ctl.rc_delrate, 9919 bbr_num_pktepo_for_del_limit); 9920 reset_time_small(&bbr->r_ctl.rc_rttprop, 9921 (bbr_filter_len_sec * USECS_IN_SECOND)); 9922 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9923 } 9924 /* 9925 * Return 0 on success, non-zero on failure 9926 * which indicates the error (usually no memory). 9927 */ 9928 static int 9929 bbr_init(struct tcpcb *tp) 9930 { 9931 struct inpcb *inp = tptoinpcb(tp); 9932 struct tcp_bbr *bbr = NULL; 9933 uint32_t cts; 9934 9935 tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); 9936 if (tp->t_fb_ptr == NULL) { 9937 /* 9938 * We need to allocate memory but cant. The INP and INP_INFO 9939 * locks and they are recursive (happens during setup. So a 9940 * scheme to drop the locks fails :( 9941 * 9942 */ 9943 return (ENOMEM); 9944 } 9945 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9946 bbr->rtt_valid = 0; 9947 inp->inp_flags2 |= INP_CANNOT_DO_ECN; 9948 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; 9949 TAILQ_INIT(&bbr->r_ctl.rc_map); 9950 TAILQ_INIT(&bbr->r_ctl.rc_free); 9951 TAILQ_INIT(&bbr->r_ctl.rc_tmap); 9952 bbr->rc_tp = tp; 9953 bbr->rc_inp = inp; 9954 cts = tcp_get_usecs(&bbr->rc_tv); 9955 tp->t_acktime = 0; 9956 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; 9957 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; 9958 bbr->rc_tlp_threshold = bbr_tlp_thresh; 9959 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; 9960 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; 9961 bbr->r_ctl.rc_min_to = bbr_min_to; 9962 bbr->rc_bbr_state = BBR_STATE_STARTUP; 9963 bbr->r_ctl.bbr_lost_at_state = 0; 9964 bbr->r_ctl.rc_lost_at_startup = 0; 9965 bbr->rc_all_timers_stopped = 0; 9966 bbr->r_ctl.rc_bbr_lastbtlbw = 0; 9967 bbr->r_ctl.rc_pkt_epoch_del = 0; 9968 bbr->r_ctl.rc_pkt_epoch = 0; 9969 bbr->r_ctl.rc_lowest_rtt = 0xffffffff; 9970 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; 9971 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 9972 bbr->r_ctl.rc_went_idle_time = cts; 9973 bbr->rc_pacer_started = cts; 9974 bbr->r_ctl.rc_pkt_epoch_time = cts; 9975 bbr->r_ctl.rc_rcvtime = cts; 9976 bbr->r_ctl.rc_bbr_state_time = cts; 9977 bbr->r_ctl.rc_del_time = cts; 9978 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 9979 bbr->r_ctl.last_in_probertt = cts; 9980 bbr->skip_gain = 0; 9981 bbr->gain_is_limited = 0; 9982 bbr->no_pacing_until = bbr_no_pacing_until; 9983 if (bbr->no_pacing_until) 9984 bbr->rc_no_pacing = 1; 9985 if (bbr_use_google_algo) { 9986 bbr->rc_no_pacing = 0; 9987 bbr->rc_use_google = 1; 9988 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9989 bbr->r_use_policer = bbr_policer_detection_enabled; 9990 } else { 9991 bbr->rc_use_google = 0; 9992 bbr->r_ctl.bbr_google_discount = 0; 9993 bbr->r_use_policer = 0; 9994 } 9995 if (bbr_ts_limiting) 9996 bbr->rc_use_ts_limit = 1; 9997 else 9998 bbr->rc_use_ts_limit = 0; 9999 if (bbr_ts_can_raise) 10000 bbr->ts_can_raise = 1; 10001 else 10002 bbr->ts_can_raise = 0; 10003 if (V_tcp_delack_enabled == 1) 10004 tp->t_delayed_ack = 2; 10005 else if (V_tcp_delack_enabled == 0) 10006 tp->t_delayed_ack = 0; 10007 else if (V_tcp_delack_enabled < 100) 10008 tp->t_delayed_ack = V_tcp_delack_enabled; 10009 else 10010 tp->t_delayed_ack = 2; 10011 if (bbr->rc_use_google == 0) 10012 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10013 else 10014 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 10015 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; 10016 bbr->rc_max_rto_sec = bbr_rto_max_sec; 10017 bbr->rc_init_win = bbr_def_init_win; 10018 if (tp->t_flags & TF_REQ_TSTMP) 10019 bbr->rc_last_options = TCP_TS_OVERHEAD; 10020 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; 10021 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; 10022 bbr->r_init_rtt = 1; 10023 10024 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 10025 if (bbr_allow_hdwr_pacing) 10026 bbr->bbr_hdw_pace_ena = 1; 10027 else 10028 bbr->bbr_hdw_pace_ena = 0; 10029 if (bbr_sends_full_iwnd) 10030 bbr->bbr_init_win_cheat = 1; 10031 else 10032 bbr->bbr_init_win_cheat = 0; 10033 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; 10034 bbr->r_ctl.rc_drain_pg = bbr_drain_gain; 10035 bbr->r_ctl.rc_startup_pg = bbr_high_gain; 10036 bbr->rc_loss_exit = bbr_exit_startup_at_loss; 10037 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; 10038 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; 10039 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; 10040 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; 10041 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; 10042 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; 10043 bbr->r_ctl.bbr_cross_over = bbr_cross_over; 10044 bbr->r_ctl.rc_rtt_shrinks = cts; 10045 if (bbr->rc_use_google) { 10046 setup_time_filter(&bbr->r_ctl.rc_delrate, 10047 FILTER_TYPE_MAX, 10048 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 10049 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10050 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); 10051 } else { 10052 setup_time_filter(&bbr->r_ctl.rc_delrate, 10053 FILTER_TYPE_MAX, 10054 bbr_num_pktepo_for_del_limit); 10055 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10056 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); 10057 } 10058 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); 10059 if (bbr_uses_idle_restart) 10060 bbr->rc_use_idle_restart = 1; 10061 else 10062 bbr->rc_use_idle_restart = 0; 10063 bbr->r_ctl.rc_bbr_cur_del_rate = 0; 10064 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; 10065 if (bbr_resends_use_tso) 10066 bbr->rc_resends_use_tso = 1; 10067 #ifdef NETFLIX_PEAKRATE 10068 tp->t_peakrate_thr = tp->t_maxpeakrate; 10069 #endif 10070 if (tp->snd_una != tp->snd_max) { 10071 /* Create a send map for the current outstanding data */ 10072 struct bbr_sendmap *rsm; 10073 10074 rsm = bbr_alloc(bbr); 10075 if (rsm == NULL) { 10076 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10077 tp->t_fb_ptr = NULL; 10078 return (ENOMEM); 10079 } 10080 rsm->r_rtt_not_allowed = 1; 10081 rsm->r_tim_lastsent[0] = cts; 10082 rsm->r_rtr_cnt = 1; 10083 rsm->r_rtr_bytes = 0; 10084 rsm->r_start = tp->snd_una; 10085 rsm->r_end = tp->snd_max; 10086 rsm->r_dupack = 0; 10087 rsm->r_delivered = bbr->r_ctl.rc_delivered; 10088 rsm->r_ts_valid = 0; 10089 rsm->r_del_ack_ts = tp->ts_recent; 10090 rsm->r_del_time = cts; 10091 if (bbr->r_ctl.r_app_limited_until) 10092 rsm->r_app_limited = 1; 10093 else 10094 rsm->r_app_limited = 0; 10095 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 10096 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 10097 rsm->r_in_tmap = 1; 10098 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 10099 rsm->r_bbr_state = bbr_state_val(bbr); 10100 else 10101 rsm->r_bbr_state = 8; 10102 } 10103 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) 10104 bbr->bbr_use_rack_cheat = 1; 10105 if (bbr_incr_timers && (bbr->rc_use_google == 0)) 10106 bbr->r_ctl.rc_incr_tmrs = 1; 10107 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) 10108 bbr->r_ctl.rc_inc_tcp_oh = 1; 10109 if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) 10110 bbr->r_ctl.rc_inc_ip_oh = 1; 10111 if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) 10112 bbr->r_ctl.rc_inc_enet_oh = 1; 10113 10114 bbr_log_type_statechange(bbr, cts, __LINE__); 10115 if (TCPS_HAVEESTABLISHED(tp->t_state) && 10116 (tp->t_srtt)) { 10117 uint32_t rtt; 10118 10119 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 10120 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 10121 } 10122 /* announce the settings and state */ 10123 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); 10124 tcp_bbr_tso_size_check(bbr, cts); 10125 /* 10126 * Now call the generic function to start a timer. This will place 10127 * the TCB on the hptsi wheel if a timer is needed with appropriate 10128 * flags. 10129 */ 10130 bbr_stop_all_timers(tp); 10131 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); 10132 return (0); 10133 } 10134 10135 /* 10136 * Return 0 if we can accept the connection. Return 10137 * non-zero if we can't handle the connection. A EAGAIN 10138 * means you need to wait until the connection is up. 10139 * a EADDRNOTAVAIL means we can never handle the connection 10140 * (no SACK). 10141 */ 10142 static int 10143 bbr_handoff_ok(struct tcpcb *tp) 10144 { 10145 if ((tp->t_state == TCPS_CLOSED) || 10146 (tp->t_state == TCPS_LISTEN)) { 10147 /* Sure no problem though it may not stick */ 10148 return (0); 10149 } 10150 if ((tp->t_state == TCPS_SYN_SENT) || 10151 (tp->t_state == TCPS_SYN_RECEIVED)) { 10152 /* 10153 * We really don't know you have to get to ESTAB or beyond 10154 * to tell. 10155 */ 10156 return (EAGAIN); 10157 } 10158 if (tp->t_flags & TF_SENTFIN) 10159 return (EINVAL); 10160 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { 10161 return (0); 10162 } 10163 /* 10164 * If we reach here we don't do SACK on this connection so we can 10165 * never do rack. 10166 */ 10167 return (EINVAL); 10168 } 10169 10170 static void 10171 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) 10172 { 10173 if (tp->t_fb_ptr) { 10174 struct inpcb *inp = tptoinpcb(tp); 10175 uint32_t calc; 10176 struct tcp_bbr *bbr; 10177 struct bbr_sendmap *rsm; 10178 10179 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10180 if (bbr->r_ctl.crte) 10181 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 10182 bbr_log_flowend(bbr); 10183 bbr->rc_tp = NULL; 10184 /* Backout any flags2 we applied */ 10185 inp->inp_flags2 &= ~INP_CANNOT_DO_ECN; 10186 inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; 10187 inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 10188 if (bbr->bbr_hdrw_pacing) 10189 counter_u64_add(bbr_flows_whdwr_pacing, -1); 10190 else 10191 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 10192 if (bbr->r_ctl.crte != NULL) { 10193 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 10194 bbr->r_ctl.crte = NULL; 10195 } 10196 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10197 while (rsm) { 10198 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 10199 uma_zfree(bbr_zone, rsm); 10200 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10201 } 10202 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10203 while (rsm) { 10204 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 10205 uma_zfree(bbr_zone, rsm); 10206 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10207 } 10208 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; 10209 if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) 10210 BBR_STAT_INC(bbr_dynamic_rwnd); 10211 else 10212 BBR_STAT_INC(bbr_static_rwnd); 10213 bbr->r_ctl.rc_free_cnt = 0; 10214 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10215 tp->t_fb_ptr = NULL; 10216 } 10217 /* Make sure snd_nxt is correctly set */ 10218 tp->snd_nxt = tp->snd_max; 10219 } 10220 10221 static void 10222 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) 10223 { 10224 switch (tp->t_state) { 10225 case TCPS_SYN_SENT: 10226 bbr->r_state = TCPS_SYN_SENT; 10227 bbr->r_substate = bbr_do_syn_sent; 10228 break; 10229 case TCPS_SYN_RECEIVED: 10230 bbr->r_state = TCPS_SYN_RECEIVED; 10231 bbr->r_substate = bbr_do_syn_recv; 10232 break; 10233 case TCPS_ESTABLISHED: 10234 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); 10235 bbr->r_state = TCPS_ESTABLISHED; 10236 bbr->r_substate = bbr_do_established; 10237 break; 10238 case TCPS_CLOSE_WAIT: 10239 bbr->r_state = TCPS_CLOSE_WAIT; 10240 bbr->r_substate = bbr_do_close_wait; 10241 break; 10242 case TCPS_FIN_WAIT_1: 10243 bbr->r_state = TCPS_FIN_WAIT_1; 10244 bbr->r_substate = bbr_do_fin_wait_1; 10245 break; 10246 case TCPS_CLOSING: 10247 bbr->r_state = TCPS_CLOSING; 10248 bbr->r_substate = bbr_do_closing; 10249 break; 10250 case TCPS_LAST_ACK: 10251 bbr->r_state = TCPS_LAST_ACK; 10252 bbr->r_substate = bbr_do_lastack; 10253 break; 10254 case TCPS_FIN_WAIT_2: 10255 bbr->r_state = TCPS_FIN_WAIT_2; 10256 bbr->r_substate = bbr_do_fin_wait_2; 10257 break; 10258 case TCPS_LISTEN: 10259 case TCPS_CLOSED: 10260 case TCPS_TIME_WAIT: 10261 default: 10262 break; 10263 }; 10264 } 10265 10266 static void 10267 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) 10268 { 10269 /* 10270 * Now what state are we going into now? Is there adjustments 10271 * needed? 10272 */ 10273 int32_t old_state; 10274 10275 old_state = bbr_state_val(bbr); 10276 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { 10277 /* Save the lowest srtt we saw in our end of the sub-state */ 10278 bbr->rc_hit_state_1 = 0; 10279 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) 10280 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; 10281 } 10282 bbr->rc_bbr_substate++; 10283 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { 10284 /* Cycle back to first state-> gain */ 10285 bbr->rc_bbr_substate = 0; 10286 } 10287 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10288 /* 10289 * We enter the gain(5/4) cycle (possibly less if 10290 * shallow buffer detection is enabled) 10291 */ 10292 if (bbr->skip_gain) { 10293 /* 10294 * Hardware pacing has set our rate to 10295 * the max and limited our b/w just 10296 * do level i.e. no gain. 10297 */ 10298 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; 10299 } else if (bbr->gain_is_limited && 10300 bbr->bbr_hdrw_pacing && 10301 bbr->r_ctl.crte) { 10302 /* 10303 * We can't gain above the hardware pacing 10304 * rate which is less than our rate + the gain 10305 * calculate the gain needed to reach the hardware 10306 * pacing rate.. 10307 */ 10308 uint64_t bw, rate, gain_calc; 10309 10310 bw = bbr_get_bw(bbr); 10311 rate = bbr->r_ctl.crte->rate; 10312 if ((rate > bw) && 10313 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { 10314 gain_calc = (rate * BBR_UNIT) / bw; 10315 if (gain_calc < BBR_UNIT) 10316 gain_calc = BBR_UNIT; 10317 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; 10318 } else { 10319 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10320 } 10321 } else 10322 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10323 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { 10324 bbr->r_ctl.rc_bbr_state_atflight = cts; 10325 } else 10326 bbr->r_ctl.rc_bbr_state_atflight = 0; 10327 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10328 bbr->rc_hit_state_1 = 1; 10329 bbr->r_ctl.rc_exta_time_gd = 0; 10330 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10331 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10332 if (bbr_state_drain_2_tar) { 10333 bbr->r_ctl.rc_bbr_state_atflight = 0; 10334 } else 10335 bbr->r_ctl.rc_bbr_state_atflight = cts; 10336 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; 10337 } else { 10338 /* All other cycles hit here 2-7 */ 10339 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { 10340 if (bbr_sub_drain_slam_cwnd && 10341 (bbr->rc_use_google == 0) && 10342 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10343 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10344 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10345 } 10346 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) 10347 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - 10348 bbr_get_rtt(bbr, BBR_RTT_PROP)); 10349 else 10350 bbr->r_ctl.rc_exta_time_gd = 0; 10351 if (bbr->r_ctl.rc_exta_time_gd) { 10352 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; 10353 /* Now chop up the time for each state (div by 7) */ 10354 bbr->r_ctl.rc_level_state_extra /= 7; 10355 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { 10356 /* Add a randomization */ 10357 bbr_randomize_extra_state_time(bbr); 10358 } 10359 } 10360 } 10361 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10362 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; 10363 } 10364 if (bbr->rc_use_google) { 10365 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10366 } 10367 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10368 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10369 if (dolog) 10370 bbr_log_type_statechange(bbr, cts, line); 10371 10372 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10373 uint32_t time_in; 10374 10375 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10376 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 10377 counter_u64_add(bbr_state_time[(old_state + 5)], time_in); 10378 } else { 10379 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10380 } 10381 } 10382 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 10383 bbr_set_state_target(bbr, __LINE__); 10384 if (bbr_sub_drain_slam_cwnd && 10385 (bbr->rc_use_google == 0) && 10386 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10387 /* Slam down the cwnd */ 10388 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10389 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10390 if (bbr_sub_drain_app_limit) { 10391 /* Go app limited if we are on a long drain */ 10392 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + 10393 ctf_flight_size(bbr->rc_tp, 10394 (bbr->r_ctl.rc_sacked + 10395 bbr->r_ctl.rc_lost_bytes))); 10396 } 10397 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10398 } 10399 if (bbr->rc_lt_use_bw) { 10400 /* In policed mode we clamp pacing_gain to BBR_UNIT */ 10401 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10402 } 10403 /* Google changes TSO size every cycle */ 10404 if (bbr->rc_use_google) 10405 tcp_bbr_tso_size_check(bbr, cts); 10406 bbr->r_ctl.gain_epoch = cts; 10407 bbr->r_ctl.rc_bbr_state_time = cts; 10408 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; 10409 } 10410 10411 static void 10412 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10413 { 10414 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && 10415 (google_allow_early_out == 1) && 10416 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { 10417 /* We have reached out target flight size possibly early */ 10418 goto change_state; 10419 } 10420 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10421 return; 10422 } 10423 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { 10424 /* 10425 * Must be a rttProp movement forward before 10426 * we can change states. 10427 */ 10428 return; 10429 } 10430 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10431 /* 10432 * The needed time has passed but for 10433 * the gain cycle extra rules apply: 10434 * 1) If we have seen loss, we exit 10435 * 2) If we have not reached the target 10436 * we stay in GAIN (gain-to-target). 10437 */ 10438 if (google_consider_lost && losses) 10439 goto change_state; 10440 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { 10441 return; 10442 } 10443 } 10444 change_state: 10445 /* For gain we must reach our target, all others last 1 rttProp */ 10446 bbr_substate_change(bbr, cts, __LINE__, 1); 10447 } 10448 10449 static void 10450 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10451 { 10452 uint32_t flight, bbr_cur_cycle_time; 10453 10454 if (bbr->rc_use_google) { 10455 bbr_set_probebw_google_gains(bbr, cts, losses); 10456 return; 10457 } 10458 if (cts == 0) { 10459 /* 10460 * Never alow cts to be 0 we 10461 * do this so we can judge if 10462 * we have set a timestamp. 10463 */ 10464 cts = 1; 10465 } 10466 if (bbr_state_is_pkt_epoch) 10467 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 10468 else 10469 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); 10470 10471 if (bbr->r_ctl.rc_bbr_state_atflight == 0) { 10472 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10473 flight = ctf_flight_size(bbr->rc_tp, 10474 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10475 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { 10476 /* Keep it slam down */ 10477 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { 10478 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10479 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10480 } 10481 if (bbr_sub_drain_app_limit) { 10482 /* Go app limited if we are on a long drain */ 10483 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); 10484 } 10485 } 10486 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && 10487 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || 10488 (flight >= bbr->r_ctl.flightsize_at_drain))) { 10489 /* 10490 * Still here after the same time as 10491 * the gain. We need to drain harder 10492 * for the next srtt. Reduce by a set amount 10493 * the gain drop is capped at DRAIN states 10494 * value (88). 10495 */ 10496 bbr->r_ctl.flightsize_at_drain = flight; 10497 if (bbr_drain_drop_mul && 10498 bbr_drain_drop_div && 10499 (bbr_drain_drop_mul < bbr_drain_drop_div)) { 10500 /* Use your specific drop value (def 4/5 = 20%) */ 10501 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; 10502 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; 10503 } else { 10504 /* You get drop of 20% */ 10505 bbr->r_ctl.rc_bbr_hptsi_gain *= 4; 10506 bbr->r_ctl.rc_bbr_hptsi_gain /= 5; 10507 } 10508 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { 10509 /* Reduce our gain again to the bottom */ 10510 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 10511 } 10512 bbr_log_exit_gain(bbr, cts, 4); 10513 /* 10514 * Extend out so we wait another 10515 * epoch before dropping again. 10516 */ 10517 bbr->r_ctl.gain_epoch = cts; 10518 } 10519 if (flight <= bbr->r_ctl.rc_target_at_state) { 10520 if (bbr_sub_drain_slam_cwnd && 10521 (bbr->rc_use_google == 0) && 10522 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10523 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10524 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10525 } 10526 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10527 bbr_log_exit_gain(bbr, cts, 3); 10528 } 10529 } else { 10530 /* Its a gain */ 10531 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { 10532 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10533 goto change_state; 10534 } 10535 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || 10536 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= 10537 bbr->rc_tp->snd_wnd)) { 10538 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10539 bbr_log_exit_gain(bbr, cts, 2); 10540 } 10541 } 10542 /** 10543 * We fall through and return always one of two things has 10544 * occurred. 10545 * 1) We are still not at target 10546 * <or> 10547 * 2) We reached the target and set rc_bbr_state_atflight 10548 * which means we no longer hit this block 10549 * next time we are called. 10550 */ 10551 return; 10552 } 10553 change_state: 10554 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) 10555 return; 10556 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { 10557 /* Less than a full time-period has passed */ 10558 return; 10559 } 10560 if (bbr->r_ctl.rc_level_state_extra && 10561 (bbr_state_val(bbr) > BBR_SUB_DRAIN) && 10562 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10563 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10564 /* Less than a full time-period + extra has passed */ 10565 return; 10566 } 10567 if (bbr_gain_gets_extra_too && 10568 bbr->r_ctl.rc_level_state_extra && 10569 (bbr_state_val(bbr) == BBR_SUB_GAIN) && 10570 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10571 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10572 /* Less than a full time-period + extra has passed */ 10573 return; 10574 } 10575 bbr_substate_change(bbr, cts, __LINE__, 1); 10576 } 10577 10578 static uint32_t 10579 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) 10580 { 10581 uint32_t mss, tar; 10582 10583 if (bbr->rc_use_google) { 10584 /* Google just uses the cwnd target */ 10585 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); 10586 } else { 10587 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 10588 bbr->r_ctl.rc_pace_max_segs); 10589 /* Get the base cwnd with gain rounded to a mss */ 10590 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), 10591 gain), mss); 10592 /* Make sure it is within our min */ 10593 if (tar < get_min_cwnd(bbr)) 10594 return (get_min_cwnd(bbr)); 10595 } 10596 return (tar); 10597 } 10598 10599 static void 10600 bbr_set_state_target(struct tcp_bbr *bbr, int line) 10601 { 10602 uint32_t tar, meth; 10603 10604 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 10605 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 10606 /* Special case using old probe-rtt method */ 10607 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10608 meth = 1; 10609 } else { 10610 /* Non-probe-rtt case and reduced probe-rtt */ 10611 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 10612 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { 10613 /* For gain cycle we use the hptsi gain */ 10614 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10615 meth = 2; 10616 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { 10617 /* 10618 * If configured, or for google all other states 10619 * get BBR_UNIT. 10620 */ 10621 tar = bbr_get_a_state_target(bbr, BBR_UNIT); 10622 meth = 3; 10623 } else { 10624 /* 10625 * Or we set a target based on the pacing gain 10626 * for non-google mode and default (non-configured). 10627 * Note we don't set a target goal below drain (192). 10628 */ 10629 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { 10630 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); 10631 meth = 4; 10632 } else { 10633 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10634 meth = 5; 10635 } 10636 } 10637 } 10638 bbr_log_set_of_state_target(bbr, tar, line, meth); 10639 bbr->r_ctl.rc_target_at_state = tar; 10640 } 10641 10642 static void 10643 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 10644 { 10645 /* Change to probe_rtt */ 10646 uint32_t time_in; 10647 10648 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10649 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10650 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10651 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain 10652 + bbr->r_ctl.rc_delivered); 10653 /* Setup so we force feed the filter */ 10654 if (bbr->rc_use_google || bbr_probertt_sets_rtt) 10655 bbr->rc_prtt_set_ts = 1; 10656 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10657 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10658 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10659 } 10660 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); 10661 bbr->r_ctl.rc_rtt_shrinks = cts; 10662 bbr->r_ctl.last_in_probertt = cts; 10663 bbr->r_ctl.rc_probertt_srttchktim = cts; 10664 bbr->r_ctl.rc_bbr_state_time = cts; 10665 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; 10666 /* We need to force the filter to update */ 10667 10668 if ((bbr_sub_drain_slam_cwnd) && 10669 bbr->rc_hit_state_1 && 10670 (bbr->rc_use_google == 0) && 10671 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10672 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) 10673 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10674 } else 10675 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10676 /* Update the lost */ 10677 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10678 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ 10679 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ 10680 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10681 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10682 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10683 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10684 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); 10685 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; 10686 } else { 10687 /* 10688 * We bring it down slowly by using a hptsi gain that is 10689 * probably 75%. This will slowly float down our outstanding 10690 * without tampering with the cwnd. 10691 */ 10692 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 10693 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10694 bbr_set_state_target(bbr, __LINE__); 10695 if (bbr_prtt_slam_cwnd && 10696 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 10697 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10698 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10699 } 10700 } 10701 if (ctf_flight_size(bbr->rc_tp, 10702 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 10703 bbr->r_ctl.rc_target_at_state) { 10704 /* We are at target */ 10705 bbr->r_ctl.rc_bbr_enters_probertt = cts; 10706 } else { 10707 /* We need to come down to reach target before our time begins */ 10708 bbr->r_ctl.rc_bbr_enters_probertt = 0; 10709 } 10710 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; 10711 BBR_STAT_INC(bbr_enter_probertt); 10712 bbr_log_exit_gain(bbr, cts, 0); 10713 bbr_log_type_statechange(bbr, cts, line); 10714 } 10715 10716 static void 10717 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) 10718 { 10719 /* 10720 * Sanity check on probe-rtt intervals. 10721 * In crazy situations where we are competing 10722 * against new-reno flows with huge buffers 10723 * our rtt-prop interval could come to dominate 10724 * things if we can't get through a full set 10725 * of cycles, we need to adjust it. 10726 */ 10727 if (bbr_can_adjust_probertt && 10728 (bbr->rc_use_google == 0)) { 10729 uint16_t val = 0; 10730 uint32_t cur_rttp, fval, newval, baseval; 10731 10732 /* Are we to small and go into probe-rtt to often? */ 10733 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); 10734 cur_rttp = roundup(baseval, USECS_IN_SECOND); 10735 fval = bbr_filter_len_sec * USECS_IN_SECOND; 10736 if (bbr_is_ratio == 0) { 10737 if (fval > bbr_rtt_probe_limit) 10738 newval = cur_rttp + (fval - bbr_rtt_probe_limit); 10739 else 10740 newval = cur_rttp; 10741 } else { 10742 int mul; 10743 10744 mul = fval / bbr_rtt_probe_limit; 10745 newval = cur_rttp * mul; 10746 } 10747 if (cur_rttp > bbr->r_ctl.rc_probertt_int) { 10748 bbr->r_ctl.rc_probertt_int = cur_rttp; 10749 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10750 val = 1; 10751 } else { 10752 /* 10753 * No adjustments were made 10754 * do we need to shrink it? 10755 */ 10756 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { 10757 if (cur_rttp <= bbr_rtt_probe_limit) { 10758 /* 10759 * Things have calmed down lets 10760 * shrink all the way to default 10761 */ 10762 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10763 reset_time_small(&bbr->r_ctl.rc_rttprop, 10764 (bbr_filter_len_sec * USECS_IN_SECOND)); 10765 cur_rttp = bbr_rtt_probe_limit; 10766 newval = (bbr_filter_len_sec * USECS_IN_SECOND); 10767 val = 2; 10768 } else { 10769 /* 10770 * Well does some adjustment make sense? 10771 */ 10772 if (cur_rttp < bbr->r_ctl.rc_probertt_int) { 10773 /* We can reduce interval time some */ 10774 bbr->r_ctl.rc_probertt_int = cur_rttp; 10775 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10776 val = 3; 10777 } 10778 } 10779 } 10780 } 10781 if (val) 10782 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); 10783 } 10784 } 10785 10786 static void 10787 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 10788 { 10789 /* Exit probe-rtt */ 10790 10791 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { 10792 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10793 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10794 } 10795 bbr_log_exit_gain(bbr, cts, 1); 10796 bbr->rc_hit_state_1 = 0; 10797 bbr->r_ctl.rc_rtt_shrinks = cts; 10798 bbr->r_ctl.last_in_probertt = cts; 10799 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); 10800 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10801 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, 10802 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 10803 bbr->r_ctl.rc_delivered); 10804 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10805 uint32_t time_in; 10806 10807 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10808 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10809 } 10810 if (bbr->rc_filled_pipe) { 10811 /* Switch to probe_bw */ 10812 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 10813 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 10814 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10815 bbr_substate_change(bbr, cts, __LINE__, 0); 10816 bbr_log_type_statechange(bbr, cts, __LINE__); 10817 } else { 10818 /* Back to startup */ 10819 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10820 bbr->r_ctl.rc_bbr_state_time = cts; 10821 /* 10822 * We don't want to give a complete free 3 10823 * measurements until we exit, so we use 10824 * the number of pe's we were in probe-rtt 10825 * to add to the startup_epoch. That way 10826 * we will still retain the old state. 10827 */ 10828 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); 10829 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10830 /* Make sure to use the lower pg when shifting back in */ 10831 if (bbr->r_ctl.rc_lost && 10832 bbr_use_lower_gain_in_startup && 10833 (bbr->rc_use_google == 0)) 10834 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10835 else 10836 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 10837 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 10838 /* Probably not needed but set it anyway */ 10839 bbr_set_state_target(bbr, __LINE__); 10840 bbr_log_type_statechange(bbr, cts, __LINE__); 10841 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10842 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); 10843 } 10844 bbr_check_probe_rtt_limits(bbr, cts); 10845 } 10846 10847 static int32_t inline 10848 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) 10849 { 10850 if ((bbr->rc_past_init_win == 1) && 10851 (bbr->rc_in_persist == 0) && 10852 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { 10853 return (1); 10854 } 10855 if (bbr_can_force_probertt && 10856 (bbr->rc_in_persist == 0) && 10857 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 10858 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 10859 return (1); 10860 } 10861 return (0); 10862 } 10863 10864 static int32_t 10865 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) 10866 { 10867 uint64_t btlbw, gain; 10868 if (pkt_epoch == 0) { 10869 /* 10870 * Need to be on a pkt-epoch to continue. 10871 */ 10872 return (0); 10873 } 10874 btlbw = bbr_get_full_bw(bbr); 10875 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10876 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10877 if (btlbw >= gain) { 10878 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10879 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10880 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10881 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10882 } 10883 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) 10884 return (1); 10885 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10886 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 10887 return(0); 10888 } 10889 10890 static int32_t inline 10891 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) 10892 { 10893 /* Have we gained 25% in the last 3 packet based epoch's? */ 10894 uint64_t btlbw, gain; 10895 int do_exit; 10896 int delta, rtt_gain; 10897 10898 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 10899 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 10900 /* 10901 * This qualifies as a RTT_PROBE session since we drop the 10902 * data outstanding to nothing and waited more than 10903 * bbr_rtt_probe_time. 10904 */ 10905 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 10906 bbr_set_reduced_rtt(bbr, cts, __LINE__); 10907 } 10908 if (bbr_should_enter_probe_rtt(bbr, cts)) { 10909 bbr_enter_probe_rtt(bbr, cts, __LINE__); 10910 return (0); 10911 } 10912 if (bbr->rc_use_google) 10913 return (bbr_google_startup(bbr, cts, pkt_epoch)); 10914 10915 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10916 (bbr_use_lower_gain_in_startup)) { 10917 /* Drop to a lower gain 1.5 x since we saw loss */ 10918 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10919 } 10920 if (pkt_epoch == 0) { 10921 /* 10922 * Need to be on a pkt-epoch to continue. 10923 */ 10924 return (0); 10925 } 10926 if (bbr_rtt_gain_thresh) { 10927 /* 10928 * Do we allow a flow to stay 10929 * in startup with no loss and no 10930 * gain in rtt over a set threshold? 10931 */ 10932 if (bbr->r_ctl.rc_pkt_epoch_rtt && 10933 bbr->r_ctl.startup_last_srtt && 10934 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { 10935 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; 10936 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; 10937 } else 10938 rtt_gain = 0; 10939 if ((bbr->r_ctl.startup_last_srtt == 0) || 10940 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) 10941 /* First time or new lower value */ 10942 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 10943 10944 if ((bbr->r_ctl.rc_lost == 0) && 10945 (rtt_gain < bbr_rtt_gain_thresh)) { 10946 /* 10947 * No loss, and we are under 10948 * our gain threhold for 10949 * increasing RTT. 10950 */ 10951 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10952 bbr->r_ctl.rc_bbr_last_startup_epoch++; 10953 bbr_log_startup_event(bbr, cts, rtt_gain, 10954 delta, bbr->r_ctl.startup_last_srtt, 10); 10955 return (0); 10956 } 10957 } 10958 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && 10959 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && 10960 (!IN_RECOVERY(bbr->rc_tp->t_flags))) { 10961 /* 10962 * We only assess if we have a new measurement when 10963 * we have no loss and are not in recovery. 10964 * Drag up by one our last_startup epoch so we will hold 10965 * the number of non-gain we have already accumulated. 10966 */ 10967 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10968 bbr->r_ctl.rc_bbr_last_startup_epoch++; 10969 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10970 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); 10971 return (0); 10972 } 10973 /* Case where we reduced the lost (bad retransmit) */ 10974 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) 10975 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10976 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; 10977 btlbw = bbr_get_full_bw(bbr); 10978 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) 10979 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10980 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10981 else 10982 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10983 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10984 do_exit = 0; 10985 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) 10986 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10987 if (btlbw >= gain) { 10988 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10989 /* Update the lost so we won't exit in next set of tests */ 10990 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10991 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10992 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10993 } 10994 if ((bbr->rc_loss_exit && 10995 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10996 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && 10997 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { 10998 /* 10999 * If we had no gain, we had loss and that loss was above 11000 * our threshould, the rwnd is not constrained, and we have 11001 * had at least 3 packet epochs exit. Note that this is 11002 * switched off by sysctl. Google does not do this by the 11003 * way. 11004 */ 11005 if ((ctf_flight_size(bbr->rc_tp, 11006 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 11007 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { 11008 do_exit = 1; 11009 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11010 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); 11011 } else { 11012 /* Just record an updated loss value */ 11013 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11014 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11015 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); 11016 } 11017 } else 11018 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11019 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || 11020 do_exit) { 11021 /* Return 1 to exit the startup state. */ 11022 return (1); 11023 } 11024 /* Stay in startup */ 11025 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11026 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 11027 return (0); 11028 } 11029 11030 static void 11031 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) 11032 { 11033 /* 11034 * A tick occurred in the rtt epoch do we need to do anything? 11035 */ 11036 #ifdef BBR_INVARIANTS 11037 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && 11038 (bbr->rc_bbr_state != BBR_STATE_DRAIN) && 11039 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && 11040 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 11041 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { 11042 /* Debug code? */ 11043 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); 11044 } 11045 #endif 11046 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 11047 /* Do we exit the startup state? */ 11048 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { 11049 uint32_t time_in; 11050 11051 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11052 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); 11053 bbr->rc_filled_pipe = 1; 11054 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11055 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11056 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11057 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11058 } else 11059 time_in = 0; 11060 if (bbr->rc_no_pacing) 11061 bbr->rc_no_pacing = 0; 11062 bbr->r_ctl.rc_bbr_state_time = cts; 11063 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; 11064 bbr->rc_bbr_state = BBR_STATE_DRAIN; 11065 bbr_set_state_target(bbr, __LINE__); 11066 if ((bbr->rc_use_google == 0) && 11067 bbr_slam_cwnd_in_main_drain) { 11068 /* Here we don't have to worry about probe-rtt */ 11069 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 11070 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11071 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11072 } 11073 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 11074 bbr_log_type_statechange(bbr, cts, __LINE__); 11075 if (ctf_flight_size(bbr->rc_tp, 11076 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 11077 bbr->r_ctl.rc_target_at_state) { 11078 /* 11079 * Switch to probe_bw if we are already 11080 * there 11081 */ 11082 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11083 bbr_substate_change(bbr, cts, __LINE__, 0); 11084 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11085 bbr_log_type_statechange(bbr, cts, __LINE__); 11086 } 11087 } 11088 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { 11089 uint32_t inflight; 11090 struct tcpcb *tp; 11091 11092 tp = bbr->rc_tp; 11093 inflight = ctf_flight_size(tp, 11094 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11095 if (inflight >= bbr->r_ctl.rc_target_at_state) { 11096 /* We have reached a flight of the cwnd target */ 11097 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11098 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11099 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 11100 bbr_set_state_target(bbr, __LINE__); 11101 /* 11102 * Rig it so we don't do anything crazy and 11103 * start fresh with a new randomization. 11104 */ 11105 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 11106 bbr->rc_bbr_substate = BBR_SUB_LEVEL6; 11107 bbr_substate_change(bbr, cts, __LINE__, 1); 11108 } 11109 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { 11110 /* Has in-flight reached the bdp (or less)? */ 11111 uint32_t inflight; 11112 struct tcpcb *tp; 11113 11114 tp = bbr->rc_tp; 11115 inflight = ctf_flight_size(tp, 11116 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11117 if ((bbr->rc_use_google == 0) && 11118 bbr_slam_cwnd_in_main_drain && 11119 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11120 /* 11121 * Here we don't have to worry about probe-rtt 11122 * re-slam it, but keep it slammed down. 11123 */ 11124 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11125 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11126 } 11127 if (inflight <= bbr->r_ctl.rc_target_at_state) { 11128 /* We have drained */ 11129 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11130 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11131 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11132 uint32_t time_in; 11133 11134 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11135 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11136 } 11137 if ((bbr->rc_use_google == 0) && 11138 bbr_slam_cwnd_in_main_drain && 11139 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 11140 /* Restore the cwnd */ 11141 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 11142 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11143 } 11144 /* Setup probe-rtt has being done now RRS-HERE */ 11145 bbr->r_ctl.rc_rtt_shrinks = cts; 11146 bbr->r_ctl.last_in_probertt = cts; 11147 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); 11148 /* Randomly pick a sub-state */ 11149 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11150 bbr_substate_change(bbr, cts, __LINE__, 0); 11151 bbr_log_type_statechange(bbr, cts, __LINE__); 11152 } 11153 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { 11154 uint32_t flight; 11155 11156 flight = ctf_flight_size(bbr->rc_tp, 11157 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11158 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); 11159 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && 11160 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11161 /* 11162 * We must keep cwnd at the desired MSS. 11163 */ 11164 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 11165 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11166 } else if ((bbr_prtt_slam_cwnd) && 11167 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11168 /* Re-slam it */ 11169 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11170 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11171 } 11172 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { 11173 /* Has outstanding reached our target? */ 11174 if (flight <= bbr->r_ctl.rc_target_at_state) { 11175 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); 11176 bbr->r_ctl.rc_bbr_enters_probertt = cts; 11177 /* If time is exactly 0, be 1usec off */ 11178 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) 11179 bbr->r_ctl.rc_bbr_enters_probertt = 1; 11180 if (bbr->rc_use_google == 0) { 11181 /* 11182 * Restore any lowering that as occurred to 11183 * reach here 11184 */ 11185 if (bbr->r_ctl.bbr_rttprobe_gain_val) 11186 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 11187 else 11188 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11189 } 11190 } 11191 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && 11192 (bbr->rc_use_google == 0) && 11193 bbr->r_ctl.bbr_rttprobe_gain_val && 11194 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || 11195 (flight >= bbr->r_ctl.flightsize_at_drain))) { 11196 /* 11197 * We have doddled with our current hptsi 11198 * gain an srtt and have still not made it 11199 * to target, or we have increased our flight. 11200 * Lets reduce the gain by xx% 11201 * flooring the reduce at DRAIN (based on 11202 * mul/div) 11203 */ 11204 int red; 11205 11206 bbr->r_ctl.flightsize_at_drain = flight; 11207 bbr->r_ctl.rc_probertt_srttchktim = cts; 11208 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); 11209 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { 11210 /* Reduce our gain again */ 11211 bbr->r_ctl.rc_bbr_hptsi_gain -= red; 11212 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); 11213 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { 11214 /* one more chance before we give up */ 11215 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 11216 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); 11217 } else { 11218 /* At the very bottom */ 11219 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); 11220 } 11221 } 11222 } 11223 if (bbr->r_ctl.rc_bbr_enters_probertt && 11224 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && 11225 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { 11226 /* Time to exit probe RTT normally */ 11227 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); 11228 } 11229 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 11230 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 11231 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 11232 /* 11233 * This qualifies as a RTT_PROBE session since we 11234 * drop the data outstanding to nothing and waited 11235 * more than bbr_rtt_probe_time. 11236 */ 11237 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 11238 bbr_set_reduced_rtt(bbr, cts, __LINE__); 11239 } 11240 if (bbr_should_enter_probe_rtt(bbr, cts)) { 11241 bbr_enter_probe_rtt(bbr, cts, __LINE__); 11242 } else { 11243 bbr_set_probebw_gains(bbr, cts, losses); 11244 } 11245 } 11246 } 11247 11248 static void 11249 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) 11250 { 11251 int32_t epoch = 0; 11252 11253 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 11254 bbr_set_epoch(bbr, cts, line); 11255 /* At each epoch doe lt bw sampling */ 11256 epoch = 1; 11257 } 11258 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); 11259 } 11260 11261 static int 11262 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, 11263 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 11264 int32_t nxt_pkt, struct timeval *tv) 11265 { 11266 struct inpcb *inp = tptoinpcb(tp); 11267 int32_t thflags, retval; 11268 uint32_t cts, lcts; 11269 uint32_t tiwin; 11270 struct tcpopt to; 11271 struct tcp_bbr *bbr; 11272 struct bbr_sendmap *rsm; 11273 struct timeval ltv; 11274 int32_t did_out = 0; 11275 uint16_t nsegs; 11276 int32_t prev_state; 11277 uint32_t lost; 11278 11279 nsegs = max(1, m->m_pkthdr.lro_nsegs); 11280 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11281 /* add in our stats */ 11282 kern_prefetch(bbr, &prev_state); 11283 prev_state = 0; 11284 thflags = tcp_get_flags(th); 11285 /* 11286 * If this is either a state-changing packet or current state isn't 11287 * established, we require a write lock on tcbinfo. Otherwise, we 11288 * allow the tcbinfo to be in either alocked or unlocked, as the 11289 * caller may have unnecessarily acquired a write lock due to a 11290 * race. 11291 */ 11292 INP_WLOCK_ASSERT(tptoinpcb(tp)); 11293 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 11294 __func__)); 11295 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 11296 __func__)); 11297 11298 tp->t_rcvtime = ticks; 11299 /* 11300 * Unscale the window into a 32-bit value. For the SYN_SENT state 11301 * the scale is zero. 11302 */ 11303 tiwin = th->th_win << tp->snd_scale; 11304 #ifdef STATS 11305 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 11306 #endif 11307 11308 if (m->m_flags & M_TSTMP) { 11309 /* Prefer the hardware timestamp if present */ 11310 struct timespec ts; 11311 11312 mbuf_tstmp2timespec(m, &ts); 11313 bbr->rc_tv.tv_sec = ts.tv_sec; 11314 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11315 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11316 } else if (m->m_flags & M_TSTMP_LRO) { 11317 /* Next the arrival timestamp */ 11318 struct timespec ts; 11319 11320 mbuf_tstmp2timespec(m, &ts); 11321 bbr->rc_tv.tv_sec = ts.tv_sec; 11322 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11323 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11324 } else { 11325 /* 11326 * Ok just get the current time. 11327 */ 11328 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); 11329 } 11330 /* 11331 * Parse options on any incoming segment. 11332 */ 11333 tcp_dooptions(&to, (u_char *)(th + 1), 11334 (th->th_off << 2) - sizeof(struct tcphdr), 11335 (thflags & TH_SYN) ? TO_SYN : 0); 11336 11337 /* 11338 * If timestamps were negotiated during SYN/ACK and a 11339 * segment without a timestamp is received, silently drop 11340 * the segment, unless it is a RST segment or missing timestamps are 11341 * tolerated. 11342 * See section 3.2 of RFC 7323. 11343 */ 11344 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && 11345 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { 11346 retval = 0; 11347 m_freem(m); 11348 goto done_with_input; 11349 } 11350 /* 11351 * If echoed timestamp is later than the current time, fall back to 11352 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 11353 * were used when this connection was established. 11354 */ 11355 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 11356 to.to_tsecr -= tp->ts_offset; 11357 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) 11358 to.to_tsecr = 0; 11359 } 11360 /* 11361 * If its the first time in we need to take care of options and 11362 * verify we can do SACK for rack! 11363 */ 11364 if (bbr->r_state == 0) { 11365 /* 11366 * Process options only when we get SYN/ACK back. The SYN 11367 * case for incoming connections is handled in tcp_syncache. 11368 * According to RFC1323 the window field in a SYN (i.e., a 11369 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 11370 * this is traditional behavior, may need to be cleaned up. 11371 */ 11372 if (bbr->rc_inp == NULL) { 11373 bbr->rc_inp = inp; 11374 } 11375 /* 11376 * We need to init rc_inp here since its not init'd when 11377 * bbr_init is called 11378 */ 11379 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 11380 if ((to.to_flags & TOF_SCALE) && 11381 (tp->t_flags & TF_REQ_SCALE)) { 11382 tp->t_flags |= TF_RCVD_SCALE; 11383 tp->snd_scale = to.to_wscale; 11384 } else 11385 tp->t_flags &= ~TF_REQ_SCALE; 11386 /* 11387 * Initial send window. It will be updated with the 11388 * next incoming segment to the scaled value. 11389 */ 11390 tp->snd_wnd = th->th_win; 11391 if ((to.to_flags & TOF_TS) && 11392 (tp->t_flags & TF_REQ_TSTMP)) { 11393 tp->t_flags |= TF_RCVD_TSTMP; 11394 tp->ts_recent = to.to_tsval; 11395 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 11396 } else 11397 tp->t_flags &= ~TF_REQ_TSTMP; 11398 if (to.to_flags & TOF_MSS) 11399 tcp_mss(tp, to.to_mss); 11400 if ((tp->t_flags & TF_SACK_PERMIT) && 11401 (to.to_flags & TOF_SACKPERM) == 0) 11402 tp->t_flags &= ~TF_SACK_PERMIT; 11403 if (IS_FASTOPEN(tp->t_flags)) { 11404 if (to.to_flags & TOF_FASTOPEN) { 11405 uint16_t mss; 11406 11407 if (to.to_flags & TOF_MSS) 11408 mss = to.to_mss; 11409 else 11410 if ((inp->inp_vflag & INP_IPV6) != 0) 11411 mss = TCP6_MSS; 11412 else 11413 mss = TCP_MSS; 11414 tcp_fastopen_update_cache(tp, mss, 11415 to.to_tfo_len, to.to_tfo_cookie); 11416 } else 11417 tcp_fastopen_disable_path(tp); 11418 } 11419 } 11420 /* 11421 * At this point we are at the initial call. Here we decide 11422 * if we are doing RACK or not. We do this by seeing if 11423 * TF_SACK_PERMIT is set, if not rack is *not* possible and 11424 * we switch to the default code. 11425 */ 11426 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 11427 /* Bail */ 11428 tcp_switch_back_to_default(tp); 11429 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 11430 tlen, iptos); 11431 return (1); 11432 } 11433 /* Set the flag */ 11434 bbr->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0; 11435 tcp_set_hpts(inp); 11436 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); 11437 } 11438 if (thflags & TH_ACK) { 11439 /* Track ack types */ 11440 if (to.to_flags & TOF_SACK) 11441 BBR_STAT_INC(bbr_acks_with_sacks); 11442 else 11443 BBR_STAT_INC(bbr_plain_acks); 11444 } 11445 /* 11446 * This is the one exception case where we set the rack state 11447 * always. All other times (timers etc) we must have a rack-state 11448 * set (so we assure we have done the checks above for SACK). 11449 */ 11450 if (thflags & TH_FIN) 11451 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); 11452 if (bbr->r_state != tp->t_state) 11453 bbr_set_state(tp, bbr, tiwin); 11454 11455 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) 11456 kern_prefetch(rsm, &prev_state); 11457 prev_state = bbr->r_state; 11458 bbr->rc_ack_was_delayed = 0; 11459 lost = bbr->r_ctl.rc_lost; 11460 bbr->rc_is_pkt_epoch_now = 0; 11461 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { 11462 /* Get the real time into lcts and figure the real delay */ 11463 lcts = tcp_get_usecs(<v); 11464 if (TSTMP_GT(lcts, cts)) { 11465 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; 11466 bbr->rc_ack_was_delayed = 1; 11467 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, 11468 bbr->r_ctl.highest_hdwr_delay)) 11469 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; 11470 } else { 11471 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11472 bbr->rc_ack_was_delayed = 0; 11473 } 11474 } else { 11475 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11476 bbr->rc_ack_was_delayed = 0; 11477 } 11478 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); 11479 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 11480 retval = 0; 11481 m_freem(m); 11482 goto done_with_input; 11483 } 11484 /* 11485 * If a segment with the ACK-bit set arrives in the SYN-SENT state 11486 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 11487 */ 11488 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 11489 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 11490 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 11491 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 11492 return (1); 11493 } 11494 if (tiwin > bbr->r_ctl.rc_high_rwnd) 11495 bbr->r_ctl.rc_high_rwnd = tiwin; 11496 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, 11497 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11498 bbr->rtt_valid = 0; 11499 if (to.to_flags & TOF_TS) { 11500 bbr->rc_ts_valid = 1; 11501 bbr->r_ctl.last_inbound_ts = to.to_tsval; 11502 } else { 11503 bbr->rc_ts_valid = 0; 11504 bbr->r_ctl.last_inbound_ts = 0; 11505 } 11506 retval = (*bbr->r_substate) (m, th, so, 11507 tp, &to, drop_hdrlen, 11508 tlen, tiwin, thflags, nxt_pkt, iptos); 11509 if (nxt_pkt == 0) 11510 BBR_STAT_INC(bbr_rlock_left_ret0); 11511 else 11512 BBR_STAT_INC(bbr_rlock_left_ret1); 11513 if (retval == 0) { 11514 /* 11515 * If retval is 1 the tcb is unlocked and most likely the tp 11516 * is gone. 11517 */ 11518 INP_WLOCK_ASSERT(inp); 11519 tcp_bbr_xmit_timer_commit(bbr, tp, cts); 11520 if (bbr->rc_is_pkt_epoch_now) 11521 bbr_set_pktepoch(bbr, cts, __LINE__); 11522 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); 11523 if (nxt_pkt == 0) { 11524 if (bbr->r_wanted_output != 0) { 11525 bbr->rc_output_starts_timer = 0; 11526 did_out = 1; 11527 if (tcp_output(tp) < 0) 11528 return (1); 11529 } else 11530 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); 11531 } 11532 if ((nxt_pkt == 0) && 11533 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 11534 (SEQ_GT(tp->snd_max, tp->snd_una) || 11535 (tp->t_flags & TF_DELACK) || 11536 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 11537 (tp->t_state <= TCPS_CLOSING)))) { 11538 /* 11539 * We could not send (probably in the hpts but 11540 * stopped the timer)? 11541 */ 11542 if ((tp->snd_max == tp->snd_una) && 11543 ((tp->t_flags & TF_DELACK) == 0) && 11544 (tcp_in_hpts(bbr->rc_inp)) && 11545 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 11546 /* 11547 * keep alive not needed if we are hptsi 11548 * output yet 11549 */ 11550 ; 11551 } else { 11552 if (tcp_in_hpts(bbr->rc_inp)) { 11553 tcp_hpts_remove(bbr->rc_inp); 11554 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11555 (TSTMP_GT(lcts, bbr->rc_pacer_started))) { 11556 uint32_t del; 11557 11558 del = lcts - bbr->rc_pacer_started; 11559 if (bbr->r_ctl.rc_last_delay_val > del) { 11560 BBR_STAT_INC(bbr_force_timer_start); 11561 bbr->r_ctl.rc_last_delay_val -= del; 11562 bbr->rc_pacer_started = lcts; 11563 } else { 11564 /* We are late */ 11565 bbr->r_ctl.rc_last_delay_val = 0; 11566 BBR_STAT_INC(bbr_force_output); 11567 if (tcp_output(tp) < 0) 11568 return (1); 11569 } 11570 } 11571 } 11572 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 11573 0); 11574 } 11575 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { 11576 /* Do we have the correct timer running? */ 11577 bbr_timer_audit(tp, bbr, lcts, &so->so_snd); 11578 } 11579 /* Do we have a new state */ 11580 if (bbr->r_state != tp->t_state) 11581 bbr_set_state(tp, bbr, tiwin); 11582 done_with_input: 11583 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); 11584 if (did_out) 11585 bbr->r_wanted_output = 0; 11586 } 11587 return (retval); 11588 } 11589 11590 static void 11591 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 11592 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 11593 { 11594 struct timeval tv; 11595 int retval; 11596 11597 /* First lets see if we have old packets */ 11598 if (tp->t_in_pkt) { 11599 if (ctf_do_queued_segments(so, tp, 1)) { 11600 m_freem(m); 11601 return; 11602 } 11603 } 11604 if (m->m_flags & M_TSTMP_LRO) { 11605 mbuf_tstmp2timeval(m, &tv); 11606 } else { 11607 /* Should not be should we kassert instead? */ 11608 tcp_get_usecs(&tv); 11609 } 11610 retval = bbr_do_segment_nounlock(m, th, so, tp, 11611 drop_hdrlen, tlen, iptos, 0, &tv); 11612 if (retval == 0) { 11613 INP_WUNLOCK(tptoinpcb(tp)); 11614 } 11615 } 11616 11617 /* 11618 * Return how much data can be sent without violating the 11619 * cwnd or rwnd. 11620 */ 11621 11622 static inline uint32_t 11623 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, 11624 uint32_t avail, int32_t sb_offset, uint32_t cts) 11625 { 11626 uint32_t len; 11627 11628 if (ctf_outstanding(tp) >= tp->snd_wnd) { 11629 /* We never want to go over our peers rcv-window */ 11630 len = 0; 11631 } else { 11632 uint32_t flight; 11633 11634 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11635 if (flight >= sendwin) { 11636 /* 11637 * We have in flight what we are allowed by cwnd (if 11638 * it was rwnd blocking it would have hit above out 11639 * >= tp->snd_wnd). 11640 */ 11641 return (0); 11642 } 11643 len = sendwin - flight; 11644 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { 11645 /* We would send too much (beyond the rwnd) */ 11646 len = tp->snd_wnd - ctf_outstanding(tp); 11647 } 11648 if ((len + sb_offset) > avail) { 11649 /* 11650 * We don't have that much in the SB, how much is 11651 * there? 11652 */ 11653 len = avail - sb_offset; 11654 } 11655 } 11656 return (len); 11657 } 11658 11659 static inline void 11660 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11661 { 11662 #ifdef NETFLIX_STATS 11663 KMOD_TCPSTAT_INC(tcps_sndpack_error); 11664 KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); 11665 #endif 11666 } 11667 11668 static inline void 11669 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11670 { 11671 if (error) { 11672 bbr_do_error_accounting(tp, bbr, rsm, len, error); 11673 return; 11674 } 11675 if (rsm) { 11676 if (rsm->r_flags & BBR_TLP) { 11677 /* 11678 * TLP should not count in retran count, but in its 11679 * own bin 11680 */ 11681 #ifdef NETFLIX_STATS 11682 KMOD_TCPSTAT_INC(tcps_tlpresends); 11683 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); 11684 #endif 11685 } else { 11686 /* Retransmit */ 11687 tp->t_sndrexmitpack++; 11688 KMOD_TCPSTAT_INC(tcps_sndrexmitpack); 11689 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); 11690 #ifdef STATS 11691 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 11692 len); 11693 #endif 11694 } 11695 /* 11696 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is 11697 * sub-state 11698 */ 11699 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); 11700 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { 11701 /* Non probe_bw log in 1, 2, or 4. */ 11702 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); 11703 } else { 11704 /* 11705 * Log our probe state 3, and log also 5-13 to show 11706 * us the recovery sub-state for the send. This 11707 * means that 3 == (5+6+7+8+9+10+11+12+13) 11708 */ 11709 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); 11710 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); 11711 } 11712 /* Place in both 16's the totals of retransmitted */ 11713 counter_u64_add(bbr_state_lost[16], len); 11714 counter_u64_add(bbr_state_resend[16], len); 11715 /* Place in 17's the total sent */ 11716 counter_u64_add(bbr_state_resend[17], len); 11717 counter_u64_add(bbr_state_lost[17], len); 11718 11719 } else { 11720 /* New sends */ 11721 KMOD_TCPSTAT_INC(tcps_sndpack); 11722 KMOD_TCPSTAT_ADD(tcps_sndbyte, len); 11723 /* Place in 17's the total sent */ 11724 counter_u64_add(bbr_state_resend[17], len); 11725 counter_u64_add(bbr_state_lost[17], len); 11726 #ifdef STATS 11727 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 11728 len); 11729 #endif 11730 } 11731 } 11732 11733 static void 11734 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) 11735 { 11736 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { 11737 /* 11738 * Limit the cwnd to not be above N x the target plus whats 11739 * is outstanding. The target is based on the current b/w 11740 * estimate. 11741 */ 11742 uint32_t target; 11743 11744 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); 11745 target += ctf_outstanding(tp); 11746 target *= bbr_target_cwnd_mult_limit; 11747 if (tp->snd_cwnd > target) 11748 tp->snd_cwnd = target; 11749 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); 11750 } 11751 } 11752 11753 static int 11754 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) 11755 { 11756 /* 11757 * "adv" is the amount we could increase the window, taking into 11758 * account that we are limited by TCP_MAXWIN << tp->rcv_scale. 11759 */ 11760 int32_t adv; 11761 int32_t oldwin; 11762 11763 adv = recwin; 11764 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 11765 oldwin = (tp->rcv_adv - tp->rcv_nxt); 11766 if (adv > oldwin) 11767 adv -= oldwin; 11768 else { 11769 /* We can't increase the window */ 11770 adv = 0; 11771 } 11772 } else 11773 oldwin = 0; 11774 11775 /* 11776 * If the new window size ends up being the same as or less 11777 * than the old size when it is scaled, then don't force 11778 * a window update. 11779 */ 11780 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 11781 return (0); 11782 11783 if (adv >= (2 * maxseg) && 11784 (adv >= (so->so_rcv.sb_hiwat / 4) || 11785 recwin <= (so->so_rcv.sb_hiwat / 8) || 11786 so->so_rcv.sb_hiwat <= 8 * maxseg)) { 11787 return (1); 11788 } 11789 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 11790 return (1); 11791 return (0); 11792 } 11793 11794 /* 11795 * Return 0 on success and a errno on failure to send. 11796 * Note that a 0 return may not mean we sent anything 11797 * if the TCB was on the hpts. A non-zero return 11798 * does indicate the error we got from ip[6]_output. 11799 */ 11800 static int 11801 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) 11802 { 11803 struct socket *so; 11804 int32_t len; 11805 uint32_t cts; 11806 uint32_t recwin, sendwin; 11807 int32_t sb_offset; 11808 int32_t flags, abandon, error = 0; 11809 struct tcp_log_buffer *lgb = NULL; 11810 struct mbuf *m; 11811 struct mbuf *mb; 11812 uint32_t if_hw_tsomaxsegcount = 0; 11813 uint32_t if_hw_tsomaxsegsize = 0; 11814 uint32_t if_hw_tsomax = 0; 11815 struct ip *ip = NULL; 11816 struct tcp_bbr *bbr; 11817 struct tcphdr *th; 11818 struct udphdr *udp = NULL; 11819 u_char opt[TCP_MAXOLEN]; 11820 unsigned ipoptlen, optlen, hdrlen; 11821 unsigned ulen; 11822 uint32_t bbr_seq; 11823 uint32_t delay_calc=0; 11824 uint8_t doing_tlp = 0; 11825 uint8_t local_options; 11826 #ifdef BBR_INVARIANTS 11827 uint8_t doing_retran_from = 0; 11828 uint8_t picked_up_retran = 0; 11829 #endif 11830 uint8_t wanted_cookie = 0; 11831 uint8_t more_to_rxt=0; 11832 int32_t prefetch_so_done = 0; 11833 int32_t prefetch_rsm = 0; 11834 uint32_t tot_len = 0; 11835 uint32_t maxseg, pace_max_segs, p_maxseg; 11836 int32_t csum_flags = 0; 11837 int32_t hw_tls; 11838 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 11839 unsigned ipsec_optlen = 0; 11840 11841 #endif 11842 volatile int32_t sack_rxmit; 11843 struct bbr_sendmap *rsm = NULL; 11844 int32_t tso, mtu; 11845 struct tcpopt to; 11846 int32_t slot = 0; 11847 struct inpcb *inp; 11848 struct sockbuf *sb; 11849 uint32_t hpts_calling; 11850 #ifdef INET6 11851 struct ip6_hdr *ip6 = NULL; 11852 int32_t isipv6; 11853 #endif 11854 uint8_t app_limited = BBR_JR_SENT_DATA; 11855 uint8_t filled_all = 0; 11856 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11857 /* We take a cache hit here */ 11858 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); 11859 cts = tcp_tv_to_usectick(&bbr->rc_tv); 11860 inp = bbr->rc_inp; 11861 so = inp->inp_socket; 11862 sb = &so->so_snd; 11863 if (sb->sb_flags & SB_TLS_IFNET) 11864 hw_tls = 1; 11865 else 11866 hw_tls = 0; 11867 kern_prefetch(sb, &maxseg); 11868 maxseg = tp->t_maxseg - bbr->rc_last_options; 11869 if (bbr_minseg(bbr) < maxseg) { 11870 tcp_bbr_tso_size_check(bbr, cts); 11871 } 11872 /* Remove any flags that indicate we are pacing on the inp */ 11873 pace_max_segs = bbr->r_ctl.rc_pace_max_segs; 11874 p_maxseg = min(maxseg, pace_max_segs); 11875 INP_WLOCK_ASSERT(inp); 11876 #ifdef TCP_OFFLOAD 11877 if (tp->t_flags & TF_TOE) 11878 return (tcp_offload_output(tp)); 11879 #endif 11880 11881 #ifdef INET6 11882 if (bbr->r_state) { 11883 /* Use the cache line loaded if possible */ 11884 isipv6 = bbr->r_is_v6; 11885 } else { 11886 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 11887 } 11888 #endif 11889 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 11890 tcp_in_hpts(inp)) { 11891 /* 11892 * We are on the hpts for some timer but not hptsi output. 11893 * Possibly remove from the hpts so we can send/recv etc. 11894 */ 11895 if ((tp->t_flags & TF_ACKNOW) == 0) { 11896 /* 11897 * No immediate demand right now to send an ack, but 11898 * the user may have read, making room for new data 11899 * (a window update). If so we may want to cancel 11900 * whatever timer is running (KEEP/DEL-ACK?) and 11901 * continue to send out a window update. Or we may 11902 * have gotten more data into the socket buffer to 11903 * send. 11904 */ 11905 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 11906 (long)TCP_MAXWIN << tp->rcv_scale); 11907 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && 11908 ((tcp_outflags[tp->t_state] & TH_RST) == 0) && 11909 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= 11910 (tp->snd_max - tp->snd_una))) { 11911 /* 11912 * Nothing new to send and no window update 11913 * is needed to send. Lets just return and 11914 * let the timer-run off. 11915 */ 11916 return (0); 11917 } 11918 } 11919 tcp_hpts_remove(inp); 11920 bbr_timer_cancel(bbr, __LINE__, cts); 11921 } 11922 if (bbr->r_ctl.rc_last_delay_val) { 11923 /* Calculate a rough delay for early escape to sending */ 11924 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11925 delay_calc = cts - bbr->rc_pacer_started; 11926 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11927 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11928 else 11929 delay_calc = 0; 11930 } 11931 /* Mark that we have called bbr_output(). */ 11932 if ((bbr->r_timer_override) || 11933 (tp->t_state < TCPS_ESTABLISHED)) { 11934 /* Timeouts or early states are exempt */ 11935 if (tcp_in_hpts(inp)) 11936 tcp_hpts_remove(inp); 11937 } else if (tcp_in_hpts(inp)) { 11938 if ((bbr->r_ctl.rc_last_delay_val) && 11939 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11940 delay_calc) { 11941 /* 11942 * We were being paced for output and the delay has 11943 * already exceeded when we were supposed to be 11944 * called, lets go ahead and pull out of the hpts 11945 * and call output. 11946 */ 11947 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); 11948 bbr->r_ctl.rc_last_delay_val = 0; 11949 tcp_hpts_remove(inp); 11950 } else if (tp->t_state == TCPS_CLOSED) { 11951 bbr->r_ctl.rc_last_delay_val = 0; 11952 tcp_hpts_remove(inp); 11953 } else { 11954 /* 11955 * On the hpts, you shall not pass! even if ACKNOW 11956 * is on, we will when the hpts fires, unless of 11957 * course we are overdue. 11958 */ 11959 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); 11960 return (0); 11961 } 11962 } 11963 bbr->rc_cwnd_limited = 0; 11964 if (bbr->r_ctl.rc_last_delay_val) { 11965 /* recalculate the real delay and deal with over/under */ 11966 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11967 delay_calc = cts - bbr->rc_pacer_started; 11968 else 11969 delay_calc = 0; 11970 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11971 /* Setup the delay which will be added in */ 11972 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11973 else { 11974 /* 11975 * We are early setup to adjust 11976 * our slot time. 11977 */ 11978 uint64_t merged_val; 11979 11980 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); 11981 bbr->r_agg_early_set = 1; 11982 if (bbr->r_ctl.rc_hptsi_agg_delay) { 11983 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { 11984 /* Nope our previous late cancels out the early */ 11985 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; 11986 bbr->r_agg_early_set = 0; 11987 bbr->r_ctl.rc_agg_early = 0; 11988 } else { 11989 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; 11990 bbr->r_ctl.rc_hptsi_agg_delay = 0; 11991 } 11992 } 11993 merged_val = bbr->rc_pacer_started; 11994 merged_val <<= 32; 11995 merged_val |= bbr->r_ctl.rc_last_delay_val; 11996 bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, 11997 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, 11998 bbr->r_agg_early_set, 3); 11999 bbr->r_ctl.rc_last_delay_val = 0; 12000 BBR_STAT_INC(bbr_early); 12001 delay_calc = 0; 12002 } 12003 } else { 12004 /* We were not delayed due to hptsi */ 12005 if (bbr->r_agg_early_set) 12006 bbr->r_ctl.rc_agg_early = 0; 12007 bbr->r_agg_early_set = 0; 12008 delay_calc = 0; 12009 } 12010 if (delay_calc) { 12011 /* 12012 * We had a hptsi delay which means we are falling behind on 12013 * sending at the expected rate. Calculate an extra amount 12014 * of data we can send, if any, to put us back on track. 12015 */ 12016 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) 12017 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; 12018 else 12019 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; 12020 } 12021 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12022 if ((tp->snd_una == tp->snd_max) && 12023 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 12024 (sbavail(sb))) { 12025 /* 12026 * Ok we have been idle with nothing outstanding 12027 * we possibly need to start fresh with either a new 12028 * suite of states or a fast-ramp up. 12029 */ 12030 bbr_restart_after_idle(bbr, 12031 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); 12032 } 12033 /* 12034 * Now was there a hptsi delay where we are behind? We only count 12035 * being behind if: a) We are not in recovery. b) There was a delay. 12036 * <and> c) We had room to send something. 12037 * 12038 */ 12039 hpts_calling = inp->inp_hpts_calls; 12040 inp->inp_hpts_calls = 0; 12041 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 12042 int retval; 12043 12044 retval = bbr_process_timers(tp, bbr, cts, hpts_calling); 12045 if (retval != 0) { 12046 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); 12047 /* 12048 * If timers want tcp_drop(), then pass error out, 12049 * otherwise suppress it. 12050 */ 12051 return (retval < 0 ? retval : 0); 12052 } 12053 } 12054 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 12055 if (hpts_calling && 12056 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 12057 bbr->r_ctl.rc_last_delay_val = 0; 12058 } 12059 bbr->r_timer_override = 0; 12060 bbr->r_wanted_output = 0; 12061 /* 12062 * For TFO connections in SYN_RECEIVED, only allow the initial 12063 * SYN|ACK and those sent by the retransmit timer. 12064 */ 12065 if (IS_FASTOPEN(tp->t_flags) && 12066 ((tp->t_state == TCPS_SYN_RECEIVED) || 12067 (tp->t_state == TCPS_SYN_SENT)) && 12068 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 12069 (tp->t_rxtshift == 0)) { /* not a retransmit */ 12070 len = 0; 12071 goto just_return_nolock; 12072 } 12073 /* 12074 * Before sending anything check for a state update. For hpts 12075 * calling without input this is important. If its input calling 12076 * then this was already done. 12077 */ 12078 if (bbr->rc_use_google == 0) 12079 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12080 again: 12081 /* 12082 * If we've recently taken a timeout, snd_max will be greater than 12083 * snd_max. BBR in general does not pay much attention to snd_nxt 12084 * for historic reasons the persist timer still uses it. This means 12085 * we have to look at it. All retransmissions that are not persits 12086 * use the rsm that needs to be sent so snd_nxt is ignored. At the 12087 * end of this routine we pull snd_nxt always up to snd_max. 12088 */ 12089 doing_tlp = 0; 12090 #ifdef BBR_INVARIANTS 12091 doing_retran_from = picked_up_retran = 0; 12092 #endif 12093 error = 0; 12094 tso = 0; 12095 slot = 0; 12096 mtu = 0; 12097 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12098 sb_offset = tp->snd_max - tp->snd_una; 12099 flags = tcp_outflags[tp->t_state]; 12100 sack_rxmit = 0; 12101 len = 0; 12102 rsm = NULL; 12103 if (flags & TH_RST) { 12104 SOCKBUF_LOCK(sb); 12105 goto send; 12106 } 12107 recheck_resend: 12108 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 12109 /* We need to always have one in reserve */ 12110 rsm = bbr_alloc(bbr); 12111 if (rsm == NULL) { 12112 error = ENOMEM; 12113 /* Lie to get on the hpts */ 12114 tot_len = tp->t_maxseg; 12115 if (hpts_calling) 12116 /* Retry in a ms */ 12117 slot = 1001; 12118 goto just_return_nolock; 12119 } 12120 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 12121 bbr->r_ctl.rc_free_cnt++; 12122 rsm = NULL; 12123 } 12124 /* What do we send, a resend? */ 12125 if (bbr->r_ctl.rc_resend == NULL) { 12126 /* Check for rack timeout */ 12127 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 12128 if (bbr->r_ctl.rc_resend) { 12129 #ifdef BBR_INVARIANTS 12130 picked_up_retran = 1; 12131 #endif 12132 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); 12133 } 12134 } 12135 if (bbr->r_ctl.rc_resend) { 12136 rsm = bbr->r_ctl.rc_resend; 12137 #ifdef BBR_INVARIANTS 12138 doing_retran_from = 1; 12139 #endif 12140 /* Remove any TLP flags its a RACK or T-O */ 12141 rsm->r_flags &= ~BBR_TLP; 12142 bbr->r_ctl.rc_resend = NULL; 12143 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 12144 #ifdef BBR_INVARIANTS 12145 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", 12146 tp, bbr, rsm, rsm->r_start, tp->snd_una); 12147 goto recheck_resend; 12148 #else 12149 /* TSNH */ 12150 rsm = NULL; 12151 goto recheck_resend; 12152 #endif 12153 } 12154 if (rsm->r_flags & BBR_HAS_SYN) { 12155 /* Only retransmit a SYN by itself */ 12156 len = 0; 12157 if ((flags & TH_SYN) == 0) { 12158 /* Huh something is wrong */ 12159 rsm->r_start++; 12160 if (rsm->r_start == rsm->r_end) { 12161 /* Clean it up, somehow we missed the ack? */ 12162 bbr_log_syn(tp, NULL); 12163 } else { 12164 /* TFO with data? */ 12165 rsm->r_flags &= ~BBR_HAS_SYN; 12166 len = rsm->r_end - rsm->r_start; 12167 } 12168 } else { 12169 /* Retransmitting SYN */ 12170 rsm = NULL; 12171 SOCKBUF_LOCK(sb); 12172 goto send; 12173 } 12174 } else 12175 len = rsm->r_end - rsm->r_start; 12176 if ((bbr->rc_resends_use_tso == 0) && 12177 (len > maxseg)) { 12178 len = maxseg; 12179 more_to_rxt = 1; 12180 } 12181 sb_offset = rsm->r_start - tp->snd_una; 12182 if (len > 0) { 12183 sack_rxmit = 1; 12184 KMOD_TCPSTAT_INC(tcps_sack_rexmits); 12185 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, 12186 min(len, maxseg)); 12187 } else { 12188 /* I dont think this can happen */ 12189 rsm = NULL; 12190 goto recheck_resend; 12191 } 12192 BBR_STAT_INC(bbr_resends_set); 12193 } else if (bbr->r_ctl.rc_tlp_send) { 12194 /* 12195 * Tail loss probe 12196 */ 12197 doing_tlp = 1; 12198 rsm = bbr->r_ctl.rc_tlp_send; 12199 bbr->r_ctl.rc_tlp_send = NULL; 12200 sack_rxmit = 1; 12201 len = rsm->r_end - rsm->r_start; 12202 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12203 len = maxseg; 12204 12205 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 12206 #ifdef BBR_INVARIANTS 12207 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", 12208 tp, bbr, tp->snd_una, rsm, rsm->r_start); 12209 #else 12210 /* TSNH */ 12211 rsm = NULL; 12212 goto recheck_resend; 12213 #endif 12214 } 12215 sb_offset = rsm->r_start - tp->snd_una; 12216 BBR_STAT_INC(bbr_tlp_set); 12217 } 12218 /* 12219 * Enforce a connection sendmap count limit if set 12220 * as long as we are not retransmiting. 12221 */ 12222 if ((rsm == NULL) && 12223 (V_tcp_map_entries_limit > 0) && 12224 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 12225 BBR_STAT_INC(bbr_alloc_limited); 12226 if (!bbr->alloc_limit_reported) { 12227 bbr->alloc_limit_reported = 1; 12228 BBR_STAT_INC(bbr_alloc_limited_conns); 12229 } 12230 goto just_return_nolock; 12231 } 12232 #ifdef BBR_INVARIANTS 12233 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { 12234 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", 12235 tp, bbr, rsm, sb_offset, len); 12236 } 12237 #endif 12238 /* 12239 * Get standard flags, and add SYN or FIN if requested by 'hidden' 12240 * state flags. 12241 */ 12242 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) 12243 flags |= TH_FIN; 12244 if (tp->t_flags & TF_NEEDSYN) 12245 flags |= TH_SYN; 12246 12247 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { 12248 /* we are retransmitting the fin */ 12249 len--; 12250 if (len) { 12251 /* 12252 * When retransmitting data do *not* include the 12253 * FIN. This could happen from a TLP probe if we 12254 * allowed data with a FIN. 12255 */ 12256 flags &= ~TH_FIN; 12257 } 12258 } else if (rsm) { 12259 if (flags & TH_FIN) 12260 flags &= ~TH_FIN; 12261 } 12262 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 12263 void *end_rsm; 12264 12265 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 12266 if (end_rsm) 12267 kern_prefetch(end_rsm, &prefetch_rsm); 12268 prefetch_rsm = 1; 12269 } 12270 SOCKBUF_LOCK(sb); 12271 /* 12272 * If snd_nxt == snd_max and we have transmitted a FIN, the 12273 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 12274 * negative length. This can also occur when TCP opens up its 12275 * congestion window while receiving additional duplicate acks after 12276 * fast-retransmit because TCP will reset snd_nxt to snd_max after 12277 * the fast-retransmit. 12278 * 12279 * In the normal retransmit-FIN-only case, however, snd_nxt will be 12280 * set to snd_una, the sb_offset will be 0, and the length may wind 12281 * up 0. 12282 * 12283 * If sack_rxmit is true we are retransmitting from the scoreboard 12284 * in which case len is already set. 12285 */ 12286 if (sack_rxmit == 0) { 12287 uint32_t avail; 12288 12289 avail = sbavail(sb); 12290 if (SEQ_GT(tp->snd_max, tp->snd_una)) 12291 sb_offset = tp->snd_max - tp->snd_una; 12292 else 12293 sb_offset = 0; 12294 if (bbr->rc_tlp_new_data) { 12295 /* TLP is forcing out new data */ 12296 uint32_t tlplen; 12297 12298 doing_tlp = 1; 12299 tlplen = maxseg; 12300 12301 if (tlplen > (uint32_t)(avail - sb_offset)) { 12302 tlplen = (uint32_t)(avail - sb_offset); 12303 } 12304 if (tlplen > tp->snd_wnd) { 12305 len = tp->snd_wnd; 12306 } else { 12307 len = tlplen; 12308 } 12309 bbr->rc_tlp_new_data = 0; 12310 } else { 12311 len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); 12312 if ((len < p_maxseg) && 12313 (bbr->rc_in_persist == 0) && 12314 (ctf_outstanding(tp) >= (2 * p_maxseg)) && 12315 ((avail - sb_offset) >= p_maxseg)) { 12316 /* 12317 * We are not completing whats in the socket 12318 * buffer (i.e. there is at least a segment 12319 * waiting to send) and we have 2 or more 12320 * segments outstanding. There is no sense 12321 * of sending a little piece. Lets defer and 12322 * and wait until we can send a whole 12323 * segment. 12324 */ 12325 len = 0; 12326 } 12327 if (bbr->rc_in_persist) { 12328 /* 12329 * We are in persists, figure out if 12330 * a retransmit is available (maybe the previous 12331 * persists we sent) or if we have to send new 12332 * data. 12333 */ 12334 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 12335 if (rsm) { 12336 len = rsm->r_end - rsm->r_start; 12337 if (rsm->r_flags & BBR_HAS_FIN) 12338 len--; 12339 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12340 len = maxseg; 12341 if (len > 1) 12342 BBR_STAT_INC(bbr_persist_reneg); 12343 /* 12344 * XXXrrs we could force the len to 12345 * 1 byte here to cause the chunk to 12346 * split apart.. but that would then 12347 * mean we always retransmit it as 12348 * one byte even after the window 12349 * opens. 12350 */ 12351 sack_rxmit = 1; 12352 sb_offset = rsm->r_start - tp->snd_una; 12353 } else { 12354 /* 12355 * First time through in persists or peer 12356 * acked our one byte. Though we do have 12357 * to have something in the sb. 12358 */ 12359 len = 1; 12360 sb_offset = 0; 12361 if (avail == 0) 12362 len = 0; 12363 } 12364 } 12365 } 12366 } 12367 if (prefetch_so_done == 0) { 12368 kern_prefetch(so, &prefetch_so_done); 12369 prefetch_so_done = 1; 12370 } 12371 /* 12372 * Lop off SYN bit if it has already been sent. However, if this is 12373 * SYN-SENT state and if segment contains data and if we don't know 12374 * that foreign host supports TAO, suppress sending segment. 12375 */ 12376 if ((flags & TH_SYN) && (rsm == NULL) && 12377 SEQ_GT(tp->snd_max, tp->snd_una)) { 12378 if (tp->t_state != TCPS_SYN_RECEIVED) 12379 flags &= ~TH_SYN; 12380 /* 12381 * When sending additional segments following a TFO SYN|ACK, 12382 * do not include the SYN bit. 12383 */ 12384 if (IS_FASTOPEN(tp->t_flags) && 12385 (tp->t_state == TCPS_SYN_RECEIVED)) 12386 flags &= ~TH_SYN; 12387 sb_offset--, len++; 12388 if (sbavail(sb) == 0) 12389 len = 0; 12390 } else if ((flags & TH_SYN) && rsm) { 12391 /* 12392 * Subtract one from the len for the SYN being 12393 * retransmitted. 12394 */ 12395 len--; 12396 } 12397 /* 12398 * Be careful not to send data and/or FIN on SYN segments. This 12399 * measure is needed to prevent interoperability problems with not 12400 * fully conformant TCP implementations. 12401 */ 12402 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 12403 len = 0; 12404 flags &= ~TH_FIN; 12405 } 12406 /* 12407 * On TFO sockets, ensure no data is sent in the following cases: 12408 * 12409 * - When retransmitting SYN|ACK on a passively-created socket 12410 * - When retransmitting SYN on an actively created socket 12411 * - When sending a zero-length cookie (cookie request) on an 12412 * actively created socket 12413 * - When the socket is in the CLOSED state (RST is being sent) 12414 */ 12415 if (IS_FASTOPEN(tp->t_flags) && 12416 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 12417 ((tp->t_state == TCPS_SYN_SENT) && 12418 (tp->t_tfo_client_cookie_len == 0)) || 12419 (flags & TH_RST))) { 12420 len = 0; 12421 sack_rxmit = 0; 12422 rsm = NULL; 12423 } 12424 /* Without fast-open there should never be data sent on a SYN */ 12425 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 12426 len = 0; 12427 if (len <= 0) { 12428 /* 12429 * If FIN has been sent but not acked, but we haven't been 12430 * called to retransmit, len will be < 0. Otherwise, window 12431 * shrank after we sent into it. If window shrank to 0, 12432 * cancel pending retransmit, pull snd_nxt back to (closed) 12433 * window, and set the persist timer if it isn't already 12434 * going. If the window didn't close completely, just wait 12435 * for an ACK. 12436 * 12437 * We also do a general check here to ensure that we will 12438 * set the persist timer when we have data to send, but a 12439 * 0-byte window. This makes sure the persist timer is set 12440 * even if the packet hits one of the "goto send" lines 12441 * below. 12442 */ 12443 len = 0; 12444 if ((tp->snd_wnd == 0) && 12445 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12446 (tp->snd_una == tp->snd_max) && 12447 (sb_offset < (int)sbavail(sb))) { 12448 /* 12449 * Not enough room in the rwnd to send 12450 * a paced segment out. 12451 */ 12452 bbr_enter_persist(tp, bbr, cts, __LINE__); 12453 } 12454 } else if ((rsm == NULL) && 12455 (doing_tlp == 0) && 12456 (len < bbr->r_ctl.rc_pace_max_segs)) { 12457 /* 12458 * We are not sending a full segment for 12459 * some reason. Should we not send anything (think 12460 * sws or persists)? 12461 */ 12462 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12463 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12464 (len < (int)(sbavail(sb) - sb_offset))) { 12465 /* 12466 * Here the rwnd is less than 12467 * the pacing size, this is not a retransmit, 12468 * we are established and 12469 * the send is not the last in the socket buffer 12470 * lets not send, and possibly enter persists. 12471 */ 12472 len = 0; 12473 if (tp->snd_max == tp->snd_una) 12474 bbr_enter_persist(tp, bbr, cts, __LINE__); 12475 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && 12476 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12477 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12478 (len < (int)(sbavail(sb) - sb_offset)) && 12479 (len < bbr_minseg(bbr))) { 12480 /* 12481 * Here we are not retransmitting, and 12482 * the cwnd is not so small that we could 12483 * not send at least a min size (rxt timer 12484 * not having gone off), We have 2 segments or 12485 * more already in flight, its not the tail end 12486 * of the socket buffer and the cwnd is blocking 12487 * us from sending out minimum pacing segment size. 12488 * Lets not send anything. 12489 */ 12490 bbr->rc_cwnd_limited = 1; 12491 len = 0; 12492 } else if (((tp->snd_wnd - ctf_outstanding(tp)) < 12493 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12494 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12495 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12496 (len < (int)(sbavail(sb) - sb_offset)) && 12497 (TCPS_HAVEESTABLISHED(tp->t_state))) { 12498 /* 12499 * Here we have a send window but we have 12500 * filled it up and we can't send another pacing segment. 12501 * We also have in flight more than 2 segments 12502 * and we are not completing the sb i.e. we allow 12503 * the last bytes of the sb to go out even if 12504 * its not a full pacing segment. 12505 */ 12506 len = 0; 12507 } 12508 } 12509 /* len will be >= 0 after this point. */ 12510 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 12511 tcp_sndbuf_autoscale(tp, so, sendwin); 12512 /* 12513 * 12514 */ 12515 if (bbr->rc_in_persist && 12516 len && 12517 (rsm == NULL) && 12518 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { 12519 /* 12520 * We are in persist, not doing a retransmit and don't have enough space 12521 * yet to send a full TSO. So is it at the end of the sb 12522 * if so we need to send else nuke to 0 and don't send. 12523 */ 12524 int sbleft; 12525 if (sbavail(sb) > sb_offset) 12526 sbleft = sbavail(sb) - sb_offset; 12527 else 12528 sbleft = 0; 12529 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { 12530 /* not at end of sb lets not send */ 12531 len = 0; 12532 } 12533 } 12534 /* 12535 * Decide if we can use TCP Segmentation Offloading (if supported by 12536 * hardware). 12537 * 12538 * TSO may only be used if we are in a pure bulk sending state. The 12539 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 12540 * options prevent using TSO. With TSO the TCP header is the same 12541 * (except for the sequence number) for all generated packets. This 12542 * makes it impossible to transmit any options which vary per 12543 * generated segment or packet. 12544 * 12545 * IPv4 handling has a clear separation of ip options and ip header 12546 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() 12547 * does the right thing below to provide length of just ip options 12548 * and thus checking for ipoptlen is enough to decide if ip options 12549 * are present. 12550 */ 12551 #ifdef INET6 12552 if (isipv6) 12553 ipoptlen = ip6_optlen(inp); 12554 else 12555 #endif 12556 if (inp->inp_options) 12557 ipoptlen = inp->inp_options->m_len - 12558 offsetof(struct ipoption, ipopt_list); 12559 else 12560 ipoptlen = 0; 12561 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12562 /* 12563 * Pre-calculate here as we save another lookup into the darknesses 12564 * of IPsec that way and can actually decide if TSO is ok. 12565 */ 12566 #ifdef INET6 12567 if (isipv6 && IPSEC_ENABLED(ipv6)) 12568 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 12569 #ifdef INET 12570 else 12571 #endif 12572 #endif /* INET6 */ 12573 #ifdef INET 12574 if (IPSEC_ENABLED(ipv4)) 12575 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 12576 #endif /* INET */ 12577 #endif /* IPSEC */ 12578 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12579 ipoptlen += ipsec_optlen; 12580 #endif 12581 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 12582 (len > maxseg) && 12583 (tp->t_port == 0) && 12584 ((tp->t_flags & TF_SIGNATURE) == 0) && 12585 tp->rcv_numsacks == 0 && 12586 ipoptlen == 0) 12587 tso = 1; 12588 12589 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12590 (long)TCP_MAXWIN << tp->rcv_scale); 12591 /* 12592 * Sender silly window avoidance. We transmit under the following 12593 * conditions when len is non-zero: 12594 * 12595 * - We have a full segment (or more with TSO) - This is the last 12596 * buffer in a write()/send() and we are either idle or running 12597 * NODELAY - we've timed out (e.g. persist timer) - we have more 12598 * then 1/2 the maximum send window's worth of data (receiver may be 12599 * limited the window size) - we need to retransmit 12600 */ 12601 if (rsm) 12602 goto send; 12603 if (len) { 12604 if (sack_rxmit) 12605 goto send; 12606 if (len >= p_maxseg) 12607 goto send; 12608 /* 12609 * NOTE! on localhost connections an 'ack' from the remote 12610 * end may occur synchronously with the output and cause us 12611 * to flush a buffer queued with moretocome. XXX 12612 * 12613 */ 12614 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ 12615 ((tp->t_flags & TF_NODELAY) || 12616 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && 12617 (tp->t_flags & TF_NOPUSH) == 0) { 12618 goto send; 12619 } 12620 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 12621 goto send; 12622 } 12623 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 12624 goto send; 12625 } 12626 } 12627 /* 12628 * Sending of standalone window updates. 12629 * 12630 * Window updates are important when we close our window due to a 12631 * full socket buffer and are opening it again after the application 12632 * reads data from it. Once the window has opened again and the 12633 * remote end starts to send again the ACK clock takes over and 12634 * provides the most current window information. 12635 * 12636 * We must avoid the silly window syndrome whereas every read from 12637 * the receive buffer, no matter how small, causes a window update 12638 * to be sent. We also should avoid sending a flurry of window 12639 * updates when the socket buffer had queued a lot of data and the 12640 * application is doing small reads. 12641 * 12642 * Prevent a flurry of pointless window updates by only sending an 12643 * update when we can increase the advertized window by more than 12644 * 1/4th of the socket buffer capacity. When the buffer is getting 12645 * full or is very small be more aggressive and send an update 12646 * whenever we can increase by two mss sized segments. In all other 12647 * situations the ACK's to new incoming data will carry further 12648 * window increases. 12649 * 12650 * Don't send an independent window update if a delayed ACK is 12651 * pending (it will get piggy-backed on it) or the remote side 12652 * already has done a half-close and won't send more data. Skip 12653 * this if the connection is in T/TCP half-open state. 12654 */ 12655 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 12656 !(tp->t_flags & TF_DELACK) && 12657 !TCPS_HAVERCVDFIN(tp->t_state)) { 12658 /* Check to see if we should do a window update */ 12659 if (bbr_window_update_needed(tp, so, recwin, maxseg)) 12660 goto send; 12661 } 12662 /* 12663 * Send if we owe the peer an ACK, RST, SYN. ACKNOW 12664 * is also a catch-all for the retransmit timer timeout case. 12665 */ 12666 if (tp->t_flags & TF_ACKNOW) { 12667 goto send; 12668 } 12669 if (flags & TH_RST) { 12670 /* Always send a RST if one is due */ 12671 goto send; 12672 } 12673 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { 12674 goto send; 12675 } 12676 /* 12677 * If our state indicates that FIN should be sent and we have not 12678 * yet done so, then we need to send. 12679 */ 12680 if (flags & TH_FIN && 12681 ((tp->t_flags & TF_SENTFIN) == 0)) { 12682 goto send; 12683 } 12684 /* 12685 * No reason to send a segment, just return. 12686 */ 12687 just_return: 12688 SOCKBUF_UNLOCK(sb); 12689 just_return_nolock: 12690 if (tot_len) 12691 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 12692 if (bbr->rc_no_pacing) 12693 slot = 0; 12694 if (tot_len == 0) { 12695 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= 12696 tp->snd_wnd) { 12697 BBR_STAT_INC(bbr_rwnd_limited); 12698 app_limited = BBR_JR_RWND_LIMITED; 12699 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12700 if ((bbr->rc_in_persist == 0) && 12701 TCPS_HAVEESTABLISHED(tp->t_state) && 12702 (tp->snd_max == tp->snd_una) && 12703 sbavail(&so->so_snd)) { 12704 /* No send window.. we must enter persist */ 12705 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 12706 } 12707 } else if (ctf_outstanding(tp) >= sbavail(sb)) { 12708 BBR_STAT_INC(bbr_app_limited); 12709 app_limited = BBR_JR_APP_LIMITED; 12710 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12711 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12712 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { 12713 BBR_STAT_INC(bbr_cwnd_limited); 12714 app_limited = BBR_JR_CWND_LIMITED; 12715 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12716 bbr->r_ctl.rc_lost_bytes))); 12717 bbr->rc_cwnd_limited = 1; 12718 } else { 12719 BBR_STAT_INC(bbr_app_limited); 12720 app_limited = BBR_JR_APP_LIMITED; 12721 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12722 } 12723 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12724 bbr->r_agg_early_set = 0; 12725 bbr->r_ctl.rc_agg_early = 0; 12726 bbr->r_ctl.rc_last_delay_val = 0; 12727 } else if (bbr->rc_use_google == 0) 12728 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12729 /* Are we app limited? */ 12730 if ((app_limited == BBR_JR_APP_LIMITED) || 12731 (app_limited == BBR_JR_RWND_LIMITED)) { 12732 /** 12733 * We are application limited. 12734 */ 12735 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12736 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); 12737 } 12738 if (tot_len == 0) 12739 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); 12740 /* Dont update the time if we did not send */ 12741 bbr->r_ctl.rc_last_delay_val = 0; 12742 bbr->rc_output_starts_timer = 1; 12743 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); 12744 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); 12745 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 12746 /* Make sure snd_nxt is drug up */ 12747 tp->snd_nxt = tp->snd_max; 12748 } 12749 return (error); 12750 12751 send: 12752 if (doing_tlp == 0) { 12753 /* 12754 * Data not a TLP, and its not the rxt firing. If it is the 12755 * rxt firing, we want to leave the tlp_in_progress flag on 12756 * so we don't send another TLP. It has to be a rack timer 12757 * or normal send (response to acked data) to clear the tlp 12758 * in progress flag. 12759 */ 12760 bbr->rc_tlp_in_progress = 0; 12761 bbr->rc_tlp_rtx_out = 0; 12762 } else { 12763 /* 12764 * Its a TLP. 12765 */ 12766 bbr->rc_tlp_in_progress = 1; 12767 } 12768 bbr_timer_cancel(bbr, __LINE__, cts); 12769 if (rsm == NULL) { 12770 if (sbused(sb) > 0) { 12771 /* 12772 * This is sub-optimal. We only send a stand alone 12773 * FIN on its own segment. 12774 */ 12775 if (flags & TH_FIN) { 12776 flags &= ~TH_FIN; 12777 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { 12778 /* Lets not send this */ 12779 slot = 0; 12780 goto just_return; 12781 } 12782 } 12783 } 12784 } else { 12785 /* 12786 * We do *not* send a FIN on a retransmit if it has data. 12787 * The if clause here where len > 1 should never come true. 12788 */ 12789 if ((len > 0) && 12790 (((rsm->r_flags & BBR_HAS_FIN) == 0) && 12791 (flags & TH_FIN))) { 12792 flags &= ~TH_FIN; 12793 len--; 12794 } 12795 } 12796 SOCKBUF_LOCK_ASSERT(sb); 12797 if (len > 0) { 12798 if ((tp->snd_una == tp->snd_max) && 12799 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 12800 /* 12801 * This qualifies as a RTT_PROBE session since we 12802 * drop the data outstanding to nothing and waited 12803 * more than bbr_rtt_probe_time. 12804 */ 12805 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 12806 bbr_set_reduced_rtt(bbr, cts, __LINE__); 12807 } 12808 if (len >= maxseg) 12809 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 12810 else 12811 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 12812 } 12813 /* 12814 * Before ESTABLISHED, force sending of initial options unless TCP 12815 * set not to do any options. NOTE: we assume that the IP/TCP header 12816 * plus TCP options always fit in a single mbuf, leaving room for a 12817 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 12818 * + optlen <= MCLBYTES 12819 */ 12820 optlen = 0; 12821 #ifdef INET6 12822 if (isipv6) 12823 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 12824 else 12825 #endif 12826 hdrlen = sizeof(struct tcpiphdr); 12827 12828 /* 12829 * Compute options for segment. We only have to care about SYN and 12830 * established connection segments. Options for SYN-ACK segments 12831 * are handled in TCP syncache. 12832 */ 12833 to.to_flags = 0; 12834 local_options = 0; 12835 if ((tp->t_flags & TF_NOOPT) == 0) { 12836 /* Maximum segment size. */ 12837 if (flags & TH_SYN) { 12838 to.to_mss = tcp_mssopt(&inp->inp_inc); 12839 if (tp->t_port) 12840 to.to_mss -= V_tcp_udp_tunneling_overhead; 12841 to.to_flags |= TOF_MSS; 12842 /* 12843 * On SYN or SYN|ACK transmits on TFO connections, 12844 * only include the TFO option if it is not a 12845 * retransmit, as the presence of the TFO option may 12846 * have caused the original SYN or SYN|ACK to have 12847 * been dropped by a middlebox. 12848 */ 12849 if (IS_FASTOPEN(tp->t_flags) && 12850 (tp->t_rxtshift == 0)) { 12851 if (tp->t_state == TCPS_SYN_RECEIVED) { 12852 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 12853 to.to_tfo_cookie = 12854 (u_int8_t *)&tp->t_tfo_cookie.server; 12855 to.to_flags |= TOF_FASTOPEN; 12856 wanted_cookie = 1; 12857 } else if (tp->t_state == TCPS_SYN_SENT) { 12858 to.to_tfo_len = 12859 tp->t_tfo_client_cookie_len; 12860 to.to_tfo_cookie = 12861 tp->t_tfo_cookie.client; 12862 to.to_flags |= TOF_FASTOPEN; 12863 wanted_cookie = 1; 12864 } 12865 } 12866 } 12867 /* Window scaling. */ 12868 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 12869 to.to_wscale = tp->request_r_scale; 12870 to.to_flags |= TOF_SCALE; 12871 } 12872 /* Timestamps. */ 12873 if ((tp->t_flags & TF_RCVD_TSTMP) || 12874 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 12875 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; 12876 to.to_tsecr = tp->ts_recent; 12877 to.to_flags |= TOF_TS; 12878 local_options += TCPOLEN_TIMESTAMP + 2; 12879 } 12880 /* Set receive buffer autosizing timestamp. */ 12881 if (tp->rfbuf_ts == 0 && 12882 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 12883 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); 12884 /* Selective ACK's. */ 12885 if (flags & TH_SYN) 12886 to.to_flags |= TOF_SACKPERM; 12887 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 12888 tp->rcv_numsacks > 0) { 12889 to.to_flags |= TOF_SACK; 12890 to.to_nsacks = tp->rcv_numsacks; 12891 to.to_sacks = (u_char *)tp->sackblks; 12892 } 12893 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 12894 /* TCP-MD5 (RFC2385). */ 12895 if (tp->t_flags & TF_SIGNATURE) 12896 to.to_flags |= TOF_SIGNATURE; 12897 #endif /* TCP_SIGNATURE */ 12898 12899 /* Processing the options. */ 12900 hdrlen += (optlen = tcp_addoptions(&to, opt)); 12901 /* 12902 * If we wanted a TFO option to be added, but it was unable 12903 * to fit, ensure no data is sent. 12904 */ 12905 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 12906 !(to.to_flags & TOF_FASTOPEN)) 12907 len = 0; 12908 } 12909 if (tp->t_port) { 12910 if (V_tcp_udp_tunneling_port == 0) { 12911 /* The port was removed?? */ 12912 SOCKBUF_UNLOCK(&so->so_snd); 12913 return (EHOSTUNREACH); 12914 } 12915 hdrlen += sizeof(struct udphdr); 12916 } 12917 #ifdef INET6 12918 if (isipv6) 12919 ipoptlen = ip6_optlen(inp); 12920 else 12921 #endif 12922 if (inp->inp_options) 12923 ipoptlen = inp->inp_options->m_len - 12924 offsetof(struct ipoption, ipopt_list); 12925 else 12926 ipoptlen = 0; 12927 ipoptlen = 0; 12928 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12929 ipoptlen += ipsec_optlen; 12930 #endif 12931 if (bbr->rc_last_options != local_options) { 12932 /* 12933 * Cache the options length this generally does not change 12934 * on a connection. We use this to calculate TSO. 12935 */ 12936 bbr->rc_last_options = local_options; 12937 } 12938 maxseg = tp->t_maxseg - (ipoptlen + optlen); 12939 p_maxseg = min(maxseg, pace_max_segs); 12940 /* 12941 * Adjust data length if insertion of options will bump the packet 12942 * length beyond the t_maxseg length. Clear the FIN bit because we 12943 * cut off the tail of the segment. 12944 */ 12945 if (len > maxseg) { 12946 if (len != 0 && (flags & TH_FIN)) { 12947 flags &= ~TH_FIN; 12948 } 12949 if (tso) { 12950 uint32_t moff; 12951 int32_t max_len; 12952 12953 /* extract TSO information */ 12954 if_hw_tsomax = tp->t_tsomax; 12955 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 12956 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 12957 KASSERT(ipoptlen == 0, 12958 ("%s: TSO can't do IP options", __func__)); 12959 12960 /* 12961 * Check if we should limit by maximum payload 12962 * length: 12963 */ 12964 if (if_hw_tsomax != 0) { 12965 /* compute maximum TSO length */ 12966 max_len = (if_hw_tsomax - hdrlen - 12967 max_linkhdr); 12968 if (max_len <= 0) { 12969 len = 0; 12970 } else if (len > max_len) { 12971 len = max_len; 12972 } 12973 } 12974 /* 12975 * Prevent the last segment from being fractional 12976 * unless the send sockbuf can be emptied: 12977 */ 12978 if ((sb_offset + len) < sbavail(sb)) { 12979 moff = len % (uint32_t)maxseg; 12980 if (moff != 0) { 12981 len -= moff; 12982 } 12983 } 12984 /* 12985 * In case there are too many small fragments don't 12986 * use TSO: 12987 */ 12988 if (len <= maxseg) { 12989 len = maxseg; 12990 tso = 0; 12991 } 12992 } else { 12993 /* Not doing TSO */ 12994 if (optlen + ipoptlen >= tp->t_maxseg) { 12995 /* 12996 * Since we don't have enough space to put 12997 * the IP header chain and the TCP header in 12998 * one packet as required by RFC 7112, don't 12999 * send it. Also ensure that at least one 13000 * byte of the payload can be put into the 13001 * TCP segment. 13002 */ 13003 SOCKBUF_UNLOCK(&so->so_snd); 13004 error = EMSGSIZE; 13005 sack_rxmit = 0; 13006 goto out; 13007 } 13008 len = maxseg; 13009 } 13010 } else { 13011 /* Not doing TSO */ 13012 if_hw_tsomaxsegcount = 0; 13013 tso = 0; 13014 } 13015 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 13016 ("%s: len > IP_MAXPACKET", __func__)); 13017 #ifdef DIAGNOSTIC 13018 #ifdef INET6 13019 if (max_linkhdr + hdrlen > MCLBYTES) 13020 #else 13021 if (max_linkhdr + hdrlen > MHLEN) 13022 #endif 13023 panic("tcphdr too big"); 13024 #endif 13025 /* 13026 * This KASSERT is here to catch edge cases at a well defined place. 13027 * Before, those had triggered (random) panic conditions further 13028 * down. 13029 */ 13030 #ifdef BBR_INVARIANTS 13031 if (sack_rxmit) { 13032 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 13033 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", 13034 rsm, tp, bbr, rsm->r_start, tp->snd_una); 13035 } 13036 } 13037 #endif 13038 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 13039 if ((len == 0) && 13040 (flags & TH_FIN) && 13041 (sbused(sb))) { 13042 /* 13043 * We have outstanding data, don't send a fin by itself!. 13044 */ 13045 slot = 0; 13046 goto just_return; 13047 } 13048 /* 13049 * Grab a header mbuf, attaching a copy of data to be transmitted, 13050 * and initialize the header from the template for sends on this 13051 * connection. 13052 */ 13053 if (len) { 13054 uint32_t moff; 13055 13056 /* 13057 * We place a limit on sending with hptsi. 13058 */ 13059 if ((rsm == NULL) && len > pace_max_segs) 13060 len = pace_max_segs; 13061 if (len <= maxseg) 13062 tso = 0; 13063 #ifdef INET6 13064 if (MHLEN < hdrlen + max_linkhdr) 13065 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 13066 else 13067 #endif 13068 m = m_gethdr(M_NOWAIT, MT_DATA); 13069 13070 if (m == NULL) { 13071 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13072 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13073 SOCKBUF_UNLOCK(sb); 13074 error = ENOBUFS; 13075 sack_rxmit = 0; 13076 goto out; 13077 } 13078 m->m_data += max_linkhdr; 13079 m->m_len = hdrlen; 13080 /* 13081 * Start the m_copy functions from the closest mbuf to the 13082 * sb_offset in the socket buffer chain. 13083 */ 13084 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { 13085 #ifdef BBR_INVARIANTS 13086 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) 13087 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", 13088 tp, bbr, len, sb_offset, sbavail(sb), rsm, 13089 doing_retran_from, 13090 picked_up_retran, 13091 doing_tlp); 13092 13093 #endif 13094 /* 13095 * In this messed up situation we have two choices, 13096 * a) pretend the send worked, and just start timers 13097 * and what not (not good since that may lead us 13098 * back here a lot). <or> b) Send the lowest segment 13099 * in the map. <or> c) Drop the connection. Lets do 13100 * <b> which if it continues to happen will lead to 13101 * <c> via timeouts. 13102 */ 13103 BBR_STAT_INC(bbr_offset_recovery); 13104 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 13105 sb_offset = 0; 13106 if (rsm == NULL) { 13107 sack_rxmit = 0; 13108 len = sbavail(sb); 13109 } else { 13110 sack_rxmit = 1; 13111 if (rsm->r_start != tp->snd_una) { 13112 /* 13113 * Things are really messed up, <c> 13114 * is the only thing to do. 13115 */ 13116 BBR_STAT_INC(bbr_offset_drop); 13117 SOCKBUF_UNLOCK(sb); 13118 (void)m_free(m); 13119 return (-EFAULT); /* tcp_drop() */ 13120 } 13121 len = rsm->r_end - rsm->r_start; 13122 } 13123 if (len > sbavail(sb)) 13124 len = sbavail(sb); 13125 if (len > maxseg) 13126 len = maxseg; 13127 } 13128 mb = sbsndptr_noadv(sb, sb_offset, &moff); 13129 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 13130 m_copydata(mb, moff, (int)len, 13131 mtod(m, caddr_t)+hdrlen); 13132 if (rsm == NULL) 13133 sbsndptr_adv(sb, mb, len); 13134 m->m_len += len; 13135 } else { 13136 struct sockbuf *msb; 13137 13138 if (rsm) 13139 msb = NULL; 13140 else 13141 msb = sb; 13142 #ifdef BBR_INVARIANTS 13143 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { 13144 if (rsm) { 13145 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ", 13146 tp, bbr, len, moff, 13147 sbavail(sb), rsm, 13148 tp->snd_una, rsm->r_flags, rsm->r_start, 13149 doing_retran_from, 13150 picked_up_retran, 13151 doing_tlp, sack_rxmit); 13152 } else { 13153 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", 13154 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); 13155 } 13156 } 13157 #endif 13158 m->m_next = tcp_m_copym( 13159 mb, moff, &len, 13160 if_hw_tsomaxsegcount, 13161 if_hw_tsomaxsegsize, msb, 13162 ((rsm == NULL) ? hw_tls : 0) 13163 #ifdef NETFLIX_COPY_ARGS 13164 , &filled_all 13165 #endif 13166 ); 13167 if (len <= maxseg) { 13168 /* 13169 * Must have ran out of mbufs for the copy 13170 * shorten it to no longer need tso. Lets 13171 * not put on sendalot since we are low on 13172 * mbufs. 13173 */ 13174 tso = 0; 13175 } 13176 if (m->m_next == NULL) { 13177 SOCKBUF_UNLOCK(sb); 13178 (void)m_free(m); 13179 error = ENOBUFS; 13180 sack_rxmit = 0; 13181 goto out; 13182 } 13183 } 13184 #ifdef BBR_INVARIANTS 13185 if (tso && len < maxseg) { 13186 panic("tp:%p tso on, but len:%d < maxseg:%d", 13187 tp, len, maxseg); 13188 } 13189 if (tso && if_hw_tsomaxsegcount) { 13190 int32_t seg_cnt = 0; 13191 struct mbuf *foo; 13192 13193 foo = m; 13194 while (foo) { 13195 seg_cnt++; 13196 foo = foo->m_next; 13197 } 13198 if (seg_cnt > if_hw_tsomaxsegcount) { 13199 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); 13200 } 13201 } 13202 #endif 13203 /* 13204 * If we're sending everything we've got, set PUSH. (This 13205 * will keep happy those implementations which only give 13206 * data to the user when a buffer fills or a PUSH comes in.) 13207 */ 13208 if (sb_offset + len == sbused(sb) && 13209 sbused(sb) && 13210 !(flags & TH_SYN)) { 13211 flags |= TH_PUSH; 13212 } 13213 SOCKBUF_UNLOCK(sb); 13214 } else { 13215 SOCKBUF_UNLOCK(sb); 13216 if (tp->t_flags & TF_ACKNOW) 13217 KMOD_TCPSTAT_INC(tcps_sndacks); 13218 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 13219 KMOD_TCPSTAT_INC(tcps_sndctrl); 13220 else 13221 KMOD_TCPSTAT_INC(tcps_sndwinup); 13222 13223 m = m_gethdr(M_NOWAIT, MT_DATA); 13224 if (m == NULL) { 13225 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13226 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13227 error = ENOBUFS; 13228 /* Fudge the send time since we could not send */ 13229 sack_rxmit = 0; 13230 goto out; 13231 } 13232 #ifdef INET6 13233 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 13234 MHLEN >= hdrlen) { 13235 M_ALIGN(m, hdrlen); 13236 } else 13237 #endif 13238 m->m_data += max_linkhdr; 13239 m->m_len = hdrlen; 13240 } 13241 SOCKBUF_UNLOCK_ASSERT(sb); 13242 m->m_pkthdr.rcvif = (struct ifnet *)0; 13243 #ifdef MAC 13244 mac_inpcb_create_mbuf(inp, m); 13245 #endif 13246 #ifdef INET6 13247 if (isipv6) { 13248 ip6 = mtod(m, struct ip6_hdr *); 13249 if (tp->t_port) { 13250 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 13251 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13252 udp->uh_dport = tp->t_port; 13253 ulen = hdrlen + len - sizeof(struct ip6_hdr); 13254 udp->uh_ulen = htons(ulen); 13255 th = (struct tcphdr *)(udp + 1); 13256 } else { 13257 th = (struct tcphdr *)(ip6 + 1); 13258 } 13259 tcpip_fillheaders(inp, tp->t_port, ip6, th); 13260 } else 13261 #endif /* INET6 */ 13262 { 13263 ip = mtod(m, struct ip *); 13264 if (tp->t_port) { 13265 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 13266 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13267 udp->uh_dport = tp->t_port; 13268 ulen = hdrlen + len - sizeof(struct ip); 13269 udp->uh_ulen = htons(ulen); 13270 th = (struct tcphdr *)(udp + 1); 13271 } else { 13272 th = (struct tcphdr *)(ip + 1); 13273 } 13274 tcpip_fillheaders(inp, tp->t_port, ip, th); 13275 } 13276 /* 13277 * If we are doing retransmissions, then snd_nxt will not reflect 13278 * the first unsent octet. For ACK only packets, we do not want the 13279 * sequence number of the retransmitted packet, we want the sequence 13280 * number of the next unsent octet. So, if there is no data (and no 13281 * SYN or FIN), use snd_max instead of snd_nxt when filling in 13282 * ti_seq. But if we are in persist state, snd_max might reflect 13283 * one byte beyond the right edge of the window, so use snd_nxt in 13284 * that case, since we know we aren't doing a retransmission. 13285 * (retransmit and persist are mutually exclusive...) 13286 */ 13287 if (sack_rxmit == 0) { 13288 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { 13289 /* New data (including new persists) */ 13290 th->th_seq = htonl(tp->snd_max); 13291 bbr_seq = tp->snd_max; 13292 } else if (flags & TH_SYN) { 13293 /* Syn's always send from iss */ 13294 th->th_seq = htonl(tp->iss); 13295 bbr_seq = tp->iss; 13296 } else if (flags & TH_FIN) { 13297 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { 13298 /* 13299 * If we sent the fin already its 1 minus 13300 * snd_max 13301 */ 13302 th->th_seq = (htonl(tp->snd_max - 1)); 13303 bbr_seq = (tp->snd_max - 1); 13304 } else { 13305 /* First time FIN use snd_max */ 13306 th->th_seq = htonl(tp->snd_max); 13307 bbr_seq = tp->snd_max; 13308 } 13309 } else { 13310 /* 13311 * len == 0 and not persist we use snd_max, sending 13312 * an ack unless we have sent the fin then its 1 13313 * minus. 13314 */ 13315 /* 13316 * XXXRRS Question if we are in persists and we have 13317 * nothing outstanding to send and we have not sent 13318 * a FIN, we will send an ACK. In such a case it 13319 * might be better to send (tp->snd_una - 1) which 13320 * would force the peer to ack. 13321 */ 13322 if (tp->t_flags & TF_SENTFIN) { 13323 th->th_seq = htonl(tp->snd_max - 1); 13324 bbr_seq = (tp->snd_max - 1); 13325 } else { 13326 th->th_seq = htonl(tp->snd_max); 13327 bbr_seq = tp->snd_max; 13328 } 13329 } 13330 } else { 13331 /* All retransmits use the rsm to guide the send */ 13332 th->th_seq = htonl(rsm->r_start); 13333 bbr_seq = rsm->r_start; 13334 } 13335 th->th_ack = htonl(tp->rcv_nxt); 13336 if (optlen) { 13337 bcopy(opt, th + 1, optlen); 13338 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 13339 } 13340 tcp_set_flags(th, flags); 13341 /* 13342 * Calculate receive window. Don't shrink window, but avoid silly 13343 * window syndrome. 13344 */ 13345 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && 13346 recwin < maxseg))) 13347 recwin = 0; 13348 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 13349 recwin < (tp->rcv_adv - tp->rcv_nxt)) 13350 recwin = (tp->rcv_adv - tp->rcv_nxt); 13351 if (recwin > TCP_MAXWIN << tp->rcv_scale) 13352 recwin = TCP_MAXWIN << tp->rcv_scale; 13353 13354 /* 13355 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 13356 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 13357 * handled in syncache. 13358 */ 13359 if (flags & TH_SYN) 13360 th->th_win = htons((u_short) 13361 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 13362 else { 13363 /* Avoid shrinking window with window scaling. */ 13364 recwin = roundup2(recwin, 1 << tp->rcv_scale); 13365 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 13366 } 13367 /* 13368 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 13369 * window. This may cause the remote transmitter to stall. This 13370 * flag tells soreceive() to disable delayed acknowledgements when 13371 * draining the buffer. This can occur if the receiver is 13372 * attempting to read more data than can be buffered prior to 13373 * transmitting on the connection. 13374 */ 13375 if (th->th_win == 0) { 13376 tp->t_sndzerowin++; 13377 tp->t_flags |= TF_RXWIN0SENT; 13378 } else 13379 tp->t_flags &= ~TF_RXWIN0SENT; 13380 /* 13381 * We don't support urgent data, but drag along 13382 * the pointer in case of a stack switch. 13383 */ 13384 tp->snd_up = tp->snd_una; 13385 13386 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13387 if (to.to_flags & TOF_SIGNATURE) { 13388 /* 13389 * Calculate MD5 signature and put it into the place 13390 * determined before. NOTE: since TCP options buffer doesn't 13391 * point into mbuf's data, calculate offset and use it. 13392 */ 13393 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 13394 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 13395 /* 13396 * Do not send segment if the calculation of MD5 13397 * digest has failed. 13398 */ 13399 goto out; 13400 } 13401 } 13402 #endif 13403 13404 /* 13405 * Put TCP length in extended header, and then checksum extended 13406 * header and data. 13407 */ 13408 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 13409 #ifdef INET6 13410 if (isipv6) { 13411 /* 13412 * ip6_plen is not need to be filled now, and will be filled 13413 * in ip6_output. 13414 */ 13415 if (tp->t_port) { 13416 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 13417 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13418 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 13419 th->th_sum = htons(0); 13420 UDPSTAT_INC(udps_opackets); 13421 } else { 13422 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 13423 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13424 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 13425 optlen + len, IPPROTO_TCP, 0); 13426 } 13427 } 13428 #endif 13429 #if defined(INET6) && defined(INET) 13430 else 13431 #endif 13432 #ifdef INET 13433 { 13434 if (tp->t_port) { 13435 m->m_pkthdr.csum_flags = CSUM_UDP; 13436 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13437 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 13438 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 13439 th->th_sum = htons(0); 13440 UDPSTAT_INC(udps_opackets); 13441 } else { 13442 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; 13443 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13444 th->th_sum = in_pseudo(ip->ip_src.s_addr, 13445 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 13446 IPPROTO_TCP + len + optlen)); 13447 } 13448 /* IP version must be set here for ipv4/ipv6 checking later */ 13449 KASSERT(ip->ip_v == IPVERSION, 13450 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 13451 } 13452 #endif 13453 13454 /* 13455 * Enable TSO and specify the size of the segments. The TCP pseudo 13456 * header checksum is always provided. XXX: Fixme: This is currently 13457 * not the case for IPv6. 13458 */ 13459 if (tso) { 13460 KASSERT(len > maxseg, 13461 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); 13462 m->m_pkthdr.csum_flags |= CSUM_TSO; 13463 csum_flags |= CSUM_TSO; 13464 m->m_pkthdr.tso_segsz = maxseg; 13465 } 13466 KASSERT(len + hdrlen == m_length(m, NULL), 13467 ("%s: mbuf chain different than expected: %d + %u != %u", 13468 __func__, len, hdrlen, m_length(m, NULL))); 13469 13470 #ifdef TCP_HHOOK 13471 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ 13472 hhook_run_tcp_est_out(tp, th, &to, len, tso); 13473 #endif 13474 13475 /* Log to the black box */ 13476 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 13477 union tcp_log_stackspecific log; 13478 13479 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 13480 /* Record info on type of transmission */ 13481 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; 13482 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); 13483 log.u_bbr.flex3 = maxseg; 13484 log.u_bbr.flex4 = delay_calc; 13485 /* Encode filled_all into the upper flex5 bit */ 13486 log.u_bbr.flex5 = bbr->rc_past_init_win; 13487 log.u_bbr.flex5 <<= 1; 13488 log.u_bbr.flex5 |= bbr->rc_no_pacing; 13489 log.u_bbr.flex5 <<= 29; 13490 if (filled_all) 13491 log.u_bbr.flex5 |= 0x80000000; 13492 log.u_bbr.flex5 |= tp->t_maxseg; 13493 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; 13494 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); 13495 /* lets poke in the low and the high here for debugging */ 13496 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 13497 if (rsm || sack_rxmit) { 13498 if (doing_tlp) 13499 log.u_bbr.flex8 = 2; 13500 else 13501 log.u_bbr.flex8 = 1; 13502 } else { 13503 log.u_bbr.flex8 = 0; 13504 } 13505 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 13506 len, &log, false, NULL, NULL, 0, tv); 13507 } else { 13508 lgb = NULL; 13509 } 13510 /* 13511 * Fill in IP length and desired time to live and send to IP level. 13512 * There should be a better way to handle ttl and tos; we could keep 13513 * them in the template, but need a way to checksum without them. 13514 */ 13515 /* 13516 * m->m_pkthdr.len should have been set before cksum calcuration, 13517 * because in6_cksum() need it. 13518 */ 13519 #ifdef INET6 13520 if (isipv6) { 13521 /* 13522 * we separately set hoplimit for every segment, since the 13523 * user might want to change the value via setsockopt. Also, 13524 * desired default hop limit might be changed via Neighbor 13525 * Discovery. 13526 */ 13527 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 13528 13529 /* 13530 * Set the packet size here for the benefit of DTrace 13531 * probes. ip6_output() will set it properly; it's supposed 13532 * to include the option header lengths as well. 13533 */ 13534 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 13535 13536 if (V_path_mtu_discovery && maxseg > V_tcp_minmss) 13537 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13538 else 13539 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13540 13541 if (tp->t_state == TCPS_SYN_SENT) 13542 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 13543 13544 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 13545 /* TODO: IPv6 IP6TOS_ECT bit on */ 13546 error = ip6_output(m, inp->in6p_outputopts, 13547 &inp->inp_route6, 13548 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 13549 NULL, NULL, inp); 13550 13551 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 13552 mtu = inp->inp_route6.ro_nh->nh_mtu; 13553 } 13554 #endif /* INET6 */ 13555 #if defined(INET) && defined(INET6) 13556 else 13557 #endif 13558 #ifdef INET 13559 { 13560 ip->ip_len = htons(m->m_pkthdr.len); 13561 #ifdef INET6 13562 if (isipv6) 13563 ip->ip_ttl = in6_selecthlim(inp, NULL); 13564 #endif /* INET6 */ 13565 /* 13566 * If we do path MTU discovery, then we set DF on every 13567 * packet. This might not be the best thing to do according 13568 * to RFC3390 Section 2. However the tcp hostcache migitates 13569 * the problem so it affects only the first tcp connection 13570 * with a host. 13571 * 13572 * NB: Don't set DF on small MTU/MSS to have a safe 13573 * fallback. 13574 */ 13575 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 13576 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13577 if (tp->t_port == 0 || len < V_tcp_minmss) { 13578 ip->ip_off |= htons(IP_DF); 13579 } 13580 } else { 13581 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13582 } 13583 13584 if (tp->t_state == TCPS_SYN_SENT) 13585 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 13586 13587 TCP_PROBE5(send, NULL, tp, ip, tp, th); 13588 13589 error = ip_output(m, inp->inp_options, &inp->inp_route, 13590 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, 13591 inp); 13592 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 13593 mtu = inp->inp_route.ro_nh->nh_mtu; 13594 } 13595 #endif /* INET */ 13596 out: 13597 13598 if (lgb) { 13599 lgb->tlb_errno = error; 13600 lgb = NULL; 13601 } 13602 /* 13603 * In transmit state, time the transmission and arrange for the 13604 * retransmit. In persist state, just set snd_max. 13605 */ 13606 if (error == 0) { 13607 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); 13608 if (TCPS_HAVEESTABLISHED(tp->t_state) && 13609 (tp->t_flags & TF_SACK_PERMIT) && 13610 tp->rcv_numsacks > 0) 13611 tcp_clean_dsack_blocks(tp); 13612 /* We sent an ack clear the bbr_segs_rcvd count */ 13613 bbr->output_error_seen = 0; 13614 bbr->oerror_cnt = 0; 13615 bbr->bbr_segs_rcvd = 0; 13616 if (len == 0) 13617 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); 13618 /* Do accounting for new sends */ 13619 if ((len > 0) && (rsm == NULL)) { 13620 int idx; 13621 if (tp->snd_una == tp->snd_max) { 13622 /* 13623 * Special case to match google, when 13624 * nothing is in flight the delivered 13625 * time does get updated to the current 13626 * time (see tcp_rate_bsd.c). 13627 */ 13628 bbr->r_ctl.rc_del_time = cts; 13629 } 13630 if (len >= maxseg) { 13631 idx = (len / maxseg) + 3; 13632 if (idx >= TCP_MSS_ACCT_ATIMER) 13633 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); 13634 else 13635 counter_u64_add(bbr_out_size[idx], 1); 13636 } else { 13637 /* smaller than a MSS */ 13638 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); 13639 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) 13640 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); 13641 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); 13642 } 13643 } 13644 } 13645 abandon = 0; 13646 /* 13647 * We must do the send accounting before we log the output, 13648 * otherwise the state of the rsm could change and we account to the 13649 * wrong bucket. 13650 */ 13651 if (len > 0) { 13652 bbr_do_send_accounting(tp, bbr, rsm, len, error); 13653 if (error == 0) { 13654 if (tp->snd_una == tp->snd_max) 13655 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 13656 } 13657 } 13658 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, 13659 cts, mb, &abandon, rsm, 0, sb); 13660 if (abandon) { 13661 /* 13662 * If bbr_log_output destroys the TCB or sees a TH_RST being 13663 * sent we should hit this condition. 13664 */ 13665 return (0); 13666 } 13667 if (bbr->rc_in_persist == 0) { 13668 /* 13669 * Advance snd_nxt over sequence space of this segment. 13670 */ 13671 if (error) 13672 /* We don't log or do anything with errors */ 13673 goto skip_upd; 13674 13675 if (tp->snd_una == tp->snd_max && 13676 (len || (flags & (TH_SYN | TH_FIN)))) { 13677 /* 13678 * Update the time we just added data since none was 13679 * outstanding. 13680 */ 13681 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13682 bbr->rc_tp->t_acktime = ticks; 13683 } 13684 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { 13685 if (flags & TH_SYN) { 13686 /* 13687 * Smack the snd_max to iss + 1 13688 * if its a FO we will add len below. 13689 */ 13690 tp->snd_max = tp->iss + 1; 13691 } 13692 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13693 tp->snd_max++; 13694 tp->t_flags |= TF_SENTFIN; 13695 } 13696 } 13697 if (sack_rxmit == 0) 13698 tp->snd_max += len; 13699 skip_upd: 13700 if ((error == 0) && len) 13701 tot_len += len; 13702 } else { 13703 /* Persists case */ 13704 int32_t xlen = len; 13705 13706 if (error) 13707 goto nomore; 13708 13709 if (flags & TH_SYN) 13710 ++xlen; 13711 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13712 ++xlen; 13713 tp->t_flags |= TF_SENTFIN; 13714 } 13715 if (xlen && (tp->snd_una == tp->snd_max)) { 13716 /* 13717 * Update the time we just added data since none was 13718 * outstanding. 13719 */ 13720 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13721 bbr->rc_tp->t_acktime = ticks; 13722 } 13723 if (sack_rxmit == 0) 13724 tp->snd_max += xlen; 13725 tot_len += (len + optlen + ipoptlen); 13726 } 13727 nomore: 13728 if (error) { 13729 /* 13730 * Failures do not advance the seq counter above. For the 13731 * case of ENOBUFS we will fall out and become ack-clocked. 13732 * capping the cwnd at the current flight. 13733 * Everything else will just have to retransmit with the timer 13734 * (no pacer). 13735 */ 13736 SOCKBUF_UNLOCK_ASSERT(sb); 13737 BBR_STAT_INC(bbr_saw_oerr); 13738 /* Clear all delay/early tracks */ 13739 bbr->r_ctl.rc_hptsi_agg_delay = 0; 13740 bbr->r_ctl.rc_agg_early = 0; 13741 bbr->r_agg_early_set = 0; 13742 bbr->output_error_seen = 1; 13743 if (bbr->oerror_cnt < 0xf) 13744 bbr->oerror_cnt++; 13745 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { 13746 /* drop the session */ 13747 return (-ENETDOWN); 13748 } 13749 switch (error) { 13750 case ENOBUFS: 13751 /* 13752 * Make this guy have to get ack's to send 13753 * more but lets make sure we don't 13754 * slam him below a T-O (1MSS). 13755 */ 13756 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 13757 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13758 bbr->r_ctl.rc_lost_bytes)) - maxseg; 13759 if (tp->snd_cwnd < maxseg) 13760 tp->snd_cwnd = maxseg; 13761 } 13762 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; 13763 BBR_STAT_INC(bbr_saw_enobuf); 13764 if (bbr->bbr_hdrw_pacing) 13765 counter_u64_add(bbr_hdwr_pacing_enobuf, 1); 13766 else 13767 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); 13768 /* 13769 * Here even in the enobuf's case we want to do our 13770 * state update. The reason being we may have been 13771 * called by the input function. If so we have had 13772 * things change. 13773 */ 13774 error = 0; 13775 goto enobufs; 13776 case EMSGSIZE: 13777 /* 13778 * For some reason the interface we used initially 13779 * to send segments changed to another or lowered 13780 * its MTU. If TSO was active we either got an 13781 * interface without TSO capabilits or TSO was 13782 * turned off. If we obtained mtu from ip_output() 13783 * then update it and try again. 13784 */ 13785 /* Turn on tracing (or try to) */ 13786 { 13787 int old_maxseg; 13788 13789 old_maxseg = tp->t_maxseg; 13790 BBR_STAT_INC(bbr_saw_emsgsiz); 13791 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); 13792 if (mtu != 0) 13793 tcp_mss_update(tp, -1, mtu, NULL, NULL); 13794 if (old_maxseg <= tp->t_maxseg) { 13795 /* Huh it did not shrink? */ 13796 tp->t_maxseg = old_maxseg - 40; 13797 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); 13798 } 13799 /* 13800 * Nuke all other things that can interfere 13801 * with slot 13802 */ 13803 if ((tot_len + len) && (len >= tp->t_maxseg)) { 13804 slot = bbr_get_pacing_delay(bbr, 13805 bbr->r_ctl.rc_bbr_hptsi_gain, 13806 (tot_len + len), cts, 0); 13807 if (slot < bbr_error_base_paceout) 13808 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13809 } else 13810 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13811 bbr->rc_output_starts_timer = 1; 13812 bbr_start_hpts_timer(bbr, tp, cts, 10, slot, 13813 tot_len); 13814 return (error); 13815 } 13816 case EPERM: 13817 tp->t_softerror = error; 13818 /* Fall through */ 13819 case EHOSTDOWN: 13820 case EHOSTUNREACH: 13821 case ENETDOWN: 13822 case ENETUNREACH: 13823 if (TCPS_HAVERCVDSYN(tp->t_state)) { 13824 tp->t_softerror = error; 13825 } 13826 /* FALLTHROUGH */ 13827 default: 13828 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; 13829 bbr->rc_output_starts_timer = 1; 13830 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); 13831 return (error); 13832 } 13833 #ifdef STATS 13834 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && 13835 len && 13836 (rsm == NULL) && 13837 (bbr->rc_in_persist == 0)) { 13838 tp->gput_seq = bbr_seq; 13839 tp->gput_ack = bbr_seq + 13840 min(sbavail(&so->so_snd) - sb_offset, sendwin); 13841 tp->gput_ts = cts; 13842 tp->t_flags |= TF_GPUTINPROG; 13843 #endif 13844 } 13845 KMOD_TCPSTAT_INC(tcps_sndtotal); 13846 if ((bbr->bbr_hdw_pace_ena) && 13847 (bbr->bbr_attempt_hdwr_pace == 0) && 13848 (bbr->rc_past_init_win) && 13849 (bbr->rc_bbr_state != BBR_STATE_STARTUP) && 13850 (get_filter_value(&bbr->r_ctl.rc_delrate)) && 13851 (inp->inp_route.ro_nh && 13852 inp->inp_route.ro_nh->nh_ifp)) { 13853 /* 13854 * We are past the initial window and 13855 * have at least one measurement so we 13856 * could use hardware pacing if its available. 13857 * We have an interface and we have not attempted 13858 * to setup hardware pacing, lets try to now. 13859 */ 13860 uint64_t rate_wanted; 13861 int err = 0; 13862 13863 rate_wanted = bbr_get_hardware_rate(bbr); 13864 bbr->bbr_attempt_hdwr_pace = 1; 13865 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, 13866 inp->inp_route.ro_nh->nh_ifp, 13867 rate_wanted, 13868 (RS_PACING_GEQ|RS_PACING_SUB_OK), 13869 &err, NULL); 13870 if (bbr->r_ctl.crte) { 13871 bbr_type_log_hdwr_pacing(bbr, 13872 bbr->r_ctl.crte->ptbl->rs_ifp, 13873 rate_wanted, 13874 bbr->r_ctl.crte->rate, 13875 __LINE__, cts, err); 13876 BBR_STAT_INC(bbr_hdwr_rl_add_ok); 13877 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 13878 counter_u64_add(bbr_flows_whdwr_pacing, 1); 13879 bbr->bbr_hdrw_pacing = 1; 13880 /* Now what is our gain status? */ 13881 if (bbr->r_ctl.crte->rate < rate_wanted) { 13882 /* We have a problem */ 13883 bbr_setup_less_of_rate(bbr, cts, 13884 bbr->r_ctl.crte->rate, rate_wanted); 13885 } else { 13886 /* We are good */ 13887 bbr->gain_is_limited = 0; 13888 bbr->skip_gain = 0; 13889 } 13890 tcp_bbr_tso_size_check(bbr, cts); 13891 } else { 13892 bbr_type_log_hdwr_pacing(bbr, 13893 inp->inp_route.ro_nh->nh_ifp, 13894 rate_wanted, 13895 0, 13896 __LINE__, cts, err); 13897 BBR_STAT_INC(bbr_hdwr_rl_add_fail); 13898 } 13899 } 13900 if (bbr->bbr_hdrw_pacing) { 13901 /* 13902 * Worry about cases where the route 13903 * changes or something happened that we 13904 * lost our hardware pacing possibly during 13905 * the last ip_output call. 13906 */ 13907 if (inp->inp_snd_tag == NULL) { 13908 /* A change during ip output disabled hw pacing? */ 13909 bbr->bbr_hdrw_pacing = 0; 13910 } else if ((inp->inp_route.ro_nh == NULL) || 13911 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { 13912 /* 13913 * We had an interface or route change, 13914 * detach from the current hdwr pacing 13915 * and setup to re-attempt next go 13916 * round. 13917 */ 13918 bbr->bbr_hdrw_pacing = 0; 13919 bbr->bbr_attempt_hdwr_pace = 0; 13920 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 13921 tcp_bbr_tso_size_check(bbr, cts); 13922 } 13923 } 13924 /* 13925 * Data sent (as far as we can tell). If this advertises a larger 13926 * window than any other segment, then remember the size of the 13927 * advertised window. Any pending ACK has now been sent. 13928 */ 13929 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 13930 tp->rcv_adv = tp->rcv_nxt + recwin; 13931 13932 tp->last_ack_sent = tp->rcv_nxt; 13933 if ((error == 0) && 13934 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && 13935 (doing_tlp == 0) && 13936 (tso == 0) && 13937 (len > 0) && 13938 ((flags & TH_RST) == 0) && 13939 ((flags & TH_SYN) == 0) && 13940 (IN_RECOVERY(tp->t_flags) == 0) && 13941 (bbr->rc_in_persist == 0) && 13942 (tot_len < bbr->r_ctl.rc_pace_max_segs)) { 13943 /* 13944 * For non-tso we need to goto again until we have sent out 13945 * enough data to match what we are hptsi out every hptsi 13946 * interval. 13947 */ 13948 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 13949 /* Make sure snd_nxt is drug up */ 13950 tp->snd_nxt = tp->snd_max; 13951 } 13952 if (rsm != NULL) { 13953 rsm = NULL; 13954 goto skip_again; 13955 } 13956 rsm = NULL; 13957 sack_rxmit = 0; 13958 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 13959 goto again; 13960 } 13961 skip_again: 13962 if ((error == 0) && (flags & TH_FIN)) 13963 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); 13964 if ((error == 0) && (flags & TH_RST)) 13965 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 13966 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { 13967 /* 13968 * Calculate/Re-Calculate the hptsi slot in usecs based on 13969 * what we have sent so far 13970 */ 13971 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 13972 if (bbr->rc_no_pacing) 13973 slot = 0; 13974 } 13975 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 13976 enobufs: 13977 if (bbr->rc_use_google == 0) 13978 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 13979 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13980 bbr->r_ctl.rc_lost_bytes))); 13981 bbr->rc_output_starts_timer = 1; 13982 if (bbr->bbr_use_rack_cheat && 13983 (more_to_rxt || 13984 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { 13985 /* Rack cheats and shotguns out all rxt's 1ms apart */ 13986 if (slot > 1000) 13987 slot = 1000; 13988 } 13989 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { 13990 /* 13991 * We don't change the tso size until some number of sends 13992 * to give the hardware commands time to get down 13993 * to the interface. 13994 */ 13995 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; 13996 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { 13997 bbr->hw_pacing_set = 1; 13998 tcp_bbr_tso_size_check(bbr, cts); 13999 } 14000 } 14001 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); 14002 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14003 /* Make sure snd_nxt is drug up */ 14004 tp->snd_nxt = tp->snd_max; 14005 } 14006 return (error); 14007 14008 } 14009 14010 /* 14011 * See bbr_output_wtime() for return values. 14012 */ 14013 static int 14014 bbr_output(struct tcpcb *tp) 14015 { 14016 int32_t ret; 14017 struct timeval tv; 14018 14019 NET_EPOCH_ASSERT(); 14020 14021 INP_WLOCK_ASSERT(tptoinpcb(tp)); 14022 (void)tcp_get_usecs(&tv); 14023 ret = bbr_output_wtime(tp, &tv); 14024 return (ret); 14025 } 14026 14027 static void 14028 bbr_mtu_chg(struct tcpcb *tp) 14029 { 14030 struct tcp_bbr *bbr; 14031 struct bbr_sendmap *rsm, *frsm = NULL; 14032 uint32_t maxseg; 14033 14034 /* 14035 * The MTU has changed. a) Clear the sack filter. b) Mark everything 14036 * over the current size as SACK_PASS so a retransmit will occur. 14037 */ 14038 14039 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14040 maxseg = tp->t_maxseg - bbr->rc_last_options; 14041 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 14042 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 14043 /* Don't mess with ones acked (by sack?) */ 14044 if (rsm->r_flags & BBR_ACKED) 14045 continue; 14046 if ((rsm->r_end - rsm->r_start) > maxseg) { 14047 /* 14048 * We mark sack-passed on all the previous large 14049 * sends we did. This will force them to retransmit. 14050 */ 14051 rsm->r_flags |= BBR_SACK_PASSED; 14052 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && 14053 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { 14054 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 14055 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 14056 rsm->r_flags |= BBR_MARKED_LOST; 14057 } 14058 if (frsm == NULL) 14059 frsm = rsm; 14060 } 14061 } 14062 if (frsm) { 14063 bbr->r_ctl.rc_resend = frsm; 14064 } 14065 } 14066 14067 static int 14068 bbr_pru_options(struct tcpcb *tp, int flags) 14069 { 14070 if (flags & PRUS_OOB) 14071 return (EOPNOTSUPP); 14072 return (0); 14073 } 14074 14075 struct tcp_function_block __tcp_bbr = { 14076 .tfb_tcp_block_name = __XSTRING(STACKNAME), 14077 .tfb_tcp_output = bbr_output, 14078 .tfb_do_queued_segments = ctf_do_queued_segments, 14079 .tfb_do_segment_nounlock = bbr_do_segment_nounlock, 14080 .tfb_tcp_do_segment = bbr_do_segment, 14081 .tfb_tcp_ctloutput = bbr_ctloutput, 14082 .tfb_tcp_fb_init = bbr_init, 14083 .tfb_tcp_fb_fini = bbr_fini, 14084 .tfb_tcp_timer_stop_all = bbr_stopall, 14085 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, 14086 .tfb_tcp_handoff_ok = bbr_handoff_ok, 14087 .tfb_tcp_mtu_chg = bbr_mtu_chg, 14088 .tfb_pru_options = bbr_pru_options, 14089 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, 14090 }; 14091 14092 /* 14093 * bbr_ctloutput() must drop the inpcb lock before performing copyin on 14094 * socket option arguments. When it re-acquires the lock after the copy, it 14095 * has to revalidate that the connection is still valid for the socket 14096 * option. 14097 */ 14098 static int 14099 bbr_set_sockopt(struct inpcb *inp, struct sockopt *sopt) 14100 { 14101 struct epoch_tracker et; 14102 struct tcpcb *tp; 14103 struct tcp_bbr *bbr; 14104 int32_t error = 0, optval; 14105 14106 switch (sopt->sopt_level) { 14107 case IPPROTO_IPV6: 14108 case IPPROTO_IP: 14109 return (tcp_default_ctloutput(inp, sopt)); 14110 } 14111 14112 switch (sopt->sopt_name) { 14113 case TCP_RACK_PACE_MAX_SEG: 14114 case TCP_RACK_MIN_TO: 14115 case TCP_RACK_REORD_THRESH: 14116 case TCP_RACK_REORD_FADE: 14117 case TCP_RACK_TLP_THRESH: 14118 case TCP_RACK_PKT_DELAY: 14119 case TCP_BBR_ALGORITHM: 14120 case TCP_BBR_TSLIMITS: 14121 case TCP_BBR_IWINTSO: 14122 case TCP_BBR_RECFORCE: 14123 case TCP_BBR_STARTUP_PG: 14124 case TCP_BBR_DRAIN_PG: 14125 case TCP_BBR_RWND_IS_APP: 14126 case TCP_BBR_PROBE_RTT_INT: 14127 case TCP_BBR_PROBE_RTT_GAIN: 14128 case TCP_BBR_PROBE_RTT_LEN: 14129 case TCP_BBR_STARTUP_LOSS_EXIT: 14130 case TCP_BBR_USEDEL_RATE: 14131 case TCP_BBR_MIN_RTO: 14132 case TCP_BBR_MAX_RTO: 14133 case TCP_BBR_PACE_PER_SEC: 14134 case TCP_DELACK: 14135 case TCP_BBR_PACE_DEL_TAR: 14136 case TCP_BBR_SEND_IWND_IN_TSO: 14137 case TCP_BBR_EXTRA_STATE: 14138 case TCP_BBR_UTTER_MAX_TSO: 14139 case TCP_BBR_MIN_TOPACEOUT: 14140 case TCP_BBR_FLOOR_MIN_TSO: 14141 case TCP_BBR_TSTMP_RAISES: 14142 case TCP_BBR_POLICER_DETECT: 14143 case TCP_BBR_USE_RACK_CHEAT: 14144 case TCP_DATA_AFTER_CLOSE: 14145 case TCP_BBR_HDWR_PACE: 14146 case TCP_BBR_PACE_SEG_MAX: 14147 case TCP_BBR_PACE_SEG_MIN: 14148 case TCP_BBR_PACE_CROSS: 14149 case TCP_BBR_PACE_OH: 14150 #ifdef NETFLIX_PEAKRATE 14151 case TCP_MAXPEAKRATE: 14152 #endif 14153 case TCP_BBR_TMR_PACE_OH: 14154 case TCP_BBR_RACK_RTT_USE: 14155 case TCP_BBR_RETRAN_WTSO: 14156 break; 14157 default: 14158 return (tcp_default_ctloutput(inp, sopt)); 14159 break; 14160 } 14161 INP_WUNLOCK(inp); 14162 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 14163 if (error) 14164 return (error); 14165 INP_WLOCK(inp); 14166 if (inp->inp_flags & INP_DROPPED) { 14167 INP_WUNLOCK(inp); 14168 return (ECONNRESET); 14169 } 14170 tp = intotcpcb(inp); 14171 if (tp->t_fb != &__tcp_bbr) { 14172 INP_WUNLOCK(inp); 14173 return (ENOPROTOOPT); 14174 } 14175 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14176 switch (sopt->sopt_name) { 14177 case TCP_BBR_PACE_PER_SEC: 14178 BBR_OPTS_INC(tcp_bbr_pace_per_sec); 14179 bbr->r_ctl.bbr_hptsi_per_second = optval; 14180 break; 14181 case TCP_BBR_PACE_DEL_TAR: 14182 BBR_OPTS_INC(tcp_bbr_pace_del_tar); 14183 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; 14184 break; 14185 case TCP_BBR_PACE_SEG_MAX: 14186 BBR_OPTS_INC(tcp_bbr_pace_seg_max); 14187 bbr->r_ctl.bbr_hptsi_segments_max = optval; 14188 break; 14189 case TCP_BBR_PACE_SEG_MIN: 14190 BBR_OPTS_INC(tcp_bbr_pace_seg_min); 14191 bbr->r_ctl.bbr_hptsi_bytes_min = optval; 14192 break; 14193 case TCP_BBR_PACE_CROSS: 14194 BBR_OPTS_INC(tcp_bbr_pace_cross); 14195 bbr->r_ctl.bbr_cross_over = optval; 14196 break; 14197 case TCP_BBR_ALGORITHM: 14198 BBR_OPTS_INC(tcp_bbr_algorithm); 14199 if (optval && (bbr->rc_use_google == 0)) { 14200 /* Turn on the google mode */ 14201 bbr_google_mode_on(bbr); 14202 if ((optval > 3) && (optval < 500)) { 14203 /* 14204 * Must be at least greater than .3% 14205 * and must be less than 50.0%. 14206 */ 14207 bbr->r_ctl.bbr_google_discount = optval; 14208 } 14209 } else if ((optval == 0) && (bbr->rc_use_google == 1)) { 14210 /* Turn off the google mode */ 14211 bbr_google_mode_off(bbr); 14212 } 14213 break; 14214 case TCP_BBR_TSLIMITS: 14215 BBR_OPTS_INC(tcp_bbr_tslimits); 14216 if (optval == 1) 14217 bbr->rc_use_ts_limit = 1; 14218 else if (optval == 0) 14219 bbr->rc_use_ts_limit = 0; 14220 else 14221 error = EINVAL; 14222 break; 14223 14224 case TCP_BBR_IWINTSO: 14225 BBR_OPTS_INC(tcp_bbr_iwintso); 14226 if ((optval >= 0) && (optval < 128)) { 14227 uint32_t twin; 14228 14229 bbr->rc_init_win = optval; 14230 twin = bbr_initial_cwnd(bbr, tp); 14231 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) 14232 tp->snd_cwnd = twin; 14233 else 14234 error = EBUSY; 14235 } else 14236 error = EINVAL; 14237 break; 14238 case TCP_BBR_STARTUP_PG: 14239 BBR_OPTS_INC(tcp_bbr_startup_pg); 14240 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { 14241 bbr->r_ctl.rc_startup_pg = optval; 14242 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 14243 bbr->r_ctl.rc_bbr_hptsi_gain = optval; 14244 } 14245 } else 14246 error = EINVAL; 14247 break; 14248 case TCP_BBR_DRAIN_PG: 14249 BBR_OPTS_INC(tcp_bbr_drain_pg); 14250 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) 14251 bbr->r_ctl.rc_drain_pg = optval; 14252 else 14253 error = EINVAL; 14254 break; 14255 case TCP_BBR_PROBE_RTT_LEN: 14256 BBR_OPTS_INC(tcp_bbr_probertt_len); 14257 if (optval <= 1) 14258 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); 14259 else 14260 error = EINVAL; 14261 break; 14262 case TCP_BBR_PROBE_RTT_GAIN: 14263 BBR_OPTS_INC(tcp_bbr_probertt_gain); 14264 if (optval <= BBR_UNIT) 14265 bbr->r_ctl.bbr_rttprobe_gain_val = optval; 14266 else 14267 error = EINVAL; 14268 break; 14269 case TCP_BBR_PROBE_RTT_INT: 14270 BBR_OPTS_INC(tcp_bbr_probe_rtt_int); 14271 if (optval > 1000) 14272 bbr->r_ctl.rc_probertt_int = optval; 14273 else 14274 error = EINVAL; 14275 break; 14276 case TCP_BBR_MIN_TOPACEOUT: 14277 BBR_OPTS_INC(tcp_bbr_topaceout); 14278 if (optval == 0) { 14279 bbr->no_pacing_until = 0; 14280 bbr->rc_no_pacing = 0; 14281 } else if (optval <= 0x00ff) { 14282 bbr->no_pacing_until = optval; 14283 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && 14284 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ 14285 /* Turn on no pacing */ 14286 bbr->rc_no_pacing = 1; 14287 } 14288 } else 14289 error = EINVAL; 14290 break; 14291 case TCP_BBR_STARTUP_LOSS_EXIT: 14292 BBR_OPTS_INC(tcp_bbr_startup_loss_exit); 14293 bbr->rc_loss_exit = optval; 14294 break; 14295 case TCP_BBR_USEDEL_RATE: 14296 error = EINVAL; 14297 break; 14298 case TCP_BBR_MIN_RTO: 14299 BBR_OPTS_INC(tcp_bbr_min_rto); 14300 bbr->r_ctl.rc_min_rto_ms = optval; 14301 break; 14302 case TCP_BBR_MAX_RTO: 14303 BBR_OPTS_INC(tcp_bbr_max_rto); 14304 bbr->rc_max_rto_sec = optval; 14305 break; 14306 case TCP_RACK_MIN_TO: 14307 /* Minimum time between rack t-o's in ms */ 14308 BBR_OPTS_INC(tcp_rack_min_to); 14309 bbr->r_ctl.rc_min_to = optval; 14310 break; 14311 case TCP_RACK_REORD_THRESH: 14312 /* RACK reorder threshold (shift amount) */ 14313 BBR_OPTS_INC(tcp_rack_reord_thresh); 14314 if ((optval > 0) && (optval < 31)) 14315 bbr->r_ctl.rc_reorder_shift = optval; 14316 else 14317 error = EINVAL; 14318 break; 14319 case TCP_RACK_REORD_FADE: 14320 /* Does reordering fade after ms time */ 14321 BBR_OPTS_INC(tcp_rack_reord_fade); 14322 bbr->r_ctl.rc_reorder_fade = optval; 14323 break; 14324 case TCP_RACK_TLP_THRESH: 14325 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14326 BBR_OPTS_INC(tcp_rack_tlp_thresh); 14327 if (optval) 14328 bbr->rc_tlp_threshold = optval; 14329 else 14330 error = EINVAL; 14331 break; 14332 case TCP_BBR_USE_RACK_CHEAT: 14333 BBR_OPTS_INC(tcp_use_rackcheat); 14334 if (bbr->rc_use_google) { 14335 error = EINVAL; 14336 break; 14337 } 14338 BBR_OPTS_INC(tcp_rack_cheat); 14339 if (optval) 14340 bbr->bbr_use_rack_cheat = 1; 14341 else 14342 bbr->bbr_use_rack_cheat = 0; 14343 break; 14344 case TCP_BBR_FLOOR_MIN_TSO: 14345 BBR_OPTS_INC(tcp_utter_max_tso); 14346 if ((optval >= 0) && (optval < 40)) 14347 bbr->r_ctl.bbr_hptsi_segments_floor = optval; 14348 else 14349 error = EINVAL; 14350 break; 14351 case TCP_BBR_UTTER_MAX_TSO: 14352 BBR_OPTS_INC(tcp_utter_max_tso); 14353 if ((optval >= 0) && (optval < 0xffff)) 14354 bbr->r_ctl.bbr_utter_max = optval; 14355 else 14356 error = EINVAL; 14357 break; 14358 14359 case TCP_BBR_EXTRA_STATE: 14360 BBR_OPTS_INC(tcp_extra_state); 14361 if (optval) 14362 bbr->rc_use_idle_restart = 1; 14363 else 14364 bbr->rc_use_idle_restart = 0; 14365 break; 14366 case TCP_BBR_SEND_IWND_IN_TSO: 14367 BBR_OPTS_INC(tcp_iwnd_tso); 14368 if (optval) { 14369 bbr->bbr_init_win_cheat = 1; 14370 if (bbr->rc_past_init_win == 0) { 14371 uint32_t cts; 14372 cts = tcp_get_usecs(&bbr->rc_tv); 14373 tcp_bbr_tso_size_check(bbr, cts); 14374 } 14375 } else 14376 bbr->bbr_init_win_cheat = 0; 14377 break; 14378 case TCP_BBR_HDWR_PACE: 14379 BBR_OPTS_INC(tcp_hdwr_pacing); 14380 if (optval){ 14381 bbr->bbr_hdw_pace_ena = 1; 14382 bbr->bbr_attempt_hdwr_pace = 0; 14383 } else { 14384 bbr->bbr_hdw_pace_ena = 0; 14385 #ifdef RATELIMIT 14386 if (bbr->r_ctl.crte != NULL) { 14387 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 14388 bbr->r_ctl.crte = NULL; 14389 } 14390 #endif 14391 } 14392 break; 14393 14394 case TCP_DELACK: 14395 BBR_OPTS_INC(tcp_delack); 14396 if (optval < 100) { 14397 if (optval == 0) /* off */ 14398 tp->t_delayed_ack = 0; 14399 else if (optval == 1) /* on which is 2 */ 14400 tp->t_delayed_ack = 2; 14401 else /* higher than 2 and less than 100 */ 14402 tp->t_delayed_ack = optval; 14403 if (tp->t_flags & TF_DELACK) { 14404 tp->t_flags &= ~TF_DELACK; 14405 tp->t_flags |= TF_ACKNOW; 14406 NET_EPOCH_ENTER(et); 14407 bbr_output(tp); 14408 NET_EPOCH_EXIT(et); 14409 } 14410 } else 14411 error = EINVAL; 14412 break; 14413 case TCP_RACK_PKT_DELAY: 14414 /* RACK added ms i.e. rack-rtt + reord + N */ 14415 BBR_OPTS_INC(tcp_rack_pkt_delay); 14416 bbr->r_ctl.rc_pkt_delay = optval; 14417 break; 14418 #ifdef NETFLIX_PEAKRATE 14419 case TCP_MAXPEAKRATE: 14420 BBR_OPTS_INC(tcp_maxpeak); 14421 error = tcp_set_maxpeakrate(tp, optval); 14422 if (!error) 14423 tp->t_peakrate_thr = tp->t_maxpeakrate; 14424 break; 14425 #endif 14426 case TCP_BBR_RETRAN_WTSO: 14427 BBR_OPTS_INC(tcp_retran_wtso); 14428 if (optval) 14429 bbr->rc_resends_use_tso = 1; 14430 else 14431 bbr->rc_resends_use_tso = 0; 14432 break; 14433 case TCP_DATA_AFTER_CLOSE: 14434 BBR_OPTS_INC(tcp_data_ac); 14435 if (optval) 14436 bbr->rc_allow_data_af_clo = 1; 14437 else 14438 bbr->rc_allow_data_af_clo = 0; 14439 break; 14440 case TCP_BBR_POLICER_DETECT: 14441 BBR_OPTS_INC(tcp_policer_det); 14442 if (bbr->rc_use_google == 0) 14443 error = EINVAL; 14444 else if (optval) 14445 bbr->r_use_policer = 1; 14446 else 14447 bbr->r_use_policer = 0; 14448 break; 14449 14450 case TCP_BBR_TSTMP_RAISES: 14451 BBR_OPTS_INC(tcp_ts_raises); 14452 if (optval) 14453 bbr->ts_can_raise = 1; 14454 else 14455 bbr->ts_can_raise = 0; 14456 break; 14457 case TCP_BBR_TMR_PACE_OH: 14458 BBR_OPTS_INC(tcp_pacing_oh_tmr); 14459 if (bbr->rc_use_google) { 14460 error = EINVAL; 14461 } else { 14462 if (optval) 14463 bbr->r_ctl.rc_incr_tmrs = 1; 14464 else 14465 bbr->r_ctl.rc_incr_tmrs = 0; 14466 } 14467 break; 14468 case TCP_BBR_PACE_OH: 14469 BBR_OPTS_INC(tcp_pacing_oh); 14470 if (bbr->rc_use_google) { 14471 error = EINVAL; 14472 } else { 14473 if (optval > (BBR_INCL_TCP_OH| 14474 BBR_INCL_IP_OH| 14475 BBR_INCL_ENET_OH)) { 14476 error = EINVAL; 14477 break; 14478 } 14479 if (optval & BBR_INCL_TCP_OH) 14480 bbr->r_ctl.rc_inc_tcp_oh = 1; 14481 else 14482 bbr->r_ctl.rc_inc_tcp_oh = 0; 14483 if (optval & BBR_INCL_IP_OH) 14484 bbr->r_ctl.rc_inc_ip_oh = 1; 14485 else 14486 bbr->r_ctl.rc_inc_ip_oh = 0; 14487 if (optval & BBR_INCL_ENET_OH) 14488 bbr->r_ctl.rc_inc_enet_oh = 1; 14489 else 14490 bbr->r_ctl.rc_inc_enet_oh = 0; 14491 } 14492 break; 14493 default: 14494 return (tcp_default_ctloutput(inp, sopt)); 14495 break; 14496 } 14497 #ifdef NETFLIX_STATS 14498 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 14499 #endif 14500 INP_WUNLOCK(inp); 14501 return (error); 14502 } 14503 14504 /* 14505 * return 0 on success, error-num on failure 14506 */ 14507 static int 14508 bbr_get_sockopt(struct inpcb *inp, struct sockopt *sopt) 14509 { 14510 struct tcpcb *tp; 14511 struct tcp_bbr *bbr; 14512 int32_t error, optval; 14513 14514 tp = intotcpcb(inp); 14515 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14516 if (bbr == NULL) { 14517 INP_WUNLOCK(inp); 14518 return (EINVAL); 14519 } 14520 /* 14521 * Because all our options are either boolean or an int, we can just 14522 * pull everything into optval and then unlock and copy. If we ever 14523 * add a option that is not a int, then this will have quite an 14524 * impact to this routine. 14525 */ 14526 switch (sopt->sopt_name) { 14527 case TCP_BBR_PACE_PER_SEC: 14528 optval = bbr->r_ctl.bbr_hptsi_per_second; 14529 break; 14530 case TCP_BBR_PACE_DEL_TAR: 14531 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; 14532 break; 14533 case TCP_BBR_PACE_SEG_MAX: 14534 optval = bbr->r_ctl.bbr_hptsi_segments_max; 14535 break; 14536 case TCP_BBR_MIN_TOPACEOUT: 14537 optval = bbr->no_pacing_until; 14538 break; 14539 case TCP_BBR_PACE_SEG_MIN: 14540 optval = bbr->r_ctl.bbr_hptsi_bytes_min; 14541 break; 14542 case TCP_BBR_PACE_CROSS: 14543 optval = bbr->r_ctl.bbr_cross_over; 14544 break; 14545 case TCP_BBR_ALGORITHM: 14546 optval = bbr->rc_use_google; 14547 break; 14548 case TCP_BBR_TSLIMITS: 14549 optval = bbr->rc_use_ts_limit; 14550 break; 14551 case TCP_BBR_IWINTSO: 14552 optval = bbr->rc_init_win; 14553 break; 14554 case TCP_BBR_STARTUP_PG: 14555 optval = bbr->r_ctl.rc_startup_pg; 14556 break; 14557 case TCP_BBR_DRAIN_PG: 14558 optval = bbr->r_ctl.rc_drain_pg; 14559 break; 14560 case TCP_BBR_PROBE_RTT_INT: 14561 optval = bbr->r_ctl.rc_probertt_int; 14562 break; 14563 case TCP_BBR_PROBE_RTT_LEN: 14564 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); 14565 break; 14566 case TCP_BBR_PROBE_RTT_GAIN: 14567 optval = bbr->r_ctl.bbr_rttprobe_gain_val; 14568 break; 14569 case TCP_BBR_STARTUP_LOSS_EXIT: 14570 optval = bbr->rc_loss_exit; 14571 break; 14572 case TCP_BBR_USEDEL_RATE: 14573 error = EINVAL; 14574 break; 14575 case TCP_BBR_MIN_RTO: 14576 optval = bbr->r_ctl.rc_min_rto_ms; 14577 break; 14578 case TCP_BBR_MAX_RTO: 14579 optval = bbr->rc_max_rto_sec; 14580 break; 14581 case TCP_RACK_PACE_MAX_SEG: 14582 /* Max segments in a pace */ 14583 optval = bbr->r_ctl.rc_pace_max_segs; 14584 break; 14585 case TCP_RACK_MIN_TO: 14586 /* Minimum time between rack t-o's in ms */ 14587 optval = bbr->r_ctl.rc_min_to; 14588 break; 14589 case TCP_RACK_REORD_THRESH: 14590 /* RACK reorder threshold (shift amount) */ 14591 optval = bbr->r_ctl.rc_reorder_shift; 14592 break; 14593 case TCP_RACK_REORD_FADE: 14594 /* Does reordering fade after ms time */ 14595 optval = bbr->r_ctl.rc_reorder_fade; 14596 break; 14597 case TCP_BBR_USE_RACK_CHEAT: 14598 /* Do we use the rack cheat for rxt */ 14599 optval = bbr->bbr_use_rack_cheat; 14600 break; 14601 case TCP_BBR_FLOOR_MIN_TSO: 14602 optval = bbr->r_ctl.bbr_hptsi_segments_floor; 14603 break; 14604 case TCP_BBR_UTTER_MAX_TSO: 14605 optval = bbr->r_ctl.bbr_utter_max; 14606 break; 14607 case TCP_BBR_SEND_IWND_IN_TSO: 14608 /* Do we send TSO size segments initially */ 14609 optval = bbr->bbr_init_win_cheat; 14610 break; 14611 case TCP_BBR_EXTRA_STATE: 14612 optval = bbr->rc_use_idle_restart; 14613 break; 14614 case TCP_RACK_TLP_THRESH: 14615 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14616 optval = bbr->rc_tlp_threshold; 14617 break; 14618 case TCP_RACK_PKT_DELAY: 14619 /* RACK added ms i.e. rack-rtt + reord + N */ 14620 optval = bbr->r_ctl.rc_pkt_delay; 14621 break; 14622 case TCP_BBR_RETRAN_WTSO: 14623 optval = bbr->rc_resends_use_tso; 14624 break; 14625 case TCP_DATA_AFTER_CLOSE: 14626 optval = bbr->rc_allow_data_af_clo; 14627 break; 14628 case TCP_DELACK: 14629 optval = tp->t_delayed_ack; 14630 break; 14631 case TCP_BBR_HDWR_PACE: 14632 optval = bbr->bbr_hdw_pace_ena; 14633 break; 14634 case TCP_BBR_POLICER_DETECT: 14635 optval = bbr->r_use_policer; 14636 break; 14637 case TCP_BBR_TSTMP_RAISES: 14638 optval = bbr->ts_can_raise; 14639 break; 14640 case TCP_BBR_TMR_PACE_OH: 14641 optval = bbr->r_ctl.rc_incr_tmrs; 14642 break; 14643 case TCP_BBR_PACE_OH: 14644 optval = 0; 14645 if (bbr->r_ctl.rc_inc_tcp_oh) 14646 optval |= BBR_INCL_TCP_OH; 14647 if (bbr->r_ctl.rc_inc_ip_oh) 14648 optval |= BBR_INCL_IP_OH; 14649 if (bbr->r_ctl.rc_inc_enet_oh) 14650 optval |= BBR_INCL_ENET_OH; 14651 break; 14652 default: 14653 return (tcp_default_ctloutput(inp, sopt)); 14654 break; 14655 } 14656 INP_WUNLOCK(inp); 14657 error = sooptcopyout(sopt, &optval, sizeof optval); 14658 return (error); 14659 } 14660 14661 /* 14662 * return 0 on success, error-num on failure 14663 */ 14664 static int 14665 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt) 14666 { 14667 if (sopt->sopt_dir == SOPT_SET) { 14668 return (bbr_set_sockopt(inp, sopt)); 14669 } else if (sopt->sopt_dir == SOPT_GET) { 14670 return (bbr_get_sockopt(inp, sopt)); 14671 } else { 14672 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); 14673 } 14674 } 14675 14676 static const char *bbr_stack_names[] = { 14677 __XSTRING(STACKNAME), 14678 #ifdef STACKALIAS 14679 __XSTRING(STACKALIAS), 14680 #endif 14681 }; 14682 14683 static bool bbr_mod_inited = false; 14684 14685 static int 14686 tcp_addbbr(module_t mod, int32_t type, void *data) 14687 { 14688 int32_t err = 0; 14689 int num_stacks; 14690 14691 switch (type) { 14692 case MOD_LOAD: 14693 printf("Attempting to load " __XSTRING(MODNAME) "\n"); 14694 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 14695 sizeof(struct bbr_sendmap), 14696 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 14697 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 14698 sizeof(struct tcp_bbr), 14699 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 14700 sysctl_ctx_init(&bbr_sysctl_ctx); 14701 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 14702 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 14703 OID_AUTO, 14704 #ifdef STACKALIAS 14705 __XSTRING(STACKALIAS), 14706 #else 14707 __XSTRING(STACKNAME), 14708 #endif 14709 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 14710 ""); 14711 if (bbr_sysctl_root == NULL) { 14712 printf("Failed to add sysctl node\n"); 14713 err = EFAULT; 14714 goto free_uma; 14715 } 14716 bbr_init_sysctls(); 14717 num_stacks = nitems(bbr_stack_names); 14718 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, 14719 bbr_stack_names, &num_stacks); 14720 if (err) { 14721 printf("Failed to register %s stack name for " 14722 "%s module\n", bbr_stack_names[num_stacks], 14723 __XSTRING(MODNAME)); 14724 sysctl_ctx_free(&bbr_sysctl_ctx); 14725 free_uma: 14726 uma_zdestroy(bbr_zone); 14727 uma_zdestroy(bbr_pcb_zone); 14728 bbr_counter_destroy(); 14729 printf("Failed to register " __XSTRING(MODNAME) 14730 " module err:%d\n", err); 14731 return (err); 14732 } 14733 tcp_lro_reg_mbufq(); 14734 bbr_mod_inited = true; 14735 printf(__XSTRING(MODNAME) " is now available\n"); 14736 break; 14737 case MOD_QUIESCE: 14738 err = deregister_tcp_functions(&__tcp_bbr, true, false); 14739 break; 14740 case MOD_UNLOAD: 14741 err = deregister_tcp_functions(&__tcp_bbr, false, true); 14742 if (err == EBUSY) 14743 break; 14744 if (bbr_mod_inited) { 14745 uma_zdestroy(bbr_zone); 14746 uma_zdestroy(bbr_pcb_zone); 14747 sysctl_ctx_free(&bbr_sysctl_ctx); 14748 bbr_counter_destroy(); 14749 printf(__XSTRING(MODNAME) 14750 " is now no longer available\n"); 14751 bbr_mod_inited = false; 14752 } 14753 tcp_lro_dereg_mbufq(); 14754 err = 0; 14755 break; 14756 default: 14757 return (EOPNOTSUPP); 14758 } 14759 return (err); 14760 } 14761 14762 static moduledata_t tcp_bbr = { 14763 .name = __XSTRING(MODNAME), 14764 .evhand = tcp_addbbr, 14765 .priv = 0 14766 }; 14767 14768 MODULE_VERSION(MODNAME, 1); 14769 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 14770 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 14771