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_tcpdebug.h" 40 #include "opt_ratelimit.h" 41 #include <sys/param.h> 42 #include <sys/arb.h> 43 #include <sys/module.h> 44 #include <sys/kernel.h> 45 #include <sys/libkern.h> 46 #ifdef TCP_HHOOK 47 #include <sys/hhook.h> 48 #endif 49 #include <sys/malloc.h> 50 #include <sys/mbuf.h> 51 #include <sys/proc.h> 52 #include <sys/socket.h> 53 #include <sys/socketvar.h> 54 #include <sys/sysctl.h> 55 #include <sys/systm.h> 56 #ifdef STATS 57 #include <sys/qmath.h> 58 #include <sys/tree.h> 59 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 60 #endif 61 #include <sys/refcount.h> 62 #include <sys/queue.h> 63 #include <sys/eventhandler.h> 64 #include <sys/smp.h> 65 #include <sys/kthread.h> 66 #include <sys/lock.h> 67 #include <sys/mutex.h> 68 #include <sys/tim_filter.h> 69 #include <sys/time.h> 70 #include <sys/protosw.h> 71 #include <vm/uma.h> 72 #include <sys/kern_prefetch.h> 73 74 #include <net/route.h> 75 #include <net/route/nhop.h> 76 #include <net/vnet.h> 77 78 #define TCPSTATES /* for logging */ 79 80 #include <netinet/in.h> 81 #include <netinet/in_kdtrace.h> 82 #include <netinet/in_pcb.h> 83 #include <netinet/ip.h> 84 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 85 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 86 #include <netinet/ip_var.h> 87 #include <netinet/ip6.h> 88 #include <netinet6/in6_pcb.h> 89 #include <netinet6/ip6_var.h> 90 #define TCPOUTFLAGS 91 #include <netinet/tcp.h> 92 #include <netinet/tcp_fsm.h> 93 #include <netinet/tcp_seq.h> 94 #include <netinet/tcp_timer.h> 95 #include <netinet/tcp_var.h> 96 #include <netinet/tcpip.h> 97 #include <netinet/tcp_hpts.h> 98 #include <netinet/cc/cc.h> 99 #include <netinet/tcp_log_buf.h> 100 #include <netinet/tcp_ratelimit.h> 101 #include <netinet/tcp_lro.h> 102 #ifdef TCPDEBUG 103 #include <netinet/tcp_debug.h> 104 #endif /* TCPDEBUG */ 105 #ifdef TCP_OFFLOAD 106 #include <netinet/tcp_offload.h> 107 #endif 108 #ifdef INET6 109 #include <netinet6/tcp6_var.h> 110 #endif 111 #include <netinet/tcp_fastopen.h> 112 113 #include <netipsec/ipsec_support.h> 114 #include <net/if.h> 115 #include <net/if_var.h> 116 #include <net/ethernet.h> 117 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 119 #include <netipsec/ipsec.h> 120 #include <netipsec/ipsec6.h> 121 #endif /* IPSEC */ 122 123 #include <netinet/udp.h> 124 #include <netinet/udp_var.h> 125 #include <machine/in_cksum.h> 126 127 #ifdef MAC 128 #include <security/mac/mac_framework.h> 129 #endif 130 131 #include "sack_filter.h" 132 #include "tcp_bbr.h" 133 #include "rack_bbr_common.h" 134 uma_zone_t bbr_zone; 135 uma_zone_t bbr_pcb_zone; 136 137 struct sysctl_ctx_list bbr_sysctl_ctx; 138 struct sysctl_oid *bbr_sysctl_root; 139 140 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ 141 (tv) = (value); \ 142 if ((u_long)(tv) < (u_long)(tvmin)) \ 143 (tv) = (tvmin); \ 144 if ((u_long)(tv) > (u_long)(tvmax)) \ 145 (tv) = (tvmax); \ 146 } while(0) 147 148 /*#define BBR_INVARIANT 1*/ 149 150 /* 151 * initial window 152 */ 153 static uint32_t bbr_def_init_win = 10; 154 static int32_t bbr_persist_min = 250000; /* 250ms */ 155 static int32_t bbr_persist_max = 1000000; /* 1 Second */ 156 static int32_t bbr_cwnd_may_shrink = 0; 157 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; 158 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; 159 static int32_t bbr_hardware_pacing_limit = 8000; 160 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ 161 static int32_t bbr_no_retran = 0; 162 163 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ 164 static int32_t bbr_max_net_error_cnt = 10; 165 /* Should the following be dynamic too -- loss wise */ 166 static int32_t bbr_rtt_gain_thresh = 0; 167 /* Measurement controls */ 168 static int32_t bbr_use_google_algo = 1; 169 static int32_t bbr_ts_limiting = 1; 170 static int32_t bbr_ts_can_raise = 0; 171 static int32_t bbr_do_red = 600; 172 static int32_t bbr_red_scale = 20000; 173 static int32_t bbr_red_mul = 1; 174 static int32_t bbr_red_div = 2; 175 static int32_t bbr_red_growth_restrict = 1; 176 static int32_t bbr_target_is_bbunit = 0; 177 static int32_t bbr_drop_limit = 0; 178 /* 179 * How much gain do we need to see to 180 * stay in startup? 181 */ 182 static int32_t bbr_marks_rxt_sack_passed = 0; 183 static int32_t bbr_start_exit = 25; 184 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ 185 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ 186 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ 187 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this 188 * if we go back ever to where the pacer 189 * has priority over timers. 190 */ 191 static int32_t bbr_policer_call_from_rack_to = 0; 192 static int32_t bbr_policer_detection_enabled = 1; 193 static int32_t bbr_min_measurements_req = 1; /* We need at least 2 194 * measurements before we are 195 * "good" note that 2 == 1. 196 * This is because we use a > 197 * comparison. This means if 198 * min_measure was 0, it takes 199 * num-measures > min(0) and 200 * you get 1 measurement and 201 * you are good. Set to 1, you 202 * have to have two 203 * measurements (this is done 204 * to prevent it from being ok 205 * to have no measurements). */ 206 static int32_t bbr_no_pacing_until = 4; 207 208 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ 209 static int32_t bbr_min_peer_delta = 20; /* 20 units */ 210 static int32_t bbr_delta_percent = 150; /* 15.0 % */ 211 212 static int32_t bbr_target_cwnd_mult_limit = 8; 213 /* 214 * bbr_cwnd_min_val is the number of 215 * segments we hold to in the RTT probe 216 * state typically 4. 217 */ 218 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; 219 220 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; 221 222 static int32_t bbr_gain_to_target = 1; 223 static int32_t bbr_gain_gets_extra_too = 1; 224 /* 225 * bbr_high_gain is the 2/ln(2) value we need 226 * to double the sending rate in startup. This 227 * is used for both cwnd and hptsi gain's. 228 */ 229 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 230 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; 231 static int32_t bbr_use_lower_gain_in_startup = 1; 232 233 /* thresholds for reduction on drain in sub-states/drain */ 234 static int32_t bbr_drain_rtt = BBR_SRTT; 235 static int32_t bbr_drain_floor = 88; 236 static int32_t google_allow_early_out = 1; 237 static int32_t google_consider_lost = 1; 238 static int32_t bbr_drain_drop_mul = 4; 239 static int32_t bbr_drain_drop_div = 5; 240 static int32_t bbr_rand_ot = 50; 241 static int32_t bbr_can_force_probertt = 0; 242 static int32_t bbr_can_adjust_probertt = 1; 243 static int32_t bbr_probertt_sets_rtt = 0; 244 static int32_t bbr_can_use_ts_for_rtt = 1; 245 static int32_t bbr_is_ratio = 0; 246 static int32_t bbr_sub_drain_app_limit = 1; 247 static int32_t bbr_prtt_slam_cwnd = 1; 248 static int32_t bbr_sub_drain_slam_cwnd = 1; 249 static int32_t bbr_slam_cwnd_in_main_drain = 1; 250 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter 251 * hold */ 252 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); 253 /* 254 * bbr_drain_gain is the reverse of the high_gain 255 * designed to drain back out the standing queue 256 * that is formed in startup by causing a larger 257 * hptsi gain and thus drainging the packets 258 * in flight. 259 */ 260 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; 261 static int32_t bbr_rttprobe_gain = 192; 262 263 /* 264 * The cwnd_gain is the default cwnd gain applied when 265 * calculating a target cwnd. Note that the cwnd is 266 * a secondary factor in the way BBR works (see the 267 * paper and think about it, it will take some time). 268 * Basically the hptsi_gain spreads the packets out 269 * so you never get more than BDP to the peer even 270 * if the cwnd is high. In our implemenation that 271 * means in non-recovery/retransmission scenarios 272 * cwnd will never be reached by the flight-size. 273 */ 274 static int32_t bbr_cwnd_gain = BBR_UNIT * 2; 275 static int32_t bbr_tlp_type_to_use = BBR_SRTT; 276 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ 277 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ 278 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ 279 static int32_t bbr_ignore_data_after_close = 1; 280 static int16_t bbr_hptsi_gain[] = { 281 (BBR_UNIT *5 / 4), 282 (BBR_UNIT * 3 / 4), 283 BBR_UNIT, 284 BBR_UNIT, 285 BBR_UNIT, 286 BBR_UNIT, 287 BBR_UNIT, 288 BBR_UNIT 289 }; 290 int32_t bbr_use_rack_resend_cheat = 1; 291 int32_t bbr_sends_full_iwnd = 1; 292 293 #define BBR_HPTSI_GAIN_MAX 8 294 /* 295 * The BBR module incorporates a number of 296 * TCP ideas that have been put out into the IETF 297 * over the last few years: 298 * - Yuchung Cheng's RACK TCP (for which its named) that 299 * will stop us using the number of dup acks and instead 300 * use time as the gage of when we retransmit. 301 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 302 * of Dukkipati et.al. 303 * - Van Jacobson's et.al BBR. 304 * 305 * RACK depends on SACK, so if an endpoint arrives that 306 * cannot do SACK the state machine below will shuttle the 307 * connection back to using the "default" TCP stack that is 308 * in FreeBSD. 309 * 310 * To implement BBR and RACK the original TCP stack was first decomposed 311 * into a functional state machine with individual states 312 * for each of the possible TCP connection states. The do_segment 313 * functions role in life is to mandate the connection supports SACK 314 * initially and then assure that the RACK state matches the conenction 315 * state before calling the states do_segment function. Data processing 316 * of inbound segments also now happens in the hpts_do_segment in general 317 * with only one exception. This is so we can keep the connection on 318 * a single CPU. 319 * 320 * Each state is simplified due to the fact that the original do_segment 321 * has been decomposed and we *know* what state we are in (no 322 * switches on the state) and all tests for SACK are gone. This 323 * greatly simplifies what each state does. 324 * 325 * TCP output is also over-written with a new version since it 326 * must maintain the new rack scoreboard and has had hptsi 327 * integrated as a requirment. Still todo is to eliminate the 328 * use of the callout_() system and use the hpts for all 329 * timers as well. 330 */ 331 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ 332 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ 333 static const int32_t bbr_min_req_free = 2; /* The min we must have on the 334 * free list */ 335 static int32_t bbr_tlp_thresh = 1; 336 static int32_t bbr_reorder_thresh = 2; 337 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def 338 * 60,000,000 - 60 seconds */ 339 static int32_t bbr_pkt_delay = 1000; 340 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ 341 static int32_t bbr_incr_timers = 1; 342 343 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ 344 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ 345 static int32_t bbr_exit_startup_at_loss = 1; 346 347 /* 348 * bbr_lt_bw_ratio is 1/8th 349 * bbr_lt_bw_diff is < 4 Kbit/sec 350 */ 351 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ 352 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ 353 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use 354 * the lt_bw for */ 355 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure 356 * lt_bw */ 357 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ 358 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ 359 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ 360 361 static int32_t bbr_verbose_logging = 0; 362 /* 363 * Currently regular tcp has a rto_min of 30ms 364 * the backoff goes 12 times so that ends up 365 * being a total of 122.850 seconds before a 366 * connection is killed. 367 */ 368 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ 369 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ 370 371 /****************************************************/ 372 /* DEFAULT TSO SIZING (cpu performance impacting) */ 373 /****************************************************/ 374 /* What amount is our formula using to get TSO size */ 375 static int32_t bbr_hptsi_per_second = 1000; 376 377 /* 378 * For hptsi under bbr_cross_over connections what is delay 379 * target 7ms (in usec) combined with a seg_max of 2 380 * gets us close to identical google behavior in 381 * TSO size selection (possibly more 1MSS sends). 382 */ 383 static int32_t bbr_hptsi_segments_delay_tar = 7000; 384 385 /* Does pacing delay include overhead's in its time calculations? */ 386 static int32_t bbr_include_enet_oh = 0; 387 static int32_t bbr_include_ip_oh = 1; 388 static int32_t bbr_include_tcp_oh = 1; 389 static int32_t bbr_google_discount = 10; 390 391 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ 392 static int32_t bbr_state_is_pkt_epoch = 0; 393 static int32_t bbr_state_drain_2_tar = 1; 394 /* What is the max the 0 - bbr_cross_over MBPS TSO target 395 * can reach using our delay target. Note that this 396 * value becomes the floor for the cross over 397 * algorithm. 398 */ 399 static int32_t bbr_hptsi_segments_max = 2; 400 static int32_t bbr_hptsi_segments_floor = 1; 401 static int32_t bbr_hptsi_utter_max = 0; 402 403 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ 404 static int32_t bbr_hptsi_bytes_min = 1460; 405 static int32_t bbr_all_get_min = 0; 406 407 /* Cross over point from algo-a to algo-b */ 408 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; 409 410 /* Do we deal with our restart state? */ 411 static int32_t bbr_uses_idle_restart = 0; 412 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ 413 414 /* Do we allow hardware pacing? */ 415 static int32_t bbr_allow_hdwr_pacing = 0; 416 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ 417 static int32_t bbr_hdwr_pace_floor = 1; 418 static int32_t bbr_hdwr_pacing_delay_cnt = 10; 419 420 /****************************************************/ 421 static int32_t bbr_resends_use_tso = 0; 422 static int32_t bbr_tlp_max_resend = 2; 423 static int32_t bbr_sack_block_limit = 128; 424 425 #define BBR_MAX_STAT 19 426 counter_u64_t bbr_state_time[BBR_MAX_STAT]; 427 counter_u64_t bbr_state_lost[BBR_MAX_STAT]; 428 counter_u64_t bbr_state_resend[BBR_MAX_STAT]; 429 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 430 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 431 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; 432 counter_u64_t bbr_flows_whdwr_pacing; 433 counter_u64_t bbr_flows_nohdwr_pacing; 434 435 counter_u64_t bbr_nohdwr_pacing_enobuf; 436 counter_u64_t bbr_hdwr_pacing_enobuf; 437 438 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); 439 440 /* 441 * Static defintions we need for forward declarations. 442 */ 443 static uint32_t 444 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, 445 uint32_t useconds_time, uint64_t bw); 446 static uint32_t 447 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); 448 static void 449 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); 450 static void 451 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); 452 static void 453 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, 454 int dolog); 455 static uint32_t 456 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); 457 static void 458 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, 459 int32_t pkt_epoch, uint32_t losses); 460 static uint32_t 461 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm); 462 static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); 463 static uint32_t 464 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 465 struct bbr_sendmap *rsm, uint32_t srtt, 466 uint32_t cts); 467 static void 468 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 469 int32_t line); 470 static void 471 bbr_set_state_target(struct tcp_bbr *bbr, int line); 472 static void 473 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); 474 475 static void 476 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line); 477 478 static void 479 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); 480 481 static void 482 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); 483 484 static void 485 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, 486 uint32_t line, uint8_t is_start, uint16_t set); 487 488 static struct bbr_sendmap * 489 bbr_find_lowest_rsm(struct tcp_bbr *bbr); 490 static __inline uint32_t 491 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); 492 static void 493 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which); 494 495 static void 496 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 497 uint32_t thresh, uint32_t to); 498 static void 499 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); 500 501 static void 502 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, 503 uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay); 504 505 static void 506 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, 507 uint32_t cts, int32_t line); 508 static void 509 bbr_stop_all_timers(struct tcpcb *tp); 510 static void 511 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); 512 static void 513 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); 514 static void 515 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); 516 517 static void 518 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 519 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod); 520 521 static int 522 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt); 523 524 static inline uint8_t 525 bbr_state_val(struct tcp_bbr *bbr) 526 { 527 return(bbr->rc_bbr_substate); 528 } 529 530 static inline uint32_t 531 get_min_cwnd(struct tcp_bbr *bbr) 532 { 533 int mss; 534 535 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 536 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) 537 return (bbr_cwnd_min_val_hs * mss); 538 else 539 return (bbr_cwnd_min_val * mss); 540 } 541 542 static uint32_t 543 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) 544 { 545 uint64_t srtt, var; 546 uint64_t ret_val; 547 548 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 549 if (tp->t_srtt == 0) { 550 srtt = (uint64_t)BBR_INITIAL_RTO; 551 var = 0; 552 } else { 553 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 554 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); 555 } 556 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), 557 bbr_persist_min, bbr_persist_max); 558 return ((uint32_t)ret_val); 559 } 560 561 static uint32_t 562 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 563 { 564 /* 565 * Start the FR timer, we do this based on getting the first one in 566 * the rc_tmap. Note that if its NULL we must stop the timer. in all 567 * events we need to stop the running timer (if its running) before 568 * starting the new one. 569 */ 570 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; 571 int32_t idx; 572 int32_t is_tlp_timer = 0; 573 struct bbr_sendmap *rsm; 574 575 if (bbr->rc_all_timers_stopped) { 576 /* All timers have been stopped none are to run */ 577 return (0); 578 } 579 if (bbr->rc_in_persist) { 580 /* We can't start any timer in persists */ 581 return (bbr_get_persists_timer_val(tp, bbr)); 582 } 583 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 584 if ((rsm == NULL) || 585 ((tp->t_flags & TF_SACK_PERMIT) == 0) || 586 (tp->t_state < TCPS_ESTABLISHED)) { 587 /* Nothing on the send map */ 588 activate_rxt: 589 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { 590 uint64_t tov; 591 592 time_since_sent = 0; 593 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 594 if (rsm) { 595 idx = rsm->r_rtr_cnt - 1; 596 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 597 tstmp_touse = rsm->r_tim_lastsent[idx]; 598 else 599 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 600 if (TSTMP_GT(tstmp_touse, cts)) 601 time_since_sent = cts - tstmp_touse; 602 } 603 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 604 if (tp->t_srtt == 0) 605 tov = BBR_INITIAL_RTO; 606 else 607 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + 608 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); 609 if (tp->t_rxtshift) 610 tov *= tcp_backoff[tp->t_rxtshift]; 611 if (tov > time_since_sent) 612 tov -= time_since_sent; 613 else 614 tov = bbr->r_ctl.rc_min_to; 615 TCPT_RANGESET_NOSLOP(to, tov, 616 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), 617 (bbr->rc_max_rto_sec * USECS_IN_SECOND)); 618 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); 619 return (to); 620 } 621 return (0); 622 } 623 if (rsm->r_flags & BBR_ACKED) { 624 rsm = bbr_find_lowest_rsm(bbr); 625 if (rsm == NULL) { 626 /* No lowest? */ 627 goto activate_rxt; 628 } 629 } 630 /* Convert from ms to usecs */ 631 if (rsm->r_flags & BBR_SACK_PASSED) { 632 if ((tp->t_flags & TF_SENTFIN) && 633 ((tp->snd_max - tp->snd_una) == 1) && 634 (rsm->r_flags & BBR_HAS_FIN)) { 635 /* 636 * We don't start a bbr rack timer if all we have is 637 * a FIN outstanding. 638 */ 639 goto activate_rxt; 640 } 641 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); 642 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); 643 idx = rsm->r_rtr_cnt - 1; 644 exp = rsm->r_tim_lastsent[idx] + thresh; 645 if (SEQ_GEQ(exp, cts)) { 646 to = exp - cts; 647 if (to < bbr->r_ctl.rc_min_to) { 648 to = bbr->r_ctl.rc_min_to; 649 } 650 } else { 651 to = bbr->r_ctl.rc_min_to; 652 } 653 } else { 654 /* Ok we need to do a TLP not RACK */ 655 if (bbr->rc_tlp_in_progress != 0) { 656 /* 657 * The previous send was a TLP. 658 */ 659 goto activate_rxt; 660 } 661 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 662 if (rsm == NULL) { 663 /* We found no rsm to TLP with. */ 664 goto activate_rxt; 665 } 666 if (rsm->r_flags & BBR_HAS_FIN) { 667 /* If its a FIN we don't do TLP */ 668 rsm = NULL; 669 goto activate_rxt; 670 } 671 time_since_sent = 0; 672 idx = rsm->r_rtr_cnt - 1; 673 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 674 tstmp_touse = rsm->r_tim_lastsent[idx]; 675 else 676 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 677 if (TSTMP_GT(tstmp_touse, cts)) 678 time_since_sent = cts - tstmp_touse; 679 is_tlp_timer = 1; 680 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); 681 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); 682 if (thresh > time_since_sent) 683 to = thresh - time_since_sent; 684 else 685 to = bbr->r_ctl.rc_min_to; 686 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 687 /* 688 * If the TLP time works out to larger than the max 689 * RTO lets not do TLP.. just RTO. 690 */ 691 goto activate_rxt; 692 } 693 if ((bbr->rc_tlp_rtx_out == 1) && 694 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { 695 /* 696 * Second retransmit of the same TLP 697 * lets not. 698 */ 699 bbr->rc_tlp_rtx_out = 0; 700 goto activate_rxt; 701 } 702 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { 703 /* 704 * The tail is no longer the last one I did a probe 705 * on 706 */ 707 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 708 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 709 } 710 } 711 if (is_tlp_timer == 0) { 712 BBR_STAT_INC(bbr_to_arm_rack); 713 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 714 } else { 715 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); 716 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 717 /* 718 * We have exceeded how many times we can retran the 719 * current TLP timer, switch to the RTO timer. 720 */ 721 goto activate_rxt; 722 } else { 723 BBR_STAT_INC(bbr_to_arm_tlp); 724 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 725 } 726 } 727 return (to); 728 } 729 730 static inline int32_t 731 bbr_minseg(struct tcp_bbr *bbr) 732 { 733 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); 734 } 735 736 static void 737 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) 738 { 739 struct inpcb *inp; 740 struct hpts_diag diag; 741 uint32_t delayed_ack = 0; 742 uint32_t left = 0; 743 uint32_t hpts_timeout; 744 uint8_t stopped; 745 int32_t delay_calc = 0; 746 uint32_t prev_delay = 0; 747 748 inp = tp->t_inpcb; 749 if (tcp_in_hpts(inp)) { 750 /* A previous call is already set up */ 751 return; 752 } 753 if ((tp->t_state == TCPS_CLOSED) || 754 (tp->t_state == TCPS_LISTEN)) { 755 return; 756 } 757 stopped = bbr->rc_tmr_stopped; 758 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 759 left = bbr->r_ctl.rc_timer_exp - cts; 760 } 761 bbr->r_ctl.rc_hpts_flags = 0; 762 bbr->r_ctl.rc_timer_exp = 0; 763 prev_delay = bbr->r_ctl.rc_last_delay_val; 764 if (bbr->r_ctl.rc_last_delay_val && 765 (slot == 0)) { 766 /* 767 * If a previous pacer delay was in place we 768 * are not coming from the output side (where 769 * we calculate a delay, more likely a timer). 770 */ 771 slot = bbr->r_ctl.rc_last_delay_val; 772 if (TSTMP_GT(cts, bbr->rc_pacer_started)) { 773 /* Compensate for time passed */ 774 delay_calc = cts - bbr->rc_pacer_started; 775 if (delay_calc <= slot) 776 slot -= delay_calc; 777 } 778 } 779 /* Do we have early to make up for by pushing out the pacing time? */ 780 if (bbr->r_agg_early_set) { 781 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); 782 slot += bbr->r_ctl.rc_agg_early; 783 bbr->r_ctl.rc_agg_early = 0; 784 bbr->r_agg_early_set = 0; 785 } 786 /* Are we running a total debt that needs to be compensated for? */ 787 if (bbr->r_ctl.rc_hptsi_agg_delay) { 788 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { 789 /* We nuke the delay */ 790 slot -= bbr->r_ctl.rc_hptsi_agg_delay; 791 bbr->r_ctl.rc_hptsi_agg_delay = 0; 792 } else { 793 /* We nuke some of the delay, put in a minimal 100usecs */ 794 bbr->r_ctl.rc_hptsi_agg_delay -= slot; 795 bbr->r_ctl.rc_last_delay_val = slot = 100; 796 } 797 } 798 bbr->r_ctl.rc_last_delay_val = slot; 799 hpts_timeout = bbr_timer_start(tp, bbr, cts); 800 if (tp->t_flags & TF_DELACK) { 801 if (bbr->rc_in_persist == 0) { 802 delayed_ack = bbr_delack_time; 803 } else { 804 /* 805 * We are in persists and have 806 * gotten a new data element. 807 */ 808 if (hpts_timeout > bbr_delack_time) { 809 /* 810 * Lets make the persists timer (which acks) 811 * be the smaller of hpts_timeout and bbr_delack_time. 812 */ 813 hpts_timeout = bbr_delack_time; 814 } 815 } 816 } 817 if (delayed_ack && 818 ((hpts_timeout == 0) || 819 (delayed_ack < hpts_timeout))) { 820 /* We need a Delayed ack timer */ 821 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 822 hpts_timeout = delayed_ack; 823 } 824 if (slot) { 825 /* Mark that we have a pacing timer up */ 826 BBR_STAT_INC(bbr_paced_segments); 827 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 828 } 829 /* 830 * If no timers are going to run and we will fall off thfe hptsi 831 * wheel, we resort to a keep-alive timer if its configured. 832 */ 833 if ((hpts_timeout == 0) && 834 (slot == 0)) { 835 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 836 (tp->t_state <= TCPS_CLOSING)) { 837 /* 838 * Ok we have no timer (persists, rack, tlp, rxt or 839 * del-ack), we don't have segments being paced. So 840 * all that is left is the keepalive timer. 841 */ 842 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 843 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); 844 } else { 845 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); 846 } 847 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 848 } 849 } 850 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 851 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 852 /* 853 * RACK, TLP, persists and RXT timers all are restartable 854 * based on actions input .. i.e we received a packet (ack 855 * or sack) and that changes things (rw, or snd_una etc). 856 * Thus we can restart them with a new value. For 857 * keep-alive, delayed_ack we keep track of what was left 858 * and restart the timer with a smaller value. 859 */ 860 if (left < hpts_timeout) 861 hpts_timeout = left; 862 } 863 if (bbr->r_ctl.rc_incr_tmrs && slot && 864 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { 865 /* 866 * If configured to do so, and the timer is either 867 * the TLP or RXT timer, we need to increase the timeout 868 * by the pacing time. Consider the bottleneck at my 869 * machine as an example, we are sending something 870 * to start a TLP on. The last packet won't be emitted 871 * fully until the pacing time (the bottleneck will hold 872 * the data in place). Once the packet is emitted that 873 * is when we want to start waiting for the TLP. This 874 * is most evident with hardware pacing (where the nic 875 * is holding the packet(s) before emitting). But it 876 * can also show up in the network so we do it for all 877 * cases. Technically we would take off one packet from 878 * this extra delay but this is easier and being more 879 * conservative is probably better. 880 */ 881 hpts_timeout += slot; 882 } 883 if (hpts_timeout) { 884 /* 885 * Hack alert for now we can't time-out over 2147 seconds (a 886 * bit more than 35min) 887 */ 888 if (hpts_timeout > 0x7ffffffe) 889 hpts_timeout = 0x7ffffffe; 890 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 891 } else 892 bbr->r_ctl.rc_timer_exp = 0; 893 if ((slot) && 894 (bbr->rc_use_google || 895 bbr->output_error_seen || 896 (slot <= hpts_timeout)) ) { 897 /* 898 * Tell LRO that it can queue packets while 899 * we pace. 900 */ 901 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 902 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 903 (bbr->rc_cwnd_limited == 0)) { 904 /* 905 * If we are not cwnd limited and we 906 * are running a rack timer we put on 907 * the do not disturbe even for sack. 908 */ 909 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 910 } else 911 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 912 bbr->rc_pacer_started = cts; 913 914 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), 915 __LINE__, &diag); 916 bbr->rc_timer_first = 0; 917 bbr->bbr_timer_src = frm; 918 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); 919 bbr_log_hpts_diag(bbr, cts, &diag); 920 } else if (hpts_timeout) { 921 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), 922 __LINE__, &diag); 923 /* 924 * We add the flag here as well if the slot is set, 925 * since hpts will call in to clear the queue first before 926 * calling the output routine (which does our timers). 927 * We don't want to set the flag if its just a timer 928 * else the arrival of data might (that causes us 929 * to send more) might get delayed. Imagine being 930 * on a keep-alive timer and a request comes in for 931 * more data. 932 */ 933 if (slot) 934 bbr->rc_pacer_started = cts; 935 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 936 (bbr->rc_cwnd_limited == 0)) { 937 /* 938 * For a rack timer, don't wake us even 939 * if a sack arrives as long as we are 940 * not cwnd limited. 941 */ 942 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; 943 inp->inp_flags2 |= INP_DONT_SACK_QUEUE; 944 } else { 945 /* All other timers wake us up */ 946 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 947 inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; 948 } 949 bbr->bbr_timer_src = frm; 950 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); 951 bbr_log_hpts_diag(bbr, cts, &diag); 952 bbr->rc_timer_first = 1; 953 } 954 bbr->rc_tmr_stopped = 0; 955 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); 956 } 957 958 static void 959 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) 960 { 961 /* 962 * We received an ack, and then did not call send or were bounced 963 * out due to the hpts was running. Now a timer is up as well, is it 964 * the right timer? 965 */ 966 struct inpcb *inp; 967 struct bbr_sendmap *rsm; 968 uint32_t hpts_timeout; 969 int tmr_up; 970 971 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 972 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 973 return; 974 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 975 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 976 (tmr_up == PACE_TMR_RXT)) { 977 /* Should be an RXT */ 978 return; 979 } 980 inp = bbr->rc_inp; 981 if (rsm == NULL) { 982 /* Nothing outstanding? */ 983 if (tp->t_flags & TF_DELACK) { 984 if (tmr_up == PACE_TMR_DELACK) 985 /* 986 * We are supposed to have delayed ack up 987 * and we do 988 */ 989 return; 990 } else if (sbavail(&inp->inp_socket->so_snd) && 991 (tmr_up == PACE_TMR_RXT)) { 992 /* 993 * if we hit enobufs then we would expect the 994 * possiblity of nothing outstanding and the RXT up 995 * (and the hptsi timer). 996 */ 997 return; 998 } else if (((V_tcp_always_keepalive || 999 inp->inp_socket->so_options & SO_KEEPALIVE) && 1000 (tp->t_state <= TCPS_CLOSING)) && 1001 (tmr_up == PACE_TMR_KEEP) && 1002 (tp->snd_max == tp->snd_una)) { 1003 /* We should have keep alive up and we do */ 1004 return; 1005 } 1006 } 1007 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { 1008 if ((tp->t_flags & TF_SENTFIN) && 1009 ((tp->snd_max - tp->snd_una) == 1) && 1010 (rsm->r_flags & BBR_HAS_FIN)) { 1011 /* needs to be a RXT */ 1012 if (tmr_up == PACE_TMR_RXT) 1013 return; 1014 else 1015 goto wrong_timer; 1016 } else if (tmr_up == PACE_TMR_RACK) 1017 return; 1018 else 1019 goto wrong_timer; 1020 } else if (rsm && (tmr_up == PACE_TMR_RACK)) { 1021 /* Rack timer has priority if we have data out */ 1022 return; 1023 } else if (SEQ_GT(tp->snd_max, tp->snd_una) && 1024 ((tmr_up == PACE_TMR_TLP) || 1025 (tmr_up == PACE_TMR_RXT))) { 1026 /* 1027 * Either a TLP or RXT is fine if no sack-passed is in place 1028 * and data is outstanding. 1029 */ 1030 return; 1031 } else if (tmr_up == PACE_TMR_DELACK) { 1032 /* 1033 * If the delayed ack was going to go off before the 1034 * rtx/tlp/rack timer were going to expire, then that would 1035 * be the timer in control. Note we don't check the time 1036 * here trusting the code is correct. 1037 */ 1038 return; 1039 } 1040 if (SEQ_GT(tp->snd_max, tp->snd_una) && 1041 ((tmr_up == PACE_TMR_RXT) || 1042 (tmr_up == PACE_TMR_TLP) || 1043 (tmr_up == PACE_TMR_RACK))) { 1044 /* 1045 * We have outstanding data and 1046 * we *do* have a RACK, TLP or RXT 1047 * timer running. We won't restart 1048 * anything here since thats probably ok we 1049 * will get called with some timer here shortly. 1050 */ 1051 return; 1052 } 1053 /* 1054 * Ok the timer originally started is not what we want now. We will 1055 * force the hpts to be stopped if any, and restart with the slot 1056 * set to what was in the saved slot. 1057 */ 1058 wrong_timer: 1059 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { 1060 if (tcp_in_hpts(inp)) 1061 tcp_hpts_remove(inp); 1062 bbr_timer_cancel(bbr, __LINE__, cts); 1063 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 1064 0); 1065 } else { 1066 /* 1067 * Output is hptsi so we just need to switch the type of 1068 * timer. We don't bother with keep-alive, since when we 1069 * jump through the output, it will start the keep-alive if 1070 * nothing is sent. 1071 * 1072 * We only need a delayed-ack added and or the hpts_timeout. 1073 */ 1074 hpts_timeout = bbr_timer_start(tp, bbr, cts); 1075 if (tp->t_flags & TF_DELACK) { 1076 if (hpts_timeout == 0) { 1077 hpts_timeout = bbr_delack_time; 1078 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1079 } 1080 else if (hpts_timeout > bbr_delack_time) { 1081 hpts_timeout = bbr_delack_time; 1082 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1083 } 1084 } 1085 if (hpts_timeout) { 1086 if (hpts_timeout > 0x7ffffffe) 1087 hpts_timeout = 0x7ffffffe; 1088 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 1089 } 1090 } 1091 } 1092 1093 int32_t bbr_clear_lost = 0; 1094 1095 /* 1096 * Considers the two time values now (cts) and earlier. 1097 * If cts is smaller than earlier, we could have 1098 * had a sequence wrap (our counter wraps every 1099 * 70 min or so) or it could be just clock skew 1100 * getting us two different time values. Clock skew 1101 * will show up within 10ms or so. So in such 1102 * a case (where cts is behind earlier time by 1103 * less than 10ms) we return 0. Otherwise we 1104 * return the true difference between them. 1105 */ 1106 static inline uint32_t 1107 bbr_calc_time(uint32_t cts, uint32_t earlier_time) { 1108 /* 1109 * Given two timestamps, the current time stamp cts, and some other 1110 * time-stamp taken in theory earlier return the difference. The 1111 * trick is here sometimes locking will get the other timestamp 1112 * after the cts. If this occurs we need to return 0. 1113 */ 1114 if (TSTMP_GEQ(cts, earlier_time)) 1115 return (cts - earlier_time); 1116 /* 1117 * cts is behind earlier_time if its less than 10ms consider it 0. 1118 * If its more than 10ms difference then we had a time wrap. Else 1119 * its just the normal locking foo. I wonder if we should not go to 1120 * 64bit TS and get rid of this issue. 1121 */ 1122 if (TSTMP_GEQ((cts + 10000), earlier_time)) 1123 return (0); 1124 /* 1125 * Ok the time must have wrapped. So we need to answer a large 1126 * amount of time, which the normal subtraction should do. 1127 */ 1128 return (cts - earlier_time); 1129 } 1130 1131 static int 1132 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) 1133 { 1134 uint32_t stat; 1135 int32_t error; 1136 1137 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); 1138 if (error || req->newptr == NULL) 1139 return error; 1140 1141 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 1142 if (error) 1143 return (error); 1144 if (stat == 1) { 1145 #ifdef BBR_INVARIANTS 1146 printf("Clearing BBR lost counters\n"); 1147 #endif 1148 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); 1149 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); 1150 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); 1151 } else if (stat == 2) { 1152 #ifdef BBR_INVARIANTS 1153 printf("Clearing BBR option counters\n"); 1154 #endif 1155 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); 1156 } else if (stat == 3) { 1157 #ifdef BBR_INVARIANTS 1158 printf("Clearing BBR stats counters\n"); 1159 #endif 1160 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); 1161 } else if (stat == 4) { 1162 #ifdef BBR_INVARIANTS 1163 printf("Clearing BBR out-size counters\n"); 1164 #endif 1165 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); 1166 } 1167 bbr_clear_lost = 0; 1168 return (0); 1169 } 1170 1171 static void 1172 bbr_init_sysctls(void) 1173 { 1174 struct sysctl_oid *bbr_probertt; 1175 struct sysctl_oid *bbr_hptsi; 1176 struct sysctl_oid *bbr_measure; 1177 struct sysctl_oid *bbr_cwnd; 1178 struct sysctl_oid *bbr_timeout; 1179 struct sysctl_oid *bbr_states; 1180 struct sysctl_oid *bbr_startup; 1181 struct sysctl_oid *bbr_policer; 1182 1183 /* Probe rtt controls */ 1184 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1185 SYSCTL_CHILDREN(bbr_sysctl_root), 1186 OID_AUTO, 1187 "probertt", 1188 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1189 ""); 1190 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1191 SYSCTL_CHILDREN(bbr_probertt), 1192 OID_AUTO, "gain", CTLFLAG_RW, 1193 &bbr_rttprobe_gain, 192, 1194 "What is the filter gain drop in probe_rtt (0=disable)?"); 1195 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1196 SYSCTL_CHILDREN(bbr_probertt), 1197 OID_AUTO, "cwnd", CTLFLAG_RW, 1198 &bbr_rtt_probe_cwndtarg, 4, 1199 "How many mss's are outstanding during probe-rtt"); 1200 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1201 SYSCTL_CHILDREN(bbr_probertt), 1202 OID_AUTO, "int", CTLFLAG_RW, 1203 &bbr_rtt_probe_limit, 4000000, 1204 "If RTT has not shrank in this many micro-seconds enter probe-rtt"); 1205 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1206 SYSCTL_CHILDREN(bbr_probertt), 1207 OID_AUTO, "mintime", CTLFLAG_RW, 1208 &bbr_rtt_probe_time, 200000, 1209 "How many microseconds in probe-rtt"); 1210 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1211 SYSCTL_CHILDREN(bbr_probertt), 1212 OID_AUTO, "filter_len_sec", CTLFLAG_RW, 1213 &bbr_filter_len_sec, 6, 1214 "How long in seconds does the rttProp filter run?"); 1215 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1216 SYSCTL_CHILDREN(bbr_probertt), 1217 OID_AUTO, "drain_rtt", CTLFLAG_RW, 1218 &bbr_drain_rtt, BBR_SRTT, 1219 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); 1220 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1221 SYSCTL_CHILDREN(bbr_probertt), 1222 OID_AUTO, "can_force", CTLFLAG_RW, 1223 &bbr_can_force_probertt, 0, 1224 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); 1225 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1226 SYSCTL_CHILDREN(bbr_probertt), 1227 OID_AUTO, "enter_sets_force", CTLFLAG_RW, 1228 &bbr_probertt_sets_rtt, 0, 1229 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); 1230 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1231 SYSCTL_CHILDREN(bbr_probertt), 1232 OID_AUTO, "can_adjust", CTLFLAG_RW, 1233 &bbr_can_adjust_probertt, 1, 1234 "Can we dynamically adjust the probe-rtt limits and times?"); 1235 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1236 SYSCTL_CHILDREN(bbr_probertt), 1237 OID_AUTO, "is_ratio", CTLFLAG_RW, 1238 &bbr_is_ratio, 0, 1239 "is the limit to filter a ratio?"); 1240 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1241 SYSCTL_CHILDREN(bbr_probertt), 1242 OID_AUTO, "use_cwnd", CTLFLAG_RW, 1243 &bbr_prtt_slam_cwnd, 0, 1244 "Should we set/recover cwnd?"); 1245 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1246 SYSCTL_CHILDREN(bbr_probertt), 1247 OID_AUTO, "can_use_ts", CTLFLAG_RW, 1248 &bbr_can_use_ts_for_rtt, 1, 1249 "Can we use the ms timestamp if available for retransmistted rtt calculations?"); 1250 1251 /* Pacing controls */ 1252 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1253 SYSCTL_CHILDREN(bbr_sysctl_root), 1254 OID_AUTO, 1255 "pacing", 1256 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1257 ""); 1258 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1259 SYSCTL_CHILDREN(bbr_hptsi), 1260 OID_AUTO, "hw_pacing", CTLFLAG_RW, 1261 &bbr_allow_hdwr_pacing, 1, 1262 "Do we allow hardware pacing?"); 1263 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1264 SYSCTL_CHILDREN(bbr_hptsi), 1265 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, 1266 &bbr_hardware_pacing_limit, 4000, 1267 "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); 1268 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1269 SYSCTL_CHILDREN(bbr_hptsi), 1270 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, 1271 &bbr_hdwr_pace_adjust, 2, 1272 "Multiplier to calculated tso size?"); 1273 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1274 SYSCTL_CHILDREN(bbr_hptsi), 1275 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, 1276 &bbr_hdwr_pace_floor, 1, 1277 "Do we invoke the hardware pacing floor?"); 1278 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1279 SYSCTL_CHILDREN(bbr_hptsi), 1280 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, 1281 &bbr_hdwr_pacing_delay_cnt, 10, 1282 "How many packets must be sent after hdwr pacing is enabled"); 1283 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1284 SYSCTL_CHILDREN(bbr_hptsi), 1285 OID_AUTO, "bw_cross", CTLFLAG_RW, 1286 &bbr_cross_over, 3000000, 1287 "What is the point where we cross over to linux like TSO size set"); 1288 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1289 SYSCTL_CHILDREN(bbr_hptsi), 1290 OID_AUTO, "seg_deltarg", CTLFLAG_RW, 1291 &bbr_hptsi_segments_delay_tar, 7000, 1292 "What is the worse case delay target for hptsi < 48Mbp connections"); 1293 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1294 SYSCTL_CHILDREN(bbr_hptsi), 1295 OID_AUTO, "enet_oh", CTLFLAG_RW, 1296 &bbr_include_enet_oh, 0, 1297 "Do we include the ethernet overhead in calculating pacing delay?"); 1298 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1299 SYSCTL_CHILDREN(bbr_hptsi), 1300 OID_AUTO, "ip_oh", CTLFLAG_RW, 1301 &bbr_include_ip_oh, 1, 1302 "Do we include the IP overhead in calculating pacing delay?"); 1303 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1304 SYSCTL_CHILDREN(bbr_hptsi), 1305 OID_AUTO, "tcp_oh", CTLFLAG_RW, 1306 &bbr_include_tcp_oh, 0, 1307 "Do we include the TCP overhead in calculating pacing delay?"); 1308 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1309 SYSCTL_CHILDREN(bbr_hptsi), 1310 OID_AUTO, "google_discount", CTLFLAG_RW, 1311 &bbr_google_discount, 10, 1312 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); 1313 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1314 SYSCTL_CHILDREN(bbr_hptsi), 1315 OID_AUTO, "all_get_min", CTLFLAG_RW, 1316 &bbr_all_get_min, 0, 1317 "If you are less than a MSS do you just get the min?"); 1318 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1319 SYSCTL_CHILDREN(bbr_hptsi), 1320 OID_AUTO, "tso_min", CTLFLAG_RW, 1321 &bbr_hptsi_bytes_min, 1460, 1322 "For 0 -> 24Mbps what is floor number of segments for TSO"); 1323 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1324 SYSCTL_CHILDREN(bbr_hptsi), 1325 OID_AUTO, "seg_tso_max", CTLFLAG_RW, 1326 &bbr_hptsi_segments_max, 6, 1327 "For 0 -> 24Mbps what is top number of segments for TSO"); 1328 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1329 SYSCTL_CHILDREN(bbr_hptsi), 1330 OID_AUTO, "seg_floor", CTLFLAG_RW, 1331 &bbr_hptsi_segments_floor, 1, 1332 "Minimum TSO size we will fall too in segments"); 1333 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1334 SYSCTL_CHILDREN(bbr_hptsi), 1335 OID_AUTO, "utter_max", CTLFLAG_RW, 1336 &bbr_hptsi_utter_max, 0, 1337 "The absolute maximum that any pacing (outside of hardware) can be"); 1338 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1339 SYSCTL_CHILDREN(bbr_hptsi), 1340 OID_AUTO, "seg_divisor", CTLFLAG_RW, 1341 &bbr_hptsi_per_second, 100, 1342 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); 1343 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1344 SYSCTL_CHILDREN(bbr_hptsi), 1345 OID_AUTO, "srtt_mul", CTLFLAG_RW, 1346 &bbr_hptsi_max_mul, 1, 1347 "The multiplier for pace len max"); 1348 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1349 SYSCTL_CHILDREN(bbr_hptsi), 1350 OID_AUTO, "srtt_div", CTLFLAG_RW, 1351 &bbr_hptsi_max_div, 2, 1352 "The divisor for pace len max"); 1353 /* Measurement controls */ 1354 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1355 SYSCTL_CHILDREN(bbr_sysctl_root), 1356 OID_AUTO, 1357 "measure", 1358 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1359 "Measurement controls"); 1360 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1361 SYSCTL_CHILDREN(bbr_measure), 1362 OID_AUTO, "min_i_bw", CTLFLAG_RW, 1363 &bbr_initial_bw_bps, 62500, 1364 "Minimum initial b/w in bytes per second"); 1365 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1366 SYSCTL_CHILDREN(bbr_measure), 1367 OID_AUTO, "no_sack_needed", CTLFLAG_RW, 1368 &bbr_sack_not_required, 0, 1369 "Do we allow bbr to run on connections not supporting SACK?"); 1370 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1371 SYSCTL_CHILDREN(bbr_measure), 1372 OID_AUTO, "use_google", CTLFLAG_RW, 1373 &bbr_use_google_algo, 0, 1374 "Use has close to google V1.0 has possible?"); 1375 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1376 SYSCTL_CHILDREN(bbr_measure), 1377 OID_AUTO, "ts_limiting", CTLFLAG_RW, 1378 &bbr_ts_limiting, 1, 1379 "Do we attempt to use the peers timestamp to limit b/w caculations?"); 1380 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1381 SYSCTL_CHILDREN(bbr_measure), 1382 OID_AUTO, "ts_can_raise", CTLFLAG_RW, 1383 &bbr_ts_can_raise, 0, 1384 "Can we raise the b/w via timestamp b/w calculation?"); 1385 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1386 SYSCTL_CHILDREN(bbr_measure), 1387 OID_AUTO, "ts_delta", CTLFLAG_RW, 1388 &bbr_min_usec_delta, 20000, 1389 "How long in usec between ts of our sends in ts validation code?"); 1390 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1391 SYSCTL_CHILDREN(bbr_measure), 1392 OID_AUTO, "ts_peer_delta", CTLFLAG_RW, 1393 &bbr_min_peer_delta, 20, 1394 "What min numerical value should be between the peer deltas?"); 1395 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1396 SYSCTL_CHILDREN(bbr_measure), 1397 OID_AUTO, "ts_delta_percent", CTLFLAG_RW, 1398 &bbr_delta_percent, 150, 1399 "What percentage (150 = 15.0) do we allow variance for?"); 1400 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1401 SYSCTL_CHILDREN(bbr_measure), 1402 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, 1403 &bbr_min_measurements_req, 1, 1404 "What is the minimum measurement count we need before we switch to our b/w estimate"); 1405 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1406 SYSCTL_CHILDREN(bbr_measure), 1407 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, 1408 &bbr_no_pacing_until, 4, 1409 "How many pkt-epoch's (0 is off) do we need before pacing is on?"); 1410 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1411 SYSCTL_CHILDREN(bbr_measure), 1412 OID_AUTO, "quanta", CTLFLAG_RW, 1413 &bbr_quanta, 2, 1414 "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); 1415 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1416 SYSCTL_CHILDREN(bbr_measure), 1417 OID_AUTO, "noretran", CTLFLAG_RW, 1418 &bbr_no_retran, 0, 1419 "Should google mode not use retransmission measurements for the b/w estimation?"); 1420 /* State controls */ 1421 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1422 SYSCTL_CHILDREN(bbr_sysctl_root), 1423 OID_AUTO, 1424 "states", 1425 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1426 "State controls"); 1427 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1428 SYSCTL_CHILDREN(bbr_states), 1429 OID_AUTO, "idle_restart", CTLFLAG_RW, 1430 &bbr_uses_idle_restart, 0, 1431 "Do we use a new special idle_restart state to ramp back up quickly?"); 1432 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1433 SYSCTL_CHILDREN(bbr_states), 1434 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, 1435 &bbr_idle_restart_threshold, 100000, 1436 "How long must we be idle before we restart??"); 1437 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1438 SYSCTL_CHILDREN(bbr_states), 1439 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, 1440 &bbr_state_is_pkt_epoch, 0, 1441 "Do we use a pkt-epoch for substate if 0 rttProp?"); 1442 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1443 SYSCTL_CHILDREN(bbr_states), 1444 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, 1445 &bbr_rtt_gain_thresh, 0, 1446 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); 1447 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1448 SYSCTL_CHILDREN(bbr_states), 1449 OID_AUTO, "drain_floor", CTLFLAG_RW, 1450 &bbr_drain_floor, 88, 1451 "What is the lowest we can drain (pg) too?"); 1452 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1453 SYSCTL_CHILDREN(bbr_states), 1454 OID_AUTO, "drain_2_target", CTLFLAG_RW, 1455 &bbr_state_drain_2_tar, 1, 1456 "Do we drain to target in drain substate?"); 1457 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1458 SYSCTL_CHILDREN(bbr_states), 1459 OID_AUTO, "gain_2_target", CTLFLAG_RW, 1460 &bbr_gain_to_target, 1, 1461 "Does probe bw gain to target??"); 1462 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1463 SYSCTL_CHILDREN(bbr_states), 1464 OID_AUTO, "gain_extra_time", CTLFLAG_RW, 1465 &bbr_gain_gets_extra_too, 1, 1466 "Does probe bw gain get the extra time too?"); 1467 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1468 SYSCTL_CHILDREN(bbr_states), 1469 OID_AUTO, "ld_div", CTLFLAG_RW, 1470 &bbr_drain_drop_div, 5, 1471 "Long drain drop divider?"); 1472 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1473 SYSCTL_CHILDREN(bbr_states), 1474 OID_AUTO, "ld_mul", CTLFLAG_RW, 1475 &bbr_drain_drop_mul, 4, 1476 "Long drain drop multiplier?"); 1477 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1478 SYSCTL_CHILDREN(bbr_states), 1479 OID_AUTO, "rand_ot_disc", CTLFLAG_RW, 1480 &bbr_rand_ot, 50, 1481 "Random discount of the ot?"); 1482 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1483 SYSCTL_CHILDREN(bbr_states), 1484 OID_AUTO, "dr_filter_life", CTLFLAG_RW, 1485 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, 1486 "How many packet-epochs does the b/w delivery rate last?"); 1487 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1488 SYSCTL_CHILDREN(bbr_states), 1489 OID_AUTO, "subdrain_applimited", CTLFLAG_RW, 1490 &bbr_sub_drain_app_limit, 0, 1491 "Does our sub-state drain invoke app limited if its long?"); 1492 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1493 SYSCTL_CHILDREN(bbr_states), 1494 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, 1495 &bbr_sub_drain_slam_cwnd, 0, 1496 "Should we set/recover cwnd for sub-state drain?"); 1497 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1498 SYSCTL_CHILDREN(bbr_states), 1499 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, 1500 &bbr_slam_cwnd_in_main_drain, 0, 1501 "Should we set/recover cwnd for main-state drain?"); 1502 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1503 SYSCTL_CHILDREN(bbr_states), 1504 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, 1505 &google_allow_early_out, 1, 1506 "Should we allow google probe-bw/drain to exit early at flight target?"); 1507 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1508 SYSCTL_CHILDREN(bbr_states), 1509 OID_AUTO, "google_exit_loss", CTLFLAG_RW, 1510 &google_consider_lost, 1, 1511 "Should we have losses exit gain of probebw in google mode??"); 1512 /* Startup controls */ 1513 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1514 SYSCTL_CHILDREN(bbr_sysctl_root), 1515 OID_AUTO, 1516 "startup", 1517 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1518 "Startup controls"); 1519 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1520 SYSCTL_CHILDREN(bbr_startup), 1521 OID_AUTO, "cheat_iwnd", CTLFLAG_RW, 1522 &bbr_sends_full_iwnd, 1, 1523 "Do we not pace but burst out initial windows has our TSO size?"); 1524 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1525 SYSCTL_CHILDREN(bbr_startup), 1526 OID_AUTO, "loss_threshold", CTLFLAG_RW, 1527 &bbr_startup_loss_thresh, 2000, 1528 "In startup what is the loss threshold in a pe that will exit us from startup?"); 1529 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1530 SYSCTL_CHILDREN(bbr_startup), 1531 OID_AUTO, "use_lowerpg", CTLFLAG_RW, 1532 &bbr_use_lower_gain_in_startup, 1, 1533 "Should we use a lower hptsi gain if we see loss in startup?"); 1534 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1535 SYSCTL_CHILDREN(bbr_startup), 1536 OID_AUTO, "gain", CTLFLAG_RW, 1537 &bbr_start_exit, 25, 1538 "What gain percent do we need to see to stay in startup??"); 1539 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1540 SYSCTL_CHILDREN(bbr_startup), 1541 OID_AUTO, "low_gain", CTLFLAG_RW, 1542 &bbr_low_start_exit, 15, 1543 "What gain percent do we need to see to stay in the lower gain startup??"); 1544 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1545 SYSCTL_CHILDREN(bbr_startup), 1546 OID_AUTO, "loss_exit", CTLFLAG_RW, 1547 &bbr_exit_startup_at_loss, 1, 1548 "Should we exit startup at loss in an epoch if we are not gaining?"); 1549 /* CWND controls */ 1550 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1551 SYSCTL_CHILDREN(bbr_sysctl_root), 1552 OID_AUTO, 1553 "cwnd", 1554 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1555 "Cwnd controls"); 1556 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1557 SYSCTL_CHILDREN(bbr_cwnd), 1558 OID_AUTO, "tar_rtt", CTLFLAG_RW, 1559 &bbr_cwndtarget_rtt_touse, 0, 1560 "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); 1561 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1562 SYSCTL_CHILDREN(bbr_cwnd), 1563 OID_AUTO, "may_shrink", CTLFLAG_RW, 1564 &bbr_cwnd_may_shrink, 0, 1565 "Can the cwnd shrink if it would grow to more than the target?"); 1566 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1567 SYSCTL_CHILDREN(bbr_cwnd), 1568 OID_AUTO, "max_target_limit", CTLFLAG_RW, 1569 &bbr_target_cwnd_mult_limit, 8, 1570 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); 1571 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1572 SYSCTL_CHILDREN(bbr_cwnd), 1573 OID_AUTO, "highspeed_min", CTLFLAG_RW, 1574 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, 1575 "What is the high-speed min cwnd (rttProp under 1ms)"); 1576 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1577 SYSCTL_CHILDREN(bbr_cwnd), 1578 OID_AUTO, "lowspeed_min", CTLFLAG_RW, 1579 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, 1580 "What is the min cwnd (rttProp > 1ms)"); 1581 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1582 SYSCTL_CHILDREN(bbr_cwnd), 1583 OID_AUTO, "initwin", CTLFLAG_RW, 1584 &bbr_def_init_win, 10, 1585 "What is the BBR initial window, if 0 use tcp version"); 1586 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1587 SYSCTL_CHILDREN(bbr_cwnd), 1588 OID_AUTO, "do_loss_red", CTLFLAG_RW, 1589 &bbr_do_red, 600, 1590 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); 1591 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1592 SYSCTL_CHILDREN(bbr_cwnd), 1593 OID_AUTO, "red_scale", CTLFLAG_RW, 1594 &bbr_red_scale, 20000, 1595 "What RTT do we scale with?"); 1596 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1597 SYSCTL_CHILDREN(bbr_cwnd), 1598 OID_AUTO, "red_growslow", CTLFLAG_RW, 1599 &bbr_red_growth_restrict, 1, 1600 "Do we restrict cwnd growth for whats in flight?"); 1601 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1602 SYSCTL_CHILDREN(bbr_cwnd), 1603 OID_AUTO, "red_div", CTLFLAG_RW, 1604 &bbr_red_div, 2, 1605 "If we reduce whats the divisor?"); 1606 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1607 SYSCTL_CHILDREN(bbr_cwnd), 1608 OID_AUTO, "red_mul", CTLFLAG_RW, 1609 &bbr_red_mul, 1, 1610 "If we reduce whats the mulitiplier?"); 1611 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1612 SYSCTL_CHILDREN(bbr_cwnd), 1613 OID_AUTO, "target_is_unit", CTLFLAG_RW, 1614 &bbr_target_is_bbunit, 0, 1615 "Is the state target the pacing_gain or BBR_UNIT?"); 1616 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1617 SYSCTL_CHILDREN(bbr_cwnd), 1618 OID_AUTO, "drop_limit", CTLFLAG_RW, 1619 &bbr_drop_limit, 0, 1620 "Number of segments limit for drop (0=use min_cwnd w/flight)?"); 1621 1622 /* Timeout controls */ 1623 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1624 SYSCTL_CHILDREN(bbr_sysctl_root), 1625 OID_AUTO, 1626 "timeout", 1627 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1628 "Time out controls"); 1629 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1630 SYSCTL_CHILDREN(bbr_timeout), 1631 OID_AUTO, "delack", CTLFLAG_RW, 1632 &bbr_delack_time, 100000, 1633 "BBR's delayed ack time"); 1634 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1635 SYSCTL_CHILDREN(bbr_timeout), 1636 OID_AUTO, "tlp_uses", CTLFLAG_RW, 1637 &bbr_tlp_type_to_use, 3, 1638 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); 1639 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1640 SYSCTL_CHILDREN(bbr_timeout), 1641 OID_AUTO, "persmin", CTLFLAG_RW, 1642 &bbr_persist_min, 250000, 1643 "What is the minimum time in microseconds between persists"); 1644 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1645 SYSCTL_CHILDREN(bbr_timeout), 1646 OID_AUTO, "persmax", CTLFLAG_RW, 1647 &bbr_persist_max, 1000000, 1648 "What is the largest delay in microseconds between persists"); 1649 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1650 SYSCTL_CHILDREN(bbr_timeout), 1651 OID_AUTO, "tlp_minto", CTLFLAG_RW, 1652 &bbr_tlp_min, 10000, 1653 "TLP Min timeout in usecs"); 1654 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1655 SYSCTL_CHILDREN(bbr_timeout), 1656 OID_AUTO, "tlp_dack_time", CTLFLAG_RW, 1657 &bbr_delayed_ack_time, 200000, 1658 "TLP delayed ack compensation value"); 1659 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1660 SYSCTL_CHILDREN(bbr_sysctl_root), 1661 OID_AUTO, "minrto", CTLFLAG_RW, 1662 &bbr_rto_min_ms, 30, 1663 "Minimum RTO in ms"); 1664 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1665 SYSCTL_CHILDREN(bbr_timeout), 1666 OID_AUTO, "maxrto", CTLFLAG_RW, 1667 &bbr_rto_max_sec, 4, 1668 "Maximum RTO in seconds -- should be at least as large as min_rto"); 1669 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1670 SYSCTL_CHILDREN(bbr_timeout), 1671 OID_AUTO, "tlp_retry", CTLFLAG_RW, 1672 &bbr_tlp_max_resend, 2, 1673 "How many times does TLP retry a single segment or multiple with no ACK"); 1674 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1675 SYSCTL_CHILDREN(bbr_timeout), 1676 OID_AUTO, "minto", CTLFLAG_RW, 1677 &bbr_min_to, 1000, 1678 "Minimum rack timeout in useconds"); 1679 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1680 SYSCTL_CHILDREN(bbr_timeout), 1681 OID_AUTO, "pktdelay", CTLFLAG_RW, 1682 &bbr_pkt_delay, 1000, 1683 "Extra RACK time (in useconds) besides reordering thresh"); 1684 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1685 SYSCTL_CHILDREN(bbr_timeout), 1686 OID_AUTO, "incr_tmrs", CTLFLAG_RW, 1687 &bbr_incr_timers, 1, 1688 "Increase the RXT/TLP timer by the pacing time used?"); 1689 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1690 SYSCTL_CHILDREN(bbr_timeout), 1691 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, 1692 &bbr_marks_rxt_sack_passed, 0, 1693 "Mark sack passed on all those not ack'd when a RXT hits?"); 1694 /* Policer controls */ 1695 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1696 SYSCTL_CHILDREN(bbr_sysctl_root), 1697 OID_AUTO, 1698 "policer", 1699 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1700 "Policer controls"); 1701 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1702 SYSCTL_CHILDREN(bbr_policer), 1703 OID_AUTO, "detect_enable", CTLFLAG_RW, 1704 &bbr_policer_detection_enabled, 1, 1705 "Is policer detection enabled??"); 1706 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1707 SYSCTL_CHILDREN(bbr_policer), 1708 OID_AUTO, "min_pes", CTLFLAG_RW, 1709 &bbr_lt_intvl_min_rtts, 4, 1710 "Minimum number of PE's?"); 1711 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1712 SYSCTL_CHILDREN(bbr_policer), 1713 OID_AUTO, "bwdiff", CTLFLAG_RW, 1714 &bbr_lt_bw_diff, (4000/8), 1715 "Minimal bw diff?"); 1716 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1717 SYSCTL_CHILDREN(bbr_policer), 1718 OID_AUTO, "bwratio", CTLFLAG_RW, 1719 &bbr_lt_bw_ratio, 8, 1720 "Minimal bw diff?"); 1721 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1722 SYSCTL_CHILDREN(bbr_policer), 1723 OID_AUTO, "from_rack_rxt", CTLFLAG_RW, 1724 &bbr_policer_call_from_rack_to, 0, 1725 "Do we call the policer detection code from a rack-timeout?"); 1726 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1727 SYSCTL_CHILDREN(bbr_policer), 1728 OID_AUTO, "false_postive", CTLFLAG_RW, 1729 &bbr_lt_intvl_fp, 0, 1730 "What packet epoch do we do false-postive detection at (0=no)?"); 1731 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1732 SYSCTL_CHILDREN(bbr_policer), 1733 OID_AUTO, "loss_thresh", CTLFLAG_RW, 1734 &bbr_lt_loss_thresh, 196, 1735 "Loss threshold 196 = 19.6%?"); 1736 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1737 SYSCTL_CHILDREN(bbr_policer), 1738 OID_AUTO, "false_postive_thresh", CTLFLAG_RW, 1739 &bbr_lt_fd_thresh, 100, 1740 "What percentage is the false detection threshold (150=15.0)?"); 1741 /* All the rest */ 1742 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1743 SYSCTL_CHILDREN(bbr_sysctl_root), 1744 OID_AUTO, "cheat_rxt", CTLFLAG_RW, 1745 &bbr_use_rack_resend_cheat, 0, 1746 "Do we burst 1ms between sends on retransmissions (like rack)?"); 1747 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1748 SYSCTL_CHILDREN(bbr_sysctl_root), 1749 OID_AUTO, "error_paceout", CTLFLAG_RW, 1750 &bbr_error_base_paceout, 10000, 1751 "When we hit an error what is the min to pace out in usec's?"); 1752 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1753 SYSCTL_CHILDREN(bbr_sysctl_root), 1754 OID_AUTO, "kill_paceout", CTLFLAG_RW, 1755 &bbr_max_net_error_cnt, 10, 1756 "When we hit this many errors in a row, kill the session?"); 1757 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1758 SYSCTL_CHILDREN(bbr_sysctl_root), 1759 OID_AUTO, "data_after_close", CTLFLAG_RW, 1760 &bbr_ignore_data_after_close, 1, 1761 "Do we hold off sending a RST until all pending data is ack'd"); 1762 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1763 SYSCTL_CHILDREN(bbr_sysctl_root), 1764 OID_AUTO, "resend_use_tso", CTLFLAG_RW, 1765 &bbr_resends_use_tso, 0, 1766 "Can resends use TSO?"); 1767 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1768 SYSCTL_CHILDREN(bbr_sysctl_root), 1769 OID_AUTO, "sblklimit", CTLFLAG_RW, 1770 &bbr_sack_block_limit, 128, 1771 "When do we start ignoring small sack blocks"); 1772 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1773 SYSCTL_CHILDREN(bbr_sysctl_root), 1774 OID_AUTO, "bb_verbose", CTLFLAG_RW, 1775 &bbr_verbose_logging, 0, 1776 "Should BBR black box logging be verbose"); 1777 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1778 SYSCTL_CHILDREN(bbr_sysctl_root), 1779 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 1780 &bbr_reorder_thresh, 2, 1781 "What factor for rack will be added when seeing reordering (shift right)"); 1782 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1783 SYSCTL_CHILDREN(bbr_sysctl_root), 1784 OID_AUTO, "reorder_fade", CTLFLAG_RW, 1785 &bbr_reorder_fade, 0, 1786 "Does reorder detection fade, if so how many ms (0 means never)"); 1787 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1788 SYSCTL_CHILDREN(bbr_sysctl_root), 1789 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 1790 &bbr_tlp_thresh, 1, 1791 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 1792 /* Stats and counters */ 1793 /* The pacing counters for hdwr/software can't be in the array */ 1794 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1795 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1796 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1797 SYSCTL_CHILDREN(bbr_sysctl_root), 1798 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, 1799 &bbr_hdwr_pacing_enobuf, 1800 "Total number of enobufs for hardware paced flows"); 1801 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1802 SYSCTL_CHILDREN(bbr_sysctl_root), 1803 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, 1804 &bbr_nohdwr_pacing_enobuf, 1805 "Total number of enobufs for non-hardware paced flows"); 1806 1807 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); 1808 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1809 SYSCTL_CHILDREN(bbr_sysctl_root), 1810 OID_AUTO, "hdwr_pacing", CTLFLAG_RD, 1811 &bbr_flows_whdwr_pacing, 1812 "Total number of hardware paced flows"); 1813 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); 1814 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1815 SYSCTL_CHILDREN(bbr_sysctl_root), 1816 OID_AUTO, "software_pacing", CTLFLAG_RD, 1817 &bbr_flows_nohdwr_pacing, 1818 "Total number of software paced flows"); 1819 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); 1820 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1821 OID_AUTO, "stats", CTLFLAG_RD, 1822 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); 1823 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); 1824 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1825 OID_AUTO, "opts", CTLFLAG_RD, 1826 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); 1827 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); 1828 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1829 OID_AUTO, "lost", CTLFLAG_RD, 1830 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); 1831 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); 1832 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1833 OID_AUTO, "stateresend", CTLFLAG_RD, 1834 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); 1835 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); 1836 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1837 OID_AUTO, "statetime", CTLFLAG_RD, 1838 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); 1839 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 1840 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1841 OID_AUTO, "outsize", CTLFLAG_RD, 1842 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); 1843 SYSCTL_ADD_PROC(&bbr_sysctl_ctx, 1844 SYSCTL_CHILDREN(bbr_sysctl_root), 1845 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1846 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); 1847 } 1848 1849 static void 1850 bbr_counter_destroy(void) 1851 { 1852 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); 1853 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); 1854 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); 1855 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); 1856 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); 1857 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); 1858 counter_u64_free(bbr_nohdwr_pacing_enobuf); 1859 counter_u64_free(bbr_hdwr_pacing_enobuf); 1860 counter_u64_free(bbr_flows_whdwr_pacing); 1861 counter_u64_free(bbr_flows_nohdwr_pacing); 1862 1863 } 1864 1865 static __inline void 1866 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) 1867 { 1868 memset(l, 0, sizeof(union tcp_log_stackspecific)); 1869 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; 1870 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); 1871 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 1872 l->bw_inuse = bbr_get_bw(bbr); 1873 l->inflight = ctf_flight_size(bbr->rc_tp, 1874 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 1875 l->applimited = bbr->r_ctl.r_app_limited_until; 1876 l->delivered = bbr->r_ctl.rc_delivered; 1877 l->timeStamp = cts; 1878 l->lost = bbr->r_ctl.rc_lost; 1879 l->bbr_state = bbr->rc_bbr_state; 1880 l->bbr_substate = bbr_state_val(bbr); 1881 l->epoch = bbr->r_ctl.rc_rtt_epoch; 1882 l->lt_epoch = bbr->r_ctl.rc_lt_epoch; 1883 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 1884 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; 1885 l->inhpts = tcp_in_hpts(bbr->rc_inp); 1886 l->use_lt_bw = bbr->rc_lt_use_bw; 1887 l->pkts_out = bbr->r_ctl.rc_flight_at_input; 1888 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; 1889 } 1890 1891 static void 1892 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) 1893 { 1894 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1895 union tcp_log_stackspecific log; 1896 1897 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1898 log.u_bbr.flex1 = 0; 1899 log.u_bbr.flex2 = 0; 1900 log.u_bbr.flex5 = 0; 1901 log.u_bbr.flex3 = 0; 1902 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; 1903 log.u_bbr.flex7 = reason; 1904 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; 1905 log.u_bbr.flex8 = 0; 1906 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1907 &bbr->rc_inp->inp_socket->so_rcv, 1908 &bbr->rc_inp->inp_socket->so_snd, 1909 BBR_LOG_BW_RED_EV, 0, 1910 0, &log, false, &bbr->rc_tv); 1911 } 1912 } 1913 1914 static void 1915 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) 1916 { 1917 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1918 union tcp_log_stackspecific log; 1919 1920 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1921 log.u_bbr.flex1 = seq; 1922 log.u_bbr.flex2 = count; 1923 log.u_bbr.flex8 = mode; 1924 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1925 &bbr->rc_inp->inp_socket->so_rcv, 1926 &bbr->rc_inp->inp_socket->so_snd, 1927 BBR_LOG_LOWGAIN, 0, 1928 0, &log, false, &bbr->rc_tv); 1929 } 1930 } 1931 1932 static void 1933 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, 1934 uint8_t reason, uint32_t p_maxseg, int len) 1935 { 1936 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1937 union tcp_log_stackspecific log; 1938 1939 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1940 log.u_bbr.flex1 = p_maxseg; 1941 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; 1942 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 1943 log.u_bbr.flex4 = reason; 1944 log.u_bbr.flex5 = bbr->rc_in_persist; 1945 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; 1946 log.u_bbr.flex7 = p_maxseg; 1947 log.u_bbr.flex8 = bbr->rc_in_persist; 1948 log.u_bbr.pkts_out = 0; 1949 log.u_bbr.applimited = len; 1950 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1951 &bbr->rc_inp->inp_socket->so_rcv, 1952 &bbr->rc_inp->inp_socket->so_snd, 1953 BBR_LOG_JUSTRET, 0, 1954 tlen, &log, false, &bbr->rc_tv); 1955 } 1956 } 1957 1958 static void 1959 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) 1960 { 1961 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1962 union tcp_log_stackspecific log; 1963 1964 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1965 log.u_bbr.flex1 = seq; 1966 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 1967 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; 1968 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1969 &bbr->rc_inp->inp_socket->so_rcv, 1970 &bbr->rc_inp->inp_socket->so_snd, 1971 BBR_LOG_ENTREC, 0, 1972 0, &log, false, &bbr->rc_tv); 1973 } 1974 } 1975 1976 static void 1977 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) 1978 { 1979 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 1980 union tcp_log_stackspecific log; 1981 1982 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1983 log.u_bbr.flex1 = tso; 1984 log.u_bbr.flex2 = maxseg; 1985 log.u_bbr.flex3 = mtu; 1986 log.u_bbr.flex4 = csum_flags; 1987 TCP_LOG_EVENTP(tp, NULL, 1988 &bbr->rc_inp->inp_socket->so_rcv, 1989 &bbr->rc_inp->inp_socket->so_snd, 1990 BBR_LOG_MSGSIZE, 0, 1991 0, &log, false, &bbr->rc_tv); 1992 } 1993 } 1994 1995 static void 1996 bbr_log_flowend(struct tcp_bbr *bbr) 1997 { 1998 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 1999 union tcp_log_stackspecific log; 2000 struct sockbuf *r, *s; 2001 struct timeval tv; 2002 2003 if (bbr->rc_inp->inp_socket) { 2004 r = &bbr->rc_inp->inp_socket->so_rcv; 2005 s = &bbr->rc_inp->inp_socket->so_snd; 2006 } else { 2007 r = s = NULL; 2008 } 2009 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); 2010 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2011 r, s, 2012 TCP_LOG_FLOWEND, 0, 2013 0, &log, false, &tv); 2014 } 2015 } 2016 2017 static void 2018 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, 2019 uint32_t lost, uint32_t del) 2020 { 2021 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2022 union tcp_log_stackspecific log; 2023 2024 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2025 log.u_bbr.flex1 = lost; 2026 log.u_bbr.flex2 = del; 2027 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; 2028 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; 2029 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2030 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2031 log.u_bbr.flex7 = line; 2032 log.u_bbr.flex8 = 0; 2033 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; 2034 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2035 &bbr->rc_inp->inp_socket->so_rcv, 2036 &bbr->rc_inp->inp_socket->so_snd, 2037 BBR_LOG_PKT_EPOCH, 0, 2038 0, &log, false, &bbr->rc_tv); 2039 } 2040 } 2041 2042 static void 2043 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) 2044 { 2045 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2046 union tcp_log_stackspecific log; 2047 2048 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2049 log.u_bbr.flex1 = bbr->r_ctl.rc_lost; 2050 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; 2051 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; 2052 log.u_bbr.flex7 = line; 2053 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2054 &bbr->rc_inp->inp_socket->so_rcv, 2055 &bbr->rc_inp->inp_socket->so_snd, 2056 BBR_LOG_TIME_EPOCH, 0, 2057 0, &log, false, &bbr->rc_tv); 2058 } 2059 } 2060 2061 static void 2062 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) 2063 { 2064 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2065 union tcp_log_stackspecific log; 2066 2067 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2068 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2069 log.u_bbr.flex2 = new_tar; 2070 log.u_bbr.flex3 = line; 2071 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2072 log.u_bbr.flex5 = bbr_quanta; 2073 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; 2074 log.u_bbr.flex7 = bbr->rc_last_options; 2075 log.u_bbr.flex8 = meth; 2076 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2077 &bbr->rc_inp->inp_socket->so_rcv, 2078 &bbr->rc_inp->inp_socket->so_snd, 2079 BBR_LOG_STATE_TARGET, 0, 2080 0, &log, false, &bbr->rc_tv); 2081 } 2082 2083 } 2084 2085 static void 2086 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2087 { 2088 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2089 union tcp_log_stackspecific log; 2090 2091 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2092 log.u_bbr.flex1 = line; 2093 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2094 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2095 if (bbr_state_is_pkt_epoch) 2096 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 2097 else 2098 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); 2099 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2100 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2101 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); 2102 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; 2103 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; 2104 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2105 &bbr->rc_inp->inp_socket->so_rcv, 2106 &bbr->rc_inp->inp_socket->so_snd, 2107 BBR_LOG_STATE, 0, 2108 0, &log, false, &bbr->rc_tv); 2109 } 2110 } 2111 2112 static void 2113 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 2114 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) 2115 { 2116 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2117 union tcp_log_stackspecific log; 2118 2119 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2120 log.u_bbr.flex1 = line; 2121 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2122 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; 2123 log.u_bbr.flex4 = applied; 2124 log.u_bbr.flex5 = rtt; 2125 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2126 log.u_bbr.flex7 = cond; 2127 log.u_bbr.flex8 = reas; 2128 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2129 &bbr->rc_inp->inp_socket->so_rcv, 2130 &bbr->rc_inp->inp_socket->so_snd, 2131 BBR_LOG_RTT_SHRINKS, 0, 2132 0, &log, false, &bbr->rc_tv); 2133 } 2134 } 2135 2136 static void 2137 bbr_log_type_exit_rec(struct tcp_bbr *bbr) 2138 { 2139 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2140 union tcp_log_stackspecific log; 2141 2142 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2143 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; 2144 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 2145 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2146 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2147 &bbr->rc_inp->inp_socket->so_rcv, 2148 &bbr->rc_inp->inp_socket->so_snd, 2149 BBR_LOG_EXITREC, 0, 2150 0, &log, false, &bbr->rc_tv); 2151 } 2152 } 2153 2154 static void 2155 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, 2156 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) 2157 { 2158 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2159 union tcp_log_stackspecific log; 2160 2161 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2162 log.u_bbr.flex1 = line; 2163 log.u_bbr.flex2 = prev_acked; 2164 log.u_bbr.flex3 = bytes_this_ack; 2165 log.u_bbr.flex4 = chg; 2166 log.u_bbr.flex5 = th_ack; 2167 log.u_bbr.flex6 = target; 2168 log.u_bbr.flex8 = meth; 2169 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2170 &bbr->rc_inp->inp_socket->so_rcv, 2171 &bbr->rc_inp->inp_socket->so_snd, 2172 BBR_LOG_CWND, 0, 2173 0, &log, false, &bbr->rc_tv); 2174 } 2175 } 2176 2177 static void 2178 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) 2179 { 2180 /* 2181 * Log the rtt sample we are applying to the srtt algorithm in 2182 * useconds. 2183 */ 2184 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2185 union tcp_log_stackspecific log; 2186 2187 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2188 log.u_bbr.flex1 = rtt; 2189 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; 2190 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; 2191 log.u_bbr.flex4 = bbr->rc_tp->ts_offset; 2192 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2193 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); 2194 log.u_bbr.flex6 = tsin; 2195 log.u_bbr.flex7 = 0; 2196 log.u_bbr.flex8 = bbr->rc_ack_was_delayed; 2197 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2198 &bbr->rc_inp->inp_socket->so_rcv, 2199 &bbr->rc_inp->inp_socket->so_snd, 2200 TCP_LOG_RTT, 0, 2201 0, &log, false, &bbr->rc_tv); 2202 } 2203 } 2204 2205 static void 2206 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) 2207 { 2208 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2209 union tcp_log_stackspecific log; 2210 2211 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2212 log.u_bbr.flex1 = time_in; 2213 log.u_bbr.flex2 = line; 2214 log.u_bbr.flex8 = enter_exit; 2215 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2216 &bbr->rc_inp->inp_socket->so_rcv, 2217 &bbr->rc_inp->inp_socket->so_snd, 2218 BBR_LOG_PERSIST, 0, 2219 0, &log, false, &bbr->rc_tv); 2220 } 2221 } 2222 static void 2223 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) 2224 { 2225 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2226 union tcp_log_stackspecific log; 2227 2228 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2229 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; 2230 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2231 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2232 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; 2233 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2234 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2235 &bbr->rc_inp->inp_socket->so_rcv, 2236 &bbr->rc_inp->inp_socket->so_snd, 2237 BBR_LOG_ACKCLEAR, 0, 2238 0, &log, false, &bbr->rc_tv); 2239 } 2240 } 2241 2242 static void 2243 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, 2244 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) 2245 { 2246 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2247 union tcp_log_stackspecific log; 2248 struct timeval tv; 2249 2250 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2251 log.u_bbr.flex1 = nsegs; 2252 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; 2253 if (m) { 2254 struct timespec ts; 2255 2256 log.u_bbr.flex3 = m->m_flags; 2257 if (m->m_flags & M_TSTMP) { 2258 mbuf_tstmp2timespec(m, &ts); 2259 tv.tv_sec = ts.tv_sec; 2260 tv.tv_usec = ts.tv_nsec / 1000; 2261 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); 2262 } else { 2263 log.u_bbr.lt_epoch = 0; 2264 } 2265 if (m->m_flags & M_TSTMP_LRO) { 2266 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; 2267 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000; 2268 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); 2269 } else { 2270 /* No arrival timestamp */ 2271 log.u_bbr.flex5 = 0; 2272 } 2273 2274 log.u_bbr.pkts_out = tcp_get_usecs(&tv); 2275 } else { 2276 log.u_bbr.flex3 = 0; 2277 log.u_bbr.flex5 = 0; 2278 log.u_bbr.flex6 = 0; 2279 log.u_bbr.pkts_out = 0; 2280 } 2281 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; 2282 log.u_bbr.flex7 = bbr->r_wanted_output; 2283 log.u_bbr.flex8 = bbr->rc_in_persist; 2284 TCP_LOG_EVENTP(bbr->rc_tp, th, 2285 &bbr->rc_inp->inp_socket->so_rcv, 2286 &bbr->rc_inp->inp_socket->so_snd, 2287 TCP_LOG_IN, 0, 2288 tlen, &log, true, &bbr->rc_tv); 2289 } 2290 } 2291 2292 static void 2293 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) 2294 { 2295 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2296 union tcp_log_stackspecific log; 2297 2298 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2299 log.u_bbr.flex1 = did_out; 2300 log.u_bbr.flex2 = nxt_pkt; 2301 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; 2302 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2303 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; 2304 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; 2305 log.u_bbr.flex7 = bbr->r_wanted_output; 2306 log.u_bbr.flex8 = bbr->rc_in_persist; 2307 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; 2308 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2309 &bbr->rc_inp->inp_socket->so_rcv, 2310 &bbr->rc_inp->inp_socket->so_snd, 2311 BBR_LOG_DOSEG_DONE, 0, 2312 0, &log, true, &bbr->rc_tv); 2313 } 2314 } 2315 2316 static void 2317 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, 2318 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) 2319 { 2320 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2321 union tcp_log_stackspecific log; 2322 2323 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2324 log.u_bbr.flex1 = line; 2325 log.u_bbr.flex2 = o_len; 2326 log.u_bbr.flex3 = segcnt; 2327 log.u_bbr.flex4 = segsiz; 2328 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2329 &bbr->rc_inp->inp_socket->so_rcv, 2330 &bbr->rc_inp->inp_socket->so_snd, 2331 BBR_LOG_ENOBUF_JMP, ENOBUFS, 2332 len, &log, true, &bbr->rc_tv); 2333 } 2334 } 2335 2336 static void 2337 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) 2338 { 2339 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2340 union tcp_log_stackspecific log; 2341 2342 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2343 log.u_bbr.flex1 = timers; 2344 log.u_bbr.flex2 = ret; 2345 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 2346 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2347 log.u_bbr.flex5 = cts; 2348 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2349 log.u_bbr.flex8 = hpts_calling; 2350 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2351 &bbr->rc_inp->inp_socket->so_rcv, 2352 &bbr->rc_inp->inp_socket->so_snd, 2353 BBR_LOG_TO_PROCESS, 0, 2354 0, &log, false, &bbr->rc_tv); 2355 } 2356 } 2357 2358 static void 2359 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) 2360 { 2361 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2362 union tcp_log_stackspecific log; 2363 uint64_t ar; 2364 2365 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2366 log.u_bbr.flex1 = bbr->bbr_timer_src; 2367 log.u_bbr.flex2 = 0; 2368 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2369 ar = (uint64_t)(bbr->r_ctl.rc_resend); 2370 ar >>= 32; 2371 ar &= 0x00000000ffffffff; 2372 log.u_bbr.flex4 = (uint32_t)ar; 2373 ar = (uint64_t)bbr->r_ctl.rc_resend; 2374 ar &= 0x00000000ffffffff; 2375 log.u_bbr.flex5 = (uint32_t)ar; 2376 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2377 log.u_bbr.flex8 = to_num; 2378 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2379 &bbr->rc_inp->inp_socket->so_rcv, 2380 &bbr->rc_inp->inp_socket->so_snd, 2381 BBR_LOG_RTO, 0, 2382 0, &log, false, &bbr->rc_tv); 2383 } 2384 } 2385 2386 static void 2387 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) 2388 { 2389 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2390 union tcp_log_stackspecific log; 2391 2392 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2393 log.u_bbr.flex1 = flex1; 2394 log.u_bbr.flex2 = flex2; 2395 log.u_bbr.flex3 = flex3; 2396 log.u_bbr.flex4 = 0; 2397 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2398 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2399 log.u_bbr.flex8 = reason; 2400 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; 2401 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2402 &bbr->rc_inp->inp_socket->so_rcv, 2403 &bbr->rc_inp->inp_socket->so_snd, 2404 BBR_LOG_REDUCE, 0, 2405 0, &log, false, &bbr->rc_tv); 2406 } 2407 } 2408 2409 static void 2410 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) 2411 { 2412 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2413 union tcp_log_stackspecific log; 2414 2415 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2416 log.u_bbr.flex1 = diag->p_nxt_slot; 2417 log.u_bbr.flex2 = diag->p_cur_slot; 2418 log.u_bbr.flex3 = diag->slot_req; 2419 log.u_bbr.flex4 = diag->inp_hptsslot; 2420 log.u_bbr.flex5 = diag->slot_remaining; 2421 log.u_bbr.flex6 = diag->need_new_to; 2422 log.u_bbr.flex7 = diag->p_hpts_active; 2423 log.u_bbr.flex8 = diag->p_on_min_sleep; 2424 /* Hijack other fields as needed */ 2425 log.u_bbr.epoch = diag->have_slept; 2426 log.u_bbr.lt_epoch = diag->yet_to_sleep; 2427 log.u_bbr.pkts_out = diag->co_ret; 2428 log.u_bbr.applimited = diag->hpts_sleep_time; 2429 log.u_bbr.delivered = diag->p_prev_slot; 2430 log.u_bbr.inflight = diag->p_runningslot; 2431 log.u_bbr.bw_inuse = diag->wheel_slot; 2432 log.u_bbr.rttProp = diag->wheel_cts; 2433 log.u_bbr.delRate = diag->maxslots; 2434 log.u_bbr.cur_del_rate = diag->p_curtick; 2435 log.u_bbr.cur_del_rate <<= 32; 2436 log.u_bbr.cur_del_rate |= diag->p_lasttick; 2437 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2438 &bbr->rc_inp->inp_socket->so_rcv, 2439 &bbr->rc_inp->inp_socket->so_snd, 2440 BBR_LOG_HPTSDIAG, 0, 2441 0, &log, false, &bbr->rc_tv); 2442 } 2443 } 2444 2445 static void 2446 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 2447 uint32_t thresh, uint32_t to) 2448 { 2449 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2450 union tcp_log_stackspecific log; 2451 2452 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2453 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; 2454 log.u_bbr.flex2 = time_since_sent; 2455 log.u_bbr.flex3 = srtt; 2456 log.u_bbr.flex4 = thresh; 2457 log.u_bbr.flex5 = to; 2458 log.u_bbr.flex6 = bbr->rc_tp->t_srtt; 2459 log.u_bbr.flex8 = mode; 2460 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2461 &bbr->rc_inp->inp_socket->so_rcv, 2462 &bbr->rc_inp->inp_socket->so_snd, 2463 BBR_LOG_TIMERPREP, 0, 2464 0, &log, false, &bbr->rc_tv); 2465 } 2466 } 2467 2468 static void 2469 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 2470 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) 2471 { 2472 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2473 union tcp_log_stackspecific log; 2474 2475 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2476 log.u_bbr.flex1 = usecs; 2477 log.u_bbr.flex2 = len; 2478 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); 2479 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); 2480 if (override) 2481 log.u_bbr.flex5 = (1 << 2); 2482 else 2483 log.u_bbr.flex5 = 0; 2484 log.u_bbr.flex6 = override; 2485 log.u_bbr.flex7 = gain; 2486 log.u_bbr.flex8 = mod; 2487 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2488 &bbr->rc_inp->inp_socket->so_rcv, 2489 &bbr->rc_inp->inp_socket->so_snd, 2490 BBR_LOG_HPTSI_CALC, 0, 2491 len, &log, false, &bbr->rc_tv); 2492 } 2493 } 2494 2495 static void 2496 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 2497 { 2498 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2499 union tcp_log_stackspecific log; 2500 2501 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2502 2503 log.u_bbr.flex1 = bbr->bbr_timer_src; 2504 log.u_bbr.flex2 = to; 2505 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2506 log.u_bbr.flex4 = slot; 2507 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; 2508 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2509 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; 2510 log.u_bbr.flex8 = which; 2511 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2512 &bbr->rc_inp->inp_socket->so_rcv, 2513 &bbr->rc_inp->inp_socket->so_snd, 2514 BBR_LOG_TIMERSTAR, 0, 2515 0, &log, false, &bbr->rc_tv); 2516 } 2517 } 2518 2519 static void 2520 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) 2521 { 2522 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2523 union tcp_log_stackspecific log; 2524 2525 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2526 log.u_bbr.flex1 = thresh; 2527 log.u_bbr.flex2 = lro; 2528 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; 2529 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 2530 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2531 log.u_bbr.flex6 = srtt; 2532 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; 2533 log.u_bbr.flex8 = frm; 2534 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2535 &bbr->rc_inp->inp_socket->so_rcv, 2536 &bbr->rc_inp->inp_socket->so_snd, 2537 BBR_LOG_THRESH_CALC, 0, 2538 0, &log, false, &bbr->rc_tv); 2539 } 2540 } 2541 2542 static void 2543 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) 2544 { 2545 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2546 union tcp_log_stackspecific log; 2547 2548 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2549 log.u_bbr.flex1 = line; 2550 log.u_bbr.flex2 = bbr->bbr_timer_src; 2551 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2552 log.u_bbr.flex4 = bbr->rc_in_persist; 2553 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2554 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2555 log.u_bbr.flex8 = hpts_removed; 2556 log.u_bbr.pkts_out = bbr->rc_pacer_started; 2557 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2558 &bbr->rc_inp->inp_socket->so_rcv, 2559 &bbr->rc_inp->inp_socket->so_snd, 2560 BBR_LOG_TIMERCANC, 0, 2561 0, &log, false, &bbr->rc_tv); 2562 } 2563 } 2564 2565 static void 2566 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) 2567 { 2568 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2569 union tcp_log_stackspecific log; 2570 2571 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2572 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; 2573 log.u_bbr.flex2 = (peer_delta >> 32); 2574 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); 2575 log.u_bbr.flex4 = (delta >> 32); 2576 log.u_bbr.flex5 = (delta & 0x00000000ffffffff); 2577 log.u_bbr.flex7 = bbr->rc_ts_clock_set; 2578 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; 2579 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2580 &bbr->rc_inp->inp_socket->so_rcv, 2581 &bbr->rc_inp->inp_socket->so_snd, 2582 BBR_LOG_TSTMP_VAL, 0, 2583 0, &log, false, &bbr->rc_tv); 2584 } 2585 } 2586 2587 static void 2588 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) 2589 { 2590 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2591 union tcp_log_stackspecific log; 2592 2593 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2594 log.u_bbr.flex1 = tsosz; 2595 log.u_bbr.flex2 = tls; 2596 log.u_bbr.flex3 = tcp_min_hptsi_time; 2597 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; 2598 log.u_bbr.flex5 = old_val; 2599 log.u_bbr.flex6 = maxseg; 2600 log.u_bbr.flex7 = bbr->rc_no_pacing; 2601 log.u_bbr.flex7 <<= 1; 2602 log.u_bbr.flex7 |= bbr->rc_past_init_win; 2603 if (hdwr) 2604 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; 2605 else 2606 log.u_bbr.flex8 = bbr->rc_use_google; 2607 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2608 &bbr->rc_inp->inp_socket->so_rcv, 2609 &bbr->rc_inp->inp_socket->so_snd, 2610 BBR_LOG_BBRTSO, 0, 2611 0, &log, false, &bbr->rc_tv); 2612 } 2613 } 2614 2615 static void 2616 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, 2617 uint32_t flags, uint32_t line) 2618 { 2619 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2620 union tcp_log_stackspecific log; 2621 2622 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2623 log.u_bbr.flex1 = line; 2624 log.u_bbr.flex2 = rsm->r_start; 2625 log.u_bbr.flex3 = rsm->r_end; 2626 log.u_bbr.flex4 = rsm->r_delivered; 2627 log.u_bbr.flex5 = rsm->r_rtr_cnt; 2628 log.u_bbr.flex6 = rsm->r_dupack; 2629 log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; 2630 log.u_bbr.flex8 = rsm->r_flags; 2631 /* Hijack the pkts_out fids */ 2632 log.u_bbr.applimited = flags; 2633 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2634 &bbr->rc_inp->inp_socket->so_rcv, 2635 &bbr->rc_inp->inp_socket->so_snd, 2636 BBR_RSM_CLEARED, 0, 2637 0, &log, false, &bbr->rc_tv); 2638 } 2639 } 2640 2641 static void 2642 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, 2643 uint32_t flex3, uint32_t flex2, uint32_t flex5, 2644 uint32_t flex6, uint32_t pkts_out, int flex7, 2645 uint32_t flex4, uint32_t flex1) 2646 { 2647 2648 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2649 union tcp_log_stackspecific log; 2650 2651 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2652 log.u_bbr.flex1 = flex1; 2653 log.u_bbr.flex2 = flex2; 2654 log.u_bbr.flex3 = flex3; 2655 log.u_bbr.flex4 = flex4; 2656 log.u_bbr.flex5 = flex5; 2657 log.u_bbr.flex6 = flex6; 2658 log.u_bbr.flex7 = flex7; 2659 /* Hijack the pkts_out fids */ 2660 log.u_bbr.pkts_out = pkts_out; 2661 log.u_bbr.flex8 = flex8; 2662 if (bbr->rc_ack_was_delayed) 2663 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; 2664 else 2665 log.u_bbr.epoch = 0; 2666 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2667 &bbr->rc_inp->inp_socket->so_rcv, 2668 &bbr->rc_inp->inp_socket->so_snd, 2669 BBR_LOG_BBRUPD, 0, 2670 flex2, &log, false, &bbr->rc_tv); 2671 } 2672 } 2673 2674 static void 2675 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, 2676 uint32_t newbw, uint32_t obw, uint32_t diff, 2677 uint32_t tim) 2678 { 2679 if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2680 union tcp_log_stackspecific log; 2681 2682 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2683 log.u_bbr.flex1 = reason; 2684 log.u_bbr.flex2 = newbw; 2685 log.u_bbr.flex3 = obw; 2686 log.u_bbr.flex4 = diff; 2687 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; 2688 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; 2689 log.u_bbr.flex7 = bbr->rc_lt_is_sampling; 2690 log.u_bbr.pkts_out = tim; 2691 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; 2692 if (bbr->rc_lt_use_bw == 0) 2693 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 2694 else 2695 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 2696 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2697 &bbr->rc_inp->inp_socket->so_rcv, 2698 &bbr->rc_inp->inp_socket->so_snd, 2699 BBR_LOG_BWSAMP, 0, 2700 0, &log, false, &bbr->rc_tv); 2701 } 2702 } 2703 2704 static inline void 2705 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) 2706 { 2707 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2708 union tcp_log_stackspecific log; 2709 2710 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2711 log.u_bbr.flex1 = line; 2712 log.u_bbr.flex2 = tick; 2713 log.u_bbr.flex3 = tp->t_maxunacktime; 2714 log.u_bbr.flex4 = tp->t_acktime; 2715 log.u_bbr.flex8 = event; 2716 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2717 &bbr->rc_inp->inp_socket->so_rcv, 2718 &bbr->rc_inp->inp_socket->so_snd, 2719 BBR_LOG_PROGRESS, 0, 2720 0, &log, false, &bbr->rc_tv); 2721 } 2722 } 2723 2724 static void 2725 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, 2726 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, 2727 int error) 2728 { 2729 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2730 union tcp_log_stackspecific log; 2731 2732 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2733 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); 2734 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); 2735 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); 2736 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); 2737 log.u_bbr.bw_inuse = rate; 2738 log.u_bbr.flex5 = line; 2739 log.u_bbr.flex6 = error; 2740 log.u_bbr.flex8 = bbr->skip_gain; 2741 log.u_bbr.flex8 <<= 1; 2742 log.u_bbr.flex8 |= bbr->gain_is_limited; 2743 log.u_bbr.flex8 <<= 1; 2744 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; 2745 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 2746 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2747 &bbr->rc_inp->inp_socket->so_rcv, 2748 &bbr->rc_inp->inp_socket->so_snd, 2749 BBR_LOG_HDWR_PACE, 0, 2750 0, &log, false, &bbr->rc_tv); 2751 } 2752 } 2753 2754 static void 2755 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) 2756 { 2757 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2758 union tcp_log_stackspecific log; 2759 2760 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2761 log.u_bbr.flex1 = slot; 2762 log.u_bbr.flex2 = del_by; 2763 log.u_bbr.flex3 = prev_delay; 2764 log.u_bbr.flex4 = line; 2765 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; 2766 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; 2767 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); 2768 log.u_bbr.flex8 = bbr->rc_in_persist; 2769 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2770 &bbr->rc_inp->inp_socket->so_rcv, 2771 &bbr->rc_inp->inp_socket->so_snd, 2772 BBR_LOG_BBRSND, 0, 2773 len, &log, false, &bbr->rc_tv); 2774 } 2775 } 2776 2777 static void 2778 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) 2779 { 2780 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2781 union tcp_log_stackspecific log; 2782 2783 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2784 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; 2785 log.u_bbr.flex2 = 0; 2786 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; 2787 log.u_bbr.flex4 = end; 2788 log.u_bbr.flex5 = seq; 2789 log.u_bbr.flex6 = t; 2790 log.u_bbr.flex7 = match; 2791 log.u_bbr.flex8 = flags; 2792 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2793 &bbr->rc_inp->inp_socket->so_rcv, 2794 &bbr->rc_inp->inp_socket->so_snd, 2795 BBR_LOG_BBRRTT, 0, 2796 0, &log, false, &bbr->rc_tv); 2797 } 2798 } 2799 2800 static void 2801 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) 2802 { 2803 if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { 2804 union tcp_log_stackspecific log; 2805 2806 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2807 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2808 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 2809 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; 2810 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2811 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; 2812 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; 2813 log.u_bbr.flex7 = 0; 2814 log.u_bbr.flex8 = entry_method; 2815 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2816 &bbr->rc_inp->inp_socket->so_rcv, 2817 &bbr->rc_inp->inp_socket->so_snd, 2818 BBR_LOG_EXIT_GAIN, 0, 2819 0, &log, false, &bbr->rc_tv); 2820 } 2821 } 2822 2823 static void 2824 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) 2825 { 2826 if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { 2827 union tcp_log_stackspecific log; 2828 2829 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2830 /* R-HU */ 2831 log.u_bbr.flex1 = 0; 2832 log.u_bbr.flex2 = 0; 2833 log.u_bbr.flex3 = 0; 2834 log.u_bbr.flex4 = 0; 2835 log.u_bbr.flex7 = 0; 2836 log.u_bbr.flex8 = settings_desired; 2837 2838 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2839 &bbr->rc_inp->inp_socket->so_rcv, 2840 &bbr->rc_inp->inp_socket->so_snd, 2841 BBR_LOG_SETTINGS_CHG, 0, 2842 0, &log, false, &bbr->rc_tv); 2843 } 2844 } 2845 2846 /* 2847 * Returns the bw from the our filter. 2848 */ 2849 static inline uint64_t 2850 bbr_get_full_bw(struct tcp_bbr *bbr) 2851 { 2852 uint64_t bw; 2853 2854 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2855 2856 return (bw); 2857 } 2858 2859 static inline void 2860 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2861 { 2862 uint64_t calclr; 2863 uint32_t lost, del; 2864 2865 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) 2866 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; 2867 else 2868 lost = 0; 2869 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; 2870 if (lost == 0) { 2871 calclr = 0; 2872 } else if (del) { 2873 calclr = lost; 2874 calclr *= (uint64_t)1000; 2875 calclr /= (uint64_t)del; 2876 } else { 2877 /* Nothing delivered? 100.0% loss */ 2878 calclr = 1000; 2879 } 2880 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; 2881 if (IN_RECOVERY(bbr->rc_tp->t_flags)) 2882 bbr->r_ctl.recovery_lr += (uint32_t)calclr; 2883 bbr->r_ctl.rc_pkt_epoch++; 2884 if (bbr->rc_no_pacing && 2885 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { 2886 bbr->rc_no_pacing = 0; 2887 tcp_bbr_tso_size_check(bbr, cts); 2888 } 2889 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); 2890 bbr->r_ctl.rc_pkt_epoch_time = cts; 2891 /* What was our loss rate */ 2892 bbr_log_pkt_epoch(bbr, cts, line, lost, del); 2893 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; 2894 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; 2895 } 2896 2897 static inline void 2898 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2899 { 2900 uint32_t epoch_time; 2901 2902 /* Tick the RTT clock */ 2903 bbr->r_ctl.rc_rtt_epoch++; 2904 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; 2905 bbr_log_time_epoch(bbr, cts, line, epoch_time); 2906 bbr->r_ctl.rc_rcv_epoch_start = cts; 2907 } 2908 2909 static inline void 2910 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) 2911 { 2912 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { 2913 bbr->rc_is_pkt_epoch_now = 1; 2914 } 2915 } 2916 2917 /* 2918 * Returns the bw from either the b/w filter 2919 * or from the lt_bw (if the connection is being 2920 * policed). 2921 */ 2922 static inline uint64_t 2923 __bbr_get_bw(struct tcp_bbr *bbr) 2924 { 2925 uint64_t bw, min_bw; 2926 uint64_t rtt; 2927 int gm_measure_cnt = 1; 2928 2929 /* 2930 * For startup we make, like google, a 2931 * minimum b/w. This is generated from the 2932 * IW and the rttProp. We do fall back to srtt 2933 * if for some reason (initial handshake) we don't 2934 * have a rttProp. We, in the worst case, fall back 2935 * to the configured min_bw (rc_initial_hptsi_bw). 2936 */ 2937 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 2938 /* Attempt first to use rttProp */ 2939 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2940 if (rtt && (rtt < 0xffffffff)) { 2941 measure: 2942 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2943 ((uint64_t)1000000); 2944 min_bw /= rtt; 2945 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2946 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2947 } 2948 2949 } else if (bbr->rc_tp->t_srtt != 0) { 2950 /* No rttProp, use srtt? */ 2951 rtt = bbr_get_rtt(bbr, BBR_SRTT); 2952 goto measure; 2953 } else { 2954 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2955 } 2956 } else 2957 min_bw = 0; 2958 2959 if ((bbr->rc_past_init_win == 0) && 2960 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) 2961 bbr->rc_past_init_win = 1; 2962 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) 2963 gm_measure_cnt = 0; 2964 if (gm_measure_cnt && 2965 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || 2966 (bbr->rc_past_init_win == 0))) { 2967 /* For google we use our guess rate until we get 1 measurement */ 2968 2969 use_initial_window: 2970 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2971 if (rtt && (rtt < 0xffffffff)) { 2972 /* 2973 * We have an RTT measurement. Use that in 2974 * combination with our initial window to calculate 2975 * a b/w. 2976 */ 2977 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2978 ((uint64_t)1000000); 2979 bw /= rtt; 2980 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2981 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2982 } 2983 } else { 2984 /* Drop back to the 40 and punt to a default */ 2985 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2986 } 2987 if (bw < 1) 2988 /* Probably should panic */ 2989 bw = 1; 2990 if (bw > min_bw) 2991 return (bw); 2992 else 2993 return (min_bw); 2994 } 2995 if (bbr->rc_lt_use_bw) 2996 bw = bbr->r_ctl.rc_lt_bw; 2997 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) 2998 bw = bbr->r_ctl.red_bw; 2999 else 3000 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3001 if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { 3002 /* 3003 * Enforce user set rate limit, keep in mind that 3004 * t_peakrate_thr is in B/s already 3005 */ 3006 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); 3007 } 3008 if (bw == 0) { 3009 /* We should not be at 0, go to the initial window then */ 3010 goto use_initial_window; 3011 } 3012 if (bw < 1) 3013 /* Probably should panic */ 3014 bw = 1; 3015 if (bw < min_bw) 3016 bw = min_bw; 3017 return (bw); 3018 } 3019 3020 static inline uint64_t 3021 bbr_get_bw(struct tcp_bbr *bbr) 3022 { 3023 uint64_t bw; 3024 3025 bw = __bbr_get_bw(bbr); 3026 return (bw); 3027 } 3028 3029 static inline void 3030 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) 3031 { 3032 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; 3033 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; 3034 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3035 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3036 } 3037 3038 static inline void 3039 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) 3040 { 3041 bbr->rc_lt_is_sampling = 0; 3042 bbr->rc_lt_use_bw = 0; 3043 bbr->r_ctl.rc_lt_bw = 0; 3044 bbr_reset_lt_bw_interval(bbr, cts); 3045 } 3046 3047 static inline void 3048 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) 3049 { 3050 uint64_t diff; 3051 3052 /* Do we have a previous sample? */ 3053 if (bbr->r_ctl.rc_lt_bw) { 3054 /* Get the diff in bytes per second */ 3055 if (bbr->r_ctl.rc_lt_bw > bw) 3056 diff = bbr->r_ctl.rc_lt_bw - bw; 3057 else 3058 diff = bw - bbr->r_ctl.rc_lt_bw; 3059 if ((diff <= bbr_lt_bw_diff) || 3060 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { 3061 /* Consider us policed */ 3062 uint32_t saved_bw; 3063 3064 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; 3065 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ 3066 bbr->rc_lt_use_bw = 1; 3067 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 3068 /* 3069 * Use pkt based epoch for measuring length of 3070 * policer up 3071 */ 3072 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; 3073 /* 3074 * reason 4 is we need to start consider being 3075 * policed 3076 */ 3077 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); 3078 return; 3079 } 3080 } 3081 bbr->r_ctl.rc_lt_bw = bw; 3082 bbr_reset_lt_bw_interval(bbr, cts); 3083 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); 3084 } 3085 3086 static void 3087 bbr_randomize_extra_state_time(struct tcp_bbr *bbr) 3088 { 3089 uint32_t ran, deduct; 3090 3091 ran = arc4random_uniform(bbr_rand_ot); 3092 if (ran) { 3093 deduct = bbr->r_ctl.rc_level_state_extra / ran; 3094 bbr->r_ctl.rc_level_state_extra -= deduct; 3095 } 3096 } 3097 /* 3098 * Return randomly the starting state 3099 * to use in probebw. 3100 */ 3101 static uint8_t 3102 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) 3103 { 3104 uint32_t ran; 3105 uint8_t ret_val; 3106 3107 /* Initialize the offset to 0 */ 3108 bbr->r_ctl.rc_exta_time_gd = 0; 3109 bbr->rc_hit_state_1 = 0; 3110 bbr->r_ctl.rc_level_state_extra = 0; 3111 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); 3112 /* 3113 * The math works funny here :) the return value is used to set the 3114 * substate and then the state change is called which increments by 3115 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when 3116 * we fully enter the state. Note that the (8 - 1 - ran) assures that 3117 * we return 1 - 7, so we dont return 0 and end up starting in 3118 * state 1 (DRAIN). 3119 */ 3120 ret_val = BBR_SUBSTATE_COUNT - 1 - ran; 3121 /* Set an epoch */ 3122 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) 3123 bbr_set_epoch(bbr, cts, __LINE__); 3124 3125 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 3126 return (ret_val); 3127 } 3128 3129 static void 3130 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) 3131 { 3132 uint32_t diff, d_time; 3133 uint64_t del_time, bw, lost, delivered; 3134 3135 if (bbr->r_use_policer == 0) 3136 return; 3137 if (bbr->rc_lt_use_bw) { 3138 /* We are using lt bw do we stop yet? */ 3139 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 3140 if (diff > bbr_lt_bw_max_rtts) { 3141 /* Reset it all */ 3142 reset_all: 3143 bbr_reset_lt_bw_sampling(bbr, cts); 3144 if (bbr->rc_filled_pipe) { 3145 bbr_set_epoch(bbr, cts, __LINE__); 3146 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 3147 bbr_substate_change(bbr, cts, __LINE__, 0); 3148 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 3149 bbr_log_type_statechange(bbr, cts, __LINE__); 3150 } else { 3151 /* 3152 * This should not happen really 3153 * unless we remove the startup/drain 3154 * restrictions above. 3155 */ 3156 bbr->rc_bbr_state = BBR_STATE_STARTUP; 3157 bbr_set_epoch(bbr, cts, __LINE__); 3158 bbr->r_ctl.rc_bbr_state_time = cts; 3159 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 3160 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 3161 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 3162 bbr_set_state_target(bbr, __LINE__); 3163 bbr_log_type_statechange(bbr, cts, __LINE__); 3164 } 3165 /* reason 0 is to stop using lt-bw */ 3166 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); 3167 return; 3168 } 3169 if (bbr_lt_intvl_fp == 0) { 3170 /* Not doing false-postive detection */ 3171 return; 3172 } 3173 /* False positive detection */ 3174 if (diff == bbr_lt_intvl_fp) { 3175 /* At bbr_lt_intvl_fp we record the lost */ 3176 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3177 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3178 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { 3179 /* Now is our loss rate still high? */ 3180 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3181 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3182 if ((delivered == 0) || 3183 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { 3184 /* No still below our threshold */ 3185 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); 3186 } else { 3187 /* Yikes its still high, it must be a false positive */ 3188 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); 3189 goto reset_all; 3190 } 3191 } 3192 return; 3193 } 3194 /* 3195 * Wait for the first loss before sampling, to let the policer 3196 * exhaust its tokens and estimate the steady-state rate allowed by 3197 * the policer. Starting samples earlier includes bursts that 3198 * over-estimate the bw. 3199 */ 3200 if (bbr->rc_lt_is_sampling == 0) { 3201 /* reason 1 is to begin doing the sampling */ 3202 if (loss_detected == 0) 3203 return; 3204 bbr_reset_lt_bw_interval(bbr, cts); 3205 bbr->rc_lt_is_sampling = 1; 3206 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); 3207 return; 3208 } 3209 /* Now how long were we delivering long term last> */ 3210 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) 3211 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; 3212 else 3213 d_time = 0; 3214 3215 /* To avoid underestimates, reset sampling if we run out of data. */ 3216 if (bbr->r_ctl.r_app_limited_until) { 3217 /* Can not measure in app-limited state */ 3218 bbr_reset_lt_bw_sampling(bbr, cts); 3219 /* reason 2 is to reset sampling due to app limits */ 3220 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); 3221 return; 3222 } 3223 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 3224 if (diff < bbr_lt_intvl_min_rtts) { 3225 /* 3226 * need more samples (we don't 3227 * start on a round like linux so 3228 * we need 1 more). 3229 */ 3230 /* 6 is not_enough time or no-loss */ 3231 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3232 return; 3233 } 3234 if (diff > (4 * bbr_lt_intvl_min_rtts)) { 3235 /* 3236 * For now if we wait too long, reset all sampling. We need 3237 * to do some research here, its possible that we should 3238 * base this on how much loss as occurred.. something like 3239 * if its under 10% (or some thresh) reset all otherwise 3240 * don't. Thats for phase II I guess. 3241 */ 3242 bbr_reset_lt_bw_sampling(bbr, cts); 3243 /* reason 3 is to reset sampling due too long of sampling */ 3244 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3245 return; 3246 } 3247 /* 3248 * End sampling interval when a packet is lost, so we estimate the 3249 * policer tokens were exhausted. Stopping the sampling before the 3250 * tokens are exhausted under-estimates the policed rate. 3251 */ 3252 if (loss_detected == 0) { 3253 /* 6 is not_enough time or no-loss */ 3254 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3255 return; 3256 } 3257 /* Calculate packets lost and delivered in sampling interval. */ 3258 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3259 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3260 if ((delivered == 0) || 3261 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { 3262 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); 3263 return; 3264 } 3265 if (d_time < 1000) { 3266 /* Not enough time. wait */ 3267 /* 6 is not_enough time or no-loss */ 3268 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3269 return; 3270 } 3271 if (d_time >= (0xffffffff / USECS_IN_MSEC)) { 3272 /* Too long */ 3273 bbr_reset_lt_bw_sampling(bbr, cts); 3274 /* reason 3 is to reset sampling due too long of sampling */ 3275 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3276 return; 3277 } 3278 del_time = d_time; 3279 bw = delivered; 3280 bw *= (uint64_t)USECS_IN_SECOND; 3281 bw /= del_time; 3282 bbr_lt_bw_samp_done(bbr, bw, cts, d_time); 3283 } 3284 3285 /* 3286 * Allocate a sendmap from our zone. 3287 */ 3288 static struct bbr_sendmap * 3289 bbr_alloc(struct tcp_bbr *bbr) 3290 { 3291 struct bbr_sendmap *rsm; 3292 3293 BBR_STAT_INC(bbr_to_alloc); 3294 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); 3295 if (rsm) { 3296 bbr->r_ctl.rc_num_maps_alloced++; 3297 return (rsm); 3298 } 3299 if (bbr->r_ctl.rc_free_cnt) { 3300 BBR_STAT_INC(bbr_to_alloc_emerg); 3301 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 3302 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 3303 bbr->r_ctl.rc_free_cnt--; 3304 return (rsm); 3305 } 3306 BBR_STAT_INC(bbr_to_alloc_failed); 3307 return (NULL); 3308 } 3309 3310 static struct bbr_sendmap * 3311 bbr_alloc_full_limit(struct tcp_bbr *bbr) 3312 { 3313 if ((V_tcp_map_entries_limit > 0) && 3314 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 3315 BBR_STAT_INC(bbr_alloc_limited); 3316 if (!bbr->alloc_limit_reported) { 3317 bbr->alloc_limit_reported = 1; 3318 BBR_STAT_INC(bbr_alloc_limited_conns); 3319 } 3320 return (NULL); 3321 } 3322 return (bbr_alloc(bbr)); 3323 } 3324 3325 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 3326 static struct bbr_sendmap * 3327 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) 3328 { 3329 struct bbr_sendmap *rsm; 3330 3331 if (limit_type) { 3332 /* currently there is only one limit type */ 3333 if (V_tcp_map_split_limit > 0 && 3334 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { 3335 BBR_STAT_INC(bbr_split_limited); 3336 if (!bbr->alloc_limit_reported) { 3337 bbr->alloc_limit_reported = 1; 3338 BBR_STAT_INC(bbr_alloc_limited_conns); 3339 } 3340 return (NULL); 3341 } 3342 } 3343 3344 /* allocate and mark in the limit type, if set */ 3345 rsm = bbr_alloc(bbr); 3346 if (rsm != NULL && limit_type) { 3347 rsm->r_limit_type = limit_type; 3348 bbr->r_ctl.rc_num_split_allocs++; 3349 } 3350 return (rsm); 3351 } 3352 3353 static void 3354 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 3355 { 3356 if (rsm->r_limit_type) { 3357 /* currently there is only one limit type */ 3358 bbr->r_ctl.rc_num_split_allocs--; 3359 } 3360 if (rsm->r_is_smallmap) 3361 bbr->r_ctl.rc_num_small_maps_alloced--; 3362 if (bbr->r_ctl.rc_tlp_send == rsm) 3363 bbr->r_ctl.rc_tlp_send = NULL; 3364 if (bbr->r_ctl.rc_resend == rsm) { 3365 bbr->r_ctl.rc_resend = NULL; 3366 } 3367 if (bbr->r_ctl.rc_next == rsm) 3368 bbr->r_ctl.rc_next = NULL; 3369 if (bbr->r_ctl.rc_sacklast == rsm) 3370 bbr->r_ctl.rc_sacklast = NULL; 3371 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 3372 memset(rsm, 0, sizeof(struct bbr_sendmap)); 3373 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 3374 rsm->r_limit_type = 0; 3375 bbr->r_ctl.rc_free_cnt++; 3376 return; 3377 } 3378 bbr->r_ctl.rc_num_maps_alloced--; 3379 uma_zfree(bbr_zone, rsm); 3380 } 3381 3382 /* 3383 * Returns the BDP. 3384 */ 3385 static uint64_t 3386 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { 3387 /* 3388 * Calculate the bytes in flight needed given the bw (in bytes per 3389 * second) and the specifyed rtt in useconds. We need to put out the 3390 * returned value per RTT to match that rate. Gain will normally 3391 * raise it up from there. 3392 * 3393 * This should not overflow as long as the bandwidth is below 1 3394 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller 3395 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). 3396 */ 3397 uint64_t usec_per_sec; 3398 3399 usec_per_sec = USECS_IN_SECOND; 3400 return ((rtt * bw) / usec_per_sec); 3401 } 3402 3403 /* 3404 * Return the initial cwnd. 3405 */ 3406 static uint32_t 3407 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) 3408 { 3409 uint32_t i_cwnd; 3410 3411 if (bbr->rc_init_win) { 3412 i_cwnd = bbr->rc_init_win * tp->t_maxseg; 3413 } else if (V_tcp_initcwnd_segments) 3414 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 3415 max(2 * tp->t_maxseg, 14600)); 3416 else if (V_tcp_do_rfc3390) 3417 i_cwnd = min(4 * tp->t_maxseg, 3418 max(2 * tp->t_maxseg, 4380)); 3419 else { 3420 /* Per RFC5681 Section 3.1 */ 3421 if (tp->t_maxseg > 2190) 3422 i_cwnd = 2 * tp->t_maxseg; 3423 else if (tp->t_maxseg > 1095) 3424 i_cwnd = 3 * tp->t_maxseg; 3425 else 3426 i_cwnd = 4 * tp->t_maxseg; 3427 } 3428 return (i_cwnd); 3429 } 3430 3431 /* 3432 * Given a specified gain, return the target 3433 * cwnd based on that gain. 3434 */ 3435 static uint32_t 3436 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) 3437 { 3438 uint64_t bdp, rtt; 3439 uint32_t cwnd; 3440 3441 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || 3442 (bbr_get_full_bw(bbr) == 0)) { 3443 /* No measurements yet */ 3444 return (bbr_initial_cwnd(bbr, bbr->rc_tp)); 3445 } 3446 /* 3447 * Get bytes per RTT needed (rttProp is normally in 3448 * bbr_cwndtarget_rtt_touse) 3449 */ 3450 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); 3451 /* Get the bdp from the two values */ 3452 bdp = bbr_get_bw_delay_prod(rtt, bw); 3453 /* Now apply the gain */ 3454 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); 3455 3456 return (cwnd); 3457 } 3458 3459 static uint32_t 3460 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) 3461 { 3462 uint32_t cwnd, mss; 3463 3464 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 3465 /* Get the base cwnd with gain rounded to a mss */ 3466 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); 3467 /* 3468 * Add in N (2 default since we do not have a 3469 * fq layer to trap packets in) quanta's per the I-D 3470 * section 4.2.3.2 quanta adjust. 3471 */ 3472 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); 3473 if (bbr->rc_use_google) { 3474 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3475 (bbr_state_val(bbr) == BBR_SUB_GAIN)) { 3476 /* 3477 * The linux implementation adds 3478 * an extra 2 x mss in gain cycle which 3479 * is documented no-where except in the code. 3480 * so we add more for Neal undocumented feature 3481 */ 3482 cwnd += 2 * mss; 3483 } 3484 if ((cwnd / mss) & 0x1) { 3485 /* Round up for odd num mss */ 3486 cwnd += mss; 3487 } 3488 } 3489 /* Are we below the min cwnd? */ 3490 if (cwnd < get_min_cwnd(bbr)) 3491 return (get_min_cwnd(bbr)); 3492 return (cwnd); 3493 } 3494 3495 static uint16_t 3496 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) 3497 { 3498 if (gain < 1) 3499 gain = 1; 3500 return (gain); 3501 } 3502 3503 static uint32_t 3504 bbr_get_header_oh(struct tcp_bbr *bbr) 3505 { 3506 int seg_oh; 3507 3508 seg_oh = 0; 3509 if (bbr->r_ctl.rc_inc_tcp_oh) { 3510 /* Do we include TCP overhead? */ 3511 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); 3512 } 3513 if (bbr->r_ctl.rc_inc_ip_oh) { 3514 /* Do we include IP overhead? */ 3515 #ifdef INET6 3516 if (bbr->r_is_v6) { 3517 seg_oh += sizeof(struct ip6_hdr); 3518 } else 3519 #endif 3520 { 3521 3522 #ifdef INET 3523 seg_oh += sizeof(struct ip); 3524 #endif 3525 } 3526 } 3527 if (bbr->r_ctl.rc_inc_enet_oh) { 3528 /* Do we include the ethernet overhead? */ 3529 seg_oh += sizeof(struct ether_header); 3530 } 3531 return(seg_oh); 3532 } 3533 3534 static uint32_t 3535 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) 3536 { 3537 uint64_t divor, res, tim; 3538 3539 if (useconds_time == 0) 3540 return (0); 3541 gain = bbr_gain_adjust(bbr, gain); 3542 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; 3543 tim = useconds_time; 3544 res = (tim * bw * gain) / divor; 3545 if (res == 0) 3546 res = 1; 3547 return ((uint32_t)res); 3548 } 3549 3550 /* 3551 * Given a gain and a length return the delay in useconds that 3552 * should be used to evenly space out packets 3553 * on the connection (based on the gain factor). 3554 */ 3555 static uint32_t 3556 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) 3557 { 3558 uint64_t bw, lentim, res; 3559 uint32_t usecs, srtt, over = 0; 3560 uint32_t seg_oh, num_segs, maxseg; 3561 3562 if (len == 0) 3563 return (0); 3564 3565 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 3566 num_segs = (len + maxseg - 1) / maxseg; 3567 if (bbr->rc_use_google == 0) { 3568 seg_oh = bbr_get_header_oh(bbr); 3569 len += (num_segs * seg_oh); 3570 } 3571 gain = bbr_gain_adjust(bbr, gain); 3572 bw = bbr_get_bw(bbr); 3573 if (bbr->rc_use_google) { 3574 uint64_t cbw; 3575 3576 /* 3577 * Reduce the b/w by the google discount 3578 * factor 10 = 1%. 3579 */ 3580 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); 3581 cbw /= (uint64_t)1000; 3582 /* We don't apply a discount if it results in 0 */ 3583 if (cbw > 0) 3584 bw = cbw; 3585 } 3586 lentim = ((uint64_t)len * 3587 (uint64_t)USECS_IN_SECOND * 3588 (uint64_t)BBR_UNIT); 3589 res = lentim / ((uint64_t)gain * bw); 3590 if (res == 0) 3591 res = 1; 3592 usecs = (uint32_t)res; 3593 srtt = bbr_get_rtt(bbr, BBR_SRTT); 3594 if (bbr_hptsi_max_mul && bbr_hptsi_max_div && 3595 (bbr->rc_use_google == 0) && 3596 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { 3597 /* 3598 * We cannot let the delay be more than 1/2 the srtt time. 3599 * Otherwise we cannot pace out or send properly. 3600 */ 3601 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; 3602 BBR_STAT_INC(bbr_hpts_min_time); 3603 } 3604 if (!nolog) 3605 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); 3606 return (usecs); 3607 } 3608 3609 static void 3610 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, 3611 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) 3612 { 3613 INP_WLOCK_ASSERT(tp->t_inpcb); 3614 uint64_t bw; 3615 uint32_t cwnd, target_cwnd, saved_bytes, maxseg; 3616 int32_t meth; 3617 3618 #ifdef STATS 3619 if ((tp->t_flags & TF_GPUTINPROG) && 3620 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 3621 /* 3622 * Strech acks and compressed acks will cause this to 3623 * oscillate but we are doing it the same way as the main 3624 * stack so it will be compariable (though possibly not 3625 * ideal). 3626 */ 3627 int32_t cgput; 3628 int64_t gput, time_stamp; 3629 3630 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; 3631 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); 3632 cgput = gput / time_stamp; 3633 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 3634 cgput); 3635 if (tp->t_stats_gput_prev > 0) 3636 stats_voi_update_abs_s32(tp->t_stats, 3637 VOI_TCP_GPUT_ND, 3638 ((gput - tp->t_stats_gput_prev) * 100) / 3639 tp->t_stats_gput_prev); 3640 tp->t_flags &= ~TF_GPUTINPROG; 3641 tp->t_stats_gput_prev = cgput; 3642 } 3643 #endif 3644 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3645 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 3646 /* We don't change anything in probe-rtt */ 3647 return; 3648 } 3649 maxseg = tp->t_maxseg - bbr->rc_last_options; 3650 saved_bytes = bytes_this_ack; 3651 bytes_this_ack += sack_changed; 3652 if (bytes_this_ack > prev_acked) { 3653 bytes_this_ack -= prev_acked; 3654 /* 3655 * A byte ack'd gives us a full mss 3656 * to be like linux i.e. they count packets. 3657 */ 3658 if ((bytes_this_ack < maxseg) && bbr->rc_use_google) 3659 bytes_this_ack = maxseg; 3660 } else { 3661 /* Unlikely */ 3662 bytes_this_ack = 0; 3663 } 3664 cwnd = tp->snd_cwnd; 3665 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3666 if (bw) 3667 target_cwnd = bbr_get_target_cwnd(bbr, 3668 bw, 3669 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); 3670 else 3671 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); 3672 if (IN_RECOVERY(tp->t_flags) && 3673 (bbr->bbr_prev_in_rec == 0)) { 3674 /* 3675 * We are entering recovery and 3676 * thus packet conservation. 3677 */ 3678 bbr->pkt_conservation = 1; 3679 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; 3680 cwnd = ctf_flight_size(tp, 3681 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3682 bytes_this_ack; 3683 } 3684 if (IN_RECOVERY(tp->t_flags)) { 3685 uint32_t flight; 3686 3687 bbr->bbr_prev_in_rec = 1; 3688 if (cwnd > losses) { 3689 cwnd -= losses; 3690 if (cwnd < maxseg) 3691 cwnd = maxseg; 3692 } else 3693 cwnd = maxseg; 3694 flight = ctf_flight_size(tp, 3695 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3696 bbr_log_type_cwndupd(bbr, flight, 0, 3697 losses, 10, 0, 0, line); 3698 if (bbr->pkt_conservation) { 3699 uint32_t time_in; 3700 3701 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) 3702 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; 3703 else 3704 time_in = 0; 3705 3706 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 3707 /* Clear packet conservation after an rttProp */ 3708 bbr->pkt_conservation = 0; 3709 } else { 3710 if ((flight + bytes_this_ack) > cwnd) 3711 cwnd = flight + bytes_this_ack; 3712 if (cwnd < get_min_cwnd(bbr)) 3713 cwnd = get_min_cwnd(bbr); 3714 tp->snd_cwnd = cwnd; 3715 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, 3716 prev_acked, 1, target_cwnd, th->th_ack, line); 3717 return; 3718 } 3719 } 3720 } else 3721 bbr->bbr_prev_in_rec = 0; 3722 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { 3723 bbr->r_ctl.restrict_growth--; 3724 if (bytes_this_ack > maxseg) 3725 bytes_this_ack = maxseg; 3726 } 3727 if (bbr->rc_filled_pipe) { 3728 /* 3729 * Here we have exited startup and filled the pipe. We will 3730 * thus allow the cwnd to shrink to the target. We hit here 3731 * mostly. 3732 */ 3733 uint32_t s_cwnd; 3734 3735 meth = 2; 3736 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); 3737 if (s_cwnd > cwnd) 3738 cwnd = s_cwnd; 3739 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) 3740 cwnd = s_cwnd; 3741 } else { 3742 /* 3743 * Here we are still in startup, we increase cwnd by what 3744 * has been acked. 3745 */ 3746 if ((cwnd < target_cwnd) || 3747 (bbr->rc_past_init_win == 0)) { 3748 meth = 3; 3749 cwnd += bytes_this_ack; 3750 } else { 3751 /* 3752 * Method 4 means we are at target so no gain in 3753 * startup and past the initial window. 3754 */ 3755 meth = 4; 3756 } 3757 } 3758 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); 3759 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); 3760 } 3761 3762 static void 3763 tcp_bbr_partialack(struct tcpcb *tp) 3764 { 3765 struct tcp_bbr *bbr; 3766 3767 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3768 INP_WLOCK_ASSERT(tp->t_inpcb); 3769 if (ctf_flight_size(tp, 3770 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 3771 tp->snd_cwnd) { 3772 bbr->r_wanted_output = 1; 3773 } 3774 } 3775 3776 static void 3777 bbr_post_recovery(struct tcpcb *tp) 3778 { 3779 struct tcp_bbr *bbr; 3780 uint32_t flight; 3781 3782 INP_WLOCK_ASSERT(tp->t_inpcb); 3783 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3784 /* 3785 * Here we just exit recovery. 3786 */ 3787 EXIT_RECOVERY(tp->t_flags); 3788 /* Lock in our b/w reduction for the specified number of pkt-epochs */ 3789 bbr->r_recovery_bw = 0; 3790 tp->snd_recover = tp->snd_una; 3791 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3792 bbr->pkt_conservation = 0; 3793 if (bbr->rc_use_google == 0) { 3794 /* 3795 * For non-google mode lets 3796 * go ahead and make sure we clear 3797 * the recovery state so if we 3798 * bounce back in to recovery we 3799 * will do PC. 3800 */ 3801 bbr->bbr_prev_in_rec = 0; 3802 } 3803 bbr_log_type_exit_rec(bbr); 3804 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3805 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3806 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); 3807 } else { 3808 /* For probe-rtt case lets fix up its saved_cwnd */ 3809 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { 3810 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; 3811 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); 3812 } 3813 } 3814 flight = ctf_flight_size(tp, 3815 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3816 if ((bbr->rc_use_google == 0) && 3817 bbr_do_red) { 3818 uint64_t val, lr2use; 3819 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; 3820 uint32_t *cwnd_p; 3821 3822 if (bbr_get_rtt(bbr, BBR_SRTT)) { 3823 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); 3824 val /= bbr_get_rtt(bbr, BBR_SRTT); 3825 ratio = (uint32_t)val; 3826 } else 3827 ratio = 1000; 3828 3829 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, 3830 bbr->r_ctl.recovery_lr, 21, 3831 ratio, 3832 bbr->r_ctl.rc_red_cwnd_pe, 3833 __LINE__); 3834 if ((ratio < bbr_do_red) || (bbr_do_red == 0)) 3835 goto done; 3836 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3837 bbr_prtt_slam_cwnd) || 3838 (bbr_sub_drain_slam_cwnd && 3839 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3840 bbr->rc_hit_state_1 && 3841 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || 3842 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && 3843 bbr_slam_cwnd_in_main_drain)) { 3844 /* 3845 * Here we must poke at the saved cwnd 3846 * as well as the cwnd. 3847 */ 3848 cwnd = bbr->r_ctl.rc_saved_cwnd; 3849 cwnd_p = &bbr->r_ctl.rc_saved_cwnd; 3850 } else { 3851 cwnd = tp->snd_cwnd; 3852 cwnd_p = &tp->snd_cwnd; 3853 } 3854 maxseg = tp->t_maxseg - bbr->rc_last_options; 3855 /* Add the overall lr with the recovery lr */ 3856 if (bbr->r_ctl.rc_lost == 0) 3857 lr2use = 0; 3858 else if (bbr->r_ctl.rc_delivered == 0) 3859 lr2use = 1000; 3860 else { 3861 lr2use = bbr->r_ctl.rc_lost * 1000; 3862 lr2use /= bbr->r_ctl.rc_delivered; 3863 } 3864 lr2use += bbr->r_ctl.recovery_lr; 3865 acks_inflight = (flight / (maxseg * 2)); 3866 if (bbr_red_scale) { 3867 lr2use *= bbr_get_rtt(bbr, BBR_SRTT); 3868 lr2use /= bbr_red_scale; 3869 if ((bbr_red_growth_restrict) && 3870 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) 3871 bbr->r_ctl.restrict_growth += acks_inflight; 3872 } 3873 if (lr2use) { 3874 val = (uint64_t)cwnd * lr2use; 3875 val /= 1000; 3876 if (cwnd > val) 3877 newcwnd = roundup((cwnd - val), maxseg); 3878 else 3879 newcwnd = maxseg; 3880 } else { 3881 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; 3882 val /= (uint64_t)bbr_red_div; 3883 newcwnd = roundup((uint32_t)val, maxseg); 3884 } 3885 /* with standard delayed acks how many acks can I expect? */ 3886 if (bbr_drop_limit == 0) { 3887 /* 3888 * Anticpate how much we will 3889 * raise the cwnd based on the acks. 3890 */ 3891 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { 3892 /* We do enforce the min (with the acks) */ 3893 newcwnd = (get_min_cwnd(bbr) - acks_inflight); 3894 } 3895 } else { 3896 /* 3897 * A strict drop limit of N is is inplace 3898 */ 3899 if (newcwnd < (bbr_drop_limit * maxseg)) { 3900 newcwnd = bbr_drop_limit * maxseg; 3901 } 3902 } 3903 /* For the next N acks do we restrict the growth */ 3904 *cwnd_p = newcwnd; 3905 if (tp->snd_cwnd > newcwnd) 3906 tp->snd_cwnd = newcwnd; 3907 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, 3908 (uint32_t)lr2use, 3909 bbr_get_rtt(bbr, BBR_SRTT), __LINE__); 3910 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; 3911 } 3912 done: 3913 bbr->r_ctl.recovery_lr = 0; 3914 if (flight <= tp->snd_cwnd) { 3915 bbr->r_wanted_output = 1; 3916 } 3917 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3918 } 3919 3920 static void 3921 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) 3922 { 3923 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3924 /* Limit the drop in b/w to 1/2 our current filter. */ 3925 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) 3926 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; 3927 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) 3928 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; 3929 tcp_bbr_tso_size_check(bbr, cts); 3930 } 3931 3932 static void 3933 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) 3934 { 3935 struct tcp_bbr *bbr; 3936 3937 INP_WLOCK_ASSERT(tp->t_inpcb); 3938 #ifdef STATS 3939 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); 3940 #endif 3941 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3942 switch (type) { 3943 case CC_NDUPACK: 3944 if (!IN_RECOVERY(tp->t_flags)) { 3945 tp->snd_recover = tp->snd_max; 3946 /* Start a new epoch */ 3947 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 3948 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { 3949 /* 3950 * Move forward the lt epoch 3951 * so it won't count the truncated 3952 * epoch. 3953 */ 3954 bbr->r_ctl.rc_lt_epoch++; 3955 } 3956 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 3957 /* 3958 * Just like the policer detection code 3959 * if we are in startup we must push 3960 * forward the last startup epoch 3961 * to hide the truncated PE. 3962 */ 3963 bbr->r_ctl.rc_bbr_last_startup_epoch++; 3964 } 3965 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; 3966 ENTER_RECOVERY(tp->t_flags); 3967 bbr->rc_tlp_rtx_out = 0; 3968 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; 3969 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3970 if (tcp_in_hpts(bbr->rc_inp) && 3971 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { 3972 /* 3973 * When we enter recovery, we need to restart 3974 * any timers. This may mean we gain an agg 3975 * early, which will be made up for at the last 3976 * rxt out. 3977 */ 3978 bbr->rc_timer_first = 1; 3979 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 3980 } 3981 /* 3982 * Calculate a new cwnd based on to the current 3983 * delivery rate with no gain. We get the bdp 3984 * without gaining it up like we normally would and 3985 * we use the last cur_del_rate. 3986 */ 3987 if ((bbr->rc_use_google == 0) && 3988 (bbr->r_ctl.bbr_rttprobe_gain_val || 3989 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { 3990 tp->snd_cwnd = ctf_flight_size(tp, 3991 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3992 (tp->t_maxseg - bbr->rc_last_options); 3993 if (tp->snd_cwnd < get_min_cwnd(bbr)) { 3994 /* We always gate to min cwnd */ 3995 tp->snd_cwnd = get_min_cwnd(bbr); 3996 } 3997 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); 3998 } 3999 bbr_log_type_enter_rec(bbr, rsm->r_start); 4000 } 4001 break; 4002 case CC_RTO_ERR: 4003 KMOD_TCPSTAT_INC(tcps_sndrexmitbad); 4004 /* RTO was unnecessary, so reset everything. */ 4005 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); 4006 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 4007 tp->snd_cwnd = tp->snd_cwnd_prev; 4008 tp->snd_ssthresh = tp->snd_ssthresh_prev; 4009 tp->snd_recover = tp->snd_recover_prev; 4010 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 4011 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); 4012 } 4013 tp->t_badrxtwin = 0; 4014 break; 4015 } 4016 } 4017 4018 /* 4019 * Indicate whether this ack should be delayed. We can delay the ack if 4020 * following conditions are met: 4021 * - There is no delayed ack timer in progress. 4022 * - Our last ack wasn't a 0-sized window. We never want to delay 4023 * the ack that opens up a 0-sized window. 4024 * - LRO wasn't used for this segment. We make sure by checking that the 4025 * segment size is not larger than the MSS. 4026 * - Delayed acks are enabled or this is a half-synchronized T/TCP 4027 * connection. 4028 * - The data being acked is less than a full segment (a stretch ack 4029 * of more than a segment we should ack. 4030 * - nsegs is 1 (if its more than that we received more than 1 ack). 4031 */ 4032 #define DELAY_ACK(tp, bbr, nsegs) \ 4033 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 4034 ((tp->t_flags & TF_DELACK) == 0) && \ 4035 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ 4036 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 4037 4038 /* 4039 * Return the lowest RSM in the map of 4040 * packets still in flight that is not acked. 4041 * This should normally find on the first one 4042 * since we remove packets from the send 4043 * map after they are marked ACKED. 4044 */ 4045 static struct bbr_sendmap * 4046 bbr_find_lowest_rsm(struct tcp_bbr *bbr) 4047 { 4048 struct bbr_sendmap *rsm; 4049 4050 /* 4051 * Walk the time-order transmitted list looking for an rsm that is 4052 * not acked. This will be the one that was sent the longest time 4053 * ago that is still outstanding. 4054 */ 4055 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { 4056 if (rsm->r_flags & BBR_ACKED) { 4057 continue; 4058 } 4059 goto finish; 4060 } 4061 finish: 4062 return (rsm); 4063 } 4064 4065 static struct bbr_sendmap * 4066 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 4067 { 4068 struct bbr_sendmap *prsm; 4069 4070 /* 4071 * Walk the sequence order list backward until we hit and arrive at 4072 * the highest seq not acked. In theory when this is called it 4073 * should be the last segment (which it was not). 4074 */ 4075 prsm = rsm; 4076 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4077 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { 4078 continue; 4079 } 4080 return (prsm); 4081 } 4082 return (NULL); 4083 } 4084 4085 /* 4086 * Returns to the caller the number of microseconds that 4087 * the packet can be outstanding before we think we 4088 * should have had an ack returned. 4089 */ 4090 static uint32_t 4091 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) 4092 { 4093 /* 4094 * lro is the flag we use to determine if we have seen reordering. 4095 * If it gets set we have seen reordering. The reorder logic either 4096 * works in one of two ways: 4097 * 4098 * If reorder-fade is configured, then we track the last time we saw 4099 * re-ordering occur. If we reach the point where enough time as 4100 * passed we no longer consider reordering has occuring. 4101 * 4102 * Or if reorder-face is 0, then once we see reordering we consider 4103 * the connection to alway be subject to reordering and just set lro 4104 * to 1. 4105 * 4106 * In the end if lro is non-zero we add the extra time for 4107 * reordering in. 4108 */ 4109 int32_t lro; 4110 uint32_t thresh, t_rxtcur; 4111 4112 if (srtt == 0) 4113 srtt = 1; 4114 if (bbr->r_ctl.rc_reorder_ts) { 4115 if (bbr->r_ctl.rc_reorder_fade) { 4116 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { 4117 lro = cts - bbr->r_ctl.rc_reorder_ts; 4118 if (lro == 0) { 4119 /* 4120 * No time as passed since the last 4121 * reorder, mark it as reordering. 4122 */ 4123 lro = 1; 4124 } 4125 } else { 4126 /* Negative time? */ 4127 lro = 0; 4128 } 4129 if (lro > bbr->r_ctl.rc_reorder_fade) { 4130 /* Turn off reordering seen too */ 4131 bbr->r_ctl.rc_reorder_ts = 0; 4132 lro = 0; 4133 } 4134 } else { 4135 /* Reodering does not fade */ 4136 lro = 1; 4137 } 4138 } else { 4139 lro = 0; 4140 } 4141 thresh = srtt + bbr->r_ctl.rc_pkt_delay; 4142 if (lro) { 4143 /* It must be set, if not you get 1/4 rtt */ 4144 if (bbr->r_ctl.rc_reorder_shift) 4145 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); 4146 else 4147 thresh += (srtt >> 2); 4148 } else { 4149 thresh += 1000; 4150 } 4151 /* We don't let the rack timeout be above a RTO */ 4152 if ((bbr->rc_tp)->t_srtt == 0) 4153 t_rxtcur = BBR_INITIAL_RTO; 4154 else 4155 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 4156 if (thresh > t_rxtcur) { 4157 thresh = t_rxtcur; 4158 } 4159 /* And we don't want it above the RTO max either */ 4160 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4161 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4162 } 4163 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); 4164 return (thresh); 4165 } 4166 4167 /* 4168 * Return to the caller the amount of time in mico-seconds 4169 * that should be used for the TLP timer from the last 4170 * send time of this packet. 4171 */ 4172 static uint32_t 4173 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 4174 struct bbr_sendmap *rsm, uint32_t srtt, 4175 uint32_t cts) 4176 { 4177 uint32_t thresh, len, maxseg, t_rxtcur; 4178 struct bbr_sendmap *prsm; 4179 4180 if (srtt == 0) 4181 srtt = 1; 4182 if (bbr->rc_tlp_threshold) 4183 thresh = srtt + (srtt / bbr->rc_tlp_threshold); 4184 else 4185 thresh = (srtt * 2); 4186 maxseg = tp->t_maxseg - bbr->rc_last_options; 4187 /* Get the previous sent packet, if any */ 4188 len = rsm->r_end - rsm->r_start; 4189 4190 /* 2.1 behavior */ 4191 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); 4192 if (prsm && (len <= maxseg)) { 4193 /* 4194 * Two packets outstanding, thresh should be (2*srtt) + 4195 * possible inter-packet delay (if any). 4196 */ 4197 uint32_t inter_gap = 0; 4198 int idx, nidx; 4199 4200 idx = rsm->r_rtr_cnt - 1; 4201 nidx = prsm->r_rtr_cnt - 1; 4202 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 4203 /* Yes it was sent later (or at the same time) */ 4204 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 4205 } 4206 thresh += inter_gap; 4207 } else if (len <= maxseg) { 4208 /* 4209 * Possibly compensate for delayed-ack. 4210 */ 4211 uint32_t alt_thresh; 4212 4213 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; 4214 if (alt_thresh > thresh) 4215 thresh = alt_thresh; 4216 } 4217 /* Not above the current RTO */ 4218 if (tp->t_srtt == 0) 4219 t_rxtcur = BBR_INITIAL_RTO; 4220 else 4221 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); 4222 4223 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); 4224 /* Not above an RTO */ 4225 if (thresh > t_rxtcur) { 4226 thresh = t_rxtcur; 4227 } 4228 /* Not above a RTO max */ 4229 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4230 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4231 } 4232 /* And now apply the user TLP min */ 4233 if (thresh < bbr_tlp_min) { 4234 thresh = bbr_tlp_min; 4235 } 4236 return (thresh); 4237 } 4238 4239 /* 4240 * Return one of three RTTs to use (in microseconds). 4241 */ 4242 static __inline uint32_t 4243 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) 4244 { 4245 uint32_t f_rtt; 4246 uint32_t srtt; 4247 4248 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 4249 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { 4250 /* We have no rtt at all */ 4251 if (bbr->rc_tp->t_srtt == 0) 4252 f_rtt = BBR_INITIAL_RTO; 4253 else 4254 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4255 /* 4256 * Since we don't know how good the rtt is apply a 4257 * delayed-ack min 4258 */ 4259 if (f_rtt < bbr_delayed_ack_time) { 4260 f_rtt = bbr_delayed_ack_time; 4261 } 4262 } 4263 /* Take the filter version or last measured pkt-rtt */ 4264 if (rtt_type == BBR_RTT_PROP) { 4265 srtt = f_rtt; 4266 } else if (rtt_type == BBR_RTT_PKTRTT) { 4267 if (bbr->r_ctl.rc_pkt_epoch_rtt) { 4268 srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 4269 } else { 4270 /* No pkt rtt yet */ 4271 srtt = f_rtt; 4272 } 4273 } else if (rtt_type == BBR_RTT_RACK) { 4274 srtt = bbr->r_ctl.rc_last_rtt; 4275 /* We need to add in any internal delay for our timer */ 4276 if (bbr->rc_ack_was_delayed) 4277 srtt += bbr->r_ctl.rc_ack_hdwr_delay; 4278 } else if (rtt_type == BBR_SRTT) { 4279 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4280 } else { 4281 /* TSNH */ 4282 srtt = f_rtt; 4283 #ifdef BBR_INVARIANTS 4284 panic("Unknown rtt request type %d", rtt_type); 4285 #endif 4286 } 4287 return (srtt); 4288 } 4289 4290 static int 4291 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) 4292 { 4293 uint32_t thresh; 4294 4295 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), 4296 cts, rsm); 4297 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { 4298 /* It is lost (past time) */ 4299 return (1); 4300 } 4301 return (0); 4302 } 4303 4304 /* 4305 * Return a sendmap if we need to retransmit something. 4306 */ 4307 static struct bbr_sendmap * 4308 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4309 { 4310 /* 4311 * Check to see that we don't need to fall into recovery. We will 4312 * need to do so if our oldest transmit is past the time we should 4313 * have had an ack. 4314 */ 4315 4316 struct bbr_sendmap *rsm; 4317 int32_t idx; 4318 4319 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { 4320 /* Nothing outstanding that we know of */ 4321 return (NULL); 4322 } 4323 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 4324 if (rsm == NULL) { 4325 /* Nothing in the transmit map */ 4326 return (NULL); 4327 } 4328 if (tp->t_flags & TF_SENTFIN) { 4329 /* Fin restricted, don't find anything once a fin is sent */ 4330 return (NULL); 4331 } 4332 if (rsm->r_flags & BBR_ACKED) { 4333 /* 4334 * Ok the first one is acked (this really should not happen 4335 * since we remove the from the tmap once they are acked) 4336 */ 4337 rsm = bbr_find_lowest_rsm(bbr); 4338 if (rsm == NULL) 4339 return (NULL); 4340 } 4341 idx = rsm->r_rtr_cnt - 1; 4342 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { 4343 /* Send timestamp is the same or less? can't be ready */ 4344 return (NULL); 4345 } 4346 /* Get our RTT time */ 4347 if (bbr_is_lost(bbr, rsm, cts) && 4348 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || 4349 (rsm->r_flags & BBR_SACK_PASSED))) { 4350 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4351 rsm->r_flags |= BBR_MARKED_LOST; 4352 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4353 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4354 } 4355 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); 4356 #ifdef BBR_INVARIANTS 4357 if ((rsm->r_end - rsm->r_start) == 0) 4358 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); 4359 #endif 4360 return (rsm); 4361 } 4362 return (NULL); 4363 } 4364 4365 /* 4366 * RACK Timer, here we simply do logging and house keeping. 4367 * the normal bbr_output_wtime() function will call the 4368 * appropriate thing to check if we need to do a RACK retransmit. 4369 * We return 1, saying don't proceed with bbr_output_wtime only 4370 * when all timers have been stopped (destroyed PCB?). 4371 */ 4372 static int 4373 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4374 { 4375 /* 4376 * This timer simply provides an internal trigger to send out data. 4377 * The check_recovery_mode call will see if there are needed 4378 * retransmissions, if so we will enter fast-recovery. The output 4379 * call may or may not do the same thing depending on sysctl 4380 * settings. 4381 */ 4382 uint32_t lost; 4383 4384 if (bbr->rc_all_timers_stopped) { 4385 return (1); 4386 } 4387 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4388 /* Its not time yet */ 4389 return (0); 4390 } 4391 BBR_STAT_INC(bbr_to_tot); 4392 lost = bbr->r_ctl.rc_lost; 4393 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4394 bbr_set_state(tp, bbr, 0); 4395 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); 4396 if (bbr->r_ctl.rc_resend == NULL) { 4397 /* Lets do the check here */ 4398 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 4399 } 4400 if (bbr_policer_call_from_rack_to) 4401 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4402 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 4403 return (0); 4404 } 4405 4406 static __inline void 4407 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) 4408 { 4409 int idx; 4410 4411 nrsm->r_start = start; 4412 nrsm->r_end = rsm->r_end; 4413 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4414 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; 4415 nrsm->r_flags = rsm->r_flags; 4416 /* We don't transfer forward the SYN flag */ 4417 nrsm->r_flags &= ~BBR_HAS_SYN; 4418 /* We move forward the FIN flag, not that this should happen */ 4419 rsm->r_flags &= ~BBR_HAS_FIN; 4420 nrsm->r_dupack = rsm->r_dupack; 4421 nrsm->r_rtr_bytes = 0; 4422 nrsm->r_is_gain = rsm->r_is_gain; 4423 nrsm->r_is_drain = rsm->r_is_drain; 4424 nrsm->r_delivered = rsm->r_delivered; 4425 nrsm->r_ts_valid = rsm->r_ts_valid; 4426 nrsm->r_del_ack_ts = rsm->r_del_ack_ts; 4427 nrsm->r_del_time = rsm->r_del_time; 4428 nrsm->r_app_limited = rsm->r_app_limited; 4429 nrsm->r_first_sent_time = rsm->r_first_sent_time; 4430 nrsm->r_flight_at_send = rsm->r_flight_at_send; 4431 /* We split a piece the lower section looses any just_ret flag. */ 4432 nrsm->r_bbr_state = rsm->r_bbr_state; 4433 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4434 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4435 } 4436 rsm->r_end = nrsm->r_start; 4437 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 4438 idx /= 8; 4439 /* Check if we got too small */ 4440 if ((rsm->r_is_smallmap == 0) && 4441 ((rsm->r_end - rsm->r_start) <= idx)) { 4442 bbr->r_ctl.rc_num_small_maps_alloced++; 4443 rsm->r_is_smallmap = 1; 4444 } 4445 /* Check the new one as well */ 4446 if ((nrsm->r_end - nrsm->r_start) <= idx) { 4447 bbr->r_ctl.rc_num_small_maps_alloced++; 4448 nrsm->r_is_smallmap = 1; 4449 } 4450 } 4451 4452 static int 4453 bbr_sack_mergable(struct bbr_sendmap *at, 4454 uint32_t start, uint32_t end) 4455 { 4456 /* 4457 * Given a sack block defined by 4458 * start and end, and a current postion 4459 * at. Return 1 if either side of at 4460 * would show that the block is mergable 4461 * to that side. A block to be mergable 4462 * must have overlap with the start/end 4463 * and be in the SACK'd state. 4464 */ 4465 struct bbr_sendmap *l_rsm; 4466 struct bbr_sendmap *r_rsm; 4467 4468 /* first get the either side blocks */ 4469 l_rsm = TAILQ_PREV(at, bbr_head, r_next); 4470 r_rsm = TAILQ_NEXT(at, r_next); 4471 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { 4472 /* Potentially mergeable */ 4473 if ((l_rsm->r_end == start) || 4474 (SEQ_LT(start, l_rsm->r_end) && 4475 SEQ_GT(end, l_rsm->r_end))) { 4476 /* 4477 * map blk |------| 4478 * sack blk |------| 4479 * <or> 4480 * map blk |------| 4481 * sack blk |------| 4482 */ 4483 return (1); 4484 } 4485 } 4486 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { 4487 /* Potentially mergeable */ 4488 if ((r_rsm->r_start == end) || 4489 (SEQ_LT(start, r_rsm->r_start) && 4490 SEQ_GT(end, r_rsm->r_start))) { 4491 /* 4492 * map blk |---------| 4493 * sack blk |----| 4494 * <or> 4495 * map blk |---------| 4496 * sack blk |-------| 4497 */ 4498 return (1); 4499 } 4500 } 4501 return (0); 4502 } 4503 4504 static struct bbr_sendmap * 4505 bbr_merge_rsm(struct tcp_bbr *bbr, 4506 struct bbr_sendmap *l_rsm, 4507 struct bbr_sendmap *r_rsm) 4508 { 4509 /* 4510 * We are merging two ack'd RSM's, 4511 * the l_rsm is on the left (lower seq 4512 * values) and the r_rsm is on the right 4513 * (higher seq value). The simplest way 4514 * to merge these is to move the right 4515 * one into the left. I don't think there 4516 * is any reason we need to try to find 4517 * the oldest (or last oldest retransmitted). 4518 */ 4519 l_rsm->r_end = r_rsm->r_end; 4520 if (l_rsm->r_dupack < r_rsm->r_dupack) 4521 l_rsm->r_dupack = r_rsm->r_dupack; 4522 if (r_rsm->r_rtr_bytes) 4523 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 4524 if (r_rsm->r_in_tmap) { 4525 /* This really should not happen */ 4526 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); 4527 } 4528 if (r_rsm->r_app_limited) 4529 l_rsm->r_app_limited = r_rsm->r_app_limited; 4530 /* Now the flags */ 4531 if (r_rsm->r_flags & BBR_HAS_FIN) 4532 l_rsm->r_flags |= BBR_HAS_FIN; 4533 if (r_rsm->r_flags & BBR_TLP) 4534 l_rsm->r_flags |= BBR_TLP; 4535 if (r_rsm->r_flags & BBR_RWND_COLLAPSED) 4536 l_rsm->r_flags |= BBR_RWND_COLLAPSED; 4537 if (r_rsm->r_flags & BBR_MARKED_LOST) { 4538 /* This really should not happen */ 4539 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; 4540 } 4541 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); 4542 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 4543 /* Transfer the split limit to the map we free */ 4544 r_rsm->r_limit_type = l_rsm->r_limit_type; 4545 l_rsm->r_limit_type = 0; 4546 } 4547 bbr_free(bbr, r_rsm); 4548 return(l_rsm); 4549 } 4550 4551 /* 4552 * TLP Timer, here we simply setup what segment we want to 4553 * have the TLP expire on, the normal bbr_output_wtime() will then 4554 * send it out. 4555 * 4556 * We return 1, saying don't proceed with bbr_output_wtime only 4557 * when all timers have been stopped (destroyed PCB?). 4558 */ 4559 static int 4560 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4561 { 4562 /* 4563 * Tail Loss Probe. 4564 */ 4565 struct bbr_sendmap *rsm = NULL; 4566 struct socket *so; 4567 uint32_t amm; 4568 uint32_t out, avail; 4569 uint32_t maxseg; 4570 int collapsed_win = 0; 4571 4572 if (bbr->rc_all_timers_stopped) { 4573 return (1); 4574 } 4575 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4576 /* Its not time yet */ 4577 return (0); 4578 } 4579 if (ctf_progress_timeout_check(tp, true)) { 4580 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4581 return (-ETIMEDOUT); /* tcp_drop() */ 4582 } 4583 /* Did we somehow get into persists? */ 4584 if (bbr->rc_in_persist) { 4585 return (0); 4586 } 4587 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4588 bbr_set_state(tp, bbr, 0); 4589 BBR_STAT_INC(bbr_tlp_tot); 4590 maxseg = tp->t_maxseg - bbr->rc_last_options; 4591 /* 4592 * A TLP timer has expired. We have been idle for 2 rtts. So we now 4593 * need to figure out how to force a full MSS segment out. 4594 */ 4595 so = tp->t_inpcb->inp_socket; 4596 avail = sbavail(&so->so_snd); 4597 out = ctf_outstanding(tp); 4598 if (out > tp->snd_wnd) { 4599 /* special case, we need a retransmission */ 4600 collapsed_win = 1; 4601 goto need_retran; 4602 } 4603 if (avail > out) { 4604 /* New data is available */ 4605 amm = avail - out; 4606 if (amm > maxseg) { 4607 amm = maxseg; 4608 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 4609 /* not enough to fill a MTU and no-delay is off */ 4610 goto need_retran; 4611 } 4612 /* Set the send-new override */ 4613 if ((out + amm) <= tp->snd_wnd) { 4614 bbr->rc_tlp_new_data = 1; 4615 } else { 4616 goto need_retran; 4617 } 4618 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4619 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; 4620 bbr->r_ctl.rc_tlp_send = NULL; 4621 /* cap any slots */ 4622 BBR_STAT_INC(bbr_tlp_newdata); 4623 goto send; 4624 } 4625 need_retran: 4626 /* 4627 * Ok we need to arrange the last un-acked segment to be re-sent, or 4628 * optionally the first un-acked segment. 4629 */ 4630 if (collapsed_win == 0) { 4631 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 4632 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { 4633 rsm = bbr_find_high_nonack(bbr, rsm); 4634 } 4635 if (rsm == NULL) { 4636 goto restore; 4637 } 4638 } else { 4639 /* 4640 * We must find the last segment 4641 * that was acceptable by the client. 4642 */ 4643 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4644 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { 4645 /* Found one */ 4646 break; 4647 } 4648 } 4649 if (rsm == NULL) { 4650 /* None? if so send the first */ 4651 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4652 if (rsm == NULL) 4653 goto restore; 4654 } 4655 } 4656 if ((rsm->r_end - rsm->r_start) > maxseg) { 4657 /* 4658 * We need to split this the last segment in two. 4659 */ 4660 struct bbr_sendmap *nrsm; 4661 4662 nrsm = bbr_alloc_full_limit(bbr); 4663 if (nrsm == NULL) { 4664 /* 4665 * We can't get memory to split, we can either just 4666 * not split it. Or retransmit the whole piece, lets 4667 * do the large send (BTLP :-) ). 4668 */ 4669 goto go_for_it; 4670 } 4671 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); 4672 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 4673 if (rsm->r_in_tmap) { 4674 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4675 nrsm->r_in_tmap = 1; 4676 } 4677 rsm->r_flags &= (~BBR_HAS_FIN); 4678 rsm = nrsm; 4679 } 4680 go_for_it: 4681 bbr->r_ctl.rc_tlp_send = rsm; 4682 bbr->rc_tlp_rtx_out = 1; 4683 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { 4684 bbr->r_ctl.rc_tlp_seg_send_cnt++; 4685 tp->t_rxtshift++; 4686 } else { 4687 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 4688 bbr->r_ctl.rc_tlp_seg_send_cnt = 1; 4689 } 4690 send: 4691 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 4692 /* 4693 * Can't [re]/transmit a segment we have retranmitted the 4694 * max times. We need the retransmit timer to take over. 4695 */ 4696 restore: 4697 bbr->rc_tlp_new_data = 0; 4698 bbr->r_ctl.rc_tlp_send = NULL; 4699 if (rsm) 4700 rsm->r_flags &= ~BBR_TLP; 4701 BBR_STAT_INC(bbr_tlp_retran_fail); 4702 return (0); 4703 } else if (rsm) { 4704 rsm->r_flags |= BBR_TLP; 4705 } 4706 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && 4707 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { 4708 /* 4709 * We have retransmitted to many times for TLP. Switch to 4710 * the regular RTO timer 4711 */ 4712 goto restore; 4713 } 4714 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); 4715 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 4716 return (0); 4717 } 4718 4719 /* 4720 * Delayed ack Timer, here we simply need to setup the 4721 * ACK_NOW flag and remove the DELACK flag. From there 4722 * the output routine will send the ack out. 4723 * 4724 * We only return 1, saying don't proceed, if all timers 4725 * are stopped (destroyed PCB?). 4726 */ 4727 static int 4728 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4729 { 4730 if (bbr->rc_all_timers_stopped) { 4731 return (1); 4732 } 4733 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); 4734 tp->t_flags &= ~TF_DELACK; 4735 tp->t_flags |= TF_ACKNOW; 4736 KMOD_TCPSTAT_INC(tcps_delack); 4737 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 4738 return (0); 4739 } 4740 4741 /* 4742 * Here we send a KEEP-ALIVE like probe to the 4743 * peer, we do not send data. 4744 * 4745 * We only return 1, saying don't proceed, if all timers 4746 * are stopped (destroyed PCB?). 4747 */ 4748 static int 4749 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4750 { 4751 struct tcptemp *t_template; 4752 int32_t retval = 1; 4753 4754 if (bbr->rc_all_timers_stopped) { 4755 return (1); 4756 } 4757 if (bbr->rc_in_persist == 0) 4758 return (0); 4759 KASSERT(tp->t_inpcb != NULL, 4760 ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); 4761 /* 4762 * Persistence timer into zero window. Force a byte to be output, if 4763 * possible. 4764 */ 4765 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); 4766 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 4767 KMOD_TCPSTAT_INC(tcps_persisttimeo); 4768 /* 4769 * Have we exceeded the user specified progress time? 4770 */ 4771 if (ctf_progress_timeout_check(tp, true)) { 4772 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4773 return (-ETIMEDOUT); /* tcp_drop() */ 4774 } 4775 /* 4776 * Hack: if the peer is dead/unreachable, we do not time out if the 4777 * window is closed. After a full backoff, drop the connection if 4778 * the idle time (no responses to probes) reaches the maximum 4779 * backoff that we would use if retransmitting. 4780 */ 4781 if (tp->t_rxtshift == TCP_MAXRXTSHIFT && 4782 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 4783 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 4784 KMOD_TCPSTAT_INC(tcps_persistdrop); 4785 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4786 return (-ETIMEDOUT); /* tcp_drop() */ 4787 } 4788 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && 4789 tp->snd_una == tp->snd_max) { 4790 bbr_exit_persist(tp, bbr, cts, __LINE__); 4791 retval = 0; 4792 goto out; 4793 } 4794 /* 4795 * If the user has closed the socket then drop a persisting 4796 * connection after a much reduced timeout. 4797 */ 4798 if (tp->t_state > TCPS_CLOSE_WAIT && 4799 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 4800 KMOD_TCPSTAT_INC(tcps_persistdrop); 4801 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4802 return (-ETIMEDOUT); /* tcp_drop() */ 4803 } 4804 t_template = tcpip_maketemplate(bbr->rc_inp); 4805 if (t_template) { 4806 tcp_respond(tp, t_template->tt_ipgen, 4807 &t_template->tt_t, (struct mbuf *)NULL, 4808 tp->rcv_nxt, tp->snd_una - 1, 0); 4809 /* This sends an ack */ 4810 if (tp->t_flags & TF_DELACK) 4811 tp->t_flags &= ~TF_DELACK; 4812 free(t_template, M_TEMP); 4813 } 4814 if (tp->t_rxtshift < TCP_MAXRXTSHIFT) 4815 tp->t_rxtshift++; 4816 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); 4817 out: 4818 return (retval); 4819 } 4820 4821 /* 4822 * If a keepalive goes off, we had no other timers 4823 * happening. We always return 1 here since this 4824 * routine either drops the connection or sends 4825 * out a segment with respond. 4826 */ 4827 static int 4828 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4829 { 4830 struct tcptemp *t_template; 4831 struct inpcb *inp; 4832 4833 if (bbr->rc_all_timers_stopped) { 4834 return (1); 4835 } 4836 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 4837 inp = tp->t_inpcb; 4838 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); 4839 /* 4840 * Keep-alive timer went off; send something or drop connection if 4841 * idle for too long. 4842 */ 4843 KMOD_TCPSTAT_INC(tcps_keeptimeo); 4844 if (tp->t_state < TCPS_ESTABLISHED) 4845 goto dropit; 4846 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 4847 tp->t_state <= TCPS_CLOSING) { 4848 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 4849 goto dropit; 4850 /* 4851 * Send a packet designed to force a response if the peer is 4852 * up and reachable: either an ACK if the connection is 4853 * still alive, or an RST if the peer has closed the 4854 * connection due to timeout or reboot. Using sequence 4855 * number tp->snd_una-1 causes the transmitted zero-length 4856 * segment to lie outside the receive window; by the 4857 * protocol spec, this requires the correspondent TCP to 4858 * respond. 4859 */ 4860 KMOD_TCPSTAT_INC(tcps_keepprobe); 4861 t_template = tcpip_maketemplate(inp); 4862 if (t_template) { 4863 tcp_respond(tp, t_template->tt_ipgen, 4864 &t_template->tt_t, (struct mbuf *)NULL, 4865 tp->rcv_nxt, tp->snd_una - 1, 0); 4866 free(t_template, M_TEMP); 4867 } 4868 } 4869 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); 4870 return (1); 4871 dropit: 4872 KMOD_TCPSTAT_INC(tcps_keepdrops); 4873 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); 4874 return (-ETIMEDOUT); /* tcp_drop() */ 4875 } 4876 4877 /* 4878 * Retransmit helper function, clear up all the ack 4879 * flags and take care of important book keeping. 4880 */ 4881 static void 4882 bbr_remxt_tmr(struct tcpcb *tp) 4883 { 4884 /* 4885 * The retransmit timer went off, all sack'd blocks must be 4886 * un-acked. 4887 */ 4888 struct bbr_sendmap *rsm, *trsm = NULL; 4889 struct tcp_bbr *bbr; 4890 uint32_t cts, lost; 4891 4892 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 4893 cts = tcp_get_usecs(&bbr->rc_tv); 4894 lost = bbr->r_ctl.rc_lost; 4895 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4896 bbr_set_state(tp, bbr, 0); 4897 4898 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 4899 if (rsm->r_flags & BBR_ACKED) { 4900 uint32_t old_flags; 4901 4902 rsm->r_dupack = 0; 4903 if (rsm->r_in_tmap == 0) { 4904 /* We must re-add it back to the tlist */ 4905 if (trsm == NULL) { 4906 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 4907 } else { 4908 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); 4909 } 4910 rsm->r_in_tmap = 1; 4911 } 4912 old_flags = rsm->r_flags; 4913 rsm->r_flags |= BBR_RXT_CLEARED; 4914 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); 4915 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 4916 } else { 4917 if ((tp->t_state < TCPS_ESTABLISHED) && 4918 (rsm->r_start == tp->snd_una)) { 4919 /* 4920 * Special case for TCP FO. Where 4921 * we sent more data beyond the snd_max. 4922 * We don't mark that as lost and stop here. 4923 */ 4924 break; 4925 } 4926 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4927 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4928 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4929 } 4930 if (bbr_marks_rxt_sack_passed) { 4931 /* 4932 * With this option, we will rack out 4933 * in 1ms increments the rest of the packets. 4934 */ 4935 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; 4936 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4937 } else { 4938 /* 4939 * With this option we only mark them lost 4940 * and remove all sack'd markings. We will run 4941 * another RXT or a TLP. This will cause 4942 * us to eventually send more based on what 4943 * ack's come in. 4944 */ 4945 rsm->r_flags |= BBR_MARKED_LOST; 4946 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4947 rsm->r_flags &= ~BBR_SACK_PASSED; 4948 } 4949 } 4950 trsm = rsm; 4951 } 4952 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4953 /* Clear the count (we just un-acked them) */ 4954 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); 4955 bbr->rc_tlp_new_data = 0; 4956 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4957 /* zap the behindness on a rxt */ 4958 bbr->r_ctl.rc_hptsi_agg_delay = 0; 4959 bbr->r_agg_early_set = 0; 4960 bbr->r_ctl.rc_agg_early = 0; 4961 bbr->rc_tlp_rtx_out = 0; 4962 bbr->r_ctl.rc_sacked = 0; 4963 bbr->r_ctl.rc_sacklast = NULL; 4964 bbr->r_timer_override = 1; 4965 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4966 } 4967 4968 /* 4969 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 4970 * we will setup to retransmit the lowest seq number outstanding. 4971 */ 4972 static int 4973 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4974 { 4975 int32_t rexmt; 4976 int32_t retval = 0; 4977 bool isipv6; 4978 4979 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 4980 if (bbr->rc_all_timers_stopped) { 4981 return (1); 4982 } 4983 if (TCPS_HAVEESTABLISHED(tp->t_state) && 4984 (tp->snd_una == tp->snd_max)) { 4985 /* Nothing outstanding .. nothing to do */ 4986 return (0); 4987 } 4988 /* 4989 * Retransmission timer went off. Message has not been acked within 4990 * retransmit interval. Back off to a longer retransmit interval 4991 * and retransmit one segment. 4992 */ 4993 if (ctf_progress_timeout_check(tp, true)) { 4994 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4995 return (-ETIMEDOUT); /* tcp_drop() */ 4996 } 4997 bbr_remxt_tmr(tp); 4998 if ((bbr->r_ctl.rc_resend == NULL) || 4999 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { 5000 /* 5001 * If the rwnd collapsed on 5002 * the one we are retransmitting 5003 * it does not count against the 5004 * rxt count. 5005 */ 5006 tp->t_rxtshift++; 5007 } 5008 if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { 5009 tp->t_rxtshift = TCP_MAXRXTSHIFT; 5010 KMOD_TCPSTAT_INC(tcps_timeoutdrop); 5011 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); 5012 /* XXXGL: previously t_softerror was casted to uint16_t */ 5013 MPASS(tp->t_softerror >= 0); 5014 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; 5015 return (retval); /* tcp_drop() */ 5016 } 5017 if (tp->t_state == TCPS_SYN_SENT) { 5018 /* 5019 * If the SYN was retransmitted, indicate CWND to be limited 5020 * to 1 segment in cc_conn_init(). 5021 */ 5022 tp->snd_cwnd = 1; 5023 } else if (tp->t_rxtshift == 1) { 5024 /* 5025 * first retransmit; record ssthresh and cwnd so they can be 5026 * recovered if this turns out to be a "bad" retransmit. A 5027 * retransmit is considered "bad" if an ACK for this segment 5028 * is received within RTT/2 interval; the assumption here is 5029 * that the ACK was already in flight. See "On Estimating 5030 * End-to-End Network Path Properties" by Allman and Paxson 5031 * for more details. 5032 */ 5033 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5034 if (!IN_RECOVERY(tp->t_flags)) { 5035 tp->snd_cwnd_prev = tp->snd_cwnd; 5036 tp->snd_ssthresh_prev = tp->snd_ssthresh; 5037 tp->snd_recover_prev = tp->snd_recover; 5038 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 5039 tp->t_flags |= TF_PREVVALID; 5040 } else { 5041 tp->t_flags &= ~TF_PREVVALID; 5042 } 5043 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5044 } else { 5045 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5046 tp->t_flags &= ~TF_PREVVALID; 5047 } 5048 KMOD_TCPSTAT_INC(tcps_rexmttimeo); 5049 if ((tp->t_state == TCPS_SYN_SENT) || 5050 (tp->t_state == TCPS_SYN_RECEIVED)) 5051 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; 5052 else 5053 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 5054 TCPT_RANGESET(tp->t_rxtcur, rexmt, 5055 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), 5056 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 5057 /* 5058 * We enter the path for PLMTUD if connection is established or, if 5059 * connection is FIN_WAIT_1 status, reason for the last is that if 5060 * amount of data we send is very small, we could send it in couple 5061 * of packets and process straight to FIN. In that case we won't 5062 * catch ESTABLISHED state. 5063 */ 5064 #ifdef INET6 5065 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; 5066 #else 5067 isipv6 = false; 5068 #endif 5069 if (((V_tcp_pmtud_blackhole_detect == 1) || 5070 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || 5071 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && 5072 ((tp->t_state == TCPS_ESTABLISHED) || 5073 (tp->t_state == TCPS_FIN_WAIT_1))) { 5074 /* 5075 * Idea here is that at each stage of mtu probe (usually, 5076 * 1448 -> 1188 -> 524) should be given 2 chances to recover 5077 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 5078 * should take care of that. 5079 */ 5080 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 5081 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 5082 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 5083 tp->t_rxtshift % 2 == 0)) { 5084 /* 5085 * Enter Path MTU Black-hole Detection mechanism: - 5086 * Disable Path MTU Discovery (IP "DF" bit). - 5087 * Reduce MTU to lower value than what we negotiated 5088 * with peer. 5089 */ 5090 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 5091 /* 5092 * Record that we may have found a black 5093 * hole. 5094 */ 5095 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 5096 /* Keep track of previous MSS. */ 5097 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 5098 } 5099 /* 5100 * Reduce the MSS to blackhole value or to the 5101 * default in an attempt to retransmit. 5102 */ 5103 #ifdef INET6 5104 isipv6 = bbr->r_is_v6; 5105 if (isipv6 && 5106 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 5107 /* Use the sysctl tuneable blackhole MSS. */ 5108 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 5109 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5110 } else if (isipv6) { 5111 /* Use the default MSS. */ 5112 tp->t_maxseg = V_tcp_v6mssdflt; 5113 /* 5114 * Disable Path MTU Discovery when we switch 5115 * to minmss. 5116 */ 5117 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5118 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5119 } 5120 #endif 5121 #if defined(INET6) && defined(INET) 5122 else 5123 #endif 5124 #ifdef INET 5125 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 5126 /* Use the sysctl tuneable blackhole MSS. */ 5127 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 5128 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5129 } else { 5130 /* Use the default MSS. */ 5131 tp->t_maxseg = V_tcp_mssdflt; 5132 /* 5133 * Disable Path MTU Discovery when we switch 5134 * to minmss. 5135 */ 5136 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5137 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5138 } 5139 #endif 5140 } else { 5141 /* 5142 * If further retransmissions are still unsuccessful 5143 * with a lowered MTU, maybe this isn't a blackhole 5144 * and we restore the previous MSS and blackhole 5145 * detection flags. The limit '6' is determined by 5146 * giving each probe stage (1448, 1188, 524) 2 5147 * chances to recover. 5148 */ 5149 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 5150 (tp->t_rxtshift >= 6)) { 5151 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 5152 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 5153 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 5154 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); 5155 } 5156 } 5157 } 5158 /* 5159 * Disable RFC1323 and SACK if we haven't got any response to our 5160 * third SYN to work-around some broken terminal servers (most of 5161 * which have hopefully been retired) that have bad VJ header 5162 * compression code which trashes TCP segments containing 5163 * unknown-to-them TCP options. 5164 */ 5165 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 5166 (tp->t_rxtshift == 3)) 5167 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 5168 /* 5169 * If we backed off this far, our srtt estimate is probably bogus. 5170 * Clobber it so we'll take the next rtt measurement as our srtt; 5171 * move the current srtt into rttvar to keep the current retransmit 5172 * times until then. 5173 */ 5174 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 5175 #ifdef INET6 5176 if (bbr->r_is_v6) 5177 in6_losing(tp->t_inpcb); 5178 else 5179 #endif 5180 in_losing(tp->t_inpcb); 5181 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 5182 tp->t_srtt = 0; 5183 } 5184 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 5185 tp->snd_recover = tp->snd_max; 5186 tp->t_flags |= TF_ACKNOW; 5187 tp->t_rtttime = 0; 5188 5189 return (retval); 5190 } 5191 5192 static int 5193 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) 5194 { 5195 int32_t ret = 0; 5196 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 5197 5198 if (timers == 0) { 5199 return (0); 5200 } 5201 if (tp->t_state == TCPS_LISTEN) { 5202 /* no timers on listen sockets */ 5203 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 5204 return (0); 5205 return (1); 5206 } 5207 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 5208 uint32_t left; 5209 5210 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 5211 ret = -1; 5212 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5213 return (0); 5214 } 5215 if (hpts_calling == 0) { 5216 ret = -2; 5217 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5218 return (0); 5219 } 5220 /* 5221 * Ok our timer went off early and we are not paced false 5222 * alarm, go back to sleep. 5223 */ 5224 left = bbr->r_ctl.rc_timer_exp - cts; 5225 ret = -3; 5226 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); 5227 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left)); 5228 return (1); 5229 } 5230 bbr->rc_tmr_stopped = 0; 5231 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 5232 if (timers & PACE_TMR_DELACK) { 5233 ret = bbr_timeout_delack(tp, bbr, cts); 5234 } else if (timers & PACE_TMR_PERSIT) { 5235 ret = bbr_timeout_persist(tp, bbr, cts); 5236 } else if (timers & PACE_TMR_RACK) { 5237 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5238 ret = bbr_timeout_rack(tp, bbr, cts); 5239 } else if (timers & PACE_TMR_TLP) { 5240 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5241 ret = bbr_timeout_tlp(tp, bbr, cts); 5242 } else if (timers & PACE_TMR_RXT) { 5243 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5244 ret = bbr_timeout_rxt(tp, bbr, cts); 5245 } else if (timers & PACE_TMR_KEEP) { 5246 ret = bbr_timeout_keepalive(tp, bbr, cts); 5247 } 5248 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); 5249 return (ret); 5250 } 5251 5252 static void 5253 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) 5254 { 5255 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 5256 uint8_t hpts_removed = 0; 5257 5258 if (tcp_in_hpts(bbr->rc_inp) && 5259 (bbr->rc_timer_first == 1)) { 5260 /* 5261 * If we are canceling timer's when we have the 5262 * timer ahead of the output being paced. We also 5263 * must remove ourselves from the hpts. 5264 */ 5265 hpts_removed = 1; 5266 tcp_hpts_remove(bbr->rc_inp); 5267 if (bbr->r_ctl.rc_last_delay_val) { 5268 /* Update the last hptsi delay too */ 5269 uint32_t time_since_send; 5270 5271 if (TSTMP_GT(cts, bbr->rc_pacer_started)) 5272 time_since_send = cts - bbr->rc_pacer_started; 5273 else 5274 time_since_send = 0; 5275 if (bbr->r_ctl.rc_last_delay_val > time_since_send) { 5276 /* Cut down our slot time */ 5277 bbr->r_ctl.rc_last_delay_val -= time_since_send; 5278 } else { 5279 bbr->r_ctl.rc_last_delay_val = 0; 5280 } 5281 bbr->rc_pacer_started = cts; 5282 } 5283 } 5284 bbr->rc_timer_first = 0; 5285 bbr_log_to_cancel(bbr, line, cts, hpts_removed); 5286 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 5287 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 5288 } 5289 } 5290 5291 static void 5292 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type) 5293 { 5294 struct tcp_bbr *bbr; 5295 5296 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 5297 bbr->rc_all_timers_stopped = 1; 5298 return; 5299 } 5300 5301 /* 5302 * stop all timers always returning 0. 5303 */ 5304 static int 5305 bbr_stopall(struct tcpcb *tp) 5306 { 5307 return (0); 5308 } 5309 5310 static void 5311 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) 5312 { 5313 return; 5314 } 5315 5316 /* 5317 * return true if a bbr timer (rack or tlp) is active. 5318 */ 5319 static int 5320 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type) 5321 { 5322 return (0); 5323 } 5324 5325 static uint32_t 5326 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) 5327 { 5328 struct bbr_sendmap *rsm; 5329 5330 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 5331 if ((rsm == NULL) || (u_rsm == rsm)) 5332 return (cts); 5333 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 5334 } 5335 5336 static void 5337 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, 5338 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) 5339 { 5340 int32_t idx; 5341 5342 rsm->r_rtr_cnt++; 5343 rsm->r_dupack = 0; 5344 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { 5345 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; 5346 rsm->r_flags |= BBR_OVERMAX; 5347 } 5348 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 5349 /* Take off the collapsed flag at rxt */ 5350 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 5351 } 5352 if (rsm->r_flags & BBR_MARKED_LOST) { 5353 /* We have retransmitted, its no longer lost */ 5354 rsm->r_flags &= ~BBR_MARKED_LOST; 5355 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 5356 } 5357 if (rsm->r_flags & BBR_RXT_CLEARED) { 5358 /* 5359 * We hit a RXT timer on it and 5360 * we cleared the "acked" flag. 5361 * We now have it going back into 5362 * flight, we can remove the cleared 5363 * flag and possibly do accounting on 5364 * this piece. 5365 */ 5366 rsm->r_flags &= ~BBR_RXT_CLEARED; 5367 } 5368 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { 5369 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 5370 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 5371 } 5372 idx = rsm->r_rtr_cnt - 1; 5373 rsm->r_tim_lastsent[idx] = cts; 5374 rsm->r_pacing_delay = pacing_time; 5375 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5376 rsm->r_ts_valid = bbr->rc_ts_valid; 5377 if (bbr->rc_ts_valid) 5378 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5379 if (bbr->r_ctl.r_app_limited_until) 5380 rsm->r_app_limited = 1; 5381 else 5382 rsm->r_app_limited = 0; 5383 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5384 rsm->r_bbr_state = bbr_state_val(bbr); 5385 else 5386 rsm->r_bbr_state = 8; 5387 if (rsm->r_flags & BBR_ACKED) { 5388 /* Problably MTU discovery messing with us */ 5389 uint32_t old_flags; 5390 5391 old_flags = rsm->r_flags; 5392 rsm->r_flags &= ~BBR_ACKED; 5393 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 5394 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 5395 if (bbr->r_ctl.rc_sacked == 0) 5396 bbr->r_ctl.rc_sacklast = NULL; 5397 } 5398 if (rsm->r_in_tmap) { 5399 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5400 } 5401 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5402 rsm->r_in_tmap = 1; 5403 if (rsm->r_flags & BBR_SACK_PASSED) { 5404 /* We have retransmitted due to the SACK pass */ 5405 rsm->r_flags &= ~BBR_SACK_PASSED; 5406 rsm->r_flags |= BBR_WAS_SACKPASS; 5407 } 5408 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5409 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5410 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5411 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 5412 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5413 rsm->r_is_gain = 1; 5414 rsm->r_is_drain = 0; 5415 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5416 rsm->r_is_drain = 1; 5417 rsm->r_is_gain = 0; 5418 } else { 5419 rsm->r_is_drain = 0; 5420 rsm->r_is_gain = 0; 5421 } 5422 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ 5423 } 5424 5425 /* 5426 * Returns 0, or the sequence where we stopped 5427 * updating. We also update the lenp to be the amount 5428 * of data left. 5429 */ 5430 5431 static uint32_t 5432 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, 5433 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) 5434 { 5435 /* 5436 * We (re-)transmitted starting at rsm->r_start for some length 5437 * (possibly less than r_end. 5438 */ 5439 struct bbr_sendmap *nrsm; 5440 uint32_t c_end; 5441 int32_t len; 5442 5443 len = *lenp; 5444 c_end = rsm->r_start + len; 5445 if (SEQ_GEQ(c_end, rsm->r_end)) { 5446 /* 5447 * We retransmitted the whole piece or more than the whole 5448 * slopping into the next rsm. 5449 */ 5450 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5451 if (c_end == rsm->r_end) { 5452 *lenp = 0; 5453 return (0); 5454 } else { 5455 int32_t act_len; 5456 5457 /* Hangs over the end return whats left */ 5458 act_len = rsm->r_end - rsm->r_start; 5459 *lenp = (len - act_len); 5460 return (rsm->r_end); 5461 } 5462 /* We don't get out of this block. */ 5463 } 5464 /* 5465 * Here we retransmitted less than the whole thing which means we 5466 * have to split this into what was transmitted and what was not. 5467 */ 5468 nrsm = bbr_alloc_full_limit(bbr); 5469 if (nrsm == NULL) { 5470 *lenp = 0; 5471 return (0); 5472 } 5473 /* 5474 * So here we are going to take the original rsm and make it what we 5475 * retransmitted. nrsm will be the tail portion we did not 5476 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 5477 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 5478 * 1, 6 and the new piece will be 6, 11. 5479 */ 5480 bbr_clone_rsm(bbr, nrsm, rsm, c_end); 5481 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 5482 nrsm->r_dupack = 0; 5483 if (rsm->r_in_tmap) { 5484 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 5485 nrsm->r_in_tmap = 1; 5486 } 5487 rsm->r_flags &= (~BBR_HAS_FIN); 5488 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5489 *lenp = 0; 5490 return (0); 5491 } 5492 5493 static uint64_t 5494 bbr_get_hardware_rate(struct tcp_bbr *bbr) 5495 { 5496 uint64_t bw; 5497 5498 bw = bbr_get_bw(bbr); 5499 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; 5500 bw /= (uint64_t)BBR_UNIT; 5501 return(bw); 5502 } 5503 5504 static void 5505 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, 5506 uint64_t act_rate, uint64_t rate_wanted) 5507 { 5508 /* 5509 * We could not get a full gains worth 5510 * of rate. 5511 */ 5512 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { 5513 /* we can't even get the real rate */ 5514 uint64_t red; 5515 5516 bbr->skip_gain = 1; 5517 bbr->gain_is_limited = 0; 5518 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; 5519 if (red) 5520 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); 5521 } else { 5522 /* We can use a lower gain */ 5523 bbr->skip_gain = 0; 5524 bbr->gain_is_limited = 1; 5525 } 5526 } 5527 5528 static void 5529 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) 5530 { 5531 const struct tcp_hwrate_limit_table *nrte; 5532 int error, rate = -1; 5533 5534 if (bbr->r_ctl.crte == NULL) 5535 return; 5536 if ((bbr->rc_inp->inp_route.ro_nh == NULL) || 5537 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { 5538 /* Lost our routes? */ 5539 /* Clear the way for a re-attempt */ 5540 bbr->bbr_attempt_hdwr_pace = 0; 5541 lost_rate: 5542 bbr->gain_is_limited = 0; 5543 bbr->skip_gain = 0; 5544 bbr->bbr_hdrw_pacing = 0; 5545 counter_u64_add(bbr_flows_whdwr_pacing, -1); 5546 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 5547 tcp_bbr_tso_size_check(bbr, cts); 5548 return; 5549 } 5550 rate = bbr_get_hardware_rate(bbr); 5551 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, 5552 bbr->rc_tp, 5553 bbr->rc_inp->inp_route.ro_nh->nh_ifp, 5554 rate, 5555 (RS_PACING_GEQ|RS_PACING_SUB_OK), 5556 &error, NULL); 5557 if (nrte == NULL) { 5558 goto lost_rate; 5559 } 5560 if (nrte != bbr->r_ctl.crte) { 5561 bbr->r_ctl.crte = nrte; 5562 if (error == 0) { 5563 BBR_STAT_INC(bbr_hdwr_rl_mod_ok); 5564 if (bbr->r_ctl.crte->rate < rate) { 5565 /* We have a problem */ 5566 bbr_setup_less_of_rate(bbr, cts, 5567 bbr->r_ctl.crte->rate, rate); 5568 } else { 5569 /* We are good */ 5570 bbr->gain_is_limited = 0; 5571 bbr->skip_gain = 0; 5572 } 5573 } else { 5574 /* A failure should release the tag */ 5575 BBR_STAT_INC(bbr_hdwr_rl_mod_fail); 5576 bbr->gain_is_limited = 0; 5577 bbr->skip_gain = 0; 5578 bbr->bbr_hdrw_pacing = 0; 5579 } 5580 bbr_type_log_hdwr_pacing(bbr, 5581 bbr->r_ctl.crte->ptbl->rs_ifp, 5582 rate, 5583 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), 5584 __LINE__, 5585 cts, 5586 error); 5587 } 5588 } 5589 5590 static void 5591 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) 5592 { 5593 /* 5594 * If we have hardware pacing support 5595 * we need to factor that in for our 5596 * TSO size. 5597 */ 5598 const struct tcp_hwrate_limit_table *rlp; 5599 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; 5600 5601 if ((bbr->bbr_hdrw_pacing == 0) || 5602 (IN_RECOVERY(bbr->rc_tp->t_flags)) || 5603 (bbr->r_ctl.crte == NULL)) 5604 return; 5605 if (bbr->hw_pacing_set == 0) { 5606 /* Not yet by the hdwr pacing count delay */ 5607 return; 5608 } 5609 if (bbr_hdwr_pace_adjust == 0) { 5610 /* No adjustment */ 5611 return; 5612 } 5613 rlp = bbr->r_ctl.crte; 5614 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) 5615 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5616 else 5617 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5618 /* 5619 * So lets first get the 5620 * time we will take between 5621 * TSO sized sends currently without 5622 * hardware help. 5623 */ 5624 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, 5625 bbr->r_ctl.rc_pace_max_segs, cts, 1); 5626 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; 5627 hdwr_delay *= rlp->time_between; 5628 if (cur_delay > hdwr_delay) 5629 delta = cur_delay - hdwr_delay; 5630 else 5631 delta = 0; 5632 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, 5633 (bbr->r_ctl.rc_pace_max_segs / maxseg), 5634 1); 5635 if (delta && 5636 (delta < (max(rlp->time_between, 5637 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { 5638 /* 5639 * Now lets divide by the pacing 5640 * time between each segment the 5641 * hardware sends rounding up and 5642 * derive a bytes from that. We multiply 5643 * that by bbr_hdwr_pace_adjust to get 5644 * more bang for our buck. 5645 * 5646 * The goal is to have the software pacer 5647 * waiting no more than an additional 5648 * pacing delay if we can (without the 5649 * compensation i.e. x bbr_hdwr_pace_adjust). 5650 */ 5651 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), 5652 (bbr->r_ctl.rc_pace_max_segs/maxseg)); 5653 seg_sz *= bbr_hdwr_pace_adjust; 5654 if (bbr_hdwr_pace_floor && 5655 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5656 /* Currently hardware paces 5657 * out rs_min_seg segments at a time. 5658 * We need to make sure we always send at least 5659 * a full burst of bbr_hdwr_pace_floor down. 5660 */ 5661 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5662 } 5663 seg_sz *= maxseg; 5664 } else if (delta == 0) { 5665 /* 5666 * The highest pacing rate is 5667 * above our b/w gained. This means 5668 * we probably are going quite fast at 5669 * the hardware highest rate. Lets just multiply 5670 * the calculated TSO size by the 5671 * multiplier factor (its probably 5672 * 4 segments in the default config for 5673 * mlx). 5674 */ 5675 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; 5676 if (bbr_hdwr_pace_floor && 5677 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5678 /* Currently hardware paces 5679 * out rs_min_seg segments at a time. 5680 * We need to make sure we always send at least 5681 * a full burst of bbr_hdwr_pace_floor down. 5682 */ 5683 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5684 } 5685 } else { 5686 /* 5687 * The pacing time difference is so 5688 * big that the hardware will 5689 * pace out more rapidly then we 5690 * really want and then we 5691 * will have a long delay. Lets just keep 5692 * the same TSO size so its as if 5693 * we were not using hdwr pacing (we 5694 * just gain a bit of spacing from the 5695 * hardware if seg_sz > 1). 5696 */ 5697 seg_sz = bbr->r_ctl.rc_pace_max_segs; 5698 } 5699 if (seg_sz > bbr->r_ctl.rc_pace_max_segs) 5700 new_tso = seg_sz; 5701 else 5702 new_tso = bbr->r_ctl.rc_pace_max_segs; 5703 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) 5704 new_tso = PACE_MAX_IP_BYTES - maxseg; 5705 5706 if (new_tso != bbr->r_ctl.rc_pace_max_segs) { 5707 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); 5708 bbr->r_ctl.rc_pace_max_segs = new_tso; 5709 } 5710 } 5711 5712 static void 5713 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) 5714 { 5715 uint64_t bw; 5716 uint32_t old_tso = 0, new_tso; 5717 uint32_t maxseg, bytes; 5718 uint32_t tls_seg=0; 5719 /* 5720 * Google/linux uses the following algorithm to determine 5721 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): 5722 * 5723 * bytes = bw_in_bytes_per_second / 1000 5724 * bytes = min(bytes, 64k) 5725 * tso_segs = bytes / MSS 5726 * if (bw < 1.2Mbs) 5727 * min_tso_segs = 1 5728 * else 5729 * min_tso_segs = 2 5730 * tso_segs = max(tso_segs, min_tso_segs) 5731 * 5732 * * Note apply a device specific limit (we apply this in the 5733 * tcp_m_copym). 5734 * Note that before the initial measurement is made google bursts out 5735 * a full iwnd just like new-reno/cubic. 5736 * 5737 * We do not use this algorithm. Instead we 5738 * use a two phased approach: 5739 * 5740 * if ( bw <= per-tcb-cross-over) 5741 * goal_tso = calculate how much with this bw we 5742 * can send in goal-time seconds. 5743 * if (goal_tso > mss) 5744 * seg = goal_tso / mss 5745 * tso = seg * mss 5746 * else 5747 * tso = mss 5748 * if (tso > per-tcb-max) 5749 * tso = per-tcb-max 5750 * else if ( bw > 512Mbps) 5751 * tso = max-tso (64k/mss) 5752 * else 5753 * goal_tso = bw / per-tcb-divsor 5754 * seg = (goal_tso + mss-1)/mss 5755 * tso = seg * mss 5756 * 5757 * if (tso < per-tcb-floor) 5758 * tso = per-tcb-floor 5759 * if (tso > per-tcb-utter_max) 5760 * tso = per-tcb-utter_max 5761 * 5762 * Note the default per-tcb-divisor is 1000 (same as google). 5763 * the goal cross over is 30Mbps however. To recreate googles 5764 * algorithm you need to set: 5765 * 5766 * cross-over = 23,168,000 bps 5767 * goal-time = 18000 5768 * per-tcb-max = 2 5769 * per-tcb-divisor = 1000 5770 * per-tcb-floor = 1 5771 * 5772 * This will get you "google bbr" behavior with respect to tso size. 5773 * 5774 * Note we do set anything TSO size until we are past the initial 5775 * window. Before that we gnerally use either a single MSS 5776 * or we use the full IW size (so we burst a IW at a time) 5777 */ 5778 5779 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { 5780 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5781 } else { 5782 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5783 } 5784 old_tso = bbr->r_ctl.rc_pace_max_segs; 5785 if (bbr->rc_past_init_win == 0) { 5786 /* 5787 * Not enough data has been acknowledged to make a 5788 * judgement. Set up the initial TSO based on if we 5789 * are sending a full IW at once or not. 5790 */ 5791 if (bbr->rc_use_google) 5792 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); 5793 else if (bbr->bbr_init_win_cheat) 5794 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); 5795 else 5796 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5797 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) 5798 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; 5799 if (bbr->r_ctl.rc_pace_max_segs == 0) { 5800 bbr->r_ctl.rc_pace_max_segs = maxseg; 5801 } 5802 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); 5803 bbr_adjust_for_hw_pacing(bbr, cts); 5804 return; 5805 } 5806 /** 5807 * Now lets set the TSO goal based on our delivery rate in 5808 * bytes per second. Note we only do this if 5809 * we have acked at least the initial cwnd worth of data. 5810 */ 5811 bw = bbr_get_bw(bbr); 5812 if (IN_RECOVERY(bbr->rc_tp->t_flags) && 5813 (bbr->rc_use_google == 0)) { 5814 /* We clamp to one MSS in recovery */ 5815 new_tso = maxseg; 5816 } else if (bbr->rc_use_google) { 5817 int min_tso_segs; 5818 5819 /* Google considers the gain too */ 5820 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { 5821 bw *= bbr->r_ctl.rc_bbr_hptsi_gain; 5822 bw /= BBR_UNIT; 5823 } 5824 bytes = bw / 1024; 5825 if (bytes > (64 * 1024)) 5826 bytes = 64 * 1024; 5827 new_tso = bytes / maxseg; 5828 if (bw < ONE_POINT_TWO_MEG) 5829 min_tso_segs = 1; 5830 else 5831 min_tso_segs = 2; 5832 if (new_tso < min_tso_segs) 5833 new_tso = min_tso_segs; 5834 new_tso *= maxseg; 5835 } else if (bbr->rc_no_pacing) { 5836 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; 5837 } else if (bw <= bbr->r_ctl.bbr_cross_over) { 5838 /* 5839 * Calculate the worse case b/w TSO if we are inserting no 5840 * more than a delay_target number of TSO's. 5841 */ 5842 uint32_t tso_len, min_tso; 5843 5844 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); 5845 if (tso_len > maxseg) { 5846 new_tso = tso_len / maxseg; 5847 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) 5848 new_tso = bbr->r_ctl.bbr_hptsi_segments_max; 5849 new_tso *= maxseg; 5850 } else { 5851 /* 5852 * less than a full sized frame yikes.. long rtt or 5853 * low bw? 5854 */ 5855 min_tso = bbr_minseg(bbr); 5856 if ((tso_len > min_tso) && (bbr_all_get_min == 0)) 5857 new_tso = rounddown(tso_len, min_tso); 5858 else 5859 new_tso = min_tso; 5860 } 5861 } else if (bw > FIVETWELVE_MBPS) { 5862 /* 5863 * This guy is so fast b/w wise that we can TSO as large as 5864 * possible of segments that the NIC will allow. 5865 */ 5866 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5867 } else { 5868 /* 5869 * This formula is based on attempting to send a segment or 5870 * more every bbr_hptsi_per_second. The default is 1000 5871 * which means you are targeting what you can send every 1ms 5872 * based on the peers bw. 5873 * 5874 * If the number drops to say 500, then you are looking more 5875 * at 2ms and you will raise how much we send in a single 5876 * TSO thus saving CPU (less bbr_output_wtime() calls). The 5877 * trade off of course is you will send more at once and 5878 * thus tend to clump up the sends into larger "bursts" 5879 * building a queue. 5880 */ 5881 bw /= bbr->r_ctl.bbr_hptsi_per_second; 5882 new_tso = roundup(bw, (uint64_t)maxseg); 5883 /* 5884 * Gate the floor to match what our lower than 48Mbps 5885 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus 5886 * becomes the floor for this calculation. 5887 */ 5888 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) 5889 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); 5890 } 5891 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) 5892 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; 5893 if (new_tso > PACE_MAX_IP_BYTES) 5894 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5895 /* Enforce an utter maximum. */ 5896 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { 5897 new_tso = bbr->r_ctl.bbr_utter_max * maxseg; 5898 } 5899 if (old_tso != new_tso) { 5900 /* Only log changes */ 5901 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); 5902 bbr->r_ctl.rc_pace_max_segs = new_tso; 5903 } 5904 /* We have hardware pacing! */ 5905 bbr_adjust_for_hw_pacing(bbr, cts); 5906 } 5907 5908 static void 5909 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, 5910 uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, 5911 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, 5912 struct sockbuf *sb) 5913 { 5914 5915 struct bbr_sendmap *rsm, *nrsm; 5916 register uint32_t snd_max, snd_una; 5917 uint32_t pacing_time; 5918 /* 5919 * Add to the RACK log of packets in flight or retransmitted. If 5920 * there is a TS option we will use the TS echoed, if not we will 5921 * grab a TS. 5922 * 5923 * Retransmissions will increment the count and move the ts to its 5924 * proper place. Note that if options do not include TS's then we 5925 * won't be able to effectively use the ACK for an RTT on a retran. 5926 * 5927 * Notes about r_start and r_end. Lets consider a send starting at 5928 * sequence 1 for 10 bytes. In such an example the r_start would be 5929 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 5930 * This means that r_end is actually the first sequence for the next 5931 * slot (11). 5932 * 5933 */ 5934 INP_WLOCK_ASSERT(tp->t_inpcb); 5935 if (err) { 5936 /* 5937 * We don't log errors -- we could but snd_max does not 5938 * advance in this case either. 5939 */ 5940 return; 5941 } 5942 if (th_flags & TH_RST) { 5943 /* 5944 * We don't log resets and we return immediately from 5945 * sending 5946 */ 5947 *abandon = 1; 5948 return; 5949 } 5950 snd_una = tp->snd_una; 5951 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { 5952 /* 5953 * The call to bbr_log_output is made before bumping 5954 * snd_max. This means we can record one extra byte on a SYN 5955 * or FIN if seq_out is adding more on and a FIN is present 5956 * (and we are not resending). 5957 */ 5958 if ((th_flags & TH_SYN) && (tp->iss == seq_out)) 5959 len++; 5960 if (th_flags & TH_FIN) 5961 len++; 5962 } 5963 if (SEQ_LEQ((seq_out + len), snd_una)) { 5964 /* Are sending an old segment to induce an ack (keep-alive)? */ 5965 return; 5966 } 5967 if (SEQ_LT(seq_out, snd_una)) { 5968 /* huh? should we panic? */ 5969 uint32_t end; 5970 5971 end = seq_out + len; 5972 seq_out = snd_una; 5973 len = end - seq_out; 5974 } 5975 snd_max = tp->snd_max; 5976 if (len == 0) { 5977 /* We don't log zero window probes */ 5978 return; 5979 } 5980 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); 5981 /* First question is it a retransmission? */ 5982 if (seq_out == snd_max) { 5983 again: 5984 rsm = bbr_alloc(bbr); 5985 if (rsm == NULL) { 5986 return; 5987 } 5988 rsm->r_flags = 0; 5989 if (th_flags & TH_SYN) 5990 rsm->r_flags |= BBR_HAS_SYN; 5991 if (th_flags & TH_FIN) 5992 rsm->r_flags |= BBR_HAS_FIN; 5993 rsm->r_tim_lastsent[0] = cts; 5994 rsm->r_rtr_cnt = 1; 5995 rsm->r_rtr_bytes = 0; 5996 rsm->r_start = seq_out; 5997 rsm->r_end = rsm->r_start + len; 5998 rsm->r_dupack = 0; 5999 rsm->r_delivered = bbr->r_ctl.rc_delivered; 6000 rsm->r_pacing_delay = pacing_time; 6001 rsm->r_ts_valid = bbr->rc_ts_valid; 6002 if (bbr->rc_ts_valid) 6003 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 6004 rsm->r_del_time = bbr->r_ctl.rc_del_time; 6005 if (bbr->r_ctl.r_app_limited_until) 6006 rsm->r_app_limited = 1; 6007 else 6008 rsm->r_app_limited = 0; 6009 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 6010 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 6011 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 6012 /* 6013 * Here we must also add in this rsm since snd_max 6014 * is updated after we return from a new send. 6015 */ 6016 rsm->r_flight_at_send += len; 6017 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 6018 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 6019 rsm->r_in_tmap = 1; 6020 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 6021 rsm->r_bbr_state = bbr_state_val(bbr); 6022 else 6023 rsm->r_bbr_state = 8; 6024 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 6025 rsm->r_is_gain = 1; 6026 rsm->r_is_drain = 0; 6027 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 6028 rsm->r_is_drain = 1; 6029 rsm->r_is_gain = 0; 6030 } else { 6031 rsm->r_is_drain = 0; 6032 rsm->r_is_gain = 0; 6033 } 6034 return; 6035 } 6036 /* 6037 * If we reach here its a retransmission and we need to find it. 6038 */ 6039 more: 6040 if (hintrsm && (hintrsm->r_start == seq_out)) { 6041 rsm = hintrsm; 6042 hintrsm = NULL; 6043 } else if (bbr->r_ctl.rc_next) { 6044 /* We have a hint from a previous run */ 6045 rsm = bbr->r_ctl.rc_next; 6046 } else { 6047 /* No hints sorry */ 6048 rsm = NULL; 6049 } 6050 if ((rsm) && (rsm->r_start == seq_out)) { 6051 /* 6052 * We used rc_next or hintrsm to retransmit, hopefully the 6053 * likely case. 6054 */ 6055 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6056 if (len == 0) { 6057 return; 6058 } else { 6059 goto more; 6060 } 6061 } 6062 /* Ok it was not the last pointer go through it the hard way. */ 6063 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6064 if (rsm->r_start == seq_out) { 6065 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6066 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 6067 if (len == 0) { 6068 return; 6069 } else { 6070 continue; 6071 } 6072 } 6073 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 6074 /* Transmitted within this piece */ 6075 /* 6076 * Ok we must split off the front and then let the 6077 * update do the rest 6078 */ 6079 nrsm = bbr_alloc_full_limit(bbr); 6080 if (nrsm == NULL) { 6081 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 6082 return; 6083 } 6084 /* 6085 * copy rsm to nrsm and then trim the front of rsm 6086 * to not include this part. 6087 */ 6088 bbr_clone_rsm(bbr, nrsm, rsm, seq_out); 6089 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 6090 if (rsm->r_in_tmap) { 6091 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 6092 nrsm->r_in_tmap = 1; 6093 } 6094 rsm->r_flags &= (~BBR_HAS_FIN); 6095 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); 6096 if (len == 0) { 6097 return; 6098 } 6099 } 6100 } 6101 /* 6102 * Hmm not found in map did they retransmit both old and on into the 6103 * new? 6104 */ 6105 if (seq_out == tp->snd_max) { 6106 goto again; 6107 } else if (SEQ_LT(seq_out, tp->snd_max)) { 6108 #ifdef BBR_INVARIANTS 6109 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 6110 seq_out, len, tp->snd_una, tp->snd_max); 6111 printf("Starting Dump of all rack entries\n"); 6112 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6113 printf("rsm:%p start:%u end:%u\n", 6114 rsm, rsm->r_start, rsm->r_end); 6115 } 6116 printf("Dump complete\n"); 6117 panic("seq_out not found rack:%p tp:%p", 6118 bbr, tp); 6119 #endif 6120 } else { 6121 #ifdef BBR_INVARIANTS 6122 /* 6123 * Hmm beyond sndmax? (only if we are using the new rtt-pack 6124 * flag) 6125 */ 6126 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 6127 seq_out, len, tp->snd_max, tp); 6128 #endif 6129 } 6130 } 6131 6132 static void 6133 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) 6134 { 6135 /* 6136 * Collapse timeout back the cum-ack moved. 6137 */ 6138 tp->t_rxtshift = 0; 6139 tp->t_softerror = 0; 6140 } 6141 6142 static void 6143 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) 6144 { 6145 bbr->rtt_valid = 1; 6146 bbr->r_ctl.cur_rtt = rtt_usecs; 6147 bbr->r_ctl.ts_in = tsin; 6148 if (rsm_send_time) 6149 bbr->r_ctl.cur_rtt_send_time = rsm_send_time; 6150 } 6151 6152 static void 6153 bbr_make_timestamp_determination(struct tcp_bbr *bbr) 6154 { 6155 /** 6156 * We have in our bbr control: 6157 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). 6158 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). 6159 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) 6160 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) 6161 * 6162 * Now we can calculate the time between the sends by doing: 6163 * 6164 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts 6165 * 6166 * And the peer's time between receiving them by doing: 6167 * 6168 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp 6169 * 6170 * We want to figure out if the timestamp values are in msec, 10msec or usec. 6171 * We also may find that we can't use the timestamps if say we see 6172 * that the peer_delta indicates that though we may have taken 10ms to 6173 * pace out the data, it only saw 1ms between the two packets. This would 6174 * indicate that somewhere on the path is a batching entity that is giving 6175 * out time-slices of the actual b/w. This would mean we could not use 6176 * reliably the peers timestamps. 6177 * 6178 * We expect delta > peer_delta initially. Until we figure out the 6179 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. 6180 * If we place 1000 there then its a ms vs our usec. If we place 10000 there 6181 * then its 10ms vs our usec. If the peer is running a usec clock we would 6182 * put a 1 there. If the value is faster then ours, we will disable the 6183 * use of timestamps (though we could revist this later if we find it to be not 6184 * just an isolated one or two flows)). 6185 * 6186 * To detect the batching middle boxes we will come up with our compensation and 6187 * if with it in place, we find the peer is drastically off (by some margin) in 6188 * the smaller direction, then we will assume the worst case and disable use of timestamps. 6189 * 6190 */ 6191 uint64_t delta, peer_delta, delta_up; 6192 6193 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; 6194 if (delta < bbr_min_usec_delta) { 6195 /* 6196 * Have not seen a min amount of time 6197 * between our send times so we can 6198 * make a determination of the timestamp 6199 * yet. 6200 */ 6201 return; 6202 } 6203 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; 6204 if (peer_delta < bbr_min_peer_delta) { 6205 /* 6206 * We may have enough in the form of 6207 * our delta but the peers number 6208 * has not changed that much. It could 6209 * be its clock ratio is such that 6210 * we need more data (10ms tick) or 6211 * there may be other compression scenarios 6212 * going on. In any event we need the 6213 * spread to be larger. 6214 */ 6215 return; 6216 } 6217 /* Ok lets first see which way our delta is going */ 6218 if (peer_delta > delta) { 6219 /* Very unlikely, the peer without 6220 * compensation shows that it saw 6221 * the two sends arrive further apart 6222 * then we saw then in micro-seconds. 6223 */ 6224 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { 6225 /* well it looks like the peer is a micro-second clock. */ 6226 bbr->rc_ts_clock_set = 1; 6227 bbr->r_ctl.bbr_peer_tsratio = 1; 6228 } else { 6229 bbr->rc_ts_cant_be_used = 1; 6230 bbr->rc_ts_clock_set = 1; 6231 } 6232 return; 6233 } 6234 /* Ok we know that the peer_delta is smaller than our send distance */ 6235 bbr->rc_ts_clock_set = 1; 6236 /* First question is it within the percentage that they are using usec time? */ 6237 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; 6238 if ((peer_delta + delta_up) >= delta) { 6239 /* Its a usec clock */ 6240 bbr->r_ctl.bbr_peer_tsratio = 1; 6241 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6242 return; 6243 } 6244 /* Ok if not usec, what about 10usec (though unlikely)? */ 6245 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; 6246 if (((peer_delta * 10) + delta_up) >= delta) { 6247 bbr->r_ctl.bbr_peer_tsratio = 10; 6248 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6249 return; 6250 } 6251 /* And what about 100usec (though again unlikely)? */ 6252 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; 6253 if (((peer_delta * 100) + delta_up) >= delta) { 6254 bbr->r_ctl.bbr_peer_tsratio = 100; 6255 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6256 return; 6257 } 6258 /* And how about 1 msec (the most likely one)? */ 6259 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; 6260 if (((peer_delta * 1000) + delta_up) >= delta) { 6261 bbr->r_ctl.bbr_peer_tsratio = 1000; 6262 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6263 return; 6264 } 6265 /* Ok if not msec could it be 10 msec? */ 6266 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; 6267 if (((peer_delta * 10000) + delta_up) >= delta) { 6268 bbr->r_ctl.bbr_peer_tsratio = 10000; 6269 return; 6270 } 6271 /* If we fall down here the clock tick so slowly we can't use it */ 6272 bbr->rc_ts_cant_be_used = 1; 6273 bbr->r_ctl.bbr_peer_tsratio = 0; 6274 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6275 } 6276 6277 /* 6278 * Collect new round-trip time estimate 6279 * and update averages and current timeout. 6280 */ 6281 static void 6282 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) 6283 { 6284 int32_t delta; 6285 uint32_t rtt, tsin; 6286 int32_t rtt_ticks; 6287 6288 if (bbr->rtt_valid == 0) 6289 /* No valid sample */ 6290 return; 6291 6292 rtt = bbr->r_ctl.cur_rtt; 6293 tsin = bbr->r_ctl.ts_in; 6294 if (bbr->rc_prtt_set_ts) { 6295 /* 6296 * We are to force feed the rttProp filter due 6297 * to an entry into PROBE_RTT. This assures 6298 * that the times are sync'd between when we 6299 * go into PROBE_RTT and the filter expiration. 6300 * 6301 * Google does not use a true filter, so they do 6302 * this implicitly since they only keep one value 6303 * and when they enter probe-rtt they update the 6304 * value to the newest rtt. 6305 */ 6306 uint32_t rtt_prop; 6307 6308 bbr->rc_prtt_set_ts = 0; 6309 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 6310 if (rtt > rtt_prop) 6311 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); 6312 else 6313 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6314 } 6315 if (bbr->rc_ack_was_delayed) 6316 rtt += bbr->r_ctl.rc_ack_hdwr_delay; 6317 6318 if (rtt < bbr->r_ctl.rc_lowest_rtt) 6319 bbr->r_ctl.rc_lowest_rtt = rtt; 6320 bbr_log_rtt_sample(bbr, rtt, tsin); 6321 if (bbr->r_init_rtt) { 6322 /* 6323 * The initial rtt is not-trusted, nuke it and lets get 6324 * our first valid measurement in. 6325 */ 6326 bbr->r_init_rtt = 0; 6327 tp->t_srtt = 0; 6328 } 6329 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { 6330 /* 6331 * So we have not yet figured out 6332 * what the peers TSTMP value is 6333 * in (most likely ms). We need a 6334 * series of cum-ack's to determine 6335 * this reliably. 6336 */ 6337 if (bbr->rc_ack_is_cumack) { 6338 if (bbr->rc_ts_data_set) { 6339 /* Lets attempt to determine the timestamp granularity. */ 6340 bbr_make_timestamp_determination(bbr); 6341 } else { 6342 bbr->rc_ts_data_set = 1; 6343 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; 6344 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; 6345 } 6346 } else { 6347 /* 6348 * We have to have consecutive acks 6349 * reset any "filled" state to none. 6350 */ 6351 bbr->rc_ts_data_set = 0; 6352 } 6353 } 6354 /* Round it up */ 6355 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); 6356 if (rtt_ticks == 0) 6357 rtt_ticks = 1; 6358 if (tp->t_srtt != 0) { 6359 /* 6360 * srtt is stored as fixed point with 5 bits after the 6361 * binary point (i.e., scaled by 8). The following magic is 6362 * equivalent to the smoothing algorithm in rfc793 with an 6363 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 6364 * Adjust rtt to origin 0. 6365 */ 6366 6367 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) 6368 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 6369 6370 tp->t_srtt += delta; 6371 if (tp->t_srtt <= 0) 6372 tp->t_srtt = 1; 6373 6374 /* 6375 * We accumulate a smoothed rtt variance (actually, a 6376 * smoothed mean difference), then set the retransmit timer 6377 * to smoothed rtt + 4 times the smoothed variance. rttvar 6378 * is stored as fixed point with 4 bits after the binary 6379 * point (scaled by 16). The following is equivalent to 6380 * rfc793 smoothing with an alpha of .75 (rttvar = 6381 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 6382 * wired-in beta. 6383 */ 6384 if (delta < 0) 6385 delta = -delta; 6386 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 6387 tp->t_rttvar += delta; 6388 if (tp->t_rttvar <= 0) 6389 tp->t_rttvar = 1; 6390 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) 6391 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 6392 } else { 6393 /* 6394 * No rtt measurement yet - use the unsmoothed rtt. Set the 6395 * variance to half the rtt (so our first retransmit happens 6396 * at 3*rtt). 6397 */ 6398 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; 6399 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); 6400 tp->t_rttbest = tp->t_srtt + tp->t_rttvar; 6401 } 6402 KMOD_TCPSTAT_INC(tcps_rttupdated); 6403 tp->t_rttupdated++; 6404 #ifdef STATS 6405 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); 6406 #endif 6407 /* 6408 * the retransmit should happen at rtt + 4 * rttvar. Because of the 6409 * way we do the smoothing, srtt and rttvar will each average +1/2 6410 * tick of bias. When we compute the retransmit timer, we want 1/2 6411 * tick of rounding and 1 extra tick because of +-1/2 tick 6412 * uncertainty in the firing of the timer. The bias will give us 6413 * exactly the 1.5 tick we need. But, because the bias is 6414 * statistical, we have to test that we don't drop below the minimum 6415 * feasible timer (which is 2 ticks). 6416 */ 6417 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 6418 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), 6419 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 6420 6421 /* 6422 * We received an ack for a packet that wasn't retransmitted; it is 6423 * probably safe to discard any error indications we've received 6424 * recently. This isn't quite right, but close enough for now (a 6425 * route might have failed after we sent a segment, and the return 6426 * path might not be symmetrical). 6427 */ 6428 tp->t_softerror = 0; 6429 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 6430 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) 6431 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; 6432 } 6433 6434 static void 6435 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) 6436 { 6437 bbr->r_ctl.rc_rtt_shrinks = cts; 6438 if (bbr_can_force_probertt && 6439 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 6440 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 6441 /* 6442 * We should enter probe-rtt its been too long 6443 * since we have been there. 6444 */ 6445 bbr_enter_probe_rtt(bbr, cts, __LINE__); 6446 } else 6447 bbr_check_probe_rtt_limits(bbr, cts); 6448 } 6449 6450 static void 6451 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) 6452 { 6453 uint64_t orig_bw; 6454 6455 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { 6456 /* We never apply a zero measurement */ 6457 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 6458 0, 0, 0, 0, 0, 0); 6459 return; 6460 } 6461 if (bbr->r_ctl.r_measurement_count < 0xffffffff) 6462 bbr->r_ctl.r_measurement_count++; 6463 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 6464 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); 6465 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, 6466 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 6467 0, 0, 0, 0, 0, 0); 6468 if (orig_bw && 6469 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { 6470 if (bbr->bbr_hdrw_pacing) { 6471 /* 6472 * Apply a new rate to the hardware 6473 * possibly. 6474 */ 6475 bbr_update_hardware_pacing_rate(bbr, cts); 6476 } 6477 bbr_set_state_target(bbr, __LINE__); 6478 tcp_bbr_tso_size_check(bbr, cts); 6479 if (bbr->r_recovery_bw) { 6480 bbr_setup_red_bw(bbr, cts); 6481 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); 6482 } 6483 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) 6484 tcp_bbr_tso_size_check(bbr, cts); 6485 } 6486 6487 static void 6488 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6489 { 6490 if (bbr->rc_in_persist == 0) { 6491 /* We log only when not in persist */ 6492 /* Translate to a Bytes Per Second */ 6493 uint64_t tim, bw, ts_diff, ts_bw; 6494 uint32_t delivered; 6495 6496 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6497 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6498 else 6499 tim = 1; 6500 /* 6501 * Now that we have processed the tim (skipping the sample 6502 * or possibly updating the time, go ahead and 6503 * calculate the cdr. 6504 */ 6505 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6506 bw = (uint64_t)delivered; 6507 bw *= (uint64_t)USECS_IN_SECOND; 6508 bw /= tim; 6509 if (bw == 0) { 6510 /* We must have a calculatable amount */ 6511 return; 6512 } 6513 /* 6514 * If we are using this b/w shove it in now so we 6515 * can see in the trace viewer if it gets over-ridden. 6516 */ 6517 if (rsm->r_ts_valid && 6518 bbr->rc_ts_valid && 6519 bbr->rc_ts_clock_set && 6520 (bbr->rc_ts_cant_be_used == 0) && 6521 bbr->rc_use_ts_limit) { 6522 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); 6523 ts_diff *= bbr->r_ctl.bbr_peer_tsratio; 6524 if ((delivered == 0) || 6525 (rtt < 1000)) { 6526 /* Can't use the ts */ 6527 bbr_log_type_bbrupd(bbr, 61, cts, 6528 ts_diff, 6529 bbr->r_ctl.last_inbound_ts, 6530 rsm->r_del_ack_ts, 0, 6531 0, 0, 0, delivered); 6532 } else { 6533 ts_bw = (uint64_t)delivered; 6534 ts_bw *= (uint64_t)USECS_IN_SECOND; 6535 ts_bw /= ts_diff; 6536 bbr_log_type_bbrupd(bbr, 62, cts, 6537 (ts_bw >> 32), 6538 (ts_bw & 0xffffffff), 0, 0, 6539 0, 0, ts_diff, delivered); 6540 if ((bbr->ts_can_raise) && 6541 (ts_bw > bw)) { 6542 bbr_log_type_bbrupd(bbr, 8, cts, 6543 delivered, 6544 ts_diff, 6545 (bw >> 32), 6546 (bw & 0x00000000ffffffff), 6547 0, 0, 0, 0); 6548 bw = ts_bw; 6549 } else if (ts_bw && (ts_bw < bw)) { 6550 bbr_log_type_bbrupd(bbr, 7, cts, 6551 delivered, 6552 ts_diff, 6553 (bw >> 32), 6554 (bw & 0x00000000ffffffff), 6555 0, 0, 0, 0); 6556 bw = ts_bw; 6557 } 6558 } 6559 } 6560 if (rsm->r_first_sent_time && 6561 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6562 uint64_t sbw, sti; 6563 /* 6564 * We use what was in flight at the time of our 6565 * send and the size of this send to figure 6566 * out what we have been sending at (amount). 6567 * For the time we take from the time of 6568 * the send of the first send outstanding 6569 * until this send plus this sends pacing 6570 * time. This gives us a good calculation 6571 * as to the rate we have been sending at. 6572 */ 6573 6574 sbw = (uint64_t)(rsm->r_flight_at_send); 6575 sbw *= (uint64_t)USECS_IN_SECOND; 6576 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6577 sti += rsm->r_pacing_delay; 6578 sbw /= sti; 6579 if (sbw < bw) { 6580 bbr_log_type_bbrupd(bbr, 6, cts, 6581 delivered, 6582 (uint32_t)sti, 6583 (bw >> 32), 6584 (uint32_t)bw, 6585 rsm->r_first_sent_time, 0, (sbw >> 32), 6586 (uint32_t)sbw); 6587 bw = sbw; 6588 } 6589 } 6590 /* Use the google algorithm for b/w measurements */ 6591 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6592 if ((rsm->r_app_limited == 0) || 6593 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { 6594 tcp_bbr_commit_bw(bbr, cts); 6595 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6596 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6597 } 6598 } 6599 } 6600 6601 static void 6602 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6603 { 6604 if (bbr->rc_in_persist == 0) { 6605 /* We log only when not in persist */ 6606 /* Translate to a Bytes Per Second */ 6607 uint64_t tim, bw; 6608 uint32_t delivered; 6609 int no_apply = 0; 6610 6611 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6612 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6613 else 6614 tim = 1; 6615 /* 6616 * Now that we have processed the tim (skipping the sample 6617 * or possibly updating the time, go ahead and 6618 * calculate the cdr. 6619 */ 6620 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6621 bw = (uint64_t)delivered; 6622 bw *= (uint64_t)USECS_IN_SECOND; 6623 bw /= tim; 6624 if (tim < bbr->r_ctl.rc_lowest_rtt) { 6625 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6626 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6627 6628 no_apply = 1; 6629 } 6630 /* 6631 * If we are using this b/w shove it in now so we 6632 * can see in the trace viewer if it gets over-ridden. 6633 */ 6634 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6635 /* Gate by the sending rate */ 6636 if (rsm->r_first_sent_time && 6637 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6638 uint64_t sbw, sti; 6639 /* 6640 * We use what was in flight at the time of our 6641 * send and the size of this send to figure 6642 * out what we have been sending at (amount). 6643 * For the time we take from the time of 6644 * the send of the first send outstanding 6645 * until this send plus this sends pacing 6646 * time. This gives us a good calculation 6647 * as to the rate we have been sending at. 6648 */ 6649 6650 sbw = (uint64_t)(rsm->r_flight_at_send); 6651 sbw *= (uint64_t)USECS_IN_SECOND; 6652 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6653 sti += rsm->r_pacing_delay; 6654 sbw /= sti; 6655 if (sbw < bw) { 6656 bbr_log_type_bbrupd(bbr, 6, cts, 6657 delivered, 6658 (uint32_t)sti, 6659 (bw >> 32), 6660 (uint32_t)bw, 6661 rsm->r_first_sent_time, 0, (sbw >> 32), 6662 (uint32_t)sbw); 6663 bw = sbw; 6664 } 6665 if ((sti > tim) && 6666 (sti < bbr->r_ctl.rc_lowest_rtt)) { 6667 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6668 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6669 no_apply = 1; 6670 } else 6671 no_apply = 0; 6672 } 6673 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6674 if ((no_apply == 0) && 6675 ((rsm->r_app_limited == 0) || 6676 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { 6677 tcp_bbr_commit_bw(bbr, cts); 6678 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6679 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6680 } 6681 } 6682 } 6683 6684 static void 6685 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, 6686 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) 6687 { 6688 uint64_t old_rttprop; 6689 6690 /* Update our delivery time and amount */ 6691 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); 6692 bbr->r_ctl.rc_del_time = cts; 6693 if (rtt == 0) { 6694 /* 6695 * 0 means its a retransmit, for now we don't use these for 6696 * the rest of BBR. 6697 */ 6698 return; 6699 } 6700 if ((bbr->rc_use_google == 0) && 6701 (match != BBR_RTT_BY_EXACTMATCH) && 6702 (match != BBR_RTT_BY_TIMESTAMP)){ 6703 /* 6704 * We get a lot of rtt updates, lets not pay attention to 6705 * any that are not an exact match. That way we don't have 6706 * to worry about timestamps and the whole nonsense of 6707 * unsure if its a retransmission etc (if we ever had the 6708 * timestamp fixed to always have the last thing sent this 6709 * would not be a issue). 6710 */ 6711 return; 6712 } 6713 if ((bbr_no_retran && bbr->rc_use_google) && 6714 (match != BBR_RTT_BY_EXACTMATCH) && 6715 (match != BBR_RTT_BY_TIMESTAMP)){ 6716 /* 6717 * We only do measurements in google mode 6718 * with bbr_no_retran on for sure things. 6719 */ 6720 return; 6721 } 6722 /* Only update srtt if we know by exact match */ 6723 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); 6724 if (ack_type == BBR_CUM_ACKED) 6725 bbr->rc_ack_is_cumack = 1; 6726 else 6727 bbr->rc_ack_is_cumack = 0; 6728 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); 6729 /* 6730 * Note the following code differs to the original 6731 * BBR spec. It calls for <= not <. However after a 6732 * long discussion in email with Neal, he acknowledged 6733 * that it should be < than so that we will have flows 6734 * going into probe-rtt (we were seeing cases where that 6735 * did not happen and caused ugly things to occur). We 6736 * have added this agreed upon fix to our code base. 6737 */ 6738 if (rtt < old_rttprop) { 6739 /* Update when we last saw a rtt drop */ 6740 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); 6741 bbr_set_reduced_rtt(bbr, cts, __LINE__); 6742 } 6743 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, 6744 match, rsm->r_start, rsm->r_flags); 6745 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6746 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { 6747 /* 6748 * The RTT-prop moved, reset the target (may be a 6749 * nop for some states). 6750 */ 6751 bbr_set_state_target(bbr, __LINE__); 6752 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) 6753 bbr_log_rtt_shrinks(bbr, cts, 0, 0, 6754 __LINE__, BBR_RTTS_NEW_TARGET, 0); 6755 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) 6756 /* It went up */ 6757 bbr_check_probe_rtt_limits(bbr, cts); 6758 } 6759 if ((bbr->rc_use_google == 0) && 6760 (match == BBR_RTT_BY_TIMESTAMP)) { 6761 /* 6762 * We don't do b/w update with 6763 * these since they are not really 6764 * reliable. 6765 */ 6766 return; 6767 } 6768 if (bbr->r_ctl.r_app_limited_until && 6769 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { 6770 /* We are no longer app-limited */ 6771 bbr->r_ctl.r_app_limited_until = 0; 6772 } 6773 if (bbr->rc_use_google) { 6774 bbr_google_measurement(bbr, rsm, rtt, cts); 6775 } else { 6776 bbr_nf_measurement(bbr, rsm, rtt, cts); 6777 } 6778 } 6779 6780 /* 6781 * Convert a timestamp that the main stack 6782 * uses (milliseconds) into one that bbr uses 6783 * (microseconds). Return that converted timestamp. 6784 */ 6785 static uint32_t 6786 bbr_ts_convert(uint32_t cts) { 6787 uint32_t sec, msec; 6788 6789 sec = cts / MS_IN_USEC; 6790 msec = cts - (MS_IN_USEC * sec); 6791 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); 6792 } 6793 6794 /* 6795 * Return 0 if we did not update the RTT time, return 6796 * 1 if we did. 6797 */ 6798 static int 6799 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, 6800 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) 6801 { 6802 int32_t i; 6803 uint32_t t, uts = 0; 6804 6805 if ((rsm->r_flags & BBR_ACKED) || 6806 (rsm->r_flags & BBR_WAS_RENEGED) || 6807 (rsm->r_flags & BBR_RXT_CLEARED)) { 6808 /* Already done */ 6809 return (0); 6810 } 6811 if (rsm->r_rtt_not_allowed) { 6812 /* Not allowed */ 6813 return (0); 6814 } 6815 if (rsm->r_rtr_cnt == 1) { 6816 /* 6817 * Only one transmit. Hopefully the normal case. 6818 */ 6819 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) 6820 t = cts - rsm->r_tim_lastsent[0]; 6821 else 6822 t = 1; 6823 if ((int)t <= 0) 6824 t = 1; 6825 bbr->r_ctl.rc_last_rtt = t; 6826 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6827 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); 6828 return (1); 6829 } 6830 /* Convert to usecs */ 6831 if ((bbr_can_use_ts_for_rtt == 1) && 6832 (bbr->rc_use_google == 1) && 6833 (ack_type == BBR_CUM_ACKED) && 6834 (to->to_flags & TOF_TS) && 6835 (to->to_tsecr != 0)) { 6836 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; 6837 if (t < 1) 6838 t = 1; 6839 t *= MS_IN_USEC; 6840 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6841 BBR_RTT_BY_TIMESTAMP, 6842 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], 6843 ack_type, to); 6844 return (1); 6845 } 6846 uts = bbr_ts_convert(to->to_tsecr); 6847 if ((to->to_flags & TOF_TS) && 6848 (to->to_tsecr != 0) && 6849 (ack_type == BBR_CUM_ACKED) && 6850 ((rsm->r_flags & BBR_OVERMAX) == 0)) { 6851 /* 6852 * Now which timestamp does it match? In this block the ACK 6853 * may be coming from a previous transmission. 6854 */ 6855 uint32_t fudge; 6856 6857 fudge = BBR_TIMER_FUDGE; 6858 for (i = 0; i < rsm->r_rtr_cnt; i++) { 6859 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && 6860 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { 6861 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6862 t = cts - rsm->r_tim_lastsent[i]; 6863 else 6864 t = 1; 6865 if ((int)t <= 0) 6866 t = 1; 6867 bbr->r_ctl.rc_last_rtt = t; 6868 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, 6869 rsm->r_tim_lastsent[i], ack_type, to); 6870 if ((i + 1) < rsm->r_rtr_cnt) { 6871 /* Likely */ 6872 return (0); 6873 } else if (rsm->r_flags & BBR_TLP) { 6874 bbr->rc_tlp_rtx_out = 0; 6875 } 6876 return (1); 6877 } 6878 } 6879 /* Fall through if we can't find a matching timestamp */ 6880 } 6881 /* 6882 * Ok its a SACK block that we retransmitted. or a windows 6883 * machine without timestamps. We can tell nothing from the 6884 * time-stamp since its not there or the time the peer last 6885 * recieved a segment that moved forward its cum-ack point. 6886 * 6887 * Lets look at the last retransmit and see what we can tell 6888 * (with BBR for space we only keep 2 note we have to keep 6889 * at least 2 so the map can not be condensed more). 6890 */ 6891 i = rsm->r_rtr_cnt - 1; 6892 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6893 t = cts - rsm->r_tim_lastsent[i]; 6894 else 6895 goto not_sure; 6896 if (t < bbr->r_ctl.rc_lowest_rtt) { 6897 /* 6898 * We retransmitted and the ack came back in less 6899 * than the smallest rtt we have observed in the 6900 * windowed rtt. We most likey did an improper 6901 * retransmit as outlined in 4.2 Step 3 point 2 in 6902 * the rack-draft. 6903 * 6904 * Use the prior transmission to update all the 6905 * information as long as there is only one prior 6906 * transmission. 6907 */ 6908 if ((rsm->r_flags & BBR_OVERMAX) == 0) { 6909 #ifdef BBR_INVARIANTS 6910 if (rsm->r_rtr_cnt == 1) 6911 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); 6912 #endif 6913 i = rsm->r_rtr_cnt - 2; 6914 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6915 t = cts - rsm->r_tim_lastsent[i]; 6916 else 6917 t = 1; 6918 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, 6919 rsm->r_tim_lastsent[i], ack_type, to); 6920 return (0); 6921 } else { 6922 /* 6923 * Too many prior transmissions, just 6924 * updated BBR delivered 6925 */ 6926 not_sure: 6927 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6928 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6929 } 6930 } else { 6931 /* 6932 * We retransmitted it and the retransmit did the 6933 * job. 6934 */ 6935 if (rsm->r_flags & BBR_TLP) 6936 bbr->rc_tlp_rtx_out = 0; 6937 if ((rsm->r_flags & BBR_OVERMAX) == 0) 6938 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, 6939 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); 6940 else 6941 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6942 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6943 return (1); 6944 } 6945 return (0); 6946 } 6947 6948 /* 6949 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 6950 */ 6951 static void 6952 bbr_log_sack_passed(struct tcpcb *tp, 6953 struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 6954 { 6955 struct bbr_sendmap *nrsm; 6956 6957 nrsm = rsm; 6958 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, 6959 bbr_head, r_tnext) { 6960 if (nrsm == rsm) { 6961 /* Skip orginal segment he is acked */ 6962 continue; 6963 } 6964 if (nrsm->r_flags & BBR_ACKED) { 6965 /* Skip ack'd segments */ 6966 continue; 6967 } 6968 if (nrsm->r_flags & BBR_SACK_PASSED) { 6969 /* 6970 * We found one that is already marked 6971 * passed, we have been here before and 6972 * so all others below this are marked. 6973 */ 6974 break; 6975 } 6976 BBR_STAT_INC(bbr_sack_passed); 6977 nrsm->r_flags |= BBR_SACK_PASSED; 6978 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && 6979 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { 6980 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; 6981 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; 6982 nrsm->r_flags |= BBR_MARKED_LOST; 6983 } 6984 nrsm->r_flags &= ~BBR_WAS_SACKPASS; 6985 } 6986 } 6987 6988 /* 6989 * Returns the number of bytes that were 6990 * newly ack'd by sack blocks. 6991 */ 6992 static uint32_t 6993 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, 6994 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) 6995 { 6996 int32_t times = 0; 6997 uint32_t start, end, changed = 0; 6998 struct bbr_sendmap *rsm, *nrsm; 6999 int32_t used_ref = 1; 7000 uint8_t went_back = 0, went_fwd = 0; 7001 7002 start = sack->start; 7003 end = sack->end; 7004 rsm = *prsm; 7005 if (rsm == NULL) 7006 used_ref = 0; 7007 7008 /* Do we locate the block behind where we last were? */ 7009 if (rsm && SEQ_LT(start, rsm->r_start)) { 7010 went_back = 1; 7011 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 7012 if (SEQ_GEQ(start, rsm->r_start) && 7013 SEQ_LT(start, rsm->r_end)) { 7014 goto do_rest_ofb; 7015 } 7016 } 7017 } 7018 start_at_beginning: 7019 went_fwd = 1; 7020 /* 7021 * Ok lets locate the block where this guy is fwd from rsm (if its 7022 * set) 7023 */ 7024 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { 7025 if (SEQ_GEQ(start, rsm->r_start) && 7026 SEQ_LT(start, rsm->r_end)) { 7027 break; 7028 } 7029 } 7030 do_rest_ofb: 7031 if (rsm == NULL) { 7032 /* 7033 * This happens when we get duplicate sack blocks with the 7034 * same end. For example SACK 4: 100 SACK 3: 100 The sort 7035 * will not change there location so we would just start at 7036 * the end of the first one and get lost. 7037 */ 7038 if (tp->t_flags & TF_SENTFIN) { 7039 /* 7040 * Check to see if we have not logged the FIN that 7041 * went out. 7042 */ 7043 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7044 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 7045 /* 7046 * Ok we did not get the FIN logged. 7047 */ 7048 nrsm->r_end++; 7049 rsm = nrsm; 7050 goto do_rest_ofb; 7051 } 7052 } 7053 if (times == 1) { 7054 #ifdef BBR_INVARIANTS 7055 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", 7056 tp, bbr, sack, to, prsm); 7057 #else 7058 goto out; 7059 #endif 7060 } 7061 times++; 7062 BBR_STAT_INC(bbr_sack_proc_restart); 7063 rsm = NULL; 7064 goto start_at_beginning; 7065 } 7066 /* Ok we have an ACK for some piece of rsm */ 7067 if (rsm->r_start != start) { 7068 /* 7069 * Need to split this in two pieces the before and after. 7070 */ 7071 if (bbr_sack_mergable(rsm, start, end)) 7072 nrsm = bbr_alloc_full_limit(bbr); 7073 else 7074 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7075 if (nrsm == NULL) { 7076 /* We could not allocate ignore the sack */ 7077 struct sackblk blk; 7078 7079 blk.start = start; 7080 blk.end = end; 7081 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7082 goto out; 7083 } 7084 bbr_clone_rsm(bbr, nrsm, rsm, start); 7085 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7086 if (rsm->r_in_tmap) { 7087 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7088 nrsm->r_in_tmap = 1; 7089 } 7090 rsm->r_flags &= (~BBR_HAS_FIN); 7091 rsm = nrsm; 7092 } 7093 if (SEQ_GEQ(end, rsm->r_end)) { 7094 /* 7095 * The end of this block is either beyond this guy or right 7096 * at this guy. 7097 */ 7098 if ((rsm->r_flags & BBR_ACKED) == 0) { 7099 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7100 changed += (rsm->r_end - rsm->r_start); 7101 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7102 bbr_log_sack_passed(tp, bbr, rsm); 7103 if (rsm->r_flags & BBR_MARKED_LOST) { 7104 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7105 } 7106 /* Is Reordering occuring? */ 7107 if (rsm->r_flags & BBR_SACK_PASSED) { 7108 BBR_STAT_INC(bbr_reorder_seen); 7109 bbr->r_ctl.rc_reorder_ts = cts; 7110 if (rsm->r_flags & BBR_MARKED_LOST) { 7111 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7112 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7113 /* LT sampling also needs adjustment */ 7114 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7115 } 7116 } 7117 rsm->r_flags |= BBR_ACKED; 7118 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7119 if (rsm->r_in_tmap) { 7120 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7121 rsm->r_in_tmap = 0; 7122 } 7123 } 7124 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7125 if (end == rsm->r_end) { 7126 /* This block only - done */ 7127 goto out; 7128 } 7129 /* There is more not coverend by this rsm move on */ 7130 start = rsm->r_end; 7131 nrsm = TAILQ_NEXT(rsm, r_next); 7132 rsm = nrsm; 7133 times = 0; 7134 goto do_rest_ofb; 7135 } 7136 if (rsm->r_flags & BBR_ACKED) { 7137 /* Been here done that */ 7138 goto out; 7139 } 7140 /* Ok we need to split off this one at the tail */ 7141 if (bbr_sack_mergable(rsm, start, end)) 7142 nrsm = bbr_alloc_full_limit(bbr); 7143 else 7144 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7145 if (nrsm == NULL) { 7146 /* failed XXXrrs what can we do but loose the sack info? */ 7147 struct sackblk blk; 7148 7149 blk.start = start; 7150 blk.end = end; 7151 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7152 goto out; 7153 } 7154 /* Clone it */ 7155 bbr_clone_rsm(bbr, nrsm, rsm, end); 7156 /* The sack block does not cover this guy fully */ 7157 rsm->r_flags &= (~BBR_HAS_FIN); 7158 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7159 if (rsm->r_in_tmap) { 7160 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7161 nrsm->r_in_tmap = 1; 7162 } 7163 nrsm->r_dupack = 0; 7164 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7165 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7166 changed += (rsm->r_end - rsm->r_start); 7167 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7168 bbr_log_sack_passed(tp, bbr, rsm); 7169 /* Is Reordering occuring? */ 7170 if (rsm->r_flags & BBR_MARKED_LOST) { 7171 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7172 } 7173 if (rsm->r_flags & BBR_SACK_PASSED) { 7174 BBR_STAT_INC(bbr_reorder_seen); 7175 bbr->r_ctl.rc_reorder_ts = cts; 7176 if (rsm->r_flags & BBR_MARKED_LOST) { 7177 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7178 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7179 /* LT sampling also needs adjustment */ 7180 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7181 } 7182 } 7183 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7184 rsm->r_flags |= BBR_ACKED; 7185 if (rsm->r_in_tmap) { 7186 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7187 rsm->r_in_tmap = 0; 7188 } 7189 out: 7190 if (rsm && (rsm->r_flags & BBR_ACKED)) { 7191 /* 7192 * Now can we merge this newly acked 7193 * block with either the previous or 7194 * next block? 7195 */ 7196 nrsm = TAILQ_NEXT(rsm, r_next); 7197 if (nrsm && 7198 (nrsm->r_flags & BBR_ACKED)) { 7199 /* yep this and next can be merged */ 7200 rsm = bbr_merge_rsm(bbr, rsm, nrsm); 7201 } 7202 /* Now what about the previous? */ 7203 nrsm = TAILQ_PREV(rsm, bbr_head, r_next); 7204 if (nrsm && 7205 (nrsm->r_flags & BBR_ACKED)) { 7206 /* yep the previous and this can be merged */ 7207 rsm = bbr_merge_rsm(bbr, nrsm, rsm); 7208 } 7209 } 7210 if (used_ref == 0) { 7211 BBR_STAT_INC(bbr_sack_proc_all); 7212 } else { 7213 BBR_STAT_INC(bbr_sack_proc_short); 7214 } 7215 if (went_fwd && went_back) { 7216 BBR_STAT_INC(bbr_sack_search_both); 7217 } else if (went_fwd) { 7218 BBR_STAT_INC(bbr_sack_search_fwd); 7219 } else if (went_back) { 7220 BBR_STAT_INC(bbr_sack_search_back); 7221 } 7222 /* Save off where the next seq is */ 7223 if (rsm) 7224 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 7225 else 7226 bbr->r_ctl.rc_sacklast = NULL; 7227 *prsm = rsm; 7228 return (changed); 7229 } 7230 7231 static void inline 7232 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) 7233 { 7234 struct bbr_sendmap *tmap; 7235 7236 BBR_STAT_INC(bbr_reneges_seen); 7237 tmap = NULL; 7238 while (rsm && (rsm->r_flags & BBR_ACKED)) { 7239 /* Its no longer sacked, mark it so */ 7240 uint32_t oflags; 7241 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7242 #ifdef BBR_INVARIANTS 7243 if (rsm->r_in_tmap) { 7244 panic("bbr:%p rsm:%p flags:0x%x in tmap?", 7245 bbr, rsm, rsm->r_flags); 7246 } 7247 #endif 7248 oflags = rsm->r_flags; 7249 if (rsm->r_flags & BBR_MARKED_LOST) { 7250 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7251 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7252 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7253 /* LT sampling also needs adjustment */ 7254 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7255 } 7256 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); 7257 rsm->r_flags |= BBR_WAS_RENEGED; 7258 rsm->r_flags |= BBR_RXT_CLEARED; 7259 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); 7260 /* Rebuild it into our tmap */ 7261 if (tmap == NULL) { 7262 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7263 tmap = rsm; 7264 } else { 7265 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); 7266 tmap = rsm; 7267 } 7268 tmap->r_in_tmap = 1; 7269 /* 7270 * XXXrrs Delivered? Should we do anything here? 7271 * 7272 * Of course we don't on a rxt timeout so maybe its ok that 7273 * we don't? 7274 * 7275 * For now lets not. 7276 */ 7277 rsm = TAILQ_NEXT(rsm, r_next); 7278 } 7279 /* 7280 * Now lets possibly clear the sack filter so we start recognizing 7281 * sacks that cover this area. 7282 */ 7283 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); 7284 } 7285 7286 static void 7287 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) 7288 { 7289 struct tcp_bbr *bbr; 7290 struct bbr_sendmap *rsm; 7291 uint32_t cts; 7292 7293 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7294 cts = bbr->r_ctl.rc_rcvtime; 7295 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7296 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { 7297 if ((rsm->r_end - rsm->r_start) <= 1) { 7298 /* Log out the SYN completely */ 7299 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7300 rsm->r_rtr_bytes = 0; 7301 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7302 if (rsm->r_in_tmap) { 7303 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7304 rsm->r_in_tmap = 0; 7305 } 7306 if (bbr->r_ctl.rc_next == rsm) { 7307 /* scoot along the marker */ 7308 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7309 } 7310 if (to != NULL) 7311 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); 7312 bbr_free(bbr, rsm); 7313 } else { 7314 /* There is more (Fast open)? strip out SYN. */ 7315 rsm->r_flags &= ~BBR_HAS_SYN; 7316 rsm->r_start++; 7317 } 7318 } 7319 } 7320 7321 /* 7322 * Returns the number of bytes that were 7323 * acknowledged by SACK blocks. 7324 */ 7325 7326 static uint32_t 7327 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, 7328 uint32_t *prev_acked) 7329 { 7330 uint32_t changed, last_seq, entered_recovery = 0; 7331 struct tcp_bbr *bbr; 7332 struct bbr_sendmap *rsm; 7333 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 7334 register uint32_t th_ack; 7335 int32_t i, j, k, new_sb, num_sack_blks = 0; 7336 uint32_t cts, acked, ack_point, sack_changed = 0; 7337 uint32_t p_maxseg, maxseg, p_acked = 0; 7338 7339 INP_WLOCK_ASSERT(tp->t_inpcb); 7340 if (tcp_get_flags(th) & TH_RST) { 7341 /* We don't log resets */ 7342 return (0); 7343 } 7344 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7345 cts = bbr->r_ctl.rc_rcvtime; 7346 7347 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7348 changed = 0; 7349 maxseg = tp->t_maxseg - bbr->rc_last_options; 7350 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); 7351 th_ack = th->th_ack; 7352 if (SEQ_GT(th_ack, tp->snd_una)) { 7353 acked = th_ack - tp->snd_una; 7354 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); 7355 bbr->rc_tp->t_acktime = ticks; 7356 } else 7357 acked = 0; 7358 if (SEQ_LEQ(th_ack, tp->snd_una)) { 7359 /* Only sent here for sack processing */ 7360 goto proc_sack; 7361 } 7362 if (rsm && SEQ_GT(th_ack, rsm->r_start)) { 7363 changed = th_ack - rsm->r_start; 7364 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { 7365 /* 7366 * For the SYN incoming case we will not have called 7367 * tcp_output for the sending of the SYN, so there will be 7368 * no map. All other cases should probably be a panic. 7369 */ 7370 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { 7371 /* 7372 * We have a timestamp that can be used to generate 7373 * an initial RTT. 7374 */ 7375 uint32_t ts, now, rtt; 7376 7377 ts = bbr_ts_convert(to->to_tsecr); 7378 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); 7379 rtt = now - ts; 7380 if (rtt < 1) 7381 rtt = 1; 7382 bbr_log_type_bbrrttprop(bbr, rtt, 7383 tp->iss, 0, cts, 7384 BBR_RTT_BY_TIMESTAMP, tp->iss, 0); 7385 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 7386 changed = 1; 7387 bbr->r_wanted_output = 1; 7388 goto out; 7389 } 7390 goto proc_sack; 7391 } else if (rsm == NULL) { 7392 goto out; 7393 } 7394 if (changed) { 7395 /* 7396 * The ACK point is advancing to th_ack, we must drop off 7397 * the packets in the rack log and calculate any eligble 7398 * RTT's. 7399 */ 7400 bbr->r_wanted_output = 1; 7401 more: 7402 if (rsm == NULL) { 7403 if (tp->t_flags & TF_SENTFIN) { 7404 /* if we send a FIN we will not hav a map */ 7405 goto proc_sack; 7406 } 7407 #ifdef BBR_INVARIANTS 7408 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", 7409 tp, 7410 th, tp->t_state, bbr, 7411 tp->snd_una, tp->snd_max, changed); 7412 #endif 7413 goto proc_sack; 7414 } 7415 } 7416 if (SEQ_LT(th_ack, rsm->r_start)) { 7417 /* Huh map is missing this */ 7418 #ifdef BBR_INVARIANTS 7419 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", 7420 rsm->r_start, 7421 th_ack, tp->t_state, 7422 bbr->r_state, bbr); 7423 panic("th-ack is bad bbr:%p tp:%p", bbr, tp); 7424 #endif 7425 goto proc_sack; 7426 } else if (th_ack == rsm->r_start) { 7427 /* None here to ack */ 7428 goto proc_sack; 7429 } 7430 /* 7431 * Clear the dup ack counter, it will 7432 * either be freed or if there is some 7433 * remaining we need to start it at zero. 7434 */ 7435 rsm->r_dupack = 0; 7436 /* Now do we consume the whole thing? */ 7437 if (SEQ_GEQ(th_ack, rsm->r_end)) { 7438 /* Its all consumed. */ 7439 uint32_t left; 7440 7441 if (rsm->r_flags & BBR_ACKED) { 7442 /* 7443 * It was acked on the scoreboard -- remove it from 7444 * total 7445 */ 7446 p_acked += (rsm->r_end - rsm->r_start); 7447 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7448 if (bbr->r_ctl.rc_sacked == 0) 7449 bbr->r_ctl.rc_sacklast = NULL; 7450 } else { 7451 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); 7452 if (rsm->r_flags & BBR_MARKED_LOST) { 7453 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7454 } 7455 if (rsm->r_flags & BBR_SACK_PASSED) { 7456 /* 7457 * There are acked segments ACKED on the 7458 * scoreboard further up. We are seeing 7459 * reordering. 7460 */ 7461 BBR_STAT_INC(bbr_reorder_seen); 7462 bbr->r_ctl.rc_reorder_ts = cts; 7463 if (rsm->r_flags & BBR_MARKED_LOST) { 7464 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7465 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7466 /* LT sampling also needs adjustment */ 7467 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7468 } 7469 } 7470 rsm->r_flags &= ~BBR_MARKED_LOST; 7471 } 7472 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7473 rsm->r_rtr_bytes = 0; 7474 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7475 if (rsm->r_in_tmap) { 7476 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7477 rsm->r_in_tmap = 0; 7478 } 7479 if (bbr->r_ctl.rc_next == rsm) { 7480 /* scoot along the marker */ 7481 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7482 } 7483 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7484 /* Adjust the packet counts */ 7485 left = th_ack - rsm->r_end; 7486 /* Free back to zone */ 7487 bbr_free(bbr, rsm); 7488 if (left) { 7489 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7490 goto more; 7491 } 7492 goto proc_sack; 7493 } 7494 if (rsm->r_flags & BBR_ACKED) { 7495 /* 7496 * It was acked on the scoreboard -- remove it from total 7497 * for the part being cum-acked. 7498 */ 7499 p_acked += (rsm->r_end - rsm->r_start); 7500 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 7501 if (bbr->r_ctl.rc_sacked == 0) 7502 bbr->r_ctl.rc_sacklast = NULL; 7503 } else { 7504 /* 7505 * It was acked up to th_ack point for the first time 7506 */ 7507 struct bbr_sendmap lrsm; 7508 7509 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); 7510 lrsm.r_end = th_ack; 7511 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); 7512 } 7513 if ((rsm->r_flags & BBR_MARKED_LOST) && 7514 ((rsm->r_flags & BBR_ACKED) == 0)) { 7515 /* 7516 * It was marked lost and partly ack'd now 7517 * for the first time. We lower the rc_lost_bytes 7518 * and still leave it MARKED. 7519 */ 7520 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; 7521 } 7522 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7523 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7524 rsm->r_rtr_bytes = 0; 7525 /* adjust packet count */ 7526 rsm->r_start = th_ack; 7527 proc_sack: 7528 /* Check for reneging */ 7529 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7530 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { 7531 /* 7532 * The peer has moved snd_una up to the edge of this send, 7533 * i.e. one that it had previously acked. The only way that 7534 * can be true if the peer threw away data (space issues) 7535 * that it had previously sacked (else it would have given 7536 * us snd_una up to (rsm->r_end). We need to undo the acked 7537 * markings here. 7538 * 7539 * Note we have to look to make sure th_ack is our 7540 * rsm->r_start in case we get an old ack where th_ack is 7541 * behind snd_una. 7542 */ 7543 bbr_peer_reneges(bbr, rsm, th->th_ack); 7544 } 7545 if ((to->to_flags & TOF_SACK) == 0) { 7546 /* We are done nothing left to log */ 7547 goto out; 7548 } 7549 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7550 if (rsm) { 7551 last_seq = rsm->r_end; 7552 } else { 7553 last_seq = tp->snd_max; 7554 } 7555 /* Sack block processing */ 7556 if (SEQ_GT(th_ack, tp->snd_una)) 7557 ack_point = th_ack; 7558 else 7559 ack_point = tp->snd_una; 7560 for (i = 0; i < to->to_nsacks; i++) { 7561 bcopy((to->to_sacks + i * TCPOLEN_SACK), 7562 &sack, sizeof(sack)); 7563 sack.start = ntohl(sack.start); 7564 sack.end = ntohl(sack.end); 7565 if (SEQ_GT(sack.end, sack.start) && 7566 SEQ_GT(sack.start, ack_point) && 7567 SEQ_LT(sack.start, tp->snd_max) && 7568 SEQ_GT(sack.end, ack_point) && 7569 SEQ_LEQ(sack.end, tp->snd_max)) { 7570 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && 7571 (SEQ_LT(sack.end, last_seq)) && 7572 ((sack.end - sack.start) < (p_maxseg / 8))) { 7573 /* 7574 * Not the last piece and its smaller than 7575 * 1/8th of a p_maxseg. We ignore this. 7576 */ 7577 BBR_STAT_INC(bbr_runt_sacks); 7578 continue; 7579 } 7580 sack_blocks[num_sack_blks] = sack; 7581 num_sack_blks++; 7582 } else if (SEQ_LEQ(sack.start, th_ack) && 7583 SEQ_LEQ(sack.end, th_ack)) { 7584 /* 7585 * Its a D-SACK block. 7586 */ 7587 tcp_record_dsack(tp, sack.start, sack.end, 0); 7588 } 7589 } 7590 if (num_sack_blks == 0) 7591 goto out; 7592 /* 7593 * Sort the SACK blocks so we can update the rack scoreboard with 7594 * just one pass. 7595 */ 7596 new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, 7597 num_sack_blks, th->th_ack); 7598 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); 7599 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); 7600 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); 7601 num_sack_blks = new_sb; 7602 if (num_sack_blks < 2) { 7603 goto do_sack_work; 7604 } 7605 /* Sort the sacks */ 7606 for (i = 0; i < num_sack_blks; i++) { 7607 for (j = i + 1; j < num_sack_blks; j++) { 7608 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 7609 sack = sack_blocks[i]; 7610 sack_blocks[i] = sack_blocks[j]; 7611 sack_blocks[j] = sack; 7612 } 7613 } 7614 } 7615 /* 7616 * Now are any of the sack block ends the same (yes some 7617 * implememtations send these)? 7618 */ 7619 again: 7620 if (num_sack_blks > 1) { 7621 for (i = 0; i < num_sack_blks; i++) { 7622 for (j = i + 1; j < num_sack_blks; j++) { 7623 if (sack_blocks[i].end == sack_blocks[j].end) { 7624 /* 7625 * Ok these two have the same end we 7626 * want the smallest end and then 7627 * throw away the larger and start 7628 * again. 7629 */ 7630 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 7631 /* 7632 * The second block covers 7633 * more area use that 7634 */ 7635 sack_blocks[i].start = sack_blocks[j].start; 7636 } 7637 /* 7638 * Now collapse out the dup-sack and 7639 * lower the count 7640 */ 7641 for (k = (j + 1); k < num_sack_blks; k++) { 7642 sack_blocks[j].start = sack_blocks[k].start; 7643 sack_blocks[j].end = sack_blocks[k].end; 7644 j++; 7645 } 7646 num_sack_blks--; 7647 goto again; 7648 } 7649 } 7650 } 7651 } 7652 do_sack_work: 7653 rsm = bbr->r_ctl.rc_sacklast; 7654 for (i = 0; i < num_sack_blks; i++) { 7655 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); 7656 if (acked) { 7657 bbr->r_wanted_output = 1; 7658 changed += acked; 7659 sack_changed += acked; 7660 } 7661 } 7662 out: 7663 *prev_acked = p_acked; 7664 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 7665 /* 7666 * Ok we have a high probability that we need to go in to 7667 * recovery since we have data sack'd 7668 */ 7669 struct bbr_sendmap *rsm; 7670 7671 rsm = bbr_check_recovery_mode(tp, bbr, cts); 7672 if (rsm) { 7673 /* Enter recovery */ 7674 entered_recovery = 1; 7675 bbr->r_wanted_output = 1; 7676 /* 7677 * When we enter recovery we need to assure we send 7678 * one packet. 7679 */ 7680 if (bbr->r_ctl.rc_resend == NULL) { 7681 bbr->r_ctl.rc_resend = rsm; 7682 } 7683 } 7684 } 7685 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 7686 /* 7687 * See if we need to rack-retransmit anything if so set it 7688 * up as the thing to resend assuming something else is not 7689 * already in that position. 7690 */ 7691 if (bbr->r_ctl.rc_resend == NULL) { 7692 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 7693 } 7694 } 7695 /* 7696 * We return the amount that changed via sack, this is used by the 7697 * ack-received code to augment what was changed between th_ack <-> 7698 * snd_una. 7699 */ 7700 return (sack_changed); 7701 } 7702 7703 static void 7704 bbr_strike_dupack(struct tcp_bbr *bbr) 7705 { 7706 struct bbr_sendmap *rsm; 7707 7708 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 7709 if (rsm && (rsm->r_dupack < 0xff)) { 7710 rsm->r_dupack++; 7711 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) 7712 bbr->r_wanted_output = 1; 7713 } 7714 } 7715 7716 /* 7717 * Return value of 1, we do not need to call bbr_process_data(). 7718 * return value of 0, bbr_process_data can be called. 7719 * For ret_val if its 0 the TCB is locked and valid, if its non-zero 7720 * its unlocked and probably unsafe to touch the TCB. 7721 */ 7722 static int 7723 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 7724 struct tcpcb *tp, struct tcpopt *to, 7725 uint32_t tiwin, int32_t tlen, 7726 int32_t * ofia, int32_t thflags, int32_t * ret_val) 7727 { 7728 int32_t ourfinisacked = 0; 7729 int32_t acked_amount; 7730 uint16_t nsegs; 7731 int32_t acked; 7732 uint32_t lost, sack_changed = 0; 7733 struct mbuf *mfree; 7734 struct tcp_bbr *bbr; 7735 uint32_t prev_acked = 0; 7736 7737 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7738 lost = bbr->r_ctl.rc_lost; 7739 nsegs = max(1, m->m_pkthdr.lro_nsegs); 7740 if (SEQ_GT(th->th_ack, tp->snd_max)) { 7741 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 7742 bbr->r_wanted_output = 1; 7743 return (1); 7744 } 7745 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 7746 /* Process the ack */ 7747 if (bbr->rc_in_persist) 7748 tp->t_rxtshift = 0; 7749 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) 7750 bbr_strike_dupack(bbr); 7751 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 7752 } 7753 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); 7754 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 7755 /* 7756 * Old ack, behind the last one rcv'd or a duplicate ack 7757 * with SACK info. 7758 */ 7759 if (th->th_ack == tp->snd_una) { 7760 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); 7761 if (bbr->r_state == TCPS_SYN_SENT) { 7762 /* 7763 * Special case on where we sent SYN. When 7764 * the SYN-ACK is processed in syn_sent 7765 * state it bumps the snd_una. This causes 7766 * us to hit here even though we did ack 1 7767 * byte. 7768 * 7769 * Go through the nothing left case so we 7770 * send data. 7771 */ 7772 goto nothing_left; 7773 } 7774 } 7775 return (0); 7776 } 7777 /* 7778 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 7779 * something we sent. 7780 */ 7781 if (tp->t_flags & TF_NEEDSYN) { 7782 /* 7783 * T/TCP: Connection was half-synchronized, and our SYN has 7784 * been ACK'd (so connection is now fully synchronized). Go 7785 * to non-starred state, increment snd_una for ACK of SYN, 7786 * and check if we can do window scaling. 7787 */ 7788 tp->t_flags &= ~TF_NEEDSYN; 7789 tp->snd_una++; 7790 /* Do window scaling? */ 7791 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 7792 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 7793 tp->rcv_scale = tp->request_r_scale; 7794 /* Send window already scaled. */ 7795 } 7796 } 7797 INP_WLOCK_ASSERT(tp->t_inpcb); 7798 7799 acked = BYTES_THIS_ACK(tp, th); 7800 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 7801 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 7802 7803 /* 7804 * If we just performed our first retransmit, and the ACK arrives 7805 * within our recovery window, then it was a mistake to do the 7806 * retransmit in the first place. Recover our original cwnd and 7807 * ssthresh, and proceed to transmit where we left off. 7808 */ 7809 if (tp->t_flags & TF_PREVVALID) { 7810 tp->t_flags &= ~TF_PREVVALID; 7811 if (tp->t_rxtshift == 1 && 7812 (int)(ticks - tp->t_badrxtwin) < 0) 7813 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 7814 } 7815 SOCKBUF_LOCK(&so->so_snd); 7816 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 7817 tp->snd_wnd -= acked_amount; 7818 mfree = sbcut_locked(&so->so_snd, acked_amount); 7819 /* NB: sowwakeup_locked() does an implicit unlock. */ 7820 sowwakeup_locked(so); 7821 m_freem(mfree); 7822 if (SEQ_GT(th->th_ack, tp->snd_una)) { 7823 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 7824 } 7825 tp->snd_una = th->th_ack; 7826 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); 7827 if (IN_RECOVERY(tp->t_flags)) { 7828 if (SEQ_LT(th->th_ack, tp->snd_recover) && 7829 (SEQ_LT(th->th_ack, tp->snd_max))) { 7830 tcp_bbr_partialack(tp); 7831 } else { 7832 bbr_post_recovery(tp); 7833 } 7834 } 7835 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 7836 tp->snd_recover = tp->snd_una; 7837 } 7838 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 7839 tp->snd_nxt = tp->snd_max; 7840 } 7841 if (tp->snd_una == tp->snd_max) { 7842 /* Nothing left outstanding */ 7843 nothing_left: 7844 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 7845 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) 7846 bbr->rc_tp->t_acktime = 0; 7847 if ((sbused(&so->so_snd) == 0) && 7848 (tp->t_flags & TF_SENTFIN)) { 7849 ourfinisacked = 1; 7850 } 7851 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 7852 if (bbr->rc_in_persist == 0) { 7853 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 7854 } 7855 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 7856 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 7857 /* 7858 * We invalidate the last ack here since we 7859 * don't want to transfer forward the time 7860 * for our sum's calculations. 7861 */ 7862 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 7863 (sbavail(&so->so_snd) == 0) && 7864 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 7865 /* 7866 * The socket was gone and the peer sent data, time 7867 * to reset him. 7868 */ 7869 *ret_val = 1; 7870 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 7871 /* tcp_close will kill the inp pre-log the Reset */ 7872 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 7873 tp = tcp_close(tp); 7874 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 7875 BBR_STAT_INC(bbr_dropped_af_data); 7876 return (1); 7877 } 7878 /* Set need output so persist might get set */ 7879 bbr->r_wanted_output = 1; 7880 } 7881 if (ofia) 7882 *ofia = ourfinisacked; 7883 return (0); 7884 } 7885 7886 static void 7887 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7888 { 7889 if (bbr->rc_in_persist == 0) { 7890 bbr_timer_cancel(bbr, __LINE__, cts); 7891 bbr->r_ctl.rc_last_delay_val = 0; 7892 tp->t_rxtshift = 0; 7893 bbr->rc_in_persist = 1; 7894 bbr->r_ctl.rc_went_idle_time = cts; 7895 /* We should be capped when rw went to 0 but just in case */ 7896 bbr_log_type_pesist(bbr, cts, 0, line, 1); 7897 /* Time freezes for the state, so do the accounting now */ 7898 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 7899 uint32_t time_in; 7900 7901 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 7902 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7903 int32_t idx; 7904 7905 idx = bbr_state_val(bbr); 7906 counter_u64_add(bbr_state_time[(idx + 5)], time_in); 7907 } else { 7908 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 7909 } 7910 } 7911 bbr->r_ctl.rc_bbr_state_time = cts; 7912 } 7913 } 7914 7915 static void 7916 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) 7917 { 7918 /* 7919 * Note that if idle time does not exceed our 7920 * threshold, we do nothing continuing the state 7921 * transitions we were last walking through. 7922 */ 7923 if (idle_time >= bbr_idle_restart_threshold) { 7924 if (bbr->rc_use_idle_restart) { 7925 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; 7926 /* 7927 * Set our target using BBR_UNIT, so 7928 * we increase at a dramatic rate but 7929 * we stop when we get the pipe 7930 * full again for our current b/w estimate. 7931 */ 7932 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 7933 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 7934 bbr_set_state_target(bbr, __LINE__); 7935 /* Now setup our gains to ramp up */ 7936 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 7937 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 7938 bbr_log_type_statechange(bbr, cts, __LINE__); 7939 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7940 bbr_substate_change(bbr, cts, __LINE__, 1); 7941 } 7942 } 7943 } 7944 7945 static void 7946 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7947 { 7948 uint32_t idle_time; 7949 7950 if (bbr->rc_in_persist == 0) 7951 return; 7952 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); 7953 bbr->rc_in_persist = 0; 7954 bbr->rc_hit_state_1 = 0; 7955 bbr->r_ctl.rc_del_time = cts; 7956 /* 7957 * We invalidate the last ack here since we 7958 * don't want to transfer forward the time 7959 * for our sum's calculations. 7960 */ 7961 if (tcp_in_hpts(bbr->rc_inp)) { 7962 tcp_hpts_remove(bbr->rc_inp); 7963 bbr->rc_timer_first = 0; 7964 bbr->r_ctl.rc_hpts_flags = 0; 7965 bbr->r_ctl.rc_last_delay_val = 0; 7966 bbr->r_ctl.rc_hptsi_agg_delay = 0; 7967 bbr->r_agg_early_set = 0; 7968 bbr->r_ctl.rc_agg_early = 0; 7969 } 7970 bbr_log_type_pesist(bbr, cts, idle_time, line, 0); 7971 if (idle_time >= bbr_rtt_probe_time) { 7972 /* 7973 * This qualifies as a RTT_PROBE session since we drop the 7974 * data outstanding to nothing and waited more than 7975 * bbr_rtt_probe_time. 7976 */ 7977 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); 7978 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; 7979 } 7980 tp->t_rxtshift = 0; 7981 /* 7982 * If in probeBW and we have persisted more than an RTT lets do 7983 * special handling. 7984 */ 7985 /* Force a time based epoch */ 7986 bbr_set_epoch(bbr, cts, __LINE__); 7987 /* 7988 * Setup the lost so we don't count anything against the guy 7989 * we have been stuck with during persists. 7990 */ 7991 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 7992 /* Time un-freezes for the state */ 7993 bbr->r_ctl.rc_bbr_state_time = cts; 7994 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || 7995 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { 7996 /* 7997 * If we are going back to probe-bw 7998 * or probe_rtt, we may need to possibly 7999 * do a fast restart. 8000 */ 8001 bbr_restart_after_idle(bbr, cts, idle_time); 8002 } 8003 } 8004 8005 static void 8006 bbr_collapsed_window(struct tcp_bbr *bbr) 8007 { 8008 /* 8009 * Now we must walk the 8010 * send map and divide the 8011 * ones left stranded. These 8012 * guys can't cause us to abort 8013 * the connection and are really 8014 * "unsent". However if a buggy 8015 * client actually did keep some 8016 * of the data i.e. collapsed the win 8017 * and refused to ack and then opened 8018 * the win and acked that data. We would 8019 * get into an ack war, the simplier 8020 * method then of just pretending we 8021 * did not send those segments something 8022 * won't work. 8023 */ 8024 struct bbr_sendmap *rsm, *nrsm; 8025 tcp_seq max_seq; 8026 uint32_t maxseg; 8027 int can_split = 0; 8028 int fnd = 0; 8029 8030 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 8031 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; 8032 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); 8033 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 8034 /* Find the first seq past or at maxseq */ 8035 if (rsm->r_flags & BBR_RWND_COLLAPSED) 8036 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8037 if (SEQ_GEQ(max_seq, rsm->r_start) && 8038 SEQ_GEQ(rsm->r_end, max_seq)) { 8039 fnd = 1; 8040 break; 8041 } 8042 } 8043 bbr->rc_has_collapsed = 0; 8044 if (!fnd) { 8045 /* Nothing to do strange */ 8046 return; 8047 } 8048 /* 8049 * Now can we split? 8050 * 8051 * We don't want to split if splitting 8052 * would generate too many small segments 8053 * less we let an attacker fragment our 8054 * send_map and leave us out of memory. 8055 */ 8056 if ((max_seq != rsm->r_start) && 8057 (max_seq != rsm->r_end)){ 8058 /* can we split? */ 8059 int res1, res2; 8060 8061 res1 = max_seq - rsm->r_start; 8062 res2 = rsm->r_end - max_seq; 8063 if ((res1 >= (maxseg/8)) && 8064 (res2 >= (maxseg/8))) { 8065 /* No small pieces here */ 8066 can_split = 1; 8067 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { 8068 /* We are under the limit */ 8069 can_split = 1; 8070 } 8071 } 8072 /* Ok do we need to split this rsm? */ 8073 if (max_seq == rsm->r_start) { 8074 /* It's this guy no split required */ 8075 nrsm = rsm; 8076 } else if (max_seq == rsm->r_end) { 8077 /* It's the next one no split required. */ 8078 nrsm = TAILQ_NEXT(rsm, r_next); 8079 if (nrsm == NULL) { 8080 /* Huh? */ 8081 return; 8082 } 8083 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { 8084 /* yep we need to split it */ 8085 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 8086 if (nrsm == NULL) { 8087 /* failed XXXrrs what can we do mark the whole? */ 8088 nrsm = rsm; 8089 goto no_split; 8090 } 8091 /* Clone it */ 8092 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); 8093 bbr_clone_rsm(bbr, nrsm, rsm, max_seq); 8094 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 8095 if (rsm->r_in_tmap) { 8096 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 8097 nrsm->r_in_tmap = 1; 8098 } 8099 } else { 8100 /* 8101 * Split not allowed just start here just 8102 * use this guy. 8103 */ 8104 nrsm = rsm; 8105 } 8106 no_split: 8107 BBR_STAT_INC(bbr_collapsed_win); 8108 /* reuse fnd as a count */ 8109 fnd = 0; 8110 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { 8111 nrsm->r_flags |= BBR_RWND_COLLAPSED; 8112 fnd++; 8113 bbr->rc_has_collapsed = 1; 8114 } 8115 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); 8116 } 8117 8118 static void 8119 bbr_un_collapse_window(struct tcp_bbr *bbr) 8120 { 8121 struct bbr_sendmap *rsm; 8122 int cleared = 0; 8123 8124 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 8125 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 8126 /* Clear the flag */ 8127 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8128 cleared++; 8129 } else 8130 break; 8131 } 8132 bbr_log_type_rwnd_collapse(bbr, 8133 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); 8134 bbr->rc_has_collapsed = 0; 8135 } 8136 8137 /* 8138 * Return value of 1, the TCB is unlocked and most 8139 * likely gone, return value of 0, the TCB is still 8140 * locked. 8141 */ 8142 static int 8143 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 8144 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 8145 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 8146 { 8147 /* 8148 * Update window information. Don't look at window if no ACK: TAC's 8149 * send garbage on first SYN. 8150 */ 8151 uint16_t nsegs; 8152 int32_t tfo_syn; 8153 struct tcp_bbr *bbr; 8154 8155 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8156 INP_WLOCK_ASSERT(tp->t_inpcb); 8157 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8158 if ((thflags & TH_ACK) && 8159 (SEQ_LT(tp->snd_wl1, th->th_seq) || 8160 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 8161 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 8162 /* keep track of pure window updates */ 8163 if (tlen == 0 && 8164 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 8165 KMOD_TCPSTAT_INC(tcps_rcvwinupd); 8166 tp->snd_wnd = tiwin; 8167 tp->snd_wl1 = th->th_seq; 8168 tp->snd_wl2 = th->th_ack; 8169 if (tp->snd_wnd > tp->max_sndwnd) 8170 tp->max_sndwnd = tp->snd_wnd; 8171 bbr->r_wanted_output = 1; 8172 } else if (thflags & TH_ACK) { 8173 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 8174 tp->snd_wnd = tiwin; 8175 tp->snd_wl1 = th->th_seq; 8176 tp->snd_wl2 = th->th_ack; 8177 } 8178 } 8179 if (tp->snd_wnd < ctf_outstanding(tp)) 8180 /* The peer collapsed its window on us */ 8181 bbr_collapsed_window(bbr); 8182 else if (bbr->rc_has_collapsed) 8183 bbr_un_collapse_window(bbr); 8184 /* Was persist timer active and now we have window space? */ 8185 if ((bbr->rc_in_persist != 0) && 8186 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8187 bbr_minseg(bbr)))) { 8188 /* 8189 * Make the rate persist at end of persist mode if idle long 8190 * enough 8191 */ 8192 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8193 8194 /* Make sure we output to start the timer */ 8195 bbr->r_wanted_output = 1; 8196 } 8197 /* Do we need to enter persist? */ 8198 if ((bbr->rc_in_persist == 0) && 8199 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8200 TCPS_HAVEESTABLISHED(tp->t_state) && 8201 (tp->snd_max == tp->snd_una) && 8202 sbavail(&tp->t_inpcb->inp_socket->so_snd) && 8203 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { 8204 /* No send window.. we must enter persist */ 8205 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8206 } 8207 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 8208 m_freem(m); 8209 return (0); 8210 } 8211 /* 8212 * We don't support urgent data but 8213 * drag along the up just to make sure 8214 * if there is a stack switch no one 8215 * is surprised. 8216 */ 8217 tp->rcv_up = tp->rcv_nxt; 8218 INP_WLOCK_ASSERT(tp->t_inpcb); 8219 8220 /* 8221 * Process the segment text, merging it into the TCP sequencing 8222 * queue, and arranging for acknowledgment of receipt if necessary. 8223 * This process logically involves adjusting tp->rcv_wnd as data is 8224 * presented to the user (this happens in tcp_usrreq.c, case 8225 * PRU_RCVD). If a FIN has already been received on this connection 8226 * then we just ignore the text. 8227 */ 8228 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 8229 IS_FASTOPEN(tp->t_flags)); 8230 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && 8231 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8232 tcp_seq save_start = th->th_seq; 8233 tcp_seq save_rnxt = tp->rcv_nxt; 8234 int save_tlen = tlen; 8235 8236 m_adj(m, drop_hdrlen); /* delayed header drop */ 8237 /* 8238 * Insert segment which includes th into TCP reassembly 8239 * queue with control block tp. Set thflags to whether 8240 * reassembly now includes a segment with FIN. This handles 8241 * the common case inline (segment is the next to be 8242 * received on an established connection, and the queue is 8243 * empty), avoiding linkage into and removal from the queue 8244 * and repetition of various conversions. Set DELACK for 8245 * segments received in order, but ack immediately when 8246 * segments are out of order (so fast retransmit can work). 8247 */ 8248 if (th->th_seq == tp->rcv_nxt && 8249 SEGQ_EMPTY(tp) && 8250 (TCPS_HAVEESTABLISHED(tp->t_state) || 8251 tfo_syn)) { 8252 #ifdef NETFLIX_SB_LIMITS 8253 u_int mcnt, appended; 8254 8255 if (so->so_rcv.sb_shlim) { 8256 mcnt = m_memcnt(m); 8257 appended = 0; 8258 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8259 CFO_NOSLEEP, NULL) == false) { 8260 counter_u64_add(tcp_sb_shlim_fails, 1); 8261 m_freem(m); 8262 return (0); 8263 } 8264 } 8265 8266 #endif 8267 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { 8268 bbr->bbr_segs_rcvd += max(1, nsegs); 8269 tp->t_flags |= TF_DELACK; 8270 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8271 } else { 8272 bbr->r_wanted_output = 1; 8273 tp->t_flags |= TF_ACKNOW; 8274 } 8275 tp->rcv_nxt += tlen; 8276 if (tlen && 8277 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8278 (tp->t_fbyte_in == 0)) { 8279 tp->t_fbyte_in = ticks; 8280 if (tp->t_fbyte_in == 0) 8281 tp->t_fbyte_in = 1; 8282 if (tp->t_fbyte_out && tp->t_fbyte_in) 8283 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8284 } 8285 thflags = tcp_get_flags(th) & TH_FIN; 8286 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8287 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8288 SOCKBUF_LOCK(&so->so_rcv); 8289 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 8290 m_freem(m); 8291 else 8292 #ifdef NETFLIX_SB_LIMITS 8293 appended = 8294 #endif 8295 sbappendstream_locked(&so->so_rcv, m, 0); 8296 /* NB: sorwakeup_locked() does an implicit unlock. */ 8297 sorwakeup_locked(so); 8298 #ifdef NETFLIX_SB_LIMITS 8299 if (so->so_rcv.sb_shlim && appended != mcnt) 8300 counter_fo_release(so->so_rcv.sb_shlim, 8301 mcnt - appended); 8302 #endif 8303 8304 } else { 8305 /* 8306 * XXX: Due to the header drop above "th" is 8307 * theoretically invalid by now. Fortunately 8308 * m_adj() doesn't actually frees any mbufs when 8309 * trimming from the head. 8310 */ 8311 tcp_seq temp = save_start; 8312 8313 thflags = tcp_reass(tp, th, &temp, &tlen, m); 8314 tp->t_flags |= TF_ACKNOW; 8315 if (tp->t_flags & TF_WAKESOR) { 8316 tp->t_flags &= ~TF_WAKESOR; 8317 /* NB: sorwakeup_locked() does an implicit unlock. */ 8318 sorwakeup_locked(so); 8319 } 8320 } 8321 if ((tp->t_flags & TF_SACK_PERMIT) && 8322 (save_tlen > 0) && 8323 TCPS_HAVEESTABLISHED(tp->t_state)) { 8324 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { 8325 /* 8326 * DSACK actually handled in the fastpath 8327 * above. 8328 */ 8329 tcp_update_sack_list(tp, save_start, 8330 save_start + save_tlen); 8331 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 8332 if ((tp->rcv_numsacks >= 1) && 8333 (tp->sackblks[0].end == save_start)) { 8334 /* 8335 * Partial overlap, recorded at todrop 8336 * above. 8337 */ 8338 tcp_update_sack_list(tp, 8339 tp->sackblks[0].start, 8340 tp->sackblks[0].end); 8341 } else { 8342 tcp_update_dsack_list(tp, save_start, 8343 save_start + save_tlen); 8344 } 8345 } else if (tlen >= save_tlen) { 8346 /* Update of sackblks. */ 8347 tcp_update_dsack_list(tp, save_start, 8348 save_start + save_tlen); 8349 } else if (tlen > 0) { 8350 tcp_update_dsack_list(tp, save_start, 8351 save_start + tlen); 8352 } 8353 } 8354 } else { 8355 m_freem(m); 8356 thflags &= ~TH_FIN; 8357 } 8358 8359 /* 8360 * If FIN is received ACK the FIN and let the user know that the 8361 * connection is closing. 8362 */ 8363 if (thflags & TH_FIN) { 8364 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8365 /* The socket upcall is handled by socantrcvmore. */ 8366 socantrcvmore(so); 8367 /* 8368 * If connection is half-synchronized (ie NEEDSYN 8369 * flag on) then delay ACK, so it may be piggybacked 8370 * when SYN is sent. Otherwise, since we received a 8371 * FIN then no more input can be expected, send ACK 8372 * now. 8373 */ 8374 if (tp->t_flags & TF_NEEDSYN) { 8375 tp->t_flags |= TF_DELACK; 8376 bbr_timer_cancel(bbr, 8377 __LINE__, bbr->r_ctl.rc_rcvtime); 8378 } else { 8379 tp->t_flags |= TF_ACKNOW; 8380 } 8381 tp->rcv_nxt++; 8382 } 8383 switch (tp->t_state) { 8384 /* 8385 * In SYN_RECEIVED and ESTABLISHED STATES enter the 8386 * CLOSE_WAIT state. 8387 */ 8388 case TCPS_SYN_RECEIVED: 8389 tp->t_starttime = ticks; 8390 /* FALLTHROUGH */ 8391 case TCPS_ESTABLISHED: 8392 tcp_state_change(tp, TCPS_CLOSE_WAIT); 8393 break; 8394 8395 /* 8396 * If still in FIN_WAIT_1 STATE FIN has not been 8397 * acked so enter the CLOSING state. 8398 */ 8399 case TCPS_FIN_WAIT_1: 8400 tcp_state_change(tp, TCPS_CLOSING); 8401 break; 8402 8403 /* 8404 * In FIN_WAIT_2 state enter the TIME_WAIT state, 8405 * starting the time-wait timer, turning off the 8406 * other standard timers. 8407 */ 8408 case TCPS_FIN_WAIT_2: 8409 bbr->rc_timer_first = 1; 8410 bbr_timer_cancel(bbr, 8411 __LINE__, bbr->r_ctl.rc_rcvtime); 8412 INP_WLOCK_ASSERT(tp->t_inpcb); 8413 tcp_twstart(tp); 8414 return (1); 8415 } 8416 } 8417 /* 8418 * Return any desired output. 8419 */ 8420 if ((tp->t_flags & TF_ACKNOW) || 8421 (sbavail(&so->so_snd) > ctf_outstanding(tp))) { 8422 bbr->r_wanted_output = 1; 8423 } 8424 INP_WLOCK_ASSERT(tp->t_inpcb); 8425 return (0); 8426 } 8427 8428 /* 8429 * Here nothing is really faster, its just that we 8430 * have broken out the fast-data path also just like 8431 * the fast-ack. Return 1 if we processed the packet 8432 * return 0 if you need to take the "slow-path". 8433 */ 8434 static int 8435 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 8436 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8437 uint32_t tiwin, int32_t nxt_pkt) 8438 { 8439 uint16_t nsegs; 8440 int32_t newsize = 0; /* automatic sockbuf scaling */ 8441 struct tcp_bbr *bbr; 8442 #ifdef NETFLIX_SB_LIMITS 8443 u_int mcnt, appended; 8444 #endif 8445 #ifdef TCPDEBUG 8446 /* 8447 * The size of tcp_saveipgen must be the size of the max ip header, 8448 * now IPv6. 8449 */ 8450 u_char tcp_saveipgen[IP6_HDR_LEN]; 8451 struct tcphdr tcp_savetcp; 8452 short ostate = 0; 8453 8454 #endif 8455 /* On the hpts and we would have called output */ 8456 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8457 8458 /* 8459 * If last ACK falls within this segment's sequence numbers, record 8460 * the timestamp. NOTE that the test is modified according to the 8461 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8462 */ 8463 if (bbr->r_ctl.rc_resend != NULL) { 8464 return (0); 8465 } 8466 if (tiwin && tiwin != tp->snd_wnd) { 8467 return (0); 8468 } 8469 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 8470 return (0); 8471 } 8472 if (__predict_false((to->to_flags & TOF_TS) && 8473 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 8474 return (0); 8475 } 8476 if (__predict_false((th->th_ack != tp->snd_una))) { 8477 return (0); 8478 } 8479 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 8480 return (0); 8481 } 8482 if ((to->to_flags & TOF_TS) != 0 && 8483 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8484 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 8485 tp->ts_recent = to->to_tsval; 8486 } 8487 /* 8488 * This is a pure, in-sequence data packet with nothing on the 8489 * reassembly queue and we have enough buffer space to take it. 8490 */ 8491 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8492 8493 #ifdef NETFLIX_SB_LIMITS 8494 if (so->so_rcv.sb_shlim) { 8495 mcnt = m_memcnt(m); 8496 appended = 0; 8497 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8498 CFO_NOSLEEP, NULL) == false) { 8499 counter_u64_add(tcp_sb_shlim_fails, 1); 8500 m_freem(m); 8501 return (1); 8502 } 8503 } 8504 #endif 8505 /* Clean receiver SACK report if present */ 8506 if (tp->rcv_numsacks) 8507 tcp_clean_sackreport(tp); 8508 KMOD_TCPSTAT_INC(tcps_preddat); 8509 tp->rcv_nxt += tlen; 8510 if (tlen && 8511 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8512 (tp->t_fbyte_in == 0)) { 8513 tp->t_fbyte_in = ticks; 8514 if (tp->t_fbyte_in == 0) 8515 tp->t_fbyte_in = 1; 8516 if (tp->t_fbyte_out && tp->t_fbyte_in) 8517 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8518 } 8519 /* 8520 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 8521 */ 8522 tp->snd_wl1 = th->th_seq; 8523 /* 8524 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 8525 */ 8526 tp->rcv_up = tp->rcv_nxt; 8527 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8528 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8529 #ifdef TCPDEBUG 8530 if (so->so_options & SO_DEBUG) 8531 tcp_trace(TA_INPUT, ostate, tp, 8532 (void *)tcp_saveipgen, &tcp_savetcp, 0); 8533 #endif 8534 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 8535 8536 /* Add data to socket buffer. */ 8537 SOCKBUF_LOCK(&so->so_rcv); 8538 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8539 m_freem(m); 8540 } else { 8541 /* 8542 * Set new socket buffer size. Give up when limit is 8543 * reached. 8544 */ 8545 if (newsize) 8546 if (!sbreserve_locked(&so->so_rcv, 8547 newsize, so, NULL)) 8548 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 8549 m_adj(m, drop_hdrlen); /* delayed header drop */ 8550 8551 #ifdef NETFLIX_SB_LIMITS 8552 appended = 8553 #endif 8554 sbappendstream_locked(&so->so_rcv, m, 0); 8555 ctf_calc_rwin(so, tp); 8556 } 8557 /* NB: sorwakeup_locked() does an implicit unlock. */ 8558 sorwakeup_locked(so); 8559 #ifdef NETFLIX_SB_LIMITS 8560 if (so->so_rcv.sb_shlim && mcnt != appended) 8561 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); 8562 #endif 8563 if (DELAY_ACK(tp, bbr, nsegs)) { 8564 bbr->bbr_segs_rcvd += max(1, nsegs); 8565 tp->t_flags |= TF_DELACK; 8566 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8567 } else { 8568 bbr->r_wanted_output = 1; 8569 tp->t_flags |= TF_ACKNOW; 8570 } 8571 return (1); 8572 } 8573 8574 /* 8575 * This subfunction is used to try to highly optimize the 8576 * fast path. We again allow window updates that are 8577 * in sequence to remain in the fast-path. We also add 8578 * in the __predict's to attempt to help the compiler. 8579 * Note that if we return a 0, then we can *not* process 8580 * it and the caller should push the packet into the 8581 * slow-path. If we return 1, then all is well and 8582 * the packet is fully processed. 8583 */ 8584 static int 8585 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 8586 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8587 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) 8588 { 8589 int32_t acked; 8590 uint16_t nsegs; 8591 uint32_t sack_changed; 8592 #ifdef TCPDEBUG 8593 /* 8594 * The size of tcp_saveipgen must be the size of the max ip header, 8595 * now IPv6. 8596 */ 8597 u_char tcp_saveipgen[IP6_HDR_LEN]; 8598 struct tcphdr tcp_savetcp; 8599 short ostate = 0; 8600 8601 #endif 8602 uint32_t prev_acked = 0; 8603 struct tcp_bbr *bbr; 8604 8605 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 8606 /* Old ack, behind (or duplicate to) the last one rcv'd */ 8607 return (0); 8608 } 8609 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 8610 /* Above what we have sent? */ 8611 return (0); 8612 } 8613 if (__predict_false(tiwin == 0)) { 8614 /* zero window */ 8615 return (0); 8616 } 8617 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 8618 /* We need a SYN or a FIN, unlikely.. */ 8619 return (0); 8620 } 8621 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 8622 /* Timestamp is behind .. old ack with seq wrap? */ 8623 return (0); 8624 } 8625 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 8626 /* Still recovering */ 8627 return (0); 8628 } 8629 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8630 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { 8631 /* We are retransmitting */ 8632 return (0); 8633 } 8634 if (__predict_false(bbr->rc_in_persist != 0)) { 8635 /* In persist mode */ 8636 return (0); 8637 } 8638 if (bbr->r_ctl.rc_sacked) { 8639 /* We have sack holes on our scoreboard */ 8640 return (0); 8641 } 8642 /* Ok if we reach here, we can process a fast-ack */ 8643 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8644 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 8645 /* 8646 * We never detect loss in fast ack [we can't 8647 * have a sack and can't be in recovery so 8648 * we always pass 0 (nothing detected)]. 8649 */ 8650 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); 8651 /* Did the window get updated? */ 8652 if (tiwin != tp->snd_wnd) { 8653 tp->snd_wnd = tiwin; 8654 tp->snd_wl1 = th->th_seq; 8655 if (tp->snd_wnd > tp->max_sndwnd) 8656 tp->max_sndwnd = tp->snd_wnd; 8657 } 8658 /* Do we need to exit persists? */ 8659 if ((bbr->rc_in_persist != 0) && 8660 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8661 bbr_minseg(bbr)))) { 8662 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8663 bbr->r_wanted_output = 1; 8664 } 8665 /* Do we need to enter persists? */ 8666 if ((bbr->rc_in_persist == 0) && 8667 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8668 TCPS_HAVEESTABLISHED(tp->t_state) && 8669 (tp->snd_max == tp->snd_una) && 8670 sbavail(&tp->t_inpcb->inp_socket->so_snd) && 8671 (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { 8672 /* No send window.. we must enter persist */ 8673 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8674 } 8675 /* 8676 * If last ACK falls within this segment's sequence numbers, record 8677 * the timestamp. NOTE that the test is modified according to the 8678 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8679 */ 8680 if ((to->to_flags & TOF_TS) != 0 && 8681 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8682 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; 8683 tp->ts_recent = to->to_tsval; 8684 } 8685 /* 8686 * This is a pure ack for outstanding data. 8687 */ 8688 KMOD_TCPSTAT_INC(tcps_predack); 8689 8690 /* 8691 * "bad retransmit" recovery. 8692 */ 8693 if (tp->t_flags & TF_PREVVALID) { 8694 tp->t_flags &= ~TF_PREVVALID; 8695 if (tp->t_rxtshift == 1 && 8696 (int)(ticks - tp->t_badrxtwin) < 0) 8697 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 8698 } 8699 /* 8700 * Recalculate the transmit timer / rtt. 8701 * 8702 * Some boxes send broken timestamp replies during the SYN+ACK 8703 * phase, ignore timestamps of 0 or we could calculate a huge RTT 8704 * and blow up the retransmit timer. 8705 */ 8706 acked = BYTES_THIS_ACK(tp, th); 8707 8708 #ifdef TCP_HHOOK 8709 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 8710 hhook_run_tcp_est_in(tp, th, to); 8711 #endif 8712 8713 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 8714 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 8715 sbdrop(&so->so_snd, acked); 8716 8717 if (SEQ_GT(th->th_ack, tp->snd_una)) 8718 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 8719 tp->snd_una = th->th_ack; 8720 if (tp->snd_wnd < ctf_outstanding(tp)) 8721 /* The peer collapsed its window on us */ 8722 bbr_collapsed_window(bbr); 8723 else if (bbr->rc_has_collapsed) 8724 bbr_un_collapse_window(bbr); 8725 8726 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 8727 tp->snd_recover = tp->snd_una; 8728 } 8729 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); 8730 /* 8731 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 8732 */ 8733 tp->snd_wl2 = th->th_ack; 8734 m_freem(m); 8735 /* 8736 * If all outstanding data are acked, stop retransmit timer, 8737 * otherwise restart timer using current (possibly backed-off) 8738 * value. If process is waiting for space, wakeup/selwakeup/signal. 8739 * If data are ready to send, let tcp_output decide between more 8740 * output or persist. 8741 */ 8742 #ifdef TCPDEBUG 8743 if (so->so_options & SO_DEBUG) 8744 tcp_trace(TA_INPUT, ostate, tp, 8745 (void *)tcp_saveipgen, 8746 &tcp_savetcp, 0); 8747 #endif 8748 /* Wake up the socket if we have room to write more */ 8749 sowwakeup(so); 8750 if (tp->snd_una == tp->snd_max) { 8751 /* Nothing left outstanding */ 8752 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 8753 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) 8754 bbr->rc_tp->t_acktime = 0; 8755 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8756 if (bbr->rc_in_persist == 0) { 8757 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 8758 } 8759 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 8760 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 8761 /* 8762 * We invalidate the last ack here since we 8763 * don't want to transfer forward the time 8764 * for our sum's calculations. 8765 */ 8766 bbr->r_wanted_output = 1; 8767 } 8768 if (sbavail(&so->so_snd)) { 8769 bbr->r_wanted_output = 1; 8770 } 8771 return (1); 8772 } 8773 8774 /* 8775 * Return value of 1, the TCB is unlocked and most 8776 * likely gone, return value of 0, the TCB is still 8777 * locked. 8778 */ 8779 static int 8780 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 8781 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8782 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8783 { 8784 int32_t todrop; 8785 int32_t ourfinisacked = 0; 8786 struct tcp_bbr *bbr; 8787 int32_t ret_val = 0; 8788 8789 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8790 ctf_calc_rwin(so, tp); 8791 /* 8792 * If the state is SYN_SENT: if seg contains an ACK, but not for our 8793 * SYN, drop the input. if seg contains a RST, then drop the 8794 * connection. if seg does not contain SYN, then drop it. Otherwise 8795 * this is an acceptable SYN segment initialize tp->rcv_nxt and 8796 * tp->irs if seg contains ack then advance tp->snd_una. BRR does 8797 * not support ECN so we will not say we are capable. if SYN has 8798 * been acked change to ESTABLISHED else SYN_RCVD state arrange for 8799 * segment to be acked (eventually) continue processing rest of 8800 * data/controls, beginning with URG 8801 */ 8802 if ((thflags & TH_ACK) && 8803 (SEQ_LEQ(th->th_ack, tp->iss) || 8804 SEQ_GT(th->th_ack, tp->snd_max))) { 8805 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8806 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8807 return (1); 8808 } 8809 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 8810 TCP_PROBE5(connect__refused, NULL, tp, 8811 mtod(m, const char *), tp, th); 8812 tp = tcp_drop(tp, ECONNREFUSED); 8813 ctf_do_drop(m, tp); 8814 return (1); 8815 } 8816 if (thflags & TH_RST) { 8817 ctf_do_drop(m, tp); 8818 return (1); 8819 } 8820 if (!(thflags & TH_SYN)) { 8821 ctf_do_drop(m, tp); 8822 return (1); 8823 } 8824 tp->irs = th->th_seq; 8825 tcp_rcvseqinit(tp); 8826 if (thflags & TH_ACK) { 8827 int tfo_partial = 0; 8828 8829 KMOD_TCPSTAT_INC(tcps_connects); 8830 soisconnected(so); 8831 #ifdef MAC 8832 mac_socketpeer_set_from_mbuf(m, so); 8833 #endif 8834 /* Do window scaling on this connection? */ 8835 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 8836 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 8837 tp->rcv_scale = tp->request_r_scale; 8838 } 8839 tp->rcv_adv += min(tp->rcv_wnd, 8840 TCP_MAXWIN << tp->rcv_scale); 8841 /* 8842 * If not all the data that was sent in the TFO SYN 8843 * has been acked, resend the remainder right away. 8844 */ 8845 if (IS_FASTOPEN(tp->t_flags) && 8846 (tp->snd_una != tp->snd_max)) { 8847 tp->snd_nxt = th->th_ack; 8848 tfo_partial = 1; 8849 } 8850 /* 8851 * If there's data, delay ACK; if there's also a FIN ACKNOW 8852 * will be turned on later. 8853 */ 8854 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { 8855 bbr->bbr_segs_rcvd += 1; 8856 tp->t_flags |= TF_DELACK; 8857 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8858 } else { 8859 bbr->r_wanted_output = 1; 8860 tp->t_flags |= TF_ACKNOW; 8861 } 8862 if (SEQ_GT(th->th_ack, tp->iss)) { 8863 /* 8864 * The SYN is acked 8865 * handle it specially. 8866 */ 8867 bbr_log_syn(tp, to); 8868 } 8869 if (SEQ_GT(th->th_ack, tp->snd_una)) { 8870 /* 8871 * We advance snd_una for the 8872 * fast open case. If th_ack is 8873 * acknowledging data beyond 8874 * snd_una we can't just call 8875 * ack-processing since the 8876 * data stream in our send-map 8877 * will start at snd_una + 1 (one 8878 * beyond the SYN). If its just 8879 * equal we don't need to do that 8880 * and there is no send_map. 8881 */ 8882 tp->snd_una++; 8883 } 8884 /* 8885 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 8886 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 8887 */ 8888 tp->t_starttime = ticks; 8889 if (tp->t_flags & TF_NEEDFIN) { 8890 tcp_state_change(tp, TCPS_FIN_WAIT_1); 8891 tp->t_flags &= ~TF_NEEDFIN; 8892 thflags &= ~TH_SYN; 8893 } else { 8894 tcp_state_change(tp, TCPS_ESTABLISHED); 8895 TCP_PROBE5(connect__established, NULL, tp, 8896 mtod(m, const char *), tp, th); 8897 cc_conn_init(tp); 8898 } 8899 } else { 8900 /* 8901 * Received initial SYN in SYN-SENT[*] state => simultaneous 8902 * open. If segment contains CC option and there is a 8903 * cached CC, apply TAO test. If it succeeds, connection is * 8904 * half-synchronized. Otherwise, do 3-way handshake: 8905 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 8906 * there was no CC option, clear cached CC value. 8907 */ 8908 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); 8909 tcp_state_change(tp, TCPS_SYN_RECEIVED); 8910 } 8911 INP_WLOCK_ASSERT(tp->t_inpcb); 8912 /* 8913 * Advance th->th_seq to correspond to first data byte. If data, 8914 * trim to stay within window, dropping FIN if necessary. 8915 */ 8916 th->th_seq++; 8917 if (tlen > tp->rcv_wnd) { 8918 todrop = tlen - tp->rcv_wnd; 8919 m_adj(m, -todrop); 8920 tlen = tp->rcv_wnd; 8921 thflags &= ~TH_FIN; 8922 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 8923 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 8924 } 8925 tp->snd_wl1 = th->th_seq - 1; 8926 tp->rcv_up = th->th_seq; 8927 /* 8928 * Client side of transaction: already sent SYN and data. If the 8929 * remote host used T/TCP to validate the SYN, our data will be 8930 * ACK'd; if so, enter normal data segment processing in the middle 8931 * of step 5, ack processing. Otherwise, goto step 6. 8932 */ 8933 if (thflags & TH_ACK) { 8934 if ((to->to_flags & TOF_TS) != 0) { 8935 uint32_t t, rtt; 8936 8937 t = tcp_tv_to_mssectick(&bbr->rc_tv); 8938 if (TSTMP_GEQ(t, to->to_tsecr)) { 8939 rtt = t - to->to_tsecr; 8940 if (rtt == 0) { 8941 rtt = 1; 8942 } 8943 rtt *= MS_IN_USEC; 8944 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 8945 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, 8946 rtt, bbr->r_ctl.rc_rcvtime); 8947 } 8948 } 8949 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 8950 return (ret_val); 8951 /* We may have changed to FIN_WAIT_1 above */ 8952 if (tp->t_state == TCPS_FIN_WAIT_1) { 8953 /* 8954 * In FIN_WAIT_1 STATE in addition to the processing 8955 * for the ESTABLISHED state if our FIN is now 8956 * acknowledged then enter FIN_WAIT_2. 8957 */ 8958 if (ourfinisacked) { 8959 /* 8960 * If we can't receive any more data, then 8961 * closing user can proceed. Starting the 8962 * timer is contrary to the specification, 8963 * but if we don't get a FIN we'll hang 8964 * forever. 8965 * 8966 * XXXjl: we should release the tp also, and 8967 * use a compressed state. 8968 */ 8969 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8970 soisdisconnected(so); 8971 tcp_timer_activate(tp, TT_2MSL, 8972 (tcp_fast_finwait2_recycle ? 8973 tcp_finwait2_timeout : 8974 TP_MAXIDLE(tp))); 8975 } 8976 tcp_state_change(tp, TCPS_FIN_WAIT_2); 8977 } 8978 } 8979 } 8980 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 8981 tiwin, thflags, nxt_pkt)); 8982 } 8983 8984 /* 8985 * Return value of 1, the TCB is unlocked and most 8986 * likely gone, return value of 0, the TCB is still 8987 * locked. 8988 */ 8989 static int 8990 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 8991 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8992 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8993 { 8994 int32_t ourfinisacked = 0; 8995 int32_t ret_val; 8996 struct tcp_bbr *bbr; 8997 8998 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8999 ctf_calc_rwin(so, tp); 9000 if ((thflags & TH_ACK) && 9001 (SEQ_LEQ(th->th_ack, tp->snd_una) || 9002 SEQ_GT(th->th_ack, tp->snd_max))) { 9003 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9004 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9005 return (1); 9006 } 9007 if (IS_FASTOPEN(tp->t_flags)) { 9008 /* 9009 * When a TFO connection is in SYN_RECEIVED, the only valid 9010 * packets are the initial SYN, a retransmit/copy of the 9011 * initial SYN (possibly with a subset of the original 9012 * data), a valid ACK, a FIN, or a RST. 9013 */ 9014 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 9015 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9016 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9017 return (1); 9018 } else if (thflags & TH_SYN) { 9019 /* non-initial SYN is ignored */ 9020 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 9021 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 9022 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 9023 ctf_do_drop(m, NULL); 9024 return (0); 9025 } 9026 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 9027 ctf_do_drop(m, NULL); 9028 return (0); 9029 } 9030 } 9031 if ((thflags & TH_RST) || 9032 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9033 return (ctf_process_rst(m, th, so, tp)); 9034 /* 9035 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9036 * it's less than ts_recent, drop it. 9037 */ 9038 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9039 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9040 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9041 return (ret_val); 9042 } 9043 /* 9044 * In the SYN-RECEIVED state, validate that the packet belongs to 9045 * this connection before trimming the data to fit the receive 9046 * window. Check the sequence number versus IRS since we know the 9047 * sequence numbers haven't wrapped. This is a partial fix for the 9048 * "LAND" DoS attack. 9049 */ 9050 if (SEQ_LT(th->th_seq, tp->irs)) { 9051 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 9052 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9053 return (1); 9054 } 9055 INP_WLOCK_ASSERT(tp->t_inpcb); 9056 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9057 return (ret_val); 9058 } 9059 /* 9060 * If last ACK falls within this segment's sequence numbers, record 9061 * its timestamp. NOTE: 1) That the test incorporates suggestions 9062 * from the latest proposal of the tcplw@cray.com list (Braden 9063 * 1993/04/26). 2) That updating only on newer timestamps interferes 9064 * with our earlier PAWS tests, so this check should be solely 9065 * predicated on the sequence space of this segment. 3) That we 9066 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9067 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9068 * SEG.Len, This modified check allows us to overcome RFC1323's 9069 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9070 * p.869. In such cases, we can still calculate the RTT correctly 9071 * when RCV.NXT == Last.ACK.Sent. 9072 */ 9073 if ((to->to_flags & TOF_TS) != 0 && 9074 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9075 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9076 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9077 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9078 tp->ts_recent = to->to_tsval; 9079 } 9080 tp->snd_wnd = tiwin; 9081 /* 9082 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9083 * is on (half-synchronized state), then queue data for later 9084 * processing; else drop segment and return. 9085 */ 9086 if ((thflags & TH_ACK) == 0) { 9087 if (IS_FASTOPEN(tp->t_flags)) { 9088 cc_conn_init(tp); 9089 } 9090 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9091 tiwin, thflags, nxt_pkt)); 9092 } 9093 KMOD_TCPSTAT_INC(tcps_connects); 9094 soisconnected(so); 9095 /* Do window scaling? */ 9096 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 9097 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 9098 tp->rcv_scale = tp->request_r_scale; 9099 } 9100 /* 9101 * ok for the first time in lets see if we can use the ts to figure 9102 * out what the initial RTT was. 9103 */ 9104 if ((to->to_flags & TOF_TS) != 0) { 9105 uint32_t t, rtt; 9106 9107 t = tcp_tv_to_mssectick(&bbr->rc_tv); 9108 if (TSTMP_GEQ(t, to->to_tsecr)) { 9109 rtt = t - to->to_tsecr; 9110 if (rtt == 0) { 9111 rtt = 1; 9112 } 9113 rtt *= MS_IN_USEC; 9114 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 9115 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); 9116 } 9117 } 9118 /* Drop off any SYN in the send map (probably not there) */ 9119 if (thflags & TH_ACK) 9120 bbr_log_syn(tp, to); 9121 if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { 9122 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 9123 tp->t_tfo_pending = NULL; 9124 } 9125 /* 9126 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 9127 * FIN-WAIT-1 9128 */ 9129 tp->t_starttime = ticks; 9130 if (tp->t_flags & TF_NEEDFIN) { 9131 tcp_state_change(tp, TCPS_FIN_WAIT_1); 9132 tp->t_flags &= ~TF_NEEDFIN; 9133 } else { 9134 tcp_state_change(tp, TCPS_ESTABLISHED); 9135 TCP_PROBE5(accept__established, NULL, tp, 9136 mtod(m, const char *), tp, th); 9137 /* 9138 * TFO connections call cc_conn_init() during SYN 9139 * processing. Calling it again here for such connections 9140 * is not harmless as it would undo the snd_cwnd reduction 9141 * that occurs when a TFO SYN|ACK is retransmitted. 9142 */ 9143 if (!IS_FASTOPEN(tp->t_flags)) 9144 cc_conn_init(tp); 9145 } 9146 /* 9147 * Account for the ACK of our SYN prior to 9148 * regular ACK processing below, except for 9149 * simultaneous SYN, which is handled later. 9150 */ 9151 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) 9152 tp->snd_una++; 9153 /* 9154 * If segment contains data or ACK, will call tcp_reass() later; if 9155 * not, do so now to pass queued data to user. 9156 */ 9157 if (tlen == 0 && (thflags & TH_FIN) == 0) { 9158 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 9159 (struct mbuf *)0); 9160 if (tp->t_flags & TF_WAKESOR) { 9161 tp->t_flags &= ~TF_WAKESOR; 9162 /* NB: sorwakeup_locked() does an implicit unlock. */ 9163 sorwakeup_locked(so); 9164 } 9165 } 9166 tp->snd_wl1 = th->th_seq - 1; 9167 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9168 return (ret_val); 9169 } 9170 if (tp->t_state == TCPS_FIN_WAIT_1) { 9171 /* We could have went to FIN_WAIT_1 (or EST) above */ 9172 /* 9173 * In FIN_WAIT_1 STATE in addition to the processing for the 9174 * ESTABLISHED state if our FIN is now acknowledged then 9175 * enter FIN_WAIT_2. 9176 */ 9177 if (ourfinisacked) { 9178 /* 9179 * If we can't receive any more data, then closing 9180 * user can proceed. Starting the timer is contrary 9181 * to the specification, but if we don't get a FIN 9182 * we'll hang forever. 9183 * 9184 * XXXjl: we should release the tp also, and use a 9185 * compressed state. 9186 */ 9187 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9188 soisdisconnected(so); 9189 tcp_timer_activate(tp, TT_2MSL, 9190 (tcp_fast_finwait2_recycle ? 9191 tcp_finwait2_timeout : 9192 TP_MAXIDLE(tp))); 9193 } 9194 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9195 } 9196 } 9197 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9198 tiwin, thflags, nxt_pkt)); 9199 } 9200 9201 /* 9202 * Return value of 1, the TCB is unlocked and most 9203 * likely gone, return value of 0, the TCB is still 9204 * locked. 9205 */ 9206 static int 9207 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 9208 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9209 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9210 { 9211 struct tcp_bbr *bbr; 9212 int32_t ret_val; 9213 9214 /* 9215 * Header prediction: check for the two common cases of a 9216 * uni-directional data xfer. If the packet has no control flags, 9217 * is in-sequence, the window didn't change and we're not 9218 * retransmitting, it's a candidate. If the length is zero and the 9219 * ack moved forward, we're the sender side of the xfer. Just free 9220 * the data acked & wake any higher level process that was blocked 9221 * waiting for space. If the length is non-zero and the ack didn't 9222 * move, we're the receiver side. If we're getting packets in-order 9223 * (the reassembly queue is empty), add the data toc The socket 9224 * buffer and note that we need a delayed ack. Make sure that the 9225 * hidden state-flags are also off. Since we check for 9226 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 9227 */ 9228 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9229 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { 9230 /* 9231 * If we have delived under 4 segments increase the initial 9232 * window if raised by the peer. We use this to determine 9233 * dynamic and static rwnd's at the end of a connection. 9234 */ 9235 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); 9236 } 9237 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 9238 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 9239 __predict_true(SEGQ_EMPTY(tp)) && 9240 __predict_true(th->th_seq == tp->rcv_nxt)) { 9241 if (tlen == 0) { 9242 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 9243 tiwin, nxt_pkt, iptos)) { 9244 return (0); 9245 } 9246 } else { 9247 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 9248 tiwin, nxt_pkt)) { 9249 return (0); 9250 } 9251 } 9252 } 9253 ctf_calc_rwin(so, tp); 9254 9255 if ((thflags & TH_RST) || 9256 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9257 return (ctf_process_rst(m, th, so, tp)); 9258 /* 9259 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9260 * synchronized state. 9261 */ 9262 if (thflags & TH_SYN) { 9263 ctf_challenge_ack(m, th, tp, &ret_val); 9264 return (ret_val); 9265 } 9266 /* 9267 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9268 * it's less than ts_recent, drop it. 9269 */ 9270 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9271 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9272 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9273 return (ret_val); 9274 } 9275 INP_WLOCK_ASSERT(tp->t_inpcb); 9276 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9277 return (ret_val); 9278 } 9279 /* 9280 * If last ACK falls within this segment's sequence numbers, record 9281 * its timestamp. NOTE: 1) That the test incorporates suggestions 9282 * from the latest proposal of the tcplw@cray.com list (Braden 9283 * 1993/04/26). 2) That updating only on newer timestamps interferes 9284 * with our earlier PAWS tests, so this check should be solely 9285 * predicated on the sequence space of this segment. 3) That we 9286 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9287 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9288 * SEG.Len, This modified check allows us to overcome RFC1323's 9289 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9290 * p.869. In such cases, we can still calculate the RTT correctly 9291 * when RCV.NXT == Last.ACK.Sent. 9292 */ 9293 if ((to->to_flags & TOF_TS) != 0 && 9294 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9295 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9296 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9297 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9298 tp->ts_recent = to->to_tsval; 9299 } 9300 /* 9301 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9302 * is on (half-synchronized state), then queue data for later 9303 * processing; else drop segment and return. 9304 */ 9305 if ((thflags & TH_ACK) == 0) { 9306 if (tp->t_flags & TF_NEEDSYN) { 9307 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9308 tiwin, thflags, nxt_pkt)); 9309 } else if (tp->t_flags & TF_ACKNOW) { 9310 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9311 bbr->r_wanted_output = 1; 9312 return (ret_val); 9313 } else { 9314 ctf_do_drop(m, NULL); 9315 return (0); 9316 } 9317 } 9318 /* 9319 * Ack processing. 9320 */ 9321 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9322 return (ret_val); 9323 } 9324 if (sbavail(&so->so_snd)) { 9325 if (ctf_progress_timeout_check(tp, true)) { 9326 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9327 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9328 return (1); 9329 } 9330 } 9331 /* State changes only happen in bbr_process_data() */ 9332 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9333 tiwin, thflags, nxt_pkt)); 9334 } 9335 9336 /* 9337 * Return value of 1, the TCB is unlocked and most 9338 * likely gone, return value of 0, the TCB is still 9339 * locked. 9340 */ 9341 static int 9342 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 9343 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9344 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9345 { 9346 struct tcp_bbr *bbr; 9347 int32_t ret_val; 9348 9349 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9350 ctf_calc_rwin(so, tp); 9351 if ((thflags & TH_RST) || 9352 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9353 return (ctf_process_rst(m, th, so, tp)); 9354 /* 9355 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9356 * synchronized state. 9357 */ 9358 if (thflags & TH_SYN) { 9359 ctf_challenge_ack(m, th, tp, &ret_val); 9360 return (ret_val); 9361 } 9362 /* 9363 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9364 * it's less than ts_recent, drop it. 9365 */ 9366 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9367 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9368 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9369 return (ret_val); 9370 } 9371 INP_WLOCK_ASSERT(tp->t_inpcb); 9372 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9373 return (ret_val); 9374 } 9375 /* 9376 * If last ACK falls within this segment's sequence numbers, record 9377 * its timestamp. NOTE: 1) That the test incorporates suggestions 9378 * from the latest proposal of the tcplw@cray.com list (Braden 9379 * 1993/04/26). 2) That updating only on newer timestamps interferes 9380 * with our earlier PAWS tests, so this check should be solely 9381 * predicated on the sequence space of this segment. 3) That we 9382 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9383 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9384 * SEG.Len, This modified check allows us to overcome RFC1323's 9385 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9386 * p.869. In such cases, we can still calculate the RTT correctly 9387 * when RCV.NXT == Last.ACK.Sent. 9388 */ 9389 if ((to->to_flags & TOF_TS) != 0 && 9390 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9391 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9392 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9393 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9394 tp->ts_recent = to->to_tsval; 9395 } 9396 /* 9397 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9398 * is on (half-synchronized state), then queue data for later 9399 * processing; else drop segment and return. 9400 */ 9401 if ((thflags & TH_ACK) == 0) { 9402 if (tp->t_flags & TF_NEEDSYN) { 9403 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9404 tiwin, thflags, nxt_pkt)); 9405 } else if (tp->t_flags & TF_ACKNOW) { 9406 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9407 bbr->r_wanted_output = 1; 9408 return (ret_val); 9409 } else { 9410 ctf_do_drop(m, NULL); 9411 return (0); 9412 } 9413 } 9414 /* 9415 * Ack processing. 9416 */ 9417 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9418 return (ret_val); 9419 } 9420 if (sbavail(&so->so_snd)) { 9421 if (ctf_progress_timeout_check(tp, true)) { 9422 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9423 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9424 return (1); 9425 } 9426 } 9427 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9428 tiwin, thflags, nxt_pkt)); 9429 } 9430 9431 static int 9432 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, 9433 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) 9434 { 9435 9436 if (bbr->rc_allow_data_af_clo == 0) { 9437 close_now: 9438 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 9439 /* tcp_close will kill the inp pre-log the Reset */ 9440 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 9441 tp = tcp_close(tp); 9442 KMOD_TCPSTAT_INC(tcps_rcvafterclose); 9443 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 9444 return (1); 9445 } 9446 if (sbavail(&so->so_snd) == 0) 9447 goto close_now; 9448 /* Ok we allow data that is ignored and a followup reset */ 9449 tp->rcv_nxt = th->th_seq + *tlen; 9450 tp->t_flags2 |= TF2_DROP_AF_DATA; 9451 bbr->r_wanted_output = 1; 9452 *tlen = 0; 9453 return (0); 9454 } 9455 9456 /* 9457 * Return value of 1, the TCB is unlocked and most 9458 * likely gone, return value of 0, the TCB is still 9459 * locked. 9460 */ 9461 static int 9462 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 9463 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9464 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9465 { 9466 int32_t ourfinisacked = 0; 9467 int32_t ret_val; 9468 struct tcp_bbr *bbr; 9469 9470 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9471 ctf_calc_rwin(so, tp); 9472 if ((thflags & TH_RST) || 9473 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9474 return (ctf_process_rst(m, th, so, tp)); 9475 /* 9476 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9477 * synchronized state. 9478 */ 9479 if (thflags & TH_SYN) { 9480 ctf_challenge_ack(m, th, tp, &ret_val); 9481 return (ret_val); 9482 } 9483 /* 9484 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9485 * it's less than ts_recent, drop it. 9486 */ 9487 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9488 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9489 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9490 return (ret_val); 9491 } 9492 INP_WLOCK_ASSERT(tp->t_inpcb); 9493 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9494 return (ret_val); 9495 } 9496 /* 9497 * If new data are received on a connection after the user processes 9498 * are gone, then RST the other end. 9499 */ 9500 if ((so->so_state & SS_NOFDREF) && tlen) { 9501 /* 9502 * We call a new function now so we might continue and setup 9503 * to reset at all data being ack'd. 9504 */ 9505 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9506 return (1); 9507 } 9508 /* 9509 * If last ACK falls within this segment's sequence numbers, record 9510 * its timestamp. NOTE: 1) That the test incorporates suggestions 9511 * from the latest proposal of the tcplw@cray.com list (Braden 9512 * 1993/04/26). 2) That updating only on newer timestamps interferes 9513 * with our earlier PAWS tests, so this check should be solely 9514 * predicated on the sequence space of this segment. 3) That we 9515 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9516 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9517 * SEG.Len, This modified check allows us to overcome RFC1323's 9518 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9519 * p.869. In such cases, we can still calculate the RTT correctly 9520 * when RCV.NXT == Last.ACK.Sent. 9521 */ 9522 if ((to->to_flags & TOF_TS) != 0 && 9523 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9524 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9525 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9526 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9527 tp->ts_recent = to->to_tsval; 9528 } 9529 /* 9530 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9531 * is on (half-synchronized state), then queue data for later 9532 * processing; else drop segment and return. 9533 */ 9534 if ((thflags & TH_ACK) == 0) { 9535 if (tp->t_flags & TF_NEEDSYN) { 9536 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9537 tiwin, thflags, nxt_pkt)); 9538 } else if (tp->t_flags & TF_ACKNOW) { 9539 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9540 bbr->r_wanted_output = 1; 9541 return (ret_val); 9542 } else { 9543 ctf_do_drop(m, NULL); 9544 return (0); 9545 } 9546 } 9547 /* 9548 * Ack processing. 9549 */ 9550 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9551 return (ret_val); 9552 } 9553 if (ourfinisacked) { 9554 /* 9555 * If we can't receive any more data, then closing user can 9556 * proceed. Starting the timer is contrary to the 9557 * specification, but if we don't get a FIN we'll hang 9558 * forever. 9559 * 9560 * XXXjl: we should release the tp also, and use a 9561 * compressed state. 9562 */ 9563 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9564 soisdisconnected(so); 9565 tcp_timer_activate(tp, TT_2MSL, 9566 (tcp_fast_finwait2_recycle ? 9567 tcp_finwait2_timeout : 9568 TP_MAXIDLE(tp))); 9569 } 9570 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9571 } 9572 if (sbavail(&so->so_snd)) { 9573 if (ctf_progress_timeout_check(tp, true)) { 9574 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9575 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9576 return (1); 9577 } 9578 } 9579 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9580 tiwin, thflags, nxt_pkt)); 9581 } 9582 9583 /* 9584 * Return value of 1, the TCB is unlocked and most 9585 * likely gone, return value of 0, the TCB is still 9586 * locked. 9587 */ 9588 static int 9589 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 9590 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9591 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9592 { 9593 int32_t ourfinisacked = 0; 9594 int32_t ret_val; 9595 struct tcp_bbr *bbr; 9596 9597 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9598 ctf_calc_rwin(so, tp); 9599 if ((thflags & TH_RST) || 9600 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9601 return (ctf_process_rst(m, th, so, tp)); 9602 /* 9603 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9604 * synchronized state. 9605 */ 9606 if (thflags & TH_SYN) { 9607 ctf_challenge_ack(m, th, tp, &ret_val); 9608 return (ret_val); 9609 } 9610 /* 9611 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9612 * it's less than ts_recent, drop it. 9613 */ 9614 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9615 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9616 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9617 return (ret_val); 9618 } 9619 INP_WLOCK_ASSERT(tp->t_inpcb); 9620 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9621 return (ret_val); 9622 } 9623 /* 9624 * If new data are received on a connection after the user processes 9625 * are gone, then RST the other end. 9626 */ 9627 if ((so->so_state & SS_NOFDREF) && tlen) { 9628 /* 9629 * We call a new function now so we might continue and setup 9630 * to reset at all data being ack'd. 9631 */ 9632 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9633 return (1); 9634 } 9635 /* 9636 * If last ACK falls within this segment's sequence numbers, record 9637 * its timestamp. NOTE: 1) That the test incorporates suggestions 9638 * from the latest proposal of the tcplw@cray.com list (Braden 9639 * 1993/04/26). 2) That updating only on newer timestamps interferes 9640 * with our earlier PAWS tests, so this check should be solely 9641 * predicated on the sequence space of this segment. 3) That we 9642 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9643 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9644 * SEG.Len, This modified check allows us to overcome RFC1323's 9645 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9646 * p.869. In such cases, we can still calculate the RTT correctly 9647 * when RCV.NXT == Last.ACK.Sent. 9648 */ 9649 if ((to->to_flags & TOF_TS) != 0 && 9650 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9651 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9652 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9653 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9654 tp->ts_recent = to->to_tsval; 9655 } 9656 /* 9657 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9658 * is on (half-synchronized state), then queue data for later 9659 * processing; else drop segment and return. 9660 */ 9661 if ((thflags & TH_ACK) == 0) { 9662 if (tp->t_flags & TF_NEEDSYN) { 9663 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9664 tiwin, thflags, nxt_pkt)); 9665 } else if (tp->t_flags & TF_ACKNOW) { 9666 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9667 bbr->r_wanted_output = 1; 9668 return (ret_val); 9669 } else { 9670 ctf_do_drop(m, NULL); 9671 return (0); 9672 } 9673 } 9674 /* 9675 * Ack processing. 9676 */ 9677 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9678 return (ret_val); 9679 } 9680 if (ourfinisacked) { 9681 tcp_twstart(tp); 9682 m_freem(m); 9683 return (1); 9684 } 9685 if (sbavail(&so->so_snd)) { 9686 if (ctf_progress_timeout_check(tp, true)) { 9687 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9688 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9689 return (1); 9690 } 9691 } 9692 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9693 tiwin, thflags, nxt_pkt)); 9694 } 9695 9696 /* 9697 * Return value of 1, the TCB is unlocked and most 9698 * likely gone, return value of 0, the TCB is still 9699 * locked. 9700 */ 9701 static int 9702 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 9703 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9704 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9705 { 9706 int32_t ourfinisacked = 0; 9707 int32_t ret_val; 9708 struct tcp_bbr *bbr; 9709 9710 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9711 ctf_calc_rwin(so, tp); 9712 if ((thflags & TH_RST) || 9713 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9714 return (ctf_process_rst(m, th, so, tp)); 9715 /* 9716 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9717 * synchronized state. 9718 */ 9719 if (thflags & TH_SYN) { 9720 ctf_challenge_ack(m, th, tp, &ret_val); 9721 return (ret_val); 9722 } 9723 /* 9724 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9725 * it's less than ts_recent, drop it. 9726 */ 9727 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9728 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9729 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9730 return (ret_val); 9731 } 9732 INP_WLOCK_ASSERT(tp->t_inpcb); 9733 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9734 return (ret_val); 9735 } 9736 /* 9737 * If new data are received on a connection after the user processes 9738 * are gone, then RST the other end. 9739 */ 9740 if ((so->so_state & SS_NOFDREF) && tlen) { 9741 /* 9742 * We call a new function now so we might continue and setup 9743 * to reset at all data being ack'd. 9744 */ 9745 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9746 return (1); 9747 } 9748 /* 9749 * If last ACK falls within this segment's sequence numbers, record 9750 * its timestamp. NOTE: 1) That the test incorporates suggestions 9751 * from the latest proposal of the tcplw@cray.com list (Braden 9752 * 1993/04/26). 2) That updating only on newer timestamps interferes 9753 * with our earlier PAWS tests, so this check should be solely 9754 * predicated on the sequence space of this segment. 3) That we 9755 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9756 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9757 * SEG.Len, This modified check allows us to overcome RFC1323's 9758 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9759 * p.869. In such cases, we can still calculate the RTT correctly 9760 * when RCV.NXT == Last.ACK.Sent. 9761 */ 9762 if ((to->to_flags & TOF_TS) != 0 && 9763 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9764 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9765 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9766 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9767 tp->ts_recent = to->to_tsval; 9768 } 9769 /* 9770 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9771 * is on (half-synchronized state), then queue data for later 9772 * processing; else drop segment and return. 9773 */ 9774 if ((thflags & TH_ACK) == 0) { 9775 if (tp->t_flags & TF_NEEDSYN) { 9776 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9777 tiwin, thflags, nxt_pkt)); 9778 } else if (tp->t_flags & TF_ACKNOW) { 9779 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9780 bbr->r_wanted_output = 1; 9781 return (ret_val); 9782 } else { 9783 ctf_do_drop(m, NULL); 9784 return (0); 9785 } 9786 } 9787 /* 9788 * case TCPS_LAST_ACK: Ack processing. 9789 */ 9790 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9791 return (ret_val); 9792 } 9793 if (ourfinisacked) { 9794 tp = tcp_close(tp); 9795 ctf_do_drop(m, tp); 9796 return (1); 9797 } 9798 if (sbavail(&so->so_snd)) { 9799 if (ctf_progress_timeout_check(tp, true)) { 9800 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9801 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9802 return (1); 9803 } 9804 } 9805 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9806 tiwin, thflags, nxt_pkt)); 9807 } 9808 9809 /* 9810 * Return value of 1, the TCB is unlocked and most 9811 * likely gone, return value of 0, the TCB is still 9812 * locked. 9813 */ 9814 static int 9815 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 9816 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9817 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9818 { 9819 int32_t ourfinisacked = 0; 9820 int32_t ret_val; 9821 struct tcp_bbr *bbr; 9822 9823 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9824 ctf_calc_rwin(so, tp); 9825 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 9826 if ((thflags & TH_RST) || 9827 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9828 return (ctf_process_rst(m, th, so, tp)); 9829 9830 /* 9831 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9832 * synchronized state. 9833 */ 9834 if (thflags & TH_SYN) { 9835 ctf_challenge_ack(m, th, tp, &ret_val); 9836 return (ret_val); 9837 } 9838 INP_WLOCK_ASSERT(tp->t_inpcb); 9839 /* 9840 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9841 * it's less than ts_recent, drop it. 9842 */ 9843 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9844 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9845 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9846 return (ret_val); 9847 } 9848 INP_WLOCK_ASSERT(tp->t_inpcb); 9849 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9850 return (ret_val); 9851 } 9852 /* 9853 * If new data are received on a connection after the user processes 9854 * are gone, then we may RST the other end depending on the outcome 9855 * of bbr_check_data_after_close. 9856 */ 9857 if ((so->so_state & SS_NOFDREF) && 9858 tlen) { 9859 /* 9860 * We call a new function now so we might continue and setup 9861 * to reset at all data being ack'd. 9862 */ 9863 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9864 return (1); 9865 } 9866 INP_WLOCK_ASSERT(tp->t_inpcb); 9867 /* 9868 * If last ACK falls within this segment's sequence numbers, record 9869 * its timestamp. NOTE: 1) That the test incorporates suggestions 9870 * from the latest proposal of the tcplw@cray.com list (Braden 9871 * 1993/04/26). 2) That updating only on newer timestamps interferes 9872 * with our earlier PAWS tests, so this check should be solely 9873 * predicated on the sequence space of this segment. 3) That we 9874 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9875 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9876 * SEG.Len, This modified check allows us to overcome RFC1323's 9877 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9878 * p.869. In such cases, we can still calculate the RTT correctly 9879 * when RCV.NXT == Last.ACK.Sent. 9880 */ 9881 INP_WLOCK_ASSERT(tp->t_inpcb); 9882 if ((to->to_flags & TOF_TS) != 0 && 9883 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9884 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9885 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9886 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9887 tp->ts_recent = to->to_tsval; 9888 } 9889 /* 9890 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9891 * is on (half-synchronized state), then queue data for later 9892 * processing; else drop segment and return. 9893 */ 9894 if ((thflags & TH_ACK) == 0) { 9895 if (tp->t_flags & TF_NEEDSYN) { 9896 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9897 tiwin, thflags, nxt_pkt)); 9898 } else if (tp->t_flags & TF_ACKNOW) { 9899 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9900 bbr->r_wanted_output = 1; 9901 return (ret_val); 9902 } else { 9903 ctf_do_drop(m, NULL); 9904 return (0); 9905 } 9906 } 9907 /* 9908 * Ack processing. 9909 */ 9910 INP_WLOCK_ASSERT(tp->t_inpcb); 9911 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9912 return (ret_val); 9913 } 9914 if (sbavail(&so->so_snd)) { 9915 if (ctf_progress_timeout_check(tp, true)) { 9916 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9917 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9918 return (1); 9919 } 9920 } 9921 INP_WLOCK_ASSERT(tp->t_inpcb); 9922 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9923 tiwin, thflags, nxt_pkt)); 9924 } 9925 9926 static void 9927 bbr_stop_all_timers(struct tcpcb *tp) 9928 { 9929 struct tcp_bbr *bbr; 9930 9931 /* 9932 * Assure no timers are running. 9933 */ 9934 if (tcp_timer_active(tp, TT_PERSIST)) { 9935 /* We enter in persists, set the flag appropriately */ 9936 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9937 bbr->rc_in_persist = 1; 9938 } 9939 tcp_timer_suspend(tp, TT_PERSIST); 9940 tcp_timer_suspend(tp, TT_REXMT); 9941 tcp_timer_suspend(tp, TT_KEEP); 9942 tcp_timer_suspend(tp, TT_DELACK); 9943 } 9944 9945 static void 9946 bbr_google_mode_on(struct tcp_bbr *bbr) 9947 { 9948 bbr->rc_use_google = 1; 9949 bbr->rc_no_pacing = 0; 9950 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9951 bbr->r_use_policer = bbr_policer_detection_enabled; 9952 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 9953 bbr->bbr_use_rack_cheat = 0; 9954 bbr->r_ctl.rc_incr_tmrs = 0; 9955 bbr->r_ctl.rc_inc_tcp_oh = 0; 9956 bbr->r_ctl.rc_inc_ip_oh = 0; 9957 bbr->r_ctl.rc_inc_enet_oh = 0; 9958 reset_time(&bbr->r_ctl.rc_delrate, 9959 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 9960 reset_time_small(&bbr->r_ctl.rc_rttprop, 9961 (11 * USECS_IN_SECOND)); 9962 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9963 } 9964 9965 static void 9966 bbr_google_mode_off(struct tcp_bbr *bbr) 9967 { 9968 bbr->rc_use_google = 0; 9969 bbr->r_ctl.bbr_google_discount = 0; 9970 bbr->no_pacing_until = bbr_no_pacing_until; 9971 bbr->r_use_policer = 0; 9972 if (bbr->no_pacing_until) 9973 bbr->rc_no_pacing = 1; 9974 else 9975 bbr->rc_no_pacing = 0; 9976 if (bbr_use_rack_resend_cheat) 9977 bbr->bbr_use_rack_cheat = 1; 9978 else 9979 bbr->bbr_use_rack_cheat = 0; 9980 if (bbr_incr_timers) 9981 bbr->r_ctl.rc_incr_tmrs = 1; 9982 else 9983 bbr->r_ctl.rc_incr_tmrs = 0; 9984 if (bbr_include_tcp_oh) 9985 bbr->r_ctl.rc_inc_tcp_oh = 1; 9986 else 9987 bbr->r_ctl.rc_inc_tcp_oh = 0; 9988 if (bbr_include_ip_oh) 9989 bbr->r_ctl.rc_inc_ip_oh = 1; 9990 else 9991 bbr->r_ctl.rc_inc_ip_oh = 0; 9992 if (bbr_include_enet_oh) 9993 bbr->r_ctl.rc_inc_enet_oh = 1; 9994 else 9995 bbr->r_ctl.rc_inc_enet_oh = 0; 9996 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 9997 reset_time(&bbr->r_ctl.rc_delrate, 9998 bbr_num_pktepo_for_del_limit); 9999 reset_time_small(&bbr->r_ctl.rc_rttprop, 10000 (bbr_filter_len_sec * USECS_IN_SECOND)); 10001 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 10002 } 10003 /* 10004 * Return 0 on success, non-zero on failure 10005 * which indicates the error (usually no memory). 10006 */ 10007 static int 10008 bbr_init(struct tcpcb *tp) 10009 { 10010 struct tcp_bbr *bbr = NULL; 10011 struct inpcb *inp; 10012 uint32_t cts; 10013 10014 tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); 10015 if (tp->t_fb_ptr == NULL) { 10016 /* 10017 * We need to allocate memory but cant. The INP and INP_INFO 10018 * locks and they are recursive (happens during setup. So a 10019 * scheme to drop the locks fails :( 10020 * 10021 */ 10022 return (ENOMEM); 10023 } 10024 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10025 bbr->rtt_valid = 0; 10026 inp = tp->t_inpcb; 10027 inp->inp_flags2 |= INP_CANNOT_DO_ECN; 10028 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; 10029 TAILQ_INIT(&bbr->r_ctl.rc_map); 10030 TAILQ_INIT(&bbr->r_ctl.rc_free); 10031 TAILQ_INIT(&bbr->r_ctl.rc_tmap); 10032 bbr->rc_tp = tp; 10033 if (tp->t_inpcb) { 10034 bbr->rc_inp = tp->t_inpcb; 10035 } 10036 cts = tcp_get_usecs(&bbr->rc_tv); 10037 tp->t_acktime = 0; 10038 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; 10039 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; 10040 bbr->rc_tlp_threshold = bbr_tlp_thresh; 10041 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; 10042 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; 10043 bbr->r_ctl.rc_min_to = bbr_min_to; 10044 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10045 bbr->r_ctl.bbr_lost_at_state = 0; 10046 bbr->r_ctl.rc_lost_at_startup = 0; 10047 bbr->rc_all_timers_stopped = 0; 10048 bbr->r_ctl.rc_bbr_lastbtlbw = 0; 10049 bbr->r_ctl.rc_pkt_epoch_del = 0; 10050 bbr->r_ctl.rc_pkt_epoch = 0; 10051 bbr->r_ctl.rc_lowest_rtt = 0xffffffff; 10052 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; 10053 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 10054 bbr->r_ctl.rc_went_idle_time = cts; 10055 bbr->rc_pacer_started = cts; 10056 bbr->r_ctl.rc_pkt_epoch_time = cts; 10057 bbr->r_ctl.rc_rcvtime = cts; 10058 bbr->r_ctl.rc_bbr_state_time = cts; 10059 bbr->r_ctl.rc_del_time = cts; 10060 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 10061 bbr->r_ctl.last_in_probertt = cts; 10062 bbr->skip_gain = 0; 10063 bbr->gain_is_limited = 0; 10064 bbr->no_pacing_until = bbr_no_pacing_until; 10065 if (bbr->no_pacing_until) 10066 bbr->rc_no_pacing = 1; 10067 if (bbr_use_google_algo) { 10068 bbr->rc_no_pacing = 0; 10069 bbr->rc_use_google = 1; 10070 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 10071 bbr->r_use_policer = bbr_policer_detection_enabled; 10072 } else { 10073 bbr->rc_use_google = 0; 10074 bbr->r_ctl.bbr_google_discount = 0; 10075 bbr->r_use_policer = 0; 10076 } 10077 if (bbr_ts_limiting) 10078 bbr->rc_use_ts_limit = 1; 10079 else 10080 bbr->rc_use_ts_limit = 0; 10081 if (bbr_ts_can_raise) 10082 bbr->ts_can_raise = 1; 10083 else 10084 bbr->ts_can_raise = 0; 10085 if (V_tcp_delack_enabled == 1) 10086 tp->t_delayed_ack = 2; 10087 else if (V_tcp_delack_enabled == 0) 10088 tp->t_delayed_ack = 0; 10089 else if (V_tcp_delack_enabled < 100) 10090 tp->t_delayed_ack = V_tcp_delack_enabled; 10091 else 10092 tp->t_delayed_ack = 2; 10093 if (bbr->rc_use_google == 0) 10094 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10095 else 10096 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 10097 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; 10098 bbr->rc_max_rto_sec = bbr_rto_max_sec; 10099 bbr->rc_init_win = bbr_def_init_win; 10100 if (tp->t_flags & TF_REQ_TSTMP) 10101 bbr->rc_last_options = TCP_TS_OVERHEAD; 10102 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; 10103 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; 10104 bbr->r_init_rtt = 1; 10105 10106 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 10107 if (bbr_allow_hdwr_pacing) 10108 bbr->bbr_hdw_pace_ena = 1; 10109 else 10110 bbr->bbr_hdw_pace_ena = 0; 10111 if (bbr_sends_full_iwnd) 10112 bbr->bbr_init_win_cheat = 1; 10113 else 10114 bbr->bbr_init_win_cheat = 0; 10115 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; 10116 bbr->r_ctl.rc_drain_pg = bbr_drain_gain; 10117 bbr->r_ctl.rc_startup_pg = bbr_high_gain; 10118 bbr->rc_loss_exit = bbr_exit_startup_at_loss; 10119 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; 10120 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; 10121 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; 10122 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; 10123 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; 10124 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; 10125 bbr->r_ctl.bbr_cross_over = bbr_cross_over; 10126 bbr->r_ctl.rc_rtt_shrinks = cts; 10127 if (bbr->rc_use_google) { 10128 setup_time_filter(&bbr->r_ctl.rc_delrate, 10129 FILTER_TYPE_MAX, 10130 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 10131 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10132 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); 10133 } else { 10134 setup_time_filter(&bbr->r_ctl.rc_delrate, 10135 FILTER_TYPE_MAX, 10136 bbr_num_pktepo_for_del_limit); 10137 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10138 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); 10139 } 10140 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); 10141 if (bbr_uses_idle_restart) 10142 bbr->rc_use_idle_restart = 1; 10143 else 10144 bbr->rc_use_idle_restart = 0; 10145 bbr->r_ctl.rc_bbr_cur_del_rate = 0; 10146 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; 10147 if (bbr_resends_use_tso) 10148 bbr->rc_resends_use_tso = 1; 10149 #ifdef NETFLIX_PEAKRATE 10150 tp->t_peakrate_thr = tp->t_maxpeakrate; 10151 #endif 10152 if (tp->snd_una != tp->snd_max) { 10153 /* Create a send map for the current outstanding data */ 10154 struct bbr_sendmap *rsm; 10155 10156 rsm = bbr_alloc(bbr); 10157 if (rsm == NULL) { 10158 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10159 tp->t_fb_ptr = NULL; 10160 return (ENOMEM); 10161 } 10162 rsm->r_rtt_not_allowed = 1; 10163 rsm->r_tim_lastsent[0] = cts; 10164 rsm->r_rtr_cnt = 1; 10165 rsm->r_rtr_bytes = 0; 10166 rsm->r_start = tp->snd_una; 10167 rsm->r_end = tp->snd_max; 10168 rsm->r_dupack = 0; 10169 rsm->r_delivered = bbr->r_ctl.rc_delivered; 10170 rsm->r_ts_valid = 0; 10171 rsm->r_del_ack_ts = tp->ts_recent; 10172 rsm->r_del_time = cts; 10173 if (bbr->r_ctl.r_app_limited_until) 10174 rsm->r_app_limited = 1; 10175 else 10176 rsm->r_app_limited = 0; 10177 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 10178 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 10179 rsm->r_in_tmap = 1; 10180 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 10181 rsm->r_bbr_state = bbr_state_val(bbr); 10182 else 10183 rsm->r_bbr_state = 8; 10184 } 10185 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) 10186 bbr->bbr_use_rack_cheat = 1; 10187 if (bbr_incr_timers && (bbr->rc_use_google == 0)) 10188 bbr->r_ctl.rc_incr_tmrs = 1; 10189 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) 10190 bbr->r_ctl.rc_inc_tcp_oh = 1; 10191 if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) 10192 bbr->r_ctl.rc_inc_ip_oh = 1; 10193 if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) 10194 bbr->r_ctl.rc_inc_enet_oh = 1; 10195 10196 bbr_log_type_statechange(bbr, cts, __LINE__); 10197 if (TCPS_HAVEESTABLISHED(tp->t_state) && 10198 (tp->t_srtt)) { 10199 uint32_t rtt; 10200 10201 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 10202 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 10203 } 10204 /* announce the settings and state */ 10205 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); 10206 tcp_bbr_tso_size_check(bbr, cts); 10207 /* 10208 * Now call the generic function to start a timer. This will place 10209 * the TCB on the hptsi wheel if a timer is needed with appropriate 10210 * flags. 10211 */ 10212 bbr_stop_all_timers(tp); 10213 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); 10214 return (0); 10215 } 10216 10217 /* 10218 * Return 0 if we can accept the connection. Return 10219 * non-zero if we can't handle the connection. A EAGAIN 10220 * means you need to wait until the connection is up. 10221 * a EADDRNOTAVAIL means we can never handle the connection 10222 * (no SACK). 10223 */ 10224 static int 10225 bbr_handoff_ok(struct tcpcb *tp) 10226 { 10227 if ((tp->t_state == TCPS_CLOSED) || 10228 (tp->t_state == TCPS_LISTEN)) { 10229 /* Sure no problem though it may not stick */ 10230 return (0); 10231 } 10232 if ((tp->t_state == TCPS_SYN_SENT) || 10233 (tp->t_state == TCPS_SYN_RECEIVED)) { 10234 /* 10235 * We really don't know you have to get to ESTAB or beyond 10236 * to tell. 10237 */ 10238 return (EAGAIN); 10239 } 10240 if (tp->t_flags & TF_SENTFIN) 10241 return (EINVAL); 10242 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { 10243 return (0); 10244 } 10245 /* 10246 * If we reach here we don't do SACK on this connection so we can 10247 * never do rack. 10248 */ 10249 return (EINVAL); 10250 } 10251 10252 static void 10253 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) 10254 { 10255 if (tp->t_fb_ptr) { 10256 uint32_t calc; 10257 struct tcp_bbr *bbr; 10258 struct bbr_sendmap *rsm; 10259 10260 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10261 if (bbr->r_ctl.crte) 10262 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 10263 bbr_log_flowend(bbr); 10264 bbr->rc_tp = NULL; 10265 if (tp->t_inpcb) { 10266 /* Backout any flags2 we applied */ 10267 tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN; 10268 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; 10269 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 10270 } 10271 if (bbr->bbr_hdrw_pacing) 10272 counter_u64_add(bbr_flows_whdwr_pacing, -1); 10273 else 10274 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 10275 if (bbr->r_ctl.crte != NULL) { 10276 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 10277 bbr->r_ctl.crte = NULL; 10278 } 10279 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10280 while (rsm) { 10281 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 10282 uma_zfree(bbr_zone, rsm); 10283 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10284 } 10285 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10286 while (rsm) { 10287 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 10288 uma_zfree(bbr_zone, rsm); 10289 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10290 } 10291 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; 10292 if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) 10293 BBR_STAT_INC(bbr_dynamic_rwnd); 10294 else 10295 BBR_STAT_INC(bbr_static_rwnd); 10296 bbr->r_ctl.rc_free_cnt = 0; 10297 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10298 tp->t_fb_ptr = NULL; 10299 } 10300 /* Make sure snd_nxt is correctly set */ 10301 tp->snd_nxt = tp->snd_max; 10302 } 10303 10304 static void 10305 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) 10306 { 10307 switch (tp->t_state) { 10308 case TCPS_SYN_SENT: 10309 bbr->r_state = TCPS_SYN_SENT; 10310 bbr->r_substate = bbr_do_syn_sent; 10311 break; 10312 case TCPS_SYN_RECEIVED: 10313 bbr->r_state = TCPS_SYN_RECEIVED; 10314 bbr->r_substate = bbr_do_syn_recv; 10315 break; 10316 case TCPS_ESTABLISHED: 10317 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); 10318 bbr->r_state = TCPS_ESTABLISHED; 10319 bbr->r_substate = bbr_do_established; 10320 break; 10321 case TCPS_CLOSE_WAIT: 10322 bbr->r_state = TCPS_CLOSE_WAIT; 10323 bbr->r_substate = bbr_do_close_wait; 10324 break; 10325 case TCPS_FIN_WAIT_1: 10326 bbr->r_state = TCPS_FIN_WAIT_1; 10327 bbr->r_substate = bbr_do_fin_wait_1; 10328 break; 10329 case TCPS_CLOSING: 10330 bbr->r_state = TCPS_CLOSING; 10331 bbr->r_substate = bbr_do_closing; 10332 break; 10333 case TCPS_LAST_ACK: 10334 bbr->r_state = TCPS_LAST_ACK; 10335 bbr->r_substate = bbr_do_lastack; 10336 break; 10337 case TCPS_FIN_WAIT_2: 10338 bbr->r_state = TCPS_FIN_WAIT_2; 10339 bbr->r_substate = bbr_do_fin_wait_2; 10340 break; 10341 case TCPS_LISTEN: 10342 case TCPS_CLOSED: 10343 case TCPS_TIME_WAIT: 10344 default: 10345 break; 10346 }; 10347 } 10348 10349 static void 10350 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) 10351 { 10352 /* 10353 * Now what state are we going into now? Is there adjustments 10354 * needed? 10355 */ 10356 int32_t old_state; 10357 10358 old_state = bbr_state_val(bbr); 10359 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { 10360 /* Save the lowest srtt we saw in our end of the sub-state */ 10361 bbr->rc_hit_state_1 = 0; 10362 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) 10363 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; 10364 } 10365 bbr->rc_bbr_substate++; 10366 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { 10367 /* Cycle back to first state-> gain */ 10368 bbr->rc_bbr_substate = 0; 10369 } 10370 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10371 /* 10372 * We enter the gain(5/4) cycle (possibly less if 10373 * shallow buffer detection is enabled) 10374 */ 10375 if (bbr->skip_gain) { 10376 /* 10377 * Hardware pacing has set our rate to 10378 * the max and limited our b/w just 10379 * do level i.e. no gain. 10380 */ 10381 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; 10382 } else if (bbr->gain_is_limited && 10383 bbr->bbr_hdrw_pacing && 10384 bbr->r_ctl.crte) { 10385 /* 10386 * We can't gain above the hardware pacing 10387 * rate which is less than our rate + the gain 10388 * calculate the gain needed to reach the hardware 10389 * pacing rate.. 10390 */ 10391 uint64_t bw, rate, gain_calc; 10392 10393 bw = bbr_get_bw(bbr); 10394 rate = bbr->r_ctl.crte->rate; 10395 if ((rate > bw) && 10396 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { 10397 gain_calc = (rate * BBR_UNIT) / bw; 10398 if (gain_calc < BBR_UNIT) 10399 gain_calc = BBR_UNIT; 10400 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; 10401 } else { 10402 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10403 } 10404 } else 10405 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10406 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { 10407 bbr->r_ctl.rc_bbr_state_atflight = cts; 10408 } else 10409 bbr->r_ctl.rc_bbr_state_atflight = 0; 10410 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10411 bbr->rc_hit_state_1 = 1; 10412 bbr->r_ctl.rc_exta_time_gd = 0; 10413 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10414 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10415 if (bbr_state_drain_2_tar) { 10416 bbr->r_ctl.rc_bbr_state_atflight = 0; 10417 } else 10418 bbr->r_ctl.rc_bbr_state_atflight = cts; 10419 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; 10420 } else { 10421 /* All other cycles hit here 2-7 */ 10422 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { 10423 if (bbr_sub_drain_slam_cwnd && 10424 (bbr->rc_use_google == 0) && 10425 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10426 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10427 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10428 } 10429 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) 10430 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - 10431 bbr_get_rtt(bbr, BBR_RTT_PROP)); 10432 else 10433 bbr->r_ctl.rc_exta_time_gd = 0; 10434 if (bbr->r_ctl.rc_exta_time_gd) { 10435 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; 10436 /* Now chop up the time for each state (div by 7) */ 10437 bbr->r_ctl.rc_level_state_extra /= 7; 10438 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { 10439 /* Add a randomization */ 10440 bbr_randomize_extra_state_time(bbr); 10441 } 10442 } 10443 } 10444 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10445 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; 10446 } 10447 if (bbr->rc_use_google) { 10448 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10449 } 10450 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10451 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10452 if (dolog) 10453 bbr_log_type_statechange(bbr, cts, line); 10454 10455 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10456 uint32_t time_in; 10457 10458 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10459 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 10460 counter_u64_add(bbr_state_time[(old_state + 5)], time_in); 10461 } else { 10462 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10463 } 10464 } 10465 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 10466 bbr_set_state_target(bbr, __LINE__); 10467 if (bbr_sub_drain_slam_cwnd && 10468 (bbr->rc_use_google == 0) && 10469 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10470 /* Slam down the cwnd */ 10471 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10472 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10473 if (bbr_sub_drain_app_limit) { 10474 /* Go app limited if we are on a long drain */ 10475 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + 10476 ctf_flight_size(bbr->rc_tp, 10477 (bbr->r_ctl.rc_sacked + 10478 bbr->r_ctl.rc_lost_bytes))); 10479 } 10480 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10481 } 10482 if (bbr->rc_lt_use_bw) { 10483 /* In policed mode we clamp pacing_gain to BBR_UNIT */ 10484 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10485 } 10486 /* Google changes TSO size every cycle */ 10487 if (bbr->rc_use_google) 10488 tcp_bbr_tso_size_check(bbr, cts); 10489 bbr->r_ctl.gain_epoch = cts; 10490 bbr->r_ctl.rc_bbr_state_time = cts; 10491 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; 10492 } 10493 10494 static void 10495 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10496 { 10497 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && 10498 (google_allow_early_out == 1) && 10499 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { 10500 /* We have reached out target flight size possibly early */ 10501 goto change_state; 10502 } 10503 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10504 return; 10505 } 10506 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { 10507 /* 10508 * Must be a rttProp movement forward before 10509 * we can change states. 10510 */ 10511 return; 10512 } 10513 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10514 /* 10515 * The needed time has passed but for 10516 * the gain cycle extra rules apply: 10517 * 1) If we have seen loss, we exit 10518 * 2) If we have not reached the target 10519 * we stay in GAIN (gain-to-target). 10520 */ 10521 if (google_consider_lost && losses) 10522 goto change_state; 10523 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { 10524 return; 10525 } 10526 } 10527 change_state: 10528 /* For gain we must reach our target, all others last 1 rttProp */ 10529 bbr_substate_change(bbr, cts, __LINE__, 1); 10530 } 10531 10532 static void 10533 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10534 { 10535 uint32_t flight, bbr_cur_cycle_time; 10536 10537 if (bbr->rc_use_google) { 10538 bbr_set_probebw_google_gains(bbr, cts, losses); 10539 return; 10540 } 10541 if (cts == 0) { 10542 /* 10543 * Never alow cts to be 0 we 10544 * do this so we can judge if 10545 * we have set a timestamp. 10546 */ 10547 cts = 1; 10548 } 10549 if (bbr_state_is_pkt_epoch) 10550 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 10551 else 10552 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); 10553 10554 if (bbr->r_ctl.rc_bbr_state_atflight == 0) { 10555 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10556 flight = ctf_flight_size(bbr->rc_tp, 10557 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10558 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { 10559 /* Keep it slam down */ 10560 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { 10561 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10562 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10563 } 10564 if (bbr_sub_drain_app_limit) { 10565 /* Go app limited if we are on a long drain */ 10566 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); 10567 } 10568 } 10569 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && 10570 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || 10571 (flight >= bbr->r_ctl.flightsize_at_drain))) { 10572 /* 10573 * Still here after the same time as 10574 * the gain. We need to drain harder 10575 * for the next srtt. Reduce by a set amount 10576 * the gain drop is capped at DRAIN states 10577 * value (88). 10578 */ 10579 bbr->r_ctl.flightsize_at_drain = flight; 10580 if (bbr_drain_drop_mul && 10581 bbr_drain_drop_div && 10582 (bbr_drain_drop_mul < bbr_drain_drop_div)) { 10583 /* Use your specific drop value (def 4/5 = 20%) */ 10584 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; 10585 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; 10586 } else { 10587 /* You get drop of 20% */ 10588 bbr->r_ctl.rc_bbr_hptsi_gain *= 4; 10589 bbr->r_ctl.rc_bbr_hptsi_gain /= 5; 10590 } 10591 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { 10592 /* Reduce our gain again to the bottom */ 10593 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 10594 } 10595 bbr_log_exit_gain(bbr, cts, 4); 10596 /* 10597 * Extend out so we wait another 10598 * epoch before dropping again. 10599 */ 10600 bbr->r_ctl.gain_epoch = cts; 10601 } 10602 if (flight <= bbr->r_ctl.rc_target_at_state) { 10603 if (bbr_sub_drain_slam_cwnd && 10604 (bbr->rc_use_google == 0) && 10605 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10606 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10607 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10608 } 10609 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10610 bbr_log_exit_gain(bbr, cts, 3); 10611 } 10612 } else { 10613 /* Its a gain */ 10614 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { 10615 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10616 goto change_state; 10617 } 10618 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || 10619 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= 10620 bbr->rc_tp->snd_wnd)) { 10621 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10622 bbr_log_exit_gain(bbr, cts, 2); 10623 } 10624 } 10625 /** 10626 * We fall through and return always one of two things has 10627 * occurred. 10628 * 1) We are still not at target 10629 * <or> 10630 * 2) We reached the target and set rc_bbr_state_atflight 10631 * which means we no longer hit this block 10632 * next time we are called. 10633 */ 10634 return; 10635 } 10636 change_state: 10637 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) 10638 return; 10639 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { 10640 /* Less than a full time-period has passed */ 10641 return; 10642 } 10643 if (bbr->r_ctl.rc_level_state_extra && 10644 (bbr_state_val(bbr) > BBR_SUB_DRAIN) && 10645 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10646 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10647 /* Less than a full time-period + extra has passed */ 10648 return; 10649 } 10650 if (bbr_gain_gets_extra_too && 10651 bbr->r_ctl.rc_level_state_extra && 10652 (bbr_state_val(bbr) == BBR_SUB_GAIN) && 10653 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10654 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10655 /* Less than a full time-period + extra has passed */ 10656 return; 10657 } 10658 bbr_substate_change(bbr, cts, __LINE__, 1); 10659 } 10660 10661 static uint32_t 10662 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) 10663 { 10664 uint32_t mss, tar; 10665 10666 if (bbr->rc_use_google) { 10667 /* Google just uses the cwnd target */ 10668 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); 10669 } else { 10670 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 10671 bbr->r_ctl.rc_pace_max_segs); 10672 /* Get the base cwnd with gain rounded to a mss */ 10673 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), 10674 gain), mss); 10675 /* Make sure it is within our min */ 10676 if (tar < get_min_cwnd(bbr)) 10677 return (get_min_cwnd(bbr)); 10678 } 10679 return (tar); 10680 } 10681 10682 static void 10683 bbr_set_state_target(struct tcp_bbr *bbr, int line) 10684 { 10685 uint32_t tar, meth; 10686 10687 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 10688 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 10689 /* Special case using old probe-rtt method */ 10690 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10691 meth = 1; 10692 } else { 10693 /* Non-probe-rtt case and reduced probe-rtt */ 10694 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 10695 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { 10696 /* For gain cycle we use the hptsi gain */ 10697 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10698 meth = 2; 10699 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { 10700 /* 10701 * If configured, or for google all other states 10702 * get BBR_UNIT. 10703 */ 10704 tar = bbr_get_a_state_target(bbr, BBR_UNIT); 10705 meth = 3; 10706 } else { 10707 /* 10708 * Or we set a target based on the pacing gain 10709 * for non-google mode and default (non-configured). 10710 * Note we don't set a target goal below drain (192). 10711 */ 10712 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { 10713 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); 10714 meth = 4; 10715 } else { 10716 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10717 meth = 5; 10718 } 10719 } 10720 } 10721 bbr_log_set_of_state_target(bbr, tar, line, meth); 10722 bbr->r_ctl.rc_target_at_state = tar; 10723 } 10724 10725 static void 10726 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 10727 { 10728 /* Change to probe_rtt */ 10729 uint32_t time_in; 10730 10731 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10732 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10733 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10734 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain 10735 + bbr->r_ctl.rc_delivered); 10736 /* Setup so we force feed the filter */ 10737 if (bbr->rc_use_google || bbr_probertt_sets_rtt) 10738 bbr->rc_prtt_set_ts = 1; 10739 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10740 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10741 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10742 } 10743 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); 10744 bbr->r_ctl.rc_rtt_shrinks = cts; 10745 bbr->r_ctl.last_in_probertt = cts; 10746 bbr->r_ctl.rc_probertt_srttchktim = cts; 10747 bbr->r_ctl.rc_bbr_state_time = cts; 10748 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; 10749 /* We need to force the filter to update */ 10750 10751 if ((bbr_sub_drain_slam_cwnd) && 10752 bbr->rc_hit_state_1 && 10753 (bbr->rc_use_google == 0) && 10754 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10755 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) 10756 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10757 } else 10758 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10759 /* Update the lost */ 10760 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10761 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ 10762 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ 10763 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10764 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10765 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10766 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10767 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); 10768 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; 10769 } else { 10770 /* 10771 * We bring it down slowly by using a hptsi gain that is 10772 * probably 75%. This will slowly float down our outstanding 10773 * without tampering with the cwnd. 10774 */ 10775 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 10776 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10777 bbr_set_state_target(bbr, __LINE__); 10778 if (bbr_prtt_slam_cwnd && 10779 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 10780 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10781 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10782 } 10783 } 10784 if (ctf_flight_size(bbr->rc_tp, 10785 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 10786 bbr->r_ctl.rc_target_at_state) { 10787 /* We are at target */ 10788 bbr->r_ctl.rc_bbr_enters_probertt = cts; 10789 } else { 10790 /* We need to come down to reach target before our time begins */ 10791 bbr->r_ctl.rc_bbr_enters_probertt = 0; 10792 } 10793 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; 10794 BBR_STAT_INC(bbr_enter_probertt); 10795 bbr_log_exit_gain(bbr, cts, 0); 10796 bbr_log_type_statechange(bbr, cts, line); 10797 } 10798 10799 static void 10800 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) 10801 { 10802 /* 10803 * Sanity check on probe-rtt intervals. 10804 * In crazy situations where we are competing 10805 * against new-reno flows with huge buffers 10806 * our rtt-prop interval could come to dominate 10807 * things if we can't get through a full set 10808 * of cycles, we need to adjust it. 10809 */ 10810 if (bbr_can_adjust_probertt && 10811 (bbr->rc_use_google == 0)) { 10812 uint16_t val = 0; 10813 uint32_t cur_rttp, fval, newval, baseval; 10814 10815 /* Are we to small and go into probe-rtt to often? */ 10816 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); 10817 cur_rttp = roundup(baseval, USECS_IN_SECOND); 10818 fval = bbr_filter_len_sec * USECS_IN_SECOND; 10819 if (bbr_is_ratio == 0) { 10820 if (fval > bbr_rtt_probe_limit) 10821 newval = cur_rttp + (fval - bbr_rtt_probe_limit); 10822 else 10823 newval = cur_rttp; 10824 } else { 10825 int mul; 10826 10827 mul = fval / bbr_rtt_probe_limit; 10828 newval = cur_rttp * mul; 10829 } 10830 if (cur_rttp > bbr->r_ctl.rc_probertt_int) { 10831 bbr->r_ctl.rc_probertt_int = cur_rttp; 10832 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10833 val = 1; 10834 } else { 10835 /* 10836 * No adjustments were made 10837 * do we need to shrink it? 10838 */ 10839 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { 10840 if (cur_rttp <= bbr_rtt_probe_limit) { 10841 /* 10842 * Things have calmed down lets 10843 * shrink all the way to default 10844 */ 10845 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10846 reset_time_small(&bbr->r_ctl.rc_rttprop, 10847 (bbr_filter_len_sec * USECS_IN_SECOND)); 10848 cur_rttp = bbr_rtt_probe_limit; 10849 newval = (bbr_filter_len_sec * USECS_IN_SECOND); 10850 val = 2; 10851 } else { 10852 /* 10853 * Well does some adjustment make sense? 10854 */ 10855 if (cur_rttp < bbr->r_ctl.rc_probertt_int) { 10856 /* We can reduce interval time some */ 10857 bbr->r_ctl.rc_probertt_int = cur_rttp; 10858 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10859 val = 3; 10860 } 10861 } 10862 } 10863 } 10864 if (val) 10865 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); 10866 } 10867 } 10868 10869 static void 10870 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 10871 { 10872 /* Exit probe-rtt */ 10873 10874 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { 10875 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10876 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10877 } 10878 bbr_log_exit_gain(bbr, cts, 1); 10879 bbr->rc_hit_state_1 = 0; 10880 bbr->r_ctl.rc_rtt_shrinks = cts; 10881 bbr->r_ctl.last_in_probertt = cts; 10882 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); 10883 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10884 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, 10885 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 10886 bbr->r_ctl.rc_delivered); 10887 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10888 uint32_t time_in; 10889 10890 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10891 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10892 } 10893 if (bbr->rc_filled_pipe) { 10894 /* Switch to probe_bw */ 10895 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 10896 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 10897 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10898 bbr_substate_change(bbr, cts, __LINE__, 0); 10899 bbr_log_type_statechange(bbr, cts, __LINE__); 10900 } else { 10901 /* Back to startup */ 10902 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10903 bbr->r_ctl.rc_bbr_state_time = cts; 10904 /* 10905 * We don't want to give a complete free 3 10906 * measurements until we exit, so we use 10907 * the number of pe's we were in probe-rtt 10908 * to add to the startup_epoch. That way 10909 * we will still retain the old state. 10910 */ 10911 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); 10912 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10913 /* Make sure to use the lower pg when shifting back in */ 10914 if (bbr->r_ctl.rc_lost && 10915 bbr_use_lower_gain_in_startup && 10916 (bbr->rc_use_google == 0)) 10917 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10918 else 10919 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 10920 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 10921 /* Probably not needed but set it anyway */ 10922 bbr_set_state_target(bbr, __LINE__); 10923 bbr_log_type_statechange(bbr, cts, __LINE__); 10924 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10925 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); 10926 } 10927 bbr_check_probe_rtt_limits(bbr, cts); 10928 } 10929 10930 static int32_t inline 10931 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) 10932 { 10933 if ((bbr->rc_past_init_win == 1) && 10934 (bbr->rc_in_persist == 0) && 10935 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { 10936 return (1); 10937 } 10938 if (bbr_can_force_probertt && 10939 (bbr->rc_in_persist == 0) && 10940 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 10941 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 10942 return (1); 10943 } 10944 return (0); 10945 } 10946 10947 static int32_t 10948 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) 10949 { 10950 uint64_t btlbw, gain; 10951 if (pkt_epoch == 0) { 10952 /* 10953 * Need to be on a pkt-epoch to continue. 10954 */ 10955 return (0); 10956 } 10957 btlbw = bbr_get_full_bw(bbr); 10958 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10959 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10960 if (btlbw >= gain) { 10961 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10962 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10963 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10964 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10965 } 10966 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) 10967 return (1); 10968 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10969 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 10970 return(0); 10971 } 10972 10973 static int32_t inline 10974 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) 10975 { 10976 /* Have we gained 25% in the last 3 packet based epoch's? */ 10977 uint64_t btlbw, gain; 10978 int do_exit; 10979 int delta, rtt_gain; 10980 10981 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 10982 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 10983 /* 10984 * This qualifies as a RTT_PROBE session since we drop the 10985 * data outstanding to nothing and waited more than 10986 * bbr_rtt_probe_time. 10987 */ 10988 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 10989 bbr_set_reduced_rtt(bbr, cts, __LINE__); 10990 } 10991 if (bbr_should_enter_probe_rtt(bbr, cts)) { 10992 bbr_enter_probe_rtt(bbr, cts, __LINE__); 10993 return (0); 10994 } 10995 if (bbr->rc_use_google) 10996 return (bbr_google_startup(bbr, cts, pkt_epoch)); 10997 10998 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10999 (bbr_use_lower_gain_in_startup)) { 11000 /* Drop to a lower gain 1.5 x since we saw loss */ 11001 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 11002 } 11003 if (pkt_epoch == 0) { 11004 /* 11005 * Need to be on a pkt-epoch to continue. 11006 */ 11007 return (0); 11008 } 11009 if (bbr_rtt_gain_thresh) { 11010 /* 11011 * Do we allow a flow to stay 11012 * in startup with no loss and no 11013 * gain in rtt over a set threshold? 11014 */ 11015 if (bbr->r_ctl.rc_pkt_epoch_rtt && 11016 bbr->r_ctl.startup_last_srtt && 11017 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { 11018 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; 11019 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; 11020 } else 11021 rtt_gain = 0; 11022 if ((bbr->r_ctl.startup_last_srtt == 0) || 11023 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) 11024 /* First time or new lower value */ 11025 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 11026 11027 if ((bbr->r_ctl.rc_lost == 0) && 11028 (rtt_gain < bbr_rtt_gain_thresh)) { 11029 /* 11030 * No loss, and we are under 11031 * our gain threhold for 11032 * increasing RTT. 11033 */ 11034 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 11035 bbr->r_ctl.rc_bbr_last_startup_epoch++; 11036 bbr_log_startup_event(bbr, cts, rtt_gain, 11037 delta, bbr->r_ctl.startup_last_srtt, 10); 11038 return (0); 11039 } 11040 } 11041 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && 11042 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && 11043 (!IN_RECOVERY(bbr->rc_tp->t_flags))) { 11044 /* 11045 * We only assess if we have a new measurement when 11046 * we have no loss and are not in recovery. 11047 * Drag up by one our last_startup epoch so we will hold 11048 * the number of non-gain we have already accumulated. 11049 */ 11050 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 11051 bbr->r_ctl.rc_bbr_last_startup_epoch++; 11052 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11053 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); 11054 return (0); 11055 } 11056 /* Case where we reduced the lost (bad retransmit) */ 11057 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) 11058 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11059 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; 11060 btlbw = bbr_get_full_bw(bbr); 11061 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) 11062 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 11063 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 11064 else 11065 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 11066 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 11067 do_exit = 0; 11068 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) 11069 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 11070 if (btlbw >= gain) { 11071 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 11072 /* Update the lost so we won't exit in next set of tests */ 11073 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11074 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11075 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 11076 } 11077 if ((bbr->rc_loss_exit && 11078 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 11079 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && 11080 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { 11081 /* 11082 * If we had no gain, we had loss and that loss was above 11083 * our threshould, the rwnd is not constrained, and we have 11084 * had at least 3 packet epochs exit. Note that this is 11085 * switched off by sysctl. Google does not do this by the 11086 * way. 11087 */ 11088 if ((ctf_flight_size(bbr->rc_tp, 11089 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 11090 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { 11091 do_exit = 1; 11092 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11093 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); 11094 } else { 11095 /* Just record an updated loss value */ 11096 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11097 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11098 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); 11099 } 11100 } else 11101 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11102 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || 11103 do_exit) { 11104 /* Return 1 to exit the startup state. */ 11105 return (1); 11106 } 11107 /* Stay in startup */ 11108 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11109 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 11110 return (0); 11111 } 11112 11113 static void 11114 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) 11115 { 11116 /* 11117 * A tick occurred in the rtt epoch do we need to do anything? 11118 */ 11119 #ifdef BBR_INVARIANTS 11120 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && 11121 (bbr->rc_bbr_state != BBR_STATE_DRAIN) && 11122 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && 11123 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 11124 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { 11125 /* Debug code? */ 11126 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); 11127 } 11128 #endif 11129 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 11130 /* Do we exit the startup state? */ 11131 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { 11132 uint32_t time_in; 11133 11134 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11135 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); 11136 bbr->rc_filled_pipe = 1; 11137 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11138 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11139 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11140 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11141 } else 11142 time_in = 0; 11143 if (bbr->rc_no_pacing) 11144 bbr->rc_no_pacing = 0; 11145 bbr->r_ctl.rc_bbr_state_time = cts; 11146 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; 11147 bbr->rc_bbr_state = BBR_STATE_DRAIN; 11148 bbr_set_state_target(bbr, __LINE__); 11149 if ((bbr->rc_use_google == 0) && 11150 bbr_slam_cwnd_in_main_drain) { 11151 /* Here we don't have to worry about probe-rtt */ 11152 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 11153 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11154 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11155 } 11156 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 11157 bbr_log_type_statechange(bbr, cts, __LINE__); 11158 if (ctf_flight_size(bbr->rc_tp, 11159 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 11160 bbr->r_ctl.rc_target_at_state) { 11161 /* 11162 * Switch to probe_bw if we are already 11163 * there 11164 */ 11165 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11166 bbr_substate_change(bbr, cts, __LINE__, 0); 11167 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11168 bbr_log_type_statechange(bbr, cts, __LINE__); 11169 } 11170 } 11171 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { 11172 uint32_t inflight; 11173 struct tcpcb *tp; 11174 11175 tp = bbr->rc_tp; 11176 inflight = ctf_flight_size(tp, 11177 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11178 if (inflight >= bbr->r_ctl.rc_target_at_state) { 11179 /* We have reached a flight of the cwnd target */ 11180 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11181 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11182 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 11183 bbr_set_state_target(bbr, __LINE__); 11184 /* 11185 * Rig it so we don't do anything crazy and 11186 * start fresh with a new randomization. 11187 */ 11188 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 11189 bbr->rc_bbr_substate = BBR_SUB_LEVEL6; 11190 bbr_substate_change(bbr, cts, __LINE__, 1); 11191 } 11192 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { 11193 /* Has in-flight reached the bdp (or less)? */ 11194 uint32_t inflight; 11195 struct tcpcb *tp; 11196 11197 tp = bbr->rc_tp; 11198 inflight = ctf_flight_size(tp, 11199 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11200 if ((bbr->rc_use_google == 0) && 11201 bbr_slam_cwnd_in_main_drain && 11202 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11203 /* 11204 * Here we don't have to worry about probe-rtt 11205 * re-slam it, but keep it slammed down. 11206 */ 11207 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11208 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11209 } 11210 if (inflight <= bbr->r_ctl.rc_target_at_state) { 11211 /* We have drained */ 11212 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11213 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11214 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11215 uint32_t time_in; 11216 11217 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11218 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11219 } 11220 if ((bbr->rc_use_google == 0) && 11221 bbr_slam_cwnd_in_main_drain && 11222 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 11223 /* Restore the cwnd */ 11224 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 11225 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11226 } 11227 /* Setup probe-rtt has being done now RRS-HERE */ 11228 bbr->r_ctl.rc_rtt_shrinks = cts; 11229 bbr->r_ctl.last_in_probertt = cts; 11230 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); 11231 /* Randomly pick a sub-state */ 11232 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11233 bbr_substate_change(bbr, cts, __LINE__, 0); 11234 bbr_log_type_statechange(bbr, cts, __LINE__); 11235 } 11236 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { 11237 uint32_t flight; 11238 11239 flight = ctf_flight_size(bbr->rc_tp, 11240 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11241 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); 11242 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && 11243 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11244 /* 11245 * We must keep cwnd at the desired MSS. 11246 */ 11247 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 11248 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11249 } else if ((bbr_prtt_slam_cwnd) && 11250 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11251 /* Re-slam it */ 11252 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11253 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11254 } 11255 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { 11256 /* Has outstanding reached our target? */ 11257 if (flight <= bbr->r_ctl.rc_target_at_state) { 11258 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); 11259 bbr->r_ctl.rc_bbr_enters_probertt = cts; 11260 /* If time is exactly 0, be 1usec off */ 11261 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) 11262 bbr->r_ctl.rc_bbr_enters_probertt = 1; 11263 if (bbr->rc_use_google == 0) { 11264 /* 11265 * Restore any lowering that as occurred to 11266 * reach here 11267 */ 11268 if (bbr->r_ctl.bbr_rttprobe_gain_val) 11269 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 11270 else 11271 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11272 } 11273 } 11274 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && 11275 (bbr->rc_use_google == 0) && 11276 bbr->r_ctl.bbr_rttprobe_gain_val && 11277 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || 11278 (flight >= bbr->r_ctl.flightsize_at_drain))) { 11279 /* 11280 * We have doddled with our current hptsi 11281 * gain an srtt and have still not made it 11282 * to target, or we have increased our flight. 11283 * Lets reduce the gain by xx% 11284 * flooring the reduce at DRAIN (based on 11285 * mul/div) 11286 */ 11287 int red; 11288 11289 bbr->r_ctl.flightsize_at_drain = flight; 11290 bbr->r_ctl.rc_probertt_srttchktim = cts; 11291 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); 11292 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { 11293 /* Reduce our gain again */ 11294 bbr->r_ctl.rc_bbr_hptsi_gain -= red; 11295 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); 11296 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { 11297 /* one more chance before we give up */ 11298 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 11299 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); 11300 } else { 11301 /* At the very bottom */ 11302 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); 11303 } 11304 } 11305 } 11306 if (bbr->r_ctl.rc_bbr_enters_probertt && 11307 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && 11308 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { 11309 /* Time to exit probe RTT normally */ 11310 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); 11311 } 11312 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 11313 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 11314 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 11315 /* 11316 * This qualifies as a RTT_PROBE session since we 11317 * drop the data outstanding to nothing and waited 11318 * more than bbr_rtt_probe_time. 11319 */ 11320 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 11321 bbr_set_reduced_rtt(bbr, cts, __LINE__); 11322 } 11323 if (bbr_should_enter_probe_rtt(bbr, cts)) { 11324 bbr_enter_probe_rtt(bbr, cts, __LINE__); 11325 } else { 11326 bbr_set_probebw_gains(bbr, cts, losses); 11327 } 11328 } 11329 } 11330 11331 static void 11332 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) 11333 { 11334 int32_t epoch = 0; 11335 11336 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 11337 bbr_set_epoch(bbr, cts, line); 11338 /* At each epoch doe lt bw sampling */ 11339 epoch = 1; 11340 } 11341 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); 11342 } 11343 11344 static int 11345 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, 11346 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, 11347 int32_t nxt_pkt, struct timeval *tv) 11348 { 11349 int32_t thflags, retval; 11350 uint32_t cts, lcts; 11351 uint32_t tiwin; 11352 struct tcpopt to; 11353 struct tcp_bbr *bbr; 11354 struct bbr_sendmap *rsm; 11355 struct timeval ltv; 11356 int32_t did_out = 0; 11357 uint16_t nsegs; 11358 int32_t prev_state; 11359 uint32_t lost; 11360 11361 nsegs = max(1, m->m_pkthdr.lro_nsegs); 11362 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11363 /* add in our stats */ 11364 kern_prefetch(bbr, &prev_state); 11365 prev_state = 0; 11366 thflags = tcp_get_flags(th); 11367 /* 11368 * If this is either a state-changing packet or current state isn't 11369 * established, we require a write lock on tcbinfo. Otherwise, we 11370 * allow the tcbinfo to be in either alocked or unlocked, as the 11371 * caller may have unnecessarily acquired a write lock due to a 11372 * race. 11373 */ 11374 INP_WLOCK_ASSERT(tp->t_inpcb); 11375 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 11376 __func__)); 11377 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 11378 __func__)); 11379 11380 tp->t_rcvtime = ticks; 11381 /* 11382 * Unscale the window into a 32-bit value. For the SYN_SENT state 11383 * the scale is zero. 11384 */ 11385 tiwin = th->th_win << tp->snd_scale; 11386 #ifdef STATS 11387 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 11388 #endif 11389 11390 if (m->m_flags & M_TSTMP) { 11391 /* Prefer the hardware timestamp if present */ 11392 struct timespec ts; 11393 11394 mbuf_tstmp2timespec(m, &ts); 11395 bbr->rc_tv.tv_sec = ts.tv_sec; 11396 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11397 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11398 } else if (m->m_flags & M_TSTMP_LRO) { 11399 /* Next the arrival timestamp */ 11400 struct timespec ts; 11401 11402 mbuf_tstmp2timespec(m, &ts); 11403 bbr->rc_tv.tv_sec = ts.tv_sec; 11404 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11405 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11406 } else { 11407 /* 11408 * Ok just get the current time. 11409 */ 11410 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); 11411 } 11412 /* 11413 * Parse options on any incoming segment. 11414 */ 11415 tcp_dooptions(&to, (u_char *)(th + 1), 11416 (th->th_off << 2) - sizeof(struct tcphdr), 11417 (thflags & TH_SYN) ? TO_SYN : 0); 11418 11419 /* 11420 * If timestamps were negotiated during SYN/ACK and a 11421 * segment without a timestamp is received, silently drop 11422 * the segment, unless it is a RST segment or missing timestamps are 11423 * tolerated. 11424 * See section 3.2 of RFC 7323. 11425 */ 11426 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && 11427 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { 11428 retval = 0; 11429 m_freem(m); 11430 goto done_with_input; 11431 } 11432 /* 11433 * If echoed timestamp is later than the current time, fall back to 11434 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 11435 * were used when this connection was established. 11436 */ 11437 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 11438 to.to_tsecr -= tp->ts_offset; 11439 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) 11440 to.to_tsecr = 0; 11441 } 11442 /* 11443 * If its the first time in we need to take care of options and 11444 * verify we can do SACK for rack! 11445 */ 11446 if (bbr->r_state == 0) { 11447 /* 11448 * Process options only when we get SYN/ACK back. The SYN 11449 * case for incoming connections is handled in tcp_syncache. 11450 * According to RFC1323 the window field in a SYN (i.e., a 11451 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 11452 * this is traditional behavior, may need to be cleaned up. 11453 */ 11454 if (bbr->rc_inp == NULL) { 11455 bbr->rc_inp = tp->t_inpcb; 11456 } 11457 /* 11458 * We need to init rc_inp here since its not init'd when 11459 * bbr_init is called 11460 */ 11461 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 11462 if ((to.to_flags & TOF_SCALE) && 11463 (tp->t_flags & TF_REQ_SCALE)) { 11464 tp->t_flags |= TF_RCVD_SCALE; 11465 tp->snd_scale = to.to_wscale; 11466 } else 11467 tp->t_flags &= ~TF_REQ_SCALE; 11468 /* 11469 * Initial send window. It will be updated with the 11470 * next incoming segment to the scaled value. 11471 */ 11472 tp->snd_wnd = th->th_win; 11473 if ((to.to_flags & TOF_TS) && 11474 (tp->t_flags & TF_REQ_TSTMP)) { 11475 tp->t_flags |= TF_RCVD_TSTMP; 11476 tp->ts_recent = to.to_tsval; 11477 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 11478 } else 11479 tp->t_flags &= ~TF_REQ_TSTMP; 11480 if (to.to_flags & TOF_MSS) 11481 tcp_mss(tp, to.to_mss); 11482 if ((tp->t_flags & TF_SACK_PERMIT) && 11483 (to.to_flags & TOF_SACKPERM) == 0) 11484 tp->t_flags &= ~TF_SACK_PERMIT; 11485 if (IS_FASTOPEN(tp->t_flags)) { 11486 if (to.to_flags & TOF_FASTOPEN) { 11487 uint16_t mss; 11488 11489 if (to.to_flags & TOF_MSS) 11490 mss = to.to_mss; 11491 else 11492 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) 11493 mss = TCP6_MSS; 11494 else 11495 mss = TCP_MSS; 11496 tcp_fastopen_update_cache(tp, mss, 11497 to.to_tfo_len, to.to_tfo_cookie); 11498 } else 11499 tcp_fastopen_disable_path(tp); 11500 } 11501 } 11502 /* 11503 * At this point we are at the initial call. Here we decide 11504 * if we are doing RACK or not. We do this by seeing if 11505 * TF_SACK_PERMIT is set, if not rack is *not* possible and 11506 * we switch to the default code. 11507 */ 11508 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 11509 /* Bail */ 11510 tcp_switch_back_to_default(tp); 11511 (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, 11512 tlen, iptos); 11513 return (1); 11514 } 11515 /* Set the flag */ 11516 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; 11517 tcp_set_hpts(tp->t_inpcb); 11518 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); 11519 } 11520 if (thflags & TH_ACK) { 11521 /* Track ack types */ 11522 if (to.to_flags & TOF_SACK) 11523 BBR_STAT_INC(bbr_acks_with_sacks); 11524 else 11525 BBR_STAT_INC(bbr_plain_acks); 11526 } 11527 /* 11528 * This is the one exception case where we set the rack state 11529 * always. All other times (timers etc) we must have a rack-state 11530 * set (so we assure we have done the checks above for SACK). 11531 */ 11532 if (thflags & TH_FIN) 11533 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); 11534 if (bbr->r_state != tp->t_state) 11535 bbr_set_state(tp, bbr, tiwin); 11536 11537 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) 11538 kern_prefetch(rsm, &prev_state); 11539 prev_state = bbr->r_state; 11540 bbr->rc_ack_was_delayed = 0; 11541 lost = bbr->r_ctl.rc_lost; 11542 bbr->rc_is_pkt_epoch_now = 0; 11543 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { 11544 /* Get the real time into lcts and figure the real delay */ 11545 lcts = tcp_get_usecs(<v); 11546 if (TSTMP_GT(lcts, cts)) { 11547 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; 11548 bbr->rc_ack_was_delayed = 1; 11549 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, 11550 bbr->r_ctl.highest_hdwr_delay)) 11551 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; 11552 } else { 11553 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11554 bbr->rc_ack_was_delayed = 0; 11555 } 11556 } else { 11557 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11558 bbr->rc_ack_was_delayed = 0; 11559 } 11560 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); 11561 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 11562 retval = 0; 11563 m_freem(m); 11564 goto done_with_input; 11565 } 11566 /* 11567 * If a segment with the ACK-bit set arrives in the SYN-SENT state 11568 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 11569 */ 11570 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 11571 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 11572 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 11573 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 11574 return (1); 11575 } 11576 if (tiwin > bbr->r_ctl.rc_high_rwnd) 11577 bbr->r_ctl.rc_high_rwnd = tiwin; 11578 #ifdef BBR_INVARIANTS 11579 if ((tp->t_inpcb->inp_flags & INP_DROPPED) || 11580 (tp->t_inpcb->inp_flags2 & INP_FREED)) { 11581 panic("tp:%p bbr:%p given a dropped inp:%p", 11582 tp, bbr, tp->t_inpcb); 11583 } 11584 #endif 11585 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, 11586 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11587 bbr->rtt_valid = 0; 11588 if (to.to_flags & TOF_TS) { 11589 bbr->rc_ts_valid = 1; 11590 bbr->r_ctl.last_inbound_ts = to.to_tsval; 11591 } else { 11592 bbr->rc_ts_valid = 0; 11593 bbr->r_ctl.last_inbound_ts = 0; 11594 } 11595 retval = (*bbr->r_substate) (m, th, so, 11596 tp, &to, drop_hdrlen, 11597 tlen, tiwin, thflags, nxt_pkt, iptos); 11598 #ifdef BBR_INVARIANTS 11599 if ((retval == 0) && 11600 (tp->t_inpcb == NULL)) { 11601 panic("retval:%d tp:%p t_inpcb:NULL state:%d", 11602 retval, tp, prev_state); 11603 } 11604 #endif 11605 if (nxt_pkt == 0) 11606 BBR_STAT_INC(bbr_rlock_left_ret0); 11607 else 11608 BBR_STAT_INC(bbr_rlock_left_ret1); 11609 if (retval == 0) { 11610 /* 11611 * If retval is 1 the tcb is unlocked and most likely the tp 11612 * is gone. 11613 */ 11614 INP_WLOCK_ASSERT(tp->t_inpcb); 11615 tcp_bbr_xmit_timer_commit(bbr, tp, cts); 11616 if (bbr->rc_is_pkt_epoch_now) 11617 bbr_set_pktepoch(bbr, cts, __LINE__); 11618 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); 11619 if (nxt_pkt == 0) { 11620 if (bbr->r_wanted_output != 0) { 11621 bbr->rc_output_starts_timer = 0; 11622 did_out = 1; 11623 if (tcp_output(tp) < 0) 11624 return (1); 11625 } else 11626 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); 11627 } 11628 if ((nxt_pkt == 0) && 11629 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 11630 (SEQ_GT(tp->snd_max, tp->snd_una) || 11631 (tp->t_flags & TF_DELACK) || 11632 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 11633 (tp->t_state <= TCPS_CLOSING)))) { 11634 /* 11635 * We could not send (probably in the hpts but 11636 * stopped the timer)? 11637 */ 11638 if ((tp->snd_max == tp->snd_una) && 11639 ((tp->t_flags & TF_DELACK) == 0) && 11640 (tcp_in_hpts(bbr->rc_inp)) && 11641 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 11642 /* 11643 * keep alive not needed if we are hptsi 11644 * output yet 11645 */ 11646 ; 11647 } else { 11648 if (tcp_in_hpts(bbr->rc_inp)) { 11649 tcp_hpts_remove(bbr->rc_inp); 11650 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11651 (TSTMP_GT(lcts, bbr->rc_pacer_started))) { 11652 uint32_t del; 11653 11654 del = lcts - bbr->rc_pacer_started; 11655 if (bbr->r_ctl.rc_last_delay_val > del) { 11656 BBR_STAT_INC(bbr_force_timer_start); 11657 bbr->r_ctl.rc_last_delay_val -= del; 11658 bbr->rc_pacer_started = lcts; 11659 } else { 11660 /* We are late */ 11661 bbr->r_ctl.rc_last_delay_val = 0; 11662 BBR_STAT_INC(bbr_force_output); 11663 if (tcp_output(tp) < 0) 11664 return (1); 11665 } 11666 } 11667 } 11668 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 11669 0); 11670 } 11671 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { 11672 /* Do we have the correct timer running? */ 11673 bbr_timer_audit(tp, bbr, lcts, &so->so_snd); 11674 } 11675 /* Do we have a new state */ 11676 if (bbr->r_state != tp->t_state) 11677 bbr_set_state(tp, bbr, tiwin); 11678 done_with_input: 11679 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); 11680 if (did_out) 11681 bbr->r_wanted_output = 0; 11682 #ifdef BBR_INVARIANTS 11683 if (tp->t_inpcb == NULL) { 11684 panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", 11685 did_out, 11686 retval, tp, prev_state); 11687 } 11688 #endif 11689 } 11690 return (retval); 11691 } 11692 11693 static void 11694 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, 11695 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 11696 { 11697 struct timeval tv; 11698 int retval; 11699 11700 /* First lets see if we have old packets */ 11701 if (tp->t_in_pkt) { 11702 if (ctf_do_queued_segments(so, tp, 1)) { 11703 m_freem(m); 11704 return; 11705 } 11706 } 11707 if (m->m_flags & M_TSTMP_LRO) { 11708 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; 11709 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; 11710 } else { 11711 /* Should not be should we kassert instead? */ 11712 tcp_get_usecs(&tv); 11713 } 11714 retval = bbr_do_segment_nounlock(m, th, so, tp, 11715 drop_hdrlen, tlen, iptos, 0, &tv); 11716 if (retval == 0) { 11717 INP_WUNLOCK(tp->t_inpcb); 11718 } 11719 } 11720 11721 /* 11722 * Return how much data can be sent without violating the 11723 * cwnd or rwnd. 11724 */ 11725 11726 static inline uint32_t 11727 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, 11728 uint32_t avail, int32_t sb_offset, uint32_t cts) 11729 { 11730 uint32_t len; 11731 11732 if (ctf_outstanding(tp) >= tp->snd_wnd) { 11733 /* We never want to go over our peers rcv-window */ 11734 len = 0; 11735 } else { 11736 uint32_t flight; 11737 11738 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11739 if (flight >= sendwin) { 11740 /* 11741 * We have in flight what we are allowed by cwnd (if 11742 * it was rwnd blocking it would have hit above out 11743 * >= tp->snd_wnd). 11744 */ 11745 return (0); 11746 } 11747 len = sendwin - flight; 11748 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { 11749 /* We would send too much (beyond the rwnd) */ 11750 len = tp->snd_wnd - ctf_outstanding(tp); 11751 } 11752 if ((len + sb_offset) > avail) { 11753 /* 11754 * We don't have that much in the SB, how much is 11755 * there? 11756 */ 11757 len = avail - sb_offset; 11758 } 11759 } 11760 return (len); 11761 } 11762 11763 static inline void 11764 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11765 { 11766 #ifdef NETFLIX_STATS 11767 KMOD_TCPSTAT_INC(tcps_sndpack_error); 11768 KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); 11769 #endif 11770 } 11771 11772 static inline void 11773 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11774 { 11775 if (error) { 11776 bbr_do_error_accounting(tp, bbr, rsm, len, error); 11777 return; 11778 } 11779 if (rsm) { 11780 if (rsm->r_flags & BBR_TLP) { 11781 /* 11782 * TLP should not count in retran count, but in its 11783 * own bin 11784 */ 11785 #ifdef NETFLIX_STATS 11786 KMOD_TCPSTAT_INC(tcps_tlpresends); 11787 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); 11788 #endif 11789 } else { 11790 /* Retransmit */ 11791 tp->t_sndrexmitpack++; 11792 KMOD_TCPSTAT_INC(tcps_sndrexmitpack); 11793 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); 11794 #ifdef STATS 11795 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 11796 len); 11797 #endif 11798 } 11799 /* 11800 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is 11801 * sub-state 11802 */ 11803 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); 11804 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { 11805 /* Non probe_bw log in 1, 2, or 4. */ 11806 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); 11807 } else { 11808 /* 11809 * Log our probe state 3, and log also 5-13 to show 11810 * us the recovery sub-state for the send. This 11811 * means that 3 == (5+6+7+8+9+10+11+12+13) 11812 */ 11813 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); 11814 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); 11815 } 11816 /* Place in both 16's the totals of retransmitted */ 11817 counter_u64_add(bbr_state_lost[16], len); 11818 counter_u64_add(bbr_state_resend[16], len); 11819 /* Place in 17's the total sent */ 11820 counter_u64_add(bbr_state_resend[17], len); 11821 counter_u64_add(bbr_state_lost[17], len); 11822 11823 } else { 11824 /* New sends */ 11825 KMOD_TCPSTAT_INC(tcps_sndpack); 11826 KMOD_TCPSTAT_ADD(tcps_sndbyte, len); 11827 /* Place in 17's the total sent */ 11828 counter_u64_add(bbr_state_resend[17], len); 11829 counter_u64_add(bbr_state_lost[17], len); 11830 #ifdef STATS 11831 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 11832 len); 11833 #endif 11834 } 11835 } 11836 11837 static void 11838 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) 11839 { 11840 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { 11841 /* 11842 * Limit the cwnd to not be above N x the target plus whats 11843 * is outstanding. The target is based on the current b/w 11844 * estimate. 11845 */ 11846 uint32_t target; 11847 11848 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); 11849 target += ctf_outstanding(tp); 11850 target *= bbr_target_cwnd_mult_limit; 11851 if (tp->snd_cwnd > target) 11852 tp->snd_cwnd = target; 11853 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); 11854 } 11855 } 11856 11857 static int 11858 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) 11859 { 11860 /* 11861 * "adv" is the amount we could increase the window, taking into 11862 * account that we are limited by TCP_MAXWIN << tp->rcv_scale. 11863 */ 11864 int32_t adv; 11865 int32_t oldwin; 11866 11867 adv = recwin; 11868 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 11869 oldwin = (tp->rcv_adv - tp->rcv_nxt); 11870 if (adv > oldwin) 11871 adv -= oldwin; 11872 else { 11873 /* We can't increase the window */ 11874 adv = 0; 11875 } 11876 } else 11877 oldwin = 0; 11878 11879 /* 11880 * If the new window size ends up being the same as or less 11881 * than the old size when it is scaled, then don't force 11882 * a window update. 11883 */ 11884 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 11885 return (0); 11886 11887 if (adv >= (2 * maxseg) && 11888 (adv >= (so->so_rcv.sb_hiwat / 4) || 11889 recwin <= (so->so_rcv.sb_hiwat / 8) || 11890 so->so_rcv.sb_hiwat <= 8 * maxseg)) { 11891 return (1); 11892 } 11893 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 11894 return (1); 11895 return (0); 11896 } 11897 11898 /* 11899 * Return 0 on success and a errno on failure to send. 11900 * Note that a 0 return may not mean we sent anything 11901 * if the TCB was on the hpts. A non-zero return 11902 * does indicate the error we got from ip[6]_output. 11903 */ 11904 static int 11905 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) 11906 { 11907 struct socket *so; 11908 int32_t len; 11909 uint32_t cts; 11910 uint32_t recwin, sendwin; 11911 int32_t sb_offset; 11912 int32_t flags, abandon, error = 0; 11913 struct tcp_log_buffer *lgb = NULL; 11914 struct mbuf *m; 11915 struct mbuf *mb; 11916 uint32_t if_hw_tsomaxsegcount = 0; 11917 uint32_t if_hw_tsomaxsegsize = 0; 11918 uint32_t if_hw_tsomax = 0; 11919 struct ip *ip = NULL; 11920 #ifdef TCPDEBUG 11921 struct ipovly *ipov = NULL; 11922 #endif 11923 struct tcp_bbr *bbr; 11924 struct tcphdr *th; 11925 struct udphdr *udp = NULL; 11926 u_char opt[TCP_MAXOLEN]; 11927 unsigned ipoptlen, optlen, hdrlen; 11928 unsigned ulen; 11929 uint32_t bbr_seq; 11930 uint32_t delay_calc=0; 11931 uint8_t doing_tlp = 0; 11932 uint8_t local_options; 11933 #ifdef BBR_INVARIANTS 11934 uint8_t doing_retran_from = 0; 11935 uint8_t picked_up_retran = 0; 11936 #endif 11937 uint8_t wanted_cookie = 0; 11938 uint8_t more_to_rxt=0; 11939 int32_t prefetch_so_done = 0; 11940 int32_t prefetch_rsm = 0; 11941 uint32_t tot_len = 0; 11942 uint32_t rtr_cnt = 0; 11943 uint32_t maxseg, pace_max_segs, p_maxseg; 11944 int32_t csum_flags = 0; 11945 int32_t hw_tls; 11946 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 11947 unsigned ipsec_optlen = 0; 11948 11949 #endif 11950 volatile int32_t sack_rxmit; 11951 struct bbr_sendmap *rsm = NULL; 11952 int32_t tso, mtu; 11953 struct tcpopt to; 11954 int32_t slot = 0; 11955 struct inpcb *inp; 11956 struct sockbuf *sb; 11957 uint32_t hpts_calling; 11958 #ifdef INET6 11959 struct ip6_hdr *ip6 = NULL; 11960 int32_t isipv6; 11961 #endif 11962 uint8_t app_limited = BBR_JR_SENT_DATA; 11963 uint8_t filled_all = 0; 11964 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11965 /* We take a cache hit here */ 11966 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); 11967 cts = tcp_tv_to_usectick(&bbr->rc_tv); 11968 inp = bbr->rc_inp; 11969 so = inp->inp_socket; 11970 sb = &so->so_snd; 11971 if (sb->sb_flags & SB_TLS_IFNET) 11972 hw_tls = 1; 11973 else 11974 hw_tls = 0; 11975 kern_prefetch(sb, &maxseg); 11976 maxseg = tp->t_maxseg - bbr->rc_last_options; 11977 if (bbr_minseg(bbr) < maxseg) { 11978 tcp_bbr_tso_size_check(bbr, cts); 11979 } 11980 /* Remove any flags that indicate we are pacing on the inp */ 11981 pace_max_segs = bbr->r_ctl.rc_pace_max_segs; 11982 p_maxseg = min(maxseg, pace_max_segs); 11983 INP_WLOCK_ASSERT(inp); 11984 #ifdef TCP_OFFLOAD 11985 if (tp->t_flags & TF_TOE) 11986 return (tcp_offload_output(tp)); 11987 #endif 11988 11989 #ifdef INET6 11990 if (bbr->r_state) { 11991 /* Use the cache line loaded if possible */ 11992 isipv6 = bbr->r_is_v6; 11993 } else { 11994 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 11995 } 11996 #endif 11997 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 11998 tcp_in_hpts(inp)) { 11999 /* 12000 * We are on the hpts for some timer but not hptsi output. 12001 * Possibly remove from the hpts so we can send/recv etc. 12002 */ 12003 if ((tp->t_flags & TF_ACKNOW) == 0) { 12004 /* 12005 * No immediate demand right now to send an ack, but 12006 * the user may have read, making room for new data 12007 * (a window update). If so we may want to cancel 12008 * whatever timer is running (KEEP/DEL-ACK?) and 12009 * continue to send out a window update. Or we may 12010 * have gotten more data into the socket buffer to 12011 * send. 12012 */ 12013 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12014 (long)TCP_MAXWIN << tp->rcv_scale); 12015 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && 12016 ((tcp_outflags[tp->t_state] & TH_RST) == 0) && 12017 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= 12018 (tp->snd_max - tp->snd_una))) { 12019 /* 12020 * Nothing new to send and no window update 12021 * is needed to send. Lets just return and 12022 * let the timer-run off. 12023 */ 12024 return (0); 12025 } 12026 } 12027 tcp_hpts_remove(inp); 12028 bbr_timer_cancel(bbr, __LINE__, cts); 12029 } 12030 if (bbr->r_ctl.rc_last_delay_val) { 12031 /* Calculate a rough delay for early escape to sending */ 12032 if (SEQ_GT(cts, bbr->rc_pacer_started)) 12033 delay_calc = cts - bbr->rc_pacer_started; 12034 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 12035 delay_calc -= bbr->r_ctl.rc_last_delay_val; 12036 else 12037 delay_calc = 0; 12038 } 12039 /* Mark that we have called bbr_output(). */ 12040 if ((bbr->r_timer_override) || 12041 (tp->t_state < TCPS_ESTABLISHED)) { 12042 /* Timeouts or early states are exempt */ 12043 if (tcp_in_hpts(inp)) 12044 tcp_hpts_remove(inp); 12045 } else if (tcp_in_hpts(inp)) { 12046 if ((bbr->r_ctl.rc_last_delay_val) && 12047 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 12048 delay_calc) { 12049 /* 12050 * We were being paced for output and the delay has 12051 * already exceeded when we were supposed to be 12052 * called, lets go ahead and pull out of the hpts 12053 * and call output. 12054 */ 12055 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); 12056 bbr->r_ctl.rc_last_delay_val = 0; 12057 tcp_hpts_remove(inp); 12058 } else if (tp->t_state == TCPS_CLOSED) { 12059 bbr->r_ctl.rc_last_delay_val = 0; 12060 tcp_hpts_remove(inp); 12061 } else { 12062 /* 12063 * On the hpts, you shall not pass! even if ACKNOW 12064 * is on, we will when the hpts fires, unless of 12065 * course we are overdue. 12066 */ 12067 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); 12068 return (0); 12069 } 12070 } 12071 bbr->rc_cwnd_limited = 0; 12072 if (bbr->r_ctl.rc_last_delay_val) { 12073 /* recalculate the real delay and deal with over/under */ 12074 if (SEQ_GT(cts, bbr->rc_pacer_started)) 12075 delay_calc = cts - bbr->rc_pacer_started; 12076 else 12077 delay_calc = 0; 12078 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 12079 /* Setup the delay which will be added in */ 12080 delay_calc -= bbr->r_ctl.rc_last_delay_val; 12081 else { 12082 /* 12083 * We are early setup to adjust 12084 * our slot time. 12085 */ 12086 uint64_t merged_val; 12087 12088 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); 12089 bbr->r_agg_early_set = 1; 12090 if (bbr->r_ctl.rc_hptsi_agg_delay) { 12091 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { 12092 /* Nope our previous late cancels out the early */ 12093 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; 12094 bbr->r_agg_early_set = 0; 12095 bbr->r_ctl.rc_agg_early = 0; 12096 } else { 12097 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; 12098 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12099 } 12100 } 12101 merged_val = bbr->rc_pacer_started; 12102 merged_val <<= 32; 12103 merged_val |= bbr->r_ctl.rc_last_delay_val; 12104 bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, 12105 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, 12106 bbr->r_agg_early_set, 3); 12107 bbr->r_ctl.rc_last_delay_val = 0; 12108 BBR_STAT_INC(bbr_early); 12109 delay_calc = 0; 12110 } 12111 } else { 12112 /* We were not delayed due to hptsi */ 12113 if (bbr->r_agg_early_set) 12114 bbr->r_ctl.rc_agg_early = 0; 12115 bbr->r_agg_early_set = 0; 12116 delay_calc = 0; 12117 } 12118 if (delay_calc) { 12119 /* 12120 * We had a hptsi delay which means we are falling behind on 12121 * sending at the expected rate. Calculate an extra amount 12122 * of data we can send, if any, to put us back on track. 12123 */ 12124 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) 12125 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; 12126 else 12127 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; 12128 } 12129 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12130 if ((tp->snd_una == tp->snd_max) && 12131 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 12132 (sbavail(sb))) { 12133 /* 12134 * Ok we have been idle with nothing outstanding 12135 * we possibly need to start fresh with either a new 12136 * suite of states or a fast-ramp up. 12137 */ 12138 bbr_restart_after_idle(bbr, 12139 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); 12140 } 12141 /* 12142 * Now was there a hptsi delay where we are behind? We only count 12143 * being behind if: a) We are not in recovery. b) There was a delay. 12144 * <and> c) We had room to send something. 12145 * 12146 */ 12147 hpts_calling = inp->inp_hpts_calls; 12148 inp->inp_hpts_calls = 0; 12149 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 12150 int retval; 12151 12152 retval = bbr_process_timers(tp, bbr, cts, hpts_calling); 12153 if (retval != 0) { 12154 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); 12155 /* 12156 * If timers want tcp_drop(), then pass error out, 12157 * otherwise suppress it. 12158 */ 12159 return (retval < 0 ? retval : 0); 12160 } 12161 } 12162 bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; 12163 if (hpts_calling && 12164 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 12165 bbr->r_ctl.rc_last_delay_val = 0; 12166 } 12167 bbr->r_timer_override = 0; 12168 bbr->r_wanted_output = 0; 12169 /* 12170 * For TFO connections in SYN_RECEIVED, only allow the initial 12171 * SYN|ACK and those sent by the retransmit timer. 12172 */ 12173 if (IS_FASTOPEN(tp->t_flags) && 12174 ((tp->t_state == TCPS_SYN_RECEIVED) || 12175 (tp->t_state == TCPS_SYN_SENT)) && 12176 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 12177 (tp->t_rxtshift == 0)) { /* not a retransmit */ 12178 len = 0; 12179 goto just_return_nolock; 12180 } 12181 /* 12182 * Before sending anything check for a state update. For hpts 12183 * calling without input this is important. If its input calling 12184 * then this was already done. 12185 */ 12186 if (bbr->rc_use_google == 0) 12187 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12188 again: 12189 /* 12190 * If we've recently taken a timeout, snd_max will be greater than 12191 * snd_max. BBR in general does not pay much attention to snd_nxt 12192 * for historic reasons the persist timer still uses it. This means 12193 * we have to look at it. All retransmissions that are not persits 12194 * use the rsm that needs to be sent so snd_nxt is ignored. At the 12195 * end of this routine we pull snd_nxt always up to snd_max. 12196 */ 12197 doing_tlp = 0; 12198 #ifdef BBR_INVARIANTS 12199 doing_retran_from = picked_up_retran = 0; 12200 #endif 12201 error = 0; 12202 tso = 0; 12203 slot = 0; 12204 mtu = 0; 12205 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12206 sb_offset = tp->snd_max - tp->snd_una; 12207 flags = tcp_outflags[tp->t_state]; 12208 sack_rxmit = 0; 12209 len = 0; 12210 rsm = NULL; 12211 if (flags & TH_RST) { 12212 SOCKBUF_LOCK(sb); 12213 goto send; 12214 } 12215 recheck_resend: 12216 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 12217 /* We need to always have one in reserve */ 12218 rsm = bbr_alloc(bbr); 12219 if (rsm == NULL) { 12220 error = ENOMEM; 12221 /* Lie to get on the hpts */ 12222 tot_len = tp->t_maxseg; 12223 if (hpts_calling) 12224 /* Retry in a ms */ 12225 slot = 1001; 12226 goto just_return_nolock; 12227 } 12228 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 12229 bbr->r_ctl.rc_free_cnt++; 12230 rsm = NULL; 12231 } 12232 /* What do we send, a resend? */ 12233 if (bbr->r_ctl.rc_resend == NULL) { 12234 /* Check for rack timeout */ 12235 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 12236 if (bbr->r_ctl.rc_resend) { 12237 #ifdef BBR_INVARIANTS 12238 picked_up_retran = 1; 12239 #endif 12240 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); 12241 } 12242 } 12243 if (bbr->r_ctl.rc_resend) { 12244 rsm = bbr->r_ctl.rc_resend; 12245 #ifdef BBR_INVARIANTS 12246 doing_retran_from = 1; 12247 #endif 12248 /* Remove any TLP flags its a RACK or T-O */ 12249 rsm->r_flags &= ~BBR_TLP; 12250 bbr->r_ctl.rc_resend = NULL; 12251 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 12252 #ifdef BBR_INVARIANTS 12253 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", 12254 tp, bbr, rsm, rsm->r_start, tp->snd_una); 12255 goto recheck_resend; 12256 #else 12257 /* TSNH */ 12258 rsm = NULL; 12259 goto recheck_resend; 12260 #endif 12261 } 12262 rtr_cnt++; 12263 if (rsm->r_flags & BBR_HAS_SYN) { 12264 /* Only retransmit a SYN by itself */ 12265 len = 0; 12266 if ((flags & TH_SYN) == 0) { 12267 /* Huh something is wrong */ 12268 rsm->r_start++; 12269 if (rsm->r_start == rsm->r_end) { 12270 /* Clean it up, somehow we missed the ack? */ 12271 bbr_log_syn(tp, NULL); 12272 } else { 12273 /* TFO with data? */ 12274 rsm->r_flags &= ~BBR_HAS_SYN; 12275 len = rsm->r_end - rsm->r_start; 12276 } 12277 } else { 12278 /* Retransmitting SYN */ 12279 rsm = NULL; 12280 SOCKBUF_LOCK(sb); 12281 goto send; 12282 } 12283 } else 12284 len = rsm->r_end - rsm->r_start; 12285 if ((bbr->rc_resends_use_tso == 0) && 12286 (len > maxseg)) { 12287 len = maxseg; 12288 more_to_rxt = 1; 12289 } 12290 sb_offset = rsm->r_start - tp->snd_una; 12291 if (len > 0) { 12292 sack_rxmit = 1; 12293 KMOD_TCPSTAT_INC(tcps_sack_rexmits); 12294 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, 12295 min(len, maxseg)); 12296 } else { 12297 /* I dont think this can happen */ 12298 rsm = NULL; 12299 goto recheck_resend; 12300 } 12301 BBR_STAT_INC(bbr_resends_set); 12302 } else if (bbr->r_ctl.rc_tlp_send) { 12303 /* 12304 * Tail loss probe 12305 */ 12306 doing_tlp = 1; 12307 rsm = bbr->r_ctl.rc_tlp_send; 12308 bbr->r_ctl.rc_tlp_send = NULL; 12309 sack_rxmit = 1; 12310 len = rsm->r_end - rsm->r_start; 12311 rtr_cnt++; 12312 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12313 len = maxseg; 12314 12315 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 12316 #ifdef BBR_INVARIANTS 12317 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", 12318 tp, bbr, tp->snd_una, rsm, rsm->r_start); 12319 #else 12320 /* TSNH */ 12321 rsm = NULL; 12322 goto recheck_resend; 12323 #endif 12324 } 12325 sb_offset = rsm->r_start - tp->snd_una; 12326 BBR_STAT_INC(bbr_tlp_set); 12327 } 12328 /* 12329 * Enforce a connection sendmap count limit if set 12330 * as long as we are not retransmiting. 12331 */ 12332 if ((rsm == NULL) && 12333 (V_tcp_map_entries_limit > 0) && 12334 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 12335 BBR_STAT_INC(bbr_alloc_limited); 12336 if (!bbr->alloc_limit_reported) { 12337 bbr->alloc_limit_reported = 1; 12338 BBR_STAT_INC(bbr_alloc_limited_conns); 12339 } 12340 goto just_return_nolock; 12341 } 12342 #ifdef BBR_INVARIANTS 12343 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { 12344 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", 12345 tp, bbr, rsm, sb_offset, len); 12346 } 12347 #endif 12348 /* 12349 * Get standard flags, and add SYN or FIN if requested by 'hidden' 12350 * state flags. 12351 */ 12352 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) 12353 flags |= TH_FIN; 12354 if (tp->t_flags & TF_NEEDSYN) 12355 flags |= TH_SYN; 12356 12357 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { 12358 /* we are retransmitting the fin */ 12359 len--; 12360 if (len) { 12361 /* 12362 * When retransmitting data do *not* include the 12363 * FIN. This could happen from a TLP probe if we 12364 * allowed data with a FIN. 12365 */ 12366 flags &= ~TH_FIN; 12367 } 12368 } else if (rsm) { 12369 if (flags & TH_FIN) 12370 flags &= ~TH_FIN; 12371 } 12372 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 12373 void *end_rsm; 12374 12375 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 12376 if (end_rsm) 12377 kern_prefetch(end_rsm, &prefetch_rsm); 12378 prefetch_rsm = 1; 12379 } 12380 SOCKBUF_LOCK(sb); 12381 /* 12382 * If snd_nxt == snd_max and we have transmitted a FIN, the 12383 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 12384 * negative length. This can also occur when TCP opens up its 12385 * congestion window while receiving additional duplicate acks after 12386 * fast-retransmit because TCP will reset snd_nxt to snd_max after 12387 * the fast-retransmit. 12388 * 12389 * In the normal retransmit-FIN-only case, however, snd_nxt will be 12390 * set to snd_una, the sb_offset will be 0, and the length may wind 12391 * up 0. 12392 * 12393 * If sack_rxmit is true we are retransmitting from the scoreboard 12394 * in which case len is already set. 12395 */ 12396 if (sack_rxmit == 0) { 12397 uint32_t avail; 12398 12399 avail = sbavail(sb); 12400 if (SEQ_GT(tp->snd_max, tp->snd_una)) 12401 sb_offset = tp->snd_max - tp->snd_una; 12402 else 12403 sb_offset = 0; 12404 if (bbr->rc_tlp_new_data) { 12405 /* TLP is forcing out new data */ 12406 uint32_t tlplen; 12407 12408 doing_tlp = 1; 12409 tlplen = maxseg; 12410 12411 if (tlplen > (uint32_t)(avail - sb_offset)) { 12412 tlplen = (uint32_t)(avail - sb_offset); 12413 } 12414 if (tlplen > tp->snd_wnd) { 12415 len = tp->snd_wnd; 12416 } else { 12417 len = tlplen; 12418 } 12419 bbr->rc_tlp_new_data = 0; 12420 } else { 12421 len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); 12422 if ((len < p_maxseg) && 12423 (bbr->rc_in_persist == 0) && 12424 (ctf_outstanding(tp) >= (2 * p_maxseg)) && 12425 ((avail - sb_offset) >= p_maxseg)) { 12426 /* 12427 * We are not completing whats in the socket 12428 * buffer (i.e. there is at least a segment 12429 * waiting to send) and we have 2 or more 12430 * segments outstanding. There is no sense 12431 * of sending a little piece. Lets defer and 12432 * and wait until we can send a whole 12433 * segment. 12434 */ 12435 len = 0; 12436 } 12437 if (bbr->rc_in_persist) { 12438 /* 12439 * We are in persists, figure out if 12440 * a retransmit is available (maybe the previous 12441 * persists we sent) or if we have to send new 12442 * data. 12443 */ 12444 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 12445 if (rsm) { 12446 len = rsm->r_end - rsm->r_start; 12447 if (rsm->r_flags & BBR_HAS_FIN) 12448 len--; 12449 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12450 len = maxseg; 12451 if (len > 1) 12452 BBR_STAT_INC(bbr_persist_reneg); 12453 /* 12454 * XXXrrs we could force the len to 12455 * 1 byte here to cause the chunk to 12456 * split apart.. but that would then 12457 * mean we always retransmit it as 12458 * one byte even after the window 12459 * opens. 12460 */ 12461 sack_rxmit = 1; 12462 sb_offset = rsm->r_start - tp->snd_una; 12463 } else { 12464 /* 12465 * First time through in persists or peer 12466 * acked our one byte. Though we do have 12467 * to have something in the sb. 12468 */ 12469 len = 1; 12470 sb_offset = 0; 12471 if (avail == 0) 12472 len = 0; 12473 } 12474 } 12475 } 12476 } 12477 if (prefetch_so_done == 0) { 12478 kern_prefetch(so, &prefetch_so_done); 12479 prefetch_so_done = 1; 12480 } 12481 /* 12482 * Lop off SYN bit if it has already been sent. However, if this is 12483 * SYN-SENT state and if segment contains data and if we don't know 12484 * that foreign host supports TAO, suppress sending segment. 12485 */ 12486 if ((flags & TH_SYN) && (rsm == NULL) && 12487 SEQ_GT(tp->snd_max, tp->snd_una)) { 12488 if (tp->t_state != TCPS_SYN_RECEIVED) 12489 flags &= ~TH_SYN; 12490 /* 12491 * When sending additional segments following a TFO SYN|ACK, 12492 * do not include the SYN bit. 12493 */ 12494 if (IS_FASTOPEN(tp->t_flags) && 12495 (tp->t_state == TCPS_SYN_RECEIVED)) 12496 flags &= ~TH_SYN; 12497 sb_offset--, len++; 12498 if (sbavail(sb) == 0) 12499 len = 0; 12500 } else if ((flags & TH_SYN) && rsm) { 12501 /* 12502 * Subtract one from the len for the SYN being 12503 * retransmitted. 12504 */ 12505 len--; 12506 } 12507 /* 12508 * Be careful not to send data and/or FIN on SYN segments. This 12509 * measure is needed to prevent interoperability problems with not 12510 * fully conformant TCP implementations. 12511 */ 12512 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 12513 len = 0; 12514 flags &= ~TH_FIN; 12515 } 12516 /* 12517 * On TFO sockets, ensure no data is sent in the following cases: 12518 * 12519 * - When retransmitting SYN|ACK on a passively-created socket 12520 * - When retransmitting SYN on an actively created socket 12521 * - When sending a zero-length cookie (cookie request) on an 12522 * actively created socket 12523 * - When the socket is in the CLOSED state (RST is being sent) 12524 */ 12525 if (IS_FASTOPEN(tp->t_flags) && 12526 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 12527 ((tp->t_state == TCPS_SYN_SENT) && 12528 (tp->t_tfo_client_cookie_len == 0)) || 12529 (flags & TH_RST))) { 12530 len = 0; 12531 sack_rxmit = 0; 12532 rsm = NULL; 12533 } 12534 /* Without fast-open there should never be data sent on a SYN */ 12535 if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) 12536 len = 0; 12537 if (len <= 0) { 12538 /* 12539 * If FIN has been sent but not acked, but we haven't been 12540 * called to retransmit, len will be < 0. Otherwise, window 12541 * shrank after we sent into it. If window shrank to 0, 12542 * cancel pending retransmit, pull snd_nxt back to (closed) 12543 * window, and set the persist timer if it isn't already 12544 * going. If the window didn't close completely, just wait 12545 * for an ACK. 12546 * 12547 * We also do a general check here to ensure that we will 12548 * set the persist timer when we have data to send, but a 12549 * 0-byte window. This makes sure the persist timer is set 12550 * even if the packet hits one of the "goto send" lines 12551 * below. 12552 */ 12553 len = 0; 12554 if ((tp->snd_wnd == 0) && 12555 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12556 (tp->snd_una == tp->snd_max) && 12557 (sb_offset < (int)sbavail(sb))) { 12558 /* 12559 * Not enough room in the rwnd to send 12560 * a paced segment out. 12561 */ 12562 bbr_enter_persist(tp, bbr, cts, __LINE__); 12563 } 12564 } else if ((rsm == NULL) && 12565 (doing_tlp == 0) && 12566 (len < bbr->r_ctl.rc_pace_max_segs)) { 12567 /* 12568 * We are not sending a full segment for 12569 * some reason. Should we not send anything (think 12570 * sws or persists)? 12571 */ 12572 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12573 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12574 (len < (int)(sbavail(sb) - sb_offset))) { 12575 /* 12576 * Here the rwnd is less than 12577 * the pacing size, this is not a retransmit, 12578 * we are established and 12579 * the send is not the last in the socket buffer 12580 * lets not send, and possibly enter persists. 12581 */ 12582 len = 0; 12583 if (tp->snd_max == tp->snd_una) 12584 bbr_enter_persist(tp, bbr, cts, __LINE__); 12585 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && 12586 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12587 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12588 (len < (int)(sbavail(sb) - sb_offset)) && 12589 (len < bbr_minseg(bbr))) { 12590 /* 12591 * Here we are not retransmitting, and 12592 * the cwnd is not so small that we could 12593 * not send at least a min size (rxt timer 12594 * not having gone off), We have 2 segments or 12595 * more already in flight, its not the tail end 12596 * of the socket buffer and the cwnd is blocking 12597 * us from sending out minimum pacing segment size. 12598 * Lets not send anything. 12599 */ 12600 bbr->rc_cwnd_limited = 1; 12601 len = 0; 12602 } else if (((tp->snd_wnd - ctf_outstanding(tp)) < 12603 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12604 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12605 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12606 (len < (int)(sbavail(sb) - sb_offset)) && 12607 (TCPS_HAVEESTABLISHED(tp->t_state))) { 12608 /* 12609 * Here we have a send window but we have 12610 * filled it up and we can't send another pacing segment. 12611 * We also have in flight more than 2 segments 12612 * and we are not completing the sb i.e. we allow 12613 * the last bytes of the sb to go out even if 12614 * its not a full pacing segment. 12615 */ 12616 len = 0; 12617 } 12618 } 12619 /* len will be >= 0 after this point. */ 12620 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 12621 tcp_sndbuf_autoscale(tp, so, sendwin); 12622 /* 12623 * 12624 */ 12625 if (bbr->rc_in_persist && 12626 len && 12627 (rsm == NULL) && 12628 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { 12629 /* 12630 * We are in persist, not doing a retransmit and don't have enough space 12631 * yet to send a full TSO. So is it at the end of the sb 12632 * if so we need to send else nuke to 0 and don't send. 12633 */ 12634 int sbleft; 12635 if (sbavail(sb) > sb_offset) 12636 sbleft = sbavail(sb) - sb_offset; 12637 else 12638 sbleft = 0; 12639 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { 12640 /* not at end of sb lets not send */ 12641 len = 0; 12642 } 12643 } 12644 /* 12645 * Decide if we can use TCP Segmentation Offloading (if supported by 12646 * hardware). 12647 * 12648 * TSO may only be used if we are in a pure bulk sending state. The 12649 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 12650 * options prevent using TSO. With TSO the TCP header is the same 12651 * (except for the sequence number) for all generated packets. This 12652 * makes it impossible to transmit any options which vary per 12653 * generated segment or packet. 12654 * 12655 * IPv4 handling has a clear separation of ip options and ip header 12656 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() 12657 * does the right thing below to provide length of just ip options 12658 * and thus checking for ipoptlen is enough to decide if ip options 12659 * are present. 12660 */ 12661 #ifdef INET6 12662 if (isipv6) 12663 ipoptlen = ip6_optlen(inp); 12664 else 12665 #endif 12666 if (inp->inp_options) 12667 ipoptlen = inp->inp_options->m_len - 12668 offsetof(struct ipoption, ipopt_list); 12669 else 12670 ipoptlen = 0; 12671 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12672 /* 12673 * Pre-calculate here as we save another lookup into the darknesses 12674 * of IPsec that way and can actually decide if TSO is ok. 12675 */ 12676 #ifdef INET6 12677 if (isipv6 && IPSEC_ENABLED(ipv6)) 12678 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 12679 #ifdef INET 12680 else 12681 #endif 12682 #endif /* INET6 */ 12683 #ifdef INET 12684 if (IPSEC_ENABLED(ipv4)) 12685 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 12686 #endif /* INET */ 12687 #endif /* IPSEC */ 12688 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12689 ipoptlen += ipsec_optlen; 12690 #endif 12691 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 12692 (len > maxseg) && 12693 (tp->t_port == 0) && 12694 ((tp->t_flags & TF_SIGNATURE) == 0) && 12695 tp->rcv_numsacks == 0 && 12696 ipoptlen == 0) 12697 tso = 1; 12698 12699 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12700 (long)TCP_MAXWIN << tp->rcv_scale); 12701 /* 12702 * Sender silly window avoidance. We transmit under the following 12703 * conditions when len is non-zero: 12704 * 12705 * - We have a full segment (or more with TSO) - This is the last 12706 * buffer in a write()/send() and we are either idle or running 12707 * NODELAY - we've timed out (e.g. persist timer) - we have more 12708 * then 1/2 the maximum send window's worth of data (receiver may be 12709 * limited the window size) - we need to retransmit 12710 */ 12711 if (rsm) 12712 goto send; 12713 if (len) { 12714 if (sack_rxmit) 12715 goto send; 12716 if (len >= p_maxseg) 12717 goto send; 12718 /* 12719 * NOTE! on localhost connections an 'ack' from the remote 12720 * end may occur synchronously with the output and cause us 12721 * to flush a buffer queued with moretocome. XXX 12722 * 12723 */ 12724 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ 12725 ((tp->t_flags & TF_NODELAY) || 12726 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && 12727 (tp->t_flags & TF_NOPUSH) == 0) { 12728 goto send; 12729 } 12730 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 12731 goto send; 12732 } 12733 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 12734 goto send; 12735 } 12736 } 12737 /* 12738 * Sending of standalone window updates. 12739 * 12740 * Window updates are important when we close our window due to a 12741 * full socket buffer and are opening it again after the application 12742 * reads data from it. Once the window has opened again and the 12743 * remote end starts to send again the ACK clock takes over and 12744 * provides the most current window information. 12745 * 12746 * We must avoid the silly window syndrome whereas every read from 12747 * the receive buffer, no matter how small, causes a window update 12748 * to be sent. We also should avoid sending a flurry of window 12749 * updates when the socket buffer had queued a lot of data and the 12750 * application is doing small reads. 12751 * 12752 * Prevent a flurry of pointless window updates by only sending an 12753 * update when we can increase the advertized window by more than 12754 * 1/4th of the socket buffer capacity. When the buffer is getting 12755 * full or is very small be more aggressive and send an update 12756 * whenever we can increase by two mss sized segments. In all other 12757 * situations the ACK's to new incoming data will carry further 12758 * window increases. 12759 * 12760 * Don't send an independent window update if a delayed ACK is 12761 * pending (it will get piggy-backed on it) or the remote side 12762 * already has done a half-close and won't send more data. Skip 12763 * this if the connection is in T/TCP half-open state. 12764 */ 12765 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 12766 !(tp->t_flags & TF_DELACK) && 12767 !TCPS_HAVERCVDFIN(tp->t_state)) { 12768 /* Check to see if we should do a window update */ 12769 if (bbr_window_update_needed(tp, so, recwin, maxseg)) 12770 goto send; 12771 } 12772 /* 12773 * Send if we owe the peer an ACK, RST, SYN. ACKNOW 12774 * is also a catch-all for the retransmit timer timeout case. 12775 */ 12776 if (tp->t_flags & TF_ACKNOW) { 12777 goto send; 12778 } 12779 if (flags & TH_RST) { 12780 /* Always send a RST if one is due */ 12781 goto send; 12782 } 12783 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { 12784 goto send; 12785 } 12786 /* 12787 * If our state indicates that FIN should be sent and we have not 12788 * yet done so, then we need to send. 12789 */ 12790 if (flags & TH_FIN && 12791 ((tp->t_flags & TF_SENTFIN) == 0)) { 12792 goto send; 12793 } 12794 /* 12795 * No reason to send a segment, just return. 12796 */ 12797 just_return: 12798 SOCKBUF_UNLOCK(sb); 12799 just_return_nolock: 12800 if (tot_len) 12801 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 12802 if (bbr->rc_no_pacing) 12803 slot = 0; 12804 if (tot_len == 0) { 12805 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= 12806 tp->snd_wnd) { 12807 BBR_STAT_INC(bbr_rwnd_limited); 12808 app_limited = BBR_JR_RWND_LIMITED; 12809 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12810 if ((bbr->rc_in_persist == 0) && 12811 TCPS_HAVEESTABLISHED(tp->t_state) && 12812 (tp->snd_max == tp->snd_una) && 12813 sbavail(&tp->t_inpcb->inp_socket->so_snd)) { 12814 /* No send window.. we must enter persist */ 12815 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 12816 } 12817 } else if (ctf_outstanding(tp) >= sbavail(sb)) { 12818 BBR_STAT_INC(bbr_app_limited); 12819 app_limited = BBR_JR_APP_LIMITED; 12820 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12821 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12822 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { 12823 BBR_STAT_INC(bbr_cwnd_limited); 12824 app_limited = BBR_JR_CWND_LIMITED; 12825 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12826 bbr->r_ctl.rc_lost_bytes))); 12827 bbr->rc_cwnd_limited = 1; 12828 } else { 12829 BBR_STAT_INC(bbr_app_limited); 12830 app_limited = BBR_JR_APP_LIMITED; 12831 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12832 } 12833 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12834 bbr->r_agg_early_set = 0; 12835 bbr->r_ctl.rc_agg_early = 0; 12836 bbr->r_ctl.rc_last_delay_val = 0; 12837 } else if (bbr->rc_use_google == 0) 12838 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12839 /* Are we app limited? */ 12840 if ((app_limited == BBR_JR_APP_LIMITED) || 12841 (app_limited == BBR_JR_RWND_LIMITED)) { 12842 /** 12843 * We are application limited. 12844 */ 12845 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12846 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); 12847 } 12848 if (tot_len == 0) 12849 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); 12850 /* Dont update the time if we did not send */ 12851 bbr->r_ctl.rc_last_delay_val = 0; 12852 bbr->rc_output_starts_timer = 1; 12853 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); 12854 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); 12855 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 12856 /* Make sure snd_nxt is drug up */ 12857 tp->snd_nxt = tp->snd_max; 12858 } 12859 return (error); 12860 12861 send: 12862 if (doing_tlp == 0) { 12863 /* 12864 * Data not a TLP, and its not the rxt firing. If it is the 12865 * rxt firing, we want to leave the tlp_in_progress flag on 12866 * so we don't send another TLP. It has to be a rack timer 12867 * or normal send (response to acked data) to clear the tlp 12868 * in progress flag. 12869 */ 12870 bbr->rc_tlp_in_progress = 0; 12871 bbr->rc_tlp_rtx_out = 0; 12872 } else { 12873 /* 12874 * Its a TLP. 12875 */ 12876 bbr->rc_tlp_in_progress = 1; 12877 } 12878 bbr_timer_cancel(bbr, __LINE__, cts); 12879 if (rsm == NULL) { 12880 if (sbused(sb) > 0) { 12881 /* 12882 * This is sub-optimal. We only send a stand alone 12883 * FIN on its own segment. 12884 */ 12885 if (flags & TH_FIN) { 12886 flags &= ~TH_FIN; 12887 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { 12888 /* Lets not send this */ 12889 slot = 0; 12890 goto just_return; 12891 } 12892 } 12893 } 12894 } else { 12895 /* 12896 * We do *not* send a FIN on a retransmit if it has data. 12897 * The if clause here where len > 1 should never come true. 12898 */ 12899 if ((len > 0) && 12900 (((rsm->r_flags & BBR_HAS_FIN) == 0) && 12901 (flags & TH_FIN))) { 12902 flags &= ~TH_FIN; 12903 len--; 12904 } 12905 } 12906 SOCKBUF_LOCK_ASSERT(sb); 12907 if (len > 0) { 12908 if ((tp->snd_una == tp->snd_max) && 12909 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 12910 /* 12911 * This qualifies as a RTT_PROBE session since we 12912 * drop the data outstanding to nothing and waited 12913 * more than bbr_rtt_probe_time. 12914 */ 12915 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 12916 bbr_set_reduced_rtt(bbr, cts, __LINE__); 12917 } 12918 if (len >= maxseg) 12919 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 12920 else 12921 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 12922 } 12923 /* 12924 * Before ESTABLISHED, force sending of initial options unless TCP 12925 * set not to do any options. NOTE: we assume that the IP/TCP header 12926 * plus TCP options always fit in a single mbuf, leaving room for a 12927 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 12928 * + optlen <= MCLBYTES 12929 */ 12930 optlen = 0; 12931 #ifdef INET6 12932 if (isipv6) 12933 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 12934 else 12935 #endif 12936 hdrlen = sizeof(struct tcpiphdr); 12937 12938 /* 12939 * Compute options for segment. We only have to care about SYN and 12940 * established connection segments. Options for SYN-ACK segments 12941 * are handled in TCP syncache. 12942 */ 12943 to.to_flags = 0; 12944 local_options = 0; 12945 if ((tp->t_flags & TF_NOOPT) == 0) { 12946 /* Maximum segment size. */ 12947 if (flags & TH_SYN) { 12948 to.to_mss = tcp_mssopt(&inp->inp_inc); 12949 if (tp->t_port) 12950 to.to_mss -= V_tcp_udp_tunneling_overhead; 12951 to.to_flags |= TOF_MSS; 12952 /* 12953 * On SYN or SYN|ACK transmits on TFO connections, 12954 * only include the TFO option if it is not a 12955 * retransmit, as the presence of the TFO option may 12956 * have caused the original SYN or SYN|ACK to have 12957 * been dropped by a middlebox. 12958 */ 12959 if (IS_FASTOPEN(tp->t_flags) && 12960 (tp->t_rxtshift == 0)) { 12961 if (tp->t_state == TCPS_SYN_RECEIVED) { 12962 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 12963 to.to_tfo_cookie = 12964 (u_int8_t *)&tp->t_tfo_cookie.server; 12965 to.to_flags |= TOF_FASTOPEN; 12966 wanted_cookie = 1; 12967 } else if (tp->t_state == TCPS_SYN_SENT) { 12968 to.to_tfo_len = 12969 tp->t_tfo_client_cookie_len; 12970 to.to_tfo_cookie = 12971 tp->t_tfo_cookie.client; 12972 to.to_flags |= TOF_FASTOPEN; 12973 wanted_cookie = 1; 12974 } 12975 } 12976 } 12977 /* Window scaling. */ 12978 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 12979 to.to_wscale = tp->request_r_scale; 12980 to.to_flags |= TOF_SCALE; 12981 } 12982 /* Timestamps. */ 12983 if ((tp->t_flags & TF_RCVD_TSTMP) || 12984 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 12985 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; 12986 to.to_tsecr = tp->ts_recent; 12987 to.to_flags |= TOF_TS; 12988 local_options += TCPOLEN_TIMESTAMP + 2; 12989 } 12990 /* Set receive buffer autosizing timestamp. */ 12991 if (tp->rfbuf_ts == 0 && 12992 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 12993 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); 12994 /* Selective ACK's. */ 12995 if (flags & TH_SYN) 12996 to.to_flags |= TOF_SACKPERM; 12997 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 12998 tp->rcv_numsacks > 0) { 12999 to.to_flags |= TOF_SACK; 13000 to.to_nsacks = tp->rcv_numsacks; 13001 to.to_sacks = (u_char *)tp->sackblks; 13002 } 13003 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13004 /* TCP-MD5 (RFC2385). */ 13005 if (tp->t_flags & TF_SIGNATURE) 13006 to.to_flags |= TOF_SIGNATURE; 13007 #endif /* TCP_SIGNATURE */ 13008 13009 /* Processing the options. */ 13010 hdrlen += (optlen = tcp_addoptions(&to, opt)); 13011 /* 13012 * If we wanted a TFO option to be added, but it was unable 13013 * to fit, ensure no data is sent. 13014 */ 13015 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && 13016 !(to.to_flags & TOF_FASTOPEN)) 13017 len = 0; 13018 } 13019 if (tp->t_port) { 13020 if (V_tcp_udp_tunneling_port == 0) { 13021 /* The port was removed?? */ 13022 SOCKBUF_UNLOCK(&so->so_snd); 13023 return (EHOSTUNREACH); 13024 } 13025 hdrlen += sizeof(struct udphdr); 13026 } 13027 #ifdef INET6 13028 if (isipv6) 13029 ipoptlen = ip6_optlen(tp->t_inpcb); 13030 else 13031 #endif 13032 if (tp->t_inpcb->inp_options) 13033 ipoptlen = tp->t_inpcb->inp_options->m_len - 13034 offsetof(struct ipoption, ipopt_list); 13035 else 13036 ipoptlen = 0; 13037 ipoptlen = 0; 13038 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 13039 ipoptlen += ipsec_optlen; 13040 #endif 13041 if (bbr->rc_last_options != local_options) { 13042 /* 13043 * Cache the options length this generally does not change 13044 * on a connection. We use this to calculate TSO. 13045 */ 13046 bbr->rc_last_options = local_options; 13047 } 13048 maxseg = tp->t_maxseg - (ipoptlen + optlen); 13049 p_maxseg = min(maxseg, pace_max_segs); 13050 /* 13051 * Adjust data length if insertion of options will bump the packet 13052 * length beyond the t_maxseg length. Clear the FIN bit because we 13053 * cut off the tail of the segment. 13054 */ 13055 if (len > maxseg) { 13056 if (len != 0 && (flags & TH_FIN)) { 13057 flags &= ~TH_FIN; 13058 } 13059 if (tso) { 13060 uint32_t moff; 13061 int32_t max_len; 13062 13063 /* extract TSO information */ 13064 if_hw_tsomax = tp->t_tsomax; 13065 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 13066 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 13067 KASSERT(ipoptlen == 0, 13068 ("%s: TSO can't do IP options", __func__)); 13069 13070 /* 13071 * Check if we should limit by maximum payload 13072 * length: 13073 */ 13074 if (if_hw_tsomax != 0) { 13075 /* compute maximum TSO length */ 13076 max_len = (if_hw_tsomax - hdrlen - 13077 max_linkhdr); 13078 if (max_len <= 0) { 13079 len = 0; 13080 } else if (len > max_len) { 13081 len = max_len; 13082 } 13083 } 13084 /* 13085 * Prevent the last segment from being fractional 13086 * unless the send sockbuf can be emptied: 13087 */ 13088 if ((sb_offset + len) < sbavail(sb)) { 13089 moff = len % (uint32_t)maxseg; 13090 if (moff != 0) { 13091 len -= moff; 13092 } 13093 } 13094 /* 13095 * In case there are too many small fragments don't 13096 * use TSO: 13097 */ 13098 if (len <= maxseg) { 13099 len = maxseg; 13100 tso = 0; 13101 } 13102 } else { 13103 /* Not doing TSO */ 13104 if (optlen + ipoptlen >= tp->t_maxseg) { 13105 /* 13106 * Since we don't have enough space to put 13107 * the IP header chain and the TCP header in 13108 * one packet as required by RFC 7112, don't 13109 * send it. Also ensure that at least one 13110 * byte of the payload can be put into the 13111 * TCP segment. 13112 */ 13113 SOCKBUF_UNLOCK(&so->so_snd); 13114 error = EMSGSIZE; 13115 sack_rxmit = 0; 13116 goto out; 13117 } 13118 len = maxseg; 13119 } 13120 } else { 13121 /* Not doing TSO */ 13122 if_hw_tsomaxsegcount = 0; 13123 tso = 0; 13124 } 13125 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 13126 ("%s: len > IP_MAXPACKET", __func__)); 13127 #ifdef DIAGNOSTIC 13128 #ifdef INET6 13129 if (max_linkhdr + hdrlen > MCLBYTES) 13130 #else 13131 if (max_linkhdr + hdrlen > MHLEN) 13132 #endif 13133 panic("tcphdr too big"); 13134 #endif 13135 /* 13136 * This KASSERT is here to catch edge cases at a well defined place. 13137 * Before, those had triggered (random) panic conditions further 13138 * down. 13139 */ 13140 #ifdef BBR_INVARIANTS 13141 if (sack_rxmit) { 13142 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 13143 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", 13144 rsm, tp, bbr, rsm->r_start, tp->snd_una); 13145 } 13146 } 13147 #endif 13148 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 13149 if ((len == 0) && 13150 (flags & TH_FIN) && 13151 (sbused(sb))) { 13152 /* 13153 * We have outstanding data, don't send a fin by itself!. 13154 */ 13155 slot = 0; 13156 goto just_return; 13157 } 13158 /* 13159 * Grab a header mbuf, attaching a copy of data to be transmitted, 13160 * and initialize the header from the template for sends on this 13161 * connection. 13162 */ 13163 if (len) { 13164 uint32_t moff; 13165 13166 /* 13167 * We place a limit on sending with hptsi. 13168 */ 13169 if ((rsm == NULL) && len > pace_max_segs) 13170 len = pace_max_segs; 13171 if (len <= maxseg) 13172 tso = 0; 13173 #ifdef INET6 13174 if (MHLEN < hdrlen + max_linkhdr) 13175 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 13176 else 13177 #endif 13178 m = m_gethdr(M_NOWAIT, MT_DATA); 13179 13180 if (m == NULL) { 13181 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13182 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13183 SOCKBUF_UNLOCK(sb); 13184 error = ENOBUFS; 13185 sack_rxmit = 0; 13186 goto out; 13187 } 13188 m->m_data += max_linkhdr; 13189 m->m_len = hdrlen; 13190 /* 13191 * Start the m_copy functions from the closest mbuf to the 13192 * sb_offset in the socket buffer chain. 13193 */ 13194 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { 13195 #ifdef BBR_INVARIANTS 13196 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) 13197 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", 13198 tp, bbr, len, sb_offset, sbavail(sb), rsm, 13199 doing_retran_from, 13200 picked_up_retran, 13201 doing_tlp); 13202 13203 #endif 13204 /* 13205 * In this messed up situation we have two choices, 13206 * a) pretend the send worked, and just start timers 13207 * and what not (not good since that may lead us 13208 * back here a lot). <or> b) Send the lowest segment 13209 * in the map. <or> c) Drop the connection. Lets do 13210 * <b> which if it continues to happen will lead to 13211 * <c> via timeouts. 13212 */ 13213 BBR_STAT_INC(bbr_offset_recovery); 13214 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 13215 sb_offset = 0; 13216 if (rsm == NULL) { 13217 sack_rxmit = 0; 13218 len = sbavail(sb); 13219 } else { 13220 sack_rxmit = 1; 13221 if (rsm->r_start != tp->snd_una) { 13222 /* 13223 * Things are really messed up, <c> 13224 * is the only thing to do. 13225 */ 13226 BBR_STAT_INC(bbr_offset_drop); 13227 SOCKBUF_UNLOCK(sb); 13228 (void)m_free(m); 13229 return (-EFAULT); /* tcp_drop() */ 13230 } 13231 len = rsm->r_end - rsm->r_start; 13232 } 13233 if (len > sbavail(sb)) 13234 len = sbavail(sb); 13235 if (len > maxseg) 13236 len = maxseg; 13237 } 13238 mb = sbsndptr_noadv(sb, sb_offset, &moff); 13239 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 13240 m_copydata(mb, moff, (int)len, 13241 mtod(m, caddr_t)+hdrlen); 13242 if (rsm == NULL) 13243 sbsndptr_adv(sb, mb, len); 13244 m->m_len += len; 13245 } else { 13246 struct sockbuf *msb; 13247 13248 if (rsm) 13249 msb = NULL; 13250 else 13251 msb = sb; 13252 #ifdef BBR_INVARIANTS 13253 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { 13254 if (rsm) { 13255 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 ", 13256 tp, bbr, len, moff, 13257 sbavail(sb), rsm, 13258 tp->snd_una, rsm->r_flags, rsm->r_start, 13259 doing_retran_from, 13260 picked_up_retran, 13261 doing_tlp, sack_rxmit); 13262 } else { 13263 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", 13264 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); 13265 } 13266 } 13267 #endif 13268 m->m_next = tcp_m_copym( 13269 mb, moff, &len, 13270 if_hw_tsomaxsegcount, 13271 if_hw_tsomaxsegsize, msb, 13272 ((rsm == NULL) ? hw_tls : 0) 13273 #ifdef NETFLIX_COPY_ARGS 13274 , &filled_all 13275 #endif 13276 ); 13277 if (len <= maxseg) { 13278 /* 13279 * Must have ran out of mbufs for the copy 13280 * shorten it to no longer need tso. Lets 13281 * not put on sendalot since we are low on 13282 * mbufs. 13283 */ 13284 tso = 0; 13285 } 13286 if (m->m_next == NULL) { 13287 SOCKBUF_UNLOCK(sb); 13288 (void)m_free(m); 13289 error = ENOBUFS; 13290 sack_rxmit = 0; 13291 goto out; 13292 } 13293 } 13294 #ifdef BBR_INVARIANTS 13295 if (tso && len < maxseg) { 13296 panic("tp:%p tso on, but len:%d < maxseg:%d", 13297 tp, len, maxseg); 13298 } 13299 if (tso && if_hw_tsomaxsegcount) { 13300 int32_t seg_cnt = 0; 13301 struct mbuf *foo; 13302 13303 foo = m; 13304 while (foo) { 13305 seg_cnt++; 13306 foo = foo->m_next; 13307 } 13308 if (seg_cnt > if_hw_tsomaxsegcount) { 13309 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); 13310 } 13311 } 13312 #endif 13313 /* 13314 * If we're sending everything we've got, set PUSH. (This 13315 * will keep happy those implementations which only give 13316 * data to the user when a buffer fills or a PUSH comes in.) 13317 */ 13318 if (sb_offset + len == sbused(sb) && 13319 sbused(sb) && 13320 !(flags & TH_SYN)) { 13321 flags |= TH_PUSH; 13322 } 13323 SOCKBUF_UNLOCK(sb); 13324 } else { 13325 SOCKBUF_UNLOCK(sb); 13326 if (tp->t_flags & TF_ACKNOW) 13327 KMOD_TCPSTAT_INC(tcps_sndacks); 13328 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 13329 KMOD_TCPSTAT_INC(tcps_sndctrl); 13330 else 13331 KMOD_TCPSTAT_INC(tcps_sndwinup); 13332 13333 m = m_gethdr(M_NOWAIT, MT_DATA); 13334 if (m == NULL) { 13335 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13336 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13337 error = ENOBUFS; 13338 /* Fudge the send time since we could not send */ 13339 sack_rxmit = 0; 13340 goto out; 13341 } 13342 #ifdef INET6 13343 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 13344 MHLEN >= hdrlen) { 13345 M_ALIGN(m, hdrlen); 13346 } else 13347 #endif 13348 m->m_data += max_linkhdr; 13349 m->m_len = hdrlen; 13350 } 13351 SOCKBUF_UNLOCK_ASSERT(sb); 13352 m->m_pkthdr.rcvif = (struct ifnet *)0; 13353 #ifdef MAC 13354 mac_inpcb_create_mbuf(inp, m); 13355 #endif 13356 #ifdef INET6 13357 if (isipv6) { 13358 ip6 = mtod(m, struct ip6_hdr *); 13359 if (tp->t_port) { 13360 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 13361 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13362 udp->uh_dport = tp->t_port; 13363 ulen = hdrlen + len - sizeof(struct ip6_hdr); 13364 udp->uh_ulen = htons(ulen); 13365 th = (struct tcphdr *)(udp + 1); 13366 } else { 13367 th = (struct tcphdr *)(ip6 + 1); 13368 } 13369 tcpip_fillheaders(inp, tp->t_port, ip6, th); 13370 } else 13371 #endif /* INET6 */ 13372 { 13373 ip = mtod(m, struct ip *); 13374 #ifdef TCPDEBUG 13375 ipov = (struct ipovly *)ip; 13376 #endif 13377 if (tp->t_port) { 13378 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 13379 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13380 udp->uh_dport = tp->t_port; 13381 ulen = hdrlen + len - sizeof(struct ip); 13382 udp->uh_ulen = htons(ulen); 13383 th = (struct tcphdr *)(udp + 1); 13384 } else { 13385 th = (struct tcphdr *)(ip + 1); 13386 } 13387 tcpip_fillheaders(inp, tp->t_port, ip, th); 13388 } 13389 /* 13390 * If we are doing retransmissions, then snd_nxt will not reflect 13391 * the first unsent octet. For ACK only packets, we do not want the 13392 * sequence number of the retransmitted packet, we want the sequence 13393 * number of the next unsent octet. So, if there is no data (and no 13394 * SYN or FIN), use snd_max instead of snd_nxt when filling in 13395 * ti_seq. But if we are in persist state, snd_max might reflect 13396 * one byte beyond the right edge of the window, so use snd_nxt in 13397 * that case, since we know we aren't doing a retransmission. 13398 * (retransmit and persist are mutually exclusive...) 13399 */ 13400 if (sack_rxmit == 0) { 13401 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { 13402 /* New data (including new persists) */ 13403 th->th_seq = htonl(tp->snd_max); 13404 bbr_seq = tp->snd_max; 13405 } else if (flags & TH_SYN) { 13406 /* Syn's always send from iss */ 13407 th->th_seq = htonl(tp->iss); 13408 bbr_seq = tp->iss; 13409 } else if (flags & TH_FIN) { 13410 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { 13411 /* 13412 * If we sent the fin already its 1 minus 13413 * snd_max 13414 */ 13415 th->th_seq = (htonl(tp->snd_max - 1)); 13416 bbr_seq = (tp->snd_max - 1); 13417 } else { 13418 /* First time FIN use snd_max */ 13419 th->th_seq = htonl(tp->snd_max); 13420 bbr_seq = tp->snd_max; 13421 } 13422 } else { 13423 /* 13424 * len == 0 and not persist we use snd_max, sending 13425 * an ack unless we have sent the fin then its 1 13426 * minus. 13427 */ 13428 /* 13429 * XXXRRS Question if we are in persists and we have 13430 * nothing outstanding to send and we have not sent 13431 * a FIN, we will send an ACK. In such a case it 13432 * might be better to send (tp->snd_una - 1) which 13433 * would force the peer to ack. 13434 */ 13435 if (tp->t_flags & TF_SENTFIN) { 13436 th->th_seq = htonl(tp->snd_max - 1); 13437 bbr_seq = (tp->snd_max - 1); 13438 } else { 13439 th->th_seq = htonl(tp->snd_max); 13440 bbr_seq = tp->snd_max; 13441 } 13442 } 13443 } else { 13444 /* All retransmits use the rsm to guide the send */ 13445 th->th_seq = htonl(rsm->r_start); 13446 bbr_seq = rsm->r_start; 13447 } 13448 th->th_ack = htonl(tp->rcv_nxt); 13449 if (optlen) { 13450 bcopy(opt, th + 1, optlen); 13451 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 13452 } 13453 tcp_set_flags(th, flags); 13454 /* 13455 * Calculate receive window. Don't shrink window, but avoid silly 13456 * window syndrome. 13457 */ 13458 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && 13459 recwin < maxseg))) 13460 recwin = 0; 13461 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 13462 recwin < (tp->rcv_adv - tp->rcv_nxt)) 13463 recwin = (tp->rcv_adv - tp->rcv_nxt); 13464 if (recwin > TCP_MAXWIN << tp->rcv_scale) 13465 recwin = TCP_MAXWIN << tp->rcv_scale; 13466 13467 /* 13468 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 13469 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 13470 * handled in syncache. 13471 */ 13472 if (flags & TH_SYN) 13473 th->th_win = htons((u_short) 13474 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 13475 else { 13476 /* Avoid shrinking window with window scaling. */ 13477 recwin = roundup2(recwin, 1 << tp->rcv_scale); 13478 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 13479 } 13480 /* 13481 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 13482 * window. This may cause the remote transmitter to stall. This 13483 * flag tells soreceive() to disable delayed acknowledgements when 13484 * draining the buffer. This can occur if the receiver is 13485 * attempting to read more data than can be buffered prior to 13486 * transmitting on the connection. 13487 */ 13488 if (th->th_win == 0) { 13489 tp->t_sndzerowin++; 13490 tp->t_flags |= TF_RXWIN0SENT; 13491 } else 13492 tp->t_flags &= ~TF_RXWIN0SENT; 13493 /* 13494 * We don't support urgent data, but drag along 13495 * the pointer in case of a stack switch. 13496 */ 13497 tp->snd_up = tp->snd_una; 13498 13499 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13500 if (to.to_flags & TOF_SIGNATURE) { 13501 /* 13502 * Calculate MD5 signature and put it into the place 13503 * determined before. NOTE: since TCP options buffer doesn't 13504 * point into mbuf's data, calculate offset and use it. 13505 */ 13506 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 13507 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 13508 /* 13509 * Do not send segment if the calculation of MD5 13510 * digest has failed. 13511 */ 13512 goto out; 13513 } 13514 } 13515 #endif 13516 13517 /* 13518 * Put TCP length in extended header, and then checksum extended 13519 * header and data. 13520 */ 13521 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 13522 #ifdef INET6 13523 if (isipv6) { 13524 /* 13525 * ip6_plen is not need to be filled now, and will be filled 13526 * in ip6_output. 13527 */ 13528 if (tp->t_port) { 13529 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 13530 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13531 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 13532 th->th_sum = htons(0); 13533 UDPSTAT_INC(udps_opackets); 13534 } else { 13535 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 13536 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13537 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 13538 optlen + len, IPPROTO_TCP, 0); 13539 } 13540 } 13541 #endif 13542 #if defined(INET6) && defined(INET) 13543 else 13544 #endif 13545 #ifdef INET 13546 { 13547 if (tp->t_port) { 13548 m->m_pkthdr.csum_flags = CSUM_UDP; 13549 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13550 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 13551 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 13552 th->th_sum = htons(0); 13553 UDPSTAT_INC(udps_opackets); 13554 } else { 13555 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; 13556 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13557 th->th_sum = in_pseudo(ip->ip_src.s_addr, 13558 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 13559 IPPROTO_TCP + len + optlen)); 13560 } 13561 /* IP version must be set here for ipv4/ipv6 checking later */ 13562 KASSERT(ip->ip_v == IPVERSION, 13563 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 13564 } 13565 #endif 13566 13567 /* 13568 * Enable TSO and specify the size of the segments. The TCP pseudo 13569 * header checksum is always provided. XXX: Fixme: This is currently 13570 * not the case for IPv6. 13571 */ 13572 if (tso) { 13573 KASSERT(len > maxseg, 13574 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); 13575 m->m_pkthdr.csum_flags |= CSUM_TSO; 13576 csum_flags |= CSUM_TSO; 13577 m->m_pkthdr.tso_segsz = maxseg; 13578 } 13579 KASSERT(len + hdrlen == m_length(m, NULL), 13580 ("%s: mbuf chain different than expected: %d + %u != %u", 13581 __func__, len, hdrlen, m_length(m, NULL))); 13582 13583 #ifdef TCP_HHOOK 13584 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ 13585 hhook_run_tcp_est_out(tp, th, &to, len, tso); 13586 #endif 13587 #ifdef TCPDEBUG 13588 /* 13589 * Trace. 13590 */ 13591 if (so->so_options & SO_DEBUG) { 13592 u_short save = 0; 13593 13594 #ifdef INET6 13595 if (!isipv6) 13596 #endif 13597 { 13598 save = ipov->ih_len; 13599 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + 13600 * (th->th_off << 2) */ ); 13601 } 13602 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); 13603 #ifdef INET6 13604 if (!isipv6) 13605 #endif 13606 ipov->ih_len = save; 13607 } 13608 #endif /* TCPDEBUG */ 13609 13610 /* Log to the black box */ 13611 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 13612 union tcp_log_stackspecific log; 13613 13614 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 13615 /* Record info on type of transmission */ 13616 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; 13617 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); 13618 log.u_bbr.flex3 = maxseg; 13619 log.u_bbr.flex4 = delay_calc; 13620 /* Encode filled_all into the upper flex5 bit */ 13621 log.u_bbr.flex5 = bbr->rc_past_init_win; 13622 log.u_bbr.flex5 <<= 1; 13623 log.u_bbr.flex5 |= bbr->rc_no_pacing; 13624 log.u_bbr.flex5 <<= 29; 13625 if (filled_all) 13626 log.u_bbr.flex5 |= 0x80000000; 13627 log.u_bbr.flex5 |= tp->t_maxseg; 13628 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; 13629 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); 13630 /* lets poke in the low and the high here for debugging */ 13631 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 13632 if (rsm || sack_rxmit) { 13633 if (doing_tlp) 13634 log.u_bbr.flex8 = 2; 13635 else 13636 log.u_bbr.flex8 = 1; 13637 } else { 13638 log.u_bbr.flex8 = 0; 13639 } 13640 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 13641 len, &log, false, NULL, NULL, 0, tv); 13642 } else { 13643 lgb = NULL; 13644 } 13645 /* 13646 * Fill in IP length and desired time to live and send to IP level. 13647 * There should be a better way to handle ttl and tos; we could keep 13648 * them in the template, but need a way to checksum without them. 13649 */ 13650 /* 13651 * m->m_pkthdr.len should have been set before cksum calcuration, 13652 * because in6_cksum() need it. 13653 */ 13654 #ifdef INET6 13655 if (isipv6) { 13656 /* 13657 * we separately set hoplimit for every segment, since the 13658 * user might want to change the value via setsockopt. Also, 13659 * desired default hop limit might be changed via Neighbor 13660 * Discovery. 13661 */ 13662 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 13663 13664 /* 13665 * Set the packet size here for the benefit of DTrace 13666 * probes. ip6_output() will set it properly; it's supposed 13667 * to include the option header lengths as well. 13668 */ 13669 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 13670 13671 if (V_path_mtu_discovery && maxseg > V_tcp_minmss) 13672 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13673 else 13674 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13675 13676 if (tp->t_state == TCPS_SYN_SENT) 13677 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 13678 13679 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 13680 /* TODO: IPv6 IP6TOS_ECT bit on */ 13681 error = ip6_output(m, inp->in6p_outputopts, 13682 &inp->inp_route6, 13683 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 13684 NULL, NULL, inp); 13685 13686 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 13687 mtu = inp->inp_route6.ro_nh->nh_mtu; 13688 } 13689 #endif /* INET6 */ 13690 #if defined(INET) && defined(INET6) 13691 else 13692 #endif 13693 #ifdef INET 13694 { 13695 ip->ip_len = htons(m->m_pkthdr.len); 13696 #ifdef INET6 13697 if (isipv6) 13698 ip->ip_ttl = in6_selecthlim(inp, NULL); 13699 #endif /* INET6 */ 13700 /* 13701 * If we do path MTU discovery, then we set DF on every 13702 * packet. This might not be the best thing to do according 13703 * to RFC3390 Section 2. However the tcp hostcache migitates 13704 * the problem so it affects only the first tcp connection 13705 * with a host. 13706 * 13707 * NB: Don't set DF on small MTU/MSS to have a safe 13708 * fallback. 13709 */ 13710 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 13711 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13712 if (tp->t_port == 0 || len < V_tcp_minmss) { 13713 ip->ip_off |= htons(IP_DF); 13714 } 13715 } else { 13716 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13717 } 13718 13719 if (tp->t_state == TCPS_SYN_SENT) 13720 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 13721 13722 TCP_PROBE5(send, NULL, tp, ip, tp, th); 13723 13724 error = ip_output(m, inp->inp_options, &inp->inp_route, 13725 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, 13726 inp); 13727 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 13728 mtu = inp->inp_route.ro_nh->nh_mtu; 13729 } 13730 #endif /* INET */ 13731 out: 13732 13733 if (lgb) { 13734 lgb->tlb_errno = error; 13735 lgb = NULL; 13736 } 13737 /* 13738 * In transmit state, time the transmission and arrange for the 13739 * retransmit. In persist state, just set snd_max. 13740 */ 13741 if (error == 0) { 13742 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); 13743 if (TCPS_HAVEESTABLISHED(tp->t_state) && 13744 (tp->t_flags & TF_SACK_PERMIT) && 13745 tp->rcv_numsacks > 0) 13746 tcp_clean_dsack_blocks(tp); 13747 /* We sent an ack clear the bbr_segs_rcvd count */ 13748 bbr->output_error_seen = 0; 13749 bbr->oerror_cnt = 0; 13750 bbr->bbr_segs_rcvd = 0; 13751 if (len == 0) 13752 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); 13753 /* Do accounting for new sends */ 13754 if ((len > 0) && (rsm == NULL)) { 13755 int idx; 13756 if (tp->snd_una == tp->snd_max) { 13757 /* 13758 * Special case to match google, when 13759 * nothing is in flight the delivered 13760 * time does get updated to the current 13761 * time (see tcp_rate_bsd.c). 13762 */ 13763 bbr->r_ctl.rc_del_time = cts; 13764 } 13765 if (len >= maxseg) { 13766 idx = (len / maxseg) + 3; 13767 if (idx >= TCP_MSS_ACCT_ATIMER) 13768 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); 13769 else 13770 counter_u64_add(bbr_out_size[idx], 1); 13771 } else { 13772 /* smaller than a MSS */ 13773 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); 13774 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) 13775 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); 13776 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); 13777 } 13778 } 13779 } 13780 abandon = 0; 13781 /* 13782 * We must do the send accounting before we log the output, 13783 * otherwise the state of the rsm could change and we account to the 13784 * wrong bucket. 13785 */ 13786 if (len > 0) { 13787 bbr_do_send_accounting(tp, bbr, rsm, len, error); 13788 if (error == 0) { 13789 if (tp->snd_una == tp->snd_max) 13790 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 13791 } 13792 } 13793 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, 13794 cts, mb, &abandon, rsm, 0, sb); 13795 if (abandon) { 13796 /* 13797 * If bbr_log_output destroys the TCB or sees a TH_RST being 13798 * sent we should hit this condition. 13799 */ 13800 return (0); 13801 } 13802 if (bbr->rc_in_persist == 0) { 13803 /* 13804 * Advance snd_nxt over sequence space of this segment. 13805 */ 13806 if (error) 13807 /* We don't log or do anything with errors */ 13808 goto skip_upd; 13809 13810 if (tp->snd_una == tp->snd_max && 13811 (len || (flags & (TH_SYN | TH_FIN)))) { 13812 /* 13813 * Update the time we just added data since none was 13814 * outstanding. 13815 */ 13816 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13817 bbr->rc_tp->t_acktime = ticks; 13818 } 13819 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { 13820 if (flags & TH_SYN) { 13821 /* 13822 * Smack the snd_max to iss + 1 13823 * if its a FO we will add len below. 13824 */ 13825 tp->snd_max = tp->iss + 1; 13826 } 13827 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13828 tp->snd_max++; 13829 tp->t_flags |= TF_SENTFIN; 13830 } 13831 } 13832 if (sack_rxmit == 0) 13833 tp->snd_max += len; 13834 skip_upd: 13835 if ((error == 0) && len) 13836 tot_len += len; 13837 } else { 13838 /* Persists case */ 13839 int32_t xlen = len; 13840 13841 if (error) 13842 goto nomore; 13843 13844 if (flags & TH_SYN) 13845 ++xlen; 13846 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13847 ++xlen; 13848 tp->t_flags |= TF_SENTFIN; 13849 } 13850 if (xlen && (tp->snd_una == tp->snd_max)) { 13851 /* 13852 * Update the time we just added data since none was 13853 * outstanding. 13854 */ 13855 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13856 bbr->rc_tp->t_acktime = ticks; 13857 } 13858 if (sack_rxmit == 0) 13859 tp->snd_max += xlen; 13860 tot_len += (len + optlen + ipoptlen); 13861 } 13862 nomore: 13863 if (error) { 13864 /* 13865 * Failures do not advance the seq counter above. For the 13866 * case of ENOBUFS we will fall out and become ack-clocked. 13867 * capping the cwnd at the current flight. 13868 * Everything else will just have to retransmit with the timer 13869 * (no pacer). 13870 */ 13871 SOCKBUF_UNLOCK_ASSERT(sb); 13872 BBR_STAT_INC(bbr_saw_oerr); 13873 /* Clear all delay/early tracks */ 13874 bbr->r_ctl.rc_hptsi_agg_delay = 0; 13875 bbr->r_ctl.rc_agg_early = 0; 13876 bbr->r_agg_early_set = 0; 13877 bbr->output_error_seen = 1; 13878 if (bbr->oerror_cnt < 0xf) 13879 bbr->oerror_cnt++; 13880 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { 13881 /* drop the session */ 13882 return (-ENETDOWN); 13883 } 13884 switch (error) { 13885 case ENOBUFS: 13886 /* 13887 * Make this guy have to get ack's to send 13888 * more but lets make sure we don't 13889 * slam him below a T-O (1MSS). 13890 */ 13891 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 13892 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13893 bbr->r_ctl.rc_lost_bytes)) - maxseg; 13894 if (tp->snd_cwnd < maxseg) 13895 tp->snd_cwnd = maxseg; 13896 } 13897 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; 13898 BBR_STAT_INC(bbr_saw_enobuf); 13899 if (bbr->bbr_hdrw_pacing) 13900 counter_u64_add(bbr_hdwr_pacing_enobuf, 1); 13901 else 13902 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); 13903 /* 13904 * Here even in the enobuf's case we want to do our 13905 * state update. The reason being we may have been 13906 * called by the input function. If so we have had 13907 * things change. 13908 */ 13909 error = 0; 13910 goto enobufs; 13911 case EMSGSIZE: 13912 /* 13913 * For some reason the interface we used initially 13914 * to send segments changed to another or lowered 13915 * its MTU. If TSO was active we either got an 13916 * interface without TSO capabilits or TSO was 13917 * turned off. If we obtained mtu from ip_output() 13918 * then update it and try again. 13919 */ 13920 /* Turn on tracing (or try to) */ 13921 { 13922 int old_maxseg; 13923 13924 old_maxseg = tp->t_maxseg; 13925 BBR_STAT_INC(bbr_saw_emsgsiz); 13926 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); 13927 if (mtu != 0) 13928 tcp_mss_update(tp, -1, mtu, NULL, NULL); 13929 if (old_maxseg <= tp->t_maxseg) { 13930 /* Huh it did not shrink? */ 13931 tp->t_maxseg = old_maxseg - 40; 13932 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); 13933 } 13934 /* 13935 * Nuke all other things that can interfere 13936 * with slot 13937 */ 13938 if ((tot_len + len) && (len >= tp->t_maxseg)) { 13939 slot = bbr_get_pacing_delay(bbr, 13940 bbr->r_ctl.rc_bbr_hptsi_gain, 13941 (tot_len + len), cts, 0); 13942 if (slot < bbr_error_base_paceout) 13943 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13944 } else 13945 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13946 bbr->rc_output_starts_timer = 1; 13947 bbr_start_hpts_timer(bbr, tp, cts, 10, slot, 13948 tot_len); 13949 return (error); 13950 } 13951 case EPERM: 13952 tp->t_softerror = error; 13953 /* Fall through */ 13954 case EHOSTDOWN: 13955 case EHOSTUNREACH: 13956 case ENETDOWN: 13957 case ENETUNREACH: 13958 if (TCPS_HAVERCVDSYN(tp->t_state)) { 13959 tp->t_softerror = error; 13960 } 13961 /* FALLTHROUGH */ 13962 default: 13963 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; 13964 bbr->rc_output_starts_timer = 1; 13965 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); 13966 return (error); 13967 } 13968 #ifdef STATS 13969 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && 13970 len && 13971 (rsm == NULL) && 13972 (bbr->rc_in_persist == 0)) { 13973 tp->gput_seq = bbr_seq; 13974 tp->gput_ack = bbr_seq + 13975 min(sbavail(&so->so_snd) - sb_offset, sendwin); 13976 tp->gput_ts = cts; 13977 tp->t_flags |= TF_GPUTINPROG; 13978 #endif 13979 } 13980 KMOD_TCPSTAT_INC(tcps_sndtotal); 13981 if ((bbr->bbr_hdw_pace_ena) && 13982 (bbr->bbr_attempt_hdwr_pace == 0) && 13983 (bbr->rc_past_init_win) && 13984 (bbr->rc_bbr_state != BBR_STATE_STARTUP) && 13985 (get_filter_value(&bbr->r_ctl.rc_delrate)) && 13986 (inp->inp_route.ro_nh && 13987 inp->inp_route.ro_nh->nh_ifp)) { 13988 /* 13989 * We are past the initial window and 13990 * have at least one measurement so we 13991 * could use hardware pacing if its available. 13992 * We have an interface and we have not attempted 13993 * to setup hardware pacing, lets try to now. 13994 */ 13995 uint64_t rate_wanted; 13996 int err = 0; 13997 13998 rate_wanted = bbr_get_hardware_rate(bbr); 13999 bbr->bbr_attempt_hdwr_pace = 1; 14000 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, 14001 inp->inp_route.ro_nh->nh_ifp, 14002 rate_wanted, 14003 (RS_PACING_GEQ|RS_PACING_SUB_OK), 14004 &err, NULL); 14005 if (bbr->r_ctl.crte) { 14006 bbr_type_log_hdwr_pacing(bbr, 14007 bbr->r_ctl.crte->ptbl->rs_ifp, 14008 rate_wanted, 14009 bbr->r_ctl.crte->rate, 14010 __LINE__, cts, err); 14011 BBR_STAT_INC(bbr_hdwr_rl_add_ok); 14012 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 14013 counter_u64_add(bbr_flows_whdwr_pacing, 1); 14014 bbr->bbr_hdrw_pacing = 1; 14015 /* Now what is our gain status? */ 14016 if (bbr->r_ctl.crte->rate < rate_wanted) { 14017 /* We have a problem */ 14018 bbr_setup_less_of_rate(bbr, cts, 14019 bbr->r_ctl.crte->rate, rate_wanted); 14020 } else { 14021 /* We are good */ 14022 bbr->gain_is_limited = 0; 14023 bbr->skip_gain = 0; 14024 } 14025 tcp_bbr_tso_size_check(bbr, cts); 14026 } else { 14027 bbr_type_log_hdwr_pacing(bbr, 14028 inp->inp_route.ro_nh->nh_ifp, 14029 rate_wanted, 14030 0, 14031 __LINE__, cts, err); 14032 BBR_STAT_INC(bbr_hdwr_rl_add_fail); 14033 } 14034 } 14035 if (bbr->bbr_hdrw_pacing) { 14036 /* 14037 * Worry about cases where the route 14038 * changes or something happened that we 14039 * lost our hardware pacing possibly during 14040 * the last ip_output call. 14041 */ 14042 if (inp->inp_snd_tag == NULL) { 14043 /* A change during ip output disabled hw pacing? */ 14044 bbr->bbr_hdrw_pacing = 0; 14045 } else if ((inp->inp_route.ro_nh == NULL) || 14046 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { 14047 /* 14048 * We had an interface or route change, 14049 * detach from the current hdwr pacing 14050 * and setup to re-attempt next go 14051 * round. 14052 */ 14053 bbr->bbr_hdrw_pacing = 0; 14054 bbr->bbr_attempt_hdwr_pace = 0; 14055 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 14056 tcp_bbr_tso_size_check(bbr, cts); 14057 } 14058 } 14059 /* 14060 * Data sent (as far as we can tell). If this advertises a larger 14061 * window than any other segment, then remember the size of the 14062 * advertised window. Any pending ACK has now been sent. 14063 */ 14064 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 14065 tp->rcv_adv = tp->rcv_nxt + recwin; 14066 14067 tp->last_ack_sent = tp->rcv_nxt; 14068 if ((error == 0) && 14069 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && 14070 (doing_tlp == 0) && 14071 (tso == 0) && 14072 (len > 0) && 14073 ((flags & TH_RST) == 0) && 14074 ((flags & TH_SYN) == 0) && 14075 (IN_RECOVERY(tp->t_flags) == 0) && 14076 (bbr->rc_in_persist == 0) && 14077 (tot_len < bbr->r_ctl.rc_pace_max_segs)) { 14078 /* 14079 * For non-tso we need to goto again until we have sent out 14080 * enough data to match what we are hptsi out every hptsi 14081 * interval. 14082 */ 14083 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14084 /* Make sure snd_nxt is drug up */ 14085 tp->snd_nxt = tp->snd_max; 14086 } 14087 if (rsm != NULL) { 14088 rsm = NULL; 14089 goto skip_again; 14090 } 14091 rsm = NULL; 14092 sack_rxmit = 0; 14093 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 14094 goto again; 14095 } 14096 skip_again: 14097 if ((error == 0) && (flags & TH_FIN)) 14098 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); 14099 if ((error == 0) && (flags & TH_RST)) 14100 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 14101 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { 14102 /* 14103 * Calculate/Re-Calculate the hptsi slot in usecs based on 14104 * what we have sent so far 14105 */ 14106 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 14107 if (bbr->rc_no_pacing) 14108 slot = 0; 14109 } 14110 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 14111 enobufs: 14112 if (bbr->rc_use_google == 0) 14113 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 14114 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 14115 bbr->r_ctl.rc_lost_bytes))); 14116 bbr->rc_output_starts_timer = 1; 14117 if (bbr->bbr_use_rack_cheat && 14118 (more_to_rxt || 14119 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { 14120 /* Rack cheats and shotguns out all rxt's 1ms apart */ 14121 if (slot > 1000) 14122 slot = 1000; 14123 } 14124 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { 14125 /* 14126 * We don't change the tso size until some number of sends 14127 * to give the hardware commands time to get down 14128 * to the interface. 14129 */ 14130 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; 14131 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { 14132 bbr->hw_pacing_set = 1; 14133 tcp_bbr_tso_size_check(bbr, cts); 14134 } 14135 } 14136 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); 14137 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14138 /* Make sure snd_nxt is drug up */ 14139 tp->snd_nxt = tp->snd_max; 14140 } 14141 return (error); 14142 14143 } 14144 14145 /* 14146 * See bbr_output_wtime() for return values. 14147 */ 14148 static int 14149 bbr_output(struct tcpcb *tp) 14150 { 14151 int32_t ret; 14152 struct timeval tv; 14153 14154 NET_EPOCH_ASSERT(); 14155 14156 INP_WLOCK_ASSERT(tp->t_inpcb); 14157 (void)tcp_get_usecs(&tv); 14158 ret = bbr_output_wtime(tp, &tv); 14159 return (ret); 14160 } 14161 14162 static void 14163 bbr_mtu_chg(struct tcpcb *tp) 14164 { 14165 struct tcp_bbr *bbr; 14166 struct bbr_sendmap *rsm, *frsm = NULL; 14167 uint32_t maxseg; 14168 14169 /* 14170 * The MTU has changed. a) Clear the sack filter. b) Mark everything 14171 * over the current size as SACK_PASS so a retransmit will occur. 14172 */ 14173 14174 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14175 maxseg = tp->t_maxseg - bbr->rc_last_options; 14176 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 14177 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 14178 /* Don't mess with ones acked (by sack?) */ 14179 if (rsm->r_flags & BBR_ACKED) 14180 continue; 14181 if ((rsm->r_end - rsm->r_start) > maxseg) { 14182 /* 14183 * We mark sack-passed on all the previous large 14184 * sends we did. This will force them to retransmit. 14185 */ 14186 rsm->r_flags |= BBR_SACK_PASSED; 14187 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && 14188 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { 14189 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 14190 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 14191 rsm->r_flags |= BBR_MARKED_LOST; 14192 } 14193 if (frsm == NULL) 14194 frsm = rsm; 14195 } 14196 } 14197 if (frsm) { 14198 bbr->r_ctl.rc_resend = frsm; 14199 } 14200 } 14201 14202 static int 14203 bbr_pru_options(struct tcpcb *tp, int flags) 14204 { 14205 if (flags & PRUS_OOB) 14206 return (EOPNOTSUPP); 14207 return (0); 14208 } 14209 14210 struct tcp_function_block __tcp_bbr = { 14211 .tfb_tcp_block_name = __XSTRING(STACKNAME), 14212 .tfb_tcp_output = bbr_output, 14213 .tfb_do_queued_segments = ctf_do_queued_segments, 14214 .tfb_do_segment_nounlock = bbr_do_segment_nounlock, 14215 .tfb_tcp_do_segment = bbr_do_segment, 14216 .tfb_tcp_ctloutput = bbr_ctloutput, 14217 .tfb_tcp_fb_init = bbr_init, 14218 .tfb_tcp_fb_fini = bbr_fini, 14219 .tfb_tcp_timer_stop_all = bbr_stopall, 14220 .tfb_tcp_timer_activate = bbr_timer_activate, 14221 .tfb_tcp_timer_active = bbr_timer_active, 14222 .tfb_tcp_timer_stop = bbr_timer_stop, 14223 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, 14224 .tfb_tcp_handoff_ok = bbr_handoff_ok, 14225 .tfb_tcp_mtu_chg = bbr_mtu_chg, 14226 .tfb_pru_options = bbr_pru_options, 14227 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, 14228 }; 14229 14230 /* 14231 * bbr_ctloutput() must drop the inpcb lock before performing copyin on 14232 * socket option arguments. When it re-acquires the lock after the copy, it 14233 * has to revalidate that the connection is still valid for the socket 14234 * option. 14235 */ 14236 static int 14237 bbr_set_sockopt(struct inpcb *inp, struct sockopt *sopt) 14238 { 14239 struct epoch_tracker et; 14240 struct tcpcb *tp; 14241 struct tcp_bbr *bbr; 14242 int32_t error = 0, optval; 14243 14244 switch (sopt->sopt_level) { 14245 case IPPROTO_IPV6: 14246 case IPPROTO_IP: 14247 return (tcp_default_ctloutput(inp, sopt)); 14248 } 14249 14250 switch (sopt->sopt_name) { 14251 case TCP_RACK_PACE_MAX_SEG: 14252 case TCP_RACK_MIN_TO: 14253 case TCP_RACK_REORD_THRESH: 14254 case TCP_RACK_REORD_FADE: 14255 case TCP_RACK_TLP_THRESH: 14256 case TCP_RACK_PKT_DELAY: 14257 case TCP_BBR_ALGORITHM: 14258 case TCP_BBR_TSLIMITS: 14259 case TCP_BBR_IWINTSO: 14260 case TCP_BBR_RECFORCE: 14261 case TCP_BBR_STARTUP_PG: 14262 case TCP_BBR_DRAIN_PG: 14263 case TCP_BBR_RWND_IS_APP: 14264 case TCP_BBR_PROBE_RTT_INT: 14265 case TCP_BBR_PROBE_RTT_GAIN: 14266 case TCP_BBR_PROBE_RTT_LEN: 14267 case TCP_BBR_STARTUP_LOSS_EXIT: 14268 case TCP_BBR_USEDEL_RATE: 14269 case TCP_BBR_MIN_RTO: 14270 case TCP_BBR_MAX_RTO: 14271 case TCP_BBR_PACE_PER_SEC: 14272 case TCP_DELACK: 14273 case TCP_BBR_PACE_DEL_TAR: 14274 case TCP_BBR_SEND_IWND_IN_TSO: 14275 case TCP_BBR_EXTRA_STATE: 14276 case TCP_BBR_UTTER_MAX_TSO: 14277 case TCP_BBR_MIN_TOPACEOUT: 14278 case TCP_BBR_FLOOR_MIN_TSO: 14279 case TCP_BBR_TSTMP_RAISES: 14280 case TCP_BBR_POLICER_DETECT: 14281 case TCP_BBR_USE_RACK_CHEAT: 14282 case TCP_DATA_AFTER_CLOSE: 14283 case TCP_BBR_HDWR_PACE: 14284 case TCP_BBR_PACE_SEG_MAX: 14285 case TCP_BBR_PACE_SEG_MIN: 14286 case TCP_BBR_PACE_CROSS: 14287 case TCP_BBR_PACE_OH: 14288 #ifdef NETFLIX_PEAKRATE 14289 case TCP_MAXPEAKRATE: 14290 #endif 14291 case TCP_BBR_TMR_PACE_OH: 14292 case TCP_BBR_RACK_RTT_USE: 14293 case TCP_BBR_RETRAN_WTSO: 14294 break; 14295 default: 14296 return (tcp_default_ctloutput(inp, sopt)); 14297 break; 14298 } 14299 INP_WUNLOCK(inp); 14300 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 14301 if (error) 14302 return (error); 14303 INP_WLOCK(inp); 14304 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 14305 INP_WUNLOCK(inp); 14306 return (ECONNRESET); 14307 } 14308 tp = intotcpcb(inp); 14309 if (tp->t_fb != &__tcp_bbr) { 14310 INP_WUNLOCK(inp); 14311 return (ENOPROTOOPT); 14312 } 14313 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14314 switch (sopt->sopt_name) { 14315 case TCP_BBR_PACE_PER_SEC: 14316 BBR_OPTS_INC(tcp_bbr_pace_per_sec); 14317 bbr->r_ctl.bbr_hptsi_per_second = optval; 14318 break; 14319 case TCP_BBR_PACE_DEL_TAR: 14320 BBR_OPTS_INC(tcp_bbr_pace_del_tar); 14321 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; 14322 break; 14323 case TCP_BBR_PACE_SEG_MAX: 14324 BBR_OPTS_INC(tcp_bbr_pace_seg_max); 14325 bbr->r_ctl.bbr_hptsi_segments_max = optval; 14326 break; 14327 case TCP_BBR_PACE_SEG_MIN: 14328 BBR_OPTS_INC(tcp_bbr_pace_seg_min); 14329 bbr->r_ctl.bbr_hptsi_bytes_min = optval; 14330 break; 14331 case TCP_BBR_PACE_CROSS: 14332 BBR_OPTS_INC(tcp_bbr_pace_cross); 14333 bbr->r_ctl.bbr_cross_over = optval; 14334 break; 14335 case TCP_BBR_ALGORITHM: 14336 BBR_OPTS_INC(tcp_bbr_algorithm); 14337 if (optval && (bbr->rc_use_google == 0)) { 14338 /* Turn on the google mode */ 14339 bbr_google_mode_on(bbr); 14340 if ((optval > 3) && (optval < 500)) { 14341 /* 14342 * Must be at least greater than .3% 14343 * and must be less than 50.0%. 14344 */ 14345 bbr->r_ctl.bbr_google_discount = optval; 14346 } 14347 } else if ((optval == 0) && (bbr->rc_use_google == 1)) { 14348 /* Turn off the google mode */ 14349 bbr_google_mode_off(bbr); 14350 } 14351 break; 14352 case TCP_BBR_TSLIMITS: 14353 BBR_OPTS_INC(tcp_bbr_tslimits); 14354 if (optval == 1) 14355 bbr->rc_use_ts_limit = 1; 14356 else if (optval == 0) 14357 bbr->rc_use_ts_limit = 0; 14358 else 14359 error = EINVAL; 14360 break; 14361 14362 case TCP_BBR_IWINTSO: 14363 BBR_OPTS_INC(tcp_bbr_iwintso); 14364 if ((optval >= 0) && (optval < 128)) { 14365 uint32_t twin; 14366 14367 bbr->rc_init_win = optval; 14368 twin = bbr_initial_cwnd(bbr, tp); 14369 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) 14370 tp->snd_cwnd = twin; 14371 else 14372 error = EBUSY; 14373 } else 14374 error = EINVAL; 14375 break; 14376 case TCP_BBR_STARTUP_PG: 14377 BBR_OPTS_INC(tcp_bbr_startup_pg); 14378 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { 14379 bbr->r_ctl.rc_startup_pg = optval; 14380 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 14381 bbr->r_ctl.rc_bbr_hptsi_gain = optval; 14382 } 14383 } else 14384 error = EINVAL; 14385 break; 14386 case TCP_BBR_DRAIN_PG: 14387 BBR_OPTS_INC(tcp_bbr_drain_pg); 14388 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) 14389 bbr->r_ctl.rc_drain_pg = optval; 14390 else 14391 error = EINVAL; 14392 break; 14393 case TCP_BBR_PROBE_RTT_LEN: 14394 BBR_OPTS_INC(tcp_bbr_probertt_len); 14395 if (optval <= 1) 14396 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); 14397 else 14398 error = EINVAL; 14399 break; 14400 case TCP_BBR_PROBE_RTT_GAIN: 14401 BBR_OPTS_INC(tcp_bbr_probertt_gain); 14402 if (optval <= BBR_UNIT) 14403 bbr->r_ctl.bbr_rttprobe_gain_val = optval; 14404 else 14405 error = EINVAL; 14406 break; 14407 case TCP_BBR_PROBE_RTT_INT: 14408 BBR_OPTS_INC(tcp_bbr_probe_rtt_int); 14409 if (optval > 1000) 14410 bbr->r_ctl.rc_probertt_int = optval; 14411 else 14412 error = EINVAL; 14413 break; 14414 case TCP_BBR_MIN_TOPACEOUT: 14415 BBR_OPTS_INC(tcp_bbr_topaceout); 14416 if (optval == 0) { 14417 bbr->no_pacing_until = 0; 14418 bbr->rc_no_pacing = 0; 14419 } else if (optval <= 0x00ff) { 14420 bbr->no_pacing_until = optval; 14421 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && 14422 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ 14423 /* Turn on no pacing */ 14424 bbr->rc_no_pacing = 1; 14425 } 14426 } else 14427 error = EINVAL; 14428 break; 14429 case TCP_BBR_STARTUP_LOSS_EXIT: 14430 BBR_OPTS_INC(tcp_bbr_startup_loss_exit); 14431 bbr->rc_loss_exit = optval; 14432 break; 14433 case TCP_BBR_USEDEL_RATE: 14434 error = EINVAL; 14435 break; 14436 case TCP_BBR_MIN_RTO: 14437 BBR_OPTS_INC(tcp_bbr_min_rto); 14438 bbr->r_ctl.rc_min_rto_ms = optval; 14439 break; 14440 case TCP_BBR_MAX_RTO: 14441 BBR_OPTS_INC(tcp_bbr_max_rto); 14442 bbr->rc_max_rto_sec = optval; 14443 break; 14444 case TCP_RACK_MIN_TO: 14445 /* Minimum time between rack t-o's in ms */ 14446 BBR_OPTS_INC(tcp_rack_min_to); 14447 bbr->r_ctl.rc_min_to = optval; 14448 break; 14449 case TCP_RACK_REORD_THRESH: 14450 /* RACK reorder threshold (shift amount) */ 14451 BBR_OPTS_INC(tcp_rack_reord_thresh); 14452 if ((optval > 0) && (optval < 31)) 14453 bbr->r_ctl.rc_reorder_shift = optval; 14454 else 14455 error = EINVAL; 14456 break; 14457 case TCP_RACK_REORD_FADE: 14458 /* Does reordering fade after ms time */ 14459 BBR_OPTS_INC(tcp_rack_reord_fade); 14460 bbr->r_ctl.rc_reorder_fade = optval; 14461 break; 14462 case TCP_RACK_TLP_THRESH: 14463 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14464 BBR_OPTS_INC(tcp_rack_tlp_thresh); 14465 if (optval) 14466 bbr->rc_tlp_threshold = optval; 14467 else 14468 error = EINVAL; 14469 break; 14470 case TCP_BBR_USE_RACK_CHEAT: 14471 BBR_OPTS_INC(tcp_use_rackcheat); 14472 if (bbr->rc_use_google) { 14473 error = EINVAL; 14474 break; 14475 } 14476 BBR_OPTS_INC(tcp_rack_cheat); 14477 if (optval) 14478 bbr->bbr_use_rack_cheat = 1; 14479 else 14480 bbr->bbr_use_rack_cheat = 0; 14481 break; 14482 case TCP_BBR_FLOOR_MIN_TSO: 14483 BBR_OPTS_INC(tcp_utter_max_tso); 14484 if ((optval >= 0) && (optval < 40)) 14485 bbr->r_ctl.bbr_hptsi_segments_floor = optval; 14486 else 14487 error = EINVAL; 14488 break; 14489 case TCP_BBR_UTTER_MAX_TSO: 14490 BBR_OPTS_INC(tcp_utter_max_tso); 14491 if ((optval >= 0) && (optval < 0xffff)) 14492 bbr->r_ctl.bbr_utter_max = optval; 14493 else 14494 error = EINVAL; 14495 break; 14496 14497 case TCP_BBR_EXTRA_STATE: 14498 BBR_OPTS_INC(tcp_extra_state); 14499 if (optval) 14500 bbr->rc_use_idle_restart = 1; 14501 else 14502 bbr->rc_use_idle_restart = 0; 14503 break; 14504 case TCP_BBR_SEND_IWND_IN_TSO: 14505 BBR_OPTS_INC(tcp_iwnd_tso); 14506 if (optval) { 14507 bbr->bbr_init_win_cheat = 1; 14508 if (bbr->rc_past_init_win == 0) { 14509 uint32_t cts; 14510 cts = tcp_get_usecs(&bbr->rc_tv); 14511 tcp_bbr_tso_size_check(bbr, cts); 14512 } 14513 } else 14514 bbr->bbr_init_win_cheat = 0; 14515 break; 14516 case TCP_BBR_HDWR_PACE: 14517 BBR_OPTS_INC(tcp_hdwr_pacing); 14518 if (optval){ 14519 bbr->bbr_hdw_pace_ena = 1; 14520 bbr->bbr_attempt_hdwr_pace = 0; 14521 } else { 14522 bbr->bbr_hdw_pace_ena = 0; 14523 #ifdef RATELIMIT 14524 if (bbr->r_ctl.crte != NULL) { 14525 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 14526 bbr->r_ctl.crte = NULL; 14527 } 14528 #endif 14529 } 14530 break; 14531 14532 case TCP_DELACK: 14533 BBR_OPTS_INC(tcp_delack); 14534 if (optval < 100) { 14535 if (optval == 0) /* off */ 14536 tp->t_delayed_ack = 0; 14537 else if (optval == 1) /* on which is 2 */ 14538 tp->t_delayed_ack = 2; 14539 else /* higher than 2 and less than 100 */ 14540 tp->t_delayed_ack = optval; 14541 if (tp->t_flags & TF_DELACK) { 14542 tp->t_flags &= ~TF_DELACK; 14543 tp->t_flags |= TF_ACKNOW; 14544 NET_EPOCH_ENTER(et); 14545 bbr_output(tp); 14546 NET_EPOCH_EXIT(et); 14547 } 14548 } else 14549 error = EINVAL; 14550 break; 14551 case TCP_RACK_PKT_DELAY: 14552 /* RACK added ms i.e. rack-rtt + reord + N */ 14553 BBR_OPTS_INC(tcp_rack_pkt_delay); 14554 bbr->r_ctl.rc_pkt_delay = optval; 14555 break; 14556 #ifdef NETFLIX_PEAKRATE 14557 case TCP_MAXPEAKRATE: 14558 BBR_OPTS_INC(tcp_maxpeak); 14559 error = tcp_set_maxpeakrate(tp, optval); 14560 if (!error) 14561 tp->t_peakrate_thr = tp->t_maxpeakrate; 14562 break; 14563 #endif 14564 case TCP_BBR_RETRAN_WTSO: 14565 BBR_OPTS_INC(tcp_retran_wtso); 14566 if (optval) 14567 bbr->rc_resends_use_tso = 1; 14568 else 14569 bbr->rc_resends_use_tso = 0; 14570 break; 14571 case TCP_DATA_AFTER_CLOSE: 14572 BBR_OPTS_INC(tcp_data_ac); 14573 if (optval) 14574 bbr->rc_allow_data_af_clo = 1; 14575 else 14576 bbr->rc_allow_data_af_clo = 0; 14577 break; 14578 case TCP_BBR_POLICER_DETECT: 14579 BBR_OPTS_INC(tcp_policer_det); 14580 if (bbr->rc_use_google == 0) 14581 error = EINVAL; 14582 else if (optval) 14583 bbr->r_use_policer = 1; 14584 else 14585 bbr->r_use_policer = 0; 14586 break; 14587 14588 case TCP_BBR_TSTMP_RAISES: 14589 BBR_OPTS_INC(tcp_ts_raises); 14590 if (optval) 14591 bbr->ts_can_raise = 1; 14592 else 14593 bbr->ts_can_raise = 0; 14594 break; 14595 case TCP_BBR_TMR_PACE_OH: 14596 BBR_OPTS_INC(tcp_pacing_oh_tmr); 14597 if (bbr->rc_use_google) { 14598 error = EINVAL; 14599 } else { 14600 if (optval) 14601 bbr->r_ctl.rc_incr_tmrs = 1; 14602 else 14603 bbr->r_ctl.rc_incr_tmrs = 0; 14604 } 14605 break; 14606 case TCP_BBR_PACE_OH: 14607 BBR_OPTS_INC(tcp_pacing_oh); 14608 if (bbr->rc_use_google) { 14609 error = EINVAL; 14610 } else { 14611 if (optval > (BBR_INCL_TCP_OH| 14612 BBR_INCL_IP_OH| 14613 BBR_INCL_ENET_OH)) { 14614 error = EINVAL; 14615 break; 14616 } 14617 if (optval & BBR_INCL_TCP_OH) 14618 bbr->r_ctl.rc_inc_tcp_oh = 1; 14619 else 14620 bbr->r_ctl.rc_inc_tcp_oh = 0; 14621 if (optval & BBR_INCL_IP_OH) 14622 bbr->r_ctl.rc_inc_ip_oh = 1; 14623 else 14624 bbr->r_ctl.rc_inc_ip_oh = 0; 14625 if (optval & BBR_INCL_ENET_OH) 14626 bbr->r_ctl.rc_inc_enet_oh = 1; 14627 else 14628 bbr->r_ctl.rc_inc_enet_oh = 0; 14629 } 14630 break; 14631 default: 14632 return (tcp_default_ctloutput(inp, sopt)); 14633 break; 14634 } 14635 #ifdef NETFLIX_STATS 14636 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 14637 #endif 14638 INP_WUNLOCK(inp); 14639 return (error); 14640 } 14641 14642 /* 14643 * return 0 on success, error-num on failure 14644 */ 14645 static int 14646 bbr_get_sockopt(struct inpcb *inp, struct sockopt *sopt) 14647 { 14648 struct tcpcb *tp; 14649 struct tcp_bbr *bbr; 14650 int32_t error, optval; 14651 14652 tp = intotcpcb(inp); 14653 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14654 if (bbr == NULL) { 14655 INP_WUNLOCK(inp); 14656 return (EINVAL); 14657 } 14658 /* 14659 * Because all our options are either boolean or an int, we can just 14660 * pull everything into optval and then unlock and copy. If we ever 14661 * add a option that is not a int, then this will have quite an 14662 * impact to this routine. 14663 */ 14664 switch (sopt->sopt_name) { 14665 case TCP_BBR_PACE_PER_SEC: 14666 optval = bbr->r_ctl.bbr_hptsi_per_second; 14667 break; 14668 case TCP_BBR_PACE_DEL_TAR: 14669 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; 14670 break; 14671 case TCP_BBR_PACE_SEG_MAX: 14672 optval = bbr->r_ctl.bbr_hptsi_segments_max; 14673 break; 14674 case TCP_BBR_MIN_TOPACEOUT: 14675 optval = bbr->no_pacing_until; 14676 break; 14677 case TCP_BBR_PACE_SEG_MIN: 14678 optval = bbr->r_ctl.bbr_hptsi_bytes_min; 14679 break; 14680 case TCP_BBR_PACE_CROSS: 14681 optval = bbr->r_ctl.bbr_cross_over; 14682 break; 14683 case TCP_BBR_ALGORITHM: 14684 optval = bbr->rc_use_google; 14685 break; 14686 case TCP_BBR_TSLIMITS: 14687 optval = bbr->rc_use_ts_limit; 14688 break; 14689 case TCP_BBR_IWINTSO: 14690 optval = bbr->rc_init_win; 14691 break; 14692 case TCP_BBR_STARTUP_PG: 14693 optval = bbr->r_ctl.rc_startup_pg; 14694 break; 14695 case TCP_BBR_DRAIN_PG: 14696 optval = bbr->r_ctl.rc_drain_pg; 14697 break; 14698 case TCP_BBR_PROBE_RTT_INT: 14699 optval = bbr->r_ctl.rc_probertt_int; 14700 break; 14701 case TCP_BBR_PROBE_RTT_LEN: 14702 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); 14703 break; 14704 case TCP_BBR_PROBE_RTT_GAIN: 14705 optval = bbr->r_ctl.bbr_rttprobe_gain_val; 14706 break; 14707 case TCP_BBR_STARTUP_LOSS_EXIT: 14708 optval = bbr->rc_loss_exit; 14709 break; 14710 case TCP_BBR_USEDEL_RATE: 14711 error = EINVAL; 14712 break; 14713 case TCP_BBR_MIN_RTO: 14714 optval = bbr->r_ctl.rc_min_rto_ms; 14715 break; 14716 case TCP_BBR_MAX_RTO: 14717 optval = bbr->rc_max_rto_sec; 14718 break; 14719 case TCP_RACK_PACE_MAX_SEG: 14720 /* Max segments in a pace */ 14721 optval = bbr->r_ctl.rc_pace_max_segs; 14722 break; 14723 case TCP_RACK_MIN_TO: 14724 /* Minimum time between rack t-o's in ms */ 14725 optval = bbr->r_ctl.rc_min_to; 14726 break; 14727 case TCP_RACK_REORD_THRESH: 14728 /* RACK reorder threshold (shift amount) */ 14729 optval = bbr->r_ctl.rc_reorder_shift; 14730 break; 14731 case TCP_RACK_REORD_FADE: 14732 /* Does reordering fade after ms time */ 14733 optval = bbr->r_ctl.rc_reorder_fade; 14734 break; 14735 case TCP_BBR_USE_RACK_CHEAT: 14736 /* Do we use the rack cheat for rxt */ 14737 optval = bbr->bbr_use_rack_cheat; 14738 break; 14739 case TCP_BBR_FLOOR_MIN_TSO: 14740 optval = bbr->r_ctl.bbr_hptsi_segments_floor; 14741 break; 14742 case TCP_BBR_UTTER_MAX_TSO: 14743 optval = bbr->r_ctl.bbr_utter_max; 14744 break; 14745 case TCP_BBR_SEND_IWND_IN_TSO: 14746 /* Do we send TSO size segments initially */ 14747 optval = bbr->bbr_init_win_cheat; 14748 break; 14749 case TCP_BBR_EXTRA_STATE: 14750 optval = bbr->rc_use_idle_restart; 14751 break; 14752 case TCP_RACK_TLP_THRESH: 14753 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14754 optval = bbr->rc_tlp_threshold; 14755 break; 14756 case TCP_RACK_PKT_DELAY: 14757 /* RACK added ms i.e. rack-rtt + reord + N */ 14758 optval = bbr->r_ctl.rc_pkt_delay; 14759 break; 14760 case TCP_BBR_RETRAN_WTSO: 14761 optval = bbr->rc_resends_use_tso; 14762 break; 14763 case TCP_DATA_AFTER_CLOSE: 14764 optval = bbr->rc_allow_data_af_clo; 14765 break; 14766 case TCP_DELACK: 14767 optval = tp->t_delayed_ack; 14768 break; 14769 case TCP_BBR_HDWR_PACE: 14770 optval = bbr->bbr_hdw_pace_ena; 14771 break; 14772 case TCP_BBR_POLICER_DETECT: 14773 optval = bbr->r_use_policer; 14774 break; 14775 case TCP_BBR_TSTMP_RAISES: 14776 optval = bbr->ts_can_raise; 14777 break; 14778 case TCP_BBR_TMR_PACE_OH: 14779 optval = bbr->r_ctl.rc_incr_tmrs; 14780 break; 14781 case TCP_BBR_PACE_OH: 14782 optval = 0; 14783 if (bbr->r_ctl.rc_inc_tcp_oh) 14784 optval |= BBR_INCL_TCP_OH; 14785 if (bbr->r_ctl.rc_inc_ip_oh) 14786 optval |= BBR_INCL_IP_OH; 14787 if (bbr->r_ctl.rc_inc_enet_oh) 14788 optval |= BBR_INCL_ENET_OH; 14789 break; 14790 default: 14791 return (tcp_default_ctloutput(inp, sopt)); 14792 break; 14793 } 14794 INP_WUNLOCK(inp); 14795 error = sooptcopyout(sopt, &optval, sizeof optval); 14796 return (error); 14797 } 14798 14799 /* 14800 * return 0 on success, error-num on failure 14801 */ 14802 static int 14803 bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt) 14804 { 14805 if (sopt->sopt_dir == SOPT_SET) { 14806 return (bbr_set_sockopt(inp, sopt)); 14807 } else if (sopt->sopt_dir == SOPT_GET) { 14808 return (bbr_get_sockopt(inp, sopt)); 14809 } else { 14810 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); 14811 } 14812 } 14813 14814 static const char *bbr_stack_names[] = { 14815 __XSTRING(STACKNAME), 14816 #ifdef STACKALIAS 14817 __XSTRING(STACKALIAS), 14818 #endif 14819 }; 14820 14821 static bool bbr_mod_inited = false; 14822 14823 static int 14824 tcp_addbbr(module_t mod, int32_t type, void *data) 14825 { 14826 int32_t err = 0; 14827 int num_stacks; 14828 14829 switch (type) { 14830 case MOD_LOAD: 14831 printf("Attempting to load " __XSTRING(MODNAME) "\n"); 14832 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 14833 sizeof(struct bbr_sendmap), 14834 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 14835 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 14836 sizeof(struct tcp_bbr), 14837 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 14838 sysctl_ctx_init(&bbr_sysctl_ctx); 14839 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 14840 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 14841 OID_AUTO, 14842 #ifdef STACKALIAS 14843 __XSTRING(STACKALIAS), 14844 #else 14845 __XSTRING(STACKNAME), 14846 #endif 14847 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 14848 ""); 14849 if (bbr_sysctl_root == NULL) { 14850 printf("Failed to add sysctl node\n"); 14851 err = EFAULT; 14852 goto free_uma; 14853 } 14854 bbr_init_sysctls(); 14855 num_stacks = nitems(bbr_stack_names); 14856 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, 14857 bbr_stack_names, &num_stacks); 14858 if (err) { 14859 printf("Failed to register %s stack name for " 14860 "%s module\n", bbr_stack_names[num_stacks], 14861 __XSTRING(MODNAME)); 14862 sysctl_ctx_free(&bbr_sysctl_ctx); 14863 free_uma: 14864 uma_zdestroy(bbr_zone); 14865 uma_zdestroy(bbr_pcb_zone); 14866 bbr_counter_destroy(); 14867 printf("Failed to register " __XSTRING(MODNAME) 14868 " module err:%d\n", err); 14869 return (err); 14870 } 14871 tcp_lro_reg_mbufq(); 14872 bbr_mod_inited = true; 14873 printf(__XSTRING(MODNAME) " is now available\n"); 14874 break; 14875 case MOD_QUIESCE: 14876 err = deregister_tcp_functions(&__tcp_bbr, true, false); 14877 break; 14878 case MOD_UNLOAD: 14879 err = deregister_tcp_functions(&__tcp_bbr, false, true); 14880 if (err == EBUSY) 14881 break; 14882 if (bbr_mod_inited) { 14883 uma_zdestroy(bbr_zone); 14884 uma_zdestroy(bbr_pcb_zone); 14885 sysctl_ctx_free(&bbr_sysctl_ctx); 14886 bbr_counter_destroy(); 14887 printf(__XSTRING(MODNAME) 14888 " is now no longer available\n"); 14889 bbr_mod_inited = false; 14890 } 14891 tcp_lro_dereg_mbufq(); 14892 err = 0; 14893 break; 14894 default: 14895 return (EOPNOTSUPP); 14896 } 14897 return (err); 14898 } 14899 14900 static moduledata_t tcp_bbr = { 14901 .name = __XSTRING(MODNAME), 14902 .evhand = tcp_addbbr, 14903 .priv = 0 14904 }; 14905 14906 MODULE_VERSION(MODNAME, 1); 14907 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 14908 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 14909