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 #include "opt_inet.h" 35 #include "opt_inet6.h" 36 #include "opt_ipsec.h" 37 #include "opt_ratelimit.h" 38 #include <sys/param.h> 39 #include <sys/arb.h> 40 #include <sys/module.h> 41 #include <sys/kernel.h> 42 #include <sys/libkern.h> 43 #ifdef TCP_HHOOK 44 #include <sys/hhook.h> 45 #endif 46 #include <sys/malloc.h> 47 #include <sys/mbuf.h> 48 #include <sys/proc.h> 49 #include <sys/socket.h> 50 #include <sys/socketvar.h> 51 #include <sys/sysctl.h> 52 #include <sys/systm.h> 53 #ifdef STATS 54 #include <sys/qmath.h> 55 #include <sys/tree.h> 56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */ 57 #endif 58 #include <sys/refcount.h> 59 #include <sys/queue.h> 60 #include <sys/eventhandler.h> 61 #include <sys/smp.h> 62 #include <sys/kthread.h> 63 #include <sys/lock.h> 64 #include <sys/mutex.h> 65 #include <sys/tim_filter.h> 66 #include <sys/time.h> 67 #include <sys/protosw.h> 68 #include <vm/uma.h> 69 #include <sys/kern_prefetch.h> 70 71 #include <net/route.h> 72 #include <net/route/nhop.h> 73 #include <net/vnet.h> 74 75 #define TCPSTATES /* for logging */ 76 77 #include <netinet/in.h> 78 #include <netinet/in_kdtrace.h> 79 #include <netinet/in_pcb.h> 80 #include <netinet/ip.h> 81 #include <netinet/ip_icmp.h> /* required for icmp_var.h */ 82 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */ 83 #include <netinet/ip_var.h> 84 #include <netinet/ip6.h> 85 #include <netinet6/in6_pcb.h> 86 #include <netinet6/ip6_var.h> 87 #define TCPOUTFLAGS 88 #include <netinet/tcp.h> 89 #include <netinet/tcp_fsm.h> 90 #include <netinet/tcp_seq.h> 91 #include <netinet/tcp_timer.h> 92 #include <netinet/tcp_var.h> 93 #include <netinet/tcpip.h> 94 #include <netinet/tcp_hpts.h> 95 #include <netinet/cc/cc.h> 96 #include <netinet/tcp_log_buf.h> 97 #include <netinet/tcp_ratelimit.h> 98 #include <netinet/tcp_lro.h> 99 #ifdef TCP_OFFLOAD 100 #include <netinet/tcp_offload.h> 101 #endif 102 #ifdef INET6 103 #include <netinet6/tcp6_var.h> 104 #endif 105 #include <netinet/tcp_fastopen.h> 106 107 #include <netipsec/ipsec_support.h> 108 #include <net/if.h> 109 #include <net/if_var.h> 110 #include <net/ethernet.h> 111 112 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 113 #include <netipsec/ipsec.h> 114 #include <netipsec/ipsec6.h> 115 #endif /* IPSEC */ 116 117 #include <netinet/udp.h> 118 #include <netinet/udp_var.h> 119 #include <machine/in_cksum.h> 120 121 #ifdef MAC 122 #include <security/mac/mac_framework.h> 123 #endif 124 125 #include "sack_filter.h" 126 #include "tcp_bbr.h" 127 #include "rack_bbr_common.h" 128 uma_zone_t bbr_zone; 129 uma_zone_t bbr_pcb_zone; 130 131 struct sysctl_ctx_list bbr_sysctl_ctx; 132 struct sysctl_oid *bbr_sysctl_root; 133 134 #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ 135 (tv) = (value); \ 136 if ((u_long)(tv) < (u_long)(tvmin)) \ 137 (tv) = (tvmin); \ 138 if ((u_long)(tv) > (u_long)(tvmax)) \ 139 (tv) = (tvmax); \ 140 } while(0) 141 142 /*#define BBR_INVARIANT 1*/ 143 144 /* 145 * initial window 146 */ 147 static uint32_t bbr_def_init_win = 10; 148 static int32_t bbr_persist_min = 250000; /* 250ms */ 149 static int32_t bbr_persist_max = 1000000; /* 1 Second */ 150 static int32_t bbr_cwnd_may_shrink = 0; 151 static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; 152 static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; 153 static int32_t bbr_hardware_pacing_limit = 8000; 154 static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ 155 static int32_t bbr_no_retran = 0; 156 157 static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ 158 static int32_t bbr_max_net_error_cnt = 10; 159 /* Should the following be dynamic too -- loss wise */ 160 static int32_t bbr_rtt_gain_thresh = 0; 161 /* Measurement controls */ 162 static int32_t bbr_use_google_algo = 1; 163 static int32_t bbr_ts_limiting = 1; 164 static int32_t bbr_ts_can_raise = 0; 165 static int32_t bbr_do_red = 600; 166 static int32_t bbr_red_scale = 20000; 167 static int32_t bbr_red_mul = 1; 168 static int32_t bbr_red_div = 2; 169 static int32_t bbr_red_growth_restrict = 1; 170 static int32_t bbr_target_is_bbunit = 0; 171 static int32_t bbr_drop_limit = 0; 172 /* 173 * How much gain do we need to see to 174 * stay in startup? 175 */ 176 static int32_t bbr_marks_rxt_sack_passed = 0; 177 static int32_t bbr_start_exit = 25; 178 static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ 179 static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ 180 static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ 181 static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this 182 * if we go back ever to where the pacer 183 * has priority over timers. 184 */ 185 static int32_t bbr_policer_call_from_rack_to = 0; 186 static int32_t bbr_policer_detection_enabled = 1; 187 static int32_t bbr_min_measurements_req = 1; /* We need at least 2 188 * measurements before we are 189 * "good" note that 2 == 1. 190 * This is because we use a > 191 * comparison. This means if 192 * min_measure was 0, it takes 193 * num-measures > min(0) and 194 * you get 1 measurement and 195 * you are good. Set to 1, you 196 * have to have two 197 * measurements (this is done 198 * to prevent it from being ok 199 * to have no measurements). */ 200 static int32_t bbr_no_pacing_until = 4; 201 202 static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ 203 static int32_t bbr_min_peer_delta = 20; /* 20 units */ 204 static int32_t bbr_delta_percent = 150; /* 15.0 % */ 205 206 static int32_t bbr_target_cwnd_mult_limit = 8; 207 /* 208 * bbr_cwnd_min_val is the number of 209 * segments we hold to in the RTT probe 210 * state typically 4. 211 */ 212 static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; 213 214 static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; 215 216 static int32_t bbr_gain_to_target = 1; 217 static int32_t bbr_gain_gets_extra_too = 1; 218 /* 219 * bbr_high_gain is the 2/ln(2) value we need 220 * to double the sending rate in startup. This 221 * is used for both cwnd and hptsi gain's. 222 */ 223 static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; 224 static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; 225 static int32_t bbr_use_lower_gain_in_startup = 1; 226 227 /* thresholds for reduction on drain in sub-states/drain */ 228 static int32_t bbr_drain_rtt = BBR_SRTT; 229 static int32_t bbr_drain_floor = 88; 230 static int32_t google_allow_early_out = 1; 231 static int32_t google_consider_lost = 1; 232 static int32_t bbr_drain_drop_mul = 4; 233 static int32_t bbr_drain_drop_div = 5; 234 static int32_t bbr_rand_ot = 50; 235 static int32_t bbr_can_force_probertt = 0; 236 static int32_t bbr_can_adjust_probertt = 1; 237 static int32_t bbr_probertt_sets_rtt = 0; 238 static int32_t bbr_can_use_ts_for_rtt = 1; 239 static int32_t bbr_is_ratio = 0; 240 static int32_t bbr_sub_drain_app_limit = 1; 241 static int32_t bbr_prtt_slam_cwnd = 1; 242 static int32_t bbr_sub_drain_slam_cwnd = 1; 243 static int32_t bbr_slam_cwnd_in_main_drain = 1; 244 static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter 245 * hold */ 246 static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); 247 /* 248 * bbr_drain_gain is the reverse of the high_gain 249 * designed to drain back out the standing queue 250 * that is formed in startup by causing a larger 251 * hptsi gain and thus drainging the packets 252 * in flight. 253 */ 254 static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; 255 static int32_t bbr_rttprobe_gain = 192; 256 257 /* 258 * The cwnd_gain is the default cwnd gain applied when 259 * calculating a target cwnd. Note that the cwnd is 260 * a secondary factor in the way BBR works (see the 261 * paper and think about it, it will take some time). 262 * Basically the hptsi_gain spreads the packets out 263 * so you never get more than BDP to the peer even 264 * if the cwnd is high. In our implemenation that 265 * means in non-recovery/retransmission scenarios 266 * cwnd will never be reached by the flight-size. 267 */ 268 static int32_t bbr_cwnd_gain = BBR_UNIT * 2; 269 static int32_t bbr_tlp_type_to_use = BBR_SRTT; 270 static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ 271 static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ 272 static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ 273 static int32_t bbr_ignore_data_after_close = 1; 274 static int16_t bbr_hptsi_gain[] = { 275 (BBR_UNIT *5 / 4), 276 (BBR_UNIT * 3 / 4), 277 BBR_UNIT, 278 BBR_UNIT, 279 BBR_UNIT, 280 BBR_UNIT, 281 BBR_UNIT, 282 BBR_UNIT 283 }; 284 int32_t bbr_use_rack_resend_cheat = 1; 285 int32_t bbr_sends_full_iwnd = 1; 286 287 #define BBR_HPTSI_GAIN_MAX 8 288 /* 289 * The BBR module incorporates a number of 290 * TCP ideas that have been put out into the IETF 291 * over the last few years: 292 * - Yuchung Cheng's RACK TCP (for which its named) that 293 * will stop us using the number of dup acks and instead 294 * use time as the gage of when we retransmit. 295 * - Reorder Detection of RFC4737 and the Tail-Loss probe draft 296 * of Dukkipati et.al. 297 * - Van Jacobson's et.al BBR. 298 * 299 * RACK depends on SACK, so if an endpoint arrives that 300 * cannot do SACK the state machine below will shuttle the 301 * connection back to using the "default" TCP stack that is 302 * in FreeBSD. 303 * 304 * To implement BBR and RACK the original TCP stack was first decomposed 305 * into a functional state machine with individual states 306 * for each of the possible TCP connection states. The do_segment 307 * functions role in life is to mandate the connection supports SACK 308 * initially and then assure that the RACK state matches the conenction 309 * state before calling the states do_segment function. Data processing 310 * of inbound segments also now happens in the hpts_do_segment in general 311 * with only one exception. This is so we can keep the connection on 312 * a single CPU. 313 * 314 * Each state is simplified due to the fact that the original do_segment 315 * has been decomposed and we *know* what state we are in (no 316 * switches on the state) and all tests for SACK are gone. This 317 * greatly simplifies what each state does. 318 * 319 * TCP output is also over-written with a new version since it 320 * must maintain the new rack scoreboard and has had hptsi 321 * integrated as a requirment. Still todo is to eliminate the 322 * use of the callout_() system and use the hpts for all 323 * timers as well. 324 */ 325 static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ 326 static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ 327 static const int32_t bbr_min_req_free = 2; /* The min we must have on the 328 * free list */ 329 static int32_t bbr_tlp_thresh = 1; 330 static int32_t bbr_reorder_thresh = 2; 331 static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def 332 * 60,000,000 - 60 seconds */ 333 static int32_t bbr_pkt_delay = 1000; 334 static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ 335 static int32_t bbr_incr_timers = 1; 336 337 static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ 338 static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ 339 static int32_t bbr_exit_startup_at_loss = 1; 340 341 /* 342 * bbr_lt_bw_ratio is 1/8th 343 * bbr_lt_bw_diff is < 4 Kbit/sec 344 */ 345 static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ 346 static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ 347 static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use 348 * the lt_bw for */ 349 static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure 350 * lt_bw */ 351 static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ 352 static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ 353 static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ 354 355 static int32_t bbr_verbose_logging = 0; 356 /* 357 * Currently regular tcp has a rto_min of 30ms 358 * the backoff goes 12 times so that ends up 359 * being a total of 122.850 seconds before a 360 * connection is killed. 361 */ 362 static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ 363 static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ 364 365 /****************************************************/ 366 /* DEFAULT TSO SIZING (cpu performance impacting) */ 367 /****************************************************/ 368 /* What amount is our formula using to get TSO size */ 369 static int32_t bbr_hptsi_per_second = 1000; 370 371 /* 372 * For hptsi under bbr_cross_over connections what is delay 373 * target 7ms (in usec) combined with a seg_max of 2 374 * gets us close to identical google behavior in 375 * TSO size selection (possibly more 1MSS sends). 376 */ 377 static int32_t bbr_hptsi_segments_delay_tar = 7000; 378 379 /* Does pacing delay include overhead's in its time calculations? */ 380 static int32_t bbr_include_enet_oh = 0; 381 static int32_t bbr_include_ip_oh = 1; 382 static int32_t bbr_include_tcp_oh = 1; 383 static int32_t bbr_google_discount = 10; 384 385 /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ 386 static int32_t bbr_state_is_pkt_epoch = 0; 387 static int32_t bbr_state_drain_2_tar = 1; 388 /* What is the max the 0 - bbr_cross_over MBPS TSO target 389 * can reach using our delay target. Note that this 390 * value becomes the floor for the cross over 391 * algorithm. 392 */ 393 static int32_t bbr_hptsi_segments_max = 2; 394 static int32_t bbr_hptsi_segments_floor = 1; 395 static int32_t bbr_hptsi_utter_max = 0; 396 397 /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ 398 static int32_t bbr_hptsi_bytes_min = 1460; 399 static int32_t bbr_all_get_min = 0; 400 401 /* Cross over point from algo-a to algo-b */ 402 static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; 403 404 /* Do we deal with our restart state? */ 405 static int32_t bbr_uses_idle_restart = 0; 406 static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ 407 408 /* Do we allow hardware pacing? */ 409 static int32_t bbr_allow_hdwr_pacing = 0; 410 static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ 411 static int32_t bbr_hdwr_pace_floor = 1; 412 static int32_t bbr_hdwr_pacing_delay_cnt = 10; 413 414 /****************************************************/ 415 static int32_t bbr_resends_use_tso = 0; 416 static int32_t bbr_tlp_max_resend = 2; 417 static int32_t bbr_sack_block_limit = 128; 418 419 #define BBR_MAX_STAT 19 420 counter_u64_t bbr_state_time[BBR_MAX_STAT]; 421 counter_u64_t bbr_state_lost[BBR_MAX_STAT]; 422 counter_u64_t bbr_state_resend[BBR_MAX_STAT]; 423 counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; 424 counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; 425 counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; 426 counter_u64_t bbr_flows_whdwr_pacing; 427 counter_u64_t bbr_flows_nohdwr_pacing; 428 429 counter_u64_t bbr_nohdwr_pacing_enobuf; 430 counter_u64_t bbr_hdwr_pacing_enobuf; 431 432 static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); 433 434 /* 435 * Static defintions we need for forward declarations. 436 */ 437 static uint32_t 438 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, 439 uint32_t useconds_time, uint64_t bw); 440 static uint32_t 441 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); 442 static void 443 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); 444 static void 445 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); 446 static void 447 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, 448 int dolog); 449 static uint32_t 450 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); 451 static void 452 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, 453 int32_t pkt_epoch, uint32_t losses); 454 static uint32_t 455 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, 456 struct bbr_sendmap *rsm); 457 static uint32_t 458 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); 459 static uint32_t 460 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 461 struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); 462 static void 463 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 464 int32_t line); 465 static void 466 bbr_set_state_target(struct tcp_bbr *bbr, int line); 467 static void 468 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); 469 static void 470 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, 471 int event, int line); 472 static void 473 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); 474 static void 475 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); 476 static void 477 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 478 uint32_t rtt, uint32_t line, uint8_t is_start, 479 uint16_t set); 480 static struct bbr_sendmap * 481 bbr_find_lowest_rsm(struct tcp_bbr *bbr); 482 static __inline uint32_t 483 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); 484 static void 485 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, 486 uint8_t which); 487 static void 488 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, 489 uint32_t time_since_sent, uint32_t srtt, 490 uint32_t thresh, uint32_t to); 491 static void 492 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); 493 static void 494 bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, 495 uint32_t del_by, uint32_t cts, uint32_t sloton, 496 uint32_t prev_delay); 497 static void 498 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, 499 int32_t line); 500 static void 501 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr); 502 static void 503 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); 504 static void 505 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); 506 static void 507 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); 508 static void 509 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 510 uint32_t cts, uint32_t usecs, uint64_t bw, 511 uint32_t override, int mod); 512 static int bbr_ctloutput(struct tcpcb *tp, struct sockopt *sopt); 513 514 static inline uint8_t 515 bbr_state_val(struct tcp_bbr *bbr) 516 { 517 return(bbr->rc_bbr_substate); 518 } 519 520 static inline uint32_t 521 get_min_cwnd(struct tcp_bbr *bbr) 522 { 523 int mss; 524 525 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 526 bbr->r_ctl.rc_pace_max_segs); 527 if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) 528 return (bbr_cwnd_min_val_hs * mss); 529 else 530 return (bbr_cwnd_min_val * mss); 531 } 532 533 static uint32_t 534 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) 535 { 536 uint64_t srtt, var; 537 uint64_t ret_val; 538 539 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; 540 if (tp->t_srtt == 0) { 541 srtt = (uint64_t)BBR_INITIAL_RTO; 542 var = 0; 543 } else { 544 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 545 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); 546 } 547 TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), 548 bbr_persist_min, bbr_persist_max); 549 return ((uint32_t)ret_val); 550 } 551 552 static uint32_t 553 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 554 { 555 /* 556 * Start the FR timer, we do this based on getting the first one in 557 * the rc_tmap. Note that if its NULL we must stop the timer. in all 558 * events we need to stop the running timer (if its running) before 559 * starting the new one. 560 */ 561 uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; 562 int32_t idx; 563 int32_t is_tlp_timer = 0; 564 struct bbr_sendmap *rsm; 565 566 if (bbr->rc_all_timers_stopped) { 567 /* All timers have been stopped none are to run */ 568 return (0); 569 } 570 if (bbr->rc_in_persist) { 571 /* We can't start any timer in persists */ 572 return (bbr_get_persists_timer_val(tp, bbr)); 573 } 574 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 575 if ((rsm == NULL) || 576 ((tp->t_flags & TF_SACK_PERMIT) == 0) || 577 (tp->t_state < TCPS_ESTABLISHED)) { 578 /* Nothing on the send map */ 579 activate_rxt: 580 if (SEQ_LT(tp->snd_una, tp->snd_max) || 581 sbavail(&tptosocket(tp)->so_snd)) { 582 uint64_t tov; 583 584 time_since_sent = 0; 585 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 586 if (rsm) { 587 idx = rsm->r_rtr_cnt - 1; 588 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 589 tstmp_touse = rsm->r_tim_lastsent[idx]; 590 else 591 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 592 if (TSTMP_GT(tstmp_touse, cts)) 593 time_since_sent = cts - tstmp_touse; 594 } 595 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; 596 if (tp->t_srtt == 0) 597 tov = BBR_INITIAL_RTO; 598 else 599 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + 600 ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); 601 if (tp->t_rxtshift) 602 tov *= tcp_backoff[tp->t_rxtshift]; 603 if (tov > time_since_sent) 604 tov -= time_since_sent; 605 else 606 tov = bbr->r_ctl.rc_min_to; 607 TCPT_RANGESET_NOSLOP(to, tov, 608 (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), 609 (bbr->rc_max_rto_sec * USECS_IN_SECOND)); 610 bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); 611 return (to); 612 } 613 return (0); 614 } 615 if (rsm->r_flags & BBR_ACKED) { 616 rsm = bbr_find_lowest_rsm(bbr); 617 if (rsm == NULL) { 618 /* No lowest? */ 619 goto activate_rxt; 620 } 621 } 622 /* Convert from ms to usecs */ 623 if (rsm->r_flags & BBR_SACK_PASSED) { 624 if ((tp->t_flags & TF_SENTFIN) && 625 ((tp->snd_max - tp->snd_una) == 1) && 626 (rsm->r_flags & BBR_HAS_FIN)) { 627 /* 628 * We don't start a bbr rack timer if all we have is 629 * a FIN outstanding. 630 */ 631 goto activate_rxt; 632 } 633 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); 634 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); 635 idx = rsm->r_rtr_cnt - 1; 636 exp = rsm->r_tim_lastsent[idx] + thresh; 637 if (SEQ_GEQ(exp, cts)) { 638 to = exp - cts; 639 if (to < bbr->r_ctl.rc_min_to) { 640 to = bbr->r_ctl.rc_min_to; 641 } 642 } else { 643 to = bbr->r_ctl.rc_min_to; 644 } 645 } else { 646 /* Ok we need to do a TLP not RACK */ 647 if (bbr->rc_tlp_in_progress != 0) { 648 /* 649 * The previous send was a TLP. 650 */ 651 goto activate_rxt; 652 } 653 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 654 if (rsm == NULL) { 655 /* We found no rsm to TLP with. */ 656 goto activate_rxt; 657 } 658 if (rsm->r_flags & BBR_HAS_FIN) { 659 /* If its a FIN we don't do TLP */ 660 rsm = NULL; 661 goto activate_rxt; 662 } 663 time_since_sent = 0; 664 idx = rsm->r_rtr_cnt - 1; 665 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) 666 tstmp_touse = rsm->r_tim_lastsent[idx]; 667 else 668 tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; 669 if (TSTMP_GT(tstmp_touse, cts)) 670 time_since_sent = cts - tstmp_touse; 671 is_tlp_timer = 1; 672 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); 673 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); 674 if (thresh > time_since_sent) 675 to = thresh - time_since_sent; 676 else 677 to = bbr->r_ctl.rc_min_to; 678 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 679 /* 680 * If the TLP time works out to larger than the max 681 * RTO lets not do TLP.. just RTO. 682 */ 683 goto activate_rxt; 684 } 685 if ((bbr->rc_tlp_rtx_out == 1) && 686 (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { 687 /* 688 * Second retransmit of the same TLP 689 * lets not. 690 */ 691 bbr->rc_tlp_rtx_out = 0; 692 goto activate_rxt; 693 } 694 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { 695 /* 696 * The tail is no longer the last one I did a probe 697 * on 698 */ 699 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 700 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 701 } 702 } 703 if (is_tlp_timer == 0) { 704 BBR_STAT_INC(bbr_to_arm_rack); 705 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; 706 } else { 707 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); 708 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 709 /* 710 * We have exceeded how many times we can retran the 711 * current TLP timer, switch to the RTO timer. 712 */ 713 goto activate_rxt; 714 } else { 715 BBR_STAT_INC(bbr_to_arm_tlp); 716 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; 717 } 718 } 719 return (to); 720 } 721 722 static inline int32_t 723 bbr_minseg(struct tcp_bbr *bbr) 724 { 725 return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); 726 } 727 728 static void 729 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) 730 { 731 struct inpcb *inp = tptoinpcb(tp); 732 struct hpts_diag diag; 733 uint32_t delayed_ack = 0; 734 uint32_t left = 0; 735 uint32_t hpts_timeout; 736 uint8_t stopped; 737 int32_t delay_calc = 0; 738 uint32_t prev_delay = 0; 739 740 if (tcp_in_hpts(tp)) { 741 /* A previous call is already set up */ 742 return; 743 } 744 if ((tp->t_state == TCPS_CLOSED) || 745 (tp->t_state == TCPS_LISTEN)) { 746 return; 747 } 748 stopped = bbr->rc_tmr_stopped; 749 if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 750 left = bbr->r_ctl.rc_timer_exp - cts; 751 } 752 bbr->r_ctl.rc_hpts_flags = 0; 753 bbr->r_ctl.rc_timer_exp = 0; 754 prev_delay = bbr->r_ctl.rc_last_delay_val; 755 if (bbr->r_ctl.rc_last_delay_val && 756 (slot == 0)) { 757 /* 758 * If a previous pacer delay was in place we 759 * are not coming from the output side (where 760 * we calculate a delay, more likely a timer). 761 */ 762 slot = bbr->r_ctl.rc_last_delay_val; 763 if (TSTMP_GT(cts, bbr->rc_pacer_started)) { 764 /* Compensate for time passed */ 765 delay_calc = cts - bbr->rc_pacer_started; 766 if (delay_calc <= slot) 767 slot -= delay_calc; 768 } 769 } 770 /* Do we have early to make up for by pushing out the pacing time? */ 771 if (bbr->r_agg_early_set) { 772 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); 773 slot += bbr->r_ctl.rc_agg_early; 774 bbr->r_ctl.rc_agg_early = 0; 775 bbr->r_agg_early_set = 0; 776 } 777 /* Are we running a total debt that needs to be compensated for? */ 778 if (bbr->r_ctl.rc_hptsi_agg_delay) { 779 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { 780 /* We nuke the delay */ 781 slot -= bbr->r_ctl.rc_hptsi_agg_delay; 782 bbr->r_ctl.rc_hptsi_agg_delay = 0; 783 } else { 784 /* We nuke some of the delay, put in a minimal 100usecs */ 785 bbr->r_ctl.rc_hptsi_agg_delay -= slot; 786 bbr->r_ctl.rc_last_delay_val = slot = 100; 787 } 788 } 789 bbr->r_ctl.rc_last_delay_val = slot; 790 hpts_timeout = bbr_timer_start(tp, bbr, cts); 791 if (tp->t_flags & TF_DELACK) { 792 if (bbr->rc_in_persist == 0) { 793 delayed_ack = bbr_delack_time; 794 } else { 795 /* 796 * We are in persists and have 797 * gotten a new data element. 798 */ 799 if (hpts_timeout > bbr_delack_time) { 800 /* 801 * Lets make the persists timer (which acks) 802 * be the smaller of hpts_timeout and bbr_delack_time. 803 */ 804 hpts_timeout = bbr_delack_time; 805 } 806 } 807 } 808 if (delayed_ack && 809 ((hpts_timeout == 0) || 810 (delayed_ack < hpts_timeout))) { 811 /* We need a Delayed ack timer */ 812 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 813 hpts_timeout = delayed_ack; 814 } 815 if (slot) { 816 /* Mark that we have a pacing timer up */ 817 BBR_STAT_INC(bbr_paced_segments); 818 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; 819 } 820 /* 821 * If no timers are going to run and we will fall off thfe hptsi 822 * wheel, we resort to a keep-alive timer if its configured. 823 */ 824 if ((hpts_timeout == 0) && 825 (slot == 0)) { 826 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 827 (tp->t_state <= TCPS_CLOSING)) { 828 /* 829 * Ok we have no timer (persists, rack, tlp, rxt or 830 * del-ack), we don't have segments being paced. So 831 * all that is left is the keepalive timer. 832 */ 833 if (TCPS_HAVEESTABLISHED(tp->t_state)) { 834 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); 835 } else { 836 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); 837 } 838 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; 839 } 840 } 841 if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == 842 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { 843 /* 844 * RACK, TLP, persists and RXT timers all are restartable 845 * based on actions input .. i.e we received a packet (ack 846 * or sack) and that changes things (rw, or snd_una etc). 847 * Thus we can restart them with a new value. For 848 * keep-alive, delayed_ack we keep track of what was left 849 * and restart the timer with a smaller value. 850 */ 851 if (left < hpts_timeout) 852 hpts_timeout = left; 853 } 854 if (bbr->r_ctl.rc_incr_tmrs && slot && 855 (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { 856 /* 857 * If configured to do so, and the timer is either 858 * the TLP or RXT timer, we need to increase the timeout 859 * by the pacing time. Consider the bottleneck at my 860 * machine as an example, we are sending something 861 * to start a TLP on. The last packet won't be emitted 862 * fully until the pacing time (the bottleneck will hold 863 * the data in place). Once the packet is emitted that 864 * is when we want to start waiting for the TLP. This 865 * is most evident with hardware pacing (where the nic 866 * is holding the packet(s) before emitting). But it 867 * can also show up in the network so we do it for all 868 * cases. Technically we would take off one packet from 869 * this extra delay but this is easier and being more 870 * conservative is probably better. 871 */ 872 hpts_timeout += slot; 873 } 874 if (hpts_timeout) { 875 /* 876 * Hack alert for now we can't time-out over 2147 seconds (a 877 * bit more than 35min) 878 */ 879 if (hpts_timeout > 0x7ffffffe) 880 hpts_timeout = 0x7ffffffe; 881 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 882 } else 883 bbr->r_ctl.rc_timer_exp = 0; 884 if ((slot) && 885 (bbr->rc_use_google || 886 bbr->output_error_seen || 887 (slot <= hpts_timeout)) ) { 888 /* 889 * Tell LRO that it can queue packets while 890 * we pace. 891 */ 892 bbr->rc_tp->t_flags2 |= TF2_MBUF_QUEUE_READY; 893 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 894 (bbr->rc_cwnd_limited == 0)) { 895 /* 896 * If we are not cwnd limited and we 897 * are running a rack timer we put on 898 * the do not disturbe even for sack. 899 */ 900 tp->t_flags2 |= TF2_DONT_SACK_QUEUE; 901 } else 902 tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE; 903 bbr->rc_pacer_started = cts; 904 905 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot), 906 __LINE__, &diag); 907 bbr->rc_timer_first = 0; 908 bbr->bbr_timer_src = frm; 909 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); 910 bbr_log_hpts_diag(bbr, cts, &diag); 911 } else if (hpts_timeout) { 912 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout), 913 __LINE__, &diag); 914 /* 915 * We add the flag here as well if the slot is set, 916 * since hpts will call in to clear the queue first before 917 * calling the output routine (which does our timers). 918 * We don't want to set the flag if its just a timer 919 * else the arrival of data might (that causes us 920 * to send more) might get delayed. Imagine being 921 * on a keep-alive timer and a request comes in for 922 * more data. 923 */ 924 if (slot) 925 bbr->rc_pacer_started = cts; 926 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && 927 (bbr->rc_cwnd_limited == 0)) { 928 /* 929 * For a rack timer, don't wake us even 930 * if a sack arrives as long as we are 931 * not cwnd limited. 932 */ 933 tp->t_flags2 |= (TF2_MBUF_QUEUE_READY | 934 TF2_DONT_SACK_QUEUE); 935 } else { 936 /* All other timers wake us up */ 937 tp->t_flags2 &= ~(TF2_MBUF_QUEUE_READY | 938 TF2_DONT_SACK_QUEUE); 939 } 940 bbr->bbr_timer_src = frm; 941 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); 942 bbr_log_hpts_diag(bbr, cts, &diag); 943 bbr->rc_timer_first = 1; 944 } 945 bbr->rc_tmr_stopped = 0; 946 bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); 947 } 948 949 static void 950 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) 951 { 952 /* 953 * We received an ack, and then did not call send or were bounced 954 * out due to the hpts was running. Now a timer is up as well, is it 955 * the right timer? 956 */ 957 struct inpcb *inp; 958 struct bbr_sendmap *rsm; 959 uint32_t hpts_timeout; 960 int tmr_up; 961 962 tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 963 if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) 964 return; 965 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 966 if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && 967 (tmr_up == PACE_TMR_RXT)) { 968 /* Should be an RXT */ 969 return; 970 } 971 inp = bbr->rc_inp; 972 if (rsm == NULL) { 973 /* Nothing outstanding? */ 974 if (tp->t_flags & TF_DELACK) { 975 if (tmr_up == PACE_TMR_DELACK) 976 /* 977 * We are supposed to have delayed ack up 978 * and we do 979 */ 980 return; 981 } else if (sbavail(&inp->inp_socket->so_snd) && 982 (tmr_up == PACE_TMR_RXT)) { 983 /* 984 * if we hit enobufs then we would expect the 985 * possibility of nothing outstanding and the RXT up 986 * (and the hptsi timer). 987 */ 988 return; 989 } else if (((V_tcp_always_keepalive || 990 inp->inp_socket->so_options & SO_KEEPALIVE) && 991 (tp->t_state <= TCPS_CLOSING)) && 992 (tmr_up == PACE_TMR_KEEP) && 993 (tp->snd_max == tp->snd_una)) { 994 /* We should have keep alive up and we do */ 995 return; 996 } 997 } 998 if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { 999 if ((tp->t_flags & TF_SENTFIN) && 1000 ((tp->snd_max - tp->snd_una) == 1) && 1001 (rsm->r_flags & BBR_HAS_FIN)) { 1002 /* needs to be a RXT */ 1003 if (tmr_up == PACE_TMR_RXT) 1004 return; 1005 else 1006 goto wrong_timer; 1007 } else if (tmr_up == PACE_TMR_RACK) 1008 return; 1009 else 1010 goto wrong_timer; 1011 } else if (rsm && (tmr_up == PACE_TMR_RACK)) { 1012 /* Rack timer has priority if we have data out */ 1013 return; 1014 } else if (SEQ_GT(tp->snd_max, tp->snd_una) && 1015 ((tmr_up == PACE_TMR_TLP) || 1016 (tmr_up == PACE_TMR_RXT))) { 1017 /* 1018 * Either a TLP or RXT is fine if no sack-passed is in place 1019 * and data is outstanding. 1020 */ 1021 return; 1022 } else if (tmr_up == PACE_TMR_DELACK) { 1023 /* 1024 * If the delayed ack was going to go off before the 1025 * rtx/tlp/rack timer were going to expire, then that would 1026 * be the timer in control. Note we don't check the time 1027 * here trusting the code is correct. 1028 */ 1029 return; 1030 } 1031 if (SEQ_GT(tp->snd_max, tp->snd_una) && 1032 ((tmr_up == PACE_TMR_RXT) || 1033 (tmr_up == PACE_TMR_TLP) || 1034 (tmr_up == PACE_TMR_RACK))) { 1035 /* 1036 * We have outstanding data and 1037 * we *do* have a RACK, TLP or RXT 1038 * timer running. We won't restart 1039 * anything here since thats probably ok we 1040 * will get called with some timer here shortly. 1041 */ 1042 return; 1043 } 1044 /* 1045 * Ok the timer originally started is not what we want now. We will 1046 * force the hpts to be stopped if any, and restart with the slot 1047 * set to what was in the saved slot. 1048 */ 1049 wrong_timer: 1050 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { 1051 if (tcp_in_hpts(tp)) 1052 tcp_hpts_remove(tp); 1053 bbr_timer_cancel(bbr, __LINE__, cts); 1054 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 1055 0); 1056 } else { 1057 /* 1058 * Output is hptsi so we just need to switch the type of 1059 * timer. We don't bother with keep-alive, since when we 1060 * jump through the output, it will start the keep-alive if 1061 * nothing is sent. 1062 * 1063 * We only need a delayed-ack added and or the hpts_timeout. 1064 */ 1065 hpts_timeout = bbr_timer_start(tp, bbr, cts); 1066 if (tp->t_flags & TF_DELACK) { 1067 if (hpts_timeout == 0) { 1068 hpts_timeout = bbr_delack_time; 1069 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1070 } 1071 else if (hpts_timeout > bbr_delack_time) { 1072 hpts_timeout = bbr_delack_time; 1073 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; 1074 } 1075 } 1076 if (hpts_timeout) { 1077 if (hpts_timeout > 0x7ffffffe) 1078 hpts_timeout = 0x7ffffffe; 1079 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; 1080 } 1081 } 1082 } 1083 1084 int32_t bbr_clear_lost = 0; 1085 1086 /* 1087 * Considers the two time values now (cts) and earlier. 1088 * If cts is smaller than earlier, we could have 1089 * had a sequence wrap (our counter wraps every 1090 * 70 min or so) or it could be just clock skew 1091 * getting us two different time values. Clock skew 1092 * will show up within 10ms or so. So in such 1093 * a case (where cts is behind earlier time by 1094 * less than 10ms) we return 0. Otherwise we 1095 * return the true difference between them. 1096 */ 1097 static inline uint32_t 1098 bbr_calc_time(uint32_t cts, uint32_t earlier_time) { 1099 /* 1100 * Given two timestamps, the current time stamp cts, and some other 1101 * time-stamp taken in theory earlier return the difference. The 1102 * trick is here sometimes locking will get the other timestamp 1103 * after the cts. If this occurs we need to return 0. 1104 */ 1105 if (TSTMP_GEQ(cts, earlier_time)) 1106 return (cts - earlier_time); 1107 /* 1108 * cts is behind earlier_time if its less than 10ms consider it 0. 1109 * If its more than 10ms difference then we had a time wrap. Else 1110 * its just the normal locking foo. I wonder if we should not go to 1111 * 64bit TS and get rid of this issue. 1112 */ 1113 if (TSTMP_GEQ((cts + 10000), earlier_time)) 1114 return (0); 1115 /* 1116 * Ok the time must have wrapped. So we need to answer a large 1117 * amount of time, which the normal subtraction should do. 1118 */ 1119 return (cts - earlier_time); 1120 } 1121 1122 static int 1123 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) 1124 { 1125 uint32_t stat; 1126 int32_t error; 1127 1128 error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); 1129 if (error || req->newptr == NULL) 1130 return error; 1131 1132 error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); 1133 if (error) 1134 return (error); 1135 if (stat == 1) { 1136 #ifdef BBR_INVARIANTS 1137 printf("Clearing BBR lost counters\n"); 1138 #endif 1139 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); 1140 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); 1141 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); 1142 } else if (stat == 2) { 1143 #ifdef BBR_INVARIANTS 1144 printf("Clearing BBR option counters\n"); 1145 #endif 1146 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); 1147 } else if (stat == 3) { 1148 #ifdef BBR_INVARIANTS 1149 printf("Clearing BBR stats counters\n"); 1150 #endif 1151 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); 1152 } else if (stat == 4) { 1153 #ifdef BBR_INVARIANTS 1154 printf("Clearing BBR out-size counters\n"); 1155 #endif 1156 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); 1157 } 1158 bbr_clear_lost = 0; 1159 return (0); 1160 } 1161 1162 static void 1163 bbr_init_sysctls(void) 1164 { 1165 struct sysctl_oid *bbr_probertt; 1166 struct sysctl_oid *bbr_hptsi; 1167 struct sysctl_oid *bbr_measure; 1168 struct sysctl_oid *bbr_cwnd; 1169 struct sysctl_oid *bbr_timeout; 1170 struct sysctl_oid *bbr_states; 1171 struct sysctl_oid *bbr_startup; 1172 struct sysctl_oid *bbr_policer; 1173 1174 /* Probe rtt controls */ 1175 bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1176 SYSCTL_CHILDREN(bbr_sysctl_root), 1177 OID_AUTO, 1178 "probertt", 1179 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1180 ""); 1181 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1182 SYSCTL_CHILDREN(bbr_probertt), 1183 OID_AUTO, "gain", CTLFLAG_RW, 1184 &bbr_rttprobe_gain, 192, 1185 "What is the filter gain drop in probe_rtt (0=disable)?"); 1186 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1187 SYSCTL_CHILDREN(bbr_probertt), 1188 OID_AUTO, "cwnd", CTLFLAG_RW, 1189 &bbr_rtt_probe_cwndtarg, 4, 1190 "How many mss's are outstanding during probe-rtt"); 1191 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1192 SYSCTL_CHILDREN(bbr_probertt), 1193 OID_AUTO, "int", CTLFLAG_RW, 1194 &bbr_rtt_probe_limit, 4000000, 1195 "If RTT has not shrank in this many micro-seconds enter probe-rtt"); 1196 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1197 SYSCTL_CHILDREN(bbr_probertt), 1198 OID_AUTO, "mintime", CTLFLAG_RW, 1199 &bbr_rtt_probe_time, 200000, 1200 "How many microseconds in probe-rtt"); 1201 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1202 SYSCTL_CHILDREN(bbr_probertt), 1203 OID_AUTO, "filter_len_sec", CTLFLAG_RW, 1204 &bbr_filter_len_sec, 6, 1205 "How long in seconds does the rttProp filter run?"); 1206 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1207 SYSCTL_CHILDREN(bbr_probertt), 1208 OID_AUTO, "drain_rtt", CTLFLAG_RW, 1209 &bbr_drain_rtt, BBR_SRTT, 1210 "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); 1211 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1212 SYSCTL_CHILDREN(bbr_probertt), 1213 OID_AUTO, "can_force", CTLFLAG_RW, 1214 &bbr_can_force_probertt, 0, 1215 "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); 1216 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1217 SYSCTL_CHILDREN(bbr_probertt), 1218 OID_AUTO, "enter_sets_force", CTLFLAG_RW, 1219 &bbr_probertt_sets_rtt, 0, 1220 "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); 1221 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1222 SYSCTL_CHILDREN(bbr_probertt), 1223 OID_AUTO, "can_adjust", CTLFLAG_RW, 1224 &bbr_can_adjust_probertt, 1, 1225 "Can we dynamically adjust the probe-rtt limits and times?"); 1226 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1227 SYSCTL_CHILDREN(bbr_probertt), 1228 OID_AUTO, "is_ratio", CTLFLAG_RW, 1229 &bbr_is_ratio, 0, 1230 "is the limit to filter a ratio?"); 1231 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1232 SYSCTL_CHILDREN(bbr_probertt), 1233 OID_AUTO, "use_cwnd", CTLFLAG_RW, 1234 &bbr_prtt_slam_cwnd, 0, 1235 "Should we set/recover cwnd?"); 1236 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1237 SYSCTL_CHILDREN(bbr_probertt), 1238 OID_AUTO, "can_use_ts", CTLFLAG_RW, 1239 &bbr_can_use_ts_for_rtt, 1, 1240 "Can we use the ms timestamp if available for retransmistted rtt calculations?"); 1241 1242 /* Pacing controls */ 1243 bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1244 SYSCTL_CHILDREN(bbr_sysctl_root), 1245 OID_AUTO, 1246 "pacing", 1247 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1248 ""); 1249 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1250 SYSCTL_CHILDREN(bbr_hptsi), 1251 OID_AUTO, "hw_pacing", CTLFLAG_RW, 1252 &bbr_allow_hdwr_pacing, 1, 1253 "Do we allow hardware pacing?"); 1254 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1255 SYSCTL_CHILDREN(bbr_hptsi), 1256 OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, 1257 &bbr_hardware_pacing_limit, 4000, 1258 "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); 1259 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1260 SYSCTL_CHILDREN(bbr_hptsi), 1261 OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, 1262 &bbr_hdwr_pace_adjust, 2, 1263 "Multiplier to calculated tso size?"); 1264 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1265 SYSCTL_CHILDREN(bbr_hptsi), 1266 OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, 1267 &bbr_hdwr_pace_floor, 1, 1268 "Do we invoke the hardware pacing floor?"); 1269 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1270 SYSCTL_CHILDREN(bbr_hptsi), 1271 OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, 1272 &bbr_hdwr_pacing_delay_cnt, 10, 1273 "How many packets must be sent after hdwr pacing is enabled"); 1274 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1275 SYSCTL_CHILDREN(bbr_hptsi), 1276 OID_AUTO, "bw_cross", CTLFLAG_RW, 1277 &bbr_cross_over, 3000000, 1278 "What is the point where we cross over to linux like TSO size set"); 1279 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1280 SYSCTL_CHILDREN(bbr_hptsi), 1281 OID_AUTO, "seg_deltarg", CTLFLAG_RW, 1282 &bbr_hptsi_segments_delay_tar, 7000, 1283 "What is the worse case delay target for hptsi < 48Mbp connections"); 1284 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1285 SYSCTL_CHILDREN(bbr_hptsi), 1286 OID_AUTO, "enet_oh", CTLFLAG_RW, 1287 &bbr_include_enet_oh, 0, 1288 "Do we include the ethernet overhead in calculating pacing delay?"); 1289 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1290 SYSCTL_CHILDREN(bbr_hptsi), 1291 OID_AUTO, "ip_oh", CTLFLAG_RW, 1292 &bbr_include_ip_oh, 1, 1293 "Do we include the IP overhead in calculating pacing delay?"); 1294 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1295 SYSCTL_CHILDREN(bbr_hptsi), 1296 OID_AUTO, "tcp_oh", CTLFLAG_RW, 1297 &bbr_include_tcp_oh, 0, 1298 "Do we include the TCP overhead in calculating pacing delay?"); 1299 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1300 SYSCTL_CHILDREN(bbr_hptsi), 1301 OID_AUTO, "google_discount", CTLFLAG_RW, 1302 &bbr_google_discount, 10, 1303 "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); 1304 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1305 SYSCTL_CHILDREN(bbr_hptsi), 1306 OID_AUTO, "all_get_min", CTLFLAG_RW, 1307 &bbr_all_get_min, 0, 1308 "If you are less than a MSS do you just get the min?"); 1309 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1310 SYSCTL_CHILDREN(bbr_hptsi), 1311 OID_AUTO, "tso_min", CTLFLAG_RW, 1312 &bbr_hptsi_bytes_min, 1460, 1313 "For 0 -> 24Mbps what is floor number of segments for TSO"); 1314 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1315 SYSCTL_CHILDREN(bbr_hptsi), 1316 OID_AUTO, "seg_tso_max", CTLFLAG_RW, 1317 &bbr_hptsi_segments_max, 6, 1318 "For 0 -> 24Mbps what is top number of segments for TSO"); 1319 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1320 SYSCTL_CHILDREN(bbr_hptsi), 1321 OID_AUTO, "seg_floor", CTLFLAG_RW, 1322 &bbr_hptsi_segments_floor, 1, 1323 "Minimum TSO size we will fall too in segments"); 1324 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1325 SYSCTL_CHILDREN(bbr_hptsi), 1326 OID_AUTO, "utter_max", CTLFLAG_RW, 1327 &bbr_hptsi_utter_max, 0, 1328 "The absolute maximum that any pacing (outside of hardware) can be"); 1329 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1330 SYSCTL_CHILDREN(bbr_hptsi), 1331 OID_AUTO, "seg_divisor", CTLFLAG_RW, 1332 &bbr_hptsi_per_second, 100, 1333 "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); 1334 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1335 SYSCTL_CHILDREN(bbr_hptsi), 1336 OID_AUTO, "srtt_mul", CTLFLAG_RW, 1337 &bbr_hptsi_max_mul, 1, 1338 "The multiplier for pace len max"); 1339 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1340 SYSCTL_CHILDREN(bbr_hptsi), 1341 OID_AUTO, "srtt_div", CTLFLAG_RW, 1342 &bbr_hptsi_max_div, 2, 1343 "The divisor for pace len max"); 1344 /* Measurement controls */ 1345 bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1346 SYSCTL_CHILDREN(bbr_sysctl_root), 1347 OID_AUTO, 1348 "measure", 1349 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1350 "Measurement controls"); 1351 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1352 SYSCTL_CHILDREN(bbr_measure), 1353 OID_AUTO, "min_i_bw", CTLFLAG_RW, 1354 &bbr_initial_bw_bps, 62500, 1355 "Minimum initial b/w in bytes per second"); 1356 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1357 SYSCTL_CHILDREN(bbr_measure), 1358 OID_AUTO, "no_sack_needed", CTLFLAG_RW, 1359 &bbr_sack_not_required, 0, 1360 "Do we allow bbr to run on connections not supporting SACK?"); 1361 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1362 SYSCTL_CHILDREN(bbr_measure), 1363 OID_AUTO, "use_google", CTLFLAG_RW, 1364 &bbr_use_google_algo, 0, 1365 "Use has close to google V1.0 has possible?"); 1366 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1367 SYSCTL_CHILDREN(bbr_measure), 1368 OID_AUTO, "ts_limiting", CTLFLAG_RW, 1369 &bbr_ts_limiting, 1, 1370 "Do we attempt to use the peers timestamp to limit b/w caculations?"); 1371 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1372 SYSCTL_CHILDREN(bbr_measure), 1373 OID_AUTO, "ts_can_raise", CTLFLAG_RW, 1374 &bbr_ts_can_raise, 0, 1375 "Can we raise the b/w via timestamp b/w calculation?"); 1376 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1377 SYSCTL_CHILDREN(bbr_measure), 1378 OID_AUTO, "ts_delta", CTLFLAG_RW, 1379 &bbr_min_usec_delta, 20000, 1380 "How long in usec between ts of our sends in ts validation code?"); 1381 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1382 SYSCTL_CHILDREN(bbr_measure), 1383 OID_AUTO, "ts_peer_delta", CTLFLAG_RW, 1384 &bbr_min_peer_delta, 20, 1385 "What min numerical value should be between the peer deltas?"); 1386 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1387 SYSCTL_CHILDREN(bbr_measure), 1388 OID_AUTO, "ts_delta_percent", CTLFLAG_RW, 1389 &bbr_delta_percent, 150, 1390 "What percentage (150 = 15.0) do we allow variance for?"); 1391 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1392 SYSCTL_CHILDREN(bbr_measure), 1393 OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, 1394 &bbr_min_measurements_req, 1, 1395 "What is the minimum measurement count we need before we switch to our b/w estimate"); 1396 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1397 SYSCTL_CHILDREN(bbr_measure), 1398 OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, 1399 &bbr_no_pacing_until, 4, 1400 "How many pkt-epoch's (0 is off) do we need before pacing is on?"); 1401 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1402 SYSCTL_CHILDREN(bbr_measure), 1403 OID_AUTO, "quanta", CTLFLAG_RW, 1404 &bbr_quanta, 2, 1405 "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); 1406 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1407 SYSCTL_CHILDREN(bbr_measure), 1408 OID_AUTO, "noretran", CTLFLAG_RW, 1409 &bbr_no_retran, 0, 1410 "Should google mode not use retransmission measurements for the b/w estimation?"); 1411 /* State controls */ 1412 bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1413 SYSCTL_CHILDREN(bbr_sysctl_root), 1414 OID_AUTO, 1415 "states", 1416 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1417 "State controls"); 1418 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1419 SYSCTL_CHILDREN(bbr_states), 1420 OID_AUTO, "idle_restart", CTLFLAG_RW, 1421 &bbr_uses_idle_restart, 0, 1422 "Do we use a new special idle_restart state to ramp back up quickly?"); 1423 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1424 SYSCTL_CHILDREN(bbr_states), 1425 OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, 1426 &bbr_idle_restart_threshold, 100000, 1427 "How long must we be idle before we restart??"); 1428 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1429 SYSCTL_CHILDREN(bbr_states), 1430 OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, 1431 &bbr_state_is_pkt_epoch, 0, 1432 "Do we use a pkt-epoch for substate if 0 rttProp?"); 1433 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1434 SYSCTL_CHILDREN(bbr_states), 1435 OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, 1436 &bbr_rtt_gain_thresh, 0, 1437 "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); 1438 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1439 SYSCTL_CHILDREN(bbr_states), 1440 OID_AUTO, "drain_floor", CTLFLAG_RW, 1441 &bbr_drain_floor, 88, 1442 "What is the lowest we can drain (pg) too?"); 1443 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1444 SYSCTL_CHILDREN(bbr_states), 1445 OID_AUTO, "drain_2_target", CTLFLAG_RW, 1446 &bbr_state_drain_2_tar, 1, 1447 "Do we drain to target in drain substate?"); 1448 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1449 SYSCTL_CHILDREN(bbr_states), 1450 OID_AUTO, "gain_2_target", CTLFLAG_RW, 1451 &bbr_gain_to_target, 1, 1452 "Does probe bw gain to target??"); 1453 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1454 SYSCTL_CHILDREN(bbr_states), 1455 OID_AUTO, "gain_extra_time", CTLFLAG_RW, 1456 &bbr_gain_gets_extra_too, 1, 1457 "Does probe bw gain get the extra time too?"); 1458 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1459 SYSCTL_CHILDREN(bbr_states), 1460 OID_AUTO, "ld_div", CTLFLAG_RW, 1461 &bbr_drain_drop_div, 5, 1462 "Long drain drop divider?"); 1463 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1464 SYSCTL_CHILDREN(bbr_states), 1465 OID_AUTO, "ld_mul", CTLFLAG_RW, 1466 &bbr_drain_drop_mul, 4, 1467 "Long drain drop multiplier?"); 1468 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1469 SYSCTL_CHILDREN(bbr_states), 1470 OID_AUTO, "rand_ot_disc", CTLFLAG_RW, 1471 &bbr_rand_ot, 50, 1472 "Random discount of the ot?"); 1473 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1474 SYSCTL_CHILDREN(bbr_states), 1475 OID_AUTO, "dr_filter_life", CTLFLAG_RW, 1476 &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, 1477 "How many packet-epochs does the b/w delivery rate last?"); 1478 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1479 SYSCTL_CHILDREN(bbr_states), 1480 OID_AUTO, "subdrain_applimited", CTLFLAG_RW, 1481 &bbr_sub_drain_app_limit, 0, 1482 "Does our sub-state drain invoke app limited if its long?"); 1483 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1484 SYSCTL_CHILDREN(bbr_states), 1485 OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, 1486 &bbr_sub_drain_slam_cwnd, 0, 1487 "Should we set/recover cwnd for sub-state drain?"); 1488 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1489 SYSCTL_CHILDREN(bbr_states), 1490 OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, 1491 &bbr_slam_cwnd_in_main_drain, 0, 1492 "Should we set/recover cwnd for main-state drain?"); 1493 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1494 SYSCTL_CHILDREN(bbr_states), 1495 OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, 1496 &google_allow_early_out, 1, 1497 "Should we allow google probe-bw/drain to exit early at flight target?"); 1498 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1499 SYSCTL_CHILDREN(bbr_states), 1500 OID_AUTO, "google_exit_loss", CTLFLAG_RW, 1501 &google_consider_lost, 1, 1502 "Should we have losses exit gain of probebw in google mode??"); 1503 /* Startup controls */ 1504 bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1505 SYSCTL_CHILDREN(bbr_sysctl_root), 1506 OID_AUTO, 1507 "startup", 1508 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1509 "Startup controls"); 1510 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1511 SYSCTL_CHILDREN(bbr_startup), 1512 OID_AUTO, "cheat_iwnd", CTLFLAG_RW, 1513 &bbr_sends_full_iwnd, 1, 1514 "Do we not pace but burst out initial windows has our TSO size?"); 1515 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1516 SYSCTL_CHILDREN(bbr_startup), 1517 OID_AUTO, "loss_threshold", CTLFLAG_RW, 1518 &bbr_startup_loss_thresh, 2000, 1519 "In startup what is the loss threshold in a pe that will exit us from startup?"); 1520 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1521 SYSCTL_CHILDREN(bbr_startup), 1522 OID_AUTO, "use_lowerpg", CTLFLAG_RW, 1523 &bbr_use_lower_gain_in_startup, 1, 1524 "Should we use a lower hptsi gain if we see loss in startup?"); 1525 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1526 SYSCTL_CHILDREN(bbr_startup), 1527 OID_AUTO, "gain", CTLFLAG_RW, 1528 &bbr_start_exit, 25, 1529 "What gain percent do we need to see to stay in startup??"); 1530 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1531 SYSCTL_CHILDREN(bbr_startup), 1532 OID_AUTO, "low_gain", CTLFLAG_RW, 1533 &bbr_low_start_exit, 15, 1534 "What gain percent do we need to see to stay in the lower gain startup??"); 1535 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1536 SYSCTL_CHILDREN(bbr_startup), 1537 OID_AUTO, "loss_exit", CTLFLAG_RW, 1538 &bbr_exit_startup_at_loss, 1, 1539 "Should we exit startup at loss in an epoch if we are not gaining?"); 1540 /* CWND controls */ 1541 bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1542 SYSCTL_CHILDREN(bbr_sysctl_root), 1543 OID_AUTO, 1544 "cwnd", 1545 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1546 "Cwnd controls"); 1547 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1548 SYSCTL_CHILDREN(bbr_cwnd), 1549 OID_AUTO, "tar_rtt", CTLFLAG_RW, 1550 &bbr_cwndtarget_rtt_touse, 0, 1551 "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); 1552 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1553 SYSCTL_CHILDREN(bbr_cwnd), 1554 OID_AUTO, "may_shrink", CTLFLAG_RW, 1555 &bbr_cwnd_may_shrink, 0, 1556 "Can the cwnd shrink if it would grow to more than the target?"); 1557 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1558 SYSCTL_CHILDREN(bbr_cwnd), 1559 OID_AUTO, "max_target_limit", CTLFLAG_RW, 1560 &bbr_target_cwnd_mult_limit, 8, 1561 "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); 1562 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1563 SYSCTL_CHILDREN(bbr_cwnd), 1564 OID_AUTO, "highspeed_min", CTLFLAG_RW, 1565 &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, 1566 "What is the high-speed min cwnd (rttProp under 1ms)"); 1567 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1568 SYSCTL_CHILDREN(bbr_cwnd), 1569 OID_AUTO, "lowspeed_min", CTLFLAG_RW, 1570 &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, 1571 "What is the min cwnd (rttProp > 1ms)"); 1572 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1573 SYSCTL_CHILDREN(bbr_cwnd), 1574 OID_AUTO, "initwin", CTLFLAG_RW, 1575 &bbr_def_init_win, 10, 1576 "What is the BBR initial window, if 0 use tcp version"); 1577 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1578 SYSCTL_CHILDREN(bbr_cwnd), 1579 OID_AUTO, "do_loss_red", CTLFLAG_RW, 1580 &bbr_do_red, 600, 1581 "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); 1582 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1583 SYSCTL_CHILDREN(bbr_cwnd), 1584 OID_AUTO, "red_scale", CTLFLAG_RW, 1585 &bbr_red_scale, 20000, 1586 "What RTT do we scale with?"); 1587 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1588 SYSCTL_CHILDREN(bbr_cwnd), 1589 OID_AUTO, "red_growslow", CTLFLAG_RW, 1590 &bbr_red_growth_restrict, 1, 1591 "Do we restrict cwnd growth for whats in flight?"); 1592 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1593 SYSCTL_CHILDREN(bbr_cwnd), 1594 OID_AUTO, "red_div", CTLFLAG_RW, 1595 &bbr_red_div, 2, 1596 "If we reduce whats the divisor?"); 1597 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1598 SYSCTL_CHILDREN(bbr_cwnd), 1599 OID_AUTO, "red_mul", CTLFLAG_RW, 1600 &bbr_red_mul, 1, 1601 "If we reduce whats the mulitiplier?"); 1602 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1603 SYSCTL_CHILDREN(bbr_cwnd), 1604 OID_AUTO, "target_is_unit", CTLFLAG_RW, 1605 &bbr_target_is_bbunit, 0, 1606 "Is the state target the pacing_gain or BBR_UNIT?"); 1607 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1608 SYSCTL_CHILDREN(bbr_cwnd), 1609 OID_AUTO, "drop_limit", CTLFLAG_RW, 1610 &bbr_drop_limit, 0, 1611 "Number of segments limit for drop (0=use min_cwnd w/flight)?"); 1612 1613 /* Timeout controls */ 1614 bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1615 SYSCTL_CHILDREN(bbr_sysctl_root), 1616 OID_AUTO, 1617 "timeout", 1618 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1619 "Time out controls"); 1620 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1621 SYSCTL_CHILDREN(bbr_timeout), 1622 OID_AUTO, "delack", CTLFLAG_RW, 1623 &bbr_delack_time, 100000, 1624 "BBR's delayed ack time"); 1625 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1626 SYSCTL_CHILDREN(bbr_timeout), 1627 OID_AUTO, "tlp_uses", CTLFLAG_RW, 1628 &bbr_tlp_type_to_use, 3, 1629 "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); 1630 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1631 SYSCTL_CHILDREN(bbr_timeout), 1632 OID_AUTO, "persmin", CTLFLAG_RW, 1633 &bbr_persist_min, 250000, 1634 "What is the minimum time in microseconds between persists"); 1635 SYSCTL_ADD_U32(&bbr_sysctl_ctx, 1636 SYSCTL_CHILDREN(bbr_timeout), 1637 OID_AUTO, "persmax", CTLFLAG_RW, 1638 &bbr_persist_max, 1000000, 1639 "What is the largest delay in microseconds between persists"); 1640 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1641 SYSCTL_CHILDREN(bbr_timeout), 1642 OID_AUTO, "tlp_minto", CTLFLAG_RW, 1643 &bbr_tlp_min, 10000, 1644 "TLP Min timeout in usecs"); 1645 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1646 SYSCTL_CHILDREN(bbr_timeout), 1647 OID_AUTO, "tlp_dack_time", CTLFLAG_RW, 1648 &bbr_delayed_ack_time, 200000, 1649 "TLP delayed ack compensation value"); 1650 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1651 SYSCTL_CHILDREN(bbr_sysctl_root), 1652 OID_AUTO, "minrto", CTLFLAG_RW, 1653 &bbr_rto_min_ms, 30, 1654 "Minimum RTO in ms"); 1655 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1656 SYSCTL_CHILDREN(bbr_timeout), 1657 OID_AUTO, "maxrto", CTLFLAG_RW, 1658 &bbr_rto_max_sec, 4, 1659 "Maximum RTO in seconds -- should be at least as large as min_rto"); 1660 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1661 SYSCTL_CHILDREN(bbr_timeout), 1662 OID_AUTO, "tlp_retry", CTLFLAG_RW, 1663 &bbr_tlp_max_resend, 2, 1664 "How many times does TLP retry a single segment or multiple with no ACK"); 1665 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1666 SYSCTL_CHILDREN(bbr_timeout), 1667 OID_AUTO, "minto", CTLFLAG_RW, 1668 &bbr_min_to, 1000, 1669 "Minimum rack timeout in useconds"); 1670 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1671 SYSCTL_CHILDREN(bbr_timeout), 1672 OID_AUTO, "pktdelay", CTLFLAG_RW, 1673 &bbr_pkt_delay, 1000, 1674 "Extra RACK time (in useconds) besides reordering thresh"); 1675 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1676 SYSCTL_CHILDREN(bbr_timeout), 1677 OID_AUTO, "incr_tmrs", CTLFLAG_RW, 1678 &bbr_incr_timers, 1, 1679 "Increase the RXT/TLP timer by the pacing time used?"); 1680 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1681 SYSCTL_CHILDREN(bbr_timeout), 1682 OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, 1683 &bbr_marks_rxt_sack_passed, 0, 1684 "Mark sack passed on all those not ack'd when a RXT hits?"); 1685 /* Policer controls */ 1686 bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 1687 SYSCTL_CHILDREN(bbr_sysctl_root), 1688 OID_AUTO, 1689 "policer", 1690 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 1691 "Policer controls"); 1692 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1693 SYSCTL_CHILDREN(bbr_policer), 1694 OID_AUTO, "detect_enable", CTLFLAG_RW, 1695 &bbr_policer_detection_enabled, 1, 1696 "Is policer detection enabled??"); 1697 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1698 SYSCTL_CHILDREN(bbr_policer), 1699 OID_AUTO, "min_pes", CTLFLAG_RW, 1700 &bbr_lt_intvl_min_rtts, 4, 1701 "Minimum number of PE's?"); 1702 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1703 SYSCTL_CHILDREN(bbr_policer), 1704 OID_AUTO, "bwdiff", CTLFLAG_RW, 1705 &bbr_lt_bw_diff, (4000/8), 1706 "Minimal bw diff?"); 1707 SYSCTL_ADD_U64(&bbr_sysctl_ctx, 1708 SYSCTL_CHILDREN(bbr_policer), 1709 OID_AUTO, "bwratio", CTLFLAG_RW, 1710 &bbr_lt_bw_ratio, 8, 1711 "Minimal bw diff?"); 1712 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1713 SYSCTL_CHILDREN(bbr_policer), 1714 OID_AUTO, "from_rack_rxt", CTLFLAG_RW, 1715 &bbr_policer_call_from_rack_to, 0, 1716 "Do we call the policer detection code from a rack-timeout?"); 1717 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1718 SYSCTL_CHILDREN(bbr_policer), 1719 OID_AUTO, "false_postive", CTLFLAG_RW, 1720 &bbr_lt_intvl_fp, 0, 1721 "What packet epoch do we do false-positive detection at (0=no)?"); 1722 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1723 SYSCTL_CHILDREN(bbr_policer), 1724 OID_AUTO, "loss_thresh", CTLFLAG_RW, 1725 &bbr_lt_loss_thresh, 196, 1726 "Loss threshold 196 = 19.6%?"); 1727 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1728 SYSCTL_CHILDREN(bbr_policer), 1729 OID_AUTO, "false_postive_thresh", CTLFLAG_RW, 1730 &bbr_lt_fd_thresh, 100, 1731 "What percentage is the false detection threshold (150=15.0)?"); 1732 /* All the rest */ 1733 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1734 SYSCTL_CHILDREN(bbr_sysctl_root), 1735 OID_AUTO, "cheat_rxt", CTLFLAG_RW, 1736 &bbr_use_rack_resend_cheat, 0, 1737 "Do we burst 1ms between sends on retransmissions (like rack)?"); 1738 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1739 SYSCTL_CHILDREN(bbr_sysctl_root), 1740 OID_AUTO, "error_paceout", CTLFLAG_RW, 1741 &bbr_error_base_paceout, 10000, 1742 "When we hit an error what is the min to pace out in usec's?"); 1743 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1744 SYSCTL_CHILDREN(bbr_sysctl_root), 1745 OID_AUTO, "kill_paceout", CTLFLAG_RW, 1746 &bbr_max_net_error_cnt, 10, 1747 "When we hit this many errors in a row, kill the session?"); 1748 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1749 SYSCTL_CHILDREN(bbr_sysctl_root), 1750 OID_AUTO, "data_after_close", CTLFLAG_RW, 1751 &bbr_ignore_data_after_close, 1, 1752 "Do we hold off sending a RST until all pending data is ack'd"); 1753 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1754 SYSCTL_CHILDREN(bbr_sysctl_root), 1755 OID_AUTO, "resend_use_tso", CTLFLAG_RW, 1756 &bbr_resends_use_tso, 0, 1757 "Can resends use TSO?"); 1758 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1759 SYSCTL_CHILDREN(bbr_sysctl_root), 1760 OID_AUTO, "sblklimit", CTLFLAG_RW, 1761 &bbr_sack_block_limit, 128, 1762 "When do we start ignoring small sack blocks"); 1763 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1764 SYSCTL_CHILDREN(bbr_sysctl_root), 1765 OID_AUTO, "bb_verbose", CTLFLAG_RW, 1766 &bbr_verbose_logging, 0, 1767 "Should BBR black box logging be verbose"); 1768 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1769 SYSCTL_CHILDREN(bbr_sysctl_root), 1770 OID_AUTO, "reorder_thresh", CTLFLAG_RW, 1771 &bbr_reorder_thresh, 2, 1772 "What factor for rack will be added when seeing reordering (shift right)"); 1773 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1774 SYSCTL_CHILDREN(bbr_sysctl_root), 1775 OID_AUTO, "reorder_fade", CTLFLAG_RW, 1776 &bbr_reorder_fade, 0, 1777 "Does reorder detection fade, if so how many ms (0 means never)"); 1778 SYSCTL_ADD_S32(&bbr_sysctl_ctx, 1779 SYSCTL_CHILDREN(bbr_sysctl_root), 1780 OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, 1781 &bbr_tlp_thresh, 1, 1782 "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); 1783 /* Stats and counters */ 1784 /* The pacing counters for hdwr/software can't be in the array */ 1785 bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1786 bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); 1787 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1788 SYSCTL_CHILDREN(bbr_sysctl_root), 1789 OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, 1790 &bbr_hdwr_pacing_enobuf, 1791 "Total number of enobufs for hardware paced flows"); 1792 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1793 SYSCTL_CHILDREN(bbr_sysctl_root), 1794 OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, 1795 &bbr_nohdwr_pacing_enobuf, 1796 "Total number of enobufs for non-hardware paced flows"); 1797 1798 bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); 1799 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1800 SYSCTL_CHILDREN(bbr_sysctl_root), 1801 OID_AUTO, "hdwr_pacing", CTLFLAG_RD, 1802 &bbr_flows_whdwr_pacing, 1803 "Total number of hardware paced flows"); 1804 bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); 1805 SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, 1806 SYSCTL_CHILDREN(bbr_sysctl_root), 1807 OID_AUTO, "software_pacing", CTLFLAG_RD, 1808 &bbr_flows_nohdwr_pacing, 1809 "Total number of software paced flows"); 1810 COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); 1811 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1812 OID_AUTO, "stats", CTLFLAG_RD, 1813 bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); 1814 COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); 1815 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1816 OID_AUTO, "opts", CTLFLAG_RD, 1817 bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); 1818 COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); 1819 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1820 OID_AUTO, "lost", CTLFLAG_RD, 1821 bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); 1822 COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); 1823 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1824 OID_AUTO, "stateresend", CTLFLAG_RD, 1825 bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); 1826 COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); 1827 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1828 OID_AUTO, "statetime", CTLFLAG_RD, 1829 bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); 1830 COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); 1831 SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), 1832 OID_AUTO, "outsize", CTLFLAG_RD, 1833 bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); 1834 SYSCTL_ADD_PROC(&bbr_sysctl_ctx, 1835 SYSCTL_CHILDREN(bbr_sysctl_root), 1836 OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1837 &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); 1838 } 1839 1840 static void 1841 bbr_counter_destroy(void) 1842 { 1843 COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); 1844 COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); 1845 COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); 1846 COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); 1847 COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); 1848 COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); 1849 counter_u64_free(bbr_nohdwr_pacing_enobuf); 1850 counter_u64_free(bbr_hdwr_pacing_enobuf); 1851 counter_u64_free(bbr_flows_whdwr_pacing); 1852 counter_u64_free(bbr_flows_nohdwr_pacing); 1853 1854 } 1855 1856 static __inline void 1857 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) 1858 { 1859 memset(l, 0, sizeof(union tcp_log_stackspecific)); 1860 l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; 1861 l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); 1862 l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 1863 l->bw_inuse = bbr_get_bw(bbr); 1864 l->inflight = ctf_flight_size(bbr->rc_tp, 1865 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 1866 l->applimited = bbr->r_ctl.r_app_limited_until; 1867 l->delivered = bbr->r_ctl.rc_delivered; 1868 l->timeStamp = cts; 1869 l->lost = bbr->r_ctl.rc_lost; 1870 l->bbr_state = bbr->rc_bbr_state; 1871 l->bbr_substate = bbr_state_val(bbr); 1872 l->epoch = bbr->r_ctl.rc_rtt_epoch; 1873 l->lt_epoch = bbr->r_ctl.rc_lt_epoch; 1874 l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 1875 l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; 1876 l->inhpts = tcp_in_hpts(bbr->rc_tp); 1877 l->use_lt_bw = bbr->rc_lt_use_bw; 1878 l->pkts_out = bbr->r_ctl.rc_flight_at_input; 1879 l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; 1880 } 1881 1882 static void 1883 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) 1884 { 1885 if (tcp_bblogging_on(bbr->rc_tp)) { 1886 union tcp_log_stackspecific log; 1887 1888 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1889 log.u_bbr.flex1 = 0; 1890 log.u_bbr.flex2 = 0; 1891 log.u_bbr.flex5 = 0; 1892 log.u_bbr.flex3 = 0; 1893 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; 1894 log.u_bbr.flex7 = reason; 1895 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; 1896 log.u_bbr.flex8 = 0; 1897 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1898 &bbr->rc_inp->inp_socket->so_rcv, 1899 &bbr->rc_inp->inp_socket->so_snd, 1900 BBR_LOG_BW_RED_EV, 0, 1901 0, &log, false, &bbr->rc_tv); 1902 } 1903 } 1904 1905 static void 1906 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) 1907 { 1908 if (tcp_bblogging_on(bbr->rc_tp)) { 1909 union tcp_log_stackspecific log; 1910 1911 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1912 log.u_bbr.flex1 = seq; 1913 log.u_bbr.flex2 = count; 1914 log.u_bbr.flex8 = mode; 1915 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1916 &bbr->rc_inp->inp_socket->so_rcv, 1917 &bbr->rc_inp->inp_socket->so_snd, 1918 BBR_LOG_LOWGAIN, 0, 1919 0, &log, false, &bbr->rc_tv); 1920 } 1921 } 1922 1923 static void 1924 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, 1925 uint8_t reason, uint32_t p_maxseg, int len) 1926 { 1927 if (tcp_bblogging_on(bbr->rc_tp)) { 1928 union tcp_log_stackspecific log; 1929 1930 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1931 log.u_bbr.flex1 = p_maxseg; 1932 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; 1933 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 1934 log.u_bbr.flex4 = reason; 1935 log.u_bbr.flex5 = bbr->rc_in_persist; 1936 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; 1937 log.u_bbr.flex7 = p_maxseg; 1938 log.u_bbr.flex8 = bbr->rc_in_persist; 1939 log.u_bbr.pkts_out = 0; 1940 log.u_bbr.applimited = len; 1941 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1942 &bbr->rc_inp->inp_socket->so_rcv, 1943 &bbr->rc_inp->inp_socket->so_snd, 1944 BBR_LOG_JUSTRET, 0, 1945 tlen, &log, false, &bbr->rc_tv); 1946 } 1947 } 1948 1949 static void 1950 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) 1951 { 1952 if (tcp_bblogging_on(bbr->rc_tp)) { 1953 union tcp_log_stackspecific log; 1954 1955 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 1956 log.u_bbr.flex1 = seq; 1957 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 1958 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; 1959 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 1960 &bbr->rc_inp->inp_socket->so_rcv, 1961 &bbr->rc_inp->inp_socket->so_snd, 1962 BBR_LOG_ENTREC, 0, 1963 0, &log, false, &bbr->rc_tv); 1964 } 1965 } 1966 1967 static void 1968 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) 1969 { 1970 if (tcp_bblogging_on(tp)) { 1971 union tcp_log_stackspecific log; 1972 1973 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 1974 log.u_bbr.flex1 = tso; 1975 log.u_bbr.flex2 = maxseg; 1976 log.u_bbr.flex3 = mtu; 1977 log.u_bbr.flex4 = csum_flags; 1978 TCP_LOG_EVENTP(tp, NULL, 1979 &bbr->rc_inp->inp_socket->so_rcv, 1980 &bbr->rc_inp->inp_socket->so_snd, 1981 BBR_LOG_MSGSIZE, 0, 1982 0, &log, false, &bbr->rc_tv); 1983 } 1984 } 1985 1986 static void 1987 bbr_log_flowend(struct tcp_bbr *bbr) 1988 { 1989 if (tcp_bblogging_on(bbr->rc_tp)) { 1990 union tcp_log_stackspecific log; 1991 struct sockbuf *r, *s; 1992 struct timeval tv; 1993 1994 if (bbr->rc_inp->inp_socket) { 1995 r = &bbr->rc_inp->inp_socket->so_rcv; 1996 s = &bbr->rc_inp->inp_socket->so_snd; 1997 } else { 1998 r = s = NULL; 1999 } 2000 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); 2001 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2002 r, s, 2003 TCP_LOG_FLOWEND, 0, 2004 0, &log, false, &tv); 2005 } 2006 } 2007 2008 static void 2009 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, 2010 uint32_t lost, uint32_t del) 2011 { 2012 if (tcp_bblogging_on(bbr->rc_tp)) { 2013 union tcp_log_stackspecific log; 2014 2015 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2016 log.u_bbr.flex1 = lost; 2017 log.u_bbr.flex2 = del; 2018 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; 2019 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; 2020 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2021 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2022 log.u_bbr.flex7 = line; 2023 log.u_bbr.flex8 = 0; 2024 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; 2025 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2026 &bbr->rc_inp->inp_socket->so_rcv, 2027 &bbr->rc_inp->inp_socket->so_snd, 2028 BBR_LOG_PKT_EPOCH, 0, 2029 0, &log, false, &bbr->rc_tv); 2030 } 2031 } 2032 2033 static void 2034 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) 2035 { 2036 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2037 union tcp_log_stackspecific log; 2038 2039 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2040 log.u_bbr.flex1 = bbr->r_ctl.rc_lost; 2041 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; 2042 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; 2043 log.u_bbr.flex7 = line; 2044 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2045 &bbr->rc_inp->inp_socket->so_rcv, 2046 &bbr->rc_inp->inp_socket->so_snd, 2047 BBR_LOG_TIME_EPOCH, 0, 2048 0, &log, false, &bbr->rc_tv); 2049 } 2050 } 2051 2052 static void 2053 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) 2054 { 2055 if (tcp_bblogging_on(bbr->rc_tp)) { 2056 union tcp_log_stackspecific log; 2057 2058 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2059 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2060 log.u_bbr.flex2 = new_tar; 2061 log.u_bbr.flex3 = line; 2062 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2063 log.u_bbr.flex5 = bbr_quanta; 2064 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; 2065 log.u_bbr.flex7 = bbr->rc_last_options; 2066 log.u_bbr.flex8 = meth; 2067 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2068 &bbr->rc_inp->inp_socket->so_rcv, 2069 &bbr->rc_inp->inp_socket->so_snd, 2070 BBR_LOG_STATE_TARGET, 0, 2071 0, &log, false, &bbr->rc_tv); 2072 } 2073 2074 } 2075 2076 static void 2077 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2078 { 2079 if (tcp_bblogging_on(bbr->rc_tp)) { 2080 union tcp_log_stackspecific log; 2081 2082 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2083 log.u_bbr.flex1 = line; 2084 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2085 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2086 if (bbr_state_is_pkt_epoch) 2087 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 2088 else 2089 log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); 2090 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; 2091 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2092 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); 2093 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; 2094 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; 2095 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2096 &bbr->rc_inp->inp_socket->so_rcv, 2097 &bbr->rc_inp->inp_socket->so_snd, 2098 BBR_LOG_STATE, 0, 2099 0, &log, false, &bbr->rc_tv); 2100 } 2101 } 2102 2103 static void 2104 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, 2105 uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) 2106 { 2107 if (tcp_bblogging_on(bbr->rc_tp)) { 2108 union tcp_log_stackspecific log; 2109 2110 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2111 log.u_bbr.flex1 = line; 2112 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2113 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; 2114 log.u_bbr.flex4 = applied; 2115 log.u_bbr.flex5 = rtt; 2116 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2117 log.u_bbr.flex7 = cond; 2118 log.u_bbr.flex8 = reas; 2119 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2120 &bbr->rc_inp->inp_socket->so_rcv, 2121 &bbr->rc_inp->inp_socket->so_snd, 2122 BBR_LOG_RTT_SHRINKS, 0, 2123 0, &log, false, &bbr->rc_tv); 2124 } 2125 } 2126 2127 static void 2128 bbr_log_type_exit_rec(struct tcp_bbr *bbr) 2129 { 2130 if (tcp_bblogging_on(bbr->rc_tp)) { 2131 union tcp_log_stackspecific log; 2132 2133 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2134 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; 2135 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; 2136 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2137 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2138 &bbr->rc_inp->inp_socket->so_rcv, 2139 &bbr->rc_inp->inp_socket->so_snd, 2140 BBR_LOG_EXITREC, 0, 2141 0, &log, false, &bbr->rc_tv); 2142 } 2143 } 2144 2145 static void 2146 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, 2147 uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) 2148 { 2149 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2150 union tcp_log_stackspecific log; 2151 2152 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2153 log.u_bbr.flex1 = line; 2154 log.u_bbr.flex2 = prev_acked; 2155 log.u_bbr.flex3 = bytes_this_ack; 2156 log.u_bbr.flex4 = chg; 2157 log.u_bbr.flex5 = th_ack; 2158 log.u_bbr.flex6 = target; 2159 log.u_bbr.flex8 = meth; 2160 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2161 &bbr->rc_inp->inp_socket->so_rcv, 2162 &bbr->rc_inp->inp_socket->so_snd, 2163 BBR_LOG_CWND, 0, 2164 0, &log, false, &bbr->rc_tv); 2165 } 2166 } 2167 2168 static void 2169 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) 2170 { 2171 /* 2172 * Log the rtt sample we are applying to the srtt algorithm in 2173 * useconds. 2174 */ 2175 if (tcp_bblogging_on(bbr->rc_tp)) { 2176 union tcp_log_stackspecific log; 2177 2178 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2179 log.u_bbr.flex1 = rtt; 2180 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; 2181 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; 2182 log.u_bbr.flex4 = bbr->rc_tp->ts_offset; 2183 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2184 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); 2185 log.u_bbr.flex6 = tsin; 2186 log.u_bbr.flex7 = 0; 2187 log.u_bbr.flex8 = bbr->rc_ack_was_delayed; 2188 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2189 &bbr->rc_inp->inp_socket->so_rcv, 2190 &bbr->rc_inp->inp_socket->so_snd, 2191 TCP_LOG_RTT, 0, 2192 0, &log, false, &bbr->rc_tv); 2193 } 2194 } 2195 2196 static void 2197 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) 2198 { 2199 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2200 union tcp_log_stackspecific log; 2201 2202 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2203 log.u_bbr.flex1 = time_in; 2204 log.u_bbr.flex2 = line; 2205 log.u_bbr.flex8 = enter_exit; 2206 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2207 &bbr->rc_inp->inp_socket->so_rcv, 2208 &bbr->rc_inp->inp_socket->so_snd, 2209 BBR_LOG_PERSIST, 0, 2210 0, &log, false, &bbr->rc_tv); 2211 } 2212 } 2213 static void 2214 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) 2215 { 2216 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2217 union tcp_log_stackspecific log; 2218 2219 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2220 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; 2221 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; 2222 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; 2223 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; 2224 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2225 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2226 &bbr->rc_inp->inp_socket->so_rcv, 2227 &bbr->rc_inp->inp_socket->so_snd, 2228 BBR_LOG_ACKCLEAR, 0, 2229 0, &log, false, &bbr->rc_tv); 2230 } 2231 } 2232 2233 static void 2234 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, 2235 uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) 2236 { 2237 if (tcp_bblogging_on(bbr->rc_tp)) { 2238 union tcp_log_stackspecific log; 2239 struct timeval tv; 2240 2241 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2242 log.u_bbr.flex1 = nsegs; 2243 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; 2244 if (m) { 2245 struct timespec ts; 2246 2247 log.u_bbr.flex3 = m->m_flags; 2248 if (m->m_flags & M_TSTMP) { 2249 mbuf_tstmp2timespec(m, &ts); 2250 tv.tv_sec = ts.tv_sec; 2251 tv.tv_usec = ts.tv_nsec / 1000; 2252 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); 2253 } else { 2254 log.u_bbr.lt_epoch = 0; 2255 } 2256 if (m->m_flags & M_TSTMP_LRO) { 2257 mbuf_tstmp2timeval(m, &tv); 2258 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); 2259 } else { 2260 /* No arrival timestamp */ 2261 log.u_bbr.flex5 = 0; 2262 } 2263 2264 log.u_bbr.pkts_out = tcp_get_usecs(&tv); 2265 } else { 2266 log.u_bbr.flex3 = 0; 2267 log.u_bbr.flex5 = 0; 2268 log.u_bbr.flex6 = 0; 2269 log.u_bbr.pkts_out = 0; 2270 } 2271 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; 2272 log.u_bbr.flex7 = bbr->r_wanted_output; 2273 log.u_bbr.flex8 = bbr->rc_in_persist; 2274 TCP_LOG_EVENTP(bbr->rc_tp, th, 2275 &bbr->rc_inp->inp_socket->so_rcv, 2276 &bbr->rc_inp->inp_socket->so_snd, 2277 TCP_LOG_IN, 0, 2278 tlen, &log, true, &bbr->rc_tv); 2279 } 2280 } 2281 2282 static void 2283 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) 2284 { 2285 if (tcp_bblogging_on(bbr->rc_tp)) { 2286 union tcp_log_stackspecific log; 2287 2288 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2289 log.u_bbr.flex1 = did_out; 2290 log.u_bbr.flex2 = nxt_pkt; 2291 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; 2292 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2293 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; 2294 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; 2295 log.u_bbr.flex7 = bbr->r_wanted_output; 2296 log.u_bbr.flex8 = bbr->rc_in_persist; 2297 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; 2298 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2299 &bbr->rc_inp->inp_socket->so_rcv, 2300 &bbr->rc_inp->inp_socket->so_snd, 2301 BBR_LOG_DOSEG_DONE, 0, 2302 0, &log, true, &bbr->rc_tv); 2303 } 2304 } 2305 2306 static void 2307 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, 2308 int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) 2309 { 2310 if (tcp_bblogging_on(bbr->rc_tp)) { 2311 union tcp_log_stackspecific log; 2312 2313 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2314 log.u_bbr.flex1 = line; 2315 log.u_bbr.flex2 = o_len; 2316 log.u_bbr.flex3 = segcnt; 2317 log.u_bbr.flex4 = segsiz; 2318 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2319 &bbr->rc_inp->inp_socket->so_rcv, 2320 &bbr->rc_inp->inp_socket->so_snd, 2321 BBR_LOG_ENOBUF_JMP, ENOBUFS, 2322 len, &log, true, &bbr->rc_tv); 2323 } 2324 } 2325 2326 static void 2327 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) 2328 { 2329 if (tcp_bblogging_on(bbr->rc_tp)) { 2330 union tcp_log_stackspecific log; 2331 2332 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2333 log.u_bbr.flex1 = timers; 2334 log.u_bbr.flex2 = ret; 2335 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; 2336 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; 2337 log.u_bbr.flex5 = cts; 2338 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; 2339 log.u_bbr.flex8 = hpts_calling; 2340 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2341 &bbr->rc_inp->inp_socket->so_rcv, 2342 &bbr->rc_inp->inp_socket->so_snd, 2343 BBR_LOG_TO_PROCESS, 0, 2344 0, &log, false, &bbr->rc_tv); 2345 } 2346 } 2347 2348 static void 2349 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) 2350 { 2351 if (tcp_bblogging_on(bbr->rc_tp)) { 2352 union tcp_log_stackspecific log; 2353 uint64_t ar; 2354 2355 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2356 log.u_bbr.flex1 = bbr->bbr_timer_src; 2357 log.u_bbr.flex2 = 0; 2358 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2359 ar = (uint64_t)(bbr->r_ctl.rc_resend); 2360 ar >>= 32; 2361 ar &= 0x00000000ffffffff; 2362 log.u_bbr.flex4 = (uint32_t)ar; 2363 ar = (uint64_t)bbr->r_ctl.rc_resend; 2364 ar &= 0x00000000ffffffff; 2365 log.u_bbr.flex5 = (uint32_t)ar; 2366 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2367 log.u_bbr.flex8 = to_num; 2368 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2369 &bbr->rc_inp->inp_socket->so_rcv, 2370 &bbr->rc_inp->inp_socket->so_snd, 2371 BBR_LOG_RTO, 0, 2372 0, &log, false, &bbr->rc_tv); 2373 } 2374 } 2375 2376 static void 2377 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) 2378 { 2379 if (tcp_bblogging_on(bbr->rc_tp)) { 2380 union tcp_log_stackspecific log; 2381 2382 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2383 log.u_bbr.flex1 = flex1; 2384 log.u_bbr.flex2 = flex2; 2385 log.u_bbr.flex3 = flex3; 2386 log.u_bbr.flex4 = 0; 2387 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2388 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; 2389 log.u_bbr.flex8 = reason; 2390 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; 2391 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2392 &bbr->rc_inp->inp_socket->so_rcv, 2393 &bbr->rc_inp->inp_socket->so_snd, 2394 BBR_LOG_REDUCE, 0, 2395 0, &log, false, &bbr->rc_tv); 2396 } 2397 } 2398 2399 static void 2400 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) 2401 { 2402 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2403 union tcp_log_stackspecific log; 2404 2405 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2406 log.u_bbr.flex1 = diag->p_nxt_slot; 2407 log.u_bbr.flex2 = diag->p_cur_slot; 2408 log.u_bbr.flex3 = diag->slot_req; 2409 log.u_bbr.flex4 = diag->inp_hptsslot; 2410 log.u_bbr.flex5 = diag->slot_remaining; 2411 log.u_bbr.flex6 = diag->need_new_to; 2412 log.u_bbr.flex7 = diag->p_hpts_active; 2413 log.u_bbr.flex8 = diag->p_on_min_sleep; 2414 /* Hijack other fields as needed */ 2415 log.u_bbr.epoch = diag->have_slept; 2416 log.u_bbr.lt_epoch = diag->yet_to_sleep; 2417 log.u_bbr.pkts_out = diag->co_ret; 2418 log.u_bbr.applimited = diag->hpts_sleep_time; 2419 log.u_bbr.delivered = diag->p_prev_slot; 2420 log.u_bbr.inflight = diag->p_runningslot; 2421 log.u_bbr.bw_inuse = diag->wheel_slot; 2422 log.u_bbr.rttProp = diag->wheel_cts; 2423 log.u_bbr.delRate = diag->maxslots; 2424 log.u_bbr.cur_del_rate = diag->p_curtick; 2425 log.u_bbr.cur_del_rate <<= 32; 2426 log.u_bbr.cur_del_rate |= diag->p_lasttick; 2427 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2428 &bbr->rc_inp->inp_socket->so_rcv, 2429 &bbr->rc_inp->inp_socket->so_snd, 2430 BBR_LOG_HPTSDIAG, 0, 2431 0, &log, false, &bbr->rc_tv); 2432 } 2433 } 2434 2435 static void 2436 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, 2437 uint32_t thresh, uint32_t to) 2438 { 2439 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2440 union tcp_log_stackspecific log; 2441 2442 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2443 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; 2444 log.u_bbr.flex2 = time_since_sent; 2445 log.u_bbr.flex3 = srtt; 2446 log.u_bbr.flex4 = thresh; 2447 log.u_bbr.flex5 = to; 2448 log.u_bbr.flex6 = bbr->rc_tp->t_srtt; 2449 log.u_bbr.flex8 = mode; 2450 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2451 &bbr->rc_inp->inp_socket->so_rcv, 2452 &bbr->rc_inp->inp_socket->so_snd, 2453 BBR_LOG_TIMERPREP, 0, 2454 0, &log, false, &bbr->rc_tv); 2455 } 2456 } 2457 2458 static void 2459 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, 2460 uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) 2461 { 2462 if (tcp_bblogging_on(bbr->rc_tp)) { 2463 union tcp_log_stackspecific log; 2464 2465 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2466 log.u_bbr.flex1 = usecs; 2467 log.u_bbr.flex2 = len; 2468 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); 2469 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); 2470 if (override) 2471 log.u_bbr.flex5 = (1 << 2); 2472 else 2473 log.u_bbr.flex5 = 0; 2474 log.u_bbr.flex6 = override; 2475 log.u_bbr.flex7 = gain; 2476 log.u_bbr.flex8 = mod; 2477 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2478 &bbr->rc_inp->inp_socket->so_rcv, 2479 &bbr->rc_inp->inp_socket->so_snd, 2480 BBR_LOG_HPTSI_CALC, 0, 2481 len, &log, false, &bbr->rc_tv); 2482 } 2483 } 2484 2485 static void 2486 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) 2487 { 2488 if (tcp_bblogging_on(bbr->rc_tp)) { 2489 union tcp_log_stackspecific log; 2490 2491 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2492 2493 log.u_bbr.flex1 = bbr->bbr_timer_src; 2494 log.u_bbr.flex2 = to; 2495 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2496 log.u_bbr.flex4 = slot; 2497 log.u_bbr.flex5 = bbr->rc_tp->t_hpts_slot; 2498 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2499 log.u_bbr.pkts_out = bbr->rc_tp->t_flags2; 2500 log.u_bbr.flex8 = which; 2501 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2502 &bbr->rc_inp->inp_socket->so_rcv, 2503 &bbr->rc_inp->inp_socket->so_snd, 2504 BBR_LOG_TIMERSTAR, 0, 2505 0, &log, false, &bbr->rc_tv); 2506 } 2507 } 2508 2509 static void 2510 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) 2511 { 2512 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2513 union tcp_log_stackspecific log; 2514 2515 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2516 log.u_bbr.flex1 = thresh; 2517 log.u_bbr.flex2 = lro; 2518 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; 2519 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; 2520 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2521 log.u_bbr.flex6 = srtt; 2522 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; 2523 log.u_bbr.flex8 = frm; 2524 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2525 &bbr->rc_inp->inp_socket->so_rcv, 2526 &bbr->rc_inp->inp_socket->so_snd, 2527 BBR_LOG_THRESH_CALC, 0, 2528 0, &log, false, &bbr->rc_tv); 2529 } 2530 } 2531 2532 static void 2533 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) 2534 { 2535 if (tcp_bblogging_on(bbr->rc_tp)) { 2536 union tcp_log_stackspecific log; 2537 2538 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2539 log.u_bbr.flex1 = line; 2540 log.u_bbr.flex2 = bbr->bbr_timer_src; 2541 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; 2542 log.u_bbr.flex4 = bbr->rc_in_persist; 2543 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; 2544 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 2545 log.u_bbr.flex8 = hpts_removed; 2546 log.u_bbr.pkts_out = bbr->rc_pacer_started; 2547 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2548 &bbr->rc_inp->inp_socket->so_rcv, 2549 &bbr->rc_inp->inp_socket->so_snd, 2550 BBR_LOG_TIMERCANC, 0, 2551 0, &log, false, &bbr->rc_tv); 2552 } 2553 } 2554 2555 static void 2556 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) 2557 { 2558 if (tcp_bblogging_on(bbr->rc_tp)) { 2559 union tcp_log_stackspecific log; 2560 2561 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2562 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; 2563 log.u_bbr.flex2 = (peer_delta >> 32); 2564 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); 2565 log.u_bbr.flex4 = (delta >> 32); 2566 log.u_bbr.flex5 = (delta & 0x00000000ffffffff); 2567 log.u_bbr.flex7 = bbr->rc_ts_clock_set; 2568 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; 2569 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2570 &bbr->rc_inp->inp_socket->so_rcv, 2571 &bbr->rc_inp->inp_socket->so_snd, 2572 BBR_LOG_TSTMP_VAL, 0, 2573 0, &log, false, &bbr->rc_tv); 2574 } 2575 } 2576 2577 static void 2578 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) 2579 { 2580 if (tcp_bblogging_on(bbr->rc_tp)) { 2581 union tcp_log_stackspecific log; 2582 2583 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2584 log.u_bbr.flex1 = tsosz; 2585 log.u_bbr.flex2 = tls; 2586 log.u_bbr.flex3 = tcp_min_hptsi_time; 2587 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; 2588 log.u_bbr.flex5 = old_val; 2589 log.u_bbr.flex6 = maxseg; 2590 log.u_bbr.flex7 = bbr->rc_no_pacing; 2591 log.u_bbr.flex7 <<= 1; 2592 log.u_bbr.flex7 |= bbr->rc_past_init_win; 2593 if (hdwr) 2594 log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; 2595 else 2596 log.u_bbr.flex8 = bbr->rc_use_google; 2597 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2598 &bbr->rc_inp->inp_socket->so_rcv, 2599 &bbr->rc_inp->inp_socket->so_snd, 2600 BBR_LOG_BBRTSO, 0, 2601 0, &log, false, &bbr->rc_tv); 2602 } 2603 } 2604 2605 static void 2606 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, 2607 uint32_t flags, uint32_t line) 2608 { 2609 if (tcp_bblogging_on(bbr->rc_tp)) { 2610 union tcp_log_stackspecific log; 2611 2612 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2613 log.u_bbr.flex1 = line; 2614 log.u_bbr.flex2 = rsm->r_start; 2615 log.u_bbr.flex3 = rsm->r_end; 2616 log.u_bbr.flex4 = rsm->r_delivered; 2617 log.u_bbr.flex5 = rsm->r_rtr_cnt; 2618 log.u_bbr.flex6 = rsm->r_dupack; 2619 log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; 2620 log.u_bbr.flex8 = rsm->r_flags; 2621 /* Hijack the pkts_out fids */ 2622 log.u_bbr.applimited = flags; 2623 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2624 &bbr->rc_inp->inp_socket->so_rcv, 2625 &bbr->rc_inp->inp_socket->so_snd, 2626 BBR_RSM_CLEARED, 0, 2627 0, &log, false, &bbr->rc_tv); 2628 } 2629 } 2630 2631 static void 2632 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, 2633 uint32_t flex3, uint32_t flex2, uint32_t flex5, 2634 uint32_t flex6, uint32_t pkts_out, int flex7, 2635 uint32_t flex4, uint32_t flex1) 2636 { 2637 2638 if (tcp_bblogging_on(bbr->rc_tp)) { 2639 union tcp_log_stackspecific log; 2640 2641 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2642 log.u_bbr.flex1 = flex1; 2643 log.u_bbr.flex2 = flex2; 2644 log.u_bbr.flex3 = flex3; 2645 log.u_bbr.flex4 = flex4; 2646 log.u_bbr.flex5 = flex5; 2647 log.u_bbr.flex6 = flex6; 2648 log.u_bbr.flex7 = flex7; 2649 /* Hijack the pkts_out fids */ 2650 log.u_bbr.pkts_out = pkts_out; 2651 log.u_bbr.flex8 = flex8; 2652 if (bbr->rc_ack_was_delayed) 2653 log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; 2654 else 2655 log.u_bbr.epoch = 0; 2656 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2657 &bbr->rc_inp->inp_socket->so_rcv, 2658 &bbr->rc_inp->inp_socket->so_snd, 2659 BBR_LOG_BBRUPD, 0, 2660 flex2, &log, false, &bbr->rc_tv); 2661 } 2662 } 2663 2664 static void 2665 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, 2666 uint32_t newbw, uint32_t obw, uint32_t diff, 2667 uint32_t tim) 2668 { 2669 if (/*bbr_verbose_logging && */tcp_bblogging_on(bbr->rc_tp)) { 2670 union tcp_log_stackspecific log; 2671 2672 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2673 log.u_bbr.flex1 = reason; 2674 log.u_bbr.flex2 = newbw; 2675 log.u_bbr.flex3 = obw; 2676 log.u_bbr.flex4 = diff; 2677 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; 2678 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; 2679 log.u_bbr.flex7 = bbr->rc_lt_is_sampling; 2680 log.u_bbr.pkts_out = tim; 2681 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; 2682 if (bbr->rc_lt_use_bw == 0) 2683 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 2684 else 2685 log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 2686 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2687 &bbr->rc_inp->inp_socket->so_rcv, 2688 &bbr->rc_inp->inp_socket->so_snd, 2689 BBR_LOG_BWSAMP, 0, 2690 0, &log, false, &bbr->rc_tv); 2691 } 2692 } 2693 2694 static inline void 2695 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) 2696 { 2697 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2698 union tcp_log_stackspecific log; 2699 2700 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2701 log.u_bbr.flex1 = line; 2702 log.u_bbr.flex2 = tick; 2703 log.u_bbr.flex3 = tp->t_maxunacktime; 2704 log.u_bbr.flex4 = tp->t_acktime; 2705 log.u_bbr.flex8 = event; 2706 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2707 &bbr->rc_inp->inp_socket->so_rcv, 2708 &bbr->rc_inp->inp_socket->so_snd, 2709 BBR_LOG_PROGRESS, 0, 2710 0, &log, false, &bbr->rc_tv); 2711 } 2712 } 2713 2714 static void 2715 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, 2716 uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, 2717 int error) 2718 { 2719 if (tcp_bblogging_on(bbr->rc_tp)) { 2720 union tcp_log_stackspecific log; 2721 2722 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2723 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); 2724 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); 2725 log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); 2726 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); 2727 log.u_bbr.bw_inuse = rate; 2728 log.u_bbr.flex5 = line; 2729 log.u_bbr.flex6 = error; 2730 log.u_bbr.flex8 = bbr->skip_gain; 2731 log.u_bbr.flex8 <<= 1; 2732 log.u_bbr.flex8 |= bbr->gain_is_limited; 2733 log.u_bbr.flex8 <<= 1; 2734 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; 2735 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 2736 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2737 &bbr->rc_inp->inp_socket->so_rcv, 2738 &bbr->rc_inp->inp_socket->so_snd, 2739 BBR_LOG_HDWR_PACE, 0, 2740 0, &log, false, &bbr->rc_tv); 2741 } 2742 } 2743 2744 static void 2745 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) 2746 { 2747 if (tcp_bblogging_on(bbr->rc_tp)) { 2748 union tcp_log_stackspecific log; 2749 2750 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2751 log.u_bbr.flex1 = slot; 2752 log.u_bbr.flex2 = del_by; 2753 log.u_bbr.flex3 = prev_delay; 2754 log.u_bbr.flex4 = line; 2755 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; 2756 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; 2757 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); 2758 log.u_bbr.flex8 = bbr->rc_in_persist; 2759 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2760 &bbr->rc_inp->inp_socket->so_rcv, 2761 &bbr->rc_inp->inp_socket->so_snd, 2762 BBR_LOG_BBRSND, 0, 2763 len, &log, false, &bbr->rc_tv); 2764 } 2765 } 2766 2767 static void 2768 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) 2769 { 2770 if (tcp_bblogging_on(bbr->rc_tp)) { 2771 union tcp_log_stackspecific log; 2772 2773 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2774 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; 2775 log.u_bbr.flex2 = 0; 2776 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; 2777 log.u_bbr.flex4 = end; 2778 log.u_bbr.flex5 = seq; 2779 log.u_bbr.flex6 = t; 2780 log.u_bbr.flex7 = match; 2781 log.u_bbr.flex8 = flags; 2782 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2783 &bbr->rc_inp->inp_socket->so_rcv, 2784 &bbr->rc_inp->inp_socket->so_snd, 2785 BBR_LOG_BBRRTT, 0, 2786 0, &log, false, &bbr->rc_tv); 2787 } 2788 } 2789 2790 static void 2791 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) 2792 { 2793 if (tcp_bblogging_on(bbr->rc_tp)) { 2794 union tcp_log_stackspecific log; 2795 2796 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 2797 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; 2798 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 2799 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; 2800 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; 2801 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; 2802 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; 2803 log.u_bbr.flex7 = 0; 2804 log.u_bbr.flex8 = entry_method; 2805 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2806 &bbr->rc_inp->inp_socket->so_rcv, 2807 &bbr->rc_inp->inp_socket->so_snd, 2808 BBR_LOG_EXIT_GAIN, 0, 2809 0, &log, false, &bbr->rc_tv); 2810 } 2811 } 2812 2813 static void 2814 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) 2815 { 2816 if (bbr_verbose_logging && tcp_bblogging_on(bbr->rc_tp)) { 2817 union tcp_log_stackspecific log; 2818 2819 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); 2820 /* R-HU */ 2821 log.u_bbr.flex1 = 0; 2822 log.u_bbr.flex2 = 0; 2823 log.u_bbr.flex3 = 0; 2824 log.u_bbr.flex4 = 0; 2825 log.u_bbr.flex7 = 0; 2826 log.u_bbr.flex8 = settings_desired; 2827 2828 TCP_LOG_EVENTP(bbr->rc_tp, NULL, 2829 &bbr->rc_inp->inp_socket->so_rcv, 2830 &bbr->rc_inp->inp_socket->so_snd, 2831 BBR_LOG_SETTINGS_CHG, 0, 2832 0, &log, false, &bbr->rc_tv); 2833 } 2834 } 2835 2836 /* 2837 * Returns the bw from the our filter. 2838 */ 2839 static inline uint64_t 2840 bbr_get_full_bw(struct tcp_bbr *bbr) 2841 { 2842 uint64_t bw; 2843 2844 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2845 2846 return (bw); 2847 } 2848 2849 static inline void 2850 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2851 { 2852 uint64_t calclr; 2853 uint32_t lost, del; 2854 2855 if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) 2856 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; 2857 else 2858 lost = 0; 2859 del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; 2860 if (lost == 0) { 2861 calclr = 0; 2862 } else if (del) { 2863 calclr = lost; 2864 calclr *= (uint64_t)1000; 2865 calclr /= (uint64_t)del; 2866 } else { 2867 /* Nothing delivered? 100.0% loss */ 2868 calclr = 1000; 2869 } 2870 bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; 2871 if (IN_RECOVERY(bbr->rc_tp->t_flags)) 2872 bbr->r_ctl.recovery_lr += (uint32_t)calclr; 2873 bbr->r_ctl.rc_pkt_epoch++; 2874 if (bbr->rc_no_pacing && 2875 (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { 2876 bbr->rc_no_pacing = 0; 2877 tcp_bbr_tso_size_check(bbr, cts); 2878 } 2879 bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); 2880 bbr->r_ctl.rc_pkt_epoch_time = cts; 2881 /* What was our loss rate */ 2882 bbr_log_pkt_epoch(bbr, cts, line, lost, del); 2883 bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; 2884 bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; 2885 } 2886 2887 static inline void 2888 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 2889 { 2890 uint32_t epoch_time; 2891 2892 /* Tick the RTT clock */ 2893 bbr->r_ctl.rc_rtt_epoch++; 2894 epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; 2895 bbr_log_time_epoch(bbr, cts, line, epoch_time); 2896 bbr->r_ctl.rc_rcv_epoch_start = cts; 2897 } 2898 2899 static inline void 2900 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) 2901 { 2902 if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { 2903 bbr->rc_is_pkt_epoch_now = 1; 2904 } 2905 } 2906 2907 /* 2908 * Returns the bw from either the b/w filter 2909 * or from the lt_bw (if the connection is being 2910 * policed). 2911 */ 2912 static inline uint64_t 2913 __bbr_get_bw(struct tcp_bbr *bbr) 2914 { 2915 uint64_t bw, min_bw; 2916 uint64_t rtt; 2917 int gm_measure_cnt = 1; 2918 2919 /* 2920 * For startup we make, like google, a 2921 * minimum b/w. This is generated from the 2922 * IW and the rttProp. We do fall back to srtt 2923 * if for some reason (initial handshake) we don't 2924 * have a rttProp. We, in the worst case, fall back 2925 * to the configured min_bw (rc_initial_hptsi_bw). 2926 */ 2927 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 2928 /* Attempt first to use rttProp */ 2929 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2930 if (rtt && (rtt < 0xffffffff)) { 2931 measure: 2932 min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2933 ((uint64_t)1000000); 2934 min_bw /= rtt; 2935 if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2936 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2937 } 2938 2939 } else if (bbr->rc_tp->t_srtt != 0) { 2940 /* No rttProp, use srtt? */ 2941 rtt = bbr_get_rtt(bbr, BBR_SRTT); 2942 goto measure; 2943 } else { 2944 min_bw = bbr->r_ctl.rc_initial_hptsi_bw; 2945 } 2946 } else 2947 min_bw = 0; 2948 2949 if ((bbr->rc_past_init_win == 0) && 2950 (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) 2951 bbr->rc_past_init_win = 1; 2952 if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) 2953 gm_measure_cnt = 0; 2954 if (gm_measure_cnt && 2955 ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || 2956 (bbr->rc_past_init_win == 0))) { 2957 /* For google we use our guess rate until we get 1 measurement */ 2958 2959 use_initial_window: 2960 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); 2961 if (rtt && (rtt < 0xffffffff)) { 2962 /* 2963 * We have an RTT measurement. Use that in 2964 * combination with our initial window to calculate 2965 * a b/w. 2966 */ 2967 bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * 2968 ((uint64_t)1000000); 2969 bw /= rtt; 2970 if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { 2971 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2972 } 2973 } else { 2974 /* Drop back to the 40 and punt to a default */ 2975 bw = bbr->r_ctl.rc_initial_hptsi_bw; 2976 } 2977 if (bw < 1) 2978 /* Probably should panic */ 2979 bw = 1; 2980 if (bw > min_bw) 2981 return (bw); 2982 else 2983 return (min_bw); 2984 } 2985 if (bbr->rc_lt_use_bw) 2986 bw = bbr->r_ctl.rc_lt_bw; 2987 else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) 2988 bw = bbr->r_ctl.red_bw; 2989 else 2990 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 2991 if (bw == 0) { 2992 /* We should not be at 0, go to the initial window then */ 2993 goto use_initial_window; 2994 } 2995 if (bw < 1) 2996 /* Probably should panic */ 2997 bw = 1; 2998 if (bw < min_bw) 2999 bw = min_bw; 3000 return (bw); 3001 } 3002 3003 static inline uint64_t 3004 bbr_get_bw(struct tcp_bbr *bbr) 3005 { 3006 uint64_t bw; 3007 3008 bw = __bbr_get_bw(bbr); 3009 return (bw); 3010 } 3011 3012 static inline void 3013 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) 3014 { 3015 bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; 3016 bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; 3017 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3018 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3019 } 3020 3021 static inline void 3022 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) 3023 { 3024 bbr->rc_lt_is_sampling = 0; 3025 bbr->rc_lt_use_bw = 0; 3026 bbr->r_ctl.rc_lt_bw = 0; 3027 bbr_reset_lt_bw_interval(bbr, cts); 3028 } 3029 3030 static inline void 3031 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) 3032 { 3033 uint64_t diff; 3034 3035 /* Do we have a previous sample? */ 3036 if (bbr->r_ctl.rc_lt_bw) { 3037 /* Get the diff in bytes per second */ 3038 if (bbr->r_ctl.rc_lt_bw > bw) 3039 diff = bbr->r_ctl.rc_lt_bw - bw; 3040 else 3041 diff = bw - bbr->r_ctl.rc_lt_bw; 3042 if ((diff <= bbr_lt_bw_diff) || 3043 (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { 3044 /* Consider us policed */ 3045 uint32_t saved_bw; 3046 3047 saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; 3048 bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ 3049 bbr->rc_lt_use_bw = 1; 3050 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 3051 /* 3052 * Use pkt based epoch for measuring length of 3053 * policer up 3054 */ 3055 bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; 3056 /* 3057 * reason 4 is we need to start consider being 3058 * policed 3059 */ 3060 bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); 3061 return; 3062 } 3063 } 3064 bbr->r_ctl.rc_lt_bw = bw; 3065 bbr_reset_lt_bw_interval(bbr, cts); 3066 bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); 3067 } 3068 3069 static void 3070 bbr_randomize_extra_state_time(struct tcp_bbr *bbr) 3071 { 3072 uint32_t ran, deduct; 3073 3074 ran = arc4random_uniform(bbr_rand_ot); 3075 if (ran) { 3076 deduct = bbr->r_ctl.rc_level_state_extra / ran; 3077 bbr->r_ctl.rc_level_state_extra -= deduct; 3078 } 3079 } 3080 /* 3081 * Return randomly the starting state 3082 * to use in probebw. 3083 */ 3084 static uint8_t 3085 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) 3086 { 3087 uint32_t ran; 3088 uint8_t ret_val; 3089 3090 /* Initialize the offset to 0 */ 3091 bbr->r_ctl.rc_exta_time_gd = 0; 3092 bbr->rc_hit_state_1 = 0; 3093 bbr->r_ctl.rc_level_state_extra = 0; 3094 ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); 3095 /* 3096 * The math works funny here :) the return value is used to set the 3097 * substate and then the state change is called which increments by 3098 * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when 3099 * we fully enter the state. Note that the (8 - 1 - ran) assures that 3100 * we return 1 - 7, so we dont return 0 and end up starting in 3101 * state 1 (DRAIN). 3102 */ 3103 ret_val = BBR_SUBSTATE_COUNT - 1 - ran; 3104 /* Set an epoch */ 3105 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) 3106 bbr_set_epoch(bbr, cts, __LINE__); 3107 3108 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 3109 return (ret_val); 3110 } 3111 3112 static void 3113 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) 3114 { 3115 uint32_t diff, d_time; 3116 uint64_t del_time, bw, lost, delivered; 3117 3118 if (bbr->r_use_policer == 0) 3119 return; 3120 if (bbr->rc_lt_use_bw) { 3121 /* We are using lt bw do we stop yet? */ 3122 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; 3123 if (diff > bbr_lt_bw_max_rtts) { 3124 /* Reset it all */ 3125 reset_all: 3126 bbr_reset_lt_bw_sampling(bbr, cts); 3127 if (bbr->rc_filled_pipe) { 3128 bbr_set_epoch(bbr, cts, __LINE__); 3129 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 3130 bbr_substate_change(bbr, cts, __LINE__, 0); 3131 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 3132 bbr_log_type_statechange(bbr, cts, __LINE__); 3133 } else { 3134 /* 3135 * This should not happen really 3136 * unless we remove the startup/drain 3137 * restrictions above. 3138 */ 3139 bbr->rc_bbr_state = BBR_STATE_STARTUP; 3140 bbr_set_epoch(bbr, cts, __LINE__); 3141 bbr->r_ctl.rc_bbr_state_time = cts; 3142 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 3143 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 3144 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 3145 bbr_set_state_target(bbr, __LINE__); 3146 bbr_log_type_statechange(bbr, cts, __LINE__); 3147 } 3148 /* reason 0 is to stop using lt-bw */ 3149 bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); 3150 return; 3151 } 3152 if (bbr_lt_intvl_fp == 0) { 3153 /* Not doing false-positive detection */ 3154 return; 3155 } 3156 /* False positive detection */ 3157 if (diff == bbr_lt_intvl_fp) { 3158 /* At bbr_lt_intvl_fp we record the lost */ 3159 bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; 3160 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 3161 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { 3162 /* Now is our loss rate still high? */ 3163 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3164 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3165 if ((delivered == 0) || 3166 (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { 3167 /* No still below our threshold */ 3168 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); 3169 } else { 3170 /* Yikes its still high, it must be a false positive */ 3171 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); 3172 goto reset_all; 3173 } 3174 } 3175 return; 3176 } 3177 /* 3178 * Wait for the first loss before sampling, to let the policer 3179 * exhaust its tokens and estimate the steady-state rate allowed by 3180 * the policer. Starting samples earlier includes bursts that 3181 * over-estimate the bw. 3182 */ 3183 if (bbr->rc_lt_is_sampling == 0) { 3184 /* reason 1 is to begin doing the sampling */ 3185 if (loss_detected == 0) 3186 return; 3187 bbr_reset_lt_bw_interval(bbr, cts); 3188 bbr->rc_lt_is_sampling = 1; 3189 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); 3190 return; 3191 } 3192 /* Now how long were we delivering long term last> */ 3193 if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) 3194 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; 3195 else 3196 d_time = 0; 3197 3198 /* To avoid underestimates, reset sampling if we run out of data. */ 3199 if (bbr->r_ctl.r_app_limited_until) { 3200 /* Can not measure in app-limited state */ 3201 bbr_reset_lt_bw_sampling(bbr, cts); 3202 /* reason 2 is to reset sampling due to app limits */ 3203 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); 3204 return; 3205 } 3206 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; 3207 if (diff < bbr_lt_intvl_min_rtts) { 3208 /* 3209 * need more samples (we don't 3210 * start on a round like linux so 3211 * we need 1 more). 3212 */ 3213 /* 6 is not_enough time or no-loss */ 3214 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3215 return; 3216 } 3217 if (diff > (4 * bbr_lt_intvl_min_rtts)) { 3218 /* 3219 * For now if we wait too long, reset all sampling. We need 3220 * to do some research here, its possible that we should 3221 * base this on how much loss as occurred.. something like 3222 * if its under 10% (or some thresh) reset all otherwise 3223 * don't. Thats for phase II I guess. 3224 */ 3225 bbr_reset_lt_bw_sampling(bbr, cts); 3226 /* reason 3 is to reset sampling due too long of sampling */ 3227 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3228 return; 3229 } 3230 /* 3231 * End sampling interval when a packet is lost, so we estimate the 3232 * policer tokens were exhausted. Stopping the sampling before the 3233 * tokens are exhausted under-estimates the policed rate. 3234 */ 3235 if (loss_detected == 0) { 3236 /* 6 is not_enough time or no-loss */ 3237 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3238 return; 3239 } 3240 /* Calculate packets lost and delivered in sampling interval. */ 3241 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; 3242 delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; 3243 if ((delivered == 0) || 3244 (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { 3245 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); 3246 return; 3247 } 3248 if (d_time < 1000) { 3249 /* Not enough time. wait */ 3250 /* 6 is not_enough time or no-loss */ 3251 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); 3252 return; 3253 } 3254 if (d_time >= (0xffffffff / USECS_IN_MSEC)) { 3255 /* Too long */ 3256 bbr_reset_lt_bw_sampling(bbr, cts); 3257 /* reason 3 is to reset sampling due too long of sampling */ 3258 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); 3259 return; 3260 } 3261 del_time = d_time; 3262 bw = delivered; 3263 bw *= (uint64_t)USECS_IN_SECOND; 3264 bw /= del_time; 3265 bbr_lt_bw_samp_done(bbr, bw, cts, d_time); 3266 } 3267 3268 /* 3269 * Allocate a sendmap from our zone. 3270 */ 3271 static struct bbr_sendmap * 3272 bbr_alloc(struct tcp_bbr *bbr) 3273 { 3274 struct bbr_sendmap *rsm; 3275 3276 BBR_STAT_INC(bbr_to_alloc); 3277 rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); 3278 if (rsm) { 3279 bbr->r_ctl.rc_num_maps_alloced++; 3280 return (rsm); 3281 } 3282 if (bbr->r_ctl.rc_free_cnt) { 3283 BBR_STAT_INC(bbr_to_alloc_emerg); 3284 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 3285 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 3286 bbr->r_ctl.rc_free_cnt--; 3287 return (rsm); 3288 } 3289 BBR_STAT_INC(bbr_to_alloc_failed); 3290 return (NULL); 3291 } 3292 3293 static struct bbr_sendmap * 3294 bbr_alloc_full_limit(struct tcp_bbr *bbr) 3295 { 3296 if ((V_tcp_map_entries_limit > 0) && 3297 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 3298 BBR_STAT_INC(bbr_alloc_limited); 3299 if (!bbr->alloc_limit_reported) { 3300 bbr->alloc_limit_reported = 1; 3301 BBR_STAT_INC(bbr_alloc_limited_conns); 3302 } 3303 return (NULL); 3304 } 3305 return (bbr_alloc(bbr)); 3306 } 3307 3308 /* wrapper to allocate a sendmap entry, subject to a specific limit */ 3309 static struct bbr_sendmap * 3310 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) 3311 { 3312 struct bbr_sendmap *rsm; 3313 3314 if (limit_type) { 3315 /* currently there is only one limit type */ 3316 if (V_tcp_map_split_limit > 0 && 3317 bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { 3318 BBR_STAT_INC(bbr_split_limited); 3319 if (!bbr->alloc_limit_reported) { 3320 bbr->alloc_limit_reported = 1; 3321 BBR_STAT_INC(bbr_alloc_limited_conns); 3322 } 3323 return (NULL); 3324 } 3325 } 3326 3327 /* allocate and mark in the limit type, if set */ 3328 rsm = bbr_alloc(bbr); 3329 if (rsm != NULL && limit_type) { 3330 rsm->r_limit_type = limit_type; 3331 bbr->r_ctl.rc_num_split_allocs++; 3332 } 3333 return (rsm); 3334 } 3335 3336 static void 3337 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 3338 { 3339 if (rsm->r_limit_type) { 3340 /* currently there is only one limit type */ 3341 bbr->r_ctl.rc_num_split_allocs--; 3342 } 3343 if (rsm->r_is_smallmap) 3344 bbr->r_ctl.rc_num_small_maps_alloced--; 3345 if (bbr->r_ctl.rc_tlp_send == rsm) 3346 bbr->r_ctl.rc_tlp_send = NULL; 3347 if (bbr->r_ctl.rc_resend == rsm) { 3348 bbr->r_ctl.rc_resend = NULL; 3349 } 3350 if (bbr->r_ctl.rc_next == rsm) 3351 bbr->r_ctl.rc_next = NULL; 3352 if (bbr->r_ctl.rc_sacklast == rsm) 3353 bbr->r_ctl.rc_sacklast = NULL; 3354 if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 3355 memset(rsm, 0, sizeof(struct bbr_sendmap)); 3356 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 3357 rsm->r_limit_type = 0; 3358 bbr->r_ctl.rc_free_cnt++; 3359 return; 3360 } 3361 bbr->r_ctl.rc_num_maps_alloced--; 3362 uma_zfree(bbr_zone, rsm); 3363 } 3364 3365 /* 3366 * Returns the BDP. 3367 */ 3368 static uint64_t 3369 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { 3370 /* 3371 * Calculate the bytes in flight needed given the bw (in bytes per 3372 * second) and the specifyed rtt in useconds. We need to put out the 3373 * returned value per RTT to match that rate. Gain will normally 3374 * raise it up from there. 3375 * 3376 * This should not overflow as long as the bandwidth is below 1 3377 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller 3378 * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). 3379 */ 3380 uint64_t usec_per_sec; 3381 3382 usec_per_sec = USECS_IN_SECOND; 3383 return ((rtt * bw) / usec_per_sec); 3384 } 3385 3386 /* 3387 * Return the initial cwnd. 3388 */ 3389 static uint32_t 3390 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) 3391 { 3392 uint32_t i_cwnd; 3393 3394 if (bbr->rc_init_win) { 3395 i_cwnd = bbr->rc_init_win * tp->t_maxseg; 3396 } else if (V_tcp_initcwnd_segments) 3397 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), 3398 max(2 * tp->t_maxseg, 14600)); 3399 else if (V_tcp_do_rfc3390) 3400 i_cwnd = min(4 * tp->t_maxseg, 3401 max(2 * tp->t_maxseg, 4380)); 3402 else { 3403 /* Per RFC5681 Section 3.1 */ 3404 if (tp->t_maxseg > 2190) 3405 i_cwnd = 2 * tp->t_maxseg; 3406 else if (tp->t_maxseg > 1095) 3407 i_cwnd = 3 * tp->t_maxseg; 3408 else 3409 i_cwnd = 4 * tp->t_maxseg; 3410 } 3411 return (i_cwnd); 3412 } 3413 3414 /* 3415 * Given a specified gain, return the target 3416 * cwnd based on that gain. 3417 */ 3418 static uint32_t 3419 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) 3420 { 3421 uint64_t bdp, rtt; 3422 uint32_t cwnd; 3423 3424 if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || 3425 (bbr_get_full_bw(bbr) == 0)) { 3426 /* No measurements yet */ 3427 return (bbr_initial_cwnd(bbr, bbr->rc_tp)); 3428 } 3429 /* 3430 * Get bytes per RTT needed (rttProp is normally in 3431 * bbr_cwndtarget_rtt_touse) 3432 */ 3433 rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); 3434 /* Get the bdp from the two values */ 3435 bdp = bbr_get_bw_delay_prod(rtt, bw); 3436 /* Now apply the gain */ 3437 cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); 3438 3439 return (cwnd); 3440 } 3441 3442 static uint32_t 3443 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) 3444 { 3445 uint32_t cwnd, mss; 3446 3447 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 3448 /* Get the base cwnd with gain rounded to a mss */ 3449 cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); 3450 /* 3451 * Add in N (2 default since we do not have a 3452 * fq layer to trap packets in) quanta's per the I-D 3453 * section 4.2.3.2 quanta adjust. 3454 */ 3455 cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); 3456 if (bbr->rc_use_google) { 3457 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3458 (bbr_state_val(bbr) == BBR_SUB_GAIN)) { 3459 /* 3460 * The linux implementation adds 3461 * an extra 2 x mss in gain cycle which 3462 * is documented no-where except in the code. 3463 * so we add more for Neal undocumented feature 3464 */ 3465 cwnd += 2 * mss; 3466 } 3467 if ((cwnd / mss) & 0x1) { 3468 /* Round up for odd num mss */ 3469 cwnd += mss; 3470 } 3471 } 3472 /* Are we below the min cwnd? */ 3473 if (cwnd < get_min_cwnd(bbr)) 3474 return (get_min_cwnd(bbr)); 3475 return (cwnd); 3476 } 3477 3478 static uint16_t 3479 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) 3480 { 3481 if (gain < 1) 3482 gain = 1; 3483 return (gain); 3484 } 3485 3486 static uint32_t 3487 bbr_get_header_oh(struct tcp_bbr *bbr) 3488 { 3489 int seg_oh; 3490 3491 seg_oh = 0; 3492 if (bbr->r_ctl.rc_inc_tcp_oh) { 3493 /* Do we include TCP overhead? */ 3494 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); 3495 } 3496 if (bbr->r_ctl.rc_inc_ip_oh) { 3497 /* Do we include IP overhead? */ 3498 #ifdef INET6 3499 if (bbr->r_is_v6) { 3500 seg_oh += sizeof(struct ip6_hdr); 3501 } else 3502 #endif 3503 { 3504 3505 #ifdef INET 3506 seg_oh += sizeof(struct ip); 3507 #endif 3508 } 3509 } 3510 if (bbr->r_ctl.rc_inc_enet_oh) { 3511 /* Do we include the ethernet overhead? */ 3512 seg_oh += sizeof(struct ether_header); 3513 } 3514 return(seg_oh); 3515 } 3516 3517 static uint32_t 3518 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) 3519 { 3520 uint64_t divor, res, tim; 3521 3522 if (useconds_time == 0) 3523 return (0); 3524 gain = bbr_gain_adjust(bbr, gain); 3525 divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; 3526 tim = useconds_time; 3527 res = (tim * bw * gain) / divor; 3528 if (res == 0) 3529 res = 1; 3530 return ((uint32_t)res); 3531 } 3532 3533 /* 3534 * Given a gain and a length return the delay in useconds that 3535 * should be used to evenly space out packets 3536 * on the connection (based on the gain factor). 3537 */ 3538 static uint32_t 3539 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) 3540 { 3541 uint64_t bw, lentim, res; 3542 uint32_t usecs, srtt, over = 0; 3543 uint32_t seg_oh, num_segs, maxseg; 3544 3545 if (len == 0) 3546 return (0); 3547 3548 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 3549 num_segs = (len + maxseg - 1) / maxseg; 3550 if (bbr->rc_use_google == 0) { 3551 seg_oh = bbr_get_header_oh(bbr); 3552 len += (num_segs * seg_oh); 3553 } 3554 gain = bbr_gain_adjust(bbr, gain); 3555 bw = bbr_get_bw(bbr); 3556 if (bbr->rc_use_google) { 3557 uint64_t cbw; 3558 3559 /* 3560 * Reduce the b/w by the google discount 3561 * factor 10 = 1%. 3562 */ 3563 cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); 3564 cbw /= (uint64_t)1000; 3565 /* We don't apply a discount if it results in 0 */ 3566 if (cbw > 0) 3567 bw = cbw; 3568 } 3569 lentim = ((uint64_t)len * 3570 (uint64_t)USECS_IN_SECOND * 3571 (uint64_t)BBR_UNIT); 3572 res = lentim / ((uint64_t)gain * bw); 3573 if (res == 0) 3574 res = 1; 3575 usecs = (uint32_t)res; 3576 srtt = bbr_get_rtt(bbr, BBR_SRTT); 3577 if (bbr_hptsi_max_mul && bbr_hptsi_max_div && 3578 (bbr->rc_use_google == 0) && 3579 (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { 3580 /* 3581 * We cannot let the delay be more than 1/2 the srtt time. 3582 * Otherwise we cannot pace out or send properly. 3583 */ 3584 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; 3585 BBR_STAT_INC(bbr_hpts_min_time); 3586 } 3587 if (!nolog) 3588 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); 3589 return (usecs); 3590 } 3591 3592 static void 3593 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, 3594 uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) 3595 { 3596 uint64_t bw; 3597 uint32_t cwnd, target_cwnd, saved_bytes, maxseg; 3598 int32_t meth; 3599 3600 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3601 3602 #ifdef STATS 3603 if ((tp->t_flags & TF_GPUTINPROG) && 3604 SEQ_GEQ(th->th_ack, tp->gput_ack)) { 3605 /* 3606 * Strech acks and compressed acks will cause this to 3607 * oscillate but we are doing it the same way as the main 3608 * stack so it will be compariable (though possibly not 3609 * ideal). 3610 */ 3611 int32_t cgput; 3612 int64_t gput, time_stamp; 3613 3614 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; 3615 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); 3616 cgput = gput / time_stamp; 3617 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, 3618 cgput); 3619 if (tp->t_stats_gput_prev > 0) 3620 stats_voi_update_abs_s32(tp->t_stats, 3621 VOI_TCP_GPUT_ND, 3622 ((gput - tp->t_stats_gput_prev) * 100) / 3623 tp->t_stats_gput_prev); 3624 tp->t_flags &= ~TF_GPUTINPROG; 3625 tp->t_stats_gput_prev = cgput; 3626 } 3627 #endif 3628 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3629 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 3630 /* We don't change anything in probe-rtt */ 3631 return; 3632 } 3633 maxseg = tp->t_maxseg - bbr->rc_last_options; 3634 saved_bytes = bytes_this_ack; 3635 bytes_this_ack += sack_changed; 3636 if (bytes_this_ack > prev_acked) { 3637 bytes_this_ack -= prev_acked; 3638 /* 3639 * A byte ack'd gives us a full mss 3640 * to be like linux i.e. they count packets. 3641 */ 3642 if ((bytes_this_ack < maxseg) && bbr->rc_use_google) 3643 bytes_this_ack = maxseg; 3644 } else { 3645 /* Unlikely */ 3646 bytes_this_ack = 0; 3647 } 3648 cwnd = tp->snd_cwnd; 3649 bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3650 if (bw) 3651 target_cwnd = bbr_get_target_cwnd(bbr, 3652 bw, 3653 (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); 3654 else 3655 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); 3656 if (IN_RECOVERY(tp->t_flags) && 3657 (bbr->bbr_prev_in_rec == 0)) { 3658 /* 3659 * We are entering recovery and 3660 * thus packet conservation. 3661 */ 3662 bbr->pkt_conservation = 1; 3663 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; 3664 cwnd = ctf_flight_size(tp, 3665 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3666 bytes_this_ack; 3667 } 3668 if (IN_RECOVERY(tp->t_flags)) { 3669 uint32_t flight; 3670 3671 bbr->bbr_prev_in_rec = 1; 3672 if (cwnd > losses) { 3673 cwnd -= losses; 3674 if (cwnd < maxseg) 3675 cwnd = maxseg; 3676 } else 3677 cwnd = maxseg; 3678 flight = ctf_flight_size(tp, 3679 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3680 bbr_log_type_cwndupd(bbr, flight, 0, 3681 losses, 10, 0, 0, line); 3682 if (bbr->pkt_conservation) { 3683 uint32_t time_in; 3684 3685 if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) 3686 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; 3687 else 3688 time_in = 0; 3689 3690 if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 3691 /* Clear packet conservation after an rttProp */ 3692 bbr->pkt_conservation = 0; 3693 } else { 3694 if ((flight + bytes_this_ack) > cwnd) 3695 cwnd = flight + bytes_this_ack; 3696 if (cwnd < get_min_cwnd(bbr)) 3697 cwnd = get_min_cwnd(bbr); 3698 tp->snd_cwnd = cwnd; 3699 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, 3700 prev_acked, 1, target_cwnd, th->th_ack, line); 3701 return; 3702 } 3703 } 3704 } else 3705 bbr->bbr_prev_in_rec = 0; 3706 if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { 3707 bbr->r_ctl.restrict_growth--; 3708 if (bytes_this_ack > maxseg) 3709 bytes_this_ack = maxseg; 3710 } 3711 if (bbr->rc_filled_pipe) { 3712 /* 3713 * Here we have exited startup and filled the pipe. We will 3714 * thus allow the cwnd to shrink to the target. We hit here 3715 * mostly. 3716 */ 3717 uint32_t s_cwnd; 3718 3719 meth = 2; 3720 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); 3721 if (s_cwnd > cwnd) 3722 cwnd = s_cwnd; 3723 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) 3724 cwnd = s_cwnd; 3725 } else { 3726 /* 3727 * Here we are still in startup, we increase cwnd by what 3728 * has been acked. 3729 */ 3730 if ((cwnd < target_cwnd) || 3731 (bbr->rc_past_init_win == 0)) { 3732 meth = 3; 3733 cwnd += bytes_this_ack; 3734 } else { 3735 /* 3736 * Method 4 means we are at target so no gain in 3737 * startup and past the initial window. 3738 */ 3739 meth = 4; 3740 } 3741 } 3742 tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); 3743 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); 3744 } 3745 3746 static void 3747 tcp_bbr_partialack(struct tcpcb *tp) 3748 { 3749 struct tcp_bbr *bbr; 3750 3751 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3752 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3753 if (ctf_flight_size(tp, 3754 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 3755 tp->snd_cwnd) { 3756 bbr->r_wanted_output = 1; 3757 } 3758 } 3759 3760 static void 3761 bbr_post_recovery(struct tcpcb *tp) 3762 { 3763 struct tcp_bbr *bbr; 3764 uint32_t flight; 3765 3766 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3767 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3768 /* 3769 * Here we just exit recovery. 3770 */ 3771 EXIT_RECOVERY(tp->t_flags); 3772 /* Lock in our b/w reduction for the specified number of pkt-epochs */ 3773 bbr->r_recovery_bw = 0; 3774 tp->snd_recover = tp->snd_una; 3775 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3776 bbr->pkt_conservation = 0; 3777 if (bbr->rc_use_google == 0) { 3778 /* 3779 * For non-google mode lets 3780 * go ahead and make sure we clear 3781 * the recovery state so if we 3782 * bounce back in to recovery we 3783 * will do PC. 3784 */ 3785 bbr->bbr_prev_in_rec = 0; 3786 } 3787 bbr_log_type_exit_rec(bbr); 3788 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3789 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3790 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); 3791 } else { 3792 /* For probe-rtt case lets fix up its saved_cwnd */ 3793 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { 3794 bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; 3795 bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); 3796 } 3797 } 3798 flight = ctf_flight_size(tp, 3799 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 3800 if ((bbr->rc_use_google == 0) && 3801 bbr_do_red) { 3802 uint64_t val, lr2use; 3803 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; 3804 uint32_t *cwnd_p; 3805 3806 if (bbr_get_rtt(bbr, BBR_SRTT)) { 3807 val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); 3808 val /= bbr_get_rtt(bbr, BBR_SRTT); 3809 ratio = (uint32_t)val; 3810 } else 3811 ratio = 1000; 3812 3813 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, 3814 bbr->r_ctl.recovery_lr, 21, 3815 ratio, 3816 bbr->r_ctl.rc_red_cwnd_pe, 3817 __LINE__); 3818 if ((ratio < bbr_do_red) || (bbr_do_red == 0)) 3819 goto done; 3820 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 3821 bbr_prtt_slam_cwnd) || 3822 (bbr_sub_drain_slam_cwnd && 3823 (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 3824 bbr->rc_hit_state_1 && 3825 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || 3826 ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && 3827 bbr_slam_cwnd_in_main_drain)) { 3828 /* 3829 * Here we must poke at the saved cwnd 3830 * as well as the cwnd. 3831 */ 3832 cwnd = bbr->r_ctl.rc_saved_cwnd; 3833 cwnd_p = &bbr->r_ctl.rc_saved_cwnd; 3834 } else { 3835 cwnd = tp->snd_cwnd; 3836 cwnd_p = &tp->snd_cwnd; 3837 } 3838 maxseg = tp->t_maxseg - bbr->rc_last_options; 3839 /* Add the overall lr with the recovery lr */ 3840 if (bbr->r_ctl.rc_lost == 0) 3841 lr2use = 0; 3842 else if (bbr->r_ctl.rc_delivered == 0) 3843 lr2use = 1000; 3844 else { 3845 lr2use = bbr->r_ctl.rc_lost * 1000; 3846 lr2use /= bbr->r_ctl.rc_delivered; 3847 } 3848 lr2use += bbr->r_ctl.recovery_lr; 3849 acks_inflight = (flight / (maxseg * 2)); 3850 if (bbr_red_scale) { 3851 lr2use *= bbr_get_rtt(bbr, BBR_SRTT); 3852 lr2use /= bbr_red_scale; 3853 if ((bbr_red_growth_restrict) && 3854 ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) 3855 bbr->r_ctl.restrict_growth += acks_inflight; 3856 } 3857 if (lr2use) { 3858 val = (uint64_t)cwnd * lr2use; 3859 val /= 1000; 3860 if (cwnd > val) 3861 newcwnd = roundup((cwnd - val), maxseg); 3862 else 3863 newcwnd = maxseg; 3864 } else { 3865 val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; 3866 val /= (uint64_t)bbr_red_div; 3867 newcwnd = roundup((uint32_t)val, maxseg); 3868 } 3869 /* with standard delayed acks how many acks can I expect? */ 3870 if (bbr_drop_limit == 0) { 3871 /* 3872 * Anticpate how much we will 3873 * raise the cwnd based on the acks. 3874 */ 3875 if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { 3876 /* We do enforce the min (with the acks) */ 3877 newcwnd = (get_min_cwnd(bbr) - acks_inflight); 3878 } 3879 } else { 3880 /* 3881 * A strict drop limit of N is inplace 3882 */ 3883 if (newcwnd < (bbr_drop_limit * maxseg)) { 3884 newcwnd = bbr_drop_limit * maxseg; 3885 } 3886 } 3887 /* For the next N acks do we restrict the growth */ 3888 *cwnd_p = newcwnd; 3889 if (tp->snd_cwnd > newcwnd) 3890 tp->snd_cwnd = newcwnd; 3891 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, 3892 (uint32_t)lr2use, 3893 bbr_get_rtt(bbr, BBR_SRTT), __LINE__); 3894 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; 3895 } 3896 done: 3897 bbr->r_ctl.recovery_lr = 0; 3898 if (flight <= tp->snd_cwnd) { 3899 bbr->r_wanted_output = 1; 3900 } 3901 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3902 } 3903 3904 static void 3905 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) 3906 { 3907 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 3908 /* Limit the drop in b/w to 1/2 our current filter. */ 3909 if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) 3910 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; 3911 if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) 3912 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; 3913 tcp_bbr_tso_size_check(bbr, cts); 3914 } 3915 3916 static void 3917 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) 3918 { 3919 struct tcp_bbr *bbr; 3920 3921 INP_WLOCK_ASSERT(tptoinpcb(tp)); 3922 #ifdef STATS 3923 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); 3924 #endif 3925 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 3926 switch (type) { 3927 case CC_NDUPACK: 3928 if (!IN_RECOVERY(tp->t_flags)) { 3929 tp->snd_recover = tp->snd_max; 3930 /* Start a new epoch */ 3931 bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 3932 if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { 3933 /* 3934 * Move forward the lt epoch 3935 * so it won't count the truncated 3936 * epoch. 3937 */ 3938 bbr->r_ctl.rc_lt_epoch++; 3939 } 3940 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 3941 /* 3942 * Just like the policer detection code 3943 * if we are in startup we must push 3944 * forward the last startup epoch 3945 * to hide the truncated PE. 3946 */ 3947 bbr->r_ctl.rc_bbr_last_startup_epoch++; 3948 } 3949 bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; 3950 ENTER_RECOVERY(tp->t_flags); 3951 bbr->rc_tlp_rtx_out = 0; 3952 bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; 3953 tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); 3954 if (tcp_in_hpts(bbr->rc_tp) && 3955 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { 3956 /* 3957 * When we enter recovery, we need to restart 3958 * any timers. This may mean we gain an agg 3959 * early, which will be made up for at the last 3960 * rxt out. 3961 */ 3962 bbr->rc_timer_first = 1; 3963 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 3964 } 3965 /* 3966 * Calculate a new cwnd based on to the current 3967 * delivery rate with no gain. We get the bdp 3968 * without gaining it up like we normally would and 3969 * we use the last cur_del_rate. 3970 */ 3971 if ((bbr->rc_use_google == 0) && 3972 (bbr->r_ctl.bbr_rttprobe_gain_val || 3973 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { 3974 tp->snd_cwnd = ctf_flight_size(tp, 3975 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 3976 (tp->t_maxseg - bbr->rc_last_options); 3977 if (tp->snd_cwnd < get_min_cwnd(bbr)) { 3978 /* We always gate to min cwnd */ 3979 tp->snd_cwnd = get_min_cwnd(bbr); 3980 } 3981 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); 3982 } 3983 bbr_log_type_enter_rec(bbr, rsm->r_start); 3984 } 3985 break; 3986 case CC_RTO_ERR: 3987 KMOD_TCPSTAT_INC(tcps_sndrexmitbad); 3988 /* RTO was unnecessary, so reset everything. */ 3989 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); 3990 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 3991 tp->snd_cwnd = tp->snd_cwnd_prev; 3992 tp->snd_ssthresh = tp->snd_ssthresh_prev; 3993 tp->snd_recover = tp->snd_recover_prev; 3994 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); 3995 bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); 3996 } 3997 tp->t_badrxtwin = 0; 3998 break; 3999 } 4000 } 4001 4002 /* 4003 * Indicate whether this ack should be delayed. We can delay the ack if 4004 * following conditions are met: 4005 * - There is no delayed ack timer in progress. 4006 * - Our last ack wasn't a 0-sized window. We never want to delay 4007 * the ack that opens up a 0-sized window. 4008 * - LRO wasn't used for this segment. We make sure by checking that the 4009 * segment size is not larger than the MSS. 4010 * - Delayed acks are enabled or this is a half-synchronized T/TCP 4011 * connection. 4012 * - The data being acked is less than a full segment (a stretch ack 4013 * of more than a segment we should ack. 4014 * - nsegs is 1 (if its more than that we received more than 1 ack). 4015 */ 4016 #define DELAY_ACK(tp, bbr, nsegs) \ 4017 (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ 4018 ((tp->t_flags & TF_DELACK) == 0) && \ 4019 ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ 4020 (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) 4021 4022 /* 4023 * Return the lowest RSM in the map of 4024 * packets still in flight that is not acked. 4025 * This should normally find on the first one 4026 * since we remove packets from the send 4027 * map after they are marked ACKED. 4028 */ 4029 static struct bbr_sendmap * 4030 bbr_find_lowest_rsm(struct tcp_bbr *bbr) 4031 { 4032 struct bbr_sendmap *rsm; 4033 4034 /* 4035 * Walk the time-order transmitted list looking for an rsm that is 4036 * not acked. This will be the one that was sent the longest time 4037 * ago that is still outstanding. 4038 */ 4039 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { 4040 if (rsm->r_flags & BBR_ACKED) { 4041 continue; 4042 } 4043 goto finish; 4044 } 4045 finish: 4046 return (rsm); 4047 } 4048 4049 static struct bbr_sendmap * 4050 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 4051 { 4052 struct bbr_sendmap *prsm; 4053 4054 /* 4055 * Walk the sequence order list backward until we hit and arrive at 4056 * the highest seq not acked. In theory when this is called it 4057 * should be the last segment (which it was not). 4058 */ 4059 prsm = rsm; 4060 TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4061 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { 4062 continue; 4063 } 4064 return (prsm); 4065 } 4066 return (NULL); 4067 } 4068 4069 /* 4070 * Returns to the caller the number of microseconds that 4071 * the packet can be outstanding before we think we 4072 * should have had an ack returned. 4073 */ 4074 static uint32_t 4075 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) 4076 { 4077 /* 4078 * lro is the flag we use to determine if we have seen reordering. 4079 * If it gets set we have seen reordering. The reorder logic either 4080 * works in one of two ways: 4081 * 4082 * If reorder-fade is configured, then we track the last time we saw 4083 * re-ordering occur. If we reach the point where enough time as 4084 * passed we no longer consider reordering has occuring. 4085 * 4086 * Or if reorder-face is 0, then once we see reordering we consider 4087 * the connection to alway be subject to reordering and just set lro 4088 * to 1. 4089 * 4090 * In the end if lro is non-zero we add the extra time for 4091 * reordering in. 4092 */ 4093 int32_t lro; 4094 uint32_t thresh, t_rxtcur; 4095 4096 if (srtt == 0) 4097 srtt = 1; 4098 if (bbr->r_ctl.rc_reorder_ts) { 4099 if (bbr->r_ctl.rc_reorder_fade) { 4100 if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { 4101 lro = cts - bbr->r_ctl.rc_reorder_ts; 4102 if (lro == 0) { 4103 /* 4104 * No time as passed since the last 4105 * reorder, mark it as reordering. 4106 */ 4107 lro = 1; 4108 } 4109 } else { 4110 /* Negative time? */ 4111 lro = 0; 4112 } 4113 if (lro > bbr->r_ctl.rc_reorder_fade) { 4114 /* Turn off reordering seen too */ 4115 bbr->r_ctl.rc_reorder_ts = 0; 4116 lro = 0; 4117 } 4118 } else { 4119 /* Reodering does not fade */ 4120 lro = 1; 4121 } 4122 } else { 4123 lro = 0; 4124 } 4125 thresh = srtt + bbr->r_ctl.rc_pkt_delay; 4126 if (lro) { 4127 /* It must be set, if not you get 1/4 rtt */ 4128 if (bbr->r_ctl.rc_reorder_shift) 4129 thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); 4130 else 4131 thresh += (srtt >> 2); 4132 } else { 4133 thresh += 1000; 4134 } 4135 /* We don't let the rack timeout be above a RTO */ 4136 if ((bbr->rc_tp)->t_srtt == 0) 4137 t_rxtcur = BBR_INITIAL_RTO; 4138 else 4139 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); 4140 if (thresh > t_rxtcur) { 4141 thresh = t_rxtcur; 4142 } 4143 /* And we don't want it above the RTO max either */ 4144 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4145 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4146 } 4147 bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); 4148 return (thresh); 4149 } 4150 4151 /* 4152 * Return to the caller the amount of time in mico-seconds 4153 * that should be used for the TLP timer from the last 4154 * send time of this packet. 4155 */ 4156 static uint32_t 4157 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, 4158 struct bbr_sendmap *rsm, uint32_t srtt, 4159 uint32_t cts) 4160 { 4161 uint32_t thresh, len, maxseg, t_rxtcur; 4162 struct bbr_sendmap *prsm; 4163 4164 if (srtt == 0) 4165 srtt = 1; 4166 if (bbr->rc_tlp_threshold) 4167 thresh = srtt + (srtt / bbr->rc_tlp_threshold); 4168 else 4169 thresh = (srtt * 2); 4170 maxseg = tp->t_maxseg - bbr->rc_last_options; 4171 /* Get the previous sent packet, if any */ 4172 len = rsm->r_end - rsm->r_start; 4173 4174 /* 2.1 behavior */ 4175 prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); 4176 if (prsm && (len <= maxseg)) { 4177 /* 4178 * Two packets outstanding, thresh should be (2*srtt) + 4179 * possible inter-packet delay (if any). 4180 */ 4181 uint32_t inter_gap = 0; 4182 int idx, nidx; 4183 4184 idx = rsm->r_rtr_cnt - 1; 4185 nidx = prsm->r_rtr_cnt - 1; 4186 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { 4187 /* Yes it was sent later (or at the same time) */ 4188 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; 4189 } 4190 thresh += inter_gap; 4191 } else if (len <= maxseg) { 4192 /* 4193 * Possibly compensate for delayed-ack. 4194 */ 4195 uint32_t alt_thresh; 4196 4197 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; 4198 if (alt_thresh > thresh) 4199 thresh = alt_thresh; 4200 } 4201 /* Not above the current RTO */ 4202 if (tp->t_srtt == 0) 4203 t_rxtcur = BBR_INITIAL_RTO; 4204 else 4205 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); 4206 4207 bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); 4208 /* Not above an RTO */ 4209 if (thresh > t_rxtcur) { 4210 thresh = t_rxtcur; 4211 } 4212 /* Not above a RTO max */ 4213 if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { 4214 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); 4215 } 4216 /* And now apply the user TLP min */ 4217 if (thresh < bbr_tlp_min) { 4218 thresh = bbr_tlp_min; 4219 } 4220 return (thresh); 4221 } 4222 4223 /* 4224 * Return one of three RTTs to use (in microseconds). 4225 */ 4226 static __inline uint32_t 4227 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) 4228 { 4229 uint32_t f_rtt; 4230 uint32_t srtt; 4231 4232 f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 4233 if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { 4234 /* We have no rtt at all */ 4235 if (bbr->rc_tp->t_srtt == 0) 4236 f_rtt = BBR_INITIAL_RTO; 4237 else 4238 f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4239 /* 4240 * Since we don't know how good the rtt is apply a 4241 * delayed-ack min 4242 */ 4243 if (f_rtt < bbr_delayed_ack_time) { 4244 f_rtt = bbr_delayed_ack_time; 4245 } 4246 } 4247 /* Take the filter version or last measured pkt-rtt */ 4248 if (rtt_type == BBR_RTT_PROP) { 4249 srtt = f_rtt; 4250 } else if (rtt_type == BBR_RTT_PKTRTT) { 4251 if (bbr->r_ctl.rc_pkt_epoch_rtt) { 4252 srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 4253 } else { 4254 /* No pkt rtt yet */ 4255 srtt = f_rtt; 4256 } 4257 } else if (rtt_type == BBR_RTT_RACK) { 4258 srtt = bbr->r_ctl.rc_last_rtt; 4259 /* We need to add in any internal delay for our timer */ 4260 if (bbr->rc_ack_was_delayed) 4261 srtt += bbr->r_ctl.rc_ack_hdwr_delay; 4262 } else if (rtt_type == BBR_SRTT) { 4263 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 4264 } else { 4265 /* TSNH */ 4266 srtt = f_rtt; 4267 #ifdef BBR_INVARIANTS 4268 panic("Unknown rtt request type %d", rtt_type); 4269 #endif 4270 } 4271 return (srtt); 4272 } 4273 4274 static int 4275 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) 4276 { 4277 uint32_t thresh; 4278 4279 thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), 4280 cts, rsm); 4281 if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { 4282 /* It is lost (past time) */ 4283 return (1); 4284 } 4285 return (0); 4286 } 4287 4288 /* 4289 * Return a sendmap if we need to retransmit something. 4290 */ 4291 static struct bbr_sendmap * 4292 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4293 { 4294 /* 4295 * Check to see that we don't need to fall into recovery. We will 4296 * need to do so if our oldest transmit is past the time we should 4297 * have had an ack. 4298 */ 4299 4300 struct bbr_sendmap *rsm; 4301 int32_t idx; 4302 4303 if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { 4304 /* Nothing outstanding that we know of */ 4305 return (NULL); 4306 } 4307 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 4308 if (rsm == NULL) { 4309 /* Nothing in the transmit map */ 4310 return (NULL); 4311 } 4312 if (tp->t_flags & TF_SENTFIN) { 4313 /* Fin restricted, don't find anything once a fin is sent */ 4314 return (NULL); 4315 } 4316 if (rsm->r_flags & BBR_ACKED) { 4317 /* 4318 * Ok the first one is acked (this really should not happen 4319 * since we remove the from the tmap once they are acked) 4320 */ 4321 rsm = bbr_find_lowest_rsm(bbr); 4322 if (rsm == NULL) 4323 return (NULL); 4324 } 4325 idx = rsm->r_rtr_cnt - 1; 4326 if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { 4327 /* Send timestamp is the same or less? can't be ready */ 4328 return (NULL); 4329 } 4330 /* Get our RTT time */ 4331 if (bbr_is_lost(bbr, rsm, cts) && 4332 ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || 4333 (rsm->r_flags & BBR_SACK_PASSED))) { 4334 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4335 rsm->r_flags |= BBR_MARKED_LOST; 4336 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4337 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4338 } 4339 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); 4340 #ifdef BBR_INVARIANTS 4341 if ((rsm->r_end - rsm->r_start) == 0) 4342 panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); 4343 #endif 4344 return (rsm); 4345 } 4346 return (NULL); 4347 } 4348 4349 /* 4350 * RACK Timer, here we simply do logging and house keeping. 4351 * the normal bbr_output_wtime() function will call the 4352 * appropriate thing to check if we need to do a RACK retransmit. 4353 * We return 1, saying don't proceed with bbr_output_wtime only 4354 * when all timers have been stopped (destroyed PCB?). 4355 */ 4356 static int 4357 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4358 { 4359 /* 4360 * This timer simply provides an internal trigger to send out data. 4361 * The check_recovery_mode call will see if there are needed 4362 * retransmissions, if so we will enter fast-recovery. The output 4363 * call may or may not do the same thing depending on sysctl 4364 * settings. 4365 */ 4366 uint32_t lost; 4367 4368 if (bbr->rc_all_timers_stopped) { 4369 return (1); 4370 } 4371 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4372 /* Its not time yet */ 4373 return (0); 4374 } 4375 BBR_STAT_INC(bbr_to_tot); 4376 lost = bbr->r_ctl.rc_lost; 4377 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4378 bbr_set_state(tp, bbr, 0); 4379 bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); 4380 if (bbr->r_ctl.rc_resend == NULL) { 4381 /* Lets do the check here */ 4382 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 4383 } 4384 if (bbr_policer_call_from_rack_to) 4385 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4386 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; 4387 return (0); 4388 } 4389 4390 static __inline void 4391 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) 4392 { 4393 int idx; 4394 4395 nrsm->r_start = start; 4396 nrsm->r_end = rsm->r_end; 4397 nrsm->r_rtr_cnt = rsm->r_rtr_cnt; 4398 nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; 4399 nrsm->r_flags = rsm->r_flags; 4400 /* We don't transfer forward the SYN flag */ 4401 nrsm->r_flags &= ~BBR_HAS_SYN; 4402 /* We move forward the FIN flag, not that this should happen */ 4403 rsm->r_flags &= ~BBR_HAS_FIN; 4404 nrsm->r_dupack = rsm->r_dupack; 4405 nrsm->r_rtr_bytes = 0; 4406 nrsm->r_is_gain = rsm->r_is_gain; 4407 nrsm->r_is_drain = rsm->r_is_drain; 4408 nrsm->r_delivered = rsm->r_delivered; 4409 nrsm->r_ts_valid = rsm->r_ts_valid; 4410 nrsm->r_del_ack_ts = rsm->r_del_ack_ts; 4411 nrsm->r_del_time = rsm->r_del_time; 4412 nrsm->r_app_limited = rsm->r_app_limited; 4413 nrsm->r_first_sent_time = rsm->r_first_sent_time; 4414 nrsm->r_flight_at_send = rsm->r_flight_at_send; 4415 /* We split a piece the lower section looses any just_ret flag. */ 4416 nrsm->r_bbr_state = rsm->r_bbr_state; 4417 for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { 4418 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; 4419 } 4420 rsm->r_end = nrsm->r_start; 4421 idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); 4422 idx /= 8; 4423 /* Check if we got too small */ 4424 if ((rsm->r_is_smallmap == 0) && 4425 ((rsm->r_end - rsm->r_start) <= idx)) { 4426 bbr->r_ctl.rc_num_small_maps_alloced++; 4427 rsm->r_is_smallmap = 1; 4428 } 4429 /* Check the new one as well */ 4430 if ((nrsm->r_end - nrsm->r_start) <= idx) { 4431 bbr->r_ctl.rc_num_small_maps_alloced++; 4432 nrsm->r_is_smallmap = 1; 4433 } 4434 } 4435 4436 static int 4437 bbr_sack_mergable(struct bbr_sendmap *at, 4438 uint32_t start, uint32_t end) 4439 { 4440 /* 4441 * Given a sack block defined by 4442 * start and end, and a current position 4443 * at. Return 1 if either side of at 4444 * would show that the block is mergable 4445 * to that side. A block to be mergable 4446 * must have overlap with the start/end 4447 * and be in the SACK'd state. 4448 */ 4449 struct bbr_sendmap *l_rsm; 4450 struct bbr_sendmap *r_rsm; 4451 4452 /* first get the either side blocks */ 4453 l_rsm = TAILQ_PREV(at, bbr_head, r_next); 4454 r_rsm = TAILQ_NEXT(at, r_next); 4455 if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { 4456 /* Potentially mergeable */ 4457 if ((l_rsm->r_end == start) || 4458 (SEQ_LT(start, l_rsm->r_end) && 4459 SEQ_GT(end, l_rsm->r_end))) { 4460 /* 4461 * map blk |------| 4462 * sack blk |------| 4463 * <or> 4464 * map blk |------| 4465 * sack blk |------| 4466 */ 4467 return (1); 4468 } 4469 } 4470 if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { 4471 /* Potentially mergeable */ 4472 if ((r_rsm->r_start == end) || 4473 (SEQ_LT(start, r_rsm->r_start) && 4474 SEQ_GT(end, r_rsm->r_start))) { 4475 /* 4476 * map blk |---------| 4477 * sack blk |----| 4478 * <or> 4479 * map blk |---------| 4480 * sack blk |-------| 4481 */ 4482 return (1); 4483 } 4484 } 4485 return (0); 4486 } 4487 4488 static struct bbr_sendmap * 4489 bbr_merge_rsm(struct tcp_bbr *bbr, 4490 struct bbr_sendmap *l_rsm, 4491 struct bbr_sendmap *r_rsm) 4492 { 4493 /* 4494 * We are merging two ack'd RSM's, 4495 * the l_rsm is on the left (lower seq 4496 * values) and the r_rsm is on the right 4497 * (higher seq value). The simplest way 4498 * to merge these is to move the right 4499 * one into the left. I don't think there 4500 * is any reason we need to try to find 4501 * the oldest (or last oldest retransmitted). 4502 */ 4503 l_rsm->r_end = r_rsm->r_end; 4504 if (l_rsm->r_dupack < r_rsm->r_dupack) 4505 l_rsm->r_dupack = r_rsm->r_dupack; 4506 if (r_rsm->r_rtr_bytes) 4507 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; 4508 if (r_rsm->r_in_tmap) { 4509 /* This really should not happen */ 4510 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); 4511 } 4512 if (r_rsm->r_app_limited) 4513 l_rsm->r_app_limited = r_rsm->r_app_limited; 4514 /* Now the flags */ 4515 if (r_rsm->r_flags & BBR_HAS_FIN) 4516 l_rsm->r_flags |= BBR_HAS_FIN; 4517 if (r_rsm->r_flags & BBR_TLP) 4518 l_rsm->r_flags |= BBR_TLP; 4519 if (r_rsm->r_flags & BBR_RWND_COLLAPSED) 4520 l_rsm->r_flags |= BBR_RWND_COLLAPSED; 4521 if (r_rsm->r_flags & BBR_MARKED_LOST) { 4522 /* This really should not happen */ 4523 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; 4524 } 4525 TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); 4526 if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { 4527 /* Transfer the split limit to the map we free */ 4528 r_rsm->r_limit_type = l_rsm->r_limit_type; 4529 l_rsm->r_limit_type = 0; 4530 } 4531 bbr_free(bbr, r_rsm); 4532 return(l_rsm); 4533 } 4534 4535 /* 4536 * TLP Timer, here we simply setup what segment we want to 4537 * have the TLP expire on, the normal bbr_output_wtime() will then 4538 * send it out. 4539 * 4540 * We return 1, saying don't proceed with bbr_output_wtime only 4541 * when all timers have been stopped (destroyed PCB?). 4542 */ 4543 static int 4544 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4545 { 4546 /* 4547 * Tail Loss Probe. 4548 */ 4549 struct bbr_sendmap *rsm = NULL; 4550 struct socket *so; 4551 uint32_t amm; 4552 uint32_t out, avail; 4553 uint32_t maxseg; 4554 int collapsed_win = 0; 4555 4556 if (bbr->rc_all_timers_stopped) { 4557 return (1); 4558 } 4559 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 4560 /* Its not time yet */ 4561 return (0); 4562 } 4563 if (ctf_progress_timeout_check(tp, true)) { 4564 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4565 return (-ETIMEDOUT); /* tcp_drop() */ 4566 } 4567 /* Did we somehow get into persists? */ 4568 if (bbr->rc_in_persist) { 4569 return (0); 4570 } 4571 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4572 bbr_set_state(tp, bbr, 0); 4573 BBR_STAT_INC(bbr_tlp_tot); 4574 maxseg = tp->t_maxseg - bbr->rc_last_options; 4575 /* 4576 * A TLP timer has expired. We have been idle for 2 rtts. So we now 4577 * need to figure out how to force a full MSS segment out. 4578 */ 4579 so = tptosocket(tp); 4580 avail = sbavail(&so->so_snd); 4581 out = ctf_outstanding(tp); 4582 if (out > tp->snd_wnd) { 4583 /* special case, we need a retransmission */ 4584 collapsed_win = 1; 4585 goto need_retran; 4586 } 4587 if (avail > out) { 4588 /* New data is available */ 4589 amm = avail - out; 4590 if (amm > maxseg) { 4591 amm = maxseg; 4592 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { 4593 /* not enough to fill a MTU and no-delay is off */ 4594 goto need_retran; 4595 } 4596 /* Set the send-new override */ 4597 if ((out + amm) <= tp->snd_wnd) { 4598 bbr->rc_tlp_new_data = 1; 4599 } else { 4600 goto need_retran; 4601 } 4602 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4603 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; 4604 bbr->r_ctl.rc_tlp_send = NULL; 4605 /* cap any slots */ 4606 BBR_STAT_INC(bbr_tlp_newdata); 4607 goto send; 4608 } 4609 need_retran: 4610 /* 4611 * Ok we need to arrange the last un-acked segment to be re-sent, or 4612 * optionally the first un-acked segment. 4613 */ 4614 if (collapsed_win == 0) { 4615 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 4616 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { 4617 rsm = bbr_find_high_nonack(bbr, rsm); 4618 } 4619 if (rsm == NULL) { 4620 goto restore; 4621 } 4622 } else { 4623 /* 4624 * We must find the last segment 4625 * that was acceptable by the client. 4626 */ 4627 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 4628 if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { 4629 /* Found one */ 4630 break; 4631 } 4632 } 4633 if (rsm == NULL) { 4634 /* None? if so send the first */ 4635 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4636 if (rsm == NULL) 4637 goto restore; 4638 } 4639 } 4640 if ((rsm->r_end - rsm->r_start) > maxseg) { 4641 /* 4642 * We need to split this the last segment in two. 4643 */ 4644 struct bbr_sendmap *nrsm; 4645 4646 nrsm = bbr_alloc_full_limit(bbr); 4647 if (nrsm == NULL) { 4648 /* 4649 * We can't get memory to split, we can either just 4650 * not split it. Or retransmit the whole piece, lets 4651 * do the large send (BTLP :-) ). 4652 */ 4653 goto go_for_it; 4654 } 4655 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); 4656 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 4657 if (rsm->r_in_tmap) { 4658 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 4659 nrsm->r_in_tmap = 1; 4660 } 4661 rsm->r_flags &= (~BBR_HAS_FIN); 4662 rsm = nrsm; 4663 } 4664 go_for_it: 4665 bbr->r_ctl.rc_tlp_send = rsm; 4666 bbr->rc_tlp_rtx_out = 1; 4667 if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { 4668 bbr->r_ctl.rc_tlp_seg_send_cnt++; 4669 tp->t_rxtshift++; 4670 } else { 4671 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; 4672 bbr->r_ctl.rc_tlp_seg_send_cnt = 1; 4673 } 4674 send: 4675 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { 4676 /* 4677 * Can't [re]/transmit a segment we have retransmitted the 4678 * max times. We need the retransmit timer to take over. 4679 */ 4680 restore: 4681 bbr->rc_tlp_new_data = 0; 4682 bbr->r_ctl.rc_tlp_send = NULL; 4683 if (rsm) 4684 rsm->r_flags &= ~BBR_TLP; 4685 BBR_STAT_INC(bbr_tlp_retran_fail); 4686 return (0); 4687 } else if (rsm) { 4688 rsm->r_flags |= BBR_TLP; 4689 } 4690 if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && 4691 (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { 4692 /* 4693 * We have retransmitted to many times for TLP. Switch to 4694 * the regular RTO timer 4695 */ 4696 goto restore; 4697 } 4698 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); 4699 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; 4700 return (0); 4701 } 4702 4703 /* 4704 * Delayed ack Timer, here we simply need to setup the 4705 * ACK_NOW flag and remove the DELACK flag. From there 4706 * the output routine will send the ack out. 4707 * 4708 * We only return 1, saying don't proceed, if all timers 4709 * are stopped (destroyed PCB?). 4710 */ 4711 static int 4712 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4713 { 4714 if (bbr->rc_all_timers_stopped) { 4715 return (1); 4716 } 4717 bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); 4718 tp->t_flags &= ~TF_DELACK; 4719 tp->t_flags |= TF_ACKNOW; 4720 KMOD_TCPSTAT_INC(tcps_delack); 4721 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; 4722 return (0); 4723 } 4724 4725 /* 4726 * Here we send a KEEP-ALIVE like probe to the 4727 * peer, we do not send data. 4728 * 4729 * We only return 1, saying don't proceed, if all timers 4730 * are stopped (destroyed PCB?). 4731 */ 4732 static int 4733 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4734 { 4735 struct tcptemp *t_template; 4736 int32_t retval = 1; 4737 4738 if (bbr->rc_all_timers_stopped) { 4739 return (1); 4740 } 4741 if (bbr->rc_in_persist == 0) 4742 return (0); 4743 4744 /* 4745 * Persistence timer into zero window. Force a byte to be output, if 4746 * possible. 4747 */ 4748 bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); 4749 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; 4750 KMOD_TCPSTAT_INC(tcps_persisttimeo); 4751 /* 4752 * Have we exceeded the user specified progress time? 4753 */ 4754 if (ctf_progress_timeout_check(tp, true)) { 4755 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4756 return (-ETIMEDOUT); /* tcp_drop() */ 4757 } 4758 /* 4759 * Hack: if the peer is dead/unreachable, we do not time out if the 4760 * window is closed. After a full backoff, drop the connection if 4761 * the idle time (no responses to probes) reaches the maximum 4762 * backoff that we would use if retransmitting. 4763 */ 4764 if (tp->t_rxtshift >= V_tcp_retries && 4765 (ticks - tp->t_rcvtime >= tcp_maxpersistidle || 4766 ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { 4767 KMOD_TCPSTAT_INC(tcps_persistdrop); 4768 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4769 return (-ETIMEDOUT); /* tcp_drop() */ 4770 } 4771 if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && 4772 tp->snd_una == tp->snd_max) { 4773 bbr_exit_persist(tp, bbr, cts, __LINE__); 4774 retval = 0; 4775 goto out; 4776 } 4777 /* 4778 * If the user has closed the socket then drop a persisting 4779 * connection after a much reduced timeout. 4780 */ 4781 if (tp->t_state > TCPS_CLOSE_WAIT && 4782 (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { 4783 KMOD_TCPSTAT_INC(tcps_persistdrop); 4784 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); 4785 return (-ETIMEDOUT); /* tcp_drop() */ 4786 } 4787 t_template = tcpip_maketemplate(bbr->rc_inp); 4788 if (t_template) { 4789 tcp_respond(tp, t_template->tt_ipgen, 4790 &t_template->tt_t, (struct mbuf *)NULL, 4791 tp->rcv_nxt, tp->snd_una - 1, 0); 4792 /* This sends an ack */ 4793 if (tp->t_flags & TF_DELACK) 4794 tp->t_flags &= ~TF_DELACK; 4795 free(t_template, M_TEMP); 4796 } 4797 if (tp->t_rxtshift < V_tcp_retries) 4798 tp->t_rxtshift++; 4799 bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); 4800 out: 4801 return (retval); 4802 } 4803 4804 /* 4805 * If a keepalive goes off, we had no other timers 4806 * happening. We always return 1 here since this 4807 * routine either drops the connection or sends 4808 * out a segment with respond. 4809 */ 4810 static int 4811 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4812 { 4813 struct tcptemp *t_template; 4814 struct inpcb *inp = tptoinpcb(tp); 4815 4816 if (bbr->rc_all_timers_stopped) { 4817 return (1); 4818 } 4819 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; 4820 bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); 4821 /* 4822 * Keep-alive timer went off; send something or drop connection if 4823 * idle for too long. 4824 */ 4825 KMOD_TCPSTAT_INC(tcps_keeptimeo); 4826 if (tp->t_state < TCPS_ESTABLISHED) 4827 goto dropit; 4828 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && 4829 tp->t_state <= TCPS_CLOSING) { 4830 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) 4831 goto dropit; 4832 /* 4833 * Send a packet designed to force a response if the peer is 4834 * up and reachable: either an ACK if the connection is 4835 * still alive, or an RST if the peer has closed the 4836 * connection due to timeout or reboot. Using sequence 4837 * number tp->snd_una-1 causes the transmitted zero-length 4838 * segment to lie outside the receive window; by the 4839 * protocol spec, this requires the correspondent TCP to 4840 * respond. 4841 */ 4842 KMOD_TCPSTAT_INC(tcps_keepprobe); 4843 t_template = tcpip_maketemplate(inp); 4844 if (t_template) { 4845 tcp_respond(tp, t_template->tt_ipgen, 4846 &t_template->tt_t, (struct mbuf *)NULL, 4847 tp->rcv_nxt, tp->snd_una - 1, 0); 4848 free(t_template, M_TEMP); 4849 } 4850 } 4851 bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); 4852 return (1); 4853 dropit: 4854 KMOD_TCPSTAT_INC(tcps_keepdrops); 4855 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); 4856 return (-ETIMEDOUT); /* tcp_drop() */ 4857 } 4858 4859 /* 4860 * Retransmit helper function, clear up all the ack 4861 * flags and take care of important book keeping. 4862 */ 4863 static void 4864 bbr_remxt_tmr(struct tcpcb *tp) 4865 { 4866 /* 4867 * The retransmit timer went off, all sack'd blocks must be 4868 * un-acked. 4869 */ 4870 struct bbr_sendmap *rsm, *trsm = NULL; 4871 struct tcp_bbr *bbr; 4872 uint32_t cts, lost; 4873 4874 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 4875 cts = tcp_get_usecs(&bbr->rc_tv); 4876 lost = bbr->r_ctl.rc_lost; 4877 if (bbr->r_state && (bbr->r_state != tp->t_state)) 4878 bbr_set_state(tp, bbr, 0); 4879 4880 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 4881 if (rsm->r_flags & BBR_ACKED) { 4882 uint32_t old_flags; 4883 4884 rsm->r_dupack = 0; 4885 if (rsm->r_in_tmap == 0) { 4886 /* We must re-add it back to the tlist */ 4887 if (trsm == NULL) { 4888 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 4889 } else { 4890 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); 4891 } 4892 rsm->r_in_tmap = 1; 4893 } 4894 old_flags = rsm->r_flags; 4895 rsm->r_flags |= BBR_RXT_CLEARED; 4896 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); 4897 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 4898 } else { 4899 if ((tp->t_state < TCPS_ESTABLISHED) && 4900 (rsm->r_start == tp->snd_una)) { 4901 /* 4902 * Special case for TCP FO. Where 4903 * we sent more data beyond the snd_max. 4904 * We don't mark that as lost and stop here. 4905 */ 4906 break; 4907 } 4908 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { 4909 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 4910 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 4911 } 4912 if (bbr_marks_rxt_sack_passed) { 4913 /* 4914 * With this option, we will rack out 4915 * in 1ms increments the rest of the packets. 4916 */ 4917 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; 4918 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4919 } else { 4920 /* 4921 * With this option we only mark them lost 4922 * and remove all sack'd markings. We will run 4923 * another RXT or a TLP. This will cause 4924 * us to eventually send more based on what 4925 * ack's come in. 4926 */ 4927 rsm->r_flags |= BBR_MARKED_LOST; 4928 rsm->r_flags &= ~BBR_WAS_SACKPASS; 4929 rsm->r_flags &= ~BBR_SACK_PASSED; 4930 } 4931 } 4932 trsm = rsm; 4933 } 4934 bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); 4935 /* Clear the count (we just un-acked them) */ 4936 bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); 4937 bbr->rc_tlp_new_data = 0; 4938 bbr->r_ctl.rc_tlp_seg_send_cnt = 0; 4939 /* zap the behindness on a rxt */ 4940 bbr->r_ctl.rc_hptsi_agg_delay = 0; 4941 bbr->r_agg_early_set = 0; 4942 bbr->r_ctl.rc_agg_early = 0; 4943 bbr->rc_tlp_rtx_out = 0; 4944 bbr->r_ctl.rc_sacked = 0; 4945 bbr->r_ctl.rc_sacklast = NULL; 4946 bbr->r_timer_override = 1; 4947 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); 4948 } 4949 4950 /* 4951 * Re-transmit timeout! If we drop the PCB we will return 1, otherwise 4952 * we will setup to retransmit the lowest seq number outstanding. 4953 */ 4954 static int 4955 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 4956 { 4957 struct inpcb *inp = tptoinpcb(tp); 4958 int32_t rexmt; 4959 int32_t retval = 0; 4960 bool isipv6; 4961 4962 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 4963 if (bbr->rc_all_timers_stopped) { 4964 return (1); 4965 } 4966 if (TCPS_HAVEESTABLISHED(tp->t_state) && 4967 (tp->snd_una == tp->snd_max)) { 4968 /* Nothing outstanding .. nothing to do */ 4969 return (0); 4970 } 4971 /* 4972 * Retransmission timer went off. Message has not been acked within 4973 * retransmit interval. Back off to a longer retransmit interval 4974 * and retransmit one segment. 4975 */ 4976 if (ctf_progress_timeout_check(tp, true)) { 4977 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 4978 return (-ETIMEDOUT); /* tcp_drop() */ 4979 } 4980 bbr_remxt_tmr(tp); 4981 if ((bbr->r_ctl.rc_resend == NULL) || 4982 ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { 4983 /* 4984 * If the rwnd collapsed on 4985 * the one we are retransmitting 4986 * it does not count against the 4987 * rxt count. 4988 */ 4989 tp->t_rxtshift++; 4990 } 4991 if (tp->t_rxtshift > V_tcp_retries) { 4992 tp->t_rxtshift = V_tcp_retries; 4993 KMOD_TCPSTAT_INC(tcps_timeoutdrop); 4994 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); 4995 /* XXXGL: previously t_softerror was casted to uint16_t */ 4996 MPASS(tp->t_softerror >= 0); 4997 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; 4998 return (retval); /* tcp_drop() */ 4999 } 5000 if (tp->t_state == TCPS_SYN_SENT) { 5001 /* 5002 * If the SYN was retransmitted, indicate CWND to be limited 5003 * to 1 segment in cc_conn_init(). 5004 */ 5005 tp->snd_cwnd = 1; 5006 } else if (tp->t_rxtshift == 1) { 5007 /* 5008 * first retransmit; record ssthresh and cwnd so they can be 5009 * recovered if this turns out to be a "bad" retransmit. A 5010 * retransmit is considered "bad" if an ACK for this segment 5011 * is received within RTT/2 interval; the assumption here is 5012 * that the ACK was already in flight. See "On Estimating 5013 * End-to-End Network Path Properties" by Allman and Paxson 5014 * for more details. 5015 */ 5016 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5017 if (!IN_RECOVERY(tp->t_flags)) { 5018 tp->snd_cwnd_prev = tp->snd_cwnd; 5019 tp->snd_ssthresh_prev = tp->snd_ssthresh; 5020 tp->snd_recover_prev = tp->snd_recover; 5021 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 5022 tp->t_flags |= TF_PREVVALID; 5023 } else { 5024 tp->t_flags &= ~TF_PREVVALID; 5025 } 5026 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5027 } else { 5028 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; 5029 tp->t_flags &= ~TF_PREVVALID; 5030 } 5031 KMOD_TCPSTAT_INC(tcps_rexmttimeo); 5032 if ((tp->t_state == TCPS_SYN_SENT) || 5033 (tp->t_state == TCPS_SYN_RECEIVED)) 5034 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; 5035 else 5036 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; 5037 TCPT_RANGESET(tp->t_rxtcur, rexmt, 5038 MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), 5039 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 5040 /* 5041 * We enter the path for PLMTUD if connection is established or, if 5042 * connection is FIN_WAIT_1 status, reason for the last is that if 5043 * amount of data we send is very small, we could send it in couple 5044 * of packets and process straight to FIN. In that case we won't 5045 * catch ESTABLISHED state. 5046 */ 5047 #ifdef INET6 5048 isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false; 5049 #else 5050 isipv6 = false; 5051 #endif 5052 if (((V_tcp_pmtud_blackhole_detect == 1) || 5053 (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || 5054 (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && 5055 ((tp->t_state == TCPS_ESTABLISHED) || 5056 (tp->t_state == TCPS_FIN_WAIT_1))) { 5057 /* 5058 * Idea here is that at each stage of mtu probe (usually, 5059 * 1448 -> 1188 -> 524) should be given 2 chances to recover 5060 * before further clamping down. 'tp->t_rxtshift % 2 == 0' 5061 * should take care of that. 5062 */ 5063 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == 5064 (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && 5065 (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && 5066 tp->t_rxtshift % 2 == 0)) { 5067 /* 5068 * Enter Path MTU Black-hole Detection mechanism: - 5069 * Disable Path MTU Discovery (IP "DF" bit). - 5070 * Reduce MTU to lower value than what we negotiated 5071 * with peer. 5072 */ 5073 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { 5074 /* 5075 * Record that we may have found a black 5076 * hole. 5077 */ 5078 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; 5079 /* Keep track of previous MSS. */ 5080 tp->t_pmtud_saved_maxseg = tp->t_maxseg; 5081 } 5082 /* 5083 * Reduce the MSS to blackhole value or to the 5084 * default in an attempt to retransmit. 5085 */ 5086 #ifdef INET6 5087 isipv6 = bbr->r_is_v6; 5088 if (isipv6 && 5089 tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { 5090 /* Use the sysctl tuneable blackhole MSS. */ 5091 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; 5092 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5093 } else if (isipv6) { 5094 /* Use the default MSS. */ 5095 tp->t_maxseg = V_tcp_v6mssdflt; 5096 /* 5097 * Disable Path MTU Discovery when we switch 5098 * to minmss. 5099 */ 5100 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5101 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5102 } 5103 #endif 5104 #if defined(INET6) && defined(INET) 5105 else 5106 #endif 5107 #ifdef INET 5108 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { 5109 /* Use the sysctl tuneable blackhole MSS. */ 5110 tp->t_maxseg = V_tcp_pmtud_blackhole_mss; 5111 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); 5112 } else { 5113 /* Use the default MSS. */ 5114 tp->t_maxseg = V_tcp_mssdflt; 5115 /* 5116 * Disable Path MTU Discovery when we switch 5117 * to minmss. 5118 */ 5119 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 5120 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); 5121 } 5122 #endif 5123 } else { 5124 /* 5125 * If further retransmissions are still unsuccessful 5126 * with a lowered MTU, maybe this isn't a blackhole 5127 * and we restore the previous MSS and blackhole 5128 * detection flags. The limit '6' is determined by 5129 * giving each probe stage (1448, 1188, 524) 2 5130 * chances to recover. 5131 */ 5132 if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && 5133 (tp->t_rxtshift >= 6)) { 5134 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 5135 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; 5136 tp->t_maxseg = tp->t_pmtud_saved_maxseg; 5137 if (tp->t_maxseg < V_tcp_mssdflt) { 5138 /* 5139 * The MSS is so small we should not 5140 * process incoming SACK's since we are 5141 * subject to attack in such a case. 5142 */ 5143 tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT; 5144 } else { 5145 tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT; 5146 } 5147 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); 5148 } 5149 } 5150 } 5151 /* 5152 * Disable RFC1323 and SACK if we haven't got any response to our 5153 * third SYN to work-around some broken terminal servers (most of 5154 * which have hopefully been retired) that have bad VJ header 5155 * compression code which trashes TCP segments containing 5156 * unknown-to-them TCP options. 5157 */ 5158 if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && 5159 (tp->t_rxtshift == 3)) 5160 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); 5161 /* 5162 * If we backed off this far, our srtt estimate is probably bogus. 5163 * Clobber it so we'll take the next rtt measurement as our srtt; 5164 * move the current srtt into rttvar to keep the current retransmit 5165 * times until then. 5166 */ 5167 if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { 5168 #ifdef INET6 5169 if (bbr->r_is_v6) 5170 in6_losing(inp); 5171 else 5172 #endif 5173 in_losing(inp); 5174 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); 5175 tp->t_srtt = 0; 5176 } 5177 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 5178 tp->snd_recover = tp->snd_max; 5179 tp->t_flags |= TF_ACKNOW; 5180 tp->t_rtttime = 0; 5181 5182 return (retval); 5183 } 5184 5185 static int 5186 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) 5187 { 5188 int32_t ret = 0; 5189 int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); 5190 5191 if (timers == 0) { 5192 return (0); 5193 } 5194 if (tp->t_state == TCPS_LISTEN) { 5195 /* no timers on listen sockets */ 5196 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) 5197 return (0); 5198 return (1); 5199 } 5200 if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { 5201 uint32_t left; 5202 5203 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 5204 ret = -1; 5205 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5206 return (0); 5207 } 5208 if (hpts_calling == 0) { 5209 ret = -2; 5210 bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); 5211 return (0); 5212 } 5213 /* 5214 * Ok our timer went off early and we are not paced false 5215 * alarm, go back to sleep. 5216 */ 5217 left = bbr->r_ctl.rc_timer_exp - cts; 5218 ret = -3; 5219 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); 5220 tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(left)); 5221 return (1); 5222 } 5223 bbr->rc_tmr_stopped = 0; 5224 bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; 5225 if (timers & PACE_TMR_DELACK) { 5226 ret = bbr_timeout_delack(tp, bbr, cts); 5227 } else if (timers & PACE_TMR_PERSIT) { 5228 ret = bbr_timeout_persist(tp, bbr, cts); 5229 } else if (timers & PACE_TMR_RACK) { 5230 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5231 ret = bbr_timeout_rack(tp, bbr, cts); 5232 } else if (timers & PACE_TMR_TLP) { 5233 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5234 ret = bbr_timeout_tlp(tp, bbr, cts); 5235 } else if (timers & PACE_TMR_RXT) { 5236 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 5237 ret = bbr_timeout_rxt(tp, bbr, cts); 5238 } else if (timers & PACE_TMR_KEEP) { 5239 ret = bbr_timeout_keepalive(tp, bbr, cts); 5240 } 5241 bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); 5242 return (ret); 5243 } 5244 5245 static void 5246 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) 5247 { 5248 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 5249 uint8_t hpts_removed = 0; 5250 5251 if (tcp_in_hpts(bbr->rc_tp) && 5252 (bbr->rc_timer_first == 1)) { 5253 /* 5254 * If we are canceling timer's when we have the 5255 * timer ahead of the output being paced. We also 5256 * must remove ourselves from the hpts. 5257 */ 5258 hpts_removed = 1; 5259 tcp_hpts_remove(bbr->rc_tp); 5260 if (bbr->r_ctl.rc_last_delay_val) { 5261 /* Update the last hptsi delay too */ 5262 uint32_t time_since_send; 5263 5264 if (TSTMP_GT(cts, bbr->rc_pacer_started)) 5265 time_since_send = cts - bbr->rc_pacer_started; 5266 else 5267 time_since_send = 0; 5268 if (bbr->r_ctl.rc_last_delay_val > time_since_send) { 5269 /* Cut down our slot time */ 5270 bbr->r_ctl.rc_last_delay_val -= time_since_send; 5271 } else { 5272 bbr->r_ctl.rc_last_delay_val = 0; 5273 } 5274 bbr->rc_pacer_started = cts; 5275 } 5276 } 5277 bbr->rc_timer_first = 0; 5278 bbr_log_to_cancel(bbr, line, cts, hpts_removed); 5279 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; 5280 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); 5281 } 5282 } 5283 5284 static int 5285 bbr_stopall(struct tcpcb *tp) 5286 { 5287 struct tcp_bbr *bbr; 5288 5289 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 5290 bbr->rc_all_timers_stopped = 1; 5291 5292 tcp_hpts_remove(tp); 5293 5294 return (0); 5295 } 5296 5297 static uint32_t 5298 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) 5299 { 5300 struct bbr_sendmap *rsm; 5301 5302 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 5303 if ((rsm == NULL) || (u_rsm == rsm)) 5304 return (cts); 5305 return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); 5306 } 5307 5308 static void 5309 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, 5310 struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) 5311 { 5312 int32_t idx; 5313 5314 rsm->r_rtr_cnt++; 5315 rsm->r_dupack = 0; 5316 if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { 5317 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; 5318 rsm->r_flags |= BBR_OVERMAX; 5319 } 5320 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 5321 /* Take off the collapsed flag at rxt */ 5322 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 5323 } 5324 if (rsm->r_flags & BBR_MARKED_LOST) { 5325 /* We have retransmitted, its no longer lost */ 5326 rsm->r_flags &= ~BBR_MARKED_LOST; 5327 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 5328 } 5329 if (rsm->r_flags & BBR_RXT_CLEARED) { 5330 /* 5331 * We hit a RXT timer on it and 5332 * we cleared the "acked" flag. 5333 * We now have it going back into 5334 * flight, we can remove the cleared 5335 * flag and possibly do accounting on 5336 * this piece. 5337 */ 5338 rsm->r_flags &= ~BBR_RXT_CLEARED; 5339 } 5340 if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { 5341 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); 5342 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); 5343 } 5344 idx = rsm->r_rtr_cnt - 1; 5345 rsm->r_tim_lastsent[idx] = cts; 5346 rsm->r_pacing_delay = pacing_time; 5347 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5348 rsm->r_ts_valid = bbr->rc_ts_valid; 5349 if (bbr->rc_ts_valid) 5350 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5351 if (bbr->r_ctl.r_app_limited_until) 5352 rsm->r_app_limited = 1; 5353 else 5354 rsm->r_app_limited = 0; 5355 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5356 rsm->r_bbr_state = bbr_state_val(bbr); 5357 else 5358 rsm->r_bbr_state = 8; 5359 if (rsm->r_flags & BBR_ACKED) { 5360 /* Problably MTU discovery messing with us */ 5361 uint32_t old_flags; 5362 5363 old_flags = rsm->r_flags; 5364 rsm->r_flags &= ~BBR_ACKED; 5365 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); 5366 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 5367 if (bbr->r_ctl.rc_sacked == 0) 5368 bbr->r_ctl.rc_sacklast = NULL; 5369 } 5370 if (rsm->r_in_tmap) { 5371 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5372 } 5373 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5374 rsm->r_in_tmap = 1; 5375 if (rsm->r_flags & BBR_SACK_PASSED) { 5376 /* We have retransmitted due to the SACK pass */ 5377 rsm->r_flags &= ~BBR_SACK_PASSED; 5378 rsm->r_flags |= BBR_WAS_SACKPASS; 5379 } 5380 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5381 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5382 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5383 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 5384 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5385 rsm->r_is_gain = 1; 5386 rsm->r_is_drain = 0; 5387 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 5388 rsm->r_is_drain = 1; 5389 rsm->r_is_gain = 0; 5390 } else { 5391 rsm->r_is_drain = 0; 5392 rsm->r_is_gain = 0; 5393 } 5394 rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ 5395 } 5396 5397 /* 5398 * Returns 0, or the sequence where we stopped 5399 * updating. We also update the lenp to be the amount 5400 * of data left. 5401 */ 5402 5403 static uint32_t 5404 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, 5405 struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) 5406 { 5407 /* 5408 * We (re-)transmitted starting at rsm->r_start for some length 5409 * (possibly less than r_end. 5410 */ 5411 struct bbr_sendmap *nrsm; 5412 uint32_t c_end; 5413 int32_t len; 5414 5415 len = *lenp; 5416 c_end = rsm->r_start + len; 5417 if (SEQ_GEQ(c_end, rsm->r_end)) { 5418 /* 5419 * We retransmitted the whole piece or more than the whole 5420 * slopping into the next rsm. 5421 */ 5422 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5423 if (c_end == rsm->r_end) { 5424 *lenp = 0; 5425 return (0); 5426 } else { 5427 int32_t act_len; 5428 5429 /* Hangs over the end return whats left */ 5430 act_len = rsm->r_end - rsm->r_start; 5431 *lenp = (len - act_len); 5432 return (rsm->r_end); 5433 } 5434 /* We don't get out of this block. */ 5435 } 5436 /* 5437 * Here we retransmitted less than the whole thing which means we 5438 * have to split this into what was transmitted and what was not. 5439 */ 5440 nrsm = bbr_alloc_full_limit(bbr); 5441 if (nrsm == NULL) { 5442 *lenp = 0; 5443 return (0); 5444 } 5445 /* 5446 * So here we are going to take the original rsm and make it what we 5447 * retransmitted. nrsm will be the tail portion we did not 5448 * retransmit. For example say the chunk was 1, 11 (10 bytes). And 5449 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to 5450 * 1, 6 and the new piece will be 6, 11. 5451 */ 5452 bbr_clone_rsm(bbr, nrsm, rsm, c_end); 5453 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 5454 nrsm->r_dupack = 0; 5455 if (rsm->r_in_tmap) { 5456 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 5457 nrsm->r_in_tmap = 1; 5458 } 5459 rsm->r_flags &= (~BBR_HAS_FIN); 5460 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 5461 *lenp = 0; 5462 return (0); 5463 } 5464 5465 static uint64_t 5466 bbr_get_hardware_rate(struct tcp_bbr *bbr) 5467 { 5468 uint64_t bw; 5469 5470 bw = bbr_get_bw(bbr); 5471 bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; 5472 bw /= (uint64_t)BBR_UNIT; 5473 return(bw); 5474 } 5475 5476 static void 5477 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, 5478 uint64_t act_rate, uint64_t rate_wanted) 5479 { 5480 /* 5481 * We could not get a full gains worth 5482 * of rate. 5483 */ 5484 if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { 5485 /* we can't even get the real rate */ 5486 uint64_t red; 5487 5488 bbr->skip_gain = 1; 5489 bbr->gain_is_limited = 0; 5490 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; 5491 if (red) 5492 filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); 5493 } else { 5494 /* We can use a lower gain */ 5495 bbr->skip_gain = 0; 5496 bbr->gain_is_limited = 1; 5497 } 5498 } 5499 5500 static void 5501 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) 5502 { 5503 const struct tcp_hwrate_limit_table *nrte; 5504 int error, rate = -1; 5505 5506 if (bbr->r_ctl.crte == NULL) 5507 return; 5508 if ((bbr->rc_inp->inp_route.ro_nh == NULL) || 5509 (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { 5510 /* Lost our routes? */ 5511 /* Clear the way for a re-attempt */ 5512 bbr->bbr_attempt_hdwr_pace = 0; 5513 lost_rate: 5514 bbr->gain_is_limited = 0; 5515 bbr->skip_gain = 0; 5516 bbr->bbr_hdrw_pacing = 0; 5517 counter_u64_add(bbr_flows_whdwr_pacing, -1); 5518 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 5519 tcp_bbr_tso_size_check(bbr, cts); 5520 return; 5521 } 5522 rate = bbr_get_hardware_rate(bbr); 5523 nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, 5524 bbr->rc_tp, 5525 bbr->rc_inp->inp_route.ro_nh->nh_ifp, 5526 rate, 5527 (RS_PACING_GEQ|RS_PACING_SUB_OK), 5528 &error, NULL); 5529 if (nrte == NULL) { 5530 goto lost_rate; 5531 } 5532 if (nrte != bbr->r_ctl.crte) { 5533 bbr->r_ctl.crte = nrte; 5534 if (error == 0) { 5535 BBR_STAT_INC(bbr_hdwr_rl_mod_ok); 5536 if (bbr->r_ctl.crte->rate < rate) { 5537 /* We have a problem */ 5538 bbr_setup_less_of_rate(bbr, cts, 5539 bbr->r_ctl.crte->rate, rate); 5540 } else { 5541 /* We are good */ 5542 bbr->gain_is_limited = 0; 5543 bbr->skip_gain = 0; 5544 } 5545 } else { 5546 /* A failure should release the tag */ 5547 BBR_STAT_INC(bbr_hdwr_rl_mod_fail); 5548 bbr->gain_is_limited = 0; 5549 bbr->skip_gain = 0; 5550 bbr->bbr_hdrw_pacing = 0; 5551 } 5552 bbr_type_log_hdwr_pacing(bbr, 5553 bbr->r_ctl.crte->ptbl->rs_ifp, 5554 rate, 5555 ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), 5556 __LINE__, 5557 cts, 5558 error); 5559 } 5560 } 5561 5562 static void 5563 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) 5564 { 5565 /* 5566 * If we have hardware pacing support 5567 * we need to factor that in for our 5568 * TSO size. 5569 */ 5570 const struct tcp_hwrate_limit_table *rlp; 5571 uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; 5572 5573 if ((bbr->bbr_hdrw_pacing == 0) || 5574 (IN_RECOVERY(bbr->rc_tp->t_flags)) || 5575 (bbr->r_ctl.crte == NULL)) 5576 return; 5577 if (bbr->hw_pacing_set == 0) { 5578 /* Not yet by the hdwr pacing count delay */ 5579 return; 5580 } 5581 if (bbr_hdwr_pace_adjust == 0) { 5582 /* No adjustment */ 5583 return; 5584 } 5585 rlp = bbr->r_ctl.crte; 5586 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) 5587 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5588 else 5589 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5590 /* 5591 * So lets first get the 5592 * time we will take between 5593 * TSO sized sends currently without 5594 * hardware help. 5595 */ 5596 cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, 5597 bbr->r_ctl.rc_pace_max_segs, cts, 1); 5598 hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; 5599 hdwr_delay *= rlp->time_between; 5600 if (cur_delay > hdwr_delay) 5601 delta = cur_delay - hdwr_delay; 5602 else 5603 delta = 0; 5604 bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, 5605 (bbr->r_ctl.rc_pace_max_segs / maxseg), 5606 1); 5607 if (delta && 5608 (delta < (max(rlp->time_between, 5609 bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { 5610 /* 5611 * Now lets divide by the pacing 5612 * time between each segment the 5613 * hardware sends rounding up and 5614 * derive a bytes from that. We multiply 5615 * that by bbr_hdwr_pace_adjust to get 5616 * more bang for our buck. 5617 * 5618 * The goal is to have the software pacer 5619 * waiting no more than an additional 5620 * pacing delay if we can (without the 5621 * compensation i.e. x bbr_hdwr_pace_adjust). 5622 */ 5623 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), 5624 (bbr->r_ctl.rc_pace_max_segs/maxseg)); 5625 seg_sz *= bbr_hdwr_pace_adjust; 5626 if (bbr_hdwr_pace_floor && 5627 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5628 /* Currently hardware paces 5629 * out rs_min_seg segments at a time. 5630 * We need to make sure we always send at least 5631 * a full burst of bbr_hdwr_pace_floor down. 5632 */ 5633 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5634 } 5635 seg_sz *= maxseg; 5636 } else if (delta == 0) { 5637 /* 5638 * The highest pacing rate is 5639 * above our b/w gained. This means 5640 * we probably are going quite fast at 5641 * the hardware highest rate. Lets just multiply 5642 * the calculated TSO size by the 5643 * multiplier factor (its probably 5644 * 4 segments in the default config for 5645 * mlx). 5646 */ 5647 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; 5648 if (bbr_hdwr_pace_floor && 5649 (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { 5650 /* Currently hardware paces 5651 * out rs_min_seg segments at a time. 5652 * We need to make sure we always send at least 5653 * a full burst of bbr_hdwr_pace_floor down. 5654 */ 5655 seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; 5656 } 5657 } else { 5658 /* 5659 * The pacing time difference is so 5660 * big that the hardware will 5661 * pace out more rapidly then we 5662 * really want and then we 5663 * will have a long delay. Lets just keep 5664 * the same TSO size so its as if 5665 * we were not using hdwr pacing (we 5666 * just gain a bit of spacing from the 5667 * hardware if seg_sz > 1). 5668 */ 5669 seg_sz = bbr->r_ctl.rc_pace_max_segs; 5670 } 5671 if (seg_sz > bbr->r_ctl.rc_pace_max_segs) 5672 new_tso = seg_sz; 5673 else 5674 new_tso = bbr->r_ctl.rc_pace_max_segs; 5675 if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) 5676 new_tso = PACE_MAX_IP_BYTES - maxseg; 5677 5678 if (new_tso != bbr->r_ctl.rc_pace_max_segs) { 5679 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); 5680 bbr->r_ctl.rc_pace_max_segs = new_tso; 5681 } 5682 } 5683 5684 static void 5685 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) 5686 { 5687 uint64_t bw; 5688 uint32_t old_tso = 0, new_tso; 5689 uint32_t maxseg, bytes; 5690 uint32_t tls_seg=0; 5691 /* 5692 * Google/linux uses the following algorithm to determine 5693 * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): 5694 * 5695 * bytes = bw_in_bytes_per_second / 1000 5696 * bytes = min(bytes, 64k) 5697 * tso_segs = bytes / MSS 5698 * if (bw < 1.2Mbs) 5699 * min_tso_segs = 1 5700 * else 5701 * min_tso_segs = 2 5702 * tso_segs = max(tso_segs, min_tso_segs) 5703 * 5704 * * Note apply a device specific limit (we apply this in the 5705 * tcp_m_copym). 5706 * Note that before the initial measurement is made google bursts out 5707 * a full iwnd just like new-reno/cubic. 5708 * 5709 * We do not use this algorithm. Instead we 5710 * use a two phased approach: 5711 * 5712 * if ( bw <= per-tcb-cross-over) 5713 * goal_tso = calculate how much with this bw we 5714 * can send in goal-time seconds. 5715 * if (goal_tso > mss) 5716 * seg = goal_tso / mss 5717 * tso = seg * mss 5718 * else 5719 * tso = mss 5720 * if (tso > per-tcb-max) 5721 * tso = per-tcb-max 5722 * else if ( bw > 512Mbps) 5723 * tso = max-tso (64k/mss) 5724 * else 5725 * goal_tso = bw / per-tcb-divsor 5726 * seg = (goal_tso + mss-1)/mss 5727 * tso = seg * mss 5728 * 5729 * if (tso < per-tcb-floor) 5730 * tso = per-tcb-floor 5731 * if (tso > per-tcb-utter_max) 5732 * tso = per-tcb-utter_max 5733 * 5734 * Note the default per-tcb-divisor is 1000 (same as google). 5735 * the goal cross over is 30Mbps however. To recreate googles 5736 * algorithm you need to set: 5737 * 5738 * cross-over = 23,168,000 bps 5739 * goal-time = 18000 5740 * per-tcb-max = 2 5741 * per-tcb-divisor = 1000 5742 * per-tcb-floor = 1 5743 * 5744 * This will get you "google bbr" behavior with respect to tso size. 5745 * 5746 * Note we do set anything TSO size until we are past the initial 5747 * window. Before that we gnerally use either a single MSS 5748 * or we use the full IW size (so we burst a IW at a time) 5749 */ 5750 5751 if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { 5752 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5753 } else { 5754 maxseg = BBR_MIN_SEG - bbr->rc_last_options; 5755 } 5756 old_tso = bbr->r_ctl.rc_pace_max_segs; 5757 if (bbr->rc_past_init_win == 0) { 5758 /* 5759 * Not enough data has been acknowledged to make a 5760 * judgement. Set up the initial TSO based on if we 5761 * are sending a full IW at once or not. 5762 */ 5763 if (bbr->rc_use_google) 5764 bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); 5765 else if (bbr->bbr_init_win_cheat) 5766 bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); 5767 else 5768 bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 5769 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) 5770 bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; 5771 if (bbr->r_ctl.rc_pace_max_segs == 0) { 5772 bbr->r_ctl.rc_pace_max_segs = maxseg; 5773 } 5774 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); 5775 bbr_adjust_for_hw_pacing(bbr, cts); 5776 return; 5777 } 5778 /** 5779 * Now lets set the TSO goal based on our delivery rate in 5780 * bytes per second. Note we only do this if 5781 * we have acked at least the initial cwnd worth of data. 5782 */ 5783 bw = bbr_get_bw(bbr); 5784 if (IN_RECOVERY(bbr->rc_tp->t_flags) && 5785 (bbr->rc_use_google == 0)) { 5786 /* We clamp to one MSS in recovery */ 5787 new_tso = maxseg; 5788 } else if (bbr->rc_use_google) { 5789 int min_tso_segs; 5790 5791 /* Google considers the gain too */ 5792 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { 5793 bw *= bbr->r_ctl.rc_bbr_hptsi_gain; 5794 bw /= BBR_UNIT; 5795 } 5796 bytes = bw / 1024; 5797 if (bytes > (64 * 1024)) 5798 bytes = 64 * 1024; 5799 new_tso = bytes / maxseg; 5800 if (bw < ONE_POINT_TWO_MEG) 5801 min_tso_segs = 1; 5802 else 5803 min_tso_segs = 2; 5804 if (new_tso < min_tso_segs) 5805 new_tso = min_tso_segs; 5806 new_tso *= maxseg; 5807 } else if (bbr->rc_no_pacing) { 5808 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; 5809 } else if (bw <= bbr->r_ctl.bbr_cross_over) { 5810 /* 5811 * Calculate the worse case b/w TSO if we are inserting no 5812 * more than a delay_target number of TSO's. 5813 */ 5814 uint32_t tso_len, min_tso; 5815 5816 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); 5817 if (tso_len > maxseg) { 5818 new_tso = tso_len / maxseg; 5819 if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) 5820 new_tso = bbr->r_ctl.bbr_hptsi_segments_max; 5821 new_tso *= maxseg; 5822 } else { 5823 /* 5824 * less than a full sized frame yikes.. long rtt or 5825 * low bw? 5826 */ 5827 min_tso = bbr_minseg(bbr); 5828 if ((tso_len > min_tso) && (bbr_all_get_min == 0)) 5829 new_tso = rounddown(tso_len, min_tso); 5830 else 5831 new_tso = min_tso; 5832 } 5833 } else if (bw > FIVETWELVE_MBPS) { 5834 /* 5835 * This guy is so fast b/w wise that we can TSO as large as 5836 * possible of segments that the NIC will allow. 5837 */ 5838 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5839 } else { 5840 /* 5841 * This formula is based on attempting to send a segment or 5842 * more every bbr_hptsi_per_second. The default is 1000 5843 * which means you are targeting what you can send every 1ms 5844 * based on the peers bw. 5845 * 5846 * If the number drops to say 500, then you are looking more 5847 * at 2ms and you will raise how much we send in a single 5848 * TSO thus saving CPU (less bbr_output_wtime() calls). The 5849 * trade off of course is you will send more at once and 5850 * thus tend to clump up the sends into larger "bursts" 5851 * building a queue. 5852 */ 5853 bw /= bbr->r_ctl.bbr_hptsi_per_second; 5854 new_tso = roundup(bw, (uint64_t)maxseg); 5855 /* 5856 * Gate the floor to match what our lower than 48Mbps 5857 * algorithm does. The ceiling (bbr_hptsi_segments_max) thus 5858 * becomes the floor for this calculation. 5859 */ 5860 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) 5861 new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); 5862 } 5863 if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) 5864 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; 5865 if (new_tso > PACE_MAX_IP_BYTES) 5866 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); 5867 /* Enforce an utter maximum. */ 5868 if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { 5869 new_tso = bbr->r_ctl.bbr_utter_max * maxseg; 5870 } 5871 if (old_tso != new_tso) { 5872 /* Only log changes */ 5873 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); 5874 bbr->r_ctl.rc_pace_max_segs = new_tso; 5875 } 5876 /* We have hardware pacing! */ 5877 bbr_adjust_for_hw_pacing(bbr, cts); 5878 } 5879 5880 static void 5881 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, 5882 uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, 5883 struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, 5884 struct sockbuf *sb) 5885 { 5886 5887 struct bbr_sendmap *rsm, *nrsm; 5888 register uint32_t snd_max, snd_una; 5889 uint32_t pacing_time; 5890 /* 5891 * Add to the RACK log of packets in flight or retransmitted. If 5892 * there is a TS option we will use the TS echoed, if not we will 5893 * grab a TS. 5894 * 5895 * Retransmissions will increment the count and move the ts to its 5896 * proper place. Note that if options do not include TS's then we 5897 * won't be able to effectively use the ACK for an RTT on a retran. 5898 * 5899 * Notes about r_start and r_end. Lets consider a send starting at 5900 * sequence 1 for 10 bytes. In such an example the r_start would be 5901 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. 5902 * This means that r_end is actually the first sequence for the next 5903 * slot (11). 5904 * 5905 */ 5906 INP_WLOCK_ASSERT(tptoinpcb(tp)); 5907 if (err) { 5908 /* 5909 * We don't log errors -- we could but snd_max does not 5910 * advance in this case either. 5911 */ 5912 return; 5913 } 5914 if (th_flags & TH_RST) { 5915 /* 5916 * We don't log resets and we return immediately from 5917 * sending 5918 */ 5919 *abandon = 1; 5920 return; 5921 } 5922 snd_una = tp->snd_una; 5923 if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { 5924 /* 5925 * The call to bbr_log_output is made before bumping 5926 * snd_max. This means we can record one extra byte on a SYN 5927 * or FIN if seq_out is adding more on and a FIN is present 5928 * (and we are not resending). 5929 */ 5930 if ((th_flags & TH_SYN) && (tp->iss == seq_out)) 5931 len++; 5932 if (th_flags & TH_FIN) 5933 len++; 5934 } 5935 if (SEQ_LEQ((seq_out + len), snd_una)) { 5936 /* Are sending an old segment to induce an ack (keep-alive)? */ 5937 return; 5938 } 5939 if (SEQ_LT(seq_out, snd_una)) { 5940 /* huh? should we panic? */ 5941 uint32_t end; 5942 5943 end = seq_out + len; 5944 seq_out = snd_una; 5945 len = end - seq_out; 5946 } 5947 snd_max = tp->snd_max; 5948 if (len == 0) { 5949 /* We don't log zero window probes */ 5950 return; 5951 } 5952 pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); 5953 /* First question is it a retransmission? */ 5954 if (seq_out == snd_max) { 5955 again: 5956 rsm = bbr_alloc(bbr); 5957 if (rsm == NULL) { 5958 return; 5959 } 5960 rsm->r_flags = 0; 5961 if (th_flags & TH_SYN) 5962 rsm->r_flags |= BBR_HAS_SYN; 5963 if (th_flags & TH_FIN) 5964 rsm->r_flags |= BBR_HAS_FIN; 5965 rsm->r_tim_lastsent[0] = cts; 5966 rsm->r_rtr_cnt = 1; 5967 rsm->r_rtr_bytes = 0; 5968 rsm->r_start = seq_out; 5969 rsm->r_end = rsm->r_start + len; 5970 rsm->r_dupack = 0; 5971 rsm->r_delivered = bbr->r_ctl.rc_delivered; 5972 rsm->r_pacing_delay = pacing_time; 5973 rsm->r_ts_valid = bbr->rc_ts_valid; 5974 if (bbr->rc_ts_valid) 5975 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; 5976 rsm->r_del_time = bbr->r_ctl.rc_del_time; 5977 if (bbr->r_ctl.r_app_limited_until) 5978 rsm->r_app_limited = 1; 5979 else 5980 rsm->r_app_limited = 0; 5981 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); 5982 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, 5983 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 5984 /* 5985 * Here we must also add in this rsm since snd_max 5986 * is updated after we return from a new send. 5987 */ 5988 rsm->r_flight_at_send += len; 5989 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 5990 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 5991 rsm->r_in_tmap = 1; 5992 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 5993 rsm->r_bbr_state = bbr_state_val(bbr); 5994 else 5995 rsm->r_bbr_state = 8; 5996 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { 5997 rsm->r_is_gain = 1; 5998 rsm->r_is_drain = 0; 5999 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { 6000 rsm->r_is_drain = 1; 6001 rsm->r_is_gain = 0; 6002 } else { 6003 rsm->r_is_drain = 0; 6004 rsm->r_is_gain = 0; 6005 } 6006 return; 6007 } 6008 /* 6009 * If we reach here its a retransmission and we need to find it. 6010 */ 6011 more: 6012 if (hintrsm && (hintrsm->r_start == seq_out)) { 6013 rsm = hintrsm; 6014 hintrsm = NULL; 6015 } else if (bbr->r_ctl.rc_next) { 6016 /* We have a hint from a previous run */ 6017 rsm = bbr->r_ctl.rc_next; 6018 } else { 6019 /* No hints sorry */ 6020 rsm = NULL; 6021 } 6022 if ((rsm) && (rsm->r_start == seq_out)) { 6023 /* 6024 * We used rc_next or hintrsm to retransmit, hopefully the 6025 * likely case. 6026 */ 6027 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6028 if (len == 0) { 6029 return; 6030 } else { 6031 goto more; 6032 } 6033 } 6034 /* Ok it was not the last pointer go through it the hard way. */ 6035 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6036 if (rsm->r_start == seq_out) { 6037 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); 6038 bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); 6039 if (len == 0) { 6040 return; 6041 } else { 6042 continue; 6043 } 6044 } 6045 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { 6046 /* Transmitted within this piece */ 6047 /* 6048 * Ok we must split off the front and then let the 6049 * update do the rest 6050 */ 6051 nrsm = bbr_alloc_full_limit(bbr); 6052 if (nrsm == NULL) { 6053 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); 6054 return; 6055 } 6056 /* 6057 * copy rsm to nrsm and then trim the front of rsm 6058 * to not include this part. 6059 */ 6060 bbr_clone_rsm(bbr, nrsm, rsm, seq_out); 6061 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 6062 if (rsm->r_in_tmap) { 6063 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 6064 nrsm->r_in_tmap = 1; 6065 } 6066 rsm->r_flags &= (~BBR_HAS_FIN); 6067 seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); 6068 if (len == 0) { 6069 return; 6070 } 6071 } 6072 } 6073 /* 6074 * Hmm not found in map did they retransmit both old and on into the 6075 * new? 6076 */ 6077 if (seq_out == tp->snd_max) { 6078 goto again; 6079 } else if (SEQ_LT(seq_out, tp->snd_max)) { 6080 #ifdef BBR_INVARIANTS 6081 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", 6082 seq_out, len, tp->snd_una, tp->snd_max); 6083 printf("Starting Dump of all rack entries\n"); 6084 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 6085 printf("rsm:%p start:%u end:%u\n", 6086 rsm, rsm->r_start, rsm->r_end); 6087 } 6088 printf("Dump complete\n"); 6089 panic("seq_out not found rack:%p tp:%p", 6090 bbr, tp); 6091 #endif 6092 } else { 6093 #ifdef BBR_INVARIANTS 6094 /* 6095 * Hmm beyond sndmax? (only if we are using the new rtt-pack 6096 * flag) 6097 */ 6098 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", 6099 seq_out, len, tp->snd_max, tp); 6100 #endif 6101 } 6102 } 6103 6104 static void 6105 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) 6106 { 6107 /* 6108 * Collapse timeout back the cum-ack moved. 6109 */ 6110 tp->t_rxtshift = 0; 6111 tp->t_softerror = 0; 6112 } 6113 6114 static void 6115 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) 6116 { 6117 bbr->rtt_valid = 1; 6118 bbr->r_ctl.cur_rtt = rtt_usecs; 6119 bbr->r_ctl.ts_in = tsin; 6120 if (rsm_send_time) 6121 bbr->r_ctl.cur_rtt_send_time = rsm_send_time; 6122 } 6123 6124 static void 6125 bbr_make_timestamp_determination(struct tcp_bbr *bbr) 6126 { 6127 /** 6128 * We have in our bbr control: 6129 * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). 6130 * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). 6131 * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) 6132 * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) 6133 * 6134 * Now we can calculate the time between the sends by doing: 6135 * 6136 * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts 6137 * 6138 * And the peer's time between receiving them by doing: 6139 * 6140 * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp 6141 * 6142 * We want to figure out if the timestamp values are in msec, 10msec or usec. 6143 * We also may find that we can't use the timestamps if say we see 6144 * that the peer_delta indicates that though we may have taken 10ms to 6145 * pace out the data, it only saw 1ms between the two packets. This would 6146 * indicate that somewhere on the path is a batching entity that is giving 6147 * out time-slices of the actual b/w. This would mean we could not use 6148 * reliably the peers timestamps. 6149 * 6150 * We expect delta > peer_delta initially. Until we figure out the 6151 * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. 6152 * If we place 1000 there then its a ms vs our usec. If we place 10000 there 6153 * then its 10ms vs our usec. If the peer is running a usec clock we would 6154 * put a 1 there. If the value is faster then ours, we will disable the 6155 * use of timestamps (though we could revist this later if we find it to be not 6156 * just an isolated one or two flows)). 6157 * 6158 * To detect the batching middle boxes we will come up with our compensation and 6159 * if with it in place, we find the peer is drastically off (by some margin) in 6160 * the smaller direction, then we will assume the worst case and disable use of timestamps. 6161 * 6162 */ 6163 uint64_t delta, peer_delta, delta_up; 6164 6165 delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; 6166 if (delta < bbr_min_usec_delta) { 6167 /* 6168 * Have not seen a min amount of time 6169 * between our send times so we can 6170 * make a determination of the timestamp 6171 * yet. 6172 */ 6173 return; 6174 } 6175 peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; 6176 if (peer_delta < bbr_min_peer_delta) { 6177 /* 6178 * We may have enough in the form of 6179 * our delta but the peers number 6180 * has not changed that much. It could 6181 * be its clock ratio is such that 6182 * we need more data (10ms tick) or 6183 * there may be other compression scenarios 6184 * going on. In any event we need the 6185 * spread to be larger. 6186 */ 6187 return; 6188 } 6189 /* Ok lets first see which way our delta is going */ 6190 if (peer_delta > delta) { 6191 /* Very unlikely, the peer without 6192 * compensation shows that it saw 6193 * the two sends arrive further apart 6194 * then we saw then in micro-seconds. 6195 */ 6196 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { 6197 /* well it looks like the peer is a micro-second clock. */ 6198 bbr->rc_ts_clock_set = 1; 6199 bbr->r_ctl.bbr_peer_tsratio = 1; 6200 } else { 6201 bbr->rc_ts_cant_be_used = 1; 6202 bbr->rc_ts_clock_set = 1; 6203 } 6204 return; 6205 } 6206 /* Ok we know that the peer_delta is smaller than our send distance */ 6207 bbr->rc_ts_clock_set = 1; 6208 /* First question is it within the percentage that they are using usec time? */ 6209 delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; 6210 if ((peer_delta + delta_up) >= delta) { 6211 /* Its a usec clock */ 6212 bbr->r_ctl.bbr_peer_tsratio = 1; 6213 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6214 return; 6215 } 6216 /* Ok if not usec, what about 10usec (though unlikely)? */ 6217 delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; 6218 if (((peer_delta * 10) + delta_up) >= delta) { 6219 bbr->r_ctl.bbr_peer_tsratio = 10; 6220 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6221 return; 6222 } 6223 /* And what about 100usec (though again unlikely)? */ 6224 delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; 6225 if (((peer_delta * 100) + delta_up) >= delta) { 6226 bbr->r_ctl.bbr_peer_tsratio = 100; 6227 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6228 return; 6229 } 6230 /* And how about 1 msec (the most likely one)? */ 6231 delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; 6232 if (((peer_delta * 1000) + delta_up) >= delta) { 6233 bbr->r_ctl.bbr_peer_tsratio = 1000; 6234 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6235 return; 6236 } 6237 /* Ok if not msec could it be 10 msec? */ 6238 delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; 6239 if (((peer_delta * 10000) + delta_up) >= delta) { 6240 bbr->r_ctl.bbr_peer_tsratio = 10000; 6241 return; 6242 } 6243 /* If we fall down here the clock tick so slowly we can't use it */ 6244 bbr->rc_ts_cant_be_used = 1; 6245 bbr->r_ctl.bbr_peer_tsratio = 0; 6246 bbr_log_tstmp_validation(bbr, peer_delta, delta); 6247 } 6248 6249 /* 6250 * Collect new round-trip time estimate 6251 * and update averages and current timeout. 6252 */ 6253 static void 6254 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) 6255 { 6256 int32_t delta; 6257 uint32_t rtt, tsin; 6258 int32_t rtt_ticks; 6259 6260 if (bbr->rtt_valid == 0) 6261 /* No valid sample */ 6262 return; 6263 6264 rtt = bbr->r_ctl.cur_rtt; 6265 tsin = bbr->r_ctl.ts_in; 6266 if (bbr->rc_prtt_set_ts) { 6267 /* 6268 * We are to force feed the rttProp filter due 6269 * to an entry into PROBE_RTT. This assures 6270 * that the times are sync'd between when we 6271 * go into PROBE_RTT and the filter expiration. 6272 * 6273 * Google does not use a true filter, so they do 6274 * this implicitly since they only keep one value 6275 * and when they enter probe-rtt they update the 6276 * value to the newest rtt. 6277 */ 6278 uint32_t rtt_prop; 6279 6280 bbr->rc_prtt_set_ts = 0; 6281 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); 6282 if (rtt > rtt_prop) 6283 filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); 6284 else 6285 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6286 } 6287 #ifdef STATS 6288 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rtt)); 6289 #endif 6290 if (bbr->rc_ack_was_delayed) 6291 rtt += bbr->r_ctl.rc_ack_hdwr_delay; 6292 6293 if (rtt < bbr->r_ctl.rc_lowest_rtt) 6294 bbr->r_ctl.rc_lowest_rtt = rtt; 6295 bbr_log_rtt_sample(bbr, rtt, tsin); 6296 if (bbr->r_init_rtt) { 6297 /* 6298 * The initial rtt is not-trusted, nuke it and lets get 6299 * our first valid measurement in. 6300 */ 6301 bbr->r_init_rtt = 0; 6302 tp->t_srtt = 0; 6303 } 6304 if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { 6305 /* 6306 * So we have not yet figured out 6307 * what the peers TSTMP value is 6308 * in (most likely ms). We need a 6309 * series of cum-ack's to determine 6310 * this reliably. 6311 */ 6312 if (bbr->rc_ack_is_cumack) { 6313 if (bbr->rc_ts_data_set) { 6314 /* Lets attempt to determine the timestamp granularity. */ 6315 bbr_make_timestamp_determination(bbr); 6316 } else { 6317 bbr->rc_ts_data_set = 1; 6318 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; 6319 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; 6320 } 6321 } else { 6322 /* 6323 * We have to have consecutive acks 6324 * reset any "filled" state to none. 6325 */ 6326 bbr->rc_ts_data_set = 0; 6327 } 6328 } 6329 /* Round it up */ 6330 rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); 6331 if (rtt_ticks == 0) 6332 rtt_ticks = 1; 6333 if (tp->t_srtt != 0) { 6334 /* 6335 * srtt is stored as fixed point with 5 bits after the 6336 * binary point (i.e., scaled by 8). The following magic is 6337 * equivalent to the smoothing algorithm in rfc793 with an 6338 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). 6339 * Adjust rtt to origin 0. 6340 */ 6341 6342 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) 6343 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); 6344 6345 tp->t_srtt += delta; 6346 if (tp->t_srtt <= 0) 6347 tp->t_srtt = 1; 6348 6349 /* 6350 * We accumulate a smoothed rtt variance (actually, a 6351 * smoothed mean difference), then set the retransmit timer 6352 * to smoothed rtt + 4 times the smoothed variance. rttvar 6353 * is stored as fixed point with 4 bits after the binary 6354 * point (scaled by 16). The following is equivalent to 6355 * rfc793 smoothing with an alpha of .75 (rttvar = 6356 * rttvar*3/4 + |delta| / 4). This replaces rfc793's 6357 * wired-in beta. 6358 */ 6359 if (delta < 0) 6360 delta = -delta; 6361 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); 6362 tp->t_rttvar += delta; 6363 if (tp->t_rttvar <= 0) 6364 tp->t_rttvar = 1; 6365 } else { 6366 /* 6367 * No rtt measurement yet - use the unsmoothed rtt. Set the 6368 * variance to half the rtt (so our first retransmit happens 6369 * at 3*rtt). 6370 */ 6371 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; 6372 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); 6373 } 6374 KMOD_TCPSTAT_INC(tcps_rttupdated); 6375 if (tp->t_rttupdated < UCHAR_MAX) 6376 tp->t_rttupdated++; 6377 #ifdef STATS 6378 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); 6379 #endif 6380 /* 6381 * the retransmit should happen at rtt + 4 * rttvar. Because of the 6382 * way we do the smoothing, srtt and rttvar will each average +1/2 6383 * tick of bias. When we compute the retransmit timer, we want 1/2 6384 * tick of rounding and 1 extra tick because of +-1/2 tick 6385 * uncertainty in the firing of the timer. The bias will give us 6386 * exactly the 1.5 tick we need. But, because the bias is 6387 * statistical, we have to test that we don't drop below the minimum 6388 * feasible timer (which is 2 ticks). 6389 */ 6390 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), 6391 max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), 6392 MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); 6393 6394 /* 6395 * We received an ack for a packet that wasn't retransmitted; it is 6396 * probably safe to discard any error indications we've received 6397 * recently. This isn't quite right, but close enough for now (a 6398 * route might have failed after we sent a segment, and the return 6399 * path might not be symmetrical). 6400 */ 6401 tp->t_softerror = 0; 6402 rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); 6403 if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) 6404 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; 6405 } 6406 6407 static void 6408 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) 6409 { 6410 bbr->r_ctl.rc_rtt_shrinks = cts; 6411 if (bbr_can_force_probertt && 6412 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 6413 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 6414 /* 6415 * We should enter probe-rtt its been too long 6416 * since we have been there. 6417 */ 6418 bbr_enter_probe_rtt(bbr, cts, __LINE__); 6419 } else 6420 bbr_check_probe_rtt_limits(bbr, cts); 6421 } 6422 6423 static void 6424 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) 6425 { 6426 uint64_t orig_bw; 6427 6428 if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { 6429 /* We never apply a zero measurement */ 6430 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 6431 0, 0, 0, 0, 0, 0); 6432 return; 6433 } 6434 if (bbr->r_ctl.r_measurement_count < 0xffffffff) 6435 bbr->r_ctl.r_measurement_count++; 6436 orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); 6437 apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); 6438 bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, 6439 (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 6440 0, 0, 0, 0, 0, 0); 6441 if (orig_bw && 6442 (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { 6443 if (bbr->bbr_hdrw_pacing) { 6444 /* 6445 * Apply a new rate to the hardware 6446 * possibly. 6447 */ 6448 bbr_update_hardware_pacing_rate(bbr, cts); 6449 } 6450 bbr_set_state_target(bbr, __LINE__); 6451 tcp_bbr_tso_size_check(bbr, cts); 6452 if (bbr->r_recovery_bw) { 6453 bbr_setup_red_bw(bbr, cts); 6454 bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); 6455 } 6456 } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) 6457 tcp_bbr_tso_size_check(bbr, cts); 6458 } 6459 6460 static void 6461 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6462 { 6463 if (bbr->rc_in_persist == 0) { 6464 /* We log only when not in persist */ 6465 /* Translate to a Bytes Per Second */ 6466 uint64_t tim, bw, ts_diff, ts_bw; 6467 uint32_t delivered; 6468 6469 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6470 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6471 else 6472 tim = 1; 6473 /* 6474 * Now that we have processed the tim (skipping the sample 6475 * or possibly updating the time, go ahead and 6476 * calculate the cdr. 6477 */ 6478 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6479 bw = (uint64_t)delivered; 6480 bw *= (uint64_t)USECS_IN_SECOND; 6481 bw /= tim; 6482 if (bw == 0) { 6483 /* We must have a calculatable amount */ 6484 return; 6485 } 6486 /* 6487 * If we are using this b/w shove it in now so we 6488 * can see in the trace viewer if it gets over-ridden. 6489 */ 6490 if (rsm->r_ts_valid && 6491 bbr->rc_ts_valid && 6492 bbr->rc_ts_clock_set && 6493 (bbr->rc_ts_cant_be_used == 0) && 6494 bbr->rc_use_ts_limit) { 6495 ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); 6496 ts_diff *= bbr->r_ctl.bbr_peer_tsratio; 6497 if ((delivered == 0) || 6498 (rtt < 1000)) { 6499 /* Can't use the ts */ 6500 bbr_log_type_bbrupd(bbr, 61, cts, 6501 ts_diff, 6502 bbr->r_ctl.last_inbound_ts, 6503 rsm->r_del_ack_ts, 0, 6504 0, 0, 0, delivered); 6505 } else { 6506 ts_bw = (uint64_t)delivered; 6507 ts_bw *= (uint64_t)USECS_IN_SECOND; 6508 ts_bw /= ts_diff; 6509 bbr_log_type_bbrupd(bbr, 62, cts, 6510 (ts_bw >> 32), 6511 (ts_bw & 0xffffffff), 0, 0, 6512 0, 0, ts_diff, delivered); 6513 if ((bbr->ts_can_raise) && 6514 (ts_bw > bw)) { 6515 bbr_log_type_bbrupd(bbr, 8, cts, 6516 delivered, 6517 ts_diff, 6518 (bw >> 32), 6519 (bw & 0x00000000ffffffff), 6520 0, 0, 0, 0); 6521 bw = ts_bw; 6522 } else if (ts_bw && (ts_bw < bw)) { 6523 bbr_log_type_bbrupd(bbr, 7, cts, 6524 delivered, 6525 ts_diff, 6526 (bw >> 32), 6527 (bw & 0x00000000ffffffff), 6528 0, 0, 0, 0); 6529 bw = ts_bw; 6530 } 6531 } 6532 } 6533 if (rsm->r_first_sent_time && 6534 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6535 uint64_t sbw, sti; 6536 /* 6537 * We use what was in flight at the time of our 6538 * send and the size of this send to figure 6539 * out what we have been sending at (amount). 6540 * For the time we take from the time of 6541 * the send of the first send outstanding 6542 * until this send plus this sends pacing 6543 * time. This gives us a good calculation 6544 * as to the rate we have been sending at. 6545 */ 6546 6547 sbw = (uint64_t)(rsm->r_flight_at_send); 6548 sbw *= (uint64_t)USECS_IN_SECOND; 6549 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6550 sti += rsm->r_pacing_delay; 6551 sbw /= sti; 6552 if (sbw < bw) { 6553 bbr_log_type_bbrupd(bbr, 6, cts, 6554 delivered, 6555 (uint32_t)sti, 6556 (bw >> 32), 6557 (uint32_t)bw, 6558 rsm->r_first_sent_time, 0, (sbw >> 32), 6559 (uint32_t)sbw); 6560 bw = sbw; 6561 } 6562 } 6563 /* Use the google algorithm for b/w measurements */ 6564 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6565 if ((rsm->r_app_limited == 0) || 6566 (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { 6567 tcp_bbr_commit_bw(bbr, cts); 6568 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6569 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6570 } 6571 } 6572 } 6573 6574 static void 6575 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) 6576 { 6577 if (bbr->rc_in_persist == 0) { 6578 /* We log only when not in persist */ 6579 /* Translate to a Bytes Per Second */ 6580 uint64_t tim, bw; 6581 uint32_t delivered; 6582 int no_apply = 0; 6583 6584 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) 6585 tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); 6586 else 6587 tim = 1; 6588 /* 6589 * Now that we have processed the tim (skipping the sample 6590 * or possibly updating the time, go ahead and 6591 * calculate the cdr. 6592 */ 6593 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); 6594 bw = (uint64_t)delivered; 6595 bw *= (uint64_t)USECS_IN_SECOND; 6596 bw /= tim; 6597 if (tim < bbr->r_ctl.rc_lowest_rtt) { 6598 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6599 tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6600 6601 no_apply = 1; 6602 } 6603 /* 6604 * If we are using this b/w shove it in now so we 6605 * can see in the trace viewer if it gets over-ridden. 6606 */ 6607 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6608 /* Gate by the sending rate */ 6609 if (rsm->r_first_sent_time && 6610 TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { 6611 uint64_t sbw, sti; 6612 /* 6613 * We use what was in flight at the time of our 6614 * send and the size of this send to figure 6615 * out what we have been sending at (amount). 6616 * For the time we take from the time of 6617 * the send of the first send outstanding 6618 * until this send plus this sends pacing 6619 * time. This gives us a good calculation 6620 * as to the rate we have been sending at. 6621 */ 6622 6623 sbw = (uint64_t)(rsm->r_flight_at_send); 6624 sbw *= (uint64_t)USECS_IN_SECOND; 6625 sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; 6626 sti += rsm->r_pacing_delay; 6627 sbw /= sti; 6628 if (sbw < bw) { 6629 bbr_log_type_bbrupd(bbr, 6, cts, 6630 delivered, 6631 (uint32_t)sti, 6632 (bw >> 32), 6633 (uint32_t)bw, 6634 rsm->r_first_sent_time, 0, (sbw >> 32), 6635 (uint32_t)sbw); 6636 bw = sbw; 6637 } 6638 if ((sti > tim) && 6639 (sti < bbr->r_ctl.rc_lowest_rtt)) { 6640 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, 6641 (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); 6642 no_apply = 1; 6643 } else 6644 no_apply = 0; 6645 } 6646 bbr->r_ctl.rc_bbr_cur_del_rate = bw; 6647 if ((no_apply == 0) && 6648 ((rsm->r_app_limited == 0) || 6649 (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { 6650 tcp_bbr_commit_bw(bbr, cts); 6651 bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 6652 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); 6653 } 6654 } 6655 } 6656 6657 static void 6658 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, 6659 uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) 6660 { 6661 uint64_t old_rttprop; 6662 6663 /* Update our delivery time and amount */ 6664 bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); 6665 bbr->r_ctl.rc_del_time = cts; 6666 if (rtt == 0) { 6667 /* 6668 * 0 means its a retransmit, for now we don't use these for 6669 * the rest of BBR. 6670 */ 6671 return; 6672 } 6673 if ((bbr->rc_use_google == 0) && 6674 (match != BBR_RTT_BY_EXACTMATCH) && 6675 (match != BBR_RTT_BY_TIMESTAMP)){ 6676 /* 6677 * We get a lot of rtt updates, lets not pay attention to 6678 * any that are not an exact match. That way we don't have 6679 * to worry about timestamps and the whole nonsense of 6680 * unsure if its a retransmission etc (if we ever had the 6681 * timestamp fixed to always have the last thing sent this 6682 * would not be a issue). 6683 */ 6684 return; 6685 } 6686 if ((bbr_no_retran && bbr->rc_use_google) && 6687 (match != BBR_RTT_BY_EXACTMATCH) && 6688 (match != BBR_RTT_BY_TIMESTAMP)){ 6689 /* 6690 * We only do measurements in google mode 6691 * with bbr_no_retran on for sure things. 6692 */ 6693 return; 6694 } 6695 /* Only update srtt if we know by exact match */ 6696 tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); 6697 if (ack_type == BBR_CUM_ACKED) 6698 bbr->rc_ack_is_cumack = 1; 6699 else 6700 bbr->rc_ack_is_cumack = 0; 6701 old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); 6702 /* 6703 * Note the following code differs to the original 6704 * BBR spec. It calls for <= not <. However after a 6705 * long discussion in email with Neal, he acknowledged 6706 * that it should be < than so that we will have flows 6707 * going into probe-rtt (we were seeing cases where that 6708 * did not happen and caused ugly things to occur). We 6709 * have added this agreed upon fix to our code base. 6710 */ 6711 if (rtt < old_rttprop) { 6712 /* Update when we last saw a rtt drop */ 6713 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); 6714 bbr_set_reduced_rtt(bbr, cts, __LINE__); 6715 } 6716 bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, 6717 match, rsm->r_start, rsm->r_flags); 6718 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 6719 if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { 6720 /* 6721 * The RTT-prop moved, reset the target (may be a 6722 * nop for some states). 6723 */ 6724 bbr_set_state_target(bbr, __LINE__); 6725 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) 6726 bbr_log_rtt_shrinks(bbr, cts, 0, 0, 6727 __LINE__, BBR_RTTS_NEW_TARGET, 0); 6728 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) 6729 /* It went up */ 6730 bbr_check_probe_rtt_limits(bbr, cts); 6731 } 6732 if ((bbr->rc_use_google == 0) && 6733 (match == BBR_RTT_BY_TIMESTAMP)) { 6734 /* 6735 * We don't do b/w update with 6736 * these since they are not really 6737 * reliable. 6738 */ 6739 return; 6740 } 6741 if (bbr->r_ctl.r_app_limited_until && 6742 (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { 6743 /* We are no longer app-limited */ 6744 bbr->r_ctl.r_app_limited_until = 0; 6745 } 6746 if (bbr->rc_use_google) { 6747 bbr_google_measurement(bbr, rsm, rtt, cts); 6748 } else { 6749 bbr_nf_measurement(bbr, rsm, rtt, cts); 6750 } 6751 } 6752 6753 /* 6754 * Convert a timestamp that the main stack 6755 * uses (milliseconds) into one that bbr uses 6756 * (microseconds). Return that converted timestamp. 6757 */ 6758 static uint32_t 6759 bbr_ts_convert(uint32_t cts) { 6760 uint32_t sec, msec; 6761 6762 sec = cts / MS_IN_USEC; 6763 msec = cts - (MS_IN_USEC * sec); 6764 return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); 6765 } 6766 6767 /* 6768 * Return 0 if we did not update the RTT time, return 6769 * 1 if we did. 6770 */ 6771 static int 6772 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, 6773 struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) 6774 { 6775 int32_t i; 6776 uint32_t t, uts = 0; 6777 6778 if ((rsm->r_flags & BBR_ACKED) || 6779 (rsm->r_flags & BBR_WAS_RENEGED) || 6780 (rsm->r_flags & BBR_RXT_CLEARED)) { 6781 /* Already done */ 6782 return (0); 6783 } 6784 if (rsm->r_rtt_not_allowed) { 6785 /* Not allowed */ 6786 return (0); 6787 } 6788 if (rsm->r_rtr_cnt == 1) { 6789 /* 6790 * Only one transmit. Hopefully the normal case. 6791 */ 6792 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) 6793 t = cts - rsm->r_tim_lastsent[0]; 6794 else 6795 t = 1; 6796 if ((int)t <= 0) 6797 t = 1; 6798 bbr->r_ctl.rc_last_rtt = t; 6799 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6800 BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); 6801 return (1); 6802 } 6803 /* Convert to usecs */ 6804 if ((bbr_can_use_ts_for_rtt == 1) && 6805 (bbr->rc_use_google == 1) && 6806 (ack_type == BBR_CUM_ACKED) && 6807 (to->to_flags & TOF_TS) && 6808 (to->to_tsecr != 0)) { 6809 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; 6810 if (t < 1) 6811 t = 1; 6812 t *= MS_IN_USEC; 6813 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, 6814 BBR_RTT_BY_TIMESTAMP, 6815 rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], 6816 ack_type, to); 6817 return (1); 6818 } 6819 uts = bbr_ts_convert(to->to_tsecr); 6820 if ((to->to_flags & TOF_TS) && 6821 (to->to_tsecr != 0) && 6822 (ack_type == BBR_CUM_ACKED) && 6823 ((rsm->r_flags & BBR_OVERMAX) == 0)) { 6824 /* 6825 * Now which timestamp does it match? In this block the ACK 6826 * may be coming from a previous transmission. 6827 */ 6828 uint32_t fudge; 6829 6830 fudge = BBR_TIMER_FUDGE; 6831 for (i = 0; i < rsm->r_rtr_cnt; i++) { 6832 if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && 6833 (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { 6834 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6835 t = cts - rsm->r_tim_lastsent[i]; 6836 else 6837 t = 1; 6838 if ((int)t <= 0) 6839 t = 1; 6840 bbr->r_ctl.rc_last_rtt = t; 6841 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, 6842 rsm->r_tim_lastsent[i], ack_type, to); 6843 if ((i + 1) < rsm->r_rtr_cnt) { 6844 /* Likely */ 6845 return (0); 6846 } else if (rsm->r_flags & BBR_TLP) { 6847 bbr->rc_tlp_rtx_out = 0; 6848 } 6849 return (1); 6850 } 6851 } 6852 /* Fall through if we can't find a matching timestamp */ 6853 } 6854 /* 6855 * Ok its a SACK block that we retransmitted. or a windows 6856 * machine without timestamps. We can tell nothing from the 6857 * time-stamp since its not there or the time the peer last 6858 * received a segment that moved forward its cum-ack point. 6859 * 6860 * Lets look at the last retransmit and see what we can tell 6861 * (with BBR for space we only keep 2 note we have to keep 6862 * at least 2 so the map can not be condensed more). 6863 */ 6864 i = rsm->r_rtr_cnt - 1; 6865 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6866 t = cts - rsm->r_tim_lastsent[i]; 6867 else 6868 goto not_sure; 6869 if (t < bbr->r_ctl.rc_lowest_rtt) { 6870 /* 6871 * We retransmitted and the ack came back in less 6872 * than the smallest rtt we have observed in the 6873 * windowed rtt. We most likey did an improper 6874 * retransmit as outlined in 4.2 Step 3 point 2 in 6875 * the rack-draft. 6876 * 6877 * Use the prior transmission to update all the 6878 * information as long as there is only one prior 6879 * transmission. 6880 */ 6881 if ((rsm->r_flags & BBR_OVERMAX) == 0) { 6882 #ifdef BBR_INVARIANTS 6883 if (rsm->r_rtr_cnt == 1) 6884 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); 6885 #endif 6886 i = rsm->r_rtr_cnt - 2; 6887 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) 6888 t = cts - rsm->r_tim_lastsent[i]; 6889 else 6890 t = 1; 6891 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, 6892 rsm->r_tim_lastsent[i], ack_type, to); 6893 return (0); 6894 } else { 6895 /* 6896 * Too many prior transmissions, just 6897 * updated BBR delivered 6898 */ 6899 not_sure: 6900 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6901 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6902 } 6903 } else { 6904 /* 6905 * We retransmitted it and the retransmit did the 6906 * job. 6907 */ 6908 if (rsm->r_flags & BBR_TLP) 6909 bbr->rc_tlp_rtx_out = 0; 6910 if ((rsm->r_flags & BBR_OVERMAX) == 0) 6911 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, 6912 BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); 6913 else 6914 bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, 6915 BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); 6916 return (1); 6917 } 6918 return (0); 6919 } 6920 6921 /* 6922 * Mark the SACK_PASSED flag on all entries prior to rsm send wise. 6923 */ 6924 static void 6925 bbr_log_sack_passed(struct tcpcb *tp, 6926 struct tcp_bbr *bbr, struct bbr_sendmap *rsm) 6927 { 6928 struct bbr_sendmap *nrsm; 6929 6930 nrsm = rsm; 6931 TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, 6932 bbr_head, r_tnext) { 6933 if (nrsm == rsm) { 6934 /* Skip original segment he is acked */ 6935 continue; 6936 } 6937 if (nrsm->r_flags & BBR_ACKED) { 6938 /* Skip ack'd segments */ 6939 continue; 6940 } 6941 if (nrsm->r_flags & BBR_SACK_PASSED) { 6942 /* 6943 * We found one that is already marked 6944 * passed, we have been here before and 6945 * so all others below this are marked. 6946 */ 6947 break; 6948 } 6949 BBR_STAT_INC(bbr_sack_passed); 6950 nrsm->r_flags |= BBR_SACK_PASSED; 6951 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && 6952 bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { 6953 bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; 6954 bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; 6955 nrsm->r_flags |= BBR_MARKED_LOST; 6956 } 6957 nrsm->r_flags &= ~BBR_WAS_SACKPASS; 6958 } 6959 } 6960 6961 /* 6962 * Returns the number of bytes that were 6963 * newly ack'd by sack blocks. 6964 */ 6965 static uint32_t 6966 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, 6967 struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) 6968 { 6969 int32_t times = 0; 6970 uint32_t start, end, changed = 0; 6971 struct bbr_sendmap *rsm, *nrsm; 6972 int32_t used_ref = 1; 6973 uint8_t went_back = 0, went_fwd = 0; 6974 6975 start = sack->start; 6976 end = sack->end; 6977 rsm = *prsm; 6978 if (rsm == NULL) 6979 used_ref = 0; 6980 6981 /* Do we locate the block behind where we last were? */ 6982 if (rsm && SEQ_LT(start, rsm->r_start)) { 6983 went_back = 1; 6984 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 6985 if (SEQ_GEQ(start, rsm->r_start) && 6986 SEQ_LT(start, rsm->r_end)) { 6987 goto do_rest_ofb; 6988 } 6989 } 6990 } 6991 start_at_beginning: 6992 went_fwd = 1; 6993 /* 6994 * Ok lets locate the block where this guy is fwd from rsm (if its 6995 * set) 6996 */ 6997 TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { 6998 if (SEQ_GEQ(start, rsm->r_start) && 6999 SEQ_LT(start, rsm->r_end)) { 7000 break; 7001 } 7002 } 7003 do_rest_ofb: 7004 if (rsm == NULL) { 7005 /* 7006 * This happens when we get duplicate sack blocks with the 7007 * same end. For example SACK 4: 100 SACK 3: 100 The sort 7008 * will not change there location so we would just start at 7009 * the end of the first one and get lost. 7010 */ 7011 if (tp->t_flags & TF_SENTFIN) { 7012 /* 7013 * Check to see if we have not logged the FIN that 7014 * went out. 7015 */ 7016 nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7017 if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { 7018 /* 7019 * Ok we did not get the FIN logged. 7020 */ 7021 nrsm->r_end++; 7022 rsm = nrsm; 7023 goto do_rest_ofb; 7024 } 7025 } 7026 if (times == 1) { 7027 #ifdef BBR_INVARIANTS 7028 panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", 7029 tp, bbr, sack, to, prsm); 7030 #else 7031 goto out; 7032 #endif 7033 } 7034 times++; 7035 BBR_STAT_INC(bbr_sack_proc_restart); 7036 rsm = NULL; 7037 goto start_at_beginning; 7038 } 7039 /* Ok we have an ACK for some piece of rsm */ 7040 if (rsm->r_start != start) { 7041 /* 7042 * Need to split this in two pieces the before and after. 7043 */ 7044 if (bbr_sack_mergable(rsm, start, end)) 7045 nrsm = bbr_alloc_full_limit(bbr); 7046 else 7047 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7048 if (nrsm == NULL) { 7049 /* We could not allocate ignore the sack */ 7050 struct sackblk blk; 7051 7052 blk.start = start; 7053 blk.end = end; 7054 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7055 goto out; 7056 } 7057 bbr_clone_rsm(bbr, nrsm, rsm, start); 7058 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7059 if (rsm->r_in_tmap) { 7060 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7061 nrsm->r_in_tmap = 1; 7062 } 7063 rsm->r_flags &= (~BBR_HAS_FIN); 7064 rsm = nrsm; 7065 } 7066 if (SEQ_GEQ(end, rsm->r_end)) { 7067 /* 7068 * The end of this block is either beyond this guy or right 7069 * at this guy. 7070 */ 7071 if ((rsm->r_flags & BBR_ACKED) == 0) { 7072 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7073 changed += (rsm->r_end - rsm->r_start); 7074 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7075 bbr_log_sack_passed(tp, bbr, rsm); 7076 if (rsm->r_flags & BBR_MARKED_LOST) { 7077 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7078 } 7079 /* Is Reordering occuring? */ 7080 if (rsm->r_flags & BBR_SACK_PASSED) { 7081 BBR_STAT_INC(bbr_reorder_seen); 7082 bbr->r_ctl.rc_reorder_ts = cts; 7083 if (rsm->r_flags & BBR_MARKED_LOST) { 7084 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7085 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7086 /* LT sampling also needs adjustment */ 7087 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7088 } 7089 } 7090 rsm->r_flags |= BBR_ACKED; 7091 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7092 if (rsm->r_in_tmap) { 7093 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7094 rsm->r_in_tmap = 0; 7095 } 7096 } 7097 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7098 if (end == rsm->r_end) { 7099 /* This block only - done */ 7100 goto out; 7101 } 7102 /* There is more not coverend by this rsm move on */ 7103 start = rsm->r_end; 7104 nrsm = TAILQ_NEXT(rsm, r_next); 7105 rsm = nrsm; 7106 times = 0; 7107 goto do_rest_ofb; 7108 } 7109 if (rsm->r_flags & BBR_ACKED) { 7110 /* Been here done that */ 7111 goto out; 7112 } 7113 /* Ok we need to split off this one at the tail */ 7114 if (bbr_sack_mergable(rsm, start, end)) 7115 nrsm = bbr_alloc_full_limit(bbr); 7116 else 7117 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 7118 if (nrsm == NULL) { 7119 /* failed XXXrrs what can we do but loose the sack info? */ 7120 struct sackblk blk; 7121 7122 blk.start = start; 7123 blk.end = end; 7124 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); 7125 goto out; 7126 } 7127 /* Clone it */ 7128 bbr_clone_rsm(bbr, nrsm, rsm, end); 7129 /* The sack block does not cover this guy fully */ 7130 rsm->r_flags &= (~BBR_HAS_FIN); 7131 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 7132 if (rsm->r_in_tmap) { 7133 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 7134 nrsm->r_in_tmap = 1; 7135 } 7136 nrsm->r_dupack = 0; 7137 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); 7138 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); 7139 changed += (rsm->r_end - rsm->r_start); 7140 bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); 7141 bbr_log_sack_passed(tp, bbr, rsm); 7142 /* Is Reordering occuring? */ 7143 if (rsm->r_flags & BBR_MARKED_LOST) { 7144 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7145 } 7146 if (rsm->r_flags & BBR_SACK_PASSED) { 7147 BBR_STAT_INC(bbr_reorder_seen); 7148 bbr->r_ctl.rc_reorder_ts = cts; 7149 if (rsm->r_flags & BBR_MARKED_LOST) { 7150 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7151 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7152 /* LT sampling also needs adjustment */ 7153 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7154 } 7155 } 7156 rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); 7157 rsm->r_flags |= BBR_ACKED; 7158 if (rsm->r_in_tmap) { 7159 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7160 rsm->r_in_tmap = 0; 7161 } 7162 out: 7163 if (rsm && (rsm->r_flags & BBR_ACKED)) { 7164 /* 7165 * Now can we merge this newly acked 7166 * block with either the previous or 7167 * next block? 7168 */ 7169 nrsm = TAILQ_NEXT(rsm, r_next); 7170 if (nrsm && 7171 (nrsm->r_flags & BBR_ACKED)) { 7172 /* yep this and next can be merged */ 7173 rsm = bbr_merge_rsm(bbr, rsm, nrsm); 7174 } 7175 /* Now what about the previous? */ 7176 nrsm = TAILQ_PREV(rsm, bbr_head, r_next); 7177 if (nrsm && 7178 (nrsm->r_flags & BBR_ACKED)) { 7179 /* yep the previous and this can be merged */ 7180 rsm = bbr_merge_rsm(bbr, nrsm, rsm); 7181 } 7182 } 7183 if (used_ref == 0) { 7184 BBR_STAT_INC(bbr_sack_proc_all); 7185 } else { 7186 BBR_STAT_INC(bbr_sack_proc_short); 7187 } 7188 if (went_fwd && went_back) { 7189 BBR_STAT_INC(bbr_sack_search_both); 7190 } else if (went_fwd) { 7191 BBR_STAT_INC(bbr_sack_search_fwd); 7192 } else if (went_back) { 7193 BBR_STAT_INC(bbr_sack_search_back); 7194 } 7195 /* Save off where the next seq is */ 7196 if (rsm) 7197 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); 7198 else 7199 bbr->r_ctl.rc_sacklast = NULL; 7200 *prsm = rsm; 7201 return (changed); 7202 } 7203 7204 static void inline 7205 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) 7206 { 7207 struct bbr_sendmap *tmap; 7208 7209 BBR_STAT_INC(bbr_reneges_seen); 7210 tmap = NULL; 7211 while (rsm && (rsm->r_flags & BBR_ACKED)) { 7212 /* Its no longer sacked, mark it so */ 7213 uint32_t oflags; 7214 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7215 #ifdef BBR_INVARIANTS 7216 if (rsm->r_in_tmap) { 7217 panic("bbr:%p rsm:%p flags:0x%x in tmap?", 7218 bbr, rsm, rsm->r_flags); 7219 } 7220 #endif 7221 oflags = rsm->r_flags; 7222 if (rsm->r_flags & BBR_MARKED_LOST) { 7223 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7224 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7225 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7226 /* LT sampling also needs adjustment */ 7227 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7228 } 7229 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); 7230 rsm->r_flags |= BBR_WAS_RENEGED; 7231 rsm->r_flags |= BBR_RXT_CLEARED; 7232 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); 7233 /* Rebuild it into our tmap */ 7234 if (tmap == NULL) { 7235 TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7236 tmap = rsm; 7237 } else { 7238 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); 7239 tmap = rsm; 7240 } 7241 tmap->r_in_tmap = 1; 7242 /* 7243 * XXXrrs Delivered? Should we do anything here? 7244 * 7245 * Of course we don't on a rxt timeout so maybe its ok that 7246 * we don't? 7247 * 7248 * For now lets not. 7249 */ 7250 rsm = TAILQ_NEXT(rsm, r_next); 7251 } 7252 /* 7253 * Now lets possibly clear the sack filter so we start recognizing 7254 * sacks that cover this area. 7255 */ 7256 sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); 7257 } 7258 7259 static void 7260 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) 7261 { 7262 struct tcp_bbr *bbr; 7263 struct bbr_sendmap *rsm; 7264 uint32_t cts; 7265 7266 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7267 cts = bbr->r_ctl.rc_rcvtime; 7268 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7269 if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { 7270 if ((rsm->r_end - rsm->r_start) <= 1) { 7271 /* Log out the SYN completely */ 7272 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7273 rsm->r_rtr_bytes = 0; 7274 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7275 if (rsm->r_in_tmap) { 7276 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7277 rsm->r_in_tmap = 0; 7278 } 7279 if (bbr->r_ctl.rc_next == rsm) { 7280 /* scoot along the marker */ 7281 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7282 } 7283 if (to != NULL) 7284 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); 7285 bbr_free(bbr, rsm); 7286 } else { 7287 /* There is more (Fast open)? strip out SYN. */ 7288 rsm->r_flags &= ~BBR_HAS_SYN; 7289 rsm->r_start++; 7290 } 7291 } 7292 } 7293 7294 /* 7295 * Returns the number of bytes that were 7296 * acknowledged by SACK blocks. 7297 */ 7298 7299 static uint32_t 7300 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, 7301 uint32_t *prev_acked) 7302 { 7303 uint32_t changed, last_seq, entered_recovery = 0; 7304 struct tcp_bbr *bbr; 7305 struct bbr_sendmap *rsm; 7306 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; 7307 register uint32_t th_ack; 7308 int32_t i, j, k, new_sb, num_sack_blks = 0; 7309 uint32_t cts, acked, ack_point, sack_changed = 0; 7310 uint32_t p_maxseg, maxseg, p_acked = 0; 7311 7312 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7313 if (tcp_get_flags(th) & TH_RST) { 7314 /* We don't log resets */ 7315 return (0); 7316 } 7317 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7318 cts = bbr->r_ctl.rc_rcvtime; 7319 7320 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7321 changed = 0; 7322 maxseg = tp->t_maxseg - bbr->rc_last_options; 7323 p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); 7324 th_ack = th->th_ack; 7325 if (SEQ_GT(th_ack, tp->snd_una)) { 7326 acked = th_ack - tp->snd_una; 7327 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); 7328 bbr->rc_tp->t_acktime = ticks; 7329 } else 7330 acked = 0; 7331 if (SEQ_LEQ(th_ack, tp->snd_una)) { 7332 /* Only sent here for sack processing */ 7333 goto proc_sack; 7334 } 7335 if (rsm && SEQ_GT(th_ack, rsm->r_start)) { 7336 changed = th_ack - rsm->r_start; 7337 } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { 7338 /* 7339 * For the SYN incoming case we will not have called 7340 * tcp_output for the sending of the SYN, so there will be 7341 * no map. All other cases should probably be a panic. 7342 */ 7343 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { 7344 /* 7345 * We have a timestamp that can be used to generate 7346 * an initial RTT. 7347 */ 7348 uint32_t ts, now, rtt; 7349 7350 ts = bbr_ts_convert(to->to_tsecr); 7351 now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); 7352 rtt = now - ts; 7353 if (rtt < 1) 7354 rtt = 1; 7355 bbr_log_type_bbrrttprop(bbr, rtt, 7356 tp->iss, 0, cts, 7357 BBR_RTT_BY_TIMESTAMP, tp->iss, 0); 7358 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 7359 changed = 1; 7360 bbr->r_wanted_output = 1; 7361 goto out; 7362 } 7363 goto proc_sack; 7364 } else if (rsm == NULL) { 7365 goto out; 7366 } 7367 if (changed) { 7368 /* 7369 * The ACK point is advancing to th_ack, we must drop off 7370 * the packets in the rack log and calculate any eligble 7371 * RTT's. 7372 */ 7373 bbr->r_wanted_output = 1; 7374 more: 7375 if (rsm == NULL) { 7376 if (tp->t_flags & TF_SENTFIN) { 7377 /* if we send a FIN we will not hav a map */ 7378 goto proc_sack; 7379 } 7380 #ifdef BBR_INVARIANTS 7381 panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", 7382 tp, 7383 th, tp->t_state, bbr, 7384 tp->snd_una, tp->snd_max, changed); 7385 #endif 7386 goto proc_sack; 7387 } 7388 } 7389 if (SEQ_LT(th_ack, rsm->r_start)) { 7390 /* Huh map is missing this */ 7391 #ifdef BBR_INVARIANTS 7392 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", 7393 rsm->r_start, 7394 th_ack, tp->t_state, 7395 bbr->r_state, bbr); 7396 panic("th-ack is bad bbr:%p tp:%p", bbr, tp); 7397 #endif 7398 goto proc_sack; 7399 } else if (th_ack == rsm->r_start) { 7400 /* None here to ack */ 7401 goto proc_sack; 7402 } 7403 /* 7404 * Clear the dup ack counter, it will 7405 * either be freed or if there is some 7406 * remaining we need to start it at zero. 7407 */ 7408 rsm->r_dupack = 0; 7409 /* Now do we consume the whole thing? */ 7410 if (SEQ_GEQ(th_ack, rsm->r_end)) { 7411 /* Its all consumed. */ 7412 uint32_t left; 7413 7414 if (rsm->r_flags & BBR_ACKED) { 7415 /* 7416 * It was acked on the scoreboard -- remove it from 7417 * total 7418 */ 7419 p_acked += (rsm->r_end - rsm->r_start); 7420 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); 7421 if (bbr->r_ctl.rc_sacked == 0) 7422 bbr->r_ctl.rc_sacklast = NULL; 7423 } else { 7424 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); 7425 if (rsm->r_flags & BBR_MARKED_LOST) { 7426 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; 7427 } 7428 if (rsm->r_flags & BBR_SACK_PASSED) { 7429 /* 7430 * There are acked segments ACKED on the 7431 * scoreboard further up. We are seeing 7432 * reordering. 7433 */ 7434 BBR_STAT_INC(bbr_reorder_seen); 7435 bbr->r_ctl.rc_reorder_ts = cts; 7436 if (rsm->r_flags & BBR_MARKED_LOST) { 7437 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; 7438 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) 7439 /* LT sampling also needs adjustment */ 7440 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; 7441 } 7442 } 7443 rsm->r_flags &= ~BBR_MARKED_LOST; 7444 } 7445 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7446 rsm->r_rtr_bytes = 0; 7447 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 7448 if (rsm->r_in_tmap) { 7449 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 7450 rsm->r_in_tmap = 0; 7451 } 7452 if (bbr->r_ctl.rc_next == rsm) { 7453 /* scoot along the marker */ 7454 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7455 } 7456 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7457 /* Adjust the packet counts */ 7458 left = th_ack - rsm->r_end; 7459 /* Free back to zone */ 7460 bbr_free(bbr, rsm); 7461 if (left) { 7462 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7463 goto more; 7464 } 7465 goto proc_sack; 7466 } 7467 if (rsm->r_flags & BBR_ACKED) { 7468 /* 7469 * It was acked on the scoreboard -- remove it from total 7470 * for the part being cum-acked. 7471 */ 7472 p_acked += (rsm->r_end - rsm->r_start); 7473 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); 7474 if (bbr->r_ctl.rc_sacked == 0) 7475 bbr->r_ctl.rc_sacklast = NULL; 7476 } else { 7477 /* 7478 * It was acked up to th_ack point for the first time 7479 */ 7480 struct bbr_sendmap lrsm; 7481 7482 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); 7483 lrsm.r_end = th_ack; 7484 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); 7485 } 7486 if ((rsm->r_flags & BBR_MARKED_LOST) && 7487 ((rsm->r_flags & BBR_ACKED) == 0)) { 7488 /* 7489 * It was marked lost and partly ack'd now 7490 * for the first time. We lower the rc_lost_bytes 7491 * and still leave it MARKED. 7492 */ 7493 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; 7494 } 7495 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); 7496 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; 7497 rsm->r_rtr_bytes = 0; 7498 /* adjust packet count */ 7499 rsm->r_start = th_ack; 7500 proc_sack: 7501 /* Check for reneging */ 7502 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 7503 if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { 7504 /* 7505 * The peer has moved snd_una up to the edge of this send, 7506 * i.e. one that it had previously acked. The only way that 7507 * can be true if the peer threw away data (space issues) 7508 * that it had previously sacked (else it would have given 7509 * us snd_una up to (rsm->r_end). We need to undo the acked 7510 * markings here. 7511 * 7512 * Note we have to look to make sure th_ack is our 7513 * rsm->r_start in case we get an old ack where th_ack is 7514 * behind snd_una. 7515 */ 7516 bbr_peer_reneges(bbr, rsm, th->th_ack); 7517 } 7518 if ((to->to_flags & TOF_SACK) == 0) { 7519 /* We are done nothing left to log */ 7520 goto out; 7521 } 7522 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); 7523 if (rsm) { 7524 last_seq = rsm->r_end; 7525 } else { 7526 last_seq = tp->snd_max; 7527 } 7528 /* Sack block processing */ 7529 if (SEQ_GT(th_ack, tp->snd_una)) 7530 ack_point = th_ack; 7531 else 7532 ack_point = tp->snd_una; 7533 for (i = 0; i < to->to_nsacks; i++) { 7534 bcopy((to->to_sacks + i * TCPOLEN_SACK), 7535 &sack, sizeof(sack)); 7536 sack.start = ntohl(sack.start); 7537 sack.end = ntohl(sack.end); 7538 if (SEQ_GT(sack.end, sack.start) && 7539 SEQ_GT(sack.start, ack_point) && 7540 SEQ_LT(sack.start, tp->snd_max) && 7541 SEQ_GT(sack.end, ack_point) && 7542 SEQ_LEQ(sack.end, tp->snd_max)) { 7543 if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && 7544 (SEQ_LT(sack.end, last_seq)) && 7545 ((sack.end - sack.start) < (p_maxseg / 8))) { 7546 /* 7547 * Not the last piece and its smaller than 7548 * 1/8th of a p_maxseg. We ignore this. 7549 */ 7550 BBR_STAT_INC(bbr_runt_sacks); 7551 continue; 7552 } 7553 sack_blocks[num_sack_blks] = sack; 7554 num_sack_blks++; 7555 } else if (SEQ_LEQ(sack.start, th_ack) && 7556 SEQ_LEQ(sack.end, th_ack)) { 7557 /* 7558 * Its a D-SACK block. 7559 */ 7560 tcp_record_dsack(tp, sack.start, sack.end, 0); 7561 } 7562 } 7563 if (num_sack_blks == 0) 7564 goto out; 7565 /* 7566 * Sort the SACK blocks so we can update the rack scoreboard with 7567 * just one pass. 7568 */ 7569 new_sb = sack_filter_blks(tp, &bbr->r_ctl.bbr_sf, sack_blocks, 7570 num_sack_blks, th->th_ack); 7571 ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); 7572 BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); 7573 BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); 7574 num_sack_blks = new_sb; 7575 if (num_sack_blks < 2) { 7576 goto do_sack_work; 7577 } 7578 /* Sort the sacks */ 7579 for (i = 0; i < num_sack_blks; i++) { 7580 for (j = i + 1; j < num_sack_blks; j++) { 7581 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 7582 sack = sack_blocks[i]; 7583 sack_blocks[i] = sack_blocks[j]; 7584 sack_blocks[j] = sack; 7585 } 7586 } 7587 } 7588 /* 7589 * Now are any of the sack block ends the same (yes some 7590 * implememtations send these)? 7591 */ 7592 again: 7593 if (num_sack_blks > 1) { 7594 for (i = 0; i < num_sack_blks; i++) { 7595 for (j = i + 1; j < num_sack_blks; j++) { 7596 if (sack_blocks[i].end == sack_blocks[j].end) { 7597 /* 7598 * Ok these two have the same end we 7599 * want the smallest end and then 7600 * throw away the larger and start 7601 * again. 7602 */ 7603 if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { 7604 /* 7605 * The second block covers 7606 * more area use that 7607 */ 7608 sack_blocks[i].start = sack_blocks[j].start; 7609 } 7610 /* 7611 * Now collapse out the dup-sack and 7612 * lower the count 7613 */ 7614 for (k = (j + 1); k < num_sack_blks; k++) { 7615 sack_blocks[j].start = sack_blocks[k].start; 7616 sack_blocks[j].end = sack_blocks[k].end; 7617 j++; 7618 } 7619 num_sack_blks--; 7620 goto again; 7621 } 7622 } 7623 } 7624 } 7625 do_sack_work: 7626 rsm = bbr->r_ctl.rc_sacklast; 7627 for (i = 0; i < num_sack_blks; i++) { 7628 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); 7629 if (acked) { 7630 bbr->r_wanted_output = 1; 7631 changed += acked; 7632 sack_changed += acked; 7633 } 7634 } 7635 out: 7636 *prev_acked = p_acked; 7637 if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { 7638 /* 7639 * Ok we have a high probability that we need to go in to 7640 * recovery since we have data sack'd 7641 */ 7642 struct bbr_sendmap *rsm; 7643 7644 rsm = bbr_check_recovery_mode(tp, bbr, cts); 7645 if (rsm) { 7646 /* Enter recovery */ 7647 entered_recovery = 1; 7648 bbr->r_wanted_output = 1; 7649 /* 7650 * When we enter recovery we need to assure we send 7651 * one packet. 7652 */ 7653 if (bbr->r_ctl.rc_resend == NULL) { 7654 bbr->r_ctl.rc_resend = rsm; 7655 } 7656 } 7657 } 7658 if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { 7659 /* 7660 * See if we need to rack-retransmit anything if so set it 7661 * up as the thing to resend assuming something else is not 7662 * already in that position. 7663 */ 7664 if (bbr->r_ctl.rc_resend == NULL) { 7665 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 7666 } 7667 } 7668 /* 7669 * We return the amount that changed via sack, this is used by the 7670 * ack-received code to augment what was changed between th_ack <-> 7671 * snd_una. 7672 */ 7673 return (sack_changed); 7674 } 7675 7676 static void 7677 bbr_strike_dupack(struct tcp_bbr *bbr) 7678 { 7679 struct bbr_sendmap *rsm; 7680 7681 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); 7682 if (rsm && (rsm->r_dupack < 0xff)) { 7683 rsm->r_dupack++; 7684 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) 7685 bbr->r_wanted_output = 1; 7686 } 7687 } 7688 7689 /* 7690 * Return value of 1, we do not need to call bbr_process_data(). 7691 * return value of 0, bbr_process_data can be called. 7692 * For ret_val if its 0 the TCB is locked and valid, if its non-zero 7693 * its unlocked and probably unsafe to touch the TCB. 7694 */ 7695 static int 7696 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, 7697 struct tcpcb *tp, struct tcpopt *to, 7698 uint32_t tiwin, int32_t tlen, 7699 int32_t * ofia, int32_t thflags, int32_t * ret_val) 7700 { 7701 int32_t ourfinisacked = 0; 7702 int32_t acked_amount; 7703 uint16_t nsegs; 7704 int32_t acked; 7705 uint32_t lost, sack_changed = 0; 7706 struct mbuf *mfree; 7707 struct tcp_bbr *bbr; 7708 uint32_t prev_acked = 0; 7709 7710 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 7711 lost = bbr->r_ctl.rc_lost; 7712 nsegs = max(1, m->m_pkthdr.lro_nsegs); 7713 if (SEQ_GT(th->th_ack, tp->snd_max)) { 7714 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); 7715 bbr->r_wanted_output = 1; 7716 return (1); 7717 } 7718 if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { 7719 /* Process the ack */ 7720 if (bbr->rc_in_persist) 7721 tp->t_rxtshift = 0; 7722 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) 7723 bbr_strike_dupack(bbr); 7724 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 7725 } 7726 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); 7727 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 7728 /* 7729 * Old ack, behind the last one rcv'd or a duplicate ack 7730 * with SACK info. 7731 */ 7732 if (th->th_ack == tp->snd_una) { 7733 bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); 7734 if (bbr->r_state == TCPS_SYN_SENT) { 7735 /* 7736 * Special case on where we sent SYN. When 7737 * the SYN-ACK is processed in syn_sent 7738 * state it bumps the snd_una. This causes 7739 * us to hit here even though we did ack 1 7740 * byte. 7741 * 7742 * Go through the nothing left case so we 7743 * send data. 7744 */ 7745 goto nothing_left; 7746 } 7747 } 7748 return (0); 7749 } 7750 /* 7751 * If we reach this point, ACK is not a duplicate, i.e., it ACKs 7752 * something we sent. 7753 */ 7754 if (tp->t_flags & TF_NEEDSYN) { 7755 /* 7756 * T/TCP: Connection was half-synchronized, and our SYN has 7757 * been ACK'd (so connection is now fully synchronized). Go 7758 * to non-starred state, increment snd_una for ACK of SYN, 7759 * and check if we can do window scaling. 7760 */ 7761 tp->t_flags &= ~TF_NEEDSYN; 7762 tp->snd_una++; 7763 /* Do window scaling? */ 7764 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 7765 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 7766 tp->rcv_scale = tp->request_r_scale; 7767 /* Send window already scaled. */ 7768 } 7769 } 7770 INP_WLOCK_ASSERT(tptoinpcb(tp)); 7771 7772 acked = BYTES_THIS_ACK(tp, th); 7773 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 7774 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 7775 7776 /* 7777 * If we just performed our first retransmit, and the ACK arrives 7778 * within our recovery window, then it was a mistake to do the 7779 * retransmit in the first place. Recover our original cwnd and 7780 * ssthresh, and proceed to transmit where we left off. 7781 */ 7782 if (tp->t_flags & TF_PREVVALID) { 7783 tp->t_flags &= ~TF_PREVVALID; 7784 if (tp->t_rxtshift == 1 && 7785 (int)(ticks - tp->t_badrxtwin) < 0) 7786 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 7787 } 7788 SOCKBUF_LOCK(&so->so_snd); 7789 acked_amount = min(acked, (int)sbavail(&so->so_snd)); 7790 tp->snd_wnd -= acked_amount; 7791 mfree = sbcut_locked(&so->so_snd, acked_amount); 7792 /* NB: sowwakeup_locked() does an implicit unlock. */ 7793 sowwakeup_locked(so); 7794 m_freem(mfree); 7795 if (SEQ_GT(th->th_ack, tp->snd_una)) { 7796 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 7797 } 7798 tp->snd_una = th->th_ack; 7799 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); 7800 if (IN_RECOVERY(tp->t_flags)) { 7801 if (SEQ_LT(th->th_ack, tp->snd_recover) && 7802 (SEQ_LT(th->th_ack, tp->snd_max))) { 7803 tcp_bbr_partialack(tp); 7804 } else { 7805 bbr_post_recovery(tp); 7806 } 7807 } 7808 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 7809 tp->snd_recover = tp->snd_una; 7810 } 7811 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 7812 tp->snd_nxt = tp->snd_max; 7813 } 7814 if (tp->snd_una == tp->snd_max) { 7815 /* Nothing left outstanding */ 7816 nothing_left: 7817 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 7818 if (sbavail(&so->so_snd) == 0) 7819 bbr->rc_tp->t_acktime = 0; 7820 if ((sbused(&so->so_snd) == 0) && 7821 (tp->t_flags & TF_SENTFIN)) { 7822 ourfinisacked = 1; 7823 } 7824 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 7825 if (bbr->rc_in_persist == 0) { 7826 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 7827 } 7828 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 7829 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 7830 /* 7831 * We invalidate the last ack here since we 7832 * don't want to transfer forward the time 7833 * for our sum's calculations. 7834 */ 7835 if ((tp->t_state >= TCPS_FIN_WAIT_1) && 7836 (sbavail(&so->so_snd) == 0) && 7837 (tp->t_flags2 & TF2_DROP_AF_DATA)) { 7838 /* 7839 * The socket was gone and the peer sent data, time 7840 * to reset him. 7841 */ 7842 *ret_val = 1; 7843 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 7844 /* tcp_close will kill the inp pre-log the Reset */ 7845 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 7846 tp = tcp_close(tp); 7847 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); 7848 BBR_STAT_INC(bbr_dropped_af_data); 7849 return (1); 7850 } 7851 /* Set need output so persist might get set */ 7852 bbr->r_wanted_output = 1; 7853 } 7854 if (ofia) 7855 *ofia = ourfinisacked; 7856 return (0); 7857 } 7858 7859 static void 7860 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7861 { 7862 if (bbr->rc_in_persist == 0) { 7863 bbr_timer_cancel(bbr, __LINE__, cts); 7864 bbr->r_ctl.rc_last_delay_val = 0; 7865 tp->t_rxtshift = 0; 7866 bbr->rc_in_persist = 1; 7867 bbr->r_ctl.rc_went_idle_time = cts; 7868 /* We should be capped when rw went to 0 but just in case */ 7869 bbr_log_type_pesist(bbr, cts, 0, line, 1); 7870 /* Time freezes for the state, so do the accounting now */ 7871 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 7872 uint32_t time_in; 7873 7874 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 7875 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7876 int32_t idx; 7877 7878 idx = bbr_state_val(bbr); 7879 counter_u64_add(bbr_state_time[(idx + 5)], time_in); 7880 } else { 7881 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 7882 } 7883 } 7884 bbr->r_ctl.rc_bbr_state_time = cts; 7885 } 7886 } 7887 7888 static void 7889 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) 7890 { 7891 /* 7892 * Note that if idle time does not exceed our 7893 * threshold, we do nothing continuing the state 7894 * transitions we were last walking through. 7895 */ 7896 if (idle_time >= bbr_idle_restart_threshold) { 7897 if (bbr->rc_use_idle_restart) { 7898 bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; 7899 /* 7900 * Set our target using BBR_UNIT, so 7901 * we increase at a dramatic rate but 7902 * we stop when we get the pipe 7903 * full again for our current b/w estimate. 7904 */ 7905 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 7906 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 7907 bbr_set_state_target(bbr, __LINE__); 7908 /* Now setup our gains to ramp up */ 7909 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 7910 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 7911 bbr_log_type_statechange(bbr, cts, __LINE__); 7912 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 7913 bbr_substate_change(bbr, cts, __LINE__, 1); 7914 } 7915 } 7916 } 7917 7918 static void 7919 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) 7920 { 7921 uint32_t idle_time; 7922 7923 if (bbr->rc_in_persist == 0) 7924 return; 7925 idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); 7926 bbr->rc_in_persist = 0; 7927 bbr->rc_hit_state_1 = 0; 7928 bbr->r_ctl.rc_del_time = cts; 7929 /* 7930 * We invalidate the last ack here since we 7931 * don't want to transfer forward the time 7932 * for our sum's calculations. 7933 */ 7934 if (tcp_in_hpts(bbr->rc_tp)) { 7935 tcp_hpts_remove(bbr->rc_tp); 7936 bbr->rc_timer_first = 0; 7937 bbr->r_ctl.rc_hpts_flags = 0; 7938 bbr->r_ctl.rc_last_delay_val = 0; 7939 bbr->r_ctl.rc_hptsi_agg_delay = 0; 7940 bbr->r_agg_early_set = 0; 7941 bbr->r_ctl.rc_agg_early = 0; 7942 } 7943 bbr_log_type_pesist(bbr, cts, idle_time, line, 0); 7944 if (idle_time >= bbr_rtt_probe_time) { 7945 /* 7946 * This qualifies as a RTT_PROBE session since we drop the 7947 * data outstanding to nothing and waited more than 7948 * bbr_rtt_probe_time. 7949 */ 7950 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); 7951 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; 7952 } 7953 tp->t_rxtshift = 0; 7954 /* 7955 * If in probeBW and we have persisted more than an RTT lets do 7956 * special handling. 7957 */ 7958 /* Force a time based epoch */ 7959 bbr_set_epoch(bbr, cts, __LINE__); 7960 /* 7961 * Setup the lost so we don't count anything against the guy 7962 * we have been stuck with during persists. 7963 */ 7964 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 7965 /* Time un-freezes for the state */ 7966 bbr->r_ctl.rc_bbr_state_time = cts; 7967 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || 7968 (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { 7969 /* 7970 * If we are going back to probe-bw 7971 * or probe_rtt, we may need to possibly 7972 * do a fast restart. 7973 */ 7974 bbr_restart_after_idle(bbr, cts, idle_time); 7975 } 7976 } 7977 7978 static void 7979 bbr_collapsed_window(struct tcp_bbr *bbr) 7980 { 7981 /* 7982 * Now we must walk the 7983 * send map and divide the 7984 * ones left stranded. These 7985 * guys can't cause us to abort 7986 * the connection and are really 7987 * "unsent". However if a buggy 7988 * client actually did keep some 7989 * of the data i.e. collapsed the win 7990 * and refused to ack and then opened 7991 * the win and acked that data. We would 7992 * get into an ack war, the simplier 7993 * method then of just pretending we 7994 * did not send those segments something 7995 * won't work. 7996 */ 7997 struct bbr_sendmap *rsm, *nrsm; 7998 tcp_seq max_seq; 7999 uint32_t maxseg; 8000 int can_split = 0; 8001 int fnd = 0; 8002 8003 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; 8004 max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; 8005 bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); 8006 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 8007 /* Find the first seq past or at maxseq */ 8008 if (rsm->r_flags & BBR_RWND_COLLAPSED) 8009 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8010 if (SEQ_GEQ(max_seq, rsm->r_start) && 8011 SEQ_GEQ(rsm->r_end, max_seq)) { 8012 fnd = 1; 8013 break; 8014 } 8015 } 8016 bbr->rc_has_collapsed = 0; 8017 if (!fnd) { 8018 /* Nothing to do strange */ 8019 return; 8020 } 8021 /* 8022 * Now can we split? 8023 * 8024 * We don't want to split if splitting 8025 * would generate too many small segments 8026 * less we let an attacker fragment our 8027 * send_map and leave us out of memory. 8028 */ 8029 if ((max_seq != rsm->r_start) && 8030 (max_seq != rsm->r_end)){ 8031 /* can we split? */ 8032 int res1, res2; 8033 8034 res1 = max_seq - rsm->r_start; 8035 res2 = rsm->r_end - max_seq; 8036 if ((res1 >= (maxseg/8)) && 8037 (res2 >= (maxseg/8))) { 8038 /* No small pieces here */ 8039 can_split = 1; 8040 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { 8041 /* We are under the limit */ 8042 can_split = 1; 8043 } 8044 } 8045 /* Ok do we need to split this rsm? */ 8046 if (max_seq == rsm->r_start) { 8047 /* It's this guy no split required */ 8048 nrsm = rsm; 8049 } else if (max_seq == rsm->r_end) { 8050 /* It's the next one no split required. */ 8051 nrsm = TAILQ_NEXT(rsm, r_next); 8052 if (nrsm == NULL) { 8053 /* Huh? */ 8054 return; 8055 } 8056 } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { 8057 /* yep we need to split it */ 8058 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); 8059 if (nrsm == NULL) { 8060 /* failed XXXrrs what can we do mark the whole? */ 8061 nrsm = rsm; 8062 goto no_split; 8063 } 8064 /* Clone it */ 8065 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); 8066 bbr_clone_rsm(bbr, nrsm, rsm, max_seq); 8067 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); 8068 if (rsm->r_in_tmap) { 8069 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); 8070 nrsm->r_in_tmap = 1; 8071 } 8072 } else { 8073 /* 8074 * Split not allowed just start here just 8075 * use this guy. 8076 */ 8077 nrsm = rsm; 8078 } 8079 no_split: 8080 BBR_STAT_INC(bbr_collapsed_win); 8081 /* reuse fnd as a count */ 8082 fnd = 0; 8083 TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { 8084 nrsm->r_flags |= BBR_RWND_COLLAPSED; 8085 fnd++; 8086 bbr->rc_has_collapsed = 1; 8087 } 8088 bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); 8089 } 8090 8091 static void 8092 bbr_un_collapse_window(struct tcp_bbr *bbr) 8093 { 8094 struct bbr_sendmap *rsm; 8095 int cleared = 0; 8096 8097 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { 8098 if (rsm->r_flags & BBR_RWND_COLLAPSED) { 8099 /* Clear the flag */ 8100 rsm->r_flags &= ~BBR_RWND_COLLAPSED; 8101 cleared++; 8102 } else 8103 break; 8104 } 8105 bbr_log_type_rwnd_collapse(bbr, 8106 (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); 8107 bbr->rc_has_collapsed = 0; 8108 } 8109 8110 /* 8111 * Return value of 1, the TCB is unlocked and most 8112 * likely gone, return value of 0, the TCB is still 8113 * locked. 8114 */ 8115 static int 8116 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, 8117 struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, 8118 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) 8119 { 8120 /* 8121 * Update window information. Don't look at window if no ACK: TAC's 8122 * send garbage on first SYN. 8123 */ 8124 uint16_t nsegs; 8125 int32_t tfo_syn; 8126 struct tcp_bbr *bbr; 8127 8128 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8129 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8130 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8131 if ((thflags & TH_ACK) && 8132 (SEQ_LT(tp->snd_wl1, th->th_seq) || 8133 (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || 8134 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { 8135 /* keep track of pure window updates */ 8136 if (tlen == 0 && 8137 tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) 8138 KMOD_TCPSTAT_INC(tcps_rcvwinupd); 8139 tp->snd_wnd = tiwin; 8140 tp->snd_wl1 = th->th_seq; 8141 tp->snd_wl2 = th->th_ack; 8142 if (tp->snd_wnd > tp->max_sndwnd) 8143 tp->max_sndwnd = tp->snd_wnd; 8144 bbr->r_wanted_output = 1; 8145 } else if (thflags & TH_ACK) { 8146 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { 8147 tp->snd_wnd = tiwin; 8148 tp->snd_wl1 = th->th_seq; 8149 tp->snd_wl2 = th->th_ack; 8150 } 8151 } 8152 if (tp->snd_wnd < ctf_outstanding(tp)) 8153 /* The peer collapsed its window on us */ 8154 bbr_collapsed_window(bbr); 8155 else if (bbr->rc_has_collapsed) 8156 bbr_un_collapse_window(bbr); 8157 /* Was persist timer active and now we have window space? */ 8158 if ((bbr->rc_in_persist != 0) && 8159 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8160 bbr_minseg(bbr)))) { 8161 /* 8162 * Make the rate persist at end of persist mode if idle long 8163 * enough 8164 */ 8165 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8166 8167 /* Make sure we output to start the timer */ 8168 bbr->r_wanted_output = 1; 8169 } 8170 /* Do we need to enter persist? */ 8171 if ((bbr->rc_in_persist == 0) && 8172 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8173 TCPS_HAVEESTABLISHED(tp->t_state) && 8174 (tp->snd_max == tp->snd_una) && 8175 sbavail(&so->so_snd) && 8176 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8177 /* No send window.. we must enter persist */ 8178 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8179 } 8180 if (tp->t_flags2 & TF2_DROP_AF_DATA) { 8181 m_freem(m); 8182 return (0); 8183 } 8184 /* 8185 * We don't support urgent data but 8186 * drag along the up just to make sure 8187 * if there is a stack switch no one 8188 * is surprised. 8189 */ 8190 tp->rcv_up = tp->rcv_nxt; 8191 8192 /* 8193 * Process the segment text, merging it into the TCP sequencing 8194 * queue, and arranging for acknowledgment of receipt if necessary. 8195 * This process logically involves adjusting tp->rcv_wnd as data is 8196 * presented to the user (this happens in tcp_usrreq.c, case 8197 * PRU_RCVD). If a FIN has already been received on this connection 8198 * then we just ignore the text. 8199 */ 8200 tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && 8201 (tp->t_flags & TF_FASTOPEN)); 8202 if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && 8203 TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8204 tcp_seq save_start = th->th_seq; 8205 tcp_seq save_rnxt = tp->rcv_nxt; 8206 int save_tlen = tlen; 8207 8208 m_adj(m, drop_hdrlen); /* delayed header drop */ 8209 /* 8210 * Insert segment which includes th into TCP reassembly 8211 * queue with control block tp. Set thflags to whether 8212 * reassembly now includes a segment with FIN. This handles 8213 * the common case inline (segment is the next to be 8214 * received on an established connection, and the queue is 8215 * empty), avoiding linkage into and removal from the queue 8216 * and repetition of various conversions. Set DELACK for 8217 * segments received in order, but ack immediately when 8218 * segments are out of order (so fast retransmit can work). 8219 */ 8220 if (th->th_seq == tp->rcv_nxt && 8221 SEGQ_EMPTY(tp) && 8222 (TCPS_HAVEESTABLISHED(tp->t_state) || 8223 tfo_syn)) { 8224 #ifdef NETFLIX_SB_LIMITS 8225 u_int mcnt, appended; 8226 8227 if (so->so_rcv.sb_shlim) { 8228 mcnt = m_memcnt(m); 8229 appended = 0; 8230 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8231 CFO_NOSLEEP, NULL) == false) { 8232 counter_u64_add(tcp_sb_shlim_fails, 1); 8233 m_freem(m); 8234 return (0); 8235 } 8236 } 8237 8238 #endif 8239 if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { 8240 bbr->bbr_segs_rcvd += max(1, nsegs); 8241 tp->t_flags |= TF_DELACK; 8242 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8243 } else { 8244 bbr->r_wanted_output = 1; 8245 tp->t_flags |= TF_ACKNOW; 8246 } 8247 tp->rcv_nxt += tlen; 8248 if (tlen && 8249 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8250 (tp->t_fbyte_in == 0)) { 8251 tp->t_fbyte_in = ticks; 8252 if (tp->t_fbyte_in == 0) 8253 tp->t_fbyte_in = 1; 8254 if (tp->t_fbyte_out && tp->t_fbyte_in) 8255 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8256 } 8257 thflags = tcp_get_flags(th) & TH_FIN; 8258 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8259 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8260 SOCKBUF_LOCK(&so->so_rcv); 8261 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 8262 m_freem(m); 8263 else 8264 #ifdef NETFLIX_SB_LIMITS 8265 appended = 8266 #endif 8267 sbappendstream_locked(&so->so_rcv, m, 0); 8268 /* NB: sorwakeup_locked() does an implicit unlock. */ 8269 sorwakeup_locked(so); 8270 #ifdef NETFLIX_SB_LIMITS 8271 if (so->so_rcv.sb_shlim && appended != mcnt) 8272 counter_fo_release(so->so_rcv.sb_shlim, 8273 mcnt - appended); 8274 #endif 8275 8276 } else { 8277 /* 8278 * XXX: Due to the header drop above "th" is 8279 * theoretically invalid by now. Fortunately 8280 * m_adj() doesn't actually frees any mbufs when 8281 * trimming from the head. 8282 */ 8283 tcp_seq temp = save_start; 8284 8285 thflags = tcp_reass(tp, th, &temp, &tlen, m); 8286 tp->t_flags |= TF_ACKNOW; 8287 if (tp->t_flags & TF_WAKESOR) { 8288 tp->t_flags &= ~TF_WAKESOR; 8289 /* NB: sorwakeup_locked() does an implicit unlock. */ 8290 sorwakeup_locked(so); 8291 } 8292 } 8293 if ((tp->t_flags & TF_SACK_PERMIT) && 8294 (save_tlen > 0) && 8295 TCPS_HAVEESTABLISHED(tp->t_state)) { 8296 if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { 8297 /* 8298 * DSACK actually handled in the fastpath 8299 * above. 8300 */ 8301 tcp_update_sack_list(tp, save_start, 8302 save_start + save_tlen); 8303 } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { 8304 if ((tp->rcv_numsacks >= 1) && 8305 (tp->sackblks[0].end == save_start)) { 8306 /* 8307 * Partial overlap, recorded at todrop 8308 * above. 8309 */ 8310 tcp_update_sack_list(tp, 8311 tp->sackblks[0].start, 8312 tp->sackblks[0].end); 8313 } else { 8314 tcp_update_dsack_list(tp, save_start, 8315 save_start + save_tlen); 8316 } 8317 } else if (tlen >= save_tlen) { 8318 /* Update of sackblks. */ 8319 tcp_update_dsack_list(tp, save_start, 8320 save_start + save_tlen); 8321 } else if (tlen > 0) { 8322 tcp_update_dsack_list(tp, save_start, 8323 save_start + tlen); 8324 } 8325 } 8326 } else { 8327 m_freem(m); 8328 thflags &= ~TH_FIN; 8329 } 8330 8331 /* 8332 * If FIN is received ACK the FIN and let the user know that the 8333 * connection is closing. 8334 */ 8335 if (thflags & TH_FIN) { 8336 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { 8337 /* The socket upcall is handled by socantrcvmore. */ 8338 socantrcvmore(so); 8339 /* 8340 * If connection is half-synchronized (ie NEEDSYN 8341 * flag on) then delay ACK, so it may be piggybacked 8342 * when SYN is sent. Otherwise, since we received a 8343 * FIN then no more input can be expected, send ACK 8344 * now. 8345 */ 8346 if (tp->t_flags & TF_NEEDSYN) { 8347 tp->t_flags |= TF_DELACK; 8348 bbr_timer_cancel(bbr, 8349 __LINE__, bbr->r_ctl.rc_rcvtime); 8350 } else { 8351 tp->t_flags |= TF_ACKNOW; 8352 } 8353 tp->rcv_nxt++; 8354 } 8355 switch (tp->t_state) { 8356 /* 8357 * In SYN_RECEIVED and ESTABLISHED STATES enter the 8358 * CLOSE_WAIT state. 8359 */ 8360 case TCPS_SYN_RECEIVED: 8361 tp->t_starttime = ticks; 8362 /* FALLTHROUGH */ 8363 case TCPS_ESTABLISHED: 8364 tcp_state_change(tp, TCPS_CLOSE_WAIT); 8365 break; 8366 8367 /* 8368 * If still in FIN_WAIT_1 STATE FIN has not been 8369 * acked so enter the CLOSING state. 8370 */ 8371 case TCPS_FIN_WAIT_1: 8372 tcp_state_change(tp, TCPS_CLOSING); 8373 break; 8374 8375 /* 8376 * In FIN_WAIT_2 state enter the TIME_WAIT state, 8377 * starting the time-wait timer, turning off the 8378 * other standard timers. 8379 */ 8380 case TCPS_FIN_WAIT_2: 8381 bbr->rc_timer_first = 1; 8382 bbr_timer_cancel(bbr, 8383 __LINE__, bbr->r_ctl.rc_rcvtime); 8384 tcp_twstart(tp); 8385 return (1); 8386 } 8387 } 8388 /* 8389 * Return any desired output. 8390 */ 8391 if ((tp->t_flags & TF_ACKNOW) || 8392 (sbavail(&so->so_snd) > ctf_outstanding(tp))) { 8393 bbr->r_wanted_output = 1; 8394 } 8395 return (0); 8396 } 8397 8398 /* 8399 * Here nothing is really faster, its just that we 8400 * have broken out the fast-data path also just like 8401 * the fast-ack. Return 1 if we processed the packet 8402 * return 0 if you need to take the "slow-path". 8403 */ 8404 static int 8405 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, 8406 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8407 uint32_t tiwin, int32_t nxt_pkt) 8408 { 8409 uint16_t nsegs; 8410 int32_t newsize = 0; /* automatic sockbuf scaling */ 8411 struct tcp_bbr *bbr; 8412 #ifdef NETFLIX_SB_LIMITS 8413 u_int mcnt, appended; 8414 #endif 8415 8416 /* On the hpts and we would have called output */ 8417 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8418 8419 /* 8420 * If last ACK falls within this segment's sequence numbers, record 8421 * the timestamp. NOTE that the test is modified according to the 8422 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8423 */ 8424 if (bbr->r_ctl.rc_resend != NULL) { 8425 return (0); 8426 } 8427 if (tiwin && tiwin != tp->snd_wnd) { 8428 return (0); 8429 } 8430 if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { 8431 return (0); 8432 } 8433 if (__predict_false((to->to_flags & TOF_TS) && 8434 (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { 8435 return (0); 8436 } 8437 if (__predict_false((th->th_ack != tp->snd_una))) { 8438 return (0); 8439 } 8440 if (__predict_false(tlen > sbspace(&so->so_rcv))) { 8441 return (0); 8442 } 8443 if ((to->to_flags & TOF_TS) != 0 && 8444 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8445 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 8446 tp->ts_recent = to->to_tsval; 8447 } 8448 /* 8449 * This is a pure, in-sequence data packet with nothing on the 8450 * reassembly queue and we have enough buffer space to take it. 8451 */ 8452 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8453 8454 #ifdef NETFLIX_SB_LIMITS 8455 if (so->so_rcv.sb_shlim) { 8456 mcnt = m_memcnt(m); 8457 appended = 0; 8458 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, 8459 CFO_NOSLEEP, NULL) == false) { 8460 counter_u64_add(tcp_sb_shlim_fails, 1); 8461 m_freem(m); 8462 return (1); 8463 } 8464 } 8465 #endif 8466 /* Clean receiver SACK report if present */ 8467 if (tp->rcv_numsacks) 8468 tcp_clean_sackreport(tp); 8469 KMOD_TCPSTAT_INC(tcps_preddat); 8470 tp->rcv_nxt += tlen; 8471 if (tlen && 8472 ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && 8473 (tp->t_fbyte_in == 0)) { 8474 tp->t_fbyte_in = ticks; 8475 if (tp->t_fbyte_in == 0) 8476 tp->t_fbyte_in = 1; 8477 if (tp->t_fbyte_out && tp->t_fbyte_in) 8478 tp->t_flags2 |= TF2_FBYTES_COMPLETE; 8479 } 8480 /* 8481 * Pull snd_wl1 up to prevent seq wrap relative to th_seq. 8482 */ 8483 tp->snd_wl1 = th->th_seq; 8484 /* 8485 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. 8486 */ 8487 tp->rcv_up = tp->rcv_nxt; 8488 KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 8489 KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); 8490 newsize = tcp_autorcvbuf(m, th, so, tp, tlen); 8491 8492 /* Add data to socket buffer. */ 8493 SOCKBUF_LOCK(&so->so_rcv); 8494 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8495 m_freem(m); 8496 } else { 8497 /* 8498 * Set new socket buffer size. Give up when limit is 8499 * reached. 8500 */ 8501 if (newsize) 8502 if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) 8503 so->so_rcv.sb_flags &= ~SB_AUTOSIZE; 8504 m_adj(m, drop_hdrlen); /* delayed header drop */ 8505 8506 #ifdef NETFLIX_SB_LIMITS 8507 appended = 8508 #endif 8509 sbappendstream_locked(&so->so_rcv, m, 0); 8510 ctf_calc_rwin(so, tp); 8511 } 8512 /* NB: sorwakeup_locked() does an implicit unlock. */ 8513 sorwakeup_locked(so); 8514 #ifdef NETFLIX_SB_LIMITS 8515 if (so->so_rcv.sb_shlim && mcnt != appended) 8516 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); 8517 #endif 8518 if (DELAY_ACK(tp, bbr, nsegs)) { 8519 bbr->bbr_segs_rcvd += max(1, nsegs); 8520 tp->t_flags |= TF_DELACK; 8521 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8522 } else { 8523 bbr->r_wanted_output = 1; 8524 tp->t_flags |= TF_ACKNOW; 8525 } 8526 return (1); 8527 } 8528 8529 /* 8530 * This subfunction is used to try to highly optimize the 8531 * fast path. We again allow window updates that are 8532 * in sequence to remain in the fast-path. We also add 8533 * in the __predict's to attempt to help the compiler. 8534 * Note that if we return a 0, then we can *not* process 8535 * it and the caller should push the packet into the 8536 * slow-path. If we return 1, then all is well and 8537 * the packet is fully processed. 8538 */ 8539 static int 8540 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 8541 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8542 uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) 8543 { 8544 int32_t acked; 8545 uint16_t nsegs; 8546 uint32_t sack_changed; 8547 uint32_t prev_acked = 0; 8548 struct tcp_bbr *bbr; 8549 8550 if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { 8551 /* Old ack, behind (or duplicate to) the last one rcv'd */ 8552 return (0); 8553 } 8554 if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { 8555 /* Above what we have sent? */ 8556 return (0); 8557 } 8558 if (__predict_false(tiwin == 0)) { 8559 /* zero window */ 8560 return (0); 8561 } 8562 if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { 8563 /* We need a SYN or a FIN, unlikely.. */ 8564 return (0); 8565 } 8566 if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { 8567 /* Timestamp is behind .. old ack with seq wrap? */ 8568 return (0); 8569 } 8570 if (__predict_false(IN_RECOVERY(tp->t_flags))) { 8571 /* Still recovering */ 8572 return (0); 8573 } 8574 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8575 if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { 8576 /* We are retransmitting */ 8577 return (0); 8578 } 8579 if (__predict_false(bbr->rc_in_persist != 0)) { 8580 /* In persist mode */ 8581 return (0); 8582 } 8583 if (bbr->r_ctl.rc_sacked) { 8584 /* We have sack holes on our scoreboard */ 8585 return (0); 8586 } 8587 /* Ok if we reach here, we can process a fast-ack */ 8588 nsegs = max(1, m->m_pkthdr.lro_nsegs); 8589 sack_changed = bbr_log_ack(tp, to, th, &prev_acked); 8590 /* 8591 * We never detect loss in fast ack [we can't 8592 * have a sack and can't be in recovery so 8593 * we always pass 0 (nothing detected)]. 8594 */ 8595 bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); 8596 /* Did the window get updated? */ 8597 if (tiwin != tp->snd_wnd) { 8598 tp->snd_wnd = tiwin; 8599 tp->snd_wl1 = th->th_seq; 8600 if (tp->snd_wnd > tp->max_sndwnd) 8601 tp->max_sndwnd = tp->snd_wnd; 8602 } 8603 /* Do we need to exit persists? */ 8604 if ((bbr->rc_in_persist != 0) && 8605 (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), 8606 bbr_minseg(bbr)))) { 8607 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8608 bbr->r_wanted_output = 1; 8609 } 8610 /* Do we need to enter persists? */ 8611 if ((bbr->rc_in_persist == 0) && 8612 (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 8613 TCPS_HAVEESTABLISHED(tp->t_state) && 8614 (tp->snd_max == tp->snd_una) && 8615 sbavail(&so->so_snd) && 8616 (sbavail(&so->so_snd) > tp->snd_wnd)) { 8617 /* No send window.. we must enter persist */ 8618 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 8619 } 8620 /* 8621 * If last ACK falls within this segment's sequence numbers, record 8622 * the timestamp. NOTE that the test is modified according to the 8623 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). 8624 */ 8625 if ((to->to_flags & TOF_TS) != 0 && 8626 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { 8627 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; 8628 tp->ts_recent = to->to_tsval; 8629 } 8630 /* 8631 * This is a pure ack for outstanding data. 8632 */ 8633 KMOD_TCPSTAT_INC(tcps_predack); 8634 8635 /* 8636 * "bad retransmit" recovery. 8637 */ 8638 if (tp->t_flags & TF_PREVVALID) { 8639 tp->t_flags &= ~TF_PREVVALID; 8640 if (tp->t_rxtshift == 1 && 8641 (int)(ticks - tp->t_badrxtwin) < 0) 8642 bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); 8643 } 8644 /* 8645 * Recalculate the transmit timer / rtt. 8646 * 8647 * Some boxes send broken timestamp replies during the SYN+ACK 8648 * phase, ignore timestamps of 0 or we could calculate a huge RTT 8649 * and blow up the retransmit timer. 8650 */ 8651 acked = BYTES_THIS_ACK(tp, th); 8652 8653 #ifdef TCP_HHOOK 8654 /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ 8655 hhook_run_tcp_est_in(tp, th, to); 8656 #endif 8657 8658 KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 8659 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); 8660 sbdrop(&so->so_snd, acked); 8661 8662 if (SEQ_GT(th->th_ack, tp->snd_una)) 8663 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); 8664 tp->snd_una = th->th_ack; 8665 if (tp->snd_wnd < ctf_outstanding(tp)) 8666 /* The peer collapsed its window on us */ 8667 bbr_collapsed_window(bbr); 8668 else if (bbr->rc_has_collapsed) 8669 bbr_un_collapse_window(bbr); 8670 8671 if (SEQ_GT(tp->snd_una, tp->snd_recover)) { 8672 tp->snd_recover = tp->snd_una; 8673 } 8674 bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); 8675 /* 8676 * Pull snd_wl2 up to prevent seq wrap relative to th_ack. 8677 */ 8678 tp->snd_wl2 = th->th_ack; 8679 m_freem(m); 8680 /* 8681 * If all outstanding data are acked, stop retransmit timer, 8682 * otherwise restart timer using current (possibly backed-off) 8683 * value. If process is waiting for space, wakeup/selwakeup/signal. 8684 * If data are ready to send, let tcp_output decide between more 8685 * output or persist. 8686 * Wake up the socket if we have room to write more. 8687 */ 8688 sowwakeup(so); 8689 if (tp->snd_una == tp->snd_max) { 8690 /* Nothing left outstanding */ 8691 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); 8692 if (sbavail(&so->so_snd) == 0) 8693 bbr->rc_tp->t_acktime = 0; 8694 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8695 if (bbr->rc_in_persist == 0) { 8696 bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; 8697 } 8698 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 8699 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); 8700 /* 8701 * We invalidate the last ack here since we 8702 * don't want to transfer forward the time 8703 * for our sum's calculations. 8704 */ 8705 bbr->r_wanted_output = 1; 8706 } 8707 if (sbavail(&so->so_snd)) { 8708 bbr->r_wanted_output = 1; 8709 } 8710 return (1); 8711 } 8712 8713 /* 8714 * Return value of 1, the TCB is unlocked and most 8715 * likely gone, return value of 0, the TCB is still 8716 * locked. 8717 */ 8718 static int 8719 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, 8720 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8721 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8722 { 8723 int32_t todrop; 8724 int32_t ourfinisacked = 0; 8725 struct tcp_bbr *bbr; 8726 int32_t ret_val = 0; 8727 8728 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8729 8730 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8731 ctf_calc_rwin(so, tp); 8732 /* 8733 * If the state is SYN_SENT: if seg contains an ACK, but not for our 8734 * SYN, drop the input. if seg contains a RST, then drop the 8735 * connection. if seg does not contain SYN, then drop it. Otherwise 8736 * this is an acceptable SYN segment initialize tp->rcv_nxt and 8737 * tp->irs if seg contains ack then advance tp->snd_una. BRR does 8738 * not support ECN so we will not say we are capable. if SYN has 8739 * been acked change to ESTABLISHED else SYN_RCVD state arrange for 8740 * segment to be acked (eventually) continue processing rest of 8741 * data/controls, beginning with URG 8742 */ 8743 if ((thflags & TH_ACK) && 8744 (SEQ_LEQ(th->th_ack, tp->iss) || 8745 SEQ_GT(th->th_ack, tp->snd_max))) { 8746 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8747 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8748 return (1); 8749 } 8750 if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { 8751 TCP_PROBE5(connect__refused, NULL, tp, 8752 mtod(m, const char *), tp, th); 8753 tp = tcp_drop(tp, ECONNREFUSED); 8754 ctf_do_drop(m, tp); 8755 return (1); 8756 } 8757 if (thflags & TH_RST) { 8758 ctf_do_drop(m, tp); 8759 return (1); 8760 } 8761 if (!(thflags & TH_SYN)) { 8762 ctf_do_drop(m, tp); 8763 return (1); 8764 } 8765 tp->irs = th->th_seq; 8766 tcp_rcvseqinit(tp); 8767 if (thflags & TH_ACK) { 8768 int tfo_partial = 0; 8769 8770 KMOD_TCPSTAT_INC(tcps_connects); 8771 soisconnected(so); 8772 #ifdef MAC 8773 mac_socketpeer_set_from_mbuf(m, so); 8774 #endif 8775 /* Do window scaling on this connection? */ 8776 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 8777 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 8778 tp->rcv_scale = tp->request_r_scale; 8779 } 8780 tp->rcv_adv += min(tp->rcv_wnd, 8781 TCP_MAXWIN << tp->rcv_scale); 8782 /* 8783 * If not all the data that was sent in the TFO SYN 8784 * has been acked, resend the remainder right away. 8785 */ 8786 if ((tp->t_flags & TF_FASTOPEN) && 8787 (tp->snd_una != tp->snd_max)) { 8788 tp->snd_nxt = th->th_ack; 8789 tfo_partial = 1; 8790 } 8791 /* 8792 * If there's data, delay ACK; if there's also a FIN ACKNOW 8793 * will be turned on later. 8794 */ 8795 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { 8796 bbr->bbr_segs_rcvd += 1; 8797 tp->t_flags |= TF_DELACK; 8798 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); 8799 } else { 8800 bbr->r_wanted_output = 1; 8801 tp->t_flags |= TF_ACKNOW; 8802 } 8803 if (SEQ_GT(th->th_ack, tp->iss)) { 8804 /* 8805 * The SYN is acked 8806 * handle it specially. 8807 */ 8808 bbr_log_syn(tp, to); 8809 } 8810 if (SEQ_GT(th->th_ack, tp->snd_una)) { 8811 /* 8812 * We advance snd_una for the 8813 * fast open case. If th_ack is 8814 * acknowledging data beyond 8815 * snd_una we can't just call 8816 * ack-processing since the 8817 * data stream in our send-map 8818 * will start at snd_una + 1 (one 8819 * beyond the SYN). If its just 8820 * equal we don't need to do that 8821 * and there is no send_map. 8822 */ 8823 tp->snd_una++; 8824 } 8825 /* 8826 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions: 8827 * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 8828 */ 8829 tp->t_starttime = ticks; 8830 if (tp->t_flags & TF_NEEDFIN) { 8831 tcp_state_change(tp, TCPS_FIN_WAIT_1); 8832 tp->t_flags &= ~TF_NEEDFIN; 8833 thflags &= ~TH_SYN; 8834 } else { 8835 tcp_state_change(tp, TCPS_ESTABLISHED); 8836 TCP_PROBE5(connect__established, NULL, tp, 8837 mtod(m, const char *), tp, th); 8838 cc_conn_init(tp); 8839 } 8840 } else { 8841 /* 8842 * Received initial SYN in SYN-SENT[*] state => simultaneous 8843 * open. If segment contains CC option and there is a 8844 * cached CC, apply TAO test. If it succeeds, connection is * 8845 * half-synchronized. Otherwise, do 3-way handshake: 8846 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If 8847 * there was no CC option, clear cached CC value. 8848 */ 8849 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN); 8850 tcp_state_change(tp, TCPS_SYN_RECEIVED); 8851 } 8852 /* 8853 * Advance th->th_seq to correspond to first data byte. If data, 8854 * trim to stay within window, dropping FIN if necessary. 8855 */ 8856 th->th_seq++; 8857 if (tlen > tp->rcv_wnd) { 8858 todrop = tlen - tp->rcv_wnd; 8859 m_adj(m, -todrop); 8860 tlen = tp->rcv_wnd; 8861 thflags &= ~TH_FIN; 8862 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); 8863 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); 8864 } 8865 tp->snd_wl1 = th->th_seq - 1; 8866 tp->rcv_up = th->th_seq; 8867 /* 8868 * Client side of transaction: already sent SYN and data. If the 8869 * remote host used T/TCP to validate the SYN, our data will be 8870 * ACK'd; if so, enter normal data segment processing in the middle 8871 * of step 5, ack processing. Otherwise, goto step 6. 8872 */ 8873 if (thflags & TH_ACK) { 8874 if ((to->to_flags & TOF_TS) != 0) { 8875 uint32_t t, rtt; 8876 8877 t = tcp_tv_to_mssectick(&bbr->rc_tv); 8878 if (TSTMP_GEQ(t, to->to_tsecr)) { 8879 rtt = t - to->to_tsecr; 8880 if (rtt == 0) { 8881 rtt = 1; 8882 } 8883 rtt *= MS_IN_USEC; 8884 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 8885 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, 8886 rtt, bbr->r_ctl.rc_rcvtime); 8887 } 8888 } 8889 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) 8890 return (ret_val); 8891 /* We may have changed to FIN_WAIT_1 above */ 8892 if (tp->t_state == TCPS_FIN_WAIT_1) { 8893 /* 8894 * In FIN_WAIT_1 STATE in addition to the processing 8895 * for the ESTABLISHED state if our FIN is now 8896 * acknowledged then enter FIN_WAIT_2. 8897 */ 8898 if (ourfinisacked) { 8899 /* 8900 * If we can't receive any more data, then 8901 * closing user can proceed. Starting the 8902 * timer is contrary to the specification, 8903 * but if we don't get a FIN we'll hang 8904 * forever. 8905 * 8906 * XXXjl: we should release the tp also, and 8907 * use a compressed state. 8908 */ 8909 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 8910 soisdisconnected(so); 8911 tcp_timer_activate(tp, TT_2MSL, 8912 (tcp_fast_finwait2_recycle ? 8913 tcp_finwait2_timeout : 8914 TP_MAXIDLE(tp))); 8915 } 8916 tcp_state_change(tp, TCPS_FIN_WAIT_2); 8917 } 8918 } 8919 } 8920 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 8921 tiwin, thflags, nxt_pkt)); 8922 } 8923 8924 /* 8925 * Return value of 1, the TCB is unlocked and most 8926 * likely gone, return value of 0, the TCB is still 8927 * locked. 8928 */ 8929 static int 8930 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, 8931 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 8932 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 8933 { 8934 int32_t ourfinisacked = 0; 8935 int32_t ret_val; 8936 struct tcp_bbr *bbr; 8937 8938 INP_WLOCK_ASSERT(tptoinpcb(tp)); 8939 8940 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 8941 ctf_calc_rwin(so, tp); 8942 if ((thflags & TH_RST) || 8943 (tp->t_fin_is_rst && (thflags & TH_FIN))) 8944 return (ctf_process_rst(m, th, so, tp)); 8945 if ((thflags & TH_ACK) && 8946 (SEQ_LEQ(th->th_ack, tp->snd_una) || 8947 SEQ_GT(th->th_ack, tp->snd_max))) { 8948 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8949 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8950 return (1); 8951 } 8952 if (tp->t_flags & TF_FASTOPEN) { 8953 /* 8954 * When a TFO connection is in SYN_RECEIVED, the only valid 8955 * packets are the initial SYN, a retransmit/copy of the 8956 * initial SYN (possibly with a subset of the original 8957 * data), a valid ACK, a FIN, or a RST. 8958 */ 8959 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { 8960 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8961 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8962 return (1); 8963 } else if (thflags & TH_SYN) { 8964 /* non-initial SYN is ignored */ 8965 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || 8966 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || 8967 (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { 8968 ctf_do_drop(m, NULL); 8969 return (0); 8970 } 8971 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { 8972 ctf_do_drop(m, NULL); 8973 return (0); 8974 } 8975 } 8976 /* 8977 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 8978 * it's less than ts_recent, drop it. 8979 */ 8980 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 8981 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 8982 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 8983 return (ret_val); 8984 } 8985 /* 8986 * In the SYN-RECEIVED state, validate that the packet belongs to 8987 * this connection before trimming the data to fit the receive 8988 * window. Check the sequence number versus IRS since we know the 8989 * sequence numbers haven't wrapped. This is a partial fix for the 8990 * "LAND" DoS attack. 8991 */ 8992 if (SEQ_LT(th->th_seq, tp->irs)) { 8993 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 8994 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 8995 return (1); 8996 } 8997 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 8998 return (ret_val); 8999 } 9000 /* 9001 * If last ACK falls within this segment's sequence numbers, record 9002 * its timestamp. NOTE: 1) That the test incorporates suggestions 9003 * from the latest proposal of the tcplw@cray.com list (Braden 9004 * 1993/04/26). 2) That updating only on newer timestamps interferes 9005 * with our earlier PAWS tests, so this check should be solely 9006 * predicated on the sequence space of this segment. 3) That we 9007 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9008 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9009 * SEG.Len, This modified check allows us to overcome RFC1323's 9010 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9011 * p.869. In such cases, we can still calculate the RTT correctly 9012 * when RCV.NXT == Last.ACK.Sent. 9013 */ 9014 if ((to->to_flags & TOF_TS) != 0 && 9015 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9016 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9017 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9018 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9019 tp->ts_recent = to->to_tsval; 9020 } 9021 tp->snd_wnd = tiwin; 9022 /* 9023 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9024 * is on (half-synchronized state), then queue data for later 9025 * processing; else drop segment and return. 9026 */ 9027 if ((thflags & TH_ACK) == 0) { 9028 if (tp->t_flags & TF_FASTOPEN) { 9029 cc_conn_init(tp); 9030 } 9031 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9032 tiwin, thflags, nxt_pkt)); 9033 } 9034 KMOD_TCPSTAT_INC(tcps_connects); 9035 if (tp->t_flags & TF_SONOTCONN) { 9036 tp->t_flags &= ~TF_SONOTCONN; 9037 soisconnected(so); 9038 } 9039 /* Do window scaling? */ 9040 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 9041 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 9042 tp->rcv_scale = tp->request_r_scale; 9043 } 9044 /* 9045 * ok for the first time in lets see if we can use the ts to figure 9046 * out what the initial RTT was. 9047 */ 9048 if ((to->to_flags & TOF_TS) != 0) { 9049 uint32_t t, rtt; 9050 9051 t = tcp_tv_to_mssectick(&bbr->rc_tv); 9052 if (TSTMP_GEQ(t, to->to_tsecr)) { 9053 rtt = t - to->to_tsecr; 9054 if (rtt == 0) { 9055 rtt = 1; 9056 } 9057 rtt *= MS_IN_USEC; 9058 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); 9059 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); 9060 } 9061 } 9062 /* Drop off any SYN in the send map (probably not there) */ 9063 if (thflags & TH_ACK) 9064 bbr_log_syn(tp, to); 9065 if ((tp->t_flags & TF_FASTOPEN) && tp->t_tfo_pending) { 9066 tcp_fastopen_decrement_counter(tp->t_tfo_pending); 9067 tp->t_tfo_pending = NULL; 9068 } 9069 /* 9070 * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> 9071 * FIN-WAIT-1 9072 */ 9073 tp->t_starttime = ticks; 9074 if (tp->t_flags & TF_NEEDFIN) { 9075 tcp_state_change(tp, TCPS_FIN_WAIT_1); 9076 tp->t_flags &= ~TF_NEEDFIN; 9077 } else { 9078 tcp_state_change(tp, TCPS_ESTABLISHED); 9079 TCP_PROBE5(accept__established, NULL, tp, 9080 mtod(m, const char *), tp, th); 9081 /* 9082 * TFO connections call cc_conn_init() during SYN 9083 * processing. Calling it again here for such connections 9084 * is not harmless as it would undo the snd_cwnd reduction 9085 * that occurs when a TFO SYN|ACK is retransmitted. 9086 */ 9087 if (!(tp->t_flags & TF_FASTOPEN)) 9088 cc_conn_init(tp); 9089 } 9090 /* 9091 * Account for the ACK of our SYN prior to 9092 * regular ACK processing below, except for 9093 * simultaneous SYN, which is handled later. 9094 */ 9095 if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) 9096 tp->snd_una++; 9097 /* 9098 * If segment contains data or ACK, will call tcp_reass() later; if 9099 * not, do so now to pass queued data to user. 9100 */ 9101 if (tlen == 0 && (thflags & TH_FIN) == 0) { 9102 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, 9103 (struct mbuf *)0); 9104 if (tp->t_flags & TF_WAKESOR) { 9105 tp->t_flags &= ~TF_WAKESOR; 9106 /* NB: sorwakeup_locked() does an implicit unlock. */ 9107 sorwakeup_locked(so); 9108 } 9109 } 9110 tp->snd_wl1 = th->th_seq - 1; 9111 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9112 return (ret_val); 9113 } 9114 if (tp->t_state == TCPS_FIN_WAIT_1) { 9115 /* We could have went to FIN_WAIT_1 (or EST) above */ 9116 /* 9117 * In FIN_WAIT_1 STATE in addition to the processing for the 9118 * ESTABLISHED state if our FIN is now acknowledged then 9119 * enter FIN_WAIT_2. 9120 */ 9121 if (ourfinisacked) { 9122 /* 9123 * If we can't receive any more data, then closing 9124 * user can proceed. Starting the timer is contrary 9125 * to the specification, but if we don't get a FIN 9126 * we'll hang forever. 9127 * 9128 * XXXjl: we should release the tp also, and use a 9129 * compressed state. 9130 */ 9131 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9132 soisdisconnected(so); 9133 tcp_timer_activate(tp, TT_2MSL, 9134 (tcp_fast_finwait2_recycle ? 9135 tcp_finwait2_timeout : 9136 TP_MAXIDLE(tp))); 9137 } 9138 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9139 } 9140 } 9141 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9142 tiwin, thflags, nxt_pkt)); 9143 } 9144 9145 /* 9146 * Return value of 1, the TCB is unlocked and most 9147 * likely gone, return value of 0, the TCB is still 9148 * locked. 9149 */ 9150 static int 9151 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, 9152 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9153 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9154 { 9155 struct tcp_bbr *bbr; 9156 int32_t ret_val; 9157 9158 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9159 9160 /* 9161 * Header prediction: check for the two common cases of a 9162 * uni-directional data xfer. If the packet has no control flags, 9163 * is in-sequence, the window didn't change and we're not 9164 * retransmitting, it's a candidate. If the length is zero and the 9165 * ack moved forward, we're the sender side of the xfer. Just free 9166 * the data acked & wake any higher level process that was blocked 9167 * waiting for space. If the length is non-zero and the ack didn't 9168 * move, we're the receiver side. If we're getting packets in-order 9169 * (the reassembly queue is empty), add the data toc The socket 9170 * buffer and note that we need a delayed ack. Make sure that the 9171 * hidden state-flags are also off. Since we check for 9172 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. 9173 */ 9174 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9175 if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { 9176 /* 9177 * If we have delived under 4 segments increase the initial 9178 * window if raised by the peer. We use this to determine 9179 * dynamic and static rwnd's at the end of a connection. 9180 */ 9181 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); 9182 } 9183 if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && 9184 __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && 9185 __predict_true(SEGQ_EMPTY(tp)) && 9186 __predict_true(th->th_seq == tp->rcv_nxt)) { 9187 if (tlen == 0) { 9188 if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, 9189 tiwin, nxt_pkt, iptos)) { 9190 return (0); 9191 } 9192 } else { 9193 if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, 9194 tiwin, nxt_pkt)) { 9195 return (0); 9196 } 9197 } 9198 } 9199 ctf_calc_rwin(so, tp); 9200 9201 if ((thflags & TH_RST) || 9202 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9203 return (ctf_process_rst(m, th, so, tp)); 9204 /* 9205 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9206 * synchronized state. 9207 */ 9208 if (thflags & TH_SYN) { 9209 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9210 return (ret_val); 9211 } 9212 /* 9213 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9214 * it's less than ts_recent, drop it. 9215 */ 9216 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9217 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9218 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9219 return (ret_val); 9220 } 9221 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9222 return (ret_val); 9223 } 9224 /* 9225 * If last ACK falls within this segment's sequence numbers, record 9226 * its timestamp. NOTE: 1) That the test incorporates suggestions 9227 * from the latest proposal of the tcplw@cray.com list (Braden 9228 * 1993/04/26). 2) That updating only on newer timestamps interferes 9229 * with our earlier PAWS tests, so this check should be solely 9230 * predicated on the sequence space of this segment. 3) That we 9231 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9232 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9233 * SEG.Len, This modified check allows us to overcome RFC1323's 9234 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9235 * p.869. In such cases, we can still calculate the RTT correctly 9236 * when RCV.NXT == Last.ACK.Sent. 9237 */ 9238 if ((to->to_flags & TOF_TS) != 0 && 9239 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9240 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9241 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9242 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9243 tp->ts_recent = to->to_tsval; 9244 } 9245 /* 9246 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9247 * is on (half-synchronized state), then queue data for later 9248 * processing; else drop segment and return. 9249 */ 9250 if ((thflags & TH_ACK) == 0) { 9251 if (tp->t_flags & TF_NEEDSYN) { 9252 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9253 tiwin, thflags, nxt_pkt)); 9254 } else if (tp->t_flags & TF_ACKNOW) { 9255 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9256 bbr->r_wanted_output = 1; 9257 return (ret_val); 9258 } else { 9259 ctf_do_drop(m, NULL); 9260 return (0); 9261 } 9262 } 9263 /* 9264 * Ack processing. 9265 */ 9266 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9267 return (ret_val); 9268 } 9269 if (sbavail(&so->so_snd)) { 9270 if (ctf_progress_timeout_check(tp, true)) { 9271 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9272 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9273 return (1); 9274 } 9275 } 9276 /* State changes only happen in bbr_process_data() */ 9277 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9278 tiwin, thflags, nxt_pkt)); 9279 } 9280 9281 /* 9282 * Return value of 1, the TCB is unlocked and most 9283 * likely gone, return value of 0, the TCB is still 9284 * locked. 9285 */ 9286 static int 9287 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, 9288 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9289 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9290 { 9291 struct tcp_bbr *bbr; 9292 int32_t ret_val; 9293 9294 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9295 9296 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9297 ctf_calc_rwin(so, tp); 9298 if ((thflags & TH_RST) || 9299 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9300 return (ctf_process_rst(m, th, so, tp)); 9301 /* 9302 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9303 * synchronized state. 9304 */ 9305 if (thflags & TH_SYN) { 9306 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9307 return (ret_val); 9308 } 9309 /* 9310 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9311 * it's less than ts_recent, drop it. 9312 */ 9313 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9314 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9315 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9316 return (ret_val); 9317 } 9318 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9319 return (ret_val); 9320 } 9321 /* 9322 * If last ACK falls within this segment's sequence numbers, record 9323 * its timestamp. NOTE: 1) That the test incorporates suggestions 9324 * from the latest proposal of the tcplw@cray.com list (Braden 9325 * 1993/04/26). 2) That updating only on newer timestamps interferes 9326 * with our earlier PAWS tests, so this check should be solely 9327 * predicated on the sequence space of this segment. 3) That we 9328 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9329 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9330 * SEG.Len, This modified check allows us to overcome RFC1323's 9331 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9332 * p.869. In such cases, we can still calculate the RTT correctly 9333 * when RCV.NXT == Last.ACK.Sent. 9334 */ 9335 if ((to->to_flags & TOF_TS) != 0 && 9336 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9337 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9338 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9339 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9340 tp->ts_recent = to->to_tsval; 9341 } 9342 /* 9343 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9344 * is on (half-synchronized state), then queue data for later 9345 * processing; else drop segment and return. 9346 */ 9347 if ((thflags & TH_ACK) == 0) { 9348 if (tp->t_flags & TF_NEEDSYN) { 9349 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9350 tiwin, thflags, nxt_pkt)); 9351 } else if (tp->t_flags & TF_ACKNOW) { 9352 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9353 bbr->r_wanted_output = 1; 9354 return (ret_val); 9355 } else { 9356 ctf_do_drop(m, NULL); 9357 return (0); 9358 } 9359 } 9360 /* 9361 * Ack processing. 9362 */ 9363 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { 9364 return (ret_val); 9365 } 9366 if (sbavail(&so->so_snd)) { 9367 if (ctf_progress_timeout_check(tp, true)) { 9368 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9369 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9370 return (1); 9371 } 9372 } 9373 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9374 tiwin, thflags, nxt_pkt)); 9375 } 9376 9377 static int 9378 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, 9379 struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) 9380 { 9381 9382 if (bbr->rc_allow_data_af_clo == 0) { 9383 close_now: 9384 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); 9385 /* tcp_close will kill the inp pre-log the Reset */ 9386 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 9387 tp = tcp_close(tp); 9388 KMOD_TCPSTAT_INC(tcps_rcvafterclose); 9389 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); 9390 return (1); 9391 } 9392 if (sbavail(&so->so_snd) == 0) 9393 goto close_now; 9394 /* Ok we allow data that is ignored and a followup reset */ 9395 tp->rcv_nxt = th->th_seq + *tlen; 9396 tp->t_flags2 |= TF2_DROP_AF_DATA; 9397 bbr->r_wanted_output = 1; 9398 *tlen = 0; 9399 return (0); 9400 } 9401 9402 /* 9403 * Return value of 1, the TCB is unlocked and most 9404 * likely gone, return value of 0, the TCB is still 9405 * locked. 9406 */ 9407 static int 9408 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, 9409 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9410 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9411 { 9412 int32_t ourfinisacked = 0; 9413 int32_t ret_val; 9414 struct tcp_bbr *bbr; 9415 9416 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9417 9418 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9419 ctf_calc_rwin(so, tp); 9420 if ((thflags & TH_RST) || 9421 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9422 return (ctf_process_rst(m, th, so, tp)); 9423 /* 9424 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9425 * synchronized state. 9426 */ 9427 if (thflags & TH_SYN) { 9428 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9429 return (ret_val); 9430 } 9431 /* 9432 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9433 * it's less than ts_recent, drop it. 9434 */ 9435 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9436 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9437 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9438 return (ret_val); 9439 } 9440 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9441 return (ret_val); 9442 } 9443 /* 9444 * If new data are received on a connection after the user processes 9445 * are gone, then RST the other end. 9446 * We call a new function now so we might continue and setup 9447 * to reset at all data being ack'd. 9448 */ 9449 if ((tp->t_flags & TF_CLOSED) && tlen && 9450 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9451 return (1); 9452 /* 9453 * If last ACK falls within this segment's sequence numbers, record 9454 * its timestamp. NOTE: 1) That the test incorporates suggestions 9455 * from the latest proposal of the tcplw@cray.com list (Braden 9456 * 1993/04/26). 2) That updating only on newer timestamps interferes 9457 * with our earlier PAWS tests, so this check should be solely 9458 * predicated on the sequence space of this segment. 3) That we 9459 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9460 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9461 * SEG.Len, This modified check allows us to overcome RFC1323's 9462 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9463 * p.869. In such cases, we can still calculate the RTT correctly 9464 * when RCV.NXT == Last.ACK.Sent. 9465 */ 9466 if ((to->to_flags & TOF_TS) != 0 && 9467 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9468 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9469 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9470 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9471 tp->ts_recent = to->to_tsval; 9472 } 9473 /* 9474 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9475 * is on (half-synchronized state), then queue data for later 9476 * processing; else drop segment and return. 9477 */ 9478 if ((thflags & TH_ACK) == 0) { 9479 if (tp->t_flags & TF_NEEDSYN) { 9480 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9481 tiwin, thflags, nxt_pkt)); 9482 } else if (tp->t_flags & TF_ACKNOW) { 9483 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9484 bbr->r_wanted_output = 1; 9485 return (ret_val); 9486 } else { 9487 ctf_do_drop(m, NULL); 9488 return (0); 9489 } 9490 } 9491 /* 9492 * Ack processing. 9493 */ 9494 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9495 return (ret_val); 9496 } 9497 if (ourfinisacked) { 9498 /* 9499 * If we can't receive any more data, then closing user can 9500 * proceed. Starting the timer is contrary to the 9501 * specification, but if we don't get a FIN we'll hang 9502 * forever. 9503 * 9504 * XXXjl: we should release the tp also, and use a 9505 * compressed state. 9506 */ 9507 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { 9508 soisdisconnected(so); 9509 tcp_timer_activate(tp, TT_2MSL, 9510 (tcp_fast_finwait2_recycle ? 9511 tcp_finwait2_timeout : 9512 TP_MAXIDLE(tp))); 9513 } 9514 tcp_state_change(tp, TCPS_FIN_WAIT_2); 9515 } 9516 if (sbavail(&so->so_snd)) { 9517 if (ctf_progress_timeout_check(tp, true)) { 9518 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9519 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9520 return (1); 9521 } 9522 } 9523 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9524 tiwin, thflags, nxt_pkt)); 9525 } 9526 9527 /* 9528 * Return value of 1, the TCB is unlocked and most 9529 * likely gone, return value of 0, the TCB is still 9530 * locked. 9531 */ 9532 static int 9533 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, 9534 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9535 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9536 { 9537 int32_t ourfinisacked = 0; 9538 int32_t ret_val; 9539 struct tcp_bbr *bbr; 9540 9541 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9542 9543 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9544 ctf_calc_rwin(so, tp); 9545 if ((thflags & TH_RST) || 9546 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9547 return (ctf_process_rst(m, th, so, tp)); 9548 /* 9549 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9550 * synchronized state. 9551 */ 9552 if (thflags & TH_SYN) { 9553 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9554 return (ret_val); 9555 } 9556 /* 9557 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9558 * it's less than ts_recent, drop it. 9559 */ 9560 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9561 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9562 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9563 return (ret_val); 9564 } 9565 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9566 return (ret_val); 9567 } 9568 /* 9569 * If last ACK falls within this segment's sequence numbers, record 9570 * its timestamp. NOTE: 1) That the test incorporates suggestions 9571 * from the latest proposal of the tcplw@cray.com list (Braden 9572 * 1993/04/26). 2) That updating only on newer timestamps interferes 9573 * with our earlier PAWS tests, so this check should be solely 9574 * predicated on the sequence space of this segment. 3) That we 9575 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9576 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9577 * SEG.Len, This modified check allows us to overcome RFC1323's 9578 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9579 * p.869. In such cases, we can still calculate the RTT correctly 9580 * when RCV.NXT == Last.ACK.Sent. 9581 */ 9582 if ((to->to_flags & TOF_TS) != 0 && 9583 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9584 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9585 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9586 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9587 tp->ts_recent = to->to_tsval; 9588 } 9589 /* 9590 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9591 * is on (half-synchronized state), then queue data for later 9592 * processing; else drop segment and return. 9593 */ 9594 if ((thflags & TH_ACK) == 0) { 9595 if (tp->t_flags & TF_NEEDSYN) { 9596 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9597 tiwin, thflags, nxt_pkt)); 9598 } else if (tp->t_flags & TF_ACKNOW) { 9599 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9600 bbr->r_wanted_output = 1; 9601 return (ret_val); 9602 } else { 9603 ctf_do_drop(m, NULL); 9604 return (0); 9605 } 9606 } 9607 /* 9608 * Ack processing. 9609 */ 9610 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9611 return (ret_val); 9612 } 9613 if (ourfinisacked) { 9614 tcp_twstart(tp); 9615 m_freem(m); 9616 return (1); 9617 } 9618 if (sbavail(&so->so_snd)) { 9619 if (ctf_progress_timeout_check(tp, true)) { 9620 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9621 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9622 return (1); 9623 } 9624 } 9625 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9626 tiwin, thflags, nxt_pkt)); 9627 } 9628 9629 /* 9630 * Return value of 1, the TCB is unlocked and most 9631 * likely gone, return value of 0, the TCB is still 9632 * locked. 9633 */ 9634 static int 9635 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, 9636 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9637 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9638 { 9639 int32_t ourfinisacked = 0; 9640 int32_t ret_val; 9641 struct tcp_bbr *bbr; 9642 9643 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9644 9645 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9646 ctf_calc_rwin(so, tp); 9647 if ((thflags & TH_RST) || 9648 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9649 return (ctf_process_rst(m, th, so, tp)); 9650 /* 9651 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9652 * synchronized state. 9653 */ 9654 if (thflags & TH_SYN) { 9655 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9656 return (ret_val); 9657 } 9658 /* 9659 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9660 * it's less than ts_recent, drop it. 9661 */ 9662 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9663 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9664 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9665 return (ret_val); 9666 } 9667 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9668 return (ret_val); 9669 } 9670 /* 9671 * If last ACK falls within this segment's sequence numbers, record 9672 * its timestamp. NOTE: 1) That the test incorporates suggestions 9673 * from the latest proposal of the tcplw@cray.com list (Braden 9674 * 1993/04/26). 2) That updating only on newer timestamps interferes 9675 * with our earlier PAWS tests, so this check should be solely 9676 * predicated on the sequence space of this segment. 3) That we 9677 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9678 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9679 * SEG.Len, This modified check allows us to overcome RFC1323's 9680 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9681 * p.869. In such cases, we can still calculate the RTT correctly 9682 * when RCV.NXT == Last.ACK.Sent. 9683 */ 9684 if ((to->to_flags & TOF_TS) != 0 && 9685 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9686 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9687 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9688 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9689 tp->ts_recent = to->to_tsval; 9690 } 9691 /* 9692 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9693 * is on (half-synchronized state), then queue data for later 9694 * processing; else drop segment and return. 9695 */ 9696 if ((thflags & TH_ACK) == 0) { 9697 if (tp->t_flags & TF_NEEDSYN) { 9698 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9699 tiwin, thflags, nxt_pkt)); 9700 } else if (tp->t_flags & TF_ACKNOW) { 9701 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9702 bbr->r_wanted_output = 1; 9703 return (ret_val); 9704 } else { 9705 ctf_do_drop(m, NULL); 9706 return (0); 9707 } 9708 } 9709 /* 9710 * case TCPS_LAST_ACK: Ack processing. 9711 */ 9712 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9713 return (ret_val); 9714 } 9715 if (ourfinisacked) { 9716 tp = tcp_close(tp); 9717 ctf_do_drop(m, tp); 9718 return (1); 9719 } 9720 if (sbavail(&so->so_snd)) { 9721 if (ctf_progress_timeout_check(tp, true)) { 9722 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9723 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9724 return (1); 9725 } 9726 } 9727 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9728 tiwin, thflags, nxt_pkt)); 9729 } 9730 9731 /* 9732 * Return value of 1, the TCB is unlocked and most 9733 * likely gone, return value of 0, the TCB is still 9734 * locked. 9735 */ 9736 static int 9737 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, 9738 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, 9739 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) 9740 { 9741 int32_t ourfinisacked = 0; 9742 int32_t ret_val; 9743 struct tcp_bbr *bbr; 9744 9745 INP_WLOCK_ASSERT(tptoinpcb(tp)); 9746 9747 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 9748 ctf_calc_rwin(so, tp); 9749 /* Reset receive buffer auto scaling when not in bulk receive mode. */ 9750 if ((thflags & TH_RST) || 9751 (tp->t_fin_is_rst && (thflags & TH_FIN))) 9752 return (ctf_process_rst(m, th, so, tp)); 9753 9754 /* 9755 * RFC5961 Section 4.2 Send challenge ACK for any SYN in 9756 * synchronized state. 9757 */ 9758 if (thflags & TH_SYN) { 9759 ctf_challenge_ack(m, th, tp, iptos, &ret_val); 9760 return (ret_val); 9761 } 9762 /* 9763 * RFC 1323 PAWS: If we have a timestamp reply on this segment and 9764 * it's less than ts_recent, drop it. 9765 */ 9766 if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && 9767 TSTMP_LT(to->to_tsval, tp->ts_recent)) { 9768 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) 9769 return (ret_val); 9770 } 9771 if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { 9772 return (ret_val); 9773 } 9774 /* 9775 * If new data are received on a connection after the user processes 9776 * are gone, then we may RST the other end depending on the outcome 9777 * of bbr_check_data_after_close. 9778 * We call a new function now so we might continue and setup 9779 * to reset at all data being ack'd. 9780 */ 9781 if ((tp->t_flags & TF_CLOSED) && tlen && 9782 bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) 9783 return (1); 9784 /* 9785 * If last ACK falls within this segment's sequence numbers, record 9786 * its timestamp. NOTE: 1) That the test incorporates suggestions 9787 * from the latest proposal of the tcplw@cray.com list (Braden 9788 * 1993/04/26). 2) That updating only on newer timestamps interferes 9789 * with our earlier PAWS tests, so this check should be solely 9790 * predicated on the sequence space of this segment. 3) That we 9791 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ 9792 * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + 9793 * SEG.Len, This modified check allows us to overcome RFC1323's 9794 * limitations as described in Stevens TCP/IP Illustrated Vol. 2 9795 * p.869. In such cases, we can still calculate the RTT correctly 9796 * when RCV.NXT == Last.ACK.Sent. 9797 */ 9798 if ((to->to_flags & TOF_TS) != 0 && 9799 SEQ_LEQ(th->th_seq, tp->last_ack_sent) && 9800 SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + 9801 ((thflags & (TH_SYN | TH_FIN)) != 0))) { 9802 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 9803 tp->ts_recent = to->to_tsval; 9804 } 9805 /* 9806 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag 9807 * is on (half-synchronized state), then queue data for later 9808 * processing; else drop segment and return. 9809 */ 9810 if ((thflags & TH_ACK) == 0) { 9811 if (tp->t_flags & TF_NEEDSYN) { 9812 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9813 tiwin, thflags, nxt_pkt)); 9814 } else if (tp->t_flags & TF_ACKNOW) { 9815 ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); 9816 bbr->r_wanted_output = 1; 9817 return (ret_val); 9818 } else { 9819 ctf_do_drop(m, NULL); 9820 return (0); 9821 } 9822 } 9823 /* 9824 * Ack processing. 9825 */ 9826 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { 9827 return (ret_val); 9828 } 9829 if (sbavail(&so->so_snd)) { 9830 if (ctf_progress_timeout_check(tp, true)) { 9831 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); 9832 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 9833 return (1); 9834 } 9835 } 9836 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, 9837 tiwin, thflags, nxt_pkt)); 9838 } 9839 9840 static void 9841 bbr_stop_all_timers(struct tcpcb *tp, struct tcp_bbr *bbr) 9842 { 9843 /* 9844 * Assure no timers are running. 9845 */ 9846 if (tcp_timer_active(tp, TT_PERSIST)) { 9847 /* We enter in persists, set the flag appropriately */ 9848 bbr->rc_in_persist = 1; 9849 } 9850 if (tcp_in_hpts(bbr->rc_tp)) { 9851 tcp_hpts_remove(bbr->rc_tp); 9852 } 9853 } 9854 9855 static void 9856 bbr_google_mode_on(struct tcp_bbr *bbr) 9857 { 9858 bbr->rc_use_google = 1; 9859 bbr->rc_no_pacing = 0; 9860 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9861 bbr->r_use_policer = bbr_policer_detection_enabled; 9862 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 9863 bbr->bbr_use_rack_cheat = 0; 9864 bbr->r_ctl.rc_incr_tmrs = 0; 9865 bbr->r_ctl.rc_inc_tcp_oh = 0; 9866 bbr->r_ctl.rc_inc_ip_oh = 0; 9867 bbr->r_ctl.rc_inc_enet_oh = 0; 9868 reset_time(&bbr->r_ctl.rc_delrate, 9869 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 9870 reset_time_small(&bbr->r_ctl.rc_rttprop, 9871 (11 * USECS_IN_SECOND)); 9872 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9873 } 9874 9875 static void 9876 bbr_google_mode_off(struct tcp_bbr *bbr) 9877 { 9878 bbr->rc_use_google = 0; 9879 bbr->r_ctl.bbr_google_discount = 0; 9880 bbr->no_pacing_until = bbr_no_pacing_until; 9881 bbr->r_use_policer = 0; 9882 if (bbr->no_pacing_until) 9883 bbr->rc_no_pacing = 1; 9884 else 9885 bbr->rc_no_pacing = 0; 9886 if (bbr_use_rack_resend_cheat) 9887 bbr->bbr_use_rack_cheat = 1; 9888 else 9889 bbr->bbr_use_rack_cheat = 0; 9890 if (bbr_incr_timers) 9891 bbr->r_ctl.rc_incr_tmrs = 1; 9892 else 9893 bbr->r_ctl.rc_incr_tmrs = 0; 9894 if (bbr_include_tcp_oh) 9895 bbr->r_ctl.rc_inc_tcp_oh = 1; 9896 else 9897 bbr->r_ctl.rc_inc_tcp_oh = 0; 9898 if (bbr_include_ip_oh) 9899 bbr->r_ctl.rc_inc_ip_oh = 1; 9900 else 9901 bbr->r_ctl.rc_inc_ip_oh = 0; 9902 if (bbr_include_enet_oh) 9903 bbr->r_ctl.rc_inc_enet_oh = 1; 9904 else 9905 bbr->r_ctl.rc_inc_enet_oh = 0; 9906 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 9907 reset_time(&bbr->r_ctl.rc_delrate, 9908 bbr_num_pktepo_for_del_limit); 9909 reset_time_small(&bbr->r_ctl.rc_rttprop, 9910 (bbr_filter_len_sec * USECS_IN_SECOND)); 9911 tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); 9912 } 9913 /* 9914 * Return 0 on success, non-zero on failure 9915 * which indicates the error (usually no memory). 9916 */ 9917 static int 9918 bbr_init(struct tcpcb *tp, void **ptr) 9919 { 9920 struct inpcb *inp = tptoinpcb(tp); 9921 struct tcp_bbr *bbr = NULL; 9922 uint32_t cts; 9923 9924 tcp_hpts_init(tp); 9925 9926 *ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); 9927 if (*ptr == NULL) { 9928 /* 9929 * We need to allocate memory but cant. The INP and INP_INFO 9930 * locks and they are recursive (happens during setup. So a 9931 * scheme to drop the locks fails :( 9932 * 9933 */ 9934 return (ENOMEM); 9935 } 9936 bbr = (struct tcp_bbr *)*ptr; 9937 bbr->rtt_valid = 0; 9938 tp->t_flags2 |= TF2_CANNOT_DO_ECN; 9939 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ; 9940 /* Take off any undesired flags */ 9941 tp->t_flags2 &= ~TF2_MBUF_QUEUE_READY; 9942 tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE; 9943 tp->t_flags2 &= ~TF2_MBUF_ACKCMP; 9944 tp->t_flags2 &= ~TF2_MBUF_L_ACKS; 9945 9946 TAILQ_INIT(&bbr->r_ctl.rc_map); 9947 TAILQ_INIT(&bbr->r_ctl.rc_free); 9948 TAILQ_INIT(&bbr->r_ctl.rc_tmap); 9949 bbr->rc_tp = tp; 9950 bbr->rc_inp = inp; 9951 cts = tcp_get_usecs(&bbr->rc_tv); 9952 tp->t_acktime = 0; 9953 bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; 9954 bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; 9955 bbr->rc_tlp_threshold = bbr_tlp_thresh; 9956 bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; 9957 bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; 9958 bbr->r_ctl.rc_min_to = bbr_min_to; 9959 bbr->rc_bbr_state = BBR_STATE_STARTUP; 9960 bbr->r_ctl.bbr_lost_at_state = 0; 9961 bbr->r_ctl.rc_lost_at_startup = 0; 9962 bbr->rc_all_timers_stopped = 0; 9963 bbr->r_ctl.rc_bbr_lastbtlbw = 0; 9964 bbr->r_ctl.rc_pkt_epoch_del = 0; 9965 bbr->r_ctl.rc_pkt_epoch = 0; 9966 bbr->r_ctl.rc_lowest_rtt = 0xffffffff; 9967 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; 9968 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 9969 bbr->r_ctl.rc_went_idle_time = cts; 9970 bbr->rc_pacer_started = cts; 9971 bbr->r_ctl.rc_pkt_epoch_time = cts; 9972 bbr->r_ctl.rc_rcvtime = cts; 9973 bbr->r_ctl.rc_bbr_state_time = cts; 9974 bbr->r_ctl.rc_del_time = cts; 9975 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 9976 bbr->r_ctl.last_in_probertt = cts; 9977 bbr->skip_gain = 0; 9978 bbr->gain_is_limited = 0; 9979 bbr->no_pacing_until = bbr_no_pacing_until; 9980 if (bbr->no_pacing_until) 9981 bbr->rc_no_pacing = 1; 9982 if (bbr_use_google_algo) { 9983 bbr->rc_no_pacing = 0; 9984 bbr->rc_use_google = 1; 9985 bbr->r_ctl.bbr_google_discount = bbr_google_discount; 9986 bbr->r_use_policer = bbr_policer_detection_enabled; 9987 } else { 9988 bbr->rc_use_google = 0; 9989 bbr->r_ctl.bbr_google_discount = 0; 9990 bbr->r_use_policer = 0; 9991 } 9992 if (bbr_ts_limiting) 9993 bbr->rc_use_ts_limit = 1; 9994 else 9995 bbr->rc_use_ts_limit = 0; 9996 if (bbr_ts_can_raise) 9997 bbr->ts_can_raise = 1; 9998 else 9999 bbr->ts_can_raise = 0; 10000 if (V_tcp_delack_enabled == 1) 10001 tp->t_delayed_ack = 2; 10002 else if (V_tcp_delack_enabled == 0) 10003 tp->t_delayed_ack = 0; 10004 else if (V_tcp_delack_enabled < 100) 10005 tp->t_delayed_ack = V_tcp_delack_enabled; 10006 else 10007 tp->t_delayed_ack = 2; 10008 if (bbr->rc_use_google == 0) 10009 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10010 else 10011 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); 10012 bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; 10013 bbr->rc_max_rto_sec = bbr_rto_max_sec; 10014 bbr->rc_init_win = bbr_def_init_win; 10015 if (tp->t_flags & TF_REQ_TSTMP) 10016 bbr->rc_last_options = TCP_TS_OVERHEAD; 10017 bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; 10018 bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; 10019 bbr->r_init_rtt = 1; 10020 10021 counter_u64_add(bbr_flows_nohdwr_pacing, 1); 10022 if (bbr_allow_hdwr_pacing) 10023 bbr->bbr_hdw_pace_ena = 1; 10024 else 10025 bbr->bbr_hdw_pace_ena = 0; 10026 if (bbr_sends_full_iwnd) 10027 bbr->bbr_init_win_cheat = 1; 10028 else 10029 bbr->bbr_init_win_cheat = 0; 10030 bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; 10031 bbr->r_ctl.rc_drain_pg = bbr_drain_gain; 10032 bbr->r_ctl.rc_startup_pg = bbr_high_gain; 10033 bbr->rc_loss_exit = bbr_exit_startup_at_loss; 10034 bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; 10035 bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; 10036 bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; 10037 bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; 10038 bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; 10039 bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; 10040 bbr->r_ctl.bbr_cross_over = bbr_cross_over; 10041 bbr->r_ctl.rc_rtt_shrinks = cts; 10042 if (bbr->rc_use_google) { 10043 setup_time_filter(&bbr->r_ctl.rc_delrate, 10044 FILTER_TYPE_MAX, 10045 BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); 10046 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10047 FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); 10048 } else { 10049 setup_time_filter(&bbr->r_ctl.rc_delrate, 10050 FILTER_TYPE_MAX, 10051 bbr_num_pktepo_for_del_limit); 10052 setup_time_filter_small(&bbr->r_ctl.rc_rttprop, 10053 FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); 10054 } 10055 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); 10056 if (bbr_uses_idle_restart) 10057 bbr->rc_use_idle_restart = 1; 10058 else 10059 bbr->rc_use_idle_restart = 0; 10060 bbr->r_ctl.rc_bbr_cur_del_rate = 0; 10061 bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; 10062 if (bbr_resends_use_tso) 10063 bbr->rc_resends_use_tso = 1; 10064 if (tp->snd_una != tp->snd_max) { 10065 /* Create a send map for the current outstanding data */ 10066 struct bbr_sendmap *rsm; 10067 10068 rsm = bbr_alloc(bbr); 10069 if (rsm == NULL) { 10070 uma_zfree(bbr_pcb_zone, *ptr); 10071 *ptr = NULL; 10072 return (ENOMEM); 10073 } 10074 rsm->r_rtt_not_allowed = 1; 10075 rsm->r_tim_lastsent[0] = cts; 10076 rsm->r_rtr_cnt = 1; 10077 rsm->r_rtr_bytes = 0; 10078 rsm->r_start = tp->snd_una; 10079 rsm->r_end = tp->snd_max; 10080 rsm->r_dupack = 0; 10081 rsm->r_delivered = bbr->r_ctl.rc_delivered; 10082 rsm->r_ts_valid = 0; 10083 rsm->r_del_ack_ts = tp->ts_recent; 10084 rsm->r_del_time = cts; 10085 if (bbr->r_ctl.r_app_limited_until) 10086 rsm->r_app_limited = 1; 10087 else 10088 rsm->r_app_limited = 0; 10089 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); 10090 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); 10091 rsm->r_in_tmap = 1; 10092 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) 10093 rsm->r_bbr_state = bbr_state_val(bbr); 10094 else 10095 rsm->r_bbr_state = 8; 10096 } 10097 if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) 10098 bbr->bbr_use_rack_cheat = 1; 10099 if (bbr_incr_timers && (bbr->rc_use_google == 0)) 10100 bbr->r_ctl.rc_incr_tmrs = 1; 10101 if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) 10102 bbr->r_ctl.rc_inc_tcp_oh = 1; 10103 if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) 10104 bbr->r_ctl.rc_inc_ip_oh = 1; 10105 if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) 10106 bbr->r_ctl.rc_inc_enet_oh = 1; 10107 10108 bbr_log_type_statechange(bbr, cts, __LINE__); 10109 if (TCPS_HAVEESTABLISHED(tp->t_state) && 10110 (tp->t_srtt)) { 10111 uint32_t rtt; 10112 10113 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); 10114 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); 10115 } 10116 /* announce the settings and state */ 10117 bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); 10118 tcp_bbr_tso_size_check(bbr, cts); 10119 /* 10120 * Now call the generic function to start a timer. This will place 10121 * the TCB on the hptsi wheel if a timer is needed with appropriate 10122 * flags. 10123 */ 10124 bbr_stop_all_timers(tp, bbr); 10125 /* 10126 * Validate the timers are not in usec, if they are convert. 10127 * BBR should in theory move to USEC and get rid of a 10128 * lot of the TICKS_2 calls.. but for now we stay 10129 * with tick timers. 10130 */ 10131 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS); 10132 TCPT_RANGESET(tp->t_rxtcur, 10133 ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, 10134 tp->t_rttmin, TCPTV_REXMTMAX); 10135 bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); 10136 return (0); 10137 } 10138 10139 /* 10140 * Return 0 if we can accept the connection. Return 10141 * non-zero if we can't handle the connection. A EAGAIN 10142 * means you need to wait until the connection is up. 10143 * a EADDRNOTAVAIL means we can never handle the connection 10144 * (no SACK). 10145 */ 10146 static int 10147 bbr_handoff_ok(struct tcpcb *tp) 10148 { 10149 if ((tp->t_state == TCPS_CLOSED) || 10150 (tp->t_state == TCPS_LISTEN)) { 10151 /* Sure no problem though it may not stick */ 10152 return (0); 10153 } 10154 if ((tp->t_state == TCPS_SYN_SENT) || 10155 (tp->t_state == TCPS_SYN_RECEIVED)) { 10156 /* 10157 * We really don't know you have to get to ESTAB or beyond 10158 * to tell. 10159 */ 10160 return (EAGAIN); 10161 } 10162 if (tp->t_flags & TF_SENTFIN) 10163 return (EINVAL); 10164 if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { 10165 return (0); 10166 } 10167 /* 10168 * If we reach here we don't do SACK on this connection so we can 10169 * never do rack. 10170 */ 10171 return (EINVAL); 10172 } 10173 10174 static void 10175 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) 10176 { 10177 if (tp->t_fb_ptr) { 10178 uint32_t calc; 10179 struct tcp_bbr *bbr; 10180 struct bbr_sendmap *rsm; 10181 10182 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 10183 if (bbr->r_ctl.crte) 10184 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 10185 bbr_log_flowend(bbr); 10186 bbr->rc_tp = NULL; 10187 if (bbr->bbr_hdrw_pacing) 10188 counter_u64_add(bbr_flows_whdwr_pacing, -1); 10189 else 10190 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 10191 if (bbr->r_ctl.crte != NULL) { 10192 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 10193 bbr->r_ctl.crte = NULL; 10194 } 10195 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10196 while (rsm) { 10197 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); 10198 uma_zfree(bbr_zone, rsm); 10199 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 10200 } 10201 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10202 while (rsm) { 10203 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); 10204 uma_zfree(bbr_zone, rsm); 10205 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); 10206 } 10207 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; 10208 if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) 10209 BBR_STAT_INC(bbr_dynamic_rwnd); 10210 else 10211 BBR_STAT_INC(bbr_static_rwnd); 10212 bbr->r_ctl.rc_free_cnt = 0; 10213 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); 10214 tp->t_fb_ptr = NULL; 10215 } 10216 /* Make sure snd_nxt is correctly set */ 10217 tp->snd_nxt = tp->snd_max; 10218 } 10219 10220 static void 10221 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) 10222 { 10223 switch (tp->t_state) { 10224 case TCPS_SYN_SENT: 10225 bbr->r_state = TCPS_SYN_SENT; 10226 bbr->r_substate = bbr_do_syn_sent; 10227 break; 10228 case TCPS_SYN_RECEIVED: 10229 bbr->r_state = TCPS_SYN_RECEIVED; 10230 bbr->r_substate = bbr_do_syn_recv; 10231 break; 10232 case TCPS_ESTABLISHED: 10233 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); 10234 bbr->r_state = TCPS_ESTABLISHED; 10235 bbr->r_substate = bbr_do_established; 10236 break; 10237 case TCPS_CLOSE_WAIT: 10238 bbr->r_state = TCPS_CLOSE_WAIT; 10239 bbr->r_substate = bbr_do_close_wait; 10240 break; 10241 case TCPS_FIN_WAIT_1: 10242 bbr->r_state = TCPS_FIN_WAIT_1; 10243 bbr->r_substate = bbr_do_fin_wait_1; 10244 break; 10245 case TCPS_CLOSING: 10246 bbr->r_state = TCPS_CLOSING; 10247 bbr->r_substate = bbr_do_closing; 10248 break; 10249 case TCPS_LAST_ACK: 10250 bbr->r_state = TCPS_LAST_ACK; 10251 bbr->r_substate = bbr_do_lastack; 10252 break; 10253 case TCPS_FIN_WAIT_2: 10254 bbr->r_state = TCPS_FIN_WAIT_2; 10255 bbr->r_substate = bbr_do_fin_wait_2; 10256 break; 10257 case TCPS_LISTEN: 10258 case TCPS_CLOSED: 10259 case TCPS_TIME_WAIT: 10260 default: 10261 break; 10262 }; 10263 } 10264 10265 static void 10266 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) 10267 { 10268 /* 10269 * Now what state are we going into now? Is there adjustments 10270 * needed? 10271 */ 10272 int32_t old_state; 10273 10274 old_state = bbr_state_val(bbr); 10275 if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { 10276 /* Save the lowest srtt we saw in our end of the sub-state */ 10277 bbr->rc_hit_state_1 = 0; 10278 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) 10279 bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; 10280 } 10281 bbr->rc_bbr_substate++; 10282 if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { 10283 /* Cycle back to first state-> gain */ 10284 bbr->rc_bbr_substate = 0; 10285 } 10286 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10287 /* 10288 * We enter the gain(5/4) cycle (possibly less if 10289 * shallow buffer detection is enabled) 10290 */ 10291 if (bbr->skip_gain) { 10292 /* 10293 * Hardware pacing has set our rate to 10294 * the max and limited our b/w just 10295 * do level i.e. no gain. 10296 */ 10297 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; 10298 } else if (bbr->gain_is_limited && 10299 bbr->bbr_hdrw_pacing && 10300 bbr->r_ctl.crte) { 10301 /* 10302 * We can't gain above the hardware pacing 10303 * rate which is less than our rate + the gain 10304 * calculate the gain needed to reach the hardware 10305 * pacing rate.. 10306 */ 10307 uint64_t bw, rate, gain_calc; 10308 10309 bw = bbr_get_bw(bbr); 10310 rate = bbr->r_ctl.crte->rate; 10311 if ((rate > bw) && 10312 (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { 10313 gain_calc = (rate * BBR_UNIT) / bw; 10314 if (gain_calc < BBR_UNIT) 10315 gain_calc = BBR_UNIT; 10316 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; 10317 } else { 10318 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10319 } 10320 } else 10321 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; 10322 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { 10323 bbr->r_ctl.rc_bbr_state_atflight = cts; 10324 } else 10325 bbr->r_ctl.rc_bbr_state_atflight = 0; 10326 } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10327 bbr->rc_hit_state_1 = 1; 10328 bbr->r_ctl.rc_exta_time_gd = 0; 10329 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10330 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10331 if (bbr_state_drain_2_tar) { 10332 bbr->r_ctl.rc_bbr_state_atflight = 0; 10333 } else 10334 bbr->r_ctl.rc_bbr_state_atflight = cts; 10335 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; 10336 } else { 10337 /* All other cycles hit here 2-7 */ 10338 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { 10339 if (bbr_sub_drain_slam_cwnd && 10340 (bbr->rc_use_google == 0) && 10341 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10342 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10343 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10344 } 10345 if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) 10346 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - 10347 bbr_get_rtt(bbr, BBR_RTT_PROP)); 10348 else 10349 bbr->r_ctl.rc_exta_time_gd = 0; 10350 if (bbr->r_ctl.rc_exta_time_gd) { 10351 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; 10352 /* Now chop up the time for each state (div by 7) */ 10353 bbr->r_ctl.rc_level_state_extra /= 7; 10354 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { 10355 /* Add a randomization */ 10356 bbr_randomize_extra_state_time(bbr); 10357 } 10358 } 10359 } 10360 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10361 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; 10362 } 10363 if (bbr->rc_use_google) { 10364 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); 10365 } 10366 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10367 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10368 if (dolog) 10369 bbr_log_type_statechange(bbr, cts, line); 10370 10371 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10372 uint32_t time_in; 10373 10374 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10375 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 10376 counter_u64_add(bbr_state_time[(old_state + 5)], time_in); 10377 } else { 10378 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10379 } 10380 } 10381 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 10382 bbr_set_state_target(bbr, __LINE__); 10383 if (bbr_sub_drain_slam_cwnd && 10384 (bbr->rc_use_google == 0) && 10385 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10386 /* Slam down the cwnd */ 10387 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10388 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10389 if (bbr_sub_drain_app_limit) { 10390 /* Go app limited if we are on a long drain */ 10391 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + 10392 ctf_flight_size(bbr->rc_tp, 10393 (bbr->r_ctl.rc_sacked + 10394 bbr->r_ctl.rc_lost_bytes))); 10395 } 10396 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10397 } 10398 if (bbr->rc_lt_use_bw) { 10399 /* In policed mode we clamp pacing_gain to BBR_UNIT */ 10400 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10401 } 10402 /* Google changes TSO size every cycle */ 10403 if (bbr->rc_use_google) 10404 tcp_bbr_tso_size_check(bbr, cts); 10405 bbr->r_ctl.gain_epoch = cts; 10406 bbr->r_ctl.rc_bbr_state_time = cts; 10407 bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; 10408 } 10409 10410 static void 10411 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10412 { 10413 if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && 10414 (google_allow_early_out == 1) && 10415 (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { 10416 /* We have reached out target flight size possibly early */ 10417 goto change_state; 10418 } 10419 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10420 return; 10421 } 10422 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { 10423 /* 10424 * Must be a rttProp movement forward before 10425 * we can change states. 10426 */ 10427 return; 10428 } 10429 if (bbr_state_val(bbr) == BBR_SUB_GAIN) { 10430 /* 10431 * The needed time has passed but for 10432 * the gain cycle extra rules apply: 10433 * 1) If we have seen loss, we exit 10434 * 2) If we have not reached the target 10435 * we stay in GAIN (gain-to-target). 10436 */ 10437 if (google_consider_lost && losses) 10438 goto change_state; 10439 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { 10440 return; 10441 } 10442 } 10443 change_state: 10444 /* For gain we must reach our target, all others last 1 rttProp */ 10445 bbr_substate_change(bbr, cts, __LINE__, 1); 10446 } 10447 10448 static void 10449 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) 10450 { 10451 uint32_t flight, bbr_cur_cycle_time; 10452 10453 if (bbr->rc_use_google) { 10454 bbr_set_probebw_google_gains(bbr, cts, losses); 10455 return; 10456 } 10457 if (cts == 0) { 10458 /* 10459 * Never alow cts to be 0 we 10460 * do this so we can judge if 10461 * we have set a timestamp. 10462 */ 10463 cts = 1; 10464 } 10465 if (bbr_state_is_pkt_epoch) 10466 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); 10467 else 10468 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); 10469 10470 if (bbr->r_ctl.rc_bbr_state_atflight == 0) { 10471 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { 10472 flight = ctf_flight_size(bbr->rc_tp, 10473 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10474 if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { 10475 /* Keep it slam down */ 10476 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { 10477 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10478 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10479 } 10480 if (bbr_sub_drain_app_limit) { 10481 /* Go app limited if we are on a long drain */ 10482 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); 10483 } 10484 } 10485 if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && 10486 (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || 10487 (flight >= bbr->r_ctl.flightsize_at_drain))) { 10488 /* 10489 * Still here after the same time as 10490 * the gain. We need to drain harder 10491 * for the next srtt. Reduce by a set amount 10492 * the gain drop is capped at DRAIN states 10493 * value (88). 10494 */ 10495 bbr->r_ctl.flightsize_at_drain = flight; 10496 if (bbr_drain_drop_mul && 10497 bbr_drain_drop_div && 10498 (bbr_drain_drop_mul < bbr_drain_drop_div)) { 10499 /* Use your specific drop value (def 4/5 = 20%) */ 10500 bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; 10501 bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; 10502 } else { 10503 /* You get drop of 20% */ 10504 bbr->r_ctl.rc_bbr_hptsi_gain *= 4; 10505 bbr->r_ctl.rc_bbr_hptsi_gain /= 5; 10506 } 10507 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { 10508 /* Reduce our gain again to the bottom */ 10509 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 10510 } 10511 bbr_log_exit_gain(bbr, cts, 4); 10512 /* 10513 * Extend out so we wait another 10514 * epoch before dropping again. 10515 */ 10516 bbr->r_ctl.gain_epoch = cts; 10517 } 10518 if (flight <= bbr->r_ctl.rc_target_at_state) { 10519 if (bbr_sub_drain_slam_cwnd && 10520 (bbr->rc_use_google == 0) && 10521 (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 10522 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10523 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10524 } 10525 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10526 bbr_log_exit_gain(bbr, cts, 3); 10527 } 10528 } else { 10529 /* Its a gain */ 10530 if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { 10531 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10532 goto change_state; 10533 } 10534 if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || 10535 ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= 10536 bbr->rc_tp->snd_wnd)) { 10537 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); 10538 bbr_log_exit_gain(bbr, cts, 2); 10539 } 10540 } 10541 /** 10542 * We fall through and return always one of two things has 10543 * occurred. 10544 * 1) We are still not at target 10545 * <or> 10546 * 2) We reached the target and set rc_bbr_state_atflight 10547 * which means we no longer hit this block 10548 * next time we are called. 10549 */ 10550 return; 10551 } 10552 change_state: 10553 if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) 10554 return; 10555 if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { 10556 /* Less than a full time-period has passed */ 10557 return; 10558 } 10559 if (bbr->r_ctl.rc_level_state_extra && 10560 (bbr_state_val(bbr) > BBR_SUB_DRAIN) && 10561 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10562 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10563 /* Less than a full time-period + extra has passed */ 10564 return; 10565 } 10566 if (bbr_gain_gets_extra_too && 10567 bbr->r_ctl.rc_level_state_extra && 10568 (bbr_state_val(bbr) == BBR_SUB_GAIN) && 10569 ((cts - bbr->r_ctl.rc_bbr_state_time) < 10570 (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { 10571 /* Less than a full time-period + extra has passed */ 10572 return; 10573 } 10574 bbr_substate_change(bbr, cts, __LINE__, 1); 10575 } 10576 10577 static uint32_t 10578 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) 10579 { 10580 uint32_t mss, tar; 10581 10582 if (bbr->rc_use_google) { 10583 /* Google just uses the cwnd target */ 10584 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); 10585 } else { 10586 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), 10587 bbr->r_ctl.rc_pace_max_segs); 10588 /* Get the base cwnd with gain rounded to a mss */ 10589 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), 10590 gain), mss); 10591 /* Make sure it is within our min */ 10592 if (tar < get_min_cwnd(bbr)) 10593 return (get_min_cwnd(bbr)); 10594 } 10595 return (tar); 10596 } 10597 10598 static void 10599 bbr_set_state_target(struct tcp_bbr *bbr, int line) 10600 { 10601 uint32_t tar, meth; 10602 10603 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && 10604 ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { 10605 /* Special case using old probe-rtt method */ 10606 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10607 meth = 1; 10608 } else { 10609 /* Non-probe-rtt case and reduced probe-rtt */ 10610 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && 10611 (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { 10612 /* For gain cycle we use the hptsi gain */ 10613 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10614 meth = 2; 10615 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { 10616 /* 10617 * If configured, or for google all other states 10618 * get BBR_UNIT. 10619 */ 10620 tar = bbr_get_a_state_target(bbr, BBR_UNIT); 10621 meth = 3; 10622 } else { 10623 /* 10624 * Or we set a target based on the pacing gain 10625 * for non-google mode and default (non-configured). 10626 * Note we don't set a target goal below drain (192). 10627 */ 10628 if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { 10629 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); 10630 meth = 4; 10631 } else { 10632 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); 10633 meth = 5; 10634 } 10635 } 10636 } 10637 bbr_log_set_of_state_target(bbr, tar, line, meth); 10638 bbr->r_ctl.rc_target_at_state = tar; 10639 } 10640 10641 static void 10642 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) 10643 { 10644 /* Change to probe_rtt */ 10645 uint32_t time_in; 10646 10647 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10648 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, 10649 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 10650 bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain 10651 + bbr->r_ctl.rc_delivered); 10652 /* Setup so we force feed the filter */ 10653 if (bbr->rc_use_google || bbr_probertt_sets_rtt) 10654 bbr->rc_prtt_set_ts = 1; 10655 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10656 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10657 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10658 } 10659 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); 10660 bbr->r_ctl.rc_rtt_shrinks = cts; 10661 bbr->r_ctl.last_in_probertt = cts; 10662 bbr->r_ctl.rc_probertt_srttchktim = cts; 10663 bbr->r_ctl.rc_bbr_state_time = cts; 10664 bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; 10665 /* We need to force the filter to update */ 10666 10667 if ((bbr_sub_drain_slam_cwnd) && 10668 bbr->rc_hit_state_1 && 10669 (bbr->rc_use_google == 0) && 10670 (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { 10671 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) 10672 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10673 } else 10674 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 10675 /* Update the lost */ 10676 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10677 if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ 10678 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ 10679 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 10680 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10681 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 10682 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10683 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); 10684 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; 10685 } else { 10686 /* 10687 * We bring it down slowly by using a hptsi gain that is 10688 * probably 75%. This will slowly float down our outstanding 10689 * without tampering with the cwnd. 10690 */ 10691 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 10692 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 10693 bbr_set_state_target(bbr, __LINE__); 10694 if (bbr_prtt_slam_cwnd && 10695 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 10696 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 10697 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10698 } 10699 } 10700 if (ctf_flight_size(bbr->rc_tp, 10701 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 10702 bbr->r_ctl.rc_target_at_state) { 10703 /* We are at target */ 10704 bbr->r_ctl.rc_bbr_enters_probertt = cts; 10705 } else { 10706 /* We need to come down to reach target before our time begins */ 10707 bbr->r_ctl.rc_bbr_enters_probertt = 0; 10708 } 10709 bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; 10710 BBR_STAT_INC(bbr_enter_probertt); 10711 bbr_log_exit_gain(bbr, cts, 0); 10712 bbr_log_type_statechange(bbr, cts, line); 10713 } 10714 10715 static void 10716 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) 10717 { 10718 /* 10719 * Sanity check on probe-rtt intervals. 10720 * In crazy situations where we are competing 10721 * against new-reno flows with huge buffers 10722 * our rtt-prop interval could come to dominate 10723 * things if we can't get through a full set 10724 * of cycles, we need to adjust it. 10725 */ 10726 if (bbr_can_adjust_probertt && 10727 (bbr->rc_use_google == 0)) { 10728 uint16_t val = 0; 10729 uint32_t cur_rttp, fval, newval, baseval; 10730 10731 /* Are we to small and go into probe-rtt to often? */ 10732 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); 10733 cur_rttp = roundup(baseval, USECS_IN_SECOND); 10734 fval = bbr_filter_len_sec * USECS_IN_SECOND; 10735 if (bbr_is_ratio == 0) { 10736 if (fval > bbr_rtt_probe_limit) 10737 newval = cur_rttp + (fval - bbr_rtt_probe_limit); 10738 else 10739 newval = cur_rttp; 10740 } else { 10741 int mul; 10742 10743 mul = fval / bbr_rtt_probe_limit; 10744 newval = cur_rttp * mul; 10745 } 10746 if (cur_rttp > bbr->r_ctl.rc_probertt_int) { 10747 bbr->r_ctl.rc_probertt_int = cur_rttp; 10748 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10749 val = 1; 10750 } else { 10751 /* 10752 * No adjustments were made 10753 * do we need to shrink it? 10754 */ 10755 if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { 10756 if (cur_rttp <= bbr_rtt_probe_limit) { 10757 /* 10758 * Things have calmed down lets 10759 * shrink all the way to default 10760 */ 10761 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; 10762 reset_time_small(&bbr->r_ctl.rc_rttprop, 10763 (bbr_filter_len_sec * USECS_IN_SECOND)); 10764 cur_rttp = bbr_rtt_probe_limit; 10765 newval = (bbr_filter_len_sec * USECS_IN_SECOND); 10766 val = 2; 10767 } else { 10768 /* 10769 * Well does some adjustment make sense? 10770 */ 10771 if (cur_rttp < bbr->r_ctl.rc_probertt_int) { 10772 /* We can reduce interval time some */ 10773 bbr->r_ctl.rc_probertt_int = cur_rttp; 10774 reset_time_small(&bbr->r_ctl.rc_rttprop, newval); 10775 val = 3; 10776 } 10777 } 10778 } 10779 } 10780 if (val) 10781 bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); 10782 } 10783 } 10784 10785 static void 10786 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) 10787 { 10788 /* Exit probe-rtt */ 10789 10790 if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { 10791 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 10792 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 10793 } 10794 bbr_log_exit_gain(bbr, cts, 1); 10795 bbr->rc_hit_state_1 = 0; 10796 bbr->r_ctl.rc_rtt_shrinks = cts; 10797 bbr->r_ctl.last_in_probertt = cts; 10798 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); 10799 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 10800 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, 10801 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 10802 bbr->r_ctl.rc_delivered); 10803 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 10804 uint32_t time_in; 10805 10806 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 10807 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 10808 } 10809 if (bbr->rc_filled_pipe) { 10810 /* Switch to probe_bw */ 10811 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 10812 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 10813 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; 10814 bbr_substate_change(bbr, cts, __LINE__, 0); 10815 bbr_log_type_statechange(bbr, cts, __LINE__); 10816 } else { 10817 /* Back to startup */ 10818 bbr->rc_bbr_state = BBR_STATE_STARTUP; 10819 bbr->r_ctl.rc_bbr_state_time = cts; 10820 /* 10821 * We don't want to give a complete free 3 10822 * measurements until we exit, so we use 10823 * the number of pe's we were in probe-rtt 10824 * to add to the startup_epoch. That way 10825 * we will still retain the old state. 10826 */ 10827 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); 10828 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10829 /* Make sure to use the lower pg when shifting back in */ 10830 if (bbr->r_ctl.rc_lost && 10831 bbr_use_lower_gain_in_startup && 10832 (bbr->rc_use_google == 0)) 10833 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10834 else 10835 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; 10836 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; 10837 /* Probably not needed but set it anyway */ 10838 bbr_set_state_target(bbr, __LINE__); 10839 bbr_log_type_statechange(bbr, cts, __LINE__); 10840 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10841 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); 10842 } 10843 bbr_check_probe_rtt_limits(bbr, cts); 10844 } 10845 10846 static int32_t inline 10847 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) 10848 { 10849 if ((bbr->rc_past_init_win == 1) && 10850 (bbr->rc_in_persist == 0) && 10851 (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { 10852 return (1); 10853 } 10854 if (bbr_can_force_probertt && 10855 (bbr->rc_in_persist == 0) && 10856 (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && 10857 ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { 10858 return (1); 10859 } 10860 return (0); 10861 } 10862 10863 static int32_t 10864 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) 10865 { 10866 uint64_t btlbw, gain; 10867 if (pkt_epoch == 0) { 10868 /* 10869 * Need to be on a pkt-epoch to continue. 10870 */ 10871 return (0); 10872 } 10873 btlbw = bbr_get_full_bw(bbr); 10874 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10875 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10876 if (btlbw >= gain) { 10877 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10878 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10879 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10880 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10881 } 10882 if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) 10883 return (1); 10884 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10885 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 10886 return(0); 10887 } 10888 10889 static int32_t inline 10890 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) 10891 { 10892 /* Have we gained 25% in the last 3 packet based epoch's? */ 10893 uint64_t btlbw, gain; 10894 int do_exit; 10895 int delta, rtt_gain; 10896 10897 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 10898 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 10899 /* 10900 * This qualifies as a RTT_PROBE session since we drop the 10901 * data outstanding to nothing and waited more than 10902 * bbr_rtt_probe_time. 10903 */ 10904 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 10905 bbr_set_reduced_rtt(bbr, cts, __LINE__); 10906 } 10907 if (bbr_should_enter_probe_rtt(bbr, cts)) { 10908 bbr_enter_probe_rtt(bbr, cts, __LINE__); 10909 return (0); 10910 } 10911 if (bbr->rc_use_google) 10912 return (bbr_google_startup(bbr, cts, pkt_epoch)); 10913 10914 if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10915 (bbr_use_lower_gain_in_startup)) { 10916 /* Drop to a lower gain 1.5 x since we saw loss */ 10917 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; 10918 } 10919 if (pkt_epoch == 0) { 10920 /* 10921 * Need to be on a pkt-epoch to continue. 10922 */ 10923 return (0); 10924 } 10925 if (bbr_rtt_gain_thresh) { 10926 /* 10927 * Do we allow a flow to stay 10928 * in startup with no loss and no 10929 * gain in rtt over a set threshold? 10930 */ 10931 if (bbr->r_ctl.rc_pkt_epoch_rtt && 10932 bbr->r_ctl.startup_last_srtt && 10933 (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { 10934 delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; 10935 rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; 10936 } else 10937 rtt_gain = 0; 10938 if ((bbr->r_ctl.startup_last_srtt == 0) || 10939 (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) 10940 /* First time or new lower value */ 10941 bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; 10942 10943 if ((bbr->r_ctl.rc_lost == 0) && 10944 (rtt_gain < bbr_rtt_gain_thresh)) { 10945 /* 10946 * No loss, and we are under 10947 * our gain threhold for 10948 * increasing RTT. 10949 */ 10950 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10951 bbr->r_ctl.rc_bbr_last_startup_epoch++; 10952 bbr_log_startup_event(bbr, cts, rtt_gain, 10953 delta, bbr->r_ctl.startup_last_srtt, 10); 10954 return (0); 10955 } 10956 } 10957 if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && 10958 (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && 10959 (!IN_RECOVERY(bbr->rc_tp->t_flags))) { 10960 /* 10961 * We only assess if we have a new measurement when 10962 * we have no loss and are not in recovery. 10963 * Drag up by one our last_startup epoch so we will hold 10964 * the number of non-gain we have already accumulated. 10965 */ 10966 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) 10967 bbr->r_ctl.rc_bbr_last_startup_epoch++; 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, 9); 10970 return (0); 10971 } 10972 /* Case where we reduced the lost (bad retransmit) */ 10973 if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) 10974 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10975 bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; 10976 btlbw = bbr_get_full_bw(bbr); 10977 if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) 10978 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10979 (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10980 else 10981 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * 10982 (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; 10983 do_exit = 0; 10984 if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) 10985 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; 10986 if (btlbw >= gain) { 10987 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; 10988 /* Update the lost so we won't exit in next set of tests */ 10989 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 10990 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 10991 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); 10992 } 10993 if ((bbr->rc_loss_exit && 10994 (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && 10995 (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && 10996 ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { 10997 /* 10998 * If we had no gain, we had loss and that loss was above 10999 * our threshould, the rwnd is not constrained, and we have 11000 * had at least 3 packet epochs exit. Note that this is 11001 * switched off by sysctl. Google does not do this by the 11002 * way. 11003 */ 11004 if ((ctf_flight_size(bbr->rc_tp, 11005 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + 11006 (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { 11007 do_exit = 1; 11008 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11009 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); 11010 } else { 11011 /* Just record an updated loss value */ 11012 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11013 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11014 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); 11015 } 11016 } else 11017 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; 11018 if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || 11019 do_exit) { 11020 /* Return 1 to exit the startup state. */ 11021 return (1); 11022 } 11023 /* Stay in startup */ 11024 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11025 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); 11026 return (0); 11027 } 11028 11029 static void 11030 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) 11031 { 11032 /* 11033 * A tick occurred in the rtt epoch do we need to do anything? 11034 */ 11035 #ifdef BBR_INVARIANTS 11036 if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && 11037 (bbr->rc_bbr_state != BBR_STATE_DRAIN) && 11038 (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && 11039 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 11040 (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { 11041 /* Debug code? */ 11042 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); 11043 } 11044 #endif 11045 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 11046 /* Do we exit the startup state? */ 11047 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { 11048 uint32_t time_in; 11049 11050 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, 11051 bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); 11052 bbr->rc_filled_pipe = 1; 11053 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11054 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11055 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11056 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11057 } else 11058 time_in = 0; 11059 if (bbr->rc_no_pacing) 11060 bbr->rc_no_pacing = 0; 11061 bbr->r_ctl.rc_bbr_state_time = cts; 11062 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; 11063 bbr->rc_bbr_state = BBR_STATE_DRAIN; 11064 bbr_set_state_target(bbr, __LINE__); 11065 if ((bbr->rc_use_google == 0) && 11066 bbr_slam_cwnd_in_main_drain) { 11067 /* Here we don't have to worry about probe-rtt */ 11068 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; 11069 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11070 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11071 } 11072 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; 11073 bbr_log_type_statechange(bbr, cts, __LINE__); 11074 if (ctf_flight_size(bbr->rc_tp, 11075 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= 11076 bbr->r_ctl.rc_target_at_state) { 11077 /* 11078 * Switch to probe_bw if we are already 11079 * there 11080 */ 11081 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11082 bbr_substate_change(bbr, cts, __LINE__, 0); 11083 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11084 bbr_log_type_statechange(bbr, cts, __LINE__); 11085 } 11086 } 11087 } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { 11088 uint32_t inflight; 11089 struct tcpcb *tp; 11090 11091 tp = bbr->rc_tp; 11092 inflight = ctf_flight_size(tp, 11093 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11094 if (inflight >= bbr->r_ctl.rc_target_at_state) { 11095 /* We have reached a flight of the cwnd target */ 11096 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11097 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11098 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; 11099 bbr_set_state_target(bbr, __LINE__); 11100 /* 11101 * Rig it so we don't do anything crazy and 11102 * start fresh with a new randomization. 11103 */ 11104 bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; 11105 bbr->rc_bbr_substate = BBR_SUB_LEVEL6; 11106 bbr_substate_change(bbr, cts, __LINE__, 1); 11107 } 11108 } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { 11109 /* Has in-flight reached the bdp (or less)? */ 11110 uint32_t inflight; 11111 struct tcpcb *tp; 11112 11113 tp = bbr->rc_tp; 11114 inflight = ctf_flight_size(tp, 11115 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11116 if ((bbr->rc_use_google == 0) && 11117 bbr_slam_cwnd_in_main_drain && 11118 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11119 /* 11120 * Here we don't have to worry about probe-rtt 11121 * re-slam it, but keep it slammed down. 11122 */ 11123 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11124 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11125 } 11126 if (inflight <= bbr->r_ctl.rc_target_at_state) { 11127 /* We have drained */ 11128 bbr->rc_bbr_state = BBR_STATE_PROBE_BW; 11129 bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; 11130 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { 11131 uint32_t time_in; 11132 11133 time_in = cts - bbr->r_ctl.rc_bbr_state_time; 11134 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); 11135 } 11136 if ((bbr->rc_use_google == 0) && 11137 bbr_slam_cwnd_in_main_drain && 11138 (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { 11139 /* Restore the cwnd */ 11140 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; 11141 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11142 } 11143 /* Setup probe-rtt has being done now RRS-HERE */ 11144 bbr->r_ctl.rc_rtt_shrinks = cts; 11145 bbr->r_ctl.last_in_probertt = cts; 11146 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); 11147 /* Randomly pick a sub-state */ 11148 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); 11149 bbr_substate_change(bbr, cts, __LINE__, 0); 11150 bbr_log_type_statechange(bbr, cts, __LINE__); 11151 } 11152 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { 11153 uint32_t flight; 11154 11155 flight = ctf_flight_size(bbr->rc_tp, 11156 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11157 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); 11158 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && 11159 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11160 /* 11161 * We must keep cwnd at the desired MSS. 11162 */ 11163 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); 11164 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11165 } else if ((bbr_prtt_slam_cwnd) && 11166 (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { 11167 /* Re-slam it */ 11168 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; 11169 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); 11170 } 11171 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { 11172 /* Has outstanding reached our target? */ 11173 if (flight <= bbr->r_ctl.rc_target_at_state) { 11174 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); 11175 bbr->r_ctl.rc_bbr_enters_probertt = cts; 11176 /* If time is exactly 0, be 1usec off */ 11177 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) 11178 bbr->r_ctl.rc_bbr_enters_probertt = 1; 11179 if (bbr->rc_use_google == 0) { 11180 /* 11181 * Restore any lowering that as occurred to 11182 * reach here 11183 */ 11184 if (bbr->r_ctl.bbr_rttprobe_gain_val) 11185 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; 11186 else 11187 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; 11188 } 11189 } 11190 if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && 11191 (bbr->rc_use_google == 0) && 11192 bbr->r_ctl.bbr_rttprobe_gain_val && 11193 (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || 11194 (flight >= bbr->r_ctl.flightsize_at_drain))) { 11195 /* 11196 * We have doddled with our current hptsi 11197 * gain an srtt and have still not made it 11198 * to target, or we have increased our flight. 11199 * Lets reduce the gain by xx% 11200 * flooring the reduce at DRAIN (based on 11201 * mul/div) 11202 */ 11203 int red; 11204 11205 bbr->r_ctl.flightsize_at_drain = flight; 11206 bbr->r_ctl.rc_probertt_srttchktim = cts; 11207 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); 11208 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { 11209 /* Reduce our gain again */ 11210 bbr->r_ctl.rc_bbr_hptsi_gain -= red; 11211 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); 11212 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { 11213 /* one more chance before we give up */ 11214 bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); 11215 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); 11216 } else { 11217 /* At the very bottom */ 11218 bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); 11219 } 11220 } 11221 } 11222 if (bbr->r_ctl.rc_bbr_enters_probertt && 11223 (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && 11224 ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { 11225 /* Time to exit probe RTT normally */ 11226 bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); 11227 } 11228 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { 11229 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && 11230 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 11231 /* 11232 * This qualifies as a RTT_PROBE session since we 11233 * drop the data outstanding to nothing and waited 11234 * more than bbr_rtt_probe_time. 11235 */ 11236 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 11237 bbr_set_reduced_rtt(bbr, cts, __LINE__); 11238 } 11239 if (bbr_should_enter_probe_rtt(bbr, cts)) { 11240 bbr_enter_probe_rtt(bbr, cts, __LINE__); 11241 } else { 11242 bbr_set_probebw_gains(bbr, cts, losses); 11243 } 11244 } 11245 } 11246 11247 static void 11248 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) 11249 { 11250 int32_t epoch = 0; 11251 11252 if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { 11253 bbr_set_epoch(bbr, cts, line); 11254 /* At each epoch doe lt bw sampling */ 11255 epoch = 1; 11256 } 11257 bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); 11258 } 11259 11260 static int 11261 bbr_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, 11262 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, 11263 struct timeval *tv) 11264 { 11265 struct inpcb *inp = tptoinpcb(tp); 11266 struct socket *so = tptosocket(tp); 11267 int32_t thflags, retval; 11268 uint32_t cts, lcts; 11269 uint32_t tiwin; 11270 struct tcpopt to; 11271 struct tcp_bbr *bbr; 11272 struct bbr_sendmap *rsm; 11273 struct timeval ltv; 11274 int32_t did_out = 0; 11275 uint16_t nsegs; 11276 int32_t prev_state; 11277 uint32_t lost; 11278 11279 nsegs = max(1, m->m_pkthdr.lro_nsegs); 11280 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11281 /* add in our stats */ 11282 kern_prefetch(bbr, &prev_state); 11283 prev_state = 0; 11284 thflags = tcp_get_flags(th); 11285 /* 11286 * If this is either a state-changing packet or current state isn't 11287 * established, we require a write lock on tcbinfo. Otherwise, we 11288 * allow the tcbinfo to be in either alocked or unlocked, as the 11289 * caller may have unnecessarily acquired a write lock due to a 11290 * race. 11291 */ 11292 INP_WLOCK_ASSERT(tptoinpcb(tp)); 11293 KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", 11294 __func__)); 11295 KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", 11296 __func__)); 11297 11298 tp->t_rcvtime = ticks; 11299 /* 11300 * Unscale the window into a 32-bit value. For the SYN_SENT state 11301 * the scale is zero. 11302 */ 11303 tiwin = th->th_win << tp->snd_scale; 11304 #ifdef STATS 11305 stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); 11306 #endif 11307 11308 if (m->m_flags & M_TSTMP) { 11309 /* Prefer the hardware timestamp if present */ 11310 struct timespec ts; 11311 11312 mbuf_tstmp2timespec(m, &ts); 11313 bbr->rc_tv.tv_sec = ts.tv_sec; 11314 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11315 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11316 } else if (m->m_flags & M_TSTMP_LRO) { 11317 /* Next the arrival timestamp */ 11318 struct timespec ts; 11319 11320 mbuf_tstmp2timespec(m, &ts); 11321 bbr->rc_tv.tv_sec = ts.tv_sec; 11322 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; 11323 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); 11324 } else { 11325 /* 11326 * Ok just get the current time. 11327 */ 11328 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); 11329 } 11330 /* 11331 * Parse options on any incoming segment. 11332 */ 11333 tcp_dooptions(&to, (u_char *)(th + 1), 11334 (th->th_off << 2) - sizeof(struct tcphdr), 11335 (thflags & TH_SYN) ? TO_SYN : 0); 11336 if (tp->t_flags2 & TF2_PROC_SACK_PROHIBIT) { 11337 /* 11338 * We don't look at sack's from the 11339 * peer because the MSS is too small which 11340 * can subject us to an attack. 11341 */ 11342 to.to_flags &= ~TOF_SACK; 11343 } 11344 /* 11345 * If timestamps were negotiated during SYN/ACK and a 11346 * segment without a timestamp is received, silently drop 11347 * the segment, unless it is a RST segment or missing timestamps are 11348 * tolerated. 11349 * See section 3.2 of RFC 7323. 11350 */ 11351 if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && 11352 ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { 11353 retval = 0; 11354 m_freem(m); 11355 goto done_with_input; 11356 } 11357 /* 11358 * If echoed timestamp is later than the current time, fall back to 11359 * non RFC1323 RTT calculation. Normalize timestamp if syncookies 11360 * were used when this connection was established. 11361 */ 11362 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { 11363 to.to_tsecr -= tp->ts_offset; 11364 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) 11365 to.to_tsecr = 0; 11366 } 11367 /* 11368 * If its the first time in we need to take care of options and 11369 * verify we can do SACK for rack! 11370 */ 11371 if (bbr->r_state == 0) { 11372 /* 11373 * Process options only when we get SYN/ACK back. The SYN 11374 * case for incoming connections is handled in tcp_syncache. 11375 * According to RFC1323 the window field in a SYN (i.e., a 11376 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX 11377 * this is traditional behavior, may need to be cleaned up. 11378 */ 11379 if (bbr->rc_inp == NULL) { 11380 bbr->rc_inp = inp; 11381 } 11382 /* 11383 * We need to init rc_inp here since its not init'd when 11384 * bbr_init is called 11385 */ 11386 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { 11387 if ((to.to_flags & TOF_SCALE) && 11388 (tp->t_flags & TF_REQ_SCALE)) { 11389 tp->t_flags |= TF_RCVD_SCALE; 11390 tp->snd_scale = to.to_wscale; 11391 } else 11392 tp->t_flags &= ~TF_REQ_SCALE; 11393 /* 11394 * Initial send window. It will be updated with the 11395 * next incoming segment to the scaled value. 11396 */ 11397 tp->snd_wnd = th->th_win; 11398 if ((to.to_flags & TOF_TS) && 11399 (tp->t_flags & TF_REQ_TSTMP)) { 11400 tp->t_flags |= TF_RCVD_TSTMP; 11401 tp->ts_recent = to.to_tsval; 11402 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); 11403 } else 11404 tp->t_flags &= ~TF_REQ_TSTMP; 11405 if (to.to_flags & TOF_MSS) 11406 tcp_mss(tp, to.to_mss); 11407 if ((tp->t_flags & TF_SACK_PERMIT) && 11408 (to.to_flags & TOF_SACKPERM) == 0) 11409 tp->t_flags &= ~TF_SACK_PERMIT; 11410 if (tp->t_flags & TF_FASTOPEN) { 11411 if (to.to_flags & TOF_FASTOPEN) { 11412 uint16_t mss; 11413 11414 if (to.to_flags & TOF_MSS) 11415 mss = to.to_mss; 11416 else 11417 if ((inp->inp_vflag & INP_IPV6) != 0) 11418 mss = TCP6_MSS; 11419 else 11420 mss = TCP_MSS; 11421 tcp_fastopen_update_cache(tp, mss, 11422 to.to_tfo_len, to.to_tfo_cookie); 11423 } else 11424 tcp_fastopen_disable_path(tp); 11425 } 11426 } 11427 /* 11428 * At this point we are at the initial call. Here we decide 11429 * if we are doing RACK or not. We do this by seeing if 11430 * TF_SACK_PERMIT is set, if not rack is *not* possible and 11431 * we switch to the default code. 11432 */ 11433 if ((tp->t_flags & TF_SACK_PERMIT) == 0) { 11434 /* Bail */ 11435 tcp_switch_back_to_default(tp); 11436 (*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen, 11437 tlen, iptos); 11438 return (1); 11439 } 11440 /* Set the flag */ 11441 bbr->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0; 11442 tcp_set_hpts(tp); 11443 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); 11444 } 11445 if (thflags & TH_ACK) { 11446 /* Track ack types */ 11447 if (to.to_flags & TOF_SACK) 11448 BBR_STAT_INC(bbr_acks_with_sacks); 11449 else 11450 BBR_STAT_INC(bbr_plain_acks); 11451 } 11452 /* 11453 * This is the one exception case where we set the rack state 11454 * always. All other times (timers etc) we must have a rack-state 11455 * set (so we assure we have done the checks above for SACK). 11456 */ 11457 if (thflags & TH_FIN) 11458 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); 11459 if (bbr->r_state != tp->t_state) 11460 bbr_set_state(tp, bbr, tiwin); 11461 11462 if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) 11463 kern_prefetch(rsm, &prev_state); 11464 prev_state = bbr->r_state; 11465 bbr->rc_ack_was_delayed = 0; 11466 lost = bbr->r_ctl.rc_lost; 11467 bbr->rc_is_pkt_epoch_now = 0; 11468 if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { 11469 /* Get the real time into lcts and figure the real delay */ 11470 lcts = tcp_get_usecs(<v); 11471 if (TSTMP_GT(lcts, cts)) { 11472 bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; 11473 bbr->rc_ack_was_delayed = 1; 11474 if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, 11475 bbr->r_ctl.highest_hdwr_delay)) 11476 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; 11477 } else { 11478 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11479 bbr->rc_ack_was_delayed = 0; 11480 } 11481 } else { 11482 bbr->r_ctl.rc_ack_hdwr_delay = 0; 11483 bbr->rc_ack_was_delayed = 0; 11484 } 11485 bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); 11486 if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { 11487 retval = 0; 11488 m_freem(m); 11489 goto done_with_input; 11490 } 11491 /* 11492 * If a segment with the ACK-bit set arrives in the SYN-SENT state 11493 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. 11494 */ 11495 if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && 11496 (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 11497 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); 11498 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); 11499 return (1); 11500 } 11501 if (tiwin > bbr->r_ctl.rc_high_rwnd) 11502 bbr->r_ctl.rc_high_rwnd = tiwin; 11503 bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, 11504 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11505 bbr->rtt_valid = 0; 11506 if (to.to_flags & TOF_TS) { 11507 bbr->rc_ts_valid = 1; 11508 bbr->r_ctl.last_inbound_ts = to.to_tsval; 11509 } else { 11510 bbr->rc_ts_valid = 0; 11511 bbr->r_ctl.last_inbound_ts = 0; 11512 } 11513 retval = (*bbr->r_substate) (m, th, so, 11514 tp, &to, drop_hdrlen, 11515 tlen, tiwin, thflags, nxt_pkt, iptos); 11516 if (nxt_pkt == 0) 11517 BBR_STAT_INC(bbr_rlock_left_ret0); 11518 else 11519 BBR_STAT_INC(bbr_rlock_left_ret1); 11520 if (retval == 0) { 11521 /* 11522 * If retval is 1 the tcb is unlocked and most likely the tp 11523 * is gone. 11524 */ 11525 INP_WLOCK_ASSERT(inp); 11526 tcp_bbr_xmit_timer_commit(bbr, tp, cts); 11527 if (bbr->rc_is_pkt_epoch_now) 11528 bbr_set_pktepoch(bbr, cts, __LINE__); 11529 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); 11530 if (nxt_pkt == 0) { 11531 if ((bbr->r_wanted_output != 0) || 11532 (tp->t_flags & TF_ACKNOW)) { 11533 11534 bbr->rc_output_starts_timer = 0; 11535 did_out = 1; 11536 if (tcp_output(tp) < 0) 11537 return (1); 11538 } else 11539 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); 11540 } 11541 if ((nxt_pkt == 0) && 11542 ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && 11543 (SEQ_GT(tp->snd_max, tp->snd_una) || 11544 (tp->t_flags & TF_DELACK) || 11545 ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && 11546 (tp->t_state <= TCPS_CLOSING)))) { 11547 /* 11548 * We could not send (probably in the hpts but 11549 * stopped the timer)? 11550 */ 11551 if ((tp->snd_max == tp->snd_una) && 11552 ((tp->t_flags & TF_DELACK) == 0) && 11553 (tcp_in_hpts(tp)) && 11554 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 11555 /* 11556 * keep alive not needed if we are hptsi 11557 * output yet 11558 */ 11559 ; 11560 } else { 11561 if (tcp_in_hpts(tp)) { 11562 tcp_hpts_remove(tp); 11563 if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11564 (TSTMP_GT(lcts, bbr->rc_pacer_started))) { 11565 uint32_t del; 11566 11567 del = lcts - bbr->rc_pacer_started; 11568 if (bbr->r_ctl.rc_last_delay_val > del) { 11569 BBR_STAT_INC(bbr_force_timer_start); 11570 bbr->r_ctl.rc_last_delay_val -= del; 11571 bbr->rc_pacer_started = lcts; 11572 } else { 11573 /* We are late */ 11574 bbr->r_ctl.rc_last_delay_val = 0; 11575 BBR_STAT_INC(bbr_force_output); 11576 if (tcp_output(tp) < 0) 11577 return (1); 11578 } 11579 } 11580 } 11581 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 11582 0); 11583 } 11584 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { 11585 /* Do we have the correct timer running? */ 11586 bbr_timer_audit(tp, bbr, lcts, &so->so_snd); 11587 } 11588 /* Clear the flag, it may have been cleared by output but we may not have */ 11589 if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS)) 11590 tp->t_flags2 &= ~TF2_HPTS_CALLS; 11591 /* Do we have a new state */ 11592 if (bbr->r_state != tp->t_state) 11593 bbr_set_state(tp, bbr, tiwin); 11594 done_with_input: 11595 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); 11596 if (did_out) 11597 bbr->r_wanted_output = 0; 11598 } 11599 return (retval); 11600 } 11601 11602 static void 11603 bbr_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, 11604 int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) 11605 { 11606 struct timeval tv; 11607 int retval; 11608 11609 /* First lets see if we have old packets */ 11610 if (!STAILQ_EMPTY(&tp->t_inqueue)) { 11611 if (ctf_do_queued_segments(tp, 1)) { 11612 m_freem(m); 11613 return; 11614 } 11615 } 11616 if (m->m_flags & M_TSTMP_LRO) { 11617 mbuf_tstmp2timeval(m, &tv); 11618 } else { 11619 /* Should not be should we kassert instead? */ 11620 tcp_get_usecs(&tv); 11621 } 11622 retval = bbr_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 11623 0, &tv); 11624 if (retval == 0) { 11625 INP_WUNLOCK(tptoinpcb(tp)); 11626 } 11627 } 11628 11629 /* 11630 * Return how much data can be sent without violating the 11631 * cwnd or rwnd. 11632 */ 11633 11634 static inline uint32_t 11635 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, 11636 uint32_t avail, int32_t sb_offset, uint32_t cts) 11637 { 11638 uint32_t len; 11639 11640 if (ctf_outstanding(tp) >= tp->snd_wnd) { 11641 /* We never want to go over our peers rcv-window */ 11642 len = 0; 11643 } else { 11644 uint32_t flight; 11645 11646 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); 11647 if (flight >= sendwin) { 11648 /* 11649 * We have in flight what we are allowed by cwnd (if 11650 * it was rwnd blocking it would have hit above out 11651 * >= tp->snd_wnd). 11652 */ 11653 return (0); 11654 } 11655 len = sendwin - flight; 11656 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { 11657 /* We would send too much (beyond the rwnd) */ 11658 len = tp->snd_wnd - ctf_outstanding(tp); 11659 } 11660 if ((len + sb_offset) > avail) { 11661 /* 11662 * We don't have that much in the SB, how much is 11663 * there? 11664 */ 11665 len = avail - sb_offset; 11666 } 11667 } 11668 return (len); 11669 } 11670 11671 static inline void 11672 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) 11673 { 11674 if (error) { 11675 return; 11676 } 11677 if (rsm) { 11678 if (rsm->r_flags & BBR_TLP) { 11679 /* 11680 * TLP should not count in retran count, but in its 11681 * own bin 11682 */ 11683 KMOD_TCPSTAT_INC(tcps_tlpresends); 11684 KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); 11685 } else { 11686 /* Retransmit */ 11687 tp->t_sndrexmitpack++; 11688 KMOD_TCPSTAT_INC(tcps_sndrexmitpack); 11689 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); 11690 #ifdef STATS 11691 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, 11692 len); 11693 #endif 11694 } 11695 /* 11696 * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is 11697 * sub-state 11698 */ 11699 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); 11700 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { 11701 /* Non probe_bw log in 1, 2, or 4. */ 11702 counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); 11703 } else { 11704 /* 11705 * Log our probe state 3, and log also 5-13 to show 11706 * us the recovery sub-state for the send. This 11707 * means that 3 == (5+6+7+8+9+10+11+12+13) 11708 */ 11709 counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); 11710 counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); 11711 } 11712 /* Place in both 16's the totals of retransmitted */ 11713 counter_u64_add(bbr_state_lost[16], len); 11714 counter_u64_add(bbr_state_resend[16], len); 11715 /* Place in 17's the total sent */ 11716 counter_u64_add(bbr_state_resend[17], len); 11717 counter_u64_add(bbr_state_lost[17], len); 11718 11719 } else { 11720 /* New sends */ 11721 KMOD_TCPSTAT_INC(tcps_sndpack); 11722 KMOD_TCPSTAT_ADD(tcps_sndbyte, len); 11723 /* Place in 17's the total sent */ 11724 counter_u64_add(bbr_state_resend[17], len); 11725 counter_u64_add(bbr_state_lost[17], len); 11726 #ifdef STATS 11727 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, 11728 len); 11729 #endif 11730 } 11731 } 11732 11733 static void 11734 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) 11735 { 11736 if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { 11737 /* 11738 * Limit the cwnd to not be above N x the target plus whats 11739 * is outstanding. The target is based on the current b/w 11740 * estimate. 11741 */ 11742 uint32_t target; 11743 11744 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); 11745 target += ctf_outstanding(tp); 11746 target *= bbr_target_cwnd_mult_limit; 11747 if (tp->snd_cwnd > target) 11748 tp->snd_cwnd = target; 11749 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); 11750 } 11751 } 11752 11753 static int 11754 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) 11755 { 11756 /* 11757 * "adv" is the amount we could increase the window, taking into 11758 * account that we are limited by TCP_MAXWIN << tp->rcv_scale. 11759 */ 11760 int32_t adv; 11761 int32_t oldwin; 11762 11763 adv = recwin; 11764 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { 11765 oldwin = (tp->rcv_adv - tp->rcv_nxt); 11766 if (adv > oldwin) 11767 adv -= oldwin; 11768 else { 11769 /* We can't increase the window */ 11770 adv = 0; 11771 } 11772 } else 11773 oldwin = 0; 11774 11775 /* 11776 * If the new window size ends up being the same as or less 11777 * than the old size when it is scaled, then don't force 11778 * a window update. 11779 */ 11780 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) 11781 return (0); 11782 11783 if (adv >= (2 * maxseg) && 11784 (adv >= (so->so_rcv.sb_hiwat / 4) || 11785 recwin <= (so->so_rcv.sb_hiwat / 8) || 11786 so->so_rcv.sb_hiwat <= 8 * maxseg)) { 11787 return (1); 11788 } 11789 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 11790 return (1); 11791 return (0); 11792 } 11793 11794 /* 11795 * Return 0 on success and a errno on failure to send. 11796 * Note that a 0 return may not mean we sent anything 11797 * if the TCB was on the hpts. A non-zero return 11798 * does indicate the error we got from ip[6]_output. 11799 */ 11800 static int 11801 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) 11802 { 11803 struct socket *so; 11804 int32_t len; 11805 uint32_t cts; 11806 uint32_t recwin, sendwin; 11807 int32_t sb_offset; 11808 int32_t flags, abandon, error = 0; 11809 struct tcp_log_buffer *lgb; 11810 struct mbuf *m; 11811 struct mbuf *mb; 11812 uint32_t if_hw_tsomaxsegcount = 0; 11813 uint32_t if_hw_tsomaxsegsize = 0; 11814 uint32_t if_hw_tsomax = 0; 11815 struct ip *ip = NULL; 11816 struct tcp_bbr *bbr; 11817 struct tcphdr *th; 11818 struct udphdr *udp = NULL; 11819 u_char opt[TCP_MAXOLEN]; 11820 unsigned ipoptlen, optlen, hdrlen; 11821 unsigned ulen; 11822 uint32_t bbr_seq; 11823 uint32_t delay_calc=0; 11824 uint8_t doing_tlp = 0; 11825 uint8_t local_options; 11826 #ifdef BBR_INVARIANTS 11827 uint8_t doing_retran_from = 0; 11828 uint8_t picked_up_retran = 0; 11829 #endif 11830 uint8_t wanted_cookie = 0; 11831 uint8_t more_to_rxt=0; 11832 int32_t prefetch_so_done = 0; 11833 int32_t prefetch_rsm = 0; 11834 uint32_t tot_len = 0; 11835 uint32_t maxseg, pace_max_segs, p_maxseg; 11836 int32_t csum_flags = 0; 11837 int32_t hw_tls; 11838 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 11839 unsigned ipsec_optlen = 0; 11840 11841 #endif 11842 volatile int32_t sack_rxmit; 11843 struct bbr_sendmap *rsm = NULL; 11844 int32_t tso, mtu; 11845 struct tcpopt to; 11846 int32_t slot = 0; 11847 struct inpcb *inp; 11848 struct sockbuf *sb; 11849 bool hpts_calling; 11850 #ifdef INET6 11851 struct ip6_hdr *ip6 = NULL; 11852 int32_t isipv6; 11853 #endif 11854 uint8_t app_limited = BBR_JR_SENT_DATA; 11855 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 11856 /* We take a cache hit here */ 11857 memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); 11858 cts = tcp_tv_to_usectick(&bbr->rc_tv); 11859 inp = bbr->rc_inp; 11860 hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS); 11861 tp->t_flags2 &= ~TF2_HPTS_CALLS; 11862 so = inp->inp_socket; 11863 sb = &so->so_snd; 11864 if (tp->t_nic_ktls_xmit) 11865 hw_tls = 1; 11866 else 11867 hw_tls = 0; 11868 kern_prefetch(sb, &maxseg); 11869 maxseg = tp->t_maxseg - bbr->rc_last_options; 11870 if (bbr_minseg(bbr) < maxseg) { 11871 tcp_bbr_tso_size_check(bbr, cts); 11872 } 11873 /* Remove any flags that indicate we are pacing on the inp */ 11874 pace_max_segs = bbr->r_ctl.rc_pace_max_segs; 11875 p_maxseg = min(maxseg, pace_max_segs); 11876 INP_WLOCK_ASSERT(inp); 11877 #ifdef TCP_OFFLOAD 11878 if (tp->t_flags & TF_TOE) 11879 return (tcp_offload_output(tp)); 11880 #endif 11881 11882 #ifdef INET6 11883 if (bbr->r_state) { 11884 /* Use the cache line loaded if possible */ 11885 isipv6 = bbr->r_is_v6; 11886 } else { 11887 isipv6 = (inp->inp_vflag & INP_IPV6) != 0; 11888 } 11889 #endif 11890 if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && 11891 tcp_in_hpts(tp)) { 11892 /* 11893 * We are on the hpts for some timer but not hptsi output. 11894 * Possibly remove from the hpts so we can send/recv etc. 11895 */ 11896 if ((tp->t_flags & TF_ACKNOW) == 0) { 11897 /* 11898 * No immediate demand right now to send an ack, but 11899 * the user may have read, making room for new data 11900 * (a window update). If so we may want to cancel 11901 * whatever timer is running (KEEP/DEL-ACK?) and 11902 * continue to send out a window update. Or we may 11903 * have gotten more data into the socket buffer to 11904 * send. 11905 */ 11906 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 11907 (long)TCP_MAXWIN << tp->rcv_scale); 11908 if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && 11909 ((tcp_outflags[tp->t_state] & TH_RST) == 0) && 11910 ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= 11911 (tp->snd_max - tp->snd_una))) { 11912 /* 11913 * Nothing new to send and no window update 11914 * is needed to send. Lets just return and 11915 * let the timer-run off. 11916 */ 11917 return (0); 11918 } 11919 } 11920 tcp_hpts_remove(tp); 11921 bbr_timer_cancel(bbr, __LINE__, cts); 11922 } 11923 if (bbr->r_ctl.rc_last_delay_val) { 11924 /* Calculate a rough delay for early escape to sending */ 11925 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11926 delay_calc = cts - bbr->rc_pacer_started; 11927 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11928 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11929 else 11930 delay_calc = 0; 11931 } 11932 /* Mark that we have called bbr_output(). */ 11933 if ((bbr->r_timer_override) || 11934 (tp->t_state < TCPS_ESTABLISHED)) { 11935 /* Timeouts or early states are exempt */ 11936 if (tcp_in_hpts(tp)) 11937 tcp_hpts_remove(tp); 11938 } else if (tcp_in_hpts(tp)) { 11939 if ((bbr->r_ctl.rc_last_delay_val) && 11940 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && 11941 delay_calc) { 11942 /* 11943 * We were being paced for output and the delay has 11944 * already exceeded when we were supposed to be 11945 * called, lets go ahead and pull out of the hpts 11946 * and call output. 11947 */ 11948 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); 11949 bbr->r_ctl.rc_last_delay_val = 0; 11950 tcp_hpts_remove(tp); 11951 } else if (tp->t_state == TCPS_CLOSED) { 11952 bbr->r_ctl.rc_last_delay_val = 0; 11953 tcp_hpts_remove(tp); 11954 } else { 11955 /* 11956 * On the hpts, you shall not pass! even if ACKNOW 11957 * is on, we will when the hpts fires, unless of 11958 * course we are overdue. 11959 */ 11960 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); 11961 return (0); 11962 } 11963 } 11964 bbr->rc_cwnd_limited = 0; 11965 if (bbr->r_ctl.rc_last_delay_val) { 11966 /* recalculate the real delay and deal with over/under */ 11967 if (SEQ_GT(cts, bbr->rc_pacer_started)) 11968 delay_calc = cts - bbr->rc_pacer_started; 11969 else 11970 delay_calc = 0; 11971 if (delay_calc >= bbr->r_ctl.rc_last_delay_val) 11972 /* Setup the delay which will be added in */ 11973 delay_calc -= bbr->r_ctl.rc_last_delay_val; 11974 else { 11975 /* 11976 * We are early setup to adjust 11977 * our slot time. 11978 */ 11979 uint64_t merged_val; 11980 11981 bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); 11982 bbr->r_agg_early_set = 1; 11983 if (bbr->r_ctl.rc_hptsi_agg_delay) { 11984 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { 11985 /* Nope our previous late cancels out the early */ 11986 bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; 11987 bbr->r_agg_early_set = 0; 11988 bbr->r_ctl.rc_agg_early = 0; 11989 } else { 11990 bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; 11991 bbr->r_ctl.rc_hptsi_agg_delay = 0; 11992 } 11993 } 11994 merged_val = bbr->rc_pacer_started; 11995 merged_val <<= 32; 11996 merged_val |= bbr->r_ctl.rc_last_delay_val; 11997 bbr_log_pacing_delay_calc(bbr, hpts_calling, 11998 bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, 11999 bbr->r_agg_early_set, 3); 12000 bbr->r_ctl.rc_last_delay_val = 0; 12001 BBR_STAT_INC(bbr_early); 12002 delay_calc = 0; 12003 } 12004 } else { 12005 /* We were not delayed due to hptsi */ 12006 if (bbr->r_agg_early_set) 12007 bbr->r_ctl.rc_agg_early = 0; 12008 bbr->r_agg_early_set = 0; 12009 delay_calc = 0; 12010 } 12011 if (delay_calc) { 12012 /* 12013 * We had a hptsi delay which means we are falling behind on 12014 * sending at the expected rate. Calculate an extra amount 12015 * of data we can send, if any, to put us back on track. 12016 */ 12017 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) 12018 bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; 12019 else 12020 bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; 12021 } 12022 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12023 if ((tp->snd_una == tp->snd_max) && 12024 (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && 12025 (sbavail(sb))) { 12026 /* 12027 * Ok we have been idle with nothing outstanding 12028 * we possibly need to start fresh with either a new 12029 * suite of states or a fast-ramp up. 12030 */ 12031 bbr_restart_after_idle(bbr, 12032 cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); 12033 } 12034 /* 12035 * Now was there a hptsi delay where we are behind? We only count 12036 * being behind if: a) We are not in recovery. b) There was a delay. 12037 * <and> c) We had room to send something. 12038 * 12039 */ 12040 if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 12041 int retval; 12042 12043 retval = bbr_process_timers(tp, bbr, cts, hpts_calling); 12044 if (retval != 0) { 12045 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); 12046 /* 12047 * If timers want tcp_drop(), then pass error out, 12048 * otherwise suppress it. 12049 */ 12050 return (retval < 0 ? retval : 0); 12051 } 12052 } 12053 bbr->rc_tp->t_flags2 &= ~TF2_MBUF_QUEUE_READY; 12054 if (hpts_calling && 12055 (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { 12056 bbr->r_ctl.rc_last_delay_val = 0; 12057 } 12058 bbr->r_timer_override = 0; 12059 bbr->r_wanted_output = 0; 12060 /* 12061 * For TFO connections in SYN_RECEIVED, only allow the initial 12062 * SYN|ACK and those sent by the retransmit timer. 12063 */ 12064 if ((tp->t_flags & TF_FASTOPEN) && 12065 ((tp->t_state == TCPS_SYN_RECEIVED) || 12066 (tp->t_state == TCPS_SYN_SENT)) && 12067 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ 12068 (tp->t_rxtshift == 0)) { /* not a retransmit */ 12069 len = 0; 12070 goto just_return_nolock; 12071 } 12072 /* 12073 * Before sending anything check for a state update. For hpts 12074 * calling without input this is important. If its input calling 12075 * then this was already done. 12076 */ 12077 if (bbr->rc_use_google == 0) 12078 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12079 again: 12080 /* 12081 * If we've recently taken a timeout, snd_max will be greater than 12082 * snd_max. BBR in general does not pay much attention to snd_nxt 12083 * for historic reasons the persist timer still uses it. This means 12084 * we have to look at it. All retransmissions that are not persits 12085 * use the rsm that needs to be sent so snd_nxt is ignored. At the 12086 * end of this routine we pull snd_nxt always up to snd_max. 12087 */ 12088 doing_tlp = 0; 12089 #ifdef BBR_INVARIANTS 12090 doing_retran_from = picked_up_retran = 0; 12091 #endif 12092 error = 0; 12093 tso = 0; 12094 slot = 0; 12095 mtu = 0; 12096 sendwin = min(tp->snd_wnd, tp->snd_cwnd); 12097 sb_offset = tp->snd_max - tp->snd_una; 12098 flags = tcp_outflags[tp->t_state]; 12099 sack_rxmit = 0; 12100 len = 0; 12101 rsm = NULL; 12102 if (flags & TH_RST) { 12103 SOCKBUF_LOCK(sb); 12104 goto send; 12105 } 12106 recheck_resend: 12107 while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { 12108 /* We need to always have one in reserve */ 12109 rsm = bbr_alloc(bbr); 12110 if (rsm == NULL) { 12111 error = ENOMEM; 12112 /* Lie to get on the hpts */ 12113 tot_len = tp->t_maxseg; 12114 if (hpts_calling) 12115 /* Retry in a ms */ 12116 slot = 1001; 12117 goto just_return_nolock; 12118 } 12119 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); 12120 bbr->r_ctl.rc_free_cnt++; 12121 rsm = NULL; 12122 } 12123 /* What do we send, a resend? */ 12124 if (bbr->r_ctl.rc_resend == NULL) { 12125 /* Check for rack timeout */ 12126 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); 12127 if (bbr->r_ctl.rc_resend) { 12128 #ifdef BBR_INVARIANTS 12129 picked_up_retran = 1; 12130 #endif 12131 bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); 12132 } 12133 } 12134 if (bbr->r_ctl.rc_resend) { 12135 rsm = bbr->r_ctl.rc_resend; 12136 #ifdef BBR_INVARIANTS 12137 doing_retran_from = 1; 12138 #endif 12139 /* Remove any TLP flags its a RACK or T-O */ 12140 rsm->r_flags &= ~BBR_TLP; 12141 bbr->r_ctl.rc_resend = NULL; 12142 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 12143 #ifdef BBR_INVARIANTS 12144 panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", 12145 tp, bbr, rsm, rsm->r_start, tp->snd_una); 12146 goto recheck_resend; 12147 #else 12148 /* TSNH */ 12149 rsm = NULL; 12150 goto recheck_resend; 12151 #endif 12152 } 12153 if (rsm->r_flags & BBR_HAS_SYN) { 12154 /* Only retransmit a SYN by itself */ 12155 len = 0; 12156 if ((flags & TH_SYN) == 0) { 12157 /* Huh something is wrong */ 12158 rsm->r_start++; 12159 if (rsm->r_start == rsm->r_end) { 12160 /* Clean it up, somehow we missed the ack? */ 12161 bbr_log_syn(tp, NULL); 12162 } else { 12163 /* TFO with data? */ 12164 rsm->r_flags &= ~BBR_HAS_SYN; 12165 len = rsm->r_end - rsm->r_start; 12166 } 12167 } else { 12168 /* Retransmitting SYN */ 12169 rsm = NULL; 12170 SOCKBUF_LOCK(sb); 12171 goto send; 12172 } 12173 } else 12174 len = rsm->r_end - rsm->r_start; 12175 if ((bbr->rc_resends_use_tso == 0) && 12176 (len > maxseg)) { 12177 len = maxseg; 12178 more_to_rxt = 1; 12179 } 12180 sb_offset = rsm->r_start - tp->snd_una; 12181 if (len > 0) { 12182 sack_rxmit = 1; 12183 KMOD_TCPSTAT_INC(tcps_sack_rexmits); 12184 KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, 12185 min(len, maxseg)); 12186 } else { 12187 /* I dont think this can happen */ 12188 rsm = NULL; 12189 goto recheck_resend; 12190 } 12191 BBR_STAT_INC(bbr_resends_set); 12192 } else if (bbr->r_ctl.rc_tlp_send) { 12193 /* 12194 * Tail loss probe 12195 */ 12196 doing_tlp = 1; 12197 rsm = bbr->r_ctl.rc_tlp_send; 12198 bbr->r_ctl.rc_tlp_send = NULL; 12199 sack_rxmit = 1; 12200 len = rsm->r_end - rsm->r_start; 12201 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12202 len = maxseg; 12203 12204 if (SEQ_GT(tp->snd_una, rsm->r_start)) { 12205 #ifdef BBR_INVARIANTS 12206 panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", 12207 tp, bbr, tp->snd_una, rsm, rsm->r_start); 12208 #else 12209 /* TSNH */ 12210 rsm = NULL; 12211 goto recheck_resend; 12212 #endif 12213 } 12214 sb_offset = rsm->r_start - tp->snd_una; 12215 BBR_STAT_INC(bbr_tlp_set); 12216 } 12217 /* 12218 * Enforce a connection sendmap count limit if set 12219 * as long as we are not retransmiting. 12220 */ 12221 if ((rsm == NULL) && 12222 (V_tcp_map_entries_limit > 0) && 12223 (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { 12224 BBR_STAT_INC(bbr_alloc_limited); 12225 if (!bbr->alloc_limit_reported) { 12226 bbr->alloc_limit_reported = 1; 12227 BBR_STAT_INC(bbr_alloc_limited_conns); 12228 } 12229 goto just_return_nolock; 12230 } 12231 #ifdef BBR_INVARIANTS 12232 if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { 12233 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", 12234 tp, bbr, rsm, sb_offset, len); 12235 } 12236 #endif 12237 /* 12238 * Get standard flags, and add SYN or FIN if requested by 'hidden' 12239 * state flags. 12240 */ 12241 if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) 12242 flags |= TH_FIN; 12243 if (tp->t_flags & TF_NEEDSYN) 12244 flags |= TH_SYN; 12245 12246 if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { 12247 /* we are retransmitting the fin */ 12248 len--; 12249 if (len) { 12250 /* 12251 * When retransmitting data do *not* include the 12252 * FIN. This could happen from a TLP probe if we 12253 * allowed data with a FIN. 12254 */ 12255 flags &= ~TH_FIN; 12256 } 12257 } else if (rsm) { 12258 if (flags & TH_FIN) 12259 flags &= ~TH_FIN; 12260 } 12261 if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { 12262 void *end_rsm; 12263 12264 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); 12265 if (end_rsm) 12266 kern_prefetch(end_rsm, &prefetch_rsm); 12267 prefetch_rsm = 1; 12268 } 12269 SOCKBUF_LOCK(sb); 12270 /* 12271 * If snd_nxt == snd_max and we have transmitted a FIN, the 12272 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a 12273 * negative length. This can also occur when TCP opens up its 12274 * congestion window while receiving additional duplicate acks after 12275 * fast-retransmit because TCP will reset snd_nxt to snd_max after 12276 * the fast-retransmit. 12277 * 12278 * In the normal retransmit-FIN-only case, however, snd_nxt will be 12279 * set to snd_una, the sb_offset will be 0, and the length may wind 12280 * up 0. 12281 * 12282 * If sack_rxmit is true we are retransmitting from the scoreboard 12283 * in which case len is already set. 12284 */ 12285 if (sack_rxmit == 0) { 12286 uint32_t avail; 12287 12288 avail = sbavail(sb); 12289 if (SEQ_GT(tp->snd_max, tp->snd_una)) 12290 sb_offset = tp->snd_max - tp->snd_una; 12291 else 12292 sb_offset = 0; 12293 if (bbr->rc_tlp_new_data) { 12294 /* TLP is forcing out new data */ 12295 uint32_t tlplen; 12296 12297 doing_tlp = 1; 12298 tlplen = maxseg; 12299 12300 if (tlplen > (uint32_t)(avail - sb_offset)) { 12301 tlplen = (uint32_t)(avail - sb_offset); 12302 } 12303 if (tlplen > tp->snd_wnd) { 12304 len = tp->snd_wnd; 12305 } else { 12306 len = tlplen; 12307 } 12308 bbr->rc_tlp_new_data = 0; 12309 } else { 12310 len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); 12311 if ((len < p_maxseg) && 12312 (bbr->rc_in_persist == 0) && 12313 (ctf_outstanding(tp) >= (2 * p_maxseg)) && 12314 ((avail - sb_offset) >= p_maxseg)) { 12315 /* 12316 * We are not completing whats in the socket 12317 * buffer (i.e. there is at least a segment 12318 * waiting to send) and we have 2 or more 12319 * segments outstanding. There is no sense 12320 * of sending a little piece. Lets defer and 12321 * and wait until we can send a whole 12322 * segment. 12323 */ 12324 len = 0; 12325 } 12326 if (bbr->rc_in_persist) { 12327 /* 12328 * We are in persists, figure out if 12329 * a retransmit is available (maybe the previous 12330 * persists we sent) or if we have to send new 12331 * data. 12332 */ 12333 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 12334 if (rsm) { 12335 len = rsm->r_end - rsm->r_start; 12336 if (rsm->r_flags & BBR_HAS_FIN) 12337 len--; 12338 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) 12339 len = maxseg; 12340 if (len > 1) 12341 BBR_STAT_INC(bbr_persist_reneg); 12342 /* 12343 * XXXrrs we could force the len to 12344 * 1 byte here to cause the chunk to 12345 * split apart.. but that would then 12346 * mean we always retransmit it as 12347 * one byte even after the window 12348 * opens. 12349 */ 12350 sack_rxmit = 1; 12351 sb_offset = rsm->r_start - tp->snd_una; 12352 } else { 12353 /* 12354 * First time through in persists or peer 12355 * acked our one byte. Though we do have 12356 * to have something in the sb. 12357 */ 12358 len = 1; 12359 sb_offset = 0; 12360 if (avail == 0) 12361 len = 0; 12362 } 12363 } 12364 } 12365 } 12366 if (prefetch_so_done == 0) { 12367 kern_prefetch(so, &prefetch_so_done); 12368 prefetch_so_done = 1; 12369 } 12370 /* 12371 * Lop off SYN bit if it has already been sent. However, if this is 12372 * SYN-SENT state and if segment contains data and if we don't know 12373 * that foreign host supports TAO, suppress sending segment. 12374 */ 12375 if ((flags & TH_SYN) && (rsm == NULL) && 12376 SEQ_GT(tp->snd_max, tp->snd_una)) { 12377 if (tp->t_state != TCPS_SYN_RECEIVED) 12378 flags &= ~TH_SYN; 12379 /* 12380 * When sending additional segments following a TFO SYN|ACK, 12381 * do not include the SYN bit. 12382 */ 12383 if ((tp->t_flags & TF_FASTOPEN) && 12384 (tp->t_state == TCPS_SYN_RECEIVED)) 12385 flags &= ~TH_SYN; 12386 sb_offset--, len++; 12387 if (sbavail(sb) == 0) 12388 len = 0; 12389 } else if ((flags & TH_SYN) && rsm) { 12390 /* 12391 * Subtract one from the len for the SYN being 12392 * retransmitted. 12393 */ 12394 len--; 12395 } 12396 /* 12397 * Be careful not to send data and/or FIN on SYN segments. This 12398 * measure is needed to prevent interoperability problems with not 12399 * fully conformant TCP implementations. 12400 */ 12401 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { 12402 len = 0; 12403 flags &= ~TH_FIN; 12404 } 12405 /* 12406 * On TFO sockets, ensure no data is sent in the following cases: 12407 * 12408 * - When retransmitting SYN|ACK on a passively-created socket 12409 * - When retransmitting SYN on an actively created socket 12410 * - When sending a zero-length cookie (cookie request) on an 12411 * actively created socket 12412 * - When the socket is in the CLOSED state (RST is being sent) 12413 */ 12414 if ((tp->t_flags & TF_FASTOPEN) && 12415 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || 12416 ((tp->t_state == TCPS_SYN_SENT) && 12417 (tp->t_tfo_client_cookie_len == 0)) || 12418 (flags & TH_RST))) { 12419 len = 0; 12420 sack_rxmit = 0; 12421 rsm = NULL; 12422 } 12423 /* Without fast-open there should never be data sent on a SYN */ 12424 if ((flags & TH_SYN) && !(tp->t_flags & TF_FASTOPEN)) 12425 len = 0; 12426 if (len <= 0) { 12427 /* 12428 * If FIN has been sent but not acked, but we haven't been 12429 * called to retransmit, len will be < 0. Otherwise, window 12430 * shrank after we sent into it. If window shrank to 0, 12431 * cancel pending retransmit, pull snd_nxt back to (closed) 12432 * window, and set the persist timer if it isn't already 12433 * going. If the window didn't close completely, just wait 12434 * for an ACK. 12435 * 12436 * We also do a general check here to ensure that we will 12437 * set the persist timer when we have data to send, but a 12438 * 0-byte window. This makes sure the persist timer is set 12439 * even if the packet hits one of the "goto send" lines 12440 * below. 12441 */ 12442 len = 0; 12443 if ((tp->snd_wnd == 0) && 12444 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12445 (tp->snd_una == tp->snd_max) && 12446 (sb_offset < (int)sbavail(sb))) { 12447 /* 12448 * Not enough room in the rwnd to send 12449 * a paced segment out. 12450 */ 12451 bbr_enter_persist(tp, bbr, cts, __LINE__); 12452 } 12453 } else if ((rsm == NULL) && 12454 (doing_tlp == 0) && 12455 (len < bbr->r_ctl.rc_pace_max_segs)) { 12456 /* 12457 * We are not sending a full segment for 12458 * some reason. Should we not send anything (think 12459 * sws or persists)? 12460 */ 12461 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12462 (TCPS_HAVEESTABLISHED(tp->t_state)) && 12463 (len < (int)(sbavail(sb) - sb_offset))) { 12464 /* 12465 * Here the rwnd is less than 12466 * the pacing size, this is not a retransmit, 12467 * we are established and 12468 * the send is not the last in the socket buffer 12469 * lets not send, and possibly enter persists. 12470 */ 12471 len = 0; 12472 if (tp->snd_max == tp->snd_una) 12473 bbr_enter_persist(tp, bbr, cts, __LINE__); 12474 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && 12475 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12476 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12477 (len < (int)(sbavail(sb) - sb_offset)) && 12478 (len < bbr_minseg(bbr))) { 12479 /* 12480 * Here we are not retransmitting, and 12481 * the cwnd is not so small that we could 12482 * not send at least a min size (rxt timer 12483 * not having gone off), We have 2 segments or 12484 * more already in flight, its not the tail end 12485 * of the socket buffer and the cwnd is blocking 12486 * us from sending out minimum pacing segment size. 12487 * Lets not send anything. 12488 */ 12489 bbr->rc_cwnd_limited = 1; 12490 len = 0; 12491 } else if (((tp->snd_wnd - ctf_outstanding(tp)) < 12492 min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && 12493 (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12494 bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && 12495 (len < (int)(sbavail(sb) - sb_offset)) && 12496 (TCPS_HAVEESTABLISHED(tp->t_state))) { 12497 /* 12498 * Here we have a send window but we have 12499 * filled it up and we can't send another pacing segment. 12500 * We also have in flight more than 2 segments 12501 * and we are not completing the sb i.e. we allow 12502 * the last bytes of the sb to go out even if 12503 * its not a full pacing segment. 12504 */ 12505 len = 0; 12506 } 12507 } 12508 /* len will be >= 0 after this point. */ 12509 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 12510 tcp_sndbuf_autoscale(tp, so, sendwin); 12511 /* 12512 * 12513 */ 12514 if (bbr->rc_in_persist && 12515 len && 12516 (rsm == NULL) && 12517 (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { 12518 /* 12519 * We are in persist, not doing a retransmit and don't have enough space 12520 * yet to send a full TSO. So is it at the end of the sb 12521 * if so we need to send else nuke to 0 and don't send. 12522 */ 12523 int sbleft; 12524 if (sbavail(sb) > sb_offset) 12525 sbleft = sbavail(sb) - sb_offset; 12526 else 12527 sbleft = 0; 12528 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { 12529 /* not at end of sb lets not send */ 12530 len = 0; 12531 } 12532 } 12533 /* 12534 * Decide if we can use TCP Segmentation Offloading (if supported by 12535 * hardware). 12536 * 12537 * TSO may only be used if we are in a pure bulk sending state. The 12538 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP 12539 * options prevent using TSO. With TSO the TCP header is the same 12540 * (except for the sequence number) for all generated packets. This 12541 * makes it impossible to transmit any options which vary per 12542 * generated segment or packet. 12543 * 12544 * IPv4 handling has a clear separation of ip options and ip header 12545 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() 12546 * does the right thing below to provide length of just ip options 12547 * and thus checking for ipoptlen is enough to decide if ip options 12548 * are present. 12549 */ 12550 #ifdef INET6 12551 if (isipv6) 12552 ipoptlen = ip6_optlen(inp); 12553 else 12554 #endif 12555 if (inp->inp_options) 12556 ipoptlen = inp->inp_options->m_len - 12557 offsetof(struct ipoption, ipopt_list); 12558 else 12559 ipoptlen = 0; 12560 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12561 /* 12562 * Pre-calculate here as we save another lookup into the darknesses 12563 * of IPsec that way and can actually decide if TSO is ok. 12564 */ 12565 #ifdef INET6 12566 if (isipv6 && IPSEC_ENABLED(ipv6)) 12567 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); 12568 #ifdef INET 12569 else 12570 #endif 12571 #endif /* INET6 */ 12572 #ifdef INET 12573 if (IPSEC_ENABLED(ipv4)) 12574 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); 12575 #endif /* INET */ 12576 #endif /* IPSEC */ 12577 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12578 ipoptlen += ipsec_optlen; 12579 #endif 12580 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && 12581 (len > maxseg) && 12582 (tp->t_port == 0) && 12583 ((tp->t_flags & TF_SIGNATURE) == 0) && 12584 tp->rcv_numsacks == 0 && 12585 ipoptlen == 0) 12586 tso = 1; 12587 12588 recwin = lmin(lmax(sbspace(&so->so_rcv), 0), 12589 (long)TCP_MAXWIN << tp->rcv_scale); 12590 /* 12591 * Sender silly window avoidance. We transmit under the following 12592 * conditions when len is non-zero: 12593 * 12594 * - We have a full segment (or more with TSO) - This is the last 12595 * buffer in a write()/send() and we are either idle or running 12596 * NODELAY - we've timed out (e.g. persist timer) - we have more 12597 * then 1/2 the maximum send window's worth of data (receiver may be 12598 * limited the window size) - we need to retransmit 12599 */ 12600 if (rsm) 12601 goto send; 12602 if (len) { 12603 if (sack_rxmit) 12604 goto send; 12605 if (len >= p_maxseg) 12606 goto send; 12607 /* 12608 * NOTE! on localhost connections an 'ack' from the remote 12609 * end may occur synchronously with the output and cause us 12610 * to flush a buffer queued with moretocome. XXX 12611 * 12612 */ 12613 if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ 12614 ((tp->t_flags & TF_NODELAY) || 12615 ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && 12616 (tp->t_flags & TF_NOPUSH) == 0) { 12617 goto send; 12618 } 12619 if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 12620 goto send; 12621 } 12622 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { 12623 goto send; 12624 } 12625 } 12626 /* 12627 * Sending of standalone window updates. 12628 * 12629 * Window updates are important when we close our window due to a 12630 * full socket buffer and are opening it again after the application 12631 * reads data from it. Once the window has opened again and the 12632 * remote end starts to send again the ACK clock takes over and 12633 * provides the most current window information. 12634 * 12635 * We must avoid the silly window syndrome whereas every read from 12636 * the receive buffer, no matter how small, causes a window update 12637 * to be sent. We also should avoid sending a flurry of window 12638 * updates when the socket buffer had queued a lot of data and the 12639 * application is doing small reads. 12640 * 12641 * Prevent a flurry of pointless window updates by only sending an 12642 * update when we can increase the advertized window by more than 12643 * 1/4th of the socket buffer capacity. When the buffer is getting 12644 * full or is very small be more aggressive and send an update 12645 * whenever we can increase by two mss sized segments. In all other 12646 * situations the ACK's to new incoming data will carry further 12647 * window increases. 12648 * 12649 * Don't send an independent window update if a delayed ACK is 12650 * pending (it will get piggy-backed on it) or the remote side 12651 * already has done a half-close and won't send more data. Skip 12652 * this if the connection is in T/TCP half-open state. 12653 */ 12654 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && 12655 !(tp->t_flags & TF_DELACK) && 12656 !TCPS_HAVERCVDFIN(tp->t_state)) { 12657 /* Check to see if we should do a window update */ 12658 if (bbr_window_update_needed(tp, so, recwin, maxseg)) 12659 goto send; 12660 } 12661 /* 12662 * Send if we owe the peer an ACK, RST, SYN. ACKNOW 12663 * is also a catch-all for the retransmit timer timeout case. 12664 */ 12665 if (tp->t_flags & TF_ACKNOW) { 12666 goto send; 12667 } 12668 if (flags & TH_RST) { 12669 /* Always send a RST if one is due */ 12670 goto send; 12671 } 12672 if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { 12673 goto send; 12674 } 12675 /* 12676 * If our state indicates that FIN should be sent and we have not 12677 * yet done so, then we need to send. 12678 */ 12679 if (flags & TH_FIN && 12680 ((tp->t_flags & TF_SENTFIN) == 0)) { 12681 goto send; 12682 } 12683 /* 12684 * No reason to send a segment, just return. 12685 */ 12686 just_return: 12687 SOCKBUF_UNLOCK(sb); 12688 just_return_nolock: 12689 if (tot_len) 12690 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 12691 if (bbr->rc_no_pacing) 12692 slot = 0; 12693 if (tot_len == 0) { 12694 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= 12695 tp->snd_wnd) { 12696 BBR_STAT_INC(bbr_rwnd_limited); 12697 app_limited = BBR_JR_RWND_LIMITED; 12698 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12699 if ((bbr->rc_in_persist == 0) && 12700 TCPS_HAVEESTABLISHED(tp->t_state) && 12701 (tp->snd_max == tp->snd_una) && 12702 sbavail(&so->so_snd)) { 12703 /* No send window.. we must enter persist */ 12704 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); 12705 } 12706 } else if (ctf_outstanding(tp) >= sbavail(sb)) { 12707 BBR_STAT_INC(bbr_app_limited); 12708 app_limited = BBR_JR_APP_LIMITED; 12709 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12710 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12711 bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { 12712 BBR_STAT_INC(bbr_cwnd_limited); 12713 app_limited = BBR_JR_CWND_LIMITED; 12714 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12715 bbr->r_ctl.rc_lost_bytes))); 12716 bbr->rc_cwnd_limited = 1; 12717 } else { 12718 BBR_STAT_INC(bbr_app_limited); 12719 app_limited = BBR_JR_APP_LIMITED; 12720 bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); 12721 } 12722 bbr->r_ctl.rc_hptsi_agg_delay = 0; 12723 bbr->r_agg_early_set = 0; 12724 bbr->r_ctl.rc_agg_early = 0; 12725 bbr->r_ctl.rc_last_delay_val = 0; 12726 } else if (bbr->rc_use_google == 0) 12727 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 12728 /* Are we app limited? */ 12729 if ((app_limited == BBR_JR_APP_LIMITED) || 12730 (app_limited == BBR_JR_RWND_LIMITED)) { 12731 /** 12732 * We are application limited. 12733 */ 12734 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 12735 bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); 12736 } 12737 if (tot_len == 0) 12738 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); 12739 /* Dont update the time if we did not send */ 12740 bbr->r_ctl.rc_last_delay_val = 0; 12741 bbr->rc_output_starts_timer = 1; 12742 bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); 12743 bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); 12744 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 12745 /* Make sure snd_nxt is drug up */ 12746 tp->snd_nxt = tp->snd_max; 12747 } 12748 return (error); 12749 12750 send: 12751 if (doing_tlp == 0) { 12752 /* 12753 * Data not a TLP, and its not the rxt firing. If it is the 12754 * rxt firing, we want to leave the tlp_in_progress flag on 12755 * so we don't send another TLP. It has to be a rack timer 12756 * or normal send (response to acked data) to clear the tlp 12757 * in progress flag. 12758 */ 12759 bbr->rc_tlp_in_progress = 0; 12760 bbr->rc_tlp_rtx_out = 0; 12761 } else { 12762 /* 12763 * Its a TLP. 12764 */ 12765 bbr->rc_tlp_in_progress = 1; 12766 } 12767 bbr_timer_cancel(bbr, __LINE__, cts); 12768 if (rsm == NULL) { 12769 if (sbused(sb) > 0) { 12770 /* 12771 * This is sub-optimal. We only send a stand alone 12772 * FIN on its own segment. 12773 */ 12774 if (flags & TH_FIN) { 12775 flags &= ~TH_FIN; 12776 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { 12777 /* Lets not send this */ 12778 slot = 0; 12779 goto just_return; 12780 } 12781 } 12782 } 12783 } else { 12784 /* 12785 * We do *not* send a FIN on a retransmit if it has data. 12786 * The if clause here where len > 1 should never come true. 12787 */ 12788 if ((len > 0) && 12789 (((rsm->r_flags & BBR_HAS_FIN) == 0) && 12790 (flags & TH_FIN))) { 12791 flags &= ~TH_FIN; 12792 len--; 12793 } 12794 } 12795 SOCKBUF_LOCK_ASSERT(sb); 12796 if (len > 0) { 12797 if ((tp->snd_una == tp->snd_max) && 12798 (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { 12799 /* 12800 * This qualifies as a RTT_PROBE session since we 12801 * drop the data outstanding to nothing and waited 12802 * more than bbr_rtt_probe_time. 12803 */ 12804 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); 12805 bbr_set_reduced_rtt(bbr, cts, __LINE__); 12806 } 12807 if (len >= maxseg) 12808 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; 12809 else 12810 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; 12811 } 12812 /* 12813 * Before ESTABLISHED, force sending of initial options unless TCP 12814 * set not to do any options. NOTE: we assume that the IP/TCP header 12815 * plus TCP options always fit in a single mbuf, leaving room for a 12816 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) 12817 * + optlen <= MCLBYTES 12818 */ 12819 optlen = 0; 12820 #ifdef INET6 12821 if (isipv6) 12822 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 12823 else 12824 #endif 12825 hdrlen = sizeof(struct tcpiphdr); 12826 12827 /* 12828 * Compute options for segment. We only have to care about SYN and 12829 * established connection segments. Options for SYN-ACK segments 12830 * are handled in TCP syncache. 12831 */ 12832 to.to_flags = 0; 12833 local_options = 0; 12834 if ((tp->t_flags & TF_NOOPT) == 0) { 12835 /* Maximum segment size. */ 12836 if (flags & TH_SYN) { 12837 to.to_mss = tcp_mssopt(&inp->inp_inc); 12838 if (tp->t_port) 12839 to.to_mss -= V_tcp_udp_tunneling_overhead; 12840 to.to_flags |= TOF_MSS; 12841 /* 12842 * On SYN or SYN|ACK transmits on TFO connections, 12843 * only include the TFO option if it is not a 12844 * retransmit, as the presence of the TFO option may 12845 * have caused the original SYN or SYN|ACK to have 12846 * been dropped by a middlebox. 12847 */ 12848 if ((tp->t_flags & TF_FASTOPEN) && 12849 (tp->t_rxtshift == 0)) { 12850 if (tp->t_state == TCPS_SYN_RECEIVED) { 12851 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; 12852 to.to_tfo_cookie = 12853 (u_int8_t *)&tp->t_tfo_cookie.server; 12854 to.to_flags |= TOF_FASTOPEN; 12855 wanted_cookie = 1; 12856 } else if (tp->t_state == TCPS_SYN_SENT) { 12857 to.to_tfo_len = 12858 tp->t_tfo_client_cookie_len; 12859 to.to_tfo_cookie = 12860 tp->t_tfo_cookie.client; 12861 to.to_flags |= TOF_FASTOPEN; 12862 wanted_cookie = 1; 12863 } 12864 } 12865 } 12866 /* Window scaling. */ 12867 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { 12868 to.to_wscale = tp->request_r_scale; 12869 to.to_flags |= TOF_SCALE; 12870 } 12871 /* Timestamps. */ 12872 if ((tp->t_flags & TF_RCVD_TSTMP) || 12873 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { 12874 to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; 12875 to.to_tsecr = tp->ts_recent; 12876 to.to_flags |= TOF_TS; 12877 local_options += TCPOLEN_TIMESTAMP + 2; 12878 } 12879 /* Set receive buffer autosizing timestamp. */ 12880 if (tp->rfbuf_ts == 0 && 12881 (so->so_rcv.sb_flags & SB_AUTOSIZE)) 12882 tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); 12883 /* Selective ACK's. */ 12884 if (flags & TH_SYN) 12885 to.to_flags |= TOF_SACKPERM; 12886 else if (TCPS_HAVEESTABLISHED(tp->t_state) && 12887 tp->rcv_numsacks > 0) { 12888 to.to_flags |= TOF_SACK; 12889 to.to_nsacks = tp->rcv_numsacks; 12890 to.to_sacks = (u_char *)tp->sackblks; 12891 } 12892 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 12893 /* TCP-MD5 (RFC2385). */ 12894 if (tp->t_flags & TF_SIGNATURE) 12895 to.to_flags |= TOF_SIGNATURE; 12896 #endif /* TCP_SIGNATURE */ 12897 12898 /* Processing the options. */ 12899 hdrlen += (optlen = tcp_addoptions(&to, opt)); 12900 /* 12901 * If we wanted a TFO option to be added, but it was unable 12902 * to fit, ensure no data is sent. 12903 */ 12904 if ((tp->t_flags & TF_FASTOPEN) && wanted_cookie && 12905 !(to.to_flags & TOF_FASTOPEN)) 12906 len = 0; 12907 } 12908 if (tp->t_port) { 12909 if (V_tcp_udp_tunneling_port == 0) { 12910 /* The port was removed?? */ 12911 SOCKBUF_UNLOCK(&so->so_snd); 12912 return (EHOSTUNREACH); 12913 } 12914 hdrlen += sizeof(struct udphdr); 12915 } 12916 #ifdef INET6 12917 if (isipv6) 12918 ipoptlen = ip6_optlen(inp); 12919 else 12920 #endif 12921 if (inp->inp_options) 12922 ipoptlen = inp->inp_options->m_len - 12923 offsetof(struct ipoption, ipopt_list); 12924 else 12925 ipoptlen = 0; 12926 ipoptlen = 0; 12927 #if defined(IPSEC) || defined(IPSEC_SUPPORT) 12928 ipoptlen += ipsec_optlen; 12929 #endif 12930 if (bbr->rc_last_options != local_options) { 12931 /* 12932 * Cache the options length this generally does not change 12933 * on a connection. We use this to calculate TSO. 12934 */ 12935 bbr->rc_last_options = local_options; 12936 } 12937 maxseg = tp->t_maxseg - (ipoptlen + optlen); 12938 p_maxseg = min(maxseg, pace_max_segs); 12939 /* 12940 * Adjust data length if insertion of options will bump the packet 12941 * length beyond the t_maxseg length. Clear the FIN bit because we 12942 * cut off the tail of the segment. 12943 */ 12944 if (len > maxseg) { 12945 if (len != 0 && (flags & TH_FIN)) { 12946 flags &= ~TH_FIN; 12947 } 12948 if (tso) { 12949 uint32_t moff; 12950 int32_t max_len; 12951 12952 /* extract TSO information */ 12953 if_hw_tsomax = tp->t_tsomax; 12954 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; 12955 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; 12956 KASSERT(ipoptlen == 0, 12957 ("%s: TSO can't do IP options", __func__)); 12958 12959 /* 12960 * Check if we should limit by maximum payload 12961 * length: 12962 */ 12963 if (if_hw_tsomax != 0) { 12964 /* compute maximum TSO length */ 12965 max_len = (if_hw_tsomax - hdrlen - 12966 max_linkhdr); 12967 if (max_len <= 0) { 12968 len = 0; 12969 } else if (len > max_len) { 12970 len = max_len; 12971 } 12972 } 12973 /* 12974 * Prevent the last segment from being fractional 12975 * unless the send sockbuf can be emptied: 12976 */ 12977 if ((sb_offset + len) < sbavail(sb)) { 12978 moff = len % (uint32_t)maxseg; 12979 if (moff != 0) { 12980 len -= moff; 12981 } 12982 } 12983 /* 12984 * In case there are too many small fragments don't 12985 * use TSO: 12986 */ 12987 if (len <= maxseg) { 12988 len = maxseg; 12989 tso = 0; 12990 } 12991 } else { 12992 /* Not doing TSO */ 12993 if (optlen + ipoptlen >= tp->t_maxseg) { 12994 /* 12995 * Since we don't have enough space to put 12996 * the IP header chain and the TCP header in 12997 * one packet as required by RFC 7112, don't 12998 * send it. Also ensure that at least one 12999 * byte of the payload can be put into the 13000 * TCP segment. 13001 */ 13002 SOCKBUF_UNLOCK(&so->so_snd); 13003 error = EMSGSIZE; 13004 sack_rxmit = 0; 13005 goto out; 13006 } 13007 len = maxseg; 13008 } 13009 } else { 13010 /* Not doing TSO */ 13011 if_hw_tsomaxsegcount = 0; 13012 tso = 0; 13013 } 13014 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, 13015 ("%s: len > IP_MAXPACKET", __func__)); 13016 #ifdef DIAGNOSTIC 13017 #ifdef INET6 13018 if (max_linkhdr + hdrlen > MCLBYTES) 13019 #else 13020 if (max_linkhdr + hdrlen > MHLEN) 13021 #endif 13022 panic("tcphdr too big"); 13023 #endif 13024 /* 13025 * This KASSERT is here to catch edge cases at a well defined place. 13026 * Before, those had triggered (random) panic conditions further 13027 * down. 13028 */ 13029 #ifdef BBR_INVARIANTS 13030 if (sack_rxmit) { 13031 if (SEQ_LT(rsm->r_start, tp->snd_una)) { 13032 panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", 13033 rsm, tp, bbr, rsm->r_start, tp->snd_una); 13034 } 13035 } 13036 #endif 13037 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); 13038 if ((len == 0) && 13039 (flags & TH_FIN) && 13040 (sbused(sb))) { 13041 /* 13042 * We have outstanding data, don't send a fin by itself!. 13043 */ 13044 slot = 0; 13045 goto just_return; 13046 } 13047 /* 13048 * Grab a header mbuf, attaching a copy of data to be transmitted, 13049 * and initialize the header from the template for sends on this 13050 * connection. 13051 */ 13052 if (len) { 13053 uint32_t moff; 13054 13055 /* 13056 * We place a limit on sending with hptsi. 13057 */ 13058 if ((rsm == NULL) && len > pace_max_segs) 13059 len = pace_max_segs; 13060 if (len <= maxseg) 13061 tso = 0; 13062 #ifdef INET6 13063 if (MHLEN < hdrlen + max_linkhdr) 13064 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 13065 else 13066 #endif 13067 m = m_gethdr(M_NOWAIT, MT_DATA); 13068 13069 if (m == NULL) { 13070 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13071 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13072 SOCKBUF_UNLOCK(sb); 13073 error = ENOBUFS; 13074 sack_rxmit = 0; 13075 goto out; 13076 } 13077 m->m_data += max_linkhdr; 13078 m->m_len = hdrlen; 13079 /* 13080 * Start the m_copy functions from the closest mbuf to the 13081 * sb_offset in the socket buffer chain. 13082 */ 13083 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { 13084 #ifdef BBR_INVARIANTS 13085 if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) 13086 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", 13087 tp, bbr, len, sb_offset, sbavail(sb), rsm, 13088 doing_retran_from, 13089 picked_up_retran, 13090 doing_tlp); 13091 13092 #endif 13093 /* 13094 * In this messed up situation we have two choices, 13095 * a) pretend the send worked, and just start timers 13096 * and what not (not good since that may lead us 13097 * back here a lot). <or> b) Send the lowest segment 13098 * in the map. <or> c) Drop the connection. Lets do 13099 * <b> which if it continues to happen will lead to 13100 * <c> via timeouts. 13101 */ 13102 BBR_STAT_INC(bbr_offset_recovery); 13103 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); 13104 sb_offset = 0; 13105 if (rsm == NULL) { 13106 sack_rxmit = 0; 13107 len = sbavail(sb); 13108 } else { 13109 sack_rxmit = 1; 13110 if (rsm->r_start != tp->snd_una) { 13111 /* 13112 * Things are really messed up, <c> 13113 * is the only thing to do. 13114 */ 13115 BBR_STAT_INC(bbr_offset_drop); 13116 SOCKBUF_UNLOCK(sb); 13117 (void)m_free(m); 13118 return (-EFAULT); /* tcp_drop() */ 13119 } 13120 len = rsm->r_end - rsm->r_start; 13121 } 13122 if (len > sbavail(sb)) 13123 len = sbavail(sb); 13124 if (len > maxseg) 13125 len = maxseg; 13126 } 13127 mb = sbsndptr_noadv(sb, sb_offset, &moff); 13128 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { 13129 m_copydata(mb, moff, (int)len, 13130 mtod(m, caddr_t)+hdrlen); 13131 if (rsm == NULL) 13132 sbsndptr_adv(sb, mb, len); 13133 m->m_len += len; 13134 } else { 13135 struct sockbuf *msb; 13136 13137 if (rsm) 13138 msb = NULL; 13139 else 13140 msb = sb; 13141 #ifdef BBR_INVARIANTS 13142 if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { 13143 if (rsm) { 13144 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 ", 13145 tp, bbr, len, moff, 13146 sbavail(sb), rsm, 13147 tp->snd_una, rsm->r_flags, rsm->r_start, 13148 doing_retran_from, 13149 picked_up_retran, 13150 doing_tlp, sack_rxmit); 13151 } else { 13152 panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", 13153 tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); 13154 } 13155 } 13156 #endif 13157 m->m_next = tcp_m_copym( 13158 mb, moff, &len, 13159 if_hw_tsomaxsegcount, 13160 if_hw_tsomaxsegsize, msb, 13161 ((rsm == NULL) ? hw_tls : 0) 13162 #ifdef NETFLIX_COPY_ARGS 13163 , NULL, NULL 13164 #endif 13165 ); 13166 if (len <= maxseg) { 13167 /* 13168 * Must have ran out of mbufs for the copy 13169 * shorten it to no longer need tso. Lets 13170 * not put on sendalot since we are low on 13171 * mbufs. 13172 */ 13173 tso = 0; 13174 } 13175 if (m->m_next == NULL) { 13176 SOCKBUF_UNLOCK(sb); 13177 (void)m_free(m); 13178 error = ENOBUFS; 13179 sack_rxmit = 0; 13180 goto out; 13181 } 13182 } 13183 #ifdef BBR_INVARIANTS 13184 if (tso && len < maxseg) { 13185 panic("tp:%p tso on, but len:%d < maxseg:%d", 13186 tp, len, maxseg); 13187 } 13188 if (tso && if_hw_tsomaxsegcount) { 13189 int32_t seg_cnt = 0; 13190 struct mbuf *foo; 13191 13192 foo = m; 13193 while (foo) { 13194 seg_cnt++; 13195 foo = foo->m_next; 13196 } 13197 if (seg_cnt > if_hw_tsomaxsegcount) { 13198 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); 13199 } 13200 } 13201 #endif 13202 /* 13203 * If we're sending everything we've got, set PUSH. (This 13204 * will keep happy those implementations which only give 13205 * data to the user when a buffer fills or a PUSH comes in.) 13206 */ 13207 if (sb_offset + len == sbused(sb) && 13208 sbused(sb) && 13209 !(flags & TH_SYN)) { 13210 flags |= TH_PUSH; 13211 } 13212 SOCKBUF_UNLOCK(sb); 13213 } else { 13214 SOCKBUF_UNLOCK(sb); 13215 if (tp->t_flags & TF_ACKNOW) 13216 KMOD_TCPSTAT_INC(tcps_sndacks); 13217 else if (flags & (TH_SYN | TH_FIN | TH_RST)) 13218 KMOD_TCPSTAT_INC(tcps_sndctrl); 13219 else 13220 KMOD_TCPSTAT_INC(tcps_sndwinup); 13221 13222 m = m_gethdr(M_NOWAIT, MT_DATA); 13223 if (m == NULL) { 13224 BBR_STAT_INC(bbr_failed_mbuf_aloc); 13225 bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); 13226 error = ENOBUFS; 13227 /* Fudge the send time since we could not send */ 13228 sack_rxmit = 0; 13229 goto out; 13230 } 13231 #ifdef INET6 13232 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && 13233 MHLEN >= hdrlen) { 13234 M_ALIGN(m, hdrlen); 13235 } else 13236 #endif 13237 m->m_data += max_linkhdr; 13238 m->m_len = hdrlen; 13239 } 13240 SOCKBUF_UNLOCK_ASSERT(sb); 13241 m->m_pkthdr.rcvif = (struct ifnet *)0; 13242 #ifdef MAC 13243 mac_inpcb_create_mbuf(inp, m); 13244 #endif 13245 #ifdef INET6 13246 if (isipv6) { 13247 ip6 = mtod(m, struct ip6_hdr *); 13248 if (tp->t_port) { 13249 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); 13250 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13251 udp->uh_dport = tp->t_port; 13252 ulen = hdrlen + len - sizeof(struct ip6_hdr); 13253 udp->uh_ulen = htons(ulen); 13254 th = (struct tcphdr *)(udp + 1); 13255 } else { 13256 th = (struct tcphdr *)(ip6 + 1); 13257 } 13258 tcpip_fillheaders(inp, tp->t_port, ip6, th); 13259 } else 13260 #endif /* INET6 */ 13261 { 13262 ip = mtod(m, struct ip *); 13263 if (tp->t_port) { 13264 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); 13265 udp->uh_sport = htons(V_tcp_udp_tunneling_port); 13266 udp->uh_dport = tp->t_port; 13267 ulen = hdrlen + len - sizeof(struct ip); 13268 udp->uh_ulen = htons(ulen); 13269 th = (struct tcphdr *)(udp + 1); 13270 } else { 13271 th = (struct tcphdr *)(ip + 1); 13272 } 13273 tcpip_fillheaders(inp, tp->t_port, ip, th); 13274 } 13275 /* 13276 * If we are doing retransmissions, then snd_nxt will not reflect 13277 * the first unsent octet. For ACK only packets, we do not want the 13278 * sequence number of the retransmitted packet, we want the sequence 13279 * number of the next unsent octet. So, if there is no data (and no 13280 * SYN or FIN), use snd_max instead of snd_nxt when filling in 13281 * ti_seq. But if we are in persist state, snd_max might reflect 13282 * one byte beyond the right edge of the window, so use snd_nxt in 13283 * that case, since we know we aren't doing a retransmission. 13284 * (retransmit and persist are mutually exclusive...) 13285 */ 13286 if (sack_rxmit == 0) { 13287 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { 13288 /* New data (including new persists) */ 13289 th->th_seq = htonl(tp->snd_max); 13290 bbr_seq = tp->snd_max; 13291 } else if (flags & TH_SYN) { 13292 /* Syn's always send from iss */ 13293 th->th_seq = htonl(tp->iss); 13294 bbr_seq = tp->iss; 13295 } else if (flags & TH_FIN) { 13296 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { 13297 /* 13298 * If we sent the fin already its 1 minus 13299 * snd_max 13300 */ 13301 th->th_seq = (htonl(tp->snd_max - 1)); 13302 bbr_seq = (tp->snd_max - 1); 13303 } else { 13304 /* First time FIN use snd_max */ 13305 th->th_seq = htonl(tp->snd_max); 13306 bbr_seq = tp->snd_max; 13307 } 13308 } else { 13309 /* 13310 * len == 0 and not persist we use snd_max, sending 13311 * an ack unless we have sent the fin then its 1 13312 * minus. 13313 */ 13314 /* 13315 * XXXRRS Question if we are in persists and we have 13316 * nothing outstanding to send and we have not sent 13317 * a FIN, we will send an ACK. In such a case it 13318 * might be better to send (tp->snd_una - 1) which 13319 * would force the peer to ack. 13320 */ 13321 if (tp->t_flags & TF_SENTFIN) { 13322 th->th_seq = htonl(tp->snd_max - 1); 13323 bbr_seq = (tp->snd_max - 1); 13324 } else { 13325 th->th_seq = htonl(tp->snd_max); 13326 bbr_seq = tp->snd_max; 13327 } 13328 } 13329 } else { 13330 /* All retransmits use the rsm to guide the send */ 13331 th->th_seq = htonl(rsm->r_start); 13332 bbr_seq = rsm->r_start; 13333 } 13334 th->th_ack = htonl(tp->rcv_nxt); 13335 if (optlen) { 13336 bcopy(opt, th + 1, optlen); 13337 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; 13338 } 13339 tcp_set_flags(th, flags); 13340 /* 13341 * Calculate receive window. Don't shrink window, but avoid silly 13342 * window syndrome. 13343 */ 13344 if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && 13345 recwin < maxseg))) 13346 recwin = 0; 13347 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && 13348 recwin < (tp->rcv_adv - tp->rcv_nxt)) 13349 recwin = (tp->rcv_adv - tp->rcv_nxt); 13350 if (recwin > TCP_MAXWIN << tp->rcv_scale) 13351 recwin = TCP_MAXWIN << tp->rcv_scale; 13352 13353 /* 13354 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or 13355 * <SYN,ACK>) segment itself is never scaled. The <SYN,ACK> case is 13356 * handled in syncache. 13357 */ 13358 if (flags & TH_SYN) 13359 th->th_win = htons((u_short) 13360 (min(sbspace(&so->so_rcv), TCP_MAXWIN))); 13361 else { 13362 /* Avoid shrinking window with window scaling. */ 13363 recwin = roundup2(recwin, 1 << tp->rcv_scale); 13364 th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); 13365 } 13366 /* 13367 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 13368 * window. This may cause the remote transmitter to stall. This 13369 * flag tells soreceive() to disable delayed acknowledgements when 13370 * draining the buffer. This can occur if the receiver is 13371 * attempting to read more data than can be buffered prior to 13372 * transmitting on the connection. 13373 */ 13374 if (th->th_win == 0) { 13375 tp->t_sndzerowin++; 13376 tp->t_flags |= TF_RXWIN0SENT; 13377 } else 13378 tp->t_flags &= ~TF_RXWIN0SENT; 13379 /* 13380 * We don't support urgent data, but drag along 13381 * the pointer in case of a stack switch. 13382 */ 13383 tp->snd_up = tp->snd_una; 13384 /* 13385 * Put TCP length in extended header, and then checksum extended 13386 * header and data. 13387 */ 13388 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ 13389 13390 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) 13391 if (to.to_flags & TOF_SIGNATURE) { 13392 /* 13393 * Calculate MD5 signature and put it into the place 13394 * determined before. NOTE: since TCP options buffer doesn't 13395 * point into mbuf's data, calculate offset and use it. 13396 */ 13397 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, 13398 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { 13399 /* 13400 * Do not send segment if the calculation of MD5 13401 * digest has failed. 13402 */ 13403 goto out; 13404 } 13405 } 13406 #endif 13407 13408 #ifdef INET6 13409 if (isipv6) { 13410 /* 13411 * ip6_plen is not need to be filled now, and will be filled 13412 * in ip6_output. 13413 */ 13414 if (tp->t_port) { 13415 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; 13416 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13417 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); 13418 th->th_sum = htons(0); 13419 UDPSTAT_INC(udps_opackets); 13420 } else { 13421 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 13422 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13423 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + 13424 optlen + len, IPPROTO_TCP, 0); 13425 } 13426 } 13427 #endif 13428 #if defined(INET6) && defined(INET) 13429 else 13430 #endif 13431 #ifdef INET 13432 { 13433 if (tp->t_port) { 13434 m->m_pkthdr.csum_flags = CSUM_UDP; 13435 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); 13436 udp->uh_sum = in_pseudo(ip->ip_src.s_addr, 13437 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); 13438 th->th_sum = htons(0); 13439 UDPSTAT_INC(udps_opackets); 13440 } else { 13441 csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; 13442 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); 13443 th->th_sum = in_pseudo(ip->ip_src.s_addr, 13444 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + 13445 IPPROTO_TCP + len + optlen)); 13446 } 13447 /* IP version must be set here for ipv4/ipv6 checking later */ 13448 KASSERT(ip->ip_v == IPVERSION, 13449 ("%s: IP version incorrect: %d", __func__, ip->ip_v)); 13450 } 13451 #endif 13452 13453 /* 13454 * Enable TSO and specify the size of the segments. The TCP pseudo 13455 * header checksum is always provided. XXX: Fixme: This is currently 13456 * not the case for IPv6. 13457 */ 13458 if (tso) { 13459 KASSERT(len > maxseg, 13460 ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); 13461 m->m_pkthdr.csum_flags |= CSUM_TSO; 13462 csum_flags |= CSUM_TSO; 13463 m->m_pkthdr.tso_segsz = maxseg; 13464 } 13465 KASSERT(len + hdrlen == m_length(m, NULL), 13466 ("%s: mbuf chain different than expected: %d + %u != %u", 13467 __func__, len, hdrlen, m_length(m, NULL))); 13468 13469 #ifdef TCP_HHOOK 13470 /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ 13471 hhook_run_tcp_est_out(tp, th, &to, len, tso); 13472 #endif 13473 13474 /* Log to the black box */ 13475 if (tcp_bblogging_on(tp)) { 13476 union tcp_log_stackspecific log; 13477 13478 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); 13479 /* Record info on type of transmission */ 13480 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; 13481 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); 13482 log.u_bbr.flex3 = maxseg; 13483 log.u_bbr.flex4 = delay_calc; 13484 log.u_bbr.flex5 = bbr->rc_past_init_win; 13485 log.u_bbr.flex5 <<= 1; 13486 log.u_bbr.flex5 |= bbr->rc_no_pacing; 13487 log.u_bbr.flex5 <<= 29; 13488 log.u_bbr.flex5 |= tp->t_maxseg; 13489 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; 13490 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); 13491 /* lets poke in the low and the high here for debugging */ 13492 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; 13493 if (rsm || sack_rxmit) { 13494 if (doing_tlp) 13495 log.u_bbr.flex8 = 2; 13496 else 13497 log.u_bbr.flex8 = 1; 13498 } else { 13499 log.u_bbr.flex8 = 0; 13500 } 13501 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, 13502 len, &log, false, NULL, NULL, 0, tv); 13503 } else { 13504 lgb = NULL; 13505 } 13506 /* 13507 * Fill in IP length and desired time to live and send to IP level. 13508 * There should be a better way to handle ttl and tos; we could keep 13509 * them in the template, but need a way to checksum without them. 13510 */ 13511 /* 13512 * m->m_pkthdr.len should have been set before cksum calcuration, 13513 * because in6_cksum() need it. 13514 */ 13515 #ifdef INET6 13516 if (isipv6) { 13517 /* 13518 * we separately set hoplimit for every segment, since the 13519 * user might want to change the value via setsockopt. Also, 13520 * desired default hop limit might be changed via Neighbor 13521 * Discovery. 13522 */ 13523 ip6->ip6_hlim = in6_selecthlim(inp, NULL); 13524 13525 /* 13526 * Set the packet size here for the benefit of DTrace 13527 * probes. ip6_output() will set it properly; it's supposed 13528 * to include the option header lengths as well. 13529 */ 13530 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); 13531 13532 if (V_path_mtu_discovery && maxseg > V_tcp_minmss) 13533 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13534 else 13535 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13536 13537 if (tp->t_state == TCPS_SYN_SENT) 13538 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); 13539 13540 TCP_PROBE5(send, NULL, tp, ip6, tp, th); 13541 /* TODO: IPv6 IP6TOS_ECT bit on */ 13542 error = ip6_output(m, inp->in6p_outputopts, 13543 &inp->inp_route6, 13544 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 13545 NULL, NULL, inp); 13546 13547 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) 13548 mtu = inp->inp_route6.ro_nh->nh_mtu; 13549 } 13550 #endif /* INET6 */ 13551 #if defined(INET) && defined(INET6) 13552 else 13553 #endif 13554 #ifdef INET 13555 { 13556 ip->ip_len = htons(m->m_pkthdr.len); 13557 #ifdef INET6 13558 if (isipv6) 13559 ip->ip_ttl = in6_selecthlim(inp, NULL); 13560 #endif /* INET6 */ 13561 /* 13562 * If we do path MTU discovery, then we set DF on every 13563 * packet. This might not be the best thing to do according 13564 * to RFC3390 Section 2. However the tcp hostcache migitates 13565 * the problem so it affects only the first tcp connection 13566 * with a host. 13567 * 13568 * NB: Don't set DF on small MTU/MSS to have a safe 13569 * fallback. 13570 */ 13571 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { 13572 tp->t_flags2 |= TF2_PLPMTU_PMTUD; 13573 if (tp->t_port == 0 || len < V_tcp_minmss) { 13574 ip->ip_off |= htons(IP_DF); 13575 } 13576 } else { 13577 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; 13578 } 13579 13580 if (tp->t_state == TCPS_SYN_SENT) 13581 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); 13582 13583 TCP_PROBE5(send, NULL, tp, ip, tp, th); 13584 13585 error = ip_output(m, inp->inp_options, &inp->inp_route, 13586 ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, 13587 inp); 13588 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) 13589 mtu = inp->inp_route.ro_nh->nh_mtu; 13590 } 13591 #endif /* INET */ 13592 if (lgb) { 13593 lgb->tlb_errno = error; 13594 lgb = NULL; 13595 } 13596 13597 out: 13598 /* 13599 * In transmit state, time the transmission and arrange for the 13600 * retransmit. In persist state, just set snd_max. 13601 */ 13602 if (error == 0) { 13603 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); 13604 if (TCPS_HAVEESTABLISHED(tp->t_state) && 13605 (tp->t_flags & TF_SACK_PERMIT) && 13606 tp->rcv_numsacks > 0) 13607 tcp_clean_dsack_blocks(tp); 13608 /* We sent an ack clear the bbr_segs_rcvd count */ 13609 bbr->output_error_seen = 0; 13610 bbr->oerror_cnt = 0; 13611 bbr->bbr_segs_rcvd = 0; 13612 if (len == 0) 13613 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); 13614 /* Do accounting for new sends */ 13615 if ((len > 0) && (rsm == NULL)) { 13616 int idx; 13617 if (tp->snd_una == tp->snd_max) { 13618 /* 13619 * Special case to match google, when 13620 * nothing is in flight the delivered 13621 * time does get updated to the current 13622 * time (see tcp_rate_bsd.c). 13623 */ 13624 bbr->r_ctl.rc_del_time = cts; 13625 } 13626 if (len >= maxseg) { 13627 idx = (len / maxseg) + 3; 13628 if (idx >= TCP_MSS_ACCT_ATIMER) 13629 counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); 13630 else 13631 counter_u64_add(bbr_out_size[idx], 1); 13632 } else { 13633 /* smaller than a MSS */ 13634 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); 13635 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) 13636 idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); 13637 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); 13638 } 13639 } 13640 } 13641 abandon = 0; 13642 /* 13643 * We must do the send accounting before we log the output, 13644 * otherwise the state of the rsm could change and we account to the 13645 * wrong bucket. 13646 */ 13647 if (len > 0) { 13648 bbr_do_send_accounting(tp, bbr, rsm, len, error); 13649 if (error == 0) { 13650 if (tp->snd_una == tp->snd_max) 13651 bbr->r_ctl.rc_tlp_rxt_last_time = cts; 13652 } 13653 } 13654 bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, 13655 cts, mb, &abandon, rsm, 0, sb); 13656 if (abandon) { 13657 /* 13658 * If bbr_log_output destroys the TCB or sees a TH_RST being 13659 * sent we should hit this condition. 13660 */ 13661 return (0); 13662 } 13663 if (bbr->rc_in_persist == 0) { 13664 /* 13665 * Advance snd_nxt over sequence space of this segment. 13666 */ 13667 if (error) 13668 /* We don't log or do anything with errors */ 13669 goto skip_upd; 13670 13671 if (tp->snd_una == tp->snd_max && 13672 (len || (flags & (TH_SYN | TH_FIN)))) { 13673 /* 13674 * Update the time we just added data since none was 13675 * outstanding. 13676 */ 13677 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13678 bbr->rc_tp->t_acktime = ticks; 13679 } 13680 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { 13681 if (flags & TH_SYN) { 13682 /* 13683 * Smack the snd_max to iss + 1 13684 * if its a FO we will add len below. 13685 */ 13686 tp->snd_max = tp->iss + 1; 13687 } 13688 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13689 tp->snd_max++; 13690 tp->t_flags |= TF_SENTFIN; 13691 } 13692 } 13693 if (sack_rxmit == 0) 13694 tp->snd_max += len; 13695 skip_upd: 13696 if ((error == 0) && len) 13697 tot_len += len; 13698 } else { 13699 /* Persists case */ 13700 int32_t xlen = len; 13701 13702 if (error) 13703 goto nomore; 13704 13705 if (flags & TH_SYN) 13706 ++xlen; 13707 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { 13708 ++xlen; 13709 tp->t_flags |= TF_SENTFIN; 13710 } 13711 if (xlen && (tp->snd_una == tp->snd_max)) { 13712 /* 13713 * Update the time we just added data since none was 13714 * outstanding. 13715 */ 13716 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); 13717 bbr->rc_tp->t_acktime = ticks; 13718 } 13719 if (sack_rxmit == 0) 13720 tp->snd_max += xlen; 13721 tot_len += (len + optlen + ipoptlen); 13722 } 13723 nomore: 13724 if (error) { 13725 /* 13726 * Failures do not advance the seq counter above. For the 13727 * case of ENOBUFS we will fall out and become ack-clocked. 13728 * capping the cwnd at the current flight. 13729 * Everything else will just have to retransmit with the timer 13730 * (no pacer). 13731 */ 13732 SOCKBUF_UNLOCK_ASSERT(sb); 13733 BBR_STAT_INC(bbr_saw_oerr); 13734 /* Clear all delay/early tracks */ 13735 bbr->r_ctl.rc_hptsi_agg_delay = 0; 13736 bbr->r_ctl.rc_agg_early = 0; 13737 bbr->r_agg_early_set = 0; 13738 bbr->output_error_seen = 1; 13739 if (bbr->oerror_cnt < 0xf) 13740 bbr->oerror_cnt++; 13741 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { 13742 /* drop the session */ 13743 return (-ENETDOWN); 13744 } 13745 switch (error) { 13746 case ENOBUFS: 13747 /* 13748 * Make this guy have to get ack's to send 13749 * more but lets make sure we don't 13750 * slam him below a T-O (1MSS). 13751 */ 13752 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { 13753 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13754 bbr->r_ctl.rc_lost_bytes)) - maxseg; 13755 if (tp->snd_cwnd < maxseg) 13756 tp->snd_cwnd = maxseg; 13757 } 13758 slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; 13759 BBR_STAT_INC(bbr_saw_enobuf); 13760 if (bbr->bbr_hdrw_pacing) 13761 counter_u64_add(bbr_hdwr_pacing_enobuf, 1); 13762 else 13763 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); 13764 /* 13765 * Here even in the enobuf's case we want to do our 13766 * state update. The reason being we may have been 13767 * called by the input function. If so we have had 13768 * things change. 13769 */ 13770 error = 0; 13771 goto enobufs; 13772 case EMSGSIZE: 13773 /* 13774 * For some reason the interface we used initially 13775 * to send segments changed to another or lowered 13776 * its MTU. If TSO was active we either got an 13777 * interface without TSO capabilits or TSO was 13778 * turned off. If we obtained mtu from ip_output() 13779 * then update it and try again. 13780 */ 13781 /* Turn on tracing (or try to) */ 13782 { 13783 int old_maxseg; 13784 13785 old_maxseg = tp->t_maxseg; 13786 BBR_STAT_INC(bbr_saw_emsgsiz); 13787 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); 13788 if (mtu != 0) 13789 tcp_mss_update(tp, -1, mtu, NULL, NULL); 13790 if (old_maxseg <= tp->t_maxseg) { 13791 /* Huh it did not shrink? */ 13792 tp->t_maxseg = old_maxseg - 40; 13793 if (tp->t_maxseg < V_tcp_mssdflt) { 13794 /* 13795 * The MSS is so small we should not 13796 * process incoming SACK's since we are 13797 * subject to attack in such a case. 13798 */ 13799 tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT; 13800 } else { 13801 tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT; 13802 } 13803 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); 13804 } 13805 /* 13806 * Nuke all other things that can interfere 13807 * with slot 13808 */ 13809 if ((tot_len + len) && (len >= tp->t_maxseg)) { 13810 slot = bbr_get_pacing_delay(bbr, 13811 bbr->r_ctl.rc_bbr_hptsi_gain, 13812 (tot_len + len), cts, 0); 13813 if (slot < bbr_error_base_paceout) 13814 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13815 } else 13816 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; 13817 bbr->rc_output_starts_timer = 1; 13818 bbr_start_hpts_timer(bbr, tp, cts, 10, slot, 13819 tot_len); 13820 return (error); 13821 } 13822 case EPERM: 13823 case EACCES: 13824 tp->t_softerror = error; 13825 /* FALLTHROUGH */ 13826 case EHOSTDOWN: 13827 case EHOSTUNREACH: 13828 case ENETDOWN: 13829 case ENETUNREACH: 13830 if (TCPS_HAVERCVDSYN(tp->t_state)) { 13831 tp->t_softerror = error; 13832 } 13833 /* FALLTHROUGH */ 13834 default: 13835 slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; 13836 bbr->rc_output_starts_timer = 1; 13837 bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); 13838 return (error); 13839 } 13840 #ifdef STATS 13841 } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && 13842 len && 13843 (rsm == NULL) && 13844 (bbr->rc_in_persist == 0)) { 13845 tp->gput_seq = bbr_seq; 13846 tp->gput_ack = bbr_seq + 13847 min(sbavail(&so->so_snd) - sb_offset, sendwin); 13848 tp->gput_ts = cts; 13849 tp->t_flags |= TF_GPUTINPROG; 13850 #endif 13851 } 13852 KMOD_TCPSTAT_INC(tcps_sndtotal); 13853 if ((bbr->bbr_hdw_pace_ena) && 13854 (bbr->bbr_attempt_hdwr_pace == 0) && 13855 (bbr->rc_past_init_win) && 13856 (bbr->rc_bbr_state != BBR_STATE_STARTUP) && 13857 (get_filter_value(&bbr->r_ctl.rc_delrate)) && 13858 (inp->inp_route.ro_nh && 13859 inp->inp_route.ro_nh->nh_ifp)) { 13860 /* 13861 * We are past the initial window and 13862 * have at least one measurement so we 13863 * could use hardware pacing if its available. 13864 * We have an interface and we have not attempted 13865 * to setup hardware pacing, lets try to now. 13866 */ 13867 uint64_t rate_wanted; 13868 int err = 0; 13869 13870 rate_wanted = bbr_get_hardware_rate(bbr); 13871 bbr->bbr_attempt_hdwr_pace = 1; 13872 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, 13873 inp->inp_route.ro_nh->nh_ifp, 13874 rate_wanted, 13875 (RS_PACING_GEQ|RS_PACING_SUB_OK), 13876 &err, NULL); 13877 if (bbr->r_ctl.crte) { 13878 bbr_type_log_hdwr_pacing(bbr, 13879 bbr->r_ctl.crte->ptbl->rs_ifp, 13880 rate_wanted, 13881 bbr->r_ctl.crte->rate, 13882 __LINE__, cts, err); 13883 BBR_STAT_INC(bbr_hdwr_rl_add_ok); 13884 counter_u64_add(bbr_flows_nohdwr_pacing, -1); 13885 counter_u64_add(bbr_flows_whdwr_pacing, 1); 13886 bbr->bbr_hdrw_pacing = 1; 13887 /* Now what is our gain status? */ 13888 if (bbr->r_ctl.crte->rate < rate_wanted) { 13889 /* We have a problem */ 13890 bbr_setup_less_of_rate(bbr, cts, 13891 bbr->r_ctl.crte->rate, rate_wanted); 13892 } else { 13893 /* We are good */ 13894 bbr->gain_is_limited = 0; 13895 bbr->skip_gain = 0; 13896 } 13897 tcp_bbr_tso_size_check(bbr, cts); 13898 } else { 13899 bbr_type_log_hdwr_pacing(bbr, 13900 inp->inp_route.ro_nh->nh_ifp, 13901 rate_wanted, 13902 0, 13903 __LINE__, cts, err); 13904 BBR_STAT_INC(bbr_hdwr_rl_add_fail); 13905 } 13906 } 13907 if (bbr->bbr_hdrw_pacing) { 13908 /* 13909 * Worry about cases where the route 13910 * changes or something happened that we 13911 * lost our hardware pacing possibly during 13912 * the last ip_output call. 13913 */ 13914 if (inp->inp_snd_tag == NULL) { 13915 /* A change during ip output disabled hw pacing? */ 13916 bbr->bbr_hdrw_pacing = 0; 13917 } else if ((inp->inp_route.ro_nh == NULL) || 13918 (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { 13919 /* 13920 * We had an interface or route change, 13921 * detach from the current hdwr pacing 13922 * and setup to re-attempt next go 13923 * round. 13924 */ 13925 bbr->bbr_hdrw_pacing = 0; 13926 bbr->bbr_attempt_hdwr_pace = 0; 13927 tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); 13928 tcp_bbr_tso_size_check(bbr, cts); 13929 } 13930 } 13931 /* 13932 * Data sent (as far as we can tell). If this advertises a larger 13933 * window than any other segment, then remember the size of the 13934 * advertised window. Any pending ACK has now been sent. 13935 */ 13936 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) 13937 tp->rcv_adv = tp->rcv_nxt + recwin; 13938 13939 tp->last_ack_sent = tp->rcv_nxt; 13940 if ((error == 0) && 13941 (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && 13942 (doing_tlp == 0) && 13943 (tso == 0) && 13944 (len > 0) && 13945 ((flags & TH_RST) == 0) && 13946 ((flags & TH_SYN) == 0) && 13947 (IN_RECOVERY(tp->t_flags) == 0) && 13948 (bbr->rc_in_persist == 0) && 13949 (tot_len < bbr->r_ctl.rc_pace_max_segs)) { 13950 /* 13951 * For non-tso we need to goto again until we have sent out 13952 * enough data to match what we are hptsi out every hptsi 13953 * interval. 13954 */ 13955 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 13956 /* Make sure snd_nxt is drug up */ 13957 tp->snd_nxt = tp->snd_max; 13958 } 13959 if (rsm != NULL) { 13960 rsm = NULL; 13961 goto skip_again; 13962 } 13963 rsm = NULL; 13964 sack_rxmit = 0; 13965 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 13966 goto again; 13967 } 13968 skip_again: 13969 if ((error == 0) && (flags & TH_FIN)) 13970 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); 13971 if ((error == 0) && (flags & TH_RST)) 13972 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); 13973 if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { 13974 /* 13975 * Calculate/Re-Calculate the hptsi slot in usecs based on 13976 * what we have sent so far 13977 */ 13978 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); 13979 if (bbr->rc_no_pacing) 13980 slot = 0; 13981 } 13982 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); 13983 enobufs: 13984 if (bbr->rc_use_google == 0) 13985 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); 13986 bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + 13987 bbr->r_ctl.rc_lost_bytes))); 13988 bbr->rc_output_starts_timer = 1; 13989 if (bbr->bbr_use_rack_cheat && 13990 (more_to_rxt || 13991 ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { 13992 /* Rack cheats and shotguns out all rxt's 1ms apart */ 13993 if (slot > 1000) 13994 slot = 1000; 13995 } 13996 if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { 13997 /* 13998 * We don't change the tso size until some number of sends 13999 * to give the hardware commands time to get down 14000 * to the interface. 14001 */ 14002 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; 14003 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { 14004 bbr->hw_pacing_set = 1; 14005 tcp_bbr_tso_size_check(bbr, cts); 14006 } 14007 } 14008 bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); 14009 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { 14010 /* Make sure snd_nxt is drug up */ 14011 tp->snd_nxt = tp->snd_max; 14012 } 14013 return (error); 14014 14015 } 14016 14017 /* 14018 * See bbr_output_wtime() for return values. 14019 */ 14020 static int 14021 bbr_output(struct tcpcb *tp) 14022 { 14023 int32_t ret; 14024 struct timeval tv; 14025 14026 NET_EPOCH_ASSERT(); 14027 14028 INP_WLOCK_ASSERT(tptoinpcb(tp)); 14029 (void)tcp_get_usecs(&tv); 14030 ret = bbr_output_wtime(tp, &tv); 14031 return (ret); 14032 } 14033 14034 static void 14035 bbr_mtu_chg(struct tcpcb *tp) 14036 { 14037 struct tcp_bbr *bbr; 14038 struct bbr_sendmap *rsm, *frsm = NULL; 14039 uint32_t maxseg; 14040 14041 /* 14042 * The MTU has changed. a) Clear the sack filter. b) Mark everything 14043 * over the current size as SACK_PASS so a retransmit will occur. 14044 */ 14045 14046 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14047 maxseg = tp->t_maxseg - bbr->rc_last_options; 14048 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); 14049 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { 14050 /* Don't mess with ones acked (by sack?) */ 14051 if (rsm->r_flags & BBR_ACKED) 14052 continue; 14053 if ((rsm->r_end - rsm->r_start) > maxseg) { 14054 /* 14055 * We mark sack-passed on all the previous large 14056 * sends we did. This will force them to retransmit. 14057 */ 14058 rsm->r_flags |= BBR_SACK_PASSED; 14059 if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && 14060 bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { 14061 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; 14062 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; 14063 rsm->r_flags |= BBR_MARKED_LOST; 14064 } 14065 if (frsm == NULL) 14066 frsm = rsm; 14067 } 14068 } 14069 if (frsm) { 14070 bbr->r_ctl.rc_resend = frsm; 14071 } 14072 } 14073 14074 static int 14075 bbr_pru_options(struct tcpcb *tp, int flags) 14076 { 14077 if (flags & PRUS_OOB) 14078 return (EOPNOTSUPP); 14079 return (0); 14080 } 14081 14082 static void 14083 bbr_switch_failed(struct tcpcb *tp) 14084 { 14085 /* 14086 * If a switch fails we only need to 14087 * make sure mbuf_queuing is still in place. 14088 * We also need to make sure we are still in 14089 * ticks granularity (though we should probably 14090 * change bbr to go to USECs). 14091 * 14092 * For timers we need to see if we are still in the 14093 * pacer (if our flags are up) if so we are good, if 14094 * not we need to get back into the pacer. 14095 */ 14096 struct timeval tv; 14097 uint32_t cts; 14098 uint32_t toval; 14099 struct tcp_bbr *bbr; 14100 struct hpts_diag diag; 14101 14102 tp->t_flags2 |= TF2_CANNOT_DO_ECN; 14103 tp->t_flags2 |= TF2_SUPPORTS_MBUFQ; 14104 tcp_change_time_units(tp, TCP_TMR_GRANULARITY_TICKS); 14105 if (tp->t_in_hpts > IHPTS_NONE) { 14106 return; 14107 } 14108 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14109 cts = tcp_get_usecs(&tv); 14110 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { 14111 if (TSTMP_GT(bbr->rc_pacer_started, cts)) { 14112 toval = bbr->rc_pacer_started - cts; 14113 } else { 14114 /* one slot please */ 14115 toval = HPTS_TICKS_PER_SLOT; 14116 } 14117 } else if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { 14118 if (TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { 14119 toval = bbr->r_ctl.rc_timer_exp - cts; 14120 } else { 14121 /* one slot please */ 14122 toval = HPTS_TICKS_PER_SLOT; 14123 } 14124 } else 14125 toval = HPTS_TICKS_PER_SLOT; 14126 (void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval), 14127 __LINE__, &diag); 14128 bbr_log_hpts_diag(bbr, cts, &diag); 14129 } 14130 14131 struct tcp_function_block __tcp_bbr = { 14132 .tfb_tcp_block_name = __XSTRING(STACKNAME), 14133 .tfb_tcp_output = bbr_output, 14134 .tfb_do_queued_segments = ctf_do_queued_segments, 14135 .tfb_do_segment_nounlock = bbr_do_segment_nounlock, 14136 .tfb_tcp_do_segment = bbr_do_segment, 14137 .tfb_tcp_ctloutput = bbr_ctloutput, 14138 .tfb_tcp_fb_init = bbr_init, 14139 .tfb_tcp_fb_fini = bbr_fini, 14140 .tfb_tcp_timer_stop_all = bbr_stopall, 14141 .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, 14142 .tfb_tcp_handoff_ok = bbr_handoff_ok, 14143 .tfb_tcp_mtu_chg = bbr_mtu_chg, 14144 .tfb_pru_options = bbr_pru_options, 14145 .tfb_switch_failed = bbr_switch_failed, 14146 .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, 14147 }; 14148 14149 /* 14150 * bbr_ctloutput() must drop the inpcb lock before performing copyin on 14151 * socket option arguments. When it re-acquires the lock after the copy, it 14152 * has to revalidate that the connection is still valid for the socket 14153 * option. 14154 */ 14155 static int 14156 bbr_set_sockopt(struct tcpcb *tp, struct sockopt *sopt) 14157 { 14158 struct epoch_tracker et; 14159 struct inpcb *inp = tptoinpcb(tp); 14160 struct tcp_bbr *bbr; 14161 int32_t error = 0, optval; 14162 14163 switch (sopt->sopt_level) { 14164 case IPPROTO_IPV6: 14165 case IPPROTO_IP: 14166 return (tcp_default_ctloutput(tp, sopt)); 14167 } 14168 14169 switch (sopt->sopt_name) { 14170 case TCP_RACK_PACE_MAX_SEG: 14171 case TCP_RACK_MIN_TO: 14172 case TCP_RACK_REORD_THRESH: 14173 case TCP_RACK_REORD_FADE: 14174 case TCP_RACK_TLP_THRESH: 14175 case TCP_RACK_PKT_DELAY: 14176 case TCP_BBR_ALGORITHM: 14177 case TCP_BBR_TSLIMITS: 14178 case TCP_BBR_IWINTSO: 14179 case TCP_BBR_RECFORCE: 14180 case TCP_BBR_STARTUP_PG: 14181 case TCP_BBR_DRAIN_PG: 14182 case TCP_BBR_RWND_IS_APP: 14183 case TCP_BBR_PROBE_RTT_INT: 14184 case TCP_BBR_PROBE_RTT_GAIN: 14185 case TCP_BBR_PROBE_RTT_LEN: 14186 case TCP_BBR_STARTUP_LOSS_EXIT: 14187 case TCP_BBR_USEDEL_RATE: 14188 case TCP_BBR_MIN_RTO: 14189 case TCP_BBR_MAX_RTO: 14190 case TCP_BBR_PACE_PER_SEC: 14191 case TCP_DELACK: 14192 case TCP_BBR_PACE_DEL_TAR: 14193 case TCP_BBR_SEND_IWND_IN_TSO: 14194 case TCP_BBR_EXTRA_STATE: 14195 case TCP_BBR_UTTER_MAX_TSO: 14196 case TCP_BBR_MIN_TOPACEOUT: 14197 case TCP_BBR_FLOOR_MIN_TSO: 14198 case TCP_BBR_TSTMP_RAISES: 14199 case TCP_BBR_POLICER_DETECT: 14200 case TCP_BBR_USE_RACK_CHEAT: 14201 case TCP_DATA_AFTER_CLOSE: 14202 case TCP_BBR_HDWR_PACE: 14203 case TCP_BBR_PACE_SEG_MAX: 14204 case TCP_BBR_PACE_SEG_MIN: 14205 case TCP_BBR_PACE_CROSS: 14206 case TCP_BBR_PACE_OH: 14207 case TCP_BBR_TMR_PACE_OH: 14208 case TCP_BBR_RACK_RTT_USE: 14209 case TCP_BBR_RETRAN_WTSO: 14210 break; 14211 default: 14212 return (tcp_default_ctloutput(tp, sopt)); 14213 break; 14214 } 14215 INP_WUNLOCK(inp); 14216 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); 14217 if (error) 14218 return (error); 14219 INP_WLOCK(inp); 14220 if (inp->inp_flags & INP_DROPPED) { 14221 INP_WUNLOCK(inp); 14222 return (ECONNRESET); 14223 } 14224 if (tp->t_fb != &__tcp_bbr) { 14225 INP_WUNLOCK(inp); 14226 return (ENOPROTOOPT); 14227 } 14228 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14229 switch (sopt->sopt_name) { 14230 case TCP_BBR_PACE_PER_SEC: 14231 BBR_OPTS_INC(tcp_bbr_pace_per_sec); 14232 bbr->r_ctl.bbr_hptsi_per_second = optval; 14233 break; 14234 case TCP_BBR_PACE_DEL_TAR: 14235 BBR_OPTS_INC(tcp_bbr_pace_del_tar); 14236 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; 14237 break; 14238 case TCP_BBR_PACE_SEG_MAX: 14239 BBR_OPTS_INC(tcp_bbr_pace_seg_max); 14240 bbr->r_ctl.bbr_hptsi_segments_max = optval; 14241 break; 14242 case TCP_BBR_PACE_SEG_MIN: 14243 BBR_OPTS_INC(tcp_bbr_pace_seg_min); 14244 bbr->r_ctl.bbr_hptsi_bytes_min = optval; 14245 break; 14246 case TCP_BBR_PACE_CROSS: 14247 BBR_OPTS_INC(tcp_bbr_pace_cross); 14248 bbr->r_ctl.bbr_cross_over = optval; 14249 break; 14250 case TCP_BBR_ALGORITHM: 14251 BBR_OPTS_INC(tcp_bbr_algorithm); 14252 if (optval && (bbr->rc_use_google == 0)) { 14253 /* Turn on the google mode */ 14254 bbr_google_mode_on(bbr); 14255 if ((optval > 3) && (optval < 500)) { 14256 /* 14257 * Must be at least greater than .3% 14258 * and must be less than 50.0%. 14259 */ 14260 bbr->r_ctl.bbr_google_discount = optval; 14261 } 14262 } else if ((optval == 0) && (bbr->rc_use_google == 1)) { 14263 /* Turn off the google mode */ 14264 bbr_google_mode_off(bbr); 14265 } 14266 break; 14267 case TCP_BBR_TSLIMITS: 14268 BBR_OPTS_INC(tcp_bbr_tslimits); 14269 if (optval == 1) 14270 bbr->rc_use_ts_limit = 1; 14271 else if (optval == 0) 14272 bbr->rc_use_ts_limit = 0; 14273 else 14274 error = EINVAL; 14275 break; 14276 14277 case TCP_BBR_IWINTSO: 14278 BBR_OPTS_INC(tcp_bbr_iwintso); 14279 if ((optval >= 0) && (optval < 128)) { 14280 uint32_t twin; 14281 14282 bbr->rc_init_win = optval; 14283 twin = bbr_initial_cwnd(bbr, tp); 14284 if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) 14285 tp->snd_cwnd = twin; 14286 else 14287 error = EBUSY; 14288 } else 14289 error = EINVAL; 14290 break; 14291 case TCP_BBR_STARTUP_PG: 14292 BBR_OPTS_INC(tcp_bbr_startup_pg); 14293 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { 14294 bbr->r_ctl.rc_startup_pg = optval; 14295 if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { 14296 bbr->r_ctl.rc_bbr_hptsi_gain = optval; 14297 } 14298 } else 14299 error = EINVAL; 14300 break; 14301 case TCP_BBR_DRAIN_PG: 14302 BBR_OPTS_INC(tcp_bbr_drain_pg); 14303 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) 14304 bbr->r_ctl.rc_drain_pg = optval; 14305 else 14306 error = EINVAL; 14307 break; 14308 case TCP_BBR_PROBE_RTT_LEN: 14309 BBR_OPTS_INC(tcp_bbr_probertt_len); 14310 if (optval <= 1) 14311 reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); 14312 else 14313 error = EINVAL; 14314 break; 14315 case TCP_BBR_PROBE_RTT_GAIN: 14316 BBR_OPTS_INC(tcp_bbr_probertt_gain); 14317 if (optval <= BBR_UNIT) 14318 bbr->r_ctl.bbr_rttprobe_gain_val = optval; 14319 else 14320 error = EINVAL; 14321 break; 14322 case TCP_BBR_PROBE_RTT_INT: 14323 BBR_OPTS_INC(tcp_bbr_probe_rtt_int); 14324 if (optval > 1000) 14325 bbr->r_ctl.rc_probertt_int = optval; 14326 else 14327 error = EINVAL; 14328 break; 14329 case TCP_BBR_MIN_TOPACEOUT: 14330 BBR_OPTS_INC(tcp_bbr_topaceout); 14331 if (optval == 0) { 14332 bbr->no_pacing_until = 0; 14333 bbr->rc_no_pacing = 0; 14334 } else if (optval <= 0x00ff) { 14335 bbr->no_pacing_until = optval; 14336 if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && 14337 (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ 14338 /* Turn on no pacing */ 14339 bbr->rc_no_pacing = 1; 14340 } 14341 } else 14342 error = EINVAL; 14343 break; 14344 case TCP_BBR_STARTUP_LOSS_EXIT: 14345 BBR_OPTS_INC(tcp_bbr_startup_loss_exit); 14346 bbr->rc_loss_exit = optval; 14347 break; 14348 case TCP_BBR_USEDEL_RATE: 14349 error = EINVAL; 14350 break; 14351 case TCP_BBR_MIN_RTO: 14352 BBR_OPTS_INC(tcp_bbr_min_rto); 14353 bbr->r_ctl.rc_min_rto_ms = optval; 14354 break; 14355 case TCP_BBR_MAX_RTO: 14356 BBR_OPTS_INC(tcp_bbr_max_rto); 14357 bbr->rc_max_rto_sec = optval; 14358 break; 14359 case TCP_RACK_MIN_TO: 14360 /* Minimum time between rack t-o's in ms */ 14361 BBR_OPTS_INC(tcp_rack_min_to); 14362 bbr->r_ctl.rc_min_to = optval; 14363 break; 14364 case TCP_RACK_REORD_THRESH: 14365 /* RACK reorder threshold (shift amount) */ 14366 BBR_OPTS_INC(tcp_rack_reord_thresh); 14367 if ((optval > 0) && (optval < 31)) 14368 bbr->r_ctl.rc_reorder_shift = optval; 14369 else 14370 error = EINVAL; 14371 break; 14372 case TCP_RACK_REORD_FADE: 14373 /* Does reordering fade after ms time */ 14374 BBR_OPTS_INC(tcp_rack_reord_fade); 14375 bbr->r_ctl.rc_reorder_fade = optval; 14376 break; 14377 case TCP_RACK_TLP_THRESH: 14378 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14379 BBR_OPTS_INC(tcp_rack_tlp_thresh); 14380 if (optval) 14381 bbr->rc_tlp_threshold = optval; 14382 else 14383 error = EINVAL; 14384 break; 14385 case TCP_BBR_USE_RACK_CHEAT: 14386 BBR_OPTS_INC(tcp_use_rackcheat); 14387 if (bbr->rc_use_google) { 14388 error = EINVAL; 14389 break; 14390 } 14391 BBR_OPTS_INC(tcp_rack_cheat); 14392 if (optval) 14393 bbr->bbr_use_rack_cheat = 1; 14394 else 14395 bbr->bbr_use_rack_cheat = 0; 14396 break; 14397 case TCP_BBR_FLOOR_MIN_TSO: 14398 BBR_OPTS_INC(tcp_utter_max_tso); 14399 if ((optval >= 0) && (optval < 40)) 14400 bbr->r_ctl.bbr_hptsi_segments_floor = optval; 14401 else 14402 error = EINVAL; 14403 break; 14404 case TCP_BBR_UTTER_MAX_TSO: 14405 BBR_OPTS_INC(tcp_utter_max_tso); 14406 if ((optval >= 0) && (optval < 0xffff)) 14407 bbr->r_ctl.bbr_utter_max = optval; 14408 else 14409 error = EINVAL; 14410 break; 14411 14412 case TCP_BBR_EXTRA_STATE: 14413 BBR_OPTS_INC(tcp_extra_state); 14414 if (optval) 14415 bbr->rc_use_idle_restart = 1; 14416 else 14417 bbr->rc_use_idle_restart = 0; 14418 break; 14419 case TCP_BBR_SEND_IWND_IN_TSO: 14420 BBR_OPTS_INC(tcp_iwnd_tso); 14421 if (optval) { 14422 bbr->bbr_init_win_cheat = 1; 14423 if (bbr->rc_past_init_win == 0) { 14424 uint32_t cts; 14425 cts = tcp_get_usecs(&bbr->rc_tv); 14426 tcp_bbr_tso_size_check(bbr, cts); 14427 } 14428 } else 14429 bbr->bbr_init_win_cheat = 0; 14430 break; 14431 case TCP_BBR_HDWR_PACE: 14432 BBR_OPTS_INC(tcp_hdwr_pacing); 14433 if (optval){ 14434 bbr->bbr_hdw_pace_ena = 1; 14435 bbr->bbr_attempt_hdwr_pace = 0; 14436 } else { 14437 bbr->bbr_hdw_pace_ena = 0; 14438 #ifdef RATELIMIT 14439 if (bbr->r_ctl.crte != NULL) { 14440 tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); 14441 bbr->r_ctl.crte = NULL; 14442 } 14443 #endif 14444 } 14445 break; 14446 14447 case TCP_DELACK: 14448 BBR_OPTS_INC(tcp_delack); 14449 if (optval < 100) { 14450 if (optval == 0) /* off */ 14451 tp->t_delayed_ack = 0; 14452 else if (optval == 1) /* on which is 2 */ 14453 tp->t_delayed_ack = 2; 14454 else /* higher than 2 and less than 100 */ 14455 tp->t_delayed_ack = optval; 14456 if (tp->t_flags & TF_DELACK) { 14457 tp->t_flags &= ~TF_DELACK; 14458 tp->t_flags |= TF_ACKNOW; 14459 NET_EPOCH_ENTER(et); 14460 bbr_output(tp); 14461 NET_EPOCH_EXIT(et); 14462 } 14463 } else 14464 error = EINVAL; 14465 break; 14466 case TCP_RACK_PKT_DELAY: 14467 /* RACK added ms i.e. rack-rtt + reord + N */ 14468 BBR_OPTS_INC(tcp_rack_pkt_delay); 14469 bbr->r_ctl.rc_pkt_delay = optval; 14470 break; 14471 14472 case TCP_BBR_RETRAN_WTSO: 14473 BBR_OPTS_INC(tcp_retran_wtso); 14474 if (optval) 14475 bbr->rc_resends_use_tso = 1; 14476 else 14477 bbr->rc_resends_use_tso = 0; 14478 break; 14479 case TCP_DATA_AFTER_CLOSE: 14480 BBR_OPTS_INC(tcp_data_ac); 14481 if (optval) 14482 bbr->rc_allow_data_af_clo = 1; 14483 else 14484 bbr->rc_allow_data_af_clo = 0; 14485 break; 14486 case TCP_BBR_POLICER_DETECT: 14487 BBR_OPTS_INC(tcp_policer_det); 14488 if (bbr->rc_use_google == 0) 14489 error = EINVAL; 14490 else if (optval) 14491 bbr->r_use_policer = 1; 14492 else 14493 bbr->r_use_policer = 0; 14494 break; 14495 14496 case TCP_BBR_TSTMP_RAISES: 14497 BBR_OPTS_INC(tcp_ts_raises); 14498 if (optval) 14499 bbr->ts_can_raise = 1; 14500 else 14501 bbr->ts_can_raise = 0; 14502 break; 14503 case TCP_BBR_TMR_PACE_OH: 14504 BBR_OPTS_INC(tcp_pacing_oh_tmr); 14505 if (bbr->rc_use_google) { 14506 error = EINVAL; 14507 } else { 14508 if (optval) 14509 bbr->r_ctl.rc_incr_tmrs = 1; 14510 else 14511 bbr->r_ctl.rc_incr_tmrs = 0; 14512 } 14513 break; 14514 case TCP_BBR_PACE_OH: 14515 BBR_OPTS_INC(tcp_pacing_oh); 14516 if (bbr->rc_use_google) { 14517 error = EINVAL; 14518 } else { 14519 if (optval > (BBR_INCL_TCP_OH| 14520 BBR_INCL_IP_OH| 14521 BBR_INCL_ENET_OH)) { 14522 error = EINVAL; 14523 break; 14524 } 14525 if (optval & BBR_INCL_TCP_OH) 14526 bbr->r_ctl.rc_inc_tcp_oh = 1; 14527 else 14528 bbr->r_ctl.rc_inc_tcp_oh = 0; 14529 if (optval & BBR_INCL_IP_OH) 14530 bbr->r_ctl.rc_inc_ip_oh = 1; 14531 else 14532 bbr->r_ctl.rc_inc_ip_oh = 0; 14533 if (optval & BBR_INCL_ENET_OH) 14534 bbr->r_ctl.rc_inc_enet_oh = 1; 14535 else 14536 bbr->r_ctl.rc_inc_enet_oh = 0; 14537 } 14538 break; 14539 default: 14540 return (tcp_default_ctloutput(tp, sopt)); 14541 break; 14542 } 14543 tcp_log_socket_option(tp, sopt->sopt_name, optval, error); 14544 INP_WUNLOCK(inp); 14545 return (error); 14546 } 14547 14548 /* 14549 * return 0 on success, error-num on failure 14550 */ 14551 static int 14552 bbr_get_sockopt(struct tcpcb *tp, struct sockopt *sopt) 14553 { 14554 struct inpcb *inp = tptoinpcb(tp); 14555 struct tcp_bbr *bbr; 14556 int32_t error, optval; 14557 14558 bbr = (struct tcp_bbr *)tp->t_fb_ptr; 14559 if (bbr == NULL) { 14560 INP_WUNLOCK(inp); 14561 return (EINVAL); 14562 } 14563 /* 14564 * Because all our options are either boolean or an int, we can just 14565 * pull everything into optval and then unlock and copy. If we ever 14566 * add a option that is not a int, then this will have quite an 14567 * impact to this routine. 14568 */ 14569 switch (sopt->sopt_name) { 14570 case TCP_BBR_PACE_PER_SEC: 14571 optval = bbr->r_ctl.bbr_hptsi_per_second; 14572 break; 14573 case TCP_BBR_PACE_DEL_TAR: 14574 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; 14575 break; 14576 case TCP_BBR_PACE_SEG_MAX: 14577 optval = bbr->r_ctl.bbr_hptsi_segments_max; 14578 break; 14579 case TCP_BBR_MIN_TOPACEOUT: 14580 optval = bbr->no_pacing_until; 14581 break; 14582 case TCP_BBR_PACE_SEG_MIN: 14583 optval = bbr->r_ctl.bbr_hptsi_bytes_min; 14584 break; 14585 case TCP_BBR_PACE_CROSS: 14586 optval = bbr->r_ctl.bbr_cross_over; 14587 break; 14588 case TCP_BBR_ALGORITHM: 14589 optval = bbr->rc_use_google; 14590 break; 14591 case TCP_BBR_TSLIMITS: 14592 optval = bbr->rc_use_ts_limit; 14593 break; 14594 case TCP_BBR_IWINTSO: 14595 optval = bbr->rc_init_win; 14596 break; 14597 case TCP_BBR_STARTUP_PG: 14598 optval = bbr->r_ctl.rc_startup_pg; 14599 break; 14600 case TCP_BBR_DRAIN_PG: 14601 optval = bbr->r_ctl.rc_drain_pg; 14602 break; 14603 case TCP_BBR_PROBE_RTT_INT: 14604 optval = bbr->r_ctl.rc_probertt_int; 14605 break; 14606 case TCP_BBR_PROBE_RTT_LEN: 14607 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); 14608 break; 14609 case TCP_BBR_PROBE_RTT_GAIN: 14610 optval = bbr->r_ctl.bbr_rttprobe_gain_val; 14611 break; 14612 case TCP_BBR_STARTUP_LOSS_EXIT: 14613 optval = bbr->rc_loss_exit; 14614 break; 14615 case TCP_BBR_USEDEL_RATE: 14616 error = EINVAL; 14617 break; 14618 case TCP_BBR_MIN_RTO: 14619 optval = bbr->r_ctl.rc_min_rto_ms; 14620 break; 14621 case TCP_BBR_MAX_RTO: 14622 optval = bbr->rc_max_rto_sec; 14623 break; 14624 case TCP_RACK_PACE_MAX_SEG: 14625 /* Max segments in a pace */ 14626 optval = bbr->r_ctl.rc_pace_max_segs; 14627 break; 14628 case TCP_RACK_MIN_TO: 14629 /* Minimum time between rack t-o's in ms */ 14630 optval = bbr->r_ctl.rc_min_to; 14631 break; 14632 case TCP_RACK_REORD_THRESH: 14633 /* RACK reorder threshold (shift amount) */ 14634 optval = bbr->r_ctl.rc_reorder_shift; 14635 break; 14636 case TCP_RACK_REORD_FADE: 14637 /* Does reordering fade after ms time */ 14638 optval = bbr->r_ctl.rc_reorder_fade; 14639 break; 14640 case TCP_BBR_USE_RACK_CHEAT: 14641 /* Do we use the rack cheat for rxt */ 14642 optval = bbr->bbr_use_rack_cheat; 14643 break; 14644 case TCP_BBR_FLOOR_MIN_TSO: 14645 optval = bbr->r_ctl.bbr_hptsi_segments_floor; 14646 break; 14647 case TCP_BBR_UTTER_MAX_TSO: 14648 optval = bbr->r_ctl.bbr_utter_max; 14649 break; 14650 case TCP_BBR_SEND_IWND_IN_TSO: 14651 /* Do we send TSO size segments initially */ 14652 optval = bbr->bbr_init_win_cheat; 14653 break; 14654 case TCP_BBR_EXTRA_STATE: 14655 optval = bbr->rc_use_idle_restart; 14656 break; 14657 case TCP_RACK_TLP_THRESH: 14658 /* RACK TLP theshold i.e. srtt+(srtt/N) */ 14659 optval = bbr->rc_tlp_threshold; 14660 break; 14661 case TCP_RACK_PKT_DELAY: 14662 /* RACK added ms i.e. rack-rtt + reord + N */ 14663 optval = bbr->r_ctl.rc_pkt_delay; 14664 break; 14665 case TCP_BBR_RETRAN_WTSO: 14666 optval = bbr->rc_resends_use_tso; 14667 break; 14668 case TCP_DATA_AFTER_CLOSE: 14669 optval = bbr->rc_allow_data_af_clo; 14670 break; 14671 case TCP_DELACK: 14672 optval = tp->t_delayed_ack; 14673 break; 14674 case TCP_BBR_HDWR_PACE: 14675 optval = bbr->bbr_hdw_pace_ena; 14676 break; 14677 case TCP_BBR_POLICER_DETECT: 14678 optval = bbr->r_use_policer; 14679 break; 14680 case TCP_BBR_TSTMP_RAISES: 14681 optval = bbr->ts_can_raise; 14682 break; 14683 case TCP_BBR_TMR_PACE_OH: 14684 optval = bbr->r_ctl.rc_incr_tmrs; 14685 break; 14686 case TCP_BBR_PACE_OH: 14687 optval = 0; 14688 if (bbr->r_ctl.rc_inc_tcp_oh) 14689 optval |= BBR_INCL_TCP_OH; 14690 if (bbr->r_ctl.rc_inc_ip_oh) 14691 optval |= BBR_INCL_IP_OH; 14692 if (bbr->r_ctl.rc_inc_enet_oh) 14693 optval |= BBR_INCL_ENET_OH; 14694 break; 14695 default: 14696 return (tcp_default_ctloutput(tp, sopt)); 14697 break; 14698 } 14699 INP_WUNLOCK(inp); 14700 error = sooptcopyout(sopt, &optval, sizeof optval); 14701 return (error); 14702 } 14703 14704 /* 14705 * return 0 on success, error-num on failure 14706 */ 14707 static int 14708 bbr_ctloutput(struct tcpcb *tp, struct sockopt *sopt) 14709 { 14710 if (sopt->sopt_dir == SOPT_SET) { 14711 return (bbr_set_sockopt(tp, sopt)); 14712 } else if (sopt->sopt_dir == SOPT_GET) { 14713 return (bbr_get_sockopt(tp, sopt)); 14714 } else { 14715 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); 14716 } 14717 } 14718 14719 static const char *bbr_stack_names[] = { 14720 __XSTRING(STACKNAME), 14721 #ifdef STACKALIAS 14722 __XSTRING(STACKALIAS), 14723 #endif 14724 }; 14725 14726 static bool bbr_mod_inited = false; 14727 14728 static int 14729 tcp_addbbr(module_t mod, int32_t type, void *data) 14730 { 14731 int32_t err = 0; 14732 int num_stacks; 14733 14734 switch (type) { 14735 case MOD_LOAD: 14736 printf("Attempting to load " __XSTRING(MODNAME) "\n"); 14737 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", 14738 sizeof(struct bbr_sendmap), 14739 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 14740 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", 14741 sizeof(struct tcp_bbr), 14742 NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); 14743 sysctl_ctx_init(&bbr_sysctl_ctx); 14744 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, 14745 SYSCTL_STATIC_CHILDREN(_net_inet_tcp), 14746 OID_AUTO, 14747 #ifdef STACKALIAS 14748 __XSTRING(STACKALIAS), 14749 #else 14750 __XSTRING(STACKNAME), 14751 #endif 14752 CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 14753 ""); 14754 if (bbr_sysctl_root == NULL) { 14755 printf("Failed to add sysctl node\n"); 14756 err = EFAULT; 14757 goto free_uma; 14758 } 14759 bbr_init_sysctls(); 14760 num_stacks = nitems(bbr_stack_names); 14761 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, 14762 bbr_stack_names, &num_stacks); 14763 if (err) { 14764 printf("Failed to register %s stack name for " 14765 "%s module\n", bbr_stack_names[num_stacks], 14766 __XSTRING(MODNAME)); 14767 sysctl_ctx_free(&bbr_sysctl_ctx); 14768 free_uma: 14769 uma_zdestroy(bbr_zone); 14770 uma_zdestroy(bbr_pcb_zone); 14771 bbr_counter_destroy(); 14772 printf("Failed to register " __XSTRING(MODNAME) 14773 " module err:%d\n", err); 14774 return (err); 14775 } 14776 tcp_lro_reg_mbufq(); 14777 bbr_mod_inited = true; 14778 printf(__XSTRING(MODNAME) " is now available\n"); 14779 break; 14780 case MOD_QUIESCE: 14781 err = deregister_tcp_functions(&__tcp_bbr, true, false); 14782 break; 14783 case MOD_UNLOAD: 14784 err = deregister_tcp_functions(&__tcp_bbr, false, true); 14785 if (err == EBUSY) 14786 break; 14787 if (bbr_mod_inited) { 14788 uma_zdestroy(bbr_zone); 14789 uma_zdestroy(bbr_pcb_zone); 14790 sysctl_ctx_free(&bbr_sysctl_ctx); 14791 bbr_counter_destroy(); 14792 printf(__XSTRING(MODNAME) 14793 " is now no longer available\n"); 14794 bbr_mod_inited = false; 14795 } 14796 tcp_lro_dereg_mbufq(); 14797 err = 0; 14798 break; 14799 default: 14800 return (EOPNOTSUPP); 14801 } 14802 return (err); 14803 } 14804 14805 static moduledata_t tcp_bbr = { 14806 .name = __XSTRING(MODNAME), 14807 .evhand = tcp_addbbr, 14808 .priv = 0 14809 }; 14810 14811 MODULE_VERSION(MODNAME, 1); 14812 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); 14813 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); 14814